Loading...
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#include <linux/reservation.h>
41
42void ttm_bo_free_old_node(struct ttm_buffer_object *bo)
43{
44 ttm_bo_mem_put(bo, &bo->mem);
45}
46
47int ttm_bo_move_ttm(struct ttm_buffer_object *bo,
48 bool evict,
49 bool no_wait_gpu, struct ttm_mem_reg *new_mem)
50{
51 struct ttm_tt *ttm = bo->ttm;
52 struct ttm_mem_reg *old_mem = &bo->mem;
53 int ret;
54
55 if (old_mem->mem_type != TTM_PL_SYSTEM) {
56 ttm_tt_unbind(ttm);
57 ttm_bo_free_old_node(bo);
58 ttm_flag_masked(&old_mem->placement, TTM_PL_FLAG_SYSTEM,
59 TTM_PL_MASK_MEM);
60 old_mem->mem_type = TTM_PL_SYSTEM;
61 }
62
63 ret = ttm_tt_set_placement_caching(ttm, new_mem->placement);
64 if (unlikely(ret != 0))
65 return ret;
66
67 if (new_mem->mem_type != TTM_PL_SYSTEM) {
68 ret = ttm_tt_bind(ttm, new_mem);
69 if (unlikely(ret != 0))
70 return ret;
71 }
72
73 *old_mem = *new_mem;
74 new_mem->mm_node = NULL;
75
76 return 0;
77}
78EXPORT_SYMBOL(ttm_bo_move_ttm);
79
80int ttm_mem_io_lock(struct ttm_mem_type_manager *man, bool interruptible)
81{
82 if (likely(man->io_reserve_fastpath))
83 return 0;
84
85 if (interruptible)
86 return mutex_lock_interruptible(&man->io_reserve_mutex);
87
88 mutex_lock(&man->io_reserve_mutex);
89 return 0;
90}
91EXPORT_SYMBOL(ttm_mem_io_lock);
92
93void ttm_mem_io_unlock(struct ttm_mem_type_manager *man)
94{
95 if (likely(man->io_reserve_fastpath))
96 return;
97
98 mutex_unlock(&man->io_reserve_mutex);
99}
100EXPORT_SYMBOL(ttm_mem_io_unlock);
101
102static int ttm_mem_io_evict(struct ttm_mem_type_manager *man)
103{
104 struct ttm_buffer_object *bo;
105
106 if (!man->use_io_reserve_lru || list_empty(&man->io_reserve_lru))
107 return -EAGAIN;
108
109 bo = list_first_entry(&man->io_reserve_lru,
110 struct ttm_buffer_object,
111 io_reserve_lru);
112 list_del_init(&bo->io_reserve_lru);
113 ttm_bo_unmap_virtual_locked(bo);
114
115 return 0;
116}
117
118
119int ttm_mem_io_reserve(struct ttm_bo_device *bdev,
120 struct ttm_mem_reg *mem)
121{
122 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
123 int ret = 0;
124
125 if (!bdev->driver->io_mem_reserve)
126 return 0;
127 if (likely(man->io_reserve_fastpath))
128 return bdev->driver->io_mem_reserve(bdev, mem);
129
130 if (bdev->driver->io_mem_reserve &&
131 mem->bus.io_reserved_count++ == 0) {
132retry:
133 ret = bdev->driver->io_mem_reserve(bdev, mem);
134 if (ret == -EAGAIN) {
135 ret = ttm_mem_io_evict(man);
136 if (ret == 0)
137 goto retry;
138 }
139 }
140 return ret;
141}
142EXPORT_SYMBOL(ttm_mem_io_reserve);
143
144void ttm_mem_io_free(struct ttm_bo_device *bdev,
145 struct ttm_mem_reg *mem)
146{
147 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
148
149 if (likely(man->io_reserve_fastpath))
150 return;
151
152 if (bdev->driver->io_mem_reserve &&
153 --mem->bus.io_reserved_count == 0 &&
154 bdev->driver->io_mem_free)
155 bdev->driver->io_mem_free(bdev, mem);
156
157}
158EXPORT_SYMBOL(ttm_mem_io_free);
159
160int ttm_mem_io_reserve_vm(struct ttm_buffer_object *bo)
161{
162 struct ttm_mem_reg *mem = &bo->mem;
163 int ret;
164
165 if (!mem->bus.io_reserved_vm) {
166 struct ttm_mem_type_manager *man =
167 &bo->bdev->man[mem->mem_type];
168
169 ret = ttm_mem_io_reserve(bo->bdev, mem);
170 if (unlikely(ret != 0))
171 return ret;
172 mem->bus.io_reserved_vm = true;
173 if (man->use_io_reserve_lru)
174 list_add_tail(&bo->io_reserve_lru,
175 &man->io_reserve_lru);
176 }
177 return 0;
178}
179
180void ttm_mem_io_free_vm(struct ttm_buffer_object *bo)
181{
182 struct ttm_mem_reg *mem = &bo->mem;
183
184 if (mem->bus.io_reserved_vm) {
185 mem->bus.io_reserved_vm = false;
186 list_del_init(&bo->io_reserve_lru);
187 ttm_mem_io_free(bo->bdev, mem);
188 }
189}
190
191static int ttm_mem_reg_ioremap(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem,
192 void **virtual)
193{
194 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
195 int ret;
196 void *addr;
197
198 *virtual = NULL;
199 (void) ttm_mem_io_lock(man, false);
200 ret = ttm_mem_io_reserve(bdev, mem);
201 ttm_mem_io_unlock(man);
202 if (ret || !mem->bus.is_iomem)
203 return ret;
204
205 if (mem->bus.addr) {
206 addr = mem->bus.addr;
207 } else {
208 if (mem->placement & TTM_PL_FLAG_WC)
209 addr = ioremap_wc(mem->bus.base + mem->bus.offset, mem->bus.size);
210 else
211 addr = ioremap_nocache(mem->bus.base + mem->bus.offset, mem->bus.size);
212 if (!addr) {
213 (void) ttm_mem_io_lock(man, false);
214 ttm_mem_io_free(bdev, mem);
215 ttm_mem_io_unlock(man);
216 return -ENOMEM;
217 }
218 }
219 *virtual = addr;
220 return 0;
221}
222
223static void ttm_mem_reg_iounmap(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem,
224 void *virtual)
225{
226 struct ttm_mem_type_manager *man;
227
228 man = &bdev->man[mem->mem_type];
229
230 if (virtual && mem->bus.addr == NULL)
231 iounmap(virtual);
232 (void) ttm_mem_io_lock(man, false);
233 ttm_mem_io_free(bdev, mem);
234 ttm_mem_io_unlock(man);
235}
236
237static int ttm_copy_io_page(void *dst, void *src, unsigned long page)
238{
239 uint32_t *dstP =
240 (uint32_t *) ((unsigned long)dst + (page << PAGE_SHIFT));
241 uint32_t *srcP =
242 (uint32_t *) ((unsigned long)src + (page << PAGE_SHIFT));
243
244 int i;
245 for (i = 0; i < PAGE_SIZE / sizeof(uint32_t); ++i)
246 iowrite32(ioread32(srcP++), dstP++);
247 return 0;
248}
249
250static int ttm_copy_io_ttm_page(struct ttm_tt *ttm, void *src,
251 unsigned long page,
252 pgprot_t prot)
253{
254 struct page *d = ttm->pages[page];
255 void *dst;
256
257 if (!d)
258 return -ENOMEM;
259
260 src = (void *)((unsigned long)src + (page << PAGE_SHIFT));
261
262#ifdef CONFIG_X86
263 dst = kmap_atomic_prot(d, prot);
264#else
265 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL))
266 dst = vmap(&d, 1, 0, prot);
267 else
268 dst = kmap(d);
269#endif
270 if (!dst)
271 return -ENOMEM;
272
273 memcpy_fromio(dst, src, PAGE_SIZE);
274
275#ifdef CONFIG_X86
276 kunmap_atomic(dst);
277#else
278 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL))
279 vunmap(dst);
280 else
281 kunmap(d);
282#endif
283
284 return 0;
285}
286
287static int ttm_copy_ttm_io_page(struct ttm_tt *ttm, void *dst,
288 unsigned long page,
289 pgprot_t prot)
290{
291 struct page *s = ttm->pages[page];
292 void *src;
293
294 if (!s)
295 return -ENOMEM;
296
297 dst = (void *)((unsigned long)dst + (page << PAGE_SHIFT));
298#ifdef CONFIG_X86
299 src = kmap_atomic_prot(s, prot);
300#else
301 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL))
302 src = vmap(&s, 1, 0, prot);
303 else
304 src = kmap(s);
305#endif
306 if (!src)
307 return -ENOMEM;
308
309 memcpy_toio(dst, src, PAGE_SIZE);
310
311#ifdef CONFIG_X86
312 kunmap_atomic(src);
313#else
314 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL))
315 vunmap(src);
316 else
317 kunmap(s);
318#endif
319
320 return 0;
321}
322
323int ttm_bo_move_memcpy(struct ttm_buffer_object *bo,
324 bool evict, bool no_wait_gpu,
325 struct ttm_mem_reg *new_mem)
326{
327 struct ttm_bo_device *bdev = bo->bdev;
328 struct ttm_mem_type_manager *man = &bdev->man[new_mem->mem_type];
329 struct ttm_tt *ttm = bo->ttm;
330 struct ttm_mem_reg *old_mem = &bo->mem;
331 struct ttm_mem_reg old_copy = *old_mem;
332 void *old_iomap;
333 void *new_iomap;
334 int ret;
335 unsigned long i;
336 unsigned long page;
337 unsigned long add = 0;
338 int dir;
339
340 ret = ttm_mem_reg_ioremap(bdev, old_mem, &old_iomap);
341 if (ret)
342 return ret;
343 ret = ttm_mem_reg_ioremap(bdev, new_mem, &new_iomap);
344 if (ret)
345 goto out;
346
347 /*
348 * Single TTM move. NOP.
349 */
350 if (old_iomap == NULL && new_iomap == NULL)
351 goto out2;
352
353 /*
354 * Don't move nonexistent data. Clear destination instead.
355 */
356 if (old_iomap == NULL &&
357 (ttm == NULL || (ttm->state == tt_unpopulated &&
358 !(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)))) {
359 memset_io(new_iomap, 0, new_mem->num_pages*PAGE_SIZE);
360 goto out2;
361 }
362
363 /*
364 * TTM might be null for moves within the same region.
365 */
366 if (ttm && ttm->state == tt_unpopulated) {
367 ret = ttm->bdev->driver->ttm_tt_populate(ttm);
368 if (ret)
369 goto out1;
370 }
371
372 add = 0;
373 dir = 1;
374
375 if ((old_mem->mem_type == new_mem->mem_type) &&
376 (new_mem->start < old_mem->start + old_mem->size)) {
377 dir = -1;
378 add = new_mem->num_pages - 1;
379 }
380
381 for (i = 0; i < new_mem->num_pages; ++i) {
382 page = i * dir + add;
383 if (old_iomap == NULL) {
384 pgprot_t prot = ttm_io_prot(old_mem->placement,
385 PAGE_KERNEL);
386 ret = ttm_copy_ttm_io_page(ttm, new_iomap, page,
387 prot);
388 } else if (new_iomap == NULL) {
389 pgprot_t prot = ttm_io_prot(new_mem->placement,
390 PAGE_KERNEL);
391 ret = ttm_copy_io_ttm_page(ttm, old_iomap, page,
392 prot);
393 } else
394 ret = ttm_copy_io_page(new_iomap, old_iomap, page);
395 if (ret)
396 goto out1;
397 }
398 mb();
399out2:
400 old_copy = *old_mem;
401 *old_mem = *new_mem;
402 new_mem->mm_node = NULL;
403
404 if ((man->flags & TTM_MEMTYPE_FLAG_FIXED) && (ttm != NULL)) {
405 ttm_tt_unbind(ttm);
406 ttm_tt_destroy(ttm);
407 bo->ttm = NULL;
408 }
409
410out1:
411 ttm_mem_reg_iounmap(bdev, old_mem, new_iomap);
412out:
413 ttm_mem_reg_iounmap(bdev, &old_copy, old_iomap);
414
415 /*
416 * On error, keep the mm node!
417 */
418 if (!ret)
419 ttm_bo_mem_put(bo, &old_copy);
420 return ret;
421}
422EXPORT_SYMBOL(ttm_bo_move_memcpy);
423
424static void ttm_transfered_destroy(struct ttm_buffer_object *bo)
425{
426 kfree(bo);
427}
428
429/**
430 * ttm_buffer_object_transfer
431 *
432 * @bo: A pointer to a struct ttm_buffer_object.
433 * @new_obj: A pointer to a pointer to a newly created ttm_buffer_object,
434 * holding the data of @bo with the old placement.
435 *
436 * This is a utility function that may be called after an accelerated move
437 * has been scheduled. A new buffer object is created as a placeholder for
438 * the old data while it's being copied. When that buffer object is idle,
439 * it can be destroyed, releasing the space of the old placement.
440 * Returns:
441 * !0: Failure.
442 */
443
444static int ttm_buffer_object_transfer(struct ttm_buffer_object *bo,
445 struct ttm_buffer_object **new_obj)
446{
447 struct ttm_buffer_object *fbo;
448 int ret;
449
450 fbo = kmalloc(sizeof(*fbo), GFP_KERNEL);
451 if (!fbo)
452 return -ENOMEM;
453
454 *fbo = *bo;
455
456 /**
457 * Fix up members that we shouldn't copy directly:
458 * TODO: Explicit member copy would probably be better here.
459 */
460
461 INIT_LIST_HEAD(&fbo->ddestroy);
462 INIT_LIST_HEAD(&fbo->lru);
463 INIT_LIST_HEAD(&fbo->swap);
464 INIT_LIST_HEAD(&fbo->io_reserve_lru);
465 drm_vma_node_reset(&fbo->vma_node);
466 atomic_set(&fbo->cpu_writers, 0);
467
468 kref_init(&fbo->list_kref);
469 kref_init(&fbo->kref);
470 fbo->destroy = &ttm_transfered_destroy;
471 fbo->acc_size = 0;
472 fbo->resv = &fbo->ttm_resv;
473 reservation_object_init(fbo->resv);
474 ret = ww_mutex_trylock(&fbo->resv->lock);
475 WARN_ON(!ret);
476
477 *new_obj = fbo;
478 return 0;
479}
480
481pgprot_t ttm_io_prot(uint32_t caching_flags, pgprot_t tmp)
482{
483 /* Cached mappings need no adjustment */
484 if (caching_flags & TTM_PL_FLAG_CACHED)
485 return tmp;
486
487#if defined(__i386__) || defined(__x86_64__)
488 if (caching_flags & TTM_PL_FLAG_WC)
489 tmp = pgprot_writecombine(tmp);
490 else if (boot_cpu_data.x86 > 3)
491 tmp = pgprot_noncached(tmp);
492#endif
493#if defined(__ia64__) || defined(__arm__) || defined(__aarch64__) || \
494 defined(__powerpc__)
495 if (caching_flags & TTM_PL_FLAG_WC)
496 tmp = pgprot_writecombine(tmp);
497 else
498 tmp = pgprot_noncached(tmp);
499#endif
500#if defined(__sparc__) || defined(__mips__)
501 tmp = pgprot_noncached(tmp);
502#endif
503 return tmp;
504}
505EXPORT_SYMBOL(ttm_io_prot);
506
507static int ttm_bo_ioremap(struct ttm_buffer_object *bo,
508 unsigned long offset,
509 unsigned long size,
510 struct ttm_bo_kmap_obj *map)
511{
512 struct ttm_mem_reg *mem = &bo->mem;
513
514 if (bo->mem.bus.addr) {
515 map->bo_kmap_type = ttm_bo_map_premapped;
516 map->virtual = (void *)(((u8 *)bo->mem.bus.addr) + offset);
517 } else {
518 map->bo_kmap_type = ttm_bo_map_iomap;
519 if (mem->placement & TTM_PL_FLAG_WC)
520 map->virtual = ioremap_wc(bo->mem.bus.base + bo->mem.bus.offset + offset,
521 size);
522 else
523 map->virtual = ioremap_nocache(bo->mem.bus.base + bo->mem.bus.offset + offset,
524 size);
525 }
526 return (!map->virtual) ? -ENOMEM : 0;
527}
528
529static int ttm_bo_kmap_ttm(struct ttm_buffer_object *bo,
530 unsigned long start_page,
531 unsigned long num_pages,
532 struct ttm_bo_kmap_obj *map)
533{
534 struct ttm_mem_reg *mem = &bo->mem; pgprot_t prot;
535 struct ttm_tt *ttm = bo->ttm;
536 int ret;
537
538 BUG_ON(!ttm);
539
540 if (ttm->state == tt_unpopulated) {
541 ret = ttm->bdev->driver->ttm_tt_populate(ttm);
542 if (ret)
543 return ret;
544 }
545
546 if (num_pages == 1 && (mem->placement & TTM_PL_FLAG_CACHED)) {
547 /*
548 * We're mapping a single page, and the desired
549 * page protection is consistent with the bo.
550 */
551
552 map->bo_kmap_type = ttm_bo_map_kmap;
553 map->page = ttm->pages[start_page];
554 map->virtual = kmap(map->page);
555 } else {
556 /*
557 * We need to use vmap to get the desired page protection
558 * or to make the buffer object look contiguous.
