[Intel-gfx] [PATCH v2] drm/i915: use a separate context for gpu relocs

[Daniele Ceraolo Spurio]

The CS pre-parser can pre-fetch commands across memory sync points and
starting from gen12 it is able to pre-fetch across BB_START and BB_END
boundaries as well, so when we emit gpu relocs the pre-parser might
fetch the target location of the reloc before the memory write lands.

The parser can't pre-fetch across the ctx switch, so we use a separate
context to guarantee that the memory is synchronized before the parser
can get to it.

Note that there is no risk of the CS doing a lite restore from the reloc
context to the user context, even if the two have the same hw_id,
because since gen11 the CS also checks the LRCA when deciding if it can
lite-restore.

v2: limit new context to gen12+, release in eb_destroy, add a comment
    in emit_fini_breadcrumb (Chris).

Suggested-by: Chris Wilson <chris at chris-wilson.co.uk>
Signed-off-by: Daniele Ceraolo Spurio <daniele.ceraolospurio at intel.com>
Cc: Chris Wilson <chris at chris-wilson.co.uk>
---
 .../gpu/drm/i915/gem/i915_gem_execbuffer.c    | 30 ++++++++++++++++++-
 drivers/gpu/drm/i915/gt/intel_lrc.c           | 18 +++++++++++
 2 files changed, 47 insertions(+), 1 deletion(-)

diff --git a/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c b/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c
index b5f6937369ea..427210e301ae 100644
--- a/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c
+++ b/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c
@@ -252,6 +252,7 @@ struct i915_execbuffer {
 		bool has_fence : 1;
 		bool needs_unfenced : 1;
 
+		struct intel_context *ce;
 		struct i915_request *rq;
 		u32 *rq_cmd;
 		unsigned int rq_size;
@@ -880,6 +881,9 @@ static void eb_destroy(const struct i915_execbuffer *eb)
 {
 	GEM_BUG_ON(eb->reloc_cache.rq);
 
+	if (eb->reloc_cache.ce)
+		intel_context_put(eb->reloc_cache.ce);
+
 	if (eb->lut_size > 0)
 		kfree(eb->buckets);
 }
@@ -903,6 +907,7 @@ static void reloc_cache_init(struct reloc_cache *cache,
 	cache->has_fence = cache->gen < 4;
 	cache->needs_unfenced = INTEL_INFO(i915)->unfenced_needs_alignment;
 	cache->node.allocated = false;
+	cache->ce = NULL;
 	cache->rq = NULL;
 	cache->rq_size = 0;
 }
@@ -1168,7 +1173,7 @@ static int __reloc_gpu_alloc(struct i915_execbuffer *eb,
 	if (err)
 		goto err_unmap;
 
-	rq = i915_request_create(eb->context);
+	rq = intel_context_create_request(cache->ce);
 	if (IS_ERR(rq)) {
 		err = PTR_ERR(rq);
 		goto err_unpin;
@@ -1239,6 +1244,29 @@ static u32 *reloc_gpu(struct i915_execbuffer *eb,
 		if (!intel_engine_can_store_dword(eb->engine))
 			return ERR_PTR(-ENODEV);
 
+		if (!cache->ce) {
+			struct intel_context *ce;
+
+			/*
+			 * The CS pre-parser can pre-fetch commands across
+			 * memory sync points and starting gen12 it is able to
+			 * pre-fetch across BB_START and BB_END boundaries
+			 * (within the same context). We therefore use a
+			 * separate context gen12+ to guarantee that the reloc
+			 * writes land before the parser gets to the target
+			 * memory location.
+			 */
+			if (cache->gen >= 12)
+				ce = intel_context_create(eb->context->gem_context,
+							  eb->engine);
+			else
+				ce = intel_context_get(eb->context);
+			if (IS_ERR(ce))
+				return ERR_CAST(ce);
+
+			cache->ce = ce;
+		}
+
 		err = __reloc_gpu_alloc(eb, vma, len);
 		if (unlikely(err))
 			return ERR_PTR(err);
diff --git a/drivers/gpu/drm/i915/gt/intel_lrc.c b/drivers/gpu/drm/i915/gt/intel_lrc.c
index d42584439f51..12ec2830ba27 100644
--- a/drivers/gpu/drm/i915/gt/intel_lrc.c
+++ b/drivers/gpu/drm/i915/gt/intel_lrc.c
@@ -2903,6 +2903,24 @@ gen8_emit_fini_breadcrumb_footer(struct i915_request *request,
 	return gen8_emit_wa_tail(request, cs);
 }
 
+/*
+ * Note that the CS instruction pre-parser will not stall on the breadcrumb
+ * flush and will continue pre-fetching the instructions after it before the
+ * memory sync is completed. On pre-gen12 HW, the pre-parser will stop at
+ * BB_START/END instructions, so, even though we might pre-fetch the pre-amble
+ * of the next request before the memory has been flushed, we're guaranteed that
+ * we won't access the batch itself too early.
+ * However, on gen12+ the parser can pre-fetch across the BB_START/END commands,
+ * so, if the current request is modifying an instruction in the next request on
+ * the same intel_context, we might pre-fetch and then execute the pre-update
+ * instruction. To avoid this, the users of self-modifying code should either
+ * disable the parser around the code emitting the memory writes, via a new flag
+ * added to MI_ARB_CHECK, or emit the writes from a different intel_context. For
+ * the in-kernel use-cases we've opted to use a separate context, see
+ * reloc_gpu() as an example.
+ * All the above applies only to the instructions themselves. Non-inline data
+ * used by the instructions is not pre-fetched.
+ */
 static u32 *gen8_emit_fini_breadcrumb(struct i915_request *request, u32 *cs)
 {
 	cs = gen8_emit_ggtt_write(cs,
-- 
2.23.0