| Message ID | 20240816081908.467810-3-hch@lst.de (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | [1/2] xfs: remove a stale comment in xfs_ioc_trim | expand |
On Fri, Aug 16, 2024 at 10:18:43AM +0200, Christoph Hellwig wrote: > Fix the newly added discard support to only offer a FITRIM range that > spans the RT device in addition to the main device if the RT device > actually supports discard. Without this we'll incorrectly accept > a larger range than actually supported and confuse user space if the > RT device does not support discard. This can easily happen when the > main device is a SSD but the RT device is a hard driver. > > Move the code around a bit to keep the max_blocks and granularity > assignments together and explicitly reject that case where only the > RT device supports discard, as that does not fit the way the FITRIM ABI > works very well and is a very fringe use case. Is there more to this than generic/260 failing? And if not, does the following patch things up for you? --D From: Darrick J. Wong <djwong@kernel.org> Subject: [PATCH] generic/260: fix for multi-device xfs with mixed discard support Fix this test so that it can handle XFS filesystems with a realtime volume when the data and rt devices do not both support discard. Signed-off-by: Darrick J. Wong <djwong@kernel.org> --- common/rc | 12 ++++++++++- common/xfs | 65 +++++++++++++++++++++++++++++++++++++++++++++++++----------- 2 files changed, 64 insertions(+), 13 deletions(-) diff --git a/common/rc b/common/rc index b718030a59..426f2de43b 100644 --- a/common/rc +++ b/common/rc @@ -4207,6 +4207,16 @@ _require_batched_discard() fi } +_bdev_queue_property() +{ + local dev="$1" + local property="$2" + local default="$3" + + local fname="/sys/block/$(_short_dev "$dev")/queue/$property" + cat "$fname" 2>/dev/null || echo "$default" +} + # Given a mountpoint and the device associated with that mountpoint, return the # maximum start offset that the FITRIM command will accept, in units of 1024 # byte blocks. @@ -4214,7 +4224,7 @@ _discard_max_offset_kb() { case "$FSTYP" in xfs) - _xfs_discard_max_offset_kb "$1" + _xfs_discard_max_offset_kb "$@" ;; *) $DF_PROG -k | awk -v dev="$2" -v mnt="$1" '$1 == dev && $7 == mnt { print $3 }' diff --git a/common/xfs b/common/xfs index 9501adac4c..5ad60a3ddc 100644 --- a/common/xfs +++ b/common/xfs @@ -1882,7 +1882,13 @@ _require_xfs_scratch_atomicswap() # of 1024 byte blocks. _xfs_discard_max_offset_kb() { + local mount="$1" + local dev="$2" local statfs + local datadev_discard= + local rtdev_discard= + local dev_discard_max= + local rtdev_discard_max= # Use awk to read the statfs output for the XFS filesystem, compute # the two possible FITRIM offset maximums, and then use some horrid @@ -1895,31 +1901,66 @@ _xfs_discard_max_offset_kb() # 2: Realtime volume size in fsblocks. # 3: Max FITRIM offset if we can only trim the data volume # 4: Max FITRIM offset if we can trim the data and rt volumes - readarray -t statfs < <($XFS_IO_PROG -c 'statfs' "$1" | \ - awk '{g[$1] = $3} END {printf("%d\n%d\n%d\n%d\n%d\n", + # 5: Max FITRIM offset if we can only trim the rt volume + readarray -t statfs < <($XFS_IO_PROG -c 'statfs' "$mount" | \ + awk '{g[$1] = $3} END {printf("%d\n%d\n%d\n%d\n%d\n%d\n", g["geom.bsize"], g["geom.datablocks"], g["geom.rtblocks"], g["geom.bsize"] * g["geom.datablocks"] / 1024, - g["geom.bsize"] * (g["geom.datablocks"] + g["geom.rtblocks"]) / 1024);}') + g["geom.bsize"] * (g["geom.datablocks"] + g["geom.rtblocks"]) / 1024, + g["geom.bsize"] * g["geom.rtblocks"] / 1024);}') # If the kernel supports discarding the realtime volume, then it will # not reject a FITRIM for fsblock dblks+1, even if the len/minlen # arguments are absurd. if [ "${statfs[2]}" -gt 0 ]; then - if $FSTRIM_PROG -o "$((statfs[0] * statfs[1]))" \ + case "$dev" in + "$SCRATCH_DEV") + rtdev_discard_max="$(_bdev_queue_property "$SCRATCH_RTDEV" discard_max_bytes 0)" + ;; + "$TEST_DEV") + rtdev_discard_max="$(_bdev_queue_property "$TEST_RTDEV" discard_max_bytes 0)" + ;; + *) + echo "Unrecognized device $dev" >&2 + rtdev_discard_max=0 + ;; + esac + + if [ "$rtdev_discard_max" -gt 0 ] && + $FSTRIM_PROG -o "$((statfs[0] * statfs[1]))" \ -l "${statfs[0]}" \ - -m "$((statfs[0] * 2))" "$1" &>/dev/null; then - # The kernel supports discarding the rt volume, so - # print out the second answer from above. - echo "${statfs[4]}" - return + -m "$((statfs[0] * 2))" "$mount" &>/dev/null; then + # The kernel supports discarding the rt volume + rtdev_discard=2 fi fi - # The kernel does not support discarding the rt volume or there is no - # rt volume. Print out the first answer from above. - echo "${statfs[3]}" + # The kernel supports discarding the rt volume + dev_discard_max="$(_bdev_queue_property "$dev" discard_max_bytes 0)" + if [ "$dev_discard_max" -gt 0 ]; then + datadev_discard=1 + fi + + case "$datadev_discard$rtdev_discard" in + "12") + # Both devices support it + echo "${statfs[4]}" + ;; + "1") + # Only the data device supports it + echo "${statfs[3]}" + ;; + "2") + # Only the rt device supports it + echo "${statfs[5]}" + ;; + *) + # No support at all + echo 0 + ;; + esac } # check if mkfs and the kernel support nocrc (v4) file systems
