drm/i915: Replaced Blitter ring based flips with MMIO flips
authorSourab Gupta <sourab.gupta@intel.com>
Mon, 2 Jun 2014 11:17:17 +0000 (16:47 +0530)
committerDaniel Vetter <daniel.vetter@ffwll.ch>
Tue, 17 Jun 2014 14:16:20 +0000 (16:16 +0200)
This patch enables the framework for using MMIO based flip calls,
in contrast with the CS based flip calls which are being used currently.

MMIO based flip calls can be enabled on architectures where
Render and Blitter engines reside in different power wells. The
decision to use MMIO flips can be made based on workloads to give
100% residency for Media power well.

v2: The MMIO flips now use the interrupt driven mechanism for issuing the
flips when target seqno is reached. (Incorporating Ville's idea)

v3: Rebasing on latest code. Code restructuring after incorporating
Damien's comments

v4: Addressing Ville's review comments
    -general cleanup
    -updating only base addr instead of calling update_primary_plane
    -extending patch for gen5+ platforms

v5: Addressed Ville's review comments
    -Making mmio flip vs cs flip selection based on module parameter
    -Adding check for DRIVER_MODESET feature in notify_ring before calling
     notify mmio flip.
    -Other changes mostly in function arguments

v6: -Having a seperate function to check condition for using mmio flips (Ville)
    -propogating error code from i915_gem_check_olr (Ville)

v7: -Adding __must_check with i915_gem_check_olr (Chris)
    -Renaming mmio_flip_data to mmio_flip (Chris)
    -Rebasing on latest nightly

v8: -Rebasing on latest code
    -squash 3rd patch in series(mmio setbase vs page flip race) with this patch
    -Added new tiling mode update in intel_do_mmio_flip (Chris)

v9: -check for obj->last_write_seqno being 0 instead of obj->ring being NULL in
intel_postpone_flip, as this is a more restrictive condition (Chris)

v10: -Applied Chris's suggestions for squashing patches 2,3 into this patch.
These patches make the selection of CS vs MMIO flip at the page flip time, and
make the module parameter for using mmio flips as tristate, the states being
'force CS flips', 'force mmio flips', 'driver discretion'.
Changed the logic for driver discretion (Chris)

v11: Minor code cleanup(better readability, fixing whitespace errors, using
lockdep to check mutex locked status in postpone_flip, removal of __must_check
in function definition) (Chris)

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Sourab Gupta <sourab.gupta@intel.com>
Signed-off-by: Akash Goel <akash.goel@intel.com>
Tested-by: Chris Wilson <chris@chris-wilson.co.uk> # snb, ivb
[danvet: Fix up parameter alignement checkpatch spotted.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
drivers/gpu/drm/i915/i915_dma.c
drivers/gpu/drm/i915/i915_drv.h
drivers/gpu/drm/i915/i915_gem.c
drivers/gpu/drm/i915/i915_irq.c
drivers/gpu/drm/i915/i915_params.c
drivers/gpu/drm/i915/intel_display.c
drivers/gpu/drm/i915/intel_drv.h

index 0f98957b5246fa01645f0ee700f6fea3568da2b7..10009fc6b12f628322f625d9558488731962aee6 100644 (file)
@@ -1575,6 +1575,7 @@ int i915_driver_load(struct drm_device *dev, unsigned long flags)
        spin_lock_init(&dev_priv->backlight_lock);
        spin_lock_init(&dev_priv->uncore.lock);
        spin_lock_init(&dev_priv->mm.object_stat_lock);
+       spin_lock_init(&dev_priv->mmio_flip_lock);
        mutex_init(&dev_priv->dpio_lock);
        mutex_init(&dev_priv->modeset_restore_lock);
 
index c5b02159898bc706dc1742748210722d40e2b1ab..9bfc242af87fcd583a7246a23f6f1c393c91def6 100644 (file)
@@ -1370,6 +1370,9 @@ struct drm_i915_private {
        /* protects the irq masks */
        spinlock_t irq_lock;
 
+       /* protects the mmio flip data */
+       spinlock_t mmio_flip_lock;
+
        bool display_irqs_enabled;
 
        /* To control wakeup latency, e.g. for irq-driven dp aux transfers. */
@@ -2043,6 +2046,7 @@ struct i915_params {
        bool reset;
        bool disable_display;
        bool disable_vtd_wa;
+       int use_mmio_flip;
 };
 extern struct i915_params i915 __read_mostly;
 
@@ -2239,6 +2243,8 @@ bool i915_gem_retire_requests(struct drm_device *dev);
 void i915_gem_retire_requests_ring(struct intel_engine_cs *ring);
 int __must_check i915_gem_check_wedge(struct i915_gpu_error *error,
                                      bool interruptible);
+int __must_check i915_gem_check_olr(struct intel_engine_cs *ring, u32 seqno);
+
 static inline bool i915_reset_in_progress(struct i915_gpu_error *error)
 {
        return unlikely(atomic_read(&error->reset_counter)
@@ -2608,6 +2614,8 @@ int i915_reg_read_ioctl(struct drm_device *dev, void *data,
 int i915_get_reset_stats_ioctl(struct drm_device *dev, void *data,
                               struct drm_file *file);
 
+void intel_notify_mmio_flip(struct intel_engine_cs *ring);
+
 /* overlay */
 extern struct intel_overlay_error_state *intel_overlay_capture_error_state(struct drm_device *dev);
 extern void intel_overlay_print_error_state(struct drm_i915_error_state_buf *e,
index 234aa1ca2eb8c2b9a4a7496aaa3a60343ed3f353..7f643db268295048f5efd480b89ef708891f84eb 100644 (file)
@@ -1095,7 +1095,7 @@ i915_gem_check_wedge(struct i915_gpu_error *error,
  * Compare seqno against outstanding lazy request. Emit a request if they are
  * equal.
  */
-static int
+int
 i915_gem_check_olr(struct intel_engine_cs *ring, u32 seqno)
 {
        int ret;
index 218ef085cf7339f24197b73024901fa3e109e351..c7d4cb7600e57b85d33f12b019db4818a31b0ce7 100644 (file)
@@ -1218,6 +1218,9 @@ static void notify_ring(struct drm_device *dev,
 
        trace_i915_gem_request_complete(ring);
 
+       if (drm_core_check_feature(dev, DRIVER_MODESET))
+               intel_notify_mmio_flip(ring);
+
        wake_up_all(&ring->irq_queue);
        i915_queue_hangcheck(dev);
 }
index d05a2afa17dc605acc294e0e15ae88ac200b5093..81457293cd3efb06201aa7cccfea05526d8d51b4 100644 (file)
@@ -48,6 +48,7 @@ struct i915_params i915 __read_mostly = {
        .disable_display = 0,
        .enable_cmd_parser = 1,
        .disable_vtd_wa = 0,
+       .use_mmio_flip = 0,
 };
 
 module_param_named(modeset, i915.modeset, int, 0400);
@@ -156,3 +157,7 @@ MODULE_PARM_DESC(disable_vtd_wa, "Disable all VT-d workarounds (default: false)"
 module_param_named(enable_cmd_parser, i915.enable_cmd_parser, int, 0600);
 MODULE_PARM_DESC(enable_cmd_parser,
                 "Enable command parsing (1=enabled [default], 0=disabled)");
+
+module_param_named(use_mmio_flip, i915.use_mmio_flip, int, 0600);
+MODULE_PARM_DESC(use_mmio_flip,
+                "use MMIO flips (-1=never, 0=driver discretion [default], 1=always)");
index a514779a5918b0f7dcc0d1561fbcaaa75d3ed5d6..5e8e7113b453511a839c4715f82dcdd32be6523c 100644 (file)
@@ -9248,6 +9248,147 @@ static int intel_gen7_queue_flip(struct drm_device *dev,
        return 0;
 }
 
+static bool use_mmio_flip(struct intel_engine_cs *ring,
+                         struct drm_i915_gem_object *obj)
+{
+       /*
+        * This is not being used for older platforms, because
+        * non-availability of flip done interrupt forces us to use
+        * CS flips. Older platforms derive flip done using some clever
+        * tricks involving the flip_pending status bits and vblank irqs.
+        * So using MMIO flips there would disrupt this mechanism.
+        */
+
+       if (INTEL_INFO(ring->dev)->gen < 5)
+               return false;
+
+       if (i915.use_mmio_flip < 0)
+               return false;
+       else if (i915.use_mmio_flip > 0)
+               return true;
+       else
+               return ring != obj->ring;
+}
+
+static void intel_do_mmio_flip(struct intel_crtc *intel_crtc)
+{
+       struct drm_device *dev = intel_crtc->base.dev;
+       struct drm_i915_private *dev_priv = dev->dev_private;
+       struct intel_framebuffer *intel_fb =
+               to_intel_framebuffer(intel_crtc->base.primary->fb);
+       struct drm_i915_gem_object *obj = intel_fb->obj;
+       u32 dspcntr;
+       u32 reg;
+
+       intel_mark_page_flip_active(intel_crtc);
+
+       reg = DSPCNTR(intel_crtc->plane);
+       dspcntr = I915_READ(reg);
+
+       if (INTEL_INFO(dev)->gen >= 4) {
+               if (obj->tiling_mode != I915_TILING_NONE)
+                       dspcntr |= DISPPLANE_TILED;
+               else
+                       dspcntr &= ~DISPPLANE_TILED;
+       }
+       I915_WRITE(reg, dspcntr);
+
+       I915_WRITE(DSPSURF(intel_crtc->plane),
+                  intel_crtc->unpin_work->gtt_offset);
+       POSTING_READ(DSPSURF(intel_crtc->plane));
+}
+
+static int intel_postpone_flip(struct drm_i915_gem_object *obj)
+{
+       struct intel_engine_cs *ring;
+       int ret;
+
+       lockdep_assert_held(&obj->base.dev->struct_mutex);
+
+       if (!obj->last_write_seqno)
+               return 0;
+
+       ring = obj->ring;
+
+       if (i915_seqno_passed(ring->get_seqno(ring, true),
+                             obj->last_write_seqno))
+               return 0;
+
+       ret = i915_gem_check_olr(ring, obj->last_write_seqno);
+       if (ret)
+               return ret;
+
+       if (WARN_ON(!ring->irq_get(ring)))
+               return 0;
+
+       return 1;
+}
+
+void intel_notify_mmio_flip(struct intel_engine_cs *ring)
+{
+       struct drm_i915_private *dev_priv = to_i915(ring->dev);
+       struct intel_crtc *intel_crtc;
+       unsigned long irq_flags;
+       u32 seqno;
+
+       seqno = ring->get_seqno(ring, false);
+
+       spin_lock_irqsave(&dev_priv->mmio_flip_lock, irq_flags);
+       for_each_intel_crtc(ring->dev, intel_crtc) {
+               struct intel_mmio_flip *mmio_flip;
+
+               mmio_flip = &intel_crtc->mmio_flip;
+               if (mmio_flip->seqno == 0)
+                       continue;
+
+               if (ring->id != mmio_flip->ring_id)
+                       continue;
+
+               if (i915_seqno_passed(seqno, mmio_flip->seqno)) {
+                       intel_do_mmio_flip(intel_crtc);
+                       mmio_flip->seqno = 0;
+                       ring->irq_put(ring);
+               }
+       }
+       spin_unlock_irqrestore(&dev_priv->mmio_flip_lock, irq_flags);
+}
+
+static int intel_queue_mmio_flip(struct drm_device *dev,
+                                struct drm_crtc *crtc,
+                                struct drm_framebuffer *fb,
+                                struct drm_i915_gem_object *obj,
+                                struct intel_engine_cs *ring,
+                                uint32_t flags)
+{
+       struct drm_i915_private *dev_priv = dev->dev_private;
+       struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
+       unsigned long irq_flags;
+       int ret;
+
+       if (WARN_ON(intel_crtc->mmio_flip.seqno))
+               return -EBUSY;
+
+       ret = intel_postpone_flip(obj);
+       if (ret < 0)
+               return ret;
+       if (ret == 0) {
+               intel_do_mmio_flip(intel_crtc);
+               return 0;
+       }
+
+       spin_lock_irqsave(&dev_priv->mmio_flip_lock, irq_flags);
+       intel_crtc->mmio_flip.seqno = obj->last_write_seqno;
+       intel_crtc->mmio_flip.ring_id = obj->ring->id;
+       spin_unlock_irqrestore(&dev_priv->mmio_flip_lock, irq_flags);
+
+       /*
+        * Double check to catch cases where irq fired before
+        * mmio flip data was ready
+        */
+       intel_notify_mmio_flip(obj->ring);
+       return 0;
+}
+
 static int intel_default_queue_flip(struct drm_device *dev,
                                    struct drm_crtc *crtc,
                                    struct drm_framebuffer *fb,
@@ -9358,7 +9499,12 @@ static int intel_crtc_page_flip(struct drm_crtc *crtc,
        work->gtt_offset =
                i915_gem_obj_ggtt_offset(obj) + intel_crtc->dspaddr_offset;
 
-       ret = dev_priv->display.queue_flip(dev, crtc, fb, obj, ring, page_flip_flags);
+       if (use_mmio_flip(ring, obj))
+               ret = intel_queue_mmio_flip(dev, crtc, fb, obj, ring,
+                                           page_flip_flags);
+       else
+               ret = dev_priv->display.queue_flip(dev, crtc, fb, obj, ring,
+                               page_flip_flags);
        if (ret)
                goto cleanup_unpin;
 
index 9aa95f01b1604390be3225dedcf6e656401c6aec..cac7539320327d584e1396de67b7c7a8e1732cfc 100644 (file)
@@ -358,6 +358,11 @@ struct intel_pipe_wm {
        bool sprites_scaled;
 };
 
+struct intel_mmio_flip {
+       u32 seqno;
+       u32 ring_id;
+};
+
 struct intel_crtc {
        struct drm_crtc base;
        enum pipe pipe;
@@ -411,6 +416,7 @@ struct intel_crtc {
        wait_queue_head_t vbl_wait;
 
        int scanline_offset;
+       struct intel_mmio_flip mmio_flip;
 };
 
 struct intel_plane_wm_parameters {