Re: [PATCH] drm/amdkfd: Fix dmabuf's redundant eviction when unmapping

[Eric Huang <jinhuieric.huang@xxxxxxx>]

Hi Felix,

What do you think my proposal in my previous email? that setting domain 
to CPU in kfd_mem_dmamap_dmabuf, and setting domain to GTT in 
kfd_mem_dmaunmap_dmabuf, that will be doing the similar way as userptr.

Thanks,
Eric

On 2023-04-10 14:50, Felix Kuehling wrote:
> Sorry, you're right, there is no AMDGPU_GEM_DOMAIN_PREEMPTIBLE. I 
> remembered this wrong. There is a flag called 
> AMDGPU_GEM_CREATE_PREEMPTIBLE, which changes what happens when is 
> placed in the AMDGPU_GEM_DOMAIN_GTT domain.
>
> So my proposal would need to be modified to set the flag 
> AMDGPU_GEM_CREATE_PREEMPTIBLE in the imported DMABuf BO.
>
> On 2023-04-10 14:28, Eric Huang wrote:
> > Hi Felix,
> >
> > Thanks for your review and suggestion, but unfortunately the 
> > AMDGPU_GEM_DOMAIN_PREEMPTIBLE is not defined in amdgpu_drm.h. I 
> > understand we need the memory eviction on either 
> > kfd_mem_dmamap_dmabuf() or kfd_mem_dmaunmap_dmabuf() to update DMA 
> > address, so I am thinking to do it as simply as userptr memory does.
> >
> > The purpose for this change is for non-MES HW scheduler we are using 
> > userptr/paged memory, but since GFX11 we will be using MES scheduler 
> > and it needs the memory to be allocated as GTT/non-paged memory, so 
> > we want all GPUs using GTT/non-paged memory, but there is performance 
> > drop, because of eviction in kfd_mem_dmaunmap_dmabuf.
> >
> > Currently userptr memory is evicted in kfd_mem_dmamap_userptr as 
> > changing domain to GTT before calling ttm_bo_validate, and not 
> > evicted in kfd_mem_dmamap_userptr, so I think we can do the similar 
> > way for GTT/non-paged memory that setting domain to CPU in 
> > kfd_mem_dmamap_dmabuf, which will evict memory to update DMA address, 
> > and setting domain to GTT in kfd_mem_dmaunmap_dmabuf, which will not 
> > evict memory. The performance should be the same as userptr/paged 
> > memory.
>
> This sounds backwards to me. dmaunmap should move objects to the CPU 
> domain because the GPU mapping is potentially invalid. And dmamap must 
> use move it to the GTT domain because that updates the GPU mapping and 
> allows the GPU virtual address mapping to be updated.
>
> The problem is the eviction in dmaunmap. Userptrs don't see these 
> evictions because the SG BOs we use to map them on other GPUs do set 
> the AMDGPU_GEM_CREATE_PREEMPTIBLE flag. My idea is to do the same 
> thing for DMABufs that map GTT (and VRAM) BOs to other GPUs._
>
> Now that I look at it in more detail, I see we're already doing that 
> in kfd_mem_attach_dmabuf:
>
>         *bo = gem_to_amdgpu_bo(gobj);
>         (*bo)->flags |= AMDGPU_GEM_CREATE_PREEMPTIBLE;
>
> So then the question is, why is this not working? I think that's the 
> second part of my proposal, which is still needed:
>
> > 2. Add a special case in the above if-block for old_mem->mem_type ==
> >    AMDGPU_PL_PREEMPT: use amdgpu_bo_sync_wait with
> >    owner=AMDGPU_FENCE_OWNER_KFD so that it doesn't wait for eviction 
> > fences 
>
> Regards,
>   Felix
>
>
> > Regards,
> > Eric
> >
> > On 2023-04-04 16:40, Felix Kuehling wrote:
> > > [+Christian]
> > >
> > > OK, this comes from the ttm_bo_wait_ctx call in this section of 
> > > amdgpu_bo_move:
> > >
> > >         if ((old_mem->mem_type == TTM_PL_TT ||
> > >              old_mem->mem_type == AMDGPU_PL_PREEMPT) &&
> > >              new_mem->mem_type == TTM_PL_SYSTEM) {
> > >                 r = ttm_bo_wait_ctx(bo, ctx);
> > >                 if (r)
> > >                         return r;
> > >
> > >                 amdgpu_ttm_backend_unbind(bo->bdev, bo->ttm);
> > >                 ttm_resource_free(bo, &bo->resource);
> > >                 ttm_bo_assign_mem(bo, new_mem);
> > >                 goto out;
> > >         }
> > >
> > > We can't just remove this wait. It's not even specific to KFD or 
> > > DMABuf imports. We also can't just change it to avoid waiting for 
> > > eviction fences because it's also used for GTT BOs (e.g. before a BO 
> > > gets swapped under extreme memory pressure). So we also need to 
> > > trigger the eviction fence in general case.
