Commit fe6a19d40c ("md/md-bitmap: merge md_bitmap_start_sync() into
bitmap_operations") converted still_degraded from int to bool, but
inverted the assignment in the loop that checks whether the array will
still be degraded after the current device is recovered:
"still_degraded = 1" became "still_degraded = false".
As a result, recovering a device while another mirror is still missing
calls md_bitmap_start_sync() with degraded == false, which clears bitmap
bits that the still-missing device needs. When that device is re-added,
its bitmap-based recovery finds the bits already cleared and skips every
region written while the array was degraded, so it is marked In_sync
while holding stale data: silent corruption.
Reproducer (raid10 near=2, 4 disks, internal bitmap):
- fail and remove one disk of each mirror pair
- write to the degraded array
- re-add both disks and let recovery finish
- "check" reports mismatch_cnt=262272 after 256 MiB of degraded
writes and file contents differ; the second disk's "recovery"
completes in milliseconds because everything is skipped
The same conversion in raid1 got it right (still_degraded = true).
Restore the correct value.
Fixes: fe6a19d40c ("md/md-bitmap: merge md_bitmap_start_sync() into bitmap_operations")
Cc: stable@vger.kernel.org
Signed-off-by: Yunye Zhao <yunye.zhao@linux.alibaba.com>
Reviewed-by: Mykola Marzhan <mykola@meshstor.io>
Reviewed-by: Paul Menzel <pmenzel@molgen.mpg.de>
Reviewed-by: Yu Kuai <yukuai@fygo.io>
Link: https://patch.msgid.link/20260723135535.101995-2-yunye.zhao@linux.alibaba.com
Signed-off-by: Yu Kuai <yukuai@fygo.io>
RAID5 maps array sectors through different geometries before and after the
reshape position. During llbitmap reshape, md core cannot account one bio
against both geometries as a single bitmap range, because the old and new
bitmap mappings can cover different chunks.
Split bios that cross reshape_position before md_account_bio(), so the
bitmap only sees ranges that belong to one side of the reshape boundary.
mddev_bio_split_at_reshape_offset() uses bio_submit_split_bioset(), which
submits the remainder immediately and returns the front split bio. If that
front bio later has to wait for reshape, md_handle_request() must not retry
the original bio pointer, because after the split that pointer is the
already-submitted remainder. Track whether the split happened, clear the
temporary BLK_STS_RESOURCE status after the internal clone completion, and
resubmit the front bio directly after the reshape wait. Keep the old
return-false retry path for unsplit bios, where md_handle_request() still
owns the same bio.
Tested-by: Mykola Marzhan <mykola@meshstor.io>
Link: https://patch.msgid.link/20260802195038.164272-30-yukuai@kernel.org
Signed-off-by: Yu Kuai <yukuai@fygo.io>
llbitmap reshape keeps one live bitmap and cannot safely make an
existing bitmap bit cover a smaller data range.
The llbitmap chunksize itself will not shrink when mddev->chunk_sectors
stays the same or grows. However, shrinking mddev->chunk_sectors shrinks
sectors_per_chunk used by raid5_bitmap_sector_map(). That can shrink the
effective data range covered by each bit across the old and new RAID5
geometry. Reject that reshape while llbitmap is active.
Tested-by: Mykola Marzhan <mykola@meshstor.io>
Link: https://patch.msgid.link/20260802195038.164272-28-yukuai@kernel.org
Signed-off-by: Yu Kuai <yukuai@fygo.io>
llbitmap reshape keeps one live bitmap and cannot safely make an
existing bitmap bit cover a smaller data range.
The llbitmap chunksize itself will not shrink when mddev->chunk_sectors
stays the same or grows. However, shrinking mddev->chunk_sectors can
shrink the effective data range covered by each bit for the RAID10
reshape geometry. Reject that reshape while llbitmap is active.
Tested-by: Mykola Marzhan <mykola@meshstor.io>
Link: https://patch.msgid.link/20260802195038.164272-24-yukuai@kernel.org
Signed-off-by: Yu Kuai <yukuai@fygo.io>
llbitmap_state_machine() can be called with an end bit beyond
llbitmap->chunks. In particular, llbitmap_cond_end_sync() passes
sector >> chunkshift, and sector can reach the tracked boundary
exactly.
Clamp the state-machine range to llbitmap->chunks so it cannot walk
past the tracked bitmap.
Tested-by: Mykola Marzhan <mykola@meshstor.io>
Link: https://patch.msgid.link/20260802195038.164272-23-yukuai@kernel.org
Signed-off-by: Yu Kuai <yukuai@fygo.io>
Merge checkpointed old llbitmap state forward as reshape_position advances
and record the checkpoint remap through reshape_mark().
Normal write accounting can run while the reshape thread checkpoints a new
reshape position. llbitmap_reshape_mark() reads old state bytes, merges them
into destination bits, and writes the result back. If llbitmap_start_write()
or llbitmap_start_discard() updates the same state bytes at the same time,
the two read/modify/write paths can overwrite each other and lose the state
from one side.
Serialize only this state-byte race with a rwlock. Normal I/O takes the read
side around llbitmap_state_machine(), after page active references are raised,
so concurrent normal I/O updates still run in parallel. Reshape checkpointing
takes the write side only while merging the checkpointed range, avoiding page
suspension and avoiding a sleeping mutex in the I/O accounting path.
Tested-by: Mykola Marzhan <mykola@meshstor.io>
Link: https://patch.msgid.link/20260802195038.164272-22-yukuai@kernel.org
Signed-off-by: Yu Kuai <yukuai@fygo.io>
Teach llbitmap to choose old versus new geometry during reshape and to
encode exact bitmap ranges for the active geometry. This is the mapping
groundwork for checkpoint remapping.
Range preparation now distinguishes writes from discards. Normal writes
must cover every touched bitmap chunk, while discards may only mark fully
covered chunks unwritten. Without this distinction, a discard that starts
or ends inside a chunk can make live data look unwritten after the range
has been mapped and floored.
Reproduce that with a RAID1 llbitmap using 128-sector chunks. A discard
starting halfway into chunk 8 with a 128-sector length changed clean bits
from 16352 to 16350 and unwritten bits from 0 to 2, even though no chunk
was fully discarded. With discard-specific range encoding, both counts
stay unchanged for the same test.
