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/**************************************************************************
2 *
3 * Copyright (c) 2007-2009 VMware, Inc., Palo Alto, CA., USA
4 * All Rights Reserved.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sub license, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
13 *
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
16 * of the Software.
17 *
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
19 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
20 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
21 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
22 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
23 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
24 * USE OR OTHER DEALINGS IN THE SOFTWARE.
25 *
26 **************************************************************************/
27/*
28 * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com>
29 */
30
31#include <drm/ttm/ttm_bo_driver.h>
32#include <drm/ttm/ttm_placement.h>
33#include <drm/drm_vma_manager.h>
34#include <linux/io.h>
35#include <linux/highmem.h>
36#include <linux/wait.h>
37#include <linux/slab.h>
38#include <linux/vmalloc.h>
39#include <linux/module.h>
40
41void ttm_bo_free_old_node(struct ttm_buffer_object *bo)
42{
43 ttm_bo_mem_put(bo, &bo->mem);
44}
45
46int ttm_bo_move_ttm(struct ttm_buffer_object *bo,
47 bool evict,
48 bool no_wait_gpu, struct ttm_mem_reg *new_mem)
49{
50 struct ttm_tt *ttm = bo->ttm;
51 struct ttm_mem_reg *old_mem = &bo->mem;
52 int ret;
53
54 if (old_mem->mem_type != TTM_PL_SYSTEM) {
55 ttm_tt_unbind(ttm);
56 ttm_bo_free_old_node(bo);
57 ttm_flag_masked(&old_mem->placement, TTM_PL_FLAG_SYSTEM,
58 TTM_PL_MASK_MEM);
59 old_mem->mem_type = TTM_PL_SYSTEM;
60 }
61
62 ret = ttm_tt_set_placement_caching(ttm, new_mem->placement);
63 if (unlikely(ret != 0))
64 return ret;
65
66 if (new_mem->mem_type != TTM_PL_SYSTEM) {
67 ret = ttm_tt_bind(ttm, new_mem);
68 if (unlikely(ret != 0))
69 return ret;
70 }
71
72 *old_mem = *new_mem;
73 new_mem->mm_node = NULL;
74
75 return 0;
76}
77EXPORT_SYMBOL(ttm_bo_move_ttm);
78
79int ttm_mem_io_lock(struct ttm_mem_type_manager *man, bool interruptible)
80{
81 if (likely(man->io_reserve_fastpath))
82 return 0;
83
84 if (interruptible)
85 return mutex_lock_interruptible(&man->io_reserve_mutex);
86
87 mutex_lock(&man->io_reserve_mutex);
88 return 0;
89}
90EXPORT_SYMBOL(ttm_mem_io_lock);
91
92void ttm_mem_io_unlock(struct ttm_mem_type_manager *man)
93{
94 if (likely(man->io_reserve_fastpath))
95 return;
96
97 mutex_unlock(&man->io_reserve_mutex);
98}
99EXPORT_SYMBOL(ttm_mem_io_unlock);
100
101static int ttm_mem_io_evict(struct ttm_mem_type_manager *man)
102{
103 struct ttm_buffer_object *bo;
104
105 if (!man->use_io_reserve_lru || list_empty(&man->io_reserve_lru))
106 return -EAGAIN;
107
108 bo = list_first_entry(&man->io_reserve_lru,
109 struct ttm_buffer_object,
110 io_reserve_lru);
111 list_del_init(&bo->io_reserve_lru);
112 ttm_bo_unmap_virtual_locked(bo);
113
