[PATCH 1/5] drm/amdgpu: re-apply "use the new cursor in the VM code" v2

Felix Kuehling felix.kuehling at amd.com
Fri Apr 30 14:57:18 UTC 2021


Am 2021-04-30 um 9:50 a.m. schrieb Christian König:
>
>
> Am 30.04.21 um 15:45 schrieb Felix Kuehling:
>> Am 2021-04-30 um 4:54 a.m. schrieb Christian König:
>>>
>>> Am 30.04.21 um 10:32 schrieb Felix Kuehling:
>>>> Am 2021-04-27 um 6:54 a.m. schrieb Christian König:
>>>>> Now that we found the underlying problem we can re-apply this patch.
>>>>>
>>>>> This reverts commit 867fee7f8821ff42e7308088cf0c3450ac49c17c.
>>>>>
>>>>> v2: rebase on KFD changes
>>>>>
>>>>> Signed-off-by: Christian König <christian.koenig at amd.com>
>>>> Acked-by: Felix Kuehling <Felix.Kuehling at amd.com>
>>> Thanks, just for this patch or the series?
>> I meant it for the patch. I could probably extend it to the series. But
>> I'm a bit confused how the patches relate to each other. There seems to
>> be an implication that everything leads up to enabling patch 5, i.e.
>> removing users of mm_nodes so that you can in patch 5 make the data
>> structure containing mm_nodes private.
>>
>> But it's not the same data structure. I don't think patch 5 really
>> depends on any of the previous patches, because all the other patches
>> deal with ttm_resource.mm_nodes and patch 5 deals with
>> amdgpu_vram_reservation.mm_nodes.
>
> No, you are close but still misunderstood the intention here a bit.
>
> I want to stop using drm_mm_node directly in the amdgpu code so that
> we are able to switch the VRAM management to a buddy allocator sooner
> or later.
>
> mm_nodes and amdgpu_vram_reservation.mm_nodes are both of that type.

Yes, I got that. So all of these patches are preparation for something
bigger. That's the part that wasn't clear. The series is

Acked-by: Felix Kuehling <Felix.Kuehling at amd.com>

We still have a few uses of mm_node in kfd_migrate.c. That code is about
to change as Alejandro is beginning to rework our VRAM management for
HMM migration to allow partial migration failures and mixed VRAM/System
mappings. But we'll still need a way to find VRAM addresses within a BO.

Regards,
  Felix


>
> Regards,
> Christian.
>
>>
>> Regards,
>>    Felix
>>
>>
>>> Christian.
>>>
>>>>
>>>>> ---
>>>>>    drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 55
>>>>> +++++++++-----------------
>>>>>    drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h |  2 +-
>>>>>    drivers/gpu/drm/amd/amdkfd/kfd_svm.c   |  3 +-
>>>>>    3 files changed, 20 insertions(+), 40 deletions(-)
>>>>>
>>>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>>>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>>>>> index dae51992c607..fa43d332a979 100644
>>>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>>>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>>>>> @@ -37,6 +37,7 @@
>>>>>    #include "amdgpu_gmc.h"
>>>>>    #include "amdgpu_xgmi.h"
>>>>>    #include "amdgpu_dma_buf.h"
>>>>> +#include "amdgpu_res_cursor.h"
>>>>>    #include "kfd_svm.h"
>>>>>      /**
>>>>> @@ -1606,7 +1607,7 @@ static int amdgpu_vm_update_ptes(struct
>>>>> amdgpu_vm_update_params *params,
>>>>>     * @last: last mapped entry
>>>>>     * @flags: flags for the entries
>>>>>     * @offset: offset into nodes and pages_addr
>>>>> - * @nodes: array of drm_mm_nodes with the MC addresses
>>>>> + * @res: ttm_resource to map
>>>>>     * @pages_addr: DMA addresses to use for mapping
>>>>>     * @fence: optional resulting fence
>>>>>     *
>>>>> @@ -1621,13 +1622,13 @@ int amdgpu_vm_bo_update_mapping(struct
>>>>> amdgpu_device *adev,
>>>>>                    bool unlocked, struct dma_resv *resv,
>>>>>                    uint64_t start, uint64_t last,
>>>>>                    uint64_t flags, uint64_t offset,
>>>>> -                struct drm_mm_node *nodes,
>>>>> +                struct ttm_resource *res,
>>>>>                    dma_addr_t *pages_addr,
>>>>>                    struct dma_fence **fence)
>>>>>    {
>>>>>        struct amdgpu_vm_update_params params;
>>>>> +    struct amdgpu_res_cursor cursor;
>>>>>        enum amdgpu_sync_mode sync_mode;
>>>>> -    uint64_t pfn;
>>>>>        int r;
>>>>>          memset(&params, 0, sizeof(params));
>>>>> @@ -1645,14 +1646,6 @@ int amdgpu_vm_bo_update_mapping(struct
>>>>> amdgpu_device *adev,
>>>>>        else
>>>>>            sync_mode = AMDGPU_SYNC_EXPLICIT;
>>>>>    -    pfn = offset >> PAGE_SHIFT;
>>>>> -    if (nodes) {
>>>>> -        while (pfn >= nodes->size) {
>>>>> -            pfn -= nodes->size;
>>>>> -            ++nodes;
>>>>> -        }
>>>>> -    }
>>>>> -
>>>>>        amdgpu_vm_eviction_lock(vm);
>>>>>        if (vm->evicting) {
>>>>>            r = -EBUSY;
>>>>> @@ -1671,23 +1664,17 @@ int amdgpu_vm_bo_update_mapping(struct
>>>>> amdgpu_device *adev,
>>>>>        if (r)
>>>>>            goto error_unlock;
>>>>>    -    do {
>>>>> +    amdgpu_res_first(res, offset, (last - start + 1) *
>>>>> AMDGPU_GPU_PAGE_SIZE,
>>>>> +             &cursor);
>>>>> +    while (cursor.remaining) {
>>>>>            uint64_t tmp, num_entries, addr;
>>>>>    -
>>>>> -        num_entries = last - start + 1;
>>>>> -        if (nodes) {
>>>>> -            addr = nodes->start << PAGE_SHIFT;
>>>>> -            num_entries = min((nodes->size - pfn) *
>>>>> -                AMDGPU_GPU_PAGES_IN_CPU_PAGE, num_entries);
>>>>> -        } else {
>>>>> -            addr = 0;
>>>>> -        }
>>>>> -
>>>>> +        num_entries = cursor.size >> AMDGPU_GPU_PAGE_SHIFT;
>>>>>            if (pages_addr) {
>>>>>                bool contiguous = true;
>>>>>                  if (num_entries > AMDGPU_GPU_PAGES_IN_CPU_PAGE) {
>>>>> +                uint64_t pfn = cursor.start >> PAGE_SHIFT;
>>>>>                    uint64_t count;
>>>>>                      contiguous = pages_addr[pfn + 1] ==
>>>>> @@ -1707,16 +1694,18 @@ int amdgpu_vm_bo_update_mapping(struct
>>>>> amdgpu_device *adev,
>>>>>                }
>>>>>                  if (!contiguous) {
>>>>> -                addr = pfn << PAGE_SHIFT;
>>>>> +                addr = cursor.start;
>>>>>                    params.pages_addr = pages_addr;
>>>>>                } else {
>>>>> -                addr = pages_addr[pfn];
>>>>> +                addr = pages_addr[cursor.start >> PAGE_SHIFT];
>>>>>                    params.pages_addr = NULL;
>>>>>                }
>>>>>              } else if (flags & (AMDGPU_PTE_VALID |
>>>>> AMDGPU_PTE_PRT)) {
