[PATCH] fs/writeback: Attach inode's wb to root if needed

[PATCH] fs/writeback: Attach inode's wb to root if needed

[zhangliguang]

There might have tons of files queued in the writeback, awaiting for
writing back. Unfortunately, the writeback's cgroup has been dead. In
this case, we reassociate the inode with another writeback cgroup, but
we possibly can't because the writeback associated with the dead cgroup
is the only valid one. In this case, the new writeback is allocated,
initialized and associated with the inode. It causes unnecessary high
system load and latency.

This fixes the issue by enforce moving the inode to root cgroup when the
previous binding cgroup becomes dead. With it, no more unnecessary
writebacks are created, populated and the system load decreased by about
6x in the online service we encounted:
    Without the patch: about 30% system load
    With the patch:    about  5% system load

with the patch observes significant perf graph change:

========================================================================
We record the trace with 'perf record cycles:k -g -a'.

The trace without the patch:

+  44.68%	[kernel]  [k] native_queued_spin_lock_slowpath
+   3.38%	[kernel]  [k] memset_erms
+   3.04%	[kernel]  [k] pcpu_alloc_area
+   1.46%	[kernel]  [k] __memmove
+   1.37%	[kernel]  [k] pcpu_alloc

detail information abount native_queued_spin_lock_slowpath:
44.68%       [kernel]  [k] native_queued_spin_lock_slowpath
 - native_queued_spin_lock_slowpath
    - _raw_spin_lock_irqsave
       - 68.80% pcpu_alloc
          - __alloc_percpu_gfp
             - 85.01% __percpu_counter_init
                - 66.75% wb_init
                     wb_get_create
                     inode_switch_wbs
                     wbc_attach_and_unlock_inode
                     __filemap_fdatawrite_range
                   + filemap_write_and_wait_range
                + 33.25% fprop_local_init_percpu
             + 14.99% percpu_ref_init
       + 30.77% free_percpu

system load (by top)
%Cpu(s): 31.9% sy

With the patch:

+   4.45%       [kernel]  [k] native_queued_spin_lock_slowpath
+   3.32%       [kernel]  [k] put_compound_page
+   2.20%       [kernel]  [k] gup_pte_range
+   1.91%       [kernel]  [k] kstat_irqs

system load (by top)
%Cpu(s): 5.4% sy

Signed-off-by: zhangliguang <zhangliguang@linux.alibaba.com>
---
 fs/fs-writeback.c | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/fs/fs-writeback.c b/fs/fs-writeback.c
index 36855c1..e7e19d8 100644
--- a/fs/fs-writeback.c
+++ b/fs/fs-writeback.c
@@ -696,6 +696,13 @@ void wbc_detach_inode(struct writeback_control *wbc)
 	inode->i_wb_frn_avg_time = min(avg_time, (unsigned long)U16_MAX);
 	inode->i_wb_frn_history = history;
 
+	/*
+	 * The wb is switched to the root memcg unconditionally. We expect
+	 * the correct wb (best candidate) is picked up in next round.
+	 */
+	if (wb == inode->i_wb && wb_dying(wb) && !(inode->i_state & I_DIRTY_ALL))
+		inode_switch_wbs(inode, root_mem_cgroup->css.id);
+
 	wb_put(wbc->wb);
 	wbc->wb = NULL;
 }
-- 
1.8.3.1

Re: [PATCH] fs/writeback: Attach inode's wb to root if needed

[Tejun Heo]

Hello,

On Thu, May 09, 2019 at 04:03:53PM +0800, zhangliguang wrote:
> There might have tons of files queued in the writeback, awaiting for
> writing back. Unfortunately, the writeback's cgroup has been dead. In
> this case, we reassociate the inode with another writeback cgroup, but
> we possibly can't because the writeback associated with the dead cgroup
> is the only valid one. In this case, the new writeback is allocated,
> initialized and associated with the inode. It causes unnecessary high
> system load and latency.
> 
> This fixes the issue by enforce moving the inode to root cgroup when the
> previous binding cgroup becomes dead. With it, no more unnecessary
> writebacks are created, populated and the system load decreased by about
> 6x in the online service we encounted:
>     Without the patch: about 30% system load
>     With the patch:    about  5% system load

Can you please describe the scenario with more details?  I'm having a
bit of hard time understanding the amount of cpu cycles being
consumed.

