Hi Étienne,
On Sun, Sep 13, 2020 at 12:58:41PM +0200, Étienne Mollier wrote:
> > But this seems to be kind of a random failure since unstable as well as
> > arm64 and ppc64el in testing are currently OK[3].
>
> Thanks for the notice, I tried running the test in several
> architectures as well. After running a few times autopkgtests,
> I see the intermittent error indeed, at least on amd64, and
> i386. Other architectures I tested (arm64, riscv64) didn't show
> this issue so far; I'm virtualizing them, so maybe Qemu
> interferes.
>
> My kernel traps a general protection fault in libblast.so during
> tblastn execution:
>
> $ sudo dmesg
> [...]
> [1741071.802876] traps: tblastn[3579372] general protection fault
> ip:7ff551327dde sp:7ffdb8072790 error:0 in libblast.so[7ff55130f000+63000]
>
> Looking up into straces, it seems the condition occurs when the
> program frees its memory before handing back the hand to the
> system, so maybe something in the taste of a double free while
> threads are not well aligned. I'm not sure yet of the exact
> reproducer with the tblastn command, here is what I record in my
> `ps` output; some of the files referenced are temporary
> artifacts obtained from former steps of the script:
>
> /usr/bin/tblastn -db_gencode 11 -seg no -db
> /tmp/autopkgtest.DOVsFP/autopkgtest_tmp/exact_match.fasta -query
> /tmp/autopkgtest.DOVsFP/autopkgtest_tmp/output/kaptive_temp_3600033_520819/temp_gene_seqs.fasta
> -num_threads 4 -outfmt 6 qseqid sseqid qstart qend sstart send evalue
> bitscore length pident qlen qseq sseq
Hmmmm, this all sounds complicated. I explicitly put Aaron in To since
he is an expert in the blast code and might have some idea. From my gut
feeling we seem to have face some issue inside either blast or biopython
which became visible by chance when running the kaptive test suite.
> I'm also near finishing packaging Biopython 1.78, so that update
> might add a bit more entropy into these testings.
Thumbs up for working on Biopython 1.78 - lets see what side effect it
will have on the issue above.
Kind regards
Andreas.
--
http://fam-tille.de
