From: Michal Hocko <[email protected]>
Tetsuo Handa has reported that the system might basically livelock in OOM
condition without triggering the OOM killer. The issue is caused by
internal dependency of the direct reclaim on vmstat counter updates (via
zone_reclaimable) which are performed from the workqueue context.
If all the current workers get assigned to an allocation request,
though, they will be looping inside the allocator trying to reclaim
memory but zone_reclaimable can see stalled numbers so it will consider
a zone reclaimable even though it has been scanned way too much. WQ
concurrency logic will not consider this situation as a congested workqueue
because it relies that worker would have to sleep in such a situation.
This also means that it doesn't try to spawn new workers or invoke
the rescuer thread if the one is assigned to the queue.
In order to fix this issue we need to do two things. First we have to
let wq concurrency code know that we are in trouble so we have to do
a short sleep. In order to prevent from issues handled by 0e093d99763e
("writeback: do not sleep on the congestion queue if there are no
congested BDIs or if significant congestion is not being encountered in
the current zone") we limit the sleep only to worker threads which are
the ones of the interest anyway.
The second thing to do is to create a dedicated workqueue for vmstat and
mark it WQ_MEM_RECLAIM to note it participates in the reclaim and to
have a spare worker thread for it.
Reported-by: Tetsuo Handa <[email protected]>
Cc: stable # 2.6.36+
Signed-off-by: Michal Hocko <[email protected]>
---
[I am convinced that I have posted this as a separate patch but I cannot
find it in the archive nor in my inbox, so let me try again.]
Hi Andrew,
The original issue reported by Tetsuo [1] has seen multiple attempts for
a fix. The easiest one being [2] which was targeted to the particular
problem. There was a more general concern that looping inside the
allocator without ever sleeping breaks the basic assumption of worker
concurrency logic so the fix should be more general. Another attempt [3]
therefore added a short (1 jiffy) sleep into the page allocator. This
would, however, introduce sleeping for all callers of the page allocator
which is not really needed. This patch tries to be a compromise and
introduce sleeping only where it matters - for kworkers.
Even though we haven't seen bug reports in the past I would suggest
backporting this to the stable trees. The issue is present since we have
stopped useing congestion_wait in the retry loop because WQ concurrency
is older as well as vmstat worqueue based refresh AFAICS.
[1] http://lkml.kernel.org/r/201510130025.EJF21331.FFOQJtVOMLFHSO%40I-love.SAKURA.ne.jp
[2] http://lkml.kernel.org/r/[email protected]
[3] http://lkml.kernel.org/r/[email protected]
mm/backing-dev.c | 19 ++++++++++++++++---
mm/vmstat.c | 6 ++++--
2 files changed, 20 insertions(+), 5 deletions(-)
diff --git a/mm/backing-dev.c b/mm/backing-dev.c
index 8ed2ffd963c5..7340353f8aea 100644
--- a/mm/backing-dev.c
+++ b/mm/backing-dev.c
@@ -957,8 +957,9 @@ EXPORT_SYMBOL(congestion_wait);
* jiffies for either a BDI to exit congestion of the given @sync queue
* or a write to complete.
*
- * In the absence of zone congestion, cond_resched() is called to yield
- * the processor if necessary but otherwise does not sleep.
+ * In the absence of zone congestion, a short sleep or a cond_resched is
+ * performed to yield the processor and to allow other subsystems to make
+ * a forward progress.
*
* The return value is 0 if the sleep is for the full timeout. Otherwise,
* it is the number of jiffies that were still remaining when the function
@@ -978,7 +979,19 @@ long wait_iff_congested(struct zone *zone, int sync, long timeout)
*/
if (atomic_read(&nr_wb_congested[sync]) == 0 ||
!test_bit(ZONE_CONGESTED, &zone->flags)) {
- cond_resched();
+
+ /*
+ * Memory allocation/reclaim might be called from a WQ
+ * context and the current implementation of the WQ
+ * concurrency control doesn't recognize that a particular
+ * WQ is congested if the worker thread is looping without
+ * ever sleeping. Therefore we have to do a short sleep
+ * here rather than calling cond_resched().
