Somalapuram Amaranath
2023-Jan-25 10:49 UTC
[Nouveau] [PATCH v2 1/4] drm/amdgpu: Use cursor start instead of ttm resource start
cleanup PAGE_SHIFT operation and replacing
ttm_resource resource->start with cursor start
using amdgpu_res_first API.
v1 -> v2: reorder patch sequence
Signed-off-by: Somalapuram Amaranath <Amaranath.Somalapuram at amd.com>
---
drivers/gpu/drm/amd/amdgpu/amdgpu_object.c | 11 ++++++++---
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 10 +++++++---
2 files changed, 15 insertions(+), 6 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
index 25a68d8888e0..2ab67ab204df 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
@@ -1346,6 +1346,7 @@ vm_fault_t amdgpu_bo_fault_reserve_notify(struct
ttm_buffer_object *bo)
struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
struct ttm_operation_ctx ctx = { false, false };
struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
+ struct amdgpu_res_cursor cursor;
unsigned long offset;
int r;
@@ -1355,7 +1356,8 @@ vm_fault_t amdgpu_bo_fault_reserve_notify(struct
ttm_buffer_object *bo)
if (bo->resource->mem_type != TTM_PL_VRAM)
return 0;
- offset = bo->resource->start << PAGE_SHIFT;
+ amdgpu_res_first(bo->resource, 0, bo->resource->size, &cursor);
+ offset = cursor.start;
if ((offset + bo->base.size) <= adev->gmc.visible_vram_size)
return 0;
@@ -1378,7 +1380,8 @@ vm_fault_t amdgpu_bo_fault_reserve_notify(struct
ttm_buffer_object *bo)
else if (unlikely(r))
return VM_FAULT_SIGBUS;
- offset = bo->resource->start << PAGE_SHIFT;
+ amdgpu_res_first(bo->resource, 0, bo->resource->size, &cursor);
+ offset = cursor.start;
/* this should never happen */
if (bo->resource->mem_type == TTM_PL_VRAM &&
(offset + bo->base.size) > adev->gmc.visible_vram_size)
@@ -1491,9 +1494,11 @@ u64 amdgpu_bo_gpu_offset(struct amdgpu_bo *bo)
u64 amdgpu_bo_gpu_offset_no_check(struct amdgpu_bo *bo)
{
struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
+ struct amdgpu_res_cursor cursor;
uint64_t offset;
- offset = (bo->tbo.resource->start << PAGE_SHIFT) +
+ amdgpu_res_first(bo->tbo.resource, 0, bo->tbo.resource->size,
&cursor);
+ offset = cursor.start +
amdgpu_ttm_domain_start(adev, bo->tbo.resource->mem_type);
return amdgpu_gmc_sign_extend(offset);
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
index c5ef7f7bdc15..ffe6a1ab7f9a 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
@@ -849,6 +849,7 @@ static int amdgpu_ttm_backend_bind(struct ttm_device *bdev,
{
struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm);
+ struct amdgpu_res_cursor cursor;
uint64_t flags;
int r;
@@ -896,7 +897,8 @@ static int amdgpu_ttm_backend_bind(struct ttm_device *bdev,
flags = amdgpu_ttm_tt_pte_flags(adev, ttm, bo_mem);
/* bind pages into GART page tables */
- gtt->offset = (u64)bo_mem->start << PAGE_SHIFT;
+ amdgpu_res_first(bo_mem, 0, bo_mem->size, &cursor);
+ gtt->offset = cursor.start;
amdgpu_gart_bind(adev, gtt->offset, ttm->num_pages,
gtt->ttm.dma_address, flags);
gtt->bound = true;
@@ -916,6 +918,7 @@ int amdgpu_ttm_alloc_gart(struct ttm_buffer_object *bo)
struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
struct ttm_operation_ctx ctx = { false, false };
struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(bo->ttm);
+ struct amdgpu_res_cursor cursor;
struct ttm_placement placement;
struct ttm_place placements;
struct ttm_resource *tmp;
@@ -927,7 +930,7 @@ int amdgpu_ttm_alloc_gart(struct ttm_buffer_object *bo)
addr = amdgpu_gmc_agp_addr(bo);
if (addr != AMDGPU_BO_INVALID_OFFSET) {
- bo->resource->start = addr >> PAGE_SHIFT;
+ bo->resource->start = addr;
return 0;
}
@@ -949,7 +952,8 @@ int amdgpu_ttm_alloc_gart(struct ttm_buffer_object *bo)
flags = amdgpu_ttm_tt_pte_flags(adev, bo->ttm, tmp);
/* Bind pages */
- gtt->offset = (u64)tmp->start << PAGE_SHIFT;
+ amdgpu_res_first(tmp, 0, tmp->size, &cursor);
+ gtt->offset = cursor.start;
