xref: /openbsd-src/sys/dev/pci/drm/radeon/radeon_object.c (revision 4b70baf6e17fc8b27fc1f7fa7929335753fa94c3)
1 /*
2  * Copyright 2009 Jerome Glisse.
3  * All Rights Reserved.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the
7  * "Software"), to deal in the Software without restriction, including
8  * without limitation the rights to use, copy, modify, merge, publish,
9  * distribute, sub license, and/or sell copies of the Software, and to
10  * permit persons to whom the Software is furnished to do so, subject to
11  * the following conditions:
12  *
13  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
14  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
15  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
16  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
17  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
18  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
19  * USE OR OTHER DEALINGS IN THE SOFTWARE.
20  *
21  * The above copyright notice and this permission notice (including the
22  * next paragraph) shall be included in all copies or substantial portions
23  * of the Software.
24  *
25  */
26 /*
27  * Authors:
28  *    Jerome Glisse <glisse@freedesktop.org>
29  *    Thomas Hellstrom <thomas-at-tungstengraphics-dot-com>
30  *    Dave Airlie
31  */
32 #include <linux/list.h>
33 #include <linux/slab.h>
34 #include <drm/drmP.h>
35 #include <drm/radeon_drm.h>
36 #include <drm/drm_cache.h>
37 #include "radeon.h"
38 #include "radeon_trace.h"
39 
40 
41 int radeon_ttm_init(struct radeon_device *rdev);
42 void radeon_ttm_fini(struct radeon_device *rdev);
43 static void radeon_bo_clear_surface_reg(struct radeon_bo *bo);
44 
45 /*
46  * To exclude mutual BO access we rely on bo_reserve exclusion, as all
47  * function are calling it.
48  */
49 
50 static void radeon_update_memory_usage(struct radeon_bo *bo,
51 				       unsigned mem_type, int sign)
52 {
53 	struct radeon_device *rdev = bo->rdev;
54 	u64 size = (u64)bo->tbo.num_pages << PAGE_SHIFT;
55 
56 	switch (mem_type) {
57 	case TTM_PL_TT:
58 		if (sign > 0)
59 			atomic64_add(size, &rdev->gtt_usage);
60 		else
61 			atomic64_sub(size, &rdev->gtt_usage);
62 		break;
63 	case TTM_PL_VRAM:
64 		if (sign > 0)
65 			atomic64_add(size, &rdev->vram_usage);
66 		else
67 			atomic64_sub(size, &rdev->vram_usage);
68 		break;
69 	}
70 }
71 
72 static void radeon_ttm_bo_destroy(struct ttm_buffer_object *tbo)
73 {
74 	struct radeon_bo *bo;
75 
76 	bo = container_of(tbo, struct radeon_bo, tbo);
77 
78 	radeon_update_memory_usage(bo, bo->tbo.mem.mem_type, -1);
79 
80 	mutex_lock(&bo->rdev->gem.mutex);
81 	list_del_init(&bo->list);
82 	mutex_unlock(&bo->rdev->gem.mutex);
83 	radeon_bo_clear_surface_reg(bo);
84 	WARN_ON_ONCE(!list_empty(&bo->va));
85 #ifdef notyet
86 	if (bo->gem_base.import_attach)
87 		drm_prime_gem_destroy(&bo->gem_base, bo->tbo.sg);
88 #endif
89 	drm_gem_object_release(&bo->gem_base);
90 	pool_put(&bo->rdev->ddev->objpl, bo);
91 }
92 
93 bool radeon_ttm_bo_is_radeon_bo(struct ttm_buffer_object *bo)
94 {
95 	if (bo->destroy == &radeon_ttm_bo_destroy)
96 		return true;
97 	return false;
98 }
99 
100 void radeon_ttm_placement_from_domain(struct radeon_bo *rbo, u32 domain)
101 {
102 	u32 c = 0, i;
103 
104 	rbo->placement.placement = rbo->placements;
105 	rbo->placement.busy_placement = rbo->placements;
106 	if (domain & RADEON_GEM_DOMAIN_VRAM) {
107 		/* Try placing BOs which don't need CPU access outside of the
