On Thu, Mar 30, 2017 at 8:25 AM, Michal Hocko <mhocko@xxxxxxxxxx> wrote: > On Wed 29-03-17 16:25:18, Ilya Dryomov wrote: >> On Wed, Mar 29, 2017 at 1:16 PM, Michal Hocko <mhocko@xxxxxxxxxx> wrote: >> > On Wed 29-03-17 13:10:01, Ilya Dryomov wrote: >> >> On Wed, Mar 29, 2017 at 12:55 PM, Michal Hocko <mhocko@xxxxxxxxxx> wrote: >> >> > On Wed 29-03-17 12:41:26, Michal Hocko wrote: >> >> > [...] >> >> >> > ceph_con_workfn >> >> >> > mutex_lock(&con->mutex) # ceph_connection::mutex >> >> >> > try_write >> >> >> > ceph_tcp_connect >> >> >> > sock_create_kern >> >> >> > GFP_KERNEL allocation >> >> >> > allocator recurses into XFS, more I/O is issued >> >> > >> >> > One more note. So what happens if this is a GFP_NOIO request which >> >> > cannot make any progress? Your IO thread is blocked on con->mutex >> >> > as you write below but the above thread cannot proceed as well. So I am >> >> > _really_ not sure this acutally helps. >> >> >> >> This is not the only I/O worker. A ceph cluster typically consists of >> >> at least a few OSDs and can be as large as thousands of OSDs. This is >> >> the reason we are calling sock_create_kern() on the writeback path in >> >> the first place: pre-opening thousands of sockets isn't feasible. >> > >> > Sorry for being dense here but what actually guarantees the forward >> > progress? My current understanding is that the deadlock is caused by >> > con->mutext being held while the allocation cannot make a forward >> > progress. I can imagine this would be possible if the other io flushers >> > depend on this lock. But then NOIO vs. KERNEL allocation doesn't make >> > much difference. What am I missing? >> >> con->mutex is per-ceph_connection, osdc->request_mutex is global and is >> the real problem here because we need both on the submit side, at least >> in 3.18. You are correct that even with GFP_NOIO this code may lock up >> in theory, however I think it's very unlikely in practice. > > No, it would just make such a bug more obscure. The real problem seems > to be that you rely on locks which cannot guarantee a forward progress > in the IO path. And that is a bug IMHO. Just to be clear: the "may lock up" comment above goes for 3.18, which is where these stack traces came from. osdc->request_mutex which stood in the way of other ceph_connection workers is no more. > >> We got rid of osdc->request_mutex in 4.7, so these workers are almost >> independent in newer kernels and should be able to free up memory for >> those blocked on GFP_NOIO retries with their respective con->mutex >> held. Using GFP_KERNEL and thus allowing the recursion is just asking >> for an AA deadlock on con->mutex OTOH, so it does make a difference. > > You keep saying this but so far I haven't heard how the AA deadlock is > possible. Both GFP_KERNEL and GFP_NOIO can stall for an unbounded amount > of time and that would cause you problems AFAIU. Suppose we have an I/O for OSD X, which means it's got to go through ceph_connection X: ceph_con_workfn mutex_lock(&con->mutex) try_write ceph_tcp_connect sock_create_kern GFP_KERNEL allocation Suppose that generates another I/O for OSD X and blocks on it. Well, it's got to go through the same ceph_connection: rbd_queue_workfn ceph_osdc_start_request ceph_con_send mutex_lock(&con->mutex) # deadlock, OSD X worker is knocked out Now if that was a GFP_NOIO allocation, we would simply block in the allocator. The placement algorithm distributes objects across the OSDs in a pseudo-random fashion, so even if we had a whole bunch of I/Os for that OSD, some other I/Os for other OSDs would complete in the meantime and free up memory. If we are under the kind of memory pressure that makes GFP_NOIO allocations block for an extended period of time, we are bound to have a lot of pre-open sockets, as we would have done at least some flushing by then. Thanks, Ilya