On Thu, Feb 08, 2024 at 12:04:45AM -0800, Song Liu wrote: > Hi Benjamin, > > On Mon, Feb 5, 2024 at 7:58 PM Benjamin Marzinski <bmarzins@xxxxxxxxxx> wrote: > > > > On Tue, Feb 06, 2024 at 09:36:18AM +0800, Yu Kuai wrote: > > > Hi! > > > > > > 在 2024/02/06 3:35, Benjamin Marzinski 写道: > > > > Could you run the test with something like > > > > > > > > # make check_local T=lvconvert-repair-raid.sh VERBOSE=1 > out 2>&1 > > > > > > > > and post the output. > > > > > > Attached is the output from my VM. > > > > Instead of running the tests from the lvm2 git repo, if you run > > > > # make -C test install > > > > to install the tests, and then create a results directory and run the > > test from there, do you still see the error in the 6.6 kernel? > > > > # make ~/results > > # cd ~/results > > # lvm2-testsuite --only lvconvert-repair-raid.sh > > > > Running the tests this way will test the installed lvm2 binaries on your > > system, instead of the ones in the lvm2 git repo. They may be compiled > > differently. > > I am not able to get reliable results from shell/lvconvert-repair-raid.sh > either. For 6.6.0 kernel, the test fails. On 6.8-rc1 kernel, the test fails > sometimes. > > Could you please share more information about your test setup? > Specifically: > 1. Which tree/branch/tag are you testing? > 2. What's the .config used in the tests? > 3. How do you run the test suite? One test at a time, or all of them > together? > 4. How do you handle "test passes sometimes" cases? So, I have been able to recreate the case where lvconvert-repair-raid.sh keeps failing. It happens when I tried running the reproducer on a virtual machine made using a cloud image, instead of one that I manually installed. I'm not sure why there is a difference. But I can show you how I can reliably recreate the errors I'm seeing. Create a new Fedora 39 virtual machine with the following commands (I'm not sure if it is possible to reproduce this on a machine using less memory and cpus, but I can try that if you need me to. You probably also want to pick a faster Fedora Mirror for the image location): # virt-install --name repair-test --memory 8192 --vcpus 8 --disk size=40 --graphics none --extra-args "console=ttyS0" --osinfo detect=on,name=fedora-unknown --location https://download.fedoraproject.org/pub/fedora/linux/releases/39/Server/x86_64/os/ Install to the whole virtual drive, using the default LVM partitioning. Then ssh into the VM and run the following commands to setup the lvm2-testsuite and 6.6.0 kernel: # dnf upgrade grub2 # dnf install -y git gcc bc flex make bison openssl openssl-devel dwarves zstd elfutils-libelf-devel libaio-devel lvm2 g++ # git clone git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git # git clone git://sourceware.org/git/lvm2.git # cd ~/lvm2 # ./configure # make # cd ~/linux # git checkout -b ver6.6 v6.6 # cp /boot/config-`uname -r` .config # make olddefconfig # modprobe -a dm_raid dm_delay ext4 raid1 raid10 brd # yes "" | make localmodconfig # make -j8 # make modules_install # make install # reboot ssh back into the VM, and run the following commands to run lvm2-testsuite: # mount -o remount,dev /tmp # cd ~/lvm2 # make check T=lvconvert-repair-raid.sh This should always pass. I ran it 100 times without failure. To test the patched kernel, run: # cd ~/linux # git checkout -b dmraid-fix-v5 v6.8-rc3 # git am ~/dmraid-fix-v5.mbox ***Apply the v5 patches*** # make olddefconfig # make -j8 # make modules_install # make install # reboot Rerun the lvm2-testsuite with the same commands as before: # mount -o remount,dev /tmp # cd ~/lvm2 # make check T=lvconvert-repair-raid.sh This fails about 20% of the time, usually at either line 146 or 164. You can check by running the following command when the test fails. # grep "STACKTRACE()" ~/lvm2/test/results/ndev-vanilla\:shell_lvconvert-repair-raid.sh.txt [ 0:13.152] ## 1 STACKTRACE() called from /root/lvm2/test/shell/lvconvert-repair-raid.sh:146 Let me know if you have any questions, or if this doesn't work for you. -Ben > Thanks, > Song
