Mikulas, Thanks for your pointing out. > The problem is that you are using workqueues the wrong way. You submit a > work item to a workqueue and the work item is active until the device is > unloaded. > > If you submit a work item to a workqueue, it is required that the work > item finishes in finite time. Otherwise, it may stall stall other tasks. > The deadlock when I terminate Xserver is caused by this - the nvidia > driver tries to flush system workqueue and it waits for all work items to > terminate - but your work items don't terminate. > > If you need a thread that runs for a long time, you should use > kthread_create, not workqueues (see this > http://people.redhat.com/~mpatocka/patches/kernel/dm-crypt-paralelizace/old-3/dm-crypt-encryption-threads.patch > or this > http://people.redhat.com/~mpatocka/patches/kernel/dm-crypt-paralelizace/old-3/dm-crypt-offload-writes-to-thread.patch > as an example how to use kthreads). But I see no reason why you recommend using a kthread for looping job instead of putting a looping work item into a single-threaded not-system workqueue. For me, they both seem to be working. Is it documented that looping job should not be put into any type of workqueue? You are only mentioning that putting a looping work item in system_wq is the wrong way since nvidia driver flush the workqueue. Akira On 10/4/13 10:38 PM, Mikulas Patocka wrote: > > > On Fri, 4 Oct 2013, Akira Hayakawa wrote: > >> Hi, Mikulas, >> >> I am sorry to say that >> I don't have such machines to reproduce the problem. >> >> But agree with that I am dealing with workqueue subsystem >> in a little bit weird way. >> I should clean them up. >> >> For example, >> free_cache() routine below is >> a deconstructor of the cache metadata >> including all the workqueues. >> >> void free_cache(struct wb_cache *cache) >> { >> cache->on_terminate = true; >> >> /* Kill in-kernel daemons */ >> cancel_work_sync(&cache->sync_work); >> cancel_work_sync(&cache->recorder_work); >> cancel_work_sync(&cache->modulator_work); >> >> cancel_work_sync(&cache->flush_work); >> destroy_workqueue(cache->flush_wq); >> >> cancel_work_sync(&cache->barrier_deadline_work); >> >> cancel_work_sync(&cache->migrate_work); >> destroy_workqueue(cache->migrate_wq); >> free_migration_buffer(cache); >> >> /* Destroy in-core structures */ >> free_ht(cache); >> free_segment_header_array(cache); >> >> free_rambuf_pool(cache); >> } >> >> cancel_work_sync() before destroy_workqueue() >> can probably be removed because destroy_workqueue() first >> flush all the works. >> >> Although I prepares independent workqueue >> for each flush_work and migrate_work >> other four works are queued into the system_wq >> through schedule_work() routine. >> This asymmetricity is not welcome for >> architecture-portable code. >> Dependencies to the subsystem should be minimized. >> In detail, workqueue subsystem is really changing >> about its concurrency support so >> trusting only the single threaded workqueue >> will be a good idea for stability. > > The problem is that you are using workqueues the wrong way. You submit a > work item to a workqueue and the work item is active until the device is > unloaded. > > If you submit a work item to a workqueue, it is required that the work > item finishes in finite time. Otherwise, it may stall stall other tasks. > The deadlock when I terminate Xserver is caused by this - the nvidia > driver tries to flush system workqueue and it waits for all work items to > terminate - but your work items don't terminate. > > If you need a thread that runs for a long time, you should use > kthread_create, not workqueues (see this > http://people.redhat.com/~mpatocka/patches/kernel/dm-crypt-paralelizace/old-3/dm-crypt-encryption-threads.patch > or this > http://people.redhat.com/~mpatocka/patches/kernel/dm-crypt-paralelizace/old-3/dm-crypt-offload-writes-to-thread.patch > as an example how to use kthreads). > > Mikulas > >> To begin with, >> these works are never out of queue >> until the deconstructor is called >> but they are repeating running and sleeping. >> Queuing these kind of works to system_wq >> may be unsupported. >> >> So, >> my strategy is to clean them up in a way that >> 1. all daemons are having their own workqueue >> 2. never use cancel_work_sync() but only calls destroy_workqueue() >> in the deconstructor free_cache() and error handling in resume_cache(). >> >> Could you please run the same test again >> after I fixed these points >> to see whether it is still reproducible? >> >> >>> On 3.11.3 on PA-RISC without preemption, the device unloads (although it >>> takes many seconds and vmstat shows that the machine is idle during this >>> time) >> This behavior is benign but probably should be improved. >> In said free_cache() it first turns `on_terminate` flag to true >> to notify all the daemons that we are shutting down. >> Since the `update_interval` and `sync_interval` are 60 seconds by default >> we must wait for them to finish for a while. >> >> Akira >> -- dm-devel mailing list dm-devel@xxxxxxxxxx https://www.redhat.com/mailman/listinfo/dm-devel
