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 <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->use_io_reserve_lru))
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}
105
106void ttm_mem_io_unlock(struct ttm_mem_type_manager *man)
107{
108 if (likely(!man->use_io_reserve_lru))
109 return;
110
111 mutex_unlock(&man->io_reserve_mutex);
112}
113
114static int ttm_mem_io_evict(struct ttm_mem_type_manager *man)
115{
116 struct ttm_buffer_object *bo;
117
118 bo = list_first_entry_or_null(&man->io_reserve_lru,
119 struct ttm_buffer_object,
120 io_reserve_lru);
121 if (!bo)
122 return -ENOSPC;
123
124 list_del_init(&bo->io_reserve_lru);
125 ttm_bo_unmap_virtual_locked(bo);
126 return 0;
127}
128
129int ttm_mem_io_reserve(struct ttm_bo_device *bdev,
130 struct ttm_mem_reg *mem)
131{
132 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
133 int ret;
134
135 if (mem->bus.io_reserved_count++)
136 return 0;
137
138 if (!bdev->driver->io_mem_reserve)
139 return 0;
140
141retry:
142 ret = bdev->driver->io_mem_reserve(bdev, mem);
143 if (ret == -ENOSPC) {
144 ret = ttm_mem_io_evict(man);
145 if (ret == 0)
146 goto retry;
147 }
148 return ret;
149}
150
151void ttm_mem_io_free(struct ttm_bo_device *bdev,
152 struct ttm_mem_reg *mem)
153{
154 if (--mem->bus.io_reserved_count)
155 return;
156
157 if (!bdev->driver->io_mem_free)
158 return;
159
160 bdev->driver->io_mem_free(bdev, mem);
161}
162
163int ttm_mem_io_reserve_vm(struct ttm_buffer_object *bo)
164{
165 struct ttm_mem_type_manager *man = &bo->bdev->man[bo->mem.mem_type];
166 struct ttm_mem_reg *mem = &bo->mem;
167 int ret;
168
169 if (mem->bus.io_reserved_vm)
170 return 0;
171
172 ret = ttm_mem_io_reserve(bo->bdev, mem);
173 if (unlikely(ret != 0))
174 return ret;
175 mem->bus.io_reserved_vm = true;
176 if (man->use_io_reserve_lru)
177 list_add_tail(&bo->io_reserve_lru,
178 &man->io_reserve_lru);
179 return 0;
180}
181
182void ttm_mem_io_free_vm(struct ttm_buffer_object *bo)
183{
184 struct ttm_mem_reg *mem = &bo->mem;
185
186 if (!mem->bus.io_reserved_vm)
187 return;
188
189 mem->bus.io_reserved_vm = false;
190 list_del_init(&bo->io_reserve_lru);
191 ttm_mem_io_free(bo->bdev, mem);
192}
193
194static int ttm_mem_reg_ioremap(struct ttm_bo_device *bdev,
195 struct ttm_mem_reg *mem,
196 void **virtual)
197{
198 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
199 int ret;
200 void *addr;
201
202 *virtual = NULL;
203 (void) ttm_mem_io_lock(man, false);
204 ret = ttm_mem_io_reserve(bdev, mem);
205 ttm_mem_io_unlock(man);
206 if (ret || !mem->bus.is_iomem)
207 return ret;
208
209 if (mem->bus.addr) {
210 addr = mem->bus.addr;
211 } else {
212 if (mem->placement & TTM_PL_FLAG_WC)
213 addr = ioremap_wc(mem->bus.base + mem->bus.offset,
214 mem->bus.size);
215 else
216 addr = ioremap(mem->bus.base + mem->bus.offset,
217 mem->bus.size);
218 if (!addr) {
219 (void) ttm_mem_io_lock(man, false);
220 ttm_mem_io_free(bdev, mem);
221 ttm_mem_io_unlock(man);
