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_gt.h"
11 #include "gt/intel_gt_pm.h"
12 #include "gt/intel_ring.h"
13 
14 #include "i915_selftest.h"
15 #include "selftests/i915_random.h"
16 
17 struct context {
18 	struct drm_i915_gem_object *obj;
19 	struct intel_engine_cs *engine;
20 };
21 
22 static int cpu_set(struct context *ctx, unsigned long offset, u32 v)
23 {
24 	unsigned int needs_clflush;
25 	struct page *page;
26 	void *map;
27 	u32 *cpu;
28 	int err;
29 
30 	err = i915_gem_object_prepare_write(ctx->obj, &needs_clflush);
31 	if (err)
32 		return err;
33 
34 	page = i915_gem_object_get_page(ctx->obj, offset >> PAGE_SHIFT);
35 	map = kmap_atomic(page);
36 	cpu = map + offset_in_page(offset);
37 
38 	if (needs_clflush & CLFLUSH_BEFORE)
39 		drm_clflush_virt_range(cpu, sizeof(*cpu));
40 
41 	*cpu = v;
42 
43 	if (needs_clflush & CLFLUSH_AFTER)
44 		drm_clflush_virt_range(cpu, sizeof(*cpu));
45 
46 	kunmap_atomic(map);
47 	i915_gem_object_finish_access(ctx->obj);
48 
49 	return 0;
50 }
51 
52 static int cpu_get(struct context *ctx, unsigned long offset, u32 *v)
53 {
54 	unsigned int needs_clflush;
55 	struct page *page;
56 	void *map;
57 	u32 *cpu;
58 	int err;
59 
60 	err = i915_gem_object_prepare_read(ctx->obj, &needs_clflush);
61 	if (err)
62 		return err;
63 
64 	page = i915_gem_object_get_page(ctx->obj, offset >> PAGE_SHIFT);
65 	map = kmap_atomic(page);
66 	cpu = map + offset_in_page(offset);
67 
68 	if (needs_clflush & CLFLUSH_BEFORE)
69 		drm_clflush_virt_range(cpu, sizeof(*cpu));
70 
71 	*v = *cpu;
72 
73 	kunmap_atomic(map);
74 	i915_gem_object_finish_access(ctx->obj);
75 
76 	return 0;
77 }
78 
79 static int gtt_set(struct context *ctx, unsigned long offset, u32 v)
80 {
81 	struct i915_vma *vma;
82 	u32 __iomem *map;
83 	int err = 0;
84 
85 	i915_gem_object_lock(ctx->obj);
86 	err = i915_gem_object_set_to_gtt_domain(ctx->obj, true);
87 	i915_gem_object_unlock(ctx->obj);
88 	if (err)
89 		return err;
90 
91 	vma = i915_gem_object_ggtt_pin(ctx->obj, NULL, 0, 0, PIN_MAPPABLE);
92 	if (IS_ERR(vma))
93 		return PTR_ERR(vma);
94 
95 	intel_gt_pm_get(vma->vm->gt);
96 
97 	map = i915_vma_pin_iomap(vma);
98 	i915_vma_unpin(vma);
99 	if (IS_ERR(map)) {
100 		err = PTR_ERR(map);
101 		goto out_rpm;
102 	}
103 
104 	iowrite32(v, &map[offset / sizeof(*map)]);
105 	i915_vma_unpin_iomap(vma);
106 
107 out_rpm:
108 	intel_gt_pm_put(vma->vm->gt);
109 	return err;
110 }
111 
112 static int gtt_get(struct context *ctx, unsigned long offset, u32 *v)
113 {
114 	struct i915_vma *vma;
115 	u32 __iomem *map;
116 	int err = 0;
117 
118 	i915_gem_object_lock(ctx->obj);
119 	err = i915_gem_object_set_to_gtt_domain(ctx->obj, false);
120 	i915_gem_object_unlock(ctx->obj);
121 	if (err)
122 		return err;
123 
124 	vma = i915_gem_object_ggtt_pin(ctx->obj, NULL, 0, 0, PIN_MAPPABLE);
125 	if (IS_ERR(vma))
126 		return PTR_ERR(vma);
127 
128 	intel_gt_pm_get(vma->vm->gt);
129 
130 	map = i915_vma_pin_iomap(vma);
131 	i915_vma_unpin(vma);
132 	if (IS_ERR(map)) {
133 		err = PTR_ERR(map);
134 		goto out_rpm;
135 	}
136 
137 	*v = ioread32(&map[offset / sizeof(*map)]);
138 	i915_vma_unpin_iomap(vma);
139 
140 out_rpm:
141 	intel_gt_pm_put(vma->vm->gt);
142 	return err;
143 }
144 
145 static int wc_set(struct context *ctx, unsigned long offset, u32 v)
146 {
147 	u32 *map;
148 	int err;
149 
150 	i915_gem_object_lock(ctx->obj);
151 	err = i915_gem_object_set_to_wc_domain(ctx->obj, true);
152 	i915_gem_object_unlock(ctx->obj);
153 	if (err)
154 		return err;
155 
156 	map = i915_gem_object_pin_map(ctx->obj, I915_MAP_WC);
157 	if (IS_ERR(map))
158 		return PTR_ERR(map);
159 
160 	map[offset / sizeof(*map)] = v;
161 	i915_gem_object_unpin_map(ctx->obj);
162 
163 	return 0;
164 }
165 
166 static int wc_get(struct context *ctx, unsigned long offset, u32 *v)
167 {
168 	u32 *map;
169 	int err;
170 
171 	i915_gem_object_lock(ctx->obj);
172 	err = i915_gem_object_set_to_wc_domain(ctx->obj, false);
173 	i915_gem_object_unlock(ctx->obj);
