Skip to contentSkip to conversation

drm/ttm,amdgpu: Introduce LRU bulk move functionality

PATCH Superseded Huang Rui 5 participants 13 messages Aug 17, 2018 – Aug 22, 2018 3 messages not in the archive

Subsystems (4)

v4 of 5 Superseded Latest is v5 →

lore
Share & tools
lore

Huang Rui

author
#0 lore raw

v4 cover drm/ttm,amdgpu: Introduce LRU bulk move functionality

The idea and proposal is originally from Christian, and I continue to work to deliver it.

Background:
amdgpu driver will move all PD/PT and PerVM BOs into idle list. Then move all of them on the end of LRU list one by one. Thus, that cause so many BOs moved to the end of the LRU, and impact performance seriously.

Then Christian provided a workaround to not move PD/PT BOs on LRU with below patch:
"drm/amdgpu: band aid validating VM PTs"
Commit 0bbf32026cf5ba41e9922b30e26e1bed1ecd38ae

However, the final solution should bulk move all PD/PT and PerVM BOs on the LRU instead of one by one.

Whenever amdgpu_vm_validate_pt_bos() is called and we have BOs which need to be validated we move all BOs together to the end of the LRU without dropping the lock for the LRU.

While doing so we note the beginning and end of this block in the LRU list.

Now when amdgpu_vm_validate_pt_bos() is called and we don't have anything to do, we don't move every BO one by one, but instead cut the LRU list into pieces so that we bulk move everything to the end in just one operation.

Test data:

+--------------+-----------------+-----------+---------------------------------------+
|              |The Talos        |Clpeak(OCL)|BusSpeedReadback(OCL)                  |
|              |Principle(Vulkan)|           |                                       |
+------------------------------------------------------------------------------------+
|              |                 |           |0.319 ms(1k) 0.314 ms(2K) 0.308 ms(4K) |
| Original     |  147.7 FPS      |  76.86 us |0.307 ms(8K) 0.310 ms(16K)             |
+------------------------------------------------------------------------------------+
| Orignial + WA|                 |           |0.254 ms(1K) 0.241 ms(2K)              |
|(don't move   |  162.1 FPS      |  42.15 us |0.230 ms(4K) 0.223 ms(8K) 0.204 ms(16K)|
|PT BOs on LRU)|                 |           |                                       |
+------------------------------------------------------------------------------------+
| Bulk move    |  163.1 FPS      |  40.52 us |0.244 ms(1K) 0.252 ms(2K) 0.213 ms(4K) |
|              |                 |           |0.214 ms(8K) 0.225 ms(16K)             |
+--------------+-----------------+-----------+---------------------------------------+

After test them with above three benchmarks include vulkan and opencl. We can see the visible improvement than original, and even better than original with workaround.

Changes from V1 -> V2:
- Fix to missed the BOs in relocated/moved that should be also moved to the end of LRU.

Changes from V2 -> V3:
- Remove unused parameter and use list_for_each_entry instead of the one with save entry.

Changes from V3 -> V4:
- Move the amdgpu_vm_move_to_lru_tail after command submission, at that time, all bo will be back on idle list.

Thanks,
Rui

Christian König (2):
drm/ttm: add helper structures for bulk moves on lru list drm/ttm: revise ttm_bo_move_to_lru_tail to support bulk moves

Huang Rui (3):
drm/ttm: add bulk move function on LRU
drm/amdgpu: use bulk moves for efficient VM LRU handling (v4) drm/amdgpu: move PD/PT bos on LRU again

 drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 11 +++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 73 +++++++++++++++++++++++--------
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h | 11 ++++-
 drivers/gpu/drm/ttm/ttm_bo.c           | 78 +++++++++++++++++++++++++++++++++-
 include/drm/ttm/ttm_bo_api.h           | 16 ++++++-
 include/drm/ttm/ttm_bo_driver.h        | 28 ++++++++++++
 6 files changed, 196 insertions(+), 21 deletions(-)
Signature · 6 lines
2.7.4

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
#1 lore raw

PATCH v4 1/5 drm/ttm: add helper structures for bulk moves on lru list

From: Christian König <[email protected]>

Add bulk move pos to store the pointer of first and last buffer object. The list in between will be bulk moved on lru list.

---
 include/drm/ttm/ttm_bo_driver.h | 28 ++++++++++++++++++++++++++++
 1 file changed, 28 insertions(+)
include/drm/ttm/ttm_bo_driver.h +28 -0
index 3234cc3..e4fee8e 100644
@@ -491,6 +491,34 @@ struct ttm_bo_device {
 };
 
 /**
+ * struct ttm_lru_bulk_move_pos
+ *
+ * @first: first BO in the bulk move range
+ * @last: last BO in the bulk move range
+ *
+ * Positions for a lru bulk move.
+ */
+struct ttm_lru_bulk_move_pos {
+	struct ttm_buffer_object *first;
+	struct ttm_buffer_object *last;
+};
+
+/**
+ * struct ttm_lru_bulk_move
+ *
+ * @tt: first/last lru entry for BOs in the TT domain
+ * @vram: first/last lru entry for BOs in the VRAM domain
+ * @swap: first/last lru entry for BOs on the swap list
+ *
+ * Helper structure for bulk moves on the LRU list.
+ */
+struct ttm_lru_bulk_move {
+	struct ttm_lru_bulk_move_pos tt[TTM_MAX_BO_PRIORITY];
+	struct ttm_lru_bulk_move_pos vram[TTM_MAX_BO_PRIORITY];
+	struct ttm_lru_bulk_move_pos swap[TTM_MAX_BO_PRIORITY];
+};
+
+/**
  * ttm_flag_masked
  *
  * @old: Pointer to the result and original value.
Signature · 6 lines
2.7.4

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
#2 lore raw

PATCH v4 2/5 drm/ttm: revise ttm_bo_move_to_lru_tail to support bulk moves

In reply to Huang Rui

From: Christian König <[email protected]>

When move a BO to the end of LRU, it need remember the BO positions. Make sure all moved bo in between "first" and "last". And they will be bulk moving together.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c |  8 ++++----
 drivers/gpu/drm/ttm/ttm_bo.c           | 26 +++++++++++++++++++++++++-
 include/drm/ttm/ttm_bo_api.h           |  6 +++++-
 3 files changed, 34 insertions(+), 6 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +4 -4
index 015613b..9c84770 100644
@@ -297,9 +297,9 @@ int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 				break;
 
 			spin_lock(&glob->lru_lock);
-			ttm_bo_move_to_lru_tail(&bo->tbo);
+			ttm_bo_move_to_lru_tail(&bo->tbo, NULL);
 			if (bo->shadow)
-				ttm_bo_move_to_lru_tail(&bo->shadow->tbo);
+				ttm_bo_move_to_lru_tail(&bo->shadow->tbo, NULL);
 			spin_unlock(&glob->lru_lock);
 		}
 
@@ -319,9 +319,9 @@ int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 		if (!bo->parent)
 			continue;
 
-		ttm_bo_move_to_lru_tail(&bo->tbo);
+		ttm_bo_move_to_lru_tail(&bo->tbo, NULL);
 		if (bo->shadow)
-			ttm_bo_move_to_lru_tail(&bo->shadow->tbo);
+			ttm_bo_move_to_lru_tail(&bo->shadow->tbo, NULL);
 	}
 	spin_unlock(&glob->lru_lock);
 
drivers/gpu/drm/ttm/ttm_bo.c +25 -1
index 7c48472..7117b6b 100644
@@ -214,12 +214,36 @@ void ttm_bo_del_sub_from_lru(struct ttm_buffer_object *bo)
 }
 EXPORT_SYMBOL(ttm_bo_del_sub_from_lru);
 
-void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo)
+static void ttm_bo_bulk_move_set_pos(struct ttm_lru_bulk_move_pos *pos,
+				     struct ttm_buffer_object *bo)
+{
+	if (!pos->first)
+		pos->first = bo;
+	pos->last = bo;
+}
+
+void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo,
+			     struct ttm_lru_bulk_move *bulk)
 {
 	reservation_object_assert_held(bo->resv);
 
 	ttm_bo_del_from_lru(bo);
 	ttm_bo_add_to_lru(bo);
+
+	if (bulk && !(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
+		switch (bo->mem.mem_type) {
+		case TTM_PL_TT:
+			ttm_bo_bulk_move_set_pos(&bulk->tt[bo->priority], bo);
+			break;
+
+		case TTM_PL_VRAM:
+			ttm_bo_bulk_move_set_pos(&bulk->vram[bo->priority], bo);
+			break;
+		}
+		if (bo->ttm && !(bo->ttm->page_flags &
+				 (TTM_PAGE_FLAG_SG | TTM_PAGE_FLAG_SWAPPED)))
+			ttm_bo_bulk_move_set_pos(&bulk->swap[bo->priority], bo);
+	}
 }
 EXPORT_SYMBOL(ttm_bo_move_to_lru_tail);
 
include/drm/ttm/ttm_bo_api.h +5 -1
index a01ba20..0d4eb81 100644
@@ -51,6 +51,8 @@ struct ttm_placement;
 
 struct ttm_place;
 
+struct ttm_lru_bulk_move;
+
 /**
  * struct ttm_bus_placement
  *
@@ -405,12 +407,14 @@ void ttm_bo_del_from_lru(struct ttm_buffer_object *bo);
  * ttm_bo_move_to_lru_tail
  *
  * @bo: The buffer object.
+ * @bulk: optional bulk move structure to remember BO positions
  *
  * Move this BO to the tail of all lru lists used to lookup and reserve an
  * object. This function must be called with struct ttm_bo_global::lru_lock
  * held, and is used to make a BO less likely to be considered for eviction.
  */
-void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo);
+void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo,
+			     struct ttm_lru_bulk_move *bulk);
 
 /**
  * ttm_bo_lock_delayed_workqueue
Signature · 6 lines
2.7.4

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
#3 lore raw

PATCH v4 3/5 drm/ttm: add bulk move function on LRU

In reply to Huang Rui

This function allow us to bulk move a group of BOs to the tail of their LRU. The positions of group of BOs are stored on the (first, last) bulk_move_pos structure.

---
 drivers/gpu/drm/ttm/ttm_bo.c | 52 ++++++++++++++++++++++++++++++++++++++++++++
 include/drm/ttm/ttm_bo_api.h | 10 +++++++++
 2 files changed, 62 insertions(+)
drivers/gpu/drm/ttm/ttm_bo.c +52 -0
index 7117b6b..39d9d55 100644
@@ -247,6 +247,58 @@ void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo,
 }
 EXPORT_SYMBOL(ttm_bo_move_to_lru_tail);
 
+static void ttm_bo_bulk_move_helper(struct ttm_lru_bulk_move_pos *pos,
+				    struct list_head *lru, bool is_swap)
+{
+	struct list_head entries, before;
+	struct list_head *list1, *list2;
+
+	list1 = is_swap ? &pos->last->swap : &pos->last->lru;
+	list2 = is_swap ? pos->first->swap.prev : pos->first->lru.prev;
+
+	list_cut_position(&entries, lru, list1);
+	list_cut_position(&before, &entries, list2);
+	list_splice(&before, lru);
+	list_splice_tail(&entries, lru);
+}
+
+void ttm_bo_bulk_move_lru_tail(struct ttm_lru_bulk_move *bulk)
+{
+	unsigned i;
+
+	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
+		struct ttm_mem_type_manager *man;
+
+		if (!bulk->tt[i].first)
+			continue;
+
+		man = &bulk->tt[i].first->bdev->man[TTM_PL_TT];
+		ttm_bo_bulk_move_helper(&bulk->tt[i], &man->lru[i], false);
+	}
+
+	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
+		struct ttm_mem_type_manager *man;
+
+		if (!bulk->vram[i].first)
+			continue;
+
+		man = &bulk->vram[i].first->bdev->man[TTM_PL_VRAM];
+		ttm_bo_bulk_move_helper(&bulk->vram[i], &man->lru[i], false);
+	}
+
+	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
+		struct ttm_lru_bulk_move_pos *pos = &bulk->swap[i];
+		struct list_head *lru;
+
+		if (!pos->first)
+			continue;
+
+		lru = &pos->first->bdev->glob->swap_lru[i];
+		ttm_bo_bulk_move_helper(&bulk->swap[i], lru, true);
+	}
+}
+EXPORT_SYMBOL(ttm_bo_bulk_move_lru_tail);
+
 static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
 				  struct ttm_mem_reg *mem, bool evict,
 				  struct ttm_operation_ctx *ctx)
include/drm/ttm/ttm_bo_api.h +10 -0
index 0d4eb81..8c19470 100644
@@ -417,6 +417,16 @@ void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo,
 			     struct ttm_lru_bulk_move *bulk);
 
 /**
+ * ttm_bo_bulk_move_lru_tail
+ *
+ * @bulk: bulk move structure
+ *
+ * Bulk move BOs to the LRU tail, only valid to use when driver makes sure that
+ * BO order never changes. Should be called with ttm_bo_global::lru_lock held.
+ */
+void ttm_bo_bulk_move_lru_tail(struct ttm_lru_bulk_move *bulk);
+
+/**
  * ttm_bo_lock_delayed_workqueue
  *
  * Prevent the delayed workqueue from running.
Signature · 6 lines
2.7.4

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
#4 lore raw

PATCH v4 4/5 drm/amdgpu: use bulk moves for efficient VM LRU handling (v4)

In reply to Huang Rui

I continue to work for bulk moving that based on the proposal by Christian.

Background:
amdgpu driver will move all PD/PT and PerVM BOs into idle list. Then move all of them on the end of LRU list one by one. Thus, that cause so many BOs moved to the end of the LRU, and impact performance seriously.

Then Christian provided a workaround to not move PD/PT BOs on LRU with below patch:
"drm/amdgpu: band aid validating VM PTs"
Commit 0bbf32026cf5ba41e9922b30e26e1bed1ecd38ae

However, the final solution should bulk move all PD/PT and PerVM BOs on the LRU instead of one by one.

Whenever amdgpu_vm_validate_pt_bos() is called and we have BOs which need to be validated we move all BOs together to the end of the LRU without dropping the lock for the LRU.

While doing so we note the beginning and end of this block in the LRU list.

Now when amdgpu_vm_validate_pt_bos() is called and we don't have anything to do, we don't move every BO one by one, but instead cut the LRU list into pieces so that we bulk move everything to the end in just one operation.

Test data:

+--------------+-----------------+-----------+---------------------------------------+
|              |The Talos        |Clpeak(OCL)|BusSpeedReadback(OCL)                  |
|              |Principle(Vulkan)|           |                                       |
+------------------------------------------------------------------------------------+
|              |                 |           |0.319 ms(1k) 0.314 ms(2K) 0.308 ms(4K) |
| Original     |  147.7 FPS      |  76.86 us |0.307 ms(8K) 0.310 ms(16K)             |
+------------------------------------------------------------------------------------+
| Orignial + WA|                 |           |0.254 ms(1K) 0.241 ms(2K)              |
|(don't move   |  162.1 FPS      |  42.15 us |0.230 ms(4K) 0.223 ms(8K) 0.204 ms(16K)|
|PT BOs on LRU)|                 |           |                                       |
+------------------------------------------------------------------------------------+
| Bulk move    |  163.1 FPS      |  40.52 us |0.244 ms(1K) 0.252 ms(2K) 0.213 ms(4K) |
|              |                 |           |0.214 ms(8K) 0.225 ms(16K)             |
+--------------+-----------------+-----------+---------------------------------------+

After test them with above three benchmarks include vulkan and opencl. We can see the visible improvement than original, and even better than original with workaround.

v2: move all BOs include idle, relocated, and moved list to the end of LRU and put them together.
v3: remove unused parameter and use list_for_each_entry instead of the one with save entry.
v4: move the amdgpu_vm_move_to_lru_tail after command submission, at that time, all bo will be back on idle list.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 11 ++++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 71 ++++++++++++++++++++++++++--------
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h | 11 +++++-
 3 files changed, 75 insertions(+), 18 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +11 -0
index 502b94f..9fbdf02 100644
@@ -1260,6 +1260,16 @@ static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
 	return 0;
 }
 
+static void amdgpu_cs_vm_move_on_lru(struct amdgpu_device *adev,
+				     struct amdgpu_cs_parser *p)
+{
+	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
+	struct amdgpu_vm *vm = &fpriv->vm;
+
+	if (vm->validated)
+		amdgpu_vm_move_to_lru_tail(adev, vm);
+}
+
 int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
 {
 	struct amdgpu_device *adev = dev->dev_private;
@@ -1310,6 +1320,7 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
 
 	r = amdgpu_cs_submit(&parser, cs);
 
+	amdgpu_cs_vm_move_on_lru(adev, &parser);
 out:
 	amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
 	return r;
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +54 -17

100 lines not on this page · raw message

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h +10 -1
index 67a15d4..85d57bb 100644
@@ -29,6 +29,7 @@
 #include <linux/rbtree.h>
 #include <drm/gpu_scheduler.h>
 #include <drm/drm_file.h>
+#include <drm/ttm/ttm_bo_driver.h>
 
 #include "amdgpu_sync.h"
 #include "amdgpu_ring.h"
@@ -226,6 +227,11 @@ struct amdgpu_vm {
 
 	/* Some basic info about the task */
 	struct amdgpu_task_info task_info;
+
+	/* Store positions of group of BOs */
+	struct ttm_lru_bulk_move lru_bulk_move;
+	/* mark whether has bo validated on evicted list */
+	bool			validated;
 };
 
 struct amdgpu_vm_manager {
@@ -330,8 +336,11 @@ bool amdgpu_vm_need_pipeline_sync(struct amdgpu_ring *ring,
 void amdgpu_vm_check_compute_bug(struct amdgpu_device *adev);
 
 void amdgpu_vm_get_task_info(struct amdgpu_device *adev, unsigned int pasid,
-			 struct amdgpu_task_info *task_info);
+			     struct amdgpu_task_info *task_info);
 
 void amdgpu_vm_set_task_info(struct amdgpu_vm *vm);
 
+void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev,
+				struct amdgpu_vm *vm);
+
 #endif
Signature · 6 lines
2.7.4

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
maintainer
#5 lore raw

Re: [PATCH v4 4/5] drm/amdgpu: use bulk moves for efficient VM LRU handling (v4)

Am 17.08.2018 um 12:08 schrieb Huang Rui:

Quoted 68 lines

I continue to work for bulk moving that based on the proposal by Christian.

Background:
amdgpu driver will move all PD/PT and PerVM BOs into idle list. Then move all of them on the end of LRU list one by one. Thus, that cause so many BOs moved to the end of the LRU, and impact performance seriously.

Then Christian provided a workaround to not move PD/PT BOs on LRU with below patch:
"drm/amdgpu: band aid validating VM PTs"
Commit 0bbf32026cf5ba41e9922b30e26e1bed1ecd38ae

However, the final solution should bulk move all PD/PT and PerVM BOs on the LRU instead of one by one.

Whenever amdgpu_vm_validate_pt_bos() is called and we have BOs which need to be validated we move all BOs together to the end of the LRU without dropping the lock for the LRU.

While doing so we note the beginning and end of this block in the LRU list.

Now when amdgpu_vm_validate_pt_bos() is called and we don't have anything to do, we don't move every BO one by one, but instead cut the LRU list into pieces so that we bulk move everything to the end in just one operation.

Test data:

+--------------+-----------------+-----------+---------------------------------------+
|              |The Talos        |Clpeak(OCL)|BusSpeedReadback(OCL)                  |
|              |Principle(Vulkan)|           |                                       |
+------------------------------------------------------------------------------------+
|              |                 |           |0.319 ms(1k) 0.314 ms(2K) 0.308 ms(4K) |
| Original     |  147.7 FPS      |  76.86 us |0.307 ms(8K) 0.310 ms(16K)             |
+------------------------------------------------------------------------------------+
| Orignial + WA|                 |           |0.254 ms(1K) 0.241 ms(2K)              |
|(don't move   |  162.1 FPS      |  42.15 us |0.230 ms(4K) 0.223 ms(8K) 0.204 ms(16K)|
|PT BOs on LRU)|                 |           |                                       |
+------------------------------------------------------------------------------------+
| Bulk move    |  163.1 FPS      |  40.52 us |0.244 ms(1K) 0.252 ms(2K) 0.213 ms(4K) |
|              |                 |           |0.214 ms(8K) 0.225 ms(16K)             |
+--------------+-----------------+-----------+---------------------------------------+

After test them with above three benchmarks include vulkan and opencl. We can see the visible improvement than original, and even better than original with workaround.

v2: move all BOs include idle, relocated, and moved list to the end of LRU and put them together.
v3: remove unused parameter and use list_for_each_entry instead of the one with save entry.
v4: move the amdgpu_vm_move_to_lru_tail after command submission, at that time, all bo will be back on idle list.

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 11 ++++++
  drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 71 ++++++++++++++++++++++++++--------
  drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h | 11 +++++-
  3 files changed, 75 insertions(+), 18 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +0 -0
index 502b94f..9fbdf02 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c static int amdgpu_cs_submit(struct amdgpu_cs_parser *p, · line 1269

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +7 -0
@@ -1260,6 +1260,16 @@ static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
  	return 0;
  }
 
+static void amdgpu_cs_vm_move_on_lru(struct amdgpu_device *adev,
+				     struct amdgpu_cs_parser *p)
+{
+	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
+	struct amdgpu_vm *vm = &fpriv->vm;
+
+	if (vm->validated)

That check belongs inside amdgpu_vm_move_to_lru_tail().

Quoted 21 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +4 -0
+		amdgpu_vm_move_to_lru_tail(adev, vm);
+}
+
  int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
  {
  	struct amdgpu_device *adev = dev->dev_private;
@@ -1310,6 +1320,7 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
 
  	r = amdgpu_cs_submit(&parser, cs);
 
+	amdgpu_cs_vm_move_on_lru(adev, &parser);
  out:
  	amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
  	return r;
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +0 -0
index 9c84770..037cfbc 100644
@@ -268,6 +268,53 @@ void amdgpu_vm_get_pd_bo(struct amdgpu_vm *vm,
  }
 

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c void amdgpu_vm_get_pd_bo(struct amdgpu_vm *vm, · line 279

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +9 -0
  /**
+ * amdgpu_vm_move_to_lru_tail_by_list - move one list of BOs to end of LRU
+ *
+ * @vm: vm providing the BOs
+ * @list: the list that stored BOs
+ *
+ * Move one list of BOs to the end of LRU and update the positions.
+ */
+static void
+amdgpu_vm_move_to_lru_tail_by_list(struct amdgpu_vm *vm, struct list_head *list)

I don't see much of a point having a separate function for this any more.