222 return -ENOMEM;
223 }
224 }
225 *virtual = addr;
226 return 0;
227}
228
229static void ttm_mem_reg_iounmap(struct ttm_bo_device *bdev,
230 struct ttm_mem_reg *mem,
231 void *virtual)
232{
233 struct ttm_mem_type_manager *man;
234
235 man = &bdev->man[mem->mem_type];
236
237 if (virtual && mem->bus.addr == NULL)
238 iounmap(virtual);
239 (void) ttm_mem_io_lock(man, false);
240 ttm_mem_io_free(bdev, mem);
241 ttm_mem_io_unlock(man);
242}
243
244static int ttm_copy_io_page(void *dst, void *src, unsigned long page)
245{
246 uint32_t *dstP =
247 (uint32_t *) ((unsigned long)dst + (page << PAGE_SHIFT));
248 uint32_t *srcP =
249 (uint32_t *) ((unsigned long)src + (page << PAGE_SHIFT));
250
251 int i;
252 for (i = 0; i < PAGE_SIZE / sizeof(uint32_t); ++i)
253 iowrite32(ioread32(srcP++), dstP++);
254 return 0;
255}
256
257static int ttm_copy_io_ttm_page(struct ttm_tt *ttm, void *src,
258 unsigned long page,
259 pgprot_t prot)
260{
261 struct page *d = ttm->pages[page];
262 void *dst;
263
264 if (!d)
265 return -ENOMEM;
266
267 src = (void *)((unsigned long)src + (page << PAGE_SHIFT));
268 dst = kmap_atomic_prot(d, prot);
269 if (!dst)
270 return -ENOMEM;
271
272 memcpy_fromio(dst, src, PAGE_SIZE);
273
274 kunmap_atomic(dst);
275
276 return 0;
277}
278
279static int ttm_copy_ttm_io_page(struct ttm_tt *ttm, void *dst,
280 unsigned long page,
281 pgprot_t prot)
282{
283 struct page *s = ttm->pages[page];
284 void *src;
285
286 if (!s)
287 return -ENOMEM;
288
289 dst = (void *)((unsigned long)dst + (page << PAGE_SHIFT));
290 src = kmap_atomic_prot(s, prot);
291 if (!src)
292 return -ENOMEM;
293
294 memcpy_toio(dst, src, PAGE_SIZE);
295
296 kunmap_atomic(src);
297
298 return 0;
299}
300
301int ttm_bo_move_memcpy(struct ttm_buffer_object *bo,
302 struct ttm_operation_ctx *ctx,
303 struct ttm_mem_reg *new_mem)
304{
305 struct ttm_bo_device *bdev = bo->bdev;
306 struct ttm_mem_type_manager *man = &bdev->man[new_mem->mem_type];
307 struct ttm_tt *ttm = bo->ttm;
308 struct ttm_mem_reg *old_mem = &bo->mem;
309 struct ttm_mem_reg old_copy = *old_mem;
310 void *old_iomap;
311 void *new_iomap;
312 int ret;
313 unsigned long i;
314 unsigned long page;
315 unsigned long add = 0;
316 int dir;
317
318 ret = ttm_bo_wait(bo, ctx->interruptible, ctx->no_wait_gpu);
319 if (ret)
320 return ret;
321
322 ret = ttm_mem_reg_ioremap(bdev, old_mem, &old_iomap);
323 if (ret)
324 return ret;
325 ret = ttm_mem_reg_ioremap(bdev, new_mem, &new_iomap);
326 if (ret)
327 goto out;
328
329 /*
330 * Single TTM move. NOP.
331 */
332 if (old_iomap == NULL && new_iomap == NULL)
333 goto out2;
334
335 /*
336 * Don't move nonexistent data. Clear destination instead.
337 */
338 if (old_iomap == NULL &&
339 (ttm == NULL || (ttm->state == tt_unpopulated &&
340 !(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)))) {
341 memset_io(new_iomap, 0, new_mem->num_pages*PAGE_SIZE);
342 goto out2;
343 }
344
345 /*
346 * TTM might be null for moves within the same region.
