<html><head>
<meta http-equiv="Content-Type" content="text/html; charset=utf-8">
  </head>
  <body text="#000000" bgcolor="#FFFFFF">
    <div class="moz-cite-prefix">Another more fundamental question:
      Could we get rid of the timeout job at all?<br>
      <br>
      I mean we used to give this as parameter to the scheduler callback
      because we had the timeout worker in the job, but that is no
      longer the case.<br>
      <br>
      E.g. in drm_sched_job_timedout() we do the following:<br>
      <blockquote type="cite">        job =
        list_first_entry_or_null(&sched->ring_mirror_list,<br>
                                               struct drm_sched_job,
        node);<br>
      </blockquote>
      <br>
      Why don't we just remove that here and only get the first job
      after we have stopped the scheduler?<br>
      <br>
      Regards,<br>
      Christian.<br>
      <br>
      Am 13.11.19 um 15:12 schrieb Andrey Grodzovsky:<br>
    </div>
    <blockquote type="cite" cite="mid:40bb3114-d996-10af-3140-51a4f7c212d6@amd.com">
      
      <p>This why I asked for a trace with timer enabled, but since
        there is a finite number of places we touch the timer Emily can
        just put prints there. Also, I wonder if this temp fix helps her
        with the issue or not.</p>
      <p>Andrey<br>
      </p>
      <div class="moz-cite-prefix">On 11/13/19 2:36 AM, Christian König
        wrote:<br>
      </div>
      <blockquote type="cite" cite="mid:33ffe2f1-32b6-a238-3752-cee67cd9e141@gmail.com">
        <div class="moz-cite-prefix">The question is where do we rearm
          the timer for this problem to occur?<br>
          <br>
          Regards,<br>
          Christian. <br>
          <br>
          Am 12.11.19 um 20:21 schrieb Andrey Grodzovsky:<br>
        </div>
        <blockquote type="cite" cite="mid:9269d447-ed32-81f7-ab43-cb16139096e2@amd.com">
          <p>I was able to reproduce the crash by using the attached
            simulate_crash.patch - waiting on guilty job to signal in
            reset work and artificially rearming the timeout timer just
            before the check for
            !cancel_delayed_work(&sched->work_tdr)  in
            drm_sched_cleanup_jobs - crash log attached in crash.log.
            This I think confirms my theory i described earlier in this
            thread.</p>
          <p>basic_fix.patch handles this by testing whether another
            timer already armed ob this scheduler or is there a timeout
            work in execution right now (see documentation for
            work_busy) - obviously  this is not a full solution as this
            will not protect from races if for example there is
            immediate work scheduling such as in <font size="2" face="Calibri"><span style="font-size:11pt;">drm_sched_fault
                -  so we probably need to account for this by making
                drm_sched_cleanup_jobs (at least in the part where it
                iterates ring mirror list and frees jobs) and GPU reset
                really mutually exclusive and not like now.</span></font></p>
          <p><font size="2" face="Calibri"><span style="font-size:11pt;">Andrey<br>
              </span></font></p>
          <p><br>
          </p>
          <div class="moz-cite-prefix">On 11/11/19 4:11 PM, Christian
            König wrote:<br>
          </div>
          <blockquote type="cite" cite="mid:2f035f22-4057-dd9e-27ef-0f5612113e29@amd.com">
            <div class="moz-cite-prefix">Hi Emily,<br>
              <br>
              you need to print which scheduler instance is freeing the
              jobs and which one is triggering the reset. The TID and
              PID is completely meaningless here since we are called
              from different worker threads and the TID/PID can change
              on each call.<br>
              <br>
              Apart from that I will look into this a bit deeper when I
              have time.<br>
              <br>
              Regards,<br>
              Christian.<br>
              <br>
              Am 12.11.19 um 07:02 schrieb Deng, Emily:<br>
            </div>
            <blockquote type="cite" cite="mid:MN2PR12MB29750EDB35E27DF9CD63152C8F770@MN2PR12MB2975.namprd12.prod.outlook.com">
              <meta name="Generator" content="Microsoft Exchange Server">
              <!-- converted from rtf -->
              <style><!-- .EmailQuote { margin-left: 1pt; padding-left: 4pt; border-left: #800000 2px solid; } --></style>
              <font size="2" face="Calibri"><span style="font-size:11pt;">
                  <div>Hi Christian,</div>
                  <div>    I add the follow print in function
                    drm_sched_cleanup_jobs. From the log it shows that
                    only use cancel_delayed_work could not avoid to free
                    job when the sched is in reset. But don’t know
                    exactly where it is wrong about the driver. Do you
                    have any suggestion about this?</div>
                  <div><font face="Times New Roman"> </font></div>
                  <div>+      
                    printk("Emily:drm_sched_cleanup_jobs:begin,tid:%lu,
                    pid:%lu\n", current->tgid, current->pid);</div>
                  <div> </div>
                  <div>        /*</div>
                  <div>         * Don't destroy jobs while the timeout
                    worker is running  OR thread</div>
                  <div>         * is being parked and hence assumed to
                    not touch ring_mirror_list</div>
                  <div>         */</div>
                  <div>         if ((sched->timeout !=
                    MAX_SCHEDULE_TIMEOUT &&</div>
