| Message ID | 20200317120422.3406-1-idryomov@gmail.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | rbd: fix some issues around flushing notifies | expand |
On Tue, Mar 17, 2020 at 8:06 AM Ilya Dryomov <idryomov@gmail.com> wrote: > > Hello, > > A recent snapshot-based mirroring experiment exposed a deadlock on > header_rwsem in the error path of rbd_dev_image_probe() (i.e. "rbd > map"). > > Thanks, > > Ilya > > > Ilya Dryomov (3): > rbd: avoid a deadlock on header_rwsem when flushing notifies > rbd: call rbd_dev_unprobe() after unwatching and flushing notifies > rbd: don't test rbd_dev->opts in rbd_dev_image_release() > > drivers/block/rbd.c | 23 ++++++++++++++--------- > 1 file changed, 14 insertions(+), 9 deletions(-) > > -- > 2.19.2 > The "get_snapcontext" call still going to hang (albeit in an interruptible state) if the image has > 510 snapshots, correct? Reviewed-by: Jason Dillaman <dillaman@redhat.com>
On Tue, Mar 17, 2020 at 5:41 PM Jason Dillaman <jdillama@redhat.com> wrote: > > On Tue, Mar 17, 2020 at 8:06 AM Ilya Dryomov <idryomov@gmail.com> wrote: > > > > Hello, > > > > A recent snapshot-based mirroring experiment exposed a deadlock on > > header_rwsem in the error path of rbd_dev_image_probe() (i.e. "rbd > > map"). > > > > Thanks, > > > > Ilya > > > > > > Ilya Dryomov (3): > > rbd: avoid a deadlock on header_rwsem when flushing notifies > > rbd: call rbd_dev_unprobe() after unwatching and flushing notifies > > rbd: don't test rbd_dev->opts in rbd_dev_image_release() > > > > drivers/block/rbd.c | 23 ++++++++++++++--------- > > 1 file changed, 14 insertions(+), 9 deletions(-) > > > > -- > > 2.19.2 > > > > The "get_snapcontext" call still going to hang (albeit in an > interruptible state) if the image has > 510 snapshots, correct? Yes, this has been a limitation of the messenger and its interface since day one. This is a krbd ticket specifically about snapshots, but this limitation affects a lot more than that: https://tracker.ceph.com/issues/12874 It is engraved pretty deeply, but I'm planning to address it while the messenger is opened up for surgery. Thanks, Ilya
Hello, A recent snapshot-based mirroring experiment exposed a deadlock on header_rwsem in the error path of rbd_dev_image_probe() (i.e. "rbd map"). Thanks, Ilya Ilya Dryomov (3): rbd: avoid a deadlock on header_rwsem when flushing notifies rbd: call rbd_dev_unprobe() after unwatching and flushing notifies rbd: don't test rbd_dev->opts in rbd_dev_image_release() drivers/block/rbd.c | 23 ++++++++++++++--------- 1 file changed, 14 insertions(+), 9 deletions(-)