[PATCH v2] mtd: blktrans: keep reclaim out of the ->release paths

From: Shuai Liu

Date: Tue Nov 17 2026 - 01:08:53 EST


blktrans_release() invokes tr->release() while holding dev->lock, and
translators can walk down into the flash driver from there:
mtdblock_release() flushes its write cache via write_cached_data() ->
erase_write() -> mtd_erase(), and concat_erase() allocates a struct
erase_info with GFP_KERNEL. del_mtd_blktrans_dev() calls ->release()
under the same lock on the device removal path.

Reclaim from those allocations can recurse back into the same block
queue and deadlock. lockdep reports the full cycle when a write to the
concat device is flushed at last close:

init/1 is trying to acquire lock:
fs_reclaim, at: __kmalloc_cache_noprof
but task is already holding lock:
&mtdblk->cache_mutex, at: mtdblock_release

Chain exists of:
fs_reclaim --> &new->lock --> &mtdblk->cache_mutex

the existing dependency chain (in reverse order) is:
fs_reclaim -> q_usage_counter(io) -> elevator_lock
-> set->srcu -> new->lock -> mtdblk->cache_mutex

Call Trace:
concat_erase
mtd_erase
erase_write
write_cached_data
mtdblock_release
blktrans_release
bdev_release

The io path is already protected: mtd_queue_rq() disables reclaim
across dispatch with memalloc_noio_save(), because translators may
allocate while processing requests under dev->lock and the srcu read
side. ->release() runs under the same locks, so give it the same
treatment in blktrans_release() and del_mtd_blktrans_dev().

Verified with mtdram + block2mtd + concat + mtdblock on QEMU
(CONFIG_PROVE_LOCKING): writing through /dev/mtdblockN and closing
the fd produces the warning above on the unpatched kernel and stays
silent with this patch, with the readback still succeeding.

Reported-by: syzbot+c66e7044aa73c620d862@xxxxxxxxxxxxxxxxxxxxxxxxx
Closes: https://syzkaller.appspot.com/bug?extid=c66e7044aa73c620d862
Fixes: 048d87199566 ("mtd: blktrans: Hotplug fixes")
Signed-off-by: Shuai Liu <847807867@xxxxxx>
---

Changes in v2:
- Add Fixes tag pointing to 048d87199566 ("mtd: blktrans: Hotplug fixes"),
which made ->release() run under dev->lock and added the ->release()
call to del_mtd_blktrans_dev() (Sashiko)

drivers/mtd/mtd_blkdevs.c | 13 +++++++++++--
1 file changed, 11 insertions(+), 2 deletions(-)

diff --git a/drivers/mtd/mtd_blkdevs.c b/drivers/mtd/mtd_blkdevs.c
index 42f250344f0cd..be4ead01f5210 100644
--- a/drivers/mtd/mtd_blkdevs.c
+++ b/drivers/mtd/mtd_blkdevs.c
@@ -233,6 +233,7 @@ static int blktrans_open(struct gendisk *disk, blk_mode_t mode)
static void blktrans_release(struct gendisk *disk)
{
struct mtd_blktrans_dev *dev = disk->private_data;
+ unsigned int noio_flags;

mutex_lock(&dev->lock);

@@ -242,8 +243,12 @@ static void blktrans_release(struct gendisk *disk)
module_put(dev->tr->owner);

if (dev->mtd) {
- if (dev->tr->release)
+ if (dev->tr->release) {
+ /* Reclaim must not recurse into I/O while releasing the device. */
+ noio_flags = memalloc_noio_save();
dev->tr->release(dev);
+ memalloc_noio_restore(noio_flags);
+ }
__put_mtd_device(dev->mtd);
}
unlock:
@@ -410,6 +415,7 @@ int del_mtd_blktrans_dev(struct mtd_blktrans_dev *old)
{
unsigned long flags;
unsigned int memflags;
+ unsigned int noio_flags;

lockdep_assert_held(&mtd_table_mutex);

@@ -435,8 +441,11 @@ int del_mtd_blktrans_dev(struct mtd_blktrans_dev *old)
then put mtd device, and don't touch it again */
mutex_lock(&old->lock);
if (old->open) {
- if (old->tr->release)
+ if (old->tr->release) {
+ noio_flags = memalloc_noio_save();
old->tr->release(old);
+ memalloc_noio_restore(noio_flags);
+ }
__put_mtd_device(old->mtd);
}

--
2.43.0