hardware-init-review/hardware_init_review.tex

% -*- coding: utf-8 -*-
% Copyright (C) 2024 Adrien 'neox' Bourmault <neox@gnu.org>
%
% Permission is granted to copy, distribute and/or modify this document
% under the terms of the GNU Free Documentation License, Version 1.3
% or any later version published by the Free Software Foundation;
% with no Invariant Sections, no Front-Cover Texts, and no Back-Cover Texts.
% A copy of the license is included in the section entitled "GNU
% Free Documentation License".

\input{packages.tex}

\title{Hardware initialization of modern computers}
\author{Adrien 'neox' Bourmault}
\date{\today}

% setup bibliography
\addbibresource{bibliographie.bib}

% ------------------------------------------------------------------------------
\begin{document}{
% ------------------------------------------------------------------------------

\sloppy % allow flexible margins
\input{titlepage.tex} % import titlepage
\newpage

% ------------------------------------------------------------------------------
% License header
% ------------------------------------------------------------------------------

\setcounter{page}{2}
\vspace*{\fill} % fill the page so that text is at the bottom

    This is Edition 1.0. \\

    Copyright \copyright\ 2024 Adrien 'neox' Bourmault
    \href{mailto:neox@gnu.org}{<neox@gnu.org>} \\

    Permission is granted to copy, distribute and/or modify this document
    under the terms of the GNU Free Documentation License, Version 1.3
    or any later version published by the Free Software Foundation;
    with the Invariant Sections being "GNU General Public License version 2"
    and "GNU Free Documentation License", with no Front-Cover Texts, and no
    Back-Cover Texts.
    A copy of the license is included in the section entitled "GNU
    Free Documentation License". \\

    Source-code included in this document is licensed under the GNU General
    Public License version 2 or any later version published by the
    Free Software Foundation.
    A copy of the license is included in the section entitled "GNU
    General Public License version 2". \\

\newpage

% ------------------------------------------------------------------------------
% ACKNOWLEDGMENTS
% ------------------------------------------------------------------------------
\chapter*{Acknowledgments}
\addcontentsline{toc}{chapter}{Acknowledgments}

    First and foremost, I would like to express my deep gratitude to
    Marie-Minerve Louërat, without whom
    this work would not have come to fruition. Her efforts to assist me on
    legal matters and our enriching discussions on the philosophy of free
    software have been invaluable to me. \\

    I am also thankful to Roselyne Chotin for agreeing to fund this work as
    without her support, none of this would have been possible, and
    Franck Wajsburt for his invaluable advice at the beginning of my work,
    which greatly helped me organize myself, as well as for his support and
    reviews throughout this period. \\

    I wish to express my appreciation to the Free Software Foundation for
    funding the necessary equipment for this project, especially to
    the board members of the foundation, including Richard M. Stallman and
    Odile Bénassy, for making the funding for this project possible.
    Special thanks go to Zoë Kooyman and Ian Kelling for their exceptional
    support in managing the hardware order and for their kindness and assistance
    throughout the entire process. \\

    I am deeply grateful to Denis Carikli (GNUtoo), my fellow GNU Boot
    co-maintainer, for his meticulous reviews, emotional support, and brilliant
    ideas that enriched this work, and to Richard M. Stallman for his advice and
    support throughout this journey. \\

    A big thank you to Manuel Bouyer for his infinite patience with all my
    requests regarding network, software, and hardware configurations. \\

    Also, I warmly thank my family and friends for their constant
    encouragement and for reviewing my work throughout this entire process. \\

    And finally, I would like to thank the break room and its kettle,
    without which no tea would have been possible, thereby jeopardizing this
    work. \\

\newpage

% ------------------------------------------------------------------------------
% ABSTRACT
% ------------------------------------------------------------------------------
\chapter*{Abstract}
\addcontentsline{toc}{chapter}{Abstract}

    The global trend is towards the scarcity of free software-compatible
    hardware, and soon there will be no computer that will work without
    software domination by big companies, especially involving firmware like
    BIOSes. \\

    A Basic Input Output System (BIOS) was originally a set of low-level
    functions contained in the read-only memory of a computer's mainboard,
    enabling it to perform basic operations when powered up. However, the
    definition of a BIOS has evolved to include what used to be known as Power
    On Self Test (POST) for the presence of peripherals, allocating resources
    for them to avoid conflicts, and then handing over to an operating system
    boot loader. Nowadays, the bulk of the BIOS work is the initialization
    and training of RAM. This means, for example, initializing the memory
    controller and optimizing timing and read/write voltage for optimal
    performance, making the code complex, as its role is to optimize several
    parallel buses operating at high speeds and shared by many CPU cores,
    and make them act as a homogeneous whole. \\

    This document is the product of a project hosted by the \textit{LIP6
    laboratory} and supported by the \textit{GNU Boot Project} and the
    \textit{Free Software Foundation}. It delves into the importance
    of firmware in the hardware initialization of modern computers and
    explores various aspects of firmware, such as Intel Management Engine
    (ME), AMD Platform Security Processor (PSP), Advanced Configuration and
    Power Interface (ACPI), and System Management Mode (SMM). Additionally,
    it provides an in-depth look at memory initialization and training
    algorithms, highlighting their critical role in system stability and
    performance. Examples of the implementation in the ASUS KGPE-D16 mainboard
    are presented, describing its hardware characteristics, topology, and the
    crucial role of firmware in its operation after the mainboard architecture
    is examined. Practical examples illustrate the impact of firmware on
    hardware initialization, memory optimization, resource allocation,
    power management, and security. Specific algorithms used for memory
    training and their outcomes are analyzed to demonstrate the complexity
    and importance of firmware in achieving optimal system performance.
    Furthermore, this document explores the relationship between firmware
    and hardware virtualization. Security considerations and future trends
    in firmware development are also addressed, emphasizing the need for
    continued research and advocacy for free software-compatible hardware.

\newpage

% ------------------------------------------------------------------------------
% Table of contents
% ------------------------------------------------------------------------------
\chapter*{\vspace{-\cftbeforechapskip}}
\addcontentsline{toc}{chapter}{Contents}
\tableofcontents
\newpage

% List of figures
\chapter*{\vspace{-\cftbeforechapskip}}
\addcontentsline{toc}{chapter}{List of Figures}
\listoffigures
\newpage

% List of listings
%\chapter*{\vspace{-3em}}
\listoflistings
\addcontentsline{toc}{chapter}{List of Listings}
\newpage

% ------------------------------------------------------------------------------
% CHAPTER 1: Introduction to firmware and BIOS evolution
% ------------------------------------------------------------------------------
\chapter{Introduction to firmware and BIOS evolution}

    \section{Historical context of BIOS}

        \subsection{Definition and origin}

        The BIOS (Basic Input/Output System) is firmware, which is a type of
        software that is embedded into hardware devices to control their basic
        functions, acting as a bridge between hardware and other software,
        ensuring that the hardware operates correctly. Unlike regular
        software, firmware is usually stored in a non-volatile memory like
        ROM or flash memory. The term "firmware" comes from its role: it is
        "firm" because it's more permanent than regular software (which can
        be easily changed) but not as rigid as hardware. \\

        The BIOS is used to perform initialization during the booting process
        and to provide runtime services for operating systems and programs.
        Being a critical component for the startup of personal computers,
        acting as an intermediary between the computer's hardware and its
        operating system, the BIOS is embedded on a chip on the motherboard
        and is the first code that runs when a PC is powered on. The concept
        of BIOS has its roots in the early days of personal computing. It
        was first developed by IBM for their IBM PC, which was introduced
        in 1981 \cite{freiberger2000fire}. The term BIOS itself was
        coined by Gary Kildall, who developed the CP/M (Control Program for
        Microcomputers) operating system \cite{shustek2016kildall}. In CP/M,
        BIOS was used to describe a component that interfaced directly
        with the hardware, allowing the operating system to be somewhat
        hardware-independent. \\

        \begin{figure}[H]
            \centering
            \includegraphics[width=0.5\textwidth]{images/IBM_logo.png}
            \caption{The eight-striped wordmark of IBM (1967, public domain,
                     trademarked)}
        \end{figure}

        IBM's implementation of BIOS became a de facto standard in
        the industry, as it was part of the IBM PC's open architecture
        \cite{grewal_ibm_pc}\cite{ibm_pc}, which refers to the design
        philosophy adopted by IBM when developing the IBM Personal Computer
        (PC), introduced in 1981. This architecture is characterized by the use
        of off-the-shelf components and publicly available specifications,
        which allowed other manufacturers to create compatible hardware
        and software. It was in fact a departure from the proprietary
        systems prevalent at the time, where companies closely guarded their
        designs to maintain control over the hardware and software ecosystem.
        For example, IBM used the Intel 8088 CPU, a well-documented and widely
        available processor, and also the Industry Standard Architecture
        (ISA) bus, which defined how various components like memory, storage,
        and peripherals communicated with the CPU. This open architecture
        allowed other manufacturers to create IBM-compatible computers, also
        known as "clones", which further popularized the BIOS concept. As
        a result, the IBM PC BIOS set the stage for a standardized method
        of interacting with computer hardware, which has evolved over the
        years but remains fundamentally the same in principle. IBM also
        published detailed technical documentation at that time, including
        circuit diagrams, BIOS listings, and interface specifications. This
        transparency allowed other companies to understand and replicate
        the IBM PC's functionality \cite{freiberger2000fire}.

        \subsection{Functionalities and limitations}

        When a computer is powered on, the BIOS executes a Power-On
        Self-Test (POST), a diagnostic sequence that verifies the integrity
        and functionality of critical hardware components such as the CPU,
        RAM, disk drives, keyboard, and other peripherals \cite{wiki_bios}.
        This process ensures that all essential hardware components are
        operational before the system attempts to load the operating system.
        If any issues are detected, the BIOS generates error messages or
        beep codes to alert the user. Following the successful completion
        of POST, the BIOS runs the bootstrap loader, a small program that
        identifies the operating system's bootloader on a storage device,
        such as a hard drive, floppy disk, or optical drive. The bootstrap
        loader then transfers control to the OS bootloader, initiating
        the process of loading the operating system into the computer's
        memory and starting it. This step effectively bridges the gap
        between hardware initialization and operating system execution.
        The BIOS also provides a set of low-level software routines known
        as interrupts. These routines enable software to perform basic
        input/output operations, such as reading from the keyboard, writing
        to the display, and accessing disk drives, without needing to manage
        the hardware directly. By providing standardized interfaces for
        hardware components, the BIOS simplifies software development and
        improves compatibility across different hardware configurations
        \cite{ibm_pc}. \\

        \begin{figure}[H]
            \centering
            \includegraphics[width=0.5\textwidth]{images/bios_chip.jpg}
            \caption{An AMI BIOS chip from a Dell 310, by Jud McCranie
                     (CC BY-SA 4.0, 2018)}
        \end{figure}

        Despite its essential role, the early BIOS had several limitations.
        One significant limitation was its limited storage capacity.
        Early BIOS firmware was stored in Read-Only Memory (ROM) chips with
        very limited storage, often just a few kilobytes. This constrained
        the complexity and functionality of the BIOS, limiting it to only the
        most essential tasks needed to start the system and provide basic
        hardware control. The original BIOS was also non-extensible. ROM
        chips were typically soldered onto the motherboard, making updates
        difficult and costly. Bug fixes, updates for new hardware support,
        or enhancements required replacing the ROM chip, leading to challenges
        in maintaining and upgrading systems. Furthermore, the early BIOS was
        tailored for the specific hardware configurations of the initial IBM
        PC models, which included a limited set of peripherals and expansion
        options. As new hardware components and peripherals were developed,
        the BIOS often needed to be updated to support them, which was not
        always feasible or timely. Performance bottlenecks were another
        limitation. The BIOS provided basic input/output operations that
        were often slower than direct hardware access methods. For example,
        disk I/O operations through BIOS interrupts were slower compared
        to later direct access methods provided by operating systems,
        resulting in performance bottlenecks, especially for disk-intensive
        operations \cite{osdev_uefi}. Early BIOS
        implementations also had minimal security features. There were no
        mechanisms to verify the integrity of the BIOS code or to protect
        against unauthorized modifications, leaving systems vulnerable to
        attacks that could alter the BIOS and potentially compromise the
        entire system, such as rootkits and firmware viruses. Added to that,
        the traditional BIOS operates in 16-bit real mode, a constraint that
        limits the amount of code and memory it can address. This limitation
        hinders the performance and complexity of firmware, making it less
        suitable for modern computing needs \cite{intel_uefi}. Additionally,
        BIOS relies on the Master Boot Record (MBR) partitioning scheme,
        which supports a maximum disk size of 2 terabytes and allows only
        four primary partitions \cite{uefi_spec}\cite{russinovich2012}.
        This constraint has become a significant drawback as storage
        capacities have increased. Furthermore, the traditional BIOS has
        limited flexibility and is challenging to update or extend. This
        inflexibility restricts the ability to support new hardware and
        technologies efficiently \cite{osdev_uefi}\cite{acmcs2015}.

    \section{Modern BIOS and UEFI}

        \subsection{Transition from traditional BIOS to UEFI (Unified
        Extensible Firmware Interface)}

        All the limitations listed earlier caused a transition to a more
        modern firmware interface, designed to address the shortcomings of
        the traditional BIOS. This section delves into the historical context
        of this shift, the driving factors behind it, and the advantages
        UEFI offers over the traditional BIOS. \\

        The development of UEFI began in the mid-1990s as part of the
        Intel Boot Initiative, which aimed to modernize the boot process
        and overcome the limitations of the traditional BIOS. By 2005, the
        Unified EFI Forum, a consortium of technology companies including
        Intel, AMD, and Microsoft, had formalized the UEFI specification
        \cite{uefi_spec}. UEFI was designed to address the shortcomings of
        the traditional BIOS, providing several key improvements.

        \begin{figure}[H]
            \centering
            \includegraphics[width=0.25\textwidth]{images/uefi_logo.png}
            \caption{The UEFI logo (public domain, 2010)}
        \end{figure}

        One of the most significant advancements of UEFI is its support for
        32-bit and 64-bit modes, allowing it to address more memory and
        run more complex firmware programs. This capability enables UEFI
        to handle the increased demands of modern hardware and software
        \cite{intel_uefi}\cite{shin2011}. Additionally, UEFI uses the GUID
        Partition Table (GPT) instead of the MBR, supporting disks larger
        than 2 terabytes and allowing for a nearly unlimited number of
        partitions \cite{microsoft_uefi}\cite{russinovich2012}.

        Improved boot performance is another driving factor. UEFI
        provides faster boot times compared to the traditional BIOS,
        thanks to its efficient hardware and software initialization
        processes. This improvement is particularly beneficial for systems
        with complex hardware configurations, where quick boot times
        are essential  \cite{intel_uefi}. UEFI's modular architecture
        makes it more extensible and easier to update compared to the
        traditional BIOS. This design allows for the addition of drivers,
        applications, and other components without requiring a complete
        firmware overhaul, providing greater flexibility and adaptability
        to new technologies  \cite{acmcs2015}. UEFI also includes enhanced
        security features such as \textit{Secure Boot}, which ensures that
        only trusted software can be executed during the boot process,
        thereby protecting the system from unauthorized modifications and
        malware \cite{osdev_uefi}\cite{chang2013}. \\

        The industry-wide support and standardization of UEFI have accelerated
        its adoption across various platforms and devices. Major industry
        players, including Intel, AMD, and Microsoft, have adopted UEFI as
        the new standard for firmware interfaces, ensuring broad compatibility
        and interoperability \cite{uefi_spec}.

        \subsection{An other way with \textit{coreboot}}

        While UEFI has become the dominant firmware interface for modern
        computing systems, it is not without its critics. Some of the primary
        concerns about UEFI include its complexity, potential security
        vulnerabilities, and the degree of control it provides to hardware
        manufacturers over the boot process. Originally known as LinuxBIOS,
        \textit{coreboot}, is a free firmware project initiated in 1999 by
        Ron Minnich and his team at the Los Alamos National Laboratory. The
        project's primary goal was to create a fast, lightweight, and
        flexible firmware solution that could initialize hardware and
        boot operating systems quickly, while remaining transparent and
        auditable\cite{coreboot}. As an alternative to UEFI, \textit{coreboot}
        offers a different approach to firmware that aims to address some
        of these concerns and continue the evolution of BIOS.\\

        One of the main advantages of \textit{coreboot} over UEFI is its
        simplicity, as it is designed to perform only the minimal tasks
        required to initialize hardware and pass control to a payload, such
        as a bootloader or operating system kernel. This minimalist approach
        reduces the attack surface and potential for security vulnerabilities,
        as there is less code that could be exploited by malicious actors.
        Another significant benefit of \textit{coreboot}
        is its libre nature. Unlike UEFI, which is controlled by a consortium
        of hardware and software vendors, \textit{coreboot}'s source code
        is freely available and can be audited, modified, and improved by
        anyone. This transparency ensures that security researchers and
        developers can review the code for potential vulnerabilities and
        contribute to its improvement, fostering a community-driven approach
        to firmware development\cite{coreboot}. This project also supports
        a wide range of bootloaders, called payloads, allowing users to
        customize their boot process to suit their specific needs. Popular
        payloads include SeaBIOS, which provides legacy BIOS compatibility, and
        Tianocore, which offers UEFI functionality within the \textit{coreboot}
        framework. This flexibility allows \textit{coreboot} to be used in
        a variety of environments, from embedded systems to high-performance
        servers \cite{coreboot_payloads}. \\

        \begin{figure}[H]
            \centering
            \includegraphics[width=0.3\textwidth]{images/coreboot_logo.png}
            \caption{The \textit{coreboot} logo, by Konsult Stuge \&
            coresystems
                     (coreboot logo license, 2008)}
        \end{figure}

        Despite its advantages, \textit{coreboot} is not without its
        challenges. The project relies heavily on community contributions, and
        support for new hardware often lags behind that of UEFI. Additionally,
        the minimalist design of \textit{coreboot} means that some advanced
        features provided by UEFI are not available by default. However,
        the \textit{coreboot} community continues to work on adding
        new features and improving compatibility with modern hardware or
        security issues \cite{coreboot_challenges}. For example, it provides
        a \textit{verified boot} function, allowing to prevent rootkits and
        other attacks based on firmware modifications \cite{coreboot_docs}.
        However, it's important to note that \textit{coreboot} is not entirely
        free in all aspects. Many modern processors and chipsets require
        \textit{proprietary blobs}, short for \textit{Binary Large Object},
        which is a collection of binary data stored as a single entity. These
        blobs are necessary for \textit{coreboot} to function correctly on
        a wide range of hardware, but they compromise the goal of having
        a fully free firmware one day \cite{blobs}, since these blobs are
        used for certain functionalities such as memory initialization and
        hardware management.

        \begin{figure}[H]
            \centering
            \includegraphics[width=0.25\textwidth]{images/gnuboot.png}
            \caption{The \textit{GNU Boot} logo, by Jason Self (CC0, 2020)}
        \end{figure}

        To address these concerns, the GNU Project has developed GNU Boot,
        a fully free distribution of firmware, including \textit{coreboot},
        that aims to be entirely free by avoiding the use of proprietary
        binary blobs.

        GNU Boot is only a distribution: it reuses existing software projects
        and is not very different from fully free GNU/Linux distributions like
        Trisquel or Guix, as GNU Boot is committed to use only free software
        for all aspects of firmware, making it a preferred choice for users
        and organizations that prioritize software freedom and transparency.
        Its goal include to build the software and assemble it in something
        that can be installed, and also to test it and to provide installation
        and upgrade instructions \cite{gnuboot}.

    \section{Shift in firmware responsibilities}

        Initially, the BIOS's primary function was to perform the POST, a basic
        diagnostic testing process to check the system's hardware components
        and ensure they were functioning correctly. This included verifying
        the CPU, memory, and essential peripherals before passing control to
        the operating system's bootloader. This process was relatively simple,
        given the limited capabilities and straightforward architecture of
        early computer systems \cite{osdev_uefi}.

        As computer systems advanced, particularly with the advent of more
        sophisticated memory technologies, the role of firmware expanded
        significantly. Modern memory modules operate at much higher
        speeds and capacities than their predecessors, requiring precise
        configuration to ensure stability and optimal performance. Firmware
        now plays a critical role in managing the memory controller, which is
        responsible for regulating data flow between the processor and memory
        modules. This includes configuring memory frequencies, voltage levels,
        and timing parameters to match the specifications of the installed
        memory \cite{uefi_spec}\cite{BKDG}. Beyond memory management,
        firmware responsibilities have broadened to encompass a wide range
        of system-critical tasks, and even so by including runtime components
        in addition to its initialization tasks.
        One key area is power management, where
        firmware is responsible for optimizing energy consumption across
        various components of the system. Efficient power management is
        essential not only for extending battery life in portable devices
        but also for reducing thermal output and ensuring system longevity
        in desktop and server environments. Moreover, modern firmware takes
        on significant roles in hardware initialization and configuration,
        which were traditionally handled by the operating system. For
        example, the initialization of USB controllers, network interfaces,
        and storage devices is now often managed by the firmware during
        the early stages of the boot process. This shift ensures that the
        operating system can seamlessly interact with hardware from the
        moment it takes control, reducing boot times and improving overall
        system reliability \cite{uefi_spec}. Security has also become a
        paramount concern for modern firmware. UEFI (Unified Extensible
        Firmware Interface), which has largely replaced traditional BIOS
        in modern systems, includes features which prevents unauthorized
        or malicious software from loading during the boot process. This
        helps protect the system from rootkits and other low-level malware
        that could compromise the integrity of the operating system before
        it even starts \cite{uefi_spec}. In the context of performance
        tuning, firmware sometimes also plays a key role in enabling and
        managing overclocking, particularly for the memory subsystem. By
        allowing adjustments to memory frequencies, voltages, and timings,
        firmware provides tools for enthusiasts to push their systems beyond
        default limits. At the same time, it includes safeguards to manage
        the risks of instability and hardware damage, balancing performance
        gains with system reliability \cite{osdev_uefi}. \\

        In summary, the evolution of firmware from simple hardware
        initialization routines to complex management systems reflects the
        increasing sophistication of modern computer architectures. Firmware
        is now a critical layer that not only ensures the correct functioning
        of hardware components but also optimizes performance, manages power
        consumption, and enhances system security, making it an indispensable
        part of contemporary computing. \\

        This document will focus on \textit{coreboot} during the next parts
        to study how modern firmware interact with hardware and also as a
        basis for improvements.

% ------------------------------------------------------------------------------
% CHAPTER 2: Characteristics of ASUS KGPE-D16 mainboard
% ------------------------------------------------------------------------------
\chapter{Characteristics of ASUS KGPE-D16 mainboard}

    \begin{figure}[H]
        \centering \includegraphics[width=0.9\textwidth]{images/kgpe-d16.png}
        \caption{The KGPE-D16 (CC BY-SA 4.0, 2021)}
    \end{figure}

    \newpage

    \section{Overview of ASUS KGPE-D16 hardware}

        The ASUS KGPE-D16 server mainboard is a dual-socket motherboard
        designed to support AMD Family 10h/15h series processors. Released
        in 2009, this mainboard was later awarded the \textit{Respects Your
        Freedom} (RYF) certification in March 2017, underscoring its commitment
        to fully free software compatibility \cite{fsf_ryf}. Indeed, this
        mainboard can be operated with a fully free firmware such as GNU
        Boot \cite{gnuboot_status}. \\

        This mainboard is equipped with robust hardware components designed to
        meet the demands of high-performance computing. It features 16 DDR3
        DIMM slots, capable of supporting up to 256GB of memory, although
        certain configurations may be limited to 192GB, with some reports
        suggesting the potential to support 256GB under specific conditions.
        In terms of expandability, the KGPE-D16 includes multiple PCIe
        slots, with five physical slots available, although only four
        can be used simultaneously due to slot sharing. For storage, the
        mainboard provides several SATA ports. Networking capabilities are
        enhanced by integrated dual gigabit Ethernet ports, which provide
        high-speed connectivity essential for data-intensive tasks and network
        communication \cite{asus_kgpe_d16_manual}. Additionally, the board
        is equipped with various peripheral interfaces, including USB ports,
        audio outputs, and other I/O ports, ensuring compatibility with a
        wide range of external devices. \\

        \begin{figure}[H]
            \centering
            \includegraphics[width=0.8\textwidth]{images/fig1_schema_basique.png}
            \caption{Basic schematics of the ASUS KGPE-D16 Mainboard, ASUS
            (2011)} \label{fig:d16_basic_schematics}
        \end{figure}

        The physical layout of the ASUS KGPE-D16 is meticulously designed
        to optimize airflow, cooling, and power distribution. All of this
        is critical for maintaining system stability, particularly under
        heavy computational loads, as this board was designed for server
        operations. In particular, key components such as the CPU sockets,
        memory slots, and PCIe slots are strategically positioned. \\

        \begin{figure}[H]
            \centering
            \includegraphics[width=0.8\textwidth]{images/kgpe-d16_real.png}
            \caption{The KGPE-D16, viewed from the top (CC BY-SA 4.0, 2024)}
            \label{fig:d16_top_view}
        \end{figure}

    \section{Chipset}

        Before diving into the specific components, it is essential
        to understand the roles of the northbridge and southbridge in
        traditional motherboard architecture. These chipsets historically
        managed communication between the CPU and other critical components
        of the system \cite{amd_chipsets}. \\

        The northbridge is a chipset on the motherboard that traditionally
        manages high-speed communication between the CPU, memory (RAM), and
        graphics card (if applicable). It serves as a hub for data that needs
        to move quickly between these components. On the ASUS KGPE-D16, the
        functions typically associated with the northbridge are divided between
        the CPU’s internal northbridge and an external SR5690 northbridge
        chip. The SR5690 specifically acts as a translator and switch,
        handling the HyperTransport interface, a high-speed communication
        protocol used by AMD processors, and converting it to ALink and PCIe
        interfaces, which are crucial for connecting peripherals like graphics
        cards \cite{SR5690BDG}. Additionally, the northbridge on the KGPE-D16
        incorporates the IOMMU (Input-Output Memory Management Unit), which
        is crucial for ensuring secure and efficient memory access by I/O
        devices. The IOMMU allows for the virtualization of memory addresses,
        providing device isolation and preventing unauthorized memory access,
        which is particularly important in environments that run multiple
        virtual machines \cite{amd_chipsets}\cite{northbridge_wiki}. \\

        The southbridge, on the other hand, is responsible for handling
        lower-speed, peripheral interfaces such as the PCI, USB, and
        IDE/SATA connections, as well as managing onboard audio and
        network controllers. On the KGPE-D16, these functions are managed
        by the SP5100 southbridge chip, which integrates several critical
        functions including the LPC bridge, SATA controllers, and other
        essential I/O operations \cite{amd_chipsets}\cite{southbridge_wiki}.
        It is essentially an ALink bus controller and includes the hardware
        interrupt controller, the IOAPIC. Interrupts from peripheral always
        pass through the northbridge (fig. \ref{fig:d16_ioapic}), since it
        translates ALink to HyperTransport for the CPUs and contains the
        IOMMU \cite{SR5690BDG}. \\

        \begin{figure}[H]
            \centering \includegraphics[width=0.9\textwidth]{images/ioapic.png}
            \caption{Functional diagram presenting the IOAPIC function of
            the SP5100,
                     ASUS (2011)}
            \label{fig:d16_ioapic}
        \end{figure}

        In addition to the northbridge and southbridge, the KGPE-D16 also
        contains specialized chips for managing input/output operations and
        system health monitoring. The WINBOND W83667HG-A Super I/O chip handles
        traditional I/O functions such as legacy serial and parallel ports,
        keyboard, and mouse interfaces, but also the SPI chip (Serial Peripheral
        Interface, a synchronous serial communication protocol primarily used
        to communicate between microcontrollers and peripheral devices like
        sensors or memory devices) that contains the firmware \cite{winbond}.
        Meanwhile, the Nuvoton W83795G/ADG Hardware
        Monitor oversees the system’s health by monitoring temperatures,
        voltages, and fan speeds, ensuring that the system operates within
        safe parameters \cite{nuvoton}. On the KGPE-D16, access to the Super
        I/O from a CPU core is done through the SR5690, then the SP5100,
        as that can be observed on the functional diagram of the chipset
        (fig. \ref{fig:d16_chipset}) \cite{SR5690BDG}.

        \begin{figure}[H]
            \centering
            \includegraphics[width=0.8\textwidth]{images/fig2_diagramme_chipset.png}
            \caption{Functional diagram of the KGPE-D16 chipset (CC BY-SA 4.0,
            2024)} \label{fig:d16_chipset}
        \end{figure}

    \section{Processors}

        The ASUS KGPE-D16 supports AMD Family 10h processors, but
        it is important to note that Vikings, a known vendor for
        libre-software-compatible hardware, does not recommend using the
        Opteron 6100 series due to the lack of IOMMU support, which is
        critical for security. Fortunately, AMD Family 15h processors are also
        supported. However, the Opteron 6300 series, while supported, requires
        proprietary microcode updates for stability, IOMMU functionality,
        and fixes for specific vulnerabilities, including a gain-root-
        via-NMI exploit. The Opteron 6200 series does not suffer from these
        problems and works properly without any proprietary microcode update
        needed \cite{vikings}. \\

        \begin{figure}[H]
            \centering
            \includegraphics[width=0.9\textwidth]{images/opteron6200_annoté.png}
            \caption{Annotated photography of an Opteron 6200 series
            CPU (2024), from a photography by AMD Inc. (2008)}
            \label{fig:opteron2600}
        \end{figure}

        The Opteron 6200 series, part of the Bulldozer microarchitecture,
        was designed to target high-performance server applications. These
        processors feature 16 cores, organized into 8 Bulldozer modules,
        with each module containing two integer cores that shared
        resources like the floating-point unit (FPU) and L2 cache
        (fig. \ref{fig:opteron2600}, \ref{fig:opteron2600_diagram})
        \cite{amd_6200}\cite{anandtech_bulldozer}.

        The architecture of the Opteron 6200 series is built around AMD's
        Bulldozer core design, which uses Clustered Multithreading (CMT) to
        maximize resource utilization. This is a technique where each processor
        module contains two integer cores that share certain resources like
        the floating-point unit (FPU), L2 cache, and instruction fetch/decode
        stages. Unlike traditional multithreading, where each core handles
        multiple threads, CMT allows two cores to share resources to improve
        parallel processing efficiency. This approach aims to balance
        performance and resource usage, particularly in multi- threaded
        workloads, though it can lead to some performance trade-offs in
        single-threaded tasks. In the Opteron 6272, the processor consists
        of eight modules, effectively creating 16 integer cores. Due to
        the CMT architecture, each Opteron 6272 chip functions as two CPUs
        within a single processor, each with its own set of cores, L2 caches,
        and shared L3 cache. Here, one CPU is made by four modules, each
        module in it sharing certain components, such as the FPU and L2 cache,
        between two integer cores. The L3 cache is shared across these modules.
        HyperTransport links provide high-speed communication between the two
        sockets of the KGPE-D16. Shared L3 cache and direct memory access are
        provided by each socket \cite{amd_6200}\cite{hill_impact_caching}. \\

        This architecture also integrates a quad-channel DDR3 memory
        controller directly into the processor die, which facilitates high
        bandwidth and low latency access to memory. This memory controller
        supports DDR3 memory speeds up to 1600 MHz and connects directly
        to the memory modules via the memory bus. By integrating the memory
        controller into the processor, the Opteron 6200 series reduces memory
        access latency, enhancing overall performance
        \cite{amd_6200}\cite{amd_ddr3_guide}.
        It is interesting to note that Opterons
        incorporate the internal northbridge that we cited previously. The
        traditional northbridge functions, such as memory controller and PCIe
        interface management, are partially integrated into the processor. This
        integration reduces the distance data must travel between the CPU and
        memory, decreasing latency and improving performance, particularly
        in memory-intensive applications \cite{amd_6200}. \\


        \begin{figure}[H]
            \centering \includegraphics[width=0.8\textwidth]{
                                        images/fig3_img_dual_processor_node.png}
            \caption{Functional diagram of an Opteron 6200 package
                    (CC BY-SA 4.0, 2024)}
            \label{fig:opteron2600_diagram}
        \end{figure}

        Power efficiency was a key focus in the design of the Opteron 6200
        series. Despite the high core count, the processor includes several
        power management features, such as Dynamic Power Management (DPM)
        and Turbo Core technology. These features allow the processor to
        adjust power usage based on workload demands, balancing performance
        with energy consumption. However, the Bulldozer architecture's
        focus on high clock speeds and multi-threaded performance resulted
        in higher power consumption compared to competing architectures
        \cite{anandtech_bulldozer}. A special model of the series, called
        \textit{high efficiency} models, solve a bit this problem by proposing
        a bit less performant processor but with a power consumption divided
        by a factor from 1.5 to 2.0 in some cases. \\

        The processor connected to the I/O hub is known as the Bootstrap
        Processor (BSP). The BSP is responsible for starting up the system
        by executing the initial firmware code from the reset vector,
        a specific memory address where the CPU begins execution after a
        reset \cite{amd_bsp}. Core 0 of the BSP, called the Bootstrap Core
        (BSC), initiates this process. During early initialization, the
        BSP performs several critical tasks, such as memory initialization,
        and bringing other CPU cores online. One of its duties is storing
        Built-In Self-Test (BIST) information, which involves checking the
        integrity of the processor's internal components to ensure they are
        functioning correctly. The BSP also determines the type of reset
        that has occurred whether it's a cold reset, which happens when
        the system is powered on from an off state, or a warm reset, which
        is a restart without turning off the power. Identifying the reset
        type is crucial for deciding which initialization procedures need
        to be executed \cite{amd_bsp}\cite{BKDG}.

    \section{Baseboard Management Controller}

        The Baseboard Management Controller (BMC) on the KGPE-D16 motherboard,
        specifically the ASpeed AST2050, plays a role in the server's
        architecture by managing out-of-band communication and control of
        the hardware. The AST2050 is based on an ARM926EJ-S processor,
        a low-power 32-bit ARM architecture designed for embedded systems
        \cite{ast2050_architecture}. This architecture is well-suited for BMCs
        due to its efficiency and capability to handle multiple management
        tasks concurrently without significant resource demands from the
        main system. \\

        The AST2050 features several key components that contribute to
        its functionality. It includes an integrated VGA controller,
        which enables remote graphical management through KVM-over-IP
        (Keyboard, Video, Mouse), a critical feature for administrators who
        need to interact with the system remotely, including BIOS updates
        and troubleshooting \cite{ast2050_kvm}. Additionally, the AST2050
        integrates a dedicated memory controller, which supports up to 256MB
        of DDR2 RAM. This allows it to handle complex tasks and maintain
        responsiveness during management operations \cite{ast2050_memory}.
        The BMC also features a network interface controller (NIC) dedicated to
        management traffic, ensuring that remote management does not interfere
        with the primary network traffic of the server. This separation is
        vital for maintaining secure and uninterrupted system management,
        especially in environments where uptime is critical \cite{ast2050_nic}.
        Another important architectural aspect of the AST2050 is its support
        for multiple I/O interfaces, including I2C, GPIO, UART, and USB,
        which allow it to interface with various sensors and peripherals
        on the motherboard \cite{ast2050_io}. This versatility enables
        comprehensive monitoring of hardware health, such as temperature
        sensors, fan speeds, and power supplies, all of which can be managed
        and configured through the BMC. \\

        When combined with OpenBMC \cite{openbmc_wiki}, a libre firmware
        that can be run on the AST2050 thanks to Raptor Engineering
        \cite{raptor_engineering}, the architecture of the BMC becomes even
        more powerful. OpenBMC takes advantage of the AST2050's architecture,
        providing a flexible and customizable environment that can be tailored
        to specific use cases. This includes adding or modifying features
        related to security, logging, and network management, all within
        the BMC's ARM architecture framework \cite{openbmc_customization}.

% ------------------------------------------------------------------------------
% CHAPTER 3: Key components in modern firmware
% ------------------------------------------------------------------------------
\chapter{Key components in modern firmware}

    \section{General structure of coreboot}

        The firmware of the ASUS KGPE-D16 is crucial in ensuring the proper
        functioning and optimization of the mainboard's hardware components.
        In this chapter and for the rest of this document, we're basing our
        study on the 4.11 version of \textit{coreboot} \cite{coreboot_4_11},
        which is the last version that supported the ASUS KGPE-D16 mainboard. \\

        For the initialization tasks to be done efficiently, \textit{coreboot} is
        organized in different stages (fig. \ref{fig:coreboot_stages})
        \cite{coreboot_docs}.

        \begin{figure}[H]
            \centering
            \includegraphics[width=0.9\textwidth]{
                                        images/fig9_coreboot_stages.png}
            \caption{\textit{coreboot}'s stages timeline, by
                     \textit{coreboot} project (CC BY-SA 4.0, 2009)}
            \label{fig:coreboot_stages}
        \end{figure}

        Being a complex project with ambitious goals, \textit{coreboot} decided
        early on to establish an file-system-based architecture for its images
        (also called ROMs). This special file-system is CBFS (which stands for
        coreboot file system). The CBFS architecture consists of a binary image
        that can be interpreted as a physical disk, referred to here as ROM. A
        number of independent components, each with a header added to the data,
        are located within the ROM. The components are nominally arranged
        sequentially, although they are aligned along a predefined boundary
        (fig. \ref{fig:coreboot_diagram}). \\

        Each stage is compiled as a separate binary and inserted into the CBFS
        with custom compression. The bootblock stage is usually not compressed,
        while the ramstage and the payload are compressed with LZMA. Each stage
        loads the next stage at a given address (possibly decompressing it in
        the process). \\

        Some stages are relocatable and can be placed anywhere in the RAM.
        These stages are typically cached in the CBMEM for faster loading times
        during wake-up. The CBMEM is a specific memory area used by the
        \textit{coreboot} firmware to store important data structures and logs
        during the boot process. This area is typically allocated in the
        system's RAM and is used to store various types of runtime information
        that it might need to reference after the initial boot stages. \\

        In general, \textit{coreboot} manages main memory through a structured
        memory map (fig. \ref{tab:memmap}), allocating specific address ranges
        for various hardware functions and system operations. The first 640KB
        of memory space is typically unused by coreboot due to historical
        reasons. Graphics-related operations use the VGA address range
        and the text mode address ranges. It also reserves the higher for
        operating system use, ensuring that critical system components
        like the IOAPIC and TPM registers have dedicated address spaces.
        This structured approach helps maintain system stability and
        compatibility across different platforms and allows for a reset vector
        fixed at an address (\texttt{0xFFFFFFF0}), regardless of the ROM size.
        Payloads are typically loaded into high memory, above the reserved areas
        for hardware components and system resources. The exact memory location
        can vary depending on the system's configuration, but generally,
        payloads are placed in a region of memory that does not conflict with
        the firmware code or the reserved memory map areas, such as the ROM
        mapping ranges. This placement ensures that payloads have sufficient
        space to execute without interfering with other critical memory regions
        allocated \cite{coreboot_mem_management}.

        \begin{table}[ht]
            \makebox[\textwidth][c]{%
                \begin{tabular}{
                                |>{\centering\arraybackslash}p{0.35\textwidth}
                                |>{\centering\arraybackslash}p{0.5\textwidth}|}
                    \hline
                        \path{0x00000 - 0x9FFFF}
                    &   Low memory (first 640KB). Never used. \\
                    \hline
                        \path{0xA0000 - 0xAFFFF}
                    &   VGA graphics address range. \\
                    \hline
                        \path{0xB0000 - 0xB7FFF}
                    &   Monochrome text mode address range.
                        Few motherboards use
                        it, but the KGPE-D16 does. \\
                    \hline
                        \path{0xB8000 - 0xBFFFF}
                    &   Text mode address range. \\
                    \hline
                        \path{0xFEC00000}
                    &   IOAPIC address. \\
                    \hline
                        \path{0xFED44000 - 0xFED4FFFF}
                    &   Address range for TPM registers. \\
                    \hline
                        \path{0xFF000000 - 0xFFFFFFFF}
                    &   16 MB ROM mapping address range. \\
                    \hline
                        \path{0xFF800000 - 0xFFFFFFFF}
                    &   8 MB ROM mapping address range. \\
                    \hline
                        \path{0xFFC00000 - 0xFFFFFFFF}
                    &   4 MB ROM mapping address range. \\
                    \hline
                        \path{0xFEC00000 - DEVICE MEM HIGH}
                    &   Reserved area for OS use. \\
                    \hline
                \end{tabular}}
            \caption{\textit{coreboot} memory map}
            \label{tab:memmap}
        \end{table}

        \subsection{Bootblock}

            The bootblock is the first stage executed after the CPU reset. The
            beginning of this stage is written in assembly language, and its
            main task is to set everything up for a C environment. The rest, of
            course, is written in C. This stage occupies the last 20k
            (fig. \ref{fig:coreboot_diagram}) of the image and within it is a
            main header containing information about the ROM, including the
            size, component alignment, and the offset of the start of the first
            CBFS component. This block is a mandatory component as it also
            contains the entry point of the firmware. \\

            \begin{figure}[H]
                \centering
                \includegraphics[width=0.8\textwidth]{images/fig8_coreboot_architecture.png}
                \caption{\textit{coreboot} ROM architecture
                         (CC BY-SA 4.0, 2024)}
                \label{fig:coreboot_diagram}
            \end{figure}

            Upon startup, the first responsibility of the bootblock is to
            execute the code from the reset vector located at the conventional
            reset vector in 16-bit real mode. This code is specific to the
            processor architecture and, for our board, is stored in the
            architecture-specific sources for x86 within \textit{coreboot}.
            The entry point into \textit{coreboot} code is defined in two files
            in the \path{src/cpu/x86/16bit/} directory: \path{reset16.inc}
            and \path{entry16.inc}. The first file serves as a jump to the
            \path{_start16bit} procedure defined in the second. Due to space
            constraints this function must remain below the 1MB address space
            because the IOMMU has not yet been configured to allow anything
            else. \\

            During this early initialization, the Bootstrap Core (BSC) performs
            several critical tasks while the other cores remain dormant. These
            tasks include saving the results (and displaying them if necessary)
            of the Built-in Self-Test (BIST), formerly known as POST;
            invalidating the TLB to prevent any address translation errors;
            determining the type of reset (e.g., cold start or warm start);
            creating and loading an empty Interrupt Descriptor Table (IDT) to
            prevent the use of "legacy" interrupts from real mode until
            protected mode is reached. In practice, this means that at the
            slightest exception, the BSC will halt. The code then switches to
            32-bit protected mode by mapping the first 4 GB of address space for
            code and data, and finally jumps to the 32-bit reset code labeled
            \path{_protected_start}. \\


            Once in protected mode, which constitutes the "normal" operating
            mode for the processor, the next step is to set up the execution
            environment. To achieve this, the code contained in
            \path{src/cpu/x86/32bit/entry32.inc}, followed by
            \path{src/cpu/x86/64bit/entry64.inc}, and finally
            \path{src/arch/x86/bootblock_crt0.S}, establishes a temporary
            stack, transitions to long mode (64-bit addressing) with paging
            enabled, and sets up a proper exception vector table. The execution
            then jumps to chipset-specific code via the
            \path{bootblock_pre_c_entry} procedure.
            Once these steps are completed, the bootblock has a minimal C
            environment. The procedure now involves allocating
            memory for the BSS, and decompressing and loading the next stage. \\

            The jump to \path{_bootblock_pre_entry} leads to the code files
            \path{src/soc/amd/common/block/cpu/car/cache_as_ram.S} and
            \path{src/vendorcode/amd/agesa/f15tn/gcccar.inc}, which are specific
            to AMD chipsets. It's worth noting that these files were developed by
            AMD's engineers as part of the \textit{AGESA} project. The operations
            performed at this stage are related to pre-RAM memory initialization.
            All cores of all processors (up to a limit of 64 cores) are started.
            The \textit{Cache-As-Ram} is configured using the
            Memory-type range registers. These registers allow the
            specification of a specific configuration for a given memory area
            \cite{BKDG}.
            In this case, the area that should correspond to physical memory is
            mapped to the cache, while other areas, such as PCI or other bus
            zones, are configured accordingly. A specific stack is set up for
            each core of each processor (within the arbitrary limit of 64 cores
            and 7 nodes, meaning 7 Core 0s). Core 0s receive 16KB, while the
            Bootstrap Core (BSC) gets 64KB. The other cores receive 4KB each.
            All cores except the BSC are halted and will restart during the
            romstage. Finally, the execution jumps to the entry point of the
            \textit{bootblock} written in C, labeled
            \path{bootblock_c_entry}.
            This entry point is located in
            \path{src/soc/amd/stoneyridge/bootblock/bootblock.c} and is
            specific to AMD processors. It is the first C routine executed, and
            its role is to verify that the current processor is indeed the BSC,
            allowing the function \path{bootblock_main_with_basetime}
            to be called exclusively by the BSC. \\

            We are now in the file \path{src/lib/bootblock.c}, written by
            Google's team, and entering the
            \path{bootblock_main_with_basetime} function, which immediately
            calls \path{bootblock_main_with_timestamp}. At this stage, the
            goal is to start the romstage, but a few more tasks need to be
            completed.

            The \path{bootblock_soc_early_init} function is called to
            initialize the I2C bus of the southbridge. The
            \path{bootblock_fch_early_init} function is invoked to
            initialize the SPI buses (Serial Peripheral Interface,
            allowing to access the chip that contains the ROM) and the
            serial and "legacy" buses of the southbridge. The CMOS clock is then
            initialized, followed by the pre-initialization of the serial
            console.
            The code then calls the \path{bootblock_mainboard_init}
            function, which enters, for the first time, the files specific to
            the ASUS KGPE-D16 motherboard:
            \path{src/mainboard/ASUS/kgpe-d16/bootblock.c}.
            This code performs the northbridge initialization via the
            \path{bootblock_northbridge_init} function found in
            \path{src/northbridge/amd/amdfam10/bootblock.c}. This involves
            locating the HyperTransport bus and enabling the discovery of
            devices connected to it (e.g., processors). The southbridge is
            initialized using the \path{bootblock_southbridge_init}
            function from \path{src/southbridge/amd/sb700/bootblock.c}.
            This function, largely programmed by Timothy Pearson from Raptor
            Engineering, who performed the first coreboot port for the ASUS
            KGPE-D16, finalizes the activation of the SPI bus and the connection
            to the ROM memory via SuperIO. The state of a recovery jumper is
            then checked (this jumper is intended to reset the CMOS content,
            although it is not fully functional at the moment, as indicated by
            the \path{FIXME} comment in the code). Control then returns to
            \path{bootblock_main} in \path{src/lib/bootblock.c}. \\

            At this point, everything is ready to enter the romstage.
            \textit{coreboot} has successfully started and can now continue its
            execution by calling the \path{run_romstage} function from
            \path{src/lib/prog_loaders.c}. This function begins by locating
            the corresponding segment in the ROM via the southbridge and SPI
            bus using \path{prog_locate}, which utilizes the SPI driver in
            \path{src/drivers/cbfs_spi.c}. The contents of the romstage are
            then copied into the cache-as-ram by
            \path{cbfs_prog_stage_load}. Finally, the \path{prog_run}
            function transitions to the romstage after switching back to
            32-bit mode.

        \subsection{Romstage}

            The \textit{romstage} in \textit{coreboot} serves the critical function
            of early initialization of peripherals, particularly system memory.
            This stage is crucial for setting up the necessary components for the
            platform's operation, ensuring that everything is in place for
            subsequent stages of the boot process.
            During this phase, \textit{coreboot} configures the Advanced
            Programmable Interrupt Controller (APIC), which is responsible for
            correctly handling interrupts across multiple CPUs, especially in
            systems using Symmetric Multiprocessing (SMP). This includes setting
            up the Local APIC on each processor and the IOAPIC, part of the
            southbridge, to ensure that interrupts from peripherals are routed
            to the appropriate CPUs. Additionally, the firmware configures the
            HyperTransport (HT) technology, a high-speed communication protocol
            that facilitates data exchange between the processor and the
            northbridge, ensuring smooth data flow between these components.
            During this stage, microcode patches may be loaded into CPU and
            remain resident, settings related to memory controllers and CPU
            too. \\

            The \textit{romstage} begins with a call to the
            \path{_start} function, defined in
            \path{src/cpu/x86/32bit/entry32.inc} via
            \path{src/arch/x86/assembly_entry.S}. We then enter the
            \path{cache_as_ram_setup} procedure, written in assembly
            language, located in \path{src/cpu/amd/car/cache_as_ram.inc}. This
            procedure configures the cache to load the future \textit{ramstage}
            and initialize memory based on the number of processors and cores
            present. Once this is completed, the code calls
            \path{cache_as_ram_main} in
            \path{src/mainboard/asus/kgpe-d16/romstage.c}, which serves as the
            main function of the \textit{romstage}.
            In the \path{cache_as_ram_main} function, after reducing the
            speed of the HyperTransport bus, only the Bootstrap Core (BSC)
            initializes the spinlocks for the serial console, the CMOS storage
            memory (used for saving parameters), and the ROM. At this point, the
            HyperTransport bus is enumerated, and the PCI bridges are
            temporarily disabled. The port 0x80 of the southbridge, used for
            motherboard debugging with \textit{Post Codes}, is also initialized.
            These codes indicate the status of the boot process and can be
            displayed using special PCI cards connected to the system. The
            SuperIO is then initialized to activate the serial port, allowing
            the serial console to follow \textit{coreboot}’s progress in
            real-time. If everything proceeds as expected, the code 0x30 is
            sent, and the boot process continues. \\

            If the result of the Built-in Self-Test (BIST), saved during the
            \textit{bootblock}, shows no anomalies, all cores of all nodes are
            configured, and they are placed back into sleep mode (except for the
            Core 0s). If everything goes well, the code 0x32 is sent, and the
            process continues. Using the \path{enable_sr5650_dev8} function,
            the southbridge’s P2P bridge is activated. Additionally, a check is
            performed to ensure that the number of physical processors detected
            does not exceed the number of sockets available on the board. If any
            issues were detected during the BIST, the machine will halt, and the
            error will be displayed on the console. Otherwise, the process
            continues, and the default hardware information table is
            constructed, and the microcode of the physical processors is updated
            if necessary. If everything proceeds correctly, the code 0x33 and
            then 0x34 is sent, and the process continues. The information about
            the physical processors is retrieved using \path{amd_ht_init},
            and communication between the two sockets is configured via
            \path{amd_ht_fixup}. This process includes disabling any
            defective HT links (one per socket in this AMD Family 15h chipset).
            If everything is working as expected, the code 0x35 is sent, and
            the boot process continues.
            With the \path{finalize_node_setup} function, the PCI bus is
            initialized, and a mapping is created
            (\path{setup_mb_resource_map}). If all goes well, the code 0x36
            is sent. This is done in parallel across all Core 0s, so the system
            waits for all cores to finish using the
            \path{wait_all_core0_started} function. The communication
            between the northbridge and southbridge is prepared using
            \path{sr5650_early_setup} and
            \path{sb7xx_51xx_early_setup}, followed by the activation of
            all cores on all nodes, with the system waiting for all cores to be
            fully initialized. If everything is successful, the code 0x38 is
            sent. \\

            At this point, the timer is activated, and a warm reset is performed
            via the \path{soft_reset} function to validate all configuration
            changes to the HT, PCI buses, and voltage/power settings of the
            processors and buses. This results in a system reboot, passing again
            through the \textit{bootblock}, but much faster this time since the
            system recognizes the warm reset condition. Once this reboot is
            complete, the HyperTransport bus is reconfigured into isochronous
            mode (switching from asynchronous mode), finalizing the
            configuration process. \\

            Memory training and optimization are also key functions of the
            firmware during the \textit{romstage}. This process involves
            adjusting memory settings, such as timings, frequencies, and
            voltages, to ensure that the installed memory modules operate
            efficiently and stably. This step is crucial for achieving optimal
            performance, especially when dealing with large amounts of RAM
            and many CPU cores, as supported by the KGPE-D16. We'll see that
            in detail during the next chapter. \\

            After memory initialization, the process returns to the
            \path{cache_as_ram_main} function, where a memory test is
            performed. This involves writing predefined values to specific
            memory locations and then verifying that the values can be read
            back correctly.
            If everything passes successfully, the CBMEM is initialized and
            one sends code \path{0x41}. At this point, the configuration of
            the PCI bus is prepared, which will be completed during the ramstage
            by configuring the PCI bridges. The system then exits
            \path{cache_as_ram_main} and returns to
            \path{cache_as_ram_setup} to finalize the process.

            \textit{coreboot} then transitions to the next stage, known as the
            postcar stage, where it exits the cache-as-RAM mode and
            begins using physical RAM.

        \subsection{Ramstage}

            The ramstage performs the general initialization of all peripherals,
            including the initialization of PCI devices, on-chip devices, the
            TPM (if not done by verstage), graphics (optional), and the CPU
            (setting up the System Management Mode). After this initialization,
            tables are written to inform the payload or operating system about
            the existence and current state of the hardware. These tables
            include ACPI tables (specific to x86), SMBIOS tables (specific to
            x86), coreboot tables, and updates to the device tree (specific to
            ARM). Additionally, the ramstage locks down the hardware and
            firmware by applying write protection to boot media, locking
            security-related registers, and locking SMM (specific to x86),
            which is a resident component in a protected area.
            \cite{coreboot_docs}. CBMEM data structures (like coreboot tables,
            memory map, etc.) are populated during this stage and left resident
            for the OS or payload to access, same for SMBIOS tables, ACPI tables
            and eventually option ROMS. \\

            Effective resource allocation is essential for system stability,
            particularly in complex configurations involving multiple CPUs
            and peripherals. This stage manages initial resource allocation,
            resolving any conflicts between hardware components to prevent
            resource contention and ensure smooth operation and security, which
            is a major concern in modern systems. This includes support for
            IOMMU, which is crucial for preventing unauthorized direct memory
            access (DMA) attacks, particularly in virtualized environments
            (however there are still vulnerabilities that can be exploited,
            such as sub-page or IOTLB-based attacks or even configuration
            weaknesses \cite{medeiros2017}\cite{markuze2021}). \\

            \subsubsection{Advanced Configuration and Power Interface}

                The Advanced Configuration and Power Interface (ACPI) is a
                critical component of modern computing systems, providing an
                open standard for device configuration and power management by
                the operating system (OS). Developed in 1996 by Intel,
                Microsoft, and Toshiba, ACPI replaced the older Advanced Power
                Management (APM) standard with more advanced and flexible power
                management capabilities \cite{intel_acpi_introduction_2023}.
                At its core,
                ACPI is implemented through a series of data structures and
                executable code known as ACPI tables, which are provided by the
                system firmware and interpreted by the OS at runtime. It means
                that these are components from the firmware that remain resident
                while the OS runs. These tables describe
                various aspects of the system, including hardware resources,
                device power states, and thermal zones. The ACPI Specification
                outlines these structures and provides the necessary
                standardization for interoperability across different platforms
                and operating systems \cite{acpi_os_support}. These tables are
                used by the OS to perform low-level task, including managing
                power states of the CPU, controlling the voltage and frequency
                scaling (also known as Dynamic Voltage and Frequency Scaling,
                or DVFS), and coordinating power delivery to peripherals. \\

                The ACPI Component Architecture (ACPICA) is the reference
                implementation of ACPI, providing a common codebase that can be
                used by OS developers to integrate ACPI support. ACPICA includes
                tools and libraries that allow for the parsing and execution of
                ACPI Machine Language (AML) code, which is embedded within the
                ACPI tables \cite{acpi_programming}. One of the key tools in
                ACPICA is the Intel ACPI Source Language (IASL) compiler, which
                converts ACPI Source Language (ASL) code into AML bytecode,
                allowing firmware developers to write custom ACPI
                methods \cite{intel_acpi_spec}. The triggering of ACPI events is
                managed through a combination of hardware signals and software
                routines. For example, when a user presses the power button on a
                system, an ACPI event is generated, which is then handled by the
                OS. This event might trigger the system to enter a low-power
                state, such as sleep or hibernation, depending on the
                configuration provided by the ACPI tables
                \cite{acpi_os_support}. These power states are defined in the
                ACPI specification, with global states (G0 to G3) representing
                different levels of system power consumption, and device states
                (D0 to D3) representing individual device power levels. \\

                The ASUS KGPE-D16 mainboard, which is designed for server and
                high-performance computing environments, needs ACPI for managing
                its power distribution across multiple CPUs and attached
                peripherals. ACPI is integral in controlling the power states of
                various components, thereby optimizing performance and energy
                use. Additionally, the firmware on the KGPE-D16 uses ACPI tables
                to manage system temperature and fan speed, ensuring reliable
                operation under heavy workloads \cite{asus_kgpe_d16_manual}.

