Loading...
1/* SPDX-License-Identifier: GPL-2.0 OR MIT */
2/**************************************************************************
3 *
4 * Copyright (c) 2007-2009 VMware, Inc., Palo Alto, CA., USA
5 * All Rights Reserved.
6 *
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the
9 * "Software"), to deal in the Software without restriction, including
10 * without limitation the rights to use, copy, modify, merge, publish,
11 * distribute, sub license, and/or sell copies of the Software, and to
12 * permit persons to whom the Software is furnished to do so, subject to
13 * the following conditions:
14 *
15 * The above copyright notice and this permission notice (including the
16 * next paragraph) shall be included in all copies or substantial portions
17 * of the Software.
18 *
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
22 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
23 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
24 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
25 * USE OR OTHER DEALINGS IN THE SOFTWARE.
26 *
27 **************************************************************************/
28/*
29 * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com>
30 */
31
32#include <drm/ttm/ttm_bo_driver.h>
33#include <drm/ttm/ttm_placement.h>
34#include <drm/drm_vma_manager.h>
35#include <linux/io.h>
36#include <linux/highmem.h>
37#include <linux/wait.h>
38#include <linux/slab.h>
39#include <linux/vmalloc.h>
40#include <linux/module.h>
41#include <linux/dma-resv.h>
42
43struct ttm_transfer_obj {
44 struct ttm_buffer_object base;
45 struct ttm_buffer_object *bo;
46};
47
48void ttm_bo_free_old_node(struct ttm_buffer_object *bo)
49{
50 ttm_bo_mem_put(bo, &bo->mem);
51}
52
53int ttm_bo_move_ttm(struct ttm_buffer_object *bo,
54 struct ttm_operation_ctx *ctx,
55 struct ttm_mem_reg *new_mem)
56{
57 struct ttm_tt *ttm = bo->ttm;
58 struct ttm_mem_reg *old_mem = &bo->mem;
59 int ret;
60
61 if (old_mem->mem_type != TTM_PL_SYSTEM) {
62 ret = ttm_bo_wait(bo, ctx->interruptible, ctx->no_wait_gpu);
63
64 if (unlikely(ret != 0)) {
65 if (ret != -ERESTARTSYS)
66 pr_err("Failed to expire sync object before unbinding TTM\n");
67 return ret;
68 }
69
70 ttm_tt_unbind(ttm);
71 ttm_bo_free_old_node(bo);
72 ttm_flag_masked(&old_mem->placement, TTM_PL_FLAG_SYSTEM,
73 TTM_PL_MASK_MEM);
74 old_mem->mem_type = TTM_PL_SYSTEM;
75 }
76
77 ret = ttm_tt_set_placement_caching(ttm, new_mem->placement);
78 if (unlikely(ret != 0))
79 return ret;
80
81 if (new_mem->mem_type != TTM_PL_SYSTEM) {
82 ret = ttm_tt_bind(ttm, new_mem, ctx);
83 if (unlikely(ret != 0))
84 return ret;
85 }
86
87 *old_mem = *new_mem;
88 new_mem->mm_node = NULL;
89
90 return 0;
91}
92EXPORT_SYMBOL(ttm_bo_move_ttm);
93
94int ttm_mem_io_lock(struct ttm_mem_type_manager *man, bool interruptible)
95{
96 if (likely(man->io_reserve_fastpath))
97 return 0;
98
99 if (interruptible)
100 return mutex_lock_interruptible(&man->io_reserve_mutex);
101
102 mutex_lock(&man->io_reserve_mutex);
103 return 0;
104}
105EXPORT_SYMBOL(ttm_mem_io_lock);
106
107void ttm_mem_io_unlock(struct ttm_mem_type_manager *man)
108{
109 if (likely(man->io_reserve_fastpath))
110 return;
111
112 mutex_unlock(&man->io_reserve_mutex);
113}
114EXPORT_SYMBOL(ttm_mem_io_unlock);
115
116static int ttm_mem_io_evict(struct ttm_mem_type_manager *man)
117{
118 struct ttm_buffer_object *bo;
119
120 if (!man->use_io_reserve_lru || list_empty(&man->io_reserve_lru))
121 return -EAGAIN;
122
123 bo = list_first_entry(&man->io_reserve_lru,
124 struct ttm_buffer_object,
125 io_reserve_lru);
126 list_del_init(&bo->io_reserve_lru);
127 ttm_bo_unmap_virtual_locked(bo);
128
129 return 0;
130}
131
132
133int ttm_mem_io_reserve(struct ttm_bo_device *bdev,
134 struct ttm_mem_reg *mem)
135{
136 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
137 int ret = 0;
138
139 if (!bdev->driver->io_mem_reserve)
140 return 0;
141 if (likely(man->io_reserve_fastpath))
142 return bdev->driver->io_mem_reserve(bdev, mem);
143
144 if (bdev->driver->io_mem_reserve &&
145 mem->bus.io_reserved_count++ == 0) {
146retry:
147 ret = bdev->driver->io_mem_reserve(bdev, mem);
148 if (ret == -EAGAIN) {
149 ret = ttm_mem_io_evict(man);
150 if (ret == 0)
151 goto retry;
152 }
153 }
154 return ret;
155}
156EXPORT_SYMBOL(ttm_mem_io_reserve);
157
158void ttm_mem_io_free(struct ttm_bo_device *bdev,
159 struct ttm_mem_reg *mem)
160{
161 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
162
163 if (likely(man->io_reserve_fastpath))
164 return;
165
166 if (bdev->driver->io_mem_reserve &&
167 --mem->bus.io_reserved_count == 0 &&
168 bdev->driver->io_mem_free)
169 bdev->driver->io_mem_free(bdev, mem);
170
171}
172EXPORT_SYMBOL(ttm_mem_io_free);
173
174int ttm_mem_io_reserve_vm(struct ttm_buffer_object *bo)
175{
176 struct ttm_mem_reg *mem = &bo->mem;
177 int ret;
178
179 if (!mem->bus.io_reserved_vm) {
180 struct ttm_mem_type_manager *man =
181 &bo->bdev->man[mem->mem_type];
182
183 ret = ttm_mem_io_reserve(bo->bdev, mem);
184 if (unlikely(ret != 0))
185 return ret;
186 mem->bus.io_reserved_vm = true;
187 if (man->use_io_reserve_lru)
188 list_add_tail(&bo->io_reserve_lru,
189 &man->io_reserve_lru);
190 }
191 return 0;
192}
193
194void ttm_mem_io_free_vm(struct ttm_buffer_object *bo)
195{
196 struct ttm_mem_reg *mem = &bo->mem;
197
198 if (mem->bus.io_reserved_vm) {
199 mem->bus.io_reserved_vm = false;
200 list_del_init(&bo->io_reserve_lru);
201 ttm_mem_io_free(bo->bdev, mem);
202 }
203}
204
205static int ttm_mem_reg_ioremap(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem,
206 void **virtual)
207{
208 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
209 int ret;
210 void *addr;
211
212 *virtual = NULL;
213 (void) ttm_mem_io_lock(man, false);
214 ret = ttm_mem_io_reserve(bdev, mem);
215 ttm_mem_io_unlock(man);
216 if (ret || !mem->bus.is_iomem)
217 return ret;
218
219 if (mem->bus.addr) {
220 addr = mem->bus.addr;
221 } else {
222 if (mem->placement & TTM_PL_FLAG_WC)
223 addr = ioremap_wc(mem->bus.base + mem->bus.offset, mem->bus.size);
224 else
225 addr = ioremap_nocache(mem->bus.base + mem->bus.offset, mem->bus.size);
226 if (!addr) {
227 (void) ttm_mem_io_lock(man, false);
228 ttm_mem_io_free(bdev, mem);
229 ttm_mem_io_unlock(man);
230 return -ENOMEM;
231 }
232 }
233 *virtual = addr;
234 return 0;
235}
236
237static void ttm_mem_reg_iounmap(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem,
238 void *virtual)
239{
240 struct ttm_mem_type_manager *man;
241
242 man = &bdev->man[mem->mem_type];
243
244 if (virtual && mem->bus.addr == NULL)
245 iounmap(virtual);
246 (void) ttm_mem_io_lock(man, false);
247 ttm_mem_io_free(bdev, mem);
248 ttm_mem_io_unlock(man);
249}
250
251static int ttm_copy_io_page(void *dst, void *src, unsigned long page)
252{
253 uint32_t *dstP =
254 (uint32_t *) ((unsigned long)dst + (page << PAGE_SHIFT));
255 uint32_t *srcP =
256 (uint32_t *) ((unsigned long)src + (page << PAGE_SHIFT));
257
258 int i;
259 for (i = 0; i < PAGE_SIZE / sizeof(uint32_t); ++i)
260 iowrite32(ioread32(srcP++), dstP++);
261 return 0;
262}
263
264#ifdef CONFIG_X86
265#define __ttm_kmap_atomic_prot(__page, __prot) kmap_atomic_prot(__page, __prot)
266#define __ttm_kunmap_atomic(__addr) kunmap_atomic(__addr)
267#else
268#define __ttm_kmap_atomic_prot(__page, __prot) vmap(&__page, 1, 0, __prot)
269#define __ttm_kunmap_atomic(__addr) vunmap(__addr)
270#endif
271
272
273/**
274 * ttm_kmap_atomic_prot - Efficient kernel map of a single page with
275 * specified page protection.
