rbd: fix rbd map vs notify races
authorIlya Dryomov <idryomov@gmail.com>
Fri, 15 Apr 2016 14:22:16 +0000 (16:22 +0200)
committerIlya Dryomov <idryomov@gmail.com>
Thu, 28 Apr 2016 08:07:22 +0000 (10:07 +0200)
A while ago, commit 9875201e1049 ("rbd: fix use-after free of
rbd_dev->disk") fixed rbd unmap vs notify race by introducing
an exported wrapper for flushing notifies and sticking it into
do_rbd_remove().

A similar problem exists on the rbd map path, though: the watch is
registered in rbd_dev_image_probe(), while the disk is set up quite
a few steps later, in rbd_dev_device_setup().  Nothing prevents
a notify from coming in and crashing on a NULL rbd_dev->disk:

    BUG: unable to handle kernel NULL pointer dereference at 0000000000000050
    Call Trace:
     [<ffffffffa0508344>] rbd_watch_cb+0x34/0x180 [rbd]
     [<ffffffffa04bd290>] do_event_work+0x40/0xb0 [libceph]
     [<ffffffff8109d5db>] process_one_work+0x17b/0x470
     [<ffffffff8109e3ab>] worker_thread+0x11b/0x400
     [<ffffffff8109e290>] ? rescuer_thread+0x400/0x400
     [<ffffffff810a5acf>] kthread+0xcf/0xe0
     [<ffffffff810b41b3>] ? finish_task_switch+0x53/0x170
     [<ffffffff810a5a00>] ? kthread_create_on_node+0x140/0x140
     [<ffffffff81645dd8>] ret_from_fork+0x58/0x90
     [<ffffffff810a5a00>] ? kthread_create_on_node+0x140/0x140
    RIP  [<ffffffffa050828a>] rbd_dev_refresh+0xfa/0x180 [rbd]

If an error occurs during rbd map, we have to error out, potentially
tearing down a watch.  Just like on rbd unmap, notifies have to be
flushed, otherwise rbd_watch_cb() may end up trying to read in the
image header after rbd_dev_image_release() has run:

    Assertion failure in rbd_dev_header_info() at line 4722:

     rbd_assert(rbd_image_format_valid(rbd_dev->image_format));

    Call Trace:
     [<ffffffff81cccee0>] ? rbd_parent_request_create+0x150/0x150
     [<ffffffff81cd4e59>] rbd_dev_refresh+0x59/0x390
     [<ffffffff81cd5229>] rbd_watch_cb+0x69/0x290
     [<ffffffff81fde9bf>] do_event_work+0x10f/0x1c0
     [<ffffffff81107799>] process_one_work+0x689/0x1a80
     [<ffffffff811076f7>] ? process_one_work+0x5e7/0x1a80
     [<ffffffff81132065>] ? finish_task_switch+0x225/0x640
     [<ffffffff81107110>] ? pwq_dec_nr_in_flight+0x2b0/0x2b0
     [<ffffffff81108c69>] worker_thread+0xd9/0x1320
     [<ffffffff81108b90>] ? process_one_work+0x1a80/0x1a80
     [<ffffffff8111b02d>] kthread+0x21d/0x2e0
     [<ffffffff8111ae10>] ? kthread_stop+0x550/0x550
     [<ffffffff82022802>] ret_from_fork+0x22/0x40
     [<ffffffff8111ae10>] ? kthread_stop+0x550/0x550
    RIP  [<ffffffff81ccd8f9>] rbd_dev_header_info+0xa19/0x1e30

To fix this, a) check if RBD_DEV_FLAG_EXISTS is set before calling
revalidate_disk(), b) move ceph_osdc_flush_notifies() call into
rbd_dev_header_unwatch_sync() to cover rbd map error paths and c) turn
header read-in into a critical section.  The latter also happens to
take care of rbd map foo@bar vs rbd snap rm foo@bar race.

Fixes: http://tracker.ceph.com/issues/15490

Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Josh Durgin <jdurgin@redhat.com>
drivers/block/rbd.c

index 94a1843..25d22da 100644 (file)
@@ -538,7 +538,6 @@ static int _rbd_dev_v2_snap_size(struct rbd_device *rbd_dev, u64 snap_id,
                                u8 *order, u64 *snap_size);
 static int _rbd_dev_v2_snap_features(struct rbd_device *rbd_dev, u64 snap_id,
                u64 *snap_features);
-static u64 rbd_snap_id_by_name(struct rbd_device *rbd_dev, const char *name);
 
 static int rbd_open(struct block_device *bdev, fmode_t mode)
 {
@@ -3127,9 +3126,6 @@ static void rbd_watch_cb(u64 ver, u64 notify_id, u8 opcode, void *data)
        struct rbd_device *rbd_dev = (struct rbd_device *)data;
        int ret;
 