Quoted 42 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +38 -0
+{
+	struct amdgpu_vm_bo_base *bo_base;
+
+	list_for_each_entry(bo_base, list, vm_status) {
+		struct amdgpu_bo *bo = bo_base->bo;
+
+		if (!bo->parent)
+			continue;
+
+		ttm_bo_move_to_lru_tail(&bo->tbo, &vm->lru_bulk_move);
+		if (bo->shadow)
+			ttm_bo_move_to_lru_tail(&bo->shadow->tbo,
+						&vm->lru_bulk_move);
+	}
+}
+
+/**
+ * amdgpu_vm_move_to_lru_tail - move all BOs to the end of LRU
+ *
+ * @adev: amdgpu device pointer
+ * @vm: vm providing the BOs
+ *
+ * Move all BOs to the end of LRU and remember their positions to put them
+ * together.
+ */
+void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev,
+				struct amdgpu_vm *vm)
+{
+	struct ttm_bo_global *glob = adev->mman.bdev.glob;
+
+	memset(&vm->lru_bulk_move, 0, sizeof(vm->lru_bulk_move));
+
+	spin_lock(&glob->lru_lock);
+	amdgpu_vm_move_to_lru_tail_by_list(vm, &vm->idle);
+	spin_unlock(&glob->lru_lock);
+}
+
+/**
   * amdgpu_vm_validate_pt_bos - validate the page table BOs
   *
   * @adev: amdgpu device pointer

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm, · line 338

Quoted 4 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +1 -0
@@ -288,6 +335,7 @@ int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
  	struct amdgpu_vm_bo_base *bo_base, *tmp;
  	int r = 0;
 
+	vm->validated = false;

That won't work like this. It is perfectly possible that CS is aborted because of a signal and amdgpu_vm_move_to_lru_tail() never called.

I suggest to do "vm->validated |= !list_empty(&vm->evicted);" here and then set it to false again in amdgpu_vm_move_to_lru_tail().

And maybe we need a better name than "validated", maybe "bulk_moveable" or something like that.

Quoted 29 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +1 -12
  	list_for_each_entry_safe(bo_base, tmp, &vm->evicted, vm_status) {
  		struct amdgpu_bo *bo = bo_base->bo;
 
@@ -295,14 +343,9 @@ int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
  			r = validate(param, bo);
  			if (r)
  				break;
-
-			spin_lock(&glob->lru_lock);
-			ttm_bo_move_to_lru_tail(&bo->tbo, NULL);
-			if (bo->shadow)
-				ttm_bo_move_to_lru_tail(&bo->shadow->tbo, NULL);
-			spin_unlock(&glob->lru_lock);
  		}
 
+		vm->validated = true;
  		if (bo->tbo.type != ttm_bo_type_kernel) {
  			spin_lock(&vm->moved_lock);
  			list_move(&bo_base->vm_status, &vm->moved);
@@ -312,20 +355,14 @@ int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
  		}
  	}
 
-	spin_lock(&glob->lru_lock);
-	list_for_each_entry(bo_base, &vm->idle, vm_status) {
-		struct amdgpu_bo *bo = bo_base->bo;
-
-		if (!bo->parent)
-			continue;

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm, · line 363

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +4 -4
+	if (vm->validated)
+		return 0;
 
-		ttm_bo_move_to_lru_tail(&bo->tbo, NULL);
-		if (bo->shadow)
-			ttm_bo_move_to_lru_tail(&bo->shadow->tbo, NULL);
-	}
+	spin_lock(&glob->lru_lock);
+	ttm_bo_bulk_move_lru_tail(&vm->lru_bulk_move);
  	spin_unlock(&glob->lru_lock);

Probably best to move that into amdgpu_vm_move_to_lru_tail() as well.

Regards,
Christian.

Quoted 33 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +1 -1
-	return r;
+	return 0;
  }
 
  /**
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h +6 -0
index 67a15d4..85d57bb 100644
@@ -29,6 +29,7 @@
  #include <linux/rbtree.h>
  #include <drm/gpu_scheduler.h>
  #include <drm/drm_file.h>
+#include <drm/ttm/ttm_bo_driver.h>
 
  #include "amdgpu_sync.h"
  #include "amdgpu_ring.h"
@@ -226,6 +227,11 @@ struct amdgpu_vm {
 
  	/* Some basic info about the task */
  	struct amdgpu_task_info task_info;
+
+	/* Store positions of group of BOs */
+	struct ttm_lru_bulk_move lru_bulk_move;
+	/* mark whether has bo validated on evicted list */
+	bool			validated;
  };
 
  struct amdgpu_vm_manager {
@@ -330,8 +336,11 @@ bool amdgpu_vm_need_pipeline_sync(struct amdgpu_ring *ring,
  void amdgpu_vm_check_compute_bug(struct amdgpu_device *adev);
 

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h bool amdgpu_vm_need_pipeline_sync(struct amdgpu_ring *ring, · line 346

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h +4 -1
  void amdgpu_vm_get_task_info(struct amdgpu_device *adev, unsigned int pasid,
-			 struct amdgpu_task_info *task_info);
+			     struct amdgpu_task_info *task_info);
 
  void amdgpu_vm_set_task_info(struct amdgpu_vm *vm);
 
+void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev,
+				struct amdgpu_vm *vm);
+
  #endif

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

A message that is not in the archive (sent off-list or not yet received): aca4bd35-af39-2b08-545e-38305f2ac000-5C7GfCeVMHo@public.gmane.org

author
#7 lore raw

Re: [PATCH v4 4/5] drm/amdgpu: use bulk moves for efficient VM LRU handling (v4)

Quoted 166 lines · Christian

Am 17.08.2018 um 12:08 schrieb Huang Rui:

Quoted 68 lines

I continue to work for bulk moving that based on the proposal by Christian.

Background:
amdgpu driver will move all PD/PT and PerVM BOs into idle list. Then move all of them on the end of LRU list one by one. Thus, that cause so many BOs moved to the end of the LRU, and impact performance seriously.

Then Christian provided a workaround to not move PD/PT BOs on LRU with below patch:
"drm/amdgpu: band aid validating VM PTs"
Commit 0bbf32026cf5ba41e9922b30e26e1bed1ecd38ae

However, the final solution should bulk move all PD/PT and PerVM BOs on the LRU instead of one by one.

Whenever amdgpu_vm_validate_pt_bos() is called and we have BOs which need to be validated we move all BOs together to the end of the LRU without dropping the lock for the LRU.

While doing so we note the beginning and end of this block in the LRU list.

Now when amdgpu_vm_validate_pt_bos() is called and we don't have anything to do, we don't move every BO one by one, but instead cut the LRU list into pieces so that we bulk move everything to the end in just one operation.

Test data:

+--------------+-----------------+-----------+---------------------------------------+
|              |The Talos        |Clpeak(OCL)|BusSpeedReadback(OCL)                  |
|              |Principle(Vulkan)|           |                                       |
+------------------------------------------------------------------------------------+
|              |                 |           |0.319 ms(1k) 0.314 ms(2K) 0.308 ms(4K) |
| Original     |  147.7 FPS      |  76.86 us |0.307 ms(8K) 0.310 ms(16K)             |
+------------------------------------------------------------------------------------+
| Orignial + WA|                 |           |0.254 ms(1K) 0.241 ms(2K)              |
|(don't move   |  162.1 FPS      |  42.15 us |0.230 ms(4K) 0.223 ms(8K) 0.204 ms(16K)|
|PT BOs on LRU)|                 |           |                                       |
+------------------------------------------------------------------------------------+
| Bulk move    |  163.1 FPS      |  40.52 us |0.244 ms(1K) 0.252 ms(2K) 0.213 ms(4K) |
|              |                 |           |0.214 ms(8K) 0.225 ms(16K)             |
+--------------+-----------------+-----------+---------------------------------------+

After test them with above three benchmarks include vulkan and opencl. We can see the visible improvement than original, and even better than original with workaround.

v2: move all BOs include idle, relocated, and moved list to the end of LRU and put them together.
v3: remove unused parameter and use list_for_each_entry instead of the one with save entry.
v4: move the amdgpu_vm_move_to_lru_tail after command submission, at that time, all bo will be back on idle list.

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 11 ++++++
  drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 71 ++++++++++++++++++++++++++--------
  drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h | 11 +++++-
  3 files changed, 75 insertions(+), 18 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +0 -0
index 502b94f..9fbdf02 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c static int amdgpu_cs_submit(struct amdgpu_cs_parser *p, · line 1269

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +7 -0
@@ -1260,6 +1260,16 @@ static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
  	return 0;
  }
 
+static void amdgpu_cs_vm_move_on_lru(struct amdgpu_device *adev,
+				     struct amdgpu_cs_parser *p)
+{
+	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
+	struct amdgpu_vm *vm = &fpriv->vm;
+
+	if (vm->validated)

That check belongs inside amdgpu_vm_move_to_lru_tail().

Quoted 21 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +4 -0
+		amdgpu_vm_move_to_lru_tail(adev, vm);
+}
+
  int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
  {
  	struct amdgpu_device *adev = dev->dev_private;
@@ -1310,6 +1320,7 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
 
  	r = amdgpu_cs_submit(&parser, cs);
 
+	amdgpu_cs_vm_move_on_lru(adev, &parser);
  out:
  	amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
  	return r;
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +0 -0
index 9c84770..037cfbc 100644
@@ -268,6 +268,53 @@ void amdgpu_vm_get_pd_bo(struct amdgpu_vm *vm,
  }
 

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c void amdgpu_vm_get_pd_bo(struct amdgpu_vm *vm, · line 279

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +9 -0
  /**
+ * amdgpu_vm_move_to_lru_tail_by_list - move one list of BOs to end of LRU
+ *
+ * @vm: vm providing the BOs
+ * @list: the list that stored BOs
+ *
+ * Move one list of BOs to the end of LRU and update the positions.
+ */
+static void
+amdgpu_vm_move_to_lru_tail_by_list(struct amdgpu_vm *vm, struct list_head *list)

I don't see much of a point having a separate function for this any more.

Quoted 42 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +38 -0
+{
+	struct amdgpu_vm_bo_base *bo_base;
+
+	list_for_each_entry(bo_base, list, vm_status) {
+		struct amdgpu_bo *bo = bo_base->bo;
+
+		if (!bo->parent)
+			continue;
+
+		ttm_bo_move_to_lru_tail(&bo->tbo, &vm->lru_bulk_move);
+		if (bo->shadow)
+			ttm_bo_move_to_lru_tail(&bo->shadow->tbo,
+						&vm->lru_bulk_move);
+	}
+}
+
+/**
+ * amdgpu_vm_move_to_lru_tail - move all BOs to the end of LRU
+ *
+ * @adev: amdgpu device pointer
+ * @vm: vm providing the BOs
+ *
+ * Move all BOs to the end of LRU and remember their positions to put them
+ * together.
+ */
+void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev,
+				struct amdgpu_vm *vm)
+{
+	struct ttm_bo_global *glob = adev->mman.bdev.glob;
+
+	memset(&vm->lru_bulk_move, 0, sizeof(vm->lru_bulk_move));
+
+	spin_lock(&glob->lru_lock);
+	amdgpu_vm_move_to_lru_tail_by_list(vm, &vm->idle);
+	spin_unlock(&glob->lru_lock);
+}
+
+/**
   * amdgpu_vm_validate_pt_bos - validate the page table BOs
   *
   * @adev: amdgpu device pointer

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm, · line 338

Quoted 4 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +1 -0
@@ -288,6 +335,7 @@ int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
  	struct amdgpu_vm_bo_base *bo_base, *tmp;
  	int r = 0;
 
+	vm->validated = false;

That won't work like this. It is perfectly possible that CS is aborted because of a signal and amdgpu_vm_move_to_lru_tail() never called.

