On Wed, Aug 27, 2014 at 11:36 AM, Mel Gorman <mgorman@xxxxxxxx> wrote: > On Tue, Aug 26, 2014 at 08:00:20PM -0400, Trond Myklebust wrote: >> On Tue, Aug 26, 2014 at 7:51 PM, Trond Myklebust >> <trond.myklebust@xxxxxxxxxxxxxxx> wrote: >> > On Tue, Aug 26, 2014 at 7:19 PM, Johannes Weiner <hannes@xxxxxxxxxxx> wrote: >> >> On Tue, Aug 26, 2014 at 02:26:24PM +0100, Mel Gorman wrote: >> >>> On Tue, Aug 26, 2014 at 08:58:36AM -0400, Trond Myklebust wrote: >> >>> > On Tue, Aug 26, 2014 at 6:53 AM, Mel Gorman <mgorman@xxxxxxxx> wrote: >> >>> > > On Mon, Aug 25, 2014 at 04:48:52PM +1000, NeilBrown wrote: >> >>> > >> On Fri, 22 Aug 2014 18:49:31 -0400 Trond Myklebust >> >>> > >> <trond.myklebust@xxxxxxxxxxxxxxx> wrote: >> >>> > >> >> >>> > >> > Junxiao Bi reports seeing the following deadlock: >> >>> > >> > >> >>> > >> > @ crash> bt 1539 >> >>> > >> > @ PID: 1539 TASK: ffff88178f64a040 CPU: 1 COMMAND: "rpciod/1" >> >>> > >> > @ #0 [ffff88178f64d2c0] schedule at ffffffff8145833a >> >>> > >> > @ #1 [ffff88178f64d348] io_schedule at ffffffff8145842c >> >>> > >> > @ #2 [ffff88178f64d368] sync_page at ffffffff810d8161 >> >>> > >> > @ #3 [ffff88178f64d378] __wait_on_bit at ffffffff8145895b >> >>> > >> > @ #4 [ffff88178f64d3b8] wait_on_page_bit at ffffffff810d82fe >> >>> > >> > @ #5 [ffff88178f64d418] wait_on_page_writeback at ffffffff810e2a1a >> >>> > >> > @ #6 [ffff88178f64d438] shrink_page_list at ffffffff810e34e1 >> >>> > >> > @ #7 [ffff88178f64d588] shrink_list at ffffffff810e3dbe >> >>> > >> > @ #8 [ffff88178f64d6f8] shrink_zone at ffffffff810e425e >> >>> > >> > @ #9 [ffff88178f64d7b8] do_try_to_free_pages at ffffffff810e4978 >> >>> > >> > @ #10 [ffff88178f64d828] try_to_free_pages at ffffffff810e4c31 >> >>> > >> > @ #11 [ffff88178f64d8c8] __alloc_pages_nodemask at ffffffff810de370 >> >>> > >> >> >>> > >> This stack trace (from 2.6.32) cannot happen in mainline, though it took me a >> >>> > >> while to remember/discover exactly why. >> >>> > >> >> >>> > >> try_to_free_pages() creates a 'struct scan_control' with ->target_mem_cgroup >> >>> > >> set to NULL. >> >>> > >> shrink_page_list() checks ->target_mem_cgroup using global_reclaim() and if >> >>> > >> it is NULL, wait_on_page_writeback is *not* called. >> >>> > >> >> >>> > > >> >>> > > wait_on_page_writeback has a host of other damage associated with it which >> >>> > > is why we don't do it from reclaim any more. If the storage is very slow >> >>> > > then a process can be stalled by unrelated IO to slow storage. If the >> >>> > > storage is broken and the writeback can never complete then it causes other >> >>> > > issues. That kind of thing. >> >>> > > >> >>> > >> So we can only hit this deadlock if mem-cgroup limits are imposed on a >> >>> > >> process which is using NFS - which is quite possible but probably not common. >> >>> > >> >> >>> > >> The fact that a dead-lock can happen only when memcg limits are imposed seems >> >>> > >> very fragile. People aren't going to test that case much so there could well >> >>> > >> be other deadlock possibilities lurking. >> >>> > >> >> >>> > > >> >>> > > memcgs still can call wait_on_page_writeback and this is known to be a >> >>> > > hand-grenade to the memcg people but I've never heard of them trying to >> >>> > > tackle the problem. >> >>> > > >> >>> > >> Mel: might there be some other way we could get out of this deadlock? >> >>> > >> Could the wait_on_page_writeback() in shrink_page_list() be made a timed-out >> >>> > >> wait or something? Any other wait out of this deadlock other than setting >> >>> > >> PF_MEMALLOC_NOIO everywhere? >> >>> > >> >> >>> > > >> >>> > > I don't have the full thread as it was not cc'd to lkml so I don't know >> >>> > > what circumstances reached this deadlock in the first place. If this is >> >>> > > on 2.6.32 and the deadline cannot happen during reclaim in mainline then >> >>> > > why is mainline