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 "ttm/ttm_bo_driver.h"
32#include "ttm/ttm_placement.h"
33#include <linux/io.h>
34#include <linux/highmem.h>
35#include <linux/wait.h>
36#include <linux/slab.h>
37#include <linux/vmalloc.h>
38#include <linux/module.h>
39
40void ttm_bo_free_old_node(struct ttm_buffer_object *bo)
41{
42 ttm_bo_mem_put(bo, &bo->mem);
43}
44
45int ttm_bo_move_ttm(struct ttm_buffer_object *bo,
46 bool evict, bool no_wait_reserve,
47 bool no_wait_gpu, struct ttm_mem_reg *new_mem)
48{
49 struct ttm_tt *ttm = bo->ttm;
50 struct ttm_mem_reg *old_mem = &bo->mem;
51 int ret;
52
53 if (old_mem->mem_type != TTM_PL_SYSTEM) {
54 ttm_tt_unbind(ttm);
55 ttm_bo_free_old_node(bo);
56 ttm_flag_masked(&old_mem->placement, TTM_PL_FLAG_SYSTEM,
57 TTM_PL_MASK_MEM);
58 old_mem->mem_type = TTM_PL_SYSTEM;
59 }
60
61 ret = ttm_tt_set_placement_caching(ttm, new_mem->placement);
62 if (unlikely(ret != 0))
63 return ret;
64
65 if (new_mem->mem_type != TTM_PL_SYSTEM) {
66 ret = ttm_tt_bind(ttm, new_mem);
67 if (unlikely(ret != 0))
68 return ret;
69 }
70
71 *old_mem = *new_mem;
72 new_mem->mm_node = NULL;
73
74 return 0;
75}
76EXPORT_SYMBOL(ttm_bo_move_ttm);
77
78int ttm_mem_io_lock(struct ttm_mem_type_manager *man, bool interruptible)
79{
80 if (likely(man->io_reserve_fastpath))
81 return 0;
82
83 if (interruptible)
84 return mutex_lock_interruptible(&man->io_reserve_mutex);
85
86 mutex_lock(&man->io_reserve_mutex);
87 return 0;
88}
89
90void ttm_mem_io_unlock(struct ttm_mem_type_manager *man)
91{
92 if (likely(man->io_reserve_fastpath))
93 return;
94
95 mutex_unlock(&man->io_reserve_mutex);
96}
97
98static int ttm_mem_io_evict(struct ttm_mem_type_manager *man)
99{
100 struct ttm_buffer_object *bo;
101
102 if (!man->use_io_reserve_lru || list_empty(&man->io_reserve_lru))
103 return -EAGAIN;
104
105 bo = list_first_entry(&man->io_reserve_lru,
106 struct ttm_buffer_object,
107 io_reserve_lru);
108 list_del_init(&bo->io_reserve_lru);
109 ttm_bo_unmap_virtual_locked(bo);
110
111 return 0;
112}
113
114static int ttm_mem_io_reserve(struct ttm_bo_device *bdev,
115 struct ttm_mem_reg *mem)
116{
117 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
118 int ret = 0;
119
120 if (!bdev->driver->io_mem_reserve)
121 return 0;
122 if (likely(man->io_reserve_fastpath))
123 return bdev->driver->io_mem_reserve(bdev, mem);
124
125 if (bdev->driver->io_mem_reserve &&
126 mem->bus.io_reserved_count++ == 0) {
127retry:
128 ret = bdev->driver->io_mem_reserve(bdev, mem);
129 if (ret == -EAGAIN) {
130 ret = ttm_mem_io_evict(man);
131 if (ret == 0)
132 goto retry;
133 }
134 }
135 return ret;
136}
137
138static void ttm_mem_io_free(struct ttm_bo_device *bdev,
139 struct ttm_mem_reg *mem)
140{
141 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
142
143 if (likely(man->io_reserve_fastpath))
144 return;
145
146 if (bdev->driver->io_mem_reserve &&
147 --mem->bus.io_reserved_count == 0 &&
148 bdev->driver->io_mem_free)
149 bdev->driver->io_mem_free(bdev, mem);
150
151}
152
