| Message ID | 20180523025422.32572-1-anand.jain@oracle.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Wed, May 23, 2018 at 10:54:22AM +0800, Anand Jain wrote: > btrfs_free_extra_devids() is called only in the mount context which > traverses through the fs_devices::devices and frees the orphan devices > devices in the given %fs_devices if any. As the search for the orphan > device is limited to fs_devices::devices so we don't need the global > uuid_mutex. > > There can't be any mount-point based ioctl threads in this context as > the mount thread is not yet returned. But there can be the btrfs-control > based scan ioctls thread which calls device_list_add(). > > Here in the mount thread the fs_devices::opened is incremented way before > btrfs_free_extra_devids() is called and in the scan context the fs_devices > which are already opened neither be freed or alloc-able at > device_list_add(). > > But lets say you change the device-path and call the scan again, then scan > would update the new device path and this operation could race against the > btrfs_free_extra_devids() thread, which might be in the process of > free-ing the same device. So synchronize it by using the > device_list_mutex. > > This scenario is a very corner case, and practically the scan and mount > are anyway serialized by the usage so unless the race is instrumented its > very difficult to achieve. > > Signed-off-by: Anand Jain <anand.jain@oracle.com> Thanks, this explanation is much better and addresses the questions I have while reading the code. Reviewed-by: David Sterba <dsterba@suse.com> > --- > Currently device_list_add() is very lean on its device_list_mutex usage, > a cleanup and fix is wip. I also have a WIP patch to rewrite device_list_add. There were quite some changes around the device locking so I'd need to refresh it on top of current code first, it's not in a shape to be posted yet. -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
diff --git a/fs/btrfs/volumes.c b/fs/btrfs/volumes.c index b6757b53c297..a9c1f4f7ebd0 100644 --- a/fs/btrfs/volumes.c +++ b/fs/btrfs/volumes.c @@ -925,7 +925,7 @@ void btrfs_free_extra_devids(struct btrfs_fs_devices *fs_devices, int step) struct btrfs_device *device, *next; struct btrfs_device *latest_dev = NULL; - mutex_lock(&uuid_mutex); + mutex_lock(&fs_devices->device_list_mutex); again: /* This is the initialized path, it is safe to release the devices. */ list_for_each_entry_safe(device, next, &fs_devices->devices, dev_list) { @@ -979,8 +979,7 @@ void btrfs_free_extra_devids(struct btrfs_fs_devices *fs_devices, int step) } fs_devices->latest_bdev = latest_dev->bdev; - - mutex_unlock(&uuid_mutex); + mutex_unlock(&fs_devices->device_list_mutex); } static void free_device_rcu(struct rcu_head *head)
btrfs_free_extra_devids() is called only in the mount context which traverses through the fs_devices::devices and frees the orphan devices devices in the given %fs_devices if any. As the search for the orphan device is limited to fs_devices::devices so we don't need the global uuid_mutex. There can't be any mount-point based ioctl threads in this context as the mount thread is not yet returned. But there can be the btrfs-control based scan ioctls thread which calls device_list_add(). Here in the mount thread the fs_devices::opened is incremented way before btrfs_free_extra_devids() is called and in the scan context the fs_devices which are already opened neither be freed or alloc-able at device_list_add(). But lets say you change the device-path and call the scan again, then scan would update the new device path and this operation could race against the btrfs_free_extra_devids() thread, which might be in the process of free-ing the same device. So synchronize it by using the device_list_mutex. This scenario is a very corner case, and practically the scan and mount are anyway serialized by the usage so unless the race is instrumented its very difficult to achieve. Signed-off-by: Anand Jain <anand.jain@oracle.com> --- Currently device_list_add() is very lean on its device_list_mutex usage, a cleanup and fix is wip. Given the practicality of the above race condition this patch is good to merge. fs/btrfs/volumes.c | 5 ++--- 1 file changed, 2 insertions(+), 3 deletions(-)