On Mon, Jul 18, 2022 at 02:18:23PM +0800, Qu Wenruo wrote:
> This is a future-proof test mostly for future zoned raid-stripe-tree
> (RST) and P/Q COW based RAID56 implementation.
>
> Unlike regular devices, zoned device can not do overwrite without
> resetting (reclaim) a whole zone.
>
> And for the RST and P/Q COW based RAID56, the idea is to CoW the parity
> stripe to other location.
>
> But all above behaviors introduce some limitation, if we fill the fs,
> then free half of the space interleaved.
>
> - For basic zoned btrfs (aka SINGLE profile for now)
> Normally this means we have no free space at all.
>
> Thankfully zoned btrfs has GC and reserved zones to reclaim those
> half filled zones.
> In theory we should be able to do new writes.
>
> - For future RST with P/Q CoW for RAID56, on non-zoned device.
> This is more complex, in this case, we should have the following
> full stripe layout for every full stripe:
> 0 64K
> Disk A |XXXXXXXXXXXXXXXXXXXXXXXXXXX| (Data 1)
> Disk B | | (Data 2)
> Disk C |XXXXXXXXXXXXXXXXXXXXXXXXXXX| (P stripe)
>
> Although in theory we can write into Disk B, but we have to find
> a free space for the new Parity.
>
> But all other full stripe are like this, which means we're deadlocking
> to find a pure free space without sub-stripe writing.
>
> This means, even for non-zoned btrfs, we still need GC and reserved
> space to handle P/Q CoW properly.
>
> Another thing specific to this test case is, to reduce the runtime, I
> use 256M as the mkfs size for each device.
> (A full run with KASAN enabled kernel already takes over 700 seconds)
>
> So far this can only works for non-zoned disks, as 256M is too small for
> zoned devices to have enough zones.
>
> Thus need extra advice from zoned device guys.
>
> Cc: Johannes Thumshirn <Johannes.Thumshirn@xxxxxxx>
> Signed-off-by: Qu Wenruo <wqu@xxxxxxxx>
> ---
I think this patch need more review from btrfs list. I just review this patch
from fstests side as below ...
> tests/btrfs/261 | 129 ++++++++++++++++++++++++++++++++++++++++++++
> tests/btrfs/261.out | 2 +
> 2 files changed, 131 insertions(+)
> create mode 100755 tests/btrfs/261
> create mode 100644 tests/btrfs/261.out
>
> diff --git a/tests/btrfs/261 b/tests/btrfs/261
> new file mode 100755
> index 00000000..01da4759
> --- /dev/null
> +++ b/tests/btrfs/261
> @@ -0,0 +1,129 @@
> +#! /bin/bash
> +# SPDX-License-Identifier: GPL-2.0
> +# Copyright (C) 2022 SUSE Linux Products GmbH. All Rights Reserved.
> +#
> +# FS QA Test 261
> +#
> +# Make sure all supported profiles (including future zoned RAID56) have proper
> +# way to handle fs with interleaved filled space, and can still write data
> +# into the fs.
> +#
> +# This is mostly inspired by some discussion on P/Q COW for RAID56, even for
> +# regular devices, this can be problematic if we fill the fs then delete
> +# half of the extents interleavedly. Without proper GC and extra reserved
> +# space, such CoW P/Q way should run out of space (even one data stripe is
> +# free, there is no place to CoW its P/Q).
> +#
> +. ./common/preamble
> +_begin_fstest auto enospc raid
> +
> +# Override the default cleanup function.
> +# _cleanup()
> +# {
> +# cd /
> +# rm -r -f $tmp.*
> +# }
This _cleanup looks like nothing special, you can remove it, to use the default
one.
> +
> +# Import common functions.
> +. ./common/filter
> +
> +# real QA test starts here
> +
> +# Modify as appropriate.
^^^
Remove this line please.
> +_supported_fs btrfs
> +# we check scratch dev after each loop
> +_require_scratch_nocheck
> +_require_scratch_dev_pool 4
> +
> +fill_fs()
There's a help named _fill_fs() in common/populate file. I'm not sure if there
are special things in your fill_fs function, better to check if our common
helper can help you?