153int ttm_mem_io_reserve_vm(struct ttm_buffer_object *bo)
154{
155 struct ttm_mem_reg *mem = &bo->mem;
156 int ret;
157
158 if (!mem->bus.io_reserved_vm) {
159 struct ttm_mem_type_manager *man =
160 &bo->bdev->man[mem->mem_type];
161
162 ret = ttm_mem_io_reserve(bo->bdev, mem);
163 if (unlikely(ret != 0))
164 return ret;
165 mem->bus.io_reserved_vm = true;
166 if (man->use_io_reserve_lru)
167 list_add_tail(&bo->io_reserve_lru,
168 &man->io_reserve_lru);
169 }
170 return 0;
171}
172
173void ttm_mem_io_free_vm(struct ttm_buffer_object *bo)
174{
175 struct ttm_mem_reg *mem = &bo->mem;
176
177 if (mem->bus.io_reserved_vm) {
178 mem->bus.io_reserved_vm = false;
179 list_del_init(&bo->io_reserve_lru);
180 ttm_mem_io_free(bo->bdev, mem);
181 }
182}
183
184int ttm_mem_reg_ioremap(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem,
185 void **virtual)
186{
187 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
188 int ret;
189 void *addr;
190
191 *virtual = NULL;
192 (void) ttm_mem_io_lock(man, false);
193 ret = ttm_mem_io_reserve(bdev, mem);
194 ttm_mem_io_unlock(man);
195 if (ret || !mem->bus.is_iomem)
196 return ret;
197
198 if (mem->bus.addr) {
199 addr = mem->bus.addr;
200 } else {
201 if (mem->placement & TTM_PL_FLAG_WC)
202 addr = ioremap_wc(mem->bus.base + mem->bus.offset, mem->bus.size);
203 else
204 addr = ioremap_nocache(mem->bus.base + mem->bus.offset, mem->bus.size);
205 if (!addr) {
206 (void) ttm_mem_io_lock(man, false);
207 ttm_mem_io_free(bdev, mem);
208 ttm_mem_io_unlock(man);
209 return -ENOMEM;
210 }
211 }
212 *virtual = addr;
213 return 0;
214}
215
216void ttm_mem_reg_iounmap(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem,
217 void *virtual)
218{
219 struct ttm_mem_type_manager *man;
220
221 man = &bdev->man[mem->mem_type];
222
223 if (virtual && mem->bus.addr == NULL)
224 iounmap(virtual);
225 (void) ttm_mem_io_lock(man, false);
226 ttm_mem_io_free(bdev, mem);
227 ttm_mem_io_unlock(man);
228}
229
230static int ttm_copy_io_page(void *dst, void *src, unsigned long page)
231{
232 uint32_t *dstP =
233 (uint32_t *) ((unsigned long)dst + (page << PAGE_SHIFT));
234 uint32_t *srcP =
235 (uint32_t *) ((unsigned long)src + (page << PAGE_SHIFT));
236
237 int i;
238 for (i = 0; i < PAGE_SIZE / sizeof(uint32_t); ++i)
239 iowrite32(ioread32(srcP++), dstP++);
240 return 0;
241}
242
243static int ttm_copy_io_ttm_page(struct ttm_tt *ttm, void *src,
244 unsigned long page,
245 pgprot_t prot)
246{
247 struct page *d = ttm_tt_get_page(ttm, page);
248 void *dst;
249
250 if (!d)
251 return -ENOMEM;
252
253 src = (void *)((unsigned long)src + (page << PAGE_SHIFT));
254
255#ifdef CONFIG_X86
256 dst = kmap_atomic_prot(d, prot);
257#else
258 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL))
259 dst = vmap(&d, 1, 0, prot);
260 else
261 dst = kmap(d);
262#endif
263 if (!dst)
264 return -ENOMEM;
265
266 memcpy_fromio(dst, src, PAGE_SIZE);
267
268#ifdef CONFIG_X86
269 kunmap_atomic(dst);
270#else
271 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL))
272 vunmap(dst);
273 else
274 kunmap(d);
275#endif
276
277 return 0;
278}
279
280static int ttm_copy_ttm_io_page(struct ttm_tt *ttm, void *dst,
281 unsigned long page,
282 pgprot_t prot)
283{
284 struct page *s = ttm_tt_get_page(ttm, page);
285 void *src;
286
287 if (!s)
288 return -ENOMEM;
289
290 dst = (void *)((unsigned long)dst + (page << PAGE_SHIFT));
291#ifdef CONFIG_X86
292 src = kmap_atomic_prot(s, prot);
293#else
294 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL))
295 src = vmap(&s, 1, 0, prot);
296 else
297 src = kmap(s);
298#endif
299 if (!src)
300 return -ENOMEM;
301
302 memcpy_toio(dst, src, PAGE_SIZE);
303
304#ifdef CONFIG_X86
305 kunmap_atomic(src);
306#else
307 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL))
308 vunmap(src);
309 else
310 kunmap(s);
311#endif
312
313 return 0;
314}
315
316int ttm_bo_move_memcpy(struct ttm_buffer_object *bo,
317 bool evict, bool no_wait_reserve, bool no_wait_gpu,
318 struct ttm_mem_reg *new_mem)
319{
320 struct ttm_bo_device *bdev = bo->bdev;
321 struct ttm_mem_type_manager *man = &bdev->man[new_mem->mem_type];
322 struct ttm_tt *ttm = bo->ttm;
323 struct ttm_mem_reg *old_mem = &bo->mem;
324 struct ttm_mem_reg old_copy = *old_mem;
325 void *old_iomap;
326 void *new_iomap;
327 int ret;
328 unsigned long i;
329 unsigned long page;
330 unsigned long add = 0;
331 int dir;
332
333 ret = ttm_mem_reg_ioremap(bdev, old_mem, &old_iomap);
334 if (ret)
335 return ret;
336 ret = ttm_mem_reg_ioremap(bdev, new_mem, &new_iomap);
337 if (ret)
338 goto out;
339
340 if (old_iomap == NULL && new_iomap == NULL)
341 goto out2;
342 if (old_iomap == NULL && ttm == NULL)
343 goto out2;
344
345 add = 0;
346 dir = 1;
347
348 if ((old_mem->mem_type == new_mem->mem_type) &&
349 (new_mem->start < old_mem->start + old_mem->size)) {
350 dir = -1;
351 add = new_mem->num_pages - 1;
352 }
353
354 for (i = 0; i < new_mem->num_pages; ++i) {
355 page = i * dir + add;
356 if (old_iomap == NULL) {
357 pgprot_t prot = ttm_io_prot(old_mem->placement,
358 PAGE_KERNEL);
359 ret = ttm_copy_ttm_io_page(ttm, new_iomap, page,
360 prot);
361 } else if (new_iomap == NULL) {
362 pgprot_t prot = ttm_io_prot(new_mem->placement,
363 PAGE_KERNEL);
364 ret = ttm_copy_io_ttm_page(ttm, old_iomap, page,
365 prot);
366 } else
367 ret = ttm_copy_io_page(new_iomap, old_iomap, page);
368 if (ret)
369 goto out1;
370 }
371 mb();
372out2:
373 old_copy = *old_mem;
374 *old_mem = *new_mem;
375 new_mem->mm_node = NULL;
376
377 if ((man->flags & TTM_MEMTYPE_FLAG_FIXED) && (ttm != NULL)) {
378 ttm_tt_unbind(ttm);
379 ttm_tt_destroy(ttm);
380 bo->ttm = NULL;
381 }
382
383out1:
384 ttm_mem_reg_iounmap(bdev, old_mem, new_iomap);
385out:
386 ttm_mem_reg_iounmap(bdev, &old_copy, old_iomap);
387 ttm_bo_mem_put(bo, &old_copy);
388 return ret;
389}
390EXPORT_SYMBOL(ttm_bo_move_memcpy);
391
392static void ttm_transfered_destroy(struct ttm_buffer_object *bo)
393{
394 kfree(bo);
395}
396
397/**
398 * ttm_buffer_object_transfer
399 *
400 * @bo: A pointer to a struct ttm_buffer_object.
