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
33 #include <linux/io.h>
34 #include <linux/list.h>
35 #include <linux/slab.h>
36
37 #include <drm/drm_cache.h>
38 #include <drm/drm_prime.h>
39 #include <drm/radeon_drm.h>
40
41 #include "radeon.h"
42 #include "radeon_trace.h"
43 #include "radeon_ttm.h"
44
45 static void radeon_bo_clear_surface_reg(struct radeon_bo *bo);
46
47 /*
48 * To exclude mutual BO access we rely on bo_reserve exclusion, as all
49 * function are calling it.
50 */
51
radeon_ttm_bo_destroy(struct ttm_buffer_object * tbo)52 static void radeon_ttm_bo_destroy(struct ttm_buffer_object *tbo)
53 {
54 struct radeon_bo *bo;
55
56 bo = container_of(tbo, struct radeon_bo, tbo);
57
58 mutex_lock(&bo->rdev->gem.mutex);
59 list_del_init(&bo->list);
60 mutex_unlock(&bo->rdev->gem.mutex);
61 radeon_bo_clear_surface_reg(bo);
62 WARN_ON_ONCE(!list_empty(&bo->va));
63 if (bo->tbo.base.import_attach)
64 drm_prime_gem_destroy(&bo->tbo.base, bo->tbo.sg);
65 drm_gem_object_release(&bo->tbo.base);
66 kfree(bo);
67 }
68
radeon_ttm_bo_is_radeon_bo(struct ttm_buffer_object * bo)69 bool radeon_ttm_bo_is_radeon_bo(struct ttm_buffer_object *bo)
70 {
71 if (bo->destroy == &radeon_ttm_bo_destroy)
72 return true;
73 return false;
74 }
75
radeon_ttm_placement_from_domain(struct radeon_bo * rbo,u32 domain)76 void radeon_ttm_placement_from_domain(struct radeon_bo *rbo, u32 domain)
77 {
78 u32 c = 0, i;
79
80 rbo->placement.placement = rbo->placements;
81 if (domain & RADEON_GEM_DOMAIN_VRAM) {
82 /* Try placing BOs which don't need CPU access outside of the
83 * CPU accessible part of VRAM
84 */
85 if ((rbo->flags & RADEON_GEM_NO_CPU_ACCESS) &&
86 rbo->rdev->mc.visible_vram_size < rbo->rdev->mc.real_vram_size) {
87 rbo->placements[c].fpfn =
88 rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
89 rbo->placements[c].mem_type = TTM_PL_VRAM;
90 rbo->placements[c++].flags = 0;
91 }
92
93 rbo->placements[c].fpfn = 0;
94 rbo->placements[c].mem_type = TTM_PL_VRAM;
95 rbo->placements[c++].flags = 0;
96 }
97
98 if (domain & RADEON_GEM_DOMAIN_GTT) {
99 rbo->placements[c].fpfn = 0;
100 rbo->placements[c].mem_type = TTM_PL_TT;
101 rbo->placements[c++].flags = 0;
102 }
103
104 if (domain & RADEON_GEM_DOMAIN_CPU) {
105 rbo->placements[c].fpfn = 0;
106 rbo->placements[c].mem_type = TTM_PL_SYSTEM;
107 rbo->placements[c++].flags = 0;
108 }
109 if (!c) {
110 rbo->placements[c].fpfn = 0;
111 rbo->placements[c].mem_type = TTM_PL_SYSTEM;
112 rbo->placements[c++].flags = 0;
113 }
114
115 rbo->placement.num_placement = c;
116
117 for (i = 0; i < c; ++i) {
118 if ((rbo->flags & RADEON_GEM_CPU_ACCESS) &&
119 (rbo->placements[i].mem_type == TTM_PL_VRAM) &&
120 !rbo->placements[i].fpfn)
121 rbo->placements[i].lpfn =
122 rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
123 else
124 rbo->placements[i].lpfn = 0;
125 }
126 }
127
radeon_bo_create(struct radeon_device * rdev,unsigned long size,int byte_align,bool kernel,u32 domain,u32 flags,struct sg_table * sg,struct dma_resv * resv,struct radeon_bo ** bo_ptr)128 int radeon_bo_create(struct radeon_device *rdev,
129 unsigned long size, int byte_align, bool kernel,
130 u32 domain, u32 flags, struct sg_table *sg,
131 struct dma_resv *resv,
132 struct radeon_bo **bo_ptr)
133 {
134 struct radeon_bo *bo;
135 enum ttm_bo_type type;