Quoted 6 lines

I suggest to do "vm->validated |= !list_empty(&vm->evicted);" here and then set it to false again in amdgpu_vm_move_to_lru_tail().

And maybe we need a better name than "validated", maybe "bulk_moveable" or something like that.

Actually, "validated" is opposite to "bulk_moveable".
So how about use "vm->bulk_moveable = list_empty(&vm->evicted);" here

Then check the flag after moving the ttm_bo_bulk_move_lru_tail() into amdgpu_vm_move_to_lru_tail() like below:

void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev,
struct amdgpu_vm *vm)

{
        struct ttm_bo_global *glob = adev->mman.bdev.glob;
        struct amdgpu_vm_bo_base *bo_base;

        if (vm->bulk_moveable) {
                spin_lock(&glob->lru_lock);
                ttm_bo_bulk_move_lru_tail(&vm->lru_bulk_move);
                spin_unlock(&glob->lru_lock);
                return;
        }

        memset(&vm->lru_bulk_move, 0, sizeof(vm->lru_bulk_move));

        spin_lock(&glob->lru_lock);
        list_for_each_entry(bo_base, &vm->idle, vm_status) {
                struct amdgpu_bo *bo = bo_base->bo;

                if (!bo->parent)
                        continue;

                ttm_bo_move_to_lru_tail(&bo->tbo, &vm->lru_bulk_move);
                if (bo->shadow)
                        ttm_bo_move_to_lru_tail(&bo->shadow->tbo,
                                                &vm->lru_bulk_move);
        }
        spin_unlock(&glob->lru_lock);

        vm->bulk_moveable = true;
}

Thanks,
Ray

Quoted 82 lines
Quoted 29 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +1 -12
  	list_for_each_entry_safe(bo_base, tmp, &vm->evicted, vm_status) {
  		struct amdgpu_bo *bo = bo_base->bo;
 
@@ -295,14 +343,9 @@ int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
  			r = validate(param, bo);
  			if (r)
  				break;
-
-			spin_lock(&glob->lru_lock);
-			ttm_bo_move_to_lru_tail(&bo->tbo, NULL);
-			if (bo->shadow)
-				ttm_bo_move_to_lru_tail(&bo->shadow->tbo, NULL);
-			spin_unlock(&glob->lru_lock);
  		}
 
+		vm->validated = true;
  		if (bo->tbo.type != ttm_bo_type_kernel) {
  			spin_lock(&vm->moved_lock);
  			list_move(&bo_base->vm_status, &vm->moved);
@@ -312,20 +355,14 @@ int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
  		}
  	}
 
-	spin_lock(&glob->lru_lock);
-	list_for_each_entry(bo_base, &vm->idle, vm_status) {
-		struct amdgpu_bo *bo = bo_base->bo;
-
-		if (!bo->parent)
-			continue;

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm, · line 363

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +4 -4
+	if (vm->validated)
+		return 0;
 
-		ttm_bo_move_to_lru_tail(&bo->tbo, NULL);
-		if (bo->shadow)
-			ttm_bo_move_to_lru_tail(&bo->shadow->tbo, NULL);
-	}
+	spin_lock(&glob->lru_lock);
+	ttm_bo_bulk_move_lru_tail(&vm->lru_bulk_move);
  	spin_unlock(&glob->lru_lock);

Probably best to move that into amdgpu_vm_move_to_lru_tail() as well.

Regards,
Christian.

Quoted 37 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +1 -1
-	return r;
+	return 0;
  }
 
  /**
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h +7 -1
index 67a15d4..85d57bb 100644
@@ -29,6 +29,7 @@
  #include <linux/rbtree.h>
  #include <drm/gpu_scheduler.h>
  #include <drm/drm_file.h>
+#include <drm/ttm/ttm_bo_driver.h>
 
  #include "amdgpu_sync.h"
  #include "amdgpu_ring.h"
@@ -226,6 +227,11 @@ struct amdgpu_vm {
 
  	/* Some basic info about the task */
  	struct amdgpu_task_info task_info;
+
+	/* Store positions of group of BOs */
+	struct ttm_lru_bulk_move lru_bulk_move;
+	/* mark whether has bo validated on evicted list */
+	bool			validated;
  };
 
  struct amdgpu_vm_manager {
@@ -330,8 +336,11 @@ bool amdgpu_vm_need_pipeline_sync(struct amdgpu_ring *ring,
  void amdgpu_vm_check_compute_bug(struct amdgpu_device *adev);
 
  void amdgpu_vm_get_task_info(struct amdgpu_device *adev, unsigned int pasid,
-			 struct amdgpu_task_info *task_info);
+			     struct amdgpu_task_info *task_info);
 
Quoted 7 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h bool amdgpu_vm_need_pipeline_sync(struct amdgpu_ring *ring, · line 346

Quoted 6 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h +3 -0
  void amdgpu_vm_set_task_info(struct amdgpu_vm *vm);
 
+void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev,
+				struct amdgpu_vm *vm);
+
  #endif

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

#8 lore raw

Re: [PATCH v4 4/5] drm/amdgpu: use bulk moves for efficient VM LRU handling (v4)

Am 20.08.2018 um 08:05 schrieb Huang Rui:

Quoted 164 lines · Christian
Quoted 163 lines

Am 17.08.2018 um 12:08 schrieb Huang Rui:

Quoted 68 lines

I continue to work for bulk moving that based on the proposal by Christian.

Background:
amdgpu driver will move all PD/PT and PerVM BOs into idle list. Then move all of them on the end of LRU list one by one. Thus, that cause so many BOs moved to the end of the LRU, and impact performance seriously.

Then Christian provided a workaround to not move PD/PT BOs on LRU with below patch:
"drm/amdgpu: band aid validating VM PTs"
Commit 0bbf32026cf5ba41e9922b30e26e1bed1ecd38ae

However, the final solution should bulk move all PD/PT and PerVM BOs on the LRU instead of one by one.

Whenever amdgpu_vm_validate_pt_bos() is called and we have BOs which need to be validated we move all BOs together to the end of the LRU without dropping the lock for the LRU.

While doing so we note the beginning and end of this block in the LRU list.

Now when amdgpu_vm_validate_pt_bos() is called and we don't have anything to do, we don't move every BO one by one, but instead cut the LRU list into pieces so that we bulk move everything to the end in just one operation.

Test data:

+--------------+-----------------+-----------+---------------------------------------+
|              |The Talos        |Clpeak(OCL)|BusSpeedReadback(OCL)                  |
|              |Principle(Vulkan)|           |                                       |
+------------------------------------------------------------------------------------+
|              |                 |           |0.319 ms(1k) 0.314 ms(2K) 0.308 ms(4K) |
| Original     |  147.7 FPS      |  76.86 us |0.307 ms(8K) 0.310 ms(16K)             |
+------------------------------------------------------------------------------------+
| Orignial + WA|                 |           |0.254 ms(1K) 0.241 ms(2K)              |
|(don't move   |  162.1 FPS      |  42.15 us |0.230 ms(4K) 0.223 ms(8K) 0.204 ms(16K)|
|PT BOs on LRU)|                 |           |                                       |
+------------------------------------------------------------------------------------+
| Bulk move    |  163.1 FPS      |  40.52 us |0.244 ms(1K) 0.252 ms(2K) 0.213 ms(4K) |
|              |                 |           |0.214 ms(8K) 0.225 ms(16K)             |
+--------------+-----------------+-----------+---------------------------------------+

After test them with above three benchmarks include vulkan and opencl. We can see the visible improvement than original, and even better than original with workaround.

v2: move all BOs include idle, relocated, and moved list to the end of LRU and put them together.
v3: remove unused parameter and use list_for_each_entry instead of the one with save entry.
v4: move the amdgpu_vm_move_to_lru_tail after command submission, at that time, all bo will be back on idle list.

---
   drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 11 ++++++
   drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 71 ++++++++++++++++++++++++++--------
   drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h | 11 +++++-
   3 files changed, 75 insertions(+), 18 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +0 -0
index 502b94f..9fbdf02 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c static int amdgpu_cs_submit(struct amdgpu_cs_parser *p, · line 1269

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +7 -0
@@ -1260,6 +1260,16 @@ static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
   	return 0;
   }
 
+static void amdgpu_cs_vm_move_on_lru(struct amdgpu_device *adev,
+				     struct amdgpu_cs_parser *p)
+{
+	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
+	struct amdgpu_vm *vm = &fpriv->vm;
+
+	if (vm->validated)

That check belongs inside amdgpu_vm_move_to_lru_tail().

Quoted 21 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +4 -0
+		amdgpu_vm_move_to_lru_tail(adev, vm);
+}
+
   int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
   {
   	struct amdgpu_device *adev = dev->dev_private;
@@ -1310,6 +1320,7 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
 
   	r = amdgpu_cs_submit(&parser, cs);
 
+	amdgpu_cs_vm_move_on_lru(adev, &parser);
   out:
   	amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
   	return r;
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +0 -0
index 9c84770..037cfbc 100644
@@ -268,6 +268,53 @@ void amdgpu_vm_get_pd_bo(struct amdgpu_vm *vm,
   }
 

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c void amdgpu_vm_get_pd_bo(struct amdgpu_vm *vm, · line 279

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +9 -0
   /**
+ * amdgpu_vm_move_to_lru_tail_by_list - move one list of BOs to end of LRU
+ *
+ * @vm: vm providing the BOs
+ * @list: the list that stored BOs
+ *
+ * Move one list of BOs to the end of LRU and update the positions.
+ */
+static void
+amdgpu_vm_move_to_lru_tail_by_list(struct amdgpu_vm *vm, struct list_head *list)

I don't see much of a point having a separate function for this any more.

Quoted 42 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +38 -0
+{
+	struct amdgpu_vm_bo_base *bo_base;
+
+	list_for_each_entry(bo_base, list, vm_status) {
+		struct amdgpu_bo *bo = bo_base->bo;
+
+		if (!bo->parent)
+			continue;
+
+		ttm_bo_move_to_lru_tail(&bo->tbo, &vm->lru_bulk_move);
+		if (bo->shadow)
+			ttm_bo_move_to_lru_tail(&bo->shadow->tbo,
+						&vm->lru_bulk_move);
+	}
+}
+
+/**
+ * amdgpu_vm_move_to_lru_tail - move all BOs to the end of LRU
+ *
+ * @adev: amdgpu device pointer
+ * @vm: vm providing the BOs
+ *
+ * Move all BOs to the end of LRU and remember their positions to put them
+ * together.
+ */
+void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev,
+				struct amdgpu_vm *vm)
+{
+	struct ttm_bo_global *glob = adev->mman.bdev.glob;
+
+	memset(&vm->lru_bulk_move, 0, sizeof(vm->lru_bulk_move));
+
+	spin_lock(&glob->lru_lock);
+	amdgpu_vm_move_to_lru_tail_by_list(vm, &vm->idle);
+	spin_unlock(&glob->lru_lock);
+}
+
+/**
    * amdgpu_vm_validate_pt_bos - validate the page table BOs
    *
    * @adev: amdgpu device pointer

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm, · line 338

Quoted 4 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +1 -0
@@ -288,6 +335,7 @@ int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
   	struct amdgpu_vm_bo_base *bo_base, *tmp;
   	int r = 0;
 
+	vm->validated = false;

That won't work like this. It is perfectly possible that CS is aborted because of a signal and amdgpu_vm_move_to_lru_tail() never called.

