drm/i915: Unwind incomplete legacy context switches

The legacy context switch for ringbuffer submission is multistaged,
where each of those stages may fail. However, we were updating global
state after some stages, and so we had to force the incomplete request
to be submitted because we could not unwind. Save the global state
before performing the switches, and so enable us to unwind back to the
previous global state should any phase fail. We then must cancel the
request instead of submitting it should the construction fail.

v2: s/saved_ctx/from_ctx/; s/ctx/to_ctx/ etc.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20171123152631.31385-1-chris@chris-wilson.co.uk

1 files changed, 57 insertions, 116 deletions

diff --git a/drivers/gpu/drm/i915/i915_gem_context.c b/drivers/gpu/drm/i915/i915_gem_context.c
index 6ca56e4..da4a64b 100644
--- a/drivers/gpu/drm/i915/i915_gem_context.c
+++ b/drivers/gpu/drm/i915/i915_gem_context.c
@@ -507,6 +507,7 @@ void i915_gem_contexts_lost(struct drm_i915_private *dev_priv)
 
 	for_each_engine(engine, dev_priv, id) {
 		engine->legacy_active_context = NULL;
+		engine->legacy_active_ppgtt = NULL;
 
 		if (!engine->last_retired_context)
 			continue;
@@ -681,68 +682,49 @@ static int remap_l3(struct drm_i915_gem_request *req, int slice)
 	return 0;
 }
 
-static inline bool skip_rcs_switch(struct i915_hw_ppgtt *ppgtt,
-				   struct intel_engine_cs *engine,
-				   struct i915_gem_context *to)
-{
-	if (to->remap_slice)
-		return false;
-
-	if (ppgtt && (intel_engine_flag(engine) & ppgtt->pd_dirty_rings))
-		return false;
-
-	return to == engine->legacy_active_context;
-}
-
-static bool
-needs_pd_load_pre(struct i915_hw_ppgtt *ppgtt, struct intel_engine_cs *engine)
-{
-	struct i915_gem_context *from = engine->legacy_active_context;
-
-	if (!ppgtt)
-		return false;
-
-	/* Always load the ppgtt on first use */
-	if (!from)
-		return true;
-
-	/* Same context without new entries, skip */
-	if ((!from->ppgtt || from->ppgtt == ppgtt) &&
-	    !(intel_engine_flag(engine) & ppgtt->pd_dirty_rings))
-		return false;
-
-	if (engine->id != RCS)
-		return true;
-
-	return true;
-}
-
-static int do_rcs_switch(struct drm_i915_gem_request *req)
+/**
+ * i915_switch_context() - perform a GPU context switch.
+ * @rq: request for which we'll execute the context switch
+ *
+ * The context life cycle is simple. The context refcount is incremented and
+ * decremented by 1 on create and destroy. If the context is in use by the GPU,
+ * it will have a refcount > 1. This allows us to destroy the context abstract
+ * object while letting the normal object tracking destroy the backing BO.
+ *
+ * This function should not be used in execlists mode.  Instead the context is
+ * switched by writing to the ELSP and requests keep a reference to their
+ * context.
+ */
+int i915_switch_context(struct drm_i915_gem_request *rq)
 {
-	struct i915_gem_context *to = req->ctx;
-	struct intel_engine_cs *engine = req->engine;
-	struct i915_hw_ppgtt *ppgtt = to->ppgtt ?: req->i915->mm.aliasing_ppgtt;
-	struct i915_gem_context *from = engine->legacy_active_context;
-	u32 hw_flags;
+	struct intel_engine_cs *engine = rq->engine;
+	struct i915_gem_context *to_ctx = rq->ctx;
+	struct i915_hw_ppgtt *to_mm =
+		to_ctx->ppgtt ?: rq->i915->mm.aliasing_ppgtt;
+	struct i915_gem_context *from_ctx = engine->legacy_active_context;
+	struct i915_hw_ppgtt *from_mm = engine->legacy_active_ppgtt;
+	u32 hw_flags = 0;
 	int ret, i;
 
-	GEM_BUG_ON(engine->id != RCS);
+	lockdep_assert_held(&rq->i915->drm.struct_mutex);
+	GEM_BUG_ON(HAS_EXECLISTS(rq->i915));
 
