drm/i915/gen9: Compute DDB allocation at atomic check time (v4)
authorMatt Roper <matthew.d.roper@intel.com>
Thu, 12 May 2016 14:06:03 +0000 (07:06 -0700)
committerMatt Roper <matthew.d.roper@intel.com>
Fri, 13 May 2016 14:34:00 +0000 (07:34 -0700)
Calculate the DDB blocks needed to satisfy the current atomic
transaction at atomic check time.  This is a prerequisite to calculating
SKL watermarks during the 'check' phase and rejecting any configurations
that we can't find valid watermarks for.

Due to the nature of DDB allocation, it's possible for the addition of a
new CRTC to make the watermark configuration already in use on another,
unchanged CRTC become invalid.  A change in which CRTC's are active
triggers a recompute of the entire DDB, which unfortunately means we
need to disallow any other atomic commits from racing with such an
update.  If the active CRTC's change, we need to grab the lock on all
CRTC's and run all CRTC's through their 'check' handler to recompute and
re-check their per-CRTC DDB allocations.

Note that with this patch we only compute the DDB allocation but we
don't actually use the computed values during watermark programming yet.
For ease of review/testing/bisecting, we still recompute the DDB at
watermark programming time and just WARN() if it doesn't match the
precomputed values.  A future patch will switch over to using the
precomputed values once we're sure they're being properly computed.

Another clarifying note:  DDB allocation itself shouldn't ever fail with
the algorithm we use today (i.e., we have enough DDB blocks on BXT to
support the minimum needs of the worst-case scenario of every pipe/plane
enabled at full size).  However the watermarks calculations based on the
DDB may fail and we'll be moving those to the atomic check as well in
future patches.

v2:
 - Skip DDB calculations in the rare case where our transaction doesn't
   actually touch any CRTC's at all.  Assuming at least one CRTC state
   is present in our transaction, then it means we can't race with any
   transactions that would update dev_priv->active_crtcs (which requires
   _all_ CRTC locks).

v3:
 - Also calculate DDB during initial hw readout, to prevent using
   incorrect bios values. (Maarten)

v4:
 - Use new distrust_bios_wm flag instead of skip_initial_wm (which was
   never actually set).
 - Set intel_state->active_pipe_changes instead of just realloc_pipes

Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Cc: Lyude Paul <cpaul@redhat.com>
Cc: Radhakrishna Sripada <radhakrishna.sripada@intel.com>
Signed-off-by: Matt Roper <matthew.d.roper@intel.com>
Signed-off-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Reviewed-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/1463061971-19638-10-git-send-email-matthew.d.roper@intel.com
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 985defc..1c8008b 100644 (file)
@@ -339,6 +339,10 @@ struct i915_hotplug {
 #define for_each_intel_crtc(dev, intel_crtc) \
        list_for_each_entry(intel_crtc, &dev->mode_config.crtc_list, base.head)
 
+#define for_each_intel_crtc_mask(dev, intel_crtc, crtc_mask) \
+       list_for_each_entry(intel_crtc, &dev->mode_config.crtc_list, base.head) \
+               for_each_if ((crtc_mask) & (1 << drm_crtc_index(&intel_crtc->base)))
+
 #define for_each_intel_encoder(dev, intel_encoder)             \
        list_for_each_entry(intel_encoder,                      \
                            &(dev)->mode_config.encoder_list,   \
@@ -594,6 +598,7 @@ struct drm_i915_display_funcs {
                                       struct intel_crtc_state *newstate);
        void (*initial_watermarks)(struct intel_crtc_state *cstate);
        void (*optimize_watermarks)(struct intel_crtc_state *cstate);
+       int (*compute_global_watermarks)(struct drm_atomic_state *state);
        void (*update_wm)(struct drm_crtc *crtc);
        int (*modeset_calc_cdclk)(struct drm_atomic_state *state);
        void (*modeset_commit_cdclk)(struct drm_atomic_state *state);
index 102320c..b1545a3 100644 (file)
@@ -13312,6 +13312,7 @@ static int intel_modeset_checks(struct drm_atomic_state *state)
 static void calc_watermark_data(struct drm_atomic_state *state)
 {
        struct drm_device *dev = state->dev;
+       struct drm_i915_private *dev_priv = to_i915(dev);
        struct intel_atomic_state *intel_state = to_intel_atomic_state(state);
        struct drm_crtc *crtc;
        struct drm_crtc_state *cstate;
@@ -13341,6 +13342,10 @@ static void calc_watermark_data(struct drm_atomic_state *state)
                    pstate->crtc_h != pstate->src_h >> 16)
                        intel_state->wm_config.sprites_scaled = true;
        }
+
+       /* Is there platform-specific watermark information to calculate? */
+       if (dev_priv->display.compute_global_watermarks)
+               dev_priv->display.compute_global_watermarks(state);
 }
 