Quoted 8 lines
Quoted 6 lines

I suggest to do "vm->validated |= !list_empty(&vm->evicted);" here and then set it to false again in amdgpu_vm_move_to_lru_tail().

And maybe we need a better name than "validated", maybe "bulk_moveable" or something like that.

Actually, "validated" is opposite to "bulk_moveable".
So how about use "vm->bulk_moveable = list_empty(&vm->evicted);" here

That still won't work correct. See the validation can be interrupted.

So we need something like "vm->bulk_moveable &= list_empty(&vm->evicted);" here.

Apart from that this now looks really good to me.

Regards,
Christian.

Quoted 122 lines

Then check the flag after moving the ttm_bo_bulk_move_lru_tail() into amdgpu_vm_move_to_lru_tail() like below:

void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev,
struct amdgpu_vm *vm)

{
         struct ttm_bo_global *glob = adev->mman.bdev.glob;
         struct amdgpu_vm_bo_base *bo_base;

         if (vm->bulk_moveable) {
                 spin_lock(&glob->lru_lock);
                 ttm_bo_bulk_move_lru_tail(&vm->lru_bulk_move);
                 spin_unlock(&glob->lru_lock);
                 return;
         }

         memset(&vm->lru_bulk_move, 0, sizeof(vm->lru_bulk_move));

         spin_lock(&glob->lru_lock);
         list_for_each_entry(bo_base, &vm->idle, vm_status) {
                 struct amdgpu_bo *bo = bo_base->bo;

                 if (!bo->parent)
                         continue;

                 ttm_bo_move_to_lru_tail(&bo->tbo, &vm->lru_bulk_move);
                 if (bo->shadow)
                         ttm_bo_move_to_lru_tail(&bo->shadow->tbo,
                                                 &vm->lru_bulk_move);
         }
         spin_unlock(&glob->lru_lock);

         vm->bulk_moveable = true;
}

Thanks,
Ray

Quoted 83 lines
Quoted 29 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +1 -12
   	list_for_each_entry_safe(bo_base, tmp, &vm->evicted, vm_status) {
   		struct amdgpu_bo *bo = bo_base->bo;
 
@@ -295,14 +343,9 @@ int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
   			r = validate(param, bo);
   			if (r)
   				break;
-
-			spin_lock(&glob->lru_lock);
-			ttm_bo_move_to_lru_tail(&bo->tbo, NULL);
-			if (bo->shadow)
-				ttm_bo_move_to_lru_tail(&bo->shadow->tbo, NULL);
-			spin_unlock(&glob->lru_lock);
   		}
 
+		vm->validated = true;
   		if (bo->tbo.type != ttm_bo_type_kernel) {
   			spin_lock(&vm->moved_lock);
   			list_move(&bo_base->vm_status, &vm->moved);
@@ -312,20 +355,14 @@ int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
   		}
   	}
 
-	spin_lock(&glob->lru_lock);
-	list_for_each_entry(bo_base, &vm->idle, vm_status) {
-		struct amdgpu_bo *bo = bo_base->bo;
-
-		if (!bo->parent)
-			continue;

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm, · line 363

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +4 -4
+	if (vm->validated)
+		return 0;
 
-		ttm_bo_move_to_lru_tail(&bo->tbo, NULL);
-		if (bo->shadow)
-			ttm_bo_move_to_lru_tail(&bo->shadow->tbo, NULL);
-	}
+	spin_lock(&glob->lru_lock);
+	ttm_bo_bulk_move_lru_tail(&vm->lru_bulk_move);
   	spin_unlock(&glob->lru_lock);

Probably best to move that into amdgpu_vm_move_to_lru_tail() as well.

Regards,
Christian.

Quoted 39 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +1 -1
-	return r;
+	return 0;
   }
 
   /**
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h +7 -1
index 67a15d4..85d57bb 100644
@@ -29,6 +29,7 @@
   #include <linux/rbtree.h>
   #include <drm/gpu_scheduler.h>
   #include <drm/drm_file.h>
+#include <drm/ttm/ttm_bo_driver.h>
 
   #include "amdgpu_sync.h"
   #include "amdgpu_ring.h"
@@ -226,6 +227,11 @@ struct amdgpu_vm {
 
   	/* Some basic info about the task */
   	struct amdgpu_task_info task_info;
+
+	/* Store positions of group of BOs */
+	struct ttm_lru_bulk_move lru_bulk_move;
+	/* mark whether has bo validated on evicted list */
+	bool			validated;
   };
 
   struct amdgpu_vm_manager {
@@ -330,8 +336,11 @@ bool amdgpu_vm_need_pipeline_sync(struct amdgpu_ring *ring,
   void amdgpu_vm_check_compute_bug(struct amdgpu_device *adev);
 
   void amdgpu_vm_get_task_info(struct amdgpu_device *adev, unsigned int pasid,
-			 struct amdgpu_task_info *task_info);
+			     struct amdgpu_task_info *task_info);
 
   void amdgpu_vm_set_task_info(struct amdgpu_vm *vm);
 
Quoted 8 lines
Quoted 4 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h bool amdgpu_vm_need_pipeline_sync(struct amdgpu_ring *ring, · line 346

Quoted 4 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h +3 -0
+void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev,
+				struct amdgpu_vm *vm);
+
   #endif

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

#9 lore raw

Re: [PATCH v4 4/5] drm/amdgpu: use bulk moves for efficient VM LRU handling (v4)

Quoted 171 lines · Christian König

Am 20.08.2018 um 08:05 schrieb Huang Rui:

Quoted 170 lines · Christian
Quoted 169 lines

Am 17.08.2018 um 12:08 schrieb Huang Rui:

Quoted 68 lines

I continue to work for bulk moving that based on the proposal by Christian.

Background:
amdgpu driver will move all PD/PT and PerVM BOs into idle list. Then move all of them on the end of LRU list one by one. Thus, that cause so many BOs moved to the end of the LRU, and impact performance seriously.

Then Christian provided a workaround to not move PD/PT BOs on LRU with below patch:
"drm/amdgpu: band aid validating VM PTs"
Commit 0bbf32026cf5ba41e9922b30e26e1bed1ecd38ae

However, the final solution should bulk move all PD/PT and PerVM BOs on the LRU instead of one by one.

Whenever amdgpu_vm_validate_pt_bos() is called and we have BOs which need to be validated we move all BOs together to the end of the LRU without dropping the lock for the LRU.

While doing so we note the beginning and end of this block in the LRU list.

Now when amdgpu_vm_validate_pt_bos() is called and we don't have anything to do, we don't move every BO one by one, but instead cut the LRU list into pieces so that we bulk move everything to the end in just one operation.

Test data:

+--------------+-----------------+-----------+---------------------------------------+
|              |The Talos        |Clpeak(OCL)|BusSpeedReadback(OCL)                  |
|              |Principle(Vulkan)|           |                                       |
+------------------------------------------------------------------------------------+
|              |                 |           |0.319 ms(1k) 0.314 ms(2K) 0.308 ms(4K) |
| Original     |  147.7 FPS      |  76.86 us |0.307 ms(8K) 0.310 ms(16K)             |
+------------------------------------------------------------------------------------+
| Orignial + WA|                 |           |0.254 ms(1K) 0.241 ms(2K)              |
|(don't move   |  162.1 FPS      |  42.15 us |0.230 ms(4K) 0.223 ms(8K) 0.204 ms(16K)|
|PT BOs on LRU)|                 |           |                                       |
+------------------------------------------------------------------------------------+
| Bulk move    |  163.1 FPS      |  40.52 us |0.244 ms(1K) 0.252 ms(2K) 0.213 ms(4K) |
|              |                 |           |0.214 ms(8K) 0.225 ms(16K)             |
+--------------+-----------------+-----------+---------------------------------------+

After test them with above three benchmarks include vulkan and opencl. We can see the visible improvement than original, and even better than original with workaround.

v2: move all BOs include idle, relocated, and moved list to the end of LRU and put them together.
v3: remove unused parameter and use list_for_each_entry instead of the one with save entry.
v4: move the amdgpu_vm_move_to_lru_tail after command submission, at that time, all bo will be back on idle list.

---
   drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 11 ++++++
   drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 71 ++++++++++++++++++++++++++--------
   drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h | 11 +++++-
   3 files changed, 75 insertions(+), 18 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +0 -0
index 502b94f..9fbdf02 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c static int amdgpu_cs_submit(struct amdgpu_cs_parser *p, · line 1269

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +7 -0
@@ -1260,6 +1260,16 @@ static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
   	return 0;
   }
 
+static void amdgpu_cs_vm_move_on_lru(struct amdgpu_device *adev,
+				     struct amdgpu_cs_parser *p)
+{
+	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
+	struct amdgpu_vm *vm = &fpriv->vm;
+
+	if (vm->validated)

That check belongs inside amdgpu_vm_move_to_lru_tail().

Quoted 21 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +4 -0
+		amdgpu_vm_move_to_lru_tail(adev, vm);
+}
+
   int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
   {
   	struct amdgpu_device *adev = dev->dev_private;
@@ -1310,6 +1320,7 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
 
   	r = amdgpu_cs_submit(&parser, cs);
 
+	amdgpu_cs_vm_move_on_lru(adev, &parser);
   out:
   	amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
   	return r;
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +0 -0
index 9c84770..037cfbc 100644
@@ -268,6 +268,53 @@ void amdgpu_vm_get_pd_bo(struct amdgpu_vm *vm,
   }
 

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c void amdgpu_vm_get_pd_bo(struct amdgpu_vm *vm, · line 279

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +9 -0
   /**
+ * amdgpu_vm_move_to_lru_tail_by_list - move one list of BOs to end of LRU
+ *
+ * @vm: vm providing the BOs
+ * @list: the list that stored BOs
+ *
+ * Move one list of BOs to the end of LRU and update the positions.
+ */
+static void
+amdgpu_vm_move_to_lru_tail_by_list(struct amdgpu_vm *vm, struct list_head *list)

I don't see much of a point having a separate function for this any more.

Quoted 42 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +38 -0
+{
+	struct amdgpu_vm_bo_base *bo_base;
+
+	list_for_each_entry(bo_base, list, vm_status) {
+		struct amdgpu_bo *bo = bo_base->bo;
+
+		if (!bo->parent)
+			continue;
+
+		ttm_bo_move_to_lru_tail(&bo->tbo, &vm->lru_bulk_move);
+		if (bo->shadow)
+			ttm_bo_move_to_lru_tail(&bo->shadow->tbo,
+						&vm->lru_bulk_move);
+	}
+}
+
+/**
+ * amdgpu_vm_move_to_lru_tail - move all BOs to the end of LRU
+ *
+ * @adev: amdgpu device pointer
+ * @vm: vm providing the BOs
+ *
+ * Move all BOs to the end of LRU and remember their positions to put them
+ * together.
+ */
+void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev,
+				struct amdgpu_vm *vm)
+{
+	struct ttm_bo_global *glob = adev->mman.bdev.glob;
+
+	memset(&vm->lru_bulk_move, 0, sizeof(vm->lru_bulk_move));
+
+	spin_lock(&glob->lru_lock);
+	amdgpu_vm_move_to_lru_tail_by_list(vm, &vm->idle);
+	spin_unlock(&glob->lru_lock);
+}
+
+/**
    * amdgpu_vm_validate_pt_bos - validate the page table BOs
    *
    * @adev: amdgpu device pointer

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm, · line 338

Quoted 4 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +1 -0
@@ -288,6 +335,7 @@ int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
   	struct amdgpu_vm_bo_base *bo_base, *tmp;
   	int r = 0;
 
+	vm->validated = false;

That won't work like this. It is perfectly possible that CS is aborted because of a signal and amdgpu_vm_move_to_lru_tail() never called.

