1 /*
2  * Copyright 2009 VMware, Inc.
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice shall be included in
12  * all copies or substantial portions of the Software.
13  *
14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20  * OTHER DEALINGS IN THE SOFTWARE.
21  *
22  * Authors: Michel Dänzer
23  */
24 #include <drm/drmP.h>
25 #include <drm/radeon_drm.h>
26 #include "radeon_reg.h"
27 #include "radeon.h"
28 
29 #define RADEON_TEST_COPY_BLIT 1
30 #define RADEON_TEST_COPY_DMA  0
31 
32 
33 /* Test BO GTT->VRAM and VRAM->GTT GPU copies across the whole GTT aperture */
34 static void radeon_do_test_moves(struct radeon_device *rdev, int flag)
35 {
36 	struct radeon_bo *vram_obj = NULL;
37 	struct radeon_bo **gtt_obj = NULL;
38 	uint64_t gtt_addr, vram_addr;
39 	unsigned n, size;
40 	int i, r, ring;
41 
42 	switch (flag) {
43 	case RADEON_TEST_COPY_DMA:
44 		ring = radeon_copy_dma_ring_index(rdev);
45 		break;
46 	case RADEON_TEST_COPY_BLIT:
47 		ring = radeon_copy_blit_ring_index(rdev);
48 		break;
49 	default:
50 		DRM_ERROR("Unknown copy method\n");
51 		return;
52 	}
53 
54 	size = 1024 * 1024;
55 
56 	/* Number of tests =
57 	 * (Total GTT - IB pool - writeback page - ring buffers) / test size
58 	 */
59 	n = rdev->mc.gtt_size - rdev->gart_pin_size;
60 	n /= size;
61 
62 	gtt_obj = kzalloc(n * sizeof(*gtt_obj), GFP_KERNEL);
63 	if (!gtt_obj) {
64 		DRM_ERROR("Failed to allocate %d pointers\n", n);
65 		r = 1;
66 		goto out_cleanup;
67 	}
68 
69 	r = radeon_bo_create(rdev, size, PAGE_SIZE, true, RADEON_GEM_DOMAIN_VRAM,
70 			     0, NULL, NULL, &vram_obj);
71 	if (r) {
72 		DRM_ERROR("Failed to create VRAM object\n");
73 		goto out_cleanup;
74 	}
75 	r = radeon_bo_reserve(vram_obj, false);
76 	if (unlikely(r != 0))
77 		goto out_unref;
78 	r = radeon_bo_pin(vram_obj, RADEON_GEM_DOMAIN_VRAM, &vram_addr);
79 	if (r) {
80 		DRM_ERROR("Failed to pin VRAM object\n");
81 		goto out_unres;
82 	}
83 	for (i = 0; i < n; i++) {
84 		void *gtt_map, *vram_map;
85 		void **gtt_start, **gtt_end;
86 		void **vram_start, **vram_end;
87 		struct radeon_fence *fence = NULL;
88 
89 		r = radeon_bo_create(rdev, size, PAGE_SIZE, true,
90 				     RADEON_GEM_DOMAIN_GTT, 0, NULL, NULL,
91 				     gtt_obj + i);
92 		if (r) {
93 			DRM_ERROR("Failed to create GTT object %d\n", i);
94 			goto out_lclean;
95 		}
96 
97 		r = radeon_bo_reserve(gtt_obj[i], false);
98 		if (unlikely(r != 0))
99 			goto out_lclean_unref;
100 		r = radeon_bo_pin(gtt_obj[i], RADEON_GEM_DOMAIN_GTT, &gtt_addr);
101 		if (r) {
102 			DRM_ERROR("Failed to pin GTT object %d\n", i);
103 			goto out_lclean_unres;
104 		}
105 
106 		r = radeon_bo_kmap(gtt_obj[i], &gtt_map);
107 		if (r) {
108 			DRM_ERROR("Failed to map GTT object %d\n", i);
109 			goto out_lclean_unpin;
