Skip to contentSkip to conversation

so_peek_off flakes on new NIPA systems

TEST Jakub Kicinski 2 participants 3 messages Nov 24, 2025

Subsystems (2)
lore
Share & tools
lore

Jakub Kicinski

author maintainer
#0 lore raw

Hi Kuniyuki!

We upgraded our system for NIPA recently to netdev foundation one (as you know). Looks like net/af_unix: so_peek_off is flaking on both debug and non-debug builds quite a lot with:

# # so_peek_off.c:149:two_chunks_overlap_blocking:Expected -1 (-1) != bytes (-1) # # two_chunks_overlap_blocking: Test terminated by assertion
# # FAIL so_peek_off.stream.two_chunks_overlap_blocking

https://netdev-ctrl.bots.linux.dev/logs/vmksft/net-dbg/results/399761/134-so-peek-off/stdout

The newer system is 10-20% faster it's also moved from AWS Linux to Fedora. But I suspect the real reason is that our old system had quietly broken compilation of af_unix selftests
because of Wflex-array-member-not-at-end which AWS Linux gcc doesn't understand:

gcc: error: unrecognized command-line option ‘-Wflex-array-member-not-at-end’
make: *** [../../lib.mk:222: /home/virtme/testing/wt-1/tools/testing/selftests/net/af_unix/diag_uid] Error 1

So effectively we're been running some old copy of af_unix tests since this flag was added.

#1 lore raw
Quoted 7 lines · Jakub Kicinski

Hi Kuniyuki!

We upgraded our system for NIPA recently to netdev foundation one (as you know). Looks like net/af_unix: so_peek_off is flaking on both debug and non-debug builds quite a lot with:

Quoted 5 lines

# # so_peek_off.c:149:two_chunks_overlap_blocking:Expected -1 (-1) != bytes (-1) # # two_chunks_overlap_blocking: Test terminated by assertion
# # FAIL so_peek_off.stream.two_chunks_overlap_blocking

https://netdev-ctrl.bots.linux.dev/logs/vmksft/net-dbg/results/399761/134-so-peek-off/stdout

Thanks for the heads-up!

At this point, a fork()ed child/client is expected to send() data after 1ms while the parent/receiver process blocks at recv() up to 3ms, but it seems recv() got -1 because fork()ed process was not scheduled fast enough ?

Probably 3ms is too short when the host is overloaded. I'll send a patch shortly with .gitignore updates.

Quoted 9 lines

The newer system is 10-20% faster it's also moved from AWS Linux to Fedora. But I suspect the real reason is that our old system had quietly broken compilation of af_unix selftests
because of Wflex-array-member-not-at-end which AWS Linux gcc doesn't understand:

gcc: error: unrecognized command-line option ‘-Wflex-array-member-not-at-end’
make: *** [../../lib.mk:222: /home/virtme/testing/wt-1/tools/testing/selftests/net/af_unix/diag_uid] Error 1

Oh well, maybe the package was gcc (11) instead of gcc14 ? IIRC, Amazon Linux 2023 provides the newer one as a namespaced package.

Quoted 3 lines

So effectively we're been running some old copy of af_unix tests since this flag was added.

Yeah, apparently I no longer run tests on AL :p

and this is not a problem now on Fedora, right ?

author maintainer
#2 lore raw
Quoted 4 lines · On Mon, 24 Nov 2025 11:04:02 -0800 Kuniyuki Iwashima wrote:
Quoted 2 lines

So effectively we're been running some old copy of af_unix tests since this flag was added.

Yeah, apparently I no longer run tests on AL :p

:)

Quoted 1 line

and this is not a problem now on Fedora, right ?

Right, not a problem for us any more.
I was just explaining why we're hitting it now.