> On Jan 31, 2023, at 5:02 PM, Benjamin Coddington <bcodding@xxxxxxxxxx> wrote:
>
> On 31 Jan 2023, at 16:15, Chuck Lever III wrote:
>
>> Hi-
>>
>> I upgraded my test client's kernel to v6.2-rc5 and now I get
>> failures during the git regression suite on all NFS versions.
>> I bisected to:
>>
>> 85aa8ddc3818 ("NFS: Trigger the "ls -l" readdir heuristic sooner")
>>
>> The failure looks like:
>>
>> not ok 6 - git am --skip succeeds despite D/F conflict
>> #
>> # test_when_finished "git -C df_plus_edit_edit clean -f" &&
>> # test_when_finished "git -C df_plus_edit_edit reset --hard" &&
>> # (
>> # cd df_plus_edit_edit &&
>> #
>> # git checkout f-edit^0 &&
>> # git format-patch -1 d-edit &&
>> # test_must_fail git am -3 0001*.patch &&
>> #
>> # git am --skip &&
>> # test_path_is_missing .git/rebase-apply &&
>> # git ls-files -u >conflicts &&
>> # test_must_be_empty conflicts
>> # )
>> #
>> # failed 1 among 6 test(s)
>> 1..6
>> make[2]: *** [Makefile:60: t1015-read-index-unmerged.sh] Error 1
>> make[2]: *** Waiting for unfinished jobs....
>>
>> The regression suite is run like this:
>>
>> RESULTS= some random directory under /tmp
>> RELEASE="git-2.37.1"
>>
>> rm -f ${RELEASE}.tar.gz
>> curl --no-progress-meter -O https://mirrors.edge.kernel.org/pub/software/scm/git/${RELEASE}.tar.gz
>> /usr/bin/time tar zxf ${RELEASE}.tar.gz >> ${RESULTS}/git 2>&1
>>
>> cd ${RELEASE}
>> make clean >> ${RESULTS}/git 2>&1
>> /usr/bin/time make -j${THREADS} all doc >> ${RESULTS}/git 2>&1
>>
>> /usr/bin/time make -j${THREADS} test >> ${RESULTS}/git 2>&1
>>
>> On this client, THREADS=12. A single-thread run doesn't seem to
>> trigger a problem. So unfortunately the specific data I have is
>> going to be noisy.
>
> I'll attempt to reproduce this and see what's up. This is an export of
> tmpfs?
Sorry for the delayed response... yes, it is a tmpfs export. I was
trying to recall whether this reproduced on other filesystem types,
and it was just far enough in the past that I couldn't be sure I
remembered correctly.
> If so, I suspect you might be running into tmpfs' unstable cookie
> problem when two processes race through nfs_do_filldir().. and if so, the
> cached listing of the directory on the client won't match a listing on the
> server.
Thanks !
--
Chuck Lever
