Smith Waterman Native crashes about gkl HOT 7 CLOSED

Here's a new reported additional example of smith-waterman related crash:

his issue also pops up during FilterAlignmentArtifacts in GATK 4.1.7.0 (experimental)

*** Error in `java': munmap_chunk(): invalid pointer: 0x00007fb4c8e6e540 ***
======= Backtrace: =========
/lib/x86_64-linux-gnu/libc.so.6(+0x777e5)[0x7fb4cdfee7e5]
/lib/x86_64-linux-gnu/libc.so.6(cfree+0x1a8)[0x7fb4cdffb698]
/cromwell_root/tmp.be1fb8a9/libgkl_smithwaterman4505316410124989699.so(_Z19runSWOnePairBT_avx2iiiiPhS_iiaPcPs+0x338)[0x7fb4ac3cffa8]
/cromwell_root/tmp.be1fb8a9/libgkl_smithwaterman4505316410124989699.so(Java_com_intel_gkl_smithwaterman_IntelSmithWaterman_alignNative+0xd8)[0x7fb4ac3cfbf8]
[0x7fb4b8b95f92]

from gkl.

This user comment on the original GATK ticket might provide some additional insight: broadinstitute/gatk#5690 (comment)

from gkl.

#114 has an inital patch to harden Smithwaterment against malloc faling and fix for AVX specific free.

It is not clear it will fix the core issue but it will no hurt it this class of issue.

from gkl.

The fix inital fix has been landed to master. Without a reproduction case there may be more work to do.

from gkl.

I didn't meen to close this one just yet.

Next steps:

1. see if 0.8.7 GLK fixes some of the issue 2 patches have been landed.

   • Smithwaterment needs over 8MB of memory to run and we now check for malloc failing ; Some of the reports say the issue goes away with more memory and it a BKM to do these checks anyway.
   • _mm_malloc memory is now freeded with __mm_free. the use of free before with this type is officaly "undefined"

2. Run Valgind on 0.8.7 and see if anything pops up.

from gkl.

@lbergelson with the latest versions of GATK/GKL do you still see this issue?

from gkl.

@Kmannth I can't say for sure since I was never actually to reproduce those two exactly. 0.8.7 fixed all of the similar issues that I could reproduce and I haven't heard any new complaints yet. I think we're good to close this and start clean with error reports from 0.8.7 onwards! Thank you.

from gkl.