On Fri, Aug 16, 2024 at 02:50:17PM -0700, Darrick J. Wong wrote: > Is there more to this than generic/260 failing? The only other users of the detection is generic/251, which doesn't seem to care about the exact value. > And if not, does the > following patch things up for you? This works. OTOH it will break again with the zoned RT subvolume which can't support FITRIM even on devices that claim it. And for actual users that care (and not just xfstests) these kinds of hacks don't seem very palatable..
On Mon, Aug 19, 2024 at 02:44:07PM +0200, Christoph Hellwig wrote: > On Fri, Aug 16, 2024 at 02:50:17PM -0700, Darrick J. Wong wrote: > > Is there more to this than generic/260 failing? > > The only other users of the detection is generic/251, which doesn't > seem to care about the exact value. > > > And if not, does the > > following patch things up for you? > > This works. OTOH it will break again with the zoned RT subvolume > which can't support FITRIM even on devices that claim it. And for > actual users that care (and not just xfstests) these kinds of hacks > don't seem very palatable.. What does discard do on a zoned device? Is that how you reset the write pointer? And does that mean that either you tell the device to discard everything it's written in a zone, or it will do nothing? I see that this test really just puts a lower bound on how much FITRIM can report that it discarded from an empty fs. That number is already rather meaningless on XFS because the amount "discarded" is roughly (free space - (amount of pending gc work + discard work already in progress)) and the user has no visibility into the amount of pending gc work. And you can repeatedly FITRIM and it reports the same number since we have no idea if the device actually *did* anything. Hmm. No manpage for FITRIM. Why don't we return the number of bytes in the space map that we iterated as range.len? Or perhaps leave it unchanged? --D
On Mon, Aug 19, 2024 at 08:00:30AM -0700, Darrick J. Wong wrote: > > This works. OTOH it will break again with the zoned RT subvolume > > which can't support FITRIM even on devices that claim it. And for > > actual users that care (and not just xfstests) these kinds of hacks > > don't seem very palatable.. > > What does discard do on a zoned device? Is that how you reset the write > pointer? And does that mean that either you tell the device to discard > everything it's written in a zone, or it will do nothing? On an actual zone device it will probably do nothing. But at least for NVMe the command used to implement discard is mandatory, so all devices will show support. We also support the zoned mode on conventional devices, but instead of through FITRIM we want to issue it instad of a zone reset when the whole rtg has been garbage collected. > Hmm. No manpage for FITRIM. Why don't we return the number of bytes > in the space map that we iterated as range.len? Or perhaps leave it > unchanged? The above would seem sensible. Not sure if we can still pull it off, though.
On Mon, Aug 19, 2024 at 05:08:04PM +0200, Christoph Hellwig wrote: > On Mon, Aug 19, 2024 at 08:00:30AM -0700, Darrick J. Wong wrote: > > > This works. OTOH it will break again with the zoned RT subvolume > > > which can't support FITRIM even on devices that claim it. And for > > > actual users that care (and not just xfstests) these kinds of hacks > > > don't seem very palatable.. > > > > What does discard do on a zoned device? Is that how you reset the write > > pointer? And does that mean that either you tell the device to discard > > everything it's written in a zone, or it will do nothing? > > On an actual zone device it will probably do nothing. But at least for > NVMe the command used to implement discard is mandatory, so all > devices will show support. We also support the zoned mode on > conventional devices, but instead of through FITRIM we want to issue > it instad of a zone reset when the whole rtg has been garbage collected. > > > Hmm. No manpage for FITRIM. Why don't we return the number of bytes > > in the space map that we iterated as range.len? Or perhaps leave it > > unchanged? > > The above would seem sensible. Not sure if we can still pull it > off, though. It seems to have survived testing on TOT overnight, so I'll bake it into djwong-dev when I go through and remove the rtgroups/rtsb feature bits today. And I guess the rtgroups xarray conversion too. --D
On Tue, Aug 20, 2024 at 09:19:55AM -0700, Darrick J. Wong wrote: > It seems to have survived testing on TOT overnight, so I'll bake it into > djwong-dev Note that this a fix for the new RT discard code in 6.11 and it would be kinda nice to get the fix into the 6.11 tree so that we don't have a release with the current state.