108 		 * CPU accessible part of VRAM
109 		 */
110 		if ((rbo->flags & RADEON_GEM_NO_CPU_ACCESS) &&
111 		    rbo->rdev->mc.visible_vram_size < rbo->rdev->mc.real_vram_size) {
112 			rbo->placements[c].fpfn =
113 				rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
114 			rbo->placements[c++].flags = TTM_PL_FLAG_WC |
115 						     TTM_PL_FLAG_UNCACHED |
116 						     TTM_PL_FLAG_VRAM;
117 		}
118 
119 		rbo->placements[c].fpfn = 0;
120 		rbo->placements[c++].flags = TTM_PL_FLAG_WC |
121 					     TTM_PL_FLAG_UNCACHED |
122 					     TTM_PL_FLAG_VRAM;
123 	}
124 
125 	if (domain & RADEON_GEM_DOMAIN_GTT) {
126 		if (rbo->flags & RADEON_GEM_GTT_UC) {
127 			rbo->placements[c].fpfn = 0;
128 			rbo->placements[c++].flags = TTM_PL_FLAG_UNCACHED |
129 				TTM_PL_FLAG_TT;
130 
131 		} else if ((rbo->flags & RADEON_GEM_GTT_WC) ||
132 			   (rbo->rdev->flags & RADEON_IS_AGP)) {
133 			rbo->placements[c].fpfn = 0;
134 			rbo->placements[c++].flags = TTM_PL_FLAG_WC |
135 				TTM_PL_FLAG_UNCACHED |
136 				TTM_PL_FLAG_TT;
137 		} else {
138 			rbo->placements[c].fpfn = 0;
139 			rbo->placements[c++].flags = TTM_PL_FLAG_CACHED |
140 						     TTM_PL_FLAG_TT;
141 		}
142 	}
143 
144 	if (domain & RADEON_GEM_DOMAIN_CPU) {
145 		if (rbo->flags & RADEON_GEM_GTT_UC) {
146 			rbo->placements[c].fpfn = 0;
147 			rbo->placements[c++].flags = TTM_PL_FLAG_UNCACHED |
148 				TTM_PL_FLAG_SYSTEM;
149 
150 		} else if ((rbo->flags & RADEON_GEM_GTT_WC) ||
151 		    rbo->rdev->flags & RADEON_IS_AGP) {
152 			rbo->placements[c].fpfn = 0;
153 			rbo->placements[c++].flags = TTM_PL_FLAG_WC |
154 				TTM_PL_FLAG_UNCACHED |
155 				TTM_PL_FLAG_SYSTEM;
156 		} else {
157 			rbo->placements[c].fpfn = 0;
158 			rbo->placements[c++].flags = TTM_PL_FLAG_CACHED |
159 						     TTM_PL_FLAG_SYSTEM;
160 		}
161 	}
162 	if (!c) {
163 		rbo->placements[c].fpfn = 0;
164 		rbo->placements[c++].flags = TTM_PL_MASK_CACHING |
165 					     TTM_PL_FLAG_SYSTEM;
166 	}
167 
168 	rbo->placement.num_placement = c;
169 	rbo->placement.num_busy_placement = c;
170 
171 	for (i = 0; i < c; ++i) {
172 		if ((rbo->flags & RADEON_GEM_CPU_ACCESS) &&
173 		    (rbo->placements[i].flags & TTM_PL_FLAG_VRAM) &&
174 		    !rbo->placements[i].fpfn)
175 			rbo->placements[i].lpfn =
176 				rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
177 		else
178 			rbo->placements[i].lpfn = 0;
179 	}
180 }
181 
182 int radeon_bo_create(struct radeon_device *rdev,
183 		     unsigned long size, int byte_align, bool kernel,
184 		     u32 domain, u32 flags, struct sg_table *sg,
185 		     struct reservation_object *resv,
186 		     struct radeon_bo **bo_ptr)
187 {
188 	struct radeon_bo *bo;
189 	enum ttm_bo_type type;
190 	unsigned long page_align = roundup(byte_align, PAGE_SIZE) >> PAGE_SHIFT;
191 	size_t acc_size;
192 	int r;
193 
194 	size = PAGE_ALIGN(size);
195 
196 	if (kernel) {
197 		type = ttm_bo_type_kernel;
198 	} else if (sg) {
199 		type = ttm_bo_type_sg;
200 	} else {
201 		type = ttm_bo_type_device;
202 	}
203 	*bo_ptr = NULL;
204 
205 	acc_size = ttm_bo_dma_acc_size(&rdev->mman.bdev, size,
206 				       sizeof(struct radeon_bo));
207 
208 	bo = pool_get(&rdev->ddev->objpl, PR_WAITOK | PR_ZERO);
209 	if (bo == NULL)
210 		return -ENOMEM;
211 	drm_gem_private_object_init(rdev->ddev, &bo->gem_base, size);
212 	bo->rdev = rdev;
213 	bo->surface_reg = -1;
214 	INIT_LIST_HEAD(&bo->list);
215 	INIT_LIST_HEAD(&bo->va);
216 	bo->initial_domain = domain & (RADEON_GEM_DOMAIN_VRAM |
