Summary: | dev-db/mariadb-10.6.12 fails unit.pfs test | ||
---|---|---|---|
Product: | Gentoo Linux | Reporter: | Paolo Pedroni <paolo.pedroni> |
Component: | Current packages | Assignee: | Gentoo Linux MySQL bugs team <mysql-bugs> |
Status: | UNCONFIRMED --- | ||
Severity: | normal | CC: | arkamar, esigra, hydrapolic |
Priority: | Normal | Keywords: | TESTFAILURE |
Version: | unspecified | ||
Hardware: | All | ||
OS: | Linux | ||
See Also: | https://github.com/MariaDB/server/pull/3038 | ||
Whiteboard: | |||
Package list: | Runtime testing required: | --- | |
Bug Depends on: | |||
Bug Blocks: | 915000 | ||
Attachments: |
mariadb-10.6.11:20230217-102748.log.gz
failed-test.tar.xz |
Description
Paolo Pedroni
2023-02-17 11:26:33 UTC
Created attachment 852192 [details]
failed-test.tar.xz
This archive contains the '/var/tmp/portage/dev-db/mariadb-10.6.11/temp/var-tests/log/unit.pfs/' where the build system saved all the data for the failing test.
Can you please test 10.6.12? (In reply to Tomáš Mózes from comment #2) > Can you please test 10.6.12? I already did, but I forgot to mention it: same failure. Can you please try to remove "-frecord-gcc-switches -march=znver3 -ftree-vectorize -flto=8" from FLAGS in your make.conf? CFLAGS="-O2 -pipe -frecord-gcc-switches -march=znver3 -ftree-vectorize -flto=8" CXXFLAGS="-O2 -pipe -frecord-gcc-switches -march=znver3 -ftree-vectorize -flto=8" FFLAGS="-O2 -pipe -frecord-gcc-switches -march=znver3 -ftree-vectorize -flto=8" ... -> CFLAGS="-O2 -pipe" CXXFLAGS="-O2 -pipe" FFLAGS="-O2 -pipe" ... (In reply to Tomáš Mózes from comment #4) > Can you please try to remove "-frecord-gcc-switches -march=znver3 > -ftree-vectorize -flto=8" from FLAGS in your make.conf? > > CFLAGS="-O2 -pipe -frecord-gcc-switches -march=znver3 -ftree-vectorize > -flto=8" > CXXFLAGS="-O2 -pipe -frecord-gcc-switches -march=znver3 -ftree-vectorize > -flto=8" > FFLAGS="-O2 -pipe -frecord-gcc-switches -march=znver3 -ftree-vectorize > -flto=8" > ... > > -> > > CFLAGS="-O2 -pipe" > CXXFLAGS="-O2 -pipe" > FFLAGS="-O2 -pipe" > ... I tried several combinations, and the culprit is "-march=znver3": with it, tests fail, with all other flags, tests succeed. I have to try with other -march options on other PCs, but today. That's quite concerning. (In reply to Sam James from comment #6) > That's quite concerning. -march=skylake and -march=icelake-client give the same problem on their respective PCs, while -march=amdfam10 on an old Phenom II seems to work fine. I think I found a solution: if I put '-mno-avx' in *FLAGS, tests succeed. dev-db/mariadb-10.11.6 has also this issue. I was able to trigger it with *FLAGS="-O2 -pipe -mavx" The issue affects tests only. They use different allocators in pfs unit tests, some of them unaligned, but -mavx can produce code which requires aligned memory. For me it SIGSEGV here #0 register_mutex_class (name=0x7ffd1d499960 "wait/synch/mutex/1234567890123/X", name_length=32, flags=<optimized out>) at /var/tmp/portage/dev-db/mariadb-10.11.6/work/mysql/storage/perfschema/pfs_instr_class.cc:1120 index = 0 entry = 0x55875f9b2550 entry address is stored in r14 register and the code SIGSEGVs on this instruction: => 0x000055875dd6960d <+381>: vmovdqa %ymm0,0xa0(%r14) The 0x55875f9b2550 address is not aligned to 32. |