* [PATCH 0/7] block fixes found by code inspection
@ 2026-11-19 7:05 Guixin Liu
2026-11-19 7:05 ` [PATCH 1/7] blk-mq: free the rqs arrays of tags in the SRCU callback Guixin Liu
` (6 more replies)
0 siblings, 7 replies; 15+ messages in thread
From: Guixin Liu @ 2026-11-19 7:05 UTC (permalink / raw)
To: Jens Axboe, Christoph Hellwig, Tejun Heo, Josef Bacik,
Nilay Shroff, Ming Lei, Hannes Reinecke
Cc: linux-block
Seven block layer fixes, all found by code inspection of the tags SRCU
lifetime, the hctx reuse path, and the blk-cgroup configuration writers.
Patch 1 and patch 7 close the two halves of the tags_srcu protection for
tag iterators: patch 1 frees the tags->rqs[] and tags->static_rqs[]
arrays in the SRCU callback together with the bitmaps, patch 7 unpublishes
set->tags[] before the grace period starts, so a reader can no longer
load a pointer that the grace period has already released.
Patch 2 clears a stale BLK_MQ_S_STOPPED flag when a parked hctx is
reused, so requests queued on it do not stall until the driver happens
to stop and restart the queue again.
Patch 3 and patch 4 fix the locking order of ioc_cost_model_write():
patch 3 releases rq_qos_mutex before freezing the queue like
ioc_qos_write() does, patch 4 moves the unfreeze in ioc_qos_write()
ahead of the bdev reference drop, closing the same lifetime gap in both
writers.
Patch 5 fixes the mean folding in blk_rq_stat_sum(), which silently
dropped the samples of an already folded source and kept
blk-iolatency's accumulated window mean at zero on rotating disks.
Patch 6 rejects a dead disk in the three lazy policy init helpers, so a
configuration write racing del_gendisk() can no longer attach an orphan
iocost/iolatency/throttle policy that nothing tears down again.
Found by code inspection.
Tested on a KVM q35 guest with a 7.4-rc1 based kernel (PROVE_LOCKING
enabled): hctx state flags after a stop/shrink/start/grow cycle on
nullb0 show no stale flags and IO on the reused hctx completes; iocost
cost.model and cost.qos writes against a live and a removed nullb
device behave as expected, with the removed device rejected with
-ENODEV and the live device accepting all model, qos and coefficient
writes; concurrent nullb submit_queues resizes against continuous
debugfs busy/tags reads under fio ran for six minutes without errors.
Guixin Liu (7):
blk-mq: free the rqs arrays of tags in the SRCU callback
blk-mq: clear stale stopped flag on hctx reuse
blk-iocost: don't take rq_qos_mutex before the queue is frozen in cost
model write
blk-iocost: unfreeze the queue before dropping the bdev reference in
qos write
blk-stat: account for the mean of the source in blk_rq_stat_sum()
blk-cgroup: don't attach a qos policy to a dead disk
blk-mq: unpublish the tags before deferring their free to SRCU
block/blk-iocost.c | 19 ++++++++++++++-----
block/blk-iolatency.c | 3 +++
block/blk-mq-tag.c | 2 ++
block/blk-mq.c | 24 +++++++-----------------
block/blk-mq.h | 1 -
block/blk-stat.c | 8 ++++++--
block/blk-throttle.c | 3 +++
7 files changed, 35 insertions(+), 25 deletions(-)
--
2.43.7
^ permalink raw reply [flat|nested] 15+ messages in thread
* [PATCH 1/7] blk-mq: free the rqs arrays of tags in the SRCU callback
2026-11-19 7:05 [PATCH 0/7] block fixes found by code inspection Guixin Liu
@ 2026-11-19 7:05 ` Guixin Liu
2026-11-19 7:05 ` [PATCH 2/7] blk-mq: clear stale stopped flag on hctx reuse Guixin Liu
` (5 subsequent siblings)
6 siblings, 0 replies; 15+ messages in thread
From: Guixin Liu @ 2026-11-19 7:05 UTC (permalink / raw)
To: Jens Axboe, Christoph Hellwig, Tejun Heo, Josef Bacik,
Nilay Shroff, Ming Lei, Hannes Reinecke
Cc: linux-block
struct blk_mq_tags and the requests are kept alive until the iterators
under tags_srcu are gone, and the bitmaps were recently moved into that
SRCU callback as well, but the tags->rqs[] and tags->static_rqs[] arrays
are still freed right away by blk_mq_free_rq_map(). An iterator holding
a tags pointer dereferences both arrays per set bit, e.g.
blk_mq_find_and_get_req(), so when the number of hardware queues shrinks
the arrays can be freed while blk_mq_tagset_busy_iter() is about to walk
them.
Free both arrays in the SRCU callback together with the bitmaps. With
that blk_mq_free_rq_map() is a trivial wrapper of blk_mq_free_tags(),
so drop it and call blk_mq_free_tags() directly.
Fixes: 995412e23bb2 ("blk-mq: Replace tags->lock with SRCU for tag iterators")
Fixes: ad0d05dbddc1 ("blk-mq: Defer freeing of tags page_list to SRCU callback")
Cc: stable@vger.kernel.org
Signed-off-by: Guixin Liu <kanie@linux.alibaba.com>
---
block/blk-mq-tag.c | 2 ++
block/blk-mq.c | 18 +++---------------
block/blk-mq.h | 1 -
3 files changed, 5 insertions(+), 16 deletions(-)
diff --git a/block/blk-mq-tag.c b/block/blk-mq-tag.c
index 3515031b3226..ee5413ebab01 100644
--- a/block/blk-mq-tag.c
+++ b/block/blk-mq-tag.c
@@ -611,6 +611,8 @@ static void blk_mq_free_tags_callback(struct rcu_head *head)
kmemleak_free(page_address(page));
__free_pages(page, page->private);
}
+ kfree(tags->rqs);
+ kfree(tags->static_rqs);
/* An iterator may walk the bitmaps for as long as it may see the tags */
sbitmap_queue_free(&tags->bitmap_tags);
sbitmap_queue_free(&tags->breserved_tags);
diff --git a/block/blk-mq.c b/block/blk-mq.c
index 5ae22d556540..ae12f2a52cc9 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -3532,16 +3532,6 @@ void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
*/
}
-void blk_mq_free_rq_map(struct blk_mq_tags *tags)
-{
- kfree(tags->rqs);
- tags->rqs = NULL;
- kfree(tags->static_rqs);
- tags->static_rqs = NULL;
-
- blk_mq_free_tags(tags);
-}
-
static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,
unsigned int hctx_idx)
{
@@ -3594,12 +3584,10 @@ static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
node);
if (!tags->static_rqs)
- goto err_free_rqs;
+ goto err_free_tags;
return tags;
-err_free_rqs:
- kfree(tags->rqs);
err_free_tags:
blk_mq_free_tags(tags);
return NULL;
@@ -4155,7 +4143,7 @@ struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
if (ret) {
- blk_mq_free_rq_map(tags);
+ blk_mq_free_tags(tags);
return NULL;
}
@@ -4183,7 +4171,7 @@ void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
{
if (tags) {
blk_mq_free_rqs(set, tags, hctx_idx);
- blk_mq_free_rq_map(tags);
+ blk_mq_free_tags(tags);
}
}
diff --git a/block/blk-mq.h b/block/blk-mq.h
index 1e64236d34d4..ba7bd917c361 100644
--- a/block/blk-mq.h
+++ b/block/blk-mq.h
@@ -62,7 +62,6 @@ void blk_mq_put_rq_ref(struct request *rq);
*/
void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
unsigned int hctx_idx);
-void blk_mq_free_rq_map(struct blk_mq_tags *tags);
struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
unsigned int hctx_idx, unsigned int depth);
void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
--
2.43.7
^ permalink raw reply related [flat|nested] 15+ messages in thread
* [PATCH 2/7] blk-mq: clear stale stopped flag on hctx reuse
2026-11-19 7:05 [PATCH 0/7] block fixes found by code inspection Guixin Liu
2026-11-19 7:05 ` [PATCH 1/7] blk-mq: free the rqs arrays of tags in the SRCU callback Guixin Liu
@ 2026-11-19 7:05 ` Guixin Liu
2026-11-19 7:05 ` [PATCH 3/7] blk-iocost: don't take rq_qos_mutex before the queue is frozen in cost model write Guixin Liu
` (4 subsequent siblings)
6 siblings, 0 replies; 15+ messages in thread
From: Guixin Liu @ 2026-11-19 7:05 UTC (permalink / raw)
To: Jens Axboe, Christoph Hellwig, Tejun Heo, Josef Bacik,
Nilay Shroff, Ming Lei, Hannes Reinecke
Cc: linux-block
When a driver stops a hardware queue with blk_mq_stop_hw_queues() and
the number of hardware queues shrinks before it calls
blk_mq_start_stopped_hw_queues() again, the parked hctx keeps its
BLK_MQ_S_STOPPED flag: the exit path does not clear it, and
blk_mq_start_stopped_hw_queues() only walks the live hctxs. A later
update that reuses the parked hctx via blk_mq_alloc_and_init_hctx()
inherits the flag, so blk_mq_run_hw_queue() refuses to run requests
queued on it until the driver happens to stop and restart the queue
again.
