Commit 5a836bf6b09f ("mm: slub: move flush_cpu_slab() invocations
__free_slab() invocations out of IRQ context") moved all flush_cpu_slab()
invocations to the global workqueue to avoid a problem related
with deactivate_slab()/__free_slab() being called from an IRQ context
on PREEMPT_RT kernels.
When the flush_all_cpu_locked() function is called from a task context
it may happen that a workqueue with WQ_MEM_RECLAIM bit set ends up
flushing the global workqueue, this will cause a dependency issue.
workqueue: WQ_MEM_RECLAIM nvme-delete-wq:nvme_delete_ctrl_work [nvme_core]
is flushing !WQ_MEM_RECLAIM events:flush_cpu_slab
WARNING: CPU: 37 PID: 410 at kernel/workqueue.c:2637
check_flush_dependency+0x10a/0x120
Workqueue: nvme-delete-wq nvme_delete_ctrl_work [nvme_core]
RIP: 0010:check_flush_dependency+0x10a/0x120[ 453.262125] Call Trace:
__flush_work.isra.0+0xbf/0x220
? __queue_work+0x1dc/0x420
flush_all_cpus_locked+0xfb/0x120
__kmem_cache_shutdown+0x2b/0x320
kmem_cache_destroy+0x49/0x100
bioset_exit+0x143/0x190
blk_release_queue+0xb9/0x100
kobject_cleanup+0x37/0x130
nvme_fc_ctrl_free+0xc6/0x150 [nvme_fc]
nvme_free_ctrl+0x1ac/0x2b0 [nvme_core]
Fix this bug by creating a workqueue for the flush operation with
the WQ_MEM_RECLAIM bit set.
v2: Create a workqueue with WQ_MEM_RECLAIM
instead of trying to revert the changes.
v3: replace create_workqueue() with alloc_workqueue() and BUG_ON() with
WARN_ON()
Signed-off-by: Maurizio Lombardi <[email protected]>
---
mm/slub.c | 9 ++++++++-
1 file changed, 8 insertions(+), 1 deletion(-)
diff --git a/mm/slub.c b/mm/slub.c
index 862dbd9af4f5..016da09608fb 100644
--- a/mm/slub.c
+++ b/mm/slub.c
@@ -310,6 +310,11 @@ static inline void stat(const struct kmem_cache *s, enum stat_item si)
*/
static nodemask_t slab_nodes;
+/*
+ * Workqueue used for flush_cpu_slab().
+ */
+static struct workqueue_struct *flushwq;
+
/********************************************************************
* Core slab cache functions
*******************************************************************/
@@ -2730,7 +2735,7 @@ static void flush_all_cpus_locked(struct kmem_cache *s)
INIT_WORK(&sfw->work, flush_cpu_slab);
sfw->skip = false;
sfw->s = s;
- schedule_work_on(cpu, &sfw->work);
+ queue_work_on(cpu, flushwq, &sfw->work);
}
for_each_online_cpu(cpu) {
@@ -4858,6 +4863,8 @@ void __init kmem_cache_init(void)
void __init kmem_cache_init_late(void)
{
+ flushwq = alloc_workqueue("slub_flushwq", WQ_MEM_RECLAIM, 0);
+ WARN_ON(!flushwq);
}
struct kmem_cache *
--
2.31.1
On Mon, Sep 19, 2022 at 06:39:29PM +0200, Maurizio Lombardi wrote:
> Commit 5a836bf6b09f ("mm: slub: move flush_cpu_slab() invocations
> __free_slab() invocations out of IRQ context") moved all flush_cpu_slab()
> invocations to the global workqueue to avoid a problem related
> with deactivate_slab()/__free_slab() being called from an IRQ context
> on PREEMPT_RT kernels.
>
> When the flush_all_cpu_locked() function is called from a task context
> it may happen that a workqueue with WQ_MEM_RECLAIM bit set ends up
> flushing the global workqueue, this will cause a dependency issue.
