fs: use wq_has_sleeper() in end_dir_add()

The routine is used a lot, while the wakeup almost never has anyone to
deal with.

wake_up_all() takes an irq-protected spinlock, wq_has_sleeper() "only"
contains a full fence -- not free by any means, but still cheaper.

Sample result tracing waiters using a custom probe during -j 20 kernel
build (0 - no waiters, 1 - waiters):

@[
    wakeprobe+5
    __wake_up_common+63
    __wake_up+54
    __d_add+234
    d_splice_alias+146
    ext4_lookup+439
    path_openat+1746
    do_filp_open+195
    do_sys_openat2+153
    __x64_sys_openat+86
    do_syscall_64+82
    entry_SYSCALL_64_after_hwframe+118
]:
[0, 1)             13999 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@|
[1, ...)               1 |                                                    |

Only 1 call out of 14000 with this backtrace had waiters.

Signed-off-by: Mateusz Guzik <mjguzik@gmail.com>
Link: https://lore.kernel.org/r/20250316232421.1642758-1-mjguzik@gmail.com
Signed-off-by: Christian Brauner <brauner@kernel.org>

diff --git a/fs/dcache.c b/fs/dcache.c
index 9cc0d47da321c52669e05f0e97a6ccfa2451d025..4d395b475ceb2bd7a7034b8deadbbdbec0365446 100644 (file)
--- a/fs/dcache.c
+++ b/fs/dcache.c
@@ -2480,7 +2480,8 @@ static inline void end_dir_add(struct inode *dir, unsigned int n,
 {
        smp_store_release(&dir->i_dir_seq, n + 2);
        preempt_enable_nested();
-       wake_up_all(d_wait);
+       if (wq_has_sleeper(d_wait))
+               wake_up_all(d_wait);
 }
 
 static void d_wait_lookup(struct dentry *dentry)