136 unsigned long page_align = roundup(byte_align, PAGE_SIZE) >> PAGE_SHIFT;
137 int r;
138
139 size = ALIGN(size, PAGE_SIZE);
140
141 if (kernel) {
142 type = ttm_bo_type_kernel;
143 } else if (sg) {
144 type = ttm_bo_type_sg;
145 } else {
146 type = ttm_bo_type_device;
147 }
148 *bo_ptr = NULL;
149
150 bo = kzalloc(sizeof(struct radeon_bo), GFP_KERNEL);
151 if (bo == NULL)
152 return -ENOMEM;
153 drm_gem_private_object_init(rdev_to_drm(rdev), &bo->tbo.base, size);
154 bo->tbo.base.funcs = &radeon_gem_object_funcs;
155 bo->rdev = rdev;
156 bo->surface_reg = -1;
157 INIT_LIST_HEAD(&bo->list);
158 INIT_LIST_HEAD(&bo->va);
159 bo->initial_domain = domain & (RADEON_GEM_DOMAIN_VRAM |
160 RADEON_GEM_DOMAIN_GTT |
161 RADEON_GEM_DOMAIN_CPU);
162
163 bo->flags = flags;
164 /* PCI GART is always snooped */
165 if (!(rdev->flags & RADEON_IS_PCIE))
166 bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
167
168 /* Write-combined CPU mappings of GTT cause GPU hangs with RV6xx
169 * See https://bugs.freedesktop.org/show_bug.cgi?id=91268
170 */
171 if (rdev->family >= CHIP_RV610 && rdev->family <= CHIP_RV635)
172 bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
173
174 #ifdef CONFIG_X86_32
175 /* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
176 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
177 */
178 bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
179 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
180 /* Don't try to enable write-combining when it can't work, or things
181 * may be slow
182 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
183 */
184 #ifndef CONFIG_COMPILE_TEST
185 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
186 thanks to write-combining
187 #endif
188
189 if (bo->flags & RADEON_GEM_GTT_WC)
190 DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
191 "better performance thanks to write-combining\n");
192 bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
193 #else
194 /* For architectures that don't support WC memory,
195 * mask out the WC flag from the BO
196 */
197 if (!drm_arch_can_wc_memory())
198 bo->flags &= ~RADEON_GEM_GTT_WC;
199 #endif
200
201 radeon_ttm_placement_from_domain(bo, domain);
202 /* Kernel allocation are uninterruptible */
203 down_read(&rdev->pm.mclk_lock);
204 r = ttm_bo_init_validate(&rdev->mman.bdev, &bo->tbo, type,
205 &bo->placement, page_align, !kernel, sg, resv,
206 &radeon_ttm_bo_destroy);
207 up_read(&rdev->pm.mclk_lock);
208 if (unlikely(r != 0)) {
209 return r;
210 }
211 *bo_ptr = bo;
212
213 trace_radeon_bo_create(bo);
214
215 return 0;
216 }
217
radeon_bo_kmap(struct radeon_bo * bo,void ** ptr)218 int radeon_bo_kmap(struct radeon_bo *bo, void **ptr)
219 {
220 bool is_iomem;
221 long r;
222
223 r = dma_resv_wait_timeout(bo->tbo.base.resv, DMA_RESV_USAGE_KERNEL,
224 false, MAX_SCHEDULE_TIMEOUT);
225 if (r < 0)
226 return r;
227
228 if (bo->kptr) {
229 if (ptr) {
230 *ptr = bo->kptr;
231 }
232 return 0;
233 }
234 r = ttm_bo_kmap(&bo->tbo, 0, PFN_UP(bo->tbo.base.size), &bo->kmap);
235 if (r) {
236 return r;
237 }
238 bo->kptr = ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
239 if (ptr) {
240 *ptr = bo->kptr;
241 }
242 radeon_bo_check_tiling(bo, 0, 0);
243 return 0;
244 }
245
radeon_bo_kunmap(struct radeon_bo * bo)246 void radeon_bo_kunmap(struct radeon_bo *bo)
247 {
248 if (bo->kptr == NULL)
249 return;
250 bo->kptr = NULL;
251 radeon_bo_check_tiling(bo, 0, 0);
252 ttm_bo_kunmap(&bo->kmap);
253 }
254
radeon_bo_ref(struct radeon_bo * bo)255 struct radeon_bo *radeon_bo_ref(struct radeon_bo *bo)
256 {
257 if (bo == NULL)
258 return NULL;
259
260 drm_gem_object_get(&bo->tbo.base);
261 return bo;
262 }
263
radeon_bo_unref(struct radeon_bo ** bo)264 void radeon_bo_unref(struct radeon_bo **bo)
265 {
266 if ((*bo) == NULL)
267 return;
268 drm_gem_object_put(&(*bo)->tbo.base);
269 *bo = NULL;
270 }
271
radeon_bo_pin_restricted(struct radeon_bo * bo,u32 domain,u64 max_offset,u64 * gpu_addr)272 int radeon_bo_pin_restricted(struct radeon_bo *bo, u32 domain, u64 max_offset,
273 u64 *gpu_addr)
274 {
275 struct ttm_operation_ctx ctx = { false, false };
276 int r, i;
277
278 if (radeon_ttm_tt_has_userptr(bo->rdev, bo->tbo.ttm))
279 return -EPERM;
280
281 if (bo->tbo.pin_count) {
282 ttm_bo_pin(&bo->tbo);
283 if (gpu_addr)
284 *gpu_addr = radeon_bo_gpu_offset(bo);
285
286 if (max_offset != 0) {
287 u64 domain_start;
288
289 if (domain == RADEON_GEM_DOMAIN_VRAM)
290 domain_start = bo->rdev->mc.vram_start;
291 else
292 domain_start = bo->rdev->mc.gtt_start;
293 WARN_ON_ONCE(max_offset <
294 (radeon_bo_gpu_offset(bo) - domain_start));
295 }
296
297 return 0;
298 }
299 if (bo->prime_shared_count && domain == RADEON_GEM_DOMAIN_VRAM) {
300 /* A BO shared as a dma-buf cannot be sensibly migrated to VRAM */
301 return -EINVAL;
302 }
303
304 radeon_ttm_placement_from_domain(bo, domain);
305 for (i = 0; i < bo->placement.num_placement; i++) {
306 /* force to pin into visible video ram */
307 if ((bo->placements[i].mem_type == TTM_PL_VRAM) &&
308 !(bo->flags & RADEON_GEM_NO_CPU_ACCESS) &&
309 (!max_offset || max_offset > bo->rdev->mc.visible_vram_size))
310 bo->placements[i].lpfn =
311 bo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
312 else
313 bo->placements[i].lpfn = max_offset >> PAGE_SHIFT;
314 }
315
316 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
317 if (likely(r == 0)) {
318 ttm_bo_pin(&bo->tbo);
319 if (gpu_addr != NULL)
320 *gpu_addr = radeon_bo_gpu_offset(bo);
321 if (domain == RADEON_GEM_DOMAIN_VRAM)
322 bo->rdev->vram_pin_size += radeon_bo_size(bo);
323 else
324 bo->rdev->gart_pin_size += radeon_bo_size(bo);
325 } else {
326 dev_err(bo->rdev->dev, "%p pin failed\n", bo);
327 }
328 return r;
329 }
330
radeon_bo_pin(struct radeon_bo * bo,u32 domain,u64 * gpu_addr)331 int radeon_bo_pin(struct radeon_bo *bo, u32 domain, u64 *gpu_addr)
332 {
333 return radeon_bo_pin_restricted(bo, domain, 0, gpu_addr);
334 }
335
radeon_bo_unpin(struct radeon_bo * bo)336 void radeon_bo_unpin(struct radeon_bo *bo)
337 {
338 ttm_bo_unpin(&bo->tbo);
339 if (!bo->tbo.pin_count) {
340 if (bo->tbo.resource->mem_type == TTM_PL_VRAM)
341 bo->rdev->vram_pin_size -= radeon_bo_size(bo);
342 else
343 bo->rdev->gart_pin_size -= radeon_bo_size(bo);
344 }
345 }
346
radeon_bo_evict_vram(struct radeon_device * rdev)347 int radeon_bo_evict_vram(struct radeon_device *rdev)
348 {
349 struct ttm_device *bdev = &rdev->mman.bdev;
350 struct ttm_resource_manager *man;
351
352 /* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
353 #ifndef CONFIG_HIBERNATION
354 if (rdev->flags & RADEON_IS_IGP) {
355 if (rdev->mc.igp_sideport_enabled == false)
356 /* Useless to evict on IGP chips */
357 return 0;
358 }
359 #endif