amdgpu_ttm_gart_bind(adev, bo, flags);
amdgpu_gart_invalidate_tlb(adev);
ttm_resource_free(bo, &bo->resource);
--
2.32.0
Somalapuram Amaranath
2023-Jan-25 10:49 UTC
[Nouveau] [PATCH v2 2/4] drm/amdkfd: Use cursor start instead of ttm resource start
cleanup PAGE_SHIFT operation and replacing
ttm_resource resource->start with cursor start
using amdgpu_res_first API
v1 -> v2: reorder patch sequence
Signed-off-by: Somalapuram Amaranath <Amaranath.Somalapuram at amd.com>
---
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 5 ++++-
1 file changed, 4 insertions(+), 1 deletion(-)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
index c06ada0844ba..f87ce4f1cb93 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
@@ -200,8 +200,11 @@ static int add_queue_mes(struct device_queue_manager *dqm,
struct queue *q,
queue_input.wptr_addr = (uint64_t)q->properties.write_ptr;
if (q->wptr_bo) {
+ struct amdgpu_res_cursor cursor;
wptr_addr_off = (uint64_t)q->properties.write_ptr & (PAGE_SIZE - 1);
- queue_input.wptr_mc_addr =
((uint64_t)q->wptr_bo->tbo.resource->start << PAGE_SHIFT) +
wptr_addr_off;
+ amdgpu_res_first(q->wptr_bo->tbo.resource, 0,
+ q->wptr_bo->tbo.resource->size, &cursor);
+ queue_input.wptr_mc_addr = cursor.start + wptr_addr_off;
}
queue_input.is_kfd_process = 1;
--
2.32.0
Somalapuram Amaranath
2023-Jan-25 10:49 UTC
[Nouveau] [PATCH v2 3/4] drm/amdgpu: Movie the amdgpu_gtt_mgr start and size from pages to bytes
To support GTT manager amdgpu_res_first, amdgpu_res_next from pages to bytes and clean up PAGE_SHIFT operation. v1 -> v2: reorder patch sequence Signed-off-by: Somalapuram Amaranath <Amaranath.Somalapuram at amd.com> --- drivers/gpu/drm/amd/amdgpu/amdgpu_res_cursor.h | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_res_cursor.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_res_cursor.h index 5c4f93ee0c57..5c78f0b09351 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_res_cursor.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_res_cursor.h @@ -94,8 +94,8 @@ static inline void amdgpu_res_first(struct ttm_resource *res, while (start >= node->size << PAGE_SHIFT) start -= node++->size << PAGE_SHIFT; - cur->start = (node->start << PAGE_SHIFT) + start; - cur->size = min((node->size << PAGE_SHIFT) - start, size); + cur->start = node->start + start; + cur->size = min(node->size - start, size); cur->remaining = size; cur->node = node; break; @@ -155,8 +155,8 @@ static inline void amdgpu_res_next(struct amdgpu_res_cursor *cur, uint64_t size) node = cur->node; cur->node = ++node; - cur->start = node->start << PAGE_SHIFT; - cur->size = min(node->size << PAGE_SHIFT, cur->remaining); + cur->start = node->start; + cur->size = min(node->size, cur->remaining); break; default: return; -- 2.32.0
Somalapuram Amaranath
2023-Jan-25 10:49 UTC
[Nouveau] [PATCH v2 4/4] drm/amdgpu: Support allocate of amdgpu_gtt_mgr from pages to bytes
Change the GTT manager init and allocate from pages to bytes
v1 -> v2: reorder patch sequence
Signed-off-by: Somalapuram Amaranath <Amaranath.Somalapuram at amd.com>
---
drivers/gpu/drm/amd/amdgpu/amdgpu_gtt_mgr.c | 13 +++++++------
1 file changed, 7 insertions(+), 6 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gtt_mgr.c
b/drivers/gpu/drm/amd/amdgpu/amdgpu_gtt_mgr.c
index 44367f03316f..a1fbfc5984d8 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gtt_mgr.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gtt_mgr.c
@@ -116,7 +116,6 @@ static int amdgpu_gtt_mgr_new(struct ttm_resource_manager
*man,
struct ttm_resource **res)
{
struct amdgpu_gtt_mgr *mgr = to_gtt_mgr(man);
- uint32_t num_pages = PFN_UP(tbo->base.size);
struct ttm_range_mgr_node *node;
int r;
@@ -134,8 +133,10 @@ static int amdgpu_gtt_mgr_new(struct ttm_resource_manager
*man,
if (place->lpfn) {
spin_lock(&mgr->lock);
r = drm_mm_insert_node_in_range(&mgr->mm, &node->mm_nodes[0],