110 		}
111 
112 		for (gtt_start = gtt_map, gtt_end = gtt_map + size;
113 		     gtt_start < gtt_end;
114 		     gtt_start++)
115 			*gtt_start = gtt_start;
116 
117 		radeon_bo_kunmap(gtt_obj[i]);
118 
119 		if (ring == R600_RING_TYPE_DMA_INDEX)
120 			fence = radeon_copy_dma(rdev, gtt_addr, vram_addr,
121 						size / RADEON_GPU_PAGE_SIZE,
122 						vram_obj->tbo.resv);
123 		else
124 			fence = radeon_copy_blit(rdev, gtt_addr, vram_addr,
125 						 size / RADEON_GPU_PAGE_SIZE,
126 						 vram_obj->tbo.resv);
127 		if (IS_ERR(fence)) {
128 			DRM_ERROR("Failed GTT->VRAM copy %d\n", i);
129 			r = PTR_ERR(fence);
130 			goto out_lclean_unpin;
131 		}
132 
133 		r = radeon_fence_wait(fence, false);
134 		if (r) {
135 			DRM_ERROR("Failed to wait for GTT->VRAM fence %d\n", i);
136 			goto out_lclean_unpin;
137 		}
138 
139 		radeon_fence_unref(&fence);
140 
141 		r = radeon_bo_kmap(vram_obj, &vram_map);
142 		if (r) {
143 			DRM_ERROR("Failed to map VRAM object after copy %d\n", i);
144 			goto out_lclean_unpin;
145 		}
146 
147 		for (gtt_start = gtt_map, gtt_end = gtt_map + size,
148 		     vram_start = vram_map, vram_end = vram_map + size;
149 		     vram_start < vram_end;
150 		     gtt_start++, vram_start++) {
151 			if (*vram_start != gtt_start) {
152 				DRM_ERROR("Incorrect GTT->VRAM copy %d: Got 0x%p, "
153 					  "expected 0x%p (GTT/VRAM offset "
154 					  "0x%16llx/0x%16llx)\n",
155 					  i, *vram_start, gtt_start,
156 					  (unsigned long long)
157 					  (gtt_addr - rdev->mc.gtt_start +
158 					   (void*)gtt_start - gtt_map),
159 					  (unsigned long long)
160 					  (vram_addr - rdev->mc.vram_start +
161 					   (void*)gtt_start - gtt_map));
162 				radeon_bo_kunmap(vram_obj);
163 				goto out_lclean_unpin;
164 			}
165 			*vram_start = vram_start;
166 		}
167 
168 		radeon_bo_kunmap(vram_obj);
169 
170 		if (ring == R600_RING_TYPE_DMA_INDEX)
171 			fence = radeon_copy_dma(rdev, vram_addr, gtt_addr,
172 						size / RADEON_GPU_PAGE_SIZE,
173 						vram_obj->tbo.resv);
174 		else
175 			fence = radeon_copy_blit(rdev, vram_addr, gtt_addr,
176 						 size / RADEON_GPU_PAGE_SIZE,
177 						 vram_obj->tbo.resv);
178 		if (IS_ERR(fence)) {
179 			DRM_ERROR("Failed VRAM->GTT copy %d\n", i);
180 			r = PTR_ERR(fence);
181 			goto out_lclean_unpin;
182 		}
183 
184 		r = radeon_fence_wait(fence, false);
185 		if (r) {
186 			DRM_ERROR("Failed to wait for VRAM->GTT fence %d\n", i);
187 			goto out_lclean_unpin;
188 		}
189 
190 		radeon_fence_unref(&fence);
191 
192 		r = radeon_bo_kmap(gtt_obj[i], &gtt_map);
193 		if (r) {
194 			DRM_ERROR("Failed to map GTT object after copy %d\n", i);
195 			goto out_lclean_unpin;
196 		}
197 
198 		for (gtt_start = gtt_map, gtt_end = gtt_map + size,
