1 /*
2 * SPDX-License-Identifier: MIT
3 *
4 * Copyright © 2018 Intel Corporation
5 */
6 #include "gt/intel_gt.h"
7
8 #include "gem/selftests/igt_gem_utils.h"
9
10 #include "igt_spinner.h"
11
igt_spinner_init(struct igt_spinner * spin,struct intel_gt * gt)12 int igt_spinner_init(struct igt_spinner *spin, struct intel_gt *gt)
13 {
14 unsigned int mode;
15 void *vaddr;
16 int err;
17
18 memset(spin, 0, sizeof(*spin));
19 spin->gt = gt;
20
21 spin->hws = i915_gem_object_create_internal(gt->i915, PAGE_SIZE);
22 if (IS_ERR(spin->hws)) {
23 err = PTR_ERR(spin->hws);
24 goto err;
25 }
26
27 spin->obj = i915_gem_object_create_internal(gt->i915, PAGE_SIZE);
28 if (IS_ERR(spin->obj)) {
29 err = PTR_ERR(spin->obj);
30 goto err_hws;
31 }
32
33 i915_gem_object_set_cache_coherency(spin->hws, I915_CACHE_LLC);
34 vaddr = i915_gem_object_pin_map(spin->hws, I915_MAP_WB);
35 if (IS_ERR(vaddr)) {
36 err = PTR_ERR(vaddr);
37 goto err_obj;
38 }
39 spin->seqno = memset(vaddr, 0xff, PAGE_SIZE);
40
41 mode = i915_coherent_map_type(gt->i915);
42 vaddr = i915_gem_object_pin_map(spin->obj, mode);
43 if (IS_ERR(vaddr)) {
44 err = PTR_ERR(vaddr);
45 goto err_unpin_hws;
46 }
47 spin->batch = vaddr;
48
49 return 0;
50
51 err_unpin_hws:
52 i915_gem_object_unpin_map(spin->hws);
53 err_obj:
54 i915_gem_object_put(spin->obj);
55 err_hws:
56 i915_gem_object_put(spin->hws);
57 err:
58 return err;
59 }
60
seqno_offset(u64 fence)61 static unsigned int seqno_offset(u64 fence)
62 {
63 return offset_in_page(sizeof(u32) * fence);
64 }
65
hws_address(const struct i915_vma * hws,const struct i915_request * rq)66 static u64 hws_address(const struct i915_vma *hws,
67 const struct i915_request *rq)
68 {
69 return hws->node.start + seqno_offset(rq->fence.context);
70 }
71
move_to_active(struct i915_vma * vma,struct i915_request * rq,unsigned int flags)72 static int move_to_active(struct i915_vma *vma,
73 struct i915_request *rq,
74 unsigned int flags)
75 {
76 int err;
77
78 i915_vma_lock(vma);
79 err = i915_request_await_object(rq, vma->obj,
80 flags & EXEC_OBJECT_WRITE);
81 if (err == 0)
82 err = i915_vma_move_to_active(vma, rq, flags);
83 i915_vma_unlock(vma);
84
85 return err;
86 }
87
88 struct i915_request *
igt_spinner_create_request(struct igt_spinner * spin,struct intel_context * ce,u32 arbitration_command)89 igt_spinner_create_request(struct igt_spinner *spin,
90 struct intel_context *ce,
91 u32 arbitration_command)
92 {
93 struct intel_engine_cs *engine = ce->engine;
94 struct i915_request *rq = NULL;
95 struct i915_vma *hws, *vma;
96 unsigned int flags;
97 u32 *batch;
98 int err;
99
100 GEM_BUG_ON(spin->gt != ce->vm->gt);
101
102 if (!intel_engine_can_store_dword(ce->engine))
103 return ERR_PTR(-ENODEV);
104
105 vma = i915_vma_instance(spin->obj, ce->vm, NULL);
106 if (IS_ERR(vma))
107 return ERR_CAST(vma);
108
109 hws = i915_vma_instance(spin->hws, ce->vm, NULL);
110 if (IS_ERR(hws))
111 return ERR_CAST(hws);
112
113 err = i915_vma_pin(vma, 0, 0, PIN_USER);
114 if (err)
115 return ERR_PTR(err);
116
117 err = i915_vma_pin(hws, 0, 0, PIN_USER);
118 if (err)
119 goto unpin_vma;
120
121 rq = intel_context_create_request(ce);