            \subsubsection{System Management Mode}

                System Management Mode (SMM) is a highly privileged operating
                mode provided by x86 processors for handling system-level
                functions such as power management, hardware control, and other
                critical tasks that are to be isolated from the OS and
                applications. Introduced by Intel, SMM operates in an
                environment separate from the main operating system, offering a
                controlled space for executing sensitive operations
                \cite{uefi_smm_security}. This is another firmware component
                that remains resident while the OS runs. \\

                SMM is triggered by a System Management Interrupt (SMI), which
                is a non-maskable interrupt that causes the CPU to save its
                current state and switch to executing code stored in a protected
                area of memory called System Management RAM (SMRAM). SMRAM is a
                specialized memory region that is isolated from the rest of the
                system, making it inaccessible to the OS and preventing
                tampering or interference from other software
                \cite{heasman2007}.
                Within SMM, the firmware can execute various low-level functions
                that require direct hardware control or need to be protected
                from the OS. This includes tasks such as thermal management,
                where the system monitors CPU temperature and adjusts
                performance or power levels to prevent overheating, as well as
                power management routines that enable efficient energy usage
                by adjusting power states based on system activity
                \cite{offsec_bios_smm}. One of the critical security features of
                SMM is its role in managing firmware updates and handling
                system-level security events. Because SMM operates in a
                privileged mode that is isolated from the OS, it can
                apply firmware updates and could respond to security threats
                without being affected by potentially compromised system
                software \cite{domas2015}. However, the high privilege level and
                isolation of SMM also present significant security challenges.
                If an attacker can compromise SMM, they gain full control over
                the system, bypassing all security measures implemented by the
                OS \cite{cyber_smm_hack}. Also, with a proprietary firmware,
                it means that this code with a very high priviledge level
                cannot be audited at all, nor even replaced. \\

                The ASUS KGPE-D16 mainboard needs SMM to perform critical
                management tasks that need to be done in parallel from the
                operating system. For example, SMM is used to monitor and manage
                system health by responding to thermal events and adjusting
                power levels to maintain system stability. SMM operates
                independently of the main operating system, allowing it to
                perform sensitive tasks securely. \textit{coreboot}
                supports SMM, but its implementation is typically
                minimal compared to traditional proprietary firmware. In
                \textit{coreboot}, SMM initialization involves setting
                up the System Management Interrupt (SMI) handler and configuring
                System Management RAM (SMRAM), the memory region where SMM code
                executes\cite{brown2003linuxbios}. The extent of SMM support in
                \textit{coreboot} can vary significantly depending on the
                hardware platform and the specific requirements of the system.
                \textit{coreboot}'s design philosophy emphasizes a lightweight
                and fast boot process, delegating more complex management tasks
                to payloads or the operating system itself
                \cite{reinauer2008coreboot}.

                One of the key challenges with implementing SMM in
                \textit{coreboot} is ensuring that SMI handlers are configured
                correctly to manage necessary system tasks without compromising
                security or performance. \textit{coreboot}'s approach to SMM is
                consistent with its overall goal of providing a streamlined and
                efficient firmware solution, leaving more intricate
                functionalities to be handled by subsequent software layers
                \cite{mohr2012comparative}.

        \subsection{Payload}

            The payload is the software that executes after coreboot has
            completed its initialization tasks. It resides in the CBFS and is
            predetermined at compile time, with no option to choose it at
            runtime. The primary role of the payload is to load and hand control
            over to the operating system. In some cases, the payload itself can
            be a component of the operating system \cite{coreboot_docs}.
            Examples of payloads are \textit{GNU GRUB}, \textit{SeaBIOS},
            \textit{memtest86+} or even sometimes the \textit{Linux kernel}
            itself. \\

            \textit{TianoCore}, a free implementation of the UEFI (Unified
            Extensible Firmware Interface) specification is often used as a
            payload \cite{tianocore_payload}.
            It provides a UEFI environment after \textit{coreboot} has completed
            its initial hardware initialization. This allows the system to
            benefit from the advanced features of UEFI, such as a more flexible
            boot manager, enhanced features, and support for modern
            hardware. Indeed, UEFI, and by extension \textit{TianoCore},
            includes a driver model that allows hardware manufacturers to
            provide UEFI-compatible drivers. These drivers can be loaded at
            boot time, allowing the firmware to support a wide range of modern
            devices that \textit{coreboot}, with its more minimalistic and
            custom-tailored approach, might not support out of the box.
            For example, GOP drivers are responsible for setting up the
            graphics hardware in UEFI environments. They replace the older VGA
            BIOS routines used in legacy BIOS systems. With GOP drivers,
            the system can initialize the GPU and display a graphical interface
            even before the operating system loads \cite{osdev_gop}.
            These are other examples of resident firmware components while the
            OS is running. \\
            Hardware manufacturers can distribute proprietary UEFI drivers as
            part of firmware updates, making it straightforward for end-users
            to install and use them. This is especially useful for specialized
            hardware that requires specific drivers not included in the
            free software community. It also gives hardware vendors more control
            over how their devices are initialized and used, which can be
            an advantage for vendors but is a freedom and user control
            limitation.

            Payloads are then definitely important parts of the firmware.

    \section{AMD Platform Security Processor and Intel Management Engine}

        The AMD Platform Security Processor (PSP) and Intel Management Engine
        (ME) are embedded subsystems within AMD and Intel processors,
        respectively, that handle a range of security-related tasks independent
        of the main CPU. These subsystems are fundamental to the security
        architecture of modern computing platforms, providing functions such as
        secure boot, cryptographic key management, and remote system management
        \cite{herrmann2017dissecting}.

        The AMD PSP is based on an ARM Cortex-A5 processor and is responsible
        for several security functions, including the validation of firmware
        during boot (secure boot), management of Trusted Platform Module (TPM)
        functions, and handling cryptographic operations such as key generation
        and storage. The PSP operates independently of the main x86 cores,
        which allows it to execute security functions even when the main system
        is powered off or compromised by malware \cite{herrmann2017dissecting}.
        The PSP's isolated environment ensures that sensitive operations are
        protected from threats that could affect the main OS. \\

        Similarly, the Intel Management Engine (ME) is a dedicated
        processor embedded within Intel chipsets that operates
        independently of the main CPU. The ME is a comprehensive subsystem that
        provides a variety of functions, including out-of-band system
        management, security enforcement, and support for Digital Rights
        Management (DRM) \cite{intel_csme}. The ME's firmware runs on an
        isolated environment that allows it to perform these tasks securely,
        even when the system is powered off. This capability is crucial for
        enterprise environments where administrators need to perform remote
        diagnostics, updates, and security checks without relying on the main
        OS.
        Intel ME enforces Digital Rights Management (DRM)
        through a multifaceted approach leveraging its deeply embedded,
        hardware-based capabilities. At the core is the Protected
        Execution Environment (PEE), which operates independently from the main
        CPU and operating system. This isolation allows to privately
        manage cryptographic keys, certificates, and other sensitive data
        critical for DRM, which can be very problematic from a user freedom
        perspective \cite{fsf_intel_me}. By handling encryption and decryption
        processes within this protected environment, Intel ME ensures that
        DRM-protected content, such as video streams, remains secure and
        unreachable by the user, raising concerns about the control users have
        over their own devices \cite{eff_intel_me}.
        Intel ME also plays a significant role in maintaining platform
        integrity through the secure boot process. During secure boot, Intel ME
        ensures that only digitally signed and authorized operating systems and
        applications are loaded, which can prevent users from installing
        alternative or modified software on their own hardware, further
        restricting their freedom \cite{uefi_what_is_uefi}. This is further
        reinforced by Intel ME's remote attestation capabilities, where the
        system’s state is reported to a remote server. This process verifies
        that only systems meeting specific security standards dictated by third
        parties are allowed to access DRM-protected content, potentially
        limiting users' control over their own devices \cite{proprivacy_intel_me}.
        Moreover, Intel ME supports High-bandwidth Digital Content Protection
        (HDCP), a technology that restricts how digital content is transmitted
        over interfaces like HDMI or DisplayPort. By enforcing HDCP, Intel ME
        ensures that protected digital content, such as high-definition video,
        is only transmitted to and displayed on authorized devices, effectively
        preventing users from freely using the content they have legally
        acquired \cite{phoronix_hdcp_2_2_i915}\cite{kernel_mei_hdcp}.
        Together, these features enable Intel ME to provide a comprehensive and
        robust DRM enforcement mechanism. However, this also means that users
        have less control over their own hardware and digital content, raising
        serious concerns about privacy, user autonomy, and the broader
        implications for freedom in computing
        \cite{fsf_intel_me}\cite{netgarage_intel_me}. \\

        Added to that, Intel ME has been a source of controversy due to its deep
        integration into the hardware and its potential to be exploited if
        vulnerabilities are discovered. Researchers have demonstrated ways to
        hack into the ME, potentially gaining control over a system even when
        it is powered off \cite{blackhat_me_hack}. These concerns have led to
        calls for greater transparency and security measures around the ME and
        similar subsystems. When comparing Intel ME and AMD PSP, the primary
        difference lies in their scope and functionality. Intel ME offers more
        extensive remote management capabilities, making it a more comprehensive
        tool for enterprise environments, while AMD PSP focuses more narrowly on
        core security tasks. Nonetheless, both play critical roles in ensuring
        the security and integrity of modern computing systems. \\

        The ASUS KGPE-D16 mainboard does not include AMD PSP nor Intel ME.

% ------------------------------------------------------------------------------
% CHAPTER 4: Memory initialization and training
% ------------------------------------------------------------------------------
\chapter{Memory initialization and training}

    \section{Importance of DDR3 memory initialization}

            Memory modules are designed solely for storing data. The only valid
            operations on a memory device are reading data stored in the device,
            writing (or storing) data into the device, and refreshing the data.
            Memory modules consist of large rectangular arrays of memory cells,
            including circuits used to read and write data into the arrays, and
            refresh circuits to maintain the integrity of the stored data. The
            memory arrays are organized into rows and columns of memory cells,
            known as word lines and bit lines, respectively. Each memory cell
            has a unique location or address defined by the intersection of a
            row and a column. A DRAM memory cell is a capacitor that is charged
            to produce a 1 or a 0. \\

            DDR3 (Double Data Rate Type 3) is a widely used type of
            SDRAM (Synchronous Dynamic Random-Access Memory) that offers
            significant performance improvements over its predecessors,
            DDR and DDR2. A DDR3 DIMM module contains 240 contacts.
            Key features of DDR3 include higher data rates,
            lower power consumption, and increased memory capacity, making
            it essential for high-performance computing environments
            \cite{DDR3_wiki}. One of the critical aspects of DDR3 is its
            internal architecture, which supports data rates ranging from
            800 to 1600 Mbps and operates at a lower voltage of 1.5V. This
            enables faster data processing and more efficient power usage,
            crucial for modern applications that require high-speed memory
            access \cite{samsung_ddr3}. Additionally, DDR3 memory modules are
            available in larger capacities, allowing systems to handle larger
            datasets and more complex computing tasks \cite{altera2008}.
            However, the advanced features of DDR3 come with increased
            complexity in its initialization and operation.
            The DDR3 memory interface, used by the ASUS KGPE-D16, is
            source-synchronous. Each memory module generates a Data Strobe
            (DQS) pulse simultaneously with the data (DQ) it sends during
            a memory read operation. Similarly, a DQS must be generated
            with its DQ information when writing to memory. The DQS differs
            between write and read operations. Specifically, the DQS generated
            by the system for a write operation is centered in the data bit
            period, while the DQS provided by the memory during a read operation
            is aligned with the edge of the data period \cite{samsung_ddr3}. \\

            Due to this edge alignment, the read DQS timing can be adjusted
            to meet the setup and hold requirements of the registers capturing
            the read data. To improve timing margins or reduce simultaneous
            switching noise in the system, the DDR3 memory interface also allows
            various other timing parameters to be adjusted. If the system uses
            dual-inline memory modules (DIMMs), as in our case, the interface
            provides write leveling: a timing adjustment that compensates for
            variations in signal travel time \cite{micron_ddr3}.
            To reduce simultaneous switching noise, DIMM modules feature a
            fly-by architecture for routing the address, command, and clock
            signals, which causes command signals to reach the
            different memory devices with a delay. The fly-by topology has a
            "daisy-chain" structure with either very short stubs or no stubs
            at all. This structure results in fewer branches and point-to-point
            connections: everything originates from the controller, passing
            through each module on the node, thereby increasing the throughput.
            In this topology, signals are routed sequentially
            from the memory controller to each DRAM chip, reducing signal
            reflections and improving overall signal integrity.
            It means that routing is done in the order of byte lane numbers,
            and the data byte lanes are routed on the same layer. Routing can be
            simplified by swapping data bits within a byte lane if necessary.
            The fly-by topology contrasts with the dual-T topology
            (fig. \ref{fig:fly-by}).  This design is essential for maintaining
            stability at the high speeds DDR3 operates at, but it also
            introduces timing challenges, such as timing skew, that must be
            carefully managed \cite{micron_ddr3}. \\

            \begin{figure}[H]
                \centering
                \begin{minipage}[b]{0.45\textwidth}
                    \centering
                    \includegraphics[width=0.90\textwidth]{images/fly-by.png}
                \end{minipage}%
                \begin{minipage}[b]{0.45\textwidth}
                    \centering
                    \includegraphics[width=0.824\textwidth]{images/t.png}
                \end{minipage}
                \caption{DDR3 fly-by \textit{versus} T-topology
                         (CC BY-SA 4.0, 2021)}
                \label{fig:fly-by}
            \end{figure}

            Proper memory initialization ensures that the memory controller
            and the memory modules are correctly configured to work together,
            allowing for efficient data transfer and reliable operation. The
            initialization process involves setting various parameters,
            such as memory timings, voltages, and frequencies, which are
            critical for ensuring that the memory operates within its optimal
            range \cite{samsung_ddr3}. Failure to initialize DDR3 memory
            correctly can lead to several serious consequences, including
            system instability, data corruption, and reduced performance
            \cite{SridharanVilas2015MEiM}. In the worst-case scenario, improper
            memory initialization can prevent the system from booting entirely,
            as the memory subsystem fails to function correctly.
            In the context of the ASUS KGPE-D16, a server motherboard
            designed for high-performance applications, proper DDR3 memory
            initialization is particularly important. The KGPE-D16 supports
            up to 256GB of DDR3 memory across 16 DIMM slots, and any issues
            during memory initialization, if non-fatal, could severely impact
            the system's ability to handle large datasets or maintain stable
            operation under heavy workloads \cite{asus_kgpe_d16_manual}. Given
            the critical role that memory plays in the overall performance of
            the KGPE-D16, ensuring that DDR3 memory is correctly initialized
            is essential for achieving the desired balance of performance,
            reliability, and stability in demanding server environments.

    \section{General steps for DDR3 configuration}

        DDR3 memory initialization is a detailed and essential
        process that ensures both the stability and performance of the
        system. The process involves several critical steps: detection
        and identification of memory modules, initial configuration of the
        memory controller, adjustment of timing and voltage settings, and
        the execution of training and calibration procedures. \\

        The initialization begins with the detection and identification of
        the installed memory modules. During the BIST, the firmware reads
        the Serial Presence Detect (SPD) data stored on
        each memory module. SPD data contains crucial information about
        the memory module's specifications, including size, speed, CAS
        latency (CL), RAS to CAS delay (tRCD), row precharge time (tRP),
        and row cycle time (tRC). This data allows to configure
        the memory controller for optimal compatibility and performance. \\

        Indeed, once the memory modules have been identified, the firmware
        proceeds to the initial configuration of the memory controller.
        This controller is governed by a state machine that
        manages the sequence of operations required to initialize,
        maintain, and control memory access. This state machine consists of
        multiple states that represent various phases of memory operation,
        such as reset, initialization, calibration, and data transfer.
        The transitions between these states are either automatic or
        command-driven, depending on the specific requirements of each
        phase \cite{samsung_ddr3}\cite{micron_ddr3}.
        This state machine is presented in the
        fig. \ref{fig:ddr3_state_machine}. Automatic transitions, depicted
        by thick arrows in the automaton, occur without external
        intervention. These typically include transitions that ensure
        the memory enters a stable state, such as the transition from
        power-on to initialization, or from calibration to idle states.
        These transitions are crucial for maintaining the integrity and
        stability of the memory system, as they ensure that the controller
        progresses through necessary stages like ZQ calibration and write
        leveling, which are essential for proper signal timing and
        impedance matching
        \cite{samsung_ddr3}\cite{micron_ddr3}\cite{burnett_ddr3}. \\

        On the other hand, command-driven transitions, represented by normal
        arrows in the automaton, require specific commands issued by the
        memory controller or the CPU to advance to the next state. For
        instance, the transition from the idle state to the data transfer
        state requires explicit read or write commands. Similarly,
        transitioning from the initialization state to the calibration
        state involves issuing mode register set (MRS) commands that
        configure the memory’s operating parameters. These command-driven
        transitions are integral to the dynamic operation of the memory
        system, allowing the controller to respond to the system's
        operational needs and ensuring that memory accesses are performed
        efficiently and accurately \cite{samsung_ddr3}\cite{micron_ddr3}. \\

        The memory controller configuration
        involves setting up fundamental parameters such as the memory clock
        (MEMCLK) frequency and the memory channel configuration. The MEMCLK
        frequency is derived from the SPD data, while the memory channels
        are configured to operate in single, dual, or quad-channel modes,
        depending on the system architecture and the installed modules
        \cite{burnett_ddr3}. Proper configuration of the memory controller
        is vital to ensure synchronization with the memory modules,
        establishing a stable foundation for subsequent operations. \\

        The first critical step, during the INIT phase involves the
        adjustment of timing and voltage settings. These settings are
        essential for ensuring that DDR3 memory operates efficiently and
        reliably. Key timing parameters include CAS Latency (CL), RAS to
        CAS Delay (tRCD), Row Precharge Time (tRP), and Row Cycle Time (tRC).
        These parameters are finely tuned to balance speed and stability
        \cite{samsung_ddr3}. The BIOS uses the SPD data to set these
        parameters and may also adjust them dynamically to achieve the
        best possible performance. Voltage settings, such as DRAM voltage
        (typically 1.5V for DDR3) and termination voltage (VTT), are also
        configured to maintain stable operation, especially under varying
        conditions such as temperature fluctuations \cite{micron_ddr3}. \\

        Training and calibration are among the most complex and crucial
        stages of DDR3 memory initialization. The fly-by topology used
        for address, command, and clock signals in DDR3 modules enhances
        signal integrity by reducing the number of stubs and their lengths,
        but it also introduces skew between the clock (CK) and data strobe
        (DQS) signals \cite{micron_ddr3}. This skew must be compensated to
        ensure that data is written and read correctly. The BIOS performs
        write leveling, which adjusts the timing of DQS relative to CK
        for each memory module. This process ensures that the memory
        controller can write data accurately across all modules, even
        when they exhibit slight variations in signal timing due to the
        physical layout \cite{samsung_ddr3}. \\

        \begin{figure}[H]
            \centering
            \begin{tikzpicture}[scale=0.6,
                                transform shape,
                                shorten >=1pt,
                                node distance=5cm and 5cm,
                                on grid,
                                auto]
                % States
                \node[state, initial] (reset) {RESET};
                \node[draw=none,fill=none] (any) [below=2cm of reset] {ANY};
                \node[state] (init) [right=of reset] {INIT};
                \node[state] (zqcal) [below=of init] {ZQ Calibration};
                \node[state, accepting] (idle) [right=of init] {IDLE};
                \node[state] (writelevel) [above=of idle] {WRITE LEVELING};
                \node[state] (refresh) [right=of idle] {REFRESH};
                \node[state] (activation) [below=of idle] {ACTIVATION};
                \node[state] (bankactive) [below=of activation] {BANK ACTIVE};
                \node[state] (readop) [below right=of bankactive] {READ OP};
                \node[state] (writeop) [below left=of bankactive] {WRITE OP};
                \node[state] (prechrg) [below right=of readop] {PRE-CHARGING};
                % Transitions
                \path[->, line width=0.2mm, >=stealth]
                    (reset) edge node {} (init)
                    (idle) edge [bend left=20] node {} (writelevel)
                           edge [bend left=20] node {REF} (refresh)
                           edge node {} (activation)
                           edge [bend left=10] node {ZQCL/S} (zqcal)
                    (activation) edge node {} (bankactive)
                    (bankactive) edge [bend left=30] node {PRE} (prechrg)
                                 edge [bend left=20] node {write} (writeop)
                                 edge [bend right=20] node {read} (readop)
                    (writeop) edge [loop left] node {write} (writeop)
                              edge [bend left=10] node {read\_a} (readop)
                              edge [bend right=15] node {PRE} (prechrg)
                    (readop) edge [loop right] node {read} (readop)
                             edge [bend left=10] node {write\_a} (writeop)
                             edge [bend right=15] node {PRE} (prechrg);
                % Thick transitions
                \path[->, line width=0.5mm, >=stealth]
                    (any) edge node {} (reset)
                    (init) edge node {ZQCL} (zqcal)
                    (zqcal) edge [bend left=10] node {} (idle)
                    (writelevel) edge [bend left=20] node {MRS} (idle)
                    (refresh) edge [bend left=20] node {} (idle)
                    (writeop) edge node {} (prechrg)
                              edge [bend left=20] node {} (bankactive)
                    (readop) edge [bend left=15] node {} (prechrg)
                             edge [bend right=20] node {} (bankactive)
                    (prechrg) edge [bend right=20] node {} (idle);
            \end{tikzpicture}
            \caption{DDR3 controller state machine}
            \label{fig:ddr3_state_machine}
        \end{figure}

        ZQ calibration is another vital procedure that adjusts the
        output driver impedance and on-die termination (ODT) to match
        the system’s characteristic impedance \cite{micron_ddr3}. This
        calibration is critical for maintaining signal integrity under
        different operating conditions, such as voltage and temperature
        changes. During initialization, the memory controller issues a
        ZQCL command to the DRAM modules, triggering the calibration
        sequence that optimizes impedance settings.
        This ensures that the memory system can
        operate with tight timing tolerances, which is crucial for
        systems requiring high reliability.
        Read training is also essential to ensure that data read from
        the memory modules is interpreted correctly by the memory
        controller. This process involves adjusting the timing of the
        read data strobe (DQS) to align perfectly with the data being
        received. Proper read training is necessary for reliable data
        retrieval, which directly impacts system performance and stability. \\

        ZQCS (ZQ Calibration Short) however is a procedure used
        to periodically adjust the DRAM's ODT and output driver impedance
        during normal operation. Unlike the full ZQCL (ZQ Calibration Long),
        which is performed during initial memory initialization, ZQCS is a
        quicker, less comprehensive calibration that fine-tunes the
        impedance settings in response to changes in temperature, voltage,
        or other environmental factors. This helps maintain optimal signal
        integrity and performance throughout the memory's operation without
        the need for a full recalibration. \\

        In summary, the DDR3 memory initialization process in systems
        like the ASUS KGPE-D16 involves a series of detailed and
        interdependent steps that are critical for ensuring system
        stability and performance. These include the detection and
        identification of memory modules, the initial configuration of
        the memory controller, precise adjustments of timing and voltage
        settings, and rigorous training and calibration procedures.

    \section{Memory initialization techniques}

        \subsection{Memory training algorithms}

            Memory training algorithms are designed to fine-tune the
            operational parameters of memory modules, such as timing, voltage,
            and impedance. These algorithms play a crucial role in achieving
            the optimal performance of DDR3 memory systems, particularly
            in complex multi-core environments where synchronization
            and timing are challenging. The primary algorithms used in
            memory training include ZQ calibration and write leveling.
            Optimizing timing and voltage settings is a critical aspect of
            memory training. The memory controller adjusts parameters such as
            CAS latency, RAS to CAS delay, and other timing characteristics
            to ensure that data is read and written with minimal delay
            and maximum accuracy. Voltage adjustments are also crucial,
            as they help stabilize the operation of memory modules by
            ensuring that the power supplied is within the optimal range,
            compensating for any variations due to temperature or other factors
            \cite{micron_ddr3}\cite{burnett_ddr3}\cite{gopikrishna2021novel}.
            \\

            ZQ calibration is a critical step in DDR3 memory initialization that
            ensures the proper impedance matching of the output driver and
            on-die termination (ODT) resistance. Impedance matching is crucial
            for maintaining signal integrity by minimizing reflections and
            ensuring reliable data transmission between the memory controller
            and the DRAM modules. It is initiated by sending ZQCL (ZQ
            Calibration Long) commands to the DDR3 DIMMs. Each ZQCL command
            triggers a long calibration cycle within the DRAM module. The
            purpose of this calibration is to adjust the output driver impedance
            and the ODT resistance to match the specified target impedance. This
            adjustment compensates for process variations, voltage fluctuations,
            and temperature changes that can affect the impedance
            characteristics of the DRAM module \cite{gopikrishna2021novel}. \\

            A bit in the DRAM Controller
            Timing register is set to 1 to send the ZQCL command, and an address
            bit is also set to 1 to indicate that the ZQCL command should be
            directed to the memory module. Upon receiving the ZQCL command, the
            DRAM module begins the calibration process. This involves a series
            of internal adjustments where the DRAM module measures its current
            impedance and compares it against the target impedance. The module
            then modifies its internal settings to reduce the difference between
            the current and target impedance values
            \cite{gopikrishna2021novel}\cite{samsung_ddr3}. This process is
            iterative, meaning that it may require multiple adjustments to
            converge on the optimal impedance settings. The calibration is
            designed to ensure that the DRAM module's impedance remains within
            a tight tolerance, which is critical for high-speed data
            communication. The ZQ calibration process is time-sensitive. After
            issuing the ZQCL command, the system must wait for 512 memory
            clock cycles (MEMCLKs) to allow the calibration to complete.
            This delay is necessary  because the calibration involves both
            measurement and adjustment phases, which require precise timing
            to ensure accuracy \cite{gopikrishna2021novel}. If the system does
            not wait the full 512 MEMCLKs, the calibration may be incomplete,
            leading to suboptimal impedance matching and potential signal
            integrity issues, such as reflections or noise on the data lines. \\

            During the ZQ calibration, the DRAM module adjusts its output driver
            impedance, which controls the strength of the signals it sends out.
            The stronger the signal, the less susceptible it is to noise, but if
            the impedance is too high or too low, it can cause signal distortion
            or reflections. The ODT resistance is also calibrated to properly
            terminate signals that reach the end of a data line. Proper
            termination is essential to prevent signal reflections that could
            interfere with the integrity of the data being transmitted. The ZQCL
            command adjusts these settings by fine-tuning the resistance values
            based on the module’s feedback, ensuring that the signal paths are
            optimized for both transmission and termination. Once the ZQ
            calibration is complete, the DCT register bit is reset to 0,
            indicating that the calibration command has been processed. The
            memory controller then verifies that the DRAM module has correctly
            adjusted its impedance settings. This verification process may
            involve additional test signals sent across the memory bus to
            confirm that signal integrity meets the required standards. If the
            calibration is successful, the memory subsystem is now properly
            calibrated and ready for normal operation. In systems with LRDIMMs
            or RDIMMs, additional steps may be necessary to ensure that all
            ranks and channels are calibrated correctly, particularly in
            multi-rank configurations where impedance matching can be more
            complex. However, in systems with complex memory configurations,
            such as those using multiple DIMMs per channel or operating at
            higher memory frequencies, the ZQ calibration process becomes even
            more critical. The calibration may need to be repeated at different
            operating points to ensure that the memory subsystem remains stable
            across all conditions. This could involve performing multiple ZQCL
            calibrations at different memory frequencies, or under different
            thermal conditions, to account for the dynamic nature of memory
            operation in modern systems. \\

            In seed-based algorithms, an initial "seed" value is used
            as a reference point for the calibration process. The memory
            controller iteratively adjusts the impedance based on feedback
            from the memory module, refining the calibration with each
            iteration. This method provides a more precise calibration,
            particularly in systems where fine-tuned impedance matching is
            critical for high-frequency operations \cite{kim2010design}.
            Also, while seed-based methods can accelerate the convergence
            of calibration, they require careful selection of initial seed
            values to avoid suboptimal or even faulty impedance settings
            \cite{gopikrishna2021novel}. \\

            Write leveling is another critical aspect of memory training,
            particularly in DDR3 systems that use a fly-by topology. It involves
            using the physical layer (PHY) to detect the edge of the Data Strobe
            (DQS) signal in synchronization with the clock (CK) signal on the
            DIMM (Dual In-line Memory Module) during write access. The DQS
            signal is a timing signal generated by the memory controller that
            accompanies data (DQ) during read and write operations. For write
            operations, the DQS signal must be perfectly aligned with the CK
            signal to ensure that data is correctly written to memory cells.
            Indeed, in systems using a fly-by topology, the DQS signal might
            arrive at different times for different memory devices on the same
            module due to the signal traveling through different lengths of
            trace. Write leveling compensates for this skew by adjusting the
            timing of the DQS signal relative to the CK signal for each lane
            (a group of data lines) \cite{burnett_ddr3}. This training is
            performed on a per-channel and per-DIMM basis, ensuring that each
            memory module is correctly synchronized with the memory controller,
            minimizing timing mismatches that could lead to data corruption. \\

            Write leveling implies to perform a DQS position training, a
            specific form of training focused on aligning the DQS signal with
            the data (DQ) signals during write operations. In this process,
            the memory controller adjusts the phase of the DQS signal to ensure
            that it is correctly aligned with the data signals across all data
            lanes, centering the DQS signal within the "data eye" for optimal
            timing. This ensures that all data bits are written correctly and
            consistently across the memory module, reducing the risk of timing
            errors and data corruption. Additionally, DQS receiver training is
            also needed to ensure that the memory controller can correctly
            capture the DQS signal during read operations
            \cite{micron_ddr3}.
            The core operation is to make the MCT send out specific test
            patterns to the DRAM to determine the timing relationship between
            the DQS and data signals, then the MCT adjusts the delay or phase of
            the DQS signal relative to the clock signal (CK) and the data
            signals (DQ) while checking the integrity of the test data in the
            DRAM. \\

            Using seed-based algorithms, the memory controller sets an initial
            delay value and then iteratively adjusts it based on the feedback
            received from the memory module. This process ensures that the DQS
            signal is correctly aligned with the CK signal at the memory
            module's pins, minimizing the risk of data corruption and ensuring
            reliable write operations \cite{samsung_ddr3}\cite{gopikrishna2021novel}.
            Seed-based write leveling offers improved precision but must be
            finely tuned to account for the specific characteristics of the
            memory module and the overall system architecture
            \cite{gopikrishna2021novel}. \\

            In contrast to seed-based algorithms, seedless methods do not rely on
            an initial reference value. Instead, they dynamically adjust the
            impedance and timing parameters during the calibration process.
            Seedless ZQ calibration continuously monitors the impedance of the
            memory module and makes real-time adjustments to maintain optimal
            matching. This approach can be beneficial in environments where the
            operating conditions are highly variable, as it allows for more
            flexible and adaptive calibration \cite{kim2010design}. Similarly,
            seedless write leveling dynamically adjusts the DQS timing based on
            real-time feedback from the memory module. This method is particularly
            useful in systems where the memory configuration is frequently changed
            or where the operating conditions vary significantly
            \cite{micron_ddr3}\cite{gopikrishna2021novel}. The traditional ZQ
            calibration methods, while effective, often struggle with matching
            impedance perfectly across all conditions. A master thesis by
            \textcite{gopikrishna2021novel} builds upon these traditional methods
            by proposing enhancements that involve more sophisticated calibration
            approaches, leading to better impedance matching and overall memory
            performance \cite{gopikrishna2021novel}.