> > >
> > > In the specific case of DMABuf imports, they share the reservation 
> > > object with the original BO. So waiting on the reservation triggers 
> > > the eviction fence on the original BO. I think we want to avoid the 
> > > waiting on eviction fences for all BOs where the underlying memory 
> > > is managed by some other BO, and at the same time also avoid ever 
> > > evicting the DMABuf import BO. That's what AMDGPU_PL_PREEMPT is for. 
> > > So I think a combination of two changes should to the trick:
> > >
> > > 1. Change kfd_mem_dmamap_dmabuf to use AMDGPU_GEM_DOMAIN_PREEMPTIBLE
> > > 2. Add a special case in the above if-block for old_mem->mem_type ==
> > >    AMDGPU_PL_PREEMPT: use amdgpu_bo_sync_wait with
> > >    owner=AMDGPU_FENCE_OWNER_KFD so that it doesn't wait for eviction 
> > > fences
> > >
> > > Regards,
> > >   Felix
> > >
> > >
> > > Am 2023-04-04 um 10:36 schrieb Eric Huang:
> > > > Here is the backtrace from Jira:
> > > >
> > > > Thu Nov 10 13:10:23 2022] Scheduling eviction of pid 97784 in 0 
> > > > jiffies
> > > > [Thu Nov 10 13:10:23 2022] WARNING: CPU: 173 PID: 97784 at 
> > > > /var/lib/dkms/amdgpu/5.16.9.22.20-1438746~20.04/build/amd/amdgpu/../amdkfd/kfd_device.c:878 
> > > > kgd2kfd_schedule_evict_and_restore_process+0x104/0x120 [amdgpu]
> > > > [Thu Nov 10 13:10:23 2022] Modules linked in: veth amdgpu(OE) 
> > > > amddrm_ttm_helper(OE) amdttm(OE) iommu_v2 amd_sched(OE) amdkcl(OE) 
> > > > xt_conntrack xt_MASQUERADE nf_conntrack_netlink nfnetlink xfrm_user 
> > > > xfrm_algo xt_addrtype iptable_filter iptable_nat nf_nat 
> > > > nf_conntrack nf_defrag_ipv6 nf_defrag_ipv4 bpfilter br_netfilter 
> > > > bridge stp llc aufs overlay binfmt_misc nls_iso8859_1 dm_multipath 
> > > > scsi_dh_rdac scsi_dh_emc scsi_dh_alua intel_rapl_msr 
> > > > intel_rapl_common amd64_edac edac_mce_amd kvm_amd kvm efi_pstore 
> > > > rapl ipmi_ssif ccp acpi_ipmi k10temp ipmi_si ipmi_devintf 
> > > > ipmi_msghandler mac_hid sch_fq_codel msr ip_tables x_tables autofs4 
> > > > btrfs blake2b_generic zstd_compress raid10 raid456 
> > > > async_raid6_recov async_memcpy async_pq async_xor async_tx xor 
> > > > raid6_pq libcrc32c raid1 raid0 multipath linear mlx5_ib ib_uverbs 
> > > > ib_core crct10dif_pclmul crc32_pclmul ghash_clmulni_intel 
> > > > aesni_intel crypto_simd cryptd ast drm_vram_helper drm_ttm_helper 
> > > > ttm mlx5_core drm_kms_helper syscopyarea sysfillrect sysimgblt 
> > > > fb_sys_fops
> > > > [Thu Nov 10 13:10:23 2022]  pci_hyperv_intf cec psample igb mlxfw 
> > > > rc_core dca ahci xhci_pci tls drm i2c_algo_bit libahci 
> > > > xhci_pci_renesas i2c_piix4
> > > > [Thu Nov 10 13:10:23 2022] CPU: 173 PID: 97784 Comm: 
> > > > onnxruntime_tes Tainted: G        W  OE 5.13.0-30-generic 
> > > > #33~20.04.1-Ubuntu
> > > > [Thu Nov 10 13:10:23 2022] Hardware name: GIGABYTE 