+ */
+ if (current->flags & PF_WQ_WORKER)
+ schedule_timeout(1);
+ else
+ cond_resched();
/* In case we scheduled, work out time remaining */
ret = timeout - (jiffies - start);
diff --git a/mm/vmstat.c b/mm/vmstat.c
index 45dcbcb5c594..0975da8e3432 100644
--- a/mm/vmstat.c
+++ b/mm/vmstat.c
@@ -1381,6 +1381,7 @@ static const struct file_operations proc_vmstat_file_operations = {
#endif /* CONFIG_PROC_FS */
#ifdef CONFIG_SMP
+static struct workqueue_struct *vmstat_wq;
static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
int sysctl_stat_interval __read_mostly = HZ;
static cpumask_var_t cpu_stat_off;
@@ -1393,7 +1394,7 @@ static void vmstat_update(struct work_struct *w)
* to occur in the future. Keep on running the
* update worker thread.
*/
- schedule_delayed_work_on(smp_processor_id(),
+ queue_delayed_work_on(smp_processor_id(), vmstat_wq,
this_cpu_ptr(&vmstat_work),
round_jiffies_relative(sysctl_stat_interval));
} else {
@@ -1462,7 +1463,7 @@ static void vmstat_shepherd(struct work_struct *w)
if (need_update(cpu) &&
cpumask_test_and_clear_cpu(cpu, cpu_stat_off))
- schedule_delayed_work_on(cpu,
+ queue_delayed_work_on(cpu, vmstat_wq,
&per_cpu(vmstat_work, cpu), 0);
put_online_cpus();
@@ -1551,6 +1552,7 @@ static int __init setup_vmstat(void)
start_shepherd_timer();
cpu_notifier_register_done();
+ vmstat_wq = alloc_workqueue("vmstat", WQ_FREEZABLE|WQ_MEM_RECLAIM, 0);
#endif
#ifdef CONFIG_PROC_FS
proc_create("buddyinfo", S_IRUGO, NULL, &fragmentation_file_operations);
--
2.6.2
Ping... Are there any concerns about this patch?
On Thu 19-11-15 13:30:53, Michal Hocko wrote:
> From: Michal Hocko <[email protected]>
>
> Tetsuo Handa has reported that the system might basically livelock in OOM
> condition without triggering the OOM killer. The issue is caused by
> internal dependency of the direct reclaim on vmstat counter updates (via
> zone_reclaimable) which are performed from the workqueue context.
> If all the current workers get assigned to an allocation request,
> though, they will be looping inside the allocator trying to reclaim
> memory but zone_reclaimable can see stalled numbers so it will consider
> a zone reclaimable even though it has been scanned way too much. WQ
> concurrency logic will not consider this situation as a congested workqueue
> because it relies that worker would have to sleep in such a situation.
> This also means that it doesn't try to spawn new workers or invoke
> the rescuer thread if the one is assigned to the queue.
>
> In order to fix this issue we need to do two things. First we have to
> let wq concurrency code know that we are in trouble so we have to do
> a short sleep. In order to prevent from issues handled by 0e093d99763e
> ("writeback: do not sleep on the congestion queue if there are no
> congested BDIs or if significant congestion is not being encountered in
> the current zone") we limit the sleep only to worker threads which are
> the ones of the interest anyway.
>
> The second thing to do is to create a dedicated workqueue for vmstat and
> mark it WQ_MEM_RECLAIM to note it participates in the reclaim and to
> have a spare worker thread for it.
>
> Reported-by: Tetsuo Handa <[email protected]>
> Cc: stable # 2.6.36+
> Signed-off-by: Michal Hocko <[email protected]>
> ---
> [I am convinced that I have posted this as a separate patch but I cannot
> find it in the archive nor in my inbox, so let me try again.]
>
> Hi Andrew,
>
> The original issue reported by Tetsuo [1] has seen multiple attempts for
> a fix. The easiest one being [2] which was targeted to the particular
> problem. There was a more general concern that looping inside the
> allocator without ever sleeping breaks the basic assumption of worker
> concurrency logic so the fix should be more general. Another attempt [3]
> therefore added a short (1 jiffy) sleep into the page allocator. This
> would, however, introduce sleeping for all callers of the page allocator
> which is not really needed. This patch tries to be a compromise and
> introduce sleeping only where it matters - for kworkers.
>
> Even though we haven't seen bug reports in the past I would suggest
> backporting this to the stable trees. The issue is present since we have
> stopped useing congestion_wait in the retry loop because WQ concurrency
> is older as well as vmstat worqueue based refresh AFAICS.