        \subsection{BIOS and Kernel Developer Guide (BKDG) recommendations}

            The BIOS and Kernel Developer Guide (BKDG from \textcite{BKDG}) is a
            technical manual aimed at BIOS developers and operating system kernel
            programmers. It provides in-depth documentation on the AMD
            processor architecture, system initialization processes, and
            configuration guidelines. The document is essential for
            understanding the proper initialization sequences, including
            those for DDR3 memory, to ensure system stability and
            performance, particularly for AMD Family 15h processors. \\

            The initialization of DDR3 memory begins with configuring the DDR
            supply voltage regulator, which ensures that the memory modules
            receive the correct power levels. Following this, the Northbridge
            (NB) P-state is forced to \path{NBP0}, a state that guarantees
            stable operation during the initial configuration phases. Once these
            preliminary steps are completed, the initialization of the DDR
            physical layer (PHY) begins, which is critical for setting up
            the communication interface between the memory controller and the
            DDR3 modules. PHY fence training deals with overall signal alignment
            at the physical interface, while ZQ calibration focuses on impedance
            matching, and write leveling addresses timing alignment during
            write operations. Each process involves different methods as PHY
            fence training uses iterative timing adjustments, ZQ calibration
            uses impedance adjustments via the ZQ pin, and write leveling
            adjusts DQS timing relative to CK during writes. These processes are
            critical for configuring DDR3 DIMMs and ensuring stable and reliable
            operation, especially when booting from an unpowered state such as
            ACPI S4 (hibernation), S5 (soft off), or G3 (mechanical off).

            \subsubsection{DDR3 initialization procedure}

                DDR3 initialization is a multi-step process that prepares
                both the memory controllers and the DIMMs for operation. This
                initialization is essential to set up the memory configuration
                and to ensure that the memory subsystem operates correctly
                under various conditions.

                \begin{itemize}
                    \item \textbf{Enable DRAM initialization}: The process
                        begins by
                        enabling DRAM initialization. This is done
                        by setting the \path{EnDramInit} bit in
                        the \path{D18F2x7C_dct} register to 1. The
                        \path{D18F2x7C_dct} register is a specific
                        configuration register within the memory
                        controller that controls various aspects of the
                        DRAM initialization process. Enabling this bit
                        initiates the sequence of operations required to
                        prepare the memory for use. After setting this bit,
                        the system waits for 200 microseconds to allow the
                        initialization command to propagate and stabilize.

                    \item \textbf{Deassert memory reset}: Next, the memory
                        reset
                        signal, known as \path{MemRstX}, is deasserted
                        by setting the \path{DeassertMemRstX} bit in the
                        \path{D18F2x7C_dct} register to 1. Deasserting
                        \path{MemRstX} effectively takes the memory
                        components out of their reset state, allowing them
                        to begin normal operation. The system then waits
                        for an additional 500 microseconds to ensure that
                        the memory reset is fully deasserted and the memory
                        components are stable.

                    \item \textbf{Assert clock enable (CKE)}: The next
                        step involves asserting the clock enable signal, known as
                        `CKE`, by setting the \path{AssertCke} bit in the
                        \path{D18F2x7C_dct} register to 1. The \path{CKE}
                        signal is critical because it enables the clocking
                        of the DRAM modules, allowing them to synchronize
                        with the memory controller. The system must wait
                        for 360 nanoseconds after asserting \path{CKE}
                        to ensure that the clocking is correctly established.

                    \item \textbf{Registered DIMMs and LRDIMMs initialization}:
                        For systems using registered DIMMs (RDIMMs) or Load
                        Reduced DIMMs (LRDIMMs), additional initialization
                        steps are necessary. RDIMMs and LRDIMMs have
                        buffering mechanisms that reduce electrical loading
                        and improve signal integrity, especially in systems
                        with multiple memory modules. During initialization,
                        the BIOS programs the \path{ParEn} bit in the
                        \path{D18F2x90_dct} register based on whether
                        the DIMM is buffered or unbuffered. For RDIMMs,
                        specific Register Control (RC) commands, such as RC0
                        through RC7, are sent to initialize the memory module's
                        control registers. Similarly, LRDIMMs require a series
                        of Flexible Register Control (FRC) commands, such as
                        F0RC and F1RC, to initialize their internal registers
                        according to the manufacturer’s specifications.

                    \item \textbf{Mode Register Set (MRS)}: The initialization
                        process also involves sending Mode Register Set
                        (MRS) commands. These commands are used to configure
                        various operational parameters of the DDR3 memory
                        modules, such as burst length, latency timings,
                        and operating modes. Each MRS command targets a
                        specific mode register within the memory module,
                        and the exact sequence of commands is crucial for
                        setting up the DIMMs according to the system’s
                        requirements and the DIMM manufacturer’s guidelines.
                \end{itemize}

            \subsubsection{ZQ calibration process}

                ZQ calibration is a key step in DDR3 initialization,
                responsible for calibrating the output driver impedance and
                on-die termination (ODT) resistance of the DDR3 modules. Proper
                impedance matching is essential for maintaining signal
                integrity, reducing signal reflections, and ensuring reliable
                data communication between the memory controller and the
                memory modules. It is important to note that ZQ calibration
                is done directly by the memory controller, and that the firmware
                is simply triggering it.

                \begin{itemize}
                    \item \textbf{Sending ZQCL commands}: The BIOS initiates
                    ZQ calibration by sending two ZQCL (ZQ Calibration Long)
                    commands to each DDR3 DIMM. ZQCL commands instruct the
                    memory module to perform a long calibration cycle, during
                    which the module adjusts its output driver impedance and
                    ODT resistance to match the desired target impedance. This
                    process compensates for variations due to manufacturing
                    differences, voltage fluctuations, and temperature
                    changes. To send a ZQCL command, the BIOS programs the
                    \path{SendZQCmd} bit in the \path{D18F2x7C_dct}
                    register to 1 and sets the \path{MrsAddress[10]} bit to 1,
                    indicating that the ZQCL command should be sent to the
                    memory module.

                    \item \textbf{Calibration timing}: After sending the
                    ZQCL command, the system must wait for 512 memory clock
                    cycles (MEMCLKs) to allow the calibration process to
                    complete. During this time, the memory module adjusts
                    its internal impedance to ensure it matches the specified
                    target impedance. This timing is critical, as inadequate
                    wait times could result in incomplete or inaccurate
                    calibration, leading to signal integrity issues and
                    potential data errors.

                    \item \textbf{Finalization of initialization}: Once the
                    ZQ calibration is complete, the BIOS deactivates the DRAM
                    initialization process by setting the \path{EnDramInit}
                    bit in the \path{D18F2x7C_dct} register to 0. For
                    LRDIMMs, additional configuration steps are required to
                    finalize the initialization process. These steps include
                    programming the DCT registers to monitor for errors and
                    ensure that the LRDIMMs are operating correctly.
                \end{itemize}

            \subsubsection{Write leveling process}

                The BIOS and Kernel Developer Guide (BKDG) provides
                information on the write leveling process, which is
                essential for ensuring correct data alignment during write
                operations in DDR3 memory systems. Write leveling is
                particularly crucial in systems utilizing a fly-by topology,
                where timing skew between the clock and data signals can
                introduce significant challenges. This kind of algorithms
                were not necessary for DDR2, for example.
                If the target operating
                frequency is higher than the lowest supported MEMCLK frequency,
                the BIOS must perform multiple passes to achieve proper write
                leveling. The MEMCLK is the clock signal that synchronizes the
                communication between the memory controller and the memory
                modules. \\

                During each pass, the memory subsystem is configured for a
                progressively higher operating frequency:

                \begin{itemize}
                    \item \textbf{Pass 1:} The memory subsystem is configured
                        for the lowest supported MEMCLK, ensuring that initial
                        timing adjustments are made under the most stable
                        conditions.
                    \item \textbf{Pass 2:} The subsystem is then adjusted for
                        the second-lowest MEMCLK, gradually increasing the
                        operating frequency while fine-tuning the alignment of
                        the DQS and CK signals.
                    \item \textbf{Pass N:} This process continues until the
                        highest MEMCLK supported by the system is reached,
                        ensuring that the memory operates reliably at its
                        maximum speed.
                \end{itemize}

                This step-wise configuration ensures that the memory system is
                stable across all supported operating frequencies, minimizing
                the risk of timing errors during write operations, especially
                as frequencies increase and timing margins become tighter. The
                configuration process varies depending on whether the DIMM is
                a Registered DIMM (RDIMM) or an Unregistered DIMM (UDIMM).
                RDIMMs include an additional buffer to improve signal integrity,
                which is particularly important in systems with multiple DIMMs.
                The steps common to both types include a preparation with the
                DDR3 Mode Register Commands
                (see fig. \ref{fig:ddr3_state_machine}).
                For RDIMMs, a 4-rank module is treated as two
                separate DIMMs, where each rank is essentially a separate memory
                module within the same DIMM. The first two ranks are the primary
                target for the initial configuration. The remaining two ranks
                are treated as non-target and are configured separately. \\

                Mode registers in DDR3
                memory are used to configure various operational parameters such
                as latency settings, burst length, and write leveling. One of
                the key mode registers is \path{MR1_dct}, which is specific to
                DDR3 and controls certain features of the memory module,
                including write leveling. \path{MR1_dct} is used to enable or
                disable specific functions such as write leveling and output
                driver settings. The \path{dct} suffix refers to the Data
                Control Timing that is specific to this register's function in
                managing the timing and control of data operations within the
                memory module. \\

                Then, these steps are followed, still common to both RDIMMs and
                UDIMMs:

                \begin{itemize}
                    \item \textbf{Step 1A: Output Driver and ODT configuration
                                  for target DIMM:}
                        \begin{itemize}
                            \item For the first rank (target):
                                \begin{itemize}
                                    \item Set \path{MR1_dct[1:0][Level] = 1}
                                        to enable write leveling.
                                    \item Set \path{MR1_dct[1:0][Qoff] = 0}
                                        to ensure the output drivers are active.
                                \end{itemize}
                            \item For other ranks:
                                \begin{itemize}
                                    \item Set \path{MR1_dct[1:0][Level] = 1}
                                        to prepare for write leveling.
                                    \item Set \path{MR1_dct[1:0][Qoff] = 1}
                                        to deactivate the output drivers for
                                        ranks that are not currently being
                                        leveled.
                                \end{itemize}
                            \item If there are two or more DIMMs per channel,
                                or if there is one DIMM per three channels:
                                \begin{itemize}
                                    \item Program the target rank’s
                                    \path{RttNom} (nominal termination
                                    resistance value) for \path{RttWr}
                                    termination, which helps in managing signal
                                    integrity during the write process by
                                    ensuring the correct impedance matching.
                                \end{itemize}
                        \end{itemize}

                    \item \textbf{Step 1B: Configure non-target RttNom to normal
                                  operation:}
                        \begin{itemize}
                            \item After the initial configuration, the
                                \path{RttNom} values for the non-target ranks
                                are set to their normal operating states.
                            \item A wait time of 40 MEMCLKs is observed to
                                ensure the configuration settings are stable
                                before proceeding.
                        \end{itemize}

                    \item \textbf{Step 3: PHY configuration:}
                        \begin{itemize}
                            \item The PHY is then configured to measure and
                                adjust the timing delays accurately for each
                                data lane. The PHY layer is responsible for
                                converting the signals from the memory
                                controller into a form that can be transmitted
                                over the physical connections to the memory
                                modules.
                        \end{itemize}

                    \item \textbf{Step 4: Perform write leveling:}
                        \begin{itemize}
                            \item The actual write leveling process is executed,
                                where the DQS signal timing is adjusted to
                                ensure it aligns perfectly with the CK signal at
                                the memory module’s pins, ensuring that data is
                                written accurately.
                        \end{itemize}

                    \item \textbf{Step 5: Disable PHY configuration
                                  post-measurement:}
                        \begin{itemize}
                            \item After completing the write leveling process,
                            the PHY configuration is disabled to stop further
                            timing measurements and adjustments, locking in the
                            calibrated settings.
                        \end{itemize}

                    \item \textbf{Step 6: Program the DIMM to normal operation:}
                        \begin{itemize}
                            \item Finally, the DIMM is reprogrammed to its
                            normal operational state, resetting \path{Qoff}
                            and \path{Level} to \path{0} to conclude the
                            write leveling process and return to standard
                            operation.
                        \end{itemize}
                \end{itemize}

                For each DIMM, the BIOS must calculate the coarse and fine
                delays for each lane in the DQS Write Timing register:

                \begin{itemize}
                    \item \textbf{Coarse Delay Calculation:} This involves
                        setting a basic delay based on a seed value specific to
                        the platform. The seed value is determined during
                        initial system configuration and serves as a starting
                        point for further delay adjustments.
                    \item \textbf{Critical Delay Determination:} The minimum of
                        the coarse delays for each lane and DIMM is considered
                        the critical delay. This delay is crucial for ensuring
                        that all data lanes are correctly synchronized.
                    \item \textbf{Platform-Specific Seed:} The seed ranges
                        between -1.20ns and +1.20ns, providing a small
                        adjustment range to fine-tune the timing based on the
                        specific characteristics of the platform. This seed
                        value can differ for the first pass compared to
                        subsequent passes, allowing for incremental adjustments
                        as the system stabilizes.
                \end{itemize}

    \section{Current implementation and potential improvements}

        \subsection{Current implementation in coreboot on the KGPE-D16}

            In this part as for the rest of this document, we're basing our
            study on the 4.11 version of \textit{coreboot} \cite{coreboot_4_11},
            which is the last version that supported the ASUS KGPE-D16
            mainboard. \\

            The process starts in
            \path{src/mainboard/asus/kgpe-d16/romstage.c}, in the
            \path{cache_as_ram_main} function by calling
            \path{fill_mem_ctrl} from
            \path{src/northbridge/amd/amdfam10/raminit_sysinfo_in_ram.c}
            (lst. \ref{lst:fill_mem_ctrl}).
            At this current step, only the BSC is running the firmware code.
            This function iterates over all memory controllers (one per
            node) and initializes their corresponding structures with the
            system information needed for the RAM to function. This includes
            the addresses of PCI nodes (important for DMA operations) and
            SPD addresses, which are internal ROMs in each memory slot
            containing crucial information for detecting and initializing
            memory modules. \\

            \begin{listing}[H]
                \begin{adjustwidth}{0.5cm}{0.5cm}
                    \inputminted{c}{
                        listings/src_northbridge_amd_amdfam10_raminit_sysinfo_in_ram.c}
                \end{adjustwidth}
                \caption{
                    \protect\path{fill_mem_ctrl()}, extract from
                    \protect\path{src/northbridge/amd/amdfam10/raminit_sysinfo_in_ram.c}}
                \label{lst:fill_mem_ctrl}
            \end{listing}

            If successful, the system posts codes \path{0x3D} and then
            \path{0x40}. The \path{raminit_amdmct} function from
            \path{src/northbridge/amd/amdfam10/raminit_amdmct.c} is then
            called. This function, in turn, calls \path{mctAutoInitMCT_D}
            (lst. \ref{lst:mctAutoInitMCT_D_1}) from
            \path{src/northbridge/amd/amdmct/mct_ddr3/mct_d.c},
            which is responsible for the initial memory initialization,
            predominantly written by Raptor Engineering. \\

            At this stage, it is assumed that memory has been pre-mapped
            contiguously from address 0 to 4GB and that the previous code
            has correctly mapped non-cacheable I/O areas below 4GB for the
            PCI bus and Local APIC access for processor cores. \\

            The following prerequisites must be in place from the previous
            steps:

            \begin{itemize}
                \item The HyperTransport bus configured, and its speed is
                    correctly set.
                \item The SMBus controller is configured.
                \item The BSP is in unreal mode.
                \item A stack is set up for all cores.
                \item All cores are initialized at a frequency of 2GHz.
                \item If we were using saved values, the NVRAM would have been
                    verified with checksums.
            \end{itemize}

            The memory controller for the BSP is queried to check if it can
            manage ECC memory, which is a type of memory that includes
            error-correcting code to detect and correct common types of data
            corruption (lst. \ref{lst:mctAutoInitMCT_D_2}).

            For each node available in the system, the memory controllers
            are identified and initialized using a \path{DCTStatStruc}
            structure defined in
            \path{src/northbridge/amd/amdmct/mct_ddr3/mct_d.h}. This
            structure contains all necessary fields for managing a memory
            module. The process includes:

            \begin{itemize}
                \item Retrieving the corresponding field in the sysinfo
                    structure for the node.
                \item Clearing fields with \path{zero}.
                \item Initializing basic fields.
                \item Initializing the controller linked to the current node.
                \item Verifying the presence of the node (checking if the
                    processor associated with this controller is present).
                    If yes, the SMBus is informed.
                \item Pre-initializing the memory module controller for this
                    node using \path{mct_preInitDCT}.
            \end{itemize}

            The memory modules must be initialized. All modules present on
            valid nodes are configured with 1.5V voltage
            (lst. \ref{lst:mctAutoInitMCT_D_3}). The ZQ calibration
            is triggered at this stage. \\

            Now, present memory modules are detected using \path{mct_initDCT}
            (lst. \ref{lst:mctAutoInitMCT_D_4}). The memory modules existence
            is checked and the machine halts immediately after displaying a
            message if there is no memory.
            \textit{coreboot} waits for all modules to be available using
            \path{SyncDCTsReady_D}. \\

            The firmware maps the physical memory address ranges into the
            address space with \path{HTMemMapInit_D} as contiguously as possible
            while also constructing the physical memory map. If there is an
            area occupied by something else, it is ignored, and a memory hole is
            created. \\

            Mapping the address ranges into the cache is done with
            \path{CPUMemTyping_D} either as WriteBack (cacheable) or
            Uncacheable, depending on whether the area corresponds to physical
            memory or a memory hole. \\

            The external northbridge is notified of this new memory
            configuration. \\

            The \textit{coreboot} code compensates for the delay between DQS
            and DQ signals, as well as between CMD and DQ. This is handled by
            the \path{DQSTiming_D} function (lst. \ref{lst:mctAutoInitMCT_D_5}).
            The initialization can be done again if needed after that, otherwise
            the channels and nodes are interleaved and ECC is enabled (if
            supported by every module). \\

            After that being done, the DRAM can be mapped into the address
            space with cacheability, and the init process finishes with
            validation of every populated DCT node
            (lst. \ref{lst:mctAutoInitMCT_D_6}). \\

            Finally, if the RAM is of the ECC type, error-correcting codes
            are enabled, and the function ends by activating power-saving
            features if requested by the user. \\

            \subsubsection{Details on the DQS training function}

                The \path{DQSTiming_D} function is a critical part of the
                firmware responsible for initializing and training the system's
                memory.
                The function primarily handles the DQS timing, which is
                essential for ensuring data integrity and synchronization
                between the memory controller and the DRAM. Proper DQS training
                is crucial to align the data signals correctly with the clock
                signals.

                The function begins by declaring local variables, which are
                used throughout the function for various operations. It also
                includes an early exit condition to bypass DQS training if a
                specific status flag (\path{GSB_EnDIMMSpareNW}) is set,
                indicating that a DIMM spare feature is enabled
                (lst. \ref{lst:var_decl_and_exit}). These spare DIMMs are not
                used for normal memory operations but are kept in reserve for
                redundancy. \\

                \begin{listing}[H]
                    \begin{adjustwidth}{0.5cm}{0.5cm}
                        \begin{minted}[linenos]{c}
if (pMCTstat->GStatus & (1 << GSB_EnDIMMSpareNW)) {
    return;
}
                        \end{minted}
                    \end{adjustwidth}
                    \caption{Early exit check,
                             extract from the
                             \protect\path{DQSTiming_D} function in
                             \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mct_d.c}}
                    \label{lst:var_decl_and_exit}
                \end{listing}

                Next, the function initializes the TCWL (CAS Write Latency)
                offset to zero for each node and DCT.
                This ensures that the memory write latency is properly aligned
                before the DQS training begins
                (lst. \ref{lst:set_tcwl_offset}). \\

                \begin{listing}[H]
                    \begin{adjustwidth}{0.5cm}{0.5cm}
                        \begin{minted}[linenos]{c}
for (Node = 0; Node < MAX_NODES_SUPPORTED; Node++) {
    uint8_t dct;
    struct DCTStatStruc *pDCTstat;
    pDCTstat = pDCTstatA + Node;
    for (dct = 0; dct < 2; dct++)
        pDCTstat->tcwl_delay[dct] = 0;
}
                        \end{minted}
                    \end{adjustwidth}
                    \caption{Setting initial TCWL offset to zero for all nodes and DCTs,
                             extract from the
                             \protect\path{DQSTiming_D} function in
                             \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mct_d.c}}
                    \label{lst:set_tcwl_offset}
                \end{listing}

                A retry mechanism is introduced to handle potential errors
                during DQS training and the pre-training function are called
                (lst. \ref{lst:retry_pre_training}). \\

                \begin{listing}[H]
                    \begin{adjustwidth}{0.5cm}{0.5cm}
                        \begin{minted}[linenos]{c}
retry_dqs_training_and_levelization:
    nv_DQSTrainCTL = !allow_config_restore;

    mct_BeforeDQSTrain_D(pMCTstat, pDCTstatA);
    phyAssistedMemFnceTraining(pMCTstat, pDCTstatA, -1);
                        \end{minted}
                    \end{adjustwidth}
                    \caption{Retry mechanism initialization and pre-training operations,
                             extract from the
                             \protect\path{DQSTiming_D} function in
                             \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mct_d.c}}
                    \label{lst:retry_pre_training}
                \end{listing}

                For AMD's Fam15h processors, additional PHY compensation is
                needed for each node and valid DCT
                (lst. \ref{lst:phy_compensation_init}). This is necessary to
                fine-tune the electrical characteristics of the memory
                interface. For more information about the PHY training, see
                the earlier sections about RAM training algorithm. \\

                \begin{listing}[H]
                    \begin{adjustwidth}{0.5cm}{0.5cm}
                        \begin{minted}[linenos]{c}
for (Node = 0; Node < MAX_NODES_SUPPORTED; Node++) {
    pDCTstat = pDCTstatA + Node;
    if (pDCTstat->NodePresent) {
        if (pDCTstat->DIMMValidDCT[0])
            InitPhyCompensation(pMCTstat, pDCTstat, 0);
        if (pDCTstat->DIMMValidDCT[1])
            InitPhyCompensation(pMCTstat, pDCTstat, 1);
    }
}
                        \end{minted}
                    \end{adjustwidth}
                    \caption{PHY compensation initialization,
                             extract from the
                             \protect\path{DQSTiming_D} function in
                             \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mct_d.c}}
                    \label{lst:phy_compensation_init}
                \end{listing}

                Before proceeding with the main DQS training, the function
                invokes a hook function that allows for additional
                configurations or custom operations:
                \path{mctHookBeforeAnyTraining}. \\

                The \path{nv_DQSTrainCTL} variable is
                set based on the \path{allow_config_restore} parameter,
                determining whether to restore a previous configuration or
                proceed with fresh training. This is however not working on the
                current implementation of ASUS KGPE-D16 firmware
                (lst. \ref{lst:mctAutoInitMCT_D_fixme}).
                If \path{nv_DQSTrainCTL} indicates that fresh training should
                proceed, the function performs the main DQS training in multiple
                passes, including receiver enable training with
                \path{TrainReceiverEn_D}, write leveling with
                \path{mct_WriteLevelization_HW}, DQS position
                training with \path{mct_TrainDQSPos_D} and the maximum read
                latency calculation with \path{TrainMaxRdLatency_En_D}
                (lst. \ref{lst:dqs_training_process}).
                Write leveling is done in two passes, with a DQS receiver
                training between and another pass of receiver training after.
                After that, a DQS position training is done and the process
                finished with the maximum read latency, i.e the delay between
                the request for data and the delivery of that data by the DRAM.
                \\

                \begin{listing}[H]
                    \begin{adjustwidth}{0.5cm}{0.5cm}
                        \begin{minted}[linenos]{c}
if (nv_DQSTrainCTL) {
    mct_WriteLevelization_HW(pMCTstat, pDCTstatA, FirstPass);
    TrainReceiverEn_D(pMCTstat, pDCTstatA, FirstPass);
    mct_WriteLevelization_HW(pMCTstat, pDCTstatA, SecondPass);

    /* TODO: Determine why running TrainReceiverEn_D in SecondPass mode yields
    * less stable training values than when run in FirstPass mode as in the HACK
    * below.*/
    TrainReceiverEn_D(pMCTstat, pDCTstatA, FirstPass);
    mct_TrainDQSPos_D(pMCTstat, pDCTstatA);
     ...
    TrainMaxRdLatency_En_D(pMCTstat, pDCTstatA);
} else {
    mct_WriteLevelization_HW(pMCTstat, pDCTstatA, FirstPass);
    mct_WriteLevelization_HW(pMCTstat, pDCTstatA, SecondPass);
#if CONFIG(HAVE_ACPI_RESUME)
    printk(BIOS_DEBUG, "mctAutoInitMCT_D: Restoring DIMM training configuration"
                        "from NVRAM\n");
    if (restore_mct_information_from_nvram(1) != 0)
        printk(BIOS_CRIT, "%s: ERROR: Unable to restore DCT configuration from"
                           "NVRAM\n", __func__);
#endif
    exit_training_mode_fam15(pMCTstat, pDCTstatA);
    pMCTstat->GStatus |= 1 << GSB_ConfigRestored;"
}
                        \end{minted}
                    \end{adjustwidth}
                    \caption{Main DQS training process in multiple passes,
                             extract from the
                             \protect\path{DQSTiming_D} function in
                             \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mct_d.c}}
                    \label{lst:dqs_training_process}
                \end{listing}