114 return 0;
115}
116
117
118int ttm_mem_io_reserve(struct ttm_bo_device *bdev,
119 struct ttm_mem_reg *mem)
120{
121 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
122 int ret = 0;
123
124 if (!bdev->driver->io_mem_reserve)
125 return 0;
126 if (likely(man->io_reserve_fastpath))
127 return bdev->driver->io_mem_reserve(bdev, mem);
128
129 if (bdev->driver->io_mem_reserve &&
130 mem->bus.io_reserved_count++ == 0) {
131retry:
132 ret = bdev->driver->io_mem_reserve(bdev, mem);
133 if (ret == -EAGAIN) {
134 ret = ttm_mem_io_evict(man);
135 if (ret == 0)
136 goto retry;
137 }
138 }
139 return ret;
140}
141EXPORT_SYMBOL(ttm_mem_io_reserve);
142
143void ttm_mem_io_free(struct ttm_bo_device *bdev,
144 struct ttm_mem_reg *mem)
145{
146 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
147
148 if (likely(man->io_reserve_fastpath))
149 return;
150
151 if (bdev->driver->io_mem_reserve &&
152 --mem->bus.io_reserved_count == 0 &&
153 bdev->driver->io_mem_free)
154 bdev->driver->io_mem_free(bdev, mem);
155
156}
157EXPORT_SYMBOL(ttm_mem_io_free);
158
159int ttm_mem_io_reserve_vm(struct ttm_buffer_object *bo)
160{
161 struct ttm_mem_reg *mem = &bo->mem;
162 int ret;
163
164 if (!mem->bus.io_reserved_vm) {
165 struct ttm_mem_type_manager *man =
166 &bo->bdev->man[mem->mem_type];
167
168 ret = ttm_mem_io_reserve(bo->bdev, mem);
169 if (unlikely(ret != 0))
170 return ret;
171 mem->bus.io_reserved_vm = true;
172 if (man->use_io_reserve_lru)
173 list_add_tail(&bo->io_reserve_lru,
174 &man->io_reserve_lru);
175 }
176 return 0;
177}
178
179void ttm_mem_io_free_vm(struct ttm_buffer_object *bo)
180{
181 struct ttm_mem_reg *mem = &bo->mem;
182
183 if (mem->bus.io_reserved_vm) {
184 mem->bus.io_reserved_vm = false;
185 list_del_init(&bo->io_reserve_lru);
186 ttm_mem_io_free(bo->bdev, mem);
187 }
188}
189
190static int ttm_mem_reg_ioremap(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem,
191 void **virtual)
192{
193 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
194 int ret;
195 void *addr;
196
197 *virtual = NULL;
198 (void) ttm_mem_io_lock(man, false);
199 ret = ttm_mem_io_reserve(bdev, mem);
200 ttm_mem_io_unlock(man);
201 if (ret || !mem->bus.is_iomem)
202 return ret;
203
204 if (mem->bus.addr) {
205 addr = mem->bus.addr;
206 } else {
207 if (mem->placement & TTM_PL_FLAG_WC)
208 addr = ioremap_wc(mem->bus.base + mem->bus.offset, mem->bus.size);
209 else
210 addr = ioremap_nocache(mem->bus.base + mem->bus.offset, mem->bus.size);
211 if (!addr) {
212 (void) ttm_mem_io_lock(man, false);
213 ttm_mem_io_free(bdev, mem);
214 ttm_mem_io_unlock(man);
215 return -ENOMEM;
216 }
217 }
218 *virtual = addr;
219 return 0;
220}
221
222static void ttm_mem_reg_iounmap(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem,
223 void *virtual)
224{
225 struct ttm_mem_type_manager *man;
226
227 man = &bdev->man[mem->mem_type];
228
229 if (virtual && mem->bus.addr == NULL)
230 iounmap(virtual);
231 (void) ttm_mem_io_lock(man, false);
232 ttm_mem_io_free(bdev, mem);
233 ttm_mem_io_unlock(man);
234}
235