217 				       RADEON_GEM_DOMAIN_GTT |
218 				       RADEON_GEM_DOMAIN_CPU);
219 
220 	bo->flags = flags;
221 	/* PCI GART is always snooped */
222 	if (!(rdev->flags & RADEON_IS_PCIE))
223 		bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
224 
225 	/* Write-combined CPU mappings of GTT cause GPU hangs with RV6xx
226 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=91268
227 	 */
228 	if (rdev->family >= CHIP_RV610 && rdev->family <= CHIP_RV635)
229 		bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
230 
231 #ifdef CONFIG_X86_32
232 	/* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
233 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
234 	 */
235 	bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
236 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
237 	/* Don't try to enable write-combining when it can't work, or things
238 	 * may be slow
239 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
240 	 */
241 #ifndef CONFIG_COMPILE_TEST
242 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
243 	 thanks to write-combining
244 #endif
245 
246 	if (bo->flags & RADEON_GEM_GTT_WC)
247 		DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
248 			      "better performance thanks to write-combining\n");
249 	bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
250 #else
251 	/* For architectures that don't support WC memory,
252 	 * mask out the WC flag from the BO
253 	 */
254 	if (!drm_arch_can_wc_memory())
255 		bo->flags &= ~RADEON_GEM_GTT_WC;
256 #endif
257 
258 	radeon_ttm_placement_from_domain(bo, domain);
259 	/* Kernel allocation are uninterruptible */
260 	down_read(&rdev->pm.mclk_lock);
261 	r = ttm_bo_init(&rdev->mman.bdev, &bo->tbo, size, type,
262 			&bo->placement, page_align, !kernel, acc_size,
263 			sg, resv, &radeon_ttm_bo_destroy);
264 	up_read(&rdev->pm.mclk_lock);
265 	if (unlikely(r != 0)) {
266 		return r;
267 	}
268 	*bo_ptr = bo;
269 
270 	trace_radeon_bo_create(bo);
271 
272 	return 0;
273 }
274 
275 int radeon_bo_kmap(struct radeon_bo *bo, void **ptr)
276 {
277 	bool is_iomem;
278 	int r;
279 
280 	if (bo->kptr) {
281 		if (ptr) {
282 			*ptr = bo->kptr;
283 		}
284 		return 0;
285 	}
286 	r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages, &bo->kmap);
287 	if (r) {
288 		return r;
289 	}
290 	bo->kptr = ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
291 	if (ptr) {
292 		*ptr = bo->kptr;
293 	}
294 	radeon_bo_check_tiling(bo, 0, 0);
295 	return 0;
296 }
297 
298 void radeon_bo_kunmap(struct radeon_bo *bo)
299 {
300 	if (bo->kptr == NULL)
301 		return;
302 	bo->kptr = NULL;
303 	radeon_bo_check_tiling(bo, 0, 0);
304 	ttm_bo_kunmap(&bo->kmap);
305 }
306 
307 struct radeon_bo *radeon_bo_ref(struct radeon_bo *bo)
308 {
309 	if (bo == NULL)
310 		return NULL;
311 
312 	ttm_bo_get(&bo->tbo);
313 	return bo;
314 }
315 
316 void radeon_bo_unref(struct radeon_bo **bo)
317 {
318 	struct ttm_buffer_object *tbo;
319 	struct radeon_device *rdev;
320 
321 	if ((*bo) == NULL)
322 		return;
323 	rdev = (*bo)->rdev;
324 	tbo = &((*bo)->tbo);
325 	ttm_bo_put(tbo);
326 	*bo = NULL;
327 }
328 
329 int radeon_bo_pin_restricted(struct radeon_bo *bo, u32 domain, u64 max_offset,
330 			     u64 *gpu_addr)
331 {
332 	struct ttm_operation_ctx ctx = { false, false };
333 	int r, i;
334 
335 	if (radeon_ttm_tt_has_userptr(bo->tbo.ttm))
336 		return -EPERM;
337 
338 	if (bo->pin_count) {
339 		bo->pin_count++;
340 		if (gpu_addr)