199 		     vram_start = vram_map, vram_end = vram_map + size;
200 		     gtt_start < gtt_end;
201 		     gtt_start++, vram_start++) {
202 			if (*gtt_start != vram_start) {
203 				DRM_ERROR("Incorrect VRAM->GTT copy %d: Got 0x%p, "
204 					  "expected 0x%p (VRAM/GTT offset "
205 					  "0x%16llx/0x%16llx)\n",
206 					  i, *gtt_start, vram_start,
207 					  (unsigned long long)
208 					  (vram_addr - rdev->mc.vram_start +
209 					   (void*)vram_start - vram_map),
210 					  (unsigned long long)
211 					  (gtt_addr - rdev->mc.gtt_start +
212 					   (void*)vram_start - vram_map));
213 				radeon_bo_kunmap(gtt_obj[i]);
214 				goto out_lclean_unpin;
215 			}
216 		}
217 
218 		radeon_bo_kunmap(gtt_obj[i]);
219 
220 		DRM_INFO("Tested GTT->VRAM and VRAM->GTT copy for GTT offset 0x%llx\n",
221 			 gtt_addr - rdev->mc.gtt_start);
222 		continue;
223 
224 out_lclean_unpin:
225 		radeon_bo_unpin(gtt_obj[i]);
226 out_lclean_unres:
227 		radeon_bo_unreserve(gtt_obj[i]);
228 out_lclean_unref:
229 		radeon_bo_unref(&gtt_obj[i]);
230 out_lclean:
231 		for (--i; i >= 0; --i) {
232 			radeon_bo_unpin(gtt_obj[i]);
233 			radeon_bo_unreserve(gtt_obj[i]);
234 			radeon_bo_unref(&gtt_obj[i]);
235 		}
236 		if (fence && !IS_ERR(fence))
237 			radeon_fence_unref(&fence);
238 		break;
239 	}
240 
241 	radeon_bo_unpin(vram_obj);
242 out_unres:
243 	radeon_bo_unreserve(vram_obj);
244 out_unref:
245 	radeon_bo_unref(&vram_obj);
246 out_cleanup:
247 	kfree(gtt_obj);
248 	if (r) {
249 		printk(KERN_WARNING "Error while testing BO move.\n");
250 	}
251 }
252 
253 void radeon_test_moves(struct radeon_device *rdev)
254 {
255 	if (rdev->asic->copy.dma)
256 		radeon_do_test_moves(rdev, RADEON_TEST_COPY_DMA);
257 	if (rdev->asic->copy.blit)
258 		radeon_do_test_moves(rdev, RADEON_TEST_COPY_BLIT);
259 }
260 
261 static int radeon_test_create_and_emit_fence(struct radeon_device *rdev,
262 					     struct radeon_ring *ring,
263 					     struct radeon_fence **fence)
264 {
265 	uint32_t handle = ring->idx ^ 0xdeafbeef;
266 	int r;
267 
268 	if (ring->idx == R600_RING_TYPE_UVD_INDEX) {
269 		r = radeon_uvd_get_create_msg(rdev, ring->idx, handle, NULL);
270 		if (r) {
271 			DRM_ERROR("Failed to get dummy create msg\n");
272 			return r;
273 		}
274 
275 		r = radeon_uvd_get_destroy_msg(rdev, ring->idx, handle, fence);
276 		if (r) {
277 			DRM_ERROR("Failed to get dummy destroy msg\n");
278 			return r;
279 		}
280 
281 	} else if (ring->idx == TN_RING_TYPE_VCE1_INDEX ||
282 		   ring->idx == TN_RING_TYPE_VCE2_INDEX) {
283 		r = radeon_vce_get_create_msg(rdev, ring->idx, handle, NULL);
284 		if (r) {
285 			DRM_ERROR("Failed to get dummy create msg\n");
286 			return r;
287 		}
288 
289 		r = radeon_vce_get_destroy_msg(rdev, ring->idx, handle, fence);
