drm/i915/gt: Split intel_ring_submission

Split the legacy submission backend from the common CS ring buffer
handling.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191024100344.5041-1-chris@chris-wilson.co.uk

drivers/gpu/drm/i915/Makefile

@@ -89,11 +89,12 @@ gt-y += \
 	gt/intel_gt_requests.o \
 	gt/intel_llc.o \
 	gt/intel_lrc.o \
+	gt/intel_mocs.o \
 	gt/intel_rc6.o \
 	gt/intel_renderstate.o \
 	gt/intel_reset.o \
-	gt/intel_ringbuffer.o \
+	gt/intel_ring.o \
-	gt/intel_mocs.o \
+	gt/intel_ring_submission.o \
 	gt/intel_sseu.o \
 	gt/intel_timeline.o \
 	gt/intel_workarounds.o

drivers/gpu/drm/i915/display/intel_overlay.c

@@ -30,6 +30,7 @@
 #include <drm/i915_drm.h>
 
 #include "gem/i915_gem_pm.h"
+#include "gt/intel_ring.h"
 
 #include "i915_drv.h"
 #include "i915_reg.h"

drivers/gpu/drm/i915/gem/i915_gem_context.c

@@ -69,9 +69,10 @@
 
 #include <drm/i915_drm.h>
 
-#include "gt/intel_lrc_reg.h"
 #include "gt/intel_engine_heartbeat.h"
 #include "gt/intel_engine_user.h"
+#include "gt/intel_lrc_reg.h"
+#include "gt/intel_ring.h"
 
 #include "i915_gem_context.h"
 #include "i915_globals.h"

drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c

@@ -19,6 +19,7 @@
 #include "gt/intel_engine_pool.h"
 #include "gt/intel_gt.h"
 #include "gt/intel_gt_pm.h"
+#include "gt/intel_ring.h"
 
 #include "i915_drv.h"
 #include "i915_gem_clflush.h"

drivers/gpu/drm/i915/gem/i915_gem_object_blt.c

@@ -8,6 +8,7 @@
 #include "gt/intel_engine_pm.h"
 #include "gt/intel_engine_pool.h"
 #include "gt/intel_gt.h"
+#include "gt/intel_ring.h"
 #include "i915_gem_clflush.h"
 #include "i915_gem_object_blt.h"
 

drivers/gpu/drm/i915/gem/selftests/i915_gem_coherency.c

@@ -8,6 +8,7 @@
 
 #include "gt/intel_gt.h"
 #include "gt/intel_gt_pm.h"
+#include "gt/intel_ring.h"
 
 #include "i915_selftest.h"
 #include "selftests/i915_random.h"

drivers/gpu/drm/i915/gt/intel_context.c

@@ -13,6 +13,7 @@
 #include "intel_context.h"
 #include "intel_engine.h"
 #include "intel_engine_pm.h"
+#include "intel_ring.h"
 
 static struct i915_global_context {
 	struct i915_global base;

drivers/gpu/drm/i915/gt/intel_context.h

@@ -12,6 +12,7 @@
 #include "i915_active.h"
 #include "intel_context_types.h"
 #include "intel_engine_types.h"
+#include "intel_ring_types.h"
 #include "intel_timeline_types.h"
 
 void intel_context_init(struct intel_context *ce,

drivers/gpu/drm/i915/gt/intel_engine.h

@@ -19,7 +19,6 @@
 #include "intel_workarounds.h"
 
 struct drm_printer;
-
 struct intel_gt;
 
 /* Early gen2 devices have a cacheline of just 32 bytes, using 64 is overkill,
@@ -176,122 +175,9 @@ intel_write_status_page(struct intel_engine_cs *engine, int reg, u32 value)
 #define I915_HWS_CSB_WRITE_INDEX	0x1f
 #define CNL_HWS_CSB_WRITE_INDEX		0x2f
 
-struct intel_ring *
-intel_engine_create_ring(struct intel_engine_cs *engine, int size);
-int intel_ring_pin(struct intel_ring *ring);
-void intel_ring_reset(struct intel_ring *ring, u32 tail);
-unsigned int intel_ring_update_space(struct intel_ring *ring);
-void intel_ring_unpin(struct intel_ring *ring);
-void intel_ring_free(struct kref *ref);
-
-static inline struct intel_ring *intel_ring_get(struct intel_ring *ring)
-{
-	kref_get(&ring->ref);
-	return ring;
-}
-
-static inline void intel_ring_put(struct intel_ring *ring)
-{
-	kref_put(&ring->ref, intel_ring_free);
-}
-
 void intel_engine_stop(struct intel_engine_cs *engine);
 void intel_engine_cleanup(struct intel_engine_cs *engine);
 
