Dmitry Monakhov <dmonakhov@xxxxxxxxxx> writes:
> ->s_next_generation is protected by s_next_gen_lock but it usage
> pattern is very primitive and can be replaced with atomic_ops
>
> This significantly improve creation/unlink scenario on SMP systems,
> for example lat_fs_create_unlink test [1] on x2 E5-2680 (32vcpu) system
> shows ~20% improvement.
> | nr_tsk | wo/ patch | w/ patch |
> |--------+-----------+----------|
> | 1 | 137 | 140 |
> | 2 | 224 | 233 |
> | 4 | 356 | 372 |
> | 8 | 439 | 519 |
> | 16 | 443 | 585 |
> | 32 | 598 | 695 |
> | 64 | 559 | 707 |
> | 128 | 385 | 437 |
FYI with lazytime enabled lat_fs_create_unlink is ~16x times slower.
The reason is quite obvious ext4_update_other_inodes_time() increase
lock contention for inode_hash_lock (4k/256) times.
->ext4_do_update_inode
->ext4_update_other_inodes_time
for (i = 0; i < inodes_per_block; i++, ino++, buf += inode_size)
->find_inode_nowait
->spin_lock(&inode_hash_lock) -> 16x contention increase
inode_hash_lock is known problem. I have patches to convert inode_hash_table
per bucket lock similar to dentry_hash, but this require massige changes in
various filesystems so will require a lot of time to be merged.
Currently lazytime amplify it significantly. May be it is reasonable to
use spin_trylock inside find_inode_nowait to make it true lightweight hint?
diff --git a/fs/inode.c b/fs/inode.c
index d1e35b5..a5b1cba1 100644
--- a/fs/inode.c
+++ b/fs/inode.c
@@ -1360,7 +1360,9 @@ struct inode *find_inode_nowait(struct super_block *sb,
struct inode *inode, *ret_inode = NULL;
int mval;
- spin_lock(&inode_hash_lock);
+ if (!spin_trylock(&inode_hash_lock))
+ return NULL;
+
hlist_for_each_entry(inode, head, i_hash) {
if (inode->i_sb != sb)
continue;