                  <div>           
                    !cancel_delayed_work(&sched->work_tdr)))</div>
                  <div>                return;</div>
                  <div>+      
                    printk("Emily:drm_sched_cleanup_jobs,tid:%lu,
                    pid:%lu\n", current->tgid, current->pid);</div>
                  <div><font face="Times New Roman"> </font></div>
                  <div><font face="Times New Roman"> </font></div>
                  <div><font face="Calibri">Best wishes</font></div>
                  <div><font face="Calibri">Emily Deng</font></div>
                  <div><font face="Times New Roman"> </font></div>
                  <div>Nov 12 12:58:20
                    ubuntu-drop-August-2018-rc2-gpu0-vf02 kernel:
                    [11380.695091]
                    Emily:drm_sched_cleanup_jobs:begin,tid:2262,
                    pid:2262</div>
                  <div>Nov 12 12:58:20
                    ubuntu-drop-August-2018-rc2-gpu0-vf02 kernel:
                    [11380.695104]
                    Emily:drm_sched_cleanup_jobs:begin,tid:2262,
                    pid:2262</div>
                  <div>Nov 12 12:58:20
                    ubuntu-drop-August-2018-rc2-gpu0-vf02 kernel:
                    [11380.695105]
                    Emily:drm_sched_cleanup_jobs,tid:2262, pid:2262</div>
                  <div>Nov 12 12:58:20
                    ubuntu-drop-August-2018-rc2-gpu0-vf02 kernel:
                    [11380.695107]
                    Emily:drm_sched_cleanup_jobs:begin,tid:2262,
                    pid:2262</div>
                  <div>Nov 12 12:58:20
                    ubuntu-drop-August-2018-rc2-gpu0-vf02 kernel:
                    [11380.695107]
                    Emily:drm_sched_cleanup_jobs,tid:2262, pid:2262</div>
                  <div>Nov 12 12:58:20
                    ubuntu-drop-August-2018-rc2-gpu0-vf02 kernel:
                    [11381.222954] [drm:amdgpu_job_timedout [amdgpu]]
                    *ERROR* ring sdma0 timeout, signaled seq=78585,
                    emitted seq=78587</div>
                  <div>Nov 12 12:58:20
                    ubuntu-drop-August-2018-rc2-gpu0-vf02 kernel:
                    [11381.224275] [drm:amdgpu_job_timedout [amdgpu]]
                    *ERROR* Process information: process  pid 0 thread 
                    pid 0, s_job:00000000fe75ab36,tid=15603, pid=15603</div>
                  <div>Nov 12 12:58:20
                    ubuntu-drop-August-2018-rc2-gpu0-vf02 kernel:
                    [11381.225413] amdgpu 0000:00:08.0: <font color="#00B050">GPU reset begin</font>!</div>
                  <div>Nov 12 12:58:20
                    ubuntu-drop-August-2018-rc2-gpu0-vf02 kernel:
                    [11381.225417]
                    Emily:drm_sched_cleanup_jobs:begin,tid:2262,
                    pid:2262</div>
                  <div>Nov 12 12:58:20
                    ubuntu-drop-August-2018-rc2-gpu0-vf02 kernel:
                    [11381.225425]
                    Emily:drm_sched_cleanup_jobs:begin,tid:2262,
                    pid:2262</div>
                  <div>Nov 12 12:58:20
                    ubuntu-drop-August-2018-rc2-gpu0-vf02 kernel:
                    [11381.225425]
                    Emily:drm_sched_cleanup_jobs,tid:2262, pid:2262</div>
                  <div>Nov 12 12:58:20
                    ubuntu-drop-August-2018-rc2-gpu0-vf02 kernel:
                    [11381.225428] Emily:<font color="#00B050">amdgpu_job_free_cb</font>,Process
                    information: process  pid 0 thread  pid 0,
                    s_job:00000000fe75ab36, tid:2262, pid:2262</div>
                  <div>Nov 12 12:58:20
                    ubuntu-drop-August-2018-rc2-gpu0-vf02 kernel:
                    [11381.225429]
                    Emily:drm_sched_cleanup_jobs:begin,tid:2262,
                    pid:2262</div>
                  <div>Nov 12 12:58:20
                    ubuntu-drop-August-2018-rc2-gpu0-vf02 kernel:
                    [11381.225430]
                    Emily:drm_sched_cleanup_jobs,tid:2262, pid:2262</div>
                  <div>Nov 12 12:58:20
                    ubuntu-drop-August-2018-rc2-gpu0-vf02 kernel:
                    [11381.225473]
                    Emily:drm_sched_cleanup_jobs:begin,tid:2253,
                    pid:2253</div>
                  <div>Nov 12 12:58:20
                    ubuntu-drop-August-2018-rc2-gpu0-vf02 kernel:
                    [11381.225486]
                    Emily:drm_sched_cleanup_jobs:begin,tid:2262,
                    pid:2262</div>
                  <div>Nov 12 12:58:20
                    ubuntu-drop-August-2018-rc2-gpu0-vf02 kernel:
                    [11381.225489]
                    Emily:drm_sched_cleanup_jobs,tid:2262, pid:2262</div>
                  <div>Nov 12 12:58:20
                    ubuntu-drop-August-2018-rc2-gpu0-vf02 kernel:
                    [11381.225494] Emily:amdgpu_job_free_cb,Process
                    information: process  pid 0 thread  pid 0,
                    s_job:00000000f086ec84, tid:2262, pid:2262</div>
                  <div>>-----Original Message-----</div>
                  <div>>From: Grodzovsky, Andrey <a class="moz-txt-link-rfc2396E" href="mailto:Andrey.Grodzovsky@amd.com" moz-do-not-send="true"><Andrey.Grodzovsky@amd.com></a></div>
                  <div>>Sent: Tuesday, November 12, 2019 11:28 AM</div>
                  <div>>To: Koenig, Christian <a class="moz-txt-link-rfc2396E" href="mailto:Christian.Koenig@amd.com" moz-do-not-send="true"><Christian.Koenig@amd.com></a>;
                    Deng, Emily</div>
                  <div>><a class="moz-txt-link-rfc2396E" href="mailto:Emily.Deng@amd.com" moz-do-not-send="true"><Emily.Deng@amd.com></a>;