-int __must_check intel_ring_cacheline_align(struct i915_request *rq);
-
-u32 __must_check *intel_ring_begin(struct i915_request *rq, unsigned int n);
-
-static inline void intel_ring_advance(struct i915_request *rq, u32 *cs)
-{
-	/* Dummy function.
-	 *
-	 * This serves as a placeholder in the code so that the reader
-	 * can compare against the preceding intel_ring_begin() and
-	 * check that the number of dwords emitted matches the space
-	 * reserved for the command packet (i.e. the value passed to
-	 * intel_ring_begin()).
-	 */
-	GEM_BUG_ON((rq->ring->vaddr + rq->ring->emit) != cs);
-}
-
-static inline u32 intel_ring_wrap(const struct intel_ring *ring, u32 pos)
-{
-	return pos & (ring->size - 1);
-}
-
-static inline bool
-intel_ring_offset_valid(const struct intel_ring *ring,
-			unsigned int pos)
-{
-	if (pos & -ring->size) /* must be strictly within the ring */
-		return false;
-
-	if (!IS_ALIGNED(pos, 8)) /* must be qword aligned */
-		return false;
-
-	return true;
-}
-
-static inline u32 intel_ring_offset(const struct i915_request *rq, void *addr)
-{
-	/* Don't write ring->size (equivalent to 0) as that hangs some GPUs. */
-	u32 offset = addr - rq->ring->vaddr;
-	GEM_BUG_ON(offset > rq->ring->size);
-	return intel_ring_wrap(rq->ring, offset);
-}
-
-static inline void
-assert_ring_tail_valid(const struct intel_ring *ring, unsigned int tail)
-{
-	GEM_BUG_ON(!intel_ring_offset_valid(ring, tail));
-
-	/*
-	 * "Ring Buffer Use"
-	 *	Gen2 BSpec "1. Programming Environment" / 1.4.4.6
-	 *	Gen3 BSpec "1c Memory Interface Functions" / 2.3.4.5
-	 *	Gen4+ BSpec "1c Memory Interface and Command Stream" / 5.3.4.5
-	 * "If the Ring Buffer Head Pointer and the Tail Pointer are on the
-	 * same cacheline, the Head Pointer must not be greater than the Tail
-	 * Pointer."
-	 *
-	 * We use ring->head as the last known location of the actual RING_HEAD,
-	 * it may have advanced but in the worst case it is equally the same
-	 * as ring->head and so we should never program RING_TAIL to advance
-	 * into the same cacheline as ring->head.
-	 */
-#define cacheline(a) round_down(a, CACHELINE_BYTES)
-	GEM_BUG_ON(cacheline(tail) == cacheline(ring->head) &&
-		   tail < ring->head);
-#undef cacheline
-}
-
-static inline unsigned int
-intel_ring_set_tail(struct intel_ring *ring, unsigned int tail)
-{
-	/* Whilst writes to the tail are strictly order, there is no
-	 * serialisation between readers and the writers. The tail may be
-	 * read by i915_request_retire() just as it is being updated
-	 * by execlists, as although the breadcrumb is complete, the context
-	 * switch hasn't been seen.
-	 */
-	assert_ring_tail_valid(ring, tail);
-	ring->tail = tail;
-	return tail;
-}
-
-static inline unsigned int
-__intel_ring_space(unsigned int head, unsigned int tail, unsigned int size)
-{
-	/*
-	 * "If the Ring Buffer Head Pointer and the Tail Pointer are on the
-	 * same cacheline, the Head Pointer must not be greater than the Tail
-	 * Pointer."
-	 */
-	GEM_BUG_ON(!is_power_of_2(size));
-	return (head - tail - CACHELINE_BYTES) & (size - 1);
-}
-
 int intel_engines_init_mmio(struct intel_gt *gt);
 int intel_engines_setup(struct intel_gt *gt);
 int intel_engines_init(struct intel_gt *gt);

drivers/gpu/drm/i915/gt/intel_engine_cs.c

@@ -37,6 +37,7 @@
 #include "intel_context.h"
 #include "intel_lrc.h"
 #include "intel_reset.h"
+#include "intel_ring.h"
 
 /* Haswell does have the CXT_SIZE register however it does not appear to be
  * valid. Now, docs explain in dwords what is in the context object. The full

drivers/gpu/drm/i915/gt/intel_engine_pm.c

@@ -13,6 +13,7 @@
 #include "intel_gt.h"
 #include "intel_gt_pm.h"
 #include "intel_rc6.h"
+#include "intel_ring.h"
 
 static int __engine_unpark(struct intel_wakeref *wf)
 {

drivers/gpu/drm/i915/gt/intel_engine_types.h

@@ -59,6 +59,7 @@ struct i915_gem_context;
 struct i915_request;
 struct i915_sched_attr;
 struct intel_gt;
+struct intel_ring;
 struct intel_uncore;
 
 typedef u8 intel_engine_mask_t;
@@ -77,32 +78,6 @@ struct intel_instdone {
 	u32 row[I915_MAX_SLICES][I915_MAX_SUBSLICES];
 };
 
-struct intel_ring {
-	struct kref ref;
-	struct i915_vma *vma;
-	void *vaddr;
-
-	/*
-	 * As we have two types of rings, one global to the engine used
-	 * by ringbuffer submission and those that are exclusive to a
-	 * context used by execlists, we have to play safe and allow
-	 * atomic updates to the pin_count. However, the actual pinning
-	 * of the context is either done during initialisation for
-	 * ringbuffer submission or serialised as part of the context
-	 * pinning for execlists, and so we do not need a mutex ourselves
-	 * to serialise intel_ring_pin/intel_ring_unpin.
-	 */
-	atomic_t pin_count;
-
-	u32 head;
-	u32 tail;
-	u32 emit;
-
-	u32 space;
-	u32 size;
-	u32 effective_size;
-};
-
 /*
  * we use a single page to load ctx workarounds so all of these
  * values are referred in terms of dwords

drivers/gpu/drm/i915/gt/intel_lrc.c

@@ -145,6 +145,7 @@
 #include "intel_lrc_reg.h"
 #include "intel_mocs.h"
 #include "intel_reset.h"
+#include "intel_ring.h"
 #include "intel_workarounds.h"
 
 #define RING_EXECLIST_QFULL		(1 << 0x2)

drivers/gpu/drm/i915/gt/intel_mocs.c

@@ -26,6 +26,7 @@
 #include "intel_gt.h"
 #include "intel_mocs.h"
 #include "intel_lrc.h"
+#include "intel_ring.h"
 