> > > > G482-Z53-YF/MZ52-G40-00, BIOS R12 05/13/2020
> > > > [Thu Nov 10 13:10:23 2022] RIP: 
> > > > 0010:kgd2kfd_schedule_evict_and_restore_process+0x104/0x120 [amdgpu]
> > > > [Thu Nov 10 13:10:23 2022] Code: 5e 5d c3 4c 89 e7 e8 cb c6 44 df 
> > > > eb e7 49 8b 45 60 48 89 ca 48 c7 c7 38 8b d7 c1 48 89 4d e0 8b b0 
> > > > 20 09 00 00 e8 87 ee 7e df <0f> 0b 48 8b 4d e0 eb 9f 41 be ea ff ff 
> > > > ff eb ba 41 be ed ff ff ff
> > > > [Thu Nov 10 13:10:23 2022] RSP: 0018:ffffb25f2a173978 EFLAGS: 00010086
> > > > [Thu Nov 10 13:10:23 2022] RAX: 0000000000000000 RBX: 
> > > > 0000000000000001 RCX: 0000000000000027
> > > > [Thu Nov 10 13:10:23 2022] RDX: 0000000000000027 RSI: 
> > > > 00000000fffeffff RDI: ffff95d06e4a09c8
> > > > [Thu Nov 10 13:10:23 2022] RBP: ffffb25f2a173998 R08: 
> > > > ffff95d06e4a09c0 R09: ffffb25f2a173750
> > > > [Thu Nov 10 13:10:23 2022] R10: 0000000000000001 R11: 
> > > > 0000000000000001 R12: ffff95c371d74580
> > > > [Thu Nov 10 13:10:23 2022] R13: ffff95b1cd3f2000 R14: 
> > > > 0000000000000000 R15: ffff95c371d74580
> > > > [Thu Nov 10 13:10:23 2022] FS:  00007fcaff268b00(0000) 
> > > > GS:ffff95d06e480000(0000) knlGS:0000000000000000
> > > > [Thu Nov 10 13:10:23 2022] CS:  0010 DS: 0000 ES: 0000 CR0: 
> > > > 0000000080050033
> > > > [Thu Nov 10 13:10:23 2022] CR2: 00007fc643980000 CR3: 
> > > > 00000003e9492000 CR4: 0000000000350ee0
> > > > [Thu Nov 10 13:10:23 2022] Call Trace:
> > > > [Thu Nov 10 13:10:23 2022]  <TASK>
> > > > [Thu Nov 10 13:10:23 2022]  amdkfd_fence_enable_signaling+0x46/0x50 
> > > > [amdgpu]
> > > > [Thu Nov 10 13:10:23 2022]  __dma_fence_enable_signaling+0x52/0xb0
> > > > [Thu Nov 10 13:10:23 2022]  dma_fence_default_wait+0xa9/0x200
> > > > [Thu Nov 10 13:10:23 2022]  dma_fence_wait_timeout+0xbd/0xe0
> > > > [Thu Nov 10 13:10:23 2022]  amddma_resv_wait_timeout+0x6f/0xd0 
> > > > [amdkcl]
> > > > [Thu Nov 10 13:10:23 2022]  amdttm_bo_wait+0x39/0x50 [amdttm]
> > > > [Thu Nov 10 13:10:23 2022]  amdgpu_bo_move+0x41e/0x7b0 [amdgpu]
> > > > [Thu Nov 10 13:10:23 2022]  ? down_write+0x13/0x50
> > > > [Thu Nov 10 13:10:23 2022]  ? unmap_mapping_pages+0x68/0x130
> > > > [Thu Nov 10 13:10:23 2022]  ttm_bo_handle_move_mem+0x7f/0x120 [amdttm]
> > > > [Thu Nov 10 13:10:23 2022]  amdttm_bo_validate+0xbf/0x100 [amdttm]
> > > > [Thu Nov 10 13:10:23 2022]  kfd_mem_dmaunmap_attachment+0x131/0x140 
> > > > [amdgpu]
> > > > [Thu Nov 10 13:10:23 2022]  unmap_bo_from_gpuvm+0x67/0x80 [amdgpu]
> > > > [Thu Nov 10 13:10:23 2022] 
> > > >  amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu+0x114/0x220 [amdgpu]
> > > > [Thu Nov 10 13:10:23 2022]  ? __mod_memcg_lruvec_state+0x22/0xe0
> > > > [Thu Nov 10 13:10:23 2022] 
> > > >  kfd_ioctl_unmap_memory_from_gpu+0xe8/0x270 [amdgpu]
> > > > [Thu Nov 10 13:10:23 2022]  kfd_ioctl+0x23c/0x590 [amdgpu]
> > > > [Thu Nov 10 13:10:23 2022]  ? 