347 */
348 if (ttm) {
349 ret = ttm_tt_populate(ttm, ctx);
350 if (ret)
351 goto out1;
352 }
353
354 add = 0;
355 dir = 1;
356
357 if ((old_mem->mem_type == new_mem->mem_type) &&
358 (new_mem->start < old_mem->start + old_mem->size)) {
359 dir = -1;
360 add = new_mem->num_pages - 1;
361 }
362
363 for (i = 0; i < new_mem->num_pages; ++i) {
364 page = i * dir + add;
365 if (old_iomap == NULL) {
366 pgprot_t prot = ttm_io_prot(old_mem->placement,
367 PAGE_KERNEL);
368 ret = ttm_copy_ttm_io_page(ttm, new_iomap, page,
369 prot);
370 } else if (new_iomap == NULL) {
371 pgprot_t prot = ttm_io_prot(new_mem->placement,
372 PAGE_KERNEL);
373 ret = ttm_copy_io_ttm_page(ttm, old_iomap, page,
374 prot);
375 } else {
376 ret = ttm_copy_io_page(new_iomap, old_iomap, page);
377 }
378 if (ret)
379 goto out1;
380 }
381 mb();
382out2:
383 old_copy = *old_mem;
384 *old_mem = *new_mem;
385 new_mem->mm_node = NULL;
386
387 if (man->flags & TTM_MEMTYPE_FLAG_FIXED) {
388 ttm_tt_destroy(ttm);
389 bo->ttm = NULL;
390 }
391
392out1:
393 ttm_mem_reg_iounmap(bdev, old_mem, new_iomap);
394out:
395 ttm_mem_reg_iounmap(bdev, &old_copy, old_iomap);
396
397 /*
398 * On error, keep the mm node!
399 */
400 if (!ret)
401 ttm_bo_mem_put(bo, &old_copy);
402 return ret;
403}
404EXPORT_SYMBOL(ttm_bo_move_memcpy);
405
406static void ttm_transfered_destroy(struct ttm_buffer_object *bo)
407{
408 struct ttm_transfer_obj *fbo;
409
410 fbo = container_of(bo, struct ttm_transfer_obj, base);
411 ttm_bo_put(fbo->bo);
412 kfree(fbo);
413}
414
415/**
416 * ttm_buffer_object_transfer
417 *
418 * @bo: A pointer to a struct ttm_buffer_object.
419 * @new_obj: A pointer to a pointer to a newly created ttm_buffer_object,
420 * holding the data of @bo with the old placement.
421 *
422 * This is a utility function that may be called after an accelerated move
423 * has been scheduled. A new buffer object is created as a placeholder for
424 * the old data while it's being copied. When that buffer object is idle,
425 * it can be destroyed, releasing the space of the old placement.
426 * Returns:
427 * !0: Failure.
428 */
429
430static int ttm_buffer_object_transfer(struct ttm_buffer_object *bo,
431 struct ttm_buffer_object **new_obj)
432{
433 struct ttm_transfer_obj *fbo;
434 int ret;
435
436 fbo = kmalloc(sizeof(*fbo), GFP_KERNEL);
437 if (!fbo)
438 return -ENOMEM;
439
440 fbo->base = *bo;
441 fbo->base.mem.placement |= TTM_PL_FLAG_NO_EVICT;
442
443 ttm_bo_get(bo);
444 fbo->bo = bo;
445
446 /**
447 * Fix up members that we shouldn't copy directly:
448 * TODO: Explicit member copy would probably be better here.
