Ah, crap yeah that won't work.
Yeah, we should block userptr BOs in amdgpu_ttm_bind() for now.
> There is something else still broken in our IPC test for system memory
> also, but I'll have to rewrite that anyway to work for large BOs with a
> small GART.
Hui? Well userptr BOs can't be used for IPC, so how do you run into this
issue?
Regards,
Christian.
Am 02.09.2017 um 01:46 schrieb Felix Kuehling:
> Actually, I read that wrong. It does GART bind userptr BOs. But it also
> unpins and repins them accidentally in the process, since
> ttm_bo_move_ttm calls backend_unbind and backend_bind callbacks. And
> ends up corrupting some kernel page state, because we had to move the
> pinning of user pages out of backend_bind.
>
> Maybe amdgpu_bind shouldn't be used for user mode BOs at all any more?
> Or does it need some safeguards for userptr BOs?
>
> Regards,
> Felix
>
>
> On 2017-09-01 07:35 PM, Felix Kuehling wrote:
>> It also seems to break GART binding of userptr BOs. Since backend_bind
>> doesn't do GART binding for userptr BOs, they now don't get GART bound
>> at all.
>>
>> There is something else still broken in our IPC test for system memory
>> also, but I'll have to rewrite that anyway to work for large BOs with a
>> small GART.
>>
>> Regards,
>> Felix
>>
>>
>> On 2017-09-01 03:17 AM, Christian König wrote:
>>> Yeah, the bug is really obvious after you pointed it out. We just need
>>> to keep the placement flags.
>>>
>>> Going to send a fix to the list in a minute,
>>> Christian.
>>>
>>> Am 01.09.2017 um 00:50 schrieb Felix Kuehling:
>>>> I believe this patch breaks pinned GTT BOs. It trips KFD tests that
>>>> apply a lot of memory pressure on one of our test systems.
>>>>
>>>> I understand the cause of the problem (see my analysis below), but I'm
>>>> not sure how to solve it, without reverting the patch. Christian, do you
>>>> have any ideas? Should we at least revert this as a short-term fix?
>>>>
>>>> I added some WARNs in the code to understand what's happening:
>>>>
>>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
>>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
>>>> index 72ad045..3ead155 100644
>>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
>>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
>>>> @@ -930,6 +930,7 @@ void amdgpu_bo_move_notify(struct
>>>> ttm_buffer_object *bo,
>>>> return;
>>>> abo = container_of(bo, struct amdgpu_bo, tbo);
>>>> + WARN(abo->pin_count, "Moving pinned bo");
>>>> amdgpu_vm_bo_invalidate(adev, abo);
>>>> amdgpu_bo_kunmap(abo);
>>>> diff --git a/drivers/gpu/drm/ttm/ttm_bo_util.c
>>>> b/drivers/gpu/drm/ttm/ttm_bo_util.c
>>>> index c934ad5..af9f40e 100644
>>>> --- a/drivers/gpu/drm/ttm/ttm_bo_util.c
>>>> +++ b/drivers/gpu/drm/ttm/ttm_bo_util.c
>>>> @@ -63,6 +63,7 @@ int ttm_bo_move_ttm(struct ttm_buffer_object *bo,
>>>> ttm_tt_unbind(ttm);
>>>> ttm_bo_free_old_node(bo);
>>>> + WARN(old_mem->placement & TTM_PL_FLAG_NO_EVICT,
>>>> "Stripping NO_EVICT flag from BO");
>>>> ttm_flag_masked(&old_mem->placement,
>>>> TTM_PL_FLAG_SYSTEM,
>>>> TTM_PL_MASK_MEM);
>>>> old_mem->mem_type = TTM_PL_SYSTEM;
>>>>
>>>> With that I see the following scenario unfold in my kernel log, where
>>>> pinned GTT BOs are swapped out and kunmapped under memory pressure.