 /**
@@ -13712,6 +13717,19 @@ static int intel_atomic_commit(struct drm_device *dev,
                intel_modeset_verify_crtc(crtc, old_crtc_state, crtc->state);
        }
 
+       /*
+        * Temporary sanity check: make sure our pre-computed DDB matches the
+        * one we actually wind up programming.
+        *
+        * Not a great place to put this, but the easiest place we have access
+        * to both the pre-computed and final DDB's; we'll be removing this
+        * check in the next patch anyway.
+        */
+       WARN(IS_GEN9(dev) &&
+            memcmp(&intel_state->ddb, &dev_priv->wm.skl_results.ddb,
+                   sizeof(intel_state->ddb)),
+            "Pre-computed DDB does not match final DDB!\n");
+
        if (intel_state->modeset)
                intel_display_power_put(dev_priv, POWER_DOMAIN_MODESET);
 
index 9f9c9cb..bd83eb6 100644 (file)
@@ -312,6 +312,9 @@ struct intel_atomic_state {
         * don't bother calculating intermediate watermarks.
         */
        bool skip_intermediate_wm;
+
+       /* Gen9+ only */
+       struct skl_ddb_allocation ddb;
 };
 
 struct intel_plane_state {
index 87fae2e..7b2545c 100644 (file)
@@ -3812,6 +3812,84 @@ static void skl_clear_wm(struct skl_wm_values *watermarks, enum pipe pipe)
 
 }
 
+static int
+skl_compute_ddb(struct drm_atomic_state *state)
+{
+       struct drm_device *dev = state->dev;
+       struct drm_i915_private *dev_priv = to_i915(dev);
+       struct intel_atomic_state *intel_state = to_intel_atomic_state(state);
+       struct intel_crtc *intel_crtc;
+       unsigned realloc_pipes = dev_priv->active_crtcs;
+       int ret;
+
+       /*
+        * If this is our first atomic update following hardware readout,
+        * we can't trust the DDB that the BIOS programmed for us.  Let's
+        * pretend that all pipes switched active status so that we'll
+        * ensure a full DDB recompute.
+        */
+       if (dev_priv->wm.distrust_bios_wm)
+               intel_state->active_pipe_changes = ~0;
+
+       /*
+        * If the modeset changes which CRTC's are active, we need to
+        * recompute the DDB allocation for *all* active pipes, even
+        * those that weren't otherwise being modified in any way by this
+        * atomic commit.  Due to the shrinking of the per-pipe allocations
+        * when new active CRTC's are added, it's possible for a pipe that
+        * we were already using and aren't changing at all here to suddenly
+        * become invalid if its DDB needs exceeds its new allocation.
+        *
+        * Note that if we wind up doing a full DDB recompute, we can't let
+        * any other display updates race with this transaction, so we need
+        * to grab the lock on *all* CRTC's.
+        */
+       if (intel_state->active_pipe_changes)
+               realloc_pipes = ~0;
+
+       for_each_intel_crtc_mask(dev, intel_crtc, realloc_pipes) {
+               struct intel_crtc_state *cstate;
+
+               cstate = intel_atomic_get_crtc_state(state, intel_crtc);
+               if (IS_ERR(cstate))
+                       return PTR_ERR(cstate);
+
+               ret = skl_allocate_pipe_ddb(cstate, &intel_state->ddb);
+               if (ret)
+                       return ret;
+       }
+
+       return 0;
+}
+
+static int
+skl_compute_wm(struct drm_atomic_state *state)
+{
+       struct drm_crtc *crtc;
+       struct drm_crtc_state *cstate;
+       int ret, i;
+       bool changed = false;
+
+       /*
+        * If this transaction isn't actually touching any CRTC's, don't
+        * bother with watermark calculation.  Note that if we pass this
+        * test, we're guaranteed to hold at least one CRTC state mutex,
+        * which means we can safely use values like dev_priv->active_crtcs
+        * since any racing commits that want to update them would need to
+        * hold _all_ CRTC state mutexes.
+        */
+       for_each_crtc_in_state(state, crtc, cstate, i)
+               changed = true;
+       if (!changed)
+               return 0;
+
+       ret = skl_compute_ddb(state);
+       if (ret)
+               return ret;
+
+       return 0;
+}
+
 static void skl_update_wm(struct drm_crtc *crtc)
 {
        struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
@@ -7334,6 +7412,7 @@ void intel_init_pm(struct drm_device *dev)
        if (INTEL_INFO(dev)->gen >= 9) {
                skl_setup_wm_latency(dev);
                dev_priv->display.update_wm = skl_update_wm;
+               dev_priv->display.compute_global_watermarks = skl_compute_wm;
        } else if (HAS_PCH_SPLIT(dev)) {
                ilk_setup_wm_latency(dev);