Clear BLK_MQ_S_STOPPED in blk_mq_init_hctx() so a reused hctx starts
like a freshly allocated one, matching the BLK_MQ_S_INACTIVE handling.
Fixes: 2f8f1336a48b ("blk-mq: always free hctx after request queue is freed")
Cc: stable@vger.kernel.org
Signed-off-by: Guixin Liu <kanie@linux.alibaba.com>
---
block/blk-mq.c | 1 +
1 file changed, 1 insertion(+)
diff --git a/block/blk-mq.c b/block/blk-mq.c
index ae12f2a52cc9..87c1ccdf66ab 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -4016,6 +4016,7 @@ static int blk_mq_init_hctx(struct request_queue *q,
hctx->queue_num = hctx_idx;
clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
+ clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
hctx->tags = set->tags[hctx_idx];
--
2.43.7
^ permalink raw reply related [flat|nested] 15+ messages in thread
* [PATCH 3/7] blk-iocost: don't take rq_qos_mutex before the queue is frozen in cost model write
2026-11-19 7:05 [PATCH 0/7] block fixes found by code inspection Guixin Liu
2026-11-19 7:05 ` [PATCH 1/7] blk-mq: free the rqs arrays of tags in the SRCU callback Guixin Liu
2026-11-19 7:05 ` [PATCH 2/7] blk-mq: clear stale stopped flag on hctx reuse Guixin Liu
@ 2026-11-19 7:05 ` Guixin Liu
2026-11-19 13:40 ` Tao Cui
2026-11-19 7:05 ` [PATCH 4/7] blk-iocost: unfreeze the queue before dropping the bdev reference in qos write Guixin Liu
` (3 subsequent siblings)
6 siblings, 1 reply; 15+ messages in thread
From: Guixin Liu @ 2026-11-19 7:05 UTC (permalink / raw)
To: Jens Axboe, Christoph Hellwig, Tejun Heo, Josef Bacik,
Nilay Shroff, Ming Lei, Hannes Reinecke
Cc: linux-block
ioc_cost_model_write() freezes the queue while holding rq_qos_mutex,
which blkg_conf_open_bdev() acquired first. Every other path takes
rq_qos_mutex only after the queue is frozen: ioc_qos_write() releases
and re-acquires the mutex around blk_mq_freeze_queue(), and rq_qos_add()
documents that its caller holds the mutex with the queue frozen.
Waiting for in-flight IO while holding rq_qos_mutex breaks that rule,
and on the first write to a queue without iocost the path even calls
blk_iocost_init() -> rq_qos_add() -> blk_mq_freeze_queue() while
nothing has frozen the queue yet.
Mirror the release-freeze-reacquire sequence of ioc_qos_write(). The
unfreeze moves to the close_bdev label so the error exits after the
freeze don't leave the queue frozen, and it runs before
blkg_conf_close_bdev() drops the bdev reference, while the queue is
still pinned.
Fixes: 2b2da2f6dc80 ("blk-iocost: prevent configuration update concurrent with io throttling")
Cc: stable@vger.kernel.org
Signed-off-by: Guixin Liu <kanie@linux.alibaba.com>
---
block/blk-iocost.c | 12 ++++++++++--
1 file changed, 10 insertions(+), 2 deletions(-)
diff --git a/block/blk-iocost.c b/block/blk-iocost.c
index 1c1025157d8e..8e06b83691aa 100644
--- a/block/blk-iocost.c
+++ b/block/blk-iocost.c
@@ -3589,6 +3589,15 @@ static ssize_t ioc_cost_model_write(struct kernfs_open_file *of, char *input,
ret = blkg_conf_open_bdev(&ctx);
if (ret)
return ret;
+ /*
+ * Nothing is protecting QoS state yet, so release q->rq_qos_mutex,
+ * which was first acquired in blkg_conf_open_bdev, and re-acquire it
+ * after freezing the queue to maintain the correct locking order.
+ */
+ mutex_unlock(&ctx.bdev->bd_queue->rq_qos_mutex);
+
+ memflags = blk_mq_freeze_queue(ctx.bdev->bd_queue);
+ mutex_lock(&ctx.bdev->bd_queue->rq_qos_mutex);
body = ctx.body;
q = bdev_get_queue(ctx.bdev);
@@ -3605,7 +3614,6 @@ static ssize_t ioc_cost_model_write(struct kernfs_open_file *of, char *input,
ioc = q_to_ioc(q);
}
- memflags = blk_mq_freeze_queue(q);
blk_mq_quiesce_queue(q);
spin_lock_irq(&ioc->lock);
@@ -3663,9 +3671,9 @@ static ssize_t ioc_cost_model_write(struct kernfs_open_file *of, char *input,
spin_unlock_irq(&ioc->lock);
blk_mq_unquiesce_queue(q);
- blk_mq_unfreeze_queue(q, memflags);
close_bdev:
+ blk_mq_unfreeze_queue(q, memflags);
blkg_conf_close_bdev(&ctx);
return ret ?: nbytes;
}
--
2.43.7
^ permalink raw reply related [flat|nested] 15+ messages in thread
* [PATCH 4/7] blk-iocost: unfreeze the queue before dropping the bdev reference in qos write
2026-11-19 7:05 [PATCH 0/7] block fixes found by code inspection Guixin Liu
` (2 preceding siblings ...)
