[PATCH] fs: dcache: Avoid livelock between d_alloc_parallel and __d_add

[PATCH] fs: dcache: Avoid livelock between d_alloc_parallel and __d_add

[Will Deacon]

If d_alloc_parallel runs concurrently with __d_add, it is possible for
d_alloc_parallel to continuously retry whilst i_dir_seq has been
incremented to an odd value by __d_add:

CPU0:
__d_add
	n = start_dir_add(dir);
		cmpxchg(&dir->i_dir_seq, n, n + 1) == n

CPU1:
d_alloc_parallel
retry:
	seq = smp_load_acquire(&parent->d_inode->i_dir_seq) & ~1;
	hlist_bl_lock(b);
		bit_spin_lock(0, (unsigned long *)b); // Always succeeds

CPU0:
	__d_lookup_done(dentry)
		hlist_bl_lock
			bit_spin_lock(0, (unsigned long *)b); // Never succeeds

CPU1:
	if (unlikely(parent->d_inode->i_dir_seq != seq)) {
		hlist_bl_unlock(b);
		goto retry;
	}

Since the simple bit_spin_lock used to implement hlist_bl_lock does not
provide any fairness guarantees, then CPU1 can starve CPU0 of the lock
and prevent it from reaching end_dir_add(dir), therefore CPU1 cannot
exit its retry loop because the sequence number always has the bottom
bit set.

This patch resolves the livelock by not taking hlist_bl_lock in
d_alloc_parallel if the sequence counter is odd, since any subsequent
masked comparison with i_dir_seq will fail anyway.

Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Will Deacon <will.deacon@arm.com>
---
 fs/dcache.c | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)

diff --git a/fs/dcache.c b/fs/dcache.c
index 7c38f39958bc..b243deec298c 100644
--- a/fs/dcache.c
+++ b/fs/dcache.c
@@ -2474,7 +2474,7 @@ struct dentry *d_alloc_parallel(struct dentry *parent,
 
 retry:
 	rcu_read_lock();
-	seq = smp_load_acquire(&parent->d_inode->i_dir_seq) & ~1;
+	seq = smp_load_acquire(&parent->d_inode->i_dir_seq);
 	r_seq = read_seqbegin(&rename_lock);
 	dentry = __d_lookup_rcu(parent, name, &d_seq);
 	if (unlikely(dentry)) {
@@ -2495,6 +2495,12 @@ struct dentry *d_alloc_parallel(struct dentry *parent,
 		rcu_read_unlock();
 		goto retry;
 	}
+
+	if (unlikely(seq & 1)) {
+		rcu_read_unlock();
+		goto retry;
+	}
+
 	hlist_bl_lock(b);
 	if (unlikely(parent->d_inode->i_dir_seq != seq)) {
 		hlist_bl_unlock(b);
-- 
2.1.4

Re: [PATCH] fs: dcache: Avoid livelock between d_alloc_parallel and __d_add

[Peter Zijlstra]

On Tue, Feb 13, 2018 at 12:58:51PM +0000, Will Deacon wrote:
> If d_alloc_parallel runs concurrently with __d_add, it is possible for
> d_alloc_parallel to continuously retry whilst i_dir_seq has been
> incremented to an odd value by __d_add:
> 
> CPU0:
> __d_add
> 	n = start_dir_add(dir);
> 		cmpxchg(&dir->i_dir_seq, n, n + 1) == n
> 
> CPU1:
> d_alloc_parallel
> retry:
> 	seq = smp_load_acquire(&parent->d_inode->i_dir_seq) & ~1;
> 	hlist_bl_lock(b);
> 		bit_spin_lock(0, (unsigned long *)b); // Always succeeds
> 
> CPU0:
> 	__d_lookup_done(dentry)
> 		hlist_bl_lock
> 			bit_spin_lock(0, (unsigned long *)b); // Never succeeds
> 
> CPU1:
> 	if (unlikely(parent->d_inode->i_dir_seq != seq)) {
> 		hlist_bl_unlock(b);
> 		goto retry;
> 	}
> 
> Since the simple bit_spin_lock used to implement hlist_bl_lock does not

And cannot, a single bit is just not enough state.

