--- Begin Message ---
|
Subject: |
[PATCH] wc: Fix crashes due to incomplete AVX2 enumeration |
|
Date: |
Tue, 13 Jun 2023 10:26:10 -0700 |
The AVX2 enumeration for 'wc -l' is incomplete which may cause wc to
crash.
The Intel SDM documents the whole AVX2 enumeration sequence in its
"Detection of Intel AVX2" section. There are three pieces:
1. Ensuring the CPU supports AVX2 instructions
2. Ensuring the OS has enabled XSAVE
3. Ensuring XSAVE is managing the YMM registers (XGETBV)
The existing code does #1 and #2 but misses #3. The kernel will
enable #3 in almost all situations where AVX2 is supported, but there
are situations where the kernel might disable XSAVE YMM support but
leave FP and SSE enabled. This is _unusual_ today, but can occur when
Linux notices XSAVE enumeration issues or in virtual machines with
odd/buggy XSAVE enumeration.
If wc is used on one of those systems in '-l' mode, it will crash:
# cat foo | ./wc.orig -l
Illegal instruction (core dumped)
Fortunately, gcc and llvm provide a builtin to check for CPU support
and AVX2 is supported. Use the builtin instead of using CPUID
directly. This fixes the problem and is also substantially simpler.
The result is that wc no longer crashes on systems where XSAVE YMM
support is disabled:
# cat foo | ./wc.fixed -l
30232
I've reproduced this issue in the current upstream git repo which git
describe says is v9.3-53-g378902407.
I needed to run a patched version of Linux to expose this issue. We
(the upstream Linux x86 maintainers) are considering merging a patch
to do this in the upstream kernel. It would be great if a wc fix was
available.
---
src/wc.c | 36 +-----------------------------------
1 file changed, 1 insertion(+), 35 deletions(-)
diff --git a/src/wc.c b/src/wc.c
index becceda98..944cfc426 100644
--- a/src/wc.c
+++ b/src/wc.c
@@ -157,41 +157,7 @@ static enum total_type total_mode = total_auto;
static bool
avx2_supported (void)
{
- unsigned int eax = 0;
- unsigned int ebx = 0;
- unsigned int ecx = 0;
- unsigned int edx = 0;
- bool getcpuid_ok = false;
- bool avx_enabled = false;
-
- if (__get_cpuid (1, &eax, &ebx, &ecx, &edx))
- {
- getcpuid_ok = true;
- if (ecx & bit_OSXSAVE)
- avx_enabled = true; /* Support is not disabled. */
- }
-
-
- if (avx_enabled)
- {
- eax = ebx = ecx = edx = 0;
- if (! __get_cpuid_count (7, 0, &eax, &ebx, &ecx, &edx))
- getcpuid_ok = false;
- else
- {
- if (! (ebx & bit_AVX2))
- avx_enabled = false; /* Hardware doesn't support it. */
- }
- }
-
-
- if (! getcpuid_ok)
- {
- if (debug)
- error (0, 0, "%s", _("failed to get cpuid"));
- return false;
- }
- else if (! avx_enabled)
+ if (!__builtin_cpu_supports("avx2"))
{
if (debug)
error (0, 0, "%s", _("avx2 support not detected"));
--
2.34.1
--- End Message ---
--- Begin Message ---
|
Subject: |
Re: bug#64058: [PATCH] wc: Fix crashes due to incomplete AVX2 enumeration |
|
Date: |
Thu, 15 Jun 2023 01:22:07 +0100 |
|
User-agent: |
Mozilla Thunderbird |
On 14/06/2023 23:04, Paul Eggert wrote:
On 6/14/23 03:46, Pádraig Brady wrote:
Paul you removed the "avx" check from cksum.c. Was that intended?
No, it's a typo I introduced. Thanks for catching that. Fixed in the
first attached patch.
While looking into this I noticed a couple of other cleanups, fixed in
the other attached patches.
I installed these into coreutils on Savannah.
Changes look good.
I pushed the attached to explicitly document the cksum fix,
and cleanup a syntax-check failure.
Marking this as done.
thanks,
Pádraig
cksum-sigill-news.patch
Description: Text Data
--- End Message ---