2026-11-19 7:05 ` [PATCH 3/7] blk-iocost: don't take rq_qos_mutex before the queue is frozen in cost model write Guixin Liu
@ 2026-11-19 7:05 ` Guixin Liu
2026-11-19 7:05 ` [PATCH 5/7] blk-stat: account for the mean of the source in blk_rq_stat_sum() Guixin Liu
` (2 subsequent siblings)
6 siblings, 0 replies; 15+ messages in thread
From: Guixin Liu @ 2026-11-19 7:05 UTC (permalink / raw)
To: Jens Axboe, Christoph Hellwig, Tejun Heo, Josef Bacik,
Nilay Shroff, Ming Lei, Hannes Reinecke
Cc: linux-block
ioc_qos_write() calls blkg_conf_close_bdev() before
blk_mq_unfreeze_queue(). close_bdev drops the bdev reference the
writer holds; if that is the last reference, the queue enters its
release path and only the RCU-delayed free in blk_free_queue() keeps
it alive across the unfreeze. Unfreeze before the close instead,
while the writer's own bdev reference still pins the queue, matching
ioc_cost_model_write().
Fixes: 9730763f4756 ("block: correct locking order for protecting blk-wbt parameters")
Cc: stable@vger.kernel.org
Signed-off-by: Guixin Liu <kanie@linux.alibaba.com>
---
block/blk-iocost.c | 4 +---
1 file changed, 1 insertion(+), 3 deletions(-)
diff --git a/block/blk-iocost.c b/block/blk-iocost.c
index 8e06b83691aa..b251ecc66d8a 100644
--- a/block/blk-iocost.c
+++ b/block/blk-iocost.c
@@ -3370,7 +3370,6 @@ static ssize_t ioc_qos_write(struct kernfs_open_file *of, char *input,
size_t nbytes, loff_t off)
{
struct blkg_conf_ctx ctx;
- struct request_queue *q;
struct gendisk *disk;
struct ioc *ioc;
u32 qos[NR_QOS_PARAMS];
@@ -3511,9 +3510,8 @@ static ssize_t ioc_qos_write(struct kernfs_open_file *of, char *input,
blk_mq_unquiesce_queue(disk->queue);
close_bdev:
- q = ctx.bdev->bd_queue;
+ blk_mq_unfreeze_queue(ctx.bdev->bd_queue, memflags);
blkg_conf_close_bdev(&ctx);
- blk_mq_unfreeze_queue(q, memflags);
return ret ?: nbytes;
einval:
--
2.43.7
^ permalink raw reply related [flat|nested] 15+ messages in thread
* [PATCH 5/7] blk-stat: account for the mean of the source in blk_rq_stat_sum()
2026-11-19 7:05 [PATCH 0/7] block fixes found by code inspection Guixin Liu
` (3 preceding siblings ...)
2026-11-19 7:05 ` [PATCH 4/7] blk-iocost: unfreeze the queue before dropping the bdev reference in qos write Guixin Liu
@ 2026-11-19 7:05 ` Guixin Liu
2026-11-19 13:32 ` Tao Cui
2026-11-19 7:05 ` [PATCH 6/7] blk-cgroup: don't attach a qos policy to a dead disk Guixin Liu
2026-11-19 7:05 ` [PATCH 7/7] blk-mq: unpublish the tags before deferring their free to SRCU Guixin Liu
6 siblings, 1 reply; 15+ messages in thread
From: Guixin Liu @ 2026-11-19 7:05 UTC (permalink / raw)
To: Jens Axboe, Christoph Hellwig, Tejun Heo, Josef Bacik,
Nilay Shroff, Ming Lei, Hannes Reinecke
Cc: linux-block
blk_rq_stat_sum() folds a src into a dst by combining dst->mean with
src->batch only, so it is only correct while src has never been folded
itself: its mean is then either zero or stale. blk-iolatency folds the
per-cpu stats into a window-local stat first and then folds that result
into iolat->cur_stat, so the second fold drops the whole window and
cur_stat.rqs.mean stays at zero. On rotating disks, where the mean
decides whether the accumulated window meets the latency target,
iolatency_check_latencies() therefore always considers the accumulated
window OK and scales up more aggressively than intended.
Fold src->mean * src->nr_samples into the result as well. The per-cpu
users keep their mean at zero through blk_rq_stat_init(), so their
behavior is unchanged.
Fixes: 451bb7c33197 ("blk-iolatency: keep track of previous windows stats")
Cc: stable@vger.kernel.org
Signed-off-by: Guixin Liu <kanie@linux.alibaba.com>
---
block/blk-stat.c | 8 ++++++--
1 file changed, 6 insertions(+), 2 deletions(-)
diff --git a/block/blk-stat.c b/block/blk-stat.c
index d57c2fc6bf06..6345ae366658 100644
--- a/block/blk-stat.c
+++ b/block/blk-stat.c
@@ -24,7 +24,10 @@ void blk_rq_stat_init(struct blk_rq_stat *stat)
stat->batch = 0;
}
-/* src is a per-cpu stat, mean isn't initialized */
+/*
+ * The mean of @src must have been calculated or set to zero by
+ * blk_rq_stat_init() before calling this.
+ */
void blk_rq_stat_sum(struct blk_rq_stat *dst, struct blk_rq_stat *src)
{
if (dst->nr_samples + src->nr_samples <= dst->nr_samples)
@@ -33,7 +36,8 @@ void blk_rq_stat_sum(struct blk_rq_stat *dst, struct blk_rq_stat *src)
dst->min = min(dst->min, src->min);
dst->max = max(dst->max, src->max);
- dst->mean = div_u64(src->batch + dst->mean * dst->nr_samples,
+ dst->mean = div_u64(dst->mean * dst->nr_samples +
+ src->mean * src->nr_samples + src->batch,
dst->nr_samples + src->nr_samples);
dst->nr_samples += src->nr_samples;
--
2.43.7
^ permalink raw reply related [flat|nested] 15+ messages in thread
* [PATCH 6/7] blk-cgroup: don't attach a qos policy to a dead disk
2026-11-19 7:05 [PATCH 0/7] block fixes found by code inspection Guixin Liu
` (4 preceding siblings ...)
2026-11-19 7:05 ` [PATCH 5/7] blk-stat: account for the mean of the source in blk_rq_stat_sum() Guixin Liu
@ 2026-11-19 7:05 ` Guixin Liu
2026-11-19 7:05 ` [PATCH 7/7] blk-mq: unpublish the tags before deferring their free to SRCU Guixin Liu
6 siblings, 0 replies; 15+ messages in thread
From: Guixin Liu @ 2026-11-19 7:05 UTC (permalink / raw)
To: Jens Axboe, Christoph Hellwig, Tejun Heo, Josef Bacik,
Nilay Shroff, Ming Lei, Hannes Reinecke
Cc: linux-block
blkg_conf_open_bdev() validates disk_live() once under rq_qos_mutex.
ioc_qos_write() and ioc_cost_model_write() then release that mutex,
freeze the queue and re-acquire it, so __del_gendisk() can run
rq_qos_exit() to completion in the window and detach the existing
policy. After the re-acquire, q_to_ioc() returning NULL now also means
"the policy was torn down", not just "never enabled", and the writer
calls blk_iocost_init() to attach a fresh ioc to the dead disk.
Nothing tears that ioc down again: rq_qos_exit() already ran, and for
disks on a shared queue, e.g. sd/sr, blk_mq_destroy_queue() does not
call it either. The ioc with its timer and per-cpu stats is leaked, and
QUEUE_FLAG_QOS_ENABLED is resurrected on the dead queue. The same
lazy-init pattern in blk_throtl_init() and blk_iolatency_init() is
reachable the same way once their respective config writers get the
release-freeze-reacquire treatment.
Reject a dead disk in all three init helpers.
