On Tue, Feb 22, 2022 at 11:20:53PM +0530, Ojaswin Mujoo wrote:
> Kernel currently doesn't support resize of EXT4 mounted
> with sparse_super2 option enabled. Earlier, it used to leave the resize
> incomplete and the fs would be left in an inconsistent state, however commit
> b1489186cc83[1] fixed this to avoid the fs corruption by clearly returning
> -EOPNOTSUPP.
>
> Test to ensure that kernel handles resizing with sparse_super2 correctly. Run
> resize for multiple iterations because this sometimes leads to kernel crash due
> to fs corruption, which we want to detect.
>
> Related commit in mainline:
>
> [1] commit b1489186cc8391e0c1e342f9fbc3eedf6b944c61
>
> ext4: add check to prevent attempting to resize an fs with sparse_super2
>
> Signed-off-by: Ojaswin Mujoo <ojaswin@xxxxxxxxxxxxx>
> ---
>
> I would like to add a few comments on the approach followed in this
> test:
>
> 1. So although we check the return codes of the resize operation for
> proper logging, the test is only considered to be passed if fsck
> passes after the resize. This is because resizing a patched kernel
> always keeps the fs consistent whereas resizing on unpatched kernel
> always corrupts the fs.
>
> 2. I've noticed that running mkfs + resize multiple times on unpatched
> kernel sometimes results in KASAN reporting use-after-free. Hence, if
> we detect the kernel is unpatched (doesn't return EOPNOTSUPP on
> resize) we continue iterating to capture this. In this case, we don't
> run fsck in each iteration but run it only after all iterations are
> complete because _check_scratch_fs exits the test if it fails.
>
> 3. In case we detect patched kernel, we stop iterating, and run fsck to
> confirm success
>
> tests/ext4/056 | 108 +++++++++++++++++++++++++++++++++++++++++++++
> tests/ext4/056.out | 2 +
> 2 files changed, 110 insertions(+)
> create mode 100755 tests/ext4/056
> create mode 100644 tests/ext4/056.out
>
> diff --git a/tests/ext4/056 b/tests/ext4/056
> new file mode 100755
> index 00000000..0f275dea
> --- /dev/null
> +++ b/tests/ext4/056
> @@ -0,0 +1,108 @@
> +#! /bin/bash
> +# SPDX-License-Identifier: GPL-2.0
> +# Copyright (c) 2022 IBM. All Rights Reserved.
> +#
> +# We don't currently support resize of EXT4 filesystems mounted
> +# with sparse_super2 option enabled. Earlier, kernel used to leave the resize
> +# incomplete and the fs would be left into an incomplete state, however commit
> +# b1489186cc83[1] fixed this to avoid the fs corruption by clearly returning
> +# -ENOTSUPP.
> +#
> +# This test ensures that kernel handles resizing with sparse_super2 correctly
> +#
> +# Related commit in mainline:
> +#
> +# [1] commit b1489186cc8391e0c1e342f9fbc3eedf6b944c61
> +# ext4: add check to prevent attempting to resize an fs with sparse_super2
> +#
> +
> +. ./common/preamble
> +_begin_fstest auto ioctl resize quick
> +
> +# real QA test starts here
> +
> +INITIAL_FS_SIZE=1G
> +RESIZED_FS_SIZE=$((2*1024*1024*1024)) # 2G
> +ONLINE_RESIZE_BLOCK_LIMIT=$((256*1024*1024))
> +
> +STOP_ITER=255 # Arbitrary return code
> +
> +_supported_fs ext4
> +_require_scratch_size $(($RESIZED_FS_SIZE/1024))
> +_require_test_program "ext4_resize"
> +
> +log()
> +{
> + echo "$seq: $*" >> $seqres.full 2>&1
> +}
> +
> +do_resize()
> +{
> + _mkfs_dev -E resize=$ONLINE_RESIZE_BLOCK_LIMIT -O sparse_super2 \
> + $SCRATCH_DEV $INITIAL_FS_SIZE >> $seqres.full 2>&1 \
> + || _fail "$MKFS_PROG failed. Exiting"
> +
> + _scratch_mount || _fail "Failed to mount scratch partition. Exiting"
> +
> + local BS=$(_get_block_size $SCRATCH_MNT)
> + local REQUIRED_BLOCKS=$(($RESIZED_FS_SIZE/$BS))
> +
> + local RESIZE_RET
> + local EOPNOTSUPP=95
> +
> + log "Resizing FS"
> + $here/src/ext4_resize $SCRATCH_MNT $REQUIRED_BLOCKS >> $seqres.full 2>&1
> + RESIZE_RET=$?
> +
> + # Test appears to be successful. Stop iterating and confirm if FS is
> + # consistent.
> + if [ $RESIZE_RET = $EOPNOTSUPP ]
> + then
> + log "Resize operation not supported with sparse_super2"
> + log "Threw expected error $RESIZE_RET (EOPNOTSUPP)"
> + return $STOP_ITER
> + fi
> +
> + # Test appears to be unsuccessful. Most probably, the fs is corrupt,
> + # iterate a few more times to further stress test this.
> + log "Something is wrong. Output of resize = $RESIZE_RET. \
> + Expected $EOPNOTSUPP (EOPNOTSUPP)"
> +
> + # unmount after ioctl sometimes fails with "device busy" so add a small
> + # delay
> + sleep 0.2
> + _scratch_unmount >> $seqres.full 2>&1 \
> + || _fail "$UMOUNT_PROG failed. Exiting"
> +}
> +
> +run_test()
> +{
> + local FSCK_RET
> + local ITERS=8
> + local RET=0
> +
> + for i in $(seq 1 $ITERS)
> + do
> + log "----------- Iteration: $i ------------"
> + do_resize
> + RET=$?
> +
> + [ "$RET" = "$STOP_ITER" ] && break
> + done
> +
> + log "-------- Iterations Complete ---------"
> + log "Checking if FS is in consistent state"
> + _check_scratch_fs
_check_scratch_fs will exit the test on failure and print error message,
which will break the golden image, so there's no need to check fsck ret.
> + FSCK_RET=$?
> +
> + [ "$FSCK_RET" -ne "0" ] && \
> + echo "fs corrupt. Check $seqres.full for more details"
> +
> + return $FSCK_RET
So I removed above hunk on commit.
Thanks for the test! And my apology to the HUGE delay on review..
Thanks,
Eryu
> +}
> +
> +echo "Silence is golden"
> +run_test
> +status=$?
> +
> +exit
> diff --git a/tests/ext4/056.out b/tests/ext4/056.out
> new file mode 100644
> index 00000000..6142fcd2
> --- /dev/null
> +++ b/tests/ext4/056.out
> @@ -0,0 +1,2 @@
> +QA output created by 056
> +Silence is golden
> --
> 2.27.0