                    <a class="moz-txt-link-abbreviated" href="mailto:amd-gfx@lists.freedesktop.org" moz-do-not-send="true">amd-gfx@lists.freedesktop.org</a></div>
                  <div>>Subject: Re: [PATCH] drm/amdgpu: Fix the null
                    pointer issue for tdr</div>
                  <div>></div>
                  <div>>Thinking more about this claim - we assume
                    here that if cancel_delayed_work</div>
                  <div>>returned true it guarantees that timeout work
                    is not running but, it merely</div>
                  <div>>means there was a pending timeout work which
                    was removed from the</div>
                  <div>>workqueue before it's timer elapsed and so it
                    didn't have a chance to be</div>
                  <div>>dequeued and executed, it doesn't cover
                    already executing work. So there is a</div>
                  <div>>possibility where while timeout work started
                    executing another timeout work</div>
                  <div>>already got enqueued (maybe through earlier
                    cleanup jobs or through</div>
                  <div>>drm_sched_fault) and if at this point another
                    drm_sched_cleanup_jobs runs</div>
                  <div>>cancel_delayed_work(&sched->work_tdr)
                    will return true even while there is a</div>
                  <div>>timeout job in progress.</div>
                  <div>>Unfortunately we cannot change
                    cancel_delayed_work to</div>
                  <div>>cancel_delayed_work_sync to flush the timeout
                    work as timeout work itself</div>
                  <div>>waits for schedule thread  to be parked again
                    when calling park_thread.</div>
                  <div>></div>
                  <div>>Andrey</div>
                  <div>></div>
                  <div>>________________________________________</div>
                  <div>>From: amd-gfx <a class="moz-txt-link-rfc2396E" href="mailto:amd-gfx-bounces@lists.freedesktop.org" moz-do-not-send="true"><amd-gfx-bounces@lists.freedesktop.org></a>
                    on behalf of</div>
                  <div>>Koenig, Christian <a class="moz-txt-link-rfc2396E" href="mailto:Christian.Koenig@amd.com" moz-do-not-send="true"><Christian.Koenig@amd.com></a></div>
                  <div>>Sent: 08 November 2019 05:35:18</div>
                  <div>>To: Deng, Emily; <a class="moz-txt-link-abbreviated" href="mailto:amd-gfx@lists.freedesktop.org" moz-do-not-send="true">amd-gfx@lists.freedesktop.org</a></div>
                  <div>>Subject: Re: [PATCH] drm/amdgpu: Fix the null
                    pointer issue for tdr</div>
                  <div>></div>
                  <div>>Hi Emily,</div>
                  <div>></div>
                  <div>>exactly that can't happen. See here:</div>
                  <div>></div>
                  <div>>>         /* Don't destroy jobs while the
                    timeout worker is running */</div>
                  <div>>>         if (sched->timeout !=
                    MAX_SCHEDULE_TIMEOUT &&</div>
                  <div>>>            
                    !cancel_delayed_work(&sched->work_tdr))</div>
                  <div>>>                 return NULL;</div>
                  <div>></div>
                  <div>>We never free jobs while the timeout working
                    is running to prevent exactly</div>
                  <div>>that issue.</div>
                  <div>></div>
                  <div>>Regards,</div>
                  <div>>Christian.</div>
                  <div>></div>
                  <div>>Am 08.11.19 um 11:32 schrieb Deng, Emily:</div>
                  <div>>> Hi Christian,</div>
                  <div>>>       The drm_sched_job_timedout->
                    amdgpu_job_timedout call</div>
                  <div>>amdgpu_device_gpu_recover. I mean the main
                    scheduler free the jobs while</div>
                  <div>>in amdgpu_device_gpu_recover, and before
                    calling drm_sched_stop.</div>
                  <div>>></div>
                  <div>>> Best wishes</div>
                  <div>>> Emily Deng</div>
                  <div>>></div>
                  <div>>></div>
                  <div>>></div>
                  <div>>>> -----Original Message-----</div>
                  <div>>>> From: Koenig, Christian <a class="moz-txt-link-rfc2396E" href="mailto:Christian.Koenig@amd.com" moz-do-not-send="true"><Christian.Koenig@amd.com></a></div>
                  <div>>>> Sent: Friday, November 8, 2019 6:26
                    PM</div>
                  <div>>>> To: Deng, Emily <a class="moz-txt-link-rfc2396E" href="mailto:Emily.Deng@amd.com" moz-do-not-send="true"><Emily.Deng@amd.com></a>;
                    <a class="moz-txt-link-abbreviated" href="mailto:amd-gfx@lists.freedesktop.org" moz-do-not-send="true">amd-gfx@lists.freedesktop.org</a></div>
                  <div>>>> Subject: Re: [PATCH] drm/amdgpu: Fix
                    the null pointer issue for tdr</div>
                  <div>>>></div>
                  <div>>>> Hi Emily,</div>
                  <div>>>></div>
                  <div>>>> well who is calling
                    amdgpu_device_gpu_recover() in this case?</div>
                  <div>>>></div>
                  <div>>>> When it's not the scheduler we
                    shouldn't have a guilty job in the first place.</div>
                  <div>>>></div>
                  <div>>>> Regards,</div>
                  <div>>>> Christian.</div>
                  <div>>>></div>
                  <div>>>> Am 08.11.19 um 11:22 schrieb Deng,
                    Emily:</div>
                  <div>>>>> Hi Chrisitan,</div>
                  <div>>>>>        No, I am with the new
                    branch and also has the patch. Even it</div>
                  <div>>>>> are freed by</div>
                  <div>>>> main scheduler, how we could avoid
                    main scheduler to free jobs while</div>
                  <div>>>> enter to function
                    amdgpu_device_gpu_recover?</div>
                  <div>>>>> Best wishes</div>
                  <div>>>>> Emily Deng</div>
                  <div>>>>></div>
                  <div>>>>></div>
                  <div>>>>></div>
                  <div>>>>>> -----Original Message-----</div>
                  <div>>>>>> From: Koenig, Christian <a class="moz-txt-link-rfc2396E" href="mailto:Christian.Koenig@amd.com" moz-do-not-send="true"><Christian.Koenig@amd.com></a></div>
                  <div>>>>>> Sent: Friday, November 8,
                    2019 6:15 PM</div>
                  <div>>>>>> To: Deng, Emily <a class="moz-txt-link-rfc2396E" href="mailto:Emily.Deng@amd.com" moz-do-not-send="true"><Emily.Deng@amd.com></a>;
                    amd-</div>
                  <div>><a class="moz-txt-link-abbreviated" href="mailto:gfx@lists.freedesktop.org" moz-do-not-send="true">gfx@lists.freedesktop.org</a></div>
                  <div>>>>>> Subject: Re: [PATCH]
                    drm/amdgpu: Fix the null pointer issue for tdr</div>
                  <div>>>>>></div>
                  <div>>>>>> Hi Emily,</div>
                  <div>>>>>></div>
                  <div>>>>>> in this case you are on an
                    old code branch.</div>
                  <div>>>>>></div>
                  <div>>>>>> Jobs are freed now by the
                    main scheduler thread and only if no</div>
                  <div>>>>>> timeout handler is running.</div>
                  <div>>>>>></div>
                  <div>>>>>> See this patch here:</div>
                  <div>>>>>>> commit
                    5918045c4ed492fb5813f980dcf89a90fefd0a4e</div>
                  <div>>>>>>> Author: Christian König
                    <a class="moz-txt-link-rfc2396E" href="mailto:christian.koenig@amd.com" moz-do-not-send="true"><christian.koenig@amd.com></a></div>
                  <div>>>>>>> Date:   Thu Apr 18
                    11:00:21 2019 -0400</div>
                  <div>>>>>>></div>
                  <div>>>>>>>       drm/scheduler:
                    rework job destruction</div>
                  <div>>>>>> Regards,</div>
                  <div>>>>>> Christian.</div>
                  <div>>>>>></div>
                  <div>>>>>> Am 08.11.19 um 11:11 schrieb
                    Deng, Emily:</div>
                  <div>>>>>>> Hi Christian,</div>
                  <div>>>>>>>         Please refer to
                    follow log, when it enter to</div>
                  <div>>>>>>>
                    amdgpu_device_gpu_recover</div>
                  <div>>>>>> function, the bad job
                    000000005086879e is freeing in function</div>
                  <div>>>>>> amdgpu_job_free_cb  at the
                    same time, because of the hardware fence</div>
                  <div>>>> signal.</div>
                  <div>>>>>> But
                    amdgpu_device_gpu_recover goes faster, at this case,
                    the</div>
                  <div>>>>>> s_fence is already freed,
                    but job is not freed in time. Then this issue</div>
                  <div>>occurs.</div>
                  <div>>>>>>> [  449.792189]
                    [drm:amdgpu_job_timedout [amdgpu]] *ERROR* ring</div>
                  <div>>>> sdma0</div>
                  <div>>>>>>> timeout, signaled
                    seq=2481, emitted seq=2483 [  449.793202]</div>
                  <div>>>>>>> [drm:amdgpu_job_timedout
                    [amdgpu]] *ERROR* Process information:</div>
                  <div>>>>>> process  pid 0 thread  pid
                    0, s_job:000000005086879e [  449.794163]</div>
                  <div>>>>>> amdgpu</div>
                  <div>>>>>> 0000:00:08.0: GPU reset
                    begin!</div>
                  <div>>>>>>> [  449.794175]
                    Emily:amdgpu_job_free_cb,Process information:</div>
                  <div>>>>>>> process pid 0 thread 
                    pid 0, s_job:000000005086879e [  449.794221]</div>
                  <div>>>>>>>
                    Emily:amdgpu_job_free_cb,Process information:
                    process  pid 0</div>
                  <div>>>>>>> thread pid 0,
                    s_job:0000000066eb74ab [  449.794222]</div>
                  <div>>>>>>>
                    Emily:amdgpu_job_free_cb,Process information:
                    process  pid 0</div>
                  <div>>>>>>> thread pid 0,
                    s_job:00000000d4438ad9 [  449.794255]</div>
                  <div>>>>>>>
                    Emily:amdgpu_job_free_cb,Process information:
                    process  pid 0</div>
                  <div>>>>>>> thread pid 0,
                    s_job:00000000b6d69c65 [  449.794257]</div>
                  <div>>>>>>>
                    Emily:amdgpu_job_free_cb,Process information:
                    process  pid 0</div>
                  <div>>>>>>> thread pid 0,</div>
                  <div>>>>>> s_job:00000000ea85e922 [ 
                    449.794287]</div>
                  <div>>>>>>
                    Emily:amdgpu_job_free_cb,Process</div>
                  <div>>>>>> information: process  pid 0
                    thread  pid 0, s_job:00000000ed3a5ac6 [</div>
                  <div>>>>>> 449.794366] BUG: unable to