diff --git a/fs/xfs/xfs_discard.c b/fs/xfs/xfs_discard.c index d56efe9eae2cef..77d9d2b39f9b00 100644 --- a/fs/xfs/xfs_discard.c +++ b/fs/xfs/xfs_discard.c @@ -685,27 +685,18 @@ xfs_ioc_trim( struct xfs_mount *mp, struct fstrim_range __user *urange) { - unsigned int granularity = - bdev_discard_granularity(mp->m_ddev_targp->bt_bdev); + struct block_device *bdev = mp->m_ddev_targp->bt_bdev; struct block_device *rt_bdev = NULL; + unsigned int granularity = bdev_discard_granularity(bdev); + xfs_rfsblock_t max_blocks = mp->m_sb.sb_dblocks; struct fstrim_range range; xfs_daddr_t start, end; xfs_extlen_t minlen; - xfs_rfsblock_t max_blocks; uint64_t blocks_trimmed = 0; int error, last_error = 0; if (!capable(CAP_SYS_ADMIN)) return -EPERM; - if (mp->m_rtdev_targp && - bdev_max_discard_sectors(mp->m_rtdev_targp->bt_bdev)) - rt_bdev = mp->m_rtdev_targp->bt_bdev; - if (!bdev_max_discard_sectors(mp->m_ddev_targp->bt_bdev) && !rt_bdev) - return -EOPNOTSUPP; - - if (rt_bdev) - granularity = max(granularity, - bdev_discard_granularity(rt_bdev)); /* * We haven't recovered the log, so we cannot use our bnobt-guided @@ -714,13 +705,36 @@ xfs_ioc_trim( if (xfs_has_norecovery(mp)) return -EROFS; + /* + * If the main device doesn't support discards, fail the entire ioctl + * as the RT blocks are mapped after the main device blocks, and we'd + * have to fake operation results for the main device for the ABI to + * work. + * + * Note that while a main device that supports discards (SSD) and a RT + * device that does not (HDD) is a fairly common setup, the reverse is + * theoretically possible but rather odd, so this is not much of a loss. + */ + if (!bdev_max_discard_sectors(bdev)) + return -EOPNOTSUPP; + + if (mp->m_rtdev_targp) { + rt_bdev = mp->m_rtdev_targp->bt_bdev; + if (!bdev_max_discard_sectors(rt_bdev)) + rt_bdev = NULL; + } + if (rt_bdev) { + max_blocks += mp->m_sb.sb_rblocks; + granularity = + max(granularity, bdev_discard_granularity(rt_bdev)); + } + if (copy_from_user(&range, urange, sizeof(range))) return -EFAULT; range.minlen = max_t(u64, granularity, range.minlen); minlen = XFS_B_TO_FSB(mp, range.minlen); - max_blocks = mp->m_sb.sb_dblocks + mp->m_sb.sb_rblocks; if (range.start >= XFS_FSB_TO_B(mp, max_blocks) || range.minlen > XFS_FSB_TO_B(mp, mp->m_ag_max_usable) || range.len < mp->m_sb.sb_blocksize) @@ -729,12 +743,10 @@ xfs_ioc_trim( start = BTOBB(range.start); end = start + BTOBBT(range.len) - 1; - if (bdev_max_discard_sectors(mp->m_ddev_targp->bt_bdev)) { - error = xfs_trim_datadev_extents(mp, start, end, minlen, - &blocks_trimmed); - if (error) - last_error = error; - } + error = xfs_trim_datadev_extents(mp, start, end, minlen, + &blocks_trimmed); + if (error) + last_error = error; if (rt_bdev && !xfs_trim_should_stop()) { error = xfs_trim_rtdev_extents(mp, start, end, minlen,
Fix the newly added discard support to only offer a FITRIM range that spans the RT device in addition to the main device if the RT device actually supports discard. Without this we'll incorrectly accept a larger range than actually supported and confuse user space if the RT device does not support discard. This can easily happen when the main device is a SSD but the RT device is a hard driver. Move the code around a bit to keep the max_blocks and granularity assignments together and explicitly reject that case where only the RT device supports discard, as that does not fit the way the FITRIM ABI works very well and is a very fringe use case. Fixes: 3ba3ab1f6719 ("xfs: enable FITRIM on the realtime device") Signed-off-by: Christoph Hellwig <hch@lst.de> --- fs/xfs/xfs_discard.c | 50 +++++++++++++++++++++++++++----------------- 1 file changed, 31 insertions(+), 19 deletions(-)