drm/i915/bdw: Handle context switch events
authorThomas Daniel <thomas.daniel@intel.com>
Thu, 24 Jul 2014 16:04:39 +0000 (17:04 +0100)
committerDaniel Vetter <daniel.vetter@ffwll.ch>
Thu, 14 Aug 2014 20:43:47 +0000 (22:43 +0200)
Handle all context status events in the context status buffer on every
context switch interrupt. We only remove work from the execlist queue
after a context status buffer reports that it has completed and we only
attempt to schedule new contexts on interrupt when a previously submitted
context completes (unless no contexts are queued, which means the GPU is
free).

We canot call intel_runtime_pm_get() in an interrupt (or with a spinlock
grabbed, FWIW), because it might sleep, which is not a nice thing to do.
Instead, do the runtime_pm get/put together with the create/destroy request,
and handle the forcewake get/put directly.

Signed-off-by: Thomas Daniel <thomas.daniel@intel.com>
v2: Unreferencing the context when we are freeing the request might free
the backing bo, which requires the struct_mutex to be grabbed, so defer
unreferencing and freeing to a bottom half.

v3:
- Ack the interrupt inmediately, before trying to handle it (fix for
missing interrupts by Bob Beckett <robert.beckett@intel.com>).
- Update the Context Status Buffer Read Pointer, just in case (spotted
by Damien Lespiau).

v4: New namespace and multiple rebase changes.

v5: Squash with "drm/i915/bdw: Do not call intel_runtime_pm_get() in an
interrupt", as suggested by Daniel.

Signed-off-by: Oscar Mateo <oscar.mateo@intel.com>
Reviewed-by: Damien Lespiau <damien.lespiau@intel.com>
[danvet: Checkpatch ...]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
drivers/gpu/drm/i915/i915_irq.c
drivers/gpu/drm/i915/intel_lrc.c
drivers/gpu/drm/i915/intel_lrc.h
drivers/gpu/drm/i915/intel_ringbuffer.h

index 00957fa0b87721018005b11494564d5668287896..f5d6795887d2d99c650b4774ab547aba90e50778 100644 (file)
@@ -1632,6 +1632,7 @@ static irqreturn_t gen8_gt_irq_handler(struct drm_device *dev,
                                       struct drm_i915_private *dev_priv,
                                       u32 master_ctl)
 {
+       struct intel_engine_cs *ring;
        u32 rcs, bcs, vcs;
        uint32_t tmp = 0;
        irqreturn_t ret = IRQ_NONE;
@@ -1641,14 +1642,20 @@ static irqreturn_t gen8_gt_irq_handler(struct drm_device *dev,
                if (tmp) {
                        I915_WRITE(GEN8_GT_IIR(0), tmp);
                        ret = IRQ_HANDLED;
+
                        rcs = tmp >> GEN8_RCS_IRQ_SHIFT;
-                       bcs = tmp >> GEN8_BCS_IRQ_SHIFT;
+                       ring = &dev_priv->ring[RCS];
                        if (rcs & GT_RENDER_USER_INTERRUPT)
-                               notify_ring(dev, &dev_priv->ring[RCS]);
+                               notify_ring(dev, ring);
+                       if (rcs & GT_CONTEXT_SWITCH_INTERRUPT)
+                               intel_execlists_handle_ctx_events(ring);
+
+                       bcs = tmp >> GEN8_BCS_IRQ_SHIFT;
+                       ring = &dev_priv->ring[BCS];
                        if (bcs & GT_RENDER_USER_INTERRUPT)
-                               notify_ring(dev, &dev_priv->ring[BCS]);
-                       if ((rcs | bcs) & GT_CONTEXT_SWITCH_INTERRUPT)
-                               DRM_DEBUG_DRIVER("TODO: Context switch\n");
+                               notify_ring(dev, ring);
+                       if (bcs & GT_CONTEXT_SWITCH_INTERRUPT)
+                               intel_execlists_handle_ctx_events(ring);
                } else
                        DRM_ERROR("The master control interrupt lied (GT0)!\n");
        }
@@ -1658,16 +1665,20 @@ static irqreturn_t gen8_gt_irq_handler(struct drm_device *dev,
                if (tmp) {
                        I915_WRITE(GEN8_GT_IIR(1), tmp);
                        ret = IRQ_HANDLED;
+
                        vcs = tmp >> GEN8_VCS1_IRQ_SHIFT;
+                       ring = &dev_priv->ring[VCS];
                        if (vcs & GT_RENDER_USER_INTERRUPT)
-                               notify_ring(dev, &dev_priv->ring[VCS]);
+                               notify_ring(dev, ring);
                        if (vcs & GT_CONTEXT_SWITCH_INTERRUPT)
-                               DRM_DEBUG_DRIVER("TODO: Context switch\n");
+                               intel_execlists_handle_ctx_events(ring);
+
                        vcs = tmp >> GEN8_VCS2_IRQ_SHIFT;
+                       ring = &dev_priv->ring[VCS2];
                        if (vcs & GT_RENDER_USER_INTERRUPT)
-                               notify_ring(dev, &dev_priv->ring[VCS2]);
+                               notify_ring(dev, ring);
                        if (vcs & GT_CONTEXT_SWITCH_INTERRUPT)
-                               DRM_DEBUG_DRIVER("TODO: Context switch\n");
+                               intel_execlists_handle_ctx_events(ring);
                } else
                        DRM_ERROR("The master control interrupt lied (GT1)!\n");
        }
@@ -1688,11 +1699,13 @@ static irqreturn_t gen8_gt_irq_handler(struct drm_device *dev,
                if (tmp) {
                        I915_WRITE(GEN8_GT_IIR(3), tmp);
                        ret = IRQ_HANDLED;
+
                        vcs = tmp >> GEN8_VECS_IRQ_SHIFT;
+                       ring = &dev_priv->ring[VECS];
                        if (vcs & GT_RENDER_USER_INTERRUPT)
-                               notify_ring(dev, &dev_priv->ring[VECS]);
+                               notify_ring(dev, ring);
                        if (vcs & GT_CONTEXT_SWITCH_INTERRUPT)
-                               DRM_DEBUG_DRIVER("TODO: Context switch\n");
+                               intel_execlists_handle_ctx_events(ring);
                } else
                        DRM_ERROR("The master control interrupt lied (GT3)!\n");
        }
index e81f5f6c49b908e5f2b86ed30faa4adebfde2226..22f6a7c0cb1881a2ba471bb9285e703a5d2f2f12 100644 (file)
 #define RING_ELSP(ring)                        ((ring)->mmio_base+0x230)
 #define RING_EXECLIST_STATUS(ring)     ((ring)->mmio_base+0x234)
 #define RING_CONTEXT_CONTROL(ring)     ((ring)->mmio_base+0x244)
+#define RING_CONTEXT_STATUS_BUF(ring)  ((ring)->mmio_base+0x370)
+#define RING_CONTEXT_STATUS_PTR(ring)  ((ring)->mmio_base+0x3a0)
+
+#define RING_EXECLIST_QFULL            (1 << 0x2)
+#define RING_EXECLIST1_VALID           (1 << 0x3)
+#define RING_EXECLIST0_VALID           (1 << 0x4)
+#define RING_EXECLIST_ACTIVE_STATUS    (3 << 0xE)
+#define RING_EXECLIST1_ACTIVE          (1 << 0x11)
+#define RING_EXECLIST0_ACTIVE          (1 << 0x12)
+
+#define GEN8_CTX_STATUS_IDLE_ACTIVE    (1 << 0)
+#define GEN8_CTX_STATUS_PREEMPTED      (1 << 1)
+#define GEN8_CTX_STATUS_ELEMENT_SWITCH (1 << 2)
+#define GEN8_CTX_STATUS_ACTIVE_IDLE    (1 << 3)
+#define GEN8_CTX_STATUS_COMPLETE       (1 << 4)
+#define GEN8_CTX_STATUS_LITE_RESTORE   (1 << 15)
 
 #define CTX_LRI_HEADER_0               0x01
 #define CTX_CONTEXT_CONTROL            0x02
@@ -150,6 +166,7 @@ static void execlists_elsp_write(struct intel_engine_cs *ring,
        struct drm_i915_private *dev_priv = ring->dev->dev_private;
        uint64_t temp = 0;
        uint32_t desc[4];
+       unsigned long flags;
 
        /* XXX: You must always write both descriptors in the order below. */
        if (ctx_obj1)
@@ -163,9 +180,17 @@ static void execlists_elsp_write(struct intel_engine_cs *ring,
        desc[3] = (u32)(temp >> 32);
        desc[2] = (u32)temp;
 
-       /* Set Force Wakeup bit to prevent GT from entering C6 while
-        * ELSP writes are in progress */
-       gen6_gt_force_wake_get(dev_priv, FORCEWAKE_ALL);
+       /* Set Force Wakeup bit to prevent GT from entering C6 while ELSP writes
+        * are in progress.
+        *
+        * The other problem is that we can't just call gen6_gt_force_wake_get()
+        * because that function calls intel_runtime_pm_get(), which might sleep.
+        * Instead, we do the runtime_pm_get/put when creating/destroying requests.
+        */
+       spin_lock_irqsave(&dev_priv->uncore.lock, flags);
+       if (dev_priv->uncore.forcewake_count++ == 0)
+               dev_priv->uncore.funcs.force_wake_get(dev_priv, FORCEWAKE_ALL);
+       spin_unlock_irqrestore(&dev_priv->uncore.lock, flags);
 
        I915_WRITE(RING_ELSP(ring), desc[1]);
        I915_WRITE(RING_ELSP(ring), desc[0]);
@@ -176,7 +201,11 @@ static void execlists_elsp_write(struct intel_engine_cs *ring,
        /* ELSP is a wo register, so use another nearby reg for posting instead */
        POSTING_READ(RING_EXECLIST_STATUS(ring));
 
-       gen6_gt_force_wake_put(dev_priv, FORCEWAKE_ALL);
+       /* Release Force Wakeup (see the big comment above). */
+       spin_lock_irqsave(&dev_priv->uncore.lock, flags);
+       if (--dev_priv->uncore.forcewake_count == 0)
+               dev_priv->uncore.funcs.force_wake_put(dev_priv, FORCEWAKE_ALL);
+       spin_unlock_irqrestore(&dev_priv->uncore.lock, flags);
 }
 
 static int execlists_ctx_write_tail(struct drm_i915_gem_object *ctx_obj, u32 tail)
@@ -224,6 +253,9 @@ static void execlists_context_unqueue(struct intel_engine_cs *ring)
 {
        struct intel_ctx_submit_request *req0 = NULL, *req1 = NULL;
        struct intel_ctx_submit_request *cursor = NULL, *tmp = NULL;
+       struct drm_i915_private *dev_priv = ring->dev->dev_private;
+
+       assert_spin_locked(&ring->execlist_lock);
 
        if (list_empty(&ring->execlist_queue))
                return;
@@ -237,8 +269,7 @@ static void execlists_context_unqueue(struct intel_engine_cs *ring)
                        /* Same ctx: ignore first request, as second request
                         * will update tail past first request's workload */
                        list_del(&req0->execlist_link);
-                       i915_gem_context_unreference(req0->ctx);
-                       kfree(req0);
+                       queue_work(dev_priv->wq, &req0->work);
                        req0 = cursor;
                } else {
                        req1 = cursor;
@@ -251,11 +282,97 @@ static void execlists_context_unqueue(struct intel_engine_cs *ring)
                                         req1 ? req1->tail : 0));
 }
 
+static bool execlists_check_remove_request(struct intel_engine_cs *ring,
+                                          u32 request_id)
+{
+       struct drm_i915_private *dev_priv = ring->dev->dev_private;
+       struct intel_ctx_submit_request *head_req;
+
+       assert_spin_locked(&ring->execlist_lock);
+
+       head_req = list_first_entry_or_null(&ring->execlist_queue,
+                                           struct intel_ctx_submit_request,
+                                           execlist_link);
+
+       if (head_req != NULL) {
+               struct drm_i915_gem_object *ctx_obj =
+                               head_req->ctx->engine[ring->id].state;
+               if (intel_execlists_ctx_id(ctx_obj) == request_id) {
+                       list_del(&head_req->execlist_link);
+                       queue_work(dev_priv->wq, &head_req->work);
+                       return true;
+               }
+       }
+
+       return false;
+}
+
+void intel_execlists_handle_ctx_events(struct intel_engine_cs *ring)
+{
+       struct drm_i915_private *dev_priv = ring->dev->dev_private;
+       u32 status_pointer;
+       u8 read_pointer;
+       u8 write_pointer;
+       u32 status;
+       u32 status_id;
+       u32 submit_contexts = 0;
+
+       status_pointer = I915_READ(RING_CONTEXT_STATUS_PTR(ring));
+
+       read_pointer = ring->next_context_status_buffer;
+       write_pointer = status_pointer & 0x07;
+       if (read_pointer > write_pointer)
+               write_pointer += 6;
+
+       spin_lock(&ring->execlist_lock);
+
+       while (read_pointer < write_pointer) {
+               read_pointer++;
+               status = I915_READ(RING_CONTEXT_STATUS_BUF(ring) +
+                               (read_pointer % 6) * 8);
+               status_id = I915_READ(RING_CONTEXT_STATUS_BUF(ring) +
+                               (read_pointer % 6) * 8 + 4);
+
+               if (status & GEN8_CTX_STATUS_COMPLETE) {
+                       if (execlists_check_remove_request(ring, status_id))
+                               submit_contexts++;
+               }
+       }
+
+       if (submit_contexts != 0)
+               execlists_context_unqueue(ring);
+
+       spin_unlock(&ring->execlist_lock);
+
+       WARN(submit_contexts > 2, "More than two context complete events?\n");
+       ring->next_context_status_buffer = write_pointer % 6;
+
+       I915_WRITE(RING_CONTEXT_STATUS_PTR(ring),
+                  ((u32)ring->next_context_status_buffer & 0x07) << 8);
+}
+
+static void execlists_free_request_task(struct work_struct *work)
+{
+       struct intel_ctx_submit_request *req =
+               container_of(work, struct intel_ctx_submit_request, work);
+       struct drm_device *dev = req->ring->dev;
+       struct drm_i915_private *dev_priv = dev->dev_private;
+
+       intel_runtime_pm_put(dev_priv);
+
+       mutex_lock(&dev->struct_mutex);
+       i915_gem_context_unreference(req->ctx);
+       mutex_unlock(&dev->struct_mutex);
+
+       kfree(req);
+}
+
 static int execlists_context_queue(struct intel_engine_cs *ring,
                                   struct intel_context *to,
                                   u32 tail)
 {
        struct intel_ctx_submit_request *req = NULL;
+       struct drm_i915_private *dev_priv = ring->dev->dev_private;
        unsigned long flags;
        bool was_empty;
 
@@ -266,6 +383,9 @@ static int execlists_context_queue(struct intel_engine_cs *ring,
        i915_gem_context_reference(req->ctx);
        req->ring = ring;
        req->tail = tail;
+       INIT_WORK(&req->work, execlists_free_request_task);
+
+       intel_runtime_pm_get(dev_priv);
 
        spin_lock_irqsave(&ring->execlist_lock, flags);
 
@@ -907,6 +1027,7 @@ static int logical_ring_init(struct drm_device *dev, struct intel_engine_cs *rin
 
        INIT_LIST_HEAD(&ring->execlist_queue);
        spin_lock_init(&ring->execlist_lock);
+       ring->next_context_status_buffer = 0;
 
        ret = intel_lr_context_deferred_create(dctx, ring);
        if (ret)
index 3c389b3a2b752e1d58d30d41b09574765bff9851..a3f135cf439ea159e5a84da055c68a7646291d8a 100644 (file)
@@ -65,6 +65,9 @@ struct intel_ctx_submit_request {
        u32 tail;
 
        struct list_head execlist_link;
+       struct work_struct work;
 };
 
+void intel_execlists_handle_ctx_events(struct intel_engine_cs *ring);
+
 #endif /* _INTEL_LRC_H_ */
index 670262dabb6cff4b09f313093e80da382c21618b..9cbf7b0ebc994fd6bd383fcb112d9b4413c1932f 100644 (file)
@@ -233,6 +233,7 @@ struct  intel_engine_cs {
        /* Execlists */
        spinlock_t execlist_lock;
        struct list_head execlist_queue;
+       u8 next_context_status_buffer;
        u32             irq_keep_mask; /* bitmask for interrupts that should not be masked */
        int             (*emit_request)(struct intel_ringbuffer *ringbuf);
        int             (*emit_flush)(struct intel_ringbuffer *ringbuf,