>>>> Eventually this leads to memory access faults while accessing ring
>>>> buffers or IBs:
>>>>
>>>> [ 8.475453] Stripping NO_EVICT flag from BO
>>>> [ 8.475461] ------------[ cut here ]------------
>>>> [ 8.475465] WARNING: CPU: 5 PID: 2186 at
>>>> /home/fkuehlin/compute/kernel/drivers/gpu/drm/ttm/ttm_bo_util.c:66
>>>> ttm_bo_move_ttm+0x14a/0x160 [ttm]
>>>> [ 8.475466] Modules linked in: amdkfd(E) amd_iommu_v2(E)
>>>> amdgpu(E+) radeon(E) ttm(E)
>>>> [ 8.475472] CPU: 5 PID: 2186 Comm: systemd-udevd Tainted: G
>>>> W E 4.12.0-kfd-fkuehlin #1
>>>> [ 8.475473] Hardware name: ASUS All Series/X99-E WS/USB 3.1, BIOS
>>>> 2006 04/07/2016
>>>> [ 8.475474] task: ffff91bf12a82e00 task.stack: ffffb8bd43600000
>>>> [ 8.475477] RIP: 0010:ttm_bo_move_ttm+0x14a/0x160 [ttm]
>>>> [ 8.475478] RSP: 0018:ffffb8bd43603780 EFLAGS: 00010282
>>>> [ 8.475480] RAX: 000000000000001f RBX: ffff91bf0b0ee058 RCX:
>>>> 0000000000000006
>>>> [ 8.475481] RDX: 0000000000000006 RSI: ffff91bf12a83678 RDI:
>>>> 0000000000000202
>>>> [ 8.475482] RBP: ffffb8bd436037a0 R08: 0000000000000001 R09:
>>>> 0000000000000000
>>>> [ 8.475483] R10: 0000000000000000 R11: 0000000000000000 R12:
>>>> ffffb8bd436037e8
>>>> [ 8.475484] R13: 0000000000000000 R14: ffff91bf12c1bf00 R15:
>>>> ffff91bf0b0ee000
>>>> [ 8.475485] FS: 00007f71b381b8c0(0000) GS:ffff91bf17340000(0000)
>>>> knlGS:0000000000000000
>>>> [ 8.475486] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
>>>> [ 8.475487] CR2: 0000557e2ad63a28 CR3: 000000060d153000 CR4:
>>>> 00000000001406e0
>>>> [ 8.475488] Call Trace:
>>>> [ 8.475514] amdgpu_ttm_bind+0x129/0x1a0 [amdgpu]
>>>> [ 8.475518] ? ttm_bo_mem_compat+0x20/0x50 [ttm]
>>>> [ 8.475522] ? sparse_add_one_section+0x166/0x167
>>>> [ 8.475547] amdgpu_bo_pin_restricted+0x1ca/0x2f0 [amdgpu]
>>>> [ 8.475575] amdgpu_bo_create_reserved+0x95/0x220 [amdgpu]
>>>> [ 8.475602] amdgpu_bo_create_kernel+0x10/0x70 [amdgpu]
>>>> [ 8.475632] amdgpu_gfx_compute_mqd_sw_init+0xf2/0x170 [amdgpu]
>>>> [ 8.475663] gfx_v8_0_sw_init+0xca8/0x15d0 [amdgpu]
>>>> [ 8.475690] amdgpu_device_init+0xe6f/0x1620 [amdgpu]
>>>> [ 8.475717] amdgpu_driver_load_kms+0x63/0x200 [amdgpu]
>>>> [ 8.475720] drm_dev_register+0x143/0x1d0
>>>> [ 8.475744] amdgpu_pci_probe+0x110/0x140 [amdgpu]
>>>> [ 8.475747] local_pci_probe+0x40/0xa0
>>>> [ 8.475749] ? pci_match_device+0xdb/0x100
>>>> [ 8.475752] pci_device_probe+0x136/0x160
>>>> [ 8.475755] driver_probe_device+0x299/0x440
>>>> [ 8.475757] __driver_attach+0xde/0xe0
>>>> [ 8.475759] ? driver_probe_device+0x440/0x440
>>>> [ 8.475761] bus_for_each_dev+0x61/0xa0
>>>> [ 8.475763] driver_attach+0x19/0x20
>>>> [ 8.475764] bus_add_driver+0x1f6/0x270
>>>> [ 8.475766] ? 0xffffffffc0839000
>>>> [ 8.475768] driver_register+0x5b/0xd0
>>>> [ 8.475769] ? 0xffffffffc0839000
>>>> [ 8.475771] __pci_register_driver+0x5b/0x60
>>>> [ 8.475799] amdgpu_init+0x8a/0x1000 [amdgpu]
>>>> [ 8.475801] do_one_initcall+0x4e/0x190
>>>> [ 8.475804] ? kmem_cache_alloc+0xd8/0x170
>>>> [ 8.475807] do_init_module+0x55/0x1e9
>>>> [ 8.475810] load_module+0x2771/0x29c0
>>>> [ 8.475817] SYSC_finit_module+0xbc/0xf0
>>>> [ 8.475819] ? SYSC_finit_module+0xbc/0xf0
>>>> [ 8.475823] SyS_finit_module+0x9/0x10
>>>> [ 8.475825] entry_SYSCALL_64_fastpath+0x1f/0xbe
>>>> [...]
