Skip to content

Latest commit

 

History

History
176 lines (142 loc) · 8.11 KB

fast_boot.md

File metadata and controls

176 lines (142 loc) · 8.11 KB
Raw

PC Engines fast boot

Long boot time

PC Engines apu2 platforms are booting coreboot in approximately 2.6s while the memory training takes over 2s. Dump of cbmem timestamps on v4.9.0.2 firmware on apu:

   0:1st timestamp                                     18,886
   1:start of romstage                                 18,958 (71)
   2:before ram initialization                         46,334 (27,376)
   3:after ram initialization                          2,225,788 (2,179,453)
   4:end of romstage                                   2,298,974 (73,185)
   8:starting to load ramstage                         2,299,924 (949)
  15:starting LZMA decompress (ignore for x86)         2,300,264 (340)
  16:finished LZMA decompress (ignore for x86)         2,341,499 (41,234)
   9:finished loading ramstage                         2,343,593 (2,094)
  10:start of ramstage                                 2,343,929 (335)
  30:device enumeration                                2,343,937 (8)
  40:device configuration                              2,347,291 (3,354)
  50:device enable                                     2,349,098 (1,806)
  60:device initialization                             2,381,455 (32,356)
  70:device setup done                                 2,452,309 (70,853)
  75:cbmem post                                        2,472,769 (20,460)
  80:write tables                                      2,472,773 (3)
  85:finalize chips                                    2,475,425 (2,652)
  90:load payload                                      2,475,637 (211)
  15:starting LZMA decompress (ignore for x86)         2,475,906 (269)
  16:finished LZMA decompress (ignore for x86)         2,514,778 (38,871)
  99:selfboot jump                                     2,514,808 (30)

Memory Context Restore

AGESA has an option called Memory Context Restore which restores the configuration of the memory controller from non-volatile storage. The feature should dramatically reduce of memory training time after first boot.

  IN BOOLEAN MemRestoreCtl;        ///< Memory context restore control
                                   ///<   FALSE = perform memory init as normal (AMD default)
                                   ///<   TRUE = restore memory context and skip training. This requires
                                   ///<          MemContext is valid before AmdInitPost
                                   ///<
  IN BOOLEAN SaveMemContextCtl;    ///< Control switch to save memory context at the end of MemAuto
                                   ///<   TRUE = AGESA will setup MemContext block before exit AmdInitPost
                                   ///<   FALSE = AGESA will not setup MemContext block. Platform is
                                   ///<           expected to call S3Save later in POST if it wants to
                                   ///<           use memory context restore feature.
                                   ///<
  IN OUT AMD_S3_PARAMS MemContext; ///< Memory context block describes the data that platform needs to
                                   ///< save and restore for memory context restore feature to work.
                                   ///< It uses the subset of S3Save block to save/restore. Hence platform
                                   ///< may save only S3 block and uses it for both S3 resume and
                                   ///< memory context restore.
                                   ///<  - If MemRestoreCtl is TRUE, platform needs to pass in MemContext
                                   ///<    before AmdInitPost.
                                   ///<  - If SaveMemContextCtl is TRUE, platform needs to save MemContext
                                   ///<    right after AmdInitPost.

Implementation and testing

Implementation of memory context restore feature contained:

  • passsing correct values to MemRestoreCtl and SaveMemContextCtl in following manner:
    • set MemRestoreCtl to TRUE
    • set SaveMemContextCtl to TRUE if memory context present and valid; FALSE otherwise
  • obtain memory context from AGESA in MemContext structure and save it in SPI flash

Results

At first boot platform behaved correctly, the booting contained normal memory training. However, after a reboot or second boot, the platform began to reset unexpectedly during iPXE booting or normal kernel loading from storage. Such behaviour could point to incorrect memory initialization.

Further debugging of the memory context restoring revealed that the memory context returned by AGESA is inconsistent across boots. The test conditions were as follows:

  • set MemRestoreCtl to TRUE
  • do not save MemContext in SPi flash in order to force memory training after each boot

Dumps of the memory context showed that the content is different each time platform boots. The differences always occurred on fixed offsets in the structure:

  • 0x220 (single byte changes)
  • 0x300 (multiple byte changes in 80-byte long block)
  • 0x390 (multiple byte changes in 80-byte long block)

Example diff of offset 0x300 across two random boots:

00 52 4a 00 00 00 00 00 00 5c 00 5d 00 60 00 62  .RJ......\.].`.b
00 65 00 00 00 70 00 6e 00 77 00 78 00 1b 1c 25  .e...p.n.w.x...%
24 32 30 37 35 2b 00 00 00 14 14 14 14 14 14 14  $2075+..........
14 0b 00 0a 00 13 00 11 00 19 00 00 00 20 00 20  ............. . 
00 25 00 23 00 38 03 38 03 00 00 00 00 03 00 00  .%.#.8.8........
=================
00 52 4a 00 00 00 00 00 00 5c 00 5d 00 60 00 62  .RJ......\.].`.b
00 64 00 00 00 70 00 6e 00 76 00 78 00 1a 1c 24  .d...p.n.v.x...$
24 32 30 37 35 2c 00 00 00 14 14 14 14 14 14 14  $2075,..........
14 0a 00 0a 00 12 00 11 00 1a 00 00 00 20 00 20  ............. . 
00 25 00 23 00 38 03 38 03 00 00 00 00 03 00 00  .%.#.8.8........

The conclusion is that restoring the memory context from non-volatile storage is not reliable. Platform skipped training and initialized memory controller with wrong values taken from context. Without correctly initialized memory the platform could not load any kernel, thus kept resetting unexpectedly.

AMD and community statement

The fast boot topic has been also touched by coreboot community:

https://review.coreboot.org/c/coreboot/+/22727

The most interesting are the comments:

Unfortunately, according to AMD, we can only use the saved training results for S3 resume, and not as part of a fastboot path...

AMD has taken a pretty strong position that an Intel-like "fast boot" is not practical to develop.

Given that it is impossible to use available AGESA binary to implement fast boot feature.

Providing initial training data

AGESA has more parameters that can be specified when calling AmdInitPost. Some of them are:

IN UINT8             *TableBasedAlterations;
IN PSO_TABLE         *PlatformMemoryConfiguration;
IN FORCE_TRAIN_MODE   ForceTrainMode;

PlatformMemoryConfiguration is filled with a set of macros available in AGESA.h. Things like number of DIMMs or channels supported, motherboard layers and signal maps can be set, but it doesn't have any measurable effect. Format for TableBasedAlterations isn't specified in the available documentation, but can be read from source code available in coreboot repository, assuming that AGESA blob follows the same format. ForceTrainMode can force 1D training only, which is slightly faster.

Results

Using all of those settings with data taken from full memory training seems to speed up that process, but the changes are minimal (less than 2%):

type of boot v4.9.0.2 with filled data
cold 2,544,751 2,509,498
warm 2,529,474 2,484,815
reboot 2,494,380 2,470,353

It is possible that there are more not documented, configurable options than those found which could give additional gain. Some of the entries to the TableBasedAlterations can override final data, but they are applied after the training finishes, not instead of the training.

Initial data was obtained from one platform, data obtained from different platforms can be different. Because of that results on other platforms might differ too.