Range preparation also clamps the pre-map range in the same coordinate
space as the incoming IO. RAID5 receives array-sector offsets but tracks
llbitmap sync size in component sectors, so steady-state RAID5 must use
bitmap_array_sectors() before mapping and keep the existing sync-size
clamp after mapping.
Reproduce that with a 4-disk RAID5 llbitmap created --assume-clean. A
write below dev_sectors changed dirty bits from 0 to 512, but a write at
seek=2094080 left the count at 512. With the array-sector pre-map limit,
writing at seek=component_size + 65536 increased dirty bits from 512 to
1024.
Link: https://lore.kernel.org/all/20260726185916.2223460-1-mykola@meshstor.io/
Tested-by: Mykola Marzhan <mykola@meshstor.io>
Link: https://patch.msgid.link/20260802195038.164272-20-yukuai@kernel.org
Signed-off-by: Yu Kuai <yukuai@fygo.io>
Commit the staged llbitmap geometry when reshape finishes.
When assembling a stopped reshape, md_run() creates the bitmap before
publishing mddev->pers. llbitmap_read_sb() can therefore only initialize
the reshape fields from the old on-disk sync size. Refresh the staged
reshape geometry again from llbitmap_load(), after mddev->pers is
available, and expand the in-memory page controls before replaying bitmap
state.
Reproduce on the old kernel by creating a RAID10 llbitmap with four
active disks and two spares, growing it to six disks, then stopping and
assembling while reshape is still running. The llbitmap chunk count was
32704 before grow, 49056 during reshape, then rolled back to 32704 after
reassemble.
The fixed kernel kept the target geometry across the same stop/reassemble
flow: 65440 chunks before grow, 98160 during reshape, and 98160 after
reassemble.
Link: https://lore.kernel.org/all/20260726185916.2223460-1-mykola@meshstor.io/
Tested-by: Mykola Marzhan <mykola@meshstor.io>
Link: https://patch.msgid.link/20260802195038.164272-18-yukuai@kernel.org
Signed-off-by: Yu Kuai <yukuai@fygo.io>
Allocate one llbitmap page-control object at a time and free each
object through the same model.
Let llbitmap_read_page() return a zeroed page without reading disk when
the page index is beyond the current bitmap size, so page-control
allocation no longer needs a separate read_existing flag.
This keeps the llbitmap page-control lifetime self-consistent and
prepares the page-cache code for later in-place growth.
Reviewed-by: Su Yue <glass.su@suse.com>
Tested-by: Mykola Marzhan <mykola@meshstor.io>
Link: https://patch.msgid.link/20260802195038.164272-15-yukuai@kernel.org
Signed-off-by: Yu Kuai <yukuai@fygo.io>
Add bitmap mapping and reshape hooks needed by llbitmap reshape
support without teaching md core to account a single bio against
multiple bitmap ranges.
This also adds the old/new bitmap geometry helpers used by
personalities to describe reshape mapping to llbitmap.
Tested-by: Mykola Marzhan <mykola@meshstor.io>
Link: https://patch.msgid.link/20260802195038.164272-13-yukuai@kernel.org
Signed-off-by: Yu Kuai <yukuai@fygo.io>
Add mddev_bio_split_at_reshape_offset() so personalities can share
reshape-offset bio splitting instead of open-coding the same boundary
handling in multiple places.
The helper first applies the optional max_sectors limit. If reshape is
running and the bio crosses reshape_position, it further limits the front
bio to the current reshape boundary so callers can account and submit one
side of the reshape at a time.
Snapshot reshape_position with READ_ONCE(). RAID5 and RAID10 update this
field as reshape progresses, while the I/O path only needs one consistent
decision point for the current bio. Using an explicit single load avoids a
plain lockless access and prevents the compiler from refetching a different
boundary while deciding whether and where to split.
When a split is needed, bio_submit_split_bioset() submits the remainder and
returns the front bio. Callers must therefore continue processing the
returned bio, not the original pointer.
Tested-by: Mykola Marzhan <mykola@meshstor.io>
Link: https://patch.msgid.link/20260802195038.164272-12-yukuai@kernel.org
Signed-off-by: Yu Kuai <yukuai@fygo.io>
mkfs.ext4 can submit zero-sector flush/FUA bios. These bios are WRITE
bios for md_write_start() purposes, but they do not cover any data sector
and must not dirty bitmap bits.
md bitmap accounting currently passes such bios to bitmap start_write().
For llbitmap this reaches llbitmap_start_write() with sectors == 0,
which underflows the end chunk calculation.
Personality bitmap mapping can also turn a non-empty bio into an empty
bitmap range when the requested sectors are outside the active bitmap
geometry. Treat both cases as not started, so the completion path will not
call end_write() for an empty range.
Tested-by: Mykola Marzhan <mykola@meshstor.io>
Link: https://patch.msgid.link/20260802195038.164272-11-yukuai@kernel.org
Signed-off-by: Yu Kuai <yukuai@fygo.io>
llbitmap_destroy() deletes pending_timer before flushing
md_llbitmap_io_wq. However, daemon_work can still be queued or running
after the timer has been deleted, and the daemon path can arm
pending_timer again when it finds dirty chunks that are not ready to
flush yet.
If that happens during teardown, pending_timer can remain armed after
llbitmap is freed and later dereference freed memory.
Add a BITMAP_SHUTDOWN bit to llbitmap->flags, set it before deleting
the timer, and make the timer and daemon paths stop queueing or rearming
work once teardown starts. Cancel daemon_work before flushing the shared
workqueue so no already queued daemon instance can race with the free.
Use timer_shutdown_sync() so a daemon instance that passed the shutdown
check before teardown cannot rearm the timer afterward.
BITMAP_SHUTDOWN is a runtime-only state. Mask it out when reading and
updating the llbitmap superblock so the shutdown state is never loaded
from disk or persisted to disk.
