On Fri, Jul 14, 2023 at 07:59:00AM -0700, Darrick J. Wong wrote: > From: Darrick J. Wong <djwong@xxxxxxxxxx> > > Mikulas reported that this test became a forkbomb on his system when he > tested it with bcachefs. Unlike XFS and ext4, which have large inodes > consuming hundreds of bytes, bcachefs has very tiny ones. Therefore, it > reports a large number of free inodes on a freshly mounted 1GB fs (~15 > million), which causes this test to try to create 15000 processes. > > There's really no reason to do that -- all this test wanted to do was to > exhaust the number of inodes as quickly as possible using all available > CPUs, and then it ran xfs_repair to try to reproduce a bug. Set the > number of subshells to 4x the CPU count and spread the work among them > instead of forking thousands of processes. > > Reported-by: Mikulas Patocka <mpatocka@xxxxxxxxxx> > Signed-off-by: Darrick J. Wong <djwong@xxxxxxxxxx> > Tested-by: Mikulas Patocka <mpatocka@xxxxxxxxxx> > --- > tests/generic/558 | 18 ++++++++++++------ > 1 file changed, 12 insertions(+), 6 deletions(-) > > diff --git a/tests/generic/558 b/tests/generic/558 > index 4e22ce656b..de5c28d00d 100755 > --- a/tests/generic/558 > +++ b/tests/generic/558 > @@ -39,15 +39,21 @@ _scratch_mkfs_sized $((1024 * 1024 * 1024)) >>$seqres.full 2>&1 > _scratch_mount > > i=0 > -free_inode=`_get_free_inode $SCRATCH_MNT` > -file_per_dir=1000 > -loop=$((free_inode / file_per_dir + 1)) NAK. Here, the old code effectively does: loop = howmany(free_inode, 1000); for i in 0...loop: create_file ... 1000files... & IOWs, it rounds the number of files to create up to the nearest 1000, which I overlooked because I was overloaded and words are easier than resurrecting mathematical concepts from raw formulae. If, say, the 1G fs claims to have 524,288 free inodes, the test will start *525* create_file subshells to create 1000 files each, or 525,000 files. The /new/ code does this instead: nr_cpus=(cpu count * 4) files_per_dir = howmany(free_inodes, nr_cpus) for i in 0..nr_cpus: create_file ... files_per_dir... & If nr_cpu is a factor of free_inodes, we don't do /any/ roundup at all. 524,288 free inodes with 4 CPUs gets you 16 threads and 32768 files per thread. Apparently this is significant somehow, because on a lark I decided to revert the referenced commit and the new code doesn't reliably reproduce the failure when parent pointers are enabled. Reintroducing the "rounding free_inodes up to the nearest 1000" does make it trip, though. Sooooo... I'll have a new version out tomorrow after some testing. Please do not apply this patch until then, unless you are testing bcachefs. --D > +free_inodes=$(_get_free_inode $SCRATCH_MNT) > +nr_cpus=$(( $($here/src/feature -o) * 4 * LOAD_FACTOR )) > +echo "free inodes: $free_inodes nr_cpus: $nr_cpus" >> $seqres.full > + > +if ((free_inodes <= nr_cpus)); then > + nr_cpus=1 > + files_per_dir=$free_inodes > +else > + files_per_dir=$(( (free_inodes + nr_cpus - 1) / nr_cpus )) > +fi > mkdir -p $SCRATCH_MNT/testdir > > echo "Create $((loop * file_per_dir)) files in $SCRATCH_MNT/testdir" >>$seqres.full > -while [ $i -lt $loop ]; do > - create_file $SCRATCH_MNT/testdir $file_per_dir $i >>$seqres.full 2>&1 & > - let i=$i+1 > +for ((i = 0; i < nr_cpus; i++)); do > + create_file $SCRATCH_MNT/testdir $files_per_dir $i >>$seqres.full 2>&1 & > done > wait >