401 * @new_obj: A pointer to a pointer to a newly created ttm_buffer_object,
402 * holding the data of @bo with the old placement.
403 *
404 * This is a utility function that may be called after an accelerated move
405 * has been scheduled. A new buffer object is created as a placeholder for
406 * the old data while it's being copied. When that buffer object is idle,
407 * it can be destroyed, releasing the space of the old placement.
408 * Returns:
409 * !0: Failure.
410 */
411
412static int ttm_buffer_object_transfer(struct ttm_buffer_object *bo,
413 struct ttm_buffer_object **new_obj)
414{
415 struct ttm_buffer_object *fbo;
416 struct ttm_bo_device *bdev = bo->bdev;
417 struct ttm_bo_driver *driver = bdev->driver;
418
419 fbo = kzalloc(sizeof(*fbo), GFP_KERNEL);
420 if (!fbo)
421 return -ENOMEM;
422
423 *fbo = *bo;
424
425 /**
426 * Fix up members that we shouldn't copy directly:
427 * TODO: Explicit member copy would probably be better here.
428 */
429
430 init_waitqueue_head(&fbo->event_queue);
431 INIT_LIST_HEAD(&fbo->ddestroy);
432 INIT_LIST_HEAD(&fbo->lru);
433 INIT_LIST_HEAD(&fbo->swap);
434 INIT_LIST_HEAD(&fbo->io_reserve_lru);
435 fbo->vm_node = NULL;
436 atomic_set(&fbo->cpu_writers, 0);
437
438 fbo->sync_obj = driver->sync_obj_ref(bo->sync_obj);
439 kref_init(&fbo->list_kref);
440 kref_init(&fbo->kref);
441 fbo->destroy = &ttm_transfered_destroy;
442
443 *new_obj = fbo;
444 return 0;
445}
446
447pgprot_t ttm_io_prot(uint32_t caching_flags, pgprot_t tmp)
448{
449#if defined(__i386__) || defined(__x86_64__)
450 if (caching_flags & TTM_PL_FLAG_WC)
451 tmp = pgprot_writecombine(tmp);
452 else if (boot_cpu_data.x86 > 3)
453 tmp = pgprot_noncached(tmp);
454
455#elif defined(__powerpc__)
456 if (!(caching_flags & TTM_PL_FLAG_CACHED)) {
457 pgprot_val(tmp) |= _PAGE_NO_CACHE;
458 if (caching_flags & TTM_PL_FLAG_UNCACHED)
459 pgprot_val(tmp) |= _PAGE_GUARDED;
460 }
461#endif
462#if defined(__ia64__)
463 if (caching_flags & TTM_PL_FLAG_WC)
464 tmp = pgprot_writecombine(tmp);
465 else
466 tmp = pgprot_noncached(tmp);
467#endif
468#if defined(__sparc__)
469 if (!(caching_flags & TTM_PL_FLAG_CACHED))
470 tmp = pgprot_noncached(tmp);
471#endif
472 return tmp;
473}
474EXPORT_SYMBOL(ttm_io_prot);
475
476static int ttm_bo_ioremap(struct ttm_buffer_object *bo,
477 unsigned long offset,
478 unsigned long size,
479 struct ttm_bo_kmap_obj *map)
480{
481 struct ttm_mem_reg *mem = &bo->mem;
482
483 if (bo->mem.bus.addr) {
484 map->bo_kmap_type = ttm_bo_map_premapped;
485 map->virtual = (void *)(((u8 *)bo->mem.bus.addr) + offset);
486 } else {
487 map->bo_kmap_type = ttm_bo_map_iomap;
488 if (mem->placement & TTM_PL_FLAG_WC)
489 map->virtual = ioremap_wc(bo->mem.bus.base + bo->mem.bus.offset + offset,
490 size);
491 else
492 map->virtual = ioremap_nocache(bo->mem.bus.base + bo->mem.bus.offset + offset,
493 size);
494 }
495 return (!map->virtual) ? -ENOMEM : 0;
496}
497
498static int ttm_bo_kmap_ttm(struct ttm_buffer_object *bo,
499 unsigned long start_page,
500 unsigned long num_pages,
501 struct ttm_bo_kmap_obj *map)
502{
503 struct ttm_mem_reg *mem = &bo->mem; pgprot_t prot;
504 struct ttm_tt *ttm = bo->ttm;
505 struct page *d;
506 int i;
507
508 BUG_ON(!ttm);
509 if (num_pages == 1 && (mem->placement & TTM_PL_FLAG_CACHED)) {
510 /*
511 * We're mapping a single page, and the desired
512 * page protection is consistent with the bo.