                    handle kernel NULL pointer dereference</div>
                  <div>>>>>> at</div>
                  <div>>>>>> 00000000000000c0 [ 
                    449.800818] PGD 0 P4D 0 [  449.801040] Oops:</div>
                  <div>>>>>> 0000 [#1] SMP PTI</div>
                  <div>>>>>>> [  449.801338] CPU: 3
                    PID: 55 Comm: kworker/3:1 Tainted: G           OE</div>
                  <div>>>>>> 4.18.0-15-generic
                    #16~18.04.1-Ubuntu</div>
                  <div>>>>>>> [  449.802157] Hardware
                    name: QEMU Standard PC (i440FX + PIIX,</div>
                  <div>>>>>>> 1996), BIOS
                    Ubuntu-1.8.2-1ubuntu1 04/01/2014 [  449.802944]</div>
                  <div>>>>>>> Workqueue: events
                    drm_sched_job_timedout [amd_sched] [</div>
                  <div>>>>>>> 449.803488]</div>
                  <div>>>> RIP:</div>
                  <div>>>>>>
                    0010:amdgpu_device_gpu_recover+0x1da/0xb60 [amdgpu]</div>
                  <div>>>>>>> [  449.804020] Code: dd
                    ff ff 49 39 c5 48 89 55 a8 0f 85 56 ff ff</div>
                  <div>>>>>>> ff</div>
                  <div>>>>>>> 45 85 e4 0f</div>
                  <div>>>>>> 85 a1 00 00 00 48 8b 45 b0
                    48 85 c0 0f 84 60 01 00 00 48 8b 40 10</div>
                  <div>>>>>> <48> 8b</div>
                  <div>>>> 98</div>
                  <div>>>>>> c0 00         00 00 48 85 db
                    0f 84 4c 01 00 00 48 8b 43 48 a8 01</div>
                  <div>>>>>>> [  449.805593] RSP:
                    0018:ffffb4c7c08f7d68 EFLAGS: 00010286 [</div>
                  <div>>>>>>> 449.806032] RAX:
                    0000000000000000 RBX: 0000000000000000 RCX:</div>
                  <div>>>>>>> 0000000000000000 [ 
                    449.806625] RDX: ffffb4c7c08f5ac0 RSI:</div>
                  <div>>>>>>> 0000000fffffffe0 RDI:
                    0000000000000246 [  449.807224] RBP:</div>
                  <div>>>>>>> ffffb4c7c08f7de0 R08:
                    00000068b9d54000 R09: 0000000000000000 [</div>
                  <div>>>>>>> 449.807818] R10:
                    0000000000000000 R11: 0000000000000148 R12:</div>
                  <div>>>>>>> 0000000000000000 [ 
                    449.808411] R13: ffffb4c7c08f7da0 R14:</div>
                  <div>>>>>>> ffff8d82b8525d40 R15:
                    ffff8d82b8525d40 [  449.809004] FS:</div>
                  <div>>>>>>> 0000000000000000(0000)
                    GS:ffff8d82bfd80000(0000)</div>
                  <div>>>>>>> knlGS:0000000000000000
                    [  449.809674] CS:  0010 DS: 0000 ES: 0000</div>
                  <div>>CR0:</div>
                  <div>>>>>>> 0000000080050033 [ 
                    449.810153] CR2: 00000000000000c0 CR3:</div>
                  <div>>>>>>> 000000003cc0a001 CR4:
                    00000000003606e0 [  449.810747] DR0:</div>
                  <div>>>>>> 0000000000000000 DR1:
                    0000000000000000 DR2: 0000000000000000 [</div>
                  <div>>>>>> 449.811344] DR3:
                    0000000000000000 DR6: 00000000fffe0ff0 DR7:</div>
                  <div>>>>>> 0000000000000400 [ 
                    449.811937] Call Trace:</div>
                  <div>>>>>>> [  449.812206] 
                    amdgpu_job_timedout+0x114/0x140 [amdgpu] [</div>
                  <div>>>>>>> 449.812635] 
                    drm_sched_job_timedout+0x44/0x90 [amd_sched] [</div>
                  <div>>>>>>> 449.813139]  ?
                    amdgpu_cgs_destroy_device+0x10/0x10 [amdgpu] [</div>
                  <div>>>>>>> 449.813609]  ?
                    drm_sched_job_timedout+0x44/0x90 [amd_sched] [</div>
                  <div>>>>>>> 449.814077] 
                    process_one_work+0x1fd/0x3f0 [  449.814417]</div>
                  <div>>>>>>> worker_thread+0x34/0x410
                    [  449.814728]  kthread+0x121/0x140 [</div>
                  <div>>>>>>> 449.815004]  ?
                    process_one_work+0x3f0/0x3f0 [  449.815374]  ?</div>
                  <div>>>>>>>
                    kthread_create_worker_on_cpu+0x70/0x70</div>
                  <div>>>>>>> [  449.815799] 
                    ret_from_fork+0x35/0x40</div>
                  <div>>>>>>></div>
                  <div>>>>>>>> -----Original
                    Message-----</div>
                  <div>>>>>>>> From: Koenig,
                    Christian <a class="moz-txt-link-rfc2396E" href="mailto:Christian.Koenig@amd.com" moz-do-not-send="true"><Christian.Koenig@amd.com></a></div>
                  <div>>>>>>>> Sent: Friday,
                    November 8, 2019 5:43 PM</div>
                  <div>>>>>>>> To: Deng, Emily <a class="moz-txt-link-rfc2396E" href="mailto:Emily.Deng@amd.com" moz-do-not-send="true"><Emily.Deng@amd.com></a>;
                    amd-</div>
                  <div>>>> <a class="moz-txt-link-abbreviated" href="mailto:gfx@lists.freedesktop.org" moz-do-not-send="true">gfx@lists.freedesktop.org</a></div>
                  <div>>>>>>>> Subject: Re: [PATCH]
                    drm/amdgpu: Fix the null pointer issue for</div>
                  <div>>>>>>>> tdr</div>
                  <div>>>>>>>></div>
                  <div>>>>>>>> Am 08.11.19 um 10:39
                    schrieb Deng, Emily:</div>
                  <div>>>>>>>>> Sorry, please
                    take your time.</div>
                  <div>>>>>>>> Have you seen my
                    other response a bit below?</div>
                  <div>>>>>>>></div>
                  <div>>>>>>>> I can't follow how
                    it would be possible for job->s_fence to be</div>
