<html xmlns:v="urn:schemas-microsoft-com:vml" xmlns:o="urn:schemas-microsoft-com:office:office" xmlns:w="urn:schemas-microsoft-com:office:word" xmlns:x="urn:schemas-microsoft-com:office:excel" xmlns:m="http://schemas.microsoft.com/office/2004/12/omml" xmlns="http://www.w3.org/TR/REC-html40"> <head> <meta http-equiv="Content-Type" content="text/html; charset=utf-8"> <meta name="Generator" content="Microsoft Word 15 (filtered medium)"> <style></style> </head> <body bgcolor="white" lang="ZH-CN" link="#0563C1" vlink="#954F72"> <div class="WordSection1"> 1. Drm_sched_entity_fini(): it exit right after entity->job_queue empty, [ but that time scheduler is not fast enough to deal with this entity now ] <o:p></o:p> That should never happen. The last job from the entity->job_queue is only removed after the scheduler is done with the entity (at least that was the original idea, not sure if that still works as expected).<o:p></o:p> <o:p> </o:p> [ML] no that’s not true and we already catch the kernel NULL pointer issue with a entity->last_scheduled fence get double put , exactly like the way I described in the scenario …<o:p></o:p> <o:p> </o:p> Pixel already fixed it by moving the put/get pair on entity->last_scheduled prior to spsc_queue_pop() and the race issue is therefore avoided <o:p></o:p> <o:p> </o:p> /Monk<o:p></o:p> <o:p> </o:p> <div> <div style="border:none;border-top:solid #E1E1E1 1.0pt;padding:3.0pt 0cm 0cm 0cm"> From: Christian König [mailto:ckoenig.leichtzumerken@gmail.com] Sent: 2018年4月18日 16:36 To: Liu, Monk <Monk.Liu@amd.com>; Koenig, Christian <Christian.Koenig@amd.com>; Deng, Emily <Emily.Deng@amd.com> Cc: amd-gfx@lists.freedesktop.org Subject: Re: force app kill patch<o:p></o:p> </div> </div> <o:p> </o:p> <div> <blockquote style="margin-top:5.0pt;margin-bottom:5.0pt"> See that in “sched_entity_fini”, we only call dma_fence_put(entity->last_scheduled” under the condition of “If (entity->fini_status)”, so <o:p></o:p> This way there is memory leak for the case of “entity->fini_stats ==0”<o:p></o:p> </blockquote> Good catch, we indeed should fix that. <o:p></o:p> <blockquote style="margin-top:5.0pt;margin-bottom:5.0pt"> 1. Drm_sched_entity_fini(): it exit right after entity->job_queue empty, [ but that time scheduler is not fast enough to deal with this entity now ] <o:p></o:p> </blockquote> That should never happen. The last job from the entity->job_queue is only removed after the scheduler is done with the entity (at least that was the original idea, not sure if that still works as expected). Regards, Christian. Am 18.04.2018 um 09:20 schrieb Liu, Monk:<o:p></o:p> </div> <blockquote style="margin-top:5.0pt;margin-bottom:5.0pt"> *Correctio for the scenario *<o:p></o:p> <o:p></o:p> After we move fence_put(entity->last_sched) out of the fini_status check:<o:p></o:p> <o:p></o:p> <o:p></o:p> A potential race issue for the scenario:<o:p></o:p> <o:p></o:p> <![if !supportLists]>1. <![endif]>Drm_sched_entity_fini(): it exit right after entity->job_queue empty, [ but that time scheduler is not fast enough to deal with this entity now ]<o:p></o:p> <![if !supportLists]>2. <![endif]>Drm_sched_entity_cleanup() : it call dma_fence_put(entity->last_scheduled) [ but this time entity->last_scheduled actually points to the fence prior to the real last one ]<o:p></o:p> <![if !supportLists]>3. <![endif]>Scheduler_main() now dealing with this entity: it call dma_fence_put(entity->last_scheduled) [ Now this fence get double put !!! ]<o:p></o:p> <![if !supportLists]>4. <![endif]>Scheduler_main() now call dma_fence_get() on the *real* last one !<o:p></o:p> <o:p></o:p> So eventually the real last one fence triggers memory leak and more critical the double put fence cause NULL pointer access <o:p></o:p> <o:p></o:p> /Monk<o:p></o:p> <o:p></o:p> <div> <div style="border:none;border-top:solid #E1E1E1 1.0pt;padding:3.0pt 0cm 0cm 0cm"> From: Liu, Monk Sent: 2018年4月18日 15:11 To: Koenig, Christian <a href="mailto:Christian.Koenig@amd.com"><Christian.Koenig@amd.com></a>; Deng, Emily <a href="mailto:Emily.Deng@amd.com"><Emily.Deng@amd.com></a> Cc: <a href="mailto:amd-gfx@lists.freedesktop.org">amd-gfx@lists.freedesktop.org</a> Subject: force app kill patch<o:p></o:p> </div> </div> <o:p></o:p> Hi Christian & Emily<o:p></o:p> <o:p></o:p> I think the v4 fix for “fix force app kill hang” is still not good enough:<o:p></o:p> <o:p></o:p> <o:p></o:p> First:<o:p></o:p> See that in “sched_entity_fini”, we only call dma_fence_put(entity->last_scheduled” under the condition of “If (entity->fini_status)”, so <o:p></o:p> This way there is memory leak for the case of “entity->fini_stats ==0”<o:p></o:p> <o:p></o:p> <o:p></o:p> Second:<o:p></o:p> If we move dma_fence_put(entity->last_scheduled) out of the condition of “if (entity->fini_status)”, the memory leak issue can be fixed <o:p></o:p> But there will be kernel NULL pointer access, I think the time you call dma_fence_put(entity->last_scheduled”) may actually executed *not*<o:p></o:p> On the last scheduled fence of this entity, because it is run without “thread_park/unpark” pair which to make sure scheduler not dealing this entity<o:p></o:p> <o:p></o:p> So with certain race issue, here is the scenario:<o:p></o:p> <o:p></o:p> <![if !supportLists]>1. <![endif]>scheduler is doing the dma_fence_put() on the 1st fence, <o:p></o:p> <![if !supportLists]>2. <![endif]>scheduler set entity->last_scheduled to 1st fence<o:p></o:p> <![if !supportLists]>3. <![endif]>now sched_entity_fini() run, and it call dma_fence_put() on entity->last_scheduled <o:p></o:p> <![if !supportLists]>4. <![endif]>now this 1st fence is actually put double time and the real last fence won’t get put by expected <o:p></o:p> <o:p></o:p> <o:p></o:p> any idea?<o:p></o:p> <o:p></o:p> <o:p></o:p> /Monk<o:p></o:p> <o:p></o:p> <pre>_______________________________________________<o:p></o:p></pre> <pre>amd-gfx mailing list<o:p></o:p></pre> <pre><a href="mailto:amd-gfx@lists.freedesktop.org">amd-gfx@lists.freedesktop.org</a><o:p></o:p></pre> <pre><a href="https://lists.freedesktop.org/mailman/listinfo/amd-gfx">https://lists.freedesktop.org/mailman/listinfo/amd-gfx</a><o:p></o:p></pre> </blockquote> <o:p> </o:p> </div> </body> </html>