122 if (IS_ERR(rq)) {
123 err = PTR_ERR(rq);
124 goto unpin_hws;
125 }
126
127 err = move_to_active(vma, rq, 0);
128 if (err)
129 goto cancel_rq;
130
131 err = move_to_active(hws, rq, 0);
132 if (err)
133 goto cancel_rq;
134
135 batch = spin->batch;
136
137 if (INTEL_GEN(rq->engine->i915) >= 8) {
138 *batch++ = MI_STORE_DWORD_IMM_GEN4;
139 *batch++ = lower_32_bits(hws_address(hws, rq));
140 *batch++ = upper_32_bits(hws_address(hws, rq));
141 } else if (INTEL_GEN(rq->engine->i915) >= 6) {
142 *batch++ = MI_STORE_DWORD_IMM_GEN4;
143 *batch++ = 0;
144 *batch++ = hws_address(hws, rq);
145 } else if (INTEL_GEN(rq->engine->i915) >= 4) {
146 *batch++ = MI_STORE_DWORD_IMM_GEN4 | MI_USE_GGTT;
147 *batch++ = 0;
148 *batch++ = hws_address(hws, rq);
149 } else {
150 *batch++ = MI_STORE_DWORD_IMM | MI_MEM_VIRTUAL;
151 *batch++ = hws_address(hws, rq);
152 }
153 *batch++ = rq->fence.seqno;
154
155 *batch++ = arbitration_command;
156
157 if (INTEL_GEN(rq->engine->i915) >= 8)
158 *batch++ = MI_BATCH_BUFFER_START | BIT(8) | 1;
159 else if (IS_HASWELL(rq->engine->i915))
160 *batch++ = MI_BATCH_BUFFER_START | MI_BATCH_PPGTT_HSW;
161 else if (INTEL_GEN(rq->engine->i915) >= 6)
162 *batch++ = MI_BATCH_BUFFER_START;
163 else
164 *batch++ = MI_BATCH_BUFFER_START | MI_BATCH_GTT;
165 *batch++ = lower_32_bits(vma->node.start);
166 *batch++ = upper_32_bits(vma->node.start);
167
168 *batch++ = MI_BATCH_BUFFER_END; /* not reached */
169
170 intel_gt_chipset_flush(engine->gt);
171
172 if (engine->emit_init_breadcrumb) {
173 err = engine->emit_init_breadcrumb(rq);
174 if (err)
175 goto cancel_rq;
176 }
177
178 flags = 0;
179 if (INTEL_GEN(rq->engine->i915) <= 5)
180 flags |= I915_DISPATCH_SECURE;
181 err = engine->emit_bb_start(rq, vma->node.start, PAGE_SIZE, flags);
182
183 cancel_rq:
184 if (err) {
185 i915_request_set_error_once(rq, err);
186 i915_request_add(rq);
187 }
188 unpin_hws:
189 i915_vma_unpin(hws);
190 unpin_vma:
191 i915_vma_unpin(vma);
192 return err ? ERR_PTR(err) : rq;
193 }
194
195 static u32
hws_seqno(const struct igt_spinner * spin,const struct i915_request * rq)196 hws_seqno(const struct igt_spinner *spin, const struct i915_request *rq)
197 {
198 u32 *seqno = spin->seqno + seqno_offset(rq->fence.context);
199
200 return READ_ONCE(*seqno);
201 }
202
igt_spinner_end(struct igt_spinner * spin)203 void igt_spinner_end(struct igt_spinner *spin)
204 {
205 *spin->batch = MI_BATCH_BUFFER_END;
206 intel_gt_chipset_flush(spin->gt);
207 }
208
igt_spinner_fini(struct igt_spinner * spin)209 void igt_spinner_fini(struct igt_spinner *spin)
210 {
211 igt_spinner_end(spin);
212
213 i915_gem_object_unpin_map(spin->obj);
214 i915_gem_object_put(spin->obj);
215
216 i915_gem_object_unpin_map(spin->hws);
217 i915_gem_object_put(spin->hws);
218 }
219
igt_wait_for_spinner(struct igt_spinner * spin,struct i915_request * rq)220 bool igt_wait_for_spinner(struct igt_spinner *spin, struct i915_request *rq)
221 {
222 return !(wait_for_us(i915_seqno_passed(hws_seqno(spin, rq),
223 rq->fence.seqno),
224 100) &&
225 wait_for(i915_seqno_passed(hws_seqno(spin, rq),
226 rq->fence.seqno),
227 50));
228 }
229