>>>>> -            addr += bo_adev->vm_manager.vram_base_offset;
>>>>> -            addr += pfn << PAGE_SHIFT;
>>>>> +            addr = bo_adev->vm_manager.vram_base_offset +
>>>>> +                cursor.start;
>>>>> +        } else {
>>>>> +            addr = 0;
>>>>>            }
>>>>>              tmp = start + num_entries;
>>>>> @@ -1724,14 +1713,9 @@ int amdgpu_vm_bo_update_mapping(struct
>>>>> amdgpu_device *adev,
>>>>>            if (r)
>>>>>                goto error_unlock;
>>>>>    -        pfn += num_entries / AMDGPU_GPU_PAGES_IN_CPU_PAGE;
>>>>> -        if (nodes && nodes->size == pfn) {
>>>>> -            pfn = 0;
>>>>> -            ++nodes;
>>>>> -        }
>>>>> +        amdgpu_res_next(&cursor, num_entries *
>>>>> AMDGPU_GPU_PAGE_SIZE);
>>>>>            start = tmp;
>>>>> -
>>>>> -    } while (unlikely(start != last + 1));
>>>>> +    };
>>>>>          r = vm->update_funcs->commit(&params, fence);
>>>>>    @@ -1760,7 +1744,6 @@ int amdgpu_vm_bo_update(struct amdgpu_device
>>>>> *adev, struct amdgpu_bo_va *bo_va,
>>>>>        struct amdgpu_bo_va_mapping *mapping;
>>>>>        dma_addr_t *pages_addr = NULL;
>>>>>        struct ttm_resource *mem;
>>>>> -    struct drm_mm_node *nodes;
>>>>>        struct dma_fence **last_update;
>>>>>        struct dma_resv *resv;
>>>>>        uint64_t flags;
>>>>> @@ -1769,7 +1752,6 @@ int amdgpu_vm_bo_update(struct amdgpu_device
>>>>> *adev, struct amdgpu_bo_va *bo_va,
>>>>>          if (clear || !bo) {
>>>>>            mem = NULL;
>>>>> -        nodes = NULL;
>>>>>            resv = vm->root.base.bo->tbo.base.resv;
>>>>>        } else {
>>>>>            struct drm_gem_object *obj = &bo->tbo.base;
>>>>> @@ -1784,7 +1766,6 @@ int amdgpu_vm_bo_update(struct amdgpu_device
>>>>> *adev, struct amdgpu_bo_va *bo_va,
>>>>>                    bo = gem_to_amdgpu_bo(gobj);
>>>>>            }
>>>>>            mem = &bo->tbo.mem;
>>>>> -        nodes = mem->mm_node;
>>>>>            if (mem->mem_type == TTM_PL_TT)
>>>>>                pages_addr = bo->tbo.ttm->dma_address;
>>>>>        }
>>>>> @@ -1833,7 +1814,7 @@ int amdgpu_vm_bo_update(struct amdgpu_device
>>>>> *adev, struct amdgpu_bo_va *bo_va,
>>>>>            r = amdgpu_vm_bo_update_mapping(adev, bo_adev, vm, false,
>>>>> false,
>>>>>                            resv, mapping->start,
>>>>>                            mapping->last, update_flags,
>>>>> -                        mapping->offset, nodes,
>>>>> +                        mapping->offset, mem,
>>>>>                            pages_addr, last_update);
>>>>>            if (r)
>>>>>                return r;
>>>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
>>>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
>>>>> index e5a3f18be2b7..1ae5ea8db497 100644
>>>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
>>>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
>>>>> @@ -402,7 +402,7 @@ int amdgpu_vm_bo_update_mapping(struct
>>>>> amdgpu_device *adev,
>>>>>                    bool unlocked, struct dma_resv *resv,
>>>>>                    uint64_t start, uint64_t last,
>>>>>                    uint64_t flags, uint64_t offset,
>>>>> -                struct drm_mm_node *nodes,
>>>>> +                struct ttm_resource *res,
>>>>>                    dma_addr_t *pages_addr,
>>>>>                    struct dma_fence **fence);
>>>>>    int amdgpu_vm_bo_update(struct amdgpu_device *adev,
>>>>> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
>>>>> b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
>>>>> index e4ce97ab6e26..0b0e76e16ddc 100644
>>>>> --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
>>>>> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
>>>>> @@ -1151,8 +1151,7 @@ svm_range_map_to_gpu(struct amdgpu_device
>>>>> *adev, struct amdgpu_vm *vm,
>>>>>                        prange->mapping.start,
>>>>>                        prange->mapping.last, pte_flags,
>>>>>                        prange->mapping.offset,
>>>>> -                    prange->ttm_res ?
>>>>> -                        prange->ttm_res->mm_node : NULL,
>>>>> +                    prange->ttm_res,
>>>>>                        dma_addr, &vm->last_update);
>>>>>        if (r) {
>>>>>            pr_debug("failed %d to map to gpu 0x%lx\n", r,
>>>>> prange->start);
>


More information about the amd-gfx mailing list