[PATCH] fs/dcache: optimize start_dir_add()

[PATCH] fs/dcache: optimize start_dir_add()

[Davidlohr Bueso]

Considering both end_dir_add() and d_alloc_parallel(), the
dir->i_dir_seq wants acquire/release semantics, therefore
micro-optimize for ll/sc archs and use finer grained barriers
to provide (load)-ACQUIRE ordering (L->S + L->L). This comes
at no additional cost for most of x86, as sane tso models will
have a nop for smp_rmb/smp_acquire__after_ctrl_dep.

Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
---
Alternatively I guess we could just use cmpxchg_acquire().

 fs/dcache.c | 11 ++++++++---
 1 file changed, 8 insertions(+), 3 deletions(-)

diff --git a/fs/dcache.c b/fs/dcache.c
index ea0485861d93..22738daccb9c 100644
--- a/fs/dcache.c
+++ b/fs/dcache.c
@@ -2502,13 +2502,18 @@ EXPORT_SYMBOL(d_rehash);
 
 static inline unsigned start_dir_add(struct inode *dir)
 {
+	unsigned n;
 
 	for (;;) {
-		unsigned n = dir->i_dir_seq;
-		if (!(n & 1) && cmpxchg(&dir->i_dir_seq, n, n + 1) == n)
-			return n;
+		n = READ_ONCE(dir->i_dir_seq);
+		if (!(n & 1) && cmpxchg_relaxed(&dir->i_dir_seq, n, n + 1) == n)
+			break;
 		cpu_relax();
 	}
+
+	/* create (load)-ACQUIRE ordering */
+	smp_acquire__after_ctrl_dep();
+	return n;
 }
 
 static inline void end_dir_add(struct inode *dir, unsigned n)
-- 
2.26.2

Re: [PATCH] fs/dcache: optimize start_dir_add()

[Dave Chinner]

On Thu, Oct 22, 2020 at 02:16:50PM -0700, Davidlohr Bueso wrote:
> Considering both end_dir_add() and d_alloc_parallel(), the
> dir->i_dir_seq wants acquire/release semantics, therefore
> micro-optimize for ll/sc archs and use finer grained barriers
> to provide (load)-ACQUIRE ordering (L->S + L->L). This comes
> at no additional cost for most of x86, as sane tso models will
> have a nop for smp_rmb/smp_acquire__after_ctrl_dep.
> 
> Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
> ---
> Alternatively I guess we could just use cmpxchg_acquire().

Please us cmpxchg_acquire() so that people who have no clue what the
hell smp_acquire__after_ctrl_dep() means or does have some hope of
understanding of what objects the ordering semantics in the function
actually apply to....

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com

[PATCH v2] fs/dcache: optimize start_dir_add()

[Davidlohr Bueso]

Considering both end_dir_add() and d_alloc_parallel(), the
dir->i_dir_seq wants acquire/release semantics, therefore
micro-optimize for ll/sc archs. Also add READ_ONCE around
the variable mostly for documentation purposes - either
the successful cmpxchg or the pause will avoid the tearing).

Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
---
 fs/dcache.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/fs/dcache.c b/fs/dcache.c
index ea0485861d93..9177f0d08a5a 100644
--- a/fs/dcache.c
+++ b/fs/dcache.c
@@ -2504,8 +2504,8 @@ static inline unsigned start_dir_add(struct inode *dir)
 {

	for (;;) {
-		unsigned n = dir->i_dir_seq;
-		if (!(n & 1) && cmpxchg(&dir->i_dir_seq, n, n + 1) == n)
+		unsigned n = READ_ONCE(dir->i_dir_seq);
+		if (!(n & 1) && cmpxchg_acquire(&dir->i_dir_seq, n, n + 1) == n)
			return n;
		cpu_relax();
	}
--
2.26.2