Date:   Fri, 2 Jun 2023 11:13:09 +0800
From:   Qi Zheng <qi.zheng@...ux.dev>
To:     Dave Chinner <david@...morbit.com>
Cc:     akpm@...ux-foundation.org, tkhai@...ru, roman.gushchin@...ux.dev,
        vbabka@...e.cz, viro@...iv.linux.org.uk, brauner@...nel.org,
        djwong@...nel.org, hughd@...gle.com, paulmck@...nel.org,
        muchun.song@...ux.dev, linux-mm@...ck.org,
        linux-fsdevel@...r.kernel.org, linux-xfs@...r.kernel.org,
        linux-kernel@...r.kernel.org, Qi Zheng <zhengqi.arch@...edance.com>
Subject: Re: [PATCH 6/8] xfs: introduce xfs_fs_destroy_super()

Hi Dave,

On 2023/6/2 07:06, Dave Chinner wrote:
> On Thu, Jun 01, 2023 at 04:43:32PM +0800, Qi Zheng wrote:
>> Hi Dave,
>> On 2023/6/1 07:48, Dave Chinner wrote:
>>> On Wed, May 31, 2023 at 09:57:40AM +0000, Qi Zheng wrote:
>>>> From: Kirill Tkhai <tkhai@...ru>
>>> I don't really like this ->destroy_super() callback, especially as
>>> it's completely undocumented as to why it exists. This is purely a
>>> work-around for handling extended filesystem superblock shrinker
>>> functionality, yet there's nothing that tells the reader this.
>>>
>>> It also seems to imply that the superblock shrinker can continue to
>>> run after the existing unregister_shrinker() call before ->kill_sb()
>>> is called. This violates the assumption made in filesystems that the
>>> superblock shrinkers have been stopped and will never run again
>>> before ->kill_sb() is called. Hence ->kill_sb() implementations
>>> assume there is nothing else accessing filesystem owned structures
>>> and it can tear down internal structures safely.
>>>
>>> Realistically, the days of XFS using this superblock shrinker
>>> extension are numbered. We've got a lot of the infrastructure we
>>> need in place to get rid of the background inode reclaim
>>> infrastructure that requires this shrinker extension, and it's on my
>>> list of things that need to be addressed in the near future.
>>>
>>> In fact, now that I look at it, I think the shmem usage of this
>>> superblock shrinker interface is broken - it returns SHRINK_STOP to
>>> ->free_cached_objects(), but the only valid return value is the
>>> number of objects freed (i.e. 0 is nothing freed). These special
>>> superblock extension interfaces do not work like a normal
>>> shrinker....
>>>
>>> Hence I think the shmem usage should be replaced with an separate
>>> internal shmem shrinker that is managed by the filesystem itself
>>> (similar to how XFS has multiple internal shrinkers).
>>>
>>> At this point, then the only user of this interface is (again) XFS.
>>> Given this, adding new VFS methods for a single filesystem
>>> for functionality that is planned to be removed is probably not the
>>> best approach to solving the problem.
>>
>> Thanks for such a detailed analysis. Kirill Tkhai just proposeed a
>> new method[1], I cc'd you on the email.
> 
> I;ve just read through that thread, and I've looked at the original
> patch that caused the regression.
> 
> I'm a bit annoyed right now. Nobody cc'd me on the original patches
> nor were any of the subsystems that use shrinkers were cc'd on the
> patches that changed shrinker behaviour. I only find out about this

Sorry about that, my mistake. I followed the results of
scripts/get_maintainer.pl before.

> because someone tries to fix something they broke by *breaking more
> stuff* and not even realising how broken what they are proposing is.

Yes, this slows down the speed of umount. But the benefit is that
slab shrink becomes lockless, the mount operation and slab shrink no
longer affect each other, and the IPC no longer drops significantly,
etc.

And I used bpftrace to measure the time consumption of
unregister_shrinker():

```
And I just tested it on a physical machine (Intel(R) Xeon(R) Platinum
8260 CPU @ 2.40GHz) and the results are as follows:

1) use synchronize_srcu():

@ns[umount]:
[8K, 16K)             83 |@@@@@@@       |
[16K, 32K)           578 
|@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@|
[32K, 64K)            78 |@@@@@@@       |
[64K, 128K)            6 |       |
[128K, 256K)           7 |       |
[256K, 512K)          29 |@@       |
[512K, 1M)            51 |@@@@      |
[1M, 2M)              90 |@@@@@@@@       |
[2M, 4M)              70 |@@@@@@      |
[4M, 8M)               8 |      |

2) use synchronize_srcu_expedited():

@ns[umount]:
[8K, 16K)             31 |@@       |
[16K, 32K)           803 
|@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@|
[32K, 64K)           158 |@@@@@@@@@@       |
[64K, 128K)            4 |       |
[128K, 256K)           2 |       |
[256K, 512K)           2 |       |
```

With synchronize_srcu(), most of the time consumption is between 16us 
and 32us, the worst case between 4ms and 8ms. Is this totally
unacceptable?

This performance regression report comes from a stress test. Will the
umount action be executed so frequently under real workloads?

If there are really unacceptable, after applying the newly proposed
method, umount will be as fast as before (or even faster).

Thanks,
Qi

> 
> The previous code was not broken and it provided specific guarantees
> to subsystems via unregister_shrinker(). From the above discussion,
> it appears that the original authors of these changes either did not
> know about or did not understand them, so that casts doubt in my
> mind about the attempted solution and all the proposed fixes for it.
> 
> I don't have the time right now unravel this mess and fully
> understand the original problem, changes or the band-aids that are
> being thrown around. We are also getting quite late in the cycle to
> be doing major surgery to critical infrastructure, especially as it
> gives so little time to review regression test whatever new solution
> is proposed.
> 
> Given this appears to be a change introduced in 6.4-rc1, I think the
> right thing to do is to revert the change rather than make things
> worse by trying to shove some "quick fix" into the kernel to address
> it.
> 
> Andrew, could you please sort out a series to revert this shrinker
> infrastructure change and all the dependent hacks that have been
> added to try to fix it so far?
> 
> -Dave.

-- 
Thanks,
Qi

Powered by blists - more mailing lists