>
> [1] http://lkml.kernel.org/r/201510130025.EJF21331.FFOQJtVOMLFHSO%40I-love.SAKURA.ne.jp
> [2] http://lkml.kernel.org/r/[email protected]
> [3] http://lkml.kernel.org/r/[email protected]
>
> mm/backing-dev.c | 19 ++++++++++++++++---
> mm/vmstat.c | 6 ++++--
> 2 files changed, 20 insertions(+), 5 deletions(-)
>
> diff --git a/mm/backing-dev.c b/mm/backing-dev.c
> index 8ed2ffd963c5..7340353f8aea 100644
> --- a/mm/backing-dev.c
> +++ b/mm/backing-dev.c
> @@ -957,8 +957,9 @@ EXPORT_SYMBOL(congestion_wait);
> * jiffies for either a BDI to exit congestion of the given @sync queue
> * or a write to complete.
> *
> - * In the absence of zone congestion, cond_resched() is called to yield
> - * the processor if necessary but otherwise does not sleep.
> + * In the absence of zone congestion, a short sleep or a cond_resched is
> + * performed to yield the processor and to allow other subsystems to make
> + * a forward progress.
> *
> * The return value is 0 if the sleep is for the full timeout. Otherwise,
> * it is the number of jiffies that were still remaining when the function
> @@ -978,7 +979,19 @@ long wait_iff_congested(struct zone *zone, int sync, long timeout)
> */
> if (atomic_read(&nr_wb_congested[sync]) == 0 ||
> !test_bit(ZONE_CONGESTED, &zone->flags)) {
> - cond_resched();
> +
> + /*
> + * Memory allocation/reclaim might be called from a WQ
> + * context and the current implementation of the WQ
> + * concurrency control doesn't recognize that a particular
> + * WQ is congested if the worker thread is looping without
> + * ever sleeping. Therefore we have to do a short sleep
> + * here rather than calling cond_resched().
> + */
> + if (current->flags & PF_WQ_WORKER)
> + schedule_timeout(1);
> + else
> + cond_resched();
>
> /* In case we scheduled, work out time remaining */
> ret = timeout - (jiffies - start);
> diff --git a/mm/vmstat.c b/mm/vmstat.c
> index 45dcbcb5c594..0975da8e3432 100644
> --- a/mm/vmstat.c
> +++ b/mm/vmstat.c
> @@ -1381,6 +1381,7 @@ static const struct file_operations proc_vmstat_file_operations = {
> #endif /* CONFIG_PROC_FS */
>
> #ifdef CONFIG_SMP
> +static struct workqueue_struct *vmstat_wq;
> static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
> int sysctl_stat_interval __read_mostly = HZ;
> static cpumask_var_t cpu_stat_off;
> @@ -1393,7 +1394,7 @@ static void vmstat_update(struct work_struct *w)
> * to occur in the future. Keep on running the
> * update worker thread.
> */
> - schedule_delayed_work_on(smp_processor_id(),
> + queue_delayed_work_on(smp_processor_id(), vmstat_wq,
> this_cpu_ptr(&vmstat_work),
> round_jiffies_relative(sysctl_stat_interval));
> } else {
> @@ -1462,7 +1463,7 @@ static void vmstat_shepherd(struct work_struct *w)
> if (need_update(cpu) &&
> cpumask_test_and_clear_cpu(cpu, cpu_stat_off))
>
> - schedule_delayed_work_on(cpu,
> + queue_delayed_work_on(cpu, vmstat_wq,
> &per_cpu(vmstat_work, cpu), 0);
>
> put_online_cpus();
> @@ -1551,6 +1552,7 @@ static int __init setup_vmstat(void)
>
> start_shepherd_timer();
> cpu_notifier_register_done();
> + vmstat_wq = alloc_workqueue("vmstat", WQ_FREEZABLE|WQ_MEM_RECLAIM, 0);
> #endif
> #ifdef CONFIG_PROC_FS
> proc_create("buddyinfo", S_IRUGO, NULL, &fragmentation_file_operations);
> --
> 2.6.2
>
--
Michal Hocko
SUSE Labs
Michal Hocko wrote:
> Ping... Are there any concerns about this patch?
>
I'm OK with this patch.