276 *
277 * @page: The page to map.
278 * @prot: The page protection.
279 *
280 * This function maps a TTM page using the kmap_atomic api if available,
281 * otherwise falls back to vmap. The user must make sure that the
282 * specified page does not have an aliased mapping with a different caching
283 * policy unless the architecture explicitly allows it. Also mapping and
284 * unmapping using this api must be correctly nested. Unmapping should
285 * occur in the reverse order of mapping.
286 */
287void *ttm_kmap_atomic_prot(struct page *page, pgprot_t prot)
288{
289 if (pgprot_val(prot) == pgprot_val(PAGE_KERNEL))
290 return kmap_atomic(page);
291 else
292 return __ttm_kmap_atomic_prot(page, prot);
293}
294EXPORT_SYMBOL(ttm_kmap_atomic_prot);
295
296/**
297 * ttm_kunmap_atomic_prot - Unmap a page that was mapped using
298 * ttm_kmap_atomic_prot.
299 *
300 * @addr: The virtual address from the map.
301 * @prot: The page protection.
302 */
303void ttm_kunmap_atomic_prot(void *addr, pgprot_t prot)
304{
305 if (pgprot_val(prot) == pgprot_val(PAGE_KERNEL))
306 kunmap_atomic(addr);
307 else
308 __ttm_kunmap_atomic(addr);
309}
310EXPORT_SYMBOL(ttm_kunmap_atomic_prot);
311
312static int ttm_copy_io_ttm_page(struct ttm_tt *ttm, void *src,
313 unsigned long page,
314 pgprot_t prot)
315{
316 struct page *d = ttm->pages[page];
317 void *dst;
318
319 if (!d)
320 return -ENOMEM;
321
322 src = (void *)((unsigned long)src + (page << PAGE_SHIFT));
323 dst = ttm_kmap_atomic_prot(d, prot);
324 if (!dst)
325 return -ENOMEM;
326
327 memcpy_fromio(dst, src, PAGE_SIZE);
328
329 ttm_kunmap_atomic_prot(dst, prot);
330
331 return 0;
332}
333
334static int ttm_copy_ttm_io_page(struct ttm_tt *ttm, void *dst,
335 unsigned long page,
336 pgprot_t prot)
337{
338 struct page *s = ttm->pages[page];
339 void *src;
340
341 if (!s)
342 return -ENOMEM;
343
344 dst = (void *)((unsigned long)dst + (page << PAGE_SHIFT));
345 src = ttm_kmap_atomic_prot(s, prot);
346 if (!src)
347 return -ENOMEM;
348
349 memcpy_toio(dst, src, PAGE_SIZE);
350
351 ttm_kunmap_atomic_prot(src, prot);
352
353 return 0;
354}
355
356int ttm_bo_move_memcpy(struct ttm_buffer_object *bo,
357 struct ttm_operation_ctx *ctx,
358 struct ttm_mem_reg *new_mem)
359{
360 struct ttm_bo_device *bdev = bo->bdev;
361 struct ttm_mem_type_manager *man = &bdev->man[new_mem->mem_type];
362 struct ttm_tt *ttm = bo->ttm;
363 struct ttm_mem_reg *old_mem = &bo->mem;
364 struct ttm_mem_reg old_copy = *old_mem;
365 void *old_iomap;
366 void *new_iomap;
367 int ret;
368 unsigned long i;
369 unsigned long page;
370 unsigned long add = 0;
371 int dir;
372
373 ret = ttm_bo_wait(bo, ctx->interruptible, ctx->no_wait_gpu);
374 if (ret)
375 return ret;
376
377 ret = ttm_mem_reg_ioremap(bdev, old_mem, &old_iomap);
378 if (ret)
379 return ret;
380 ret = ttm_mem_reg_ioremap(bdev, new_mem, &new_iomap);
381 if (ret)
382 goto out;
383
384 /*
385 * Single TTM move. NOP.
386 */
387 if (old_iomap == NULL && new_iomap == NULL)
388 goto out2;
389
390 /*
391 * Don't move nonexistent data. Clear destination instead.
392 */
393 if (old_iomap == NULL &&
394 (ttm == NULL || (ttm->state == tt_unpopulated &&
395 !(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)))) {
396 memset_io(new_iomap, 0, new_mem->num_pages*PAGE_SIZE);
397 goto out2;
398 }
399
400 /*
401 * TTM might be null for moves within the same region.
402 */
403 if (ttm) {
404 ret = ttm_tt_populate(ttm, ctx);
405 if (ret)
406 goto out1;
407 }
408
409 add = 0;
410 dir = 1;
411
412 if ((old_mem->mem_type == new_mem->mem_type) &&
413 (new_mem->start < old_mem->start + old_mem->size)) {
414 dir = -1;
415 add = new_mem->num_pages - 1;
416 }
417
418 for (i = 0; i < new_mem->num_pages; ++i) {
419 page = i * dir + add;
420 if (old_iomap == NULL) {
421 pgprot_t prot = ttm_io_prot(old_mem->placement,
422 PAGE_KERNEL);
423 ret = ttm_copy_ttm_io_page(ttm, new_iomap, page,
424 prot);
425 } else if (new_iomap == NULL) {
426 pgprot_t prot = ttm_io_prot(new_mem->placement,
427 PAGE_KERNEL);
428 ret = ttm_copy_io_ttm_page(ttm, old_iomap, page,
429 prot);
430 } else {
431 ret = ttm_copy_io_page(new_iomap, old_iomap, page);
432 }
433 if (ret)
434 goto out1;
435 }
436 mb();
437out2:
438 old_copy = *old_mem;
439 *old_mem = *new_mem;
440 new_mem->mm_node = NULL;
441
442 if (man->flags & TTM_MEMTYPE_FLAG_FIXED) {
443 ttm_tt_destroy(ttm);
444 bo->ttm = NULL;
445 }
446
447out1:
448 ttm_mem_reg_iounmap(bdev, old_mem, new_iomap);
449out:
450 ttm_mem_reg_iounmap(bdev, &old_copy, old_iomap);
451
452 /*
453 * On error, keep the mm node!
454 */
455 if (!ret)
456 ttm_bo_mem_put(bo, &old_copy);
457 return ret;
458}
459EXPORT_SYMBOL(ttm_bo_move_memcpy);
460
461static void ttm_transfered_destroy(struct ttm_buffer_object *bo)
462{
463 struct ttm_transfer_obj *fbo;
464
465 fbo = container_of(bo, struct ttm_transfer_obj, base);
466 ttm_bo_put(fbo->bo);
467 kfree(fbo);
468}
469
470/**
471 * ttm_buffer_object_transfer
472 *
473 * @bo: A pointer to a struct ttm_buffer_object.
