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