* [PATCH 0/2] dm thin: stop waiting for grace periods when activating and removing thins
@ 2026-11-16 5:12 Qiliang Yuan
2026-11-16 5:12 ` [PATCH 1/2] dm thin: drop the grace period from thin_ctr() Qiliang Yuan
2026-11-16 5:12 ` [PATCH 2/2] dm thin: free thins with kfree_rcu() instead of waiting in thin_dtr() Qiliang Yuan
0 siblings, 2 replies; 5+ messages in thread
From: Qiliang Yuan @ 2026-11-16 5:12 UTC (permalink / raw)
To: Alasdair Kergon, Mike Snitzer, Mikulas Patocka, Benjamin Marzinski
Cc: dm-devel, linux-kernel, Qiliang Yuan
Activating a thin device waits for an RCU grace period in thin_ctr(),
and removing one waits for another in thin_dtr(). Tools that create and
remove thin snapshots for every container, such as the containerd
devmapper snapshotter, pay for both each time.
Patch 1 drops the grace period from thin_ctr(), which isn't needed.
Patch 2 lets the active_thins walkers skip thins whose last reference is
gone, so that thin_dtr() can free the thin with kfree_rcu() instead of
waiting. Activating a thin device gets 51% faster and removing one 15%.
Signed-off-by: Qiliang Yuan <qiliang.yuan@linux.dev>
---
Qiliang Yuan (2):
dm thin: drop the grace period from thin_ctr()
dm thin: free thins with kfree_rcu() instead of waiting in thin_dtr()
drivers/md/dm-thin.c | 37 ++++++++++++++++++-------------------
1 file changed, 18 insertions(+), 19 deletions(-)
---
base-commit: 23c62304dde29abe92013d85910d03a77dada35d
change-id: 20261114-bug-dm-thin-rcu-gp-d29d4b74a0b8
Best regards,
--
Qiliang Yuan <qiliang.yuan@linux.dev>
^ permalink raw reply [flat|nested] 5+ messages in thread
* [PATCH 1/2] dm thin: drop the grace period from thin_ctr()
2026-11-16 5:12 [PATCH 0/2] dm thin: stop waiting for grace periods when activating and removing thins Qiliang Yuan
@ 2026-11-16 5:12 ` Qiliang Yuan
2026-11-18 0:51 ` Benjamin Marzinski
2026-11-16 5:12 ` [PATCH 2/2] dm thin: free thins with kfree_rcu() instead of waiting in thin_dtr() Qiliang Yuan
1 sibling, 1 reply; 5+ messages in thread
From: Qiliang Yuan @ 2026-11-16 5:12 UTC (permalink / raw)
To: Alasdair Kergon, Mike Snitzer, Mikulas Patocka, Benjamin Marzinski
Cc: dm-devel, linux-kernel, Qiliang Yuan
thin_ctr() adds the new thin to pool->active_thins with
list_add_tail_rcu() and then calls synchronize_rcu(), so that a
wake_worker() call can't miss the bios of a thin that isn't visible yet.
The grace period doesn't make the entry any more visible. A worker that
walks the list after list_add_tail_rcu() finds the thin, and the thin
gets no bios before its table is resumed. A walk already in progress may
miss it, but the wake_worker() for the first bio queues the worker
again. Still, activating a thin device waits for a full grace period,
which takes about half of the activation.
Remove the synchronize_rcu() call.
Activating 100 thin devices of an idle pool with dmsetup create takes
per device (mean of five rounds, min-max of the rounds in parentheses):
before (ms) after (ms) change
15.4 (14.9-15.9) 7.6 (7.3-7.8) -51%
Signed-off-by: Qiliang Yuan <qiliang.yuan@linux.dev>
---
drivers/md/dm-thin.c | 7 -------
1 file changed, 7 deletions(-)
diff --git a/drivers/md/dm-thin.c b/drivers/md/dm-thin.c
index 358ed77ffb2be..d5ff937eeacba 100644
--- a/drivers/md/dm-thin.c
+++ b/drivers/md/dm-thin.c
@@ -4289,13 +4289,6 @@ static int thin_ctr(struct dm_target *ti, unsigned int argc, char **argv)
init_completion(&tc->can_destroy);
list_add_tail_rcu(&tc->list, &tc->pool->active_thins);
spin_unlock_irq(&tc->pool->lock);
- /*
- * This synchronize_rcu() call is needed here otherwise we risk a
- * wake_worker() call finding no bios to process (because the newly
- * added tc isn't yet visible). So this reduces latency since we
- * aren't then dependent on the periodic commit to wake_worker().