513 */
514
515 map->bo_kmap_type = ttm_bo_map_kmap;
516 map->page = ttm_tt_get_page(ttm, start_page);
517 map->virtual = kmap(map->page);
518 } else {
519 /*
520 * Populate the part we're mapping;
521 */
522 for (i = start_page; i < start_page + num_pages; ++i) {
523 d = ttm_tt_get_page(ttm, i);
524 if (!d)
525 return -ENOMEM;
526 }
527
528 /*
529 * We need to use vmap to get the desired page protection
530 * or to make the buffer object look contiguous.
531 */
532 prot = (mem->placement & TTM_PL_FLAG_CACHED) ?
533 PAGE_KERNEL :
534 ttm_io_prot(mem->placement, PAGE_KERNEL);
535 map->bo_kmap_type = ttm_bo_map_vmap;
536 map->virtual = vmap(ttm->pages + start_page, num_pages,
537 0, prot);
538 }
539 return (!map->virtual) ? -ENOMEM : 0;
540}
541
542int ttm_bo_kmap(struct ttm_buffer_object *bo,
543 unsigned long start_page, unsigned long num_pages,
544 struct ttm_bo_kmap_obj *map)
545{
546 struct ttm_mem_type_manager *man =
547 &bo->bdev->man[bo->mem.mem_type];
548 unsigned long offset, size;
549 int ret;
550
551 BUG_ON(!list_empty(&bo->swap));
552 map->virtual = NULL;
553 map->bo = bo;
554 if (num_pages > bo->num_pages)
555 return -EINVAL;
556 if (start_page > bo->num_pages)
557 return -EINVAL;
558#if 0
559 if (num_pages > 1 && !DRM_SUSER(DRM_CURPROC))
560 return -EPERM;
561#endif
562 (void) ttm_mem_io_lock(man, false);
563 ret = ttm_mem_io_reserve(bo->bdev, &bo->mem);
564 ttm_mem_io_unlock(man);
565 if (ret)
566 return ret;
567 if (!bo->mem.bus.is_iomem) {
568 return ttm_bo_kmap_ttm(bo, start_page, num_pages, map);
569 } else {
570 offset = start_page << PAGE_SHIFT;
571 size = num_pages << PAGE_SHIFT;
572 return ttm_bo_ioremap(bo, offset, size, map);
573 }
574}
575EXPORT_SYMBOL(ttm_bo_kmap);
576
577void ttm_bo_kunmap(struct ttm_bo_kmap_obj *map)
578{
579 struct ttm_buffer_object *bo = map->bo;
580 struct ttm_mem_type_manager *man =
581 &bo->bdev->man[bo->mem.mem_type];
582
583 if (!map->virtual)
584 return;
585 switch (map->bo_kmap_type) {
586 case ttm_bo_map_iomap:
587 iounmap(map->virtual);
588 break;
589 case ttm_bo_map_vmap:
590 vunmap(map->virtual);
591 break;
592 case ttm_bo_map_kmap:
593 kunmap(map->page);
594 break;
595 case ttm_bo_map_premapped:
596 break;
597 default:
598 BUG();
599 }
600 (void) ttm_mem_io_lock(man, false);
601 ttm_mem_io_free(map->bo->bdev, &map->bo->mem);
602 ttm_mem_io_unlock(man);
603 map->virtual = NULL;
604 map->page = NULL;
605}
606EXPORT_SYMBOL(ttm_bo_kunmap);
607
608int ttm_bo_move_accel_cleanup(struct ttm_buffer_object *bo,
609 void *sync_obj,
610 void *sync_obj_arg,
611 bool evict, bool no_wait_reserve,
612 bool no_wait_gpu,
613 struct ttm_mem_reg *new_mem)
614{
615 struct ttm_bo_device *bdev = bo->bdev;
616 struct ttm_bo_driver *driver = bdev->driver;
617 struct ttm_mem_type_manager *man = &bdev->man[new_mem->mem_type];
618 struct ttm_mem_reg *old_mem = &bo->mem;
619 int ret;
620 struct ttm_buffer_object *ghost_obj;