236static int ttm_copy_io_page(void *dst, void *src, unsigned long page)
237{
238 uint32_t *dstP =
239 (uint32_t *) ((unsigned long)dst + (page << PAGE_SHIFT));
240 uint32_t *srcP =
241 (uint32_t *) ((unsigned long)src + (page << PAGE_SHIFT));
242
243 int i;
244 for (i = 0; i < PAGE_SIZE / sizeof(uint32_t); ++i)
245 iowrite32(ioread32(srcP++), dstP++);
246 return 0;
247}
248
249static int ttm_copy_io_ttm_page(struct ttm_tt *ttm, void *src,
250 unsigned long page,
251 pgprot_t prot)
252{
253 struct page *d = ttm->pages[page];
254 void *dst;
255
256 if (!d)
257 return -ENOMEM;
258
259 src = (void *)((unsigned long)src + (page << PAGE_SHIFT));
260
261#ifdef CONFIG_X86
262 dst = kmap_atomic_prot(d, prot);
263#else
264 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL))
265 dst = vmap(&d, 1, 0, prot);
266 else
267 dst = kmap(d);
268#endif
269 if (!dst)
270 return -ENOMEM;
271
272 memcpy_fromio(dst, src, PAGE_SIZE);
273
274#ifdef CONFIG_X86
275 kunmap_atomic(dst);
276#else
277 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL))
278 vunmap(dst);
279 else
280 kunmap(d);
281#endif
282
283 return 0;
284}
285
286static int ttm_copy_ttm_io_page(struct ttm_tt *ttm, void *dst,
287 unsigned long page,
288 pgprot_t prot)
289{
290 struct page *s = ttm->pages[page];
291 void *src;
292
293 if (!s)
294 return -ENOMEM;
295
296 dst = (void *)((unsigned long)dst + (page << PAGE_SHIFT));
297#ifdef CONFIG_X86
298 src = kmap_atomic_prot(s, prot);
299#else
300 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL))
301 src = vmap(&s, 1, 0, prot);
302 else
303 src = kmap(s);
304#endif
305 if (!src)
306 return -ENOMEM;
307
308 memcpy_toio(dst, src, PAGE_SIZE);
309
310#ifdef CONFIG_X86
311 kunmap_atomic(src);
312#else
313 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL))
314 vunmap(src);
315 else
316 kunmap(s);
317#endif
318
319 return 0;
320}
321
322int ttm_bo_move_memcpy(struct ttm_buffer_object *bo,
323 bool evict, bool no_wait_gpu,
324 struct ttm_mem_reg *new_mem)
325{
326 struct ttm_bo_device *bdev = bo->bdev;
327 struct ttm_mem_type_manager *man = &bdev->man[new_mem->mem_type];
328 struct ttm_tt *ttm = bo->ttm;
329 struct ttm_mem_reg *old_mem = &bo->mem;
330 struct ttm_mem_reg old_copy = *old_mem;
331 void *old_iomap;
332 void *new_iomap;
333 int ret;
334 unsigned long i;
335 unsigned long page;
336 unsigned long add = 0;
337 int dir;
338
339 ret = ttm_mem_reg_ioremap(bdev, old_mem, &old_iomap);
340 if (ret)
341 return ret;
342 ret = ttm_mem_reg_ioremap(bdev, new_mem, &new_iomap);
343 if (ret)
344 goto out;
345
346 /*
347 * Single TTM move. NOP.
348 */
349 if (old_iomap == NULL && new_iomap == NULL)
350 goto out2;
351
352 /*
353 * Don't move nonexistent data. Clear destination instead.
354 */
355 if (old_iomap == NULL &&
356 (ttm == NULL || (ttm->state == tt_unpopulated &&
357 !(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)))) {
358 memset_io(new_iomap, 0, new_mem->num_pages*PAGE_SIZE);
359 goto out2;
360 }
361
362 /*
363 * TTM might be null for moves within the same region.