>>>> [ 278.693790] Moving pinned bo (this kunmaps the BO)
>>>> [ 278.693802] ------------[ cut here ]------------
>>>> [ 278.693833] WARNING: CPU: 11 PID: 2145 at
>>>> /home/fkuehlin/compute/kernel/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c:933
>>>> amdgpu_bo_move_notify+0x7c/0x90 [amdgpu]
>>>> [ 278.693834] Modules linked in: fuse(E) x86_pkg_temp_thermal(E)
>>>> amdkfd(E) amd_iommu_v2(E) amdgpu(E) radeon(E) ttm(E)
>>>> [ 278.693843] CPU: 11 PID: 2145 Comm: kworker/11:2 Tainted: G
>>>> W E 4.12.0-kfd-fkuehlin #1
>>>> [ 278.693845] Hardware name: ASUS All Series/X99-E WS/USB 3.1, BIOS
>>>> 2006 04/07/2016
>>>> [ 278.693851] Workqueue: kfd_process_wq kfd_process_wq_release [amdkfd]
>>>> [ 278.693853] task: ffff91bf0dab1700 task.stack: ffffb8bd4495c000
>>>> [ 278.693880] RIP: 0010:amdgpu_bo_move_notify+0x7c/0x90 [amdgpu]
>>>> [ 278.693882] RSP: 0018:ffffb8bd4495fc20 EFLAGS: 00010282
>>>> [ 278.693884] RAX: 0000000000000010 RBX: ffff91bf0bec7058 RCX:
>>>> 0000000000000006
>>>> [ 278.693885] RDX: 0000000000000006 RSI: ffff91bf0dab1f50 RDI:
>>>> 0000000000000202
>>>> [ 278.693886] RBP: ffffb8bd4495fc40 R08: 0000000000000001 R09:
>>>> 0000000000000000
>>>> [ 278.693887] R10: ffffb8bd4495fb98 R11: 0000000000000000 R12:
>>>> ffff91bf0b0c2a68
>>>> [ 278.693889] R13: 0000000000000000 R14: ffff91bf0bec708c R15:
>>>> ffff91bf104bb000
>>>> [ 278.693890] FS: 0000000000000000(0000) GS:ffff91bf174c0000(0000)
>>>> knlGS:0000000000000000
>>>> [ 278.693891] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
>>>> [ 278.693892] CR2: 0000000004199000 CR3: 00000001f3e0f000 CR4:
>>>> 00000000001406e0
>>>> [ 278.693893] Call Trace:
>>>> [ 278.693898] ttm_bo_cleanup_memtype_use+0x1f/0x70 [ttm]
>>>> [ 278.693902] ttm_bo_unref+0x388/0x470 [ttm]
>>>> [ 278.693931] amdgpu_bo_unref+0x25/0x40 [amdgpu]
>>>> [ 278.693962] amdgpu_amdkfd_gpuvm_free_memory_of_gpu+0x1b5/0x260
>>>> [amdgpu]
>>>> [ 278.693974] kfd_process_free_outstanding_kfd_bos+0xab/0xe0 [amdkfd]
>>>> [ 278.693980] kfd_process_wq_release+0x56/0x110 [amdkfd]
>>>> [ 278.693983] process_one_work+0x190/0x430
>>>> [ 278.693984] ? process_one_work+0x133/0x430
>>>> [ 278.693988] worker_thread+0x46/0x400
>>>> [ 278.693991] kthread+0x10a/0x140
>>>> [ 278.693993] ? process_one_work+0x430/0x430
>>>> [ 278.693995] ? kthread_create_on_node+0x40/0x40
>>>> [ 278.693996] ? umh_complete+0x40/0x40
>>>> [ 278.693998] ? call_usermodehelper_exec_async+0x137/0x140
>>>> [ 278.694000] ret_from_fork+0x2a/0x40
>>>> [...]
