<html> <head> <meta http-equiv="Content-Type" content="text/html; charset=utf-8"> </head> <body text="#000000" bgcolor="#FFFFFF"> <div class="moz-cite-prefix">On 08/06/2018 04:14 AM, Christian König wrote: </div> <blockquote type="cite" cite="mid:f6df105d-764b-f2a9-6e1a-b30d106a1988@gmail.com"> <meta http-equiv="Content-Type" content="text/html; charset=utf-8"> <div class="moz-cite-prefix">Am 03.08.2018 um 20:41 schrieb Andrey Grodzovsky: </div> <blockquote type="cite" cite="mid:0bb326cb-7753-c19a-b4a9-af3268095dd9@amd.com"> <div class="moz-cite-prefix">On 08/06/2018 08:44 AM, Christian König wrote: </div> <blockquote type="cite" cite="mid:40743441-ccbb-335d-a086-d8057aecc9e6@amd.com"> <div class="moz-cite-prefix">Am 03.08.2018 um 16:54 schrieb Andrey Grodzovsky: </div> <blockquote type="cite" cite="mid:e1cd09f1-287e-ab42-243a-133e67a4f3b2@amd.com"> [SNIP] <blockquote type="cite" cite="mid:0a7c28f9-e60e-74de-1c48-0089935bc4f7@amd.com"> <blockquote type="cite" cite="mid:5d3ec8da-db2e-0aa2-0a7c-a55f27b27a04@amd.com"> <blockquote type="cite" cite="mid:CAFd4ddwKJvfrTu5uAub1K517DhdcOZbZ8k=gyzPy6GVqx_PhtA@mail.gmail.com"> <div dir="ltr"> <div class="gmail_quote"> <blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"> > > Second of all, even after we removed the entity from rq in > drm_sched_entity_flush to terminate any subsequent submissions > > to the queue the other thread doing push job can just acquire again a > run queue > > from drm_sched_entity_get_free_sched and continue submission </blockquote> <div>Hi Christian </div> <blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"> That is actually desired. When another process is now using the entity to submit jobs adding it back to the rq is actually the right thing to do cause the entity is still in use. </blockquote> </div> </div> </blockquote> Yes, no problem if it's another process. But what about another thread from same process ? Is it a possible use case that 2 threads from same process submit to same entity concurrently ? If so and we specifically kill one, the other will not stop event if we want him to because current code makes him just require a rq for him self. </blockquote> Well you can't kill a single thread of a process (you can only interrupt it), a SIGKILL will always kill the whole process. </blockquote> Is the following scenario possible and acceptable ? 2 threads from same process working on same queue where thread A currently in drm_sched_entity_flush->wait_event_timeout (the process getting shut down because of SIGKILL sent by user) while thread B still inside drm_sched_entity_push_job before 'if (reschedule)'. 'A' stopped waiting because queue became empty and then removes the entity queue from scheduler's run queue while B goes inside 'reschedule' because it evaluates to true ('first' is true and all the rest of the conditions), acquires new rq, and later adds it back to scheduler (different one maybe) and keeps submitting jobs as much as he likes and then can be stack for up to 'timeout' time in his drm_sched_entity_flush waiting for them. </blockquote> I'm not 100% sure but I don't think that can happen. See flushing the fd is done while dropping the fd, which happens only after all threads of the process in question are killed. </blockquote> Yea, this FDs handling is indeed a lot of gray area for me but as far as I remember flushing is done per each thread when exits (possibly due to a signal). Now signals interception and processing (as a result of which .flush will get called if SIGKILL received) is done in some points amongst which is when returning from IOCTL. So if first thread was at the very end of the CS ioctl when SIGKILL was received while the other one at the beginning then I think we might see something like the scenario above. </blockquote> SIGKILL isn't processed as long as any thread of the application is still inside the kernel. That's why we have wait_event_killable(). </blockquote> Can you tell me where is this happening ? What i see is in the code is that do_group_exit calls zap_other_threads which just adds SIGKILL to signal sets of other threads in group and sends a wake up. Then do_exit will close all FDs for current thread and so .flush will be called, when last thread drops it's refcount for the FD .release will be called. Andrey <blockquote type="cite" cite="mid:f6df105d-764b-f2a9-6e1a-b30d106a1988@gmail.com"> So I don't think that the scenario above is possible, but I'm really not 100% sure either. Christian. <blockquote type="cite" cite="mid:0bb326cb-7753-c19a-b4a9-af3268095dd9@amd.com"> Andrey <blockquote type="cite" cite="mid:40743441-ccbb-335d-a086-d8057aecc9e6@amd.com"> Otherwise the flushing wouldn't make to much sense. In other words imagine an application where a thread does a write() on a fd which is killed. The idea of the flush is to preserve the data and that won't work if that isn't correctly ordered. <blockquote type="cite" cite="mid:e1cd09f1-287e-ab42-243a-133e67a4f3b2@amd.com"> My understanding was that introduction of entity->last is to force immediate termination job submissions by any thread from the terminating process. </blockquote> We could consider reordering that once more. Going to play out all scenarios in my head over the weekend :) Christian. <blockquote type="cite" cite="mid:e1cd09f1-287e-ab42-243a-133e67a4f3b2@amd.com"> Andrey </blockquote> </blockquote> <fieldset class="mimeAttachmentHeader"></fieldset> <pre wrap="">_______________________________________________ dri-devel mailing list <a class="moz-txt-link-abbreviated" href="mailto:dri-devel@lists.freedesktop.org" moz-do-not-send="true">dri-devel@lists.freedesktop.org</a> <a class="moz-txt-link-freetext" href="https://lists.freedesktop.org/mailman/listinfo/dri-devel" moz-do-not-send="true">https://lists.freedesktop.org/mailman/listinfo/dri-devel</a> </pre> </blockquote> </blockquote> </body> </html>