449 */
450
451 atomic_inc(&ttm_bo_glob.bo_count);
452 INIT_LIST_HEAD(&fbo->base.ddestroy);
453 INIT_LIST_HEAD(&fbo->base.lru);
454 INIT_LIST_HEAD(&fbo->base.swap);
455 INIT_LIST_HEAD(&fbo->base.io_reserve_lru);
456 fbo->base.moving = NULL;
457 drm_vma_node_reset(&fbo->base.base.vma_node);
458
459 kref_init(&fbo->base.kref);
460 fbo->base.destroy = &ttm_transfered_destroy;
461 fbo->base.acc_size = 0;
462 if (bo->type != ttm_bo_type_sg)
463 fbo->base.base.resv = &fbo->base.base._resv;
464
465 dma_resv_init(&fbo->base.base._resv);
466 fbo->base.base.dev = NULL;
467 ret = dma_resv_trylock(&fbo->base.base._resv);
468 WARN_ON(!ret);
469
470 *new_obj = &fbo->base;
471 return 0;
472}
473
474pgprot_t ttm_io_prot(uint32_t caching_flags, pgprot_t tmp)
475{
476 /* Cached mappings need no adjustment */
477 if (caching_flags & TTM_PL_FLAG_CACHED)
478 return tmp;
479
480#if defined(__i386__) || defined(__x86_64__)
481 if (caching_flags & TTM_PL_FLAG_WC)
482 tmp = pgprot_writecombine(tmp);
483 else if (boot_cpu_data.x86 > 3)
484 tmp = pgprot_noncached(tmp);
485#endif
486#if defined(__ia64__) || defined(__arm__) || defined(__aarch64__) || \
487 defined(__powerpc__) || defined(__mips__)
488 if (caching_flags & TTM_PL_FLAG_WC)
489 tmp = pgprot_writecombine(tmp);
490 else
491 tmp = pgprot_noncached(tmp);
492#endif
493#if defined(__sparc__)
494 tmp = pgprot_noncached(tmp);
495#endif
496 return tmp;
497}
498EXPORT_SYMBOL(ttm_io_prot);
499
500static int ttm_bo_ioremap(struct ttm_buffer_object *bo,
501 unsigned long offset,
502 unsigned long size,
503 struct ttm_bo_kmap_obj *map)
504{
505 struct ttm_mem_reg *mem = &bo->mem;
506
507 if (bo->mem.bus.addr) {
508 map->bo_kmap_type = ttm_bo_map_premapped;
509 map->virtual = (void *)(((u8 *)bo->mem.bus.addr) + offset);
510 } else {
511 map->bo_kmap_type = ttm_bo_map_iomap;
512 if (mem->placement & TTM_PL_FLAG_WC)
513 map->virtual = ioremap_wc(bo->mem.bus.base +
514 bo->mem.bus.offset + offset,
515 size);
516 else
517 map->virtual = ioremap(bo->mem.bus.base +
518 bo->mem.bus.offset + offset,
519 size);
520 }
521 return (!map->virtual) ? -ENOMEM : 0;
522}
523
524static int ttm_bo_kmap_ttm(struct ttm_buffer_object *bo,
525 unsigned long start_page,
526 unsigned long num_pages,
527 struct ttm_bo_kmap_obj *map)
528{
529 struct ttm_mem_reg *mem = &bo->mem;
530 struct ttm_operation_ctx ctx = {
531 .interruptible = false,
532 .no_wait_gpu = false
533 };
534 struct ttm_tt *ttm = bo->ttm;
535 pgprot_t prot;
536 int ret;
537
538 BUG_ON(!ttm);
539
540 ret = ttm_tt_populate(ttm, &ctx);
541 if (ret)
542 return ret;
543
544 if (num_pages == 1 && (mem->placement & TTM_PL_FLAG_CACHED)) {
545 /*
546 * We're mapping a single page, and the desired
547 * page protection is consistent with the bo.
548 */
549
550 map->bo_kmap_type = ttm_bo_map_kmap;
551 map->page = ttm->pages[start_page];
552 map->virtual = kmap(map->page);
553 } else {
554 /*
555 * We need to use vmap to get the desired page protection
556 * or to make the buffer object look contiguous.