- */
- synchronize_rcu();
dm_put(pool_md);
--
2.43.0
^ permalink raw reply [flat|nested] 5+ messages in thread
* [PATCH 2/2] dm thin: free thins with kfree_rcu() instead of waiting in thin_dtr()
2026-11-16 5:12 [PATCH 0/2] dm thin: stop waiting for grace periods when activating and removing thins Qiliang Yuan
2026-11-16 5:12 ` [PATCH 1/2] dm thin: drop the grace period from thin_ctr() Qiliang Yuan
@ 2026-11-16 5:12 ` Qiliang Yuan
2026-11-18 1:13 ` Benjamin Marzinski
1 sibling, 1 reply; 5+ messages in thread
From: Qiliang Yuan @ 2026-11-16 5:12 UTC (permalink / raw)
To: Alasdair Kergon, Mike Snitzer, Mikulas Patocka, Benjamin Marzinski
Cc: dm-devel, linux-kernel, Qiliang Yuan
get_first_thin() and get_next_thin() find a thin on pool->active_thins
under rcu_read_lock() and take a reference on it. thin_dtr() removes the
thin from the list and waits for a grace period before dropping its own
reference, so that no walker takes one after the last is gone.
Removing a thin device thus waits for a full grace period in thin_dtr(),
before it waits for the references that are still held.
Take the references with refcount_inc_not_zero() and skip the thins
whose last reference is gone, so that thin_dtr() only waits for the
references already taken. Free the thin with kfree_rcu(), as
requeue_bios() and error_retry_list_with_code() may still look at it
under rcu_read_lock().
Removing 100 thin devices of an idle pool with dmsetup remove takes per
device (mean of five rounds, min-max of the rounds in parentheses). The
queue freeze in del_gendisk() waits for a grace period of its own, which
the one in thin_dtr() used to overlap, so only part of it is saved:
before (ms) after (ms) change
16.9 (15.8-17.3) 14.4 (13.8-15.3) -15%
Signed-off-by: Qiliang Yuan <qiliang.yuan@linux.dev>
---
drivers/md/dm-thin.c | 30 ++++++++++++++++++------------
1 file changed, 18 insertions(+), 12 deletions(-)
diff --git a/drivers/md/dm-thin.c b/drivers/md/dm-thin.c
index d5ff937eeacba..8386215101eb2 100644
--- a/drivers/md/dm-thin.c
+++ b/drivers/md/dm-thin.c
@@ -360,6 +360,7 @@ struct thin_c {
*/
refcount_t refcount;
struct completion can_destroy;
+ struct rcu_head rcu;
};
/*----------------------------------------------------------------*/
@@ -2317,25 +2318,29 @@ static void process_thin_deferred_cells(struct thin_c *tc)
} while (!list_empty(&cells));
}
-static void thin_get(struct thin_c *tc);
+static bool thin_tryget(struct thin_c *tc);
static void thin_put(struct thin_c *tc);
/*
* We can't hold rcu_read_lock() around code that can block. So we
* find a thin with the rcu lock held; bump a refcount; then drop
- * the lock.
+ * the lock. Skip the thins whose refcount thin_dtr() already dropped
+ * to zero.
*/
static struct thin_c *get_first_thin(struct pool *pool)
{
- struct thin_c *tc = NULL;
+ struct thin_c *tc;
rcu_read_lock();
- tc = list_first_or_null_rcu(&pool->active_thins, struct thin_c, list);
- if (tc)
- thin_get(tc);
+ list_for_each_entry_rcu(tc, &pool->active_thins, list) {
+ if (thin_tryget(tc)) {
+ rcu_read_unlock();
+ return tc;
+ }
+ }
rcu_read_unlock();
- return tc;
+ return NULL;
}
static struct thin_c *get_next_thin(struct pool *pool, struct thin_c *tc)
@@ -2344,7 +2349,8 @@ static struct thin_c *get_next_thin(struct pool *pool, struct thin_c *tc)
rcu_read_lock();
list_for_each_entry_continue_rcu(tc, &pool->active_thins, list) {
- thin_get(tc);
+ if (!thin_tryget(tc))
+ continue;
thin_put(old_tc);
rcu_read_unlock();
return tc;
@@ -4128,9 +4134,9 @@ static struct target_type pool_target = {
* Thin target methods
*--------------------------------------------------------------
*/
-static void thin_get(struct thin_c *tc)
+static bool thin_tryget(struct thin_c *tc)
{
- refcount_inc(&tc->refcount);
+ return refcount_inc_not_zero(&tc->refcount);
}
static void thin_put(struct thin_c *tc)
@@ -4146,7 +4152,6 @@ static void thin_dtr(struct dm_target *ti)
spin_lock_irq(&tc->pool->lock);
list_del_rcu(&tc->list);
spin_unlock_irq(&tc->pool->lock);
- synchronize_rcu();
thin_put(tc);
wait_for_completion(&tc->can_destroy);