                The function checks for any errors during the DQS training. If
                errors are detected, it may request a retrain, reset certain
                parameters, and restart the training process and even restart
                the whole system if needed (lst. \ref{lst:error_handling}).
                If the training process it to be restarted, the firmware
                sets the DIMMs frequencies to minimum and applies timing changes
                to DIMMs before jumping to the retry label
                (lst. \ref{lst:retry_pre_training}). \\

                Once the training is successfully completed without errors, the
                function finalizes the process by setting the maximum read
                latency and exiting the training mode. For systems with
                \path{allow_config_restore} enabled, it restores the previous
                configuration from NVRAM instead of performing a fresh training
                (lst. \ref{lst:dqs_training_process}). \\

                Finally, the function performs a cleanup operation specific to
                Fam15h processors, where it switches the DCT control register
                as required by a known erratum from AMD for the BKDG
                (Erratum 505) \cite{amd_fam15h_revision_guide}.
                This is followed by a post-training hook that
                allows for any additional necessary actions
                (lst. \ref{lst:post_training_cleanup}). \\

                \begin{listing}[htpb]
                    \begin{adjustwidth}{0.5cm}{0.5cm}
                        \begin{minted}[linenos]{c}
for (Node = 0; Node < MAX_NODES_SUPPORTED; Node++) {
    pDCTstat = pDCTstatA + Node;
    if (pDCTstat->NodePresent) {
        fam15h_switch_dct(pDCTstat->dev_map, 0);
    }
}

/* FIXME - currently uses calculated value
 * TrainMaxReadLatency_D(pMCTstat, pDCTstatA); */
mctHookAfterAnyTraining();
                        \end{minted}
                    \end{adjustwidth}
                    \caption{Post-training cleanup and final hook execution,
                             extract from the
                             \protect\path{DQSTiming_D} function in
                             \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mct_d.c}}
                    \label{lst:post_training_cleanup}
                \end{listing}

            \subsubsection{Details on the write leveling implementation}

                The \path{WriteLevelization_HW} function is responsible for
                performing hardware-level write leveling on DRAM modules during
                the memory initialization process. Write leveling ensures that
                the DQS signals are correctly aligned with the clock signals,
                preventing timing mismatches during write operations. \\

                The function begins by initializing pointers to key data
                structures, linking the memory controller (MCT) and DRAM
                controller timing (DCT) data for subsequent operations. \\

                Auto-refresh and short ZQ calibration are temporarily disabled
                to prevent interference during the critical timing adjustments
                of write leveling.
                The memory controller is prepared for write leveling by
                configuring necessary parameters with \path{PrepareC_MCT},
                then the main operation can begin. \\

                In the first pass (lst. \ref{lst:write_level_first_pass}),
                the function repeatedly attempts to align
                the DQS signals with \path{PhyWLPass1}, retrying if invalid
                values are detected. This phase ensures basic alignment for
                further fine-tuning. The function retries up to 8 times if it
                detects invalid timing values. \\

                During the second pass (lst. \ref{lst:write_level_second_pass}),
                the function first checks if the target memory frequency
                (\path{TargetFreq}) is higher than the minimum memory clock
                frequency stored in the non-volatile bits
                (\path{NV_MIN_MEMCLK}). If so, the memory frequency is
                incrementally adjusted toward the final target f requency.
                This step-by-step approach is crucial, especially for AMD Fam15h
                processors, where the frequency must be gradually stepped up to
                avoid instability. \\

                For each frequency step, the write leveling process is
                recalibrated by invoking the \path{PhyWLPass2} function. This
                function adjusts the DQS timing for each data channel (DCT) and
                validates the results. The function retries up to 8 times if it
                detects invalid timing values. The global status
                (\path{global_phy_training_status}) aggregates the results of
                each step, tracking any persistent issues. \\

                The \path{PhyWLPass1} and \path{PhyWLPass2} function relyon
                \path{AgesaHwWlPhase1}, \path{AgesaHwWlPhase2} and
                \path{AgesaHwWlPhase3} for this. \\

                Once the target frequency is reached and all write leveling
                adjustments are made, the final timing values are stored.
                The gross and fine delays from the previous passes are copied
                into the final pass structures. This ensures that the DQS
                timings are consistent and stable across all data channels. \\

                If any issues persist after retries, the function logs a
                warning. This indicates that the system may continue to operate,
                but with a potential risk of instability due to imperfect
                write leveling calibration. \\

                After leveling, the function re-enables auto-refresh and short
                ZQ calibration, ensuring the memory subsystem is correctly
                configured for normal operation. \\

                \begin{listing}[htpb]
                    \begin{adjustwidth}{0.5cm}{0.5cm}
                        \begin{minted}[linenos]{c}
if (Pass == FirstPass) {
    timeout = 0;
    do {
        status = 0;
        timeout++;
        status |= PhyWLPass1(pMCTstat, pDCTstat, 0);
        status |= PhyWLPass1(pMCTstat, pDCTstat, 1);
        if (status)
            printk(BIOS_INFO, "%s: Retrying write levelling due to invalid "
                              "value(s) detected in first phase\n", __func__);
    } while (status && (timeout < 8));
    if (status)
        printk(BIOS_INFO, "%s: Uncorrectable invalid value(s) detected in first "
                          "phase of write levelling\n", __func__);
}
                        \end{minted}
                    \end{adjustwidth}
                    \caption{Write leveling (first pass),
                             extract from the
                             \protect\path{WriteLevelization_HW} function in
                             \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mcthwl.c}}
                    \label{lst:write_level_first_pass}
                \end{listing}

                    The detailled write leveling process is divided into three
                    distinct phases, each managed by a specific function:
                    \path{AgesaHwWlPhase1}, \path{AgesaHwWlPhase2}, and
                    \path{AgesaHwWlPhase3} from \path{mcthwl.c}.
                    These phases work together to
                    fine-tune the timing delays (gross and fine) for each byte
                    lane, ensuring reliable data transmission. \\

                    The write leveling process begins by selecting the target
                    DIMM. This is accomplished by programming the
                    \path{TrDimmSel} register to ensure that the subsequent
                    operations apply to the correct DIMM
                    (lst. \ref{lst:target_dimm_selection}) \\

                    \begin{listing}
                        \begin{adjustwidth}{0.5cm}{0.5cm}
                            \begin{minted}[linenos]{c}
set_DCT_ADDR_Bits(pDCTData, dct, pDCTData->NodeId, FUN_DCT,
        DRAM_ADD_DCT_PHY_CONTROL_REG, TrDimmSelStart,
        TrDimmSelEnd, (u32)dimm);
                            \end{minted}
                        \end{adjustwidth}
                    \caption{Target DIMM selection for write leveling,
                        extract from
                        \protect\path{AgesaHwWlPhase1} function in
                        \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mhwlc_d.c}}
                    \label{lst:target_dimm_selection}
                    \end{listing}

                    In the case of x4 DIMMs, which are common in high-density
                    memory configurations, write leveling must be performed
                    separately for each nibble (4-bit group). The function
                    checks if x4 DIMMs are present and, if so, prepares to train
                    both nibbles (lst. \ref{lst:x4_dimm_handling}). \\

                    \begin{listing}
                        \begin{adjustwidth}{0.5cm}{0.5cm}
                            \begin{minted}[linenos]{c}
train_both_nibbles = 0;
if (pDCTstat->Dimmx4Present)
    if (is_fam15h())
        train_both_nibbles = 1;
                            \end{minted}
                        \end{adjustwidth}
                    \caption{Handling of x4 DIMMs and nibble training,
                        extract from
                        \protect\path{AgesaHwWlPhase1} function in
                        \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mhwlc_d.c}}
                    \label{lst:x4_dimm_handling}
                    \end{listing}

                    The DIMMs are prepared for write leveling by issuing Mode
                    Register (MR) commands. These commands configure the DIMMs
                    to enter a state where write leveling can be performed
                    (lst. \ref{lst:prepare_dimms}). \\

                    \begin{listing}
                    \begin{adjustwidth}{0.5cm}{0.5cm}
                    \begin{minted}[linenos]{c}
prepareDimms(pMCTstat, pDCTstat, dct, dimm, TRUE);
                    \end{minted}
                    \end{adjustwidth}
                    \caption{Preparing DIMMs for write leveling,
                        extract from
                        \protect\path{AgesaHwWlPhase1} function in
                        \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mhwlc_d.c}}
                    \label{lst:prepare_dimms}
                    \end{listing}

                    The \path{procConfig} function is called to configure the
                    processor's DDR PHY (Physical Layer) for write leveling.
                    This configuration includes setting initial seed values for
                    gross and fine delays, which are essential for the
                    subsequent timing adjustments. \\

                    \path{procConfig} generates initial seed values
                    (lst. \ref{lst:seed_generation}) for gross
                    and fine delays. These seeds are calculated based on several
                    factors:

                    \begin{itemize}
                        \item \textbf{Processor Type:} For Fam15h processors,
                            specific tables from the Fam15h BKDG \cite{BKDG} are
                            referenced to select appropriate seed values for
                            different package types (e.g., Socket G34, Socket
                            C32).
                        \item \textbf{DIMM Type:} The seed values are adjusted
                            based on whether the RDIMMs are registered or
                            load-reduced, with different base values used for
                            these configurations.
                        \item \textbf{Memory Clock Frequency:} The seeds are
                            further adjusted based on the current memory clock
                            frequency (\path{MemClkFreq}), ensuring that the
                            timing is correct for the operating speed of the
                            memory.
                    \end{itemize}

                    The calculated seed values are then scaled to the minimum
                    supported memory frequency and stored in the
                    \path{WLSeedGrossDelay} and \path{WLSeedFineDelay} arrays
                    for each byte lane. \\

                    \begin{listing}
                    \begin{adjustwidth}{0.5cm}{0.5cm}
                    \begin{minted}[linenos]{c}
Seed_Total = (int32_t) (((((int64_t) Seed_Total) *
    fam15h_freq_tab[MemClkFreq] * 100) / (mctGet_NVbits(NV_MIN_MEMCLK) * 100)));

Seed_Gross = (Seed_Total >> 5) & 0x1f;
Seed_Fine = Seed_Total & 0x1f;
                    \end{minted}
                    \end{adjustwidth}
                    \caption{Seed generation,
                        extract from
                        \protect\path{procConfig} function in
                        \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mhwlc_d.c}}
                    \label{lst:seed_generation}
                    \end{listing}

                    Write leveling is initiated by enabling the
                    \path{WrtLvTrEn} bit. This allows the DDR PHY to begin
                    adjusting the DQS signals relative to the clock signals
                    (lst. \ref{lst:initiate_write_leveling}). \\

                    \begin{listing}
                    \begin{adjustwidth}{0.5cm}{0.5cm}
                    \begin{minted}[linenos]{c}
set_DCT_ADDR_Bits(pDCTData, dct, pDCTData->NodeId, FUN_DCT,
        DRAM_ADD_DCT_PHY_CONTROL_REG, WrtLvTrEn, WrtLvTrEn, 1);
                    \end{minted}
                    \end{adjustwidth}
                    \caption{Initiating write leveling training,
                        extract from
                        \protect\path{AgesaHwWlPhase1} function in
                        \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mhwlc_d.c}}
                    \label{lst:initiate_write_leveling}
                    \end{listing}

                    If the DIMM is not x4, the function skips the nibble
                    training loop, as it is unnecessary
                    (lst. \ref{lst:exit_non_x4}). \\

                    \begin{listing}
                    \begin{adjustwidth}{0.5cm}{0.5cm}
                    \begin{minted}[linenos]{c}
if ((pDCTstat->Dimmx4Present & (1 << (dimm + dct))) == 0)
    break;
                    \end{minted}
                    \end{adjustwidth}
                    \caption{Exit for non-x4 DIMMs,
                        extract from
                        \protect\path{AgesaHwWlPhase2} function in
                        \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mhwlc_d.c}}
                    \label{lst:exit_non_x4}
                    \end{listing}

                    After a delay to allow the leveling process to stabilize,
                    the function reads the gross and fine delay values from the
                    relevant registers and stores them
                    (lst. \ref{lst:finalize_write_leveling}). These values
                    represent the initial timing adjustments necessary for
                    correct DQS alignment. \\

                    \begin{listing}
                    \begin{adjustwidth}{0.5cm}{0.5cm}
                    \begin{minted}[linenos]{c}
for (ByteLane = 0; ByteLane < lane_count; ByteLane++) {
    getWLByteDelay(pDCTstat, dct, ByteLane, dimm, pass, nibble, lane_count);
}
                    \end{minted}
                    \end{adjustwidth}
                    \caption{Reading and storing delay values after write leveling,
                        extract from
                        \protect\path{AgesaHwWlPhase2} function in
                        \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mhwlc_d.c}}
                    \label{lst:finalize_write_leveling}
                    \end{listing}

            \subsubsection{Details on the DQS position training function}

                The DQS position training is a crucial step in the memory
                initialization process, ensuring that both read and write
                operations are correctly aligned with the clock signal. \\

                The function \path{TrainDQSRdWrPos_D_Fam15} orchestrates this
                process by iterating over memory lanes and adjusting timing
                parameters to find optimal settings
                (lst. \ref{lst:dqs_train_init}). It is called by
                \path{mct_TrainDQSPos_D}. \\

                The function begins by initializing several variables and
                settings necessary for the training process. These include:

                \begin{itemize}
                    \item \path{Errors}: A variable to track any errors
                        encountered during the training.
                    \item \path{dual_rank}: A flag to indicate whether the
                        current DIMM has two ranks.
                    \item \path{passing_dqs_delay_found}: An array to track
                        whether a passing DQS delay has been found for each lane.
                    \item \path{dqs_results_array}: A multi-dimensional array to
                        store the results of the DQS delay tests across
                        different write and read steps.
                \end{itemize}

                The function then loops over each receiver (loosely associated
                with chip selects) to perform the training for each rank within
                each DIMM. \\

                \begin{listing}
                    \begin{adjustwidth}{0.5cm}{0.5cm}
                        \begin{minted}[linenos]{c}
for (Receiver = receiver_start; Receiver < receiver_end; Receiver++) {
    dimm = (Receiver >> 1);
    ...
    if (!mct_RcvrRankEnabled_D(pMCTstat, pDCTstat, dct, Receiver)) {
        continue;
    }
                        \end{minted}
                    \end{adjustwidth}
                \caption{Looping over each receiver,
                    extract from
                    \protect\path{TrainDQSRdWrPos_D_Fam15} function in
                    \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mctdqs_d.c}}
                \label{lst:dqs_train_init}
                \end{listing}

                For each lane in the memory channel, the function iterates over
                possible write and read delay values to find the optimal
                configuration (lst. \ref{lst:dqs_train_iteration}).
                This is done by:

                \begin{enumerate}
                    \item Iterating over the write data delay values from the
                        initial value to the initial value plus 1 UI
                        (Unit Interval).
                    \item For each write data delay, iterating over possible
                        read DQS delay values from 0 to 1 UI.
                    \item For each combination of write and read delays, testing
                        the configuration by writing a training pattern to the
                        memory and reading it back to check if it passes or
                        fails.
                \end{enumerate}

                \begin{listing}
                    \begin{adjustwidth}{0.5cm}{0.5cm}
                        \begin{minted}[linenos]{c}
for (current_write_data_delay[lane] = initial_write_dqs_delay[lane];
     current_write_data_delay[lane] < (initial_write_dqs_delay[lane] + 0x20);
     current_write_data_delay[lane]++) {
    ...
    for (current_read_dqs_delay[lane] = 0;
         current_read_dqs_delay[lane] < 0x20;
         current_read_dqs_delay[lane]++) {
        ...
        write_dqs_read_data_timing_registers(
            current_read_dqs_delay, dev, dct, dimm, index_reg);
        read_dram_dqs_training_pattern_fam15(
            pMCTstat, pDCTstat, dct, Receiver, lane, ((check_antiphase == 0)?1:0));
        ...
    }
}
                        \end{minted}
                    \end{adjustwidth}
                \caption{Iteration over write and read delay values for each lane,
                    extract from
                    \protect\path{TrainDQSRdWrPos_D_Fam15} function in
                    \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mctdqs_d.c}}
                \label{lst:dqs_train_iteration}
                \end{listing}

                During each iteration, the results are recorded in the
                \path{dqs_results_array}, which tracks whether the combination
                of write and read delays was successful (pass) or not (fail).
                The results are stored for both the primary rank and, if
                applicable, the secondary rank when dual rank DIMMs are used.
                \\

                After iterating over all possible delay values, the function
                processes the results to determine the best DQS delay settings
                (lst. \ref{lst:dqs_train_results}). \\

                This is done by:

                \begin{itemize}
                    \item Finding the longest consecutive string of passing
                        values for both read and write operations.
                    \item Calculating the center of the passing region and using
                        this as the optimal delay setting.
                    \item If the center of the region is below a threshold,
                        issuing a warning that a negative DQS recovery delay
                        was detected, which could lead to instability.
                \end{itemize}

                \begin{listing}
                    \begin{adjustwidth}{0.5cm}{0.5cm}
                        \begin{minted}[linenos]{c}
if (best_count > 2) {
    uint16_t region_center = (best_pos + (best_count / 2));
    if (region_center < 16) {
        printk(BIOS_WARNING,
            "TrainDQSRdWrPos: negative DQS recovery delay detected!");
        region_center = 0;
    } else {
        region_center -= 16;
    }
    ...
    current_read_dqs_delay[lane] = region_center;
    passing_dqs_delay_found[lane] = 1;
    write_dqs_read_data_timing_registers(
        current_read_dqs_delay, dev, dct, dimm, index_reg);
}
                        \end{minted}
                    \end{adjustwidth}
                \caption{Processing the results to determine the best DQS delay settings,
                    extract from
                    \protect\path{TrainDQSRdWrPos_D_Fam15} function in
                    \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mctdqs_d.c}}
                \label{lst:dqs_train_results}
                \end{listing}

                Finally, the function checks if any lane did not find a valid
                passing region (lst. \ref{lst:dqs_train_finalize}).
                If any lanes failed to find a passing DQS delay,
                the \path{Errors} flag is set, and this error is propagated
                through the \path{pDCTstat->TrainErrors} and
                \path{pDCTstat->ErrStatus} variables.
                \\

                The function returns \path{1} if no errors were encountered,
                and \texttt{0} otherwise, which is unusual. \\

                \begin{listing}
                    \begin{adjustwidth}{0.5cm}{0.5cm}
                        \begin{minted}[linenos]{c}
for (lane = lane_start; lane < lane_end; lane++) {
    if (!passing_dqs_delay_found[lane]) {
        Errors |= 1 << SB_NODQSPOS;
    }
}
pDCTstat->TrainErrors |= Errors;
pDCTstat->ErrStatus |= Errors;
return !Errors;
                        \end{minted}
                    \end{adjustwidth}
                \caption{Final error handling and return value,
                    extract from
                    \protect\path{TrainDQSRdWrPos_D_Fam15} function in
                    \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mctdqs_d.c}}
                \label{lst:dqs_train_finalize}
                \end{listing}

            \subsubsection{Details on the DQS receiver training function}

                In AMD Fam15h G34 processors, the DQS receiver enable training
                is a critical step in ensuring that the memory subsystem operates
                correctly and reliably. This training aligns the DQS signal with
                the clock signal, ensuring proper data capture during memory reads.
                \\

                The DQS receiver enable training algorithm is executed twice:
                first at the lowest supported MEMCLK frequency and then at the
                highest supported MEMCLK frequency. The purpose of this training
                is to fine-tune the timing parameters so that the memory
                controller can reliably read data from the memory modules.
                The algorithm is implemented in the function
                \path{dqsTrainRcvrEn_SW_Fam15} from
                \path{src/northbridge/amd/amdmct/mct_ddr3/mctsrc.c}, which
                orchestrates the
                entire process, called by the \path{mct_TrainRcvrEn_D} function,
                which has been called itself by \path{TrainReceiverEn_D} from
                \path{src/northbridge/amd/amdmct/mct_ddr3/mctdqs_d.c}. \\

                Here, seeds are initial delay values used to set
                up the memory controller's timing parameters. These seeds are
                generated based on the specific characteristics of the memory
                configuration, such as the package type (e.g., G34, C32), the
                type of DIMMs installed (Registered, Load Reduced, etc.), and
                the maximum number of DIMMs that can be installed in a channel.
                \\

                The seed generation is handled by the function
                \path{fam15_receiver_enable_training_seed}. This function
                generates a base seed value for each memory channel, based on
                predefined tables in the BKDG \cite{BKDG}. The base seed values
                are specific to the memory configuration and are adjusted based
                on the type of DIMM and the number of DIMMs in each channel. \\

                The generated seed values are then adjusted
                (lst. \ref{lst:seed_adjustment}) based on the
                operating frequency of the memory (MEMCLK). The adjustment
                scales the seed values to account for the difference between
                the current memory frequency and the minimum supported
                frequency. This ensures that the training can be accurately
                performed across different operating conditions. \\

                \begin{listing}
                \begin{adjustwidth}{0.5cm}{0.5cm}
                \begin{minted}[linenos]{c}
initial_seed = (uint16_t) (((((uint64_t) initial_seed) *
    fam15h_freq_tab[mem_clk] * 100) / (min_mem_clk * 100)));
                \end{minted}
                \end{adjustwidth}
                \caption{Adjusting the seed values based on the operating
                    frequency of the memory,
                    extract from
                    \protect\path{dqsTrainRcvrEn_SW_Fam15} function in
                    \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mctsrc.c}}
                \label{lst:seed_adjustment}
                \end{listing}

                Once the seeds are generated and adjusted, they are used to set
                the initial delay values for the DQS receiver enable training
                (lst. \ref{lst:initial_delay_values}).
                The delay values are split into two components: gross delay and
                fine delay. The gross delay determines the overall timing
                offset, while the fine delay adjusts the timing with finer
                granularity. \\

                \begin{listing}
                \begin{adjustwidth}{0.5cm}{0.5cm}
                \begin{minted}[linenos]{c}
for (lane = 0; lane < lane_count; lane++) {
    seed_gross[lane] = (seed[lane] >> 5) & 0x1f;
    seed_fine[lane] = seed[lane] & 0x1f;

    if (seed_gross[lane] & 0x1)
        seed_pre_gross[lane] = 1;
    else
        seed_pre_gross[lane] = 2;

    // Set the gross delay
    current_total_delay[lane] = ((seed_gross[lane] & 0x1f) << 5);
}
                \end{minted}
                \end{adjustwidth}
                \caption{Setting initial delay values based on the generated
                    seed values,
                    extract from
                    \protect\path{dqsTrainRcvrEn_SW_Fam15} function in
                    \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mctsrc.c}}
                \label{lst:initial_delay_values}
                \end{listing}

                These delay values are then written to the appropriate registers
                to configure the memory controller for the DQS receiver enable
                training. The training is performed in multiple steps,
                iteratively refining the delay values until the DQS signal is
                correctly aligned with the clock signal. \\

                During the initialization phase, the memory controller is
                prepared for training. This includes enabling the training mode,
                configuring the memory channels, and disabling certain features
                such as ECC (Error-Correcting Code) to prevent interference
                during training (lst. \ref{lst:initialization_phase}). \\

                \begin{listing}
                    \begin{adjustwidth}{0.5cm}{0.5cm}
                        \begin{minted}[linenos]{c}
fam15EnableTrainingMode(pMCTstat, pDCTstat, ch, 1);
_DisableDramECC = mct_DisableDimmEccEn_D(pMCTstat, pDCTstat);
                        \end{minted}
                    \end{adjustwidth}
                \caption{Enabling training mode and disabling ECC,
                    extract from
                    \protect\path{dqsTrainRcvrEn_SW_Fam15} function in
                    \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mctsrc.c}}
                \label{lst:initialization_phase}
                \end{listing}

                The training phase is where the actual alignment of the DQS
                signal occurs. The memory controller iterates over each DIMM and
                each lane (lst. \ref{lst:training_phase}),
                applying the seed values and adjusting the delay
                registers accordingly. For each DIMM, the training is performed
                twice: once for the first nibble (lower 4 bits) and once for
                the second nibble (upper 4 bits) if the DIMM is x4. \\

                \begin{listing}
                    \begin{adjustwidth}{0.5cm}{0.5cm}
                        \begin{minted}[linenos]{c}
for (rank = 0; rank < (_2Ranks + 1); rank++) {
    for (nibble = 0; nibble < (train_both_nibbles + 1); nibble++) {
        ...
        write_dqs_receiver_enable_control_registers(
            current_total_delay, dev, Channel, dimm, index_reg);
        ...
    }
}
                        \end{minted}
                    \end{adjustwidth}
                \caption{Iterating over ranks and nibbles to apply delay values,
                    extract from
                    \protect\path{dqsTrainRcvrEn_SW_Fam15} function in
                    \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mctsrc.c}}
                \label{lst:training_phase}
                \end{listing}

                During the training, the controller issues read requests to the
                memory to observe the timing of the DQS signal. The observed
                delays are then averaged and adjusted to ensure the DQS signal
                is correctly aligned across all lanes and ranks. \\

                In the finalization phase, the memory controller exits the
                training mode (lst. \ref{lst:finalization_phase}),
                and the computed delay values are written back to
                the appropriate registers. This ensures that the DQS signal
                remains correctly aligned during normal operation. \\

                \begin{listing}
                    \begin{adjustwidth}{0.5cm}{0.5cm}
                        \begin{minted}[linenos]{c}
Calc_SetMaxRdLatency_D_Fam15(pMCTstat, pDCTstat, 0, 0);
Calc_SetMaxRdLatency_D_Fam15(pMCTstat, pDCTstat, 1, 0);
if (Pass == FirstPass) {
    mct_DisableDQSRcvEn_D(pDCTstat);
}
                        \end{minted}
                    \end{adjustwidth}
                \caption{Exiting training mode and setting read latency,
                    extract from
                    \protect\path{dqsTrainRcvrEn_SW_Fam15} function in
                    \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mctsrc.c}}
                \label{lst:finalization_phase}
                \end{listing}

        \subsection{Potential enhancements}

            \subsubsection{DQS receiver training}

                While the DQS receiver enable training implementation for AMD
                Fam15h G34 processors can perform its intended function in some
                cases, there are several areas where the code is either
                incomplete, suboptimal, or potentially problematic. \\

                The presence of \path{TODO} comments in the code indicates areas
                where the implementation is either incomplete or lacks certain
                necessary functionality. These unaddressed tasks can lead to
                performance issues, potential bugs, or incomplete training,
                which could compromise the stability and reliability of the
                memory subsystem. \\

                In the seed adjustment section for the second pass of training,
                the code includes a \path{TODO} comment regarding fetching the
                correct value from \path{RC2[0]} for the \path{addr_prelaunch}
                variable (lst. \ref{lst:todo_rc2}).

                \begin{listing}
                    \begin{adjustwidth}{0.5cm}{0.5cm}
                        \begin{minted}[linenos]{c}
uint8_t addr_prelaunch = 0; /* TODO: Fetch the correct value from RC2[0] */
                        \end{minted}
                    \end{adjustwidth}
                \caption{An unimplemented feature in the seed adjustment logic,
                    extract from
                    \protect\path{dqsTrainRcvrEn_SW_Fam15} function in
                    \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mcrsrc.c}}
                \label{lst:todo_rc2}
                \end{listing}

                This unimplemented feature suggests that the training process
                may not be fully optimized, as the correct prelaunch address
                setting is not being applied. This could result in incorrect
                seed values being used during the training, leading to
                suboptimal alignment of the DQS signal. Also, the comment
                is unclear about what RC2[0] really means. \\

                The code contains another \path{TODO} comment indicating that
                the support for Load Reduced DIMMs (LRDIMMs) is unimplemented
                (lst. \ref{lst:todo_lrdimm}).