> provide any fairness guarantees, then CPU1 can starve CPU0 of the lock
> and prevent it from reaching end_dir_add(dir), therefore CPU1 cannot
> exit its retry loop because the sequence number always has the bottom
> bit set.
> 
> This patch resolves the livelock by not taking hlist_bl_lock in
> d_alloc_parallel if the sequence counter is odd, since any subsequent
> masked comparison with i_dir_seq will fail anyway.
> 

Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>

> Cc: Al Viro <viro@zeniv.linux.org.uk>
> Signed-off-by: Will Deacon <will.deacon@arm.com>
> ---
>  fs/dcache.c | 8 +++++++-
>  1 file changed, 7 insertions(+), 1 deletion(-)
> 
> diff --git a/fs/dcache.c b/fs/dcache.c
> index 7c38f39958bc..b243deec298c 100644
> --- a/fs/dcache.c
> +++ b/fs/dcache.c
> @@ -2474,7 +2474,7 @@ struct dentry *d_alloc_parallel(struct dentry *parent,
>  
>  retry:
>  	rcu_read_lock();
> -	seq = smp_load_acquire(&parent->d_inode->i_dir_seq) & ~1;
> +	seq = smp_load_acquire(&parent->d_inode->i_dir_seq);
>  	r_seq = read_seqbegin(&rename_lock);
>  	dentry = __d_lookup_rcu(parent, name, &d_seq);
>  	if (unlikely(dentry)) {
> @@ -2495,6 +2495,12 @@ struct dentry *d_alloc_parallel(struct dentry *parent,
>  		rcu_read_unlock();
>  		goto retry;
>  	}
> +
> +	if (unlikely(seq & 1)) {
> +		rcu_read_unlock();
> +		goto retry;
> +	}
> +
>  	hlist_bl_lock(b);
>  	if (unlikely(parent->d_inode->i_dir_seq != seq)) {

Also, should that not read:

	if (unlikely(READ_ONCE(parent->d_inode->i_dir_seq) != seq)) {

I mean, load-tearing can only result in additional failure, but still.

>  		hlist_bl_unlock(b);

Re: [PATCH] fs: dcache: Avoid livelock between d_alloc_parallel and __d_add

[Matthew Wilcox]

On Tue, Feb 13, 2018 at 12:58:51PM +0000, Will Deacon wrote:
> This patch resolves the livelock by not taking hlist_bl_lock in
> d_alloc_parallel if the sequence counter is odd, since any subsequent
> masked comparison with i_dir_seq will fail anyway.
> 
> Cc: Peter Zijlstra <peterz@infradead.org>
> Cc: Al Viro <viro@zeniv.linux.org.uk>
> Signed-off-by: Will Deacon <will.deacon@arm.com>

Reviewed-by: Matthew Wilcox <mawilcox@microsoft.com>

I wonder whether it makes sense to turn i_dir_seq into a seqcount_t,
which would give us the lockdep checking as well.

Re: [PATCH] fs: dcache: Avoid livelock between d_alloc_parallel and __d_add

[Will Deacon]

Hi Matthew,

On Tue, Feb 13, 2018 at 07:16:08AM -0800, Matthew Wilcox wrote:
> On Tue, Feb 13, 2018 at 12:58:51PM +0000, Will Deacon wrote:
> > This patch resolves the livelock by not taking hlist_bl_lock in
> > d_alloc_parallel if the sequence counter is odd, since any subsequent
> > masked comparison with i_dir_seq will fail anyway.
> > 
> > Cc: Peter Zijlstra <peterz@infradead.org>
> > Cc: Al Viro <viro@zeniv.linux.org.uk>
> > Signed-off-by: Will Deacon <will.deacon@arm.com>
> 
> Reviewed-by: Matthew Wilcox <mawilcox@microsoft.com>

Thanks!

> I wonder whether it makes sense to turn i_dir_seq into a seqcount_t,
> which would give us the lockdep checking as well.

I'm not sure it's quite as simple as that. start_dir_add looks very much
like it's intended to run concurrently, so we'd need a write_seqcount
implementation that provides the same atomicity guarantees.

Will