[PATCH 22/22] drm/amdgpu: fix reservation obj shared count bug
Christian König
christian.koenig at amd.com
Wed Feb 28 12:38:49 UTC 2018
Ok, I've figured out what is going on here. The problem was that we
tried to clear VM BOs with the standard background clear function.
Because of this you needed not only one, but two fence slots for newly
allocated page directories/tables.
But I've already fixed this bug quite a while ago, make sure you have
this patch in your branch:
> commit 13307f7e1d0c05a68f4ba19193cbd213573a8680
> Author: Christian König <christian.koenig at amd.com>
> Date: Wed Jan 24 17:19:04 2018 +0100
>
> drm/amdgpu: revert "drm/amdgpu: use AMDGPU_GEM_CREATE_VRAM_CLEARED
> for VM PD/PTs" v2
>
> Using the standard clear turned out to be to inflexible.
>
> First of all it is executed on the system queue, together with buffer
> moves instead on the per VM queue.
>
> And second we need to fill in the page tables with more than just
> zero.
>
> We keep the new functionality of initializing the PDEs/PTEs with ATC
> routing entries intact.
>
> v2: update commit message.
>
> Signed-off-by: Christian König <christian.koenig at amd.com>
> Signed-off-by: Alex Deucher <alexander.deucher at amd.com>
And also this follow up commit:
> commit 29e8357b4cbbfcee6d375f2d183b674b678923d7
> Author: Christian König <christian.koenig at amd.com>
> Date: Sun Feb 4 19:36:52 2018 +0100
>
> drm/amdgpu: sync the VM PD/PT before clearing it
>
> Otherwise we might overwrite stuff which is still in use.
>
> Signed-off-by: Christian König <christian.koenig at amd.com>
> Reviewed-by: Felix Kuehling <Felix.Kuehling at amd.com>
> Signed-off-by: Alex Deucher <alexander.deucher at amd.com>
Regards,
Christian.
Am 27.02.2018 um 10:17 schrieb Liu, Monk:
> I know in theoretically this issue should be equal for sriov or not
> But since the gpu performance under SRIOV is only one quarter compared with bare-metal (assume 4VF)
> And the CPU's performance is also limited, so there are chance that the fence is signaled not as quick as bare-metal
> And lead to the scenario that four different context fence are living in resv,
>
> If you do it on bare-metal maybe you need some tricks, at least on my sriov platform I can duplicate this BUG()
> With 4 repeating loop on VK_EXAMPLE every time
>
> /Monk
>
> -----Original Message-----
> From: Koenig, Christian
> Sent: 2018年2月27日 16:34
> To: Liu, Monk <Monk.Liu at amd.com>; amd-gfx at lists.freedesktop.org; Chris Wilson <chris at chris-wilson.co.uk>
> Subject: Re: [PATCH 22/22] drm/amdgpu: fix reservation obj shared count bug
>
> Am 27.02.2018 um 04:36 schrieb Liu, Monk:
>>> Well then there is something else broken and I actually think that this is the root cause here.
>> What if this VM submits four different jobs on four different rings,
>> e.g. gfx/compute1/compute2/vce So the fences for those jobs are from
>> different context, and now you get a mapping on this VM So the shared
>> count is now 4
> Yeah, but each VM submission reserves a shared slot before adding the fence. So even if you have 4 submissions we still have a guaranteed free slot.
>
>> I ran VK_example and it always can hit this reservation BUG(), just
>> repeat running it 4 loops and The BUG() hit, but I did it under SRIOV
>> case so the CPU&GPU performance is different with bare-metal
>>
>> Anyway do you think above scenario exist ?
> Well you somehow run into this problem, I just don't understand how.
>
> It shouldn't matter if you run on bare-metal or SRIOV, the VM handling is the same as far as I know.
>
> Going to give it a try today to create some unit tests exercising the problem.
>
> Christian.