> > > > kfd_ioctl_get_process_apertures_new+0x330/0x330 [amdgpu]
> > > > [Thu Nov 10 13:10:23 2022]  ? exit_to_user_mode_prepare+0x3d/0x1c0
> > > > [Thu Nov 10 13:10:23 2022]  ? __fget_files+0xa7/0xd0
> > > > [Thu Nov 10 13:10:23 2022]  __x64_sys_ioctl+0x91/0xc0
> > > > [Thu Nov 10 13:10:23 2022]  do_syscall_64+0x61/0xb0
> > > > [Thu Nov 10 13:10:23 2022]  ? do_syscall_64+0x6e/0xb0
> > > > [Thu Nov 10 13:10:23 2022]  ? do_syscall_64+0x6e/0xb0
> > > > [Thu Nov 10 13:10:23 2022]  ? do_syscall_64+0x6e/0xb0
> > > > [Thu Nov 10 13:10:23 2022]  ? do_syscall_64+0x6e/0xb0
> > > > [Thu Nov 10 13:10:23 2022]  ? asm_sysvec_apic_timer_interrupt+0xa/0x20
> > > > [Thu Nov 10 13:10:23 2022]  entry_SYSCALL_64_after_hwframe+0x44/0xae
> > > > [Thu Nov 10 13:10:23 2022] RIP: 0033:0x7fcaff57b3ab
> > > > [Thu Nov 10 13:10:23 2022] Code: 0f 1e fa 48 8b 05 e5 7a 0d 00 64 
> > > > c7 00 26 00 00 00 48 c7 c0 ff ff ff ff c3 66 0f 1f 44 00 00 f3 0f 
> > > > 1e fa b8 10 00 00 00 0f 05 <48> 3d 01 f0 ff ff 73 01 c3 48 8b 0d b5 
> > > > 7a 0d 00 f7 d8 64 89 01 48
> > > > [Thu Nov 10 13:10:23 2022] RSP: 002b:00007fffe41e0098 EFLAGS: 
> > > > 00000206 ORIG_RAX: 0000000000000010
> > > > [Thu Nov 10 13:10:23 2022] RAX: ffffffffffffffda RBX: 
> > > > 00007fcacc7f7f80 RCX: 00007fcaff57b3ab
> > > > [Thu Nov 10 13:10:23 2022] RDX: 00007fffe41e0120 RSI: 
> > > > 00000000c0184b19 RDI: 0000000000000003
> > > > [Thu Nov 10 13:10:23 2022] RBP: 00007fffe41e00d0 R08: 
> > > > 0000562e2d5730d0 R09: 0000000000000000
> > > > [Thu Nov 10 13:10:23 2022] R10: 0000562e2c928ec0 R11: 
> > > > 0000000000000206 R12: 0000000000000001
> > > > [Thu Nov 10 13:10:23 2022] R13: 00007fffe41e04b0 R14: 
> > > > 0000000000000000 R15: 0000562e2d3f5b20
> > > > [Thu Nov 10 13:10:23 2022]  </TASK>
> > > > [Thu Nov 10 13:10:23 2022] ---[ end trace 1464f08f6be60b30 ]---
> > > >
> > > > Regards,
> > > > Eric
> > > >
> > > > On 2023-04-04 10:11, Felix Kuehling wrote:
> > > > > If we keep the BO in the GTT domain, it means it will not be 
> > > > > updated if we validate it again later in kfd_mem_dmamap_dmabuf. 
> > > > > This means we'll use stale DMA addresses when we update the page 
> > > > > tables after evictions.
> > > > >
> > > > > I think we'll need to find a different way to avoid triggering the 
> > > > > eviction fence on the original BO when changing the placement of 
> > > > > the DMABuf import here. If you need help brainstorming here, 
> > > > > please share a backtrace from the eviction generated with the 
> > > > > debug_evictions module param.
> > > > >
> > > > > Regards,
> > > > >   Felix
> > > > >
> > > > >
> > > > > Am 2023-04-03 um 13:59 schrieb Eric Huang:
> > > > > > dmabuf is allocated/mapped as GTT domain, when dma-unmapping dmabuf
> > > > > > changing placement to CPU will trigger memory eviction after calling
> > > > > > ttm_bo_validate, and the eviction will cause performance drop.
> > > > > > Keeping the correct domain will solve the issue.
> > > > > >
> > > > > > Signed-off-by: Eric Huang <jinhuieric.huang@xxxxxxx>
> > > > > > ---
> > > > > >   drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 2 +-
> > > > > >   1 file changed, 1 insertion(+), 1 deletion(-)
> > > > > >
> > > > > > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c 
> > > > > > b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
> > > > > > index a3b09edfd1bf..17b708acb447 100644
> > > > > > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
> > > > > > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
> > > > > > @@ -642,7 +642,7 @@ kfd_mem_dmaunmap_dmabuf(struct 
> > > > > > kfd_mem_attachment *attachment)
> > > > > >       struct ttm_operation_ctx ctx = {.interruptible = true};
> > > > > >       struct amdgpu_bo *bo = attachment->bo_va->base.bo;
> > > > > >   -    amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
> > > > > > +    amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
> > > > > >       ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
> > > > > >   }