In do_dentry_open() the usage is:
f->f_op = fops_get(inode->i_fop);
In generated asm the compiler emits 2 reads from inode->i_fop instead of
just one.
This popped up due to false-sharing where loads from that offset end up
bouncing a cacheline during parallel open. While this is going to be fixed,
the spurious load does not need to be there.
This makes do_dentry_open() go down from 1177 to 1154 bytes.
fops_put() is patched to maintain some consistency.
No functional changes.
Reviewed-by: Josef Bacik <josef@xxxxxxxxxxxxxx>
Signed-off-by: Mateusz Guzik <mjguzik@xxxxxxxxx>
---
This is the same as v1 except for the commit message, which on second
look might have failed to convey what's up.
That said please replace the patch, thanks and sorry for the churn :)
include/linux/fs.h | 15 +++++++++++----
1 file changed, 11 insertions(+), 4 deletions(-)
diff --git a/include/linux/fs.h b/include/linux/fs.h
index ef5ada9d5e33..87d191798454 100644
--- a/include/linux/fs.h
+++ b/include/linux/fs.h
@@ -2565,10 +2565,17 @@ struct super_block *sget(struct file_system_type *type,
struct super_block *sget_dev(struct fs_context *fc, dev_t dev);
/* Alas, no aliases. Too much hassle with bringing module.h everywhere */
-#define fops_get(fops) \
- (((fops) && try_module_get((fops)->owner) ? (fops) : NULL))
-#define fops_put(fops) \
- do { if (fops) module_put((fops)->owner); } while(0)
+#define fops_get(fops) ({ \
+ const struct file_operations *_fops = (fops); \
+ (((_fops) && try_module_get((_fops)->owner) ? (_fops) : NULL)); \
+})
+
+#define fops_put(fops) ({ \
+ const struct file_operations *_fops = (fops); \
+ if (_fops) \
+ module_put((_fops)->owner); \
+})
+
/*
* This one is to be used *ONLY* from ->open() instances.
* fops must be non-NULL, pinned down *and* module dependencies
--
2.43.0