364 */
365 if (ttm && ttm->state == tt_unpopulated) {
366 ret = ttm->bdev->driver->ttm_tt_populate(ttm);
367 if (ret)
368 goto out1;
369 }
370
371 add = 0;
372 dir = 1;
373
374 if ((old_mem->mem_type == new_mem->mem_type) &&
375 (new_mem->start < old_mem->start + old_mem->size)) {
376 dir = -1;
377 add = new_mem->num_pages - 1;
378 }
379
380 for (i = 0; i < new_mem->num_pages; ++i) {
381 page = i * dir + add;
382 if (old_iomap == NULL) {
383 pgprot_t prot = ttm_io_prot(old_mem->placement,
384 PAGE_KERNEL);
385 ret = ttm_copy_ttm_io_page(ttm, new_iomap, page,
386 prot);
387 } else if (new_iomap == NULL) {
388 pgprot_t prot = ttm_io_prot(new_mem->placement,
389 PAGE_KERNEL);
390 ret = ttm_copy_io_ttm_page(ttm, old_iomap, page,
391 prot);
392 } else
393 ret = ttm_copy_io_page(new_iomap, old_iomap, page);
394 if (ret)
395 goto out1;
396 }
397 mb();
398out2:
399 old_copy = *old_mem;
400 *old_mem = *new_mem;
401 new_mem->mm_node = NULL;
402
403 if ((man->flags & TTM_MEMTYPE_FLAG_FIXED) && (ttm != NULL)) {
404 ttm_tt_unbind(ttm);
405 ttm_tt_destroy(ttm);
406 bo->ttm = NULL;
407 }
408
409out1:
410 ttm_mem_reg_iounmap(bdev, old_mem, new_iomap);
411out:
412 ttm_mem_reg_iounmap(bdev, &old_copy, old_iomap);
413
414 /*
415 * On error, keep the mm node!
416 */
417 if (!ret)
418 ttm_bo_mem_put(bo, &old_copy);
419 return ret;
420}
421EXPORT_SYMBOL(ttm_bo_move_memcpy);
422
423static void ttm_transfered_destroy(struct ttm_buffer_object *bo)
424{
425 kfree(bo);
426}
427
428/**
429 * ttm_buffer_object_transfer
430 *
431 * @bo: A pointer to a struct ttm_buffer_object.
432 * @new_obj: A pointer to a pointer to a newly created ttm_buffer_object,
433 * holding the data of @bo with the old placement.
434 *
435 * This is a utility function that may be called after an accelerated move
436 * has been scheduled. A new buffer object is created as a placeholder for
437 * the old data while it's being copied. When that buffer object is idle,
438 * it can be destroyed, releasing the space of the old placement.
439 * Returns:
440 * !0: Failure.
441 */
442
443static int ttm_buffer_object_transfer(struct ttm_buffer_object *bo,
444 struct ttm_buffer_object **new_obj)
445{
446 struct ttm_buffer_object *fbo;
447 struct ttm_bo_device *bdev = bo->bdev;
448 struct ttm_bo_driver *driver = bdev->driver;
449 int ret;
450
451 fbo = kmalloc(sizeof(*fbo), GFP_KERNEL);
452 if (!fbo)
453 return -ENOMEM;
454
455 *fbo = *bo;
456
457 /**
458 * Fix up members that we shouldn't copy directly:
459 * TODO: Explicit member copy would probably be better here.
460 */
461
462 INIT_LIST_HEAD(&fbo->ddestroy);
463 INIT_LIST_HEAD(&fbo->lru);
464 INIT_LIST_HEAD(&fbo->swap);
465 INIT_LIST_HEAD(&fbo->io_reserve_lru);
466 drm_vma_node_reset(&fbo->vma_node);
467 atomic_set(&fbo->cpu_writers, 0);
468
469 spin_lock(&bdev->fence_lock);
470 if (bo->sync_obj)