being patched? >> >>> > > >> >>> > > Do not alter wait_on_page_writeback() to timeout as it will blow >> >>> > > up spectacularly -- swap unuse races, data would not longer be synced >> >>> > > correctly to disk, sync IO would be flaky, stable page writes would be >> >>> > > fired out the window etc. >> >>> > >> >>> > Hi Mel, >> >>> > >> >>> > The above stack trace really is the entire deadlock: the rpciod work >> >>> > queue, which drives I/O on behalf of NFS, gets caught in a >> >>> > shrink_page_list() situation where it ends up waiting on page >> >>> > writeback. Boom.... >> >>> > >> >>> > Even if this can only happen for non-trivial memcg situations, then it >> >>> > still needs to be addressed: if rpciod blocks, then all NFS I/O will >> >>> > block and we can no longer write out the dirty pages. This is why we >> >>> > need a mainline fix. >> >>> > >> >>> >> >>> In that case I'm adding the memcg people. I recognise that rpciod should >> >>> never block on writeback for similar reasons why flushers should never block. >> >>> memcg blocking on writeback is dangerous for reasons other than NFS but >> >>> adding a variant that times out just means that on occasion processes get >> >>> stalled for long periods of time timing out on these writeback pages. In >> >>> that case, forward progress of rpciod would be painfully slow. >> >>> >> >>> On the other hand, forcing PF_MEMALLOC_NOIO for all rpciod allocations in >> >>> an ideal world is massive overkill and while it will work, there will be >> >>> other consequences -- unable to swap pages for example, unable to release >> >>> buffers to free clean pages etc. >> >>> >> >>> It'd be nice of the memcg people could comment on whether they plan to >> >>> handle the fact that memcg is the only called of wait_on_page_writeback >> >>> in direct reclaim paths. >> >> >> >> wait_on_page_writeback() is a hammer, and we need to be better about >> >> this once we have per-memcg dirty writeback and throttling, but I >> >> think that really misses the point. Even if memcg writeback waiting >> >> were smarter, any length of time spent waiting for yourself to make >> >> progress is absurd. We just shouldn't be solving deadlock scenarios >> >> through arbitrary timeouts on one side. If you can't wait for IO to >> >> finish, you shouldn't be passing __GFP_IO. >> >> >> >> Can't you use mempools like the other IO paths? >> > >> > There is no way to pass any allocation flags at all to an operation >> > such as __sock_create() (which may be needed if the server >> > disconnects). So in general, the answer is no. >> > >> >> Actually, one question that should probably be raised before anything >> else: is it at all possible for a workqueue like rpciod to have a >> non-trivial setting for ->target_mem_cgroup? If not, then the whole >> question is moot. >> > > AFAIK, today it's not possible to add kernel threads (which rpciod is one) > to a memcg so the issue is entirely theoritical at the moment. Even if > this was to change, it's not clear to me what adding kernel threads to a > memcg would mean as kernel threads have no RSS. Even if kernel resources > were accounted for, I cannot see why a kernel thread would join a memcg. > > I expec that it's currently impossible for rpciod to have a non-trivial > target_mem_cgroup. The memcg folk will correct me if I'm wrong or if there > are plans to change that for some reason. > Thanks! Then I'll assume that the problem is nonexistent in upstream for now, and drop the idea of using PF_MEMALLOC_NOIO. Perhaps we can then encourage Junxiao to look into backporting some of the VM changes in order to fix his Oracle legacy kernel issues? Cheers Trond -- Trond Myklebust Linux NFS client maintainer, PrimaryData trond.myklebust@xxxxxxxxxxxxxxx -- To unsubscribe from this list: send the line "unsubscribe linux-nfs" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