621 void *tmp_obj = NULL;
622
623 spin_lock(&bdev->fence_lock);
624 if (bo->sync_obj) {
625 tmp_obj = bo->sync_obj;
626 bo->sync_obj = NULL;
627 }
628 bo->sync_obj = driver->sync_obj_ref(sync_obj);
629 bo->sync_obj_arg = sync_obj_arg;
630 if (evict) {
631 ret = ttm_bo_wait(bo, false, false, false);
632 spin_unlock(&bdev->fence_lock);
633 if (tmp_obj)
634 driver->sync_obj_unref(&tmp_obj);
635 if (ret)
636 return ret;
637
638 if ((man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
639 (bo->ttm != NULL)) {
640 ttm_tt_unbind(bo->ttm);
641 ttm_tt_destroy(bo->ttm);
642 bo->ttm = NULL;
643 }
644 ttm_bo_free_old_node(bo);
645 } else {
646 /**
647 * This should help pipeline ordinary buffer moves.
648 *
649 * Hang old buffer memory on a new buffer object,
650 * and leave it to be released when the GPU
651 * operation has completed.
652 */
653
654 set_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
655 spin_unlock(&bdev->fence_lock);
656 if (tmp_obj)
657 driver->sync_obj_unref(&tmp_obj);
658
659 ret = ttm_buffer_object_transfer(bo, &ghost_obj);
660 if (ret)
661 return ret;
662
663 /**
664 * If we're not moving to fixed memory, the TTM object
665 * needs to stay alive. Otherwhise hang it on the ghost
666 * bo to be unbound and destroyed.
667 */
668
669 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED))
670 ghost_obj->ttm = NULL;
671 else
672 bo->ttm = NULL;
673
674 ttm_bo_unreserve(ghost_obj);
675 ttm_bo_unref(&ghost_obj);
676 }
677
678 *old_mem = *new_mem;
679 new_mem->mm_node = NULL;
680
681 return 0;
682}
683EXPORT_SYMBOL(ttm_bo_move_accel_cleanup);
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 interruptible, bool no_wait_gpu,
49 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 ret = ttm_bo_wait(bo, interruptible, no_wait_gpu);
57
58 if (unlikely(ret != 0)) {
59 if (ret != -ERESTARTSYS)
60 pr_err("Failed to expire sync object before unbinding TTM\n");
61 return ret;
62 }
63
64 ttm_tt_unbind(ttm);
65 ttm_bo_free_old_node(bo);
66 ttm_flag_masked(&old_mem->placement, TTM_PL_FLAG_SYSTEM,
67 TTM_PL_MASK_MEM);
68 old_mem->mem_type = TTM_PL_SYSTEM;
69 }
70
71 ret = ttm_tt_set_placement_caching(ttm, new_mem->placement);
72 if (unlikely(ret != 0))
73 return ret;
74
75 if (new_mem->mem_type != TTM_PL_SYSTEM) {
76 ret = ttm_tt_bind(ttm, new_mem);
77 if (unlikely(ret != 0))
78 return ret;
79 }
80
81 *old_mem = *new_mem;
82 new_mem->mm_node = NULL;
83
84 return 0;
85}
86EXPORT_SYMBOL(ttm_bo_move_ttm);
87
88int ttm_mem_io_lock(struct ttm_mem_type_manager *man, bool interruptible)
89{
90 if (likely(man->io_reserve_fastpath))
91 return 0;
92
93 if (interruptible)
94 return mutex_lock_interruptible(&man->io_reserve_mutex);
95
96 mutex_lock(&man->io_reserve_mutex);
97 return 0;
98}
99EXPORT_SYMBOL(ttm_mem_io_lock);
100
101void ttm_mem_io_unlock(struct ttm_mem_type_manager *man)
102{
103 if (likely(man->io_reserve_fastpath))