Thanks.

-- 
tejun

Re: [PATCH] fs/writeback: Attach inode's wb to root if needed

[乱石]

Hi Tejun,

在 2019/5/10 0:48, Tejun Heo 写道:
> Hi Tejun,
>
> On Thu, May 09, 2019 at 04:03:53PM +0800, zhangliguang wrote:
>> There might have tons of files queued in the writeback, awaiting for
>> writing back. Unfortunately, the writeback's cgroup has been dead. In
>> this case, we reassociate the inode with another writeback cgroup, but
>> we possibly can't because the writeback associated with the dead cgroup
>> is the only valid one. In this case, the new writeback is allocated,
>> initialized and associated with the inode. It causes unnecessary high
>> system load and latency.
>>
>> This fixes the issue by enforce moving the inode to root cgroup when the
>> previous binding cgroup becomes dead. With it, no more unnecessary
>> writebacks are created, populated and the system load decreased by about
>> 6x in the online service we encounted:
>>      Without the patch: about 30% system load
>>      With the patch:    about  5% system load
> Can you please describe the scenario with more details?  I'm having a
> bit of hard time understanding the amount of cpu cycles being
> consumed.
>
> Thanks.

Our search line reported a problem, when containerA was removed,
containerB and containerC's system load were up to 30%.

We record the trace with 'perf record cycles:k -g -a', found that wb_init
was the hotspot function.

Function call:

generic_file_direct_write
    filemap_write_and_wait_range
       __filemap_fdatawrite_range
          wbc_attach_fdatawrite_inode
             inode_attach_wb
                __inode_attach_wb
                   wb_get_create
             wbc_attach_and_unlock_inode
                if (unlikely(wb_dying(wbc->wb)))
                   inode_switch_wbs
                      wb_get_create
                         ; Search bdi->cgwb_tree from memcg_css->id
                         ; OR cgwb_create
                            kmalloc
                            wb_init       // hot spot
                            ; Insert to bdi->cgwb_tree, mmecg_css->id as key

We discussed it through, base on the analysis:  When we running into the
issue, there is cgroups are being deleted. The inodes (files) that were
associated with these cgroups have to switch into another newly created
writeback. We think there are huge amount of inodes in the writeback list
that time. So we don't think there is anything abnormal. However, one
thing we possibly can do: enforce these inodes to BDI embedded wirteback
and we needn't create huge amount of writebacks in that case, to avoid
the high system load phenomenon. We expect correct wb (best candidate) is
picked up in next round.

Thanks,
Liguang

>

Re: [PATCH] fs/writeback: Attach inode's wb to root if needed

[Dennis Zhou]