341 			*gpu_addr = radeon_bo_gpu_offset(bo);
342 
343 		if (max_offset != 0) {
344 			u64 domain_start;
345 
346 			if (domain == RADEON_GEM_DOMAIN_VRAM)
347 				domain_start = bo->rdev->mc.vram_start;
348 			else
349 				domain_start = bo->rdev->mc.gtt_start;
350 			WARN_ON_ONCE(max_offset <
351 				     (radeon_bo_gpu_offset(bo) - domain_start));
352 		}
353 
354 		return 0;
355 	}
356 	if (bo->prime_shared_count && domain == RADEON_GEM_DOMAIN_VRAM) {
357 		/* A BO shared as a dma-buf cannot be sensibly migrated to VRAM */
358 		return -EINVAL;
359 	}
360 
361 	radeon_ttm_placement_from_domain(bo, domain);
362 	for (i = 0; i < bo->placement.num_placement; i++) {
363 		/* force to pin into visible video ram */
364 		if ((bo->placements[i].flags & TTM_PL_FLAG_VRAM) &&
365 		    !(bo->flags & RADEON_GEM_NO_CPU_ACCESS) &&
366 		    (!max_offset || max_offset > bo->rdev->mc.visible_vram_size))
367 			bo->placements[i].lpfn =
368 				bo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
369 		else
370 			bo->placements[i].lpfn = max_offset >> PAGE_SHIFT;
371 
372 		bo->placements[i].flags |= TTM_PL_FLAG_NO_EVICT;
373 	}
374 
375 	r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
376 	if (likely(r == 0)) {
377 		bo->pin_count = 1;
378 		if (gpu_addr != NULL)
379 			*gpu_addr = radeon_bo_gpu_offset(bo);
380 		if (domain == RADEON_GEM_DOMAIN_VRAM)
381 			bo->rdev->vram_pin_size += radeon_bo_size(bo);
382 		else
383 			bo->rdev->gart_pin_size += radeon_bo_size(bo);
384 	} else {
385 		dev_err(bo->rdev->dev, "%p pin failed\n", bo);
386 	}
387 	return r;
388 }
389 
390 int radeon_bo_pin(struct radeon_bo *bo, u32 domain, u64 *gpu_addr)
391 {
392 	return radeon_bo_pin_restricted(bo, domain, 0, gpu_addr);
393 }
394 
395 int radeon_bo_unpin(struct radeon_bo *bo)
396 {
397 	struct ttm_operation_ctx ctx = { false, false };
398 	int r, i;
399 
400 	if (!bo->pin_count) {
401 		dev_warn(bo->rdev->dev, "%p unpin not necessary\n", bo);
402 		return 0;
403 	}
404 	bo->pin_count--;
405 	if (bo->pin_count)
406 		return 0;
407 	for (i = 0; i < bo->placement.num_placement; i++) {
408 		bo->placements[i].lpfn = 0;
409 		bo->placements[i].flags &= ~TTM_PL_FLAG_NO_EVICT;
410 	}
411 	r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
412 	if (likely(r == 0)) {
413 		if (bo->tbo.mem.mem_type == TTM_PL_VRAM)
414 			bo->rdev->vram_pin_size -= radeon_bo_size(bo);
415 		else
416 			bo->rdev->gart_pin_size -= radeon_bo_size(bo);
417 	} else {
418 		dev_err(bo->rdev->dev, "%p validate failed for unpin\n", bo);
419 	}
420 	return r;
421 }
422 
423 int radeon_bo_evict_vram(struct radeon_device *rdev)
424 {
425 	/* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
426 	if (0 && (rdev->flags & RADEON_IS_IGP)) {
427 		if (rdev->mc.igp_sideport_enabled == false)
428 			/* Useless to evict on IGP chips */
429 			return 0;
430 	}
431 	return ttm_bo_evict_mm(&rdev->mman.bdev, TTM_PL_VRAM);
432 }
433 
434 void radeon_bo_force_delete(struct radeon_device *rdev)
435 {
436 	struct radeon_bo *bo, *n;
437 
438 	if (list_empty(&rdev->gem.objects)) {
439 		return;
440 	}
441 	dev_err(rdev->dev, "Userspace still has active objects !\n");
442 	list_for_each_entry_safe(bo, n, &rdev->gem.objects, list) {
443 		dev_err(rdev->dev, "%p %p %lu %lu force free\n",
444 			&bo->gem_base, bo, (unsigned long)bo->gem_base.size,
445 			*((unsigned long *)&bo->gem_base.refcount));
446 		mutex_lock(&bo->rdev->gem.mutex);
447 		list_del_init(&bo->list);
448 		mutex_unlock(&bo->rdev->gem.mutex);