104 return;
105
106 mutex_unlock(&man->io_reserve_mutex);
107}
108EXPORT_SYMBOL(ttm_mem_io_unlock);
109
110static int ttm_mem_io_evict(struct ttm_mem_type_manager *man)
111{
112 struct ttm_buffer_object *bo;
113
114 if (!man->use_io_reserve_lru || list_empty(&man->io_reserve_lru))
115 return -EAGAIN;
116
117 bo = list_first_entry(&man->io_reserve_lru,
118 struct ttm_buffer_object,
119 io_reserve_lru);
120 list_del_init(&bo->io_reserve_lru);
121 ttm_bo_unmap_virtual_locked(bo);
122
123 return 0;
124}
125
126
127int ttm_mem_io_reserve(struct ttm_bo_device *bdev,
128 struct ttm_mem_reg *mem)
129{
130 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
131 int ret = 0;
132
133 if (!bdev->driver->io_mem_reserve)
134 return 0;
135 if (likely(man->io_reserve_fastpath))
136 return bdev->driver->io_mem_reserve(bdev, mem);
137
138 if (bdev->driver->io_mem_reserve &&
139 mem->bus.io_reserved_count++ == 0) {
140retry:
141 ret = bdev->driver->io_mem_reserve(bdev, mem);
142 if (ret == -EAGAIN) {
143 ret = ttm_mem_io_evict(man);
144 if (ret == 0)
145 goto retry;
146 }
147 }
148 return ret;
149}
150EXPORT_SYMBOL(ttm_mem_io_reserve);
151
152void ttm_mem_io_free(struct ttm_bo_device *bdev,
153 struct ttm_mem_reg *mem)
154{
155 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
156
157 if (likely(man->io_reserve_fastpath))
158 return;
159
160 if (bdev->driver->io_mem_reserve &&
161 --mem->bus.io_reserved_count == 0 &&
162 bdev->driver->io_mem_free)
163 bdev->driver->io_mem_free(bdev, mem);
164
165}
166EXPORT_SYMBOL(ttm_mem_io_free);
167
168int ttm_mem_io_reserve_vm(struct ttm_buffer_object *bo)
169{
170 struct ttm_mem_reg *mem = &bo->mem;
171 int ret;
172
173 if (!mem->bus.io_reserved_vm) {
174 struct ttm_mem_type_manager *man =
175 &bo->bdev->man[mem->mem_type];
176
177 ret = ttm_mem_io_reserve(bo->bdev, mem);
178 if (unlikely(ret != 0))
179 return ret;
180 mem->bus.io_reserved_vm = true;
181 if (man->use_io_reserve_lru)
182 list_add_tail(&bo->io_reserve_lru,
183 &man->io_reserve_lru);
184 }
185 return 0;
186}
187
188void ttm_mem_io_free_vm(struct ttm_buffer_object *bo)
189{
190 struct ttm_mem_reg *mem = &bo->mem;
191
192 if (mem->bus.io_reserved_vm) {
193 mem->bus.io_reserved_vm = false;
194 list_del_init(&bo->io_reserve_lru);
195 ttm_mem_io_free(bo->bdev, mem);
196 }
197}
198
199static int ttm_mem_reg_ioremap(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem,
200 void **virtual)
201{
202 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
203 int ret;
204 void *addr;
205
206 *virtual = NULL;
207 (void) ttm_mem_io_lock(man, false);
208 ret = ttm_mem_io_reserve(bdev, mem);
209 ttm_mem_io_unlock(man);
210 if (ret || !mem->bus.is_iomem)
211 return ret;
212
213 if (mem->bus.addr) {
214 addr = mem->bus.addr;
215 } else {
216 if (mem->placement & TTM_PL_FLAG_WC)
217 addr = ioremap_wc(mem->bus.base + mem->bus.offset, mem->bus.size);