 /* structures required */
 struct drm_i915_mocs_entry {

drivers/gpu/drm/i915/gt/intel_renderstate.c

@@ -27,6 +27,7 @@
 
 #include "i915_drv.h"
 #include "intel_renderstate.h"
+#include "intel_ring.h"
 
 struct intel_renderstate {
 	const struct intel_renderstate_rodata *rodata;

drivers/gpu/drm/i915/gt/intel_ring.c

@@ -0,0 +1,321 @@
+/*
+ * SPDX-License-Identifier: MIT
+ *
+ * Copyright © 2019 Intel Corporation
+ */
+
+#include "gem/i915_gem_object.h"
+#include "i915_drv.h"
+#include "i915_vma.h"
+#include "intel_engine.h"
+#include "intel_ring.h"
+#include "intel_timeline.h"
+
+unsigned int intel_ring_update_space(struct intel_ring *ring)
+{
+	unsigned int space;
+
+	space = __intel_ring_space(ring->head, ring->emit, ring->size);
+
+	ring->space = space;
+	return space;
+}
+
+int intel_ring_pin(struct intel_ring *ring)
+{
+	struct i915_vma *vma = ring->vma;
+	unsigned int flags;
+	void *addr;
+	int ret;
+
+	if (atomic_fetch_inc(&ring->pin_count))
+		return 0;
+
+	flags = PIN_GLOBAL;
+
+	/* Ring wraparound at offset 0 sometimes hangs. No idea why. */
+	flags |= PIN_OFFSET_BIAS | i915_ggtt_pin_bias(vma);
+
+	if (vma->obj->stolen)
+		flags |= PIN_MAPPABLE;
+	else
+		flags |= PIN_HIGH;
+
+	ret = i915_vma_pin(vma, 0, 0, flags);
+	if (unlikely(ret))
+		goto err_unpin;
+
+	if (i915_vma_is_map_and_fenceable(vma))
+		addr = (void __force *)i915_vma_pin_iomap(vma);
+	else
+		addr = i915_gem_object_pin_map(vma->obj,
+					       i915_coherent_map_type(vma->vm->i915));
+	if (IS_ERR(addr)) {
+		ret = PTR_ERR(addr);
+		goto err_ring;
+	}
+
+	i915_vma_make_unshrinkable(vma);
+
+	GEM_BUG_ON(ring->vaddr);
+	ring->vaddr = addr;
+
+	return 0;
+
+err_ring:
+	i915_vma_unpin(vma);
+err_unpin:
+	atomic_dec(&ring->pin_count);
+	return ret;
+}
+
+void intel_ring_reset(struct intel_ring *ring, u32 tail)
+{
+	tail = intel_ring_wrap(ring, tail);
+	ring->tail = tail;
+	ring->head = tail;
+	ring->emit = tail;
+	intel_ring_update_space(ring);
+}
+
+void intel_ring_unpin(struct intel_ring *ring)
+{
+	struct i915_vma *vma = ring->vma;
+
+	if (!atomic_dec_and_test(&ring->pin_count))
+		return;
+
+	/* Discard any unused bytes beyond that submitted to hw. */
+	intel_ring_reset(ring, ring->emit);
+
+	i915_vma_unset_ggtt_write(vma);
+	if (i915_vma_is_map_and_fenceable(vma))
+		i915_vma_unpin_iomap(vma);
+	else
+		i915_gem_object_unpin_map(vma->obj);
+
+	GEM_BUG_ON(!ring->vaddr);
+	ring->vaddr = NULL;
+
+	i915_vma_unpin(vma);
+	i915_vma_make_purgeable(vma);
+}
+
+static struct i915_vma *create_ring_vma(struct i915_ggtt *ggtt, int size)
+{
+	struct i915_address_space *vm = &ggtt->vm;
+	struct drm_i915_private *i915 = vm->i915;
+	struct drm_i915_gem_object *obj;
+	struct i915_vma *vma;
+
+	obj = i915_gem_object_create_stolen(i915, size);
+	if (IS_ERR(obj))
+		obj = i915_gem_object_create_internal(i915, size);
+	if (IS_ERR(obj))
+		return ERR_CAST(obj);
+
+	/*
+	 * Mark ring buffers as read-only from GPU side (so no stray overwrites)
+	 * if supported by the platform's GGTT.
+	 */
+	if (vm->has_read_only)
+		i915_gem_object_set_readonly(obj);
+
+	vma = i915_vma_instance(obj, vm, NULL);
+	if (IS_ERR(vma))
+		goto err;
+
+	return vma;
+
+err:
+	i915_gem_object_put(obj);
+	return vma;
+}
+
+struct intel_ring *
+intel_engine_create_ring(struct intel_engine_cs *engine, int size)
+{
+	struct drm_i915_private *i915 = engine->i915;
+	struct intel_ring *ring;
+	struct i915_vma *vma;
+
+	GEM_BUG_ON(!is_power_of_2(size));
+	GEM_BUG_ON(RING_CTL_SIZE(size) & ~RING_NR_PAGES);
+
+	ring = kzalloc(sizeof(*ring), GFP_KERNEL);
+	if (!ring)
+		return ERR_PTR(-ENOMEM);
+
+	kref_init(&ring->ref);
+	ring->size = size;
+
+	/*
+	 * Workaround an erratum on the i830 which causes a hang if
+	 * the TAIL pointer points to within the last 2 cachelines
+	 * of the buffer.
+	 */
+	ring->effective_size = size;
+	if (IS_I830(i915) || IS_I845G(i915))
+		ring->effective_size -= 2 * CACHELINE_BYTES;
+
+	intel_ring_update_space(ring);
+
+	vma = create_ring_vma(engine->gt->ggtt, size);
+	if (IS_ERR(vma)) {
+		kfree(ring);
+		return ERR_CAST(vma);
+	}
+	ring->vma = vma;
+
+	return ring;
+}
+
+void intel_ring_free(struct kref *ref)
+{
+	struct intel_ring *ring = container_of(ref, typeof(*ring), ref);
+
+	i915_vma_put(ring->vma);
+	kfree(ring);
+}
+
+static noinline int
+wait_for_space(struct intel_ring *ring,
+	       struct intel_timeline *tl,
+	       unsigned int bytes)
+{
+	struct i915_request *target;
+	long timeout;
+
+	if (intel_ring_update_space(ring) >= bytes)
+		return 0;
+
+	GEM_BUG_ON(list_empty(&tl->requests));
+	list_for_each_entry(target, &tl->requests, link) {
+		if (target->ring != ring)
+			continue;
+
+		/* Would completion of this request free enough space? */
+		if (bytes <= __intel_ring_space(target->postfix,
+						ring->emit, ring->size))
+			break;
+	}
+
+	if (GEM_WARN_ON(&target->link == &tl->requests))
+		return -ENOSPC;
+
+	timeout = i915_request_wait(target,
+				    I915_WAIT_INTERRUPTIBLE,
+				    MAX_SCHEDULE_TIMEOUT);
+	if (timeout < 0)
+		return timeout;
+
+	i915_request_retire_upto(target);
+
+	intel_ring_update_space(ring);
+	GEM_BUG_ON(ring->space < bytes);
+	return 0;
+}
+
+u32 *intel_ring_begin(struct i915_request *rq, unsigned int num_dwords)
+{
+	struct intel_ring *ring = rq->ring;
+	const unsigned int remain_usable = ring->effective_size - ring->emit;
+	const unsigned int bytes = num_dwords * sizeof(u32);
+	unsigned int need_wrap = 0;
+	unsigned int total_bytes;
+	u32 *cs;
+
+	/* Packets must be qword aligned. */
+	GEM_BUG_ON(num_dwords & 1);
+
+	total_bytes = bytes + rq->reserved_space;
+	GEM_BUG_ON(total_bytes > ring->effective_size);
+
+	if (unlikely(total_bytes > remain_usable)) {
+		const int remain_actual = ring->size - ring->emit;
+
+		if (bytes > remain_usable) {
+			/*
+			 * Not enough space for the basic request. So need to
+			 * flush out the remainder and then wait for
+			 * base + reserved.
+			 */
+			total_bytes += remain_actual;
+			need_wrap = remain_actual | 1;
+		} else  {
+			/*
+			 * The base request will fit but the reserved space
+			 * falls off the end. So we don't need an immediate
+			 * wrap and only need to effectively wait for the
+			 * reserved size from the start of ringbuffer.
+			 */
+			total_bytes = rq->reserved_space + remain_actual;
+		}
+	}
+
+	if (unlikely(total_bytes > ring->space)) {
+		int ret;
+
+		/*
+		 * Space is reserved in the ringbuffer for finalising the
+		 * request, as that cannot be allowed to fail. During request
+		 * finalisation, reserved_space is set to 0 to stop the
+		 * overallocation and the assumption is that then we never need
+		 * to wait (which has the risk of failing with EINTR).
+		 *
+		 * See also i915_request_alloc() and i915_request_add().
+		 */
+		GEM_BUG_ON(!rq->reserved_space);
+
+		ret = wait_for_space(ring,
+				     i915_request_timeline(rq),
+				     total_bytes);
+		if (unlikely(ret))
+			return ERR_PTR(ret);
+	}
+
+	if (unlikely(need_wrap)) {
+		need_wrap &= ~1;
+		GEM_BUG_ON(need_wrap > ring->space);
+		GEM_BUG_ON(ring->emit + need_wrap > ring->size);
+		GEM_BUG_ON(!IS_ALIGNED(need_wrap, sizeof(u64)));
+
+		/* Fill the tail with MI_NOOP */
+		memset64(ring->vaddr + ring->emit, 0, need_wrap / sizeof(u64));
+		ring->space -= need_wrap;
+		ring->emit = 0;
+	}
+
+	GEM_BUG_ON(ring->emit > ring->size - bytes);
+	GEM_BUG_ON(ring->space < bytes);
+	cs = ring->vaddr + ring->emit;
+	GEM_DEBUG_EXEC(memset32(cs, POISON_INUSE, bytes / sizeof(*cs)));
+	ring->emit += bytes;
+	ring->space -= bytes;
+
+	return cs;
+}
+
+/* Align the ring tail to a cacheline boundary */
+int intel_ring_cacheline_align(struct i915_request *rq)
+{
+	int num_dwords;
+	void *cs;
+
+	num_dwords = (rq->ring->emit & (CACHELINE_BYTES - 1)) / sizeof(u32);
+	if (num_dwords == 0)
+		return 0;
+
+	num_dwords = CACHELINE_DWORDS - num_dwords;
+	GEM_BUG_ON(num_dwords & 1);
+
+	cs = intel_ring_begin(rq, num_dwords);
+	if (IS_ERR(cs))
+		return PTR_ERR(cs);
+
+	memset64(cs, (u64)MI_NOOP << 32 | MI_NOOP, num_dwords / 2);
+	intel_ring_advance(rq, cs + num_dwords);
+
+	GEM_BUG_ON(rq->ring->emit & (CACHELINE_BYTES - 1));
+	return 0;
+}

