Re: [PATCH v2] ns: do not block exit_task_namespaces() for a long time

["Myklebust, Trond" <Trond.Myklebust@netapp.com>]

[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #1: Type: text/plain; charset="utf-8", Size: 5376 bytes --]

On Mon, 2012-07-16 at 19:39 +0300, Kirill A. Shutemov wrote:
> On Mon, Jul 16, 2012 at 03:39:36PM +0000, Myklebust, Trond wrote:
> > On Mon, 2012-07-16 at 18:09 +0300, Kirill A. Shutemov wrote:
> > > From: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
> > > 
> > > On exiting of the last task in a namespace we need to trigger freeing of
> > > the namespace. Currently, we call synchronize_rcu() and free_nsproxy()
> > > directly on do_exit() path.
> > > 
> > > On my machine synchronize_rcu() blocks for about 0.01 seconds. For
> > > comparing: normal exit_group() syscall takes less than 0.0003 seconds.
> > > 
> > > Let's offload synchronize_rcu() and free_nsproxy() to a workqueue.
> > > 
> > > I also move synchronize_rcu() inside free_nsproxy(). It fixes racy
> > > put_nsproxy() which calls free_nsproxy() without synchronize_rcu().
> > > I guess it was missed during switch to RCU (see cf7b708).
> > > 
> > > Microbenchmark:
> > > 
> > > : #define _GNU_SOURCE
> > > : #include <unistd.h>
> > > : #include <sched.h>
> > > : #include <stdlib.h>
> > > : #include <sys/wait.h>
> > > :
> > > : int
> > > : main(void)
> > > : {
> > > :       int i;
> > > :       for (i = 0; i < 1024; i++) {
> > > :               if (fork()) {
> > > :                       wait(NULL);
> > > :                       continue;
> > > :               }
> > > :               unshare(CLONE_NEWIPC);
> > > :               exit(0);
> > > :       }
> > > :       return 0;
> > > : }
> > > 
> > > Before the patch:
> > > 
> > > real    0m8.335s
> > > user    0m0.000s
> > > sys     0m0.265s
> > > 
> > > After:
> > > 
> > > real    0m0.569s
> > > user    0m0.001s
> > > sys     0m0.154s
> > > 
> > > Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
> > > Acked-by: Serge E. Hallyn <serge.hallyn@ubuntu.com>
> > > ---
> > > 
> > > v2:
> > >  - Updated description.
> > > 
> > > ---
> > >  include/linux/nsproxy.h |    1 +
> > >  kernel/nsproxy.c        |   34 +++++++++++++++++++++++-----------
> > >  2 files changed, 24 insertions(+), 11 deletions(-)
> > > 
> > > diff --git a/include/linux/nsproxy.h b/include/linux/nsproxy.h
> > > index cc37a55..1d26be7 100644
> > > --- a/include/linux/nsproxy.h
> > > +++ b/include/linux/nsproxy.h
> > > @@ -24,6 +24,7 @@ struct fs_struct;
> > >   */
> > >  struct nsproxy {
> > >  	atomic_t count;
> > > +	struct work_struct free_nsproxy_work;
> > >  	struct uts_namespace *uts_ns;
> > >  	struct ipc_namespace *ipc_ns;
> > >  	struct mnt_namespace *mnt_ns;
> > > diff --git a/kernel/nsproxy.c b/kernel/nsproxy.c
> > > index b576f7f..ebc7d40 100644
> > > --- a/kernel/nsproxy.c
> > > +++ b/kernel/nsproxy.c
> > > @@ -41,13 +41,17 @@ struct nsproxy init_nsproxy = {
> > >  #endif
> > >  };
> > >  
> > > +static void free_nsproxy_work(struct work_struct *work);
> > > +
> > >  static inline struct nsproxy *create_nsproxy(void)
> > >  {
> > >  	struct nsproxy *nsproxy;
> > >  
> > >  	nsproxy = kmem_cache_alloc(nsproxy_cachep, GFP_KERNEL);
> > > -	if (nsproxy)
> > > +	if (nsproxy) {
> > >  		atomic_set(&nsproxy->count, 1);
> > > +		INIT_WORK(&nsproxy->free_nsproxy_work, free_nsproxy_work);
> > > +	}
> > >  	return nsproxy;
> > >  }
> > >  
> > > @@ -166,6 +170,14 @@ out:
> > >  
> > >  void free_nsproxy(struct nsproxy *ns)
> > >  {
> > > +	/*
> > > +	 * wait for others to get what they want from this nsproxy.
> > > +	 *
> > > +	 * cannot release this nsproxy via the call_rcu() since
> > > +	 * put_mnt_ns() will want to sleep
> > > +	 */
> > > +	synchronize_rcu();
> > > +
> > >  	if (ns->mnt_ns)
> > >  		put_mnt_ns(ns->mnt_ns);
> > >  	if (ns->uts_ns)
> > > @@ -178,6 +190,14 @@ void free_nsproxy(struct nsproxy *ns)
> > >  	kmem_cache_free(nsproxy_cachep, ns);
> > >  }
> > >  
> > > +static void free_nsproxy_work(struct work_struct *work)
> > > +{
> > > +	struct nsproxy *ns = container_of(work, struct nsproxy,
> > > +			free_nsproxy_work);
> > > +
> > > +	free_nsproxy(ns);
> > > +}
> > > +
> > >  /*
> > >   * Called from unshare. Unshare all the namespaces part of nsproxy.
> > >   * On success, returns the new nsproxy.
> > > @@ -215,16 +235,8 @@ void switch_task_namespaces(struct task_struct *p, struct nsproxy *new)
> > >  
> > >  	rcu_assign_pointer(p->nsproxy, new);
> > >  
> > > -	if (ns && atomic_dec_and_test(&ns->count)) {
> > > -		/*
> > > -		 * wait for others to get what they want from this nsproxy.
> > > -		 *
> > > -		 * cannot release this nsproxy via the call_rcu() since
> > > -		 * put_mnt_ns() will want to sleep
> > > -		 */
> > > -		synchronize_rcu();
> > > -		free_nsproxy(ns);
> > > -	}
> > > +	if (ns && atomic_dec_and_test(&ns->count))
> > > +		schedule_work(&ns->free_nsproxy_work);
> > 
> > What's wrong with using call_rcu()? The above will cause a workqueue
> > thread to block for no good reason.
> 
> See comment to synchronize_rcu(). free_nsproxy() might sleep.
> call_rcu() callback invocation might happen from either softirq or process
> context, so we can't use it.

But call_rcu() should be allowed to call schedule_work(). At least you'd
be able to get rid of the 0.01s synchronize_rcu() sleep inside keventd.

-- 
Trond Myklebust
Linux NFS client maintainer

NetApp
Trond.Myklebust@netapp.com
www.netapp.com

ÿôèº{.nÇ+‰·Ÿ®‰­†+%ŠËÿ±éݶ\x17¥Šwÿº{.nÇ+‰·¥Š{±þG«éÿŠ{ayº\x1dʇڙë,j\a­¢f£¢·hšïêÿ‘êçz_è®\x03(­éšŽŠÝ¢j"ú\x1a¶^[m§ÿÿ¾\a«þG«éÿ¢¸?™¨è­Ú&£ø§~á¶iO•æ¬z·švØ^\x14\x04\x1a¶^[m§ÿÿÃ\fÿ¶ìÿ¢¸?–I¥