Obvious leak, which cannot detected unless you try to unload
fuse module. Found accidentally searching for reasons why
fuse_dev_find_request take 7% of cpu in profiles.
Also, make it in more optimal way, fget() is too expensive and
well seen on profiles.
NOTE: this place also needs rcufication. I did that, it is simple
but patch is too intrusive and suprizingly does not show too
much of win. It looks like cache bouncing between krpc thread
and client thread sending request is dominating rather
than spinlock contention/bouncing.
Signed-off-by: Alexey Kuznetsov <kuz...@virtuozzo.com>
---
fs/fuse/dev.c | 20 ++++++++++++++++----
1 file changed, 16 insertions(+), 4 deletions(-)
diff --git a/fs/fuse/dev.c b/fs/fuse/dev.c
index f5594e4..5967248 100644
--- a/fs/fuse/dev.c
+++ b/fs/fuse/dev.c
@@ -2678,16 +2678,28 @@ int fuse_dev_release(struct inode *inode, struct file
*file)
struct fuse_req *fuse_dev_find_request(int fd, u64 unique)
{
- struct file *f = fget(fd);
- struct fuse_dev *fud = fuse_get_dev(f);
- struct fuse_pqueue *fpq = &fud->pq;
+ struct file * file;
+ struct fuse_dev *fud;
+ struct fuse_pqueue *fpq;
struct fuse_req *req = NULL;
spin_lock(&fpq->lock);
+ file = files_lookup_fd_rcu(current->files, fd);
+ if (!file)
+ return NULL;
+
+ if (file->f_op != &fuse_dev_operations)
+ return NULL;
+
+ fud = fuse_get_dev(file);
+ fpq = &fud->pq;
+
if (fpq->connected)
req = request_find(&fud->pq, unique);
- spin_unlock(&fpq->lock);
+ if (req && hlist_unhashed_lockless(&req->hlist))
+ req = NULL;
+ spin_lock(&fpq->lock);
return req;
}
EXPORT_SYMBOL_GPL(fuse_dev_find_request);
--
1.8.3.1
_______________________________________________
Devel mailing list
Devel@openvz.org
https://lists.openvz.org/mailman/listinfo/devel