On Thu, 19 Nov 2015 13:30:53 +0100 Michal Hocko <[email protected]> wrote:
> From: Michal Hocko <[email protected]>
>
> Tetsuo Handa has reported that the system might basically livelock in OOM
> condition without triggering the OOM killer. The issue is caused by
> internal dependency of the direct reclaim on vmstat counter updates (via
> zone_reclaimable) which are performed from the workqueue context.
> If all the current workers get assigned to an allocation request,
> though, they will be looping inside the allocator trying to reclaim
> memory but zone_reclaimable can see stalled numbers so it will consider
> a zone reclaimable even though it has been scanned way too much. WQ
> concurrency logic will not consider this situation as a congested workqueue
> because it relies that worker would have to sleep in such a situation.
> This also means that it doesn't try to spawn new workers or invoke
> the rescuer thread if the one is assigned to the queue.
>
> In order to fix this issue we need to do two things. First we have to
> let wq concurrency code know that we are in trouble so we have to do
> a short sleep. In order to prevent from issues handled by 0e093d99763e
> ("writeback: do not sleep on the congestion queue if there are no
> congested BDIs or if significant congestion is not being encountered in
> the current zone") we limit the sleep only to worker threads which are
> the ones of the interest anyway.
>
> The second thing to do is to create a dedicated workqueue for vmstat and
> mark it WQ_MEM_RECLAIM to note it participates in the reclaim and to
> have a spare worker thread for it.
This vmstat update thing is being a problem. Please see Joonsoo's
"mm/vmstat: retrieve more accurate vmstat value".
Joonsoo, might this patch help with that issue?
>
> The original issue reported by Tetsuo [1] has seen multiple attempts for
> a fix. The easiest one being [2] which was targeted to the particular
> problem. There was a more general concern that looping inside the
> allocator without ever sleeping breaks the basic assumption of worker
> concurrency logic so the fix should be more general. Another attempt [3]
> therefore added a short (1 jiffy) sleep into the page allocator. This
> would, however, introduce sleeping for all callers of the page allocator
> which is not really needed. This patch tries to be a compromise and
> introduce sleeping only where it matters - for kworkers.
>
> Even though we haven't seen bug reports in the past I would suggest
> backporting this to the stable trees. The issue is present since we have
> stopped useing congestion_wait in the retry loop because WQ concurrency
> is older as well as vmstat worqueue based refresh AFAICS.
hm, I'm reluctant. If the patch fixes something that real people are
really hurting from then yes. But I suspect this is just one fly-swat
amongst many.
On Tue, Nov 24, 2015 at 03:44:48PM -0800, Andrew Morton wrote:
> On Thu, 19 Nov 2015 13:30:53 +0100 Michal Hocko <[email protected]> wrote:
>
> > From: Michal Hocko <[email protected]>
> >
> > Tetsuo Handa has reported that the system might basically livelock in OOM
> > condition without triggering the OOM killer. The issue is caused by
> > internal dependency of the direct reclaim on vmstat counter updates (via
> > zone_reclaimable) which are performed from the workqueue context.
> > If all the current workers get assigned to an allocation request,
> > though, they will be looping inside the allocator trying to reclaim
> > memory but zone_reclaimable can see stalled numbers so it will consider
> > a zone reclaimable even though it has been scanned way too much. WQ
> > concurrency logic will not consider this situation as a congested workqueue
> > because it relies that worker would have to sleep in such a situation.
> > This also means that it doesn't try to spawn new workers or invoke
> > the rescuer thread if the one is assigned to the queue.
> >
> > In order to fix this issue we need to do two things. First we have to
> > let wq concurrency code know that we are in trouble so we have to do
> > a short sleep. In order to prevent from issues handled by 0e093d99763e
> > ("writeback: do not sleep on the congestion queue if there are no
> > congested BDIs or if significant congestion is not being encountered in
> > the current zone") we limit the sleep only to worker threads which are
> > the ones of the interest anyway.
> >
> > The second thing to do is to create a dedicated workqueue for vmstat and
> > mark it WQ_MEM_RECLAIM to note it participates in the reclaim and to
> > have a spare worker thread for it.
>
> This vmstat update thing is being a problem. Please see Joonsoo's
> "mm/vmstat: retrieve more accurate vmstat value".
>
> Joonsoo, might this patch help with that issue?
