Hi Jeremy,
Thanks for posting this. I know that you're planning on doing a presentation about this in this week's leadership meeting and look forward to that. https://coreboot.org/calendar.html
A few questions:
- How does the uGOP driver work with libgfxinit? Does using uGOP mean
that the full GOP driver then needs to be used, or can the system transition back to libgfxinit after memory is initialized.
uGOP does not interact with libgfxinit. However, even though we could not test it, there is no reason to think libgfxinit would not work properly if run after uGOP.
- When is the Graphics Programmer Reference Manuals going to be
published so that the support can be added? Is this planned for next month, next year, or not currently planned, but hoped for?
We do not have a clear commitment that the Meteor Lake Programmer Manual are going to published publicly. i915 public driver code can be used as a reference once it fully supports Meteor Lake graphics generation.
- Is there a reason that the uGOP driver can't be open sourced, at
least once the Graphics Programmer Reference Manuals are released?
We cannot open-source the code for platforms for which the PRMs are not publicly published and we currently do not have a commitment for having RPMs for Meteor Lake.
- When you talk about the differences in time between the uGOP driver
and libgfxinit, is that strictly due to when they are called, or is there some further difference that the uGOP driver is able to accomplish that libgfxinit wouldn't ever be able to do?
This is not due to the way they are being called. We have not investigated the reasons behind why uGOP and GOP also are faster to bring up graphics. This is just an observation.
- Is there a reason that Intel is unwilling to add (or help add) the
required code to libgfxinit, an open source solution that according to your notes should be comparable to the uGOP binary solution? Would Intel be willing to help once the reference manuals are released, or is any cooperation between intel and the community on libgfxinit just not able to happen?
libgfxinit presents some challenges for us: - This component is only used by coreboot which does fit our software platform convergence goal - it is open-source which prevent us from using it as a solution until PRMs are publicly published - It is written in SPARK making it very specific and also most Intel GFX engineers are not familiar with ADA.
- I assume that the uGOP driver is completely optional, and is only
needed to show early signs of life. Is that correct, or could the uGOP driver become mandatory at some point?
uGOP is only an option we want to make available for pre-memory Sign-of-Life use-cases. uGOP, even though it looks like yet another blob, is actually mostly just sharing what is already in FSP-S. We are offering to make it available outside of FSP-M to offer more flexibility in the use-cases for coreboot and chromebook devices.
Please understand that any unhappiness about this plan is not directed at you personally (or *should* not be), but just the idea of adding yet another binary blob to the coreboot boot flow. I've been in this same spot and understand the frustration of just trying to get your work done, while the community is unhappy about the direction of the work being done.
Thanks very much. Martin
Aug 14, 2023, 14:53 by jeremy.compostella@intel.com:
Dear coreboot developers,
With Raptor Lake, we introduced the Pre-Memory Sign-of-Life feature
which displays an on-screen message while firmware components such as coreboot, Firmware Support Package Memory (FSP-M) or, CSME perform long time operations during pre-memory stages.
We propose to take advantage of a proprietary driver Intel already
supports, validates and includes in FSP silicon: the Intel Graphics PEIM (Pre-EFI Initialization Module) driver also known as the GOP (Graphical Output Protocol) driver.
This driver is designed to run in post-memory initialization
stages. Therefore, we derived a new version capable of running in pre-memory stages which we called µGOP. This version is specifically designed to perform graphics legacy VGA initialization.
We intend to keep providing such a binary base solution on the long
run as it addresses our software convergence goals and is compatible with early platform development stage constraints. > libgfxinit https://github.com/coreboot/libgfxinit> supports can always be added later by the open-source community once the Graphics Programmer Reference Manuals are published.
Below, we present the work we performed to run this µGOP driver from
coreboot romstage. It allows to initialize graphics with a very similar flow compared to > libgfxinit https://github.com/coreboot/libgfxinit> use.
Our goal is to start collecting feedback. We will release all the
patches on coreboot.org under the > ugop https://review.coreboot.org/q/topic:ugop> topic soon.
1.> µGOP driver interface
The uGOP PEIM provides the following PEIM-to-PEIM protocol under the 31a4622d-0e21-40a2-80db-c44208fce1b5> GUID.
#define> > PEI_PREMEM_GRAPHICS_PPI_GUID> { \ 0x31a4622d, 0x0e21,
0x40a2, 0x80, 0xdb, 0xc4, 0x42, 0x08, 0xfc, 0xe1, 0xb5 };
The protocol is composed of three fields.
struct> { > UINT32> > Version> ; > PREMEM_PEI_GRAPHICS_INIT> >
PreMemGraphicsPpiInit> ; > PREMEM_PEI_GRAPHICS_EXIT> > PreMemGraphicsPpiExit> ;} > PEI_MICRO_GRAPHICS_PPI> ;
The current > Version> is > 0x00010000> . Where the upper 16 bits
represent the major (1) and the lower 16 bits represent the minor number.
PreMemGraphicsPpiInit()
typedef> EFI_STATUS> (> EFIAPI> *> PREMEM_PEI_GRAPHICS_INIT> ) ( IN VOID> *> Vbt> );
The > PreMemGraphicsPpiInit()> should be supplied with a pointer to
the Video BIOS Table.
PreMemGraphicsPpiExit()> does not take any parameters. This function
must be called to disable VGA graphics configuration once not necessary anymore. Not performing this operation may lead to undesirable behaviour when other graphics stack starts (GOP in FSP-S or Operating System driver).
