Message-Id: <20080701000422.67e211c3.akpm@linux-foundation.org>
Date:	Tue, 1 Jul 2008 00:04:22 -0700
From:	Andrew Morton <akpm@...ux-foundation.org>
To:	Ian Kent <raven@...maw.net>
Cc:	Kernel Mailing List <linux-kernel@...r.kernel.org>,
	autofs mailing list <autofs@...ux.kernel.org>,
	linux-fsdevel <linux-fsdevel@...r.kernel.org>
Subject: Re: [PATCH 7/7] autofs4 - fix pending mount race.

On Tue, 17 Jun 2008 20:24:12 +0800 Ian Kent <raven@...maw.net> wrote:

> Close a race between a pending mount that is about to finish and a new
> lookup for the same directory.
> 
> Process P1 triggers a mount of directory foo.
> It sets DCACHE_AUTOFS_PENDING in the ->lookup routine, creates a waitq
> entry for 'foo', and calls out to the daemon to perform the mount.
> The autofs daemon will then create the directory 'foo', using a new dentry
> that will be hashed in the dcache.
> 
> Before the mount completes, another process, P2, tries to walk into the
> 'foo' directory. The vfs path walking code finds an entry for 'foo' and
> calls the revalidate method. Revalidate finds that the entry is not
> PENDING (because PENDING was never set on the dentry created by the mkdir),
> but it does find the directory is empty. Revalidate calls try_to_fill_dentry,
> which sets the PENDING flag and then calls into the autofs4 wait code to
> trigger or wait for a mount of 'foo'. The wait code finds the entry for
> 'foo' and goes to sleep waiting for the completion of the mount.
> 
> Yet another process, P3, tries to walk into the 'foo' directory. This
> process again finds a dentry in the dcache for 'foo', and calls into
> the autofs revalidate code.
> 
> The revalidate code finds that the PENDING flag is set, and so calls
> try_to_fill_dentry.
> 
> a) try_to_fill_dentry sets the PENDING flag redundantly for this dentry,
>    then calls into the autofs4 wait code.
> b) the autofs4 wait code takes the waitq mutex and searches for an entry
>    for 'foo'
> 
> Between a and b, P1 is woken up because the mount completed.
> P1 takes the wait queue mutex, clears the PENDING flag from the dentry,
> and removes the waitqueue entry for 'foo' from the list.
> 
> When it releases the waitq mutex, P3 (eventually) acquires it.  At this
> time, it looks for an existing waitq for 'foo', finds none, and so
> creates a new one and calls out to the daemon to mount the 'foo' directory.
> 
> Now, the reason that three processes are required to trigger this race
> is that, because the PENDING flag is not set on the dentry created by
> mkdir, the window for the race would be way to slim for it to ever occur.
> Basically, between the testing of d_mountpoint(dentry) and the taking of
> the waitq mutex, the mount would have to complete and the daemon would
> have to be woken up, and that in turn would have to wake up P1.  This is
> simply impossible.  Add the third process, though, and it becomes slightly
> more likely.
> 
> ...
>
> diff --git a/fs/autofs4/waitq.c b/fs/autofs4/waitq.c
> index 5208cfb..cd21fd4 100644
> --- a/fs/autofs4/waitq.c
> +++ b/fs/autofs4/waitq.c
> @@ -207,19 +207,106 @@ autofs4_find_wait(struct autofs_sb_info *sbi, struct qstr *qstr)
>  	return wq;
>  }
>  
> +/*
> + * Check if we have a valid request.
> + * Returns
> + * 1 if the request should continue.
> + *   In this case we can return an autofs_wait_queue entry if one is
> + *   found or NULL to idicate a new wait needs to be created.
> + * 0 or a negative errno if the request shouldn't continue.
> + */
> +static int validate_request(struct autofs_wait_queue **wait,
> +			    struct autofs_sb_info *sbi,
> +			    struct qstr *qstr,
> +			    struct dentry*dentry, enum autofs_notify notify)
> +{
> +	struct autofs_wait_queue *wq;
> +	struct autofs_info *ino;
> +
> +	/* Wait in progress, continue; */
> +	wq = autofs4_find_wait(sbi, qstr);
> +	if (wq) {
> +		*wait = wq;
> +		return 1;

Returns 1 with the mutex held.

> +	}
> +
> +	*wait = NULL;
> +
> +	/* If we don't yet have any info this is a new request */
> +	ino = autofs4_dentry_ino(dentry);
> +	if (!ino)
> +		return 1;
> +
> +	/*
> +	 * If we've been asked to wait on an existing expire (NFY_NONE)
> +	 * but there is no wait in the queue ...
> +	 */
> +	if (notify == NFY_NONE) {
> +		/*
> +		 * Either we've betean the pending expire to post it's
> +		 * wait or it finished while we waited on the mutex.
> +		 * So we need to wait till either, the wait appears
> +		 * or the expire finishes.
> +		 */

Wanna have another go at that comment?  The grammar and spelling should
cause an oops or something.

> +		while (ino->flags & AUTOFS_INF_EXPIRING) {
> +			mutex_unlock(&sbi->wq_mutex);
> +			schedule_timeout_interruptible(HZ/10);
> +			if (mutex_lock_interruptible(&sbi->wq_mutex))
> +				return -EINTR;

Returns -EFOO with the mutex not held.

> +
> +			wq = autofs4_find_wait(sbi, qstr);
> +			if (wq) {
> +				*wait = wq;
> +				return 1;
> +			}
> +		}
> +
> +		/*
> +		 * Not ideal but the status has already gone. Of the two
> +		 * cases where we wait on NFY_NONE neither depend on the
> +		 * return status of the wait.
> +		 */
> +		return 0;

Returns zero with the mutex held.

> +	}
> +
> +	/*
> +	 * If we've been asked to trigger a mount and the request
> +	 * completed while we waited on the mutex ...
> +	 */
> +	if (notify == NFY_MOUNT) {
> +		/*
> +		 * If the dentry isn't hashed just go ahead and try the
> +		 * mount again with a new wait (not much else we can do).
> +		*/
> +		if (!d_unhashed(dentry)) {
> +			/*
> +			 * But if the dentry is hashed, that means that we
> +			 * got here through the revalidate path.  Thus, we
> +			 * need to check if the dentry has been mounted
> +			 * while we waited on the wq_mutex. If it has,
> +			 * simply return success.
> +			 */
> +			if (d_mountpoint(dentry))
> +				return 0;
> +		}
> +	}
> +
> +	return 1;
> +}
>
> ...
>
> +	ret = validate_request(&wq, sbi, &qstr, dentry, notify);
> +	if (ret <= 0) {
> +		if (ret == 0)
>  			mutex_unlock(&sbi->wq_mutex);
> -			return 0;
> -		}
> +		kfree(qstr.name);
> +		return ret;
>  	}

Leave the mutex held if it returned 1.  Doesn't unlock the mutex if it
returned -EFOO.  Presumably callers of this function will unlock the
mutex if it returned zero.

Or something like that.  My brain just exploded.

Please double-check the locking protocol here and then document the
sorry thing.

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists