[PATCH] drm/amd/amdgpu: Code comments for the amdgpu_ttm.c driver. (v2)

tstdenis@xxxxxxx (Tom St Denis) · Fri, 18 May 2018 06:09:49 -0400

On 05/18/2018 05:52 AM, Christian KÃ¶nig wrote:
> Am 17.05.2018 um 17:34 schrieb Alex Deucher:
>> On Tue, May 15, 2018 at 10:02 AM, Tom St Denis <tom.stdenis at amd.com> 
>> wrote:
>>> NFC just comments.
>>>
>>> (v2):Â  Updated based on feedback from Alex Deucher.
>>>
>>> Signed-off-by: Tom St Denis <tom.stdenis at amd.com>
>> Reviewed-by: Alex Deucher <alexander.deucher at amd.com>
> 
> Just one comment "Pin pages of memory pointed to..." better write "Grab 
> a reference to the memory pointed to...".
> 
> get_user_pages() does not pin anything! I've heard that misconception so 
> many times now that I can't remember how often we had to explain it and 
> we should definitely not leak it into the documentation.
> 
> With that fixed the patch is Reviewed-by: Christian KÃ¶nig 
> <christian.koenig at amd.com>.

Hi Christian,

Well it's been pushed but as per Alex's email last night I'll swing back 
to add Sphinx style comments to the file and fix that in the same patch.

Thanks,
tom

> 
> Regards,
> Christian.
> 
>>
>>> ---
>>> Â  drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 347 
>>> +++++++++++++++++++++++++++++++-
>>> Â  1 file changed, 340 insertions(+), 7 deletions(-)
>>>
>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c 
>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
>>> index dfd22db13fb1..2eaaa1fb7b59 100644
>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
>>> @@ -63,16 +63,44 @@ static void amdgpu_ttm_debugfs_fini(struct 
>>> amdgpu_device *adev);
>>> Â  /*
>>> Â Â  * Global memory.
>>> Â Â  */
>>> +
>>> +/**
>>> + * amdgpu_ttm_mem_global_init - Initialize and acquire reference to
>>> + * memory object
>>> + *
>>> + * @ref: Object for initialization.
>>> + *
>>> + * This is called by drm_global_item_ref() when an object is being
>>> + * initialized.
>>> + */
>>> Â  static int amdgpu_ttm_mem_global_init(struct drm_global_reference 
>>> *ref)
>>> Â  {
>>> Â Â Â Â Â Â Â Â  return ttm_mem_global_init(ref->object);
>>> Â  }
>>>
>>> +/**
>>> + * amdgpu_ttm_mem_global_release - Drop reference to a memory object
>>> + *
>>> + * @ref: Object being removed
>>> + *
>>> + * This is called by drm_global_item_unref() when an object is being
>>> + * released.
>>> + */
>>> Â  static void amdgpu_ttm_mem_global_release(struct 
>>> drm_global_reference *ref)
>>> Â  {
>>> Â Â Â Â Â Â Â Â  ttm_mem_global_release(ref->object);
>>> Â  }
>>>
>>> +/**
>>> + * amdgpu_ttm_global_init - Initialize global TTM memory reference
>>> + *Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  structures.
>>> + *
>>> + * @adev:Â Â Â Â Â  AMDGPU device for which the global structures need to be
>>> + *Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  registered.
>>> + *
>>> + * This is called as part of the AMDGPU ttm init from amdgpu_ttm_init()
>>> + * during bring up.
>>> + */
>>> Â  static int amdgpu_ttm_global_init(struct amdgpu_device *adev)
>>> Â  {
>>> Â Â Â Â Â Â Â Â  struct drm_global_reference *global_ref;
>>> @@ -80,7 +108,9 @@ static int amdgpu_ttm_global_init(struct 
>>> amdgpu_device *adev)
>>> Â Â Â Â Â Â Â Â  struct drm_sched_rq *rq;
>>> Â Â Â Â Â Â Â Â  int r;
>>>
>>> +Â Â Â Â Â Â  /* ensure reference is false in case init fails */
>>> Â Â Â Â Â Â Â Â  adev->mman.mem_global_referenced = false;
>>> +
>>> Â Â Â Â Â Â Â Â  global_ref = &adev->mman.mem_global_ref;
>>> Â Â Â Â Â Â Â Â  global_ref->global_type = DRM_GLOBAL_TTM_MEM;
>>> Â Â Â Â Â Â Â Â  global_ref->size = sizeof(struct ttm_mem_global);
>>> @@ -146,6 +176,18 @@ static int amdgpu_invalidate_caches(struct 
>>> ttm_bo_device *bdev, uint32_t flags)
>>> Â Â Â Â Â Â Â Â  return 0;
>>> Â  }
>>>
>>> +/**
>>> + * amdgpu_init_mem_type -Â Â Â Â Â  Initialize a memory manager for a 
>>> specific
>>> + *Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  type of 
>>> memory request.
>>> + *
>>> + * @bdev:Â Â Â Â Â  The TTM BO device object (contains a reference to
>>> + *Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  amdgpu_device)
>>> + * @type:Â Â Â Â Â  The type of memory requested
>>> + * @man:
>>> + *
>>> + * This is called by ttm_bo_init_mm() when a buffer object is being
>>> + * initialized.