Fixes: 5ab829f197 ("md/md-llbitmap: introduce new lockless bitmap")
Tested-by: Mykola Marzhan <mykola@meshstor.io>
Link: https://patch.msgid.link/20260802195038.164272-10-yukuai@kernel.org
Signed-off-by: Yu Kuai <yukuai@fygo.io>
llbitmap_create() publishes mddev->bitmap before reading the bitmap
superblock. This is needed because llbitmap_read_sb() can initialize a
new bitmap and flush it through helpers that use mddev->bitmap.
If llbitmap_read_sb() fails, the old cleanup dropped bitmap_info.mutex
and freed llbitmap before clearing mddev->bitmap. Readers such as
/proc/mdstat rely on bitmap_info.mutex to keep the bitmap pointer stable
while collecting bitmap stats, so they could observe the stale pointer
after the failed create path released the mutex.
Clear mddev->bitmap while still holding bitmap_info.mutex, then free the
failed llbitmap after dropping the mutex. This makes mutex-protected
readers see either a live bitmap or no bitmap.
Fixes: 5ab829f197 ("md/md-llbitmap: introduce new lockless bitmap")
Tested-by: Mykola Marzhan <mykola@meshstor.io>
Link: https://patch.msgid.link/20260802195038.164272-9-yukuai@kernel.org
Signed-off-by: Yu Kuai <yukuai@fygo.io>
md_clone_bio() always allocates the clone from mddev->io_clone_set, even
when queue I/O stats are disabled. In that case it does not call
bio_start_io_acct(), but it also left md_io_clone->start_time untouched.
The clone private data comes from a mempool and can contain data from a
previous user. md_end_clone_io() checks start_time to decide whether it
needs to call bio_end_io_acct(), so a stale non-zero value can make the
completion path end accounting that was never started for this bio.
Set start_time to 0 in the no-stats branch. This keeps the end path tied
to whether bio_start_io_acct() actually ran.
Fixes: c687297b88 ("md: also clone new io if io accounting is disabled")
Tested-by: Mykola Marzhan <mykola@meshstor.io>
Link: https://patch.msgid.link/20260802195038.164272-8-yukuai@kernel.org
Signed-off-by: Yu Kuai <yukuai@fygo.io>
__md_stop() destroyed the bitmap before calling mddev_detach(). That made
mddev_detach() skip bitmap_ops->wait_behind_writes(), because the bitmap
was already disconnected from mddev.
This was still safe for the legacy bitmap because bitmap_destroy() waits
for behind writes itself. llbitmap keeps that wait in its
->wait_behind_writes() operation instead, while ->destroy() tears down the
llbitmap storage. With the old ordering, RAID1 behind-write completions
could still run after llbitmap storage had been freed.
Call mddev_detach() before md_bitmap_destroy() so the common detach path
can wait for behind writes while the bitmap is still alive. Only destroy
the bitmap after those users are gone.
Fixes: 5ab829f197 ("md/md-llbitmap: introduce new lockless bitmap")
Tested-by: Mykola Marzhan <mykola@meshstor.io>
Link: https://patch.msgid.link/20260802195038.164272-7-yukuai@kernel.org
Signed-off-by: Yu Kuai <yukuai@fygo.io>
raid5_bitmap_sector_map() aligns the array range to full RAID5 stripe
widths before converting it to component sectors. That width is
chunk_sectors multiplied by the number of data disks, and it is not
always a power of two.
Reproduce with a 4-disk RAID5, 1024-sector chunks, and three data disks.
The full-stripe width is 3072 sectors. For a one-sector write at array
sector 3072, correct rounding gives array range [3072, 6144), which maps
to component range [1024, 2048). The old round_down()/round_up() logic
instead gives [1024, 4096), which maps to [0, 1024).
Use sector_div() based arithmetic so the rounded range is aligned to the
actual RAID5 stripe width.
The deterministic mapper test now reports the fixed component range as
[1024, 2048), while the old mask-based range was [0, 1024).
Fixes: 9c89f60447 ("md/raid5: implement pers->bitmap_sector()")
Reported-by: Mykola Marzhan <mykola@meshstor.io>
Link: https://lore.kernel.org/all/20260726185916.2223460-1-mykola@meshstor.io/
Tested-by: Mykola Marzhan <mykola@meshstor.io>
Link: https://patch.msgid.link/20260802195038.164272-6-yukuai@kernel.org
Signed-off-by: Yu Kuai <yukuai@fygo.io>
Sashiko reported that RAID5 can accept a reshape chunk size that becomes
zero sectors. chunk_size_store() stores the sysfs byte value as n >> 9, so
writing a value below 512 bytes sets mddev->new_chunk_sectors to zero.
RAID5 then accepted that pending reshape geometry and raid5_start_reshape()
installed it into conf->chunk_sectors, letting reshape code divide by zero.
Reject zero-sector chunks both in check_reshape(), where normal sysfs
requests are validated, and in raid5_start_reshape(), so assembly/resume
paths also cannot install zero chunk geometry.
Test script: in QEMU, create a plain three-disk RAID5 array with 64K
chunks, write/read back a small pattern, write 1 to
/sys/block/md0/md/chunk_size, add a fourth disk, and run mdadm --grow
--raid-devices=4 --backup-file=... . The script scans dmesg for divide
error/Oops/KASAN signatures.
Bad kernel, eb29914412c3:
echo 1 > /sys/block/md0/md/chunk_size
mdadm --grow /dev/md0 --raid-devices=4 --backup-file=/root/md0-grow.bak
Oops: divide error: 0000 [#1] SMP KASAN NOPTI
RIP: raid5_get_active_stripe+0x863/0xc10
Call Trace:
raid5_sync_request
md_do_sync
md_thread
Kernel panic - not syncing: Fatal exception
Fixed kernel: echo 1 > /sys/block/md0/md/chunk_size bash: echo: write
error: Invalid argument chunk_write_rc=1 grow_rc=skipped RESULT:
REJECTED_ZERO_CHUNK_NO_OOPS
Tested-by: Mykola Marzhan <mykola@meshstor.io>
Link: https://patch.msgid.link/20260802195038.164272-5-yukuai@kernel.org
Signed-off-by: Yu Kuai <yukuai@fygo.io>
llbitmap_cond_end_sync() is called with the sync thread's current sector.
That value is an exclusive progress boundary: sectors below it have
completed, but the llbitmap chunk containing it can still be in progress.
