| Message ID | 1425744599-4934-5-git-send-email-jeff.layton@primarydata.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
diff --git a/fs/locks.c b/fs/locks.c index 4347f3dc17cc..22c0785c00ed 100644 --- a/fs/locks.c +++ b/fs/locks.c @@ -223,14 +223,7 @@ locks_get_lock_context(struct inode *inode, int type) * Assign the pointer if it's not already assigned. If it is, then * free the context we just allocated. */ - spin_lock(&inode->i_lock); - if (likely(!inode->i_flctx)) { - inode->i_flctx = new; - new = NULL; - } - spin_unlock(&inode->i_lock); - - if (new) + if (cmpxchg(&inode->i_flctx, NULL, new)) kmem_cache_free(flctx_cache, new); out: return inode->i_flctx;
During the v3.20/v4.0 cycle, I had originally had the code manage the inode->i_flctx pointer using a compare-and-swap operation instead of the i_lock. Sasha Levin though hit a problem while testing trinity that made me believe that that wasn't safe. I now think though that I completely misread the problem, even though it seemed like it went away when we started using the i_lock to protect this pointer. The issue was likely the same race that Kirill Shutemov hit while testing the pre-rc1 v4.0 kernel and that Linus spotted. Due to the way that the spinlock was dropped in the middle of flock_lock_file, you could end up with multiple flock locks for the same struct file on the inode. Reinstate the use of a CAS operation to assign this pointer since it's likely to be more efficient and gets the i_lock completely out of the file locking business. Signed-off-by: Jeff Layton <jeff.layton@primarydata.com> --- fs/locks.c | 9 +-------- 1 file changed, 1 insertion(+), 8 deletions(-)