557 */
558 prot = ttm_io_prot(mem->placement, PAGE_KERNEL);
559 map->bo_kmap_type = ttm_bo_map_vmap;
560 map->virtual = vmap(ttm->pages + start_page, num_pages,
561 0, prot);
562 }
563 return (!map->virtual) ? -ENOMEM : 0;
564}
565
566int ttm_bo_kmap(struct ttm_buffer_object *bo,
567 unsigned long start_page, unsigned long num_pages,
568 struct ttm_bo_kmap_obj *map)
569{
570 struct ttm_mem_type_manager *man =
571 &bo->bdev->man[bo->mem.mem_type];
572 unsigned long offset, size;
573 int ret;
574
575 map->virtual = NULL;
576 map->bo = bo;
577 if (num_pages > bo->num_pages)
578 return -EINVAL;
579 if (start_page > bo->num_pages)
580 return -EINVAL;
581
582 (void) ttm_mem_io_lock(man, false);
583 ret = ttm_mem_io_reserve(bo->bdev, &bo->mem);
584 ttm_mem_io_unlock(man);
585 if (ret)
586 return ret;
587 if (!bo->mem.bus.is_iomem) {
588 return ttm_bo_kmap_ttm(bo, start_page, num_pages, map);
589 } else {
590 offset = start_page << PAGE_SHIFT;
591 size = num_pages << PAGE_SHIFT;
592 return ttm_bo_ioremap(bo, offset, size, map);
593 }
594}
595EXPORT_SYMBOL(ttm_bo_kmap);
596
597void ttm_bo_kunmap(struct ttm_bo_kmap_obj *map)
598{
599 struct ttm_buffer_object *bo = map->bo;
600 struct ttm_mem_type_manager *man =
601 &bo->bdev->man[bo->mem.mem_type];
602
603 if (!map->virtual)
604 return;
605 switch (map->bo_kmap_type) {
606 case ttm_bo_map_iomap:
607 iounmap(map->virtual);
608 break;
609 case ttm_bo_map_vmap:
610 vunmap(map->virtual);
611 break;
612 case ttm_bo_map_kmap:
613 kunmap(map->page);
614 break;
615 case ttm_bo_map_premapped:
616 break;
617 default:
618 BUG();
619 }
620 (void) ttm_mem_io_lock(man, false);
621 ttm_mem_io_free(map->bo->bdev, &map->bo->mem);
622 ttm_mem_io_unlock(man);
623 map->virtual = NULL;
624 map->page = NULL;
625}
626EXPORT_SYMBOL(ttm_bo_kunmap);
627
628int ttm_bo_move_accel_cleanup(struct ttm_buffer_object *bo,
629 struct dma_fence *fence,
630 bool evict,
631 struct ttm_mem_reg *new_mem)
632{
633 struct ttm_bo_device *bdev = bo->bdev;
634 struct ttm_mem_type_manager *man = &bdev->man[new_mem->mem_type];
635 struct ttm_mem_reg *old_mem = &bo->mem;
636 int ret;
637 struct ttm_buffer_object *ghost_obj;
638
639 dma_resv_add_excl_fence(bo->base.resv, fence);
640 if (evict) {
641 ret = ttm_bo_wait(bo, false, false);
642 if (ret)
643 return ret;
644
645 if (man->flags & TTM_MEMTYPE_FLAG_FIXED) {
646 ttm_tt_destroy(bo->ttm);
647 bo->ttm = NULL;
648 }
649 ttm_bo_free_old_node(bo);
650 } else {
651 /**
652 * This should help pipeline ordinary buffer moves.
653 *
654 * Hang old buffer memory on a new buffer object,
655 * and leave it to be released when the GPU
656 * operation has completed.
657 */
658
659 dma_fence_put(bo->moving);
660 bo->moving = dma_fence_get(fence);
661
662 ret = ttm_buffer_object_transfer(bo, &ghost_obj);
663 if (ret)
664 return ret;
665
666 dma_resv_add_excl_fence(&ghost_obj->base._resv, fence);
667
668 /**
669 * If we're not moving to fixed memory, the TTM object
670 * needs to stay alive. Otherwhise hang it on the ghost
671 * bo to be unbound and destroyed.