The old code converted that sector directly to the last bit passed to
BitmapActionEndsync. If resync had only advanced part-way into a large
llbitmap chunk, the in-progress chunk was marked synced and flushed before
the rest of the chunk was repaired. A later bitmap-assisted RAID1 resync
could then skip the remainder of that chunk and leave stale mirror data
behind.
This can be reproduced without editing bitmap metadata by creating a large
RAID1 with a lockless bitmap so llbitmap naturally selects a 524288-sector
chunk (with the default 128 KiB bitmap area, an array just over 16 TiB is
enough), making one mirror stale through the normal degraded write/re-add
path, and throttling resync so the daemon checkpoint runs while resync is
still inside the first chunk. On the bad kernel, bit 0 is ended early and a
stale sector later in the same chunk is skipped. With this fix, bit 0
remains Syncing until resync reaches the next chunk boundary.
Round the exclusive progress sector down to the nearest llbitmap chunk
boundary and end only chunks strictly below that boundary. Also honor the
force argument so callers that need an immediate checkpoint are not
suppressed by daemon_sleep.
Tested-by: Mykola Marzhan <mykola@meshstor.io>
Link: https://patch.msgid.link/20260802195038.164272-4-yukuai@kernel.org
Signed-off-by: Yu Kuai <yukuai@fygo.io>
llbitmap allocates its in-memory page cache and page-control structures from
paths that can already be holding MD reconfiguration or bitmap state locks.
For example, component_size_store() takes mddev_lock(), update_size() calls
the personality resize method, and llbitmap_resize() can grow the page cache
through llbitmap_prepare_resize().
Using GFP_KERNEL in those paths allows direct reclaim to enter filesystem or
block I/O while MD resize state is locked. That can recurse back into the
same array and wait on state that cannot make progress until the resize path
finishes.
Use GFP_NOIO for the llbitmap object, cached bitmap pages, page controls,
page-control arrays, and percpu_ref initialization. Leave the explicit
metadata zeroout path unchanged because it is intentional bitmap I/O rather
than reclaim-driven allocation.
Tested-by: Mykola Marzhan <mykola@meshstor.io>
Link: https://patch.msgid.link/20260802195038.164272-3-yukuai@kernel.org
Signed-off-by: Yu Kuai <yukuai@fygo.io>
llbitmap_flush() sets LLPageFlush on each bitmap page before it queues the
daemon worker. The flag tells md_llbitmap_daemon_fn() to ignore the normal
barrier_idle expiry check and clean the page immediately.
The daemon only tested LLPageFlush. Once a page had been flushed explicitly,
the flag stayed set, so later dirty bits on that page also bypassed
barrier_idle and were cleaned the next time the daemon ran. That can make a
new write look clean much earlier than the configured idle window.
Consume LLPageFlush in md_llbitmap_daemon_fn() with test_and_clear_bit() and
use the returned value for the current expiry check. The explicit flush still
forces the current daemon pass, while later writes on the same page wait for
barrier_idle again.
This can be reproduced through normal sysfs operations:
1. Create a small RAID1 with --bitmap=lockless and --assume-clean.
2. Set llbitmap/daemon_sleep=1 and llbitmap/barrier_idle=10.
3. Toggle md/array_state from active to readonly and back to active to call
llbitmap_flush() without destroying the in-memory bitmap.
4. Write one sector and read llbitmap/bits immediately, after 2 seconds,
and after 12 seconds.
On the bad kernel the dirty bit is already clean after 2 seconds. With this
change it remains dirty until the barrier_idle window expires.
Tested-by: Mykola Marzhan <mykola@meshstor.io>
Link: https://patch.msgid.link/20260802195038.164272-2-yukuai@kernel.org
Signed-off-by: Yu Kuai <yukuai@fygo.io>
The following bug has been observed with kernel 7.1.3 after adding a new
rdev to an existing RAID1 array with serialize_policy enabled:
Oops: 0002 [#1]
CPU: 0 UID: 0 PID: 19639 Comm: ext4lazyinit Not tainted 7.1.3-1-default
RIP: _raw_spin_lock_irqsave+0x27/0x50
CR2: 0000000000004960
Call Trace:
wait_for_serialization+0xb9/0x260 [raid1]
raid1_make_request+0x762/0xaff [raid1]
md_handle_request+0x1c9/0x2e0 [md_mod]
The raid1.c code calls wait_for_serialization() if the MD_SERIALIZE_POLICY
is set, and wait_for_serialization assumes that rdev->serial is
initialized. Normally this will be the case for arrays that have
the serialize_policy sysfs attribute set to 1.
But when a new rdev is added to an existing array in bind_rdev_to_array(),
the condition at mddev_create_serial_pool() causes creation of rdev->serial
to be skipped. Fix it.
Fixes: 69b00b5bb2 ("md: introduce a new struct for IO serialization")
Signed-off-by: Martin Wilck <mwilck@suse.com>
Reviewed-by: Mykola Marzhan <mykola@meshstor.io>
Link: https://patch.msgid.link/20260723112741.1206836-1-mwilck@suse.com
Signed-off-by: Yu Kuai <yukuai@fygo.io>
In super_1_load(), sb->bblog_shift is an __u8 type value loaded from on-
disk superblock. It is used for badblocks API badblocks_set() by the
following sequence,
1930 rdev->badblocks.shift = sb->bblog_shift;
1931 for (i = 0 ; i < (sectors << (9-3)) ; i++, bbp++) {
1932 u64 bb = le64_to_cpu(*bbp);
1933 int count = bb & (0x3ff);
1934 u64 sector = bb >> 10;
1935 sector <<= sb->bblog_shift;
1936 count <<= sb->bblog_shift;
1937 if (bb + 1 == 0)
1938 break;
1939 if (!badblocks_set(&rdev->badblocks, sector, count, 1))
1940 return -EINVAL;
1941 }
bb->bblog_shit is in range of 0-255, variable sector is 64bit width, for
an invalid bb->bblog_shit, it is possible to make sector be overflowed
by the following calculation,
1935 sector <<= sb->bblog_shift;
Then in turn when call badblocks_set() at line 1939 with the invalid
rdev->badblocks.shift set at line 1930, may result an overflow inside
_badblocks_clear() in block/badblocks.c.