218 else
219 addr = ioremap_nocache(mem->bus.base + mem->bus.offset, mem->bus.size);
220 if (!addr) {
221 (void) ttm_mem_io_lock(man, false);
222 ttm_mem_io_free(bdev, mem);
223 ttm_mem_io_unlock(man);
224 return -ENOMEM;
225 }
226 }
227 *virtual = addr;
228 return 0;
229}
230
231static void ttm_mem_reg_iounmap(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem,
232 void *virtual)
233{
234 struct ttm_mem_type_manager *man;
235
236 man = &bdev->man[mem->mem_type];
237
238 if (virtual && mem->bus.addr == NULL)
239 iounmap(virtual);
240 (void) ttm_mem_io_lock(man, false);
241 ttm_mem_io_free(bdev, mem);
242 ttm_mem_io_unlock(man);
243}
244
245static int ttm_copy_io_page(void *dst, void *src, unsigned long page)
246{
247 uint32_t *dstP =
248 (uint32_t *) ((unsigned long)dst + (page << PAGE_SHIFT));
249 uint32_t *srcP =
250 (uint32_t *) ((unsigned long)src + (page << PAGE_SHIFT));
251
252 int i;
253 for (i = 0; i < PAGE_SIZE / sizeof(uint32_t); ++i)
254 iowrite32(ioread32(srcP++), dstP++);
255 return 0;
256}
257
258static int ttm_copy_io_ttm_page(struct ttm_tt *ttm, void *src,
259 unsigned long page,
260 pgprot_t prot)
261{
262 struct page *d = ttm->pages[page];
263 void *dst;
264
265 if (!d)
266 return -ENOMEM;
267
268 src = (void *)((unsigned long)src + (page << PAGE_SHIFT));
269
270#ifdef CONFIG_X86
271 dst = kmap_atomic_prot(d, prot);
272#else
273 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL))
274 dst = vmap(&d, 1, 0, prot);
275 else
276 dst = kmap(d);
277#endif
278 if (!dst)
279 return -ENOMEM;
280
281 memcpy_fromio(dst, src, PAGE_SIZE);
282
283#ifdef CONFIG_X86
284 kunmap_atomic(dst);
285#else
286 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL))
287 vunmap(dst);
288 else
289 kunmap(d);
290#endif
291
292 return 0;
293}
294
295static int ttm_copy_ttm_io_page(struct ttm_tt *ttm, void *dst,
296 unsigned long page,
297 pgprot_t prot)
298{
299 struct page *s = ttm->pages[page];
300 void *src;
301
302 if (!s)
303 return -ENOMEM;
304
305 dst = (void *)((unsigned long)dst + (page << PAGE_SHIFT));
306#ifdef CONFIG_X86
307 src = kmap_atomic_prot(s, prot);
308#else
309 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL))
310 src = vmap(&s, 1, 0, prot);
311 else
312 src = kmap(s);
313#endif
314 if (!src)
315 return -ENOMEM;
316
317 memcpy_toio(dst, src, PAGE_SIZE);
318
319#ifdef CONFIG_X86
320 kunmap_atomic(src);
321#else
322 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL))
323 vunmap(src);
324 else
325 kunmap(s);
326#endif
327
328 return 0;
329}
330
331int ttm_bo_move_memcpy(struct ttm_buffer_object *bo,
332 bool interruptible, bool no_wait_gpu,
333 struct ttm_mem_reg *new_mem)
334{
335 struct ttm_bo_device *bdev = bo->bdev;
336 struct ttm_mem_type_manager *man = &bdev->man[new_mem->mem_type];
337 struct ttm_tt *ttm = bo->ttm;
338 struct ttm_mem_reg *old_mem = &bo->mem;
339 struct ttm_mem_reg old_copy = *old_mem;