672 */
673
674 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED))
675 ghost_obj->ttm = NULL;
676 else
677 bo->ttm = NULL;
678
679 dma_resv_unlock(&ghost_obj->base._resv);
680 ttm_bo_put(ghost_obj);
681 }
682
683 *old_mem = *new_mem;
684 new_mem->mm_node = NULL;
685
686 return 0;
687}
688EXPORT_SYMBOL(ttm_bo_move_accel_cleanup);
689
690int ttm_bo_pipeline_move(struct ttm_buffer_object *bo,
691 struct dma_fence *fence, bool evict,
692 struct ttm_mem_reg *new_mem)
693{
694 struct ttm_bo_device *bdev = bo->bdev;
695 struct ttm_mem_reg *old_mem = &bo->mem;
696
697 struct ttm_mem_type_manager *from = &bdev->man[old_mem->mem_type];
698 struct ttm_mem_type_manager *to = &bdev->man[new_mem->mem_type];
699
700 int ret;
701
702 dma_resv_add_excl_fence(bo->base.resv, fence);
703
704 if (!evict) {
705 struct ttm_buffer_object *ghost_obj;
706
707 /**
708 * This should help pipeline ordinary buffer moves.
709 *
710 * Hang old buffer memory on a new buffer object,
711 * and leave it to be released when the GPU
712 * operation has completed.
713 */
714
715 dma_fence_put(bo->moving);
716 bo->moving = dma_fence_get(fence);
717
718 ret = ttm_buffer_object_transfer(bo, &ghost_obj);
719 if (ret)
720 return ret;
721
722 dma_resv_add_excl_fence(&ghost_obj->base._resv, fence);
723
724 /**
725 * If we're not moving to fixed memory, the TTM object
726 * needs to stay alive. Otherwhise hang it on the ghost
727 * bo to be unbound and destroyed.
728 */
729
730 if (!(to->flags & TTM_MEMTYPE_FLAG_FIXED))
731 ghost_obj->ttm = NULL;
732 else
733 bo->ttm = NULL;
734
735 dma_resv_unlock(&ghost_obj->base._resv);
736 ttm_bo_put(ghost_obj);
737
738 } else if (from->flags & TTM_MEMTYPE_FLAG_FIXED) {
739
740 /**
741 * BO doesn't have a TTM we need to bind/unbind. Just remember
742 * this eviction and free up the allocation
743 */
744
745 spin_lock(&from->move_lock);
746 if (!from->move || dma_fence_is_later(fence, from->move)) {
747 dma_fence_put(from->move);
748 from->move = dma_fence_get(fence);
749 }
750 spin_unlock(&from->move_lock);
751
752 ttm_bo_free_old_node(bo);
753
754 dma_fence_put(bo->moving);
755 bo->moving = dma_fence_get(fence);
756
757 } else {
758 /**
759 * Last resort, wait for the move to be completed.
760 *
761 * Should never happen in pratice.
762 */
763
764 ret = ttm_bo_wait(bo, false, false);
765 if (ret)
766 return ret;
767
768 if (to->flags & TTM_MEMTYPE_FLAG_FIXED) {
769 ttm_tt_destroy(bo->ttm);
770 bo->ttm = NULL;
771 }
772 ttm_bo_free_old_node(bo);
773 }
774
775 *old_mem = *new_mem;
776 new_mem->mm_node = NULL;
777
778 return 0;
779}
780EXPORT_SYMBOL(ttm_bo_pipeline_move);
781
782int ttm_bo_pipeline_gutting(struct ttm_buffer_object *bo)
783{
784 struct ttm_buffer_object *ghost;
785 int ret;
786
787 ret = ttm_buffer_object_transfer(bo, &ghost);
788 if (ret)
789 return ret;
790
791 ret = dma_resv_copy_fences(&ghost->base._resv, bo->base.resv);
792 /* Last resort, wait for the BO to be idle when we are OOM */
793 if (ret)
794 ttm_bo_wait(bo, false, false);
795
796 memset(&bo->mem, 0, sizeof(bo->mem));
797 bo->mem.mem_type = TTM_PL_SYSTEM;
798 bo->ttm = NULL;
799
800 dma_resv_unlock(&ghost->base._resv);
801 ttm_bo_put(ghost);
802
803 return 0;
804}