Hi Liguang,

On Fri, May 10, 2019 at 09:54:27AM +0800, 乱石 wrote:
> Hi Tejun,
> 
> 在 2019/5/10 0:48, Tejun Heo 写道:
> > Hi Tejun,
> > 
> > On Thu, May 09, 2019 at 04:03:53PM +0800, zhangliguang wrote:
> > > There might have tons of files queued in the writeback, awaiting for
> > > writing back. Unfortunately, the writeback's cgroup has been dead. In
> > > this case, we reassociate the inode with another writeback cgroup, but
> > > we possibly can't because the writeback associated with the dead cgroup
> > > is the only valid one. In this case, the new writeback is allocated,
> > > initialized and associated with the inode. It causes unnecessary high
> > > system load and latency.
> > > 
> > > This fixes the issue by enforce moving the inode to root cgroup when the
> > > previous binding cgroup becomes dead. With it, no more unnecessary
> > > writebacks are created, populated and the system load decreased by about
> > > 6x in the online service we encounted:
> > >      Without the patch: about 30% system load
> > >      With the patch:    about  5% system load
> > Can you please describe the scenario with more details?  I'm having a
> > bit of hard time understanding the amount of cpu cycles being
> > consumed.
> > 
> > Thanks.
> 
> Our search line reported a problem, when containerA was removed,
> containerB and containerC's system load were up to 30%.
> 
> We record the trace with 'perf record cycles:k -g -a', found that wb_init
> was the hotspot function.
> 
> Function call:
> 
> generic_file_direct_write
>    filemap_write_and_wait_range
>       __filemap_fdatawrite_range
>          wbc_attach_fdatawrite_inode
>             inode_attach_wb
>                __inode_attach_wb
>                   wb_get_create
>             wbc_attach_and_unlock_inode
>                if (unlikely(wb_dying(wbc->wb)))
>                   inode_switch_wbs
>                      wb_get_create
>                         ; Search bdi->cgwb_tree from memcg_css->id
>                         ; OR cgwb_create
>                            kmalloc
>                            wb_init       // hot spot
>                            ; Insert to bdi->cgwb_tree, mmecg_css->id as key
> 
> We discussed it through, base on the analysis:  When we running into the
> issue, there is cgroups are being deleted. The inodes (files) that were
> associated with these cgroups have to switch into another newly created
> writeback. We think there are huge amount of inodes in the writeback list
> that time. So we don't think there is anything abnormal. However, one
> thing we possibly can do: enforce these inodes to BDI embedded wirteback
> and we needn't create huge amount of writebacks in that case, to avoid
> the high system load phenomenon. We expect correct wb (best candidate) is
> picked up in next round.
> 
> Thanks,
> Liguang
> 
> > 
> 

If I understand correctly, this is mostlikely caused by a file shared by
cgroup A and cgroup B. This means cgroup B is doing direct io against
the file owned by the dying cgroup A. In this case, the code tries to do
a wb switch. However, it fails to reallocate the wb as it's deleted and
for the original cgrouip A's memcg id.

I think the below may be a better solution. Could you please test it? If
it works, I'll spin a patch with a more involved description.

Thanks,
Dennis

---
diff --git a/fs/fs-writeback.c b/fs/fs-writeback.c
index 36855c1f8daf..fb331ea2a626 100644
--- a/fs/fs-writeback.c
+++ b/fs/fs-writeback.c
@@ -577,7 +577,7 @@ void wbc_attach_and_unlock_inode(struct writeback_control *wbc,
 	 * A dying wb indicates that the memcg-blkcg mapping has changed
 	 * and a new wb is already serving the memcg.  Switch immediately.
 	 */
-	if (unlikely(wb_dying(wbc->wb)))
+	if (unlikely(wb_dying(wbc->wb)) && !css_is_dying(wbc->wb->memcg_css))
 		inode_switch_wbs(inode, wbc->wb_id);
 }
 
diff --git a/mm/backing-dev.c b/mm/backing-dev.c
index 72e6d0c55cfa..685563ed9788 100644
--- a/mm/backing-dev.c
+++ b/mm/backing-dev.c
@@ -659,7 +659,7 @@ struct bdi_writeback *wb_get_create(struct backing_dev_info *bdi,
 
 	might_sleep_if(gfpflags_allow_blocking(gfp));
 
-	if (!memcg_css->parent)
+	if (!memcg_css->parent || css_is_dying(memcg_css))
 		return &bdi->wb;
 
 	do {

Re: [PATCH] fs/writeback: Attach inode's wb to root if needed

[乱石]

Hi Dennis,

Sorry for the later reply. Becase I cann't reproduce this problem by 
local test,

and online environment is not allowed to operate, I am constructing 
scenario

to reproduce it in recent days.


