Linux-Fsdevel Archive on lore.kernel.org
help / color / mirror / Atom feed
From: ebiederm@xmission.com (Eric W. Biederman)
To: Junxiao Bi <junxiao.bi@oracle.com>
Cc: Matthew Wilcox <willy@infradead.org>,
	linux-kernel@vger.kernel.org, linux-fsdevel@vger.kernel.org,
	Matthew Wilcox <matthew.wilcox@oracle.com>,
	Srinivas Eeda <SRINIVAS.EEDA@oracle.com>,
	"joe.jin\@oracle.com" <joe.jin@oracle.com>
Subject: Re: severe proc dentry lock contention
Date: Thu, 18 Jun 2020 22:30:15 -0500	[thread overview]
Message-ID: <87sger91h4.fsf@x220.int.ebiederm.org> (raw)
In-Reply-To: <2cf6af59-e86b-f6cc-06d3-84309425bd1d@oracle.com> (Junxiao Bi's message of "Thu, 18 Jun 2020 17:27:43 -0700")

Junxiao Bi <junxiao.bi@oracle.com> writes:

> On 6/18/20 5:02 PM, ebiederm@xmission.com wrote:
>
>> Matthew Wilcox <willy@infradead.org> writes:
>>
>>> On Thu, Jun 18, 2020 at 03:17:33PM -0700, Junxiao Bi wrote:
>>>> When debugging some performance issue, i found that thousands of threads
>>>> exit around same time could cause a severe spin lock contention on proc
>>>> dentry "/proc/$parent_process_pid/task/", that's because threads needs to
>>>> clean up their pid file from that dir when exit. Check the following
>>>> standalone test case that simulated the case and perf top result on v5.7
>>>> kernel. Any idea on how to fix this?
>>> Thanks, Junxiao.
>>>
>>> We've looked at a few different ways of fixing this problem.
>>>
>>> Even though the contention is within the dcache, it seems like a usecase
>>> that the dcache shouldn't be optimised for -- generally we do not have
>>> hundreds of CPUs removing dentries from a single directory in parallel.
>>>
>>> We could fix this within procfs.  We don't have a great patch yet, but
>>> the current approach we're looking at allows only one thread at a time
>>> to call dput() on any /proc/*/task directory.
>>>
>>> We could also look at fixing this within the scheduler.  Only allowing
>>> one CPU to run the threads of an exiting process would fix this particular
>>> problem, but might have other consequences.
>>>
>>> I was hoping that 7bc3e6e55acf would fix this, but that patch is in 5.7,
>>> so that hope is ruled out.
>> Does anyone know if problem new in v5.7?  I am wondering if I introduced
>> this problem when I refactored the code or if I simply churned the code
>> but the issue remains effectively the same.
> It's not new issue, we see it in old kernel like v4.14
>>
>> Can you try only flushing entries when the last thread of the process is
>> reaped?  I think in practice we would want to be a little more
>> sophisticated but it is a good test case to see if it solves the issue.
>
> Thank you. i will try and let you know.

Assuming this works we can remove the hacking part of always doing
this by only doing this if SIGNAL_GROUP_EXIT when we know this
thundering herd will appear.

We still need to do something with the list however.

If your testing works out performance wise I will figure out what
we need to make the hack generale and safe.

Eric



> Thanks,
>
> Junxiao.
>
>>
>> diff --git a/kernel/exit.c b/kernel/exit.c
>> index cebae77a9664..d56e4eb60bdd 100644
>> --- a/kernel/exit.c
>> +++ b/kernel/exit.c
>> @@ -152,7 +152,7 @@ void put_task_struct_rcu_user(struct task_struct *task)
>>   void release_task(struct task_struct *p)
>>   {
>>   	struct task_struct *leader;
>> -	struct pid *thread_pid;
>> +	struct pid *thread_pid = NULL;
>>   	int zap_leader;
>>   repeat:
>>   	/* don't need to get the RCU readlock here - the process is dead and
>> @@ -165,7 +165,8 @@ void release_task(struct task_struct *p)
>>     	write_lock_irq(&tasklist_lock);
>>   	ptrace_release_task(p);
>> -	thread_pid = get_pid(p->thread_pid);
>> +	if (p == p->group_leader)
>> +		thread_pid = get_pid(p->thread_pid);
>>   	__exit_signal(p);
>>     	/*
>> @@ -188,8 +189,10 @@ void release_task(struct task_struct *p)
>>   	}
>>     	write_unlock_irq(&tasklist_lock);
>> -	proc_flush_pid(thread_pid);
>> -	put_pid(thread_pid);
>> +	if (thread_pid) {
>> +		proc_flush_pid(thread_pid);
>> +		put_pid(thread_pid);
>> +	}
>>   	release_thread(p);
>>   	put_task_struct_rcu_user(p);
>>   

  reply	other threads:[~2020-06-19  3:34 UTC|newest]

Thread overview: 20+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2020-06-18 22:17 Junxiao Bi
2020-06-18 23:39 ` Matthew Wilcox
2020-06-19  0:02   ` Eric W. Biederman
2020-06-19  0:27     ` Junxiao Bi
2020-06-19  3:30       ` Eric W. Biederman [this message]
2020-06-19 14:09       ` [PATCH] proc: Avoid a thundering herd of threads freeing proc dentries Eric W. Biederman
2020-06-19 15:56         ` Junxiao Bi
2020-06-19 17:24           ` Eric W. Biederman
2020-06-19 21:56             ` Junxiao Bi
2020-06-19 22:42               ` Eric W. Biederman
2020-06-20 16:27                 ` Matthew Wilcox
2020-06-22  5:15                   ` Junxiao Bi
2020-06-22 15:20                     ` Eric W. Biederman
2020-06-22 15:48                       ` willy
2020-08-17 12:19                         ` Eric W. Biederman
2020-06-22 17:16                       ` Junxiao Bi
2020-06-23  0:47                     ` Matthew Wilcox
2020-06-25 22:11                       ` Junxiao Bi
2020-06-22  5:33         ` Masahiro Yamada
2020-06-22 15:13           ` Eric W. Biederman

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=87sger91h4.fsf@x220.int.ebiederm.org \
    --to=ebiederm@xmission.com \
    --cc=SRINIVAS.EEDA@oracle.com \
    --cc=joe.jin@oracle.com \
    --cc=junxiao.bi@oracle.com \
    --cc=linux-fsdevel@vger.kernel.org \
    --cc=linux-kernel@vger.kernel.org \
    --cc=matthew.wilcox@oracle.com \
    --cc=willy@infradead.org \
    --subject='Re: severe proc dentry lock contention' \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).