471 fbo->sync_obj = driver->sync_obj_ref(bo->sync_obj);
472 else
473 fbo->sync_obj = NULL;
474 spin_unlock(&bdev->fence_lock);
475 kref_init(&fbo->list_kref);
476 kref_init(&fbo->kref);
477 fbo->destroy = &ttm_transfered_destroy;
478 fbo->acc_size = 0;
479 fbo->resv = &fbo->ttm_resv;
480 reservation_object_init(fbo->resv);
481 ret = ww_mutex_trylock(&fbo->resv->lock);
482 WARN_ON(!ret);
483
484 *new_obj = fbo;
485 return 0;
486}
487
488pgprot_t ttm_io_prot(uint32_t caching_flags, pgprot_t tmp)
489{
490#if defined(__i386__) || defined(__x86_64__)
491 if (caching_flags & TTM_PL_FLAG_WC)
492 tmp = pgprot_writecombine(tmp);
493 else if (boot_cpu_data.x86 > 3)
494 tmp = pgprot_noncached(tmp);
495
496#elif defined(__powerpc__)
497 if (!(caching_flags & TTM_PL_FLAG_CACHED)) {
498 pgprot_val(tmp) |= _PAGE_NO_CACHE;
499 if (caching_flags & TTM_PL_FLAG_UNCACHED)
500 pgprot_val(tmp) |= _PAGE_GUARDED;
501 }
502#endif
503#if defined(__ia64__)
504 if (caching_flags & TTM_PL_FLAG_WC)
505 tmp = pgprot_writecombine(tmp);
506 else
507 tmp = pgprot_noncached(tmp);
508#endif
509#if defined(__sparc__) || defined(__mips__)
510 if (!(caching_flags & TTM_PL_FLAG_CACHED))
511 tmp = pgprot_noncached(tmp);
512#endif
513 return tmp;
514}
515EXPORT_SYMBOL(ttm_io_prot);
516
517static int ttm_bo_ioremap(struct ttm_buffer_object *bo,
518 unsigned long offset,
519 unsigned long size,
520 struct ttm_bo_kmap_obj *map)
521{
522 struct ttm_mem_reg *mem = &bo->mem;
523
524 if (bo->mem.bus.addr) {
525 map->bo_kmap_type = ttm_bo_map_premapped;
526 map->virtual = (void *)(((u8 *)bo->mem.bus.addr) + offset);
527 } else {
528 map->bo_kmap_type = ttm_bo_map_iomap;
529 if (mem->placement & TTM_PL_FLAG_WC)
530 map->virtual = ioremap_wc(bo->mem.bus.base + bo->mem.bus.offset + offset,
531 size);
532 else
533 map->virtual = ioremap_nocache(bo->mem.bus.base + bo->mem.bus.offset + offset,
534 size);
535 }
536 return (!map->virtual) ? -ENOMEM : 0;
537}
538
539static int ttm_bo_kmap_ttm(struct ttm_buffer_object *bo,
540 unsigned long start_page,
541 unsigned long num_pages,
542 struct ttm_bo_kmap_obj *map)
543{
544 struct ttm_mem_reg *mem = &bo->mem; pgprot_t prot;
545 struct ttm_tt *ttm = bo->ttm;
546 int ret;
547
548 BUG_ON(!ttm);
549
550 if (ttm->state == tt_unpopulated) {
551 ret = ttm->bdev->driver->ttm_tt_populate(ttm);
552 if (ret)
553 return ret;
554 }
555
556 if (num_pages == 1 && (mem->placement & TTM_PL_FLAG_CACHED)) {
557 /*
558 * We're mapping a single page, and the desired
559 * page protection is consistent with the bo.
560 */
561
562 map->bo_kmap_type = ttm_bo_map_kmap;
563 map->page = ttm->pages[start_page];
564 map->virtual = kmap(map->page);
565 } else {
566 /*
567 * We need to use vmap to get the desired page protection
568 * or to make the buffer object look contiguous.
569 */
570 prot = (mem->placement & TTM_PL_FLAG_CACHED) ?