                \begin{listing}
                    \begin{adjustwidth}{0.5cm}{0.5cm}
                        \begin{minted}[linenos]{c}
else if ((pDCTstat->Status & (1 << SB_LoadReduced))) {
    /* TODO
     * Load reduced DIMM support unimplemented
     */
    register_delay = 0x0;
}
                        \end{minted}
                    \end{adjustwidth}
                \caption{LRDIMM support is
                    unimplemented,
                    extract from
                    \protect\path{dqsTrainRcvrEn_SW_Fam15} function in
                    \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mcrsrc.c}}
                \label{lst:todo_lrdimm}
                \end{listing}

                This omission is significant because LRDIMMs are commonly used
                in server environments where high memory capacity is required.
                The lack of support for LRDIMMs could lead to incorrect training
                or even failures when such DIMMs are installed, severely
                impacting the reliability of the system. \\

                \path{FIXME} comments in the code are often indicators of known
                issues or temporary workarounds that need to be addressed. In
                this implementation, there are several such comments that
                highlight critical areas where the current approach may be
                flawed or incomplete. \\

                The first \path{FIXME} comment questions the usage of the
                \path{SSEDIS} setting during the training process
                (lst. \ref{lst:fixme_ssedis}).

                \begin{listing}
                    \begin{adjustwidth}{0.5cm}{0.5cm}
                        \begin{minted}[linenos]{c}
msr = HWCR_MSR;
_RDMSR(msr, &lo, &hi);
/* FIXME: Why use SSEDIS */
if (lo & (1 << 17)) {   /* save the old value */
    _Wrap32Dis = 1;
}
lo |= (1 << 17);        /* HWCR.wrap32dis */
lo &= ~(1 << 15);       /* SSEDIS */
_WRMSR(msr, lo, hi);    /* Setting wrap32dis allows 64-bit memory
                         * references in real mode                */
                        \end{minted}
                    \end{adjustwidth}
                \caption{Questioning the use of
                    \texttt{SSEDIS} in the MSR setting,
                    extract from
                    \protect\path{dqsTrainRcvrEn_SW_Fam15} function in
                    \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mcrsrc.c}}
                \label{lst:fixme_ssedis}
                \end{listing}

                The concern here is that disabling the \path{SSEDIS}
                (SSE Disable) bit could have unintended side effects,
                particularly in environments where SSE instructions are
                expected to be enabled. This could impact the performance of
                the system during training and potentially lead to instability.
                \\

                The code also highlights a potential misprint in the BKDG
                regarding the \path{WrDqDqsEarly} value
                (lst. \ref{lst:fixme_misprint}).

                \begin{listing}
                    \begin{adjustwidth}{0.5cm}{0.5cm}
                        \begin{minted}[linenos]{c}
/* NOTE: While the BKDG states to only program DqsRcvEnGrossDelay, this appears
 * to have been a misprint as DqsRcvEnFineDelay should be set to zero as well.
 */
                        \end{minted}
                    \end{adjustwidth}
                \caption{A possible misprint
                    in the BKDG regarding delay settings,
                    extract from
                    \protect\path{dqsTrainRcvrEn_SW_Fam15} function in
                    \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mcrsrc.c}}
                \label{lst:fixme_misprint}
                \end{listing}

                This indicates that the implementation may be based on incorrect
                or incomplete documentation, leading to potential errors in
                setting the delay values. If this is indeed a misprint in the
                BKDG, the correction should be verified with updated
                documentation, and the implementation should be adjusted
                accordingly. \\

                In addition to the explicit \path{TODO} and \path{FIXME}
                comments, there are other aspects of the implementation that
                could impact performance and stability. \\

                The logic for adjusting the seed values based on the memory
                frequency and the platform's minimum supported frequency is
                complex and prone to errors
                (lst. \ref{lst:seed_adjustment_logic}),
                especially when combined with the
                incomplete \path{TODO} features.

                \begin{listing}
                    \begin{adjustwidth}{0.5cm}{0.5cm}
                        \begin{minted}[linenos]{c}
initial_seed = (uint16_t) (((((uint64_t) initial_seed) *
    fam15h_freq_tab[mem_clk] * 100) / (min_mem_clk * 100)));
                        \end{minted}
                    \end{adjustwidth}
                \caption{Complex seed adjustment logic,
                    extract from
                    \protect\path{dqsTrainRcvrEn_SW_Fam15} function in
                    \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mcrsrc.c}}
                \label{lst:seed_adjustment_logic}
                \end{listing}

                The risk here is that incorrect
                seed values could be used, leading to timing mismatches during
                the training process. \\

                Added to that, stock seeds from the BKDG are used
                (lst. \ref{lst:dqs_receiver_training_seeds}).
                However, it seems that that seeds for used for DQS
                training should be extensively determined for each motherboard,
                and the BKDG \cite{BKDG} does not tell otherwise. Moreover,
                seeds can be configured uniquely for every possible socket,
                channel, DIMM module, and even byte lane combination. The current
                implementation is here only using the recommended seeds from
                the table 99 of the BKDG \cite{BKDG}, which is not sufficient
                and absolutely not adapted to every DIMM module in the market.
                \\

                \begin{listing}
                    \begin{adjustwidth}{0.5cm}{0.5cm}
                        \begin{minted}[linenos]{c}
if (pDCTstat->Status & (1 << SB_Registered)) {
    if (package_type == PT_GR) {
        /* Socket G34: Fam15h BKDG v3.14 Table 99 */
        if (MaxDimmsInstallable == 1) {
            if (channel == 0)
                seed = 0x43;
            else if (channel == 1)
                seed = 0x3f;
            else if (channel == 2)
                seed = 0x3a;
            else if (channel == 3)
                seed = 0x35;
        } else if (MaxDimmsInstallable == 2) {
            if (channel == 0)
                seed = 0x54;
            else if (channel == 1)
                seed = 0x4d;
            else if (channel == 2)
                seed = 0x45;
            else if (channel == 3)
                seed = 0x40;
        } else if (MaxDimmsInstallable == 3) {
            if (channel == 0)
                seed = 0x6b;
            else if (channel == 1)
                seed = 0x5e;
            else if (channel == 2)
                seed = 0x4b;
            else if (channel == 3)
                seed = 0x3d;
        }
                        \end{minted}
                    \end{adjustwidth}
                \caption{Seeds used for DQS Receiver training,
                    extract from
                    \protect\path{dqsTrainRcvrEn_SW_Fam15} function in
                    \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mcrsrc.c}}
                \label{lst:dqs_receiver_training_seeds}
                \end{listing}

                The current implementation also has limited error handling and
                reporting. While some errors are detected during training, the
                code does not have robust mechanisms for recovering from or
                correcting these errors. \\

                This approach might lead to further complications in high-load
                scenarios or when the memory configuration changes, as the
                underlying issues are not resolved. \\

            \subsubsection{Write leveling}

                While the current implementation of write leveling on AMD Fam15h
                G34 processors with RDIMMs can be functional in some cases and
                provides the necessary steps to align DQS signals correctly
                during write operations, there are several areas where the
                implementation is either incomplete, relies on temporary
                workarounds, or may introduce stability and performance issues.
                \\

                One of the most significant concerns with the current
                implementation is the presence of unresolved \path{TODO} and
                \path{FIXME} comments throughout the code. These comments
                indicate areas where the implementation is either incomplete or
                has known issues that have not been fully resolved. \\

                In the \path{procConfig} function, a \path{TODO} comment
                mentions that the current implementation may not be using
                the correct or final value for this variable, once again because
                of a value from RC2[0] that isn't fetched, potentially
                leading to inaccuracies in the seed values used during write
                leveling (lst. \ref{lst:todo_seed_generation}).
                This inaccuracy can result in timing mismatches, which
                may cause data corruption or other stability issues. \\

                \begin{listing}
                    \begin{adjustwidth}{0.5cm}{0.5cm}
                     \begin{minted}[linenos]{c}
uint8_t AddrCmdPrelaunch = 0;   /* TODO: Fetch the correct value from RC2[0] */
                        \end{minted}
                    \end{adjustwidth}
                \caption{Incomplete seed generation
                    implementation,
                    extract from
                    \protect\path{procConfig} function in
                    \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mhwlc_d.c}}
                \label{lst:todo_seed_generation}
                \end{listing}

                Another \path{FIXME} in the code indicates that the
                \path{WrDqDqsEarly} parameter, which is critical for fine-tuning
                the DQS signal’s timing during write operations, is being
                ignored due to unresolved issues
                (lst. \ref{lst:fixme_wrdqdqs_early}). This omission can result in
                less accurate timing adjustments, leading to potential marginal
                instability in systems where tight timing margins are critical.
                \\

                \begin{listing}
                    \begin{adjustwidth}{0.5cm}{0.5cm}
                        \begin{minted}[linenos]{c}
/* FIXME: Ignore WrDqDqsEarly for now to work around training issues */
                        \end{minted}
                    \end{adjustwidth}
                \caption{Omission of
                    \texttt{WrDqDqsEarly} parameter,
                    extract from
                    \protect\path{procConfig} function in
                    \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mhwlc_d.c}}
                \label{lst:fixme_wrdqdqs_early}
                \end{listing}}

                The current implementation uses generic or "stock" seed values
                for certain configurations, such as Socket G34
                (lst. \ref{lst:fixme_mainboard_specific_overrides}). Without
                mainboard-specific overrides, the memory initialization process
                might not be fully optimized for the particular motherboard in
                use. This could result in suboptimal performance or stability
                issues in specific environments, particularly in server
                applications where memory performance is critical. \\

                \begin{listing}
                    \begin{adjustwidth}{0.5cm}{0.5cm}
                        \begin{minted}[linenos]{c}
/* FIXME: Implement mainboard-specific seed and WrDqsGrossDly base overrides.
 * 0x41 and 0x0 are the "stock" values */
                        \end{minted}
                    \end{adjustwidth}
                \caption{Lack of
                    mainboard-specific seed overrides,
                    extract from
                    \protect\path{procConfig} function in
                    \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mhwlc_d.c}}
                \label{lst:fixme_mainboard_specific_overrides}
                \end{listing}

                In \path{AgesaHwWlPhase2}, there is a \path{FIXME} comment that
                suggests that the Critical Gross Delay adjustment has been
                temporarily disabled due to conflicts with RDIMM training
                (lst. \ref{lst:fixme_cgd_adjustment}).
                Disabling this adjustment can lead to less precise DQS alignment,
                especially in complex memory configurations like those using
                RDIMMs, potentially causing instability or degraded performance.
                \\

                \begin{listing}
                    \begin{adjustwidth}{0.5cm}{0.5cm}
                        \begin{minted}[linenos]{c}
/* FIXME: For now, disable CGD adjustment as it seems to interfere with
 * registered DIMM training */
                        \end{minted}
                    \end{adjustwidth}
                \caption{Disabled CGD adjustment due
                    to conflicts,
                    extract from
                    \protect\path{AgesaHwWlPhase2} function in
                    \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mhwlc_d.c}}
                \label{lst:fixme_cgd_adjustment}
                \end{listing}

                The function also bypasses (lst. \ref{lst:fixme_bypass_critical_adjustments})
                certain critical adjustments if the memory speed is being tuned (e.g.,
                during frequency stepping). This bypass is noted as a temporary
                measure due to problems encountered during testing, where the
                first pass values were found to cause issues with PHY training
                on all Family 15h processors tested. This approach indicates a
                lack of robustness in the implementation, particularly in
                handling dynamic changes in memory frequency, which is essential
                for server environments where performance tuning is common. \\

                \begin{listing}
                \begin{adjustwidth}{0.5cm}{0.5cm}
                \begin{minted}[linenos]{c}
/* FIXME: Using the Pass 1 training values causes major phy training problems on
 * all Family 15h processors I tested (Pass 1 values are randomly too high,
 * and Pass 2 cannot lock). Figure out why this is and fix it, then remove
 * the bypass code below... */
                \end{minted}
                \end{adjustwidth}
                \caption{Bypass of critical
                    adjustments during speed tuning,
                    extract from
                    \protect\path{AgesaHwWlPhase2} function in
                    \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mhwlc_d.c}}
                \label{lst:fixme_bypass_critical_adjustments}
                \end{listing}

                The current implementation attempts to compensate for noise and
                instability by overriding faulty values with seed values in
                \path{AgesaHwWlPhase2} (lst. \ref{lst:reactive_error_handling}).
                However, this approach is somewhat blunt
                and reactive, addressing the symptoms rather than the underlying
                causes of instability. This method does not ensure that noise or
                instability is sufficiently mitigated, potentially leading to
                marginal or sporadic failures during normal operation. \\

                \begin{listing}
                    \begin{adjustwidth}{0.5cm}{0.5cm}
                        \begin{minted}[linenos]{c}
if (faulty_value_detected) {
    pDCTData->WLGrossDelay[index+ByteLane] =
        pDCTData->WLSeedGrossDelay[index+ByteLane];
    pDCTData->WLFineDelay[index+ByteLane] =
        pDCTData->WLSeedFineDelay[index+ByteLane];
    status = 1;
}
                        \end{minted}
                    \end{adjustwidth}
                \caption{Blunt error handling to compensate for noise and
                    instability,
                    extract from
                    \protect\path{AgesaHwWlPhase2} function in
                    \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mhwlc_d.c}}
                \label{lst:reactive_error_handling}
                \end{listing}

                The handling of x4 DIMMs, with separate training for each nibble,
                introduces additional complexity. While necessary for these
                configurations, the logic is fragmented, with several points
                where the function branches based on whether the DIMM is x4.
                This complexity increases the risk of bugs or missed conditions,
                particularly if future changes or enhancements are made to the
                code. The overcomplicated logic can also make the code more
                difficult to maintain and extend. \\

            \subsubsection{DQS position training}

                While the DQS position training algorithm implemented in the
                \path{TrainDQSRdWrPos_D_Fam15} function may work in some
                cased to ensure optimal data strobe alignment, there are
                several critical flaws and issues within the implementation
                that could impact its effectiveness and reliability. \\

                Throughout the function, there is an overreliance on hardcoded
                constants and magic numbers, such as:

                \begin{itemize}
                    \item The use of \texttt{0x20} to represent 1 UI (Unit
                        Interval) in multiple places.
                    \item The constant \texttt{16} used in the adjustment of
                        \texttt{region\_center} during the processing of results.
                    \item Magic numbers like \texttt{32} and \texttt{48} in the
                        array dimensions for \texttt{dqs\_results\_array}.
                \end{itemize}

                These values should be replaced with named constants or
                variables that clearly indicate their purpose, improving code
                readability and maintainability. Additionally, using
                well-defined constants would allow easier adjustments if the
                algorithm needs to be adapted for different hardware
                configurations or future revisions of the architecture. \\

                The error handling within the function is rudimentary, with
                errors being flagged primarily by setting bits in the
                \texttt{Errors} variable. However, the function does not
                provide detailed diagnostics or recovery strategies when an
                error occurs. For example:

                \begin{itemize}
                    \item If no passing DQS delay is found for a lane, the
                        function simply sets an error bit without attempting any
                        corrective actions or providing detailed information on
                        what went wrong.
                    \item The early abort mechanism based on the value read from
                        the \texttt{0x264} register does not offer a robust
                        fallback or retry mechanism, which could lead to
                        situations where minor, recoverable issues cause the
                        entire training process to fail.
                \end{itemize}

                Improving the error handling to include detailed diagnostics,
                logging, and potentially corrective actions (such as retrying
                the training with adjusted parameters) would make the function
                more resilient and reliable. \\

                The function contains several areas where the logic is more
                complex than necessary, which can lead to difficulties in
                understanding and maintaining the code. Examples include:

                \begin{itemize}
                    \item The nested loops for iterating over write and read
                        delays are deeply nested, making it challenging to
                        follow the flow of the code and understand the
                        interactions between different parts of the algorithm.
                    \item The use of multiple copies of delay settings (e.g.,
                        \texttt{current\_write\_data\_delay},
                        \texttt{initial\_write\_data\_timing}, and
                        \texttt{initial\_write\_dqs\_delay}) introduces
                        redundancy and increases the likelihood of errors
                        or inconsistencies.
                \end{itemize}

                Refactoring the code to simplify the logic, reduce redundancy,
                and make the flow of operations clearer would improve both the
                readability and reliability of the implementation. \\

                The current implementation does not adequately handle edge cases
                and boundary conditions, such as:

                \begin{itemize}
                    \item The warning issued when a negative DQS recovery delay
                        is detected suggests that the function continues despite
                        recognizing a potentially critical issue, which could
                        lead to system instability
                        (lst. \ref{lst:dqs_train_negative_delay}).
                    \item The averaging of delay values for dual-rank DIMMs does
                        not account for the possibility of significant
                        discrepancies between the ranks, which could result in
                        suboptimal or unstable settings.
                    \item The function does not include comprehensive checks for
                        situations where the calculated delay settings might
                        exceed hardware limitations or cause timing violations.
                \end{itemize}


                \begin{listing}
                    \begin{adjustwidth}{0.5cm}{0.5cm}
                        \begin{minted}[linenos]{c}
if (best_count > 2) {
    uint16_t region_center = (best_pos + (best_count / 2));
    if (region_center < 16) {
        printk(BIOS_WARNING,
            "TrainDQSRdWrPos: negative DQS recovery delay detected!");
        region_center = 0;
    } else {
        region_center -= 16;
    }
    ...
    current_read_dqs_delay[lane] = region_center;
    passing_dqs_delay_found[lane] = 1;
    write_dqs_read_data_timing_registers(current_read_dqs_delay, dev, dct, dimm, index_reg);
}
                        \end{minted}
                    \end{adjustwidth}
                \caption{Allowing a negative DQS recovery delay measurement,
                    extract from
                    \protect\path{TrainDQSRdWrPos_D_Fam15} function in
                    \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mctdqs_d.c}}
                \label{lst:dqs_train_negative_delay}
                \end{listing}

                Improving the handling of edge cases and boundary conditions,
                possibly by incorporating more robust validation checks and
                conservative fallback mechanisms, would make the algorithm more
                reliable in a wider range of scenarios. \\

                The code contains several \texttt{TODO} and \texttt{FIXME}
                comments that indicate incomplete or problematic parts of
                the implementation:

                \begin{itemize}
                    \item The comment \texttt{TODO: Fetch the correct value
                        from RC2[0]} suggests that critical configuration values
                        are not correctly initialized, which could compromise
                        the entire training process.
                    \item The \texttt{FIXME} comments related to early abort
                        checks and DQS recovery delay calculations indicate that
                        there are known issues with the current approach that
                        have not been resolved, potentially leading to incorrect
                        or unstable results.
                    \item The handling of antiphase results, particularly with
                        respect to checking for early aborts, is incomplete and
                        could lead to situations where incorrect results are
                        accepted without proper validation.
                \end{itemize}

                The current implementation's approach to iterating over every
                possible combination of write and read delays is exhaustive but
                may be inefficient. The function performs multiple reads and
                writes to hardware registers for every iteration, which could
                be time-consuming, especially on systems with a large number
                of lanes or complex memory configurations. \\

                Consideration should be given to optimizing the algorithm,
                possibly by narrowing the search space based on prior knowledge
                or implementing more efficient search techniques, to reduce
                the time required for DQS position training without compromising
                accuracy. \\

            \subsubsection{On saving training values in NVRAM}

                The function \path{mctAutoInitMCT_D} is responsible for
                automatically initializing the memory controller training (MCT)
                process, which involves configuring various memory parameters
                and performing training routines to ensure stable and efficient
                memory operation. However, the fact that
                \path{mctAutoInitMCT\_D} does not allow for the restoration of
                training data from NVRAM (lst. \ref{lst:mctAutoInitMCT_D_fixme})
                poses several significant problems. \\

                Memory training is a time-consuming process that involves
                multiple iterations of read/write operations, delay adjustments,
                and calibration steps. By not restoring previously saved
                training data from NVRAM, the system is forced to re-run the
                full training sequence every time it boots up. This leads to
                longer boot times, which can be particularly problematic in
                environments where quick system restarts are critical, such
                as in servers or embedded systems. \\

                Each time memory training is performed, it puts additional
                stress on the memory modules and the memory controller.
                Repeatedly executing the training process at every boot can
                contribute to the wear and tear of hardware components,
                potentially reducing their lifespan. This issue is especially
                concerning in systems that frequently power cycle or reboot. \\

                Memory training is sensitive to various factors, such as
                temperature, voltage, and load conditions. As a result, the
                training results can vary slightly between different boot
                cycles. Without the ability to restore previously validated
                training data, there is a risk of inconsistency in memory
                performance across reboots. This could lead to instability
                or suboptimal memory operation, affecting the overall
                performance of the system. \\

                If the memory training process fails during boot, the system
                may be unable to operate properly or may fail to boot entirely.
                By restoring validated training data from NVRAM, the system
                can bypass the training process altogether, reducing the risk
                of boot failures caused by training issues. Without this
                feature, any minor issue that affects training could result
                in system downtime. \\

                Finally, modern memory controllers often include power-saving
                features that are fine-tuned during the training process. By
                reusing validated training data from NVRAM, the system can
                quickly return to an optimized state with lower power
                consumption.
                The inability to restore this data forces the system to
                operate at a potentially less efficient state until training
                is complete, leading to higher power consumption during the
                boot process. \\

            \subsubsection{A seedless DQS position training algorithm}

                An algorithm to find the best timing for the DQS so that the
                memory controller can reliably read data from the memory
                could be done without relying on any pre-known starting
                values (seeds). This would allow for better reliability and
                wider support for different situations. The algorithm
                could be describe as follows. \\

                \begin{itemize}
                    \item Prepare Memory Controller:
                    The memory controller needs to be in a state where it can
                    safely adjust the DQS timing without affecting the normal
                    operation of the system. By blocking the DQS signal locking,
                    we ensure that the adjustments made during training do not
                    interfere with the controller’s ability to capture data
                    until the optimal settings are found.

                    \item Initialize Variables:
                        Set up variables to store the various timing settings and
                        test results for each bytelane. This setup is crucial
                        because each bytelane might require a different optimal
                        timing, and keeping track of these values ensures that the
                        algorithm can correctly determine the best delay settings
                        later.
                 \end{itemize}

                The main loop is the core of the algorithm, where different
                timing settings are systematically explored. By looping
                through possible delay settings, the algorithm ensures
                that it doesn't miss any potential optimal timings. The
                loop structure allows a methodical test of a range of
                delays to find the most reliable one. \\

                The gross delay is here the coarse adjustment to the timing
                of the DQS signal. It shifts the timing window by a large
                amount, helping to broadly align the DQS with the data
                lines (DQ). The fine delay, which is the smaller, more
                precise change to the timing of the DQS signal once the
                coarse alignment (through gross delay) has been achieved,
                would then be computed. \\

                To compute a delay, here would be the steps:

                \begin{itemize}
                    \item Set a delay:
                    Setting an initial delay allows the algorithm to start
                    testing. The initial delay might be zero or another default
                    value, providing a baseline from which to begin the search
                    for the optimal timing.

                    \item Test it:
                    After setting the delay, it is essential to test whether the
                    memory controller can read data correctly. This step is
                    critical because it indicates whether the current delay
                    setting is within the acceptable range for reliable data
                    capture.

                    \item Check the result:
                    If the memory controller successfully reads data, it means
                    the current delay setting is valid. This information is
                    crucial because it helps define the range of acceptable
                    timings. If the test fails, it indicates that the curren
                    t delay setting is outside the range where the memory
                    controller can reliably capture data.

                    \item Increase/decrease delay:
                    By incrementally adjusting the delay, either increasing or
                    decreasing, the algorithm can explore different timing
                    settings in a controlled manner. This ensures that the
                    entire range of possible delays is covered without skipping
                    over any potential good delays.

                    \item Test again:
                    Re-testing after each adjustment ensures that the exact
                    point where the DQS timing goes from acceptable (pass) to
                    unacceptable (fail) is caught. This step helps in
                    identifying the transition point, which is often the optimal
                    place to set the DQS delay.

                    \item Look for a transition:
                    The transition from pass to fail is where the DQS timing
                    crosses the boundary of the valid timing window. This
                    transition is crucial because it marks the end of the
                    reliable range. The best timing is usually just before
                    this transition.

                    \item Record the best setting:
                    Storing the best delay setting for each bytelane ensures
                    that a reliable timing configuration is available when the
                    training is complete.

                    \item Confirm all bytelanes:
                    Before finalizing the settings, it is important to ensure
                    that the chosen delays work for all bytelanes. This step
                    serves as a final safeguard against errors, ensuring that
                    every part of the data bus is correctly aligned.
                \end{itemize}

                Each bytelane (8-bit segment of data) may require a
                different optimal delay setting. By repeating the process
                for all bytelanes, the algorithm ensures that the entire
                data bus is correctly timed. Misalignment in even one
                bytelane can lead to data errors, making it essential to
                tune every bytelane individually. \\

                Once the best settings are confirmed, they need to be
                applied to the memory controller for use during normal
                operation. This step locks in the most reliable timing
                configuration found during the training process. \\

                After the optimal settings are applied, it is necessary
                to allow the DQS signal locking mechanism to resume. This
                locks in the delay settings, ensuring stable operation going
                forward. \\

                Finally, the algorithm needs to indicate whether it was
                successful in finding reliable timing settings for all
                bytelanes. This feedback is crucial for determining whether
                the memory system is correctly configured or if further
                adjustments or troubleshooting are needed. \\

% ------------------------------------------------------------------------------
% CHAPTER 5: Virtualization of the operating system through firmware abstraction
% ------------------------------------------------------------------------------
\chapter{Virtualization of the operating system through firmware abstraction}

    In contemporary computing systems, the operating system (OS) no longer
    interacts directly with hardware in the same way it did in earlier computing
    architectures. Instead, the OS operates within a highly abstracted
    environment, where critical functions are managed by various firmware
    components such as ACPI, SMM, UEFI, Intel Management Engine (ME), and AMD
    Platform Security Processor (PSP). This layered abstraction has led to the
    argument that the OS is effectively running in a virtualized environment,
    akin to a virtual machine (VM).

    \section{ACPI and abstraction of hardware control}

        The Advanced Configuration and Power Interface (ACPI) provides a
        standardized method for the OS to manage hardware configuration and
        power states, effectively abstracting the underlying hardware
        complexities. ACPI abstracts hardware details, allowing the OS to
        interact with hardware components without needing direct control over
        them. This abstraction is similar to how a hypervisor abstracts physical
        hardware for VMs, enabling a consistent interface regardless of the
        underlying hardware specifics. \\

        According to \textcite{bellosa2010}, the abstraction provided by ACPI
        not only simplifies the OS's interaction with hardware but also limits
        the OS's ability to fully control the hardware, which is instead managed
        by ACPI-compliant firmware. This layer of abstraction contributes to the
        virtualization-like environment in which the OS operates. \\

        More importantly, the ACPI Component Architecture (ACPICA) is a critical
        component integrated into the Linux kernel, serving as the foundation
        for the system's ACPI implementation \cite{intel_acpi_programming_2023}.
        ACPICA provides the core ACPI functionalities, such as hardware
        configuration, power management, and thermal management, which are
        essential for modern computing platforms. However, its integration into
        the Linux kernel has brought significant complexity and code overhead,
        making Linux heavily dependent on ACPICA for managing ACPI-related
        tasks.

        ACPICA is a large and complex project, with its codebase encompassing
        a wide range of functionalities required to implement ACPI standards.
        The integration of ACPICA into the Linux kernel significantly increases
        the kernel's overall code size. An example of that can easily be
        reproduced with a small experiment (lst. \ref{lst:acpica_in_linux}).

        \begin{listing}[H]
            \begin{adjustwidth}{0.5cm}{0.5cm}
                \inputminted{sh}{listings/acpica_size.sh}
            \end{adjustwidth}
            \caption{How to estimate the impact of ACPICA in Linux}
            \label{lst:acpica_in_linux}
        \end{listing}

        As of recent statistics, ACPICA comprises between 100,000 to 200,000
        lines of code, making it one of the larger subsystems within the Linux
        kernel. This size is indicative of the extensive range of features
        and capabilities ACPICA must support, including but not limited to the
        ACPI interpreter, AML (ACPI Machine Language) parser, and various
        hardware-specific drivers. The ACPICA codebase is not monolithic; it is
        highly modular and consists of various components, each responsible for
        specific ACPI functions. For instance, ACPICA includes components for
        managing ACPI tables, interpreting AML bytecode, handling events, and
        interacting with hardware. This modularity, while beneficial for
        isolating different functionalities, also contributes to the overall
        complexity of the system. The separation of ACPICA into multiple modules
        necessitates careful coordination and integration with the rest of the
        Linux kernel, adding to the kernel's complexity. \\

        ACPICA's integration into the Linux kernel is designed to maintain a
        clear separation between the core ACPI functionalities and the kernel's
        other subsystems \cite{intel_acpi_programming_2023}. This separation is
        achieved through well-defined interfaces and abstraction layers,
        allowing the Linux kernel to interact with ACPICA without being tightly
        coupled to its internal implementation details. For example, ACPICA
        provides an API that the Linux kernel can use to interact with ACPI
        tables, execute ACPI methods, and manage power states. This API
        abstracts the underlying complexity of the ACPI implementation, making
        it easier for kernel developers to incorporate ACPI support without
        delving into the intricacies of ACPICA's internals.
        Moreover, ACPICA's role in interpreting AML bytecode, which is
        essentially a form of low-level programming language embedded in ACPI
        tables, adds a layer of abstraction. The Linux kernel relies on ACPICA
        to execute AML methods and manage hardware resources according to the
        ACPI specifications. This reliance further underscores the idea that
        ACPI acts as a virtualizing environment, shielding the kernel from
        the complexities of directly interfacing with hardware components.