559 */
560 prot = ttm_io_prot(mem->placement, PAGE_KERNEL);
561 map->bo_kmap_type = ttm_bo_map_vmap;
562 map->virtual = vmap(ttm->pages + start_page, num_pages,
563 0, prot);
564 }
565 return (!map->virtual) ? -ENOMEM : 0;
566}
567
568int ttm_bo_kmap(struct ttm_buffer_object *bo,
569 unsigned long start_page, unsigned long num_pages,
570 struct ttm_bo_kmap_obj *map)
571{
572 struct ttm_mem_type_manager *man =
573 &bo->bdev->man[bo->mem.mem_type];
574 unsigned long offset, size;
575 int ret;
576
577 BUG_ON(!list_empty(&bo->swap));
578 map->virtual = NULL;
579 map->bo = bo;
580 if (num_pages > bo->num_pages)
581 return -EINVAL;
582 if (start_page > bo->num_pages)
583 return -EINVAL;
584#if 0
585 if (num_pages > 1 && !capable(CAP_SYS_ADMIN))
586 return -EPERM;
587#endif
588 (void) ttm_mem_io_lock(man, false);
589 ret = ttm_mem_io_reserve(bo->bdev, &bo->mem);
590 ttm_mem_io_unlock(man);
591 if (ret)
592 return ret;
593 if (!bo->mem.bus.is_iomem) {
594 return ttm_bo_kmap_ttm(bo, start_page, num_pages, map);
595 } else {
596 offset = start_page << PAGE_SHIFT;
597 size = num_pages << PAGE_SHIFT;
598 return ttm_bo_ioremap(bo, offset, size, map);
599 }
600}
601EXPORT_SYMBOL(ttm_bo_kmap);
602
603void ttm_bo_kunmap(struct ttm_bo_kmap_obj *map)
604{
605 struct ttm_buffer_object *bo = map->bo;
606 struct ttm_mem_type_manager *man =
607 &bo->bdev->man[bo->mem.mem_type];
608
609 if (!map->virtual)
610 return;
611 switch (map->bo_kmap_type) {
612 case ttm_bo_map_iomap:
613 iounmap(map->virtual);
614 break;
615 case ttm_bo_map_vmap:
616 vunmap(map->virtual);
617 break;
618 case ttm_bo_map_kmap:
619 kunmap(map->page);
620 break;
621 case ttm_bo_map_premapped:
622 break;
623 default:
624 BUG();
625 }
626 (void) ttm_mem_io_lock(man, false);
627 ttm_mem_io_free(map->bo->bdev, &map->bo->mem);
628 ttm_mem_io_unlock(man);
629 map->virtual = NULL;
630 map->page = NULL;
631}
632EXPORT_SYMBOL(ttm_bo_kunmap);
633
634int ttm_bo_move_accel_cleanup(struct ttm_buffer_object *bo,
635 struct fence *fence,
636 bool evict,
637 bool no_wait_gpu,
638 struct ttm_mem_reg *new_mem)
639{
640 struct ttm_bo_device *bdev = bo->bdev;
641 struct ttm_mem_type_manager *man = &bdev->man[new_mem->mem_type];
642 struct ttm_mem_reg *old_mem = &bo->mem;
643 int ret;
644 struct ttm_buffer_object *ghost_obj;
645
646 reservation_object_add_excl_fence(bo->resv, fence);
647 if (evict) {
648 ret = ttm_bo_wait(bo, false, false, false);
649 if (ret)
650 return ret;
651
652 if ((man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
653 (bo->ttm != NULL)) {
654 ttm_tt_unbind(bo->ttm);
655 ttm_tt_destroy(bo->ttm);
656 bo->ttm = NULL;
657 }
658 ttm_bo_free_old_node(bo);
659 } else {
660 /**
661 * This should help pipeline ordinary buffer moves.
662 *
663 * Hang old buffer memory on a new buffer object,
664 * and leave it to be released when the GPU
665 * operation has completed.
666 */
667
668 set_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
669
670 ret = ttm_buffer_object_transfer(bo, &ghost_obj);
671 if (ret)
672 return ret;
673
674 reservation_object_add_excl_fence(ghost_obj->resv, fence);
675
676 /**
677 * If we're not moving to fixed memory, the TTM object
678 * needs to stay alive. Otherwhise hang it on the ghost
679 * bo to be unbound and destroyed.
680 */
681
682 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED))
683 ghost_obj->ttm = NULL;
684 else
685 bo->ttm = NULL;
686
687 ttm_bo_unreserve(ghost_obj);
688 ttm_bo_unref(&ghost_obj);
689 }
690
691 *old_mem = *new_mem;
692 new_mem->mm_node = NULL;
693
694 return 0;
695}
696EXPORT_SYMBOL(ttm_bo_move_accel_cleanup);
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}