Yocto Autobuilder: Latency Monitor and AB-INT - Meeting notes: June 10, 2021
(Buried on my desktop, better late than never)
Attendees: Alex, Richard, Saul, Randy, Tony, Trevor
- taskset work-around really helps.
- working on stack_changes bug for qemuarm43
- reproduced on master, hardknott but not gatesgarth
- Let's drop it for now to reduce the AB noise.
task summary script:
- procpath compared to top -bn1
- about the same number of syscalls (strace)
- about 5x more cpu and 2x more ram.
- could be a problem in that we'd need to boostrap since
procpath-native would need to be built at the beginning of each
bitbake run or be added to buildtools or hacked into y-ab-helper.
- it might be simpler to screen-scrape top output...
make: job server
- working and able to build images now.
- testing to confirm that it actually limits the load
- we'll submit for master-next once that's confirmed.
There's a problem in qemu/kernel corruption apparently from
an ltp test. Richard has done lots of work to go back to the
near pristine 5.10 kernel and the problem still happens so
that seems to eliminate the possibility that it's a linux-yocto
Richard has a patch in here:
and to reproduce the problem you should be able to just run:
IMAGE_INSTALL_append = ' ltp'
TEST_SUITES = 'ping ssh ltp' then bitbake core-image-sato; bitbake core-image-sato -c testimage
( qemux86-64, with kvm)
then look for BUG: in log in core-image-sato WORKDIR/testimage/qemu*
Reproducible about 70% of the time.
RP: revert qemu-6.0 to 5.x: The problem persists.
Remaining more frequent issues:
There are 3 patches in master-next that we need to decide on:
- library preload: read qemu libs into memory.
- helpful but maybe treat as a bandaid until we are able to
reduce the IO and cpu spikes.
- qemu-runner? timeout increase 120 -> 240
- ptest timeouts 300 -> 450?
The additional (top and qemu log) output when qemu hangs is merged.
Plans for the week:
Richard: build M1, debug ltp issue.
Alex: full day on ltp tomorrow?
Sakib: task summary
Trevor: make job server
Tony: more valgrind bugs/work-arounds.
Saul: have QMP deal with sigusr1 to close the QMP socket
Randy: test ltp failure