1 /*
2 * SPDX-License-Identifier: MIT
3 *
4 * Copyright © 2017 Intel Corporation
5 */
6
7 #include <linux/prime_numbers.h>
8
9 #include "gt/intel_engine_pm.h"
10 #include "gt/intel_gpu_commands.h"
11 #include "gt/intel_gt.h"
12 #include "gt/intel_gt_pm.h"
13 #include "gt/intel_ring.h"
14
15 #include "i915_selftest.h"
16 #include "selftests/i915_random.h"
17
18 struct context {
19 struct drm_i915_gem_object *obj;
20 struct intel_engine_cs *engine;
21 };
22
cpu_set(struct context * ctx,unsigned long offset,u32 v)23 static int cpu_set(struct context *ctx, unsigned long offset, u32 v)
24 {
25 unsigned int needs_clflush;
26 struct page *page;
27 void *map;
28 u32 *cpu;
29 int err;
30
31 i915_gem_object_lock(ctx->obj, NULL);
32 err = i915_gem_object_prepare_write(ctx->obj, &needs_clflush);
33 if (err)
34 goto out;
35
36 page = i915_gem_object_get_page(ctx->obj, offset >> PAGE_SHIFT);
37 map = kmap_atomic(page);
38 cpu = map + offset_in_page(offset);
39
40 if (needs_clflush & CLFLUSH_BEFORE)
41 drm_clflush_virt_range(cpu, sizeof(*cpu));
42
43 *cpu = v;
44
45 if (needs_clflush & CLFLUSH_AFTER)
46 drm_clflush_virt_range(cpu, sizeof(*cpu));
47
48 kunmap_atomic(map);
49 i915_gem_object_finish_access(ctx->obj);
50
51 out:
52 i915_gem_object_unlock(ctx->obj);
53 return err;
54 }
55
cpu_get(struct context * ctx,unsigned long offset,u32 * v)56 static int cpu_get(struct context *ctx, unsigned long offset, u32 *v)
57 {
58 unsigned int needs_clflush;
59 struct page *page;
60 void *map;
61 u32 *cpu;
62 int err;
63
64 i915_gem_object_lock(ctx->obj, NULL);
65 err = i915_gem_object_prepare_read(ctx->obj, &needs_clflush);
66 if (err)
67 goto out;
68
69 page = i915_gem_object_get_page(ctx->obj, offset >> PAGE_SHIFT);
70 map = kmap_atomic(page);
71 cpu = map + offset_in_page(offset);
72
73 if (needs_clflush & CLFLUSH_BEFORE)
74 drm_clflush_virt_range(cpu, sizeof(*cpu));
75
76 *v = *cpu;
77
78 kunmap_atomic(map);
79 i915_gem_object_finish_access(ctx->obj);
80
81 out:
82 i915_gem_object_unlock(ctx->obj);
83 return err;
84 }
85
gtt_set(struct context * ctx,unsigned long offset,u32 v)86 static int gtt_set(struct context *ctx, unsigned long offset, u32 v)
87 {
88 struct i915_vma *vma;
89 u32 __iomem *map;
90 int err = 0;
91
92 i915_gem_object_lock(ctx->obj, NULL);
93 err = i915_gem_object_set_to_gtt_domain(ctx->obj, true);
94 i915_gem_object_unlock(ctx->obj);
95 if (err)
96 return err;
97
98 vma = i915_gem_object_ggtt_pin(ctx->obj, NULL, 0, 0, PIN_MAPPABLE);
99 if (IS_ERR(vma))
100 return PTR_ERR(vma);
101
102 intel_gt_pm_get(vma->vm->gt);
103
104 map = i915_vma_pin_iomap(vma);
105 i915_vma_unpin(vma);
106 if (IS_ERR(map)) {
107 err = PTR_ERR(map);
108 goto out_rpm;
109 }
110
111 iowrite32(v, &map[offset / sizeof(*map)]);
112 i915_vma_unpin_iomap(vma);
113
114 out_rpm:
115 intel_gt_pm_put(vma->vm->gt);
116 return err;
117 }
118