360 man = ttm_manager_type(bdev, TTM_PL_VRAM);
361 if (!man)
362 return 0;
363 return ttm_resource_manager_evict_all(bdev, man);
364 }
365
radeon_bo_force_delete(struct radeon_device * rdev)366 void radeon_bo_force_delete(struct radeon_device *rdev)
367 {
368 struct radeon_bo *bo, *n;
369
370 if (list_empty(&rdev->gem.objects)) {
371 return;
372 }
373 dev_err(rdev->dev, "Userspace still has active objects !\n");
374 list_for_each_entry_safe(bo, n, &rdev->gem.objects, list) {
375 dev_err(rdev->dev, "%p %p %lu %lu force free\n",
376 &bo->tbo.base, bo, (unsigned long)bo->tbo.base.size,
377 *((unsigned long *)&bo->tbo.base.refcount));
378 mutex_lock(&bo->rdev->gem.mutex);
379 list_del_init(&bo->list);
380 mutex_unlock(&bo->rdev->gem.mutex);
381 /* this should unref the ttm bo */
382 drm_gem_object_put(&bo->tbo.base);
383 }
384 }
385
radeon_bo_init(struct radeon_device * rdev)386 int radeon_bo_init(struct radeon_device *rdev)
387 {
388 /* reserve PAT memory space to WC for VRAM */
389 arch_io_reserve_memtype_wc(rdev->mc.aper_base,
390 rdev->mc.aper_size);
391
392 /* Add an MTRR for the VRAM */
393 if (!rdev->fastfb_working) {
394 rdev->mc.vram_mtrr = arch_phys_wc_add(rdev->mc.aper_base,
395 rdev->mc.aper_size);
396 }
397 DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
398 rdev->mc.mc_vram_size >> 20,
399 (unsigned long long)rdev->mc.aper_size >> 20);
400 DRM_INFO("RAM width %dbits %cDR\n",
401 rdev->mc.vram_width, rdev->mc.vram_is_ddr ? 'D' : 'S');
402 return radeon_ttm_init(rdev);
403 }
404
radeon_bo_fini(struct radeon_device * rdev)405 void radeon_bo_fini(struct radeon_device *rdev)
406 {
407 radeon_ttm_fini(rdev);
408 arch_phys_wc_del(rdev->mc.vram_mtrr);
409 arch_io_free_memtype_wc(rdev->mc.aper_base, rdev->mc.aper_size);
410 }
411
412 /* Returns how many bytes TTM can move per IB.
413 */
radeon_bo_get_threshold_for_moves(struct radeon_device * rdev)414 static u64 radeon_bo_get_threshold_for_moves(struct radeon_device *rdev)
415 {
416 u64 real_vram_size = rdev->mc.real_vram_size;
417 struct ttm_resource_manager *man =
418 ttm_manager_type(&rdev->mman.bdev, TTM_PL_VRAM);
419 u64 vram_usage = ttm_resource_manager_usage(man);
420
421 /* This function is based on the current VRAM usage.
422 *
423 * - If all of VRAM is free, allow relocating the number of bytes that
424 * is equal to 1/4 of the size of VRAM for this IB.
425
426 * - If more than one half of VRAM is occupied, only allow relocating
427 * 1 MB of data for this IB.
428 *
429 * - From 0 to one half of used VRAM, the threshold decreases
430 * linearly.
431 * __________________
432 * 1/4 of -|\ |
433 * VRAM | \ |
434 * | \ |
435 * | \ |
436 * | \ |
437 * | \ |
438 * | \ |
439 * | \________|1 MB
440 * |----------------|
441 * VRAM 0 % 100 %
442 * used used
443 *
444 * Note: It's a threshold, not a limit. The threshold must be crossed
445 * for buffer relocations to stop, so any buffer of an arbitrary size
446 * can be moved as long as the threshold isn't crossed before
447 * the relocation takes place. We don't want to disable buffer
448 * relocations completely.
449 *
450 * The idea is that buffers should be placed in VRAM at creation time
451 * and TTM should only do a minimum number of relocations during
452 * command submission. In practice, you need to submit at least
453 * a dozen IBs to move all buffers to VRAM if they are in GTT.