Although there are many places to call badblocks APIs, the non-zero
shift value is only used in super_1_load(), other places always use 0 as
the shift value. Therefore it is unnecessary to do a general shift value
overflow check inside badblock API, and just check here as the caller.
This may avoid unnecessary check, make the badblocks API code more simple
and elegant.
Fixes: 2699b67223 ("md: load/store badblock list from v1.x metadata")
Fixes: 1726c77467 ("badblocks: improve badblocks_set() for multiple ranges handling")
Cc: stable@vger.kernel.org
Cc: Ramesh Adhikari <adhikari.resume@gmail.com>
Signed-off-by: Coly Li <colyli@fygo.io>
Reviewed-by: Yu Kuai <yukuai@fygo.io>
Link: https://patch.msgid.link/20260720111400.2120834-1-colyli@fygo.io
Signed-off-by: Yu Kuai <yukuai@fygo.io>
Storing a memalloc_noio_save() token in mddev->noio_flags lets one task
save the token and another task restore it. With concurrent suspend sysfs
writes, task A can enter PF_MEMALLOC_NOIO, return to userspace still in
that scope, and later task B can restore A's saved token.
Avoid tying the token lifetime to mddev. Keep mddev_suspend() and
mddev_resume() only responsible for array suspension, and enter
PF_MEMALLOC_NOIO only in the MD paths that allocate memory after the array
has been suspended. Restore the token before resuming the array.
A reproducer repeatedly writes suspend_lo and suspend_hi from concurrent
workers and checks each worker's /proc/self/stat flags before and after the
sysfs write.
Link: https://github.com/chencheng-fnnas/reproducer/blob/main/repro-md-noio-token-leak.sh
Fixes: 78f57ef9d5 ("md: use memalloc scope APIs in mddev_suspend()/mddev_resume()")
Signed-off-by: Chen Cheng <chencheng@fnnas.com>
Reviewed-by: Yu Kuai <yukuai@fygo.io>
Link: https://patch.msgid.link/20260718084218.417895-1-chencheng@fnnas.com
Signed-off-by: Yu Kuai <yukuai@fygo.io>
backlog_store() suspends the array before checking whether a write-mostly
device exists. If no such device exists, the error path only unlocks
reconfig_mutex and leaves the array suspended, blocking subsequent I/O.
Use mddev_unlock_and_resume() to release both states.
Fixes: 58226942ad ("md: use new apis to suspend array before mddev_create/destroy_serial_pool")
Signed-off-by: Chen Cheng <chencheng@fnnas.com>
Reviewed-by: Yu Kuai <yukuai@fygo.io>
Link: https://patch.msgid.link/20260718034236.4119093-1-chencheng@fnnas.com
Signed-off-by: Yu Kuai <yukuai@fygo.io>
make_discard_request() returns without completing the bio when reshape
is in progress. Discard callers block in submit_bio_wait()
waiting for a completion that never arrives. The caller hangs in
uninterruptible sleep, and this does not resolve when reshape finishes.
Complete the bio with BLK_STS_AGAIN so userspace can retry after reshape,
consistent with the existing policy of not processing discard during
reshape.
Tested on a loop-backed RAID5 array during mdadm --grow: without this
patch, blkdiscard hangs in bio_await() and remains in uninterruptible
sleep after md reports "reshape done"; with this patch it returns
-EAGAIN instead.
Signed-off-by: Genjian Zhang <zhanggenjian@kylinos.cn>
Reviewed-by: Yu Kuai <yukuai@fygo.io>
Link: https://patch.msgid.link/20260711161326.962336-1-zhanggenjian@126.com
Signed-off-by: Yu Kuai <yukuai@fygo.io>
origin flow:
bio_endio(master_bio); /* may drop active_io to zero */
allow_barrier(conf);
free_r10bio(r10_bio); /* reads conf->geo, returns to pool */
one scenario is:
CPU A (softirq, raid_end_bio_io) CPU B (action_store) --> reshape
================================ ===============================
bio_endio(master_bio)
md_end_clone_io
percpu_ref_put -> 0
wait_event wakeup, and,
mddev_suspend return
raid10_start_reshape:
setup_geo(&conf->geo, new)
...
mempool_destroy(old_pool)
conf->r10bio_pool = new_pool
allow_barrier(conf)
free_r10bio(r10_bio)
put_all_bios:
for (i=0; i<conf->geo.raid_disks; i++)
==> old obj, new geo, OOB
mempool_free(r10_bio, conf->r10bio_pool)
==> old-geometry obj freed into new pool
so .. fix by reorder the flow:
free_r10bio(r10_bio)
bio_endio(master_bio)
allow_barrier(conf)
raid_end_discard_bio() is exactly the same.
Signed-off-by: Chen Cheng <chencheng@fnnas.com>
Link: https://patch.msgid.link/20260711100352.425177-4-chencheng@fnnas.com
Signed-off-by: Yu Kuai <yukuai@fygo.io>
When reshape grows raid_disks, the pool must also switch to new geometry
object size , and allocate a new geometry size pool and replace the old.
But not for shrinking reshape, because regular I/O can still use the
prev geo for sectors that have not crossed reshape_progress yet.
Signed-off-by: Chen Cheng <chencheng@fnnas.com>
Link: https://patch.msgid.link/20260711100352.425177-3-chencheng@fnnas.com
Signed-off-by: Yu Kuai <yukuai@fygo.io>
raid10 needs to resize/swap r10bio_pool when reshape changes
raid_disks, and, don't let new requests keep allocating r10bio
objects from the old pool while that transition is in progress.
suspend and lock array before mddev_start_reshape(), and resume
it on exit.
Other sync_action ops are unchanged.
Signed-off-by: Chen Cheng <chencheng@fnnas.com>
Link: https://patch.msgid.link/20260711100352.425177-2-chencheng@fnnas.com
Signed-off-by: Yu Kuai <yukuai@fygo.io>
The badblocks core API -- badblocks_set(), badblocks_clear() and
badblocks_check() -- and the is_badblock() helper all take the range
length as sector_t. The md wrappers rdev_set_badblocks(),
rdev_clear_badblocks() and rdev_has_badblock(), however, declared the
same length as int, narrowing sector_t to int and back again in the
middle of an otherwise 64-bit clean path.
