xref: /linux/drivers/gpu/drm/i915/gt/intel_ring.h (revision 4b132aacb0768ac1e652cf517097ea6f237214b9)
1 /* SPDX-License-Identifier: MIT */
2 /*
3  * Copyright © 2019 Intel Corporation
4  */
5 
6 #ifndef INTEL_RING_H
7 #define INTEL_RING_H
8 
9 #include "i915_gem.h" /* GEM_BUG_ON */
10 #include "i915_request.h"
11 #include "intel_ring_types.h"
12 
13 struct intel_engine_cs;
14 
15 struct intel_ring *
16 intel_engine_create_ring(struct intel_engine_cs *engine, int size);
17 
18 u32 *intel_ring_begin(struct i915_request *rq, unsigned int num_dwords);
19 int intel_ring_cacheline_align(struct i915_request *rq);
20 
21 unsigned int intel_ring_update_space(struct intel_ring *ring);
22 
23 void __intel_ring_pin(struct intel_ring *ring);
24 int intel_ring_pin(struct intel_ring *ring, struct i915_gem_ww_ctx *ww);
25 void intel_ring_unpin(struct intel_ring *ring);
26 void intel_ring_reset(struct intel_ring *ring, u32 tail);
27 
28 void intel_ring_free(struct kref *ref);
29 
30 static inline struct intel_ring *intel_ring_get(struct intel_ring *ring)
31 {
32 	kref_get(&ring->ref);
33 	return ring;
34 }
35 
36 static inline void intel_ring_put(struct intel_ring *ring)
37 {
38 	kref_put(&ring->ref, intel_ring_free);
39 }
40 
41 static inline void intel_ring_advance(struct i915_request *rq, u32 *cs)
42 {
43 	/* Dummy function.
44 	 *
45 	 * This serves as a placeholder in the code so that the reader
46 	 * can compare against the preceding intel_ring_begin() and
47 	 * check that the number of dwords emitted matches the space
48 	 * reserved for the command packet (i.e. the value passed to
49 	 * intel_ring_begin()).
50 	 */
51 	GEM_BUG_ON((rq->ring->vaddr + rq->ring->emit) != cs);
52 	GEM_BUG_ON(!IS_ALIGNED(rq->ring->emit, 8)); /* RING_TAIL qword align */
53 }
54 
55 static inline u32 intel_ring_wrap(const struct intel_ring *ring, u32 pos)
56 {
57 	return pos & (ring->size - 1);
58 }
59 
60 static inline int intel_ring_direction(const struct intel_ring *ring,
61 				       u32 next, u32 prev)
62 {
63 	typecheck(typeof(ring->size), next);
64 	typecheck(typeof(ring->size), prev);
65 	return (next - prev) << ring->wrap;
66 }
67 
68 static inline bool
69 intel_ring_offset_valid(const struct intel_ring *ring,
70 			unsigned int pos)
71 {
72 	if (pos & -ring->size) /* must be strictly within the ring */
73 		return false;
74 
75 	if (!IS_ALIGNED(pos, 8)) /* must be qword aligned */
76 		return false;
77 
78 	return true;
79 }
80 
81 static inline u32 intel_ring_offset(const struct i915_request *rq, void *addr)
82 {
83 	/* Don't write ring->size (equivalent to 0) as that hangs some GPUs. */
84 	u32 offset = addr - rq->ring->vaddr;
85 
86 	GEM_BUG_ON(offset > rq->ring->size);
87 	return intel_ring_wrap(rq->ring, offset);
88 }
89 
90 static inline void
91 assert_ring_tail_valid(const struct intel_ring *ring, unsigned int tail)
92 {
93 	unsigned int head = READ_ONCE(ring->head);
94 
95 	GEM_BUG_ON(!intel_ring_offset_valid(ring, tail));
96 
97 	/*
98 	 * "Ring Buffer Use"
99 	 *	Gen2 BSpec "1. Programming Environment" / 1.4.4.6
100 	 *	Gen3 BSpec "1c Memory Interface Functions" / 2.3.4.5
101 	 *	Gen4+ BSpec "1c Memory Interface and Command Stream" / 5.3.4.5
102 	 * "If the Ring Buffer Head Pointer and the Tail Pointer are on the
103 	 * same cacheline, the Head Pointer must not be greater than the Tail
104 	 * Pointer."
105 	 *
106 	 * We use ring->head as the last known location of the actual RING_HEAD,
107 	 * it may have advanced but in the worst case it is equally the same
108 	 * as ring->head and so we should never program RING_TAIL to advance
109 	 * into the same cacheline as ring->head.
110 	 */
111 #define cacheline(a) round_down(a, CACHELINE_BYTES)
112 	GEM_BUG_ON(cacheline(tail) == cacheline(head) && tail < head);
113 #undef cacheline
114 }
115 
116 static inline unsigned int
117 intel_ring_set_tail(struct intel_ring *ring, unsigned int tail)
118 {
119 	/* Whilst writes to the tail are strictly order, there is no
120 	 * serialisation between readers and the writers. The tail may be
121 	 * read by i915_request_retire() just as it is being updated
122 	 * by execlists, as although the breadcrumb is complete, the context
123 	 * switch hasn't been seen.
124 	 */
125 	assert_ring_tail_valid(ring, tail);
126 	ring->tail = tail;
127 	return tail;
128 }
129 
130 static inline unsigned int
131 __intel_ring_space(unsigned int head, unsigned int tail, unsigned int size)
132 {
133 	/*
134 	 * "If the Ring Buffer Head Pointer and the Tail Pointer are on the
135 	 * same cacheline, the Head Pointer must not be greater than the Tail
136 	 * Pointer."
137 	 */
138 	GEM_BUG_ON(!is_power_of_2(size));
139 	return (head - tail - CACHELINE_BYTES) & (size - 1);
140 }
141 
142 #endif /* INTEL_RING_H */
143