2.> Integration
As we intend to run µGOP driver in romstage, we want to keep the
required coreboot code as small and efficient as possible. For this reason, we discarded re-using the EDK2 code which would have a major impact on the romstage binary size in addition to adding complication to the build scripts. Instead, we implemented a limited set of Pre-EFI Initialization services. The code is small and designed to accommodate a simple PEIM driver such as µGOP.
2.1.> PEI services
µGOP depends on a limited of PEI services:
InstallPpi()> to install the PEIM Graphics PPI LocatePpi()> to access PEIM-to-PEIM Interface (PPI) Dependencies AllocatePool()> to dynamically allocate memory to handle internal
data structure such as display information …
GetHobList()> and > CreateHob()> to access Hand Off Blocks (HOB) holding runtime data ReportStatusCode()> to report debug information which coreboot
prints using > printk> .
Those services implemented in coreboot are pretty straightforward
and fit in less than 300 lines of code.
2.2.> PEI services pointer
µGOP expects to find the PEI services pointer in the architecture
size word immediately preceding the Interrupt Descriptor Table (IDT) (cf. > Platform Initialization (PI) Specification https://uefi.org/sites/default/files/resources/PI_Spec_1_6.pdf> > 5.4 PEI Services Table Retrieval> ). Since > coreboot x86/exception https://github.com/coreboot/coreboot/blob/master/src/arch/x86/exception.c> module already sets up the IDT and as we do not want to disrupt this configuration we create a copy of the IDT. But we include an extra architecture size word preceding the table to store the PEI services pointer.
Note that FSP memory installs its own IDT but it backups and
restores the one we have set up. Therefore, there is no risk of having PEI services pointer conflicts.
2.3.> Portable Executable Relocation
As we need to execute the µGOP binary in place, we need to perform a
relocation operation of the Portable Executable binary. Since memory space is limited in the pre-memory stages, it is preferable to perform a static relocation operation during the firmware stitching operation.
Fortunately, most of the logic and code is already available as this
operation is performed on the FSP-M binary. We only have to add explicit support for EFI binaries (cf. > 76762 cbfstool: Add relocation support for EFI binaries https://review.coreboot.org/c/coreboot/+/76762> ).
2.4.> Uncompressed VBT
As µGOP requires the Video BIOS Table (VBT) and since memory space
is limited in the pre-memory stages, it is preferable to keep VBT in uncompressed form in CBFS. We introduced the > CONFIG_VBT_CBFS_COMPRESSION> configuration entry to allow this (cf. > 76816 drivers/intel/gma/Kconfig: Add VBT compression configuration entry https://review.coreboot.org/c/coreboot/+/76816> ).
3.> Code flow
4.> Performances analysis
The analysis below is based on a µGOP binary with eDP and HDMI support.
4.1.> Size impact
When > CONFIG_UGOP_EARLY_GRAPHICS> is set
ugop.efi> is included as a CBFS file romstage> includes extra code: > pei.c> , > ugop.c> and > ux.c vbt.bin> is stored uncompressed instead of lzma compressed CBFS File UGOP_EARLY_GRAPHICS=n (bytes) UGOP_EARLY_GRAPHICS=y (bytes) Delta (bytes) µGOP 0 68448 68448 romstage 126128 136256 10128 vbt.bin 1264 9216 7952 Total 127392 213920 86528
The use of µGOP in coreboot represents a size increase of around 84
KB per region (> COREBOOT> , > FW_MAIN_A> and > FW_MAIN_B> ).
4.2.> Regular Boot time impact (no Sign-of-Life)
On a Meteor Lake Google Rex board, we performed 5 warm reset cycles
(without and with > CONFIG_UGOP_EARLY_GRAPHICS> set) and we collected the > cbmem -t> outputs. We computed the median time of each duration (time between two timestamps) and then performed a comparison with a threshold of 0.5 ms.
Start ID Start Description End ID End Description Delta (ms) 947 CSE received 'CPU Reset Done Ack sent' from PMC 991 Die Management Unit (DMU) load completed +1.0 507 starting to verify body (load+SHA2+RSA) 508 finished loading body +4.7 510 finished verifying body signature (RSA) 511 starting TPM PCR extend -0.8 1030 finished EC verification 1040 finished storage device initialization +1.4
The only relevant impact is the verification of the image (507 →
508): +4.7 ms and it can be explained by the 84 KB size increase of the image.
Overall the boot time impact is about 5 ms and concentrated on > verstage> .
4.3.> Cache-As-Ram
For µGOP to execute properly, we have to provide some memory
allocation services (> allocate_pool> and > create_hob> ). These services relies on cache-as-ram memory reservation (> .bss> section). We looked at the two new object files:
Object file .bss section size (bytes) pei.o 6730 ugop.o 9 6739
With µGOP sign-of-life, there is an extra 7 KB cache-as-ram use.
4.4.> Conclusion
The SPINOR and cache-as-RAM space use along with the boot
performance penalty are limited and comparable to what it would be with libgfxinit.
We also noticed that µGOP is faster to bring-up graphics than
libgfxinit. Indeed, according to previously captured numbers on Raptor Lake compared to some number of µGOP on Meteor Lake, µGOP is three times faster to bring up graphics than libgfxinit on an eDP panel (119 ms vs 373 ms).
5.> Summary
This Sign-of-Life µGOP driver based implementation presents the following advantages:
it needs a limited code addition it has a limited impact on the performance its flow and boot performance impact is comparable to libgfxinit solution it is compatible with our software convergence goals it can be available during new platform development early stages
which help our partners to test the feature and stabilize the platform
Regards,
– Jeremy One Emacs to rule them all