CentOS Stream 10 guest kernel panic (invalid opcode: 0000 in chacha_8block_xor_avx512vl)
#3022
Comments
|
Is this the version where they bumped the microarch? So that it now requires v3 - but not v4 afaik? https://www.centos.org/centos10/ If I understand correctly, that only means they can guarantee AVX2 - but still need to probe AVX512VL if (boot_cpu_has(X86_FEATURE_AVX) &&
boot_cpu_has(X86_FEATURE_AVX2) &&
cpu_has_xfeatures(XFEATURE_MASK_SSE | XFEATURE_MASK_YMM, NULL)) {
static_branch_enable(&chacha_use_avx2);
if (IS_ENABLED(CONFIG_AS_AVX512) &&
boot_cpu_has(X86_FEATURE_AVX512VL) &&
boot_cpu_has(X86_FEATURE_AVX512BW)) /* kmovq */
static_branch_enable(&chacha_use_avx512vl);
}Then again, it might be one of those AVX/AVX2 that are crashing - earlier versions did check for those too. https://developers.redhat.com/articles/2024/01/02/exploring-x86-64-v3-red-hat-enterprise-linux-10 |
This comment was marked as resolved.
This comment was marked as resolved.
|
There is now a 10.0 beta out, not sure it works better than stream 10 did? https://almalinux.org/blog/2024-12-10-almalinux-10-0-beta-now-available/ No cloud images yet, but maybe they will come later? |
This comment was marked as resolved.
This comment was marked as resolved.
According to this page, it comes in two different CPU variants (wrt AVX512): Can you It seems there is some mis-detection of the CPU features (or it has been disabled) https://kitten.repo.almalinux.org/10-kitten/BaseOS/x86_64/os/Packages/ (v3) |
Workaround: specify the cpu type as "Haswell-v4" |
|
Most likely there is some opcode that needs to be in that kernel check, Crashes on first # x12 += counter values 0-3
vpaddd CTR8BL(%rip),%ymm12,%ymm12
vmovdqa64 %ymm0,%ymm16
vmovdqa64 %ymm1,%ymm17
vmovdqa64 %ymm2,%ymm18Also the CentOS Stream kernel (6.12) is different from the EL10 kernel, which for some reason used a non-lts version? https://www.theregister.com/2024/12/13/rhel_almalinux_betas/ The chacha crypto is unmodified, though. (v6.11-rc5 --> v6.12) |
|
I guess RHEL 10 will switch to 6.12 too before its GA? |
|
Some interesting details in the golang issues about AVX512 support:
The instruction fault handler is exactly what we are seeing here... do_error_trap(regs, 0, "invalid opcode", X86_TRAP_UD, SIGILL,
ILL_ILLOPN, error_get_trap_addr(regs));Apparently this is matched with cpuid returning all those features that could be used if you do that song-and-dance, and not just the ones that are available without it (such as SSE and AVX2, i.e. v2 and v3). Maybe it can be masked out? |
|
i.e. we would do the code equivalent of: cpuType:
x86_64: "host,-avx512vl"And then macOS would run without AVX512? Which is better than kernel panic, at least. Can't try it here, but it "worked" for removing AVX2:
i.e. this old host CPU doesn't have AVX512 support (v4) So as a test I (tried to) ran centos-stream-10 with "host,-avx2" |
CentOS Stream 10 guest kernel panics:
lima.yamlserial.logThe text was updated successfully, but these errors were encountered: