Use an LRU policy to map usermode rings to HW compute queues.
Most compute clients use one queue, and usually the first queue
available. This results in poor pipe/queue work distribution when
multiple compute apps are running. In most cases pipe 0 queue 0 is
the only queue that gets used.
In order to better distribute work across multiple HW queues, we adopt
a policy to map the usermode ring ids to the LRU HW queue.
This fixes a large majority of multi-app compute workloads sharing the
same HW queue, even though 7 other queues are available.
v2: use ring->funcs->type instead of ring->hw_ip
Signed-off-by: Andres Rodriguez <andresx7 at gmail.com>
---
drivers/gpu/drm/amd/amdgpu/amdgpu.h | 3 ++
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 3 ++
drivers/gpu/drm/amd/amdgpu/amdgpu_queue_mgr.c | 52 +++++++++++++++++++++++-
drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c | 57 +++++++++++++++++++++++++++
drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h | 4 ++
5 files changed, 117 insertions(+), 2 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
index dc79c0e..a4c16a8 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
@@ -1517,6 +1517,9 @@ struct amdgpu_device {
/* link all gtt */
spinlock_t gtt_list_lock;
struct list_head gtt_list;
+ /* keep an lru list of rings by HW IP */
+ struct list_head ring_lru_list;
+ struct mutex ring_lru_list_lock;
/* record hw reset is performed */
bool has_hw_reset;
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
index 6abb238..954e3b9 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
@@ -1719,6 +1719,9 @@ int amdgpu_device_init(struct amdgpu_device *adev,
INIT_LIST_HEAD(&adev->gtt_list);
spin_lock_init(&adev->gtt_list_lock);
+ INIT_LIST_HEAD(&adev->ring_lru_list);
+ mutex_init(&adev->ring_lru_list_lock);
+
if (adev->asic_type >= CHIP_BONAIRE) {
adev->rmmio_base = pci_resource_start(adev->pdev, 5);
adev->rmmio_size = pci_resource_len(adev->pdev, 5);
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_queue_mgr.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_queue_mgr.c
index 3918bdb..0cf5d24 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_queue_mgr.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_queue_mgr.c
@@ -97,6 +97,44 @@ static struct amdgpu_queue_mapper_funcs identity_mapper = {
.map = amdgpu_identity_map
};
+static enum amdgpu_ring_type amdgpu_hw_ip_to_ring_type(int hw_ip)
+{
+ switch (hw_ip) {
+ case AMDGPU_HW_IP_GFX:
+ return AMDGPU_RING_TYPE_GFX;
+ case AMDGPU_HW_IP_COMPUTE:
+ return AMDGPU_RING_TYPE_COMPUTE;
+ case AMDGPU_HW_IP_DMA:
+ return AMDGPU_RING_TYPE_SDMA;
+ case AMDGPU_HW_IP_UVD:
+ return AMDGPU_RING_TYPE_UVD;
+ case AMDGPU_HW_IP_VCE:
+ return AMDGPU_RING_TYPE_VCE;
+ default:
+ DRM_ERROR("Invalid HW IP specified %d\n", hw_ip);
+ return -1;
+ }
+}
+
+static int amdgpu_lru_map(struct amdgpu_device *adev,
+ struct amdgpu_queue_mapper *mapper,
+ int user_ring,
+ struct amdgpu_ring **out_ring)
+{
+ int r;
+ int ring_type = amdgpu_hw_ip_to_ring_type(mapper->hw_ip);
+
+ r = amdgpu_ring_lru_get(adev, ring_type, out_ring);
+ if (r)
+ return r;
+
+ return update_cached_map(mapper, user_ring, *out_ring);
+}
+
+static struct amdgpu_queue_mapper_funcs lru_mapper = {
+ .map = amdgpu_lru_map
+};
+
int amdgpu_queue_mgr_init(struct amdgpu_device *adev,
struct amdgpu_queue_mgr *mgr)
{
@@ -107,8 +145,18 @@ int amdgpu_queue_mgr_init(struct amdgpu_device *adev,
memset(mgr, 0, sizeof(*mgr));
- for (i = 0; i < AMDGPU_MAX_IP_NUM; ++i)
- amdgpu_queue_mapper_init(&mgr->mapper[i], i, &identity_mapper);
+ for (i = 0; i < AMDGPU_MAX_IP_NUM; ++i) {
+ switch (i) {
+ case AMDGPU_HW_IP_COMPUTE:
+ amdgpu_queue_mapper_init(&mgr->mapper[i], i,
+ &lru_mapper);
+ break;
+ default:
+ amdgpu_queue_mapper_init(&mgr->mapper[i], i,
+ &identity_mapper);
+ break;
+ }
+ }
return 0;
}
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c
index 43cd539..31c6274 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c
@@ -180,6 +180,8 @@ void amdgpu_ring_commit(struct amdgpu_ring *ring)
if (ring->funcs->end_use)
ring->funcs->end_use(ring);
+
+ amdgpu_ring_lru_touch(ring->adev, ring);
}
/**
@@ -279,6 +281,8 @@ int amdgpu_ring_init(struct amdgpu_device *adev, struct amdgpu_ring *ring,
}
ring->ptr_mask = (ring->ring_size / 4) - 1;
ring->max_dw = max_dw;
+ INIT_LIST_HEAD(&ring->lru_list);
+ amdgpu_ring_lru_touch(adev, ring);
if (amdgpu_debugfs_ring_init(adev, ring)) {
DRM_ERROR("Failed to register debugfs file for rings !\n");
@@ -312,6 +316,59 @@ void amdgpu_ring_fini(struct amdgpu_ring *ring)
ring->adev->rings[ring->idx] = NULL;
}
+/**
+ * amdgpu_ring_lru_get - get the least recently used ring for a HW IP block
+ *
+ * @adev: amdgpu_device pointer
+ * @type: amdgpu_ring_type enum
+ * @ring: output ring
+ *
+ * Retreive the amdgpu_ring structure for the least recently used ring of
+ * a specific IP block (all asics).
+ * Returns 0 on success, error on failure.
+ */
+int amdgpu_ring_lru_get(struct amdgpu_device *adev, int type,
+ struct amdgpu_ring **ring)
+{
+ struct amdgpu_ring *entry;
+
+ /* List is sorted in LRU order, find first entry corresponding
+ * to the desired HW IP */
+ *ring = NULL;
+ mutex_lock(&adev->ring_lru_list_lock);
+ list_for_each_entry(entry, &adev->ring_lru_list, lru_list) {
+ if (entry->funcs->type == type) {
+ *ring = entry;
+ break;
+ }
+ }
+ mutex_unlock(&adev->ring_lru_list_lock);
+
+ if (!*ring) {
+ DRM_ERROR("Ring LRU contains no entries for ring type:%d\n", type);
+ return -EINVAL;
+ }
+
+ amdgpu_ring_lru_touch(adev, entry);
+ return 0;
+}
+
+/**
+ * amdgpu_ring_lru_touch - mark a ring as recently being used
+ *
+ * @adev: amdgpu_device pointer
+ * @ring: ring to touch
+ *
+ * Move @ring to the the tail of the lru list
+ */
+void amdgpu_ring_lru_touch(struct amdgpu_device *adev, struct amdgpu_ring *ring)
+{
+ /* list_move_tail handles the case where ring isn't part of the list */
+ mutex_lock(&adev->ring_lru_list_lock);
+ list_move_tail(&ring->lru_list, &adev->ring_lru_list);
+ mutex_unlock(&adev->ring_lru_list_lock);
+}
+
/*
* Debugfs info
*/
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
index 35da5c5..b51bdcd 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
@@ -144,6 +144,7 @@ struct amdgpu_ring {
const struct amdgpu_ring_funcs *funcs;
struct amdgpu_fence_driver fence_drv;
struct amd_gpu_scheduler sched;
+ struct list_head lru_list;
struct amdgpu_bo *ring_obj;
volatile uint32_t *ring;
@@ -186,5 +187,8 @@ int amdgpu_ring_init(struct amdgpu_device *adev, struct amdgpu_ring *ring,
unsigned ring_size, struct amdgpu_irq_src *irq_src,
unsigned irq_type);
void amdgpu_ring_fini(struct amdgpu_ring *ring);
+int amdgpu_ring_lru_get(struct amdgpu_device *adev, int hw_ip,
+ struct amdgpu_ring **ring);
+void amdgpu_ring_lru_touch(struct amdgpu_device *adev, struct amdgpu_ring *ring);
#endif
--
2.9.3