drivers/gpu/drm/i915/gt/intel_ring.h

@@ -0,0 +1,131 @@
+/*
+ * SPDX-License-Identifier: MIT
+ *
+ * Copyright © 2019 Intel Corporation
+ */
+
+#ifndef INTEL_RING_H
+#define INTEL_RING_H
+
+#include "i915_gem.h" /* GEM_BUG_ON */
+#include "i915_request.h"
+#include "intel_ring_types.h"
+
+struct intel_engine_cs;
+
+struct intel_ring *
+intel_engine_create_ring(struct intel_engine_cs *engine, int size);
+
+u32 *intel_ring_begin(struct i915_request *rq, unsigned int num_dwords);
+int intel_ring_cacheline_align(struct i915_request *rq);
+
+unsigned int intel_ring_update_space(struct intel_ring *ring);
+
+int intel_ring_pin(struct intel_ring *ring);
+void intel_ring_unpin(struct intel_ring *ring);
+void intel_ring_reset(struct intel_ring *ring, u32 tail);
+
+void intel_ring_free(struct kref *ref);
+
+static inline struct intel_ring *intel_ring_get(struct intel_ring *ring)
+{
+	kref_get(&ring->ref);
+	return ring;
+}
+
+static inline void intel_ring_put(struct intel_ring *ring)
+{
+	kref_put(&ring->ref, intel_ring_free);
+}
+
+static inline void intel_ring_advance(struct i915_request *rq, u32 *cs)
+{
+	/* Dummy function.
+	 *
+	 * This serves as a placeholder in the code so that the reader
+	 * can compare against the preceding intel_ring_begin() and
+	 * check that the number of dwords emitted matches the space
+	 * reserved for the command packet (i.e. the value passed to
+	 * intel_ring_begin()).
+	 */
+	GEM_BUG_ON((rq->ring->vaddr + rq->ring->emit) != cs);
+}
+
+static inline u32 intel_ring_wrap(const struct intel_ring *ring, u32 pos)
+{
+	return pos & (ring->size - 1);
+}
+
+static inline bool
+intel_ring_offset_valid(const struct intel_ring *ring,
+			unsigned int pos)
+{
+	if (pos & -ring->size) /* must be strictly within the ring */
+		return false;
+
+	if (!IS_ALIGNED(pos, 8)) /* must be qword aligned */
+		return false;
+
+	return true;
+}
+
+static inline u32 intel_ring_offset(const struct i915_request *rq, void *addr)
+{
+	/* Don't write ring->size (equivalent to 0) as that hangs some GPUs. */
+	u32 offset = addr - rq->ring->vaddr;
+	GEM_BUG_ON(offset > rq->ring->size);
+	return intel_ring_wrap(rq->ring, offset);
+}
+
+static inline void
+assert_ring_tail_valid(const struct intel_ring *ring, unsigned int tail)
+{
+	GEM_BUG_ON(!intel_ring_offset_valid(ring, tail));
+
+	/*
+	 * "Ring Buffer Use"
+	 *	Gen2 BSpec "1. Programming Environment" / 1.4.4.6
+	 *	Gen3 BSpec "1c Memory Interface Functions" / 2.3.4.5
+	 *	Gen4+ BSpec "1c Memory Interface and Command Stream" / 5.3.4.5
+	 * "If the Ring Buffer Head Pointer and the Tail Pointer are on the
+	 * same cacheline, the Head Pointer must not be greater than the Tail
+	 * Pointer."
+	 *
+	 * We use ring->head as the last known location of the actual RING_HEAD,
+	 * it may have advanced but in the worst case it is equally the same
+	 * as ring->head and so we should never program RING_TAIL to advance
+	 * into the same cacheline as ring->head.
+	 */
+#define cacheline(a) round_down(a, CACHELINE_BYTES)
+	GEM_BUG_ON(cacheline(tail) == cacheline(ring->head) &&
+		   tail < ring->head);
+#undef cacheline
+}
+
+static inline unsigned int
+intel_ring_set_tail(struct intel_ring *ring, unsigned int tail)
+{
+	/* Whilst writes to the tail are strictly order, there is no
+	 * serialisation between readers and the writers. The tail may be
+	 * read by i915_request_retire() just as it is being updated
+	 * by execlists, as although the breadcrumb is complete, the context
+	 * switch hasn't been seen.
+	 */
+	assert_ring_tail_valid(ring, tail);
+	ring->tail = tail;
+	return tail;
+}
+
+static inline unsigned int
+__intel_ring_space(unsigned int head, unsigned int tail, unsigned int size)
+{
+	/*
+	 * "If the Ring Buffer Head Pointer and the Tail Pointer are on the
+	 * same cacheline, the Head Pointer must not be greater than the Tail
+	 * Pointer."
+	 */
+	GEM_BUG_ON(!is_power_of_2(size));
+	return (head - tail - CACHELINE_BYTES) & (size - 1);
+}
+
+#endif /* INTEL_RING_H */

drivers/gpu/drm/i915/gt/intel_ring_submission.c

@@ -40,6 +40,7 @@
 #include "intel_gt_irq.h"
 #include "intel_gt_pm_irq.h"
 #include "intel_reset.h"
+#include "intel_ring.h"
 #include "intel_workarounds.h"
 
 /* Rough estimate of the typical request size, performing a flush,
@@ -47,16 +48,6 @@
  */
 #define LEGACY_REQUEST_SIZE 200
 