I suggest to do "vm->validated |= !list_empty(&vm->evicted);" here and then set it to false again in amdgpu_vm_move_to_lru_tail().

And maybe we need a better name than "validated", maybe "bulk_moveable" or something like that.

Quoted 7 lines
Quoted 2 lines

Actually, "validated" is opposite to "bulk_moveable".
So how about use "vm->bulk_moveable = list_empty(&vm->evicted);" here

That still won't work correct. See the validation can be interrupted.

So we need something like "vm->bulk_moveable &=
list_empty(&vm->evicted);" here.

Thanks. I found if we use vm->bulk_moveable, it will have a corruption of list.

[ 2763.834228] ------------[ cut here ]------------
[ 2763.839221] list_del corruption. prev->next should be ffff9137dae028f8, but was ffffa59b82003ae0
[ 2763.848330] WARNING: CPU: 1 PID: 281 at lib/list_debug.c:53 __list_del_entry_valid+0x7c/0xa0
...
[ 2764.053257] Call Trace:
[ 2764.055748]  ttm_bo_del_from_lru+0x77/0xc0 [ttm]
[ 2764.060437]  ttm_bo_release+0x23a/0x2c0 [ttm]
[ 2764.064924]  amdgpu_bo_destroy+0x93/0x140 [amdgpu]
[ 2764.069769]  ttm_bo_release_list+0x113/0x170 [ttm]
[ 2764.074695]  amdgpu_bo_destroy+0x93/0x140 [amdgpu]
[ 2764.079544]  ttm_bo_release_list+0x113/0x170 [ttm]

However, if use vm->bulk_not_moveable (like vm->validated), the issue is gone. I didn't confirm the reason yet.

Thanks,
Ray

Quoted 137 lines

Apart from that this now looks really good to me.

Regards,
Christian.

Quoted 130 lines

Then check the flag after moving the ttm_bo_bulk_move_lru_tail() into amdgpu_vm_move_to_lru_tail() like below:

void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev,
struct amdgpu_vm *vm)

{
         struct ttm_bo_global *glob = adev->mman.bdev.glob;
         struct amdgpu_vm_bo_base *bo_base;

         if (vm->bulk_moveable) {
                 spin_lock(&glob->lru_lock);
                 ttm_bo_bulk_move_lru_tail(&vm->lru_bulk_move);
                 spin_unlock(&glob->lru_lock);
                 return;
         }

         memset(&vm->lru_bulk_move, 0, sizeof(vm->lru_bulk_move));

         spin_lock(&glob->lru_lock);
         list_for_each_entry(bo_base, &vm->idle, vm_status) {
                 struct amdgpu_bo *bo = bo_base->bo;

                 if (!bo->parent)
                         continue;

                 ttm_bo_move_to_lru_tail(&bo->tbo, &vm->lru_bulk_move);
                 if (bo->shadow)
                         ttm_bo_move_to_lru_tail(&bo->shadow->tbo,
                                                 &vm->lru_bulk_move);
         }
         spin_unlock(&glob->lru_lock);

         vm->bulk_moveable = true;
}

Thanks,
Ray

Quoted 81 lines
Quoted 29 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +1 -12
   	list_for_each_entry_safe(bo_base, tmp, &vm->evicted, vm_status) {
   		struct amdgpu_bo *bo = bo_base->bo;
 
@@ -295,14 +343,9 @@ int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
   			r = validate(param, bo);
   			if (r)
   				break;
-
-			spin_lock(&glob->lru_lock);
-			ttm_bo_move_to_lru_tail(&bo->tbo, NULL);
-			if (bo->shadow)
-				ttm_bo_move_to_lru_tail(&bo->shadow->tbo, NULL);
-			spin_unlock(&glob->lru_lock);
   		}
 
+		vm->validated = true;
   		if (bo->tbo.type != ttm_bo_type_kernel) {
   			spin_lock(&vm->moved_lock);
   			list_move(&bo_base->vm_status, &vm->moved);
@@ -312,20 +355,14 @@ int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
   		}
   	}
 
-	spin_lock(&glob->lru_lock);
-	list_for_each_entry(bo_base, &vm->idle, vm_status) {
-		struct amdgpu_bo *bo = bo_base->bo;
-
-		if (!bo->parent)
-			continue;

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm, · line 363

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +4 -4
+	if (vm->validated)
+		return 0;
 
-		ttm_bo_move_to_lru_tail(&bo->tbo, NULL);
-		if (bo->shadow)
-			ttm_bo_move_to_lru_tail(&bo->shadow->tbo, NULL);
-	}
+	spin_lock(&glob->lru_lock);
+	ttm_bo_bulk_move_lru_tail(&vm->lru_bulk_move);
   	spin_unlock(&glob->lru_lock);

Probably best to move that into amdgpu_vm_move_to_lru_tail() as well.

Regards,
Christian.

Quoted 37 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +1 -1
-	return r;
+	return 0;
   }
 
   /**
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h +7 -1
index 67a15d4..85d57bb 100644
@@ -29,6 +29,7 @@
   #include <linux/rbtree.h>
   #include <drm/gpu_scheduler.h>
   #include <drm/drm_file.h>
+#include <drm/ttm/ttm_bo_driver.h>
 
   #include "amdgpu_sync.h"
   #include "amdgpu_ring.h"
@@ -226,6 +227,11 @@ struct amdgpu_vm {
 
   	/* Some basic info about the task */
   	struct amdgpu_task_info task_info;
+
+	/* Store positions of group of BOs */
+	struct ttm_lru_bulk_move lru_bulk_move;
+	/* mark whether has bo validated on evicted list */
+	bool			validated;
   };
 
   struct amdgpu_vm_manager {
@@ -330,8 +336,11 @@ bool amdgpu_vm_need_pipeline_sync(struct amdgpu_ring *ring,
   void amdgpu_vm_check_compute_bug(struct amdgpu_device *adev);
 
   void amdgpu_vm_get_task_info(struct amdgpu_device *adev, unsigned int pasid,
-			 struct amdgpu_task_info *task_info);
+			     struct amdgpu_task_info *task_info);
 
Quoted 6 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h bool amdgpu_vm_need_pipeline_sync(struct amdgpu_ring *ring, · line 346

Quoted 6 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h +3 -0
   void amdgpu_vm_set_task_info(struct amdgpu_vm *vm);
 
+void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev,
+				struct amdgpu_vm *vm);
+
   #endif

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

#10 lore raw

Re: [PATCH v4 4/5] drm/amdgpu: use bulk moves for efficient VM LRU handling (v4)

Am 21.08.2018 um 15:43 schrieb Huang Rui:

Quoted 192 lines · Christian König
Quoted 171 lines

Am 20.08.2018 um 08:05 schrieb Huang Rui:

Quoted 170 lines · Christian
Quoted 169 lines

Am 17.08.2018 um 12:08 schrieb Huang Rui:

Quoted 68 lines

I continue to work for bulk moving that based on the proposal by Christian.

Background:
amdgpu driver will move all PD/PT and PerVM BOs into idle list. Then move all of them on the end of LRU list one by one. Thus, that cause so many BOs moved to the end of the LRU, and impact performance seriously.

Then Christian provided a workaround to not move PD/PT BOs on LRU with below patch:
"drm/amdgpu: band aid validating VM PTs"
Commit 0bbf32026cf5ba41e9922b30e26e1bed1ecd38ae

However, the final solution should bulk move all PD/PT and PerVM BOs on the LRU instead of one by one.

Whenever amdgpu_vm_validate_pt_bos() is called and we have BOs which need to be validated we move all BOs together to the end of the LRU without dropping the lock for the LRU.

While doing so we note the beginning and end of this block in the LRU list.

Now when amdgpu_vm_validate_pt_bos() is called and we don't have anything to do, we don't move every BO one by one, but instead cut the LRU list into pieces so that we bulk move everything to the end in just one operation.

Test data:

+--------------+-----------------+-----------+---------------------------------------+
|              |The Talos        |Clpeak(OCL)|BusSpeedReadback(OCL)                  |
|              |Principle(Vulkan)|           |                                       |
+------------------------------------------------------------------------------------+
|              |                 |           |0.319 ms(1k) 0.314 ms(2K) 0.308 ms(4K) |
| Original     |  147.7 FPS      |  76.86 us |0.307 ms(8K) 0.310 ms(16K)             |
+------------------------------------------------------------------------------------+
| Orignial + WA|                 |           |0.254 ms(1K) 0.241 ms(2K)              |
|(don't move   |  162.1 FPS      |  42.15 us |0.230 ms(4K) 0.223 ms(8K) 0.204 ms(16K)|
|PT BOs on LRU)|                 |           |                                       |
+------------------------------------------------------------------------------------+
| Bulk move    |  163.1 FPS      |  40.52 us |0.244 ms(1K) 0.252 ms(2K) 0.213 ms(4K) |
|              |                 |           |0.214 ms(8K) 0.225 ms(16K)             |
+--------------+-----------------+-----------+---------------------------------------+

After test them with above three benchmarks include vulkan and opencl. We can see the visible improvement than original, and even better than original with workaround.

v2: move all BOs include idle, relocated, and moved list to the end of LRU and put them together.
v3: remove unused parameter and use list_for_each_entry instead of the one with save entry.
v4: move the amdgpu_vm_move_to_lru_tail after command submission, at that time, all bo will be back on idle list.

---
    drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 11 ++++++
    drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 71 ++++++++++++++++++++++++++--------
    drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h | 11 +++++-
    3 files changed, 75 insertions(+), 18 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +0 -0
index 502b94f..9fbdf02 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c static int amdgpu_cs_submit(struct amdgpu_cs_parser *p, · line 1269

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +7 -0
@@ -1260,6 +1260,16 @@ static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
    	return 0;
    }
 
+static void amdgpu_cs_vm_move_on_lru(struct amdgpu_device *adev,
+				     struct amdgpu_cs_parser *p)
+{
+	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
+	struct amdgpu_vm *vm = &fpriv->vm;
+
+	if (vm->validated)

That check belongs inside amdgpu_vm_move_to_lru_tail().

Quoted 21 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +4 -0
+		amdgpu_vm_move_to_lru_tail(adev, vm);
+}
+
    int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
    {
    	struct amdgpu_device *adev = dev->dev_private;
@@ -1310,6 +1320,7 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
 
    	r = amdgpu_cs_submit(&parser, cs);
 
+	amdgpu_cs_vm_move_on_lru(adev, &parser);
    out:
    	amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
    	return r;
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +0 -0
index 9c84770..037cfbc 100644
@@ -268,6 +268,53 @@ void amdgpu_vm_get_pd_bo(struct amdgpu_vm *vm,
    }
 

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c void amdgpu_vm_get_pd_bo(struct amdgpu_vm *vm, · line 279

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +9 -0
    /**
+ * amdgpu_vm_move_to_lru_tail_by_list - move one list of BOs to end of LRU
+ *
+ * @vm: vm providing the BOs
+ * @list: the list that stored BOs
+ *
+ * Move one list of BOs to the end of LRU and update the positions.
+ */
+static void
+amdgpu_vm_move_to_lru_tail_by_list(struct amdgpu_vm *vm, struct list_head *list)

I don't see much of a point having a separate function for this any more.