290 		if (r) {
291 			DRM_ERROR("Failed to get dummy destroy msg\n");
292 			return r;
293 		}
294 
295 	} else {
296 		r = radeon_ring_lock(rdev, ring, 64);
297 		if (r) {
298 			DRM_ERROR("Failed to lock ring A %d\n", ring->idx);
299 			return r;
300 		}
301 		radeon_fence_emit(rdev, fence, ring->idx);
302 		radeon_ring_unlock_commit(rdev, ring, false);
303 	}
304 	return 0;
305 }
306 
307 void radeon_test_ring_sync(struct radeon_device *rdev,
308 			   struct radeon_ring *ringA,
309 			   struct radeon_ring *ringB)
310 {
311 	struct radeon_fence *fence1 = NULL, *fence2 = NULL;
312 	struct radeon_semaphore *semaphore = NULL;
313 	int r;
314 
315 	r = radeon_semaphore_create(rdev, &semaphore);
316 	if (r) {
317 		DRM_ERROR("Failed to create semaphore\n");
318 		goto out_cleanup;
319 	}
320 
321 	r = radeon_ring_lock(rdev, ringA, 64);
322 	if (r) {
323 		DRM_ERROR("Failed to lock ring A %d\n", ringA->idx);
324 		goto out_cleanup;
325 	}
326 	radeon_semaphore_emit_wait(rdev, ringA->idx, semaphore);
327 	radeon_ring_unlock_commit(rdev, ringA, false);
328 
329 	r = radeon_test_create_and_emit_fence(rdev, ringA, &fence1);
330 	if (r)
331 		goto out_cleanup;
332 
333 	r = radeon_ring_lock(rdev, ringA, 64);
334 	if (r) {
335 		DRM_ERROR("Failed to lock ring A %d\n", ringA->idx);
336 		goto out_cleanup;
337 	}
338 	radeon_semaphore_emit_wait(rdev, ringA->idx, semaphore);
339 	radeon_ring_unlock_commit(rdev, ringA, false);
340 
341 	r = radeon_test_create_and_emit_fence(rdev, ringA, &fence2);
342 	if (r)
343 		goto out_cleanup;
344 
345 	mdelay(1000);
346 
347 	if (radeon_fence_signaled(fence1)) {
348 		DRM_ERROR("Fence 1 signaled without waiting for semaphore.\n");
349 		goto out_cleanup;
350 	}
351 
352 	r = radeon_ring_lock(rdev, ringB, 64);
353 	if (r) {
354 		DRM_ERROR("Failed to lock ring B %p\n", ringB);
355 		goto out_cleanup;
356 	}
357 	radeon_semaphore_emit_signal(rdev, ringB->idx, semaphore);
358 	radeon_ring_unlock_commit(rdev, ringB, false);
359 
360 	r = radeon_fence_wait(fence1, false);
361 	if (r) {
362 		DRM_ERROR("Failed to wait for sync fence 1\n");
363 		goto out_cleanup;
364 	}
365 
366 	mdelay(1000);
367 
368 	if (radeon_fence_signaled(fence2)) {
369 		DRM_ERROR("Fence 2 signaled without waiting for semaphore.\n");
370 		goto out_cleanup;
371 	}
372 
373 	r = radeon_ring_lock(rdev, ringB, 64);
374 	if (r) {
375 		DRM_ERROR("Failed to lock ring B %p\n", ringB);
376 		goto out_cleanup;
377 	}
378 	radeon_semaphore_emit_signal(rdev, ringB->idx, semaphore);
379 	radeon_ring_unlock_commit(rdev, ringB, false);
380 
381 	r = radeon_fence_wait(fence2, false);
382 	if (r) {
383 		DRM_ERROR("Failed to wait for sync fence 1\n");
384 		goto out_cleanup;
385 	}
386 
387 out_cleanup:
388 	radeon_semaphore_free(rdev, &semaphore, NULL);