-unsigned int intel_ring_update_space(struct intel_ring *ring)
-{
-	unsigned int space;
-
-	space = __intel_ring_space(ring->head, ring->emit, ring->size);
-
-	ring->space = space;
-	return space;
-}
-
 static int
 gen2_render_ring_flush(struct i915_request *rq, u32 mode)
 {
@@ -1186,162 +1177,6 @@ i915_emit_bb_start(struct i915_request *rq,
 	return 0;
 }
 
-int intel_ring_pin(struct intel_ring *ring)
-{
-	struct i915_vma *vma = ring->vma;
-	unsigned int flags;
-	void *addr;
-	int ret;
-
-	if (atomic_fetch_inc(&ring->pin_count))
-		return 0;
-
-	flags = PIN_GLOBAL;
-
-	/* Ring wraparound at offset 0 sometimes hangs. No idea why. */
-	flags |= PIN_OFFSET_BIAS | i915_ggtt_pin_bias(vma);
-
-	if (vma->obj->stolen)
-		flags |= PIN_MAPPABLE;
-	else
-		flags |= PIN_HIGH;
-
-	ret = i915_vma_pin(vma, 0, 0, flags);
-	if (unlikely(ret))
-		goto err_unpin;
-
-	if (i915_vma_is_map_and_fenceable(vma))
-		addr = (void __force *)i915_vma_pin_iomap(vma);
-	else
-		addr = i915_gem_object_pin_map(vma->obj,
-					       i915_coherent_map_type(vma->vm->i915));
-	if (IS_ERR(addr)) {
-		ret = PTR_ERR(addr);
-		goto err_ring;
-	}
-
-	i915_vma_make_unshrinkable(vma);
-
-	GEM_BUG_ON(ring->vaddr);
-	ring->vaddr = addr;
-
-	return 0;
-
-err_ring:
-	i915_vma_unpin(vma);
-err_unpin:
-	atomic_dec(&ring->pin_count);
-	return ret;
-}
-
-void intel_ring_reset(struct intel_ring *ring, u32 tail)
-{
-	tail = intel_ring_wrap(ring, tail);
-	ring->tail = tail;
-	ring->head = tail;
-	ring->emit = tail;
-	intel_ring_update_space(ring);
-}
-
-void intel_ring_unpin(struct intel_ring *ring)
-{
-	struct i915_vma *vma = ring->vma;
-
-	if (!atomic_dec_and_test(&ring->pin_count))
-		return;
-
-	/* Discard any unused bytes beyond that submitted to hw. */
-	intel_ring_reset(ring, ring->emit);
-
-	i915_vma_unset_ggtt_write(vma);
-	if (i915_vma_is_map_and_fenceable(vma))
-		i915_vma_unpin_iomap(vma);
-	else
-		i915_gem_object_unpin_map(vma->obj);
-
-	GEM_BUG_ON(!ring->vaddr);
-	ring->vaddr = NULL;
-
-	i915_vma_unpin(vma);
-	i915_vma_make_purgeable(vma);
-}
-
-static struct i915_vma *create_ring_vma(struct i915_ggtt *ggtt, int size)
-{
-	struct i915_address_space *vm = &ggtt->vm;
-	struct drm_i915_private *i915 = vm->i915;
-	struct drm_i915_gem_object *obj;
-	struct i915_vma *vma;
-
-	obj = i915_gem_object_create_stolen(i915, size);
-	if (IS_ERR(obj))
-		obj = i915_gem_object_create_internal(i915, size);
-	if (IS_ERR(obj))
-		return ERR_CAST(obj);
-
-	/*
-	 * Mark ring buffers as read-only from GPU side (so no stray overwrites)
-	 * if supported by the platform's GGTT.
-	 */
-	if (vm->has_read_only)
-		i915_gem_object_set_readonly(obj);
-
-	vma = i915_vma_instance(obj, vm, NULL);
-	if (IS_ERR(vma))
-		goto err;
-
-	return vma;
-
-err:
-	i915_gem_object_put(obj);
-	return vma;
-}
-
-struct intel_ring *
-intel_engine_create_ring(struct intel_engine_cs *engine, int size)
-{
-	struct drm_i915_private *i915 = engine->i915;
-	struct intel_ring *ring;
-	struct i915_vma *vma;
-
-	GEM_BUG_ON(!is_power_of_2(size));
-	GEM_BUG_ON(RING_CTL_SIZE(size) & ~RING_NR_PAGES);
-
-	ring = kzalloc(sizeof(*ring), GFP_KERNEL);
-	if (!ring)
-		return ERR_PTR(-ENOMEM);
-
-	kref_init(&ring->ref);
-
-	ring->size = size;
-	/* Workaround an erratum on the i830 which causes a hang if
-	 * the TAIL pointer points to within the last 2 cachelines
-	 * of the buffer.
-	 */
-	ring->effective_size = size;
-	if (IS_I830(i915) || IS_I845G(i915))
-		ring->effective_size -= 2 * CACHELINE_BYTES;
-
-	intel_ring_update_space(ring);
-
-	vma = create_ring_vma(engine->gt->ggtt, size);
-	if (IS_ERR(vma)) {
-		kfree(ring);
-		return ERR_CAST(vma);
-	}
-	ring->vma = vma;
-
-	return ring;
-}
-
-void intel_ring_free(struct kref *ref)
-{
-	struct intel_ring *ring = container_of(ref, typeof(*ring), ref);
-
-	i915_vma_put(ring->vma);
-	kfree(ring);
-}
-
 static void __ring_context_fini(struct intel_context *ce)
 {
 	i915_vma_put(ce->state);
@@ -1836,148 +1671,6 @@ static int ring_request_alloc(struct i915_request *request)
 	return 0;
 }
 