在 2019/5/14 2:30, Dennis Zhou 写道:
> Hi Liguang,
>
> On Fri, May 10, 2019 at 09:54:27AM +0800, 乱石 wrote:
>> Hi Tejun,
>>
>> 在 2019/5/10 0:48, Tejun Heo 写道:
>>> Hi Tejun,
>>>
>>> On Thu, May 09, 2019 at 04:03:53PM +0800, zhangliguang wrote:
>>>> There might have tons of files queued in the writeback, awaiting for
>>>> writing back. Unfortunately, the writeback's cgroup has been dead. In
>>>> this case, we reassociate the inode with another writeback cgroup, but
>>>> we possibly can't because the writeback associated with the dead cgroup
>>>> is the only valid one. In this case, the new writeback is allocated,
>>>> initialized and associated with the inode. It causes unnecessary high
>>>> system load and latency.
>>>>
>>>> This fixes the issue by enforce moving the inode to root cgroup when the
>>>> previous binding cgroup becomes dead. With it, no more unnecessary
>>>> writebacks are created, populated and the system load decreased by about
>>>> 6x in the online service we encounted:
>>>>       Without the patch: about 30% system load
>>>>       With the patch:    about  5% system load
>>> Can you please describe the scenario with more details?  I'm having a
>>> bit of hard time understanding the amount of cpu cycles being
>>> consumed.
>>>
>>> Thanks.
>> Our search line reported a problem, when containerA was removed,
>> containerB and containerC's system load were up to 30%.
>>
>> We record the trace with 'perf record cycles:k -g -a', found that wb_init
>> was the hotspot function.
>>
>> Function call:
>>
>> generic_file_direct_write
>>     filemap_write_and_wait_range
>>        __filemap_fdatawrite_range
>>           wbc_attach_fdatawrite_inode
>>              inode_attach_wb
>>                 __inode_attach_wb
>>                    wb_get_create
>>              wbc_attach_and_unlock_inode
>>                 if (unlikely(wb_dying(wbc->wb)))
>>                    inode_switch_wbs
>>                       wb_get_create
>>                          ; Search bdi->cgwb_tree from memcg_css->id
>>                          ; OR cgwb_create
>>                             kmalloc
>>                             wb_init       // hot spot
>>                             ; Insert to bdi->cgwb_tree, mmecg_css->id as key
>>
>> We discussed it through, base on the analysis:  When we running into the
>> issue, there is cgroups are being deleted. The inodes (files) that were
>> associated with these cgroups have to switch into another newly created
>> writeback. We think there are huge amount of inodes in the writeback list
>> that time. So we don't think there is anything abnormal. However, one
>> thing we possibly can do: enforce these inodes to BDI embedded wirteback
>> and we needn't create huge amount of writebacks in that case, to avoid
>> the high system load phenomenon. We expect correct wb (best candidate) is
>> picked up in next round.
>>
>> Thanks,
>> Liguang
>>
> If I understand correctly, this is mostlikely caused by a file shared by
> cgroup A and cgroup B. This means cgroup B is doing direct io against
> the file owned by the dying cgroup A. In this case, the code tries to do
> a wb switch. However, it fails to reallocate the wb as it's deleted and
> for the original cgrouip A's memcg id.
>
> I think the below may be a better solution. Could you please test it? If
> it works, I'll spin a patch with a more involved description.
>
> Thanks,
> Dennis
>
> ---
> diff --git a/fs/fs-writeback.c b/fs/fs-writeback.c
> index 36855c1f8daf..fb331ea2a626 100644
> --- a/fs/fs-writeback.c
> +++ b/fs/fs-writeback.c
> @@ -577,7 +577,7 @@ void wbc_attach_and_unlock_inode(struct writeback_control *wbc,
>   	 * A dying wb indicates that the memcg-blkcg mapping has changed
>   	 * and a new wb is already serving the memcg.  Switch immediately.
>   	 */
> -	if (unlikely(wb_dying(wbc->wb)))
> +	if (unlikely(wb_dying(wbc->wb)) && !css_is_dying(wbc->wb->memcg_css))
>   		inode_switch_wbs(inode, wbc->wb_id);
>   }
>   
> diff --git a/mm/backing-dev.c b/mm/backing-dev.c
> index 72e6d0c55cfa..685563ed9788 100644
> --- a/mm/backing-dev.c
> +++ b/mm/backing-dev.c
> @@ -659,7 +659,7 @@ struct bdi_writeback *wb_get_create(struct backing_dev_info *bdi,
>   
>   	might_sleep_if(gfpflags_allow_blocking(gfp));
>   
> -	if (!memcg_css->parent)
> +	if (!memcg_css->parent || css_is_dying(memcg_css))
>   		return &bdi->wb;
>   
>   	do {