Fixes: 7caa47151ab2 ("blkcg: implement blk-iocost")
Fixes: a3166c51702b ("blk-throttle: delay initialization until configuration")
Fixes: a13696b83da4 ("blk-iolatency: Make initialization lazy")
Cc: stable@vger.kernel.org
Signed-off-by: Guixin Liu <kanie@linux.alibaba.com>
---
block/blk-iocost.c | 3 +++
block/blk-iolatency.c | 3 +++
block/blk-throttle.c | 3 +++
3 files changed, 9 insertions(+)
diff --git a/block/blk-iocost.c b/block/blk-iocost.c
index b251ecc66d8a..f71d997f232c 100644
--- a/block/blk-iocost.c
+++ b/block/blk-iocost.c
@@ -3009,6 +3009,9 @@ static int blk_iocost_init(struct gendisk *disk)
struct ioc *ioc;
int i, cpu, ret;
+ if (test_bit(GD_DEAD, &disk->state))
+ return -ENODEV;
+
/* The caller holds rq_qos_mutex, which is taken with the queue frozen */
ioc = kzalloc_obj(*ioc, GFP_NOIO);
if (!ioc)
diff --git a/block/blk-iolatency.c b/block/blk-iolatency.c
index a3bd153c67be..553f2091730f 100644
--- a/block/blk-iolatency.c
+++ b/block/blk-iolatency.c
@@ -761,6 +761,9 @@ static int blk_iolatency_init(struct gendisk *disk)
struct blk_iolatency *blkiolat;
int ret;
+ if (test_bit(GD_DEAD, &disk->state))
+ return -ENODEV;
+
/* The caller holds rq_qos_mutex, which is taken with the queue frozen */
blkiolat = kzalloc_obj(*blkiolat, GFP_NOIO);
if (!blkiolat)
diff --git a/block/blk-throttle.c b/block/blk-throttle.c
index c5e26551956a..374ac75832ea 100644
--- a/block/blk-throttle.c
+++ b/block/blk-throttle.c
@@ -1322,6 +1322,9 @@ static int blk_throtl_init(struct gendisk *disk)
unsigned int memflags;
int ret;
+ if (test_bit(GD_DEAD, &disk->state))
+ return -ENODEV;
+
/* The caller holds rq_qos_mutex, which is taken with the queue frozen */
td = kzalloc_node(sizeof(*td), GFP_NOIO, q->node);
if (!td)
--
2.43.7
^ permalink raw reply related [flat|nested] 15+ messages in thread
* [PATCH 7/7] blk-mq: unpublish the tags before deferring their free to SRCU
2026-11-19 7:05 [PATCH 0/7] block fixes found by code inspection Guixin Liu
` (5 preceding siblings ...)
2026-11-19 7:05 ` [PATCH 6/7] blk-cgroup: don't attach a qos policy to a dead disk Guixin Liu
@ 2026-11-19 7:05 ` Guixin Liu
6 siblings, 0 replies; 15+ messages in thread
From: Guixin Liu @ 2026-11-19 7:05 UTC (permalink / raw)
To: Jens Axboe, Christoph Hellwig, Tejun Heo, Josef Bacik,
Nilay Shroff, Ming Lei, Hannes Reinecke
Cc: linux-block
__blk_mq_free_map_and_rqs() frees set->tags[hctx_idx] first, which
starts the tags_srcu grace period, and only then stores NULL into the
slot. The store is a plain store, unordered against the plain load in
blk_mq_tagset_busy_iter(), which reads set->tags[i] to find the tags to
walk. A reader that enters the SRCU read side after the grace period
snapshot, but still loads the old, non-NULL pointer, is not waited for
by that grace period, so blk_mq_free_tags_callback() can free the tags
and everything they point to while the iterator is about to walk them.
blk_mq_tagset_busy_iter() is reachable from the nvme and scsi timeout
and error recovery paths without any queue freeze gating the free, so
nothing else serializes the two.
Clear set->tags[hctx_idx] before blk_mq_free_map_and_rqs(), so a reader
either sees NULL and skips the tags, or loads the pointer before
entering the SRCU read side, in which case the grace period waits for
it.
Fixes: 995412e23bb25 ("blk-mq: Replace tags->lock with SRCU for tag iterators")
Cc: stable@vger.kernel.org
Signed-off-by: Guixin Liu <kanie@linux.alibaba.com>
---
block/blk-mq.c | 5 +++--
1 file changed, 3 insertions(+), 2 deletions(-)
diff --git a/block/blk-mq.c b/block/blk-mq.c
index 87c1ccdf66ab..b58f73256202 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -4179,10 +4179,11 @@ void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
unsigned int hctx_idx)
{
- if (!blk_mq_is_shared_tags(set->flags))
- blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
+ struct blk_mq_tags *tags = set->tags[hctx_idx];
set->tags[hctx_idx] = NULL;
+ if (!blk_mq_is_shared_tags(set->flags))
+ blk_mq_free_map_and_rqs(set, tags, hctx_idx);
}
static void blk_mq_map_swqueue(struct request_queue *q)
--
2.43.7
^ permalink raw reply related [flat|nested] 15+ messages in thread
* Re: [PATCH 5/7] blk-stat: account for the mean of the source in blk_rq_stat_sum()
2026-11-19 7:05 ` [PATCH 5/7] blk-stat: account for the mean of the source in blk_rq_stat_sum() Guixin Liu
@ 2026-11-19 13:32 ` Tao Cui
2026-11-20 2:04 ` Guixin Liu
0 siblings, 1 reply; 15+ messages in thread
From: Tao Cui @ 2026-11-19 13:32 UTC (permalink / raw)
To: Guixin Liu, Jens Axboe, Christoph Hellwig, Tejun Heo, Josef Bacik,
Nilay Shroff, Ming Lei, Hannes Reinecke, Tang Yizhou
Cc: cui.tao, linux-block
Hello, Guixin,
在 2026/11/19 15:05, Guixin Liu 写道:
> blk_rq_stat_sum() folds a src into a dst by combining dst->mean with
> src->batch only, so it is only correct while src has never been folded
> itself: its mean is then either zero or stale. blk-iolatency folds the
> per-cpu stats into a window-local stat first and then folds that result
> into iolat->cur_stat, so the second fold drops the whole window and
> cur_stat.rqs.mean stays at zero. On rotating disks, where the mean
> decides whether the accumulated window meets the latency target,
> iolatency_check_latencies() therefore always considers the accumulated
> window OK and scales up more aggressively than intended.
>
> Fold src->mean * src->nr_samples into the result as well. The per-cpu
> users keep their mean at zero through blk_rq_stat_init(), so their
> behavior is unchanged.
>
I posted a fix for the same issue before, which Tang Yizhou reviewed:
https://lore.kernel.org/all/20260904065426.1281914-1-cui.tao@linux.dev/
Adding him here to take a look at your approach as well. Hope the link helps.
Thanks,
Tao
> Fixes: 451bb7c33197 ("blk-iolatency: keep track of previous windows stats")
> Cc: stable@vger.kernel.org
> Signed-off-by: Guixin Liu <kanie@linux.alibaba.com>
> ---
> block/blk-stat.c | 8 ++++++--
> 1 file changed, 6 insertions(+), 2 deletions(-)
>
> diff --git a/block/blk-stat.c b/block/blk-stat.c
> index d57c2fc6bf06..6345ae366658 100644
> --- a/block/blk-stat.c
> +++ b/block/blk-stat.c
> @@ -24,7 +24,10 @@ void blk_rq_stat_init(struct blk_rq_stat *stat)
> stat->batch = 0;
> }
>
> -/* src is a per-cpu stat, mean isn't initialized */
> +/*
> + * The mean of @src must have been calculated or set to zero by
> + * blk_rq_stat_init() before calling this.
