drm/i915: CHV DDR DVFS support and another watermark rewrite
authorVille Syrjälä <ville.syrjala@linux.intel.com>
Wed, 24 Jun 2015 19:00:04 +0000 (22:00 +0300)
committerDaniel Vetter <daniel.vetter@ffwll.ch>
Mon, 29 Jun 2015 08:55:13 +0000 (10:55 +0200)
Turns out the VLV/CHV system agent doesn't understand memory
latencies, so trying to rely on the PND deadline mechanism is not
going to fly especially when DDR DVFS is enabled. Currently we try to
avoid the problems by lying to the system agent about the deadlines
and setting the FIFO watermarks to 8 cachelines. This however leads to
bad memory self refresh residency.

So in order to satosfy everyone we'll just give up on the deadline
scheme and program the watermarks old school based on the worst case
memory latency.

I've modelled this a bit on the ILK+ approach where we compute multiple
sets of watermarks for each pipe (PM2,PM5,DDR DVFS) and when merge thet
appropriate one later with the watermarks from other pipes. There isn't
too much to merge actually since each pipe has a totally independent
FIFO (well apart from the mess with the partially shared DSPARB
registers), but still decopuling the pipes from each other seems like a
good idea.

Eventually we'll want to perform the watermark update in two phases
around the plane update to avoid underruns due to the single buffered
watermark registers. But that's still in limbo for ILK+ too, so I've not
gone that far yet for VLV/CHV either.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Clint Taylor <Clinton.A.Taylor@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
drivers/gpu/drm/i915/i915_drv.h
drivers/gpu/drm/i915/intel_display.c
drivers/gpu/drm/i915/intel_drv.h
drivers/gpu/drm/i915/intel_pm.c

index 2009ba5163346882ac83ab0f33d32b7e7ba2a18d..6acc6504863b62bb4680ba9192114de566cec808 100644 (file)
@@ -276,6 +276,12 @@ struct i915_hotplug {
                            &dev->mode_config.plane_list,       \
                            base.head)
 
+#define for_each_intel_plane_on_crtc(dev, intel_crtc, intel_plane)     \
+       list_for_each_entry(intel_plane,                                \
+                           &(dev)->mode_config.plane_list,             \
+                           base.head)                                  \
+               if ((intel_plane)->pipe == (intel_crtc)->pipe)
+
 #define for_each_intel_crtc(dev, intel_crtc) \
        list_for_each_entry(intel_crtc, &dev->mode_config.crtc_list, base.head)
 
@@ -1498,18 +1504,20 @@ struct ilk_wm_values {
        enum intel_ddb_partitioning partitioning;
 };
 
-struct vlv_wm_values {
-       struct {
-               uint16_t primary;
-               uint16_t sprite[2];
-               uint8_t cursor;
-       } pipe[3];
+struct vlv_pipe_wm {
+       uint16_t primary;
+       uint16_t sprite[2];
+       uint8_t cursor;
+};
 
-       struct {
-               uint16_t plane;
-               uint8_t cursor;
-       } sr;
+struct vlv_sr_wm {
+       uint16_t plane;
+       uint8_t cursor;
+};
 
+struct vlv_wm_values {
+       struct vlv_pipe_wm pipe[3];
+       struct vlv_sr_wm sr;
        struct {
                uint8_t cursor;
                uint8_t sprite[2];
index 2295f08ac0b6bb0b6ec37da32c6d98773acc1d56..7baa45db975639acd37bbc933524d4d2f271f1da 100644 (file)
@@ -4699,8 +4699,11 @@ intel_pre_disable_primary(struct drm_crtc *crtc)
         * event which is after the vblank start event, so we need to have a
         * wait-for-vblank between disabling the plane and the pipe.
         */
-       if (HAS_GMCH_DISPLAY(dev))
+       if (HAS_GMCH_DISPLAY(dev)) {
                intel_set_memory_cxsr(dev_priv, false);
+               dev_priv->wm.vlv.cxsr = false;
+               intel_wait_for_vblank(dev, pipe);
+       }
 
        /*
         * FIXME IPS should be fine as long as one plane is
@@ -6017,7 +6020,6 @@ static void valleyview_crtc_enable(struct drm_crtc *crtc)
 
        intel_crtc_load_lut(crtc);
 
-       intel_update_watermarks(crtc);
        intel_enable_pipe(intel_crtc);
 
        assert_vblank_disabled(crtc);
index e7a82dcfda2459e9daacf15a1f2e110c404cdeca..8397d1c9005bfe504155bbc37ccb504865a5bccb 100644 (file)
@@ -462,6 +462,15 @@ struct intel_crtc_state {
        enum pipe hsw_workaround_pipe;
 };
 
+struct vlv_wm_state {
+       struct vlv_pipe_wm wm[3];
+       struct vlv_sr_wm sr[3];
+       uint8_t num_active_planes;
+       uint8_t num_levels;
+       uint8_t level;
+       bool cxsr;
+};
+
 struct intel_pipe_wm {
        struct intel_wm_level wm[5];
        uint32_t linetime;
@@ -565,6 +574,8 @@ struct intel_crtc {
 
        /* scalers available on this crtc */
        int num_scalers;
+
+       struct vlv_wm_state wm_state;
 };
 
 struct intel_plane_wm_parameters {
index 8ea4768dcf10c7a7f889265666a7fc832071244a..23e5be9887a3785a15ae89d180ea93f7104c7216 100644 (file)
@@ -335,8 +335,6 @@ void intel_set_memory_cxsr(struct drm_i915_private *dev_priv, bool enable)
        if (IS_VALLEYVIEW(dev)) {
                I915_WRITE(FW_BLC_SELF_VLV, enable ? FW_CSPWRDWNEN : 0);
                POSTING_READ(FW_BLC_SELF_VLV);
-               if (IS_CHERRYVIEW(dev))
-                       chv_set_memory_pm5(dev_priv, enable);
        } else if (IS_G4X(dev) || IS_CRESTLINE(dev)) {
                I915_WRITE(FW_BLC_SELF, enable ? FW_BLC_SELF_EN : 0);
                POSTING_READ(FW_BLC_SELF);
@@ -929,8 +927,6 @@ static void vlv_write_wm_values(struct intel_crtc *crtc,
        }
 
        POSTING_READ(DSPFW1);
-
-       dev_priv->wm.vlv = *wm;
 }
 
 #undef FW_WM_VLV
@@ -1014,6 +1010,72 @@ enum vlv_wm_level {
        VLV_WM_NUM_LEVELS = 1,
 };
 
+/* latency must be in 0.1us units. */
+static unsigned int vlv_wm_method2(unsigned int pixel_rate,
+                                  unsigned int pipe_htotal,
+                                  unsigned int horiz_pixels,
+                                  unsigned int bytes_per_pixel,
+                                  unsigned int latency)
+{
+       unsigned int ret;
+
+       ret = (latency * pixel_rate) / (pipe_htotal * 10000);
+       ret = (ret + 1) * horiz_pixels * bytes_per_pixel;
+       ret = DIV_ROUND_UP(ret, 64);
+
+       return ret;
+}
+
+static void vlv_setup_wm_latency(struct drm_device *dev)
+{
+       struct drm_i915_private *dev_priv = dev->dev_private;
+
+       /* all latencies in usec */
+       dev_priv->wm.pri_latency[VLV_WM_LEVEL_PM2] = 3;
+
+       if (IS_CHERRYVIEW(dev_priv)) {
+               dev_priv->wm.pri_latency[VLV_WM_LEVEL_PM5] = 12;
+               dev_priv->wm.pri_latency[VLV_WM_LEVEL_DDR_DVFS] = 33;
+       }
+}
+
+static uint16_t vlv_compute_wm_level(struct intel_plane *plane,
+                                    struct intel_crtc *crtc,
+                                    const struct intel_plane_state *state,
+                                    int level)
+{
+       struct drm_i915_private *dev_priv = to_i915(plane->base.dev);
+       int clock, htotal, pixel_size, width, wm;
+
+       if (dev_priv->wm.pri_latency[level] == 0)
+               return USHRT_MAX;
+
+       if (!state->visible)
+               return 0;
+
+       pixel_size = drm_format_plane_cpp(state->base.fb->pixel_format, 0);
+       clock = crtc->config->base.adjusted_mode.crtc_clock;
+       htotal = crtc->config->base.adjusted_mode.crtc_htotal;
+       width = crtc->config->pipe_src_w;
+       if (WARN_ON(htotal == 0))
+               htotal = 1;
+
+       if (plane->base.type == DRM_PLANE_TYPE_CURSOR) {
+               /*
+                * FIXME the formula gives values that are
+                * too big for the cursor FIFO, and hence we
+                * would never be able to use cursors. For
+                * now just hardcode the watermark.
+                */
+               wm = 63;
+       } else {
+               wm = vlv_wm_method2(clock, htotal, width, pixel_size,
+                                   dev_priv->wm.pri_latency[level] * 10);
+       }
+
+       return min_t(int, wm, USHRT_MAX);
+}
+
 static bool vlv_compute_sr_wm(struct drm_device *dev,
                              struct vlv_wm_values *wm)
 {
@@ -1105,6 +1167,249 @@ static void valleyview_update_wm(struct drm_crtc *crtc)
 
        if (cxsr_enabled)
                intel_set_memory_cxsr(dev_priv, true);
+
+       dev_priv->wm.vlv = wm;
+}
+
+static void vlv_invert_wms(struct intel_crtc *crtc)
+{
+       struct vlv_wm_state *wm_state = &crtc->wm_state;
+       int level;
+
+       for (level = 0; level < wm_state->num_levels; level++) {
+               struct drm_device *dev = crtc->base.dev;
+               const int sr_fifo_size = INTEL_INFO(dev)->num_pipes * 512 - 1;
+               struct intel_plane *plane;
+
+               wm_state->sr[level].plane = sr_fifo_size - wm_state->sr[level].plane;
+               wm_state->sr[level].cursor = 63 - wm_state->sr[level].cursor;
+
+               for_each_intel_plane_on_crtc(dev, crtc, plane) {
+                       switch (plane->base.type) {
+                               int sprite;
+                       case DRM_PLANE_TYPE_CURSOR:
+                               wm_state->wm[level].cursor = plane->wm.fifo_size -
+                                       wm_state->wm[level].cursor;
+                               break;
+                       case DRM_PLANE_TYPE_PRIMARY:
+                               wm_state->wm[level].primary = plane->wm.fifo_size -
+                                       wm_state->wm[level].primary;
+                               break;
+                       case DRM_PLANE_TYPE_OVERLAY:
+                               sprite = plane->plane;
+                               wm_state->wm[level].sprite[sprite] = plane->wm.fifo_size -
+                                       wm_state->wm[level].sprite[sprite];
+                               break;
+                       }
+               }
+       }
+}
+
+static void _vlv_compute_wm(struct intel_crtc *crtc)
+{
+       struct drm_device *dev = crtc->base.dev;
+       struct vlv_wm_state *wm_state = &crtc->wm_state;
+       struct intel_plane *plane;
+       int sr_fifo_size = INTEL_INFO(dev)->num_pipes * 512 - 1;
+       int level;
+
+       memset(wm_state, 0, sizeof(*wm_state));
+
+       wm_state->cxsr = crtc->pipe != PIPE_C;
+       if (IS_CHERRYVIEW(dev))
+               wm_state->num_levels = CHV_WM_NUM_LEVELS;
+       else
+               wm_state->num_levels = VLV_WM_NUM_LEVELS;
+
+       wm_state->num_active_planes = 0;
+       for_each_intel_plane_on_crtc(dev, crtc, plane) {
+               struct intel_plane_state *state =
+                       to_intel_plane_state(plane->base.state);
+
+               if (plane->base.type == DRM_PLANE_TYPE_CURSOR)
+                       continue;
+
+               if (state->visible)
+                       wm_state->num_active_planes++;
+       }
+
+       if (wm_state->num_active_planes != 1)
+               wm_state->cxsr = false;
+
+       if (wm_state->cxsr) {
+               for (level = 0; level < wm_state->num_levels; level++) {
+                       wm_state->sr[level].plane = sr_fifo_size;
+                       wm_state->sr[level].cursor = 63;
+               }
+       }
+
+       for_each_intel_plane_on_crtc(dev, crtc, plane) {
+               struct intel_plane_state *state =
+                       to_intel_plane_state(plane->base.state);
+
+               if (!state->visible)
+                       continue;
+
+               /* normal watermarks */
+               for (level = 0; level < wm_state->num_levels; level++) {
+                       int wm = vlv_compute_wm_level(plane, crtc, state, level);
+                       int max_wm = plane->base.type == DRM_PLANE_TYPE_CURSOR ? 63 : 511;
+
+                       /* hack */
+                       if (WARN_ON(level == 0 && wm > max_wm))
+                               wm = max_wm;
+
+                       if (wm > plane->wm.fifo_size)
+                               break;
+
+                       switch (plane->base.type) {
+                               int sprite;
+                       case DRM_PLANE_TYPE_CURSOR:
+                               wm_state->wm[level].cursor = wm;
+                               break;
+                       case DRM_PLANE_TYPE_PRIMARY:
+                               wm_state->wm[level].primary = wm;
+                               break;
+                       case DRM_PLANE_TYPE_OVERLAY:
+                               sprite = plane->plane;
+                               wm_state->wm[level].sprite[sprite] = wm;
+                               break;
+                       }
+               }
+
+               wm_state->num_levels = level;
+
+               if (!wm_state->cxsr)
+                       continue;
+
+               /* maxfifo watermarks */
+               switch (plane->base.type) {
+                       int sprite, level;
+               case DRM_PLANE_TYPE_CURSOR:
+                       for (level = 0; level < wm_state->num_levels; level++)
+                               wm_state->sr[level].cursor =
+                                       wm_state->sr[level].cursor;
+                       break;
+               case DRM_PLANE_TYPE_PRIMARY:
+                       for (level = 0; level < wm_state->num_levels; level++)
+                               wm_state->sr[level].plane =
+                                       min(wm_state->sr[level].plane,
+                                           wm_state->wm[level].primary);
+                       break;
+               case DRM_PLANE_TYPE_OVERLAY:
+                       sprite = plane->plane;
+                       for (level = 0; level < wm_state->num_levels; level++)
+                               wm_state->sr[level].plane =
+                                       min(wm_state->sr[level].plane,
+                                           wm_state->wm[level].sprite[sprite]);
+                       break;
+               }
+       }
+
+       /* clear any (partially) filled invalid levels */
+       for (level = wm_state->num_levels; level < CHV_WM_NUM_LEVELS; level++) {
+               memset(&wm_state->wm[level], 0, sizeof(wm_state->wm[level]));
+               memset(&wm_state->sr[level], 0, sizeof(wm_state->sr[level]));
+       }
+
+       vlv_invert_wms(crtc);
+}
+
+static void vlv_merge_wm(struct drm_device *dev,
+                        struct vlv_wm_values *wm)
+{
+       struct intel_crtc *crtc;
+       int num_active_crtcs = 0;
+
+       if (IS_CHERRYVIEW(dev))
+               wm->level = VLV_WM_LEVEL_DDR_DVFS;
+       else
+               wm->level = VLV_WM_LEVEL_PM2;
+       wm->cxsr = true;
+
+       for_each_intel_crtc(dev, crtc) {
+               const struct vlv_wm_state *wm_state = &crtc->wm_state;
+
+               if (!crtc->active)
+                       continue;
+
+               if (!wm_state->cxsr)
+                       wm->cxsr = false;
+
+               num_active_crtcs++;
+               wm->level = min_t(int, wm->level, wm_state->num_levels - 1);
+       }
+
+       if (num_active_crtcs != 1)
+               wm->cxsr = false;
+
+       for_each_intel_crtc(dev, crtc) {
+               struct vlv_wm_state *wm_state = &crtc->wm_state;
+               enum pipe pipe = crtc->pipe;
+
+               if (!crtc->active)
+                       continue;
+
+               wm->pipe[pipe] = wm_state->wm[wm->level];
+               if (wm->cxsr)
+                       wm->sr = wm_state->sr[wm->level];
+
+               wm->ddl[pipe].primary = DDL_PRECISION_HIGH | 2;
+               wm->ddl[pipe].sprite[0] = DDL_PRECISION_HIGH | 2;
+               wm->ddl[pipe].sprite[1] = DDL_PRECISION_HIGH | 2;
+               wm->ddl[pipe].cursor = DDL_PRECISION_HIGH | 2;
+       }
+}
+
+static void vlv_update_wm(struct drm_crtc *crtc)
+{
+       struct drm_device *dev = crtc->dev;
+       struct drm_i915_private *dev_priv = dev->dev_private;
+       struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
+       enum pipe pipe = intel_crtc->pipe;
+       struct vlv_wm_values wm = {};
+
+       _vlv_compute_wm(intel_crtc);
+       vlv_merge_wm(dev, &wm);
+
+       if (memcmp(&dev_priv->wm.vlv, &wm, sizeof(wm)) == 0)
+               return;
+
+       if (wm.level < VLV_WM_LEVEL_DDR_DVFS &&
+           dev_priv->wm.vlv.level >= VLV_WM_LEVEL_DDR_DVFS)
+               chv_set_memory_dvfs(dev_priv, false);
+
+       if (wm.level < VLV_WM_LEVEL_PM5 &&
+           dev_priv->wm.vlv.level >= VLV_WM_LEVEL_PM5)
+               chv_set_memory_pm5(dev_priv, false);
+
+       if (!wm.cxsr && dev_priv->wm.vlv.cxsr) {
+               intel_set_memory_cxsr(dev_priv, false);
+               intel_wait_for_vblank(dev, pipe);
+       }
+
+       vlv_write_wm_values(intel_crtc, &wm);
+
+       DRM_DEBUG_KMS("Setting FIFO watermarks - %c: plane=%d, cursor=%d, "
+                     "sprite0=%d, sprite1=%d, SR: plane=%d, cursor=%d level=%d cxsr=%d\n",
+                     pipe_name(pipe), wm.pipe[pipe].primary, wm.pipe[pipe].cursor,
+                     wm.pipe[pipe].sprite[0], wm.pipe[pipe].sprite[1],
+                     wm.sr.plane, wm.sr.cursor, wm.level, wm.cxsr);
+
+       if (wm.cxsr && !dev_priv->wm.vlv.cxsr) {
+               intel_wait_for_vblank(dev, pipe);
+               intel_set_memory_cxsr(dev_priv, true);
+       }
+
+       if (wm.level >= VLV_WM_LEVEL_PM5 &&
+           dev_priv->wm.vlv.level < VLV_WM_LEVEL_PM5)
+               chv_set_memory_pm5(dev_priv, true);
+
+       if (wm.level >= VLV_WM_LEVEL_DDR_DVFS &&
+           dev_priv->wm.vlv.level < VLV_WM_LEVEL_DDR_DVFS)
+               chv_set_memory_dvfs(dev_priv, true);
+
+       dev_priv->wm.vlv = wm;
 }
 
 static void valleyview_update_sprite_wm(struct drm_plane *plane,
@@ -6831,8 +7136,9 @@ void intel_init_pm(struct drm_device *dev)
                else if (INTEL_INFO(dev)->gen == 8)
                        dev_priv->display.init_clock_gating = broadwell_init_clock_gating;
        } else if (IS_CHERRYVIEW(dev)) {
-               dev_priv->display.update_wm = valleyview_update_wm;
-               dev_priv->display.update_sprite_wm = valleyview_update_sprite_wm;
+               vlv_setup_wm_latency(dev);
+
+               dev_priv->display.update_wm = vlv_update_wm;
                dev_priv->display.init_clock_gating =
                        cherryview_init_clock_gating;
        } else if (IS_VALLEYVIEW(dev)) {