454 *
455 * Also, things can get pretty crazy under memory pressure and actual
456 * VRAM usage can change a lot, so playing safe even at 50% does
457 * consistently increase performance.
458 */
459
460 u64 half_vram = real_vram_size >> 1;
461 u64 half_free_vram = vram_usage >= half_vram ? 0 : half_vram - vram_usage;
462 u64 bytes_moved_threshold = half_free_vram >> 1;
463 return max(bytes_moved_threshold, 1024*1024ull);
464 }
465
radeon_bo_list_validate(struct radeon_device * rdev,struct drm_exec * exec,struct list_head * head,int ring)466 int radeon_bo_list_validate(struct radeon_device *rdev,
467 struct drm_exec *exec,
468 struct list_head *head, int ring)
469 {
470 struct ttm_operation_ctx ctx = { true, false };
471 struct radeon_bo_list *lobj;
472 u64 bytes_moved = 0, initial_bytes_moved;
473 u64 bytes_moved_threshold = radeon_bo_get_threshold_for_moves(rdev);
474 int r;
475
476 drm_exec_until_all_locked(exec) {
477 list_for_each_entry(lobj, head, list) {
478 r = drm_exec_prepare_obj(exec, &lobj->robj->tbo.base,
479 1);
480 drm_exec_retry_on_contention(exec);
481 if (unlikely(r && r != -EALREADY))
482 return r;
483 }
484 }
485
486 list_for_each_entry(lobj, head, list) {
487 struct radeon_bo *bo = lobj->robj;
488 if (!bo->tbo.pin_count) {
489 u32 domain = lobj->preferred_domains;
490 u32 allowed = lobj->allowed_domains;
491 u32 current_domain =
492 radeon_mem_type_to_domain(bo->tbo.resource->mem_type);
493
494 /* Check if this buffer will be moved and don't move it
495 * if we have moved too many buffers for this IB already.
496 *
497 * Note that this allows moving at least one buffer of
498 * any size, because it doesn't take the current "bo"
499 * into account. We don't want to disallow buffer moves
500 * completely.
501 */
502 if ((allowed & current_domain) != 0 &&
503 (domain & current_domain) == 0 && /* will be moved */
504 bytes_moved > bytes_moved_threshold) {
505 /* don't move it */
506 domain = current_domain;
507 }
508
509 retry:
510 radeon_ttm_placement_from_domain(bo, domain);
511 if (ring == R600_RING_TYPE_UVD_INDEX)
512 radeon_uvd_force_into_uvd_segment(bo, allowed);
513
514 initial_bytes_moved = atomic64_read(&rdev->num_bytes_moved);
515 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
516 bytes_moved += atomic64_read(&rdev->num_bytes_moved) -
517 initial_bytes_moved;
518
519 if (unlikely(r)) {
520 if (r != -ERESTARTSYS &&
521 domain != lobj->allowed_domains) {
522 domain = lobj->allowed_domains;
523 goto retry;
524 }
525 return r;
526 }
527 }
528 lobj->gpu_offset = radeon_bo_gpu_offset(bo);
529 lobj->tiling_flags = bo->tiling_flags;
530 }
531
532 return 0;
533 }
534
radeon_bo_get_surface_reg(struct radeon_bo * bo)535 int radeon_bo_get_surface_reg(struct radeon_bo *bo)
536 {
537 struct radeon_device *rdev = bo->rdev;
538 struct radeon_surface_reg *reg;
539 struct radeon_bo *old_object;
540 int steal;
541 int i;
542
543 dma_resv_assert_held(bo->tbo.base.resv);
544
545 if (!bo->tiling_flags)
546 return 0;
547
548 if (bo->surface_reg >= 0) {
549 i = bo->surface_reg;
550 goto out;
551 }
552
553 steal = -1;
554 for (i = 0; i < RADEON_GEM_MAX_SURFACES; i++) {
555
556 reg = &rdev->surface_regs[i];
557 if (!reg->bo)
558 break;
559
560 old_object = reg->bo;
561 if (old_object->tbo.pin_count == 0)
562 steal = i;
563 }
564
