commit 9c1f41ca655d33a4897b1cf2a798284b4795b3c4
author Venkatesh Srinivas <venkateshs@google.com> Tue Oct 27 15:54:16 2020 +0000
committer CQ Bot <commit-bot@chromium.org> Tue Oct 27 15:54:16 2020 +0000
tree d75cabbcbfff641132c0866d3b230d239eb0b736
parent 5f5abb149735c201fc72b7adb5609174d95a9b44

[kernel][x86] Mitigate Intel Skylake Jcc errata

Intel Skylake and Skylake-derived cores have an errata that cause
them to mis-execute control transfer instructions when the instruction
ends on or crosses 32-byte boundaries and the instructions are
executed from the Decode Stream Buffer (DSB). All control transfer
instructions are affected, including macro-fused arithmetic + Jcc
instructions that span a boundary.

Updated microcode has been released, but it comes with a performance
penalty. To avoid the performance penalty and to avoid the errata
where the update is not yet loaded, prevent the compiler from emitting
jumps that would cause the errata.

For more details:
* https://www.intel.com/content/dam/support/us/en/documents/processors/mitigations-jump-conditional-code-erratum.pdf
* https://access.redhat.com/solutions/2019-microcode-nov

This CR only covers the kernel; userland will need to be recompiled to
avoid the errata or performance impact there.

Tested:
0) Compared text size before / after. This increases .text by 55 KB:
readelf -t out/default.zircon/kernel-x64-clang/obj/kernel/zircon.elf

before:
  [ 4] .text
       PROGBITS         ffffffff80101170  0000000000002170  0
       0000000000190140 0000000000000000  0                 16
       [0000000000000006]: ALLOC, EXEC

after:
  [ 4] .text
       PROGBITS         ffffffff80101180  0000000000002180  0
       000000000019daf0 0000000000000000  0                 32
       [0000000000000006]: ALLOC, EXEC

1) Booted on AMD Family 15h; kernel still boots

2) Booted on Intel Skylake-U (Core i5-6260U)
   Tested microcode 0xd6 (Jcc fix, no alignment) vs.
   0xd6 (Jcc fix, alignment):
780 test cases in total
775 test cases had no significant difference (no_sig_diff)
4 test cases got faster
1 test case got slower
0 test cases added
0 test cases removed

Vmo/Clone/ReadCloneAll/512kbytes.clone	 faster	0.609-0.984 2314 +/- 380 ns	1772 +/- 130 ns
Vmo/Clone/ReadCloneAll/512kbytes.close	 faster	0.611-0.939 2208 +/- 298 ns	1662 +/- 131 ns
Vmo/Clone/ReadCloneSome/2048kbytes.close faster	0.567-0.995 2066 +/- 307 ns	1548 +/- 202 ns
Vmo/Clone/ReadOrigAll/512kbytes.close	 faster	0.549-0.995 2272 +/- 474 ns	1648 +/- 140 ns

https://docs.google.com/spreadsheets/d/1FJQ4trU8uJT65Zz7ljMriEhGC-6eCWTSKz6KQIIVvm0/
This is how much performance the s/w fix recovers; it is quite
substantial on a handful of tests.

3) Booted on AMD Zen (R7 1700), perfcompare:
Summary counts:
792 test cases in total
772 test cases had no significant difference (no_sig_diff)
9 test cases got faster
11 test cases got slower
0 test cases added
0 test cases removed

https://docs.google.com/spreadsheets/d/1FJQ4trU8uJT65Zz7ljMriEhGC-6eCWTSKz6KQIIVvm0/
See page two - some tests are slightly affected positively, some slightly negatively.
Largest factor change range was 1.071 - 1.114.

Bug: 44204 Consider building fuchsia w/ 32-byte Jcc fix

Change-Id: I34a558e9f20818c04c20c32c05767326be07cf7d
Reviewed-on: https://fuchsia-review.googlesource.com/c/fuchsia/+/362255
Commit-Queue: Venkatesh Srinivas <venkateshs@google.com>
Reviewed-by: Roland McGrath <mcgrathr@google.com>
Testability-Review: Roland McGrath <mcgrathr@google.com>