gtt_get(struct context * ctx,unsigned long offset,u32 * v)119 static int gtt_get(struct context *ctx, unsigned long offset, u32 *v)
120 {
121 struct i915_vma *vma;
122 u32 __iomem *map;
123 int err = 0;
124
125 i915_gem_object_lock(ctx->obj, NULL);
126 err = i915_gem_object_set_to_gtt_domain(ctx->obj, false);
127 i915_gem_object_unlock(ctx->obj);
128 if (err)
129 return err;
130
131 vma = i915_gem_object_ggtt_pin(ctx->obj, NULL, 0, 0, PIN_MAPPABLE);
132 if (IS_ERR(vma))
133 return PTR_ERR(vma);
134
135 intel_gt_pm_get(vma->vm->gt);
136
137 map = i915_vma_pin_iomap(vma);
138 i915_vma_unpin(vma);
139 if (IS_ERR(map)) {
140 err = PTR_ERR(map);
141 goto out_rpm;
142 }
143
144 *v = ioread32(&map[offset / sizeof(*map)]);
145 i915_vma_unpin_iomap(vma);
146
147 out_rpm:
148 intel_gt_pm_put(vma->vm->gt);
149 return err;
150 }
151
wc_set(struct context * ctx,unsigned long offset,u32 v)152 static int wc_set(struct context *ctx, unsigned long offset, u32 v)
153 {
154 u32 *map;
155 int err;
156
157 i915_gem_object_lock(ctx->obj, NULL);
158 err = i915_gem_object_set_to_wc_domain(ctx->obj, true);
159 i915_gem_object_unlock(ctx->obj);
160 if (err)
161 return err;
162
163 map = i915_gem_object_pin_map_unlocked(ctx->obj, I915_MAP_WC);
164 if (IS_ERR(map))
165 return PTR_ERR(map);
166
167 map[offset / sizeof(*map)] = v;
168
169 __i915_gem_object_flush_map(ctx->obj, offset, sizeof(*map));
170 i915_gem_object_unpin_map(ctx->obj);
171
172 return 0;
173 }
174
wc_get(struct context * ctx,unsigned long offset,u32 * v)175 static int wc_get(struct context *ctx, unsigned long offset, u32 *v)
176 {
177 u32 *map;
178 int err;
179
180 i915_gem_object_lock(ctx->obj, NULL);
181 err = i915_gem_object_set_to_wc_domain(ctx->obj, false);
182 i915_gem_object_unlock(ctx->obj);
183 if (err)
184 return err;
185
186 map = i915_gem_object_pin_map_unlocked(ctx->obj, I915_MAP_WC);
187 if (IS_ERR(map))
188 return PTR_ERR(map);
189
190 *v = map[offset / sizeof(*map)];
191 i915_gem_object_unpin_map(ctx->obj);
192
193 return 0;
194 }
195
gpu_set(struct context * ctx,unsigned long offset,u32 v)196 static int gpu_set(struct context *ctx, unsigned long offset, u32 v)
197 {
198 struct i915_request *rq;
199 struct i915_vma *vma;
200 u32 *cs;
201 int err;
202
203 vma = i915_gem_object_ggtt_pin(ctx->obj, NULL, 0, 0, 0);
204 if (IS_ERR(vma))
205 return PTR_ERR(vma);
206
207 i915_gem_object_lock(ctx->obj, NULL);
208 err = i915_gem_object_set_to_gtt_domain(ctx->obj, true);
209 if (err)
210 goto out_unlock;
211
212 rq = intel_engine_create_kernel_request(ctx->engine);
213 if (IS_ERR(rq)) {
214 err = PTR_ERR(rq);
215 goto out_unpin;
216 }
217
218 cs = intel_ring_begin(rq, 4);
219 if (IS_ERR(cs)) {
220 err = PTR_ERR(cs);
221 goto out_rq;
222 }
223
224 if (GRAPHICS_VER(ctx->engine->i915) >= 8) {
225 *cs++ = MI_STORE_DWORD_IMM_GEN4 | 1 << 22;
226 *cs++ = lower_32_bits(i915_ggtt_offset(vma) + offset);
227 *cs++ = upper_32_bits(i915_ggtt_offset(vma) + offset);