174 	if (err)
175 		return err;
176 
177 	map = i915_gem_object_pin_map(ctx->obj, I915_MAP_WC);
178 	if (IS_ERR(map))
179 		return PTR_ERR(map);
180 
181 	*v = map[offset / sizeof(*map)];
182 	i915_gem_object_unpin_map(ctx->obj);
183 
184 	return 0;
185 }
186 
187 static int gpu_set(struct context *ctx, unsigned long offset, u32 v)
188 {
189 	struct i915_request *rq;
190 	struct i915_vma *vma;
191 	u32 *cs;
192 	int err;
193 
194 	i915_gem_object_lock(ctx->obj);
195 	err = i915_gem_object_set_to_gtt_domain(ctx->obj, true);
196 	i915_gem_object_unlock(ctx->obj);
197 	if (err)
198 		return err;
199 
200 	vma = i915_gem_object_ggtt_pin(ctx->obj, NULL, 0, 0, 0);
201 	if (IS_ERR(vma))
202 		return PTR_ERR(vma);
203 
204 	rq = intel_engine_create_kernel_request(ctx->engine);
205 	if (IS_ERR(rq)) {
206 		i915_vma_unpin(vma);
207 		return PTR_ERR(rq);
208 	}
209 
210 	cs = intel_ring_begin(rq, 4);
211 	if (IS_ERR(cs)) {
212 		i915_request_add(rq);
213 		i915_vma_unpin(vma);
214 		return PTR_ERR(cs);
215 	}
216 
217 	if (INTEL_GEN(ctx->engine->i915) >= 8) {
218 		*cs++ = MI_STORE_DWORD_IMM_GEN4 | 1 << 22;
219 		*cs++ = lower_32_bits(i915_ggtt_offset(vma) + offset);
220 		*cs++ = upper_32_bits(i915_ggtt_offset(vma) + offset);
221 		*cs++ = v;
222 	} else if (INTEL_GEN(ctx->engine->i915) >= 4) {
223 		*cs++ = MI_STORE_DWORD_IMM_GEN4 | MI_USE_GGTT;
224 		*cs++ = 0;
225 		*cs++ = i915_ggtt_offset(vma) + offset;
226 		*cs++ = v;
227 	} else {
228 		*cs++ = MI_STORE_DWORD_IMM | MI_MEM_VIRTUAL;
229 		*cs++ = i915_ggtt_offset(vma) + offset;
230 		*cs++ = v;
231 		*cs++ = MI_NOOP;
232 	}
233 	intel_ring_advance(rq, cs);
234 
235 	i915_vma_lock(vma);
236 	err = i915_request_await_object(rq, vma->obj, true);
237 	if (err == 0)
238 		err = i915_vma_move_to_active(vma, rq, EXEC_OBJECT_WRITE);
239 	i915_vma_unlock(vma);
240 	i915_vma_unpin(vma);
241 
242 	i915_request_add(rq);
243 
244 	return err;
245 }
246 
247 static bool always_valid(struct context *ctx)
248 {
249 	return true;
250 }
251 
252 static bool needs_fence_registers(struct context *ctx)
253 {
254 	struct intel_gt *gt = ctx->engine->gt;
255 
256 	if (intel_gt_is_wedged(gt))
257 		return false;
258 
259 	return gt->ggtt->num_fences;
260 }
261 
262 static bool needs_mi_store_dword(struct context *ctx)
263 {
264 	if (intel_gt_is_wedged(ctx->engine->gt))
265 		return false;
266 
267 	return intel_engine_can_store_dword(ctx->engine);
268 }
269 
270 static const struct igt_coherency_mode {
271 	const char *name;
272 	int (*set)(struct context *ctx, unsigned long offset, u32 v);
273 	int (*get)(struct context *ctx, unsigned long offset, u32 *v);
274 	bool (*valid)(struct context *ctx);
275 } igt_coherency_mode[] = {
276 	{ "cpu", cpu_set, cpu_get, always_valid },
277 	{ "gtt", gtt_set, gtt_get, needs_fence_registers },
278 	{ "wc", wc_set, wc_get, always_valid },
279 	{ "gpu", gpu_set, NULL, needs_mi_store_dword },
280 	{ },
281 };
282 
283 static struct intel_engine_cs *
284 random_engine(struct drm_i915_private *i915, struct rnd_state *prng)
285 {
286 	struct intel_engine_cs *engine;
287 	unsigned int count;
288 
289 	count = 0;
290 	for_each_uabi_engine(engine, i915)
291 		count++;
292 
293 	count = i915_prandom_u32_max_state(count, prng);
294 	for_each_uabi_engine(engine, i915)
295 		if (count-- == 0)
296 			return engine;
297 
298 	return NULL;
299 }
300 
301 static int igt_gem_coherency(void *arg)
302 {
303 	const unsigned int ncachelines = PAGE_SIZE/64;
304 	struct drm_i915_private *i915 = arg;
305 	const struct igt_coherency_mode *read, *write, *over;
306 	unsigned long count, n;
307 	u32 *offsets, *values;
308 	I915_RND_STATE(prng);
309 	struct context ctx;
310 	int err = 0;
311 
312 	/*
313 	 * We repeatedly write, overwrite and read from a sequence of
314 	 * cachelines in order to try and detect incoherency (unflushed writes
315 	 * from either the CPU or GPU). Each setter/getter uses our cache
316 	 * domain API which should prevent incoherency.