474 * @new_obj: A pointer to a pointer to a newly created ttm_buffer_object,
475 * holding the data of @bo with the old placement.
476 *
477 * This is a utility function that may be called after an accelerated move
478 * has been scheduled. A new buffer object is created as a placeholder for
479 * the old data while it's being copied. When that buffer object is idle,
480 * it can be destroyed, releasing the space of the old placement.
481 * Returns:
482 * !0: Failure.
483 */
484
485static int ttm_buffer_object_transfer(struct ttm_buffer_object *bo,
486 struct ttm_buffer_object **new_obj)
487{
488 struct ttm_transfer_obj *fbo;
489 int ret;
490
491 fbo = kmalloc(sizeof(*fbo), GFP_KERNEL);
492 if (!fbo)
493 return -ENOMEM;
494
495 fbo->base = *bo;
496 fbo->base.mem.placement |= TTM_PL_FLAG_NO_EVICT;
497
498 ttm_bo_get(bo);
499 fbo->bo = bo;
500
501 /**
502 * Fix up members that we shouldn't copy directly:
503 * TODO: Explicit member copy would probably be better here.
504 */
505
506 atomic_inc(&bo->bdev->glob->bo_count);
507 INIT_LIST_HEAD(&fbo->base.ddestroy);
508 INIT_LIST_HEAD(&fbo->base.lru);
509 INIT_LIST_HEAD(&fbo->base.swap);
510 INIT_LIST_HEAD(&fbo->base.io_reserve_lru);
511 mutex_init(&fbo->base.wu_mutex);
512 fbo->base.moving = NULL;
513 drm_vma_node_reset(&fbo->base.base.vma_node);
514 atomic_set(&fbo->base.cpu_writers, 0);
515
516 kref_init(&fbo->base.list_kref);
517 kref_init(&fbo->base.kref);
518 fbo->base.destroy = &ttm_transfered_destroy;
519 fbo->base.acc_size = 0;
520 fbo->base.base.resv = &fbo->base.base._resv;
521 dma_resv_init(fbo->base.base.resv);
522 ret = dma_resv_trylock(fbo->base.base.resv);
523 WARN_ON(!ret);
524
525 *new_obj = &fbo->base;
526 return 0;
527}
528
529pgprot_t ttm_io_prot(uint32_t caching_flags, pgprot_t tmp)
530{
531 /* Cached mappings need no adjustment */
532 if (caching_flags & TTM_PL_FLAG_CACHED)
533 return tmp;
534
535#if defined(__i386__) || defined(__x86_64__)
536 if (caching_flags & TTM_PL_FLAG_WC)
537 tmp = pgprot_writecombine(tmp);
538 else if (boot_cpu_data.x86 > 3)
539 tmp = pgprot_noncached(tmp);
540#endif
541#if defined(__ia64__) || defined(__arm__) || defined(__aarch64__) || \
542 defined(__powerpc__) || defined(__mips__)
543 if (caching_flags & TTM_PL_FLAG_WC)
544 tmp = pgprot_writecombine(tmp);
545 else
546 tmp = pgprot_noncached(tmp);
547#endif
548#if defined(__sparc__)
549 tmp = pgprot_noncached(tmp);
550#endif
551 return tmp;
552}
553EXPORT_SYMBOL(ttm_io_prot);
554
555static int ttm_bo_ioremap(struct ttm_buffer_object *bo,
556 unsigned long offset,
557 unsigned long size,
558 struct ttm_bo_kmap_obj *map)
559{
560 struct ttm_mem_reg *mem = &bo->mem;
561
562 if (bo->mem.bus.addr) {
563 map->bo_kmap_type = ttm_bo_map_premapped;
564 map->virtual = (void *)(((u8 *)bo->mem.bus.addr) + offset);
565 } else {
566 map->bo_kmap_type = ttm_bo_map_iomap;
567 if (mem->placement & TTM_PL_FLAG_WC)
568 map->virtual = ioremap_wc(bo->mem.bus.base + bo->mem.bus.offset + offset,
569 size);
570 else
571 map->virtual = ioremap_nocache(bo->mem.bus.base + bo->mem.bus.offset + offset,
572 size);
573 }
574 return (!map->virtual) ? -ENOMEM : 0;
575}
576
577static int ttm_bo_kmap_ttm(struct ttm_buffer_object *bo,
578 unsigned long start_page,
579 unsigned long num_pages,
580 struct ttm_bo_kmap_obj *map)
581{
582 struct ttm_mem_reg *mem = &bo->mem;
583 struct ttm_operation_ctx ctx = {
584 .interruptible = false,
585 .no_wait_gpu = false
586 };
587 struct ttm_tt *ttm = bo->ttm;
588 pgprot_t prot;
589 int ret;
590
591 BUG_ON(!ttm);
592
593 ret = ttm_tt_populate(ttm, &ctx);
594 if (ret)
595 return ret;
596
597 if (num_pages == 1 && (mem->placement & TTM_PL_FLAG_CACHED)) {
598 /*
599 * We're mapping a single page, and the desired
600 * page protection is consistent with the bo.
601 */
602
603 map->bo_kmap_type = ttm_bo_map_kmap;
604 map->page = ttm->pages[start_page];
605 map->virtual = kmap(map->page);
606 } else {
607 /*
608 * We need to use vmap to get the desired page protection
609 * or to make the buffer object look contiguous.
610 */
611 prot = ttm_io_prot(mem->placement, PAGE_KERNEL);
612 map->bo_kmap_type = ttm_bo_map_vmap;
613 map->virtual = vmap(ttm->pages + start_page, num_pages,
614 0, prot);
615 }
616 return (!map->virtual) ? -ENOMEM : 0;
617}
618
619int ttm_bo_kmap(struct ttm_buffer_object *bo,
620 unsigned long start_page, unsigned long num_pages,
621 struct ttm_bo_kmap_obj *map)
622{
623 struct ttm_mem_type_manager *man =
624 &bo->bdev->man[bo->mem.mem_type];
625 unsigned long offset, size;
626 int ret;
627
628 map->virtual = NULL;
629 map->bo = bo;
630 if (num_pages > bo->num_pages)
631 return -EINVAL;
632 if (start_page > bo->num_pages)
633 return -EINVAL;
634
635 (void) ttm_mem_io_lock(man, false);
636 ret = ttm_mem_io_reserve(bo->bdev, &bo->mem);
637 ttm_mem_io_unlock(man);
638 if (ret)
639 return ret;
640 if (!bo->mem.bus.is_iomem) {
641 return ttm_bo_kmap_ttm(bo, start_page, num_pages, map);
642 } else {
643 offset = start_page << PAGE_SHIFT;
644 size = num_pages << PAGE_SHIFT;
645 return ttm_bo_ioremap(bo, offset, size, map);
646 }
647}
648EXPORT_SYMBOL(ttm_bo_kmap);
649
650void ttm_bo_kunmap(struct ttm_bo_kmap_obj *map)
651{
652 struct ttm_buffer_object *bo = map->bo;
653 struct ttm_mem_type_manager *man =
654 &bo->bdev->man[bo->mem.mem_type];
655
656 if (!map->virtual)
657 return;
658 switch (map->bo_kmap_type) {
659 case ttm_bo_map_iomap:
660 iounmap(map->virtual);
661 break;
662 case ttm_bo_map_vmap:
663 vunmap(map->virtual);
664 break;
665 case ttm_bo_map_kmap:
666 kunmap(map->page);
667 break;
668 case ttm_bo_map_premapped:
669 break;
670 default:
671 BUG();
672 }
673 (void) ttm_mem_io_lock(man, false);
674 ttm_mem_io_free(map->bo->bdev, &map->bo->mem);
675 ttm_mem_io_unlock(man);
676 map->virtual = NULL;
677 map->page = NULL;
678}
679EXPORT_SYMBOL(ttm_bo_kunmap);
680
681int ttm_bo_move_accel_cleanup(struct ttm_buffer_object *bo,
682 struct dma_fence *fence,
683 bool evict,
684 struct ttm_mem_reg *new_mem)
685{
686 struct ttm_bo_device *bdev = bo->bdev;
687 struct ttm_mem_type_manager *man = &bdev->man[new_mem->mem_type];
688 struct ttm_mem_reg *old_mem = &bo->mem;
689 int ret;
690 struct ttm_buffer_object *ghost_obj;
691
692 dma_resv_add_excl_fence(bo->base.resv, fence);
693 if (evict) {
694 ret = ttm_bo_wait(bo, false, false);
695 if (ret)
696 return ret;
697
698 if (man->flags & TTM_MEMTYPE_FLAG_FIXED) {
699 ttm_tt_destroy(bo->ttm);
700 bo->ttm = NULL;
701 }
702 ttm_bo_free_old_node(bo);
703 } else {
704 /**
705 * This should help pipeline ordinary buffer moves.