>
> workqueue: WQ_MEM_RECLAIM nvme-delete-wq:nvme_delete_ctrl_work [nvme_core]
> is flushing !WQ_MEM_RECLAIM events:flush_cpu_slab
> WARNING: CPU: 37 PID: 410 at kernel/workqueue.c:2637
> check_flush_dependency+0x10a/0x120
> Workqueue: nvme-delete-wq nvme_delete_ctrl_work [nvme_core]
> RIP: 0010:check_flush_dependency+0x10a/0x120[ 453.262125] Call Trace:
> __flush_work.isra.0+0xbf/0x220
> ? __queue_work+0x1dc/0x420
> flush_all_cpus_locked+0xfb/0x120
> __kmem_cache_shutdown+0x2b/0x320
> kmem_cache_destroy+0x49/0x100
> bioset_exit+0x143/0x190
> blk_release_queue+0xb9/0x100
> kobject_cleanup+0x37/0x130
> nvme_fc_ctrl_free+0xc6/0x150 [nvme_fc]
> nvme_free_ctrl+0x1ac/0x2b0 [nvme_core]
>
> Fix this bug by creating a workqueue for the flush operation with
> the WQ_MEM_RECLAIM bit set.
>
> v2: Create a workqueue with WQ_MEM_RECLAIM
> instead of trying to revert the changes.
>
> v3: replace create_workqueue() with alloc_workqueue() and BUG_ON() with
> WARN_ON()
>
> Signed-off-by: Maurizio Lombardi <[email protected]>
> ---
> mm/slub.c | 9 ++++++++-
> 1 file changed, 8 insertions(+), 1 deletion(-)
>
> diff --git a/mm/slub.c b/mm/slub.c
> index 862dbd9af4f5..016da09608fb 100644
> --- a/mm/slub.c
> +++ b/mm/slub.c
> @@ -310,6 +310,11 @@ static inline void stat(const struct kmem_cache *s, enum stat_item si)
> */
> static nodemask_t slab_nodes;
>
> +/*
> + * Workqueue used for flush_cpu_slab().
> + */
> +static struct workqueue_struct *flushwq;
> +
> /********************************************************************
> * Core slab cache functions
> *******************************************************************/
> @@ -2730,7 +2735,7 @@ static void flush_all_cpus_locked(struct kmem_cache *s)
> INIT_WORK(&sfw->work, flush_cpu_slab);
> sfw->skip = false;
> sfw->s = s;
> - schedule_work_on(cpu, &sfw->work);
> + queue_work_on(cpu, flushwq, &sfw->work);
Hi. what happens here if flushwq failed?
I think avoiding BUG_ON() makes sense,
but shouldn't we have fallback method?
> }
>
> for_each_online_cpu(cpu) {
> @@ -4858,6 +4863,8 @@ void __init kmem_cache_init(void)
>
> void __init kmem_cache_init_late(void)
> {
> + flushwq = alloc_workqueue("slub_flushwq", WQ_MEM_RECLAIM, 0);
> + WARN_ON(!flushwq);
> }
>
> struct kmem_cache *
> --
> 2.31.1
>
>
--
Thanks,
Hyeonggon
On 2022-09-20 16:46:41 [+0900], Hyeonggon Yoo wrote:
> > @@ -2730,7 +2735,7 @@ static void flush_all_cpus_locked(struct kmem_cache *s)
> > INIT_WORK(&sfw->work, flush_cpu_slab);
> > sfw->skip = false;
> > sfw->s = s;
> > - schedule_work_on(cpu, &sfw->work);
> > + queue_work_on(cpu, flushwq, &sfw->work);
>
> Hi. what happens here if flushwq failed?
>
> I think avoiding BUG_ON() makes sense,
> but shouldn't we have fallback method?
You get an output to act on and fix. The point is that it shouldn't have
happen in the first place. With the bug_on() that early, chances are
that you never see anything but a blank screen. So with the warn_on you
get probably to see the warn_on before you get here.
Sebastian
On Mon, Sep 19, 2022 at 06:39:29PM +0200, Maurizio Lombardi wrote:
> Commit 5a836bf6b09f ("mm: slub: move flush_cpu_slab() invocations
> __free_slab() invocations out of IRQ context") moved all flush_cpu_slab()
> invocations to the global workqueue to avoid a problem related
> with deactivate_slab()/__free_slab() being called from an IRQ context
> on PREEMPT_RT kernels.
>
> When the flush_all_cpu_locked() function is called from a task context
> it may happen that a workqueue with WQ_MEM_RECLAIM bit set ends up
> flushing the global workqueue, this will cause a dependency issue.