-static noinline int
-wait_for_space(struct intel_ring *ring,
-	       struct intel_timeline *tl,
-	       unsigned int bytes)
-{
-	struct i915_request *target;
-	long timeout;
-
-	if (intel_ring_update_space(ring) >= bytes)
-		return 0;
-
-	GEM_BUG_ON(list_empty(&tl->requests));
-	list_for_each_entry(target, &tl->requests, link) {
-		if (target->ring != ring)
-			continue;
-
-		/* Would completion of this request free enough space? */
-		if (bytes <= __intel_ring_space(target->postfix,
-						ring->emit, ring->size))
-			break;
-	}
-
-	if (GEM_WARN_ON(&target->link == &tl->requests))
-		return -ENOSPC;
-
-	timeout = i915_request_wait(target,
-				    I915_WAIT_INTERRUPTIBLE,
-				    MAX_SCHEDULE_TIMEOUT);
-	if (timeout < 0)
-		return timeout;
-
-	i915_request_retire_upto(target);
-
-	intel_ring_update_space(ring);
-	GEM_BUG_ON(ring->space < bytes);
-	return 0;
-}
-
-u32 *intel_ring_begin(struct i915_request *rq, unsigned int num_dwords)
-{
-	struct intel_ring *ring = rq->ring;
-	const unsigned int remain_usable = ring->effective_size - ring->emit;
-	const unsigned int bytes = num_dwords * sizeof(u32);
-	unsigned int need_wrap = 0;
-	unsigned int total_bytes;
-	u32 *cs;
-
-	/* Packets must be qword aligned. */
-	GEM_BUG_ON(num_dwords & 1);
-
-	total_bytes = bytes + rq->reserved_space;
-	GEM_BUG_ON(total_bytes > ring->effective_size);
-
-	if (unlikely(total_bytes > remain_usable)) {
-		const int remain_actual = ring->size - ring->emit;
-
-		if (bytes > remain_usable) {
-			/*
-			 * Not enough space for the basic request. So need to
-			 * flush out the remainder and then wait for
-			 * base + reserved.
-			 */
-			total_bytes += remain_actual;
-			need_wrap = remain_actual | 1;
-		} else  {
-			/*
-			 * The base request will fit but the reserved space
-			 * falls off the end. So we don't need an immediate
-			 * wrap and only need to effectively wait for the
-			 * reserved size from the start of ringbuffer.
-			 */
-			total_bytes = rq->reserved_space + remain_actual;
-		}
-	}
-
-	if (unlikely(total_bytes > ring->space)) {
-		int ret;
-
-		/*
-		 * Space is reserved in the ringbuffer for finalising the
-		 * request, as that cannot be allowed to fail. During request
-		 * finalisation, reserved_space is set to 0 to stop the
-		 * overallocation and the assumption is that then we never need
-		 * to wait (which has the risk of failing with EINTR).
-		 *
-		 * See also i915_request_alloc() and i915_request_add().
-		 */
-		GEM_BUG_ON(!rq->reserved_space);
-
-		ret = wait_for_space(ring,
-				     i915_request_timeline(rq),
-				     total_bytes);
-		if (unlikely(ret))
-			return ERR_PTR(ret);
-	}
-
-	if (unlikely(need_wrap)) {
-		need_wrap &= ~1;
-		GEM_BUG_ON(need_wrap > ring->space);
-		GEM_BUG_ON(ring->emit + need_wrap > ring->size);
-		GEM_BUG_ON(!IS_ALIGNED(need_wrap, sizeof(u64)));
-
-		/* Fill the tail with MI_NOOP */
-		memset64(ring->vaddr + ring->emit, 0, need_wrap / sizeof(u64));
-		ring->space -= need_wrap;
-		ring->emit = 0;
-	}
-
-	GEM_BUG_ON(ring->emit > ring->size - bytes);
-	GEM_BUG_ON(ring->space < bytes);
-	cs = ring->vaddr + ring->emit;
-	GEM_DEBUG_EXEC(memset32(cs, POISON_INUSE, bytes / sizeof(*cs)));
-	ring->emit += bytes;
-	ring->space -= bytes;
-
-	return cs;
-}
-
-/* Align the ring tail to a cacheline boundary */
-int intel_ring_cacheline_align(struct i915_request *rq)
-{
-	int num_dwords;
-	void *cs;
-
-	num_dwords = (rq->ring->emit & (CACHELINE_BYTES - 1)) / sizeof(u32);
-	if (num_dwords == 0)
-		return 0;
-
-	num_dwords = CACHELINE_DWORDS - num_dwords;
-	GEM_BUG_ON(num_dwords & 1);
-
-	cs = intel_ring_begin(rq, num_dwords);
-	if (IS_ERR(cs))
-		return PTR_ERR(cs);
-
-	memset64(cs, (u64)MI_NOOP << 32 | MI_NOOP, num_dwords / 2);
-	intel_ring_advance(rq, cs);
-
-	GEM_BUG_ON(rq->ring->emit & (CACHELINE_BYTES - 1));
-	return 0;
-}
-
 static void gen6_bsd_submit_request(struct i915_request *request)
 {
 	struct intel_uncore *uncore = request->engine->uncore;

