drm/i915: Extract context switch skip and add pd load logic
authorBen Widawsky <benjamin.widawsky@intel.com>
Mon, 16 Mar 2015 16:00:55 +0000 (16:00 +0000)
committerDaniel Vetter <daniel.vetter@ffwll.ch>
Fri, 20 Mar 2015 10:48:17 +0000 (11:48 +0100)
In Gen8, PDPs are saved and restored with legacy contexts (legacy contexts
only exist on the render ring). So change the ordering of LRI vs MI_SET_CONTEXT
for the initialization of the context. Also the only cases in which we
need to manually update the PDPs are when MI_RESTORE_INHIBIT has been
set in MI_SET_CONTEXT (i.e. when the context is not yet initialized or
it is the default context).

Legacy submission is not available post GEN8, so it isn't necessary to
add extra checks for newer generations.

v2: Use new functions to replace the logic right away (Daniel)
v3: Add missing pd load logic.
v4: Add warning in case pd_load_pre & pd_load_post are true, and add
missing trace_switch_mm. Cleaned up pd_load conditions. Add more
information about when is pd_load_post needed. (Mika)

Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Michel Thierry <michel.thierry@intel.com> (v2+)
Reviewed-by: Mika Kuoppala <mika.kuoppala@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
drivers/gpu/drm/i915/i915_gem_context.c

index 70346b0028f9def73db1dc32831a8f2d7b1d7dcf..b6ea85dcf9e6967a907776f4b4d46e0bc80a908b 100644 (file)
@@ -569,6 +569,56 @@ mi_set_context(struct intel_engine_cs *ring,
        return ret;
 }
 
+static inline bool should_skip_switch(struct intel_engine_cs *ring,
+                                     struct intel_context *from,
+                                     struct intel_context *to)
+{
+       if (from == to && !to->remap_slice)
+               return true;
+
+       return false;
+}
+
+static bool
+needs_pd_load_pre(struct intel_engine_cs *ring, struct intel_context *to)
+{
+       struct drm_i915_private *dev_priv = ring->dev->dev_private;
+
+       if (!to->ppgtt)
+               return false;
+
+       if (INTEL_INFO(ring->dev)->gen < 8)
+               return true;
+
+       if (ring != &dev_priv->ring[RCS])
+               return true;
+
+       return false;
+}
+
+static bool
+needs_pd_load_post(struct intel_engine_cs *ring, struct intel_context *to)
+{
+       struct drm_i915_private *dev_priv = ring->dev->dev_private;
+
+       if (!to->ppgtt)
+               return false;
+
+       if (!IS_GEN8(ring->dev))
+               return false;
+
+       if (ring != &dev_priv->ring[RCS])
+               return false;
+
+       if (!to->legacy_hw_ctx.initialized)
+               return true;
+
+       if (i915_gem_context_is_default(to))
+               return true;
+
+       return false;
+}
+
 static int do_switch(struct intel_engine_cs *ring,
                     struct intel_context *to)
 {
@@ -584,7 +634,7 @@ static int do_switch(struct intel_engine_cs *ring,
                BUG_ON(!i915_gem_obj_is_pinned(from->legacy_hw_ctx.rcs_state));
        }
 
-       if (from == to && !to->remap_slice)
+       if (should_skip_switch(ring, from, to))
                return 0;
 
        /* Trying to pin first makes error handling easier. */
@@ -602,7 +652,14 @@ static int do_switch(struct intel_engine_cs *ring,
         */
        from = ring->last_context;
 
-       if (to->ppgtt) {
+       /* We should never emit switch_mm more than once */
+       WARN_ON(needs_pd_load_pre(ring, to) && needs_pd_load_post(ring, to));
+
+       if (needs_pd_load_pre(ring, to)) {
+               /* Older GENs and non render rings still want the load first,
+                * "PP_DCLV followed by PP_DIR_BASE register through Load
+                * Register Immediate commands in Ring Buffer before submitting
+                * a context."*/
                trace_switch_mm(ring, to);
                ret = to->ppgtt->switch_mm(to->ppgtt, ring);
                if (ret)
@@ -644,6 +701,20 @@ static int do_switch(struct intel_engine_cs *ring,
        if (ret)
                goto unpin_out;
 
+       if (needs_pd_load_post(ring, to)) {
+               trace_switch_mm(ring, to);
+               ret = to->ppgtt->switch_mm(to->ppgtt, ring);
+               /* The hardware context switch is emitted, but we haven't
+                * actually changed the state - so it's probably safe to bail
+                * here. Still, let the user know something dangerous has
+                * happened.
+                */
+               if (ret) {
+                       DRM_ERROR("Failed to change address space on context switch\n");
+                       goto unpin_out;
+               }
+       }
+
        for (i = 0; i < MAX_L3_SLICES; i++) {
                if (!(to->remap_slice & (1<<i)))
                        continue;