@@ -4158,7 +4163,8 @@ static void thin_dtr(struct dm_target *ti)
dm_put_device(ti, tc->pool_dev);
if (tc->origin_dev)
dm_put_device(ti, tc->origin_dev);
- kfree(tc);
+ /* Let RCU readers of active_thins finish with tc before freeing it */
+ kfree_rcu(tc, rcu);
mutex_unlock(&dm_thin_pool_table.mutex);
}
--
2.43.0
^ permalink raw reply [flat|nested] 5+ messages in thread
* Re: [PATCH 1/2] dm thin: drop the grace period from thin_ctr()
2026-11-16 5:12 ` [PATCH 1/2] dm thin: drop the grace period from thin_ctr() Qiliang Yuan
@ 2026-11-18 0:51 ` Benjamin Marzinski
0 siblings, 0 replies; 5+ messages in thread
From: Benjamin Marzinski @ 2026-11-18 0:51 UTC (permalink / raw)
To: Qiliang Yuan
Cc: Alasdair Kergon, Mike Snitzer, Mikulas Patocka, dm-devel, linux-kernel
On Mon, Nov 16, 2026 at 12:12:03AM -0500, Qiliang Yuan wrote:
> thin_ctr() adds the new thin to pool->active_thins with
> list_add_tail_rcu() and then calls synchronize_rcu(), so that a
> wake_worker() call can't miss the bios of a thin that isn't visible yet.
>
> The grace period doesn't make the entry any more visible. A worker that
> walks the list after list_add_tail_rcu() finds the thin, and the thin
> gets no bios before its table is resumed. A walk already in progress may
> miss it, but the wake_worker() for the first bio queues the worker
> again. Still, activating a thin device waits for a full grace period,
> which takes about half of the activation.
>
> Remove the synchronize_rcu() call.
>
> Activating 100 thin devices of an idle pool with dmsetup create takes
> per device (mean of five rounds, min-max of the rounds in parentheses):
>
> before (ms) after (ms) change
> 15.4 (14.9-15.9) 7.6 (7.3-7.8) -51%
>
> Signed-off-by: Qiliang Yuan <qiliang.yuan@linux.dev>
Reviewed-by: Benjamin Marzinski <bmarzins@redhat.com>
^ permalink raw reply [flat|nested] 5+ messages in thread
* Re: [PATCH 2/2] dm thin: free thins with kfree_rcu() instead of waiting in thin_dtr()
2026-11-16 5:12 ` [PATCH 2/2] dm thin: free thins with kfree_rcu() instead of waiting in thin_dtr() Qiliang Yuan
@ 2026-11-18 1:13 ` Benjamin Marzinski
0 siblings, 0 replies; 5+ messages in thread
From: Benjamin Marzinski @ 2026-11-18 1:13 UTC (permalink / raw)
To: Qiliang Yuan
Cc: Alasdair Kergon, Mike Snitzer, Mikulas Patocka, dm-devel, linux-kernel
On Mon, Nov 16, 2026 at 12:12:04AM -0500, Qiliang Yuan wrote:
> get_first_thin() and get_next_thin() find a thin on pool->active_thins
> under rcu_read_lock() and take a reference on it. thin_dtr() removes the
> thin from the list and waits for a grace period before dropping its own
> reference, so that no walker takes one after the last is gone.
>
> Removing a thin device thus waits for a full grace period in thin_dtr(),
> before it waits for the references that are still held.
>
> Take the references with refcount_inc_not_zero() and skip the thins
> whose last reference is gone, so that thin_dtr() only waits for the
> references already taken. Free the thin with kfree_rcu(), as
> requeue_bios() and error_retry_list_with_code() may still look at it
> under rcu_read_lock().
>
> Removing 100 thin devices of an idle pool with dmsetup remove takes per
> device (mean of five rounds, min-max of the rounds in parentheses). The
> queue freeze in del_gendisk() waits for a grace period of its own, which
> the one in thin_dtr() used to overlap, so only part of it is saved:
>
> before (ms) after (ms) change
> 16.9 (15.8-17.3) 14.4 (13.8-15.3) -15%
>
> Signed-off-by: Qiliang Yuan <qiliang.yuan@linux.dev>
So, this isn't directly a problem with your patch, but in reviewing it,
I noticed a bug in the interaction between thin_dtr() and
get_next_thin(). thin_dtr() can take tc off the active_thins list, while
it's being held by a worker. The refcount keeps it from being deleted.
But once tc is removed from the list, there's nothing that guarantees
that tc->list.next is still meaningful. If "dmsetup remove" is called on
the thin device that is next in the active_thins list, that thin_dtr()
call can free the thin pointed to by the original tc->list.next pointer.