>>>> [ 694.740232] BUG: unable to handle kernel paging request at
>>>> ffffb8bd4354d540
>>>> [ 694.740387] IP: sdma_v3_0_ring_set_wptr+0x64/0x80 [amdgpu]
>>>> [ 694.740430] PGD 616c23067
>>>> [ 694.740432] P4D 616c23067
>>>> [ 694.740454] PUD 616c24067
>>>> [ 694.740476] PMD 613704067
>>>> [ 694.740498] PTE 0
>>>>
>>>> [ 694.740551] Oops: 0002 [#1] SMP
>>>> [ 694.740576] Modules linked in: fuse(E) x86_pkg_temp_thermal(E)
>>>> amdkfd(E) amd_iommu_v2(E) amdgpu(E) radeon(E) ttm(E)
>>>> [ 694.740667] CPU: 10 PID: 2213 Comm: sdma0 Tainted: G W
>>>> E 4.12.0-kfd-fkuehlin #1
>>>> [ 694.740727] Hardware name: ASUS All Series/X99-E WS/USB 3.1, BIOS
>>>> 2006 04/07/2016
>>>> [ 694.740782] task: ffff91bf0d989700 task.stack: ffffb8bd43730000
>>>> [ 694.740895] RIP: 0010:sdma_v3_0_ring_set_wptr+0x64/0x80 [amdgpu]
>>>> [ 694.740941] RSP: 0018:ffffb8bd43733dc0 EFLAGS: 00010202
>>>> [ 694.740982] RAX: ffff91bf0b0c7770 RBX: ffff91bf0b0c7770 RCX:
>>>> ffffb8bd4354d540
>>>> [ 694.741035] RDX: 0000000000001ac0 RSI: 00000000000006b0 RDI:
>>>> ffff91bf0b0c0000
>>>> [ 694.741087] RBP: ffffb8bd43733dc0 R08: ffffffffc072bd60 R09:
>>>> 0000000000000000
>>>> [ 694.741140] R10: ffff91bf0b0c0000 R11: ffffffffc064ef47 R12:
>>>> ffff91bef6166000
>>>> [ 694.741192] R13: ffff91bef6166368 R14: 0000000000000001 R15:
>>>> ffff91bf0b0c7770
>>>> [ 694.741246] FS: 0000000000000000(0000) GS:ffff91bf17480000(0000)
>>>> knlGS:0000000000000000
>>>> [ 694.741305] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
>>>> [ 694.741349] CR2: ffffb8bd4354d540 CR3: 000000060e4a6000 CR4:
>>>> 00000000001406e0
>>>> [ 694.741401] Call Trace:
>>>> [ 694.741480] amdgpu_ring_commit+0x3a/0x60 [amdgpu]
>>>> [ 694.741575] amdgpu_ib_schedule+0x2c7/0x450 [amdgpu]
>>>> [ 694.741687] amdgpu_job_run+0x72/0x110 [amdgpu]
>>>> [ 694.741793] amd_sched_main+0x343/0x470 [amdgpu]
>>>> [ 694.741838] ? wake_atomic_t_function+0x60/0x60
>>>> [ 694.741876] kthread+0x10a/0x140
>>>> [ 694.741969] ? amd_sched_process_job+0x60/0x60 [amdgpu]
>>>> [ 694.742010] ? kthread_create_on_node+0x40/0x40
>>>> [ 694.742048] ret_from_fork+0x2a/0x40
>>>>
>>>>
>>>>
>>>> On 2017-08-23 05:02 AM, Christian König wrote:
>>>>> From: Christian König <christian.koenig at amd.com>
>>>>>
>>>>> Use ttm_bo_mem_space instead of manually allocating GART space.
>>>>>
>>>>> This allows us to evict BOs when there isn't enought GART space any
>>>>> more.
>>>>>
>>>>> Signed-off-by: Christian König <christian.koenig at amd.com>
>>>>> ---
>>>>> drivers/gpu/drm/amd/amdgpu/amdgpu_gtt_mgr.c | 14 +++++--------
>>>>> drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 31
>>>>> +++++++++++++++++++++++------
>>>>> drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h | 4 ----
>>>>> 3 files changed, 30 insertions(+), 19 deletions(-)
>>>>>
>>>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gtt_mgr.c
>>>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_gtt_mgr.c
>>>>> index 9e05e25..0d15eb7 100644
>>>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gtt_mgr.c
>>>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gtt_mgr.c
>>>>> @@ -108,10 +108,10 @@ bool amdgpu_gtt_mgr_is_allocated(struct
>>>>> ttm_mem_reg *mem)
>>>>> *
>>>>> * Allocate the address space for a node.