565 /* if we are all out */
566 if (i == RADEON_GEM_MAX_SURFACES) {
567 if (steal == -1)
568 return -ENOMEM;
569 /* find someone with a surface reg and nuke their BO */
570 reg = &rdev->surface_regs[steal];
571 old_object = reg->bo;
572 /* blow away the mapping */
573 DRM_DEBUG("stealing surface reg %d from %p\n", steal, old_object);
574 ttm_bo_unmap_virtual(&old_object->tbo);
575 old_object->surface_reg = -1;
576 i = steal;
577 }
578
579 bo->surface_reg = i;
580 reg->bo = bo;
581
582 out:
583 radeon_set_surface_reg(rdev, i, bo->tiling_flags, bo->pitch,
584 bo->tbo.resource->start << PAGE_SHIFT,
585 bo->tbo.base.size);
586 return 0;
587 }
588
radeon_bo_clear_surface_reg(struct radeon_bo * bo)589 static void radeon_bo_clear_surface_reg(struct radeon_bo *bo)
590 {
591 struct radeon_device *rdev = bo->rdev;
592 struct radeon_surface_reg *reg;
593
594 if (bo->surface_reg == -1)
595 return;
596
597 reg = &rdev->surface_regs[bo->surface_reg];
598 radeon_clear_surface_reg(rdev, bo->surface_reg);
599
600 reg->bo = NULL;
601 bo->surface_reg = -1;
602 }
603
radeon_bo_set_tiling_flags(struct radeon_bo * bo,uint32_t tiling_flags,uint32_t pitch)604 int radeon_bo_set_tiling_flags(struct radeon_bo *bo,
605 uint32_t tiling_flags, uint32_t pitch)
606 {
607 struct radeon_device *rdev = bo->rdev;
608 int r;
609
610 if (rdev->family >= CHIP_CEDAR) {
611 unsigned bankw, bankh, mtaspect, tilesplit, stilesplit;
612
613 bankw = (tiling_flags >> RADEON_TILING_EG_BANKW_SHIFT) & RADEON_TILING_EG_BANKW_MASK;
614 bankh = (tiling_flags >> RADEON_TILING_EG_BANKH_SHIFT) & RADEON_TILING_EG_BANKH_MASK;
615 mtaspect = (tiling_flags >> RADEON_TILING_EG_MACRO_TILE_ASPECT_SHIFT) & RADEON_TILING_EG_MACRO_TILE_ASPECT_MASK;
616 tilesplit = (tiling_flags >> RADEON_TILING_EG_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_TILE_SPLIT_MASK;
617 stilesplit = (tiling_flags >> RADEON_TILING_EG_STENCIL_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_STENCIL_TILE_SPLIT_MASK;
618 switch (bankw) {
619 case 0:
620 case 1:
621 case 2:
622 case 4:
623 case 8:
624 break;
625 default:
626 return -EINVAL;
627 }
628 switch (bankh) {
629 case 0:
630 case 1:
631 case 2:
632 case 4:
633 case 8:
634 break;
635 default:
636 return -EINVAL;
637 }
638 switch (mtaspect) {
639 case 0:
640 case 1:
641 case 2:
642 case 4:
643 case 8:
644 break;
645 default:
646 return -EINVAL;
647 }
648 if (tilesplit > 6) {
649 return -EINVAL;
650 }
651 if (stilesplit > 6) {
652 return -EINVAL;
653 }
654 }
655 r = radeon_bo_reserve(bo, false);
656 if (unlikely(r != 0))
657 return r;
658 bo->tiling_flags = tiling_flags;
659 bo->pitch = pitch;
660 radeon_bo_unreserve(bo);
661 return 0;
662 }
663
radeon_bo_get_tiling_flags(struct radeon_bo * bo,uint32_t * tiling_flags,uint32_t * pitch)664 void radeon_bo_get_tiling_flags(struct radeon_bo *bo,
665 uint32_t *tiling_flags,
666 uint32_t *pitch)
667 {
668 dma_resv_assert_held(bo->tbo.base.resv);
669
670 if (tiling_flags)
671 *tiling_flags = bo->tiling_flags;
672 if (pitch)
673 *pitch = bo->pitch;
674 }
675
radeon_bo_check_tiling(struct radeon_bo * bo,bool has_moved,bool force_drop)676 int radeon_bo_check_tiling(struct radeon_bo *bo, bool has_moved,
677 bool force_drop)
678 {
679 if (!force_drop)
680 dma_resv_assert_held(bo->tbo.base.resv);
681
682 if (!(bo->tiling_flags & RADEON_TILING_SURFACE))
