]> git.karo-electronics.de Git - karo-tx-linux.git/commitdiff
drm/i915: Execlists small cleanups and micro-optimisations
authorTvrtko Ursulin <tvrtko.ursulin@intel.com>
Fri, 26 Feb 2016 16:58:32 +0000 (16:58 +0000)
committerTvrtko Ursulin <tvrtko.ursulin@intel.com>
Tue, 1 Mar 2016 10:36:02 +0000 (10:36 +0000)
Assorted changes in the areas of code cleanup, reduction of
invariant conditional in the interrupt handler and lock
contention and MMIO access optimisation.

 * Remove needless initialization.
 * Improve cache locality by reorganizing code and/or using
   branch hints to keep unexpected or error conditions out
   of line.
 * Favor busy submit path vs. empty queue.
 * Less branching in hot-paths.

v2:

 * Avoid mmio reads when possible. (Chris Wilson)
 * Use natural integer size for csb indices.
 * Remove useless return value from execlists_update_context.
 * Extract 32-bit ppgtt PDPs update so it is out of line and
   shared with two callers.
 * Grab forcewake across all mmio operations to ease the
   load on uncore lock and use chepear mmio ops.

v3:

 * Removed some more pointless u8 data types.
 * Removed unused return from execlists_context_queue.
 * Commit message updates.

v4:
 * Unclumsify the unqueue if statement. (Chris Wilson)
 * Hide forcewake from the queuing function. (Chris Wilson)

Version 3 now makes the irq handling code path ~20% smaller on
48-bit PPGTT hardware, and a little bit less elsewhere. Hot
paths are mostly in-line now and hammering on the uncore
spinlock is greatly reduced together with mmio traffic to an
extent.

Benchmarking with "gem_latency -n 100" (keep submitting
batches with 100 nop instruction) shows approximately 4% higher
throughput, 2% less CPU time and 22% smaller latencies. This was
on a big-core while small-cores could benefit even more.

Most likely reason for the improvements are the MMIO
optimization and uncore lock traffic reduction.

One odd result is with "gem_latency -n 0" (dispatching empty
batches) which shows 5% more throughput, 8% less CPU time,
25% better producer and consumer latencies, but 15% higher
dispatch latency which is yet unexplained.

Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Link: http://patchwork.freedesktop.org/patch/msgid/1456505912-22286-1-git-send-email-tvrtko.ursulin@linux.intel.com
drivers/gpu/drm/i915/intel_lrc.c
drivers/gpu/drm/i915/intel_ringbuffer.h

index 6a978ce8024436251009979cf1b1b47fa14ff549..27c9ee3f7372f72c4bbd3786b86ce221aafdac19 100644 (file)
@@ -270,6 +270,9 @@ logical_ring_init_platform_invariants(struct intel_engine_cs *ring)
 {
        struct drm_device *dev = ring->dev;
 
+       if (IS_GEN8(dev) || IS_GEN9(dev))
+               ring->idle_lite_restore_wa = ~0;
+
        ring->disable_lite_restore_wa = (IS_SKL_REVID(dev, 0, SKL_REVID_B0) ||
                                        IS_BXT_REVID(dev, 0, BXT_REVID_A1)) &&
                                        (ring->id == VCS || ring->id == VCS2);
@@ -373,8 +376,6 @@ static void execlists_elsp_write(struct drm_i915_gem_request *rq0,
        rq0->elsp_submitted++;
 
        /* You must always write both descriptors in the order below. */
-       spin_lock(&dev_priv->uncore.lock);
-       intel_uncore_forcewake_get__locked(dev_priv, FORCEWAKE_ALL);
        I915_WRITE_FW(RING_ELSP(ring), upper_32_bits(desc[1]));
        I915_WRITE_FW(RING_ELSP(ring), lower_32_bits(desc[1]));
 
@@ -384,11 +385,18 @@ static void execlists_elsp_write(struct drm_i915_gem_request *rq0,
 
        /* ELSP is a wo register, use another nearby reg for posting */
        POSTING_READ_FW(RING_EXECLIST_STATUS_LO(ring));
-       intel_uncore_forcewake_put__locked(dev_priv, FORCEWAKE_ALL);
-       spin_unlock(&dev_priv->uncore.lock);
 }
 
-static int execlists_update_context(struct drm_i915_gem_request *rq)
+static void
+execlists_update_context_pdps(struct i915_hw_ppgtt *ppgtt, u32 *reg_state)
+{
+       ASSIGN_CTX_PDP(ppgtt, reg_state, 3);
+       ASSIGN_CTX_PDP(ppgtt, reg_state, 2);
+       ASSIGN_CTX_PDP(ppgtt, reg_state, 1);
+       ASSIGN_CTX_PDP(ppgtt, reg_state, 0);
+}
+
+static void execlists_update_context(struct drm_i915_gem_request *rq)
 {
        struct intel_engine_cs *ring = rq->ring;
        struct i915_hw_ppgtt *ppgtt = rq->ctx->ppgtt;
@@ -396,19 +404,13 @@ static int execlists_update_context(struct drm_i915_gem_request *rq)
 
        reg_state[CTX_RING_TAIL+1] = rq->tail;
 
-       if (ppgtt && !USES_FULL_48BIT_PPGTT(ppgtt->base.dev)) {
-               /* True 32b PPGTT with dynamic page allocation: update PDP
-                * registers and point the unallocated PDPs to scratch page.
-                * PML4 is allocated during ppgtt init, so this is not needed
-                * in 48-bit mode.
-                */
-               ASSIGN_CTX_PDP(ppgtt, reg_state, 3);
-               ASSIGN_CTX_PDP(ppgtt, reg_state, 2);
-               ASSIGN_CTX_PDP(ppgtt, reg_state, 1);
-               ASSIGN_CTX_PDP(ppgtt, reg_state, 0);
-       }
-
-       return 0;
+       /* True 32b PPGTT with dynamic page allocation: update PDP
+        * registers and point the unallocated PDPs to scratch page.
+        * PML4 is allocated during ppgtt init, so this is not needed
+        * in 48-bit mode.
+        */
+       if (ppgtt && !USES_FULL_48BIT_PPGTT(ppgtt->base.dev))
+               execlists_update_context_pdps(ppgtt, reg_state);
 }
 
 static void execlists_submit_requests(struct drm_i915_gem_request *rq0,
@@ -422,10 +424,10 @@ static void execlists_submit_requests(struct drm_i915_gem_request *rq0,
        execlists_elsp_write(rq0, rq1);
 }
 
-static void execlists_context_unqueue(struct intel_engine_cs *ring)
+static void execlists_context_unqueue__locked(struct intel_engine_cs *ring)
 {
        struct drm_i915_gem_request *req0 = NULL, *req1 = NULL;
-       struct drm_i915_gem_request *cursor = NULL, *tmp = NULL;
+       struct drm_i915_gem_request *cursor, *tmp;
 
        assert_spin_locked(&ring->execlist_lock);
 
@@ -435,9 +437,6 @@ static void execlists_context_unqueue(struct intel_engine_cs *ring)
         */
        WARN_ON(!intel_irqs_enabled(ring->dev->dev_private));
 
-       if (list_empty(&ring->execlist_queue))
-               return;
-
        /* Try to read in pairs */
        list_for_each_entry_safe(cursor, tmp, &ring->execlist_queue,
                                 execlist_link) {
@@ -452,37 +451,48 @@ static void execlists_context_unqueue(struct intel_engine_cs *ring)
                        req0 = cursor;
                } else {
                        req1 = cursor;
+                       WARN_ON(req1->elsp_submitted);
                        break;
                }
        }
 
-       if (IS_GEN8(ring->dev) || IS_GEN9(ring->dev)) {
+       if (unlikely(!req0))
+               return;
+
+       if (req0->elsp_submitted & ring->idle_lite_restore_wa) {
                /*
-                * WaIdleLiteRestore: make sure we never cause a lite
-                * restore with HEAD==TAIL
+                * WaIdleLiteRestore: make sure we never cause a lite restore
+                * with HEAD==TAIL.
+                *
+                * Apply the wa NOOPS to prevent ring:HEAD == req:TAIL as we
+                * resubmit the request. See gen8_emit_request() for where we
+                * prepare the padding after the end of the request.
                 */
-               if (req0->elsp_submitted) {
-                       /*
-                        * Apply the wa NOOPS to prevent ring:HEAD == req:TAIL
-                        * as we resubmit the request. See gen8_emit_request()
-                        * for where we prepare the padding after the end of the
-                        * request.
-                        */
-                       struct intel_ringbuffer *ringbuf;
+               struct intel_ringbuffer *ringbuf;
 
-                       ringbuf = req0->ctx->engine[ring->id].ringbuf;
-                       req0->tail += 8;
-                       req0->tail &= ringbuf->size - 1;
-               }
+               ringbuf = req0->ctx->engine[ring->id].ringbuf;
+               req0->tail += 8;
+               req0->tail &= ringbuf->size - 1;
        }
 
-       WARN_ON(req1 && req1->elsp_submitted);
-
        execlists_submit_requests(req0, req1);
 }
 
-static bool execlists_check_remove_request(struct intel_engine_cs *ring,
-                                          u32 request_id)
+static void execlists_context_unqueue(struct intel_engine_cs *ring)
+{
+       struct drm_i915_private *dev_priv = ring->dev->dev_private;
+
+       spin_lock(&dev_priv->uncore.lock);
+       intel_uncore_forcewake_get__locked(dev_priv, FORCEWAKE_ALL);
+
+       execlists_context_unqueue__locked(ring);
+
+       intel_uncore_forcewake_put__locked(dev_priv, FORCEWAKE_ALL);
+       spin_unlock(&dev_priv->uncore.lock);
+}
+
+static unsigned int
+execlists_check_remove_request(struct intel_engine_cs *ring, u32 request_id)
 {
        struct drm_i915_gem_request *head_req;
 
@@ -492,33 +502,41 @@ static bool execlists_check_remove_request(struct intel_engine_cs *ring,
                                            struct drm_i915_gem_request,
                                            execlist_link);
 
-       if (head_req != NULL) {
-               if (intel_execlists_ctx_id(head_req->ctx, ring) == request_id) {
-                       WARN(head_req->elsp_submitted == 0,
-                            "Never submitted head request\n");
+       if (!head_req)
+               return 0;
 
-                       if (--head_req->elsp_submitted <= 0) {
-                               list_move_tail(&head_req->execlist_link,
-                                              &ring->execlist_retired_req_list);
-                               return true;
-                       }
-               }
-       }
+       if (unlikely(intel_execlists_ctx_id(head_req->ctx, ring) != request_id))
+               return 0;
+
+       WARN(head_req->elsp_submitted == 0, "Never submitted head request\n");
+
+       if (--head_req->elsp_submitted > 0)
+               return 0;
+
+       list_move_tail(&head_req->execlist_link,
+                      &ring->execlist_retired_req_list);
 
-       return false;
+       return 1;
 }
 
-static void get_context_status(struct intel_engine_cs *ring,
-                              u8 read_pointer,
-                              u32 *status, u32 *context_id)
+static u32
+get_context_status(struct intel_engine_cs *ring, unsigned int read_pointer,
+                  u32 *context_id)
 {
        struct drm_i915_private *dev_priv = ring->dev->dev_private;
+       u32 status;
 
-       if (WARN_ON(read_pointer >= GEN8_CSB_ENTRIES))
-               return;
+       read_pointer %= GEN8_CSB_ENTRIES;
+
+       status = I915_READ_FW(RING_CONTEXT_STATUS_BUF_LO(ring, read_pointer));
+
+       if (status & GEN8_CTX_STATUS_IDLE_ACTIVE)
+               return 0;
 
-       *status = I915_READ(RING_CONTEXT_STATUS_BUF_LO(ring, read_pointer));
-       *context_id = I915_READ(RING_CONTEXT_STATUS_BUF_HI(ring, read_pointer));
+       *context_id = I915_READ_FW(RING_CONTEXT_STATUS_BUF_HI(ring,
+                                                             read_pointer));
+
+       return status;
 }
 
 /**
@@ -532,30 +550,27 @@ void intel_lrc_irq_handler(struct intel_engine_cs *ring)
 {
        struct drm_i915_private *dev_priv = ring->dev->dev_private;
        u32 status_pointer;
-       u8 read_pointer;
-       u8 write_pointer;
+       unsigned int read_pointer, write_pointer;
        u32 status = 0;
        u32 status_id;
-       u32 submit_contexts = 0;
+       unsigned int submit_contexts = 0;
+
+       spin_lock(&ring->execlist_lock);
 
-       status_pointer = I915_READ(RING_CONTEXT_STATUS_PTR(ring));
+       spin_lock(&dev_priv->uncore.lock);
+       intel_uncore_forcewake_get__locked(dev_priv, FORCEWAKE_ALL);
+
+       status_pointer = I915_READ_FW(RING_CONTEXT_STATUS_PTR(ring));
 
        read_pointer = ring->next_context_status_buffer;
        write_pointer = GEN8_CSB_WRITE_PTR(status_pointer);
        if (read_pointer > write_pointer)
                write_pointer += GEN8_CSB_ENTRIES;
 
-       spin_lock(&ring->execlist_lock);
-
        while (read_pointer < write_pointer) {
+               status = get_context_status(ring, ++read_pointer, &status_id);
 
-               get_context_status(ring, ++read_pointer % GEN8_CSB_ENTRIES,
-                                  &status, &status_id);
-
-               if (status & GEN8_CTX_STATUS_IDLE_ACTIVE)
-                       continue;
-
-               if (status & GEN8_CTX_STATUS_PREEMPTED) {
+               if (unlikely(status & GEN8_CTX_STATUS_PREEMPTED)) {
                        if (status & GEN8_CTX_STATUS_LITE_RESTORE) {
                                if (execlists_check_remove_request(ring, status_id))
                                        WARN(1, "Lite Restored request removed from queue\n");
@@ -563,37 +578,36 @@ void intel_lrc_irq_handler(struct intel_engine_cs *ring)
                                WARN(1, "Preemption without Lite Restore\n");
                }
 
-               if ((status & GEN8_CTX_STATUS_ACTIVE_IDLE) ||
-                   (status & GEN8_CTX_STATUS_ELEMENT_SWITCH)) {
-                       if (execlists_check_remove_request(ring, status_id))
-                               submit_contexts++;
-               }
+               if (status & (GEN8_CTX_STATUS_ACTIVE_IDLE |
+                   GEN8_CTX_STATUS_ELEMENT_SWITCH))
+                       submit_contexts +=
+                               execlists_check_remove_request(ring, status_id);
        }
 
-       if (ring->disable_lite_restore_wa) {
-               /* Prevent a ctx to preempt itself */
-               if ((status & GEN8_CTX_STATUS_ACTIVE_IDLE) &&
-                   (submit_contexts != 0))
-                       execlists_context_unqueue(ring);
-       } else if (submit_contexts != 0) {
-               execlists_context_unqueue(ring);
+       if (submit_contexts) {
+               if (!ring->disable_lite_restore_wa ||
+                   (status & GEN8_CTX_STATUS_ACTIVE_IDLE))
+                       execlists_context_unqueue__locked(ring);
        }
 
-       spin_unlock(&ring->execlist_lock);
-
-       if (unlikely(submit_contexts > 2))
-               DRM_ERROR("More than two context complete events?\n");
-
        ring->next_context_status_buffer = write_pointer % GEN8_CSB_ENTRIES;
 
        /* Update the read pointer to the old write pointer. Manual ringbuffer
         * management ftw </sarcasm> */
-       I915_WRITE(RING_CONTEXT_STATUS_PTR(ring),
-                  _MASKED_FIELD(GEN8_CSB_READ_PTR_MASK,
-                                ring->next_context_status_buffer << 8));
+       I915_WRITE_FW(RING_CONTEXT_STATUS_PTR(ring),
+                     _MASKED_FIELD(GEN8_CSB_READ_PTR_MASK,
+                                   ring->next_context_status_buffer << 8));
+
+       intel_uncore_forcewake_put__locked(dev_priv, FORCEWAKE_ALL);
+       spin_unlock(&dev_priv->uncore.lock);
+
+       spin_unlock(&ring->execlist_lock);
+
+       if (unlikely(submit_contexts > 2))
+               DRM_ERROR("More than two context complete events?\n");
 }
 
-static int execlists_context_queue(struct drm_i915_gem_request *request)
+static void execlists_context_queue(struct drm_i915_gem_request *request)
 {
        struct intel_engine_cs *ring = request->ring;
        struct drm_i915_gem_request *cursor;
@@ -630,8 +644,6 @@ static int execlists_context_queue(struct drm_i915_gem_request *request)
                execlists_context_unqueue(ring);
 
        spin_unlock_irq(&ring->execlist_lock);
-
-       return 0;
 }
 
 static int logical_ring_invalidate_all_caches(struct drm_i915_gem_request *req)
@@ -1550,7 +1562,7 @@ static int gen8_init_common_ring(struct intel_engine_cs *ring)
 {
        struct drm_device *dev = ring->dev;
        struct drm_i915_private *dev_priv = dev->dev_private;
-       u8 next_context_status_buffer_hw;
+       unsigned int next_context_status_buffer_hw;
 
        lrc_setup_hardware_status_page(ring,
                                dev_priv->kernel_context->engine[ring->id].state);
@@ -2013,6 +2025,7 @@ void intel_logical_ring_cleanup(struct intel_engine_cs *ring)
                ring->status_page.obj = NULL;
        }
 
+       ring->idle_lite_restore_wa = 0;
        ring->disable_lite_restore_wa = false;
        ring->ctx_desc_template = 0;
 
@@ -2439,10 +2452,7 @@ populate_lr_context(struct intel_context *ctx, struct drm_i915_gem_object *ctx_o
                 * With dynamic page allocation, PDPs may not be allocated at
                 * this point. Point the unallocated PDPs to the scratch page
                 */
-               ASSIGN_CTX_PDP(ppgtt, reg_state, 3);
-               ASSIGN_CTX_PDP(ppgtt, reg_state, 2);
-               ASSIGN_CTX_PDP(ppgtt, reg_state, 1);
-               ASSIGN_CTX_PDP(ppgtt, reg_state, 0);
+               execlists_update_context_pdps(ppgtt, reg_state);
        }
 
        if (ring->id == RCS) {
index 566b0ae10ce00f6a3eec840734c0c26c30aef27e..dd910d30a380570cafd0107b5a03069571af3fe2 100644 (file)
@@ -271,7 +271,8 @@ struct  intel_engine_cs {
        spinlock_t execlist_lock;
        struct list_head execlist_queue;
        struct list_head execlist_retired_req_list;
-       u8 next_context_status_buffer;
+       unsigned int next_context_status_buffer;
+       unsigned int idle_lite_restore_wa;
        bool disable_lite_restore_wa;
        u32 ctx_desc_template;
        u32             irq_keep_mask; /* bitmask for interrupts that should not be masked */