- num_pages, tbo->page_alignment,
- 0, place->fpfn, place->lpfn,
+ tbo->base.size,
+ tbo->page_alignment << PAGE_SHIFT, 0,
+ place->fpfn << PAGE_SHIFT,
+ place->lpfn << PAGE_SHIFT,
DRM_MM_INSERT_BEST);
spin_unlock(&mgr->lock);
if (unlikely(r))
@@ -144,7 +145,7 @@ static int amdgpu_gtt_mgr_new(struct ttm_resource_manager
*man,
node->base.start = node->mm_nodes[0].start;
} else {
node->mm_nodes[0].start = 0;
- node->mm_nodes[0].size = PFN_UP(node->base.size);
+ node->mm_nodes[0].size = node->base.size;
node->base.start = AMDGPU_BO_INVALID_OFFSET;
}
@@ -285,8 +286,8 @@ int amdgpu_gtt_mgr_init(struct amdgpu_device *adev, uint64_t
gtt_size)
ttm_resource_manager_init(man, &adev->mman.bdev, gtt_size);
- start = AMDGPU_GTT_MAX_TRANSFER_SIZE * AMDGPU_GTT_NUM_TRANSFER_WINDOWS;
- size = (adev->gmc.gart_size >> PAGE_SHIFT) - start;
+ start = (AMDGPU_GTT_MAX_TRANSFER_SIZE * AMDGPU_GTT_NUM_TRANSFER_WINDOWS)
<< PAGE_SHIFT;
+ size = adev->gmc.gart_size - start;
drm_mm_init(&mgr->mm, start, size);
spin_lock_init(&mgr->lock);
--
2.32.0
Christian König
2023-Jan-25 11:41 UTC
[Nouveau] [PATCH v2 1/4] drm/amdgpu: Use cursor start instead of ttm resource start
Am 25.01.23 um 11:48 schrieb Somalapuram Amaranath:> cleanup PAGE_SHIFT operation and replacing > ttm_resource resource->start with cursor start > using amdgpu_res_first API. > v1 -> v2: reorder patch sequence > > Signed-off-by: Somalapuram Amaranath <Amaranath.Somalapuram at amd.com> > --- > drivers/gpu/drm/amd/amdgpu/amdgpu_object.c | 11 ++++++++--- > drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 10 +++++++--- > 2 files changed, 15 insertions(+), 6 deletions(-) > > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c > index 25a68d8888e0..2ab67ab204df 100644 > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c > @@ -1346,6 +1346,7 @@ vm_fault_t amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo) > struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev); > struct ttm_operation_ctx ctx = { false, false }; > struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo); > + struct amdgpu_res_cursor cursor; > unsigned long offset; > int r; > > @@ -1355,7 +1356,8 @@ vm_fault_t amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo) > if (bo->resource->mem_type != TTM_PL_VRAM) > return 0; > > - offset = bo->resource->start << PAGE_SHIFT; > + amdgpu_res_first(bo->resource, 0, bo->resource->size, &cursor); > + offset = cursor.start;That won't work like this. We use a virtual resource start for this check here.> if ((offset + bo->base.size) <= adev->gmc.visible_vram_size) > return 0;Just replace this whole test with calling amdgpu_bo_in_cpu_visible_vram().> > @@ -1378,7 +1380,8 @@ vm_fault_t amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo) > else if (unlikely(r)) > return VM_FAULT_SIGBUS; > > - offset = bo->resource->start << PAGE_SHIFT; > + amdgpu_res_first(bo->resource, 0, bo->resource->size, &cursor); > + offset = cursor.start; > /* this should never happen */ > if (bo->resource->mem_type == TTM_PL_VRAM && > (offset + bo->base.size) > adev->gmc.visible_vram_size)Same here, just call amdgpu_bo_in_cpu_visible_vram() instead.> @@ -1491,9 +1494,11 @@ u64 amdgpu_bo_gpu_offset(struct amdgpu_bo *bo) > u64 amdgpu_bo_gpu_offset_no_check(struct amdgpu_bo *bo) > { > struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev); > + struct amdgpu_res_cursor cursor; > uint64_t offset; > > - offset = (bo->tbo.resource->start << PAGE_SHIFT) + > + amdgpu_res_first(bo->tbo.resource, 0, bo->tbo.resource->size, &cursor); > + offset = cursor.start + > amdgpu_ttm_domain_start(adev, bo->tbo.resource->mem_type); > > return amdgpu_gmc_sign_extend(offset); > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c > index c5ef7f7bdc15..ffe6a1ab7f9a 100644 > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c > @@ -849,6 +849,7 @@ static int amdgpu_ttm_backend_bind(struct ttm_device *bdev, > { > struct amdgpu_device *adev = amdgpu_ttm_adev(bdev); > struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(ttm); > + struct amdgpu_res_cursor cursor; > uint64_t flags; > int r; > > @@ -896,7 +897,8 @@ static int amdgpu_ttm_backend_bind(struct ttm_device *bdev, > flags = amdgpu_ttm_tt_pte_flags(adev, ttm, bo_mem); > > /* bind pages into GART page tables */ > - gtt->offset = (u64)bo_mem->start << PAGE_SHIFT; > + amdgpu_res_first(bo_mem, 0, bo_mem->size, &cursor); > + gtt->offset = cursor.start; > amdgpu_gart_bind(adev, gtt->offset, ttm->num_pages, > gtt->ttm.dma_address, flags); > gtt->bound = true; > @@ -916,6 +918,7 @@ int amdgpu_ttm_alloc_gart(struct ttm_buffer_object *bo) > struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev); > struct ttm_operation_ctx ctx = { false, false }; > struct amdgpu_ttm_tt *gtt = ttm_to_amdgpu_ttm_tt(bo->ttm); > + struct amdgpu_res_cursor cursor; > struct ttm_placement placement; > struct ttm_place placements; > struct ttm_resource *tmp; > @@ -927,7 +930,7 @@ int amdgpu_ttm_alloc_gart(struct ttm_buffer_object *bo) > > addr = amdgpu_gmc_agp_addr(bo); > if (addr != AMDGPU_BO_INVALID_OFFSET) { > - bo->resource->start = addr >> PAGE_SHIFT; > + bo->resource->start = addr;This doesn't belong into this patch and would break if only this patch is applied. Please move it into a later patch. Apart from those comments that looks good now, Christian.> return 0; > } > > @@ -949,7 +952,8 @@ int amdgpu_ttm_alloc_gart(struct ttm_buffer_object *bo) > flags = amdgpu_ttm_tt_pte_flags(adev, bo->ttm, tmp); > > /* Bind pages */ > - gtt->offset = (u64)tmp->start << PAGE_SHIFT; > + amdgpu_res_first(tmp, 0, tmp->size, &cursor); > + gtt->offset = cursor.start; > amdgpu_ttm_gart_bind(adev, bo, flags); > amdgpu_gart_invalidate_tlb(adev); > ttm_resource_free(bo, &bo->resource);
Christian König
2023-Jan-25 11:43 UTC
[Nouveau] [PATCH v2 2/4] drm/amdkfd: Use cursor start instead of ttm resource start
Am 25.01.23 um 11:48 schrieb Somalapuram Amaranath:> cleanup PAGE_SHIFT operation and replacing > ttm_resource resource->start with cursor start > using amdgpu_res_first API > v1 -> v2: reorder patch sequence > > Signed-off-by: Somalapuram Amaranath <Amaranath.Somalapuram at amd.com> > --- > drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 5 ++++- > 1 file changed, 4 insertions(+), 1 deletion(-) > > diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c > index c06ada0844ba..f87ce4f1cb93 100644 > --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c > +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c > @@ -200,8 +200,11 @@ static int add_queue_mes(struct device_queue_manager *dqm, struct queue *q, > queue_input.wptr_addr = (uint64_t)q->properties.write_ptr; > > if (q->wptr_bo) { > + struct amdgpu_res_cursor cursor; > wptr_addr_off = (uint64_t)q->properties.write_ptr & (PAGE_SIZE - 1);Add an empty line between declaration and code or otherwise the automated checkers will complain. Apart from this nit pick the patch is Reviewed-by: Christian K?nig <christian.koenig at amd.com> Regards, Christian.> - queue_input.wptr_mc_addr = ((uint64_t)q->wptr_bo->tbo.resource->start << PAGE_SHIFT) + wptr_addr_off; > + amdgpu_res_first(q->wptr_bo->tbo.resource, 0, > + q->wptr_bo->tbo.resource->size, &cursor); > + queue_input.wptr_mc_addr = cursor.start + wptr_addr_off; > } > > queue_input.is_kfd_process = 1;
Possibly Parallel Threads
- [PATCH v2 1/4] drm/amdgpu: Use cursor start instead of ttm resource start
- [PATCH v3 1/4] drm/amdgpu: Use cursor start instead of ttm resource start
- [PATCH v3 3/4] drm/amdgpu: Movie the amdgpu_gtt_mgr start and size from pages to bytes
- [PATCH] Change the meaning of the fields in the ttm_place structure from pfn to bytes
- [PATCH 15/15] amdgpu: remove CONFIG_DRM_AMDGPU_USERPTR