>
> workqueue: WQ_MEM_RECLAIM nvme-delete-wq:nvme_delete_ctrl_work [nvme_core]
> is flushing !WQ_MEM_RECLAIM events:flush_cpu_slab
> WARNING: CPU: 37 PID: 410 at kernel/workqueue.c:2637
> check_flush_dependency+0x10a/0x120
> Workqueue: nvme-delete-wq nvme_delete_ctrl_work [nvme_core]
> RIP: 0010:check_flush_dependency+0x10a/0x120[ 453.262125] Call Trace:
> __flush_work.isra.0+0xbf/0x220
> ? __queue_work+0x1dc/0x420
> flush_all_cpus_locked+0xfb/0x120
> __kmem_cache_shutdown+0x2b/0x320
> kmem_cache_destroy+0x49/0x100
> bioset_exit+0x143/0x190
> blk_release_queue+0xb9/0x100
> kobject_cleanup+0x37/0x130
> nvme_fc_ctrl_free+0xc6/0x150 [nvme_fc]
> nvme_free_ctrl+0x1ac/0x2b0 [nvme_core]
>
> Fix this bug by creating a workqueue for the flush operation with
> the WQ_MEM_RECLAIM bit set.
>
> v2: Create a workqueue with WQ_MEM_RECLAIM
> instead of trying to revert the changes.
>
> v3: replace create_workqueue() with alloc_workqueue() and BUG_ON() with
> WARN_ON()
>
> Signed-off-by: Maurizio Lombardi <[email protected]>
> ---
> mm/slub.c | 9 ++++++++-
> 1 file changed, 8 insertions(+), 1 deletion(-)
>
> diff --git a/mm/slub.c b/mm/slub.c
> index 862dbd9af4f5..016da09608fb 100644
> --- a/mm/slub.c
> +++ b/mm/slub.c
> @@ -310,6 +310,11 @@ static inline void stat(const struct kmem_cache *s, enum stat_item si)
> */
> static nodemask_t slab_nodes;
>
> +/*
> + * Workqueue used for flush_cpu_slab().
> + */
> +static struct workqueue_struct *flushwq;
> +
> /********************************************************************
> * Core slab cache functions
> *******************************************************************/
> @@ -2730,7 +2735,7 @@ static void flush_all_cpus_locked(struct kmem_cache *s)
> INIT_WORK(&sfw->work, flush_cpu_slab);
> sfw->skip = false;
> sfw->s = s;
> - schedule_work_on(cpu, &sfw->work);
> + queue_work_on(cpu, flushwq, &sfw->work);
> }
>
> for_each_online_cpu(cpu) {
> @@ -4858,6 +4863,8 @@ void __init kmem_cache_init(void)
>
> void __init kmem_cache_init_late(void)
> {
> + flushwq = alloc_workqueue("slub_flushwq", WQ_MEM_RECLAIM, 0);
> + WARN_ON(!flushwq);
> }
>
> struct kmem_cache *
> --
> 2.31.1
>
Looks good to me.
Reviewed-by: Hyeonggon Yoo <[email protected]>
--
Thanks,
Hyeonggon
On Tue, Sep 20, 2022 at 09:56:59AM +0200, Sebastian Andrzej Siewior wrote:
> On 2022-09-20 16:46:41 [+0900], Hyeonggon Yoo wrote:
> > > @@ -2730,7 +2735,7 @@ static void flush_all_cpus_locked(struct kmem_cache *s)
> > > INIT_WORK(&sfw->work, flush_cpu_slab);
> > > sfw->skip = false;
> > > sfw->s = s;
> > > - schedule_work_on(cpu, &sfw->work);
> > > + queue_work_on(cpu, flushwq, &sfw->work);
> >
> > Hi. what happens here if flushwq failed?
> >
> > I think avoiding BUG_ON() makes sense,
> > but shouldn't we have fallback method?
>
> You get an output to act on and fix. The point is that it shouldn't have
> happen in the first place. With the bug_on() that early, chances are
> that you never see anything but a blank screen. So with the warn_on you
> get probably to see the warn_on before you get here.
>
> Sebastian
Thank you for kind explanation.
Makes sense!
--
Thanks,
Hyeonggon