                  <div>>>>>>>> NULL without the job
                    also being freed.</div>
                  <div>>>>>>>></div>
                  <div>>>>>>>> So it looks like
                    this patch is just papering over some bigger issues.</div>
                  <div>>>>>>>></div>
                  <div>>>>>>>> Regards,</div>
                  <div>>>>>>>> Christian.</div>
                  <div>>>>>>>></div>
                  <div>>>>>>>>> Best wishes</div>
                  <div>>>>>>>>> Emily Deng</div>
                  <div>>>>>>>>></div>
                  <div>>>>>>>>></div>
                  <div>>>>>>>>></div>
                  <div>>>>>>>>>>
                    -----Original Message-----</div>
                  <div>>>>>>>>>> From:
                    Koenig, Christian <a class="moz-txt-link-rfc2396E" href="mailto:Christian.Koenig@amd.com" moz-do-not-send="true"><Christian.Koenig@amd.com></a></div>
                  <div>>>>>>>>>> Sent:
                    Friday, November 8, 2019 5:08 PM</div>
                  <div>>>>>>>>>> To: Deng,
                    Emily <a class="moz-txt-link-rfc2396E" href="mailto:Emily.Deng@amd.com" moz-do-not-send="true"><Emily.Deng@amd.com></a>;
                    amd-</div>
                  <div>>>>>> <a class="moz-txt-link-abbreviated" href="mailto:gfx@lists.freedesktop.org" moz-do-not-send="true">gfx@lists.freedesktop.org</a></div>
                  <div>>>>>>>>>> Subject: Re:
                    [PATCH] drm/amdgpu: Fix the null pointer issue for</div>
                  <div>>>>>>>>>> tdr</div>
                  <div>>>>>>>>>></div>
                  <div>>>>>>>>>> Am 08.11.19
                    um 09:52 schrieb Deng, Emily:</div>
                  <div>>>>>>>>>>>
                    Ping.....</div>
                  <div>>>>>>>>>> You need to
                    give me at least enough time to wake up :)</div>
                  <div>>>>>>>>>></div>
                  <div>>>>>>>>>>> Best
                    wishes</div>
                  <div>>>>>>>>>>> Emily
                    Deng</div>
                  <div>>>>>>>>>>></div>
                  <div>>>>>>>>>>></div>
                  <div>>>>>>>>>>></div>
                  <div>>>>>>>>>>>>
                    -----Original Message-----</div>
                  <div>>>>>>>>>>>>
                    From: amd-gfx <a class="moz-txt-link-rfc2396E" href="mailto:amd-gfx-bounces@lists.freedesktop.org" moz-do-not-send="true"><amd-gfx-bounces@lists.freedesktop.org></a>
                    On</div>
                  <div>>>> Behalf</div>
                  <div>>>>>>>>>>>> Of
                    Deng, Emily</div>
                  <div>>>>>>>>>>>>
                    Sent: Friday, November 8, 2019 10:56 AM</div>
                  <div>>>>>>>>>>>> To:
                    Koenig, Christian <a class="moz-txt-link-rfc2396E" href="mailto:Christian.Koenig@amd.com" moz-do-not-send="true"><Christian.Koenig@amd.com></a>;
                    amd-</div>
                  <div>>>>>>>>>>>> <a class="moz-txt-link-abbreviated" href="mailto:gfx@lists.freedesktop.org" moz-do-not-send="true">gfx@lists.freedesktop.org</a></div>
                  <div>>>>>>>>>>>>
                    Subject: RE: [PATCH] drm/amdgpu: Fix the null
                    pointer issue</div>
                  <div>>>>>>>>>>>> for
                    tdr</div>
                  <div>>>>>>>>>>>></div>
                  <div>>>>>>>>>>>>>
                    -----Original Message-----</div>
                  <div>>>>>>>>>>>>>
                    From: Christian König <a class="moz-txt-link-rfc2396E" href="mailto:ckoenig.leichtzumerken@gmail.com" moz-do-not-send="true"><ckoenig.leichtzumerken@gmail.com></a></div>
                  <div>>>>>>>>>>>>>
                    Sent: Thursday, November 7, 2019 7:28 PM</div>
                  <div>>>>>>>>>>>>>
                    To: Deng, Emily <a class="moz-txt-link-rfc2396E" href="mailto:Emily.Deng@amd.com" moz-do-not-send="true"><Emily.Deng@amd.com></a>;</div>
                  <div>>>>>>>>>>>>>
                    <a class="moz-txt-link-abbreviated" href="mailto:amd-gfx@lists.freedesktop.org" moz-do-not-send="true">amd-gfx@lists.freedesktop.org</a></div>
                  <div>>>>>>>>>>>>>
                    Subject: Re: [PATCH] drm/amdgpu: Fix the null
                    pointer issue</div>
                  <div>>>>>>>>>>>>>
                    for tdr</div>
                  <div>>>>>>>>>>>>></div>
                  <div>>>>>>>>>>>>>
                    Am 07.11.19 um 11:25 schrieb Emily Deng:</div>
                  <div>>>>>>>>>>>>>>
                    When the job is already signaled, the s_fence is
                    freed.</div>
                  <div>>>>>>>>>>>>>>
                    Then it will has null pointer in
                    amdgpu_device_gpu_recover.</div>
                  <div>>>>>>>>>>>>>
                    NAK, the s_fence is only set to NULL when the job is
                    destroyed.</div>
                  <div>>>>>>>>>>>>>
                    See drm_sched_job_cleanup().</div>
                  <div>>>>>>>>>>>> I
                    know it is set to NULL in drm_sched_job_cleanup. But
                    in one</div>
                  <div>>>>>>>>>>>>
                    case, when it enter into the
                    amdgpu_device_gpu_recover, it</div>
                  <div>>>>>>>>>>>>
                    already in drm_sched_job_cleanup, and at this time,
                    it will</div>
                  <div>>>>>>>>>>>> go
                    to free</div>
                  <div>>>>>> job.</div>
                  <div>>>>>>>>>>>> But
                    the amdgpu_device_gpu_recover sometimes is faster.
                    At</div>
                  <div>>>>>>>>>>>> that
                    time, job is not freed, but s_fence is already NULL.</div>
                  <div>>>>>>>>>> No, that
                    case can't happen. See here:</div>
                  <div>>>>>>>>>></div>
                  <div>>>>>>>>>>>            
                    drm_sched_job_cleanup(s_job);</div>
                  <div>>>>>>>>>>></div>
                  <div>>>>>>>>>>>            
                    amdgpu_ring_priority_put(ring,
                    s_job->s_priority);</div>
                  <div>>>>>>>>>>>            
                    dma_fence_put(job->fence);</div>
                  <div>>>>>>>>>>>            
                    amdgpu_sync_free(&job->sync);</div>
                  <div>>>>>>>>>>>            
                    amdgpu_sync_free(&job->sched_sync);</div>
                  <div>>>>>>>>>>>            
                    kfree(job);</div>
                  <div>>>>>>>>>> The job
                    itself is freed up directly after freeing the
                    reference</div>
                  <div>>>>>>>>>> to the</div>
                  <div>>>>>> s_fence.</div>
                  <div>>>>>>>>>> So you are
                    just papering over a much bigger problem here. This</div>
                  <div>>>>>>>>>> patch is a
                    clear NAK.</div>
                  <div>>>>>>>>>></div>
                  <div>>>>>>>>>> Regards,</div>
                  <div>>>>>>>>>> Christian.</div>
                  <div>>>>>>>>>></div>
                  <div>>>>>>>>>>>>>
                    When you see a job without an s_fence then that
                    means the</div>
                  <div>>>>>>>>>>>>>
                    problem is somewhere else.</div>
                  <div>>>>>>>>>>>>></div>
                  <div>>>>>>>>>>>>>
                    Regards,</div>
                  <div>>>>>>>>>>>>>
                    Christian.</div>
                  <div>>>>>>>>>>>>></div>
                  <div>>>>>>>>>>>>>>
                    Signed-off-by: Emily Deng <a class="moz-txt-link-rfc2396E" href="mailto:Emily.Deng@amd.com" moz-do-not-send="true"><Emily.Deng@amd.com></a></div>
                  <div>>>>>>>>>>>>>>
                    ---</div>
                  <div>>>>>>>>>>>>>>       
                    drivers/gpu/drm/amd/amdgpu/amdgpu_device.c |  2 +-</div>
                  <div>>>>>>>>>>>>>>       
                    drivers/gpu/drm/scheduler/sched_main.c     | 11
                    ++++++---</div>
                  <div>>--</div>
                  <div>>>>>>>>>>>>>>       
                    2 files changed, 7 insertions(+), 6 deletions(-)</div>
                  <div>>>>>>>>>>>>>></div>
                  <div>>>>>>>>>>>>>>
                    diff --git
                    a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c</div>
                  <div>>>>>>>>>>>>>>
                    b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c</div>
                  <div>>>>>>>>>>>>>>
                    index e6ce949..5a8f08e 100644</div>
                  <div>>>>>>>>>>>>>>
                    --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c</div>
                  <div>>>>>>>>>>>>>>
                    +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c</div>
                  <div>>>>>>>>>>>>>>
                    @@ -4075,7 +4075,7 @@ int</div>
                  <div>>>> amdgpu_device_gpu_recover(struct</div>
                  <div>>>>>>>>>>>>>
                    amdgpu_device *adev,</div>
                  <div>>>>>>>>>>>>>>            
                    *</div>
                  <div>>>>>>>>>>>>>>            
                    * job->base holds a reference to parent fence</div>
                  <div>>>>>>>>>>>>>>            
                    */</div>
                  <div>>>>>>>>>>>>>>
                    -  if (job &&
                    job->base.s_fence->parent &&</div>
                  <div>>>>>>>>>>>>>>
                    +  if (job && job->base.s_fence
                    &&</div>
                  <div>>>>>>>>>>>>>>
                    + job->base.s_fence->parent</div>
                  <div>>>>>>>> &&</div>
                  <div>>>>>>>>>>>>>>               
dma_fence_is_signaled(job->base.s_fence->parent))</div>
                  <div>>>>>>>>>>>>>>                   
                    job_signaled = true;</div>
                  <div>>>>>>>>>>>>>></div>
                  <div>>>>>>>>>>>>>>
                    diff --git a/drivers/gpu/drm/scheduler/sched_main.c</div>
                  <div>>>>>>>>>>>>>>
                    b/drivers/gpu/drm/scheduler/sched_main.c</div>
                  <div>>>>>>>>>>>>>>
                    index 31809ca..56cc10e 100644</div>
                  <div>>>>>>>>>>>>>>
                    --- a/drivers/gpu/drm/scheduler/sched_main.c</div>
                  <div>>>>>>>>>>>>>>
                    +++ b/drivers/gpu/drm/scheduler/sched_main.c</div>
                  <div>>>>>>>>>>>>>>
                    @@ -334,8 +334,8 @@ void</div>
                  <div>>drm_sched_increase_karma(struct</div>
                  <div>>>>>>>>>>>>>
                    drm_sched_job</div>
                  <div>>>>>>>>>>>>>>
                    *bad)</div>
                  <div>>>>>>>>>>>>>></div>
                  <div>>>>>>>>>>>>>>                           
                    spin_lock(&rq->lock);</div>
                  <div>>>>>>>>>>>>>>                           
                    list_for_each_entry_safe(entity,</div>
                  <div>>>>>>>>>>>>>>
                    tmp,</div>
                  <div>>>> &rq-</div>
                  <div>>>>>>>>> entities,</div>
                  <div>>>>>>>>>>>>>
                    list) {</div>
                  <div>>>>>>>>>>>>>>
                    -                          if
                    (bad->s_fence->scheduled.context</div>
                  <div>>>>>>>> ==</div>
                  <div>>>>>>>>>>>>>>
                    -                             
                    entity->fence_context) {</div>
                  <div>>>>>>>>>>>>>>
                    +                          if (bad->s_fence
                    &&</div>
                  <div>>>>>>>>>>>>>>
                    + (bad->s_fence-</div>
                  <div>>>>>>>>>>>>>>
                    scheduled.context ==</div>
                  <div>>>>>>>>>>>>>>
                    +                             
                    entity->fence_context)) {</div>
                  <div>>>>>>>>>>>>>>                                           
                    if</div>
                  <div>>>>>>>>>>>>>>
                    (atomic_read(&bad-</div>
                  <div>>>>>>>>> karma) ></div>
                  <div>>>>>>>>>>>>>>                                               
                    bad->sched-</div>
                  <div>>>>> hang_limit)</div>