449 		/* this should unref the ttm bo */
450 		drm_gem_object_put_unlocked(&bo->gem_base);
451 	}
452 }
453 
454 int radeon_bo_init(struct radeon_device *rdev)
455 {
456 	paddr_t start, end;
457 
458 #ifdef __linux__
459 	/* reserve PAT memory space to WC for VRAM */
460 	arch_io_reserve_memtype_wc(rdev->mc.aper_base,
461 				   rdev->mc.aper_size);
462 #endif
463 
464 	/* Add an MTRR for the VRAM */
465 	if (!rdev->fastfb_working) {
466 #ifdef __linux__
467 		rdev->mc.vram_mtrr = arch_phys_wc_add(rdev->mc.aper_base,
468 						      rdev->mc.aper_size);
469 #else
470 		drm_mtrr_add(rdev->mc.aper_base, rdev->mc.aper_size, DRM_MTRR_WC);
471 		/* fake a 'cookie', seems to be unused? */
472 		rdev->mc.vram_mtrr = 1;
473 #endif
474 	}
475 
476 	start = atop(bus_space_mmap(rdev->memt, rdev->mc.aper_base, 0, 0, 0));
477 	end = start + atop(rdev->mc.aper_size);
478 	uvm_page_physload(start, end, start, end, PHYSLOAD_DEVICE);
479 
480 	DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
481 		rdev->mc.mc_vram_size >> 20,
482 		(unsigned long long)rdev->mc.aper_size >> 20);
483 	DRM_INFO("RAM width %dbits %cDR\n",
484 			rdev->mc.vram_width, rdev->mc.vram_is_ddr ? 'D' : 'S');
485 	return radeon_ttm_init(rdev);
486 }
487 
488 void radeon_bo_fini(struct radeon_device *rdev)
489 {
490 	radeon_ttm_fini(rdev);
491 #ifdef __linux__
492 	arch_phys_wc_del(rdev->mc.vram_mtrr);
493 	arch_io_free_memtype_wc(rdev->mc.aper_base, rdev->mc.aper_size);
494 #else
495 	drm_mtrr_del(0, rdev->mc.aper_base, rdev->mc.aper_size, DRM_MTRR_WC);
496 #endif
497 }
498 
499 /* Returns how many bytes TTM can move per IB.
500  */
501 static u64 radeon_bo_get_threshold_for_moves(struct radeon_device *rdev)
502 {
503 	u64 real_vram_size = rdev->mc.real_vram_size;
504 	u64 vram_usage = atomic64_read(&rdev->vram_usage);
505 
506 	/* This function is based on the current VRAM usage.
507 	 *
508 	 * - If all of VRAM is free, allow relocating the number of bytes that
509 	 *   is equal to 1/4 of the size of VRAM for this IB.
510 
511 	 * - If more than one half of VRAM is occupied, only allow relocating
512 	 *   1 MB of data for this IB.
513 	 *
514 	 * - From 0 to one half of used VRAM, the threshold decreases
515 	 *   linearly.
516 	 *         __________________
517 	 * 1/4 of -|\               |
518 	 * VRAM    | \              |
519 	 *         |  \             |
520 	 *         |   \            |
521 	 *         |    \           |
522 	 *         |     \          |
523 	 *         |      \         |
524 	 *         |       \________|1 MB
525 	 *         |----------------|
526 	 *    VRAM 0 %             100 %
527 	 *         used            used
528 	 *
529 	 * Note: It's a threshold, not a limit. The threshold must be crossed
530 	 * for buffer relocations to stop, so any buffer of an arbitrary size
531 	 * can be moved as long as the threshold isn't crossed before
532 	 * the relocation takes place. We don't want to disable buffer
533 	 * relocations completely.
534 	 *
535 	 * The idea is that buffers should be placed in VRAM at creation time
536 	 * and TTM should only do a minimum number of relocations during
537 	 * command submission. In practice, you need to submit at least
538 	 * a dozen IBs to move all buffers to VRAM if they are in GTT.
539 	 *
540 	 * Also, things can get pretty crazy under memory pressure and actual
541 	 * VRAM usage can change a lot, so playing safe even at 50% does
542 	 * consistently increase performance.
