The original report is actually one real deadlock:
[ 106.132865] Possible interrupt unsafe locking scenario:
[ 106.132865]
[ 106.133659] CPU0 CPU1
[ 106.134194] ---- ----
[ 106.134733] lock(&(&sb->map[i].swap_lock)->rlock);
[ 106.135318] local_irq_disable();
[ 106.136014] lock(&sbq->ws[i].wait);
[ 106.136747] lock(&(&hctx->dispatch_wait_lock)->rlock);
[ 106.137742] <Interrupt>
[ 106.138110] lock(&sbq->ws[i].wait);
Because we may call blk_mq_get_driver_tag() directly from
blk_mq_dispatch_rq_list() without holding any lock, then HARDIRQ may come
and the above DEADLOCK is triggered.
ab53dcfb3e7b ("sbitmap: Protect swap_lock from hardirq") tries to fix
this issue by using 'spin_lock_bh', which isn't enough because we complete
request from hardirq context direclty in case of multiqueue.
Cc: Clark Williams <[email protected]>
Fixes: ab53dcfb3e7b ("sbitmap: Protect swap_lock from hardirq")
Cc: Jens Axboe <[email protected]>
Cc: Ming Lei <[email protected]>
Cc: Guenter Roeck <[email protected]>
Cc: Steven Rostedt (VMware) <[email protected]>
Cc: Linus Torvalds <[email protected]>
Cc: [email protected]
Signed-off-by: Ming Lei <[email protected]>
---
lib/sbitmap.c | 5 +++--
1 file changed, 3 insertions(+), 2 deletions(-)
diff --git a/lib/sbitmap.c b/lib/sbitmap.c
index 864354000e04..5b382c1244ed 100644
--- a/lib/sbitmap.c
+++ b/lib/sbitmap.c
@@ -27,8 +27,9 @@ static inline bool sbitmap_deferred_clear(struct sbitmap *sb, int index)
{
unsigned long mask, val;
bool ret = false;
+ unsigned long flags;
- spin_lock_bh(&sb->map[index].swap_lock);
+ spin_lock_irqsave(&sb->map[index].swap_lock, flags);
if (!sb->map[index].cleared)
goto out_unlock;
@@ -49,7 +50,7 @@ static inline bool sbitmap_deferred_clear(struct sbitmap *sb, int index)
ret = true;
out_unlock:
- spin_unlock_bh(&sb->map[index].swap_lock);
+ spin_unlock_irqrestore(&sb->map[index].swap_lock, flags);
return ret;
}
--
2.14.4
On 1/14/19 8:59 PM, Ming Lei wrote:
> The original report is actually one real deadlock:
>
> [ 106.132865] Possible interrupt unsafe locking scenario:
> [ 106.132865]
> [ 106.133659] CPU0 CPU1
> [ 106.134194] ---- ----
> [ 106.134733] lock(&(&sb->map[i].swap_lock)->rlock);
> [ 106.135318] local_irq_disable();
> [ 106.136014] lock(&sbq->ws[i].wait);
> [ 106.136747] lock(&(&hctx->dispatch_wait_lock)->rlock);
> [ 106.137742] <Interrupt>
> [ 106.138110] lock(&sbq->ws[i].wait);
>
> Because we may call blk_mq_get_driver_tag() directly from
> blk_mq_dispatch_rq_list() without holding any lock, then HARDIRQ may come
> and the above DEADLOCK is triggered.
>
> ab53dcfb3e7b ("sbitmap: Protect swap_lock from hardirq") tries to fix
> this issue by using 'spin_lock_bh', which isn't enough because we complete
> request from hardirq context direclty in case of multiqueue.
Thanks Ming, I'll queue this up for shipping this week.
--
Jens Axboe
On Tue, Jan 15, 2019 at 4:28 PM Jens Axboe <[email protected]> wrote:
>
> Thanks Ming, I'll queue this up for shipping this week.
Oops. I _just_ applied it to my tree as a follow-up to Steven's
softirq version. I just hadn't had time to build test and push out
yet.
Linus
On 1/14/19 9:31 PM, Linus Torvalds wrote:
> On Tue, Jan 15, 2019 at 4:28 PM Jens Axboe <[email protected]> wrote:
>>
>> Thanks Ming, I'll queue this up for shipping this week.
>
> Oops. I _just_ applied it to my tree as a follow-up to Steven's
> softirq version. I just hadn't had time to build test and push out
> yet.
No big deal, fwiw, this is what I queued up:
http://git.kernel.dk/cgit/linux-block/commit/?h=for-linus&id=8218a55b6b911d396565da4ed5ca8b18bf0d38fb
which has a spelling error fixed, and the indentation a bit nicer
for the locking scenario. But I can just drop it.
--
Jens Axboe