>>> + */
>>> Â  static int amdgpu_init_mem_type(struct ttm_bo_device *bdev, 
>>> uint32_t type,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  struct ttm_mem_type_manager *man)
>>> Â  {
>>> @@ -161,6 +203,7 @@ static int amdgpu_init_mem_type(struct 
>>> ttm_bo_device *bdev, uint32_t type,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  man->default_caching = TTM_PL_FLAG_CACHED;
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  break;
>>> Â Â Â Â Â Â Â Â  case TTM_PL_TT:
>>> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â  /* GTT memoryÂ  */
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  man->func = &amdgpu_gtt_mgr_func;
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  man->gpu_offset = adev->gmc.gart_start;
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  man->available_caching = TTM_PL_MASK_CACHING;
>>> @@ -193,6 +236,14 @@ static int amdgpu_init_mem_type(struct 
>>> ttm_bo_device *bdev, uint32_t type,
>>> Â Â Â Â Â Â Â Â  return 0;
>>> Â  }
>>>
>>> +/**
>>> + * amdgpu_evict_flags - Compute placement flags
>>> + *
>>> + * @bo: The buffer object to evict
>>> + * @placement: Possible destination(s) for evicted BO
>>> + *
>>> + * Fill in placement data when ttm_bo_evict() is called
>>> + */
>>> Â  static void amdgpu_evict_flags(struct ttm_buffer_object *bo,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  struct ttm_placement *placement)
>>> Â  {
>>> @@ -204,12 +255,14 @@ static void amdgpu_evict_flags(struct 
>>> ttm_buffer_object *bo,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  .flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM
>>> Â Â Â Â Â Â Â Â  };
>>>
>>> +Â Â Â Â Â Â  /* Don't handle scatter gather BOs */
>>> Â Â Â Â Â Â Â Â  if (bo->type == ttm_bo_type_sg) {
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  placement->num_placement = 0;
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  placement->num_busy_placement = 0;
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  return;
>>> Â Â Â Â Â Â Â Â  }
>>>
>>> +Â Â Â Â Â Â  /* Object isn't an AMDGPU object so ignore */
>>> Â Â Â Â Â Â Â Â  if (!amdgpu_ttm_bo_is_amdgpu_bo(bo)) {
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  placement->placement = &placements;
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  placement->busy_placement = &placements;
>>> @@ -217,10 +270,12 @@ static void amdgpu_evict_flags(struct 
>>> ttm_buffer_object *bo,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  placement->num_busy_placement = 1;
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  return;
>>> Â Â Â Â Â Â Â Â  }
>>> +
>>> Â Â Â Â Â Â Â Â  abo = ttm_to_amdgpu_bo(bo);
>>> Â Â Â Â Â Â Â Â  switch (bo->mem.mem_type) {
>>> Â Â Â Â Â Â Â Â  case TTM_PL_VRAM:
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  if (!adev->mman.buffer_funcs_enabled) {
>>> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  /* Move to system memory */
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  amdgpu_ttm_placement_from_domain(abo, 
>>> AMDGPU_GEM_DOMAIN_CPU);
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  } else if (adev->gmc.visible_vram_size < 
>>> adev->gmc.real_vram_size &&
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  !(abo->flags & 
>>> AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED) &&
>>> @@ -238,6 +293,7 @@ static void amdgpu_evict_flags(struct 
>>> ttm_buffer_object *bo,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  abo->placement.busy_placement = 
>>> &abo->placements[1];
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  abo->placement.num_busy_placement = 1;
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  } else {
>>> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  /* Move to GTT memory */
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  amdgpu_ttm_placement_from_domain(abo, 
>>> AMDGPU_GEM_DOMAIN_GTT);
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  }
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  break;
>>> @@ -248,6 +304,15 @@ static void amdgpu_evict_flags(struct 
>>> ttm_buffer_object *bo,
>>> Â Â Â Â Â Â Â Â  *placement = abo->placement;
>>> Â  }
>>>
>>> +/**
>>> + * amdgpu_verify_access - Verify access for a mmap call
>>> + *
>>> + * @bo:Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  The buffer object to map
>>> + * @filp:Â Â Â Â Â  The file pointer from the process performing the mmap
>>> + *
>>> + * This is called by ttm_bo_mmap() to verify whether a process
>>> + * has the right to mmap a BO to their process space.
>>> + */
>>> Â  static int amdgpu_verify_access(struct ttm_buffer_object *bo, 
>>> struct file *filp)
>>> Â  {
>>> Â Â Â Â Â Â Â Â  struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
>>> @@ -265,6 +330,15 @@ static int amdgpu_verify_access(struct 
>>> ttm_buffer_object *bo, struct file *filp)
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  filp->private_data);
>>> Â  }
>>>
>>> +/**
>>> + * amdgpu_move_null - Register memory for a buffer object
>>> + *
>>> + * @bo:Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  The bo to assign the memory to
>>> + * @new_mem:Â Â  The memory to be assigned.
>>> + *
>>> + * Assign the memory from new_mem to the memory of the buffer object
>>> + * bo.
>>> + */
>>> Â  static void amdgpu_move_null(struct ttm_buffer_object *bo,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  struct ttm_mem_reg *new_mem)
>>> Â  {
>>> @@ -275,6 +349,10 @@ static void amdgpu_move_null(struct 
>>> ttm_buffer_object *bo,
>>> Â Â Â Â Â Â Â Â  new_mem->mm_node = NULL;
>>> Â  }
>>>
>>> +/**
>>> + * amdgpu_mm_node_addr -Â Â Â Â Â Â  Compute the GPU relative offset of a GTT
>>> + *Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  buffer.
>>> + */
>>> Â  static uint64_t amdgpu_mm_node_addr(struct ttm_buffer_object *bo,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  struct drm_mm_node *mm_node,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  struct ttm_mem_reg *mem)
>>> @@ -289,9 +367,10 @@ static uint64_t amdgpu_mm_node_addr(struct 
>>> ttm_buffer_object *bo,
>>> Â  }
>>>
>>> Â  /**
>>> - * amdgpu_find_mm_node - Helper function finds the drm_mm_node
>>> - *Â  corresponding to @offset. It also modifies the offset to be
>>> - *Â  within the drm_mm_node returned
>>> + * amdgpu_find_mm_node -Â Â Â Â Â Â  Helper function finds the drm_mm_node
>>> + *Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  corresponding to 
>>> @offset. It also modifies
>>> + *Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  the offset to 
>>> be within the drm_mm_node
>>> + *Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  returned
>>> Â Â  */
>>> Â  static struct drm_mm_node *amdgpu_find_mm_node(struct ttm_mem_reg 
>>> *mem,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  unsigned long *offset)
>>> @@ -430,7 +509,12 @@ int amdgpu_ttm_copy_mem_to_mem(struct 
>>> amdgpu_device *adev,
>>> Â Â Â Â Â Â Â Â  return r;
>>> Â  }
>>>
>>> -
>>> +/**
>>> + * amdgpu_move_blit - Copy an entire buffer to another buffer
>>> + *
>>> + * This is a helper called by amdgpu_bo_move() and
>>> + * amdgpu_move_vram_ram() to help move buffers to and from VRAM.
>>> + */
>>> Â  static int amdgpu_move_blit(struct ttm_buffer_object *bo,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  bool evict, bool no_wait_gpu,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  struct ttm_mem_reg *new_mem,
>>> @@ -465,6 +549,11 @@ static int amdgpu_move_blit(struct 
>>> ttm_buffer_object *bo,
>>> Â Â Â Â Â Â Â Â  return r;
>>> Â  }
>>>
>>> +/**
>>> + * amdgpu_move_vram_ram - Copy VRAM buffer to RAM buffer
>>> + *
>>> + * Called by amdgpu_bo_move().