Change the sectors parameter to sector_t in these three wrappers so it
matches the core API and is_badblock(). No functional change: current
callers pass per-I/O or per-resync-chunk lengths well within int range.
This just removes a gratuitous truncation point and keeps the type
consistent end to end.
Signed-off-by: Hiroshi Nishida <nishidafmly@gmail.com>
Link: https://patch.msgid.link/20260710132329.7273-3-nishidafmly@gmail.com
Signed-off-by: Yu Kuai <yukuai@fygo.io>
raid10_write_request() drops the barrier before calling
bio_submit_split_bioset() and reacquires it afterwards. This is no
longer necessary because the split bio cannot re-enter
raid10_write_request() while the barrier is held.
The allow_barrier()/wait_barrier() pair was introduced by commit
e820d55cb9 ("md: fix raid10 hang issue caused by barrier") when
submit_flushes() called md_handle_request() directly, allowing re-entry
into raid10_write_request(). Since v5.2, submit_flushes() has instead
gone through submit_bio(), eliminating that recursion. submit_flushes()
was later removed entirely by commit b75197e86e ("md: Remove flush
handling").
Currently, raid10_write_request() is only entered from the bio
submission path, so the split bio submitted by bio_submit_split_bioset()
cannot recurse back into wait_barrier().
Remove the redundant allow_barrier()/wait_barrier() pair around
bio_submit_split_bioset().
Signed-off-by: Abd-Alrhman Masalkhi <abd.masalkhi@gmail.com>
Reviewed-by: Yu Kuai <yukuai@fygo.io>
Link: https://patch.msgid.link/20260710101521.1714-5-abd.masalkhi@gmail.com
Signed-off-by: Yu Kuai <yukuai@fygo.io>
RAID10 currently handles one badblock path explicitly by failing atomic
writes with EIO. However, another badblock path can also reduce the
writable range and force the bio through bio_submit_split_bioset(),
which implicitly completes the bio with EINVAL.
Fix this by handling atomic writes in the common split check. If RAID10
determines that an atomic write would require splitting, complete the
bio with EIO.
Fixes: a1d9b4fd42 ("md/raid10: Atomic write support")
Signed-off-by: Abd-Alrhman Masalkhi <abd.masalkhi@gmail.com>
Reviewed-by: Yu Kuai <yukuai@fygo.io>
Reviewed-by: John Garry <john.g.garry@oracle.com>
Link: https://patch.msgid.link/20260710101521.1714-4-abd.masalkhi@gmail.com
Signed-off-by: Yu Kuai <yukuai@fygo.io>
Restrict the RAID1 atomic write limits by setting chunk_sectors to
BARRIER_UNIT_SECTOR_SIZE so that atomic writes never straddle a barrier
unit.
A bio that passes block-layer validation may still become unserviceable
within RAID1 due to bad blocks or write-behind constraints. In the former
case, complete the bio with EIO. In the latter case, disable
write-behind rather than failing the bio with EIO.
Fixes: f2a38abf5f ("md/raid1: Atomic write support")
Fixes: a4c55c9026 ("md/raid1: simplify raid1_write_request() error handling")
Reviewed-by: John Garry <john.g.garry@oracle.com>
Signed-off-by: Abd-Alrhman Masalkhi <abd.masalkhi@gmail.com>
Reviewed-by: Yu Kuai <yukuai@fygo.io>
Link: https://patch.msgid.link/20260710101521.1714-3-abd.masalkhi@gmail.com
Signed-off-by: Yu Kuai <yukuai@fygo.io>
remove_spares() and remove_and_add_spares() modify the array's rdev
configuration. These operations are only safe after the array has been
suspended.
md_start_sync() checks whether spare configuration changes are needed
before taking reconfig_mutex. However, the rdev state can change before
the mutex is acquired, so the initial check can become stale. In that
case, md_choose_sync_action() may remove or replace rdevs while normal
I/O is still accessing them.
The race can occur as follows:
raid10d Worker Normal IO
____________ _______________________ ______________________
raid10_write_request()
wait_blocked_dev()
set Blocked
set Faulty
Skip Faulty rdev
rrdev->nr_pending++
.repl_bio = bio
removeable_rdev = false .
array not suspended .
lock mddev goto err_handle
lock mddev (wait)
.
update sb .
clear Blocked .
.
unlock mddev .
lock mddev (acquires)
remove_spares()
removeable_rdev = true
raid10_remove_disk()
rdev = replacement
replacement = NULL
rdev_dec_pending(NULL)
unlock mddev (NULL)->nr_pending--
In this case, rdev_dec_pending() is called with a NULL pointer,
resulting in a NULL pointer dereference when attempting to decrement
nr_pending.
Fix this by suspending the array when spare configuration changes are
needed, including for non-read-write arrays, and checking again after
taking reconfig_mutex. If the array was not already suspended and a
change is now needed, release the mutex, suspend the array, and
reacquire the mutex before continuing.
Fixes: bc08041b32 ("md: suspend array in md_start_sync() if array need reconfiguration")
Reported-by: sashiko-bot <sashiko-bot@kernel.org>
Closes: https://sashiko.dev/#/patchset/20260628142420.1051027-1-abd.masalkhi@gmail.com?part=3
Signed-off-by: Abd-Alrhman Masalkhi <abd.masalkhi@gmail.com>
Reviewed-by: Yu Kuai <yukuai@fygo.io>
Link: https://patch.msgid.link/20260708112003.474537-1-abd.masalkhi@gmail.com
Signed-off-by: Yu Kuai <yukuai@fygo.io>
REQ_NOWAIT support in md personalities that can block internally is
fundamentally incomplete. While reads can avoid some blocking paths,
write requests can still encounter cases where one mirror succeeds while
another returns -EAGAIN. At that point md cannot distinguish queue
pressure from a real device failure, so it can neither record a bad
block nor safely retry the write without REQ_NOWAIT, leaving mirrors
with divergent data.