543 	 */
544 
545 	u64 half_vram = real_vram_size >> 1;
546 	u64 half_free_vram = vram_usage >= half_vram ? 0 : half_vram - vram_usage;
547 	u64 bytes_moved_threshold = half_free_vram >> 1;
548 	return max(bytes_moved_threshold, 1024*1024ull);
549 }
550 
551 int radeon_bo_list_validate(struct radeon_device *rdev,
552 			    struct ww_acquire_ctx *ticket,
553 			    struct list_head *head, int ring)
554 {
555 	struct ttm_operation_ctx ctx = { true, false };
556 	struct radeon_bo_list *lobj;
557 	struct list_head duplicates;
558 	int r;
559 	u64 bytes_moved = 0, initial_bytes_moved;
560 	u64 bytes_moved_threshold = radeon_bo_get_threshold_for_moves(rdev);
561 
562 	INIT_LIST_HEAD(&duplicates);
563 	r = ttm_eu_reserve_buffers(ticket, head, true, &duplicates);
564 	if (unlikely(r != 0)) {
565 		return r;
566 	}
567 
568 	list_for_each_entry(lobj, head, tv.head) {
569 		struct radeon_bo *bo = lobj->robj;
570 		if (!bo->pin_count) {
571 			u32 domain = lobj->preferred_domains;
572 			u32 allowed = lobj->allowed_domains;
573 			u32 current_domain =
574 				radeon_mem_type_to_domain(bo->tbo.mem.mem_type);
575 
576 			/* Check if this buffer will be moved and don't move it
577 			 * if we have moved too many buffers for this IB already.
578 			 *
579 			 * Note that this allows moving at least one buffer of
580 			 * any size, because it doesn't take the current "bo"
581 			 * into account. We don't want to disallow buffer moves
582 			 * completely.
583 			 */
584 			if ((allowed & current_domain) != 0 &&
585 			    (domain & current_domain) == 0 && /* will be moved */
586 			    bytes_moved > bytes_moved_threshold) {
587 				/* don't move it */
588 				domain = current_domain;
589 			}
590 
591 		retry:
592 			radeon_ttm_placement_from_domain(bo, domain);
593 			if (ring == R600_RING_TYPE_UVD_INDEX)
594 				radeon_uvd_force_into_uvd_segment(bo, allowed);
595 
596 			initial_bytes_moved = atomic64_read(&rdev->num_bytes_moved);
597 			r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
598 			bytes_moved += atomic64_read(&rdev->num_bytes_moved) -
599 				       initial_bytes_moved;
600 
601 			if (unlikely(r)) {
602 				if (r != -ERESTARTSYS &&
603 				    domain != lobj->allowed_domains) {
604 					domain = lobj->allowed_domains;
605 					goto retry;
606 				}
607 				ttm_eu_backoff_reservation(ticket, head);
608 				return r;
609 			}
610 		}
611 		lobj->gpu_offset = radeon_bo_gpu_offset(bo);
612 		lobj->tiling_flags = bo->tiling_flags;
613 	}
614 
615 	list_for_each_entry(lobj, &duplicates, tv.head) {
616 		lobj->gpu_offset = radeon_bo_gpu_offset(lobj->robj);
617 		lobj->tiling_flags = lobj->robj->tiling_flags;
618 	}
619 
620 	return 0;
621 }
622 
623 int radeon_bo_get_surface_reg(struct radeon_bo *bo)
624 {
625 	struct radeon_device *rdev = bo->rdev;
626 	struct radeon_surface_reg *reg;
627 	struct radeon_bo *old_object;
628 	int steal;
629 	int i;
630 
631 #ifdef notyet
632 	lockdep_assert_held(&bo->tbo.resv->lock.base);
633 #endif
634 
635 	if (!bo->tiling_flags)
636 		return 0;
637 
638 	if (bo->surface_reg >= 0) {
639 		reg = &rdev->surface_regs[bo->surface_reg];
640 		i = bo->surface_reg;
641 		goto out;
642 	}
643 
644 	steal = -1;
645 	for (i = 0; i < RADEON_GEM_MAX_SURFACES; i++) {
646 
647 		reg = &rdev->surface_regs[i];
648 		if (!reg->bo)
649 			break;
650 
651 		old_object = reg->bo;
652 		if (old_object->pin_count == 0)
653 			steal = i;
654 	}