228 *cs++ = v;
229 } else if (GRAPHICS_VER(ctx->engine->i915) >= 4) {
230 *cs++ = MI_STORE_DWORD_IMM_GEN4 | MI_USE_GGTT;
231 *cs++ = 0;
232 *cs++ = i915_ggtt_offset(vma) + offset;
233 *cs++ = v;
234 } else {
235 *cs++ = MI_STORE_DWORD_IMM | MI_MEM_VIRTUAL;
236 *cs++ = i915_ggtt_offset(vma) + offset;
237 *cs++ = v;
238 *cs++ = MI_NOOP;
239 }
240 intel_ring_advance(rq, cs);
241
242 err = i915_request_await_object(rq, vma->obj, true);
243 if (err == 0)
244 err = i915_vma_move_to_active(vma, rq, EXEC_OBJECT_WRITE);
245
246 out_rq:
247 i915_request_add(rq);
248 out_unpin:
249 i915_vma_unpin(vma);
250 out_unlock:
251 i915_gem_object_unlock(ctx->obj);
252
253 return err;
254 }
255
always_valid(struct context * ctx)256 static bool always_valid(struct context *ctx)
257 {
258 return true;
259 }
260
needs_fence_registers(struct context * ctx)261 static bool needs_fence_registers(struct context *ctx)
262 {
263 struct intel_gt *gt = ctx->engine->gt;
264
265 if (intel_gt_is_wedged(gt))
266 return false;
267
268 return gt->ggtt->num_fences;
269 }
270
needs_mi_store_dword(struct context * ctx)271 static bool needs_mi_store_dword(struct context *ctx)
272 {
273 if (intel_gt_is_wedged(ctx->engine->gt))
274 return false;
275
276 return intel_engine_can_store_dword(ctx->engine);
277 }
278
279 static const struct igt_coherency_mode {
280 const char *name;
281 int (*set)(struct context *ctx, unsigned long offset, u32 v);
282 int (*get)(struct context *ctx, unsigned long offset, u32 *v);
283 bool (*valid)(struct context *ctx);
284 } igt_coherency_mode[] = {
285 { "cpu", cpu_set, cpu_get, always_valid },
286 { "gtt", gtt_set, gtt_get, needs_fence_registers },
287 { "wc", wc_set, wc_get, always_valid },
288 { "gpu", gpu_set, NULL, needs_mi_store_dword },
289 { },
290 };
291
292 static struct intel_engine_cs *
random_engine(struct drm_i915_private * i915,struct rnd_state * prng)293 random_engine(struct drm_i915_private *i915, struct rnd_state *prng)
294 {
295 struct intel_engine_cs *engine;
296 unsigned int count;
297
298 count = 0;
299 for_each_uabi_engine(engine, i915)
300 count++;
301
302 count = i915_prandom_u32_max_state(count, prng);
303 for_each_uabi_engine(engine, i915)
304 if (count-- == 0)
305 return engine;
306
307 return NULL;
308 }
309
igt_gem_coherency(void * arg)310 static int igt_gem_coherency(void *arg)
311 {
312 const unsigned int ncachelines = PAGE_SIZE/64;
313 struct drm_i915_private *i915 = arg;
314 const struct igt_coherency_mode *read, *write, *over;
315 unsigned long count, n;
316 u32 *offsets, *values;
317 I915_RND_STATE(prng);
318 struct context ctx;
319 int err = 0;
320
321 /*
322 * We repeatedly write, overwrite and read from a sequence of
323 * cachelines in order to try and detect incoherency (unflushed writes
324 * from either the CPU or GPU). Each setter/getter uses our cache
325 * domain API which should prevent incoherency.