Rather than continue advertising REQ_NOWAIT support for personalities
that cannot implement it correctly, remove it from raid1, raid10 and
raid456. Keep REQ_NOWAIT for linear and raid0, which only remap bios to
their underlying devices; stacked limits will still clear the feature if
any component device lacks REQ_NOWAIT support.
Fixes: bf2c411bb1 ("md: raid456 add nowait support")
Fixes: c9aa889b03 ("md: raid10 add nowait support")
Fixes: 5aa705039c ("md: raid1 add nowait support")
Fixes: f51d46d0e7 ("md: add support for REQ_NOWAIT")
Suggested-by: Yu Kuai <yukuai@fygo.io>
Signed-off-by: Abd-Alrhman Masalkhi <abd.masalkhi@gmail.com>
Reviewed-by: Yu Kuai <yukuai@fygo.io>
Link: https://patch.msgid.link/20260628142737.1051059-1-abd.masalkhi@gmail.com
Signed-off-by: Yu Kuai <yukuai@fygo.io>
During reshape:
- reshape_request() advances rdev->recovery_offset for non-In_sync
devices locklessly.
- analyse_stripe() reads rdev->recovery_offset locklessly to decide:
a. use a replacement device to read ?
b. a device can already be treated as in-sync for the current
stripe ?
one possible scenario is:
CPU1 CPU2
reshape_request()
-> mddev->curr_resync_completed = sector_nr
-> if (!mddev->reshape_backwards)
-> rdev->recovery_offset = sector_nr
analyse_stripe(sh)
-> rdev = conf->disks[i].replacement
-> if (rdev->recovery_offset >=
sh->sector + stripe_sectors)
set_bit(R5_ReadRepl)
-> or
-> if (sh->sector + stripe_sectors <=
rdev->recovery_offset)
set_bit(R5_Insync)
And it could be:
- reading from a replacement before it is recovered far enough; or
- treating a not-yet-recovered device as in-sync for the current stripe.
Fixes: db0505d320 ("md: be cautious about using ->curr_resync_completed for ->recovery_offset")
The race report:
==================================================================
BUG: KCSAN: data-race in ops_run_io / reshape_request
write to 0xffff8bdee168b270 of 8 bytes by task 1704 on cpu 10:
reshape_request+0x1292/0x17b0
raid5_sync_request+0x815/0xa00
md_do_sync.cold+0xf8d/0x1516
[......]
read to 0xffff8bdee168b270 of 8 bytes by task 1696 on cpu 9:
ops_run_io+0xc25/0x1960
handle_stripe+0x2273/0x4570
handle_active_stripes.isra.0+0x6e0/0xa50
raid5d+0x7d5/0xb90
[......]
value changed: 0x0000000000091a00 -> 0x0000000000091b00
==================================================================
Signed-off-by: Chen Cheng <chencheng@fnnas.com>
Reviewed-by: Yu Kuai <yukuai@fygo.io>
Link: https://patch.msgid.link/20260627102519.136940-1-chencheng@fnnas.com
Signed-off-by: Yu Kuai <yukuai@fygo.io>
reshape stripe lifetime:
- start reshape ==> reshape_request():
* get destination stripe,
- if need to copy source data chunks, set STRIPE_EXPANDING;
- or, if new regions past the old end of the array, zero-filled,
no need source data, set STRIPE_EXPANDING | STRIPE_READY
* get source stripe,
- set STRIPE_EXPAND_SOURCE
- handle expand stripe ==> handle_stripe():
reshape use reconstruct-write to construct stripe,
four stages:
1. prepare source data chunks for old geometry stripe
- fill source stripe data by read or compute
2. move data from old geometry source stripe to new geometry
destination stripe
- source stripe clear STRIPE_EXPAND_SOURCE
- drain data from source to destination stripe
- mark stripe chunk as R5_Expanded|R5_UPTODATE when the
drain from source chunk to destination chunk is completed
- all stripe chunks drain are completed, then mark
STRIPE_EXPAND_READY
3. calculate p/q chunks for destination stripe
- if destination stripe doesn't depends on source dstripe,
then we can clear STRIPE_EXPANDING
4. write-out to disks and release
- set R5_Wantwrite|R5_Locked, writeout to disk
- if write-out succeeded, clear STRIPE_EXPAND_READY, and
decrement reshape_stripe, call md_done_sync() to report
reshape progress.
1. cleanup the following kinds of **destination stripe**
when failed device more than max degraded:
- new regions past the old end of the array, zero-filled in place,
requires no source data.
(STRIPE_EXPANDING | STRIPE_EXPAND_READY)
- prepare source data chunks already done, and writeout failed
(STRIPE_EXPAND_READY)
2. destination stripes that need source data
(STRIPE_EXPANDING, no STRIPE_HANDLE)
- these kind of stripes sit idle in the stripe cache and are never seen
by handle_stripe(). So clean up indirectly when their source stripe
(type 3) is processed.
3. source stripes (STRIPE_EXPAND_SOURCE)
- hit handle_stripe() after their member disks are marked Faulty.
- clear STRIPE_EXPAND_SOURCE, finds and cleanup all dependent destination
stripes that were waiting for data.
- walks the source's data disks, compute the corresponding destination
sector, looks up the destination stripe, and do cleanup(clear flags,
dec counters, call md_done_sync())
Reproducer:
- Create a 4-disk RAID5 with mdadm on top of 5 disposable test disks
wrapped by dm targets.
- Add the 5th device as a spare and start a 4 -> 5 reshape.
- Wait until /sys/block/mdX/md/sync_action reports "reshape".
- Inject failures on two members so reshape exceeds max_degraded.
- After a few seconds, write "frozen" to /sys/block/mdX/md/sync_action.
Before this fix, the write blocks indefinitely.
Read-error variant:
- Use dm-dust on /dev/sd[b-f].
- Preload bad blocks on two source members, e.g. dust0 and dust1:
dmsetup message dust0 0 addbadblock <range>
dmsetup message dust1 0 addbadblock <range>
- Start reshape:
mdadm -C /dev/mdX -e 1.2 -l 5 -n 4 -c 64 \
--assume-clean /dev/mapper/dust{0..3}
mdadm --manage /dev/mdX --add /dev/mapper/dust4
mdadm --grow /dev/mdX -n 5 --backup-file=/tmp/grow.backup &
- Once reshape starts, enable the injected read failures:
dmsetup message dust0 0 enable
dmsetup message dust1 0 enable
- Then:
echo frozen > /sys/block/mdX/md/sync_action
hangs forever before the fix.