    \section{SMM as a hidden execution layer}

        System Management Mode (SMM) is a special-purpose operating mode
        provided by x86 processors, designed to handle system-wide functions
        such as power management, thermal monitoring, and hardware control,
        independent of the OS. SMM operates transparently to the OS, executing
        code that the OS cannot detect or control, similar to how a hypervisor
        controls the execution environment of VMs. \\

        Research by \textcite{huang2009invisible} argues that SMM introduces a
        hidden layer of execution that diminishes the OS's control over the
        hardware, creating a virtualized environment where the OS is unaware of
        and unable to influence certain system-level operations. This hidden
        execution layer reinforces the idea that the OS runs in an environment
        similar to a VM, with the firmware acting as a hypervisor. \\

    \section{UEFI and persistence}

        The Unified Extensible Firmware Interface (UEFI) has largely replaced
        the traditional BIOS in modern systems, providing a sophisticated
        environment that includes a kernel-like structure capable of running
        drivers and applications independently of the OS. UEFI remains active
        even after the OS has booted, continuing to manage certain hardware
        functions, which abstracts these functions away from the OS. \\

        \textcite{mcclean2017uefi} discusses how UEFI creates a persistent
        execution environment that overlaps with the OS's operation, effectively
        placing the OS in a position where it runs on top of another controlling
        layer, much like a guest OS in a VM. This persistence and the ability of
        UEFI to manage hardware resources independently further blur the lines
        between traditional OS operation and virtualized environments.
        Indeed, as we studied in a precedent chapter, UEFI is designed as a
        modular and extensible firmware interface that sits between the
        computer's hardware and the operating system. Unlike the monolithic
        BIOS, UEFI is composed of several layers and components, each
        responsible for different aspects of the system's boot and runtime
        processes. The core components of UEFI include the Pre-EFI
        Initialization (PEI), Driver Execution Environment (DXE),
        Boot Device Selection (BDS), and Runtime Services. Each of these
        components plays a critical role in initializing the hardware,
        managing drivers, selecting boot devices, and providing runtime
        services to the OS. \\

        The PEI (Pre-EFI Initialization) phase is responsible for initializing
        the CPU, memory, and other essential hardware components. It ensures
        that the system is in a stable state before handing control to the
        DXE phase. In the DXE phase, the system loads and initializes various
        drivers required for the OS to interact with the hardware. The DXE phase
        also constructs the UEFI Boot Services, which provide the OS with
        interfaces to the hardware during the boot process. The BDS (Boot Device
        Selection) phase is responsible for selecting the device from which the
        OS will boot. It interacts with the UEFI Boot Manager to determine the
        correct boot path and load the OS. After the OS has booted, UEFI
        provides Runtime Services that remain accessible to the OS. These
        services include interfaces for managing system variables, time, and
        hardware. UEFI also supports the execution of standalone applications,
        which can be used for system diagnostics, firmware updates, or other
        tasks. These applications operate independently of the OS, highlighting
        UEFI's capabilities as a minimalistic OS. \\

        UEFI abstracts the underlying hardware from the OS, providing a
        standardized interface for the OS to interact with different hardware
        components. This abstraction simplifies the development of OSes and
        drivers, as they do not need to be tailored for specific hardware
        configurations. UEFI's hardware abstraction is one of the key features
        that enable it to act as a virtualizing environment for the OS
        \cite{mcclean2017uefi}.

        \subsection{Memory Management}

            UEFI provides a detailed memory map to the OS during the boot process,
            which includes information about available, reserved, and used memory
            regions. The OS uses this memory map to manage its own memory allocation
            and paging mechanisms. The overlap in memory management functions
            highlights UEFI's role in preparing the system for OS operation.
            This memory map includes all the memory regions in the system,
            categorized into different types, such as usable memory, reserved
            memory, and memory-mapped I/O. The OS relies on this map to understand
            the system's memory layout and avoid conflicts \cite{osdev_uefi_memory}.
            The OS extends UEFI's memory
            management by implementing its own memory allocation, paging, and
            virtual memory mechanisms. However, the OS's memory management is
            built on the foundation provided by UEFI, demonstrating the close
            relationship between the two.

        \subsection{File System Management}

            UEFI includes its own file system management capabilities, which overlap
            with those of the OS. The most notable example is the EFI System
            Partition (ESP), a special partition formatted with the FAT file system
            that UEFI uses to store bootloaders, drivers, and other critical files
            \cite{uefi_spec}. The ESP is a mandatory partition in UEFI systems,
            containing the bootloaders, firmware updates, and other files
            necessary for system initialization. UEFI accesses the ESP
            independently of the OS, but the OS can also access and manage files
            on the ESP, creating an overlap in file system management functions
            \cite{uefi_smm_security}. UEFI natively supports the FAT file
            system, allowing it to read and write files on the ESP. This support
            overlaps with the OS's file system management, as both UEFI and the
            OS can manipulate files on the ESP.

        \subsection{Device Drivers}

            As we studied in an earlier chapter, UEFI includes its own driver
            model, allowing it to load and execute drivers independently of the
            OS. This capability overlaps with the OS's driver management
            functions, as both UEFI and the OS manage hardware devices through
            drivers.
            UEFI drivers are typically used during
            the boot process to initialize and control hardware devices. These
            drivers provide the necessary interfaces for the OS to interact with
            the hardware once it has booted \cite{uefi_smm_security}.
            After the OS has booted, it loads its own drivers for hardware
            devices. However, the OS often relies on the initial hardware setup
            performed by UEFI drivers.

        \subsection{Power Management}

            UEFI provides power management services that overlap with the OS's
            power management functions. These services allow UEFI to manage
            power states and transitions independently of the OS \cite{uefi_spec}.
            These services ensure that the system conserves power during periods
            of inactivity and can quickly resume operation when needed
            The OS extends UEFI's power management by implementing its own
            power-saving mechanisms, such as CPU throttling and dynamic voltage
            scaling.

    \section{Intel and AMD: control beyond the OS}

        Intel Management Engine (ME) and AMD Platform Security Processor (PSP)
        are embedded microcontrollers within Intel and AMD processors,
        respectively. These components run their own firmware and operate
        independently of the main CPU, handling tasks such as security
        enforcement, remote management, and digital rights management (DRM). \\

        \textcite{bulygin2013chipset} highlights how these microcontrollers have
        control over the system that supersedes the OS, managing hardware and
        security functions without the OS's knowledge or consent. This level of
        control is reminiscent of a hypervisor that manages the resources and
        security of VMs. The OS, in this context, operates similarly to a VM
        that does not have full control over the hardware it ostensibly manages. \\

    \section{Processors microcode}

        Modern CPUs are incredibly complex, with their functionality relying
        heavily on microcode to interpret and execute instructions. Microcode
        acts as a translation layer between the high-level instructions that
        software provides and the lower-level operations that the hardware
        can execute. Microcode operates directly within the CPU. \\

        CPU microcode is a set of low-level firmware instructions embedded
        within the processor. It translates complex machine instructions into
        simpler, executable sequences of operations that the CPU's hardware
        can directly perform \cite{Intel2018}. This layer of abstraction allows
        CPU manufacturers to update or patch the behavior of the processor
        post-manufacturing, which is crucial for addressing bugs, optimizing
        performance, and applying security patches \cite{Wilcox2018}.

        In a sense, microcode can be seen as an argument for the CPU running
        a form of low-level virtual machine. Just as a VM abstracts and manages
        hardware resources for a guest OS, microcode abstracts and manages the
        complexity of CPU hardware for machine-level instructions. This
        virtualization enables the CPU to support a wide variety of instructions
        and operational modes without needing to change the underlying hardware
        \cite{Abraham1983}.

    \section{The OS as a virtualized environment}

        The combined effect of these firmware components (ACPI, SMM, UEFI,
        Intel ME, and AMD PSP) creates an environment where the OS operates in
        a virtualized or highly abstracted layer. The OS does not directly
        manage the hardware; instead, it interfaces with these firmware
        components, which themselves control the hardware resources. This
        situation is analogous to a virtual machine, where the guest OS
        operates on virtualized hardware managed by a hypervisor. \\

        \textcite{smith2019firmware} argues that modern OS environments,
        influenced by these firmware components, should be considered
        virtualized environments. The firmware acts as an intermediary layer
        that abstracts and controls hardware resources, thereby limiting the
        OS's direct access and control. \\

        The presence and operation of modern firmware components such as ACPI,
        SMM, UEFI, Intel ME, and AMD PSP and even CPU microcode contribute to
        a significant abstraction of hardware from the OS.
        This abstraction creates an environment that
        parallels the operation of a virtual machine, where the OS functions
        within a controlled, virtualized layer managed by these firmware
        systems. The growing body of research supports this perspective,
        suggesting that the traditional notion of an OS directly managing
        hardware is increasingly outdated in the face of these complex,
        autonomous firmware components.

\chapter*{Conclusion}
\addcontentsline{toc}{chapter}{Conclusion}

    This document has explored the evolution and current state of firmware,
    particularly focusing on the transition from traditional BIOS to more
    advanced firmware interfaces such as UEFI and \textit{coreboot}. The
    evolution from a simple set of routines stored in ROM to complex systems
    like UEFI and \textit{coreboot} highlights the growing importance of
    firmware in modern computing.

    Firmware now plays a critical role not
    only in hardware initialization but also in memory management, security,
    and system performance optimization. \\

    The study of the ASUS KGPE-D16 mainboard illustrates how firmware,
    particularly \textit{coreboot}, plays a crucial role in the efficient
    and secure operation of high-performance systems. The KGPE-D16, with its
    support for free software-compatible firmware, exemplifies the potential
    of libre firmware to deliver both high performance and freedom from
    proprietary constraints. However, it is important to acknowledge that
    the KGPE-D16 is not without its imperfections. The detailed analysis of
    firmware components, such as the bootblock, romstage, and especially the
    RAM initialization and training algorithms, reveals areas where the
    firmware can be further refined to enhance system stability and
    performance. These improvements are not only beneficial for the KGPE-D16
    but can also be applied to other boards, extending the impact of these
    optimizations across a broader range of hardware. \\

    Moreover, the discussion on modern firmware components such as ACPI,
    SMM, UEFI, Intel ME, and AMD PSP demonstrates how these elements
    abstract hardware from the operating system, creating a virtualized
    environment where the OS operates more like a guest in a
    hypervisor-controlled system. This abstraction raises important
    considerations about control, security, and user freedom in contemporary
    computing.
    As we continue to witness the increasing complexity and influence of
    firmware in computing, it becomes crucial to advocate for free
    software-compatible hardware. The dependence on proprietary firmware and
    the associated restrictions on user freedom are growing concerns that
    need to be addressed. The development and adoption of libre firmware
    solutions, such as \textit{coreboot} and GNU Boot, are essential steps
    towards ensuring that users retain control over their hardware and
    software environments. \\

    It is imperative that the community of developers, researchers, and
    users come together to support and contribute to the development of
    free firmware. By fostering innovation and collaboration in this field,
    we can advance towards a future where free software-compatible hardware
    becomes the norm, ensuring that computing remains open, secure, and
    under the control of its users. The significance of a libre BIOS cannot
    be overstated, it is the foundation upon which a truly free and open
    computing ecosystem can be built \cite{coreboot_fsf}.
    The importance of the GNU Boot project cannot be
    overstated. As a fully free firmware initiative, GNU Boot represents a
    critical step towards achieving truly libre BIOSes, ensuring that users
    can maintain full control over their hardware and firmware environments.
    The continued development and support of GNU Boot are essential for
    advancing the goals of free software and protecting user freedoms in the
    increasingly complex landscape of modern computing. \\

\newpage

% Bibliography
\nocite{*}
\printbibliography
\addcontentsline{toc}{chapter}{Bibliography}
\newpage

\chapter*{Appendix: Long code listings}
\addcontentsline{toc}{chapter}{Appendix: Long code listings}
\renewcommand{\thelisting}{L.\arabic{listing}}

    \begin{listing}
    \begin{adjustwidth}{0.5cm}{0.5cm}
    \inputminted{c}{
        listings/src_northbridge_amd_amdmct_mct_ddr3_mct_d_1.c}
    \end{adjustwidth}
    \caption{
        Beginning of
        \protect\path{mctAutoInitMCT_D()}, extract from
        \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mct_d.c}}
    \label{lst:mctAutoInitMCT_D_1}
    \end{listing}

    \begin{listing}
    \begin{adjustwidth}{0.5cm}{0.5cm}
    \inputminted{c}{
        listings/src_northbridge_amd_amdmct_mct_ddr3_mct_d_2.c}
    \end{adjustwidth}
    \caption{
        DIMM initialization in
        \protect\path{mctAutoInitMCT_D()}, extract from
        \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mct_d.c}}
    \label{lst:mctAutoInitMCT_D_2}
    \end{listing}

    \begin{listing}
    \begin{adjustwidth}{0.5cm}{0.5cm}
    \inputminted{c}{
        listings/src_northbridge_amd_amdmct_mct_ddr3_mct_d_3.c}
    \end{adjustwidth}
    \caption{
        Voltage control in
        \protect\path{mctAutoInitMCT_D()}, extract from
        \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mct_d.c}}
    \label{lst:mctAutoInitMCT_D_3}
    \end{listing}

    \begin{listing}
    \begin{adjustwidth}{0.5cm}{0.5cm}
    \inputminted{c}{
        listings/src_northbridge_amd_amdmct_mct_ddr3_mct_d_fixme.c}
    \end{adjustwidth}
    \caption{
        \protect\path{mctAutoInitMCT_D()} does not allow restoring
        previous training values, extract from
        \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mct_d.c}}
    \label{lst:mctAutoInitMCT_D_fixme}
    \end{listing}

    \begin{listing}
    \begin{adjustwidth}{0.5cm}{0.5cm}
    \inputminted{c}{
        listings/src_northbridge_amd_amdmct_mct_ddr3_mct_d_4.c}
    \end{adjustwidth}
    \caption{
        Preparing SMBus, DCTs and NB in
        \protect\path{mctAutoInitMCT_D()} from
        \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mct_d.c}}
    \label{lst:mctAutoInitMCT_D_4}
    \end{listing}

    \begin{listing}
    \begin{adjustwidth}{0.5cm}{0.5cm}
    \inputminted{c}{
        listings/src_northbridge_amd_amdmct_mct_ddr3_mct_d_5.c}
    \end{adjustwidth}
    \caption{
        Get DQS, reset and activate ECC in
        \protect\path{mctAutoInitMCT_D()} from
        \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mct_d.c}}
    \label{lst:mctAutoInitMCT_D_5}
    \end{listing}

    \begin{listing}
    \begin{adjustwidth}{0.5cm}{0.5cm}
    \inputminted{c}{
        listings/src_northbridge_amd_amdmct_mct_ddr3_mct_d_6.c}
    \end{adjustwidth}
    \caption{
        Mapping DRAM with cache, validating DCT nodes
        and finishing the init process in
        \protect\path{mctAutoInitMCT_D()} from
        \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mct_d.c}}
    \label{lst:mctAutoInitMCT_D_6}
    \end{listing}

    \begin{listing}[H]
    \begin{adjustwidth}{0.5cm}{0.5cm}
    \begin{minted}[linenos]{c}
retry_requested = 0;
for (Node = 0; Node < MAX_NODES_SUPPORTED; Node++) {
    struct DCTStatStruc *pDCTstat;
    pDCTstat = pDCTstatA + Node;

    if (pDCTstat->NodePresent) {
        if (pDCTstat->TrainErrors & (1 << SB_FatalError)) {
            printk(BIOS_ERR, "DIMM training FAILED!  Restarting system...");
            soft_reset();
        }
        if (pDCTstat->TrainErrors & (1 << SB_RetryConfigTrain)) {
            retry_requested = 1;

            pDCTstat->TrainErrors &= ~(1 << SB_RetryConfigTrain);
            pDCTstat->TrainErrors &= ~(1 << SB_NODQSPOS);
            pDCTstat->ErrStatus &= ~(1 << SB_RetryConfigTrain);
            pDCTstat->ErrStatus &= ~(1 << SB_NODQSPOS);
        }
    }
}

if (retry_requested) {
    printk(BIOS_DEBUG, "%s: Restarting training on algorithm request\n",
                        __func__);
    /* Reset frequency to minimum */
    for (Node = 0; Node < MAX_NODES_SUPPORTED; Node++) {
        struct DCTStatStruc *pDCTstat;
        pDCTstat = pDCTstatA + Node;
        if (pDCTstat->NodePresent) {
            uint8_t original_target_freq = pDCTstat->TargetFreq;
            uint8_t original_auto_speed = pDCTstat->DIMMAutoSpeed;
            pDCTstat->TargetFreq = mhz_to_memclk_config(mctGet_NVbits(NV_MIN_MEMCLK));
            pDCTstat->Speed = pDCTstat->DIMMAutoSpeed = pDCTstat->TargetFreq;
            SetTargetFreq(pMCTstat, pDCTstatA, Node);
            pDCTstat->TargetFreq = original_target_freq;
            pDCTstat->DIMMAutoSpeed = original_auto_speed;
        }
    }
    /* Apply any DIMM timing changes */
    for (Node = 0; Node < MAX_NODES_SUPPORTED; Node++) {
        struct DCTStatStruc *pDCTstat;
        pDCTstat = pDCTstatA + Node;
        if (pDCTstat->NodePresent) {
            AutoCycTiming_D(pMCTstat, pDCTstat, 0);
            if (!pDCTstat->GangedMode)
                if (pDCTstat->DIMMValidDCT[1] > 0)
                    AutoCycTiming_D(pMCTstat, pDCTstat, 1);
        }
    }
    goto retry_dqs_training_and_levelization;
}
    \end{minted}
    \end{adjustwidth}
    \caption{Error detection and retry mechanism during DQS training,
             extract from the
             \protect\path{DQSTiming_D} function in
             \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mct_d.c}}
    \label{lst:error_handling}
    \end{listing}

    \begin{listing}[H]
    \begin{adjustwidth}{0.5cm}{0.5cm}
    \begin{minted}[linenos]{c}
if (Pass == SecondPass) {
    if (pDCTstat->TargetFreq > mhz_to_memclk_config(mctGet_NVbits(NV_MIN_MEMCLK))) {
        uint8_t global_phy_training_status = 0;
        final_target_freq = pDCTstat->TargetFreq;

        while (pDCTstat->Speed != final_target_freq) {
            if (is_fam15h())
                pDCTstat->TargetFreq =
                    fam15h_next_highest_memclk_freq(pDCTstat->Speed);
            else
                pDCTstat->TargetFreq = final_target_freq;
            SetTargetFreq(pMCTstat, pDCTstatA, Node);
            timeout = 0;
            do {
                status = 0;
                timeout++;
                status |= PhyWLPass2(pMCTstat, pDCTstat, 0,
                                        (pDCTstat->TargetFreq == final_target_freq));
                status |= PhyWLPass2(pMCTstat, pDCTstat, 1,
                                        (pDCTstat->TargetFreq == final_target_freq));
                if (status)
                    printk(BIOS_INFO,
                        "%s: Retrying write levelling due to invalid value(s) "
                        "detected in last phase\n",
                        __func__);
            } while (status && (timeout < 8));
            global_phy_training_status |= status;
        }

        pDCTstat->TargetFreq = final_target_freq;

        if (global_phy_training_status)
            printk(BIOS_WARNING,
                "%s: Uncorrectable invalid value(s) detected in second phase of "
                "write levelling; "
                "continuing but system may be unstable!\n",
                __func__);

        uint8_t dct;
        for (dct = 0; dct < 2; dct++) {
            sDCTStruct *pDCTData = pDCTstat->C_DCTPtr[dct];
            memcpy(pDCTData->WLGrossDelayFinalPass,
                   pDCTData->WLGrossDelayPrevPass,
                   sizeof(pDCTData->WLGrossDelayPrevPass));
            memcpy(pDCTData->WLFineDelayFinalPass,
                   pDCTData->WLFineDelayPrevPass,
                   sizeof(pDCTData->WLFineDelayPrevPass));
            pDCTData->WLCriticalGrossDelayFinalPass =
                pDCTData->WLCriticalGrossDelayPrevPass;
        }
    }
}
    \end{minted}
    \end{adjustwidth}
    \caption{Write Leveling (second pass), extract from the
             \texttt{WriteLevelization\_HW} function in
             \texttt{src/northbridge/amd/amdmct/mct\_ddr3/mcthwl.c}.}
    \label{lst:write_level_second_pass}
    \end{listing}

    \begin{listing}
    \begin{adjustwidth}{0.5cm}{0.5cm}
    \begin{minted}[linenos]{c}
uint8_t MaxDimmsInstallable = mctGet_NVbits(NV_MAX_DIMMS_PER_CH);

if (pDCTstat->Status & (1 << SB_Registered)) {
    if (package_type == PT_GR) {
        // Socket G34: Fam15h BKDG v3.14 Table 99
        if (MaxDimmsInstallable == 1) {
            if (channel == 0)
                seed = 0x43;
            else if (channel == 1)
                seed = 0x3f;
            else if (channel == 2)
                seed = 0x3a;
            else if (channel == 3)
                seed = 0x35;
        }
        ...
    }
    ...
} else if (pDCTstat->Status & (1 << SB_LoadReduced)) {
    // Load Reduced DIMM configuration
    if (package_type == PT_GR) {
        // Socket G34: Fam15h BKDG v3.14 Table 99
        if (MaxDimmsInstallable == 1) {
            if (channel == 0)
                seed = 0x123;
            ...
        }
    }
}
    \end{minted}
    \end{adjustwidth}
    \caption{Seed generation for DQS receiver enable training based on DIMM type
        and configuration,
        extract from
        \protect\path{fam15_receiver_enable_training_seed} function in
        \protect\path{src/northbridge/amd/amdmct/mct_ddr3/mctsrc.c}}
    \label{lst:seed_generation}
    \end{listing}

\newpage

% ------------------------------------------------------------------------------
%
%
%                                LICENSES
%
%
%
% ------------------------------------------------------------------------------
\chapter*{\center\rlap{GNU General Public License version 2}}
\addcontentsline{toc}{chapter}{GNU General Public License version 2}

    \parindent 0in

    Version 2, June 1991

    Copyright \copyright\ 1989, 1991 Free Software Foundation, Inc.

    \bigskip

    51 Franklin Street, Fifth Floor, Boston, MA  02110-1301, USA

    \bigskip

    Everyone is permitted to copy and distribute verbatim copies
    of this license document, but changing it is not allowed.

    \bigskip{\bf\large Preamble}\bigskip


    The licenses for most software are designed to take away your freedom to
    share and change it.  By contrast, the GNU General Public License is
    intended to guarantee your freedom to share and change free software---to
    make sure the software is free for all its users.  This General Public
    License applies to most of the Free Software Foundation's software and to
    any other program whose authors commit to using it.  (Some other Free
    Software Foundation software is covered by the GNU Library General Public
    License instead.)  You can apply it to your programs, too.

    When we speak of free software, we are referring to freedom, not price.
    Our General Public Licenses are designed to make sure that you have the
    freedom to distribute copies of free software (and charge for this service
    if you wish), that you receive source code or can get it if you want it,
    that you can change the software or use pieces of it in new free programs;
    and that you know you can do these things.

    To protect your rights, we need to make restrictions that forbid anyone to
    deny you these rights or to ask you to surrender the rights.  These
    restrictions translate to certain responsibilities for you if you
    distribute copies of the software, or if you modify it.

    For example, if you distribute copies of such a program, whether gratis or
    for a fee, you must give the recipients all the rights that you have.  You
    must make sure that they, too, receive or can get the source code.  And
    you must show them these terms so they know their rights.

    We protect your rights with two steps: (1) copyright the software, and (2)
    offer you this license which gives you legal permission to copy,
    distribute and/or modify the software.

    Also, for each author's protection and ours, we want to make certain that
    everyone understands that there is no warranty for this free software.  If
    the software is modified by someone else and passed on, we want its
    recipients to know that what they have is not the original, so that any
    problems introduced by others will not reflect on the original authors'
    reputations.

    Finally, any free program is threatened constantly by software patents.
    We wish to avoid the danger that redistributors of a free program will
    individually obtain patent licenses, in effect making the program
    proprietary.  To prevent this, we have made it clear that any patent must
    be licensed for everyone's free use or not licensed at all.

    The precise terms and conditions for copying, distribution and
    modification follow.

    \bigskip{\Large \sc Terms and Conditions For Copying, Distribution and
      Modification}\bigskip


    %\renewcommand{\theenumi}{\alpha{enumi}}
    \begin{enumerate}

    \addtocounter{enumi}{-1}

    \item

    This License applies to any program or other work which contains a notice
    placed by the copyright holder saying it may be distributed under the
    terms of this General Public License.  The ``Program'', below, refers to
    any such program or work, and a ``work based on the Program'' means either
    the Program or any derivative work under copyright law: that is to say, a
    work containing the Program or a portion of it, either verbatim or with
    modifications and/or translated into another language.  (Hereinafter,
    translation is included without limitation in the term ``modification''.)
    Each licensee is addressed as ``you''.

    Activities other than copying, distribution and modification are not
    covered by this License; they are outside its scope.  The act of
    running the Program is not restricted, and the output from the Program
    is covered only if its contents constitute a work based on the
    Program (independent of having been made by running the Program).
    Whether that is true depends on what the Program does.

    \item You may copy and distribute verbatim copies of the Program's source
      code as you receive it, in any medium, provided that you conspicuously
      and appropriately publish on each copy an appropriate copyright notice
      and disclaimer of warranty; keep intact all the notices that refer to
      this License and to the absence of any warranty; and give any other
      recipients of the Program a copy of this License along with the Program.

    You may charge a fee for the physical act of transferring a copy, and you
    may at your option offer warranty protection in exchange for a fee.

    \item

    You may modify your copy or copies of the Program or any portion
    of it, thus forming a work based on the Program, and copy and
    distribute such modifications or work under the terms of Section 1
    above, provided that you also meet all of these conditions:

    \begin{enumerate}

    \item

    You must cause the modified files to carry prominent notices stating that
    you changed the files and the date of any change.

    \item

    You must cause any work that you distribute or publish, that in
    whole or in part contains or is derived from the Program or any
    part thereof, to be licensed as a whole at no charge to all third
    parties under the terms of this License.

    \item
    If the modified program normally reads commands interactively
    when run, you must cause it, when started running for such
    interactive use in the most ordinary way, to print or display an
    announcement including an appropriate copyright notice and a
    notice that there is no warranty (or else, saying that you provide
    a warranty) and that users may redistribute the program under
    these conditions, and telling the user how to view a copy of this
    License.  (Exception: if the Program itself is interactive but
    does not normally print such an announcement, your work based on
    the Program is not required to print an announcement.)

    \end{enumerate}


    These requirements apply to the modified work as a whole.  If
    identifiable sections of that work are not derived from the Program,
    and can be reasonably considered independent and separate works in
    themselves, then this License, and its terms, do not apply to those
    sections when you distribute them as separate works.  But when you
    distribute the same sections as part of a whole which is a work based
    on the Program, the distribution of the whole must be on the terms of
    this License, whose permissions for other licensees extend to the
    entire whole, and thus to each and every part regardless of who wrote it.

    Thus, it is not the intent of this section to claim rights or contest
    your rights to work written entirely by you; rather, the intent is to
    exercise the right to control the distribution of derivative or
    collective works based on the Program.