Quoted 42 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +38 -0
+{
+	struct amdgpu_vm_bo_base *bo_base;
+
+	list_for_each_entry(bo_base, list, vm_status) {
+		struct amdgpu_bo *bo = bo_base->bo;
+
+		if (!bo->parent)
+			continue;
+
+		ttm_bo_move_to_lru_tail(&bo->tbo, &vm->lru_bulk_move);
+		if (bo->shadow)
+			ttm_bo_move_to_lru_tail(&bo->shadow->tbo,
+						&vm->lru_bulk_move);
+	}
+}
+
+/**
+ * amdgpu_vm_move_to_lru_tail - move all BOs to the end of LRU
+ *
+ * @adev: amdgpu device pointer
+ * @vm: vm providing the BOs
+ *
+ * Move all BOs to the end of LRU and remember their positions to put them
+ * together.
+ */
+void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev,
+				struct amdgpu_vm *vm)
+{
+	struct ttm_bo_global *glob = adev->mman.bdev.glob;
+
+	memset(&vm->lru_bulk_move, 0, sizeof(vm->lru_bulk_move));
+
+	spin_lock(&glob->lru_lock);
+	amdgpu_vm_move_to_lru_tail_by_list(vm, &vm->idle);
+	spin_unlock(&glob->lru_lock);
+}
+
+/**
     * amdgpu_vm_validate_pt_bos - validate the page table BOs
     *
     * @adev: amdgpu device pointer

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm, · line 338

Quoted 4 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +1 -0
@@ -288,6 +335,7 @@ int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
    	struct amdgpu_vm_bo_base *bo_base, *tmp;
    	int r = 0;
 
+	vm->validated = false;

That won't work like this. It is perfectly possible that CS is aborted because of a signal and amdgpu_vm_move_to_lru_tail() never called.

I suggest to do "vm->validated |= !list_empty(&vm->evicted);" here and then set it to false again in amdgpu_vm_move_to_lru_tail().

And maybe we need a better name than "validated", maybe "bulk_moveable" or something like that.

Quoted 6 lines
Quoted 2 lines

Actually, "validated" is opposite to "bulk_moveable".
So how about use "vm->bulk_moveable = list_empty(&vm->evicted);" here

That still won't work correct. See the validation can be interrupted.

So we need something like "vm->bulk_moveable &=
list_empty(&vm->evicted);" here.

Thanks. I found if we use vm->bulk_moveable, it will have a corruption of list.

[ 2763.834228] ------------[ cut here ]------------
[ 2763.839221] list_del corruption. prev->next should be ffff9137dae028f8, but was ffffa59b82003ae0
[ 2763.848330] WARNING: CPU: 1 PID: 281 at lib/list_debug.c:53 __list_del_entry_valid+0x7c/0xa0
...
[ 2764.053257] Call Trace:
[ 2764.055748]  ttm_bo_del_from_lru+0x77/0xc0 [ttm]
[ 2764.060437]  ttm_bo_release+0x23a/0x2c0 [ttm]
[ 2764.064924]  amdgpu_bo_destroy+0x93/0x140 [amdgpu]
[ 2764.069769]  ttm_bo_release_list+0x113/0x170 [ttm]
[ 2764.074695]  amdgpu_bo_destroy+0x93/0x140 [amdgpu]
[ 2764.079544]  ttm_bo_release_list+0x113/0x170 [ttm]
Quoted 2 lines

However, if use vm->bulk_not_moveable (like vm->validated), the issue is gone. I didn't confirm the reason yet.

Make sure that you don't try to move the root PD as well. E.g. test for bo->parent and skip if it is NULL.

Christian.

Quoted 4 lines

Thanks,
Ray

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): d527210c-33c3-d5e6-ccf7-ac308139d1e2-5C7GfCeVMHo@public.gmane.org

author
#12 lore raw

Re: [PATCH v4 4/5] drm/amdgpu: use bulk moves for efficient VM LRU handling (v4)

Quoted 188 lines · Christian König

Am 21.08.2018 um 15:43 schrieb Huang Rui:

Quoted 187 lines · Christian König
Quoted 166 lines

Am 20.08.2018 um 08:05 schrieb Huang Rui:

Quoted 165 lines · Christian
Quoted 164 lines

Am 17.08.2018 um 12:08 schrieb Huang Rui:

Quoted 68 lines

I continue to work for bulk moving that based on the proposal by Christian.

Background:
amdgpu driver will move all PD/PT and PerVM BOs into idle list. Then move all of them on the end of LRU list one by one. Thus, that cause so many BOs moved to the end of the LRU, and impact performance seriously.

Then Christian provided a workaround to not move PD/PT BOs on LRU with below patch:
"drm/amdgpu: band aid validating VM PTs"
Commit 0bbf32026cf5ba41e9922b30e26e1bed1ecd38ae

However, the final solution should bulk move all PD/PT and PerVM BOs on the LRU instead of one by one.

Whenever amdgpu_vm_validate_pt_bos() is called and we have BOs which need to be validated we move all BOs together to the end of the LRU without dropping the lock for the LRU.

While doing so we note the beginning and end of this block in the LRU list.

Now when amdgpu_vm_validate_pt_bos() is called and we don't have anything to do, we don't move every BO one by one, but instead cut the LRU list into pieces so that we bulk move everything to the end in just one operation.

Test data:

+--------------+-----------------+-----------+---------------------------------------+
|              |The Talos        |Clpeak(OCL)|BusSpeedReadback(OCL)                  |
|              |Principle(Vulkan)|           |                                       |
+------------------------------------------------------------------------------------+
|              |                 |           |0.319 ms(1k) 0.314 ms(2K) 0.308 ms(4K) |
| Original     |  147.7 FPS      |  76.86 us |0.307 ms(8K) 0.310 ms(16K)             |
+------------------------------------------------------------------------------------+
| Orignial + WA|                 |           |0.254 ms(1K) 0.241 ms(2K)              |
|(don't move   |  162.1 FPS      |  42.15 us |0.230 ms(4K) 0.223 ms(8K) 0.204 ms(16K)|
|PT BOs on LRU)|                 |           |                                       |
+------------------------------------------------------------------------------------+
| Bulk move    |  163.1 FPS      |  40.52 us |0.244 ms(1K) 0.252 ms(2K) 0.213 ms(4K) |
|              |                 |           |0.214 ms(8K) 0.225 ms(16K)             |
+--------------+-----------------+-----------+---------------------------------------+

After test them with above three benchmarks include vulkan and opencl. We can see the visible improvement than original, and even better than original with workaround.

v2: move all BOs include idle, relocated, and moved list to the end of LRU and put them together.
v3: remove unused parameter and use list_for_each_entry instead of the one with save entry.
v4: move the amdgpu_vm_move_to_lru_tail after command submission, at that time, all bo will be back on idle list.

---
   drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 11 ++++++
   drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 71 ++++++++++++++++++++++++++--------
   drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h | 11 +++++-
   3 files changed, 75 insertions(+), 18 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +0 -0
index 502b94f..9fbdf02 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c static int amdgpu_cs_submit(struct amdgpu_cs_parser *p, · line 1268

Quoted 9 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +7 -0
@@ -1260,6 +1260,16 @@ static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
   	return 0;
   }
+static void amdgpu_cs_vm_move_on_lru(struct amdgpu_device *adev,
+				     struct amdgpu_cs_parser *p)
+{
+	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
+	struct amdgpu_vm *vm = &fpriv->vm;
+
+	if (vm->validated)

That check belongs inside amdgpu_vm_move_to_lru_tail().

Quoted 18 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +4 -0
+		amdgpu_vm_move_to_lru_tail(adev, vm);
+}
+
   int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
   {
   	struct amdgpu_device *adev = dev->dev_private;
@@ -1310,6 +1320,7 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
   	r = amdgpu_cs_submit(&parser, cs);
+	amdgpu_cs_vm_move_on_lru(adev, &parser);
   out:
   	amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
   	return r;
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +0 -0
index 9c84770..037cfbc 100644
@@ -268,6 +268,53 @@ void amdgpu_vm_get_pd_bo(struct amdgpu_vm *vm,
   }

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c void amdgpu_vm_get_pd_bo(struct amdgpu_vm *vm, · line 278

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +9 -0
   /**
+ * amdgpu_vm_move_to_lru_tail_by_list - move one list of BOs to end of LRU
+ *
+ * @vm: vm providing the BOs
+ * @list: the list that stored BOs
+ *
+ * Move one list of BOs to the end of LRU and update the positions.
+ */
+static void
+amdgpu_vm_move_to_lru_tail_by_list(struct amdgpu_vm *vm, struct list_head *list)

I don't see much of a point having a separate function for this any more.

Quoted 42 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +38 -0
+{
+	struct amdgpu_vm_bo_base *bo_base;
+
+	list_for_each_entry(bo_base, list, vm_status) {
+		struct amdgpu_bo *bo = bo_base->bo;
+
+		if (!bo->parent)
+			continue;
+
+		ttm_bo_move_to_lru_tail(&bo->tbo, &vm->lru_bulk_move);
+		if (bo->shadow)
+			ttm_bo_move_to_lru_tail(&bo->shadow->tbo,
+						&vm->lru_bulk_move);
+	}
+}
+
+/**
+ * amdgpu_vm_move_to_lru_tail - move all BOs to the end of LRU
+ *
+ * @adev: amdgpu device pointer
+ * @vm: vm providing the BOs
+ *
+ * Move all BOs to the end of LRU and remember their positions to put them
+ * together.
+ */
+void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev,
+				struct amdgpu_vm *vm)
+{
+	struct ttm_bo_global *glob = adev->mman.bdev.glob;
+
+	memset(&vm->lru_bulk_move, 0, sizeof(vm->lru_bulk_move));
+
+	spin_lock(&glob->lru_lock);
+	amdgpu_vm_move_to_lru_tail_by_list(vm, &vm->idle);
+	spin_unlock(&glob->lru_lock);
+}
+
+/**
    * amdgpu_vm_validate_pt_bos - validate the page table BOs
    *
    * @adev: amdgpu device pointer

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm, · line 337

Quoted 3 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +1 -0
@@ -288,6 +335,7 @@ int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
   	struct amdgpu_vm_bo_base *bo_base, *tmp;
   	int r = 0;
+	vm->validated = false;

That won't work like this. It is perfectly possible that CS is aborted because of a signal and amdgpu_vm_move_to_lru_tail() never called.

I suggest to do "vm->validated |= !list_empty(&vm->evicted);" here and then set it to false again in amdgpu_vm_move_to_lru_tail().

And maybe we need a better name than "validated", maybe "bulk_moveable" or something like that.

Quoted 6 lines
Quoted 2 lines

Actually, "validated" is opposite to "bulk_moveable".
So how about use "vm->bulk_moveable = list_empty(&vm->evicted);" here

That still won't work correct. See the validation can be interrupted.

So we need something like "vm->bulk_moveable &=
list_empty(&vm->evicted);" here.

Thanks. I found if we use vm->bulk_moveable, it will have a corruption of list.