706 *
707 * Hang old buffer memory on a new buffer object,
708 * and leave it to be released when the GPU
709 * operation has completed.
710 */
711
712 dma_fence_put(bo->moving);
713 bo->moving = dma_fence_get(fence);
714
715 ret = ttm_buffer_object_transfer(bo, &ghost_obj);
716 if (ret)
717 return ret;
718
719 dma_resv_add_excl_fence(ghost_obj->base.resv, fence);
720
721 /**
722 * If we're not moving to fixed memory, the TTM object
723 * needs to stay alive. Otherwhise hang it on the ghost
724 * bo to be unbound and destroyed.
725 */
726
727 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED))
728 ghost_obj->ttm = NULL;
729 else
730 bo->ttm = NULL;
731
732 ttm_bo_unreserve(ghost_obj);
733 ttm_bo_put(ghost_obj);
734 }
735
736 *old_mem = *new_mem;
737 new_mem->mm_node = NULL;
738
739 return 0;
740}
741EXPORT_SYMBOL(ttm_bo_move_accel_cleanup);
742
743int ttm_bo_pipeline_move(struct ttm_buffer_object *bo,
744 struct dma_fence *fence, bool evict,
745 struct ttm_mem_reg *new_mem)
746{
747 struct ttm_bo_device *bdev = bo->bdev;
748 struct ttm_mem_reg *old_mem = &bo->mem;
749
750 struct ttm_mem_type_manager *from = &bdev->man[old_mem->mem_type];
751 struct ttm_mem_type_manager *to = &bdev->man[new_mem->mem_type];
752
753 int ret;
754
755 dma_resv_add_excl_fence(bo->base.resv, fence);
756
757 if (!evict) {
758 struct ttm_buffer_object *ghost_obj;
759
760 /**
761 * This should help pipeline ordinary buffer moves.
762 *
763 * Hang old buffer memory on a new buffer object,
764 * and leave it to be released when the GPU
765 * operation has completed.
766 */
767
768 dma_fence_put(bo->moving);
769 bo->moving = dma_fence_get(fence);
770
771 ret = ttm_buffer_object_transfer(bo, &ghost_obj);
772 if (ret)
773 return ret;
774
775 dma_resv_add_excl_fence(ghost_obj->base.resv, fence);
776
777 /**
778 * If we're not moving to fixed memory, the TTM object
779 * needs to stay alive. Otherwhise hang it on the ghost
780 * bo to be unbound and destroyed.
781 */
782
783 if (!(to->flags & TTM_MEMTYPE_FLAG_FIXED))
784 ghost_obj->ttm = NULL;
785 else
786 bo->ttm = NULL;
787
788 ttm_bo_unreserve(ghost_obj);
789 ttm_bo_put(ghost_obj);
790
791 } else if (from->flags & TTM_MEMTYPE_FLAG_FIXED) {
792
793 /**
794 * BO doesn't have a TTM we need to bind/unbind. Just remember
795 * this eviction and free up the allocation
796 */
797
798 spin_lock(&from->move_lock);
799 if (!from->move || dma_fence_is_later(fence, from->move)) {
800 dma_fence_put(from->move);
801 from->move = dma_fence_get(fence);
802 }
803 spin_unlock(&from->move_lock);
804
805 ttm_bo_free_old_node(bo);
806
807 dma_fence_put(bo->moving);
808 bo->moving = dma_fence_get(fence);
809
810 } else {
811 /**
812 * Last resort, wait for the move to be completed.
813 *
814 * Should never happen in pratice.
815 */
816
817 ret = ttm_bo_wait(bo, false, false);
818 if (ret)
819 return ret;
820
821 if (to->flags & TTM_MEMTYPE_FLAG_FIXED) {
822 ttm_tt_destroy(bo->ttm);
823 bo->ttm = NULL;
824 }
825 ttm_bo_free_old_node(bo);
826 }
827
828 *old_mem = *new_mem;
829 new_mem->mm_node = NULL;
830
831 return 0;
832}
833EXPORT_SYMBOL(ttm_bo_pipeline_move);
834
835int ttm_bo_pipeline_gutting(struct ttm_buffer_object *bo)
836{
837 struct ttm_buffer_object *ghost;
838 int ret;
839
840 ret = ttm_buffer_object_transfer(bo, &ghost);
841 if (ret)
842 return ret;
843
844 ret = dma_resv_copy_fences(ghost->base.resv, bo->base.resv);
845 /* Last resort, wait for the BO to be idle when we are OOM */
846 if (ret)
847 ttm_bo_wait(bo, false, false);
848
849 memset(&bo->mem, 0, sizeof(bo->mem));
850 bo->mem.mem_type = TTM_PL_SYSTEM;
851 bo->ttm = NULL;
852
853 ttm_bo_unreserve(ghost);
854 ttm_bo_put(ghost);
855
856 return 0;
857}
1/* SPDX-License-Identifier: GPL-2.0 OR MIT */
2/**************************************************************************
3 *
4 * Copyright (c) 2007-2009 VMware, Inc., Palo Alto, CA., USA
5 * All Rights Reserved.
6 *
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the
9 * "Software"), to deal in the Software without restriction, including
10 * without limitation the rights to use, copy, modify, merge, publish,
11 * distribute, sub license, and/or sell copies of the Software, and to
12 * permit persons to whom the Software is furnished to do so, subject to
13 * the following conditions:
14 *
15 * The above copyright notice and this permission notice (including the
16 * next paragraph) shall be included in all copies or substantial portions
17 * of the Software.
18 *
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
22 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
23 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
24 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
25 * USE OR OTHER DEALINGS IN THE SOFTWARE.
26 *
27 **************************************************************************/
28/*
29 * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com>
30 */
31
32#include <linux/vmalloc.h>
33
34#include <drm/ttm/ttm_bo.h>
35#include <drm/ttm/ttm_placement.h>
36#include <drm/ttm/ttm_tt.h>
37
38#include <drm/drm_cache.h>
39
40struct ttm_transfer_obj {
41 struct ttm_buffer_object base;
42 struct ttm_buffer_object *bo;
43};
44
45int ttm_mem_io_reserve(struct ttm_device *bdev,
46 struct ttm_resource *mem)
47{
48 if (mem->bus.offset || mem->bus.addr)
49 return 0;
50
51 mem->bus.is_iomem = false;
52 if (!bdev->funcs->io_mem_reserve)
53 return 0;
54
55 return bdev->funcs->io_mem_reserve(bdev, mem);
56}
57
58void ttm_mem_io_free(struct ttm_device *bdev,
59 struct ttm_resource *mem)
60{
61 if (!mem)
62 return;
63
64 if (!mem->bus.offset && !mem->bus.addr)
65 return;
66
67 if (bdev->funcs->io_mem_free)
68 bdev->funcs->io_mem_free(bdev, mem);
69
70 mem->bus.offset = 0;
71 mem->bus.addr = NULL;
72}
73
74/**
75 * ttm_move_memcpy - Helper to perform a memcpy ttm move operation.