389 
390 	if (fence1)
391 		radeon_fence_unref(&fence1);
392 
393 	if (fence2)
394 		radeon_fence_unref(&fence2);
395 
396 	if (r)
397 		printk(KERN_WARNING "Error while testing ring sync (%d).\n", r);
398 }
399 
400 static void radeon_test_ring_sync2(struct radeon_device *rdev,
401 			    struct radeon_ring *ringA,
402 			    struct radeon_ring *ringB,
403 			    struct radeon_ring *ringC)
404 {
405 	struct radeon_fence *fenceA = NULL, *fenceB = NULL;
406 	struct radeon_semaphore *semaphore = NULL;
407 	bool sigA, sigB;
408 	int i, r;
409 
410 	r = radeon_semaphore_create(rdev, &semaphore);
411 	if (r) {
412 		DRM_ERROR("Failed to create semaphore\n");
413 		goto out_cleanup;
414 	}
415 
416 	r = radeon_ring_lock(rdev, ringA, 64);
417 	if (r) {
418 		DRM_ERROR("Failed to lock ring A %d\n", ringA->idx);
419 		goto out_cleanup;
420 	}
421 	radeon_semaphore_emit_wait(rdev, ringA->idx, semaphore);
422 	radeon_ring_unlock_commit(rdev, ringA, false);
423 
424 	r = radeon_test_create_and_emit_fence(rdev, ringA, &fenceA);
425 	if (r)
426 		goto out_cleanup;
427 
428 	r = radeon_ring_lock(rdev, ringB, 64);
429 	if (r) {
430 		DRM_ERROR("Failed to lock ring B %d\n", ringB->idx);
431 		goto out_cleanup;
432 	}
433 	radeon_semaphore_emit_wait(rdev, ringB->idx, semaphore);
434 	radeon_ring_unlock_commit(rdev, ringB, false);
435 	r = radeon_test_create_and_emit_fence(rdev, ringB, &fenceB);
436 	if (r)
437 		goto out_cleanup;
438 
439 	mdelay(1000);
440 
441 	if (radeon_fence_signaled(fenceA)) {
442 		DRM_ERROR("Fence A signaled without waiting for semaphore.\n");
443 		goto out_cleanup;
444 	}
445 	if (radeon_fence_signaled(fenceB)) {
446 		DRM_ERROR("Fence B signaled without waiting for semaphore.\n");
447 		goto out_cleanup;
448 	}
449 
450 	r = radeon_ring_lock(rdev, ringC, 64);
451 	if (r) {
452 		DRM_ERROR("Failed to lock ring B %p\n", ringC);
453 		goto out_cleanup;
454 	}
455 	radeon_semaphore_emit_signal(rdev, ringC->idx, semaphore);
456 	radeon_ring_unlock_commit(rdev, ringC, false);
457 
458 	for (i = 0; i < 30; ++i) {
459 		mdelay(100);
460 		sigA = radeon_fence_signaled(fenceA);
461 		sigB = radeon_fence_signaled(fenceB);
462 		if (sigA || sigB)
463 			break;
464 	}
465 
466 	if (!sigA && !sigB) {
467 		DRM_ERROR("Neither fence A nor B has been signaled\n");
468 		goto out_cleanup;
469 	} else if (sigA && sigB) {
470 		DRM_ERROR("Both fence A and B has been signaled\n");
471 		goto out_cleanup;
472 	}
473 
474 	DRM_INFO("Fence %c was first signaled\n", sigA ? 'A' : 'B');
475 
476 	r = radeon_ring_lock(rdev, ringC, 64);
477 	if (r) {
478 		DRM_ERROR("Failed to lock ring B %p\n", ringC);
479 		goto out_cleanup;
480 	}
481 	radeon_semaphore_emit_signal(rdev, ringC->idx, semaphore);