655 
656 	/* if we are all out */
657 	if (i == RADEON_GEM_MAX_SURFACES) {
658 		if (steal == -1)
659 			return -ENOMEM;
660 		/* find someone with a surface reg and nuke their BO */
661 		reg = &rdev->surface_regs[steal];
662 		old_object = reg->bo;
663 		/* blow away the mapping */
664 		DRM_DEBUG("stealing surface reg %d from %p\n", steal, old_object);
665 		ttm_bo_unmap_virtual(&old_object->tbo);
666 		old_object->surface_reg = -1;
667 		i = steal;
668 	}
669 
670 	bo->surface_reg = i;
671 	reg->bo = bo;
672 
673 out:
674 	radeon_set_surface_reg(rdev, i, bo->tiling_flags, bo->pitch,
675 			       bo->tbo.mem.start << PAGE_SHIFT,
676 			       bo->tbo.num_pages << PAGE_SHIFT);
677 	return 0;
678 }
679 
680 static void radeon_bo_clear_surface_reg(struct radeon_bo *bo)
681 {
682 	struct radeon_device *rdev = bo->rdev;
683 	struct radeon_surface_reg *reg;
684 
685 	if (bo->surface_reg == -1)
686 		return;
687 
688 	reg = &rdev->surface_regs[bo->surface_reg];
689 	radeon_clear_surface_reg(rdev, bo->surface_reg);
690 
691 	reg->bo = NULL;
692 	bo->surface_reg = -1;
693 }
694 
695 int radeon_bo_set_tiling_flags(struct radeon_bo *bo,
696 				uint32_t tiling_flags, uint32_t pitch)
697 {
698 	struct radeon_device *rdev = bo->rdev;
699 	int r;
700 
701 	if (rdev->family >= CHIP_CEDAR) {
702 		unsigned bankw, bankh, mtaspect, tilesplit, stilesplit;
703 
704 		bankw = (tiling_flags >> RADEON_TILING_EG_BANKW_SHIFT) & RADEON_TILING_EG_BANKW_MASK;
705 		bankh = (tiling_flags >> RADEON_TILING_EG_BANKH_SHIFT) & RADEON_TILING_EG_BANKH_MASK;
706 		mtaspect = (tiling_flags >> RADEON_TILING_EG_MACRO_TILE_ASPECT_SHIFT) & RADEON_TILING_EG_MACRO_TILE_ASPECT_MASK;
707 		tilesplit = (tiling_flags >> RADEON_TILING_EG_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_TILE_SPLIT_MASK;
708 		stilesplit = (tiling_flags >> RADEON_TILING_EG_STENCIL_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_STENCIL_TILE_SPLIT_MASK;
709 		switch (bankw) {
710 		case 0:
711 		case 1:
712 		case 2:
713 		case 4:
714 		case 8:
715 			break;
716 		default:
717 			return -EINVAL;
718 		}
719 		switch (bankh) {
720 		case 0:
721 		case 1:
722 		case 2:
723 		case 4:
724 		case 8:
725 			break;
726 		default:
727 			return -EINVAL;
728 		}
729 		switch (mtaspect) {
730 		case 0:
731 		case 1:
732 		case 2:
733 		case 4:
734 		case 8:
735 			break;
736 		default:
737 			return -EINVAL;
738 		}
739 		if (tilesplit > 6) {
740 			return -EINVAL;
741 		}
742 		if (stilesplit > 6) {
743 			return -EINVAL;
744 		}
745 	}
746 	r = radeon_bo_reserve(bo, false);
747 	if (unlikely(r != 0))
748 		return r;
749 	bo->tiling_flags = tiling_flags;
750 	bo->pitch = pitch;
751 	radeon_bo_unreserve(bo);
752 	return 0;
753 }
754 
755 void radeon_bo_get_tiling_flags(struct radeon_bo *bo,
756 				uint32_t *tiling_flags,
757 				uint32_t *pitch)
758 {
759 #ifdef notyet
760 	lockdep_assert_held(&bo->tbo.resv->lock.base);
761 #endif
762 
763 	if (tiling_flags)
764 		*tiling_flags = bo->tiling_flags;
765 	if (pitch)
766 		*pitch = bo->pitch;
767 }
768 
769 int radeon_bo_check_tiling(struct radeon_bo *bo, bool has_moved,
770 				bool force_drop)
771 {
772 #ifdef notyet
773 	if (!force_drop)
774 		lockdep_assert_held(&bo->tbo.resv->lock.base);
775 #endif
776 
777 	if (!(bo->tiling_flags & RADEON_TILING_SURFACE))
778 		return 0;
779 
780 	if (force_drop) {
781 		radeon_bo_clear_surface_reg(bo);
782 		return 0;
783 	}
784 
