SPI accesses can be slow depending on the setup and the access pattern.
The current SPI hardware setup to cache and prefetch. The alternative
cbfs_load_payload() function takes advantage of the caching in the CPU
because the ROM is cached as write protected as well as the SPI's
hardware's caching/prefetching implementation. The CPU will fetch
consecutive aligned cachelines which will hit the ROM as
cacheline-aligned addresses. Once the payload is mirrored into RAM the
segment loading can take place by reading RAM instead of ROM.
With the alternative cbfs_load_payload() the boot time on a baskingridge
board saves ~100ms. This savings is observed using cbmem.py after
performing warm reboots and looking at TS_SELFBOOT_JUMP (99) entries.
This is booting with a depthcharge payload whose payload file fits
within the SMM_DEFAULT_SIZE (0x10000 bytes).
Datapoints with TS_LOAD_PAYLOAD (90) & TS_SELFBOOT_JUMP (99) cbmem entries:
Baseline Alt
-------- --------
90:3,859,310 (473) 90:3,863,647 (454)
99:3,989,578 (130,268) 99:3,888,709 (25,062)
90:3,899,450 (477) 90:3,860,926 (463)
99:4,029,459 (130,008) 99:3,890,583 (29,657)
90:3,834,600 (466) 90:3,890,564 (465)
99:3,964,535 (129,934) 99:3,920,213 (29,649)
Booted baskingridge many times and observed 100ms reduction in
TS_SELFBOOT_JUMP times (time to load payload).
Change-Id: I27b2dec59ecd469a4906b4179b39928e9201db81
Signed-off-by: Aaron Durbin <adurbin@chromium.org>
Reviewed-on: http://review.coreboot.org/2783
Tested-by: build bot (Jenkins)
Reviewed-by: Stefan Reinauer <stefan.reinauer@coreboot.org>
94998c4d3f