683 return 0;
684
685 if (force_drop) {
686 radeon_bo_clear_surface_reg(bo);
687 return 0;
688 }
689
690 if (bo->tbo.resource->mem_type != TTM_PL_VRAM) {
691 if (!has_moved)
692 return 0;
693
694 if (bo->surface_reg >= 0)
695 radeon_bo_clear_surface_reg(bo);
696 return 0;
697 }
698
699 if ((bo->surface_reg >= 0) && !has_moved)
700 return 0;
701
702 return radeon_bo_get_surface_reg(bo);
703 }
704
radeon_bo_move_notify(struct ttm_buffer_object * bo)705 void radeon_bo_move_notify(struct ttm_buffer_object *bo)
706 {
707 struct radeon_bo *rbo;
708
709 if (!radeon_ttm_bo_is_radeon_bo(bo))
710 return;
711
712 rbo = container_of(bo, struct radeon_bo, tbo);
713 radeon_bo_check_tiling(rbo, 0, 1);
714 radeon_vm_bo_invalidate(rbo->rdev, rbo);
715 }
716
radeon_bo_fault_reserve_notify(struct ttm_buffer_object * bo)717 vm_fault_t radeon_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
718 {
719 struct ttm_operation_ctx ctx = { false, false };
720 struct radeon_device *rdev;
721 struct radeon_bo *rbo;
722 unsigned long offset, size, lpfn;
723 int i, r;
724
725 if (!radeon_ttm_bo_is_radeon_bo(bo))
726 return 0;
727 rbo = container_of(bo, struct radeon_bo, tbo);
728 radeon_bo_check_tiling(rbo, 0, 0);
729 rdev = rbo->rdev;
730 if (bo->resource->mem_type != TTM_PL_VRAM)
731 return 0;
732
733 size = bo->resource->size;
734 offset = bo->resource->start << PAGE_SHIFT;
735 if ((offset + size) <= rdev->mc.visible_vram_size)
736 return 0;
737
738 /* Can't move a pinned BO to visible VRAM */
739 if (rbo->tbo.pin_count > 0)
740 return VM_FAULT_SIGBUS;
741
742 /* hurrah the memory is not visible ! */
743 radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_VRAM);
744 lpfn = rdev->mc.visible_vram_size >> PAGE_SHIFT;
745 for (i = 0; i < rbo->placement.num_placement; i++) {
746 /* Force into visible VRAM */
747 if ((rbo->placements[i].mem_type == TTM_PL_VRAM) &&
748 (!rbo->placements[i].lpfn || rbo->placements[i].lpfn > lpfn))
749 rbo->placements[i].lpfn = lpfn;
750 }
751 r = ttm_bo_validate(bo, &rbo->placement, &ctx);
752 if (unlikely(r == -ENOMEM)) {
753 radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_GTT);
754 r = ttm_bo_validate(bo, &rbo->placement, &ctx);
755 } else if (likely(!r)) {
756 offset = bo->resource->start << PAGE_SHIFT;
757 /* this should never happen */
758 if ((offset + size) > rdev->mc.visible_vram_size)
759 return VM_FAULT_SIGBUS;
760 }
761
762 if (unlikely(r == -EBUSY || r == -ERESTARTSYS))
763 return VM_FAULT_NOPAGE;
764 else if (unlikely(r))
765 return VM_FAULT_SIGBUS;
766
767 ttm_bo_move_to_lru_tail_unlocked(bo);
768 return 0;
769 }
770
771 /**
772 * radeon_bo_fence - add fence to buffer object
773 *
774 * @bo: buffer object in question
775 * @fence: fence to add
776 * @shared: true if fence should be added shared
777 *
778 */
radeon_bo_fence(struct radeon_bo * bo,struct radeon_fence * fence,bool shared)779 void radeon_bo_fence(struct radeon_bo *bo, struct radeon_fence *fence,
780 bool shared)
781 {
782 struct dma_resv *resv = bo->tbo.base.resv;
783 int r;
784
785 r = dma_resv_reserve_fences(resv, 1);
786 if (r) {
787 /* As last resort on OOM we block for the fence */
788 dma_fence_wait(&fence->base, false);
789 return;
790 }
791
792 dma_resv_add_fence(resv, &fence->base, shared ?
793 DMA_RESV_USAGE_READ : DMA_RESV_USAGE_WRITE);
794 }
795