> + */
> void blk_rq_stat_sum(struct blk_rq_stat *dst, struct blk_rq_stat *src)
> {
> if (dst->nr_samples + src->nr_samples <= dst->nr_samples)
> @@ -33,7 +36,8 @@ void blk_rq_stat_sum(struct blk_rq_stat *dst, struct blk_rq_stat *src)
> dst->min = min(dst->min, src->min);
> dst->max = max(dst->max, src->max);
>
> - dst->mean = div_u64(src->batch + dst->mean * dst->nr_samples,
> + dst->mean = div_u64(dst->mean * dst->nr_samples +
> + src->mean * src->nr_samples + src->batch,
> dst->nr_samples + src->nr_samples);
>
> dst->nr_samples += src->nr_samples;
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [PATCH 3/7] blk-iocost: don't take rq_qos_mutex before the queue is frozen in cost model write
2026-11-19 7:05 ` [PATCH 3/7] blk-iocost: don't take rq_qos_mutex before the queue is frozen in cost model write Guixin Liu
@ 2026-11-19 13:40 ` Tao Cui
2026-11-20 2:14 ` Guixin Liu
0 siblings, 1 reply; 15+ messages in thread
From: Tao Cui @ 2026-11-19 13:40 UTC (permalink / raw)
To: Guixin Liu, Jens Axboe, Christoph Hellwig, Tejun Heo, Josef Bacik,
Nilay Shroff, Ming Lei, Hannes Reinecke
Cc: cui.tao, linux-block
Hello, Guixin,
在 2026/11/19 15:05, Guixin Liu 写道:
> ioc_cost_model_write() freezes the queue while holding rq_qos_mutex,
> which blkg_conf_open_bdev() acquired first. Every other path takes
> rq_qos_mutex only after the queue is frozen: ioc_qos_write() releases
> and re-acquires the mutex around blk_mq_freeze_queue(), and rq_qos_add()
> documents that its caller holds the mutex with the queue frozen.
> Waiting for in-flight IO while holding rq_qos_mutex breaks that rule,
> and on the first write to a queue without iocost the path even calls
> blk_iocost_init() -> rq_qos_add() -> blk_mq_freeze_queue() while
> nothing has frozen the queue yet.
>
> Mirror the release-freeze-reacquire sequence of ioc_qos_write(). The
> unfreeze moves to the close_bdev label so the error exits after the
> freeze don't leave the queue frozen, and it runs before
> blkg_conf_close_bdev() drops the bdev reference, while the queue is
> still pinned.
>
I tried the same change in cost model write recently and it turned out flawed, these two reviews explain it better than I can [1][2].
In short: the disk_live() check in blkg_conf_open_bdev() only holds while rq_qos_mutex is kept, and dropping it around the freeze lets the disk and the ioc go away; Tejun reproduced a use-after-free that way.
I've been busy with another thread lately, so I'm glad to see you proposing a fix for this. Hope the links help.
Thanks,
Tao
[1] https://lore.kernel.org/all/20261111163306.GA17267@lst.de/
[2] https://lore.kernel.org/all/bfe5007b2758bb06504b6966abaf1f7e@kernel.org/
> Fixes: 2b2da2f6dc80 ("blk-iocost: prevent configuration update concurrent with io throttling")
> Cc: stable@vger.kernel.org
> Signed-off-by: Guixin Liu <kanie@linux.alibaba.com>
> ---
> block/blk-iocost.c | 12 ++++++++++--
> 1 file changed, 10 insertions(+), 2 deletions(-)
>
> diff --git a/block/blk-iocost.c b/block/blk-iocost.c
> index 1c1025157d8e..8e06b83691aa 100644
> --- a/block/blk-iocost.c
> +++ b/block/blk-iocost.c
> @@ -3589,6 +3589,15 @@ static ssize_t ioc_cost_model_write(struct kernfs_open_file *of, char *input,
> ret = blkg_conf_open_bdev(&ctx);
> if (ret)
> return ret;
> + /*
> + * Nothing is protecting QoS state yet, so release q->rq_qos_mutex,
> + * which was first acquired in blkg_conf_open_bdev, and re-acquire it
> + * after freezing the queue to maintain the correct locking order.
> + */
> + mutex_unlock(&ctx.bdev->bd_queue->rq_qos_mutex);
> +
> + memflags = blk_mq_freeze_queue(ctx.bdev->bd_queue);
> + mutex_lock(&ctx.bdev->bd_queue->rq_qos_mutex);
>
> body = ctx.body;
> q = bdev_get_queue(ctx.bdev);
> @@ -3605,7 +3614,6 @@ static ssize_t ioc_cost_model_write(struct kernfs_open_file *of, char *input,
> ioc = q_to_ioc(q);
> }
>
> - memflags = blk_mq_freeze_queue(q);
> blk_mq_quiesce_queue(q);
>
> spin_lock_irq(&ioc->lock);
> @@ -3663,9 +3671,9 @@ static ssize_t ioc_cost_model_write(struct kernfs_open_file *of, char *input,
> spin_unlock_irq(&ioc->lock);
>
> blk_mq_unquiesce_queue(q);
> - blk_mq_unfreeze_queue(q, memflags);
>
> close_bdev:
> + blk_mq_unfreeze_queue(q, memflags);
> blkg_conf_close_bdev(&ctx);
> return ret ?: nbytes;
> }
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [PATCH 5/7] blk-stat: account for the mean of the source in blk_rq_stat_sum()
2026-11-19 13:32 ` Tao Cui
@ 2026-11-20 2:04 ` Guixin Liu
2026-11-20 2:50 ` Tao Cui
0 siblings, 1 reply; 15+ messages in thread
From: Guixin Liu @ 2026-11-20 2:04 UTC (permalink / raw)
To: Tao Cui, Jens Axboe, Christoph Hellwig, Tejun Heo, Josef Bacik,
Nilay Shroff, Ming Lei, Hannes Reinecke, Tang Yizhou
Cc: linux-block
在 2026/11/19 21:32, Tao Cui 写道:
> Hello, Guixin,
>
> 在 2026/11/19 15:05, Guixin Liu 写道:
>> blk_rq_stat_sum() folds a src into a dst by combining dst->mean with
>> src->batch only, so it is only correct while src has never been folded
>> itself: its mean is then either zero or stale. blk-iolatency folds the
>> per-cpu stats into a window-local stat first and then folds that result
>> into iolat->cur_stat, so the second fold drops the whole window and
>> cur_stat.rqs.mean stays at zero. On rotating disks, where the mean
>> decides whether the accumulated window meets the latency target,
>> iolatency_check_latencies() therefore always considers the accumulated
>> window OK and scales up more aggressively than intended.
>>
>> Fold src->mean * src->nr_samples into the result as well. The per-cpu
>> users keep their mean at zero through blk_rq_stat_init(), so their
>> behavior is unchanged.
>>
> I posted a fix for the same issue before, which Tang Yizhou reviewed:
>
> https://lore.kernel.org/all/20260904065426.1281914-1-cui.tao@linux.dev/
>
> Adding him here to take a look at your approach as well. Hope the link helps.
>
> Thanks,
> Tao
Hi Tao,
My thanks for the review, your v3 is better, I will drop my patch5 in
next version.
