Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
MIME-Version: 1.0
References: <CAH1Ww+TPCSyiaC3oeoWPtsi-vDfDY=K4ByoLD37-onMvsAB5Rg@mail.gmail.com>
 <e62c6c4a-b8d2-4f7e-09c2-b5151fe12046@amd.com>
In-Reply-To: <e62c6c4a-b8d2-4f7e-09c2-b5151fe12046@amd.com>
From:   Daniel Gomez <daniel@qtec.com>
Date:   Wed, 3 Feb 2021 15:54:36 +0100
Message-ID: <CAH1Ww+QJAjzek18Zh8O7m+twdjV2uBBeaWYeCS5D8B4As3WWCw@mail.gmail.com>
Subject: Re: [amdgpu] deadlock
To:     =?UTF-8?Q?Christian_K=C3=B6nig?= <christian.koenig@amd.com>
Cc:     amd-gfx list <amd-gfx@lists.freedesktop.org>,
        dri-devel <dri-devel@lists.freedesktop.org>,
        Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
        Alex Deucher <alexander.deucher@amd.com>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
Precedence: bulk

On Wed, 3 Feb 2021 at 15:37, Christian K=C3=B6nig <christian.koenig@amd.com=
> wrote:
>
> Hi Daniel,
>
> I've talked a bit with our internal team.
>
> The problem is that the 20.20 release still uses the older OpenCL stack
> which obviously has a bug here and causes a hang.
>
> The best approach I can give you is to switch to the ROCm stack instead.
Thanks Christian. I'll try with the ROCm stack then. As far as I understood=
,
it should work because the part of the code where it now hangs is not actua=
lly
used by the ROCm stack, is that correct? However, the hang/bug will
still be there
even though it is not used in that stack.
Anyway, I'll keep you guys posted with this change.

>
> Regards,
> Christian.
>
> Am 03.02.21 um 09:33 schrieb Daniel Gomez:
> > Hi all,
> >
> > I have a deadlock with the amdgpu mainline driver when running in paral=
lel two
> > OpenCL applications. So far, we've been able to replicate it easily by =
executing
> > clinfo and MatrixMultiplication (from AMD opencl-samples). It's quite o=
ld the
> > opencl-samples so, if you have any other suggestion for testing I'd be =
very
> > happy to test it as well.
> >
> > How to replicate the issue:
> >
> > # while true; do /usr/bin/MatrixMultiplication --device gpu \
> >      --deviceId 0 -x 1000 -y 1000 -z 1000 -q -t -i 50; done
> > # while true; do clinfo; done
> >
> > Output:
> >
> > After a minute or less (sometimes could be more) I can see that
> > MatrixMultiplication and clinfo hang. In addition, with radeontop you c=
an see
> > how the Graphics pipe goes from ~50% to 100%. Also the shader clocks
> > goes up from ~35% to ~96%.
> >
> > clinfo keeps printing:
> > ioctl(7, DRM_IOCTL_SYNCOBJ_WAIT, 0x7ffe46e5f950) =3D -1 ETIME (Timer ex=
pired)
> >
> > And MatrixMultiplication prints the following (strace) if you try to
> > kill the process:
> >
> > sched_yield()                           =3D 0
> > futex(0x557e945343b8, FUTEX_WAIT_BITSET_PRIVATE|FUTEX_CLOCK_REALTIME, 0=
,
> > NULL, FUTEX_BITSET_MATCH_ANYstrace: Process 651 detached
> >   <detached ...>
> >
> > After this, the gpu is not functional at all and you'd need a power cyc=
le reset
> > to restore the system.
> >
> > Hardware info:
> > CPU: AMD Ryzen Embedded V1605B with Radeon Vega Gfx (8) @ 2.000GHz
> > GPU: AMD ATI Radeon Vega Series / Radeon Vega Mobile Series
> >
> > 03:00.0 VGA compatible controller: Advanced Micro Devices, Inc.
> > [AMD/ATI] Raven Ridge [Radeon Vega Series / Radeon Vega Mobile Series]
> > (rev 83)
> >      DeviceName: Broadcom 5762
> >      Subsystem: Advanced Micro Devices, Inc. [AMD/ATI] Raven Ridge
> > [Radeon Vega Series / Radeon Vega Mobile Series]
> >      Kernel driver in use: amdgpu
> >      Kernel modules: amdgpu
> >
> > Linux kernel info:
> >
> > root@qt5222:~# uname -a
> > Linux qt5222 5.11.0-rc6-qtec-standard #2 SMP Tue Feb 2 09:41:46 UTC
> > 2021 x86_64 x86_64 x86_64 GNU/Linux
> >
> > By enabling the kernel locks stats I could see the MatrixMultiplication=
 is
> > hanged in the amdgpu_mn_invalidate_gfx function:
> >
> > [  738.359202] 1 lock held by MatrixMultiplic/653:
> > [  738.359206]  #0: ffff88810e364fe0
> > (&adev->notifier_lock){+.+.}-{3:3}, at:
> > amdgpu_mn_invalidate_gfx+0x34/0xa0 [amdgpu]
> >
> > I can see in the the amdgpu_mn_invalidate_gfx function: the
> > dma_resv_wait_timeout_rcu uses wait_all (fences) and MAX_SCHEDULE_TIMEO=
UT so, I
> > guess the code gets stuck there waiting forever. According to the
> > documentation: "When somebody tries to invalidate the page tables we bl=
ock the
> > update until all operations on the pages in question are completed, the=
n those
> > pages are marked  as accessed and also dirty if it wasn=E2=80=99t a rea=
d only access."