>>>>> */
>>>>> -int amdgpu_gtt_mgr_alloc(struct ttm_mem_type_manager *man,
>>>>> - struct ttm_buffer_object *tbo,
>>>>> - const struct ttm_place *place,
>>>>> - struct ttm_mem_reg *mem)
>>>>> +static int amdgpu_gtt_mgr_alloc(struct ttm_mem_type_manager *man,
>>>>> + struct ttm_buffer_object *tbo,
>>>>> + const struct ttm_place *place,
>>>>> + struct ttm_mem_reg *mem)
>>>>> {
>>>>> struct amdgpu_device *adev = amdgpu_ttm_adev(man->bdev);
>>>>> struct amdgpu_gtt_mgr *mgr = man->priv;
>>>>> @@ -143,12 +143,8 @@ int amdgpu_gtt_mgr_alloc(struct
>>>>> ttm_mem_type_manager *man,
>>>>> fpfn, lpfn, mode);
>>>>> spin_unlock(&mgr->lock);
>>>>> - if (!r) {
>>>>> + if (!r)
>>>>> mem->start = node->start;
>>>>> - if (&tbo->mem == mem)
>>>>> - tbo->offset = (tbo->mem.start << PAGE_SHIFT) +
>>>>> - tbo->bdev->man[tbo->mem.mem_type].gpu_offset;
>>>>> - }
>>>>> return r;
>>>>> }
>>>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
>>>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
>>>>> index 38d26a7..aa0b7f5 100644
>>>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
>>>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
>>>>> @@ -809,20 +809,39 @@ bool amdgpu_ttm_is_bound(struct ttm_tt *ttm)
>>>>> int amdgpu_ttm_bind(struct ttm_buffer_object *bo, struct
>>>>> ttm_mem_reg *bo_mem)
>>>>> {
>>>>> + struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
>>>>> struct ttm_tt *ttm = bo->ttm;
>>>>> + struct ttm_mem_reg tmp;
>>>>> +
>>>>> + struct ttm_placement placement;
>>>>> + struct ttm_place placements;
>>>>> int r;
>>>>> if (!ttm || amdgpu_ttm_is_bound(ttm))
>>>>> return 0;
>>>>> - r = amdgpu_gtt_mgr_alloc(&bo->bdev->man[TTM_PL_TT], bo,
>>>>> - NULL, bo_mem);
>>>>> - if (r) {
>>>>> - DRM_ERROR("Failed to allocate GTT address space (%d)\n", r);
>>>>> + tmp = bo->mem;
>>>>> + tmp.mm_node = NULL;
>>>>> + placement.num_placement = 1;
>>>>> + placement.placement = &placements;
>>>>> + placement.num_busy_placement = 1;
>>>>> + placement.busy_placement = &placements;
>>>>> + placements.fpfn = 0;
>>>>> + placements.lpfn = adev->mc.gart_size >> PAGE_SHIFT;
>>>>> + placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
>>>>> +
>>>>> + r = ttm_bo_mem_space(bo, &placement, &tmp, true, false);
>>>>> + if (unlikely(r))
>>>>> return r;
>>>>> - }
>>>>> - return amdgpu_ttm_do_bind(ttm, bo_mem);
>>>>> + r = ttm_bo_move_ttm(bo, true, false, &tmp);
>>>>> + if (unlikely(r))
>>>>> + ttm_bo_mem_put(bo, &tmp);
>>>>> + else
>>>>> + bo->offset = (bo->mem.start << PAGE_SHIFT) +
>>>>> + bo->bdev->man[bo->mem.mem_type].gpu_offset;
>>>>> +
>>>>> + return r;
>>>>> }
>>>>> int amdgpu_ttm_recover_gart(struct amdgpu_device *adev)
>>>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h
>>>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h
>>>>> index f22a475..43093bf 100644
>>>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h
>>>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h
>>>>> @@ -62,10 +62,6 @@ extern const struct ttm_mem_type_manager_func
>>>>> amdgpu_gtt_mgr_func;
>>>>> extern const struct ttm_mem_type_manager_func amdgpu_vram_mgr_func;
>>>>> bool amdgpu_gtt_mgr_is_allocated(struct ttm_mem_reg *mem);
>>>>> -int amdgpu_gtt_mgr_alloc(struct ttm_mem_type_manager *man,
>>>>> - struct ttm_buffer_object *tbo,
>>>>> - const struct ttm_place *place,
>>>>> - struct ttm_mem_reg *mem);
>>>>> uint64_t amdgpu_gtt_mgr_usage(struct ttm_mem_type_manager *man);
>>>>> uint64_t amdgpu_vram_mgr_usage(struct ttm_mem_type_manager *man);
>>>> _______________________________________________
>>>> amd-gfx mailing list
>>>> amd-gfx at lists.freedesktop.org
>>>> https://lists.freedesktop.org/mailman/listinfo/amd-gfx
>> _______________________________________________
>> amd-gfx mailing list
>> amd-gfx at lists.freedesktop.org
>> https://lists.freedesktop.org/mailman/listinfo/amd-gfx