326 */
327
328 offsets = kmalloc_array(ncachelines, 2*sizeof(u32), GFP_KERNEL);
329 if (!offsets)
330 return -ENOMEM;
331 for (count = 0; count < ncachelines; count++)
332 offsets[count] = count * 64 + 4 * (count % 16);
333
334 values = offsets + ncachelines;
335
336 ctx.engine = random_engine(i915, &prng);
337 if (!ctx.engine) {
338 err = -ENODEV;
339 goto out_free;
340 }
341 pr_info("%s: using %s\n", __func__, ctx.engine->name);
342 intel_engine_pm_get(ctx.engine);
343
344 for (over = igt_coherency_mode; over->name; over++) {
345 if (!over->set)
346 continue;
347
348 if (!over->valid(&ctx))
349 continue;
350
351 for (write = igt_coherency_mode; write->name; write++) {
352 if (!write->set)
353 continue;
354
355 if (!write->valid(&ctx))
356 continue;
357
358 for (read = igt_coherency_mode; read->name; read++) {
359 if (!read->get)
360 continue;
361
362 if (!read->valid(&ctx))
363 continue;
364
365 for_each_prime_number_from(count, 1, ncachelines) {
366 ctx.obj = i915_gem_object_create_internal(i915, PAGE_SIZE);
367 if (IS_ERR(ctx.obj)) {
368 err = PTR_ERR(ctx.obj);
369 goto out_pm;
370 }
371
372 i915_random_reorder(offsets, ncachelines, &prng);
373 for (n = 0; n < count; n++)
374 values[n] = prandom_u32_state(&prng);
375
376 for (n = 0; n < count; n++) {
377 err = over->set(&ctx, offsets[n], ~values[n]);
378 if (err) {
379 pr_err("Failed to set stale value[%ld/%ld] in object using %s, err=%d\n",
380 n, count, over->name, err);
381 goto put_object;
382 }
383 }
384
385 for (n = 0; n < count; n++) {
386 err = write->set(&ctx, offsets[n], values[n]);
387 if (err) {
388 pr_err("Failed to set value[%ld/%ld] in object using %s, err=%d\n",
389 n, count, write->name, err);
390 goto put_object;
391 }
392 }
393
394 for (n = 0; n < count; n++) {
395 u32 found;
396
397 err = read->get(&ctx, offsets[n], &found);
398 if (err) {
399 pr_err("Failed to get value[%ld/%ld] in object using %s, err=%d\n",
400 n, count, read->name, err);
401 goto put_object;
402 }
403
404 if (found != values[n]) {
405 pr_err("Value[%ld/%ld] mismatch, (overwrite with %s) wrote [%s] %x read [%s] %x (inverse %x), at offset %x\n",
406 n, count, over->name,
407 write->name, values[n],
408 read->name, found,
409 ~values[n], offsets[n]);
410 err = -EINVAL;
411 goto put_object;
412 }
413 }
414
415 i915_gem_object_put(ctx.obj);
416 }
417 }
418 }
419 }
420 out_pm:
421 intel_engine_pm_put(ctx.engine);
422 out_free:
423 kfree(offsets);
424 return err;
425
426 put_object:
427 i915_gem_object_put(ctx.obj);
428 goto out_pm;
429 }
430
i915_gem_coherency_live_selftests(struct drm_i915_private * i915)431 int i915_gem_coherency_live_selftests(struct drm_i915_private *i915)
432 {
433 static const struct i915_subtest tests[] = {
434 SUBTEST(igt_gem_coherency),
435 };
436
437 return i915_live_subtests(tests, i915);
438 }
439