571 PAGE_KERNEL :
572 ttm_io_prot(mem->placement, PAGE_KERNEL);
573 map->bo_kmap_type = ttm_bo_map_vmap;
574 map->virtual = vmap(ttm->pages + start_page, num_pages,
575 0, prot);
576 }
577 return (!map->virtual) ? -ENOMEM : 0;
578}
579
580int ttm_bo_kmap(struct ttm_buffer_object *bo,
581 unsigned long start_page, unsigned long num_pages,
582 struct ttm_bo_kmap_obj *map)
583{
584 struct ttm_mem_type_manager *man =
585 &bo->bdev->man[bo->mem.mem_type];
586 unsigned long offset, size;
587 int ret;
588
589 BUG_ON(!list_empty(&bo->swap));
590 map->virtual = NULL;
591 map->bo = bo;
592 if (num_pages > bo->num_pages)
593 return -EINVAL;
594 if (start_page > bo->num_pages)
595 return -EINVAL;
596#if 0
597 if (num_pages > 1 && !capable(CAP_SYS_ADMIN))
598 return -EPERM;
599#endif
600 (void) ttm_mem_io_lock(man, false);
601 ret = ttm_mem_io_reserve(bo->bdev, &bo->mem);
602 ttm_mem_io_unlock(man);
603 if (ret)
604 return ret;
605 if (!bo->mem.bus.is_iomem) {
606 return ttm_bo_kmap_ttm(bo, start_page, num_pages, map);
607 } else {
608 offset = start_page << PAGE_SHIFT;
609 size = num_pages << PAGE_SHIFT;
610 return ttm_bo_ioremap(bo, offset, size, map);
611 }
612}
613EXPORT_SYMBOL(ttm_bo_kmap);
614
615void ttm_bo_kunmap(struct ttm_bo_kmap_obj *map)
616{
617 struct ttm_buffer_object *bo = map->bo;
618 struct ttm_mem_type_manager *man =
619 &bo->bdev->man[bo->mem.mem_type];
620
621 if (!map->virtual)
622 return;
623 switch (map->bo_kmap_type) {
624 case ttm_bo_map_iomap:
625 iounmap(map->virtual);
626 break;
627 case ttm_bo_map_vmap:
628 vunmap(map->virtual);
629 break;
630 case ttm_bo_map_kmap:
631 kunmap(map->page);
632 break;
633 case ttm_bo_map_premapped:
634 break;
635 default:
636 BUG();
637 }
638 (void) ttm_mem_io_lock(man, false);
639 ttm_mem_io_free(map->bo->bdev, &map->bo->mem);
640 ttm_mem_io_unlock(man);
641 map->virtual = NULL;
642 map->page = NULL;
643}
644EXPORT_SYMBOL(ttm_bo_kunmap);
645
646int ttm_bo_move_accel_cleanup(struct ttm_buffer_object *bo,
647 void *sync_obj,
648 bool evict,
649 bool no_wait_gpu,
650 struct ttm_mem_reg *new_mem)
651{
652 struct ttm_bo_device *bdev = bo->bdev;
653 struct ttm_bo_driver *driver = bdev->driver;
654 struct ttm_mem_type_manager *man = &bdev->man[new_mem->mem_type];
655 struct ttm_mem_reg *old_mem = &bo->mem;
656 int ret;
657 struct ttm_buffer_object *ghost_obj;
658 void *tmp_obj = NULL;
659
660 spin_lock(&bdev->fence_lock);
661 if (bo->sync_obj) {
662 tmp_obj = bo->sync_obj;
663 bo->sync_obj = NULL;
664 }
665 bo->sync_obj = driver->sync_obj_ref(sync_obj);
666 if (evict) {
667 ret = ttm_bo_wait(bo, false, false, false);
668 spin_unlock(&bdev->fence_lock);
669 if (tmp_obj)
670 driver->sync_obj_unref(&tmp_obj);
671 if (ret)
672 return ret;
673
674 if ((man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
675 (bo->ttm != NULL)) {
676 ttm_tt_unbind(bo->ttm);
677 ttm_tt_destroy(bo->ttm);
678 bo->ttm = NULL;
679 }
680 ttm_bo_free_old_node(bo);
681 } else {
682 /**
683 * This should help pipeline ordinary buffer moves.
684 *
685 * Hang old buffer memory on a new buffer object,
686 * and leave it to be released when the GPU
687 * operation has completed.
688 */
689
690 set_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
691 spin_unlock(&bdev->fence_lock);
692 if (tmp_obj)
693 driver->sync_obj_unref(&tmp_obj);
694
695 ret = ttm_buffer_object_transfer(bo, &ghost_obj);
696 if (ret)
697 return ret;
698
699 /**
700 * If we're not moving to fixed memory, the TTM object
701 * needs to stay alive. Otherwhise hang it on the ghost
702 * bo to be unbound and destroyed.
703 */
704
705 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED))
706 ghost_obj->ttm = NULL;
707 else
708 bo->ttm = NULL;
709
710 ttm_bo_unreserve(ghost_obj);
711 ttm_bo_unref(&ghost_obj);
712 }
713
714 *old_mem = *new_mem;
715 new_mem->mm_node = NULL;
716
717 return 0;
718}
719EXPORT_SYMBOL(ttm_bo_move_accel_cleanup);