-       if (!rbd_dev)
-               return;
-
        dout("%s: \"%s\" notify_id %llu opcode %u\n", __func__,
                rbd_dev->header_name, (unsigned long long)notify_id,
                (unsigned int)opcode);
@@ -3263,6 +3259,9 @@ static void rbd_dev_header_unwatch_sync(struct rbd_device *rbd_dev)
 
        ceph_osdc_cancel_event(rbd_dev->watch_event);
        rbd_dev->watch_event = NULL;
+
+       dout("%s flushing notifies\n", __func__);
+       ceph_osdc_flush_notifies(&rbd_dev->rbd_client->client->osdc);
 }
 
 /*
@@ -3642,21 +3641,14 @@ static void rbd_exists_validate(struct rbd_device *rbd_dev)
 static void rbd_dev_update_size(struct rbd_device *rbd_dev)
 {
        sector_t size;
-       bool removing;
 
        /*
-        * Don't hold the lock while doing disk operations,
-        * or lock ordering will conflict with the bdev mutex via:
-        * rbd_add() -> blkdev_get() -> rbd_open()
+        * If EXISTS is not set, rbd_dev->disk may be NULL, so don't
+        * try to update its size.  If REMOVING is set, updating size
+        * is just useless work since the device can't be opened.
         */
-       spin_lock_irq(&rbd_dev->lock);
-       removing = test_bit(RBD_DEV_FLAG_REMOVING, &rbd_dev->flags);
-       spin_unlock_irq(&rbd_dev->lock);
-       /*
-        * If the device is being removed, rbd_dev->disk has
-        * been destroyed, so don't try to update its size
-        */
-       if (!removing) {
+       if (test_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags) &&
+           !test_bit(RBD_DEV_FLAG_REMOVING, &rbd_dev->flags)) {
                size = (sector_t)rbd_dev->mapping.size / SECTOR_SIZE;
                dout("setting size to %llu sectors", (unsigned long long)size);
                set_capacity(rbd_dev->disk, size);
@@ -5187,6 +5179,10 @@ out_err:
        return ret;
 }
 
+/*
+ * rbd_dev->header_rwsem must be locked for write and will be unlocked
+ * upon return.
+ */
 static int rbd_dev_device_setup(struct rbd_device *rbd_dev)
 {
        int ret;
@@ -5195,7 +5191,7 @@ static int rbd_dev_device_setup(struct rbd_device *rbd_dev)
 
        ret = rbd_dev_id_get(rbd_dev);
        if (ret)
-               return ret;
+               goto err_out_unlock;
 
        BUILD_BUG_ON(DEV_NAME_LEN
                        < sizeof (RBD_DRV_NAME) + MAX_INT_FORMAT_WIDTH);
@@ -5236,8 +5232,9 @@ static int rbd_dev_device_setup(struct rbd_device *rbd_dev)
        /* Everything's ready.  Announce the disk to the world. */
 
        set_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags);
-       add_disk(rbd_dev->disk);
+       up_write(&rbd_dev->header_rwsem);
 
+       add_disk(rbd_dev->disk);
        pr_info("%s: added with size 0x%llx\n", rbd_dev->disk->disk_name,
                (unsigned long long) rbd_dev->mapping.size);
 
@@ -5252,6 +5249,8 @@ err_out_blkdev:
                unregister_blkdev(rbd_dev->major, rbd_dev->name);
 err_out_id:
        rbd_dev_id_put(rbd_dev);
+err_out_unlock:
+       up_write(&rbd_dev->header_rwsem);
        return ret;
 }
 
@@ -5442,6 +5441,7 @@ static ssize_t do_rbd_add(struct bus_type *bus,
        spec = NULL;            /* rbd_dev now owns this */
        rbd_opts = NULL;        /* rbd_dev now owns this */
 
+       down_write(&rbd_dev->header_rwsem);
        rc = rbd_dev_image_probe(rbd_dev, 0);
        if (rc < 0)
                goto err_out_rbd_dev;
@@ -5471,6 +5471,7 @@ out:
        return rc;
 
 err_out_rbd_dev:
+       up_write(&rbd_dev->header_rwsem);
        rbd_dev_destroy(rbd_dev);
 err_out_client:
        rbd_put_client(rbdc);
@@ -5577,12 +5578,6 @@ static ssize_t do_rbd_remove(struct bus_type *bus,
                return ret;
 
        rbd_dev_header_unwatch_sync(rbd_dev);
-       /*
-        * flush remaining watch callbacks - these must be complete
-        * before the osd_client is shutdown
-        */
-       dout("%s: flushing notifies", __func__);
-       ceph_osdc_flush_notifies(&rbd_dev->rbd_client->client->osdc);
 
        /*
         * Don't free anything from rbd_dev->disk until after all