drivers/gpu/drm/i915/gt/intel_ring_types.h

@@ -0,0 +1,51 @@
+/*
+ * SPDX-License-Identifier: MIT
+ *
+ * Copyright © 2019 Intel Corporation
+ */
+
+#ifndef INTEL_RING_TYPES_H
+#define INTEL_RING_TYPES_H
+
+#include <linux/atomic.h>
+#include <linux/kref.h>
+#include <linux/types.h>
+
+/*
+ * Early gen2 devices have a cacheline of just 32 bytes, using 64 is overkill,
+ * but keeps the logic simple. Indeed, the whole purpose of this macro is just
+ * to give some inclination as to some of the magic values used in the various
+ * workarounds!
+ */
+#define CACHELINE_BYTES 64
+#define CACHELINE_DWORDS (CACHELINE_BYTES / sizeof(u32))
+
+struct i915_vma;
+
+struct intel_ring {
+	struct kref ref;
+	struct i915_vma *vma;
+	void *vaddr;
+
+	/*
+	 * As we have two types of rings, one global to the engine used
+	 * by ringbuffer submission and those that are exclusive to a
+	 * context used by execlists, we have to play safe and allow
+	 * atomic updates to the pin_count. However, the actual pinning
+	 * of the context is either done during initialisation for
+	 * ringbuffer submission or serialised as part of the context
+	 * pinning for execlists, and so we do not need a mutex ourselves
+	 * to serialise intel_ring_pin/intel_ring_unpin.
+	 */
+	atomic_t pin_count;
+
+	u32 head;
+	u32 tail;
+	u32 emit;
+
+	u32 space;
+	u32 size;
+	u32 effective_size;
+};
+
+#endif /* INTEL_RING_TYPES_H */

drivers/gpu/drm/i915/gt/intel_timeline.c

@@ -4,13 +4,13 @@
  * Copyright © 2016-2018 Intel Corporation
  */
 
-#include "gt/intel_gt_types.h"
-
 #include "i915_drv.h"
 
 #include "i915_active.h"
 #include "i915_syncmap.h"
-#include "gt/intel_timeline.h"
+#include "intel_gt.h"
+#include "intel_ring.h"
+#include "intel_timeline.h"
 
 #define ptr_set_bit(ptr, bit) ((typeof(ptr))((unsigned long)(ptr) | BIT(bit)))
 #define ptr_test_bit(ptr, bit) ((unsigned long)(ptr) & BIT(bit))

drivers/gpu/drm/i915/gt/intel_workarounds.c

@@ -7,6 +7,7 @@
 #include "i915_drv.h"
 #include "intel_context.h"
 #include "intel_gt.h"
+#include "intel_ring.h"
 #include "intel_workarounds.h"
 
 /**

drivers/gpu/drm/i915/gt/mock_engine.c

@@ -23,6 +23,7 @@
  */
 
 #include "gem/i915_gem_context.h"
+#include "gt/intel_ring.h"
 
 #include "i915_drv.h"
 #include "intel_context.h"

drivers/gpu/drm/i915/gt/selftest_timeline.c

@@ -9,6 +9,7 @@
 #include "intel_engine_pm.h"
 #include "intel_gt.h"
 #include "intel_gt_requests.h"
+#include "intel_ring.h"
 
 #include "../selftests/i915_random.h"
 #include "../i915_selftest.h"

drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c

@@ -6,12 +6,13 @@
 #include <linux/circ_buf.h>
 
 #include "gem/i915_gem_context.h"
-
 #include "gt/intel_context.h"
 #include "gt/intel_engine_pm.h"
 #include "gt/intel_gt.h"
 #include "gt/intel_gt_pm.h"
 #include "gt/intel_lrc_reg.h"
+#include "gt/intel_ring.h"
+
 #include "intel_guc_submission.h"
 
 #include "i915_drv.h"

drivers/gpu/drm/i915/gvt/cmd_parser.c

@@ -35,7 +35,9 @@
  */
 
 #include <linux/slab.h>
+
 #include "i915_drv.h"
+#include "gt/intel_ring.h"
 #include "gvt.h"
 #include "i915_pvinfo.h"
 #include "trace.h"

drivers/gpu/drm/i915/gvt/mmio_context.c

@@ -35,6 +35,7 @@
 
 #include "i915_drv.h"
 #include "gt/intel_context.h"
+#include "gt/intel_ring.h"
 #include "gvt.h"
 #include "trace.h"
 

drivers/gpu/drm/i915/gvt/scheduler.c

@@ -38,6 +38,7 @@
 #include "gem/i915_gem_context.h"
 #include "gem/i915_gem_pm.h"
 #include "gt/intel_context.h"
+#include "gt/intel_ring.h"
 
 #include "i915_drv.h"
 #include "gvt.h"

drivers/gpu/drm/i915/i915_active.c

@@ -7,6 +7,7 @@
 #include <linux/debugobjects.h>
 
 #include "gt/intel_engine_pm.h"
+#include "gt/intel_ring.h"
 
 #include "i915_drv.h"
 #include "i915_active.h"

drivers/gpu/drm/i915/i915_perf.c

@@ -200,6 +200,7 @@
 #include "gt/intel_engine_user.h"
 #include "gt/intel_gt.h"
 #include "gt/intel_lrc_reg.h"
+#include "gt/intel_ring.h"
 
 #include "i915_drv.h"
 #include "i915_perf.h"

drivers/gpu/drm/i915/i915_request.c

@@ -31,6 +31,7 @@
 
 #include "gem/i915_gem_context.h"
 #include "gt/intel_context.h"
+#include "gt/intel_ring.h"
 
 #include "i915_active.h"
 #include "i915_drv.h"