The placement calculation logic in cbfs_add_component() has become quite
a mess, and this patch can only fix that to a limited degree. The
interaction between all the different pathways of how the `offset`
variable can be set and at what point exactly the final placement offset
is decided can get quite convoluted. In particular, one existing problem
is that the offset for a file added with the --align flag is decided
before the convert() function is called, which may change the form (and
thereby the size) of the file again after its location was found --
resulting in a location that ends up being too small, or being unable to
find a location for a file that should fit. This used to be okay under
the assumption that forced alignment should really only be necessary for
use cases like XIP where the file is directly "used" straight from its
location on flash in some way, and those cases can never be compressed
-- however, recent AMD platforms have started using the --align flag to
meet the requirements of their SPI DMA controller and broken this
assumption.
This patch fixes that particular problem and hopefully eliminates a bit
of the convolution by moving the offset decision point in the --align
case after the convert() step. This is safe when the steps in-between
(add_topswap_bootblock() and convert() itself) do not rely on the
location having already been decided by --align before that point. For
the topswap case this is easy, because in practice we always call it
with --base-address (and as far as I can tell that's the only way it was
ever meant to work?) -- so codify that assumption in the function. For
convert() this mostly means that the implementations that do touch the
offset variable (mkstage and FSP) need to ensure they take care of the
alignment themselves. The FSP case is particularly complex so I tried to
rewrite the code in a slightly more straight-forward way and clearly
document the supported cases, which should hopefully make it easier to
see that the offset variable is handled correctly in all of them. For
mkstage the best solution seems to be to only have it touch the offset
variable in the XIP case (where we know compression must be disabled, so
we can rely on it not changing the file size later), and have the extra
space for the stage header directly taken care of by do_cbfs_locate() so
that can happen after convert().
NOTE: This is changing the behavior of `cbfstool add -t fsp` when
neither --base-address nor --xip are passed (e.g. FSP-S). Previously,
cbfstool would implicitly force an alignment of 4K. As far as I can tell
from the comments, this is unnecessary because this binary is loaded
into RAM and CBFS placement does not matter, so I assume this is an
oversight caused by accidentally reusing code that was only meant for
the XIP case.
Signed-off-by: Julius Werner <jwerner@chromium.org>
Change-Id: Ia49a585988f7a74944a6630b77b3ebd79b3a9897
Reviewed-on: https://review.coreboot.org/c/coreboot/+/59877
Tested-by: build bot (Jenkins) <no-reply@coreboot.org>
Reviewed-by: Patrick Georgi <patrick@coreboot.org>
Reviewed-by: Raul Rangel <rrangel@chromium.org>
Reviewed-by: Tim Wawrzynczak <twawrzynczak@chromium.org>