    In addition, mere aggregation of another work not based on the Program
    with the Program (or with a work based on the Program) on a volume of
    a storage or distribution medium does not bring the other work under
    the scope of this License.

    \item
    You may copy and distribute the Program (or a work based on it,
    under Section 2) in object code or executable form under the terms of
    Sections 1 and 2 above provided that you also do one of the following:

    \begin{enumerate}

    \item

    Accompany it with the complete corresponding machine-readable
    source code, which must be distributed under the terms of Sections
    1 and 2 above on a medium customarily used for software interchange; or,

    \item

    Accompany it with a written offer, valid for at least three
    years, to give any third party, for a charge no more than your
    cost of physically performing source distribution, a complete
    machine-readable copy of the corresponding source code, to be
    distributed under the terms of Sections 1 and 2 above on a medium
    customarily used for software interchange; or,

    \item

    Accompany it with the information you received as to the offer
    to distribute corresponding source code.  (This alternative is
    allowed only for noncommercial distribution and only if you
    received the program in object code or executable form with such
    an offer, in accord with Subsection b above.)

    \end{enumerate}


    The source code for a work means the preferred form of the work for
    making modifications to it.  For an executable work, complete source
    code means all the source code for all modules it contains, plus any
    associated interface definition files, plus the scripts used to
    control compilation and installation of the executable.  However, as a
    special exception, the source code distributed need not include
    anything that is normally distributed (in either source or binary
    form) with the major components (compiler, kernel, and so on) of the
    operating system on which the executable runs, unless that component
    itself accompanies the executable.

    If distribution of executable or object code is made by offering
    access to copy from a designated place, then offering equivalent
    access to copy the source code from the same place counts as
    distribution of the source code, even though third parties are not
    compelled to copy the source along with the object code.

    \item
    You may not copy, modify, sublicense, or distribute the Program
    except as expressly provided under this License.  Any attempt
    otherwise to copy, modify, sublicense or distribute the Program is
    void, and will automatically terminate your rights under this License.
    However, parties who have received copies, or rights, from you under
    this License will not have their licenses terminated so long as such
    parties remain in full compliance.

    \item
    You are not required to accept this License, since you have not
    signed it.  However, nothing else grants you permission to modify or
    distribute the Program or its derivative works.  These actions are
    prohibited by law if you do not accept this License.  Therefore, by
    modifying or distributing the Program (or any work based on the
    Program), you indicate your acceptance of this License to do so, and
    all its terms and conditions for copying, distributing or modifying
    the Program or works based on it.

    \item
    Each time you redistribute the Program (or any work based on the
    Program), the recipient automatically receives a license from the
    original licensor to copy, distribute or modify the Program subject to
    these terms and conditions.  You may not impose any further
    restrictions on the recipients' exercise of the rights granted herein.
    You are not responsible for enforcing compliance by third parties to
    this License.

    \item
    If, as a consequence of a court judgment or allegation of patent
    infringement or for any other reason (not limited to patent issues),
    conditions are imposed on you (whether by court order, agreement or
    otherwise) that contradict the conditions of this License, they do not
    excuse you from the conditions of this License.  If you cannot
    distribute so as to satisfy simultaneously your obligations under this
    License and any other pertinent obligations, then as a consequence you
    may not distribute the Program at all.  For example, if a patent
    license would not permit royalty-free redistribution of the Program by
    all those who receive copies directly or indirectly through you, then
    the only way you could satisfy both it and this License would be to
    refrain entirely from distribution of the Program.

    If any portion of this section is held invalid or unenforceable under
    any particular circumstance, the balance of the section is intended to
    apply and the section as a whole is intended to apply in other
    circumstances.

    It is not the purpose of this section to induce you to infringe any
    patents or other property right claims or to contest validity of any
    such claims; this section has the sole purpose of protecting the
    integrity of the free software distribution system, which is
    implemented by public license practices.  Many people have made
    generous contributions to the wide range of software distributed
    through that system in reliance on consistent application of that
    system; it is up to the author/donor to decide if he or she is willing
    to distribute software through any other system and a licensee cannot
    impose that choice.

    This section is intended to make thoroughly clear what is believed to
    be a consequence of the rest of this License.

    \item
    If the distribution and/or use of the Program is restricted in
    certain countries either by patents or by copyrighted interfaces, the
    original copyright holder who places the Program under this License
    may add an explicit geographical distribution limitation excluding
    those countries, so that distribution is permitted only in or among
    countries not thus excluded.  In such case, this License incorporates
    the limitation as if written in the body of this License.

    \item
    The Free Software Foundation may publish revised and/or new versions
    of the General Public License from time to time.  Such new versions will
    be similar in spirit to the present version, but may differ in detail to
    address new problems or concerns.

    Each version is given a distinguishing version number.  If the Program
    specifies a version number of this License which applies to it and ``any
    later version'', you have the option of following the terms and conditions
    either of that version or of any later version published by the Free
    Software Foundation.  If the Program does not specify a version number of
    this License, you may choose any version ever published by the Free Software
    Foundation.

    \item
    If you wish to incorporate parts of the Program into other free
    programs whose distribution conditions are different, write to the author
    to ask for permission.  For software which is copyrighted by the Free
    Software Foundation, write to the Free Software Foundation; we sometimes
    make exceptions for this.  Our decision will be guided by the two goals
    of preserving the free status of all derivatives of our free software and
    of promoting the sharing and reuse of software generally.

    \bigskip{\Large\sc
    No Warranty
    }\bigskip

    \item
    {\sc Because the program is licensed free of charge, there is no warranty
    for the program, to the extent permitted by applicable law.  Except when
    otherwise stated in writing the copyright holders and/or other parties
    provide the program ``as is'' without warranty of any kind, either expressed
    or implied, including, but not limited to, the implied warranties of
    merchantability and fitness for a particular purpose.  The entire risk as
    to the quality and performance of the program is with you.  Should the
    program prove defective, you assume the cost of all necessary servicing,
    repair or correction.}

    \item
    {\sc In no event unless required by applicable law or agreed to in writing
    will any copyright holder, or any other party who may modify and/or
    redistribute the program as permitted above, be liable to you for damages,
    including any general, special, incidental or consequential damages arising
    out of the use or inability to use the program (including but not limited
    to loss of data or data being rendered inaccurate or losses sustained by
    you or third parties or a failure of the program to operate with any other
    programs), even if such holder or other party has been advised of the
    possibility of such damages.}

    \end{enumerate}


    \bigskip{\Large\sc End of Terms and Conditions}\bigskip


    \pagebreak[2]

    \section*{Appendix: How to Apply These Terms to Your New Programs}

    If you develop a new program, and you want it to be of the greatest
    possible use to the public, the best way to achieve this is to make it
    free software which everyone can redistribute and change under these
    terms.

      To do so, attach the following notices to the program.  It is safest to
      attach them to the start of each source file to most effectively convey
      the exclusion of warranty; and each file should have at least the
      ``copyright'' line and a pointer to where the full notice is found.

    \begin{quote}
    one line to give the program's name and a brief idea of what it does. \\
    Copyright (C) yyyy  name of author \\

    This program is free software; you can redistribute it and/or modify
    it under the terms of the GNU General Public License as published by
    the Free Software Foundation; either version 2 of the License, or
    (at your option) any later version.

    This program is distributed in the hope that it will be useful,
    but WITHOUT ANY WARRANTY; without even the implied warranty of
    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
    GNU General Public License for more details.

    You should have received a copy of the GNU General Public License
    along with this program; if not, write to the Free Software
    Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301, USA.
    \end{quote}

    Also add information on how to contact you by electronic and paper mail.

    If the program is interactive, make it output a short notice like this
    when it starts in an interactive mode:

    \begin{quote}
    Gnomovision version 69, Copyright (C) yyyy  name of author \\
    Gnomovision comes with ABSOLUTELY NO WARRANTY; for details type `show w'. \\
    This is free software, and you are welcome to redistribute it
    under certain conditions; type `show c' for details.
    \end{quote}


    The hypothetical commands {\tt show w} and {\tt show c} should show the
    appropriate parts of the General Public License.  Of course, the commands
    you use may be called something other than {\tt show w} and {\tt show c};
    they could even be mouse-clicks or menu items---whatever suits your
    program.

    You should also get your employer (if you work as a programmer) or your
    school, if any, to sign a ``copyright disclaimer'' for the program, if
    necessary.  Here is a sample; alter the names:

    \begin{quote}
    Yoyodyne, Inc., hereby disclaims all copyright interest in the program \\
    `Gnomovision' (which makes passes at compilers) written by James Hacker. \\

    signature of Ty Coon, 1 April 1989 \\
    Ty Coon, President of Vice
    \end{quote}


    This General Public License does not permit incorporating your program
    into proprietary programs.  If your program is a subroutine library, you
    may consider it more useful to permit linking proprietary applications
    with the library.  If this is what you want to do, use the GNU Library
    General Public License instead of this License.

\chapter*{\center\rlap{GNU Free Documentation License}}
\addcontentsline{toc}{chapter}{GNU Free Documentation License}

           Version 1.3, 3 November 2008


     Copyright \copyright{} 2000, 2001, 2002, 2007, 2008  Free Software Foundation, Inc.

     \bigskip

         \path{<https://fsf.org/>}

     \bigskip

     Everyone is permitted to copy and distribute verbatim copies
     of this license document, but changing it is not allowed.




    \bigskip\bigskip{\bf\large Preamble}\bigskip


    The purpose of this License is to make a manual, textbook, or other
    functional and useful document ``free'' in the sense of freedom: to
    assure everyone the effective freedom to copy and redistribute it,
    with or without modifying it, either commercially or noncommercially.
    Secondarily, this License preserves for the author and publisher a way
    to get credit for their work, while not being considered responsible
    for modifications made by others.

    This License is a kind of ``copyleft'', which means that derivative
    works of the document must themselves be free in the same sense. It
    complements the GNU General Public License, which is a copyleft
    license designed for free software.

    We have designed this License in order to use it for manuals for free
    software, because free software needs free documentation: a free
    program should come with manuals providing the same freedoms that the
    software does. But this License is not limited to software manuals;
    it can be used for any textual work, regardless of subject matter or
    whether it is published as a printed book. We recommend this License
    principally for works whose purpose is instruction or reference.



    \bigskip\bigskip{\Large\bf 1. APPLICABILITY AND DEFINITIONS\par}\bigskip


    This License applies to any manual or other work, in any medium, that
    contains a notice placed by the copyright holder saying it can be
    distributed under the terms of this License. Such a notice grants a
    world-wide, royalty-free license, unlimited in duration, to use that
    work under the conditions stated herein. The ``\textbf{Document}'', below,
    refers to any such manual or work. Any member of the public is a
    licensee, and is addressed as ``\textbf{you}''. You accept the license if you
    copy, modify or distribute the work in a way requiring permission
    under copyright law.

    A ``\textbf{Modified Version}'' of the Document means any work containing the
    Document or a portion of it, either copied verbatim, or with
    modifications and/or translated into another language.

    A ``\textbf{Secondary Section}'' is a named appendix or a front-matter section of
    the Document that deals exclusively with the relationship of the
    publishers or authors of the Document to the Document's overall subject
    (or to related matters) and contains nothing that could fall directly
    within that overall subject. (Thus, if the Document is in part a
    textbook of mathematics, a Secondary Section may not explain any
    mathematics.)  The relationship could be a matter of historical
    connection with the subject or with related matters, or of legal,
    commercial, philosophical, ethical or political position regarding
    them.

    The ``\textbf{Invariant Sections}'' are certain Secondary Sections whose titles
    are designated, as being those of Invariant Sections, in the notice
    that says that the Document is released under this License. If a
    section does not fit the above definition of Secondary then it is not
    allowed to be designated as Invariant. The Document may contain zero
    Invariant Sections. If the Document does not identify any Invariant
    Sections then there are none.

    The ``\textbf{Cover Texts}'' are certain short passages of text that are listed,
    as Front-Cover Texts or Back-Cover Texts, in the notice that says that
    the Document is released under this License. A Front-Cover Text may
    be at most 5 words, and a Back-Cover Text may be at most 25 words.

    A ``\textbf{Transparent}'' copy of the Document means a machine-readable copy,
    represented in a format whose specification is available to the
    general public, that is suitable for revising the document
    straightforwardly with generic text editors or (for images composed of
    pixels) generic paint programs or (for drawings) some widely available
    drawing editor, and that is suitable for input to text formatters or
    for automatic translation to a variety of formats suitable for input
    to text formatters. A copy made in an otherwise Transparent file
    format whose markup, or absence of markup, has been arranged to thwart
    or discourage subsequent modification by readers is not Transparent.
    An image format is not Transparent if used for any substantial amount
    of text. A copy that is not ``Transparent'' is called ``\textbf{Opaque}''.

    Examples of suitable formats for Transparent copies include plain
    ASCII without markup, Texinfo input format, LaTeX input format, SGML
    or XML using a publicly available DTD, and standard-conforming simple
    HTML, PostScript or PDF designed for human modification. Examples of
    transparent image formats include PNG, XCF and JPG. Opaque formats
    include proprietary formats that can be read and edited only by
    proprietary word processors, SGML or XML for which the DTD and/or
    processing tools are not generally available, and the
    machine-generated HTML, PostScript or PDF produced by some word
    processors for output purposes only.

    The ``\textbf{Title Page}'' means, for a printed book, the title page itself,
    plus such following pages as are needed to hold, legibly, the material
    this License requires to appear in the title page. For works in
    formats which do not have any title page as such, ``Title Page'' means
    the text near the most prominent appearance of the work's title,
    preceding the beginning of the body of the text.

    The ``\textbf{publisher}'' means any person or entity that distributes
    copies of the Document to the public.

    A section ``\textbf{Entitled XYZ}'' means a named subunit of the Document whose
    title either is precisely XYZ or contains XYZ in parentheses following
    text that translates XYZ in another language. (Here XYZ stands for a
    specific section name mentioned below, such as ``\textbf{Acknowledgements}'',
    ``\textbf{Dedications}'', ``\textbf{Endorsements}'', or ``\textbf{History}''.)
    To ``\textbf{Preserve the Title}''
    of such a section when you modify the Document means that it remains a
    section ``Entitled XYZ'' according to this definition.

    The Document may include Warranty Disclaimers next to the notice which
    states that this License applies to the Document. These Warranty
    Disclaimers are considered to be included by reference in this
    License, but only as regards disclaiming warranties: any other
    implication that these Warranty Disclaimers may have is void and has
    no effect on the meaning of this License.



    \bigskip\bigskip{\Large\bf 2. VERBATIM COPYING\par}\bigskip


    You may copy and distribute the Document in any medium, either
    commercially or noncommercially, provided that this License, the
    copyright notices, and the license notice saying this License applies
    to the Document are reproduced in all copies, and that you add no other
    conditions whatsoever to those of this License. You may not use
    technical measures to obstruct or control the reading or further
    copying of the copies you make or distribute. However, you may accept
    compensation in exchange for copies. If you distribute a large enough
    number of copies you must also follow the conditions in section~3.

    You may also lend copies, under the same conditions stated above, and
    you may publicly display copies.



    \bigskip\bigskip{\Large\bf 3. COPYING IN QUANTITY\par}\bigskip



    If you publish printed copies (or copies in media that commonly have
    printed covers) of the Document, numbering more than 100, and the
    Document's license notice requires Cover Texts, you must enclose the
    copies in covers that carry, clearly and legibly, all these Cover
    Texts: Front-Cover Texts on the front cover, and Back-Cover Texts on
    the back cover. Both covers must also clearly and legibly identify
    you as the publisher of these copies. The front cover must present
    the full title with all words of the title equally prominent and
    visible. You may add other material on the covers in addition.
    Copying with changes limited to the covers, as long as they preserve
    the title of the Document and satisfy these conditions, can be treated
    as verbatim copying in other respects.

    If the required texts for either cover are too voluminous to fit
    legibly, you should put the first ones listed (as many as fit
    reasonably) on the actual cover, and continue the rest onto adjacent
    pages.

    If you publish or distribute Opaque copies of the Document numbering
    more than 100, you must either include a machine-readable Transparent
    copy along with each Opaque copy, or state in or with each Opaque copy
    a computer-network location from which the general network-using
    public has access to download using public-standard network protocols
    a complete Transparent copy of the Document, free of added material.
    If you use the latter option, you must take reasonably prudent steps,
    when you begin distribution of Opaque copies in quantity, to ensure
    that this Transparent copy will remain thus accessible at the stated
    location until at least one year after the last time you distribute an
    Opaque copy (directly or through your agents or retailers) of that
    edition to the public.

    It is requested, but not required, that you contact the authors of the
    Document well before redistributing any large number of copies, to give
    them a chance to provide you with an updated version of the Document.



    \bigskip\bigskip{\Large\bf 4. MODIFICATIONS\par}\bigskip


    You may copy and distribute a Modified Version of the Document under
    the conditions of sections 2 and 3 above, provided that you release
    the Modified Version under precisely this License, with the Modified
    Version filling the role of the Document, thus licensing distribution
    and modification of the Modified Version to whoever possesses a copy
    of it. In addition, you must do these things in the Modified Version:

    \begin{itemize}
    \item[A.]
       Use in the Title Page (and on the covers, if any) a title distinct
       from that of the Document, and from those of previous versions
       (which should, if there were any, be listed in the History section
       of the Document). You may use the same title as a previous version
       if the original publisher of that version gives permission.

    \item[B.]
       List on the Title Page, as authors, one or more persons or entities
       responsible for authorship of the modifications in the Modified
       Version, together with at least five of the principal authors of the
       Document (all of its principal authors, if it has fewer than five),
       unless they release you from this requirement.

    \item[C.]
       State on the Title page the name of the publisher of the
       Modified Version, as the publisher.

    \item[D.]
       Preserve all the copyright notices of the Document.

    \item[E.]
       Add an appropriate copyright notice for your modifications
       adjacent to the other copyright notices.

    \item[F.]
       Include, immediately after the copyright notices, a license notice
       giving the public permission to use the Modified Version under the
       terms of this License, in the form shown in the Addendum below.

    \item[G.]
       Preserve in that license notice the full lists of Invariant Sections
       and required Cover Texts given in the Document's license notice.

    \item[H.]
       Include an unaltered copy of this License.

    \item[I.]
       Preserve the section Entitled ``History'', Preserve its Title, and add
       to it an item stating at least the title, year, new authors, and
       publisher of the Modified Version as given on the Title Page. If
       there is no section Entitled ``History'' in the Document, create one
       stating the title, year, authors, and publisher of the Document as
       given on its Title Page, then add an item describing the Modified
       Version as stated in the previous sentence.

    \item[J.]
       Preserve the network location, if any, given in the Document for
       public access to a Transparent copy of the Document, and likewise
       the network locations given in the Document for previous versions
       it was based on. These may be placed in the ``History'' section.
       You may omit a network location for a work that was published at
       least four years before the Document itself, or if the original
       publisher of the version it refers to gives permission.

    \item[K.]
       For any section Entitled ``Acknowledgements'' or ``Dedications'',
       Preserve the Title of the section, and preserve in the section all
       the substance and tone of each of the contributor acknowledgements
       and/or dedications given therein.

    \item[L.]
       Preserve all the Invariant Sections of the Document,
       unaltered in their text and in their titles. Section numbers
       or the equivalent are not considered part of the section titles.

    \item[M.]
       Delete any section Entitled ``Endorsements''. Such a section
       may not be included in the Modified Version.

    \item[N.]
       Do not retitle any existing section to be Entitled ``Endorsements''
       or to conflict in title with any Invariant Section.

    \item[O.]
       Preserve any Warranty Disclaimers.
    \end{itemize}

    If the Modified Version includes new front-matter sections or
    appendices that qualify as Secondary Sections and contain no material
    copied from the Document, you may at your option designate some or all
    of these sections as invariant. To do this, add their titles to the
    list of Invariant Sections in the Modified Version's license notice.
    These titles must be distinct from any other section titles.

    You may add a section Entitled ``Endorsements'', provided it contains
    nothing but endorsements of your Modified Version by various
    parties---for example, statements of peer review or that the text has
    been approved by an organization as the authoritative definition of a
    standard.

    You may add a passage of up to five words as a Front-Cover Text, and a
    passage of up to 25 words as a Back-Cover Text, to the end of the list
    of Cover Texts in the Modified Version. Only one passage of
    Front-Cover Text and one of Back-Cover Text may be added by (or
    through arrangements made by) any one entity. If the Document already
    includes a cover text for the same cover, previously added by you or
    by arrangement made by the same entity you are acting on behalf of,
    you may not add another; but you may replace the old one, on explicit
    permission from the previous publisher that added the old one.

    The author(s) and publisher(s) of the Document do not by this License
    give permission to use their names for publicity for or to assert or
    imply endorsement of any Modified Version.



    \bigskip\bigskip{\Large\bf 5. COMBINING DOCUMENTS\par}\bigskip



    You may combine the Document with other documents released under this
    License, under the terms defined in section~4 above for modified
    versions, provided that you include in the combination all of the
    Invariant Sections of all of the original documents, unmodified, and
    list them all as Invariant Sections of your combined work in its
    license notice, and that you preserve all their Warranty Disclaimers.

    The combined work need only contain one copy of this License, and
    multiple identical Invariant Sections may be replaced with a single
    copy. If there are multiple Invariant Sections with the same name but
    different contents, make the title of each such section unique by
    adding at the end of it, in parentheses, the name of the original
    author or publisher of that section if known, or else a unique number.
    Make the same adjustment to the section titles in the list of
    Invariant Sections in the license notice of the combined work.

    In the combination, you must combine any sections Entitled ``History''
    in the various original documents, forming one section Entitled
    ``History''; likewise combine any sections Entitled ``Acknowledgements'',
    and any sections Entitled ``Dedications''. You must delete all sections
    Entitled ``Endorsements''.


    \bigskip\bigskip{\Large\bf 6. COLLECTIONS OF DOCUMENTS\par}\bigskip


    You may make a collection consisting of the Document and other documents
    released under this License, and replace the individual copies of this
    License in the various documents with a single copy that is included in
    the collection, provided that you follow the rules of this License for
    verbatim copying of each of the documents in all other respects.

    You may extract a single document from such a collection, and distribute
    it individually under this License, provided you insert a copy of this
    License into the extracted document, and follow this License in all
    other respects regarding verbatim copying of that document.



    \bigskip\bigskip{\Large\bf 7. AGGREGATION WITH INDEPENDENT WORKS\par}\bigskip



    A compilation of the Document or its derivatives with other separate
    and independent documents or works, in or on a volume of a storage or
    distribution medium, is called an ``aggregate'' if the copyright
    resulting from the compilation is not used to limit the legal rights
    of the compilation's users beyond what the individual works permit.
    When the Document is included in an aggregate, this License does not
    apply to the other works in the aggregate which are not themselves
    derivative works of the Document.

    If the Cover Text requirement of section~3 is applicable to these
    copies of the Document, then if the Document is less than one half of
    the entire aggregate, the Document's Cover Texts may be placed on
    covers that bracket the Document within the aggregate, or the
    electronic equivalent of covers if the Document is in electronic form.
    Otherwise they must appear on printed covers that bracket the whole
    aggregate.



    \bigskip\bigskip{\Large\bf 8. TRANSLATION\par}\bigskip



    Translation is considered a kind of modification, so you may
    distribute translations of the Document under the terms of section~4.
    Replacing Invariant Sections with translations requires special
    permission from their copyright holders, but you may include
    translations of some or all Invariant Sections in addition to the
    original versions of these Invariant Sections. You may include a
    translation of this License, and all the license notices in the
    Document, and any Warranty Disclaimers, provided that you also include
    the original English version of this License and the original versions
    of those notices and disclaimers. In case of a disagreement between
    the translation and the original version of this License or a notice
    or disclaimer, the original version will prevail.

    If a section in the Document is Entitled ``Acknowledgements'',
    ``Dedications'', or ``History'', the requirement (section~4) to Preserve
    its Title (section~1) will typically require changing the actual
    title.



    \bigskip\bigskip{\Large\bf 9. TERMINATION\par}\bigskip



    You may not copy, modify, sublicense, or distribute the Document
    except as expressly provided under this License. Any attempt
    otherwise to copy, modify, sublicense, or distribute it is void, and
    will automatically terminate your rights under this License.

    However, if you cease all violation of this License, then your license
    from a particular copyright holder is reinstated (a) provisionally,
    unless and until the copyright holder explicitly and finally
    terminates your license, and (b) permanently, if the copyright holder
    fails to notify you of the violation by some reasonable means prior to
    60 days after the cessation.

    Moreover, your license from a particular copyright holder is
    reinstated permanently if the copyright holder notifies you of the
    violation by some reasonable means, this is the first time you have
    received notice of violation of this License (for any work) from that
    copyright holder, and you cure the violation prior to 30 days after
    your receipt of the notice.

    Termination of your rights under this section does not terminate the
    licenses of parties who have received copies or rights from you under
    this License. If your rights have been terminated and not permanently
    reinstated, receipt of a copy of some or all of the same material does
    not give you any rights to use it.



    \bigskip\bigskip{\Large\bf 10. FUTURE REVISIONS OF THIS LICENSE\par}\bigskip



    The Free Software Foundation may publish new, revised versions
    of the GNU Free Documentation License from time to time. Such new
    versions will be similar in spirit to the present version, but may
    differ in detail to address new problems or concerns. See
    \path{https://www.gnu.org/licenses/}.

    Each version of the License is given a distinguishing version number.
    If the Document specifies that a particular numbered version of this
    License ``or any later version'' applies to it, you have the option of
    following the terms and conditions either of that specified version or
    of any later version that has been published (not as a draft) by the
    Free Software Foundation. If the Document does not specify a version
    number of this License, you may choose any version ever published (not
    as a draft) by the Free Software Foundation. If the Document
    specifies that a proxy can decide which future versions of this
    License can be used, that proxy's public statement of acceptance of a
    version permanently authorizes you to choose that version for the
    Document.



    \bigskip\bigskip{\Large\bf 11. RELICENSING\par}\bigskip



    ``Massive Multiauthor Collaboration Site'' (or ``MMC Site'') means any
    World Wide Web server that publishes copyrightable works and also
    provides prominent facilities for anybody to edit those works. A
    public wiki that anybody can edit is an example of such a server. A
    ``Massive Multiauthor Collaboration'' (or ``MMC'') contained in the
    site means any set of copyrightable works thus published on the MMC
    site.

    ``CC-BY-SA'' means the Creative Commons Attribution-Share Alike 3.0
    license published by Creative Commons Corporation, a not-for-profit
    corporation with a principal place of business in San Francisco,
    California, as well as future copyleft versions of that license
    published by that same organization.

    ``Incorporate'' means to publish or republish a Document, in whole or
    in part, as part of another Document.

    An MMC is ``eligible for relicensing'' if it is licensed under this
    License, and if all works that were first published under this License
    somewhere other than this MMC, and subsequently incorporated in whole
    or in part into the MMC, (1) had no cover texts or invariant sections,
    and (2) were thus incorporated prior to November 1, 2008.

    The operator of an MMC Site may republish an MMC contained in the site
    under CC-BY-SA on the same site at any time before August 1, 2009,
    provided the MMC is eligible for relicensing.



    \bigskip\bigskip{\Large\bf ADDENDUM: How to use this License for your documents\par}\bigskip


    To use this License in a document you have written, include a copy of
    the License in the document and put the following copyright and
    license notices just after the title page:

    \bigskip
    \begin{quote}
        Copyright \copyright{}  YEAR  YOUR NAME.
        Permission is granted to copy, distribute and/or modify this document
        under the terms of the GNU Free Documentation License, Version 1.3
        or any later version published by the Free Software Foundation;
        with no Invariant Sections, no Front-Cover Texts, and no Back-Cover Texts.
        A copy of the license is included in the section entitled ``GNU
        Free Documentation License''.
    \end{quote}
    \bigskip

    If you have Invariant Sections, Front-Cover Texts and Back-Cover Texts,
    replace the ``with \dots\ Texts.''\ line with this:

    \bigskip
    \begin{quote}
        with the Invariant Sections being LIST THEIR TITLES, with the
        Front-Cover Texts being LIST, and with the Back-Cover Texts being LIST.
    \end{quote}
    \bigskip

    If you have Invariant Sections without Cover Texts, or some other
    combination of the three, merge those two alternatives to suit the
    situation.

    If your document contains nontrivial examples of program code, we
    recommend releasing these examples in parallel under your choice of
    free software license, such as the GNU General Public License,
    to permit their use in free software.

\end{document}