That issue cannot be solved by this patch. This patch solves blocking
vmstat updator problem but that issue is caused by long update delay
(not blocking). In there, update happens every 1 sec as usuall.
Thanks.
On Tue 24-11-15 15:44:48, Andrew Morton wrote:
[...]
> > Even though we haven't seen bug reports in the past I would suggest
> > backporting this to the stable trees. The issue is present since we have
> > stopped useing congestion_wait in the retry loop because WQ concurrency
> > is older as well as vmstat worqueue based refresh AFAICS.
>
> hm, I'm reluctant. If the patch fixes something that real people are
> really hurting from then yes. But I suspect this is just one fly-swat
> amongst many.
Arkadiusz was seeing reclaim issues [1] on 4.1 kernel. I didn't have
time to look deeper in that report but vmstat counters seemed terribly
outdated and the issue went away when this patch was used. The thing is
that there were others in the bundle so it is not 100% clear whether the
patch alone helped or it was just a part of the puzzle.
Anyway I think that the issue is not solely theoretical. WQ_MEM_RECLAIM
is simply not working if the allocation path doesn't sleep currently and
my understanding of what Tejun claims [2] is that that reimplementing WQ
concurrency would be too intrusive and lacks sufficient justification
because other kernel paths do sleep. This patch tries to reduce the
sleep only to worker threads which should not cause any problems to
regular tasks.
I am open to any other suggestions. I do not like artificial sleep as
well but this sounds like the most practical way to go now.
[1] http://lkml.kernel.org/r/[email protected]
[2] http://lkml.kernel.org/r/[email protected]
--
Michal Hocko
SUSE Labs
Michal Hocko wrote:
> Anyway I think that the issue is not solely theoretical. WQ_MEM_RECLAIM
> is simply not working if the allocation path doesn't sleep currently and
> my understanding of what Tejun claims [2] is that that reimplementing WQ
> concurrency would be too intrusive and lacks sufficient justification
> because other kernel paths do sleep. This patch tries to reduce the
> sleep only to worker threads which should not cause any problems to
> regular tasks.
I received many unexplained hangup/reboot reports from customers when I was
working at support center. But we can't answer whether real people ever hit
this problem because we have no watchdog for memory allocation stalls.
I want one like http://lkml.kernel.org/r/[email protected]
as I wrote off-list ( "mm,oom: The reason why I continue proposing timeout
based approach." ). It will help with judging when we tackle TIF_MEMDIE
livelock problem.
What I can say is that RHEL6 (a 2.6.32-based distro) backported the
wait_iff_congested() changes and therefore people might really hit
this problem.
On Wed, 25 Nov 2015 12:07:05 +0100 Michal Hocko <[email protected]> wrote:
> On Tue 24-11-15 15:44:48, Andrew Morton wrote:
> [...]
> > > Even though we haven't seen bug reports in the past I would suggest
> > > backporting this to the stable trees. The issue is present since we have
> > > stopped useing congestion_wait in the retry loop because WQ concurrency
> > > is older as well as vmstat worqueue based refresh AFAICS.
> >
> > hm, I'm reluctant. If the patch fixes something that real people are
> > really hurting from then yes. But I suspect this is just one fly-swat
> > amongst many.
>
> Arkadiusz was seeing reclaim issues [1] on 4.1 kernel. I didn't have
> time to look deeper in that report but vmstat counters seemed terribly
> outdated and the issue went away when this patch was used. The thing is
> that there were others in the bundle so it is not 100% clear whether the
> patch alone helped or it was just a part of the puzzle.
>
> Anyway I think that the issue is not solely theoretical. WQ_MEM_RECLAIM
> is simply not working if the allocation path doesn't sleep currently and
> my understanding of what Tejun claims [2] is that that reimplementing WQ
> concurrency would be too intrusive and lacks sufficient justification
> because other kernel paths do sleep. This patch tries to reduce the
> sleep only to worker threads which should not cause any problems to
> regular tasks.
>
> I am open to any other suggestions. I do not like artificial sleep as
> well but this sounds like the most practical way to go now.
>
> [1] http://lkml.kernel.org/r/[email protected]
> [2] http://lkml.kernel.org/r/[email protected]
hmpf, OK, I stuck a cc:stable in there. It looks like the current
changelog is sufficient to explain to Greg (and others) why we think
backporting is needed.