Re: [PATCH] nfsd: free nfsd_file by gc after adding it to lru list

[Jeff Layton <jlayton@kernel.org>]

On Wed, 2025-01-22 at 14:48 +1100, NeilBrown wrote:
> On Wed, 22 Jan 2025, NeilBrown wrote:
> > On Wed, 22 Jan 2025, Jeff Layton wrote:
> > > To be clear, I think we need to drop e57420be100ab from your nfsd-
> > > testing branch. The race I identified above is quite likely to occur
> > > and could lead to leaks.
> > > 
> > > If Li Lingfeng doesn't propose a patch, I'll spin one up tomorrow. I
> > > think the RCU approach is safe.
> > 
> > I'm not convinced this is the right approach.
> > I cannot see how nfsd_file_put() can race with unhashing.  If it cannot
> > then we can simply unconditionally call nfsd_file_schedule_laundrette().
> > 
> > Can describe how the race can happen - if indeed it can.
> 
> I thought I should explore this more and explain what I think actually
> happens ...
> 
> Certainly nfsd_file_unhash() might race with nfsd_file_put().  At this
> point in nfsd_file_put() we have the only reference but a hash lookup
> could gain another reference and the immediately unhash it.
> nfsd_file_queue_for_close() can do this.  There might be other paths.
> 
> But why does this mean we need to remove it from the lru and free it
> immediately?  If we leave it on the lru it will be freed in a couple of
> seconds.
> 
> The reason might be nfsd_file_close_inode_sync().  This needs to close
> files before returning.
> But if nfsd_file_close_inode_sync() is called while some other thread
> holds a reference to the file and might want to call nfsd_file_put(),
> then it isn't going to succeed anyway so any race here doesn't make any
> difference.
> 
> So I think the following might be the best fix
> 
> ???
> 
> Thanks,
> NeilBrown
> 
> diff --git a/fs/nfsd/filecache.c b/fs/nfsd/filecache.c
> index fcd751cb7c76..773788a50e56 100644
> --- a/fs/nfsd/filecache.c
> +++ b/fs/nfsd/filecache.c
> @@ -322,10 +322,13 @@ nfsd_file_check_writeback(struct nfsd_file *nf)
>  static bool nfsd_file_lru_add(struct nfsd_file *nf)
>  {
>  	set_bit(NFSD_FILE_REFERENCED, &nf->nf_flags);
> +	rcu_read_lock();
>  	if (list_lru_add_obj(&nfsd_file_lru, &nf->nf_lru)) {
>  		trace_nfsd_file_lru_add(nf);
> +		rcu_read_unlock();
>  		return true;
>  	}
> +	rcu_read_unlock();
>  	return false;
>  }
>  

I think that I'm now convinced that it's OK to remove the code in the
if block below. But if we do that, then I don't think you need the
rcu_read_lock() in nfsd_file_lru_add(). It's just handing off the
reference to the LRU at that point and once that's done, it doesn't
need to look at it again. That makes the rcu_read_lock() unnecessary.

Given that, Li Lingfeng's original patch is OK after all.

Am I missing something?


> @@ -371,19 +374,8 @@ nfsd_file_put(struct nfsd_file *nf)
>  
>  		/* Try to add it to the LRU.  If that fails, decrement. */
>  		if (nfsd_file_lru_add(nf)) {
> -			/* If it's still hashed, we're done */
> -			if (test_bit(NFSD_FILE_HASHED, &nf->nf_flags)) {
> -				nfsd_file_schedule_laundrette();
> -				return;
> -			}
> -
> -			/*
> -			 * We're racing with unhashing, so try to remove it from
> -			 * the LRU. If removal fails, then someone else already
> -			 * has our reference.
> -			 */
> -			if (!nfsd_file_lru_remove(nf))
> -				return;
> +			nfsd_file_schedule_laundrette();
> +			return;
>  		}
>  	}
>  	if (refcount_dec_and_test(&nf->nf_ref))
> 

-- 
Jeff Layton <jlayton@kernel.org>