785 	if (bo->tbo.mem.mem_type != TTM_PL_VRAM) {
786 		if (!has_moved)
787 			return 0;
788 
789 		if (bo->surface_reg >= 0)
790 			radeon_bo_clear_surface_reg(bo);
791 		return 0;
792 	}
793 
794 	if ((bo->surface_reg >= 0) && !has_moved)
795 		return 0;
796 
797 	return radeon_bo_get_surface_reg(bo);
798 }
799 
800 void radeon_bo_move_notify(struct ttm_buffer_object *bo,
801 			   bool evict,
802 			   struct ttm_mem_reg *new_mem)
803 {
804 	struct radeon_bo *rbo;
805 
806 	if (!radeon_ttm_bo_is_radeon_bo(bo))
807 		return;
808 
809 	rbo = container_of(bo, struct radeon_bo, tbo);
810 	radeon_bo_check_tiling(rbo, 0, 1);
811 	radeon_vm_bo_invalidate(rbo->rdev, rbo);
812 
813 	/* update statistics */
814 	if (!new_mem)
815 		return;
816 
817 	radeon_update_memory_usage(rbo, bo->mem.mem_type, -1);
818 	radeon_update_memory_usage(rbo, new_mem->mem_type, 1);
819 }
820 
821 int radeon_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
822 {
823 	struct ttm_operation_ctx ctx = { false, false };
824 	struct radeon_device *rdev;
825 	struct radeon_bo *rbo;
826 	unsigned long offset, size, lpfn;
827 	int i, r;
828 
829 	if (!radeon_ttm_bo_is_radeon_bo(bo))
830 		return 0;
831 	rbo = container_of(bo, struct radeon_bo, tbo);
832 	radeon_bo_check_tiling(rbo, 0, 0);
833 	rdev = rbo->rdev;
834 	if (bo->mem.mem_type != TTM_PL_VRAM)
835 		return 0;
836 
837 	size = bo->mem.num_pages << PAGE_SHIFT;
838 	offset = bo->mem.start << PAGE_SHIFT;
839 	if ((offset + size) <= rdev->mc.visible_vram_size)
840 		return 0;
841 
842 	/* Can't move a pinned BO to visible VRAM */
843 	if (rbo->pin_count > 0)
844 		return -EINVAL;
845 
846 	/* hurrah the memory is not visible ! */
847 	radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_VRAM);
848 	lpfn =	rdev->mc.visible_vram_size >> PAGE_SHIFT;
849 	for (i = 0; i < rbo->placement.num_placement; i++) {
850 		/* Force into visible VRAM */
851 		if ((rbo->placements[i].flags & TTM_PL_FLAG_VRAM) &&
852 		    (!rbo->placements[i].lpfn || rbo->placements[i].lpfn > lpfn))
853 			rbo->placements[i].lpfn = lpfn;
854 	}
855 	r = ttm_bo_validate(bo, &rbo->placement, &ctx);
856 	if (unlikely(r == -ENOMEM)) {
857 		radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_GTT);
858 		return ttm_bo_validate(bo, &rbo->placement, &ctx);
859 	} else if (unlikely(r != 0)) {
860 		return r;
861 	}
862 
863 	offset = bo->mem.start << PAGE_SHIFT;
864 	/* this should never happen */
865 	if ((offset + size) > rdev->mc.visible_vram_size)
866 		return -EINVAL;
867 
868 	return 0;
869 }
870 
871 int radeon_bo_wait(struct radeon_bo *bo, u32 *mem_type, bool no_wait)
872 {
873 	int r;
874 
875 	r = ttm_bo_reserve(&bo->tbo, true, no_wait, NULL);
876 	if (unlikely(r != 0))
877 		return r;
878 	if (mem_type)
879 		*mem_type = bo->tbo.mem.mem_type;
880 
881 	r = ttm_bo_wait(&bo->tbo, true, no_wait);
882 	ttm_bo_unreserve(&bo->tbo);
883 	return r;
884 }
885 
886 /**
887  * radeon_bo_fence - add fence to buffer object
888  *
889  * @bo: buffer object in question
890  * @fence: fence to add
891  * @shared: true if fence should be added shared
892  *
893  */
894 void radeon_bo_fence(struct radeon_bo *bo, struct radeon_fence *fence,
895 		     bool shared)
896 {
897 	struct reservation_object *resv = bo->tbo.resv;
898 
899 	if (shared)
900 		reservation_object_add_shared_fence(resv, &fence->base);
901 	else
902 		reservation_object_add_excl_fence(resv, &fence->base);
903 }
904