virtiofs: Do not end request in submission context
authorVivek Goyal <vgoyal@redhat.com>
Tue, 15 Oct 2019 17:46:22 +0000 (13:46 -0400)
committerMiklos Szeredi <mszeredi@redhat.com>
Mon, 21 Oct 2019 13:57:07 +0000 (15:57 +0200)
Submission context can hold some locks which end request code tries to hold
again and deadlock can occur. For example, fc->bg_lock. If a background
request is being submitted, it might hold fc->bg_lock and if we could not
submit request (because device went away) and tried to end request, then
deadlock happens. During testing, I also got a warning from deadlock
detection code.

So put requests on a list and end requests from a worker thread.

I got following warning from deadlock detector.

[  603.137138] WARNING: possible recursive locking detected
[  603.137142] --------------------------------------------
[  603.137144] blogbench/2036 is trying to acquire lock:
[  603.137149] 00000000f0f51107 (&(&fc->bg_lock)->rlock){+.+.}, at: fuse_request_end+0xdf/0x1c0 [fuse]
[  603.140701]
[  603.140701] but task is already holding lock:
[  603.140703] 00000000f0f51107 (&(&fc->bg_lock)->rlock){+.+.}, at: fuse_simple_background+0x92/0x1d0 [fuse]
[  603.140713]
[  603.140713] other info that might help us debug this:
[  603.140714]  Possible unsafe locking scenario:
[  603.140714]
[  603.140715]        CPU0
[  603.140716]        ----
[  603.140716]   lock(&(&fc->bg_lock)->rlock);
[  603.140718]   lock(&(&fc->bg_lock)->rlock);
[  603.140719]
[  603.140719]  *** DEADLOCK ***

Signed-off-by: Vivek Goyal <vgoyal@redhat.com>
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
fs/fuse/virtio_fs.c

index e22a0c003c3d24628a611277afc65669cc28d7be..7ea58606cc1d3e971e156364be44daf6b3ec6727 100644 (file)
@@ -30,6 +30,7 @@ struct virtio_fs_vq {
        struct virtqueue *vq;     /* protected by ->lock */
        struct work_struct done_work;
        struct list_head queued_reqs;
+       struct list_head end_reqs;      /* End these requests */
        struct delayed_work dispatch_work;
        struct fuse_dev *fud;
        bool connected;
@@ -259,8 +260,27 @@ static void virtio_fs_hiprio_done_work(struct work_struct *work)
        spin_unlock(&fsvq->lock);
 }
 
-static void virtio_fs_dummy_dispatch_work(struct work_struct *work)
+static void virtio_fs_request_dispatch_work(struct work_struct *work)
 {
+       struct fuse_req *req;
+       struct virtio_fs_vq *fsvq = container_of(work, struct virtio_fs_vq,
+                                                dispatch_work.work);
+       struct fuse_conn *fc = fsvq->fud->fc;
+
+       pr_debug("virtio-fs: worker %s called.\n", __func__);
+       while (1) {
+               spin_lock(&fsvq->lock);
+               req = list_first_entry_or_null(&fsvq->end_reqs, struct fuse_req,
+                                              list);
+               if (!req) {
+                       spin_unlock(&fsvq->lock);
+                       return;
+               }
+
+               list_del_init(&req->list);
+               spin_unlock(&fsvq->lock);
+               fuse_request_end(fc, req);
+       }
 }
 
 static void virtio_fs_hiprio_dispatch_work(struct work_struct *work)
@@ -502,6 +522,7 @@ static int virtio_fs_setup_vqs(struct virtio_device *vdev,
        names[VQ_HIPRIO] = fs->vqs[VQ_HIPRIO].name;
        INIT_WORK(&fs->vqs[VQ_HIPRIO].done_work, virtio_fs_hiprio_done_work);
        INIT_LIST_HEAD(&fs->vqs[VQ_HIPRIO].queued_reqs);
+       INIT_LIST_HEAD(&fs->vqs[VQ_HIPRIO].end_reqs);
        INIT_DELAYED_WORK(&fs->vqs[VQ_HIPRIO].dispatch_work,
                        virtio_fs_hiprio_dispatch_work);
        spin_lock_init(&fs->vqs[VQ_HIPRIO].lock);
@@ -511,8 +532,9 @@ static int virtio_fs_setup_vqs(struct virtio_device *vdev,
                spin_lock_init(&fs->vqs[i].lock);
                INIT_WORK(&fs->vqs[i].done_work, virtio_fs_requests_done_work);
                INIT_DELAYED_WORK(&fs->vqs[i].dispatch_work,
-                                       virtio_fs_dummy_dispatch_work);
+                                 virtio_fs_request_dispatch_work);
                INIT_LIST_HEAD(&fs->vqs[i].queued_reqs);
+               INIT_LIST_HEAD(&fs->vqs[i].end_reqs);
                snprintf(fs->vqs[i].name, sizeof(fs->vqs[i].name),
                         "requests.%u", i - VQ_REQUEST);
                callbacks[i] = virtio_fs_vq_done;
@@ -918,6 +940,7 @@ __releases(fiq->lock)
        struct fuse_conn *fc;
        struct fuse_req *req;
        struct fuse_pqueue *fpq;
+       struct virtio_fs_vq *fsvq;
        int ret;
 
        WARN_ON(list_empty(&fiq->pending));
@@ -951,7 +974,8 @@ __releases(fiq->lock)
        smp_mb__after_atomic();
 
 retry:
-       ret = virtio_fs_enqueue_req(&fs->vqs[queue_id], req);
+       fsvq = &fs->vqs[queue_id];
+       ret = virtio_fs_enqueue_req(fsvq, req);
        if (ret < 0) {
                if (ret == -ENOMEM || ret == -ENOSPC) {
                        /* Virtqueue full. Retry submission */
@@ -965,7 +989,12 @@ retry:
                clear_bit(FR_SENT, &req->flags);
                list_del_init(&req->list);
                spin_unlock(&fpq->lock);
-               fuse_request_end(fc, req);
+
+               /* Can't end request in submission context. Use a worker */
+               spin_lock(&fsvq->lock);
+               list_add_tail(&req->list, &fsvq->end_reqs);
+               schedule_delayed_work(&fsvq->dispatch_work, 0);
+               spin_unlock(&fsvq->lock);
                return;
        }
 }