> +{
> + for (( i = 0;; i += 2 )); do
> + $XFS_IO_PROG -f -c "pwrite 0 64K" $SCRATCH_MNT/file_$i \
> + &> /dev/null
> + if [ $? -ne 0 ]; then
> + break
> + fi
> + $XFS_IO_PROG -f -c "pwrite 0 64K" $SCRATCH_MNT/file_$(($i + 1)) \
> + &> /dev/null
> + if [ $? -ne 0 ]; then
> + break
> + fi
> +
> + # Only sync after data 1M writes.
> + if [ $(( $i % 8)) -eq 0 ]; then
> + sync
> + fi
> + done
> +
> + # Sync what hasn't yet synced.
> + sync
> +
> + echo "fs filled with $i full stripe write" >> $seqres.full
> +
> + # Delete half of the files created above, which should leave
> + # the fs half empty. For RAID56 this would leave all of its full
> + # stripes to be have one full data stripe, one free data stripe,
> + # and one P/Q stripe still in use.
> + rm -rf -- $SCRATCH_MNT/file_*[02468]
> +
> + # Sync to make sure above deleted files really got freed.
> + sync
> +}
> +
> +run_test()
> +{
> + local profile=$1
> + local nr_dev=$2
> +
> + echo "=== profile=$profile nr_dev=$nr_dev ===" >> $seqres.full
> + _scratch_dev_pool_get $nr_dev
> + # -b is for each device.
> + # Here we use 256M to reduce the runtime.
> + _scratch_pool_mkfs -b 256M -m$profile -d$profile >>$seqres.full 2>&1
Do you need to make sure this mkfs successed at here?
> + # make sure we created btrfs with desired options
> + if [ $? -ne 0 ]; then
> + echo "mkfs $mkfs_opts failed"
> + return
> + fi
> + _scratch_mount >>$seqres.full 2>&1
If _scratch_mount fails, the testing will exit directly. So generally we don't
need to fill out stdout/stderr. Or you actually want to use _try_scratch_mount
at here?
> +
> + fill_fs
> +
> + # Now try to write 4M data, with the fs half empty we should be
> + # able to do that.
> + # For zoned devices, this will test if the GC and reserved zones
> + # can handle such cases properly.
> + $XFS_IO_PROG -f -c "pwrite 0 4M" -c sync $SCRATCH_MNT/final_write \
> + >> $seqres.full 2>&1
> + if [ $? -ne 0 ]; then
> + echo "The final write failed"
> + fi
> +
> + _scratch_unmount
> + # we called _require_scratch_nocheck instead of _require_scratch
> + # do check after test for each profile config
> + _check_scratch_fs
> + echo >> $seqres.full
> + _scratch_dev_pool_put
> +}
> +
> +# Here we don't use _btrfs_profile_configs as that doesn't include
> +# the number of devices, but for full stripe writes for RAID56, we
> +# need to ensure nr_data must be 2, so here we manually specify
> +# the profile and number of devices.
> +run_test "single" "1"
> +
> +# Zoned only support
> +if _scratch_btrfs_is_zoned; then
> + exit
I think this "exit" will fail this test directly, due to status=1 currectly.
You can use _require_non_zoned_device() to run this case for non-zoned device
only. Or
if ! _scratch_btrfs_is_zoned;then
run_test "raid0" "2"
run_test "raid1" "2"
run_test "raid10" "4"
run_test "raid5" "3"
run_test "raid6" "4"
fi
As this case is "Silence is golden".
I'm not sure what do you really need at here, can these help?
Thanks,
Zorro
> +fi
> +
> +run_test "raid0" "2"
> +run_test "raid1" "2"
> +run_test "raid10" "4"
> +run_test "raid5" "3"
> +run_test "raid6" "4"
> +
> +echo "Silence is golden"
> +status=0
> +exit
> diff --git a/tests/btrfs/261.out b/tests/btrfs/261.out
> new file mode 100644
> index 00000000..679ddc0f
> --- /dev/null
> +++ b/tests/btrfs/261.out
> @@ -0,0 +1,2 @@
> +QA output created by 261
> +Silence is golden
> --
> 2.36.1
>