76 * @clear: Whether to clear rather than copy.
77 * @num_pages: Number of pages of the operation.
78 * @dst_iter: A struct ttm_kmap_iter representing the destination resource.
79 * @src_iter: A struct ttm_kmap_iter representing the source resource.
80 *
81 * This function is intended to be able to move out async under a
82 * dma-fence if desired.
83 */
84void ttm_move_memcpy(bool clear,
85 u32 num_pages,
86 struct ttm_kmap_iter *dst_iter,
87 struct ttm_kmap_iter *src_iter)
88{
89 const struct ttm_kmap_iter_ops *dst_ops = dst_iter->ops;
90 const struct ttm_kmap_iter_ops *src_ops = src_iter->ops;
91 struct iosys_map src_map, dst_map;
92 pgoff_t i;
93
94 /* Single TTM move. NOP */
95 if (dst_ops->maps_tt && src_ops->maps_tt)
96 return;
97
98 /* Don't move nonexistent data. Clear destination instead. */
99 if (clear) {
100 for (i = 0; i < num_pages; ++i) {
101 dst_ops->map_local(dst_iter, &dst_map, i);
102 if (dst_map.is_iomem)
103 memset_io(dst_map.vaddr_iomem, 0, PAGE_SIZE);
104 else
105 memset(dst_map.vaddr, 0, PAGE_SIZE);
106 if (dst_ops->unmap_local)
107 dst_ops->unmap_local(dst_iter, &dst_map);
108 }
109 return;
110 }
111
112 for (i = 0; i < num_pages; ++i) {
113 dst_ops->map_local(dst_iter, &dst_map, i);
114 src_ops->map_local(src_iter, &src_map, i);
115
116 drm_memcpy_from_wc(&dst_map, &src_map, PAGE_SIZE);
117
118 if (src_ops->unmap_local)
119 src_ops->unmap_local(src_iter, &src_map);
120 if (dst_ops->unmap_local)
121 dst_ops->unmap_local(dst_iter, &dst_map);
122 }
123}
124EXPORT_SYMBOL(ttm_move_memcpy);
125
126/**
127 * ttm_bo_move_memcpy
128 *
129 * @bo: A pointer to a struct ttm_buffer_object.
130 * @ctx: operation context
131 * @dst_mem: struct ttm_resource indicating where to move.
132 *
133 * Fallback move function for a mappable buffer object in mappable memory.
134 * The function will, if successful,
135 * free any old aperture space, and set (@new_mem)->mm_node to NULL,
136 * and update the (@bo)->mem placement flags. If unsuccessful, the old
137 * data remains untouched, and it's up to the caller to free the
138 * memory space indicated by @new_mem.
139 * Returns:
140 * !0: Failure.
141 */
142int ttm_bo_move_memcpy(struct ttm_buffer_object *bo,
143 struct ttm_operation_ctx *ctx,
144 struct ttm_resource *dst_mem)
145{
146 struct ttm_device *bdev = bo->bdev;
147 struct ttm_resource_manager *dst_man =
148 ttm_manager_type(bo->bdev, dst_mem->mem_type);
149 struct ttm_tt *ttm = bo->ttm;
150 struct ttm_resource *src_mem = bo->resource;
151 struct ttm_resource_manager *src_man;
152 union {
153 struct ttm_kmap_iter_tt tt;
154 struct ttm_kmap_iter_linear_io io;
155 } _dst_iter, _src_iter;
156 struct ttm_kmap_iter *dst_iter, *src_iter;
157 bool clear;
158 int ret = 0;
159
160 if (WARN_ON(!src_mem))
161 return -EINVAL;
162
163 src_man = ttm_manager_type(bdev, src_mem->mem_type);
164 if (ttm && ((ttm->page_flags & TTM_TT_FLAG_SWAPPED) ||
165 dst_man->use_tt)) {
166 ret = ttm_bo_populate(bo, ctx);
167 if (ret)
168 return ret;
169 }
170
171 dst_iter = ttm_kmap_iter_linear_io_init(&_dst_iter.io, bdev, dst_mem);
172 if (PTR_ERR(dst_iter) == -EINVAL && dst_man->use_tt)
173 dst_iter = ttm_kmap_iter_tt_init(&_dst_iter.tt, bo->ttm);
174 if (IS_ERR(dst_iter))
175 return PTR_ERR(dst_iter);
176
177 src_iter = ttm_kmap_iter_linear_io_init(&_src_iter.io, bdev, src_mem);
178 if (PTR_ERR(src_iter) == -EINVAL && src_man->use_tt)
179 src_iter = ttm_kmap_iter_tt_init(&_src_iter.tt, bo->ttm);
180 if (IS_ERR(src_iter)) {
181 ret = PTR_ERR(src_iter);
182 goto out_src_iter;
183 }
184
185 clear = src_iter->ops->maps_tt && (!ttm || !ttm_tt_is_populated(ttm));
186 if (!(clear && ttm && !(ttm->page_flags & TTM_TT_FLAG_ZERO_ALLOC)))
187 ttm_move_memcpy(clear, PFN_UP(dst_mem->size), dst_iter, src_iter);
188
189 if (!src_iter->ops->maps_tt)
190 ttm_kmap_iter_linear_io_fini(&_src_iter.io, bdev, src_mem);
191 ttm_bo_move_sync_cleanup(bo, dst_mem);
192
193out_src_iter:
194 if (!dst_iter->ops->maps_tt)
195 ttm_kmap_iter_linear_io_fini(&_dst_iter.io, bdev, dst_mem);
196
197 return ret;
198}
199EXPORT_SYMBOL(ttm_bo_move_memcpy);
200
201static void ttm_transfered_destroy(struct ttm_buffer_object *bo)
202{
203 struct ttm_transfer_obj *fbo;
204
205 fbo = container_of(bo, struct ttm_transfer_obj, base);
206 dma_resv_fini(&fbo->base.base._resv);
207 ttm_bo_put(fbo->bo);
208 kfree(fbo);
209}
210
211/**
212 * ttm_buffer_object_transfer
213 *
214 * @bo: A pointer to a struct ttm_buffer_object.
215 * @new_obj: A pointer to a pointer to a newly created ttm_buffer_object,
216 * holding the data of @bo with the old placement.
217 *
218 * This is a utility function that may be called after an accelerated move
219 * has been scheduled. A new buffer object is created as a placeholder for
220 * the old data while it's being copied. When that buffer object is idle,
221 * it can be destroyed, releasing the space of the old placement.
222 * Returns:
223 * !0: Failure.
224 */
225
226static int ttm_buffer_object_transfer(struct ttm_buffer_object *bo,
227 struct ttm_buffer_object **new_obj)
228{
229 struct ttm_transfer_obj *fbo;
230 int ret;
231
232 fbo = kmalloc(sizeof(*fbo), GFP_KERNEL);
233 if (!fbo)
234 return -ENOMEM;
235
236 fbo->base = *bo;
237
238 /**
239 * Fix up members that we shouldn't copy directly:
240 * TODO: Explicit member copy would probably be better here.