                  <div>>>>>>>>>>>>>>                                                   
                    if</div>
                  <div>>>>>>>>>>>>>>
                    (entity-</div>
                  <div>>>>> guilty) @@ -376,7 +376,7 @@ void</div>
                  <div>>>>>>>>>>>>>>
                    drm_sched_stop(struct</div>
                  <div>>>>>>>> drm_gpu_scheduler</div>
                  <div>>>>>>>>>>>>>
                    *sched, struct drm_sched_job *bad)</div>
                  <div>>>>>>>>>>>>>>            
                    * This iteration is thread safe as sched thread</div>
                  <div>>>>>>>>>>>>>>
                    is</div>
                  <div>>>> stopped.</div>
                  <div>>>>>>>>>>>>>>            
                    */</div>
                  <div>>>>>>>>>>>>>>           
                    list_for_each_entry_safe_reverse(s_job, tmp,</div>
                  <div>>>>>>>>>>>>>>
                    &sched- ring_mirror_list, node) {</div>
                  <div>>>>>>>>>>>>>>
                    -          if (s_job->s_fence->parent
                    &&</div>
                  <div>>>>>>>>>>>>>>
                    +          if (s_job->s_fence &&
                    s_job->s_fence->parent &&</div>
                  <div>>>>>>>>>>>>>>                       
                    dma_fence_remove_callback(s_job-</div>
                  <div>>>>> s_fence-</div>
                  <div>>>>>>>>> parent,</div>
                  <div>>>>>>>>>>>>>>                                                 
                    &s_job->cb)) {</div>
                  <div>>>>>>>>>>>>>>                           
                    atomic_dec(&sched->hw_rq_count);</div>
                  <div>>>> @@ -</div>
                  <div>>>>>>>> 395,7</div>
                  <div>>>>>>>>>>>>
                    +395,8 @@ void</div>
                  <div>>>>>>>>>>>>>>
                    drm_sched_stop(struct drm_gpu_scheduler</div>
                  <div>>>>>>>>>>>>>
                    *sched, struct drm_sched_job *bad)</div>
                  <div>>>>>>>>>>>>>>                            
                    *</div>
                  <div>>>>>>>>>>>>>>                            
                    * Job is still alive so fence</div>
                  <div>>>>>>>>>>>>>>
                    refcount at</div>
                  <div>>>> least 1</div>
                  <div>>>>>>>>>>>>>>                            
                    */</div>
                  <div>>>>>>>>>>>>>>
                    -                 
                    dma_fence_wait(&s_job->s_fence->finished,</div>
                  <div>>>>>>>> false);</div>
                  <div>>>>>>>>>>>>>>
                    +                  if (s_job->s_fence)</div>
                  <div>>>>>>>>>>>>>>
                    +                         
                    dma_fence_wait(&s_job->s_fence-</div>
                  <div>>>>>>>>> finished,</div>
                  <div>>>>>>>>>>>>>
                    false);</div>
                  <div>>>>>>>>>>>>>>                           
                    /*</div>
                  <div>>>>>>>>>>>>>>                            
                    * We must keep bad job alive</div>
                  <div>>>>>>>>>>>>>>
                    for later</div>
                  <div>>>> use</div>
                  <div>>>>>>>> during @@</div>
                  <div>>>>>>>>>>>>>
                    -438,7</div>
                  <div>>>>>>>>>>>>>>
                    +439,7 @@ void drm_sched_start(struct
                    drm_gpu_scheduler</div>
                  <div>>>>>> *sched,</div>
                  <div>>>>>>>>>>>>>>
                    +bool</div>
                  <div>>>>>>>>>>>>>
                    full_recovery)</div>
                  <div>>>>>>>>>>>>>>            
                    * GPU recovers can't run in parallel.</div>
                  <div>>>>>>>>>>>>>>            
                    */</div>
                  <div>>>>>>>>>>>>>>           
                    list_for_each_entry_safe(s_job, tmp,</div>
                  <div>>>>>>>>>>>>>>
                    &sched->ring_mirror_list,</div>
                  <div>>>>>>>>>>>>>>
                    node)</div>
                  <div>>>>>>>>>>>>>
                    {</div>
                  <div>>>>>>>>>>>>>>
                    -          struct dma_fence *fence =
                    s_job->s_fence->parent;</div>
                  <div>>>>>>>>>>>>>>
                    +          struct dma_fence *fence =
                    s_job->s_fence ?</div>
                  <div>>>>>>>>>>>>>>
                    + s_job-</div>
                  <div>>>>>>>>> s_fence-</div>
                  <div>>>>>>>>>>>>>>
                    parent :</div>
                  <div>>>>>>>>>>>>>>
                    +NULL;</div>
                  <div>>>>>>>>>>>>>></div>
                  <div>>>>>>>>>>>>>>                   
                    atomic_inc(&sched->hw_rq_count);</div>
                  <div>>>>>>>>>>>>>></div>
                  <div>>>>>>>>>>>>
                    _______________________________________________</div>
                  <div>>>>>>>>>>>>
                    amd-gfx mailing list</div>
                  <div>>>>>>>>>>>> <a class="moz-txt-link-abbreviated" href="mailto:amd-gfx@lists.freedesktop.org" moz-do-not-send="true">amd-gfx@lists.freedesktop.org</a></div>
                  <div>>>>>>>>>>>> <a href="https://lists.freedesktop.org/mailman/listinfo/amd-gfx" moz-do-not-send="true">
                      https://lists.freedesktop.org/mailman/listinfo/amd-gfx</a></div>
                  <div>></div>
                  <div>>_______________________________________________</div>
                  <div>>amd-gfx mailing list</div>
                  <div>><a class="moz-txt-link-abbreviated" href="mailto:amd-gfx@lists.freedesktop.org" moz-do-not-send="true">amd-gfx@lists.freedesktop.org</a></div>
                  <div>><a href="https://lists.freedesktop.org/mailman/listinfo/amd-gfx" moz-do-not-send="true">https://lists.freedesktop.org/mailman/listinfo/amd-gfx</a></div>
                  <div><font face="Times New Roman"> </font></div>
                </span></font> </blockquote>
            <br>
          </blockquote>
          <br>
          <fieldset class="mimeAttachmentHeader"></fieldset>
          <pre class="moz-quote-pre" wrap="">_______________________________________________
amd-gfx mailing list
<a class="moz-txt-link-abbreviated" href="mailto:amd-gfx@lists.freedesktop.org" moz-do-not-send="true">amd-gfx@lists.freedesktop.org</a>
<a class="moz-txt-link-freetext" href="https://lists.freedesktop.org/mailman/listinfo/amd-gfx" moz-do-not-send="true">https://lists.freedesktop.org/mailman/listinfo/amd-gfx</a></pre>
        </blockquote>
        <br>
      </blockquote>
    </blockquote>
    <br>
  </body>
</html>