>
>>
>> -----Original Message-----
>> From: Koenig, Christian
>> Sent: 2018年2月26日 19:19
>> To: Liu, Monk <Monk.Liu at amd.com>; amd-gfx at lists.freedesktop.org; Chris
>> Wilson <chris at chris-wilson.co.uk>
>> Subject: Re: [PATCH 22/22] drm/amdgpu: fix reservation obj shared
>> count bug
>>
>>> What if all those context are not equal ? and accumulated finally to 4 ?
>> Well then there is something else broken and I actually think that this is the root cause here.
>>
>>> You can try VK_EXAMPLE test yourself, see if you can hit that BUG()
>>> in reservation.c file,
>> Could you generate a simpler unit test to trigger the problem?
>>
>>
>> Thanks,
>> Christian.
>>
>> Am 26.02.2018 um 11:57 schrieb Liu, Monk:
>>>> So we call reserve_shared only once and then call reservation_object_add_shared_fence() multiple times, always with the same context.
>>>> should replace the fence when it has the same context as a previously added fence.
>>> What if all those context are not equal ? and accumulated finally to 4 ?
>>> You can try VK_EXAMPLE test yourself, see if you can hit that BUG()
>>> in reservation.c file, I hit it and it is fixed by this patch, I don't think there is other reason.
>>>
>>> -----Original Message-----
>>> From: amd-gfx [mailto:amd-gfx-bounces at lists.freedesktop.org] On
>>> Behalf Of Christian K?nig
>>> Sent: 2018年2月26日 18:45
>>> To: Liu, Monk <Monk.Liu at amd.com>; amd-gfx at lists.freedesktop.org;
>>> Chris Wilson <chris at chris-wilson.co.uk>
>>> Subject: Re: [PATCH 22/22] drm/amdgpu: fix reservation obj shared
>>> count bug
>>>
>>>> We must call reserve_shared before amdgpu_bo_fence....
>>> Actually that's not correct. See
>>> reservation_object_add_shared_fence()
>>> should replace the fence when it has the same context as a previously added fence.
>>>
>>> So we call reserve_shared only once and then call
>>> reservation_object_add_shared_fence() multiple times, always with the same context.
>>>
>>> Writing this, Chris actually had patches which may have broken that assumption, but as far as I know those aren't merged yet.
>>>
>>> Regards,
>>> Christian.
>>>
>>> Am 26.02.2018 um 11:36 schrieb Liu, Monk:
>>>> amdgpu_vm_get_pd_bo() should already reserves a shared slot for the fence:
>>>>> entry->tv.shared = true;
>>>> That's not related with the bug I fixed! The root cause is there is
>>>> some sequence that When code path comes to amdgpu_bo_fence of
>>>> vm_update_directories, no one The shared_cnt is by coincidence
>>>> already the max value of resv obj, so the BUG() in Reservation.c
>>>> file will hit,
>>>>
>>>> We must call reserve_shared before amdgpu_bo_fence....
>>>>
>>>>
>>>>
>>>> -----Original Message-----
>>>> From: Christian König [mailto:ckoenig.leichtzumerken at gmail.com]
>>>> Sent: 2018年2月26日 17:48
>>>> To: Liu, Monk <Monk.Liu at amd.com>; amd-gfx at lists.freedesktop.org
>>>> Subject: Re: [PATCH 22/22] drm/amdgpu: fix reservation obj shared
>>>> count bug
>>>>
>>>> That fix is incorrect.
>>>>
>>>> amdgpu_vm_get_pd_bo() should already reserves a shared slot for the fence:
>>>>> entry->tv.shared = true;
>>>> So you must run into this issue because of something else.
>>>>
>>>> Regards,
>>>> Christian.
>>>>
>>>> Am 26.02.2018 um 06:35 schrieb Monk Liu:
>>>>> should call reservation_object_reserve_shared before
>>>>> amdgpu_bo_fence(), otherwise there are odds kernel hit BUG in
>>>>> reversation.c
>>>>>
>>>>> bug:
>>>>> [12622.076435] ------------[ cut here ]------------ [12622.076438]
>>>>> kernel BUG at drivers/dma-buf/reservation.c:233!
