soc/amd/stoneyridge: Revert CAR teardown wbinvd

Change the cache-as-ram teardown to use invd instead of wbinvd.
Save the return and recover the call's return address in
chipset_teardown_car.

CAR teardown had been modified to use wbinvd to send CAR contents
to DRAM backing prior to teardown.  This allowed CAR variables,
stack, and local variables to be preserved while running the
AMD_DISABLE_STACK macro.

Using the wbinvd instruction has the side effect of sending all
dirty cache contents to DRAM and not only our CAR data.  This
would likely cause corruption, e.g. during S3 resume.

Stoney Ridge now uses a postcar stage and this is no longer a
requirement.

BUG=b:64768556

Change-Id: I8e6bcb3947f508b1db1a42fd0714bba70074837a
Signed-off-by: Marshall Dawson <marshalldawson3rd@gmail.com>
Reviewed-on: https://review.coreboot.org/20967
Reviewed-by: Aaron Durbin <adurbin@chromium.org>
Tested-by: build bot (Jenkins) <no-reply@coreboot.org>
This commit is contained in:
Marshall Dawson 2017-08-09 19:59:20 -06:00 committed by Aaron Durbin
parent 18b477ea41
commit d61e832e53
3 changed files with 9 additions and 31 deletions

View File

@ -3,6 +3,8 @@ config SOC_AMD_COMMON_BLOCK_CAR
default n
help
This option allows the SOC to use a standard AMD cache-as-ram (CAR)
implementation. CAR setup is built into bootblock and teardown in
romstage. If it is not used the system must implement these functions
separately.
implementation. CAR setup is built into bootblock and teardown is
in postcar. The teardown procedure does not preserve the stack so
it may not be appropriate for a romstage implementation without
additional consideration. If this option is not used, the SOC must
implement these functions separately.

View File

@ -20,6 +20,7 @@
.globl chipset_teardown_car
chipset_teardown_car:
pop %esp
/* Disable cache */
movl %cr0, %eax
@ -33,4 +34,4 @@ chipset_teardown_car:
andl $(~(CR0_CD | CR0_NW)), %eax
movl %eax, %cr0
ret
jmp *%esp

View File

@ -401,16 +401,6 @@ fam15_enable_stack_hook_exit:
* Return any family specific controls to their 'standard'
* settings for using cache with main memory.
*
* Note: Customized for coreboot:
* A wbinvd is used to send cache to memory. The existing stack is preserved
* at its original location and additional information is preserved (e.g.
* coreboot CAR globals, heap structures, etc.). This implementation should
* NOT be used with S3 resume IF the stack/cache area is not reserved and
* over system memory.
*
* This CPU resume path doesn't use CAR, but be careful if porting to
* other CPUs.
*
* Inputs:
* ESI - [31:24] flags; [15,8]= Node#; [7,0]= core#
* Outputs:
@ -644,12 +634,7 @@ fam15_disable_stack_remote_read_exit:
# Begin critical sequence in which EAX, BX, ECX, and EDX must be preserved.
#--------------------------------------------------------------------------
#--------------------------------------------------------------------------
# Send cache to memory. Preserve stack and coreboot CAR globals.
# This shouldn't be used with S3 resume IF the stack/cache area is
# not reserved and over system memory.
#--------------------------------------------------------------------------
wbinvd
invd
# #.if (bh == 01h) || (bh == 03h) ; Is this TN or KM?
# cmp $01, %bh
@ -1289,17 +1274,7 @@ ClearTheStack: # Stack base is in SS, stack pointer is
.endm
/*****************************************************************************
* AMD_DISABLE_STACK: Implementation is modified for coreboot from
* the original AMD intent. A WBINVD is used in the HOOK
* to send dirty cache contents to DRAM backing before
* disabling cache-as-ram. This is not safe for S3 resume.
*
* todo:
* * rework PI/AGESA source to set DRAM to UC to send
* writes directly to memory
* * move DCACHE_BASE or use postcar stage for teardown
* to eliminate car_migrated problem that will occur
* after wbinvd is changed back to invd
* AMD_DISABLE_STACK:
*
* In:
* none