[ 2763.834228] ------------[ cut here ]------------
[ 2763.839221] list_del corruption. prev->next should be ffff9137dae028f8, but was ffffa59b82003ae0
[ 2763.848330] WARNING: CPU: 1 PID: 281 at lib/list_debug.c:53 __list_del_entry_valid+0x7c/0xa0
...
[ 2764.053257] Call Trace:
[ 2764.055748]  ttm_bo_del_from_lru+0x77/0xc0 [ttm]
[ 2764.060437]  ttm_bo_release+0x23a/0x2c0 [ttm]
[ 2764.064924]  amdgpu_bo_destroy+0x93/0x140 [amdgpu]
[ 2764.069769]  ttm_bo_release_list+0x113/0x170 [ttm]
[ 2764.074695]  amdgpu_bo_destroy+0x93/0x140 [amdgpu]
[ 2764.079544]  ttm_bo_release_list+0x113/0x170 [ttm]
Quoted 6 lines
Quoted 2 lines

However, if use vm->bulk_not_moveable (like vm->validated), the issue is gone. I didn't confirm the reason yet.

Make sure that you don't try to move the root PD as well. E.g. test for bo->parent and skip if it is NULL.

Yes, I have skipped the root PD, but the issue is still existed.

list_for_each_entry(bo_base, &vm->idle, vm_status) {
        struct amdgpu_bo *bo = bo_base->bo;

        if (!bo->parent)
                continue;

        ttm_bo_move_to_lru_tail(&bo->tbo, &vm->lru_bulk_move);
        if (bo->shadow)
                ttm_bo_move_to_lru_tail(&bo->shadow->tbo,
                                        &vm->lru_bulk_move);
}

Let me continue to narrow down this issue.

Thanks,
Ray
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

author
#13 lore raw

Re: [PATCH v4 4/5] drm/amdgpu: use bulk moves for efficient VM LRU handling (v4)

Quoted 210 lines · Huang Rui
Quoted 188 lines · Christian König

Am 21.08.2018 um 15:43 schrieb Huang Rui:

Quoted 187 lines · Christian König
Quoted 166 lines

Am 20.08.2018 um 08:05 schrieb Huang Rui:

Quoted 165 lines · Christian
Quoted 164 lines

Am 17.08.2018 um 12:08 schrieb Huang Rui:

Quoted 68 lines

I continue to work for bulk moving that based on the proposal by Christian.

Background:
amdgpu driver will move all PD/PT and PerVM BOs into idle list. Then move all of them on the end of LRU list one by one. Thus, that cause so many BOs moved to the end of the LRU, and impact performance seriously.

Then Christian provided a workaround to not move PD/PT BOs on LRU with below patch:
"drm/amdgpu: band aid validating VM PTs"
Commit 0bbf32026cf5ba41e9922b30e26e1bed1ecd38ae

However, the final solution should bulk move all PD/PT and PerVM BOs on the LRU instead of one by one.

Whenever amdgpu_vm_validate_pt_bos() is called and we have BOs which need to be validated we move all BOs together to the end of the LRU without dropping the lock for the LRU.

While doing so we note the beginning and end of this block in the LRU list.

Now when amdgpu_vm_validate_pt_bos() is called and we don't have anything to do, we don't move every BO one by one, but instead cut the LRU list into pieces so that we bulk move everything to the end in just one operation.

Test data:

+--------------+-----------------+-----------+---------------------------------------+
|              |The Talos        |Clpeak(OCL)|BusSpeedReadback(OCL)                  |
|              |Principle(Vulkan)|           |                                       |
+------------------------------------------------------------------------------------+
|              |                 |           |0.319 ms(1k) 0.314 ms(2K) 0.308 ms(4K) |
| Original     |  147.7 FPS      |  76.86 us |0.307 ms(8K) 0.310 ms(16K)             |
+------------------------------------------------------------------------------------+
| Orignial + WA|                 |           |0.254 ms(1K) 0.241 ms(2K)              |
|(don't move   |  162.1 FPS      |  42.15 us |0.230 ms(4K) 0.223 ms(8K) 0.204 ms(16K)|
|PT BOs on LRU)|                 |           |                                       |
+------------------------------------------------------------------------------------+
| Bulk move    |  163.1 FPS      |  40.52 us |0.244 ms(1K) 0.252 ms(2K) 0.213 ms(4K) |
|              |                 |           |0.214 ms(8K) 0.225 ms(16K)             |
+--------------+-----------------+-----------+---------------------------------------+

After test them with above three benchmarks include vulkan and opencl. We can see the visible improvement than original, and even better than original with workaround.

v2: move all BOs include idle, relocated, and moved list to the end of LRU and put them together.
v3: remove unused parameter and use list_for_each_entry instead of the one with save entry.
v4: move the amdgpu_vm_move_to_lru_tail after command submission, at that time, all bo will be back on idle list.

---
   drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 11 ++++++
   drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 71 ++++++++++++++++++++++++++--------
   drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h | 11 +++++-
   3 files changed, 75 insertions(+), 18 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +0 -0
index 502b94f..9fbdf02 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c static int amdgpu_cs_submit(struct amdgpu_cs_parser *p, · line 1268

Quoted 9 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +7 -0
@@ -1260,6 +1260,16 @@ static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
   	return 0;
   }
+static void amdgpu_cs_vm_move_on_lru(struct amdgpu_device *adev,
+				     struct amdgpu_cs_parser *p)
+{
+	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
+	struct amdgpu_vm *vm = &fpriv->vm;
+
+	if (vm->validated)

That check belongs inside amdgpu_vm_move_to_lru_tail().

Quoted 18 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +4 -0
+		amdgpu_vm_move_to_lru_tail(adev, vm);
+}
+
   int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
   {
   	struct amdgpu_device *adev = dev->dev_private;
@@ -1310,6 +1320,7 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
   	r = amdgpu_cs_submit(&parser, cs);
+	amdgpu_cs_vm_move_on_lru(adev, &parser);
   out:
   	amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
   	return r;
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +0 -0
index 9c84770..037cfbc 100644
@@ -268,6 +268,53 @@ void amdgpu_vm_get_pd_bo(struct amdgpu_vm *vm,
   }

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c void amdgpu_vm_get_pd_bo(struct amdgpu_vm *vm, · line 278

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +9 -0
   /**
+ * amdgpu_vm_move_to_lru_tail_by_list - move one list of BOs to end of LRU
+ *
+ * @vm: vm providing the BOs
+ * @list: the list that stored BOs
+ *
+ * Move one list of BOs to the end of LRU and update the positions.
+ */
+static void
+amdgpu_vm_move_to_lru_tail_by_list(struct amdgpu_vm *vm, struct list_head *list)

I don't see much of a point having a separate function for this any more.

Quoted 42 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +38 -0
+{
+	struct amdgpu_vm_bo_base *bo_base;
+
+	list_for_each_entry(bo_base, list, vm_status) {
+		struct amdgpu_bo *bo = bo_base->bo;
+
+		if (!bo->parent)
+			continue;
+
+		ttm_bo_move_to_lru_tail(&bo->tbo, &vm->lru_bulk_move);
+		if (bo->shadow)
+			ttm_bo_move_to_lru_tail(&bo->shadow->tbo,
+						&vm->lru_bulk_move);
+	}
+}
+
+/**
+ * amdgpu_vm_move_to_lru_tail - move all BOs to the end of LRU
+ *
+ * @adev: amdgpu device pointer
+ * @vm: vm providing the BOs
+ *
+ * Move all BOs to the end of LRU and remember their positions to put them
+ * together.
+ */
+void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev,
+				struct amdgpu_vm *vm)
+{
+	struct ttm_bo_global *glob = adev->mman.bdev.glob;
+
+	memset(&vm->lru_bulk_move, 0, sizeof(vm->lru_bulk_move));
+
+	spin_lock(&glob->lru_lock);
+	amdgpu_vm_move_to_lru_tail_by_list(vm, &vm->idle);
+	spin_unlock(&glob->lru_lock);
+}
+
+/**
    * amdgpu_vm_validate_pt_bos - validate the page table BOs
    *
    * @adev: amdgpu device pointer

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm, · line 337

Quoted 3 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +1 -0
@@ -288,6 +335,7 @@ int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
   	struct amdgpu_vm_bo_base *bo_base, *tmp;
   	int r = 0;
+	vm->validated = false;

That won't work like this. It is perfectly possible that CS is aborted because of a signal and amdgpu_vm_move_to_lru_tail() never called.

I suggest to do "vm->validated |= !list_empty(&vm->evicted);" here and then set it to false again in amdgpu_vm_move_to_lru_tail().

And maybe we need a better name than "validated", maybe "bulk_moveable" or something like that.

Quoted 6 lines
Quoted 2 lines

Actually, "validated" is opposite to "bulk_moveable".
So how about use "vm->bulk_moveable = list_empty(&vm->evicted);" here

That still won't work correct. See the validation can be interrupted.

So we need something like "vm->bulk_moveable &=
list_empty(&vm->evicted);" here.

Thanks. I found if we use vm->bulk_moveable, it will have a corruption of list.

[ 2763.834228] ------------[ cut here ]------------
[ 2763.839221] list_del corruption. prev->next should be ffff9137dae028f8, but was ffffa59b82003ae0
[ 2763.848330] WARNING: CPU: 1 PID: 281 at lib/list_debug.c:53 __list_del_entry_valid+0x7c/0xa0
...
[ 2764.053257] Call Trace:
[ 2764.055748]  ttm_bo_del_from_lru+0x77/0xc0 [ttm]
[ 2764.060437]  ttm_bo_release+0x23a/0x2c0 [ttm]
[ 2764.064924]  amdgpu_bo_destroy+0x93/0x140 [amdgpu]
[ 2764.069769]  ttm_bo_release_list+0x113/0x170 [ttm]
[ 2764.074695]  amdgpu_bo_destroy+0x93/0x140 [amdgpu]
[ 2764.079544]  ttm_bo_release_list+0x113/0x170 [ttm]
Quoted 6 lines
Quoted 2 lines

However, if use vm->bulk_not_moveable (like vm->validated), the issue is gone. I didn't confirm the reason yet.

Make sure that you don't try to move the root PD as well. E.g. test for bo->parent and skip if it is NULL.

Yes, I have skipped the root PD, but the issue is still existed.

list_for_each_entry(bo_base, &vm->idle, vm_status) {
        struct amdgpu_bo *bo = bo_base->bo;

        if (!bo->parent)
                continue;

        ttm_bo_move_to_lru_tail(&bo->tbo, &vm->lru_bulk_move);
        if (bo->shadow)
                ttm_bo_move_to_lru_tail(&bo->shadow->tbo,
                                        &vm->lru_bulk_move);
}
Quoted 2 lines

Let me continue to narrow down this issue.

I found the cause of this corruption.
The init value of vm->bulk_moveable is actually false.
Then after vm->bulk_moveable &= list_empty(&vm->evicted), the bulk_moveable is still false even the evicted list is empty at that time. That's not expected.

So we just set vm->bulk_moveable as true at amdgpu_vm_init() during vm init, the issue is fixed.

Thanks,
Ray
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 1534500481-15123-1-git-send-email-ray.huang-5C7GfCeVMHo@public.gmane.org

author
#15 lore raw

PATCH v4 5/5 drm/amdgpu: move PD/PT bos on LRU again

The new bulk moving functionality is ready, the overhead of moving PD/PT bos to LRU is fixed. So move them on LRU again.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +1 -1
index 037cfbc..f498b7c 100644
@@ -1121,7 +1121,7 @@ int amdgpu_vm_update_directories(struct amdgpu_device *adev,
 					   struct amdgpu_vm_bo_base,
 					   vm_status);
 		bo_base->moved = false;
-		list_del_init(&bo_base->vm_status);
+		list_move(&bo_base->vm_status, &vm->idle);
 
 		bo = bo_base->bo->parent;
 		if (!bo)
Signature · 1 line
2.7.4

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx