drm/i915: Correctly map DBUF slices to pipes
authorStanislav Lisovskiy <stanislav.lisovskiy@intel.com>
Sun, 2 Feb 2020 23:06:30 +0000 (01:06 +0200)
committerVille Syrjälä <ville.syrjala@linux.intel.com>
Wed, 5 Feb 2020 17:22:11 +0000 (19:22 +0200)
Added proper DBuf slice mapping to correspondent
pipes, depending on pipe configuration as stated
in BSpec.

v2:
    - Remove unneeded braces
    - Stop using macro for DBuf assignments as
      it seems to reduce readability.

v3: Start using enabled slices mask in dev_priv

v4: Renamed "enabled_slices" used in dev_priv
    to "enabled_dbuf_slices_mask"(Matt Roper)

v5: - Removed redundant parameters from
      intel_get_ddb_size function.(Matt Roper)
    - Made i915_possible_dbuf_slices static(Matt Roper)
    - Renamed total_width into total_width_in_range
      so that it now reflects that this is not
      a total pipe width but the one in current
      dbuf slice allowed range for pipe.(Matt Roper)
    - Removed 4th pipe for ICL in DBuf assignment
      table(Matt Roper)
    - Fixed wrong DBuf slice in DBuf table for TGL
      (Matt Roper)
    - Added comment regarding why we currently not
      using pipe ratio for DBuf assignment for ICL

v6: - Changed u32 to unsigned int in
      icl_get_first_dbuf_slice_offset function signature
      (Ville Syrjälä)
    - Changed also u32 to u8 in dbuf slice mask structure
      (Ville Syrjälä)
    - Switched from DBUF_S1_BIT to enum + explicit
      BIT(DBUF_S1) access(Ville Syrjälä)
    - Switched to named initializers in DBuf assignment
      arrays(Ville Syrjälä)
    - DBuf assignment arrays now use autogeneration tool
      from
      https://patchwork.freedesktop.org/series/70493/
      to avoid typos.
    - Renamed i915_find_pipe_conf to *_compute_dbuf_slices
      (Ville Syrjälä)
    - Changed platforms ordering in skl_compute_dbuf_slices
      to be from newest to oldest(Ville Syrjälä)

v7: - Now ORing assigned DBuf slice config always with DBUF_S1
      because slice 1 has to be constantly powered on.
      (Ville Syrjälä)

v8: - Added pipe_name for neater printing(Ville Syrjälä)
    - Renamed width_before_pipe to width_before_pipe_in_range,
      to better reflect that now all the calculations are happening
      inside DBuf range allowed by current pipe configuration mask
      (Ville Syrjälä)
    - Shortened FIXME comment message, regarding constant ORing with
      DBUF_S1(Ville Syrjälä)
    - Added .dbuf_mask named initializer to pipe assignment array
      (Ville Syrjälä)
    - Edited pipe assignment array to use only single DBuf slice
      for gen11 single pipe configurations, until "pipe ratio"
      thing is finally sorted out(Ville Syrjälä)
    - Removed unused parameter crtc_state for now(Ville Syrjälä)
      from icl/tgl_compute_dbuf_slices function

Reviewed-by: Matt Roper <matthew.d.roper@intel.com>
Signed-off-by: Stanislav Lisovskiy <stanislav.lisovskiy@intel.com>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200202230630.8975-7-stanislav.lisovskiy@intel.com
drivers/gpu/drm/i915/intel_pm.c

index 4c87abc..ffac0b8 100644 (file)
@@ -3809,13 +3809,29 @@ bool intel_can_enable_sagv(struct intel_atomic_state *state)
        return true;
 }
 
-static u16 intel_get_ddb_size(struct drm_i915_private *dev_priv,
-                             const struct intel_crtc_state *crtc_state,
-                             const u64 total_data_rate,
-                             const int num_active)
+/*
+ * Calculate initial DBuf slice offset, based on slice size
+ * and mask(i.e if slice size is 1024 and second slice is enabled
+ * offset would be 1024)
+ */
+static unsigned int
+icl_get_first_dbuf_slice_offset(u32 dbuf_slice_mask,
+                               u32 slice_size,
+                               u32 ddb_size)
+{
+       unsigned int offset = 0;
+
+       if (!dbuf_slice_mask)
+               return 0;
+
+       offset = (ffs(dbuf_slice_mask) - 1) * slice_size;
+
+       WARN_ON(offset >= ddb_size);
+       return offset;
+}
+
+static u16 intel_get_ddb_size(struct drm_i915_private *dev_priv)
 {
-       struct drm_atomic_state *state = crtc_state->uapi.state;
-       struct intel_atomic_state *intel_state = to_intel_atomic_state(state);
        u16 ddb_size = INTEL_INFO(dev_priv)->ddb_size;
 
        drm_WARN_ON(&dev_priv->drm, ddb_size == 0);
@@ -3823,12 +3839,12 @@ static u16 intel_get_ddb_size(struct drm_i915_private *dev_priv,
        if (INTEL_GEN(dev_priv) < 11)
                return ddb_size - 4; /* 4 blocks for bypass path allocation */
 
-       intel_state->enabled_dbuf_slices_mask = BIT(DBUF_S1);
-       ddb_size /= 2;
-
        return ddb_size;
 }
 
+static u8 skl_compute_dbuf_slices(const struct intel_crtc_state *crtc_state,
+                                 u32 active_pipes);
+
 static void
 skl_ddb_get_pipe_allocation_limits(struct drm_i915_private *dev_priv,
                                   const struct intel_crtc_state *crtc_state,
@@ -3840,10 +3856,17 @@ skl_ddb_get_pipe_allocation_limits(struct drm_i915_private *dev_priv,
        struct intel_atomic_state *intel_state = to_intel_atomic_state(state);
        struct drm_crtc *for_crtc = crtc_state->uapi.crtc;
        const struct intel_crtc *crtc;
-       u32 pipe_width = 0, total_width = 0, width_before_pipe = 0;
+       u32 pipe_width = 0, total_width_in_range = 0, width_before_pipe_in_range = 0;
        enum pipe for_pipe = to_intel_crtc(for_crtc)->pipe;
        u16 ddb_size;
+       u32 ddb_range_size;
        u32 i;
+       u32 dbuf_slice_mask;
+       u32 active_pipes;
+       u32 offset;
+       u32 slice_size;
+       u32 total_slice_mask;
+       u32 start, end;
 
        if (drm_WARN_ON(&dev_priv->drm, !state) || !crtc_state->hw.active) {
                alloc->start = 0;
@@ -3853,12 +3876,15 @@ skl_ddb_get_pipe_allocation_limits(struct drm_i915_private *dev_priv,
        }
 
        if (intel_state->active_pipe_changes)
-               *num_active = hweight8(intel_state->active_pipes);
+               active_pipes = intel_state->active_pipes;
        else
-               *num_active = hweight8(dev_priv->active_pipes);
+               active_pipes = dev_priv->active_pipes;
+
+       *num_active = hweight8(active_pipes);
+
+       ddb_size = intel_get_ddb_size(dev_priv);
 
-       ddb_size = intel_get_ddb_size(dev_priv, crtc_state, total_data_rate,
-                                     *num_active);
+       slice_size = ddb_size / INTEL_INFO(dev_priv)->num_supported_dbuf_slices;
 
        /*
         * If the state doesn't change the active CRTC's or there is no
@@ -3878,30 +3904,95 @@ skl_ddb_get_pipe_allocation_limits(struct drm_i915_private *dev_priv,
        }
 
        /*
+        * Get allowed DBuf slices for correspondent pipe and platform.
+        */
+       dbuf_slice_mask = skl_compute_dbuf_slices(crtc_state, active_pipes);
+
+       DRM_DEBUG_KMS("DBuf slice mask %x pipe %c active pipes %x\n",
+                     dbuf_slice_mask,
+                     pipe_name(for_pipe), active_pipes);
+
+       /*
+        * Figure out at which DBuf slice we start, i.e if we start at Dbuf S2
+        * and slice size is 1024, the offset would be 1024
+        */
+       offset = icl_get_first_dbuf_slice_offset(dbuf_slice_mask,
+                                                slice_size, ddb_size);
+
+       /*
+        * Figure out total size of allowed DBuf slices, which is basically
+        * a number of allowed slices for that pipe multiplied by slice size.
+        * Inside of this
+        * range ddb entries are still allocated in proportion to display width.
+        */
+       ddb_range_size = hweight8(dbuf_slice_mask) * slice_size;
+
+       /*
         * Watermark/ddb requirement highly depends upon width of the
         * framebuffer, So instead of allocating DDB equally among pipes
         * distribute DDB based on resolution/width of the display.
         */
+       total_slice_mask = dbuf_slice_mask;
        for_each_new_intel_crtc_in_state(intel_state, crtc, crtc_state, i) {
                const struct drm_display_mode *adjusted_mode =
                        &crtc_state->hw.adjusted_mode;
                enum pipe pipe = crtc->pipe;
                int hdisplay, vdisplay;
+               u32 pipe_dbuf_slice_mask;
 
-               if (!crtc_state->hw.enable)
+               if (!crtc_state->hw.active)
+                       continue;
+
+               pipe_dbuf_slice_mask = skl_compute_dbuf_slices(crtc_state,
+                                                              active_pipes);
+
+               /*
+                * According to BSpec pipe can share one dbuf slice with another
+                * pipes or pipe can use multiple dbufs, in both cases we
+                * account for other pipes only if they have exactly same mask.
+                * However we need to account how many slices we should enable
+                * in total.
+                */
+               total_slice_mask |= pipe_dbuf_slice_mask;
+
+               /*
+                * Do not account pipes using other slice sets
+                * luckily as of current BSpec slice sets do not partially
+                * intersect(pipes share either same one slice or same slice set
+                * i.e no partial intersection), so it is enough to check for
+                * equality for now.
+                */
+               if (dbuf_slice_mask != pipe_dbuf_slice_mask)
                        continue;
 
                drm_mode_get_hv_timing(adjusted_mode, &hdisplay, &vdisplay);
-               total_width += hdisplay;
+
+               total_width_in_range += hdisplay;
 
                if (pipe < for_pipe)
-                       width_before_pipe += hdisplay;
+                       width_before_pipe_in_range += hdisplay;
                else if (pipe == for_pipe)
                        pipe_width = hdisplay;
        }
 
-       alloc->start = ddb_size * width_before_pipe / total_width;
-       alloc->end = ddb_size * (width_before_pipe + pipe_width) / total_width;
+       /*
+        * FIXME: For now we always enable slice S1 as per
+        * the Bspec display initialization sequence.
+        */
+       intel_state->enabled_dbuf_slices_mask = total_slice_mask | BIT(DBUF_S1);
+
+       start = ddb_range_size * width_before_pipe_in_range / total_width_in_range;
+       end = ddb_range_size *
+               (width_before_pipe_in_range + pipe_width) / total_width_in_range;
+
+       alloc->start = offset + start;
+       alloc->end = offset + end;
+
+       DRM_DEBUG_KMS("Pipe %d ddb %d-%d\n", for_pipe,
+                     alloc->start, alloc->end);
+       DRM_DEBUG_KMS("Enabled ddb slices mask %x num supported %d\n",
+                     intel_state->enabled_dbuf_slices_mask,
+                     INTEL_INFO(dev_priv)->num_supported_dbuf_slices);
 }
 
 static int skl_compute_wm_params(const struct intel_crtc_state *crtc_state,
@@ -4072,6 +4163,262 @@ skl_plane_downscale_amount(const struct intel_crtc_state *crtc_state,
        return mul_fixed16(downscale_w, downscale_h);
 }
 
+struct dbuf_slice_conf_entry {
+       u8 active_pipes;
+       u8 dbuf_mask[I915_MAX_PIPES];
+};
+
+/*
+ * Table taken from Bspec 12716
+ * Pipes do have some preferred DBuf slice affinity,
+ * plus there are some hardcoded requirements on how
+ * those should be distributed for multipipe scenarios.
+ * For more DBuf slices algorithm can get even more messy
+ * and less readable, so decided to use a table almost
+ * as is from BSpec itself - that way it is at least easier
+ * to compare, change and check.
+ */
+static struct dbuf_slice_conf_entry icl_allowed_dbufs[] =
+/* Autogenerated with igt/tools/intel_dbuf_map tool: */
+{
+       {
+               .active_pipes = BIT(PIPE_A),
+               .dbuf_mask = {
+                       [PIPE_A] = BIT(DBUF_S1)
+               }
+       },
+       {
+               .active_pipes = BIT(PIPE_B),
+               .dbuf_mask = {
+                       [PIPE_B] = BIT(DBUF_S1)
+               }
+       },
+       {
+               .active_pipes = BIT(PIPE_A) | BIT(PIPE_B),
+               .dbuf_mask = {
+                       [PIPE_A] = BIT(DBUF_S1),
+                       [PIPE_B] = BIT(DBUF_S2)
+               }
+       },
+       {
+               .active_pipes = BIT(PIPE_C),
+               .dbuf_mask = {
+                       [PIPE_C] = BIT(DBUF_S2)
+               }
+       },
+       {
+               .active_pipes = BIT(PIPE_A) | BIT(PIPE_C),
+               .dbuf_mask = {
+                       [PIPE_A] = BIT(DBUF_S1),
+                       [PIPE_C] = BIT(DBUF_S2)
+               }
+       },
+       {
+               .active_pipes = BIT(PIPE_B) | BIT(PIPE_C),
+               .dbuf_mask = {
+                       [PIPE_B] = BIT(DBUF_S1),
+                       [PIPE_C] = BIT(DBUF_S2)
+               }
+       },
+       {
+               .active_pipes = BIT(PIPE_A) | BIT(PIPE_B) | BIT(PIPE_C),
+               .dbuf_mask = {
+                       [PIPE_A] = BIT(DBUF_S1),
+                       [PIPE_B] = BIT(DBUF_S1),
+                       [PIPE_C] = BIT(DBUF_S2)
+               }
+       },
+};
+
+/*
+ * Table taken from Bspec 49255
+ * Pipes do have some preferred DBuf slice affinity,
+ * plus there are some hardcoded requirements on how
+ * those should be distributed for multipipe scenarios.
+ * For more DBuf slices algorithm can get even more messy
+ * and less readable, so decided to use a table almost
+ * as is from BSpec itself - that way it is at least easier
+ * to compare, change and check.
+ */
+static struct dbuf_slice_conf_entry tgl_allowed_dbufs[] =
+/* Autogenerated with igt/tools/intel_dbuf_map tool: */
+{
+       {
+               .active_pipes = BIT(PIPE_A),
+               .dbuf_mask = {
+                       [PIPE_A] = BIT(DBUF_S1) | BIT(DBUF_S2)
+               }
+       },
+       {
+               .active_pipes = BIT(PIPE_B),
+               .dbuf_mask = {
+                       [PIPE_B] = BIT(DBUF_S1) | BIT(DBUF_S2)
+               }
+       },
+       {
+               .active_pipes = BIT(PIPE_A) | BIT(PIPE_B),
+               .dbuf_mask = {
+                       [PIPE_A] = BIT(DBUF_S2),
+                       [PIPE_B] = BIT(DBUF_S1)
+               }
+       },
+       {
+               .active_pipes = BIT(PIPE_C),
+               .dbuf_mask = {
+                       [PIPE_C] = BIT(DBUF_S2) | BIT(DBUF_S1)
+               }
+       },
+       {
+               .active_pipes = BIT(PIPE_A) | BIT(PIPE_C),
+               .dbuf_mask = {
+                       [PIPE_A] = BIT(DBUF_S1),
+                       [PIPE_C] = BIT(DBUF_S2)
+               }
+       },
+       {
+               .active_pipes = BIT(PIPE_B) | BIT(PIPE_C),
+               .dbuf_mask = {
+                       [PIPE_B] = BIT(DBUF_S1),
+                       [PIPE_C] = BIT(DBUF_S2)
+               }
+       },
+       {
+               .active_pipes = BIT(PIPE_A) | BIT(PIPE_B) | BIT(PIPE_C),
+               .dbuf_mask = {
+                       [PIPE_A] = BIT(DBUF_S1),
+                       [PIPE_B] = BIT(DBUF_S1),
+                       [PIPE_C] = BIT(DBUF_S2)
+               }
+       },
+       {
+               .active_pipes = BIT(PIPE_D),
+               .dbuf_mask = {
+                       [PIPE_D] = BIT(DBUF_S2) | BIT(DBUF_S1)
+               }
+       },
+       {
+               .active_pipes = BIT(PIPE_A) | BIT(PIPE_D),
+               .dbuf_mask = {
+                       [PIPE_A] = BIT(DBUF_S1),
+                       [PIPE_D] = BIT(DBUF_S2)
+               }
+       },
+       {
+               .active_pipes = BIT(PIPE_B) | BIT(PIPE_D),
+               .dbuf_mask = {
+                       [PIPE_B] = BIT(DBUF_S1),
+                       [PIPE_D] = BIT(DBUF_S2)
+               }
+       },
+       {
+               .active_pipes = BIT(PIPE_A) | BIT(PIPE_B) | BIT(PIPE_D),
+               .dbuf_mask = {
+                       [PIPE_A] = BIT(DBUF_S1),
+                       [PIPE_B] = BIT(DBUF_S1),
+                       [PIPE_D] = BIT(DBUF_S2)
+               }
+       },
+       {
+               .active_pipes = BIT(PIPE_C) | BIT(PIPE_D),
+               .dbuf_mask = {
+                       [PIPE_C] = BIT(DBUF_S1),
+                       [PIPE_D] = BIT(DBUF_S2)
+               }
+       },
+       {
+               .active_pipes = BIT(PIPE_A) | BIT(PIPE_C) | BIT(PIPE_D),
+               .dbuf_mask = {
+                       [PIPE_A] = BIT(DBUF_S1),
+                       [PIPE_C] = BIT(DBUF_S2),
+                       [PIPE_D] = BIT(DBUF_S2)
+               }
+       },
+       {
+               .active_pipes = BIT(PIPE_B) | BIT(PIPE_C) | BIT(PIPE_D),
+               .dbuf_mask = {
+                       [PIPE_B] = BIT(DBUF_S1),
+                       [PIPE_C] = BIT(DBUF_S2),
+                       [PIPE_D] = BIT(DBUF_S2)
+               }
+       },
+       {
+               .active_pipes = BIT(PIPE_A) | BIT(PIPE_B) | BIT(PIPE_C) | BIT(PIPE_D),
+               .dbuf_mask = {
+                       [PIPE_A] = BIT(DBUF_S1),
+                       [PIPE_B] = BIT(DBUF_S1),
+                       [PIPE_C] = BIT(DBUF_S2),
+                       [PIPE_D] = BIT(DBUF_S2)
+               }
+       },
+};
+
+static u8 compute_dbuf_slices(enum pipe pipe,
+                             u32 active_pipes,
+                             const struct dbuf_slice_conf_entry *dbuf_slices,
+                             int size)
+{
+       int i;
+
+       for (i = 0; i < size; i++) {
+               if (dbuf_slices[i].active_pipes == active_pipes)
+                       return dbuf_slices[i].dbuf_mask[pipe];
+       }
+       return 0;
+}
+
+/*
+ * This function finds an entry with same enabled pipe configuration and
+ * returns correspondent DBuf slice mask as stated in BSpec for particular
+ * platform.
+ */
+static u32 icl_compute_dbuf_slices(enum pipe pipe,
+                                  u32 active_pipes)
+{
+       /*
+        * FIXME: For ICL this is still a bit unclear as prev BSpec revision
+        * required calculating "pipe ratio" in order to determine
+        * if one or two slices can be used for single pipe configurations
+        * as additional constraint to the existing table.
+        * However based on recent info, it should be not "pipe ratio"
+        * but rather ratio between pixel_rate and cdclk with additional
+        * constants, so for now we are using only table until this is
+        * clarified. Also this is the reason why crtc_state param is
+        * still here - we will need it once those additional constraints
+        * pop up.
+        */
+       return compute_dbuf_slices(pipe, active_pipes,
+                                  icl_allowed_dbufs,
+                                  ARRAY_SIZE(icl_allowed_dbufs));
+}
+
+static u32 tgl_compute_dbuf_slices(enum pipe pipe,
+                                  u32 active_pipes)
+{
+       return compute_dbuf_slices(pipe, active_pipes,
+                                  tgl_allowed_dbufs,
+                                  ARRAY_SIZE(tgl_allowed_dbufs));
+}
+
+static u8 skl_compute_dbuf_slices(const struct intel_crtc_state *crtc_state,
+                                 u32 active_pipes)
+{
+       struct intel_crtc *crtc = to_intel_crtc(crtc_state->uapi.crtc);
+       struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
+       enum pipe pipe = crtc->pipe;
+
+       if (IS_GEN(dev_priv, 12))
+               return tgl_compute_dbuf_slices(pipe,
+                                              active_pipes);
+       else if (IS_GEN(dev_priv, 11))
+               return icl_compute_dbuf_slices(pipe,
+                                              active_pipes);
+       /*
+        * For anything else just return one slice yet.
+        * Should be extended for other platforms.
+        */
+       return BIT(DBUF_S1);
+}
+
 static u64
 skl_plane_relative_data_rate(const struct intel_crtc_state *crtc_state,
                             const struct intel_plane_state *plane_state,