>>> + */
>>> Â  static int amdgpu_move_vram_ram(struct ttm_buffer_object *bo, bool 
>>> evict,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  struct ttm_operation_ctx *ctx,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  struct ttm_mem_reg *new_mem)
>>> @@ -477,6 +566,8 @@ static int amdgpu_move_vram_ram(struct 
>>> ttm_buffer_object *bo, bool evict,
>>> Â Â Â Â Â Â Â Â  int r;
>>>
>>> Â Â Â Â Â Â Â Â  adev = amdgpu_ttm_adev(bo->bdev);
>>> +
>>> +Â Â Â Â Â Â  /* create space/pages for new_mem in GTT space */
>>> Â Â Â Â Â Â Â Â  tmp_mem = *new_mem;
>>> Â Â Â Â Â Â Â Â  tmp_mem.mm_node = NULL;
>>> Â Â Â Â Â Â Â Â  placement.num_placement = 1;
>>> @@ -491,25 +582,36 @@ static int amdgpu_move_vram_ram(struct 
>>> ttm_buffer_object *bo, bool evict,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  return r;
>>> Â Â Â Â Â Â Â Â  }
>>>
>>> +Â Â Â Â Â Â  /* set caching flags */
>>> Â Â Â Â Â Â Â Â  r = ttm_tt_set_placement_caching(bo->ttm, tmp_mem.placement);
>>> Â Â Â Â Â Â Â Â  if (unlikely(r)) {
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  goto out_cleanup;
>>> Â Â Â Â Â Â Â Â  }
>>>
>>> +Â Â Â Â Â Â  /* Bind the memory to the GTT space */
>>> Â Â Â Â Â Â Â Â  r = ttm_tt_bind(bo->ttm, &tmp_mem, ctx);
>>> Â Â Â Â Â Â Â Â  if (unlikely(r)) {
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  goto out_cleanup;
>>> Â Â Â Â Â Â Â Â  }
>>> +
>>> +Â Â Â Â Â Â  /* blit VRAM to GTT */
>>> Â Â Â Â Â Â Â Â  r = amdgpu_move_blit(bo, true, ctx->no_wait_gpu, &tmp_mem, 
>>> old_mem);
>>> Â Â Â Â Â Â Â Â  if (unlikely(r)) {
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  goto out_cleanup;
>>> Â Â Â Â Â Â Â Â  }
>>> +
>>> +Â Â Â Â Â Â  /* move BO (in tmp_mem) to new_mem */
>>> Â Â Â Â Â Â Â Â  r = ttm_bo_move_ttm(bo, ctx, new_mem);
>>> Â  out_cleanup:
>>> Â Â Â Â Â Â Â Â  ttm_bo_mem_put(bo, &tmp_mem);
>>> Â Â Â Â Â Â Â Â  return r;
>>> Â  }
>>>
>>> +/**
>>> + * amdgpu_move_ram_vram - Copy buffer from RAM to VRAM
>>> + *
>>> + * Called by amdgpu_bo_move().
>>> + */
>>> Â  static int amdgpu_move_ram_vram(struct ttm_buffer_object *bo, bool 
>>> evict,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  struct ttm_operation_ctx *ctx,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  struct ttm_mem_reg *new_mem)
>>> @@ -522,6 +624,8 @@ static int amdgpu_move_ram_vram(struct 
>>> ttm_buffer_object *bo, bool evict,
>>> Â Â Â Â Â Â Â Â  int r;
>>>
>>> Â Â Â Â Â Â Â Â  adev = amdgpu_ttm_adev(bo->bdev);
>>> +
>>> +Â Â Â Â Â Â  /* make space in GTT for old_mem buffer */
>>> Â Â Â Â Â Â Â Â  tmp_mem = *new_mem;
>>> Â Â Â Â Â Â Â Â  tmp_mem.mm_node = NULL;
>>> Â Â Â Â Â Â Â Â  placement.num_placement = 1;
>>> @@ -535,10 +639,14 @@ static int amdgpu_move_ram_vram(struct 
>>> ttm_buffer_object *bo, bool evict,
>>> Â Â Â Â Â Â Â Â  if (unlikely(r)) {
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  return r;
>>> Â Â Â Â Â Â Â Â  }
>>> +
>>> +Â Â Â Â Â Â  /* move/bind old memory to GTT space */
>>> Â Â Â Â Â Â Â Â  r = ttm_bo_move_ttm(bo, ctx, &tmp_mem);
>>> Â Â Â Â Â Â Â Â  if (unlikely(r)) {
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  goto out_cleanup;
>>> Â Â Â Â Â Â Â Â  }
>>> +
>>> +Â Â Â Â Â Â  /* copy to VRAM */
>>> Â Â Â Â Â Â Â Â  r = amdgpu_move_blit(bo, true, ctx->no_wait_gpu, new_mem, 
>>> old_mem);
>>> Â Â Â Â Â Â Â Â  if (unlikely(r)) {
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  goto out_cleanup;
>>> @@ -548,6 +656,11 @@ static int amdgpu_move_ram_vram(struct 
>>> ttm_buffer_object *bo, bool evict,
>>> Â Â Â Â Â Â Â Â  return r;
>>> Â  }
>>>
>>> +/**
>>> + * amdgpu_bo_move - Move a buffer object to a new memory location
>>> + *
>>> + * Called by ttm_bo_handle_move_mem()
>>> + */
>>> Â  static int amdgpu_bo_move(struct ttm_buffer_object *bo, bool evict,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  struct ttm_operation_ctx *ctx,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  struct ttm_mem_reg *new_mem)
>>> @@ -613,6 +726,11 @@ static int amdgpu_bo_move(struct 
>>> ttm_buffer_object *bo, bool evict,
>>> Â Â Â Â Â Â Â Â  return 0;
>>> Â  }
>>>
>>> +/**
>>> + * amdgpu_ttm_io_mem_reserve - Reserve a block of memory during a fault
>>> + *
>>> + * Called by ttm_mem_io_reserve() ultimately via ttm_bo_vm_fault()
>>> + */
>>> Â  static int amdgpu_ttm_io_mem_reserve(struct ttm_bo_device *bdev, 
>>> struct ttm_mem_reg *mem)
>>> Â  {
>>> Â Â Â Â Â Â Â Â  struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
>>> @@ -690,6 +808,14 @@ struct amdgpu_ttm_tt {
>>> Â Â Â Â Â Â Â Â  uint32_tÂ Â Â Â Â Â Â Â Â Â Â Â Â Â Â  last_set_pages;
>>> Â  };
>>>
>>> +/**
>>> + * amdgpu_ttm_tt_get_user_pages -Â Â Â Â Â  Pin pages of memory pointed to
>>> + 
>>> *                                                                     
>>> by a USERPTR pointer to memory
>>> + *
>>> + * Called by amdgpu_gem_userptr_ioctl() and amdgpu_cs_parser_bos().
>>> + * This provides a wrapper around the get_user_pages() call to provide
>>> + * device accessible pages that back user memory.