317 	 */
318 
319 	offsets = kmalloc_array(ncachelines, 2*sizeof(u32), GFP_KERNEL);
320 	if (!offsets)
321 		return -ENOMEM;
322 	for (count = 0; count < ncachelines; count++)
323 		offsets[count] = count * 64 + 4 * (count % 16);
324 
325 	values = offsets + ncachelines;
326 
327 	ctx.engine = random_engine(i915, &prng);
328 	if (!ctx.engine) {
329 		err = -ENODEV;
330 		goto out_free;
331 	}
332 	pr_info("%s: using %s\n", __func__, ctx.engine->name);
333 	intel_engine_pm_get(ctx.engine);
334 
335 	for (over = igt_coherency_mode; over->name; over++) {
336 		if (!over->set)
337 			continue;
338 
339 		if (!over->valid(&ctx))
340 			continue;
341 
342 		for (write = igt_coherency_mode; write->name; write++) {
343 			if (!write->set)
344 				continue;
345 
346 			if (!write->valid(&ctx))
347 				continue;
348 
349 			for (read = igt_coherency_mode; read->name; read++) {
350 				if (!read->get)
351 					continue;
352 
353 				if (!read->valid(&ctx))
354 					continue;
355 
356 				for_each_prime_number_from(count, 1, ncachelines) {
357 					ctx.obj = i915_gem_object_create_internal(i915, PAGE_SIZE);
358 					if (IS_ERR(ctx.obj)) {
359 						err = PTR_ERR(ctx.obj);
360 						goto out_pm;
361 					}
362 
363 					i915_random_reorder(offsets, ncachelines, &prng);
364 					for (n = 0; n < count; n++)
365 						values[n] = prandom_u32_state(&prng);
366 
367 					for (n = 0; n < count; n++) {
368 						err = over->set(&ctx, offsets[n], ~values[n]);
369 						if (err) {
370 							pr_err("Failed to set stale value[%ld/%ld] in object using %s, err=%d\n",
371 							       n, count, over->name, err);
372 							goto put_object;
373 						}
374 					}
375 
376 					for (n = 0; n < count; n++) {
377 						err = write->set(&ctx, offsets[n], values[n]);
378 						if (err) {
379 							pr_err("Failed to set value[%ld/%ld] in object using %s, err=%d\n",
380 							       n, count, write->name, err);
381 							goto put_object;
382 						}
383 					}
384 
385 					for (n = 0; n < count; n++) {
386 						u32 found;
387 
388 						err = read->get(&ctx, offsets[n], &found);
389 						if (err) {
390 							pr_err("Failed to get value[%ld/%ld] in object using %s, err=%d\n",
391 							       n, count, read->name, err);
392 							goto put_object;
393 						}
394 
395 						if (found != values[n]) {
396 							pr_err("Value[%ld/%ld] mismatch, (overwrite with %s) wrote [%s] %x read [%s] %x (inverse %x), at offset %x\n",
397 							       n, count, over->name,
398 							       write->name, values[n],
399 							       read->name, found,
400 							       ~values[n], offsets[n]);
401 							err = -EINVAL;
402 							goto put_object;
403 						}
404 					}
405 
406 					i915_gem_object_put(ctx.obj);
407 				}
408 			}
409 		}
410 	}
411 out_pm:
412 	intel_engine_pm_put(ctx.engine);
413 out_free:
414 	kfree(offsets);
415 	return err;
416 
417 put_object:
418 	i915_gem_object_put(ctx.obj);
419 	goto out_pm;
420 }
421 
422 int i915_gem_coherency_live_selftests(struct drm_i915_private *i915)
423 {
424 	static const struct i915_subtest tests[] = {
425 		SUBTEST(igt_gem_coherency),
426 	};
427 
428 	return i915_subtests(tests, i915);
429 }
430