241 */
242
243 atomic_inc(&ttm_glob.bo_count);
244 drm_vma_node_reset(&fbo->base.base.vma_node);
245
246 kref_init(&fbo->base.kref);
247 fbo->base.destroy = &ttm_transfered_destroy;
248 fbo->base.pin_count = 0;
249 if (bo->type != ttm_bo_type_sg)
250 fbo->base.base.resv = &fbo->base.base._resv;
251
252 dma_resv_init(&fbo->base.base._resv);
253 fbo->base.base.dev = NULL;
254 ret = dma_resv_trylock(&fbo->base.base._resv);
255 WARN_ON(!ret);
256
257 if (fbo->base.resource) {
258 ttm_resource_set_bo(fbo->base.resource, &fbo->base);
259 bo->resource = NULL;
260 ttm_bo_set_bulk_move(&fbo->base, NULL);
261 } else {
262 fbo->base.bulk_move = NULL;
263 }
264
265 ret = dma_resv_reserve_fences(&fbo->base.base._resv, 1);
266 if (ret) {
267 kfree(fbo);
268 return ret;
269 }
270
271 ttm_bo_get(bo);
272 fbo->bo = bo;
273
274 ttm_bo_move_to_lru_tail_unlocked(&fbo->base);
275
276 *new_obj = &fbo->base;
277 return 0;
278}
279
280/**
281 * ttm_io_prot
282 *
283 * @bo: ttm buffer object
284 * @res: ttm resource object
285 * @tmp: Page protection flag for a normal, cached mapping.
286 *
287 * Utility function that returns the pgprot_t that should be used for
288 * setting up a PTE with the caching model indicated by @c_state.
289 */
290pgprot_t ttm_io_prot(struct ttm_buffer_object *bo, struct ttm_resource *res,
291 pgprot_t tmp)
292{
293 struct ttm_resource_manager *man;
294 enum ttm_caching caching;
295
296 man = ttm_manager_type(bo->bdev, res->mem_type);
297 if (man->use_tt) {
298 caching = bo->ttm->caching;
299 if (bo->ttm->page_flags & TTM_TT_FLAG_DECRYPTED)
300 tmp = pgprot_decrypted(tmp);
301 } else {
302 caching = res->bus.caching;
303 }
304
305 return ttm_prot_from_caching(caching, tmp);
306}
307EXPORT_SYMBOL(ttm_io_prot);
308
309static int ttm_bo_ioremap(struct ttm_buffer_object *bo,
310 unsigned long offset,
311 unsigned long size,
312 struct ttm_bo_kmap_obj *map)
313{
314 struct ttm_resource *mem = bo->resource;
315
316 if (bo->resource->bus.addr) {
317 map->bo_kmap_type = ttm_bo_map_premapped;
318 map->virtual = ((u8 *)bo->resource->bus.addr) + offset;
319 } else {
320 resource_size_t res = bo->resource->bus.offset + offset;
321
322 map->bo_kmap_type = ttm_bo_map_iomap;
323 if (mem->bus.caching == ttm_write_combined)
324 map->virtual = ioremap_wc(res, size);
325#ifdef CONFIG_X86
326 else if (mem->bus.caching == ttm_cached)
327 map->virtual = ioremap_cache(res, size);
328#endif
329 else
330 map->virtual = ioremap(res, size);
331 }
332 return (!map->virtual) ? -ENOMEM : 0;
333}
334
335static int ttm_bo_kmap_ttm(struct ttm_buffer_object *bo,
336 unsigned long start_page,
337 unsigned long num_pages,
338 struct ttm_bo_kmap_obj *map)
339{
340 struct ttm_resource *mem = bo->resource;
341 struct ttm_operation_ctx ctx = {
342 .interruptible = false,
343 .no_wait_gpu = false
344 };
345 struct ttm_tt *ttm = bo->ttm;
346 struct ttm_resource_manager *man =
347 ttm_manager_type(bo->bdev, bo->resource->mem_type);
348 pgprot_t prot;
349 int ret;
350
351 BUG_ON(!ttm);
352
353 ret = ttm_bo_populate(bo, &ctx);
354 if (ret)
355 return ret;
356
357 if (num_pages == 1 && ttm->caching == ttm_cached &&
358 !(man->use_tt && (ttm->page_flags & TTM_TT_FLAG_DECRYPTED))) {
359 /*
360 * We're mapping a single page, and the desired
361 * page protection is consistent with the bo.
362 */
363
364 map->bo_kmap_type = ttm_bo_map_kmap;
365 map->page = ttm->pages[start_page];
366 map->virtual = kmap(map->page);
367 } else {
368 /*
369 * We need to use vmap to get the desired page protection
370 * or to make the buffer object look contiguous.
371 */
372 prot = ttm_io_prot(bo, mem, PAGE_KERNEL);
373 map->bo_kmap_type = ttm_bo_map_vmap;
374 map->virtual = vmap(ttm->pages + start_page, num_pages,
375 0, prot);
376 }
377 return (!map->virtual) ? -ENOMEM : 0;
378}
379
380/**
381 * ttm_bo_kmap
382 *
383 * @bo: The buffer object.
384 * @start_page: The first page to map.
385 * @num_pages: Number of pages to map.
386 * @map: pointer to a struct ttm_bo_kmap_obj representing the map.
387 *
388 * Sets up a kernel virtual mapping, using ioremap, vmap or kmap to the
389 * data in the buffer object. The ttm_kmap_obj_virtual function can then be
390 * used to obtain a virtual address to the data.
391 *
392 * Returns
393 * -ENOMEM: Out of memory.
394 * -EINVAL: Invalid range.
395 */
396int ttm_bo_kmap(struct ttm_buffer_object *bo,
397 unsigned long start_page, unsigned long num_pages,
398 struct ttm_bo_kmap_obj *map)
399{
400 unsigned long offset, size;
401 int ret;
402
403 map->virtual = NULL;
404 map->bo = bo;
405 if (num_pages > PFN_UP(bo->resource->size))
406 return -EINVAL;
407 if ((start_page + num_pages) > PFN_UP(bo->resource->size))
408 return -EINVAL;
409
410 ret = ttm_mem_io_reserve(bo->bdev, bo->resource);
411 if (ret)
412 return ret;
413 if (!bo->resource->bus.is_iomem) {
414 return ttm_bo_kmap_ttm(bo, start_page, num_pages, map);
415 } else {
416 offset = start_page << PAGE_SHIFT;
417 size = num_pages << PAGE_SHIFT;
418 return ttm_bo_ioremap(bo, offset, size, map);
419 }
420}
421EXPORT_SYMBOL(ttm_bo_kmap);
422
423/**
424 * ttm_bo_kunmap
425 *
426 * @map: Object describing the map to unmap.
427 *
428 * Unmaps a kernel map set up by ttm_bo_kmap.
429 */
430void ttm_bo_kunmap(struct ttm_bo_kmap_obj *map)
431{
432 if (!map->virtual)
433 return;
434 switch (map->bo_kmap_type) {
435 case ttm_bo_map_iomap:
436 iounmap(map->virtual);
437 break;
438 case ttm_bo_map_vmap:
439 vunmap(map->virtual);
440 break;
441 case ttm_bo_map_kmap:
442 kunmap(map->page);
443 break;
444 case ttm_bo_map_premapped:
445 break;
446 default:
447 BUG();
448 }
449 ttm_mem_io_free(map->bo->bdev, map->bo->resource);
450 map->virtual = NULL;
451 map->page = NULL;
452}
453EXPORT_SYMBOL(ttm_bo_kunmap);
454
455/**
456 * ttm_bo_vmap
457 *
458 * @bo: The buffer object.
459 * @map: pointer to a struct iosys_map representing the map.
460 *
461 * Sets up a kernel virtual mapping, using ioremap or vmap to the
462 * data in the buffer object. The parameter @map returns the virtual
463 * address as struct iosys_map. Unmap the buffer with ttm_bo_vunmap().
464 *
465 * Returns
466 * -ENOMEM: Out of memory.
467 * -EINVAL: Invalid range.
