fuse: limit queued background requests
Miklos Szeredi [Wed, 6 Feb 2008 09:38:39 +0000 (01:38 -0800)]
Libfuse basically creates a new thread for each new request.  This is fine for
synchronous requests, which are naturally limited.  However background
requests (especially writepage) can cause a thread creation storm.

To avoid this, limit the number of background requests available to userspace.

This is done by introducing another queue for background requests, and a
counter for the number of "active" requests, which are currently available for
userspace.

Signed-off-by: Miklos Szeredi <mszeredi@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

fs/fuse/dev.c
fs/fuse/fuse_i.h
fs/fuse/inode.c

index db534bc..af63980 100644 (file)
@@ -201,6 +201,55 @@ void fuse_put_request(struct fuse_conn *fc, struct fuse_req *req)
        }
 }
 
+static unsigned len_args(unsigned numargs, struct fuse_arg *args)
+{
+       unsigned nbytes = 0;
+       unsigned i;
+
+       for (i = 0; i < numargs; i++)
+               nbytes += args[i].size;
+
+       return nbytes;
+}
+
+static u64 fuse_get_unique(struct fuse_conn *fc)
+{
+       fc->reqctr++;
+       /* zero is special */
+       if (fc->reqctr == 0)
+               fc->reqctr = 1;
+
+       return fc->reqctr;
+}
+
+static void queue_request(struct fuse_conn *fc, struct fuse_req *req)
+{
+       req->in.h.unique = fuse_get_unique(fc);
+       req->in.h.len = sizeof(struct fuse_in_header) +
+               len_args(req->in.numargs, (struct fuse_arg *) req->in.args);
+       list_add_tail(&req->list, &fc->pending);
+       req->state = FUSE_REQ_PENDING;
+       if (!req->waiting) {
+               req->waiting = 1;
+               atomic_inc(&fc->num_waiting);
+       }
+       wake_up(&fc->waitq);
+       kill_fasync(&fc->fasync, SIGIO, POLL_IN);
+}
+
+static void flush_bg_queue(struct fuse_conn *fc)
+{
+       while (fc->active_background < FUSE_MAX_BACKGROUND &&
+              !list_empty(&fc->bg_queue)) {
+               struct fuse_req *req;
+
+               req = list_entry(fc->bg_queue.next, struct fuse_req, list);
+               list_del(&req->list);
+               fc->active_background++;
+               queue_request(fc, req);
+       }
+}
+
 /*
  * This function is called when a request is finished.  Either a reply
  * has arrived or it was aborted (and not yet sent) or some error
@@ -229,6 +278,8 @@ static void request_end(struct fuse_conn *fc, struct fuse_req *req)
                        clear_bdi_congested(&fc->bdi, WRITE);
                }
                fc->num_background--;
+               fc->active_background--;
+               flush_bg_queue(fc);
        }
        spin_unlock(&fc->lock);
        wake_up(&req->waitq);
@@ -320,42 +371,6 @@ static void request_wait_answer(struct fuse_conn *fc, struct fuse_req *req)
        }
 }
 
-static unsigned len_args(unsigned numargs, struct fuse_arg *args)
-{
-       unsigned nbytes = 0;
-       unsigned i;
-
-       for (i = 0; i < numargs; i++)
-               nbytes += args[i].size;
-
-       return nbytes;
-}
-
-static u64 fuse_get_unique(struct fuse_conn *fc)
- {
-       fc->reqctr++;
-       /* zero is special */
-       if (fc->reqctr == 0)
-               fc->reqctr = 1;
-
-       return fc->reqctr;
-}
-
-static void queue_request(struct fuse_conn *fc, struct fuse_req *req)
-{
-       req->in.h.unique = fuse_get_unique(fc);
-       req->in.h.len = sizeof(struct fuse_in_header) +
-               len_args(req->in.numargs, (struct fuse_arg *) req->in.args);
-       list_add_tail(&req->list, &fc->pending);
-       req->state = FUSE_REQ_PENDING;
-       if (!req->waiting) {
-               req->waiting = 1;
-               atomic_inc(&fc->num_waiting);
-       }
-       wake_up(&fc->waitq);
-       kill_fasync(&fc->fasync, SIGIO, POLL_IN);
-}
-
 void request_send(struct fuse_conn *fc, struct fuse_req *req)
 {
        req->isreply = 1;
@@ -375,20 +390,26 @@ void request_send(struct fuse_conn *fc, struct fuse_req *req)
        spin_unlock(&fc->lock);
 }
 
+static void request_send_nowait_locked(struct fuse_conn *fc,
+                                      struct fuse_req *req)
+{
+       req->background = 1;
+       fc->num_background++;
+       if (fc->num_background == FUSE_MAX_BACKGROUND)
+               fc->blocked = 1;
+       if (fc->num_background == FUSE_CONGESTION_THRESHOLD) {
+               set_bdi_congested(&fc->bdi, READ);
+               set_bdi_congested(&fc->bdi, WRITE);
+       }
+       list_add_tail(&req->list, &fc->bg_queue);
+       flush_bg_queue(fc);
+}
+
 static void request_send_nowait(struct fuse_conn *fc, struct fuse_req *req)
 {
        spin_lock(&fc->lock);
        if (fc->connected) {
-               req->background = 1;
-               fc->num_background++;
-               if (fc->num_background == FUSE_MAX_BACKGROUND)
-                       fc->blocked = 1;
-               if (fc->num_background == FUSE_CONGESTION_THRESHOLD) {
-                       set_bdi_congested(&fc->bdi, READ);
-                       set_bdi_congested(&fc->bdi, WRITE);
-               }
-
-               queue_request(fc, req);
+               request_send_nowait_locked(fc, req);
                spin_unlock(&fc->lock);
        } else {
                req->out.h.error = -ENOTCONN;
index c5c1ebf..67aaf6e 100644 (file)
@@ -296,6 +296,12 @@ struct fuse_conn {
        /** Number of requests currently in the background */
        unsigned num_background;
 
+       /** Number of background requests currently queued for userspace */
+       unsigned active_background;
+
+       /** The list of background requests set aside for later queuing */
+       struct list_head bg_queue;
+
        /** Pending interrupts */
        struct list_head interrupts;
 
index e5e80d1..c90f633 100644 (file)
@@ -465,6 +465,7 @@ static struct fuse_conn *new_conn(void)
                INIT_LIST_HEAD(&fc->processing);
                INIT_LIST_HEAD(&fc->io);
                INIT_LIST_HEAD(&fc->interrupts);
+               INIT_LIST_HEAD(&fc->bg_queue);
                atomic_set(&fc->num_waiting, 0);
                fc->bdi.ra_pages = (VM_MAX_READAHEAD * 1024) / PAGE_CACHE_SIZE;
                fc->bdi.unplug_io_fn = default_unplug_io_fn;