>>> + */
>>> Â  int amdgpu_ttm_tt_get_user_pages(struct ttm_tt *ttm, struct page 
>>> **pages)
>>> Â  {
>>> Â Â Â Â Â Â Â Â  struct amdgpu_ttm_tt *gtt = (void *)ttm;
>>> @@ -715,6 +841,7 @@ int amdgpu_ttm_tt_get_user_pages(struct ttm_tt 
>>> *ttm, struct page **pages)
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  }
>>> Â Â Â Â Â Â Â Â  }
>>>
>>> +Â Â Â Â Â Â  /* loop enough times using contiguous pages of memory */
>>> Â Â Â Â Â Â Â Â  do {
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  unsigned num_pages = ttm->num_pages - pinned;
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  uint64_t userptr = gtt->userptr + pinned * PAGE_SIZE;
>>> @@ -748,6 +875,14 @@ int amdgpu_ttm_tt_get_user_pages(struct ttm_tt 
>>> *ttm, struct page **pages)
>>> Â Â Â Â Â Â Â Â  return r;
>>> Â  }
>>>
>>> +/**
>>> + * amdgpu_ttm_tt_set_user_pages -Â Â Â Â Â  Copy pages in, putting old pages
>>> + 
>>> *                                                                     
>>> as necessary.
>>> + *
>>> + * Called by amdgpu_cs_list_validate().Â  This creates the page list
>>> + * that backs user memory and will ultimately be mapped into the device
>>> + * address space.
>>> + */
>>> Â  void amdgpu_ttm_tt_set_user_pages(struct ttm_tt *ttm, struct page 
>>> **pages)
>>> Â  {
>>> Â Â Â Â Â Â Â Â  struct amdgpu_ttm_tt *gtt = (void *)ttm;
>>> @@ -762,6 +897,11 @@ void amdgpu_ttm_tt_set_user_pages(struct ttm_tt 
>>> *ttm, struct page **pages)
>>> Â Â Â Â Â Â Â Â  }
>>> Â  }
>>>
>>> +/**
>>> + * amdgpu_ttm_tt_mark_user_page - Mark pages as dirty
>>> + *
>>> + * Called while unpinning userptr pages
>>> + */
>>> Â  void amdgpu_ttm_tt_mark_user_pages(struct ttm_tt *ttm)
>>> Â  {
>>> Â Â Â Â Â Â Â Â  struct amdgpu_ttm_tt *gtt = (void *)ttm;
>>> @@ -780,7 +920,12 @@ void amdgpu_ttm_tt_mark_user_pages(struct ttm_tt 
>>> *ttm)
>>> Â Â Â Â Â Â Â Â  }
>>> Â  }
>>>
>>> -/* prepare the sg table with the user pages */
>>> +/**
>>> + * amdgpu_ttm_tt_pin_userptr -Â Â Â Â Â Â Â Â  prepare the sg table with the
>>> + *Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  user 
>>> pages
>>> + *
>>> + * Called by amdgpu_ttm_backend_bind()
>>> + **/
>>> Â  static int amdgpu_ttm_tt_pin_userptr(struct ttm_tt *ttm)
>>> Â  {
>>> Â Â Â Â Â Â Â Â  struct amdgpu_device *adev = amdgpu_ttm_adev(ttm->bdev);
>>> @@ -792,17 +937,20 @@ static int amdgpu_ttm_tt_pin_userptr(struct 
>>> ttm_tt *ttm)
>>> Â Â Â Â Â Â Â Â  enum dma_data_direction direction = write ?
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
>>>
>>> +Â Â Â Â Â Â  /* Allocate an SG array and squash pages into it */
>>> Â Â Â Â Â Â Â Â  r = sg_alloc_table_from_pages(ttm->sg, ttm->pages, 
>>> ttm->num_pages, 0,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  ttm->num_pages << PAGE_SHIFT,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  GFP_KERNEL);
>>> Â Â Â Â Â Â Â Â  if (r)
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  goto release_sg;
>>>
>>> +Â Â Â Â Â Â  /* Map SG to device */
>>> Â Â Â Â Â Â Â Â  r = -ENOMEM;
>>> Â Â Â Â Â Â Â Â  nents = dma_map_sg(adev->dev, ttm->sg->sgl, ttm->sg->nents, 
>>> direction);
>>> Â Â Â Â Â Â Â Â  if (nents != ttm->sg->nents)
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  goto release_sg;
>>>
>>> +Â Â Â Â Â Â  /* convert SG to linear array of pages and dma addresses */
>>> Â Â Â Â Â Â Â Â  drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  gtt->ttm.dma_address, 
>>> ttm->num_pages);
>>>
>>> @@ -813,6 +961,9 @@ static int amdgpu_ttm_tt_pin_userptr(struct 
>>> ttm_tt *ttm)
>>> Â Â Â Â Â Â Â Â  return r;
>>> Â  }
>>>
>>> +/**
>>> + * amdgpu_ttm_tt_unpin_userptr - Unpin and unmap userptr pages
>>> + */
>>> Â  static void amdgpu_ttm_tt_unpin_userptr(struct ttm_tt *ttm)
>>> Â  {
>>> Â Â Â Â Â Â Â Â  struct amdgpu_device *adev = amdgpu_ttm_adev(ttm->bdev);
>>> @@ -826,14 +977,22 @@ static void amdgpu_ttm_tt_unpin_userptr(struct 
>>> ttm_tt *ttm)
>>> Â Â Â Â Â Â Â Â  if (!ttm->sg->sgl)
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  return;
>>>
>>> -Â Â Â Â Â Â  /* free the sg table and pages again */
>>> +Â Â Â Â Â Â  /* unmap the pages mapped to the device */
>>> Â Â Â Â Â Â Â Â  dma_unmap_sg(adev->dev, ttm->sg->sgl, ttm->sg->nents, 
>>> direction);
>>>
>>> +Â Â Â Â Â Â  /* mark the pages as dirty */
>>> Â Â Â Â Â Â Â Â  amdgpu_ttm_tt_mark_user_pages(ttm);
>>>
>>> Â Â Â Â Â Â Â Â  sg_free_table(ttm->sg);
>>> Â  }
>>>
>>> +
>>> +/**
>>> + * amdgpu_ttm_backend_bind - Bind GTT memory
>>> + *
>>> + * Called by ttm_tt_bind() on behalf of ttm_bo_handle_move_mem().
>>> + * This handles binding GTT memory to the device address space.