468 */
469int ttm_bo_vmap(struct ttm_buffer_object *bo, struct iosys_map *map)
470{
471 struct ttm_resource *mem = bo->resource;
472 int ret;
473
474 dma_resv_assert_held(bo->base.resv);
475
476 ret = ttm_mem_io_reserve(bo->bdev, mem);
477 if (ret)
478 return ret;
479
480 if (mem->bus.is_iomem) {
481 void __iomem *vaddr_iomem;
482
483 if (mem->bus.addr)
484 vaddr_iomem = (void __iomem *)mem->bus.addr;
485 else if (mem->bus.caching == ttm_write_combined)
486 vaddr_iomem = ioremap_wc(mem->bus.offset,
487 bo->base.size);
488#ifdef CONFIG_X86
489 else if (mem->bus.caching == ttm_cached)
490 vaddr_iomem = ioremap_cache(mem->bus.offset,
491 bo->base.size);
492#endif
493 else
494 vaddr_iomem = ioremap(mem->bus.offset, bo->base.size);
495
496 if (!vaddr_iomem)
497 return -ENOMEM;
498
499 iosys_map_set_vaddr_iomem(map, vaddr_iomem);
500
501 } else {
502 struct ttm_operation_ctx ctx = {
503 .interruptible = false,
504 .no_wait_gpu = false
505 };
506 struct ttm_tt *ttm = bo->ttm;
507 pgprot_t prot;
508 void *vaddr;
509
510 ret = ttm_bo_populate(bo, &ctx);
511 if (ret)
512 return ret;
513
514 /*
515 * We need to use vmap to get the desired page protection
516 * or to make the buffer object look contiguous.
517 */
518 prot = ttm_io_prot(bo, mem, PAGE_KERNEL);
519 vaddr = vmap(ttm->pages, ttm->num_pages, 0, prot);
520 if (!vaddr)
521 return -ENOMEM;
522
523 iosys_map_set_vaddr(map, vaddr);
524 }
525
526 return 0;
527}
528EXPORT_SYMBOL(ttm_bo_vmap);
529
530/**
531 * ttm_bo_vunmap
532 *
533 * @bo: The buffer object.
534 * @map: Object describing the map to unmap.
535 *
536 * Unmaps a kernel map set up by ttm_bo_vmap().
537 */
538void ttm_bo_vunmap(struct ttm_buffer_object *bo, struct iosys_map *map)
539{
540 struct ttm_resource *mem = bo->resource;
541
542 dma_resv_assert_held(bo->base.resv);
543
544 if (iosys_map_is_null(map))
545 return;
546
547 if (!map->is_iomem)
548 vunmap(map->vaddr);
549 else if (!mem->bus.addr)
550 iounmap(map->vaddr_iomem);
551 iosys_map_clear(map);
552
553 ttm_mem_io_free(bo->bdev, bo->resource);
554}
555EXPORT_SYMBOL(ttm_bo_vunmap);
556
557static int ttm_bo_wait_free_node(struct ttm_buffer_object *bo,
558 bool dst_use_tt)
559{
560 long ret;
561
562 ret = dma_resv_wait_timeout(bo->base.resv, DMA_RESV_USAGE_BOOKKEEP,
563 false, 15 * HZ);
564 if (ret == 0)
565 return -EBUSY;
566 if (ret < 0)
567 return ret;
568
569 if (!dst_use_tt)
570 ttm_bo_tt_destroy(bo);
571 ttm_resource_free(bo, &bo->resource);
572 return 0;
573}
574
575static int ttm_bo_move_to_ghost(struct ttm_buffer_object *bo,
576 struct dma_fence *fence,
577 bool dst_use_tt)
578{
579 struct ttm_buffer_object *ghost_obj;
580 int ret;
581
582 /**
583 * This should help pipeline ordinary buffer moves.
584 *
585 * Hang old buffer memory on a new buffer object,
586 * and leave it to be released when the GPU
587 * operation has completed.
588 */
589
590 ret = ttm_buffer_object_transfer(bo, &ghost_obj);
591 if (ret)
592 return ret;
593
594 dma_resv_add_fence(&ghost_obj->base._resv, fence,
595 DMA_RESV_USAGE_KERNEL);
596
597 /**
598 * If we're not moving to fixed memory, the TTM object
599 * needs to stay alive. Otherwhise hang it on the ghost
600 * bo to be unbound and destroyed.
601 */
602
603 if (dst_use_tt)
604 ghost_obj->ttm = NULL;
605 else
606 bo->ttm = NULL;
607
608 dma_resv_unlock(&ghost_obj->base._resv);
609 ttm_bo_put(ghost_obj);
610 return 0;
611}
612
613static void ttm_bo_move_pipeline_evict(struct ttm_buffer_object *bo,
614 struct dma_fence *fence)
615{
616 struct ttm_device *bdev = bo->bdev;
617 struct ttm_resource_manager *from;
618
619 from = ttm_manager_type(bdev, bo->resource->mem_type);
620
621 /**
622 * BO doesn't have a TTM we need to bind/unbind. Just remember
623 * this eviction and free up the allocation
624 */
625 spin_lock(&from->move_lock);
626 if (!from->move || dma_fence_is_later(fence, from->move)) {
627 dma_fence_put(from->move);
628 from->move = dma_fence_get(fence);
629 }
630 spin_unlock(&from->move_lock);
631
632 ttm_resource_free(bo, &bo->resource);
633}
634
635/**
636 * ttm_bo_move_accel_cleanup - cleanup helper for hw copies
637 *
638 * @bo: A pointer to a struct ttm_buffer_object.
639 * @fence: A fence object that signals when moving is complete.
640 * @evict: This is an evict move. Don't return until the buffer is idle.
641 * @pipeline: evictions are to be pipelined.
642 * @new_mem: struct ttm_resource indicating where to move.
643 *
644 * Accelerated move function to be called when an accelerated move
645 * has been scheduled. The function will create a new temporary buffer object
646 * representing the old placement, and put the sync object on both buffer
647 * objects. After that the newly created buffer object is unref'd to be
648 * destroyed when the move is complete. This will help pipeline
649 * buffer moves.
650 */
651int ttm_bo_move_accel_cleanup(struct ttm_buffer_object *bo,
652 struct dma_fence *fence,
653 bool evict,
654 bool pipeline,
655 struct ttm_resource *new_mem)
656{
657 struct ttm_device *bdev = bo->bdev;
658 struct ttm_resource_manager *from = ttm_manager_type(bdev, bo->resource->mem_type);
659 struct ttm_resource_manager *man = ttm_manager_type(bdev, new_mem->mem_type);
660 int ret = 0;
661
662 dma_resv_add_fence(bo->base.resv, fence, DMA_RESV_USAGE_KERNEL);
663 if (!evict)
664 ret = ttm_bo_move_to_ghost(bo, fence, man->use_tt);
665 else if (!from->use_tt && pipeline)
666 ttm_bo_move_pipeline_evict(bo, fence);
667 else
668 ret = ttm_bo_wait_free_node(bo, man->use_tt);
669
670 if (ret)
671 return ret;
672
673 ttm_bo_assign_mem(bo, new_mem);
674
675 return 0;
676}
677EXPORT_SYMBOL(ttm_bo_move_accel_cleanup);
678
679/**
680 * ttm_bo_move_sync_cleanup - cleanup by waiting for the move to finish
681 *
682 * @bo: A pointer to a struct ttm_buffer_object.
683 * @new_mem: struct ttm_resource indicating where to move.
684 *
685 * Special case of ttm_bo_move_accel_cleanup where the bo is guaranteed
686 * by the caller to be idle. Typically used after memcpy buffer moves.
687 */
688void ttm_bo_move_sync_cleanup(struct ttm_buffer_object *bo,
689 struct ttm_resource *new_mem)
690{
691 struct ttm_device *bdev = bo->bdev;
692 struct ttm_resource_manager *man = ttm_manager_type(bdev, new_mem->mem_type);
693 int ret;
694
695 ret = ttm_bo_wait_free_node(bo, man->use_tt);
696 if (WARN_ON(ret))
697 return;
698
699 ttm_bo_assign_mem(bo, new_mem);
700}
701EXPORT_SYMBOL(ttm_bo_move_sync_cleanup);
702
703/**
704 * ttm_bo_pipeline_gutting - purge the contents of a bo
705 * @bo: The buffer object
706 *
707 * Purge the contents of a bo, async if the bo is not idle.
