On Tue, Jan 31, 2012 at 08:27:56PM -0600, Eric Sandeen wrote: > On 1/31/12 4:41 PM, Dave Chinner wrote: > > On Mon, Jan 30, 2012 at 04:27:51PM -0600, Eric Sandeen wrote: > >> -dd if=/dev/zero of=tmp1 bs=1M >/dev/null 2>&1 > >> -dd if=/dev/zero of=tmp2 bs=4K >/dev/null 2>&1 > >> +# Fill the rest of the fs completely > >> +dd if=/dev/zero of=$SCRATCH_MNT/tmp1 bs=1M >>$seq.full 2>&1 > >> +dd if=/dev/zero of=$SCRATCH_MNT/tmp2 bs=4K >>$seq.full 2>&1 > >> sync > >> +# Last effort, use O_SYNC > >> +dd if=/dev/zero of=$SCRATCH_MNT/tmp3 bs=4K oflag=sync >>$seq.full 2>&1 > >> +# Save space usage info > >> +echo "Post-fill space:" >> $seq.full > >> +df $SCRATCH_MNT >>$seq.full 2>&1 > >> > >> -dd if=/dev/zero of=test seek=1 bs=4K count=2 conv=notrunc >/dev/null 2>&1 > >> +# Now attempt a write into all of the preallocated space > >> +dd if=/dev/zero of=$SCRATCH_MNT/test seek=1 bs=4K count=1024 conv=notrunc >>$seq.full 2>&1 > >> if [ $? -ne 0 ] > >> then > >> - echo "fill prealloc range err" > >> + echo "fill prealloc range error" > >> status=1 > >> exit > >> fi > > > > I'd still like to see this write attempt to trigger nasty behaviours > > like needing to allocate a metadata block for the extent list. I > > suggested randholes, but perhaps this would be easier: > > > > for i in `seq 1 2 1023`; do > > dd if=/dev/zero of=$SCRATCH_MNT/test seek=$i bs=4K count=1 conv=notrunc \ > > >>$seq.full 2>&1 || _fail "failed to write test file" > > done > > > > which will write every second block and so only convert every second > > block from unwritten to written and hence blow out the size of the > > extent list and require extent map block allocation and potentially > > trigger ENOSPC that way.... > > but will also only write half the actual preallocated space. I guess it > depends on what we are trying to test - reservations for metadata or proper > accounting of prealloc'd space itself...? So then go back and fill in the unwritten blocks. i.e: for i in `seq 1 2 1023`; do dd if=/dev/zero of=$SCRATCH_MNT/test seek=$i bs=4K count=1 conv=notrunc \ >>$seq.full 2>&1 || _fail "failed to write test file" done sync for i in `seq 2 2 1023`; do dd if=/dev/zero of=$SCRATCH_MNT/test seek=$i bs=4K count=1 conv=notrunc \ >>$seq.full 2>&1 || _fail "failed to fill test file" done sync > (and really, what is the proper amount of preallocated metadata space; if > I subbed 2G for 4M, would anything survive the test above?) Given that the test is only using a 1GB filesystem.... ;) But even if it was 2GB, then I'm pretty sure that XFS would handle it fine. Take a long to run, though, with all those individual 4k IOs. My point was simply to determine if the filesystem will be able to hold up the guarantees posix_fallocate() tries to give applications, and so enough extents simple to force a single metadata block allocation is all that is required. Having a 1000 extent state changes shoul dbe sufficient to cause that.... Cheers, Dave. -- Dave Chinner david@xxxxxxxxxxxxx _______________________________________________ xfs mailing list xfs@xxxxxxxxxxx http://oss.sgi.com/mailman/listinfo/xfs