>>> + */
>>> Â  static int amdgpu_ttm_backend_bind(struct ttm_tt *ttm,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  struct ttm_mem_reg *bo_mem)
>>> Â  {
>>> @@ -864,7 +1023,10 @@ static int amdgpu_ttm_backend_bind(struct 
>>> ttm_tt *ttm,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  return 0;
>>> Â Â Â Â Â Â Â Â  }
>>>
>>> +Â Â Â Â Â Â  /* compute PTE flags relevant to this BO memory */
>>> Â Â Â Â Â Â Â Â  flags = amdgpu_ttm_tt_pte_flags(adev, ttm, bo_mem);
>>> +
>>> +Â Â Â Â Â Â  /* bind pages into GART page tables */
>>> Â Â Â Â Â Â Â Â  gtt->offset = (u64)bo_mem->start << PAGE_SHIFT;
>>> Â Â Â Â Â Â Â Â  r = amdgpu_gart_bind(adev, gtt->offset, ttm->num_pages,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  ttm->pages, gtt->ttm.dma_address, flags);
>>> @@ -875,6 +1037,9 @@ static int amdgpu_ttm_backend_bind(struct ttm_tt 
>>> *ttm,
>>> Â Â Â Â Â Â Â Â  return r;
>>> Â  }
>>>
>>> +/**
>>> + * amdgpu_ttm_alloc_gart - Allocate GART memory for buffer object
>>> + */
>>> Â  int amdgpu_ttm_alloc_gart(struct ttm_buffer_object *bo)
>>> Â  {
>>> Â Â Â Â Â Â Â Â  struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
>>> @@ -890,6 +1055,7 @@ int amdgpu_ttm_alloc_gart(struct 
>>> ttm_buffer_object *bo)
>>> Â Â Â Â Â Â Â Â Â Â Â Â  amdgpu_gtt_mgr_has_gart_addr(&bo->mem))
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  return 0;
>>>
>>> +Â Â Â Â Â Â  /* allocate GTT space */
>>> Â Â Â Â Â Â Â Â  tmp = bo->mem;
>>> Â Â Â Â Â Â Â Â  tmp.mm_node = NULL;
>>> Â Â Â Â Â Â Â Â  placement.num_placement = 1;
>>> @@ -905,7 +1071,10 @@ int amdgpu_ttm_alloc_gart(struct 
>>> ttm_buffer_object *bo)
>>> Â Â Â Â Â Â Â Â  if (unlikely(r))
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  return r;
>>>
>>> +Â Â Â Â Â Â  /* compute PTE flags for this buffer object */
>>> Â Â Â Â Â Â Â Â  flags = amdgpu_ttm_tt_pte_flags(adev, bo->ttm, &tmp);
>>> +
>>> +Â Â Â Â Â Â  /* Bind pages */
>>> Â Â Â Â Â Â Â Â  gtt->offset = (u64)tmp.start << PAGE_SHIFT;
>>> Â Â Â Â Â Â Â Â  r = amdgpu_gart_bind(adev, gtt->offset, bo->ttm->num_pages,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  bo->ttm->pages, gtt->ttm.dma_address, 
>>> flags);
>>> @@ -922,6 +1091,12 @@ int amdgpu_ttm_alloc_gart(struct 
>>> ttm_buffer_object *bo)
>>> Â Â Â Â Â Â Â Â  return 0;
>>> Â  }
>>>
>>> +/**
>>> + * amdgpu_ttm_recover_gart - Rebind GTT pages
>>> + *
>>> + * Called by amdgpu_gtt_mgr_recover() from amdgpu_device_reset() to
>>> + * rebind GTT pages during a GPU reset.
>>> + */
>>> Â  int amdgpu_ttm_recover_gart(struct ttm_buffer_object *tbo)
>>> Â  {
>>> Â Â Â Â Â Â Â Â  struct amdgpu_device *adev = amdgpu_ttm_adev(tbo->bdev);
>>> @@ -941,12 +1116,19 @@ int amdgpu_ttm_recover_gart(struct 
>>> ttm_buffer_object *tbo)
>>> Â Â Â Â Â Â Â Â  return r;
>>> Â  }
>>>
>>> +/**
>>> + * amdgpu_ttm_backend_unbind - Unbind GTT mapped pages
>>> + *
>>> + * Called by ttm_tt_unbind() on behalf of ttm_bo_move_ttm() and
>>> + * ttm_tt_destroy().
>>> + */
>>> Â  static int amdgpu_ttm_backend_unbind(struct ttm_tt *ttm)
>>> Â  {
>>> Â Â Â Â Â Â Â Â  struct amdgpu_device *adev = amdgpu_ttm_adev(ttm->bdev);
>>> Â Â Â Â Â Â Â Â  struct amdgpu_ttm_tt *gtt = (void *)ttm;
>>> Â Â Â Â Â Â Â Â  int r;
>>>
>>> +Â Â Â Â Â Â  /* if the pages have userptr pinning then clear that first */
>>> Â Â Â Â Â Â Â Â  if (gtt->userptr)
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  amdgpu_ttm_tt_unpin_userptr(ttm);
>>>
>>> @@ -975,6 +1157,13 @@ static struct ttm_backend_func 
>>> amdgpu_backend_func = {
>>> Â Â Â Â Â Â Â Â  .destroy = &amdgpu_ttm_backend_destroy,
>>> Â  };
>>>
>>> +/**
>>> + * amdgpu_ttm_tt_create - Create a ttm_tt object for a given BO
>>> + *
>>> + * @bo: The buffer object to create a GTT ttm_tt object around
>>> + *
>>> + * Called by ttm_tt_create().
>>> + */
>>> Â  static struct ttm_tt *amdgpu_ttm_tt_create(struct ttm_buffer_object 
>>> *bo,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  uint32_t page_flags)
>>> Â  {
>>> @@ -988,6 +1177,8 @@ static struct ttm_tt 
>>> *amdgpu_ttm_tt_create(struct ttm_buffer_object *bo,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  return NULL;
>>> Â Â Â Â Â Â Â Â  }
>>> Â Â Â Â Â Â Â Â  gtt->ttm.ttm.func = &amdgpu_backend_func;
>>> +
>>> +Â Â Â Â Â Â  /* allocate space for the uninitialized page entries */
>>> Â Â Â Â Â Â Â Â  if (ttm_sg_tt_init(&gtt->ttm, bo, page_flags)) {
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  kfree(gtt);
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  return NULL;
>>> @@ -995,6 +1186,12 @@ static struct ttm_tt 
>>> *amdgpu_ttm_tt_create(struct ttm_buffer_object *bo,
>>> Â Â Â Â Â Â Â Â  return &gtt->ttm.ttm;
>>> Â  }
>>>
>>> +/**
>>> + * amdgpu_ttm_tt_populate - Map GTT pages visible to the device
>>> + *
>>> + * Map the pages of a ttm_tt object to an address space visible
>>> + * to the underlying device.
>>> + */
>>> Â  static int amdgpu_ttm_tt_populate(struct ttm_tt *ttm,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  struct ttm_operation_ctx *ctx)
>>> Â  {
>>> @@ -1002,6 +1199,7 @@ static int amdgpu_ttm_tt_populate(struct ttm_tt 
>>> *ttm,
>>> Â Â Â Â Â Â Â Â  struct amdgpu_ttm_tt *gtt = (void *)ttm;
>>> Â Â Â Â Â Â Â Â  bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
>>>
>>> +Â Â Â Â Â Â  /* user pages are bound by amdgpu_ttm_tt_pin_userptr() */
>>> Â Â Â Â Â Â Â Â  if (gtt && gtt->userptr) {
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  ttm->sg = kzalloc(sizeof(struct sg_table), GFP_KERNEL);
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  if (!ttm->sg)
>>> @@ -1026,9 +1224,17 @@ static int amdgpu_ttm_tt_populate(struct 
>>> ttm_tt *ttm,
>>> Â Â Â Â Â Â Â Â  }
>>> Â  #endif
>>>
>>> +Â Â Â Â Â Â  /* fall back to generic helper to populate the page array
>>> +Â Â Â Â Â Â Â  * and map them to the device */
>>> Â Â Â Â Â Â Â Â  return ttm_populate_and_map_pages(adev->dev, &gtt->ttm, ctx);
>>> Â  }
>>>
>>> +/**
>>> + * amdgpu_ttm_tt_unpopulate - unmap GTT pages and unpopulate page 
>>> arrays
>>> + *
>>> + * Unmaps pages of a ttm_tt object from the device address space and
>>> + * unpopulates the page array backing it.
>>> + */
>>> Â  static void amdgpu_ttm_tt_unpopulate(struct ttm_tt *ttm)
>>> Â  {
>>> Â Â Â Â Â Â Â Â  struct amdgpu_device *adev;
>>> @@ -1054,9 +1260,21 @@ static void amdgpu_ttm_tt_unpopulate(struct 
>>> ttm_tt *ttm)
>>> Â Â Â Â Â Â Â Â  }
>>> Â  #endif
>>>
>>> +Â Â Â Â Â Â  /* fall back to generic helper to unmap and unpopulate array */
>>> Â Â Â Â Â Â Â Â  ttm_unmap_and_unpopulate_pages(adev->dev, &gtt->ttm);
>>> Â  }
>>>
>>> +/**
>>> + * amdgpu_ttm_tt_set_userptr - Initialize userptr GTT ttm_tt
>>> + *Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  for 
>>> the current task
>>> + *
>>> + * @ttm: The ttm_tt object to bind this userptr object to
>>> + * @addr:Â  The address in the current tasks VM space to use
>>> + * @flags: Requirements of userptr object.
>>> + *
>>> + * Called by amdgpu_gem_userptr_ioctl() to bind userptr pages
>>> + * to current task
>>> + */
>>> Â  int amdgpu_ttm_tt_set_userptr(struct ttm_tt *ttm, uint64_t addr,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  uint32_t flags)
>>> Â  {
>>> @@ -1076,6 +1294,9 @@ int amdgpu_ttm_tt_set_userptr(struct ttm_tt 
>>> *ttm, uint64_t addr,
>>> Â Â Â Â Â Â Â Â  return 0;
>>> Â  }
>>>
>>> +/**
>>> + * amdgpu_ttm_tt_get_usermm - Return memory manager for ttm_tt object
>>> + */
>>> Â  struct mm_struct *amdgpu_ttm_tt_get_usermm(struct ttm_tt *ttm)
>>> Â  {
>>> Â Â Â Â Â Â Â Â  struct amdgpu_ttm_tt *gtt = (void *)ttm;
>>> @@ -1086,6 +1307,12 @@ struct mm_struct 
>>> *amdgpu_ttm_tt_get_usermm(struct ttm_tt *ttm)
>>> Â Â Â Â Â Â Â Â  return gtt->usermm;
>>> Â  }
>>>
>>> +/**
>>> + * amdgpu_ttm_tt_affect_userptr -Â Â Â Â Â  Determine if a ttm_tt object 
>>> lays
>>> + 
>>> *                                                                     
>>> inside an address range for the
>>> + 
>>> *                                                                     
>>> current task.
>>> + *
>>> + */
>>> Â  bool amdgpu_ttm_tt_affect_userptr(struct ttm_tt *ttm, unsigned long 
>>> start,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  unsigned long end)
>>> Â  {
>>> @@ -1096,10 +1323,14 @@ bool amdgpu_ttm_tt_affect_userptr(struct 
>>> ttm_tt *ttm, unsigned long start,
>>> Â Â Â Â Â Â Â Â  if (gtt == NULL || !gtt->userptr)
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  return false;
>>>
>>> +Â Â Â Â Â Â  /* does the address lie within the span for the userptr 
>>> memory? */
>>> Â Â Â Â Â Â Â Â  size = (unsigned long)gtt->ttm.ttm.num_pages * PAGE_SIZE;
>>> Â Â Â Â Â Â Â Â  if (gtt->userptr > end || gtt->userptr + size <= start)
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  return false;
>>>
>>> +Â Â Â Â Â Â  /* Search the lists of tasks that hold this mapping and see
>>> +Â Â Â Â Â Â Â  * if current is one of them.Â  If it is return false.
>>> +Â Â Â Â Â Â Â  */
>>> Â Â Â Â Â Â Â Â  spin_lock(&gtt->guptasklock);
>>> Â Â Â Â Â Â Â Â  list_for_each_entry(entry, &gtt->guptasks, list) {
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  if (entry->task == current) {
>>> @@ -1114,6 +1345,10 @@ bool amdgpu_ttm_tt_affect_userptr(struct 
>>> ttm_tt *ttm, unsigned long start,
>>> Â Â Â Â Â Â Â Â  return true;
>>> Â  }
>>>
>>> +/**
>>> + * amdgpu_ttm_tt_userptr_invalidated - Has the ttm_tt object been
>>> + 
>>> *                                                                             
>>> invalidated?
>>> + */
>>> Â  bool amdgpu_ttm_tt_userptr_invalidated(struct ttm_tt *ttm,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  int *last_invalidated)
>>> Â  {
>>> @@ -1124,6 +1359,12 @@ bool amdgpu_ttm_tt_userptr_invalidated(struct 
>>> ttm_tt *ttm,
>>> Â Â Â Â Â Â Â Â  return prev_invalidated != *last_invalidated;
>>> Â  }
>>>
>>> +/**
>>> + * amdgpu_ttm_tt_userptr_needs_pages - Have the pages backing this
>>> + 
>>> *                                                                             
>>> ttm_tt object been invalidated
>>> + 
>>> *                                                                             
>>> since the last time they've
>>> + 
>>> *                                                                             
>>> been set?
>>> + */
>>> Â  bool amdgpu_ttm_tt_userptr_needs_pages(struct ttm_tt *ttm)
>>> Â  {
>>> Â Â Â Â Â Â Â Â  struct amdgpu_ttm_tt *gtt = (void *)ttm;
>>> @@ -1134,6 +1375,9 @@ bool amdgpu_ttm_tt_userptr_needs_pages(struct 
>>> ttm_tt *ttm)
>>> Â Â Â Â Â Â Â Â  return atomic_read(&gtt->mmu_invalidations) != 
>>> gtt->last_set_pages;
>>> Â  }
>>>
>>> +/**
>>> + * amdgpu_ttm_tt_is_readonly - Is the ttm_tt object read only?
>>> + */
>>> Â  bool amdgpu_ttm_tt_is_readonly(struct ttm_tt *ttm)
>>> Â  {
>>> Â Â Â Â Â Â Â Â  struct amdgpu_ttm_tt *gtt = (void *)ttm;
>>> @@ -1144,6 +1388,12 @@ bool amdgpu_ttm_tt_is_readonly(struct ttm_tt 
>>> *ttm)
>>> Â Â Â Â Â Â Â Â  return !!(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
>>> Â  }
>>>
>>> +/**
>>> + * amdgpu_ttm_tt_pte_flags - Compute PTE flags for ttm_tt object
>>> + *
>>> + * @ttm: The ttm_tt object to compute the flags for
>>> + * @mem: The memory registry backing this ttm_tt object
>>> + */
>>> Â  uint64_t amdgpu_ttm_tt_pte_flags(struct amdgpu_device *adev, struct 
>>> ttm_tt *ttm,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  struct ttm_mem_reg *mem)
>>> Â  {
>>> @@ -1168,6 +1418,16 @@ uint64_t amdgpu_ttm_tt_pte_flags(struct 
>>> amdgpu_device *adev, struct ttm_tt *ttm,
>>> Â Â Â Â Â Â Â Â  return flags;
>>> Â  }
>>>
>>> +/**
>>> + * amdgpu_ttm_bo_eviction_valuable -Â Â  Check to see if we can evict
>>> + 
>>> *                                                                             
>>> a buffer object.
>>> + *
>>> + * Return true if eviction is sensible.Â  Called by
>>> + * ttm_mem_evict_first() on behalf of ttm_bo_mem_force_space()
>>> + * which tries to evict buffer objects until it can find space
>>> + * for a new object and by ttm_bo_force_list_clean() which is
>>> + * used to clean out a memory space.
>>> + */
>>> Â  static bool amdgpu_ttm_bo_eviction_valuable(struct 
>>> ttm_buffer_object *bo,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  const struct ttm_place 
>>> *place)
>>> Â  {
>>> @@ -1214,6 +1474,19 @@ static bool 
>>> amdgpu_ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
>>> Â Â Â Â Â Â Â Â  return ttm_bo_eviction_valuable(bo, place);
>>> Â  }
>>>
>>> +/**
>>> + * amdgpu_ttm_access_memory -Â  Read or Write memory that backs a
>>> + *                                                             
>>> buffer object.
>>> + *
>>> + * @bo:Â  The buffer object to read/write
>>> + * @offset:Â  Offset into buffer object
>>> + * @buf:Â  Secondary buffer to write/read from
>>> + * @len: Length in bytes of access
>>> + * @write:Â  true if writing
>>> + *
>>> + * This is used to access VRAM that backs a buffer object via MMIO
>>> + * access for debugging purposes.
>>> + */
>>> Â  static int amdgpu_ttm_access_memory(struct ttm_buffer_object *bo,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  unsigned long offset,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  void *buf, int len, int write)
>>> @@ -1390,13 +1663,22 @@ static int 
>>> amdgpu_ttm_fw_reserve_vram_init(struct amdgpu_device *adev)
>>> Â Â Â Â Â Â Â Â  adev->fw_vram_usage.reserved_bo = NULL;
>>> Â Â Â Â Â Â Â Â  return r;
>>> Â  }
>>> -
>>> +/**
>>> + * amdgpu_ttm_init -Â Â  Init the memory management (ttm) as well as
>>> + *Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  various gtt/vram 
>>> related fields.
>>> + *
>>> + * This initializes all of the memory space pools that the TTM layer
>>> + * will need such as the GTT space (system memory mapped to the 
>>> device),
>>> + * VRAM (on-board memory), and on-chip memories (GDS, GWS, OA) which
>>> + * can be mapped per VMID.
>>> + */
>>> Â  int amdgpu_ttm_init(struct amdgpu_device *adev)
>>> Â  {
>>> Â Â Â Â Â Â Â Â  uint64_t gtt_size;
>>> Â Â Â Â Â Â Â Â  int r;
>>> Â Â Â Â Â Â Â Â  u64 vis_vram_limit;
>>>
>>> +Â Â Â Â Â Â  /* initialize global references for vram/gtt */
>>> Â Â Â Â Â Â Â Â  r = amdgpu_ttm_global_init(adev);
>>> Â Â Â Â Â Â Â Â  if (r) {
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  return r;
>>> @@ -1417,6 +1699,7 @@ int amdgpu_ttm_init(struct amdgpu_device *adev)
>>> Â Â Â Â Â Â Â Â  /* We opt to avoid OOM on system pages allocations */
>>> Â Â Â Â Â Â Â Â  adev->mman.bdev.no_retry = true;
>>>
>>> +Â Â Â Â Â Â  /* Initialize VRAM pool with all of VRAM divided into pages */
>>> Â Â Â Â Â Â Â Â  r = ttm_bo_init_mm(&adev->mman.bdev, TTM_PL_VRAM,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  adev->gmc.real_vram_size >> 
>>> PAGE_SHIFT);
>>> Â Â Â Â Â Â Â Â  if (r) {
>>> @@ -1446,6 +1729,10 @@ int amdgpu_ttm_init(struct amdgpu_device *adev)
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  return r;
>>> Â Â Â Â Â Â Â Â  }
>>>
>>> +Â Â Â Â Â Â  /* allocate memory as required for VGA
>>> +Â Â Â Â Â Â Â  * This is used for VGA emulation and pre-OS scanout buffers to
>>> +Â Â Â Â Â Â Â  * avoid display artifacts while transitioning between pre-OS
>>> +Â Â Â Â Â Â Â  * and driver.Â  */
>>> Â Â Â Â Â Â Â Â  if (adev->gmc.stolen_size) {
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  r = amdgpu_bo_create_kernel(adev, 
>>> adev->gmc.stolen_size, PAGE_SIZE,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  AMDGPU_GEM_DOMAIN_VRAM,
>>> @@ -1457,6 +1744,8 @@ int amdgpu_ttm_init(struct amdgpu_device *adev)
>>> Â Â Â Â Â Â Â Â  DRM_INFO("amdgpu: %uM of VRAM memory ready\n",
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  (unsigned) (adev->gmc.real_vram_size / (1024 * 
>>> 1024)));
>>>
>>> +Â Â Â Â Â Â  /* Compute GTT size, either bsaed on 3/4th the size of RAM size
>>> +Â Â Â Â Â Â Â  * or whatever the user passed on module init */
>>> Â Â Â Â Â Â Â Â  if (amdgpu_gtt_size == -1) {
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  struct sysinfo si;
>>>
>>> @@ -1467,6 +1756,8 @@ int amdgpu_ttm_init(struct amdgpu_device *adev)
>>> Â Â Â Â Â Â Â Â  }
>>> Â Â Â Â Â Â Â Â  else
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  gtt_size = (uint64_t)amdgpu_gtt_size << 20;
>>> +
>>> +Â Â Â Â Â Â  /* Initialize GTT memory pool */
>>> Â Â Â Â Â Â Â Â  r = ttm_bo_init_mm(&adev->mman.bdev, TTM_PL_TT, gtt_size >> 
>>> PAGE_SHIFT);
>>> Â Â Â Â Â Â Â Â  if (r) {
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  DRM_ERROR("Failed initializing GTT heap.\n");
>>> @@ -1475,6 +1766,7 @@ int amdgpu_ttm_init(struct amdgpu_device *adev)
>>> Â Â Â Â Â Â Â Â  DRM_INFO("amdgpu: %uM of GTT memory ready.\n",
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  (unsigned)(gtt_size / (1024 * 1024)));
>>>
>>> +Â Â Â Â Â Â  /* Initialize various on-chip memory pools */
>>> Â Â Â Â Â Â Â Â  adev->gds.mem.total_size = adev->gds.mem.total_size << 
>>> AMDGPU_GDS_SHIFT;
>>> Â Â Â Â Â Â Â Â  adev->gds.mem.gfx_partition_size = 
>>> adev->gds.mem.gfx_partition_size << AMDGPU_GDS_SHIFT;
>>> Â Â Â Â Â Â Â Â  adev->gds.mem.cs_partition_size = 
>>> adev->gds.mem.cs_partition_size << AMDGPU_GDS_SHIFT;
>>> @@ -1514,6 +1806,7 @@ int amdgpu_ttm_init(struct amdgpu_device *adev)
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  }
>>> Â Â Â Â Â Â Â Â  }
>>>
>>> +Â Â Â Â Â Â  /* Register debugfs entries for amdgpu_ttm */
>>> Â Â Â Â Â Â Â Â  r = amdgpu_ttm_debugfs_init(adev);
>>> Â Â Â Â Â Â Â Â  if (r) {
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  DRM_ERROR("Failed to init debugfs\n");
>>> @@ -1522,11 +1815,19 @@ int amdgpu_ttm_init(struct amdgpu_device *adev)
>>> Â Â Â Â Â Â Â Â  return 0;
>>> Â  }
>>>
>>> +/**
>>> + * amdgpu_ttm_late_init -Â Â Â Â Â  Handle any late initialization for
>>> + *Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  amdgpu_ttm
>>> + */
>>> Â  void amdgpu_ttm_late_init(struct amdgpu_device *adev)
>>> Â  {
>>> +Â Â Â Â Â Â  /* return the VGA stolen memory (if any) back to VRAM */
>>> Â Â Â Â Â Â Â Â  amdgpu_bo_free_kernel(&adev->stolen_vga_memory, NULL, NULL);
>>> Â  }
>>>
>>> +/**
>>> + * amdgpu_ttm_fini - De-initialize the TTM memory pools
>>> + */
>>> Â  void amdgpu_ttm_fini(struct amdgpu_device *adev)
>>> Â  {
>>> Â Â Â Â Â Â Â Â  if (!adev->mman.initialized)
>>> @@ -1854,6 +2155,11 @@ static const struct drm_info_list 
>>> amdgpu_ttm_debugfs_list[] = {
>>> Â  #endif
>>> Â  };
>>>
>>> +/**
>>> + * amdgpu_ttm_vram_read - Linear read access to VRAM
>>> + *
>>> + * Accesses VRAM via MMIO for debugging purposes.
>>> + */
>>> Â  static ssize_t amdgpu_ttm_vram_read(struct file *f, char __user *buf,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  size_t size, loff_t *pos)
>>> Â  {
>>> @@ -1893,6 +2199,11 @@ static ssize_t amdgpu_ttm_vram_read(struct 
>>> file *f, char __user *buf,
>>> Â Â Â Â Â Â Â Â  return result;
>>> Â  }
>>>
>>> +/**
>>> + * amdgpu_ttm_vram_write - Linear write access to VRAM
>>> + *
>>> + * Accesses VRAM via MMIO for debugging purposes.
>>> + */
>>> Â  static ssize_t amdgpu_ttm_vram_write(struct file *f, const char 
>>> __user *buf,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  size_t size, loff_t *pos)
>>> Â  {
>>> @@ -1941,6 +2252,9 @@ static const struct file_operations 
>>> amdgpu_ttm_vram_fops = {
>>>
>>> Â  #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS
>>>
>>> +/**
>>> + * amdgpu_ttm_gtt_read - Linear read access to GTT memory
>>> + */
>>> Â  static ssize_t amdgpu_ttm_gtt_read(struct file *f, char __user *buf,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  size_t size, loff_t *pos)
>>> Â  {
>>> @@ -1988,6 +2302,13 @@ static const struct file_operations 
>>> amdgpu_ttm_gtt_fops = {
>>>
>>> Â  #endif
>>>
>>> +/**
>>> + * amdgpu_iomem_read - Virtual read access to GPU mapped memory
>>> + *
>>> + * This function is used to read memory that has been mapped to the
>>> + * GPU and the known addresses are not physical addresses but instead
>>> + * bus addresses (e.g., what you'd put in an IB or ring buffer).
>>> + */
>>> Â  static ssize_t amdgpu_iomem_read(struct file *f, char __user *buf,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  size_t size, loff_t *pos)
>>> Â  {
>>> @@ -1996,6 +2317,7 @@ static ssize_t amdgpu_iomem_read(struct file 
>>> *f, char __user *buf,
>>> Â Â Â Â Â Â Â Â  ssize_t result = 0;
>>> Â Â Â Â Â Â Â Â  int r;
>>>
>>> +Â Â Â Â Â Â  /* retrieve the IOMMU domain if any for this device */
>>> Â Â Â Â Â Â Â Â  dom = iommu_get_domain_for_dev(adev->dev);
>>>
>>> Â Â Â Â Â Â Â Â  while (size) {
>>> @@ -2008,6 +2330,10 @@ static ssize_t amdgpu_iomem_read(struct file 
>>> *f, char __user *buf,
>>>
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  bytes = bytes < size ? bytes : size;
>>>
>>> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â  /* Translate the bus address to a physical address.Â  If
>>> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  * the domain is NULL it means there is no IOMMU active
>>> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  * and the address translation is the identity
>>> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  */
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  addr = dom ? iommu_iova_to_phys(dom, addr) : addr;
>>>
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  pfn = addr >> PAGE_SHIFT;
>>> @@ -2032,6 +2358,13 @@ static ssize_t amdgpu_iomem_read(struct file 
>>> *f, char __user *buf,
>>> Â Â Â Â Â Â Â Â  return result;
>>> Â  }
>>>
>>> +/**
>>> + * amdgpu_iomem_write - Virtual write access to GPU mapped memory
>>> + *
>>> + * This function is used to write memory that has been mapped to the
>>> + * GPU and the known addresses are not physical addresses but instead
>>> + * bus addresses (e.g., what you'd put in an IB or ring buffer).
>>> + */
>>> Â  static ssize_t amdgpu_iomem_write(struct file *f, const char __user 
>>> *buf,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  size_t size, loff_t *pos)
>>> Â  {
>>> -- 
>>> 2.14.3
>>>
>>> _______________________________________________
>>> 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
> 