>>>>> [12622.078046] invalid opcode: 0000 [#1] SMP KASAN [12622.078345]
>>>>> Modules linked in: amdgpu(E) chash(E) gpu_sched(E) ttm(E) drm_kms_helper(E) drm(E) i2c_algo_bit(E) fb_sys_fops(E) syscopyarea(E) sysfillrect(E) sysimgblt(E) amdkfd(E) amd_iommu_v2(E) snd_hda_codec_generic(E) snd_hda_intel(E) snd_hda_codec(E) snd_hda_core(E) snd_hwdep(E) snd_pcm(E) crct10dif_pclmul(E) crc32_pclmul(E) ghash_clmulni_intel(E) pcbc(E) snd_seq_midi(E) snd_seq_midi_event(E) snd_rawmidi(E) snd_seq(E) aesni_intel(E) snd_seq_device(E) aes_x86_64(E) crypto_simd(E) glue_helper(E) cryptd(E) snd_timer(E) serio_raw(E) snd(E) soundcore(E) i2c_piix4(E) mac_hid(E) binfmt_misc(E) nfsd(E) auth_rpcgss(E) nfs_acl(E) lockd(E) grace(E) sunrpc(E) parport_pc(E) ppdev(E) lp(E) parport(E) autofs4(E) 8139too(E) psmouse(E) floppy(E) 8139cp(E) mii(E) pata_acpi(E)
>>>>> [12622.082664] CPU: 2 PID: 6075 Comm: ShaderImageLoad Tainted: G E 4.13.0-feb15-2018-guest-12 #8
>>>>> [12622.083278] Hardware name: QEMU Standard PC (i440FX + PIIX,
>>>>> 1996), BIOS Ubuntu-1.8.2-1ubuntu1 04/01/2014 [12622.083880] task:
>>>>> ffff8801314f8000 task.stack: ffff8801e6df0000 [12622.084267] RIP:
>>>>> 0010:reservation_object_add_shared_fence+0x2bc/0x2c0
>>>>> [12622.084680] RSP: 0018:ffff8801e6df7ae8 EFLAGS: 00010246
>>>>> [12622.085016] RAX: 0000000000000004 RBX: ffff8801f2542a80 RCX:
>>>>> ffff8801f2542b58 [12622.085471] RDX: ffff880044791428 RSI:
>>>>> ffff8801e2657d20 RDI: ffff880044791428 [12622.085995] RBP:
>>>>> ffff8801e6df7b28 R08: 0000000000000000 R09: 0000000000000000
>>>>> [12622.086451] R10: 000000009fe2fe07 R11: 0000000000000002 R12:
>>>>> ffff880044791200 [12622.086905] R13: ffff8801e1da6850 R14:
>>>>> 0000000000000000 R15: ffff8801e9303000 [12622.087359] FS:
>>>>> 00007f70109ff700(0000) GS:ffff8801f7500000(0000)
>>>>> knlGS:0000000000000000 [12622.087874] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [12622.088242] CR2: 00007f6fe1a49000 CR3: 00000001e8377000 CR4: 00000000001406e0 [12622.088698] Call Trace:
>>>>> [12622.088934] amdgpu_bo_fence+0x25/0x30 [amdgpu] [12622.089271]
>>>>> amdgpu_vm_update_directories+0x3d7/0x450 [amdgpu] [12622.089721] ?
>>>>> amdgpu_vm_free_mapping.isra.20+0x30/0x30 [amdgpu] [12622.090184]
>>>>> amdgpu_gem_va_ioctl+0x409/0x4f0 [amdgpu] [12622.090569] ?
>>>>> amdgpu_gem_metadata_ioctl+0x1b0/0x1b0 [amdgpu] [12622.090959]
>>>>> drm_ioctl_kernel+0x6a/0xb0 [drm] [12622.091247] ?