708 * After a successful call, the bo is left unpopulated in
709 * system placement. The function may wait uninterruptible
710 * for idle on OOM.
711 *
712 * Return: 0 if successful, negative error code on failure.
713 */
714int ttm_bo_pipeline_gutting(struct ttm_buffer_object *bo)
715{
716 struct ttm_buffer_object *ghost;
717 struct ttm_tt *ttm;
718 int ret;
719
720 /* If already idle, no need for ghost object dance. */
721 if (dma_resv_test_signaled(bo->base.resv, DMA_RESV_USAGE_BOOKKEEP)) {
722 if (!bo->ttm) {
723 /* See comment below about clearing. */
724 ret = ttm_tt_create(bo, true);
725 if (ret)
726 return ret;
727 } else {
728 ttm_tt_unpopulate(bo->bdev, bo->ttm);
729 if (bo->type == ttm_bo_type_device)
730 ttm_tt_mark_for_clear(bo->ttm);
731 }
732 ttm_resource_free(bo, &bo->resource);
733 return 0;
734 }
735
736 /*
737 * We need an unpopulated ttm_tt after giving our current one,
738 * if any, to the ghost object. And we can't afford to fail
739 * creating one *after* the operation. If the bo subsequently gets
740 * resurrected, make sure it's cleared (if ttm_bo_type_device)
741 * to avoid leaking sensitive information to user-space.
742 */
743
744 ttm = bo->ttm;
745 bo->ttm = NULL;
746 ret = ttm_tt_create(bo, true);
747 swap(bo->ttm, ttm);
748 if (ret)
749 return ret;
750
751 ret = ttm_buffer_object_transfer(bo, &ghost);
752 if (ret)
753 goto error_destroy_tt;
754
755 ret = dma_resv_copy_fences(&ghost->base._resv, bo->base.resv);
756 /* Last resort, wait for the BO to be idle when we are OOM */
757 if (ret) {
758 dma_resv_wait_timeout(bo->base.resv, DMA_RESV_USAGE_BOOKKEEP,
759 false, MAX_SCHEDULE_TIMEOUT);
760 }
761
762 dma_resv_unlock(&ghost->base._resv);
763 ttm_bo_put(ghost);
764 bo->ttm = ttm;
765 return 0;
766
767error_destroy_tt:
768 ttm_tt_destroy(bo->bdev, ttm);
769 return ret;
770}
771
772static bool ttm_lru_walk_trylock(struct ttm_lru_walk *walk,
773 struct ttm_buffer_object *bo,
774 bool *needs_unlock)
775{
776 struct ttm_operation_ctx *ctx = walk->ctx;
777
778 *needs_unlock = false;
779
780 if (dma_resv_trylock(bo->base.resv)) {
781 *needs_unlock = true;
782 return true;
783 }
784
785 if (bo->base.resv == ctx->resv && ctx->allow_res_evict) {
786 dma_resv_assert_held(bo->base.resv);
787 return true;
788 }
789
790 return false;
791}
792
793static int ttm_lru_walk_ticketlock(struct ttm_lru_walk *walk,
794 struct ttm_buffer_object *bo,
795 bool *needs_unlock)
796{
797 struct dma_resv *resv = bo->base.resv;
798 int ret;
799
800 if (walk->ctx->interruptible)
801 ret = dma_resv_lock_interruptible(resv, walk->ticket);
802 else
803 ret = dma_resv_lock(resv, walk->ticket);
804
805 if (!ret) {
806 *needs_unlock = true;
807 /*
808 * Only a single ticketlock per loop. Ticketlocks are prone
809 * to return -EDEADLK causing the eviction to fail, so
810 * after waiting for the ticketlock, revert back to
811 * trylocking for this walk.
812 */
813 walk->ticket = NULL;
814 } else if (ret == -EDEADLK) {
815 /* Caller needs to exit the ww transaction. */
816 ret = -ENOSPC;
817 }
818
819 return ret;
820}
821
822static void ttm_lru_walk_unlock(struct ttm_buffer_object *bo, bool locked)
823{
824 if (locked)
825 dma_resv_unlock(bo->base.resv);
826}
827
828/**
829 * ttm_lru_walk_for_evict() - Perform a LRU list walk, with actions taken on
830 * valid items.
831 * @walk: describe the walks and actions taken
832 * @bdev: The TTM device.
833 * @man: The struct ttm_resource manager whose LRU lists we're walking.
834 * @target: The end condition for the walk.
835 *
836 * The LRU lists of @man are walk, and for each struct ttm_resource encountered,
837 * the corresponding ttm_buffer_object is locked and taken a reference on, and
838 * the LRU lock is dropped. the LRU lock may be dropped before locking and, in
839 * that case, it's verified that the item actually remains on the LRU list after
840 * the lock, and that the buffer object didn't switch resource in between.
841 *
842 * With a locked object, the actions indicated by @walk->process_bo are
843 * performed, and after that, the bo is unlocked, the refcount dropped and the
844 * next struct ttm_resource is processed. Here, the walker relies on
845 * TTM's restartable LRU list implementation.
846 *
847 * Typically @walk->process_bo() would return the number of pages evicted,
848 * swapped or shrunken, so that when the total exceeds @target, or when the
849 * LRU list has been walked in full, iteration is terminated. It's also terminated
850 * on error. Note that the definition of @target is done by the caller, it
851 * could have a different meaning than the number of pages.
852 *
853 * Note that the way dma_resv individualization is done, locking needs to be done
854 * either with the LRU lock held (trylocking only) or with a reference on the
855 * object.
856 *
857 * Return: The progress made towards target or negative error code on error.
858 */
859s64 ttm_lru_walk_for_evict(struct ttm_lru_walk *walk, struct ttm_device *bdev,
860 struct ttm_resource_manager *man, s64 target)
861{
862 struct ttm_resource_cursor cursor;
863 struct ttm_resource *res;
864 s64 progress = 0;
865 s64 lret;
866
867 spin_lock(&bdev->lru_lock);
868 ttm_resource_manager_for_each_res(man, &cursor, res) {
869 struct ttm_buffer_object *bo = res->bo;
870 bool bo_needs_unlock = false;
871 bool bo_locked = false;
872 int mem_type;
873
874 /*
875 * Attempt a trylock before taking a reference on the bo,
876 * since if we do it the other way around, and the trylock fails,
877 * we need to drop the lru lock to put the bo.
878 */
879 if (ttm_lru_walk_trylock(walk, bo, &bo_needs_unlock))
880 bo_locked = true;
881 else if (!walk->ticket || walk->ctx->no_wait_gpu ||
882 walk->trylock_only)
883 continue;
884
885 if (!ttm_bo_get_unless_zero(bo)) {
886 ttm_lru_walk_unlock(bo, bo_needs_unlock);
887 continue;
888 }
889
890 mem_type = res->mem_type;
891 spin_unlock(&bdev->lru_lock);
892
893 lret = 0;
894 if (!bo_locked)
895 lret = ttm_lru_walk_ticketlock(walk, bo, &bo_needs_unlock);
896
897 /*
898 * Note that in between the release of the lru lock and the
899 * ticketlock, the bo may have switched resource,
900 * and also memory type, since the resource may have been
901 * freed and allocated again with a different memory type.
902 * In that case, just skip it.
903 */
904 if (!lret && bo->resource && bo->resource->mem_type == mem_type)
905 lret = walk->ops->process_bo(walk, bo);
906
907 ttm_lru_walk_unlock(bo, bo_needs_unlock);
908 ttm_bo_put(bo);
909 if (lret == -EBUSY || lret == -EALREADY)
910 lret = 0;
911 progress = (lret < 0) ? lret : progress + lret;
912
913 spin_lock(&bdev->lru_lock);
914 if (progress < 0 || progress >= target)
915 break;
916 }
917 ttm_resource_cursor_fini(&cursor);
918 spin_unlock(&bdev->lru_lock);
919
920 return progress;
921}