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