>>>>> kasan_check_write+0x14/0x20 [12622.091517] drm_ioctl+0x2ea/0x3b0
>>>>> [drm] [12622.091805] ? amdgpu_gem_metadata_ioctl+0x1b0/0x1b0
>>>>> [amdgpu] [12622.092212] amdgpu_drm_ioctl+0x4b/0x80 [amdgpu]
>>>>> [12622.092519]
>>>>> do_vfs_ioctl+0x96/0x5b0 [12622.092754] ?
>>>>> kvm_sched_clock_read+0x1e/0x30 [12622.093036] ?
>>>>> kvm_sched_clock_read+0x1e/0x30 [12622.093322] ?
>>>>> sched_clock+0x9/0x10 [12622.093550] SyS_ioctl+0x79/0x90 [12622.093812] ?
>>>>> vtime_user_exit+0x29/0x70 [12622.094108] do_syscall_64+0x6e/0x160
>>>>> [12622.094350] entry_SYSCALL64_slow_path+0x25/0x25
>>>>> [12622.094651] RIP: 0033:0x7f7012ebff07 [12622.094884] RSP:
>>>>> 002b:00007f70109fdae8 EFLAGS: 00000206 ORIG_RAX: 0000000000000010
>>>>> [12622.095369] RAX: ffffffffffffffda RBX: 00007f6fe100fe30 RCX:
>>>>> 00007f7012ebff07 [12622.095827] RDX: 00007f70109fdbb0 RSI:
>>>>> 00000000c0286448 RDI: 0000000000000004 [12622.096283] RBP:
>>>>> 00007f70109fdb20 R08: 0000000332000000 R09: 000000000000000e
>>>>> [12622.096739] R10: 0000000002e2e628 R11: 0000000000000206 R12:
>>>>> 0000000002ebb8f0 [12622.097195] R13: 00007f6fe100fe30 R14:
>>>>> 0000000002fe2bf0 R15: 0000000000000000 [12622.097676] Code: 2b e6
>>>>> ff ff 41 8b 4f 10 48 8b 75 c8 48 8b 55 d0 e9 06 ff ff ff 49 8d 44
>>>>> cf 10
>>>>> 45 31 ed 48 89 70 08 41 83 47 10 01 e9 12 ff ff ff <0f> 0b 90 90 0f
>>>>> 1f
>>>>> 44 00 00 55 31 c0 48 81 7f 08 00 df a9 ac 48 [12622.098919] RIP:
>>>>> reservation_object_add_shared_fence+0x2bc/0x2c0 RSP:
>>>>> ffff8801e6df7ae8 [12622.099527] ---[ end trace 8235addd8e22e444
>>>>> ]---
>>>>>
>>>>> Change-Id: I1de84f60c75a0d78b0cad244b8e2bff02aced9fa
>>>>> Signed-off-by: Monk Liu <Monk.Liu at amd.com>
>>>>> ---
>>>>> drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 5 +++++
>>>>> 1 file changed, 5 insertions(+)
>>>>>
>>>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>>>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>>>>> index 0b237e0..e6f159f 100644
>>>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>>>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>>>>> @@ -989,6 +989,11 @@ int amdgpu_vm_update_directories(struct amdgpu_device *adev,
>>>>> amdgpu_ring_pad_ib(ring, params.ib);
>>>>> amdgpu_sync_resv(adev, &job->sync, root->tbo.resv,
>>>>> AMDGPU_FENCE_OWNER_VM, false);
>>>>> +
>>>>> + r = reservation_object_reserve_shared(root->tbo.resv);
>>>>> + if (r)
>>>>> + goto error;
>>>>> +
>>>>> WARN_ON(params.ib->length_dw > ndw);
>>>>> r = amdgpu_job_submit(job, ring, &vm->entity,
>>>>> AMDGPU_FENCE_OWNER_VM, &fence);
>>> _______________________________________________
>>> amd-gfx mailing list
>>> amd-gfx at lists.freedesktop.org
>>> https://lists.freedesktop.org/mailman/listinfo/amd-gfx
More information about the amd-gfx
mailing list