> > Looks like the fences are deadlocked and therefore, it never returns. C=
ould it
> > be possible? any hint to where can I look to fix this?
> >
> > Thank you  in advance.
> >
> > Here the full dmesg output:
> >
> > [  738.337726] INFO: task MatrixMultiplic:653 blocked for more than 122=
 seconds.
> > [  738.344937]       Not tainted 5.11.0-rc6-qtec-standard #2
> > [  738.350384] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs"
> > disables this message.
> > [  738.358240] task:MatrixMultiplic state:D stack:    0 pid:  653
> > ppid:     1 flags:0x00004000
> > [  738.358254] Call Trace:
> > [  738.358261]  ? dma_fence_default_wait+0x1eb/0x230
> > [  738.358276]  __schedule+0x370/0x960
> > [  738.358291]  ? dma_fence_default_wait+0x117/0x230
> > [  738.358297]  ? dma_fence_default_wait+0x1eb/0x230
> > [  738.358305]  schedule+0x51/0xc0
> > [  738.358312]  schedule_timeout+0x275/0x380
> > [  738.358324]  ? dma_fence_default_wait+0x1eb/0x230
> > [  738.358332]  ? mark_held_locks+0x4f/0x70
> > [  738.358341]  ? dma_fence_default_wait+0x117/0x230
> > [  738.358347]  ? lockdep_hardirqs_on_prepare+0xd4/0x180
> > [  738.358353]  ? _raw_spin_unlock_irqrestore+0x39/0x40
> > [  738.358362]  ? dma_fence_default_wait+0x117/0x230
> > [  738.358370]  ? dma_fence_default_wait+0x1eb/0x230
> > [  738.358375]  dma_fence_default_wait+0x214/0x230
> > [  738.358384]  ? dma_fence_release+0x1a0/0x1a0
> > [  738.358396]  dma_fence_wait_timeout+0x105/0x200
> > [  738.358405]  dma_resv_wait_timeout_rcu+0x1aa/0x5e0
> > [  738.358421]  amdgpu_mn_invalidate_gfx+0x55/0xa0 [amdgpu]
> > [  738.358688]  __mmu_notifier_release+0x1bb/0x210
> > [  738.358710]  exit_mmap+0x2f/0x1e0
> > [  738.358723]  ? find_held_lock+0x34/0xa0
> > [  738.358746]  mmput+0x39/0xe0
> > [  738.358756]  do_exit+0x5c3/0xc00
> > [  738.358763]  ? find_held_lock+0x34/0xa0
> > [  738.358780]  do_group_exit+0x47/0xb0
> > [  738.358791]  get_signal+0x15b/0xc50
> > [  738.358807]  arch_do_signal_or_restart+0xaf/0x710
> > [  738.358816]  ? lockdep_hardirqs_on_prepare+0xd4/0x180
> > [  738.358822]  ? _raw_spin_unlock_irqrestore+0x39/0x40
> > [  738.358831]  ? ktime_get_mono_fast_ns+0x50/0xa0
> > [  738.358844]  ? amdgpu_drm_ioctl+0x6b/0x80 [amdgpu]
> > [  738.359044]  exit_to_user_mode_prepare+0xf2/0x1b0
> > [  738.359054]  syscall_exit_to_user_mode+0x19/0x60
> > [  738.359062]  entry_SYSCALL_64_after_hwframe+0x44/0xa9
> > [  738.359069] RIP: 0033:0x7f6b89a51887
> > [  738.359076] RSP: 002b:00007f6b82b54b18 EFLAGS: 00000246 ORIG_RAX:
> > 0000000000000010
> > [  738.359086] RAX: fffffffffffffe00 RBX: 00007f6b82b54b50 RCX: 00007f6=
b89a51887
> > [  738.359091] RDX: 00007f6b82b54b50 RSI: 00000000c02064c3 RDI: 0000000=
000000007
> > [  738.359096] RBP: 00000000c02064c3 R08: 0000000000000003 R09: 00007f6=
b82b54bbc
> > [  738.359101] R10: 0000000000000001 R11: 0000000000000246 R12: 0000000=
165a0bc00
> > [  738.359106] R13: 0000000000000007 R14: 0000000000000001 R15: 0000000=
000000000
> > [  738.359129]
> >                 Showing all locks held in the system:
> > [  738.359141] 1 lock held by khungtaskd/54:
> > [  738.359148]  #0: ffffffff829f6840 (rcu_read_lock){....}-{1:2}, at:
> > debug_show_all_locks+0x15/0x183
> > [  738.359187] 1 lock held by systemd-journal/174:
> > [  738.359202] 1 lock held by MatrixMultiplic/653:
> > [  738.359206]  #0: ffff88810e364fe0
> > (&adev->notifier_lock){+.+.}-{3:3}, at:
> > amdgpu_mn_invalidate_gfx+0x34/0xa0 [amdgpu]
> >
> > Daniel
>