Write-error variant:
- Use dm-flakey on /dev/sd[b-f].
- Start the same 4 -> 5 reshape on flakey0..flakey4.
- Once reshape starts, switch two members, e.g. flakey3 and flakey4,
to error_writes.
- Then:
echo frozen > /sys/block/mdX/md/sync_action
hangs forever before the fix.
md_do_sync() exits its main loop on MD_RECOVERY_INTR but then blocks
forever at:
wait_event(mddev->recovery_wait,
!atomic_read(&mddev->recovery_active));
After the fix recovery_active drains to zero, md_do_sync() prints
md/raid:md0: Cannot continue operation (2/5 failed).
md: md0: reshape interrupted.
Signed-off-by: Chen Cheng <chencheng@fnnas.com>
Reviewed-by: Yu Kuai <yukuai@fygo.io>
Link: https://patch.msgid.link/20260624075824.2601110-1-chencheng@fnnas.com
Signed-off-by: Yu Kuai <yukuai@fygo.io>
max_nr_stripes is updated under cache_size_mutex in the stripe cache
grow/shrink paths, while is_inactive_blocked() and
raid5_end_read_request() read it without that lock.
Use READ_ONCE() for those reads in lockless path to match the WRITE_ONCE()
updates and avoid KCSAN data race reports.
A similar issue was previously fixed in commit-id:
dfd2bf4367.
Fixes: 0009fad033 ("raid5 improve too many read errors msg by adding limits")
Fixes: 3514da58be ("md/raid5: Make is_inactive_blocked() helper")
KCSAN report:
=================
BUG: KCSAN: data-race in grow_one_stripe / is_inactive_blocked
write (marked) to 0xffff8f01f0b5a268 of 4 bytes by task 12616 on cpu 9:
grow_one_stripe+0x2d8/0x320
raid5d+0xb57/0xba0
md_thread+0x15a/0x2d0
[..........]
read to 0xffff8f01f0b5a268 of 4 bytes by task 12670 on cpu 11:
is_inactive_blocked+0x97/0xc0
raid5_get_active_stripe+0x2fd/0xa70
raid5_make_request+0x4aa/0x2940
[..........]
value changed: 0x000003b9 -> 0x000003ba
Signed-off-by: Chen Cheng <chencheng@fnnas.com>
Reviewed-by: Yu Kuai <yukuai@fygo.io>
Link: https://patch.msgid.link/20260624024042.2561803-1-chencheng@fnnas.com
Signed-off-by: Yu Kuai <yukuai@fygo.io>
The patch just suppress KCSAN noise. No functional change.
KCSAN reports a race, point to update_read_sectors() update next_seq_sect vs.
read next_seq_sect.
Protect next_seq_sect and seq_start with READ_ONCE/WRITE_ONCE, otherwise,
read balance see stale sequential-read hints.
KCSAN report:
==============
BUG: KCSAN: data-race in raid1_read_request / raid1_read_request
write to 0xffff8e3a2d6736d0 of 8 bytes by task 593784 on cpu 10:
raid1_read_request+0xe5a/0x19f0
raid1_make_request+0xdf/0x1990
md_handle_request+0x4a2/0xa40
[...]
read to 0xffff8e3a2d6736d0 of 8 bytes by task 593776 on cpu 11:
raid1_read_request+0xe3f/0x19f0
raid1_make_request+0xdf/0x1990
md_handle_request+0x4a2/0xa40
[...]
value changed: 0x0000000000356368 -> 0x0000000000356370
Signed-off-by: Chen Cheng <chencheng@fnnas.com>
Reviewed-by: Yu Kuai <yukuai@fygo.io>
Link: https://patch.msgid.link/20260623075940.2476255-1-chencheng@fnnas.com
Signed-off-by: Yu Kuai <yukuai@fygo.io>
kcsan detect race :
- raid5d() closes the current bitmap batch by updating
conf->seq_flush under conf->device_lock.
- __add_stripe_bio() read conf->seq_flush without that
lock when assigning sh->bm_seq.
so, protect seq_flush/seq_write consistency for multiple CPUs by
READ_ONCE()/WRITE_ONCE() under the path without held device_lock.
re-explain the stripe batch sequence number update flow:
1. sh->bm_seq declare which batch number the stripe belongs to
when perform bitmap-related write.
==> bm_seq = seq_flush+1
2. stripe be handled,
* if sh->bm_seq - conf->seq_write > 0, means the
batch stripes **newer than** the last written
batch, it cannot proceed yet, queued on bitmap_list.
* otherwise , has already proceed.
3. raid5d() `++seq_flush` to closes the current batch, means
* no more stripes join that old batch
* just-closed batch ready to write-out to disk
4. raid5d() calls bitmap hooks unplug() or writeout, then,
`++seq_write` to the same as bm_seq.
- seq_flush - for producer, to close batches.
- seq_write - for consumer, the checkpoint number.
the report:
====================================
BUG: KCSAN: data-race in __add_stripe_bio / raid5d
write to 0xffff88ba5625d470 of 4 bytes by task 82401 on cpu 0:
raid5d+0x1d9/0xba0
[.....]
read to 0xffff88ba5625d470 of 4 bytes by task 82421 on cpu 8:
__add_stripe_bio+0x332/0x400
raid5_make_request+0x6ac/0x2930
md_handle_request+0x4a2/0xa40
md_submit_bio+0x109/0x1a0
__submit_bio+0x2ec/0x390
[.....]
Fixes: 7c13edc875 ("md: incorporate new plugging into raid5.")
v1 -> v2:
- remove WRITE_ONCE(conf->seq_write) in held device_lock path.
- remove READ_ONCE(conf->seq_flush) in held device_lock path.
Signed-off-by: Chen Cheng <chencheng@fnnas.com>
Reviewed-by: Yu Kuai <yukuai@fygo.io>
Link: https://patch.msgid.link/20260622124649.1780233-1-chencheng@fnnas.com
Signed-off-by: Yu Kuai <yukuai@fygo.io>