482 	radeon_ring_unlock_commit(rdev, ringC, false);
483 
484 	mdelay(1000);
485 
486 	r = radeon_fence_wait(fenceA, false);
487 	if (r) {
488 		DRM_ERROR("Failed to wait for sync fence A\n");
489 		goto out_cleanup;
490 	}
491 	r = radeon_fence_wait(fenceB, false);
492 	if (r) {
493 		DRM_ERROR("Failed to wait for sync fence B\n");
494 		goto out_cleanup;
495 	}
496 
497 out_cleanup:
498 	radeon_semaphore_free(rdev, &semaphore, NULL);
499 
500 	if (fenceA)
501 		radeon_fence_unref(&fenceA);
502 
503 	if (fenceB)
504 		radeon_fence_unref(&fenceB);
505 
506 	if (r)
507 		printk(KERN_WARNING "Error while testing ring sync (%d).\n", r);
508 }
509 
510 static bool radeon_test_sync_possible(struct radeon_ring *ringA,
511 				      struct radeon_ring *ringB)
512 {
513 	if (ringA->idx == TN_RING_TYPE_VCE2_INDEX &&
514 	    ringB->idx == TN_RING_TYPE_VCE1_INDEX)
515 		return false;
516 
517 	return true;
518 }
519 
520 void radeon_test_syncing(struct radeon_device *rdev)
521 {
522 	int i, j, k;
523 
524 	for (i = 1; i < RADEON_NUM_RINGS; ++i) {
525 		struct radeon_ring *ringA = &rdev->ring[i];
526 		if (!ringA->ready)
527 			continue;
528 
529 		for (j = 0; j < i; ++j) {
530 			struct radeon_ring *ringB = &rdev->ring[j];
531 			if (!ringB->ready)
532 				continue;
533 
534 			if (!radeon_test_sync_possible(ringA, ringB))
535 				continue;
536 
537 			DRM_INFO("Testing syncing between rings %d and %d...\n", i, j);
538 			radeon_test_ring_sync(rdev, ringA, ringB);
539 
540 			DRM_INFO("Testing syncing between rings %d and %d...\n", j, i);
541 			radeon_test_ring_sync(rdev, ringB, ringA);
542 
543 			for (k = 0; k < j; ++k) {
544 				struct radeon_ring *ringC = &rdev->ring[k];
545 				if (!ringC->ready)
546 					continue;
547 
548 				if (!radeon_test_sync_possible(ringA, ringC))
549 					continue;
550 
551 				if (!radeon_test_sync_possible(ringB, ringC))
552 					continue;
553 
554 				DRM_INFO("Testing syncing between rings %d, %d and %d...\n", i, j, k);
555 				radeon_test_ring_sync2(rdev, ringA, ringB, ringC);
556 
557 				DRM_INFO("Testing syncing between rings %d, %d and %d...\n", i, k, j);
558 				radeon_test_ring_sync2(rdev, ringA, ringC, ringB);
559 
560 				DRM_INFO("Testing syncing between rings %d, %d and %d...\n", j, i, k);
561 				radeon_test_ring_sync2(rdev, ringB, ringA, ringC);
562 
563 				DRM_INFO("Testing syncing between rings %d, %d and %d...\n", j, k, i);
564 				radeon_test_ring_sync2(rdev, ringB, ringC, ringA);
565 
566 				DRM_INFO("Testing syncing between rings %d, %d and %d...\n", k, i, j);
567 				radeon_test_ring_sync2(rdev, ringC, ringA, ringB);
568 
569 				DRM_INFO("Testing syncing between rings %d, %d and %d...\n", k, j, i);
570 				radeon_test_ring_sync2(rdev, ringC, ringB, ringA);
571 			}
572 		}
573 	}
574 }
575