Best Regards,
Guixin Liu
>
>> Fixes: 451bb7c33197 ("blk-iolatency: keep track of previous windows stats")
>> Cc: stable@vger.kernel.org
>> Signed-off-by: Guixin Liu <kanie@linux.alibaba.com>
>> ---
>> block/blk-stat.c | 8 ++++++--
>> 1 file changed, 6 insertions(+), 2 deletions(-)
>>
>> diff --git a/block/blk-stat.c b/block/blk-stat.c
>> index d57c2fc6bf06..6345ae366658 100644
>> --- a/block/blk-stat.c
>> +++ b/block/blk-stat.c
>> @@ -24,7 +24,10 @@ void blk_rq_stat_init(struct blk_rq_stat *stat)
>> stat->batch = 0;
>> }
>>
>> -/* src is a per-cpu stat, mean isn't initialized */
>> +/*
>> + * The mean of @src must have been calculated or set to zero by
>> + * blk_rq_stat_init() before calling this.
>> + */
>> void blk_rq_stat_sum(struct blk_rq_stat *dst, struct blk_rq_stat *src)
>> {
>> if (dst->nr_samples + src->nr_samples <= dst->nr_samples)
>> @@ -33,7 +36,8 @@ void blk_rq_stat_sum(struct blk_rq_stat *dst, struct blk_rq_stat *src)
>> dst->min = min(dst->min, src->min);
>> dst->max = max(dst->max, src->max);
>>
>> - dst->mean = div_u64(src->batch + dst->mean * dst->nr_samples,
>> + dst->mean = div_u64(dst->mean * dst->nr_samples +
>> + src->mean * src->nr_samples + src->batch,
>> dst->nr_samples + src->nr_samples);
>>
>> dst->nr_samples += src->nr_samples;
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [PATCH 3/7] blk-iocost: don't take rq_qos_mutex before the queue is frozen in cost model write
2026-11-19 13:40 ` Tao Cui
@ 2026-11-20 2:14 ` Guixin Liu
2026-11-20 2:56 ` Tao Cui
0 siblings, 1 reply; 15+ messages in thread
From: Guixin Liu @ 2026-11-20 2:14 UTC (permalink / raw)
To: Tao Cui, Jens Axboe, Christoph Hellwig, Tejun Heo, Josef Bacik,
Nilay Shroff, Ming Lei, Hannes Reinecke
Cc: linux-block
在 2026/11/19 21:40, Tao Cui 写道:
> Hello, Guixin,
>
> 在 2026/11/19 15:05, Guixin Liu 写道:
>> ioc_cost_model_write() freezes the queue while holding rq_qos_mutex,
>> which blkg_conf_open_bdev() acquired first. Every other path takes
>> rq_qos_mutex only after the queue is frozen: ioc_qos_write() releases
>> and re-acquires the mutex around blk_mq_freeze_queue(), and rq_qos_add()
>> documents that its caller holds the mutex with the queue frozen.
>> Waiting for in-flight IO while holding rq_qos_mutex breaks that rule,
>> and on the first write to a queue without iocost the path even calls
>> blk_iocost_init() -> rq_qos_add() -> blk_mq_freeze_queue() while
>> nothing has frozen the queue yet.
>>
>> Mirror the release-freeze-reacquire sequence of ioc_qos_write(). The
>> unfreeze moves to the close_bdev label so the error exits after the
>> freeze don't leave the queue frozen, and it runs before
>> blkg_conf_close_bdev() drops the bdev reference, while the queue is
>> still pinned.
>>
> I tried the same change in cost model write recently and it turned out flawed, these two reviews explain it better than I can [1][2].
>
> In short: the disk_live() check in blkg_conf_open_bdev() only holds while rq_qos_mutex is kept, and dropping it around the freeze lets the disk and the ioc go away; Tejun reproduced a use-after-free that way.
>
> I've been busy with another thread lately, so I'm glad to see you proposing a fix for this. Hope the links help.
>
> Thanks,
> Tao
>
> [1] https://lore.kernel.org/all/20261111163306.GA17267@lst.de/
> [2] https://lore.kernel.org/all/bfe5007b2758bb06504b6966abaf1f7e@kernel.org/
>
Hi Tao,
Thanks for the review, right, this patch enter the same issue, I will drop this
in next version.
And about the ioc_qos_write()'s disk_live race window, My patch 6 fixed the issue,
could you please take a look?
Best Regards,
Guixin Liu
>> Fixes: 2b2da2f6dc80 ("blk-iocost: prevent configuration update concurrent with io throttling")
>> Cc: stable@vger.kernel.org
>> Signed-off-by: Guixin Liu <kanie@linux.alibaba.com>
>> ---
>> block/blk-iocost.c | 12 ++++++++++--
>> 1 file changed, 10 insertions(+), 2 deletions(-)
>>
>> diff --git a/block/blk-iocost.c b/block/blk-iocost.c
>> index 1c1025157d8e..8e06b83691aa 100644
>> --- a/block/blk-iocost.c
>> +++ b/block/blk-iocost.c
>> @@ -3589,6 +3589,15 @@ static ssize_t ioc_cost_model_write(struct kernfs_open_file *of, char *input,
>> ret = blkg_conf_open_bdev(&ctx);
>> if (ret)
>> return ret;
>> + /*
>> + * Nothing is protecting QoS state yet, so release q->rq_qos_mutex,
>> + * which was first acquired in blkg_conf_open_bdev, and re-acquire it
>> + * after freezing the queue to maintain the correct locking order.
>> + */
>> + mutex_unlock(&ctx.bdev->bd_queue->rq_qos_mutex);
>> +
>> + memflags = blk_mq_freeze_queue(ctx.bdev->bd_queue);
>> + mutex_lock(&ctx.bdev->bd_queue->rq_qos_mutex);
>>
>> body = ctx.body;
>> q = bdev_get_queue(ctx.bdev);
>> @@ -3605,7 +3614,6 @@ static ssize_t ioc_cost_model_write(struct kernfs_open_file *of, char *input,
>> ioc = q_to_ioc(q);
>> }
>>
>> - memflags = blk_mq_freeze_queue(q);
>> blk_mq_quiesce_queue(q);
>>
>> spin_lock_irq(&ioc->lock);
>> @@ -3663,9 +3671,9 @@ static ssize_t ioc_cost_model_write(struct kernfs_open_file *of, char *input,
>> spin_unlock_irq(&ioc->lock);
>>
>> blk_mq_unquiesce_queue(q);
>> - blk_mq_unfreeze_queue(q, memflags);
>>
>> close_bdev:
>> + blk_mq_unfreeze_queue(q, memflags);
>> blkg_conf_close_bdev(&ctx);
>> return ret ?: nbytes;
>> }
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [PATCH 5/7] blk-stat: account for the mean of the source in blk_rq_stat_sum()
2026-11-20 2:04 ` Guixin Liu
@ 2026-11-20 2:50 ` Tao Cui
2026-11-20 6:22 ` Guixin Liu
0 siblings, 1 reply; 15+ messages in thread
From: Tao Cui @ 2026-11-20 2:50 UTC (permalink / raw)
To: Guixin Liu, Jens Axboe, Christoph Hellwig, Tejun Heo, Josef Bacik,
Nilay Shroff, Ming Lei, Hannes Reinecke, Tang Yizhou
Cc: cui.tao, linux-block
Hello, Guixin
在 2026/11/20 10:04, Guixin Liu 写道:
>
>
> 在 2026/11/19 21:32, Tao Cui 写道:
>> Hello, Guixin,
>>
>> 在 2026/11/19 15:05, Guixin Liu 写道:
>>> blk_rq_stat_sum() folds a src into a dst by combining dst->mean with
>>> src->batch only, so it is only correct while src has never been folded
>>> itself: its mean is then either zero or stale. blk-iolatency folds the
>>> per-cpu stats into a window-local stat first and then folds that result
>>> into iolat->cur_stat, so the second fold drops the whole window and
>>> cur_stat.rqs.mean stays at zero. On rotating disks, where the mean
>>> decides whether the accumulated window meets the latency target,
>>> iolatency_check_latencies() therefore always considers the accumulated
>>> window OK and scales up more aggressively than intended.
>>>
>>> Fold src->mean * src->nr_samples into the result as well. The per-cpu
>>> users keep their mean at zero through blk_rq_stat_init(), so their
>>> behavior is unchanged.
>>>
>> I posted a fix for the same issue before, which Tang Yizhou reviewed:
>>
>> https://lore.kernel.org/all/20260904065426.1281914-1-cui.tao@linux.dev/
>>
>> Adding him here to take a look at your approach as well. Hope the link helps.
>>
>> Thanks,
>> Tao
> Hi Tao,
> My thanks for the review, your v3 is better, I will drop my patch5 in
> next version.
>
Sorry, that's not what I meant. The issue is real and your patch works; I just felt changing the contract of blk_rq_stat_sum() has a wider impact, so I adapted the fix to the contract in the caller instead. That said, my v3 has been pending for months without review, so it may well have problems I haven't noticed myself, seeing your patch I also realized my Fixes tag is wrong.
Thanks,
Tao
> Best Regards,
> Guixin Liu
>>
>>> Fixes: 451bb7c33197 ("blk-iolatency: keep track of previous windows stats")
>>> Cc: stable@vger.kernel.org
>>> Signed-off-by: Guixin Liu <kanie@linux.alibaba.com>
>>> ---
>>> block/blk-stat.c | 8 ++++++--
>>> 1 file changed, 6 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/block/blk-stat.c b/block/blk-stat.c
>>> index d57c2fc6bf06..6345ae366658 100644
>>> --- a/block/blk-stat.c
>>> +++ b/block/blk-stat.c
>>> @@ -24,7 +24,10 @@ void blk_rq_stat_init(struct blk_rq_stat *stat)
>>> stat->batch = 0;
>>> }
>>>
>>> -/* src is a per-cpu stat, mean isn't initialized */
>>> +/*
>>> + * The mean of @src must have been calculated or set to zero by
>>> + * blk_rq_stat_init() before calling this.
>>> + */
>>> void blk_rq_stat_sum(struct blk_rq_stat *dst, struct blk_rq_stat *src)
>>> {
>>> if (dst->nr_samples + src->nr_samples <= dst->nr_samples)
>>> @@ -33,7 +36,8 @@ void blk_rq_stat_sum(struct blk_rq_stat *dst, struct blk_rq_stat *src)
>>> dst->min = min(dst->min, src->min);
>>> dst->max = max(dst->max, src->max);
>>>
>>> - dst->mean = div_u64(src->batch + dst->mean * dst->nr_samples,
>>> + dst->mean = div_u64(dst->mean * dst->nr_samples +
>>> + src->mean * src->nr_samples + src->batch,
>>> dst->nr_samples + src->nr_samples);
>>>
>>> dst->nr_samples += src->nr_samples;
>
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [PATCH 3/7] blk-iocost: don't take rq_qos_mutex before the queue is frozen in cost model write
2026-11-20 2:14 ` Guixin Liu
@ 2026-11-20 2:56 ` Tao Cui
0 siblings, 0 replies; 15+ messages in thread
From: Tao Cui @ 2026-11-20 2:56 UTC (permalink / raw)
To: Guixin Liu, Jens Axboe, Christoph Hellwig, Tejun Heo, Josef Bacik,
Nilay Shroff, Ming Lei, Hannes Reinecke
Cc: cui.tao, linux-block
Hello, Guixin,
在 2026/11/20 10:14, Guixin Liu 写道:
>
>
> 在 2026/11/19 21:40, Tao Cui 写道:
>> Hello, Guixin,
>>
>> 在 2026/11/19 15:05, Guixin Liu 写道:
>>> ioc_cost_model_write() freezes the queue while holding rq_qos_mutex,
>>> which blkg_conf_open_bdev() acquired first. Every other path takes
>>> rq_qos_mutex only after the queue is frozen: ioc_qos_write() releases
>>> and re-acquires the mutex around blk_mq_freeze_queue(), and rq_qos_add()
>>> documents that its caller holds the mutex with the queue frozen.
>>> Waiting for in-flight IO while holding rq_qos_mutex breaks that rule,
>>> and on the first write to a queue without iocost the path even calls
>>> blk_iocost_init() -> rq_qos_add() -> blk_mq_freeze_queue() while
>>> nothing has frozen the queue yet.
>>>
>>> Mirror the release-freeze-reacquire sequence of ioc_qos_write(). The
>>> unfreeze moves to the close_bdev label so the error exits after the
>>> freeze don't leave the queue frozen, and it runs before
>>> blkg_conf_close_bdev() drops the bdev reference, while the queue is
>>> still pinned.
>>>
>> I tried the same change in cost model write recently and it turned out flawed, these two reviews explain it better than I can [1][2].
>>
>> In short: the disk_live() check in blkg_conf_open_bdev() only holds while rq_qos_mutex is kept, and dropping it around the freeze lets the disk and the ioc go away; Tejun reproduced a use-after-free that way.
>>
>> I've been busy with another thread lately, so I'm glad to see you proposing a fix for this. Hope the links help.
>>
>> Thanks,
>> Tao
>>
>> [1] https://lore.kernel.org/all/20261111163306.GA17267@lst.de/
>> [2] https://lore.kernel.org/all/bfe5007b2758bb06504b6966abaf1f7e@kernel.org/
>>
> Hi Tao,
> Thanks for the review, right, this patch enter the same issue, I will drop this
> in next version.
>
Thanks for checking the history before reworking. I'm still busy with another thread, so if you have time, feel free to take this forward.
Thanks,
Tao
> And about the ioc_qos_write()'s disk_live race window, My patch 6 fixed the issue,
> could you please take a look?
>
> Best Regards,
> Guixin Liu
>>> Fixes: 2b2da2f6dc80 ("blk-iocost: prevent configuration update concurrent with io throttling")
>>> Cc: stable@vger.kernel.org
>>> Signed-off-by: Guixin Liu <kanie@linux.alibaba.com>
>>> ---
>>> block/blk-iocost.c | 12 ++++++++++--
>>> 1 file changed, 10 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/block/blk-iocost.c b/block/blk-iocost.c
>>> index 1c1025157d8e..8e06b83691aa 100644
>>> --- a/block/blk-iocost.c
>>> +++ b/block/blk-iocost.c
>>> @@ -3589,6 +3589,15 @@ static ssize_t ioc_cost_model_write(struct kernfs_open_file *of, char *input,
>>> ret = blkg_conf_open_bdev(&ctx);
>>> if (ret)
>>> return ret;
>>> + /*
>>> + * Nothing is protecting QoS state yet, so release q->rq_qos_mutex,
>>> + * which was first acquired in blkg_conf_open_bdev, and re-acquire it
>>> + * after freezing the queue to maintain the correct locking order.
>>> + */
>>> + mutex_unlock(&ctx.bdev->bd_queue->rq_qos_mutex);
>>> +
>>> + memflags = blk_mq_freeze_queue(ctx.bdev->bd_queue);
>>> + mutex_lock(&ctx.bdev->bd_queue->rq_qos_mutex);
>>>
>>> body = ctx.body;
>>> q = bdev_get_queue(ctx.bdev);
>>> @@ -3605,7 +3614,6 @@ static ssize_t ioc_cost_model_write(struct kernfs_open_file *of, char *input,
>>> ioc = q_to_ioc(q);
>>> }
>>>
>>> - memflags = blk_mq_freeze_queue(q);
>>> blk_mq_quiesce_queue(q);
>>>
>>> spin_lock_irq(&ioc->lock);
>>> @@ -3663,9 +3671,9 @@ static ssize_t ioc_cost_model_write(struct kernfs_open_file *of, char *input,
>>> spin_unlock_irq(&ioc->lock);
>>>
>>> blk_mq_unquiesce_queue(q);
>>> - blk_mq_unfreeze_queue(q, memflags);
>>>
>>> close_bdev:
>>> + blk_mq_unfreeze_queue(q, memflags);
>>> blkg_conf_close_bdev(&ctx);
>>> return ret ?: nbytes;
>>> }
>
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [PATCH 5/7] blk-stat: account for the mean of the source in blk_rq_stat_sum()
2026-11-20 2:50 ` Tao Cui
@ 2026-11-20 6:22 ` Guixin Liu
0 siblings, 0 replies; 15+ messages in thread
From: Guixin Liu @ 2026-11-20 6:22 UTC (permalink / raw)
To: Tao Cui, Jens Axboe, Christoph Hellwig, Tejun Heo, Josef Bacik,
Nilay Shroff, Hannes Reinecke, Tang Yizhou
Cc: linux-block
在 2026/11/20 10:50, Tao Cui 写道:
> Hello, Guixin
>
> 在 2026/11/20 10:04, Guixin Liu 写道:
>>
>> 在 2026/11/19 21:32, Tao Cui 写道:
>>> Hello, Guixin,
>>>
>>> 在 2026/11/19 15:05, Guixin Liu 写道:
>>>> blk_rq_stat_sum() folds a src into a dst by combining dst->mean with
>>>> src->batch only, so it is only correct while src has never been folded
>>>> itself: its mean is then either zero or stale. blk-iolatency folds the
>>>> per-cpu stats into a window-local stat first and then folds that result
>>>> into iolat->cur_stat, so the second fold drops the whole window and
>>>> cur_stat.rqs.mean stays at zero. On rotating disks, where the mean
>>>> decides whether the accumulated window meets the latency target,
>>>> iolatency_check_latencies() therefore always considers the accumulated
>>>> window OK and scales up more aggressively than intended.
>>>>
>>>> Fold src->mean * src->nr_samples into the result as well. The per-cpu
>>>> users keep their mean at zero through blk_rq_stat_init(), so their
>>>> behavior is unchanged.
>>>>
>>> I posted a fix for the same issue before, which Tang Yizhou reviewed:
>>>
>>> https://lore.kernel.org/all/20260904065426.1281914-1-cui.tao@linux.dev/
>>>
>>> Adding him here to take a look at your approach as well. Hope the link helps.
>>>
>>> Thanks,
>>> Tao
>> Hi Tao,
>> My thanks for the review, your v3 is better, I will drop my patch5 in
>> next version.
>>
> Sorry, that's not what I meant. The issue is real and your patch works; I just felt changing the contract of blk_rq_stat_sum() has a wider impact, so I adapted the fix to the contract in the caller instead. That said, my v3 has been pending for months without review, so it may well have problems I haven't noticed myself, seeing your patch I also realized my Fixes tag is wrong.
OK, I will hold, waiting for other miantainers to review this patch.
Best Regards,
Guixin Liu
>
> Thanks,
> Tao
>> Best Regards,
>> Guixin Liu
>>>> Fixes: 451bb7c33197 ("blk-iolatency: keep track of previous windows stats")
>>>> Cc: stable@vger.kernel.org
>>>> Signed-off-by: Guixin Liu <kanie@linux.alibaba.com>
>>>> ---
>>>> block/blk-stat.c | 8 ++++++--
>>>> 1 file changed, 6 insertions(+), 2 deletions(-)
>>>>
>>>> diff --git a/block/blk-stat.c b/block/blk-stat.c
>>>> index d57c2fc6bf06..6345ae366658 100644
>>>> --- a/block/blk-stat.c
>>>> +++ b/block/blk-stat.c
>>>> @@ -24,7 +24,10 @@ void blk_rq_stat_init(struct blk_rq_stat *stat)
>>>> stat->batch = 0;
>>>> }
>>>>
>>>> -/* src is a per-cpu stat, mean isn't initialized */
>>>> +/*
>>>> + * The mean of @src must have been calculated or set to zero by
>>>> + * blk_rq_stat_init() before calling this.
>>>> + */
>>>> void blk_rq_stat_sum(struct blk_rq_stat *dst, struct blk_rq_stat *src)
>>>> {
>>>> if (dst->nr_samples + src->nr_samples <= dst->nr_samples)
>>>> @@ -33,7 +36,8 @@ void blk_rq_stat_sum(struct blk_rq_stat *dst, struct blk_rq_stat *src)
>>>> dst->min = min(dst->min, src->min);
>>>> dst->max = max(dst->max, src->max);
>>>>
>>>> - dst->mean = div_u64(src->batch + dst->mean * dst->nr_samples,
>>>> + dst->mean = div_u64(dst->mean * dst->nr_samples +
>>>> + src->mean * src->nr_samples + src->batch,
>>>> dst->nr_samples + src->nr_samples);
>>>>
>>>> dst->nr_samples += src->nr_samples;
^ permalink raw reply [flat|nested] 15+ messages in thread
end of thread, other threads:[~2026-11-20 6:22 UTC | newest]
Thread overview: 15+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2026-11-19 7:05 [PATCH 0/7] block fixes found by code inspection Guixin Liu
2026-11-19 7:05 ` [PATCH 1/7] blk-mq: free the rqs arrays of tags in the SRCU callback Guixin Liu
2026-11-19 7:05 ` [PATCH 2/7] blk-mq: clear stale stopped flag on hctx reuse Guixin Liu
2026-11-19 7:05 ` [PATCH 3/7] blk-iocost: don't take rq_qos_mutex before the queue is frozen in cost model write Guixin Liu
2026-11-19 13:40 ` Tao Cui
2026-11-20 2:14 ` Guixin Liu
2026-11-20 2:56 ` Tao Cui
2026-11-19 7:05 ` [PATCH 4/7] blk-iocost: unfreeze the queue before dropping the bdev reference in qos write Guixin Liu
2026-11-19 7:05 ` [PATCH 5/7] blk-stat: account for the mean of the source in blk_rq_stat_sum() Guixin Liu
2026-11-19 13:32 ` Tao Cui
2026-11-20 2:04 ` Guixin Liu
2026-11-20 2:50 ` Tao Cui
2026-11-20 6:22 ` Guixin Liu
2026-11-19 7:05 ` [PATCH 6/7] blk-cgroup: don't attach a qos policy to a dead disk Guixin Liu
2026-11-19 7:05 ` [PATCH 7/7] blk-mq: unpublish the tags before deferring their free to SRCU Guixin Liu
This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.