Add estimate of how much vram we need to reserve for RAS when caculating the total available vram. v2: apply the change to MP0 v13_0_2 and v13_0_14 Signed-off-by: Hawking Zhang <Hawking.Zhang@amd.com> Reviewed-by: Tao Zhou <tao.zhou1@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
With commit 89773b85599a
("drm/amdkfd: Let VRAM allocations go to GTT domain on small APUs")
big and small APU "VRAM" handling in KFD was unified. Since AMD_IS_APU
is set for both big and small APUs, we can simplify the checks in
the code.
v2: clean up a few more places (Lang)
Acked-by: Felix Kuehling <felix.kuehling@amd.com>
Reviewed-by: Lang Yu <Lang.Yu@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
This reverts commit 8a774fe912ff ("drm/amdgpu: avoid restore process run into dead loop")
since buffer got pinned is not related whether it needs mapping
And skip buffer validation at kfd driver if the buffer has been pinned.
Signed-off-by: Xiaogang Chen <Xiaogang.Chen@amd.com>
Reviewed-by: Felix Kuehling <felix.kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
On system with khugepaged enabled and user cases with THP buffer, the hmm_range_fault may takes > 15 seconds to return -EBUSY, the arbitrary timeout value is not accurate, cause memory allocation failure. Remove the arbitrary timeout value, return EAGAIN to application if hmm_range_fault return EBUSY, then userspace libdrm and Thunk will call ioctl again. Change EAGAIN to debug message as this is not error. Signed-off-by: Philip Yang <Philip.Yang@amd.com> Reviewed-by: Felix Kuehling <felix.kuehling@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Small APUs(i.e., consumer, embedded products) usually have a small carveout device memory which can't satisfy most compute workloads memory allocation requirements. We can't even run a Basic MNIST Example with a default 512MB carveout. https://github.com/pytorch/examples/tree/main/mnist. Error Log: "torch.cuda.OutOfMemoryError: HIP out of memory. Tried to allocate 84.00 MiB. GPU 0 has a total capacity of 512.00 MiB of which 0 bytes is free. Of the allocated memory 103.83 MiB is allocated by PyTorch, and 22.17 MiB is reserved by PyTorch but unallocated" Though we can change BIOS settings to enlarge carveout size, which is inflexible and may bring complaint. On the other hand, the memory resource can't be effectively used between host and device. The solution is MI300A approach, i.e., let VRAM allocations go to GTT. Then device and host can flexibly and effectively share memory resource. v2: Report local_mem_size_private as 0. (Felix) Signed-off-by: Lang Yu <Lang.Yu@amd.com> Reviewed-by: Felix Kuehling <felix.kuehling@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Observed on gfx8 ASIC where KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM is used. Two attachments use the same VM, root PD would be locked twice. [ 57.910418] Call Trace: [ 57.793726] ? reserve_bo_and_cond_vms+0x111/0x1c0 [amdgpu] [ 57.793820] amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu+0x6c/0x1c0 [amdgpu] [ 57.793923] ? idr_get_next_ul+0xbe/0x100 [ 57.793933] kfd_process_device_free_bos+0x7e/0xf0 [amdgpu] [ 57.794041] kfd_process_wq_release+0x2ae/0x3c0 [amdgpu] [ 57.794141] ? process_scheduled_works+0x29c/0x580 [ 57.794147] process_scheduled_works+0x303/0x580 [ 57.794157] ? __pfx_worker_thread+0x10/0x10 [ 57.794160] worker_thread+0x1a2/0x370 [ 57.794165] ? __pfx_worker_thread+0x10/0x10 [ 57.794167] kthread+0x11b/0x150 [ 57.794172] ? __pfx_kthread+0x10/0x10 [ 57.794177] ret_from_fork+0x3d/0x60 [ 57.794181] ? __pfx_kthread+0x10/0x10 [ 57.794184] ret_from_fork_asm+0x1b/0x30 Signed-off-by: Lang Yu <Lang.Yu@amd.com> Reviewed-by: Felix Kuehling <felix.kuehling@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
If the BO pages pinned for RDMA is not contiguous on VRAM, evict it to system memory first to free the VRAM space, then allocate contiguous VRAM space, and then move it from system memory back to VRAM. v6: user context should use interruptible call (Felix) Signed-off-by: Philip Yang <Philip.Yang@amd.com> Reviewed-by: Felix Kuehling <felix.kuehling@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
RDMA device with limited scatter-gather ability requires contiguous VRAM buffer allocation for RDMA peer direct support. Add a new KFD alloc memory flag and store as bo alloc flag AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS. When pin this bo to export for RDMA peerdirect access, this will set TTM_PL_FLAG_CONTIFUOUS flag, and ask VRAM buddy allocator to get contiguous VRAM. Signed-off-by: Philip Yang <Philip.Yang@amd.com> Reviewed-by: Felix Kuehling <felix.kuehling@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Subtract the VRAM pinned memory when checking for available memory in amdgpu_amdkfd_reserve_mem_limit function since that memory is not available for use. Signed-off-by: Mukul Joshi <mukul.joshi@amd.com> Reviewed-by: Felix Kuehling <felix.kuehling@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
When page table BOs were evicted but not validated before
updating page tables, VM is still in evicting state,
amdgpu_vm_update_range returns -EBUSY and
restore_process_worker runs into a dead loop.
v2: Split the BO validation and page table update into two
separate loops in amdgpu_amdkfd_restore_process_bos. (Felix)
1.Validate BOs
2.Validate VM (and DMABuf attachments)
3.Update page tables for the BOs validated above
Fixes: 50661eb1a2c8 ("drm/amdgpu: Auto-validate DMABuf imports in compute VMs")
Signed-off-by: Lang Yu <Lang.Yu@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Felix Kuehling <felix.kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>