-	if (skip_rcs_switch(ppgtt, engine, to))
-		return 0;
-
-	if (needs_pd_load_pre(ppgtt, engine)) {
-		/* 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(engine, to);
-		ret = ppgtt->switch_mm(ppgtt, req);
+	if (to_mm != from_mm ||
+	    (to_mm && intel_engine_flag(engine) & to_mm->pd_dirty_rings)) {
+		trace_switch_mm(engine, to_ctx);
+		ret = to_mm->switch_mm(to_mm, rq);
 		if (ret)
-			return ret;
+			goto err;
+
+		to_mm->pd_dirty_rings &= ~intel_engine_flag(engine);
+		engine->legacy_active_ppgtt = to_mm;
+		hw_flags = MI_FORCE_RESTORE;
 	}
 
-	if (i915_gem_context_is_kernel(to))
+	if (to_ctx->engine[engine->id].state &&
+	    (to_ctx != from_ctx || hw_flags & MI_FORCE_RESTORE)) {
+		GEM_BUG_ON(engine->id != RCS);
+
 		/*
 		 * The kernel context(s) is treated as pure scratch and is not
 		 * expected to retain any state (as we sacrifice it during
@@ -750,78 +732,37 @@ static int do_rcs_switch(struct drm_i915_gem_request *req)
 		 * as nothing actually executes using the kernel context; it
 		 * is purely used for flushing user contexts.
 		 */
-		hw_flags = MI_RESTORE_INHIBIT;
-	else if (ppgtt && intel_engine_flag(engine) & ppgtt->pd_dirty_rings)
-		hw_flags = MI_FORCE_RESTORE;
-	else
-		hw_flags = 0;
+		if (i915_gem_context_is_kernel(to_ctx))
+			hw_flags = MI_RESTORE_INHIBIT;
 
-	if (to != from || (hw_flags & MI_FORCE_RESTORE)) {
-		ret = mi_set_context(req, hw_flags);
+		ret = mi_set_context(rq, hw_flags);
 		if (ret)
-			return ret;
+			goto err_mm;
 
-		engine->legacy_active_context = to;
+		engine->legacy_active_context = to_ctx;
 	}
 
-	if (ppgtt)
-		ppgtt->pd_dirty_rings &= ~intel_engine_flag(engine);
+	if (to_ctx->remap_slice) {
+		for (i = 0; i < MAX_L3_SLICES; i++) {
+			if (!(to_ctx->remap_slice & BIT(i)))
+				continue;
 
-	for (i = 0; i < MAX_L3_SLICES; i++) {
-		if (!(to->remap_slice & (1<<i)))
-			continue;
-
-		ret = remap_l3(req, i);
-		if (ret)
-			return ret;
-
-		to->remap_slice &= ~(1<<i);
-	}
-
-	return 0;
-}
-
-/**
- * i915_switch_context() - perform a GPU context switch.
- * @req: request for which we'll execute the context switch
- *
- * The context life cycle is simple. The context refcount is incremented and
- * decremented by 1 and create and destroy. If the context is in use by the GPU,
- * it will have a refcount > 1. This allows us to destroy the context abstract
- * object while letting the normal object tracking destroy the backing BO.
- *
- * This function should not be used in execlists mode.  Instead the context is
- * switched by writing to the ELSP and requests keep a reference to their
- * context.
- */
-int i915_switch_context(struct drm_i915_gem_request *req)
-{
-	struct intel_engine_cs *engine = req->engine;
-
-	lockdep_assert_held(&req->i915->drm.struct_mutex);
-	GEM_BUG_ON(HAS_EXECLISTS(req->i915));
-
-	if (!req->ctx->engine[engine->id].state) {
-		struct i915_gem_context *to = req->ctx;
-		struct i915_hw_ppgtt *ppgtt =
-			to->ppgtt ?: req->i915->mm.aliasing_ppgtt;
-
-		if (needs_pd_load_pre(ppgtt, engine)) {
-			int ret;
-
-			trace_switch_mm(engine, to);
-			ret = ppgtt->switch_mm(ppgtt, req);
+			ret = remap_l3(rq, i);
 			if (ret)
-				return ret;
-
-			ppgtt->pd_dirty_rings &= ~intel_engine_flag(engine);
+				goto err_ctx;
 		}
 
-		engine->legacy_active_context = to;
-		return 0;
+		to_ctx->remap_slice = 0;
 	}
 
-	return do_rcs_switch(req);
+	return 0;
+
+err_ctx:
+	engine->legacy_active_context = from_ctx;
+err_mm:
+	engine->legacy_active_ppgtt = from_mm;
+err:
+	return ret;
 }
 
 static bool engine_has_idle_kernel_context(struct intel_engine_cs *engine)