So when the original thin_dtr() calls get_next_thin(), it will access
freed memory.
I think the solution is for thin_dtr to mark tc as dead somehow, instead
of removing it from active_thins where we currently do. Once the
refcount drops to 0, we can remove it, wait for a grace period, and free
it.
Or am I missing something subtle that keeps this issue from happening?
-Ben
> ---
> drivers/md/dm-thin.c | 30 ++++++++++++++++++------------
> 1 file changed, 18 insertions(+), 12 deletions(-)
>
> diff --git a/drivers/md/dm-thin.c b/drivers/md/dm-thin.c
> index d5ff937eeacba..8386215101eb2 100644
> --- a/drivers/md/dm-thin.c
> +++ b/drivers/md/dm-thin.c
> @@ -360,6 +360,7 @@ struct thin_c {
> */
> refcount_t refcount;
> struct completion can_destroy;
> + struct rcu_head rcu;
> };
>
> /*----------------------------------------------------------------*/
> @@ -2317,25 +2318,29 @@ static void process_thin_deferred_cells(struct thin_c *tc)
> } while (!list_empty(&cells));
> }
>
> -static void thin_get(struct thin_c *tc);
> +static bool thin_tryget(struct thin_c *tc);
> static void thin_put(struct thin_c *tc);
>
> /*
> * We can't hold rcu_read_lock() around code that can block. So we
> * find a thin with the rcu lock held; bump a refcount; then drop
> - * the lock.
> + * the lock. Skip the thins whose refcount thin_dtr() already dropped
> + * to zero.
> */
> static struct thin_c *get_first_thin(struct pool *pool)
> {
> - struct thin_c *tc = NULL;
> + struct thin_c *tc;
>
> rcu_read_lock();
> - tc = list_first_or_null_rcu(&pool->active_thins, struct thin_c, list);
> - if (tc)
> - thin_get(tc);
> + list_for_each_entry_rcu(tc, &pool->active_thins, list) {
> + if (thin_tryget(tc)) {
> + rcu_read_unlock();
> + return tc;
> + }
> + }
> rcu_read_unlock();
>
> - return tc;
> + return NULL;
> }
>
> static struct thin_c *get_next_thin(struct pool *pool, struct thin_c *tc)
> @@ -2344,7 +2349,8 @@ static struct thin_c *get_next_thin(struct pool *pool, struct thin_c *tc)
>
> rcu_read_lock();
> list_for_each_entry_continue_rcu(tc, &pool->active_thins, list) {
> - thin_get(tc);
> + if (!thin_tryget(tc))
> + continue;
> thin_put(old_tc);
> rcu_read_unlock();
> return tc;
> @@ -4128,9 +4134,9 @@ static struct target_type pool_target = {
> * Thin target methods
> *--------------------------------------------------------------
> */
> -static void thin_get(struct thin_c *tc)
> +static bool thin_tryget(struct thin_c *tc)
> {
> - refcount_inc(&tc->refcount);
> + return refcount_inc_not_zero(&tc->refcount);
> }
>
> static void thin_put(struct thin_c *tc)
> @@ -4146,7 +4152,6 @@ static void thin_dtr(struct dm_target *ti)
> spin_lock_irq(&tc->pool->lock);
> list_del_rcu(&tc->list);
> spin_unlock_irq(&tc->pool->lock);
> - synchronize_rcu();
>
> thin_put(tc);
> wait_for_completion(&tc->can_destroy);
> @@ -4158,7 +4163,8 @@ static void thin_dtr(struct dm_target *ti)
> dm_put_device(ti, tc->pool_dev);
> if (tc->origin_dev)
> dm_put_device(ti, tc->origin_dev);
> - kfree(tc);
> + /* Let RCU readers of active_thins finish with tc before freeing it */
> + kfree_rcu(tc, rcu);
>
> mutex_unlock(&dm_thin_pool_table.mutex);
> }
>
> --
> 2.43.0
^ permalink raw reply [flat|nested] 5+ messages in thread
end of thread, other threads:[~2026-11-18 1:13 UTC | newest]
Thread overview: 5+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2026-11-16 5:12 [PATCH 0/2] dm thin: stop waiting for grace periods when activating and removing thins Qiliang Yuan
2026-11-16 5:12 ` [PATCH 1/2] dm thin: drop the grace period from thin_ctr() Qiliang Yuan
2026-11-18 0:51 ` Benjamin Marzinski
2026-11-16 5:12 ` [PATCH 2/2] dm thin: free thins with kfree_rcu() instead of waiting in thin_dtr() Qiliang Yuan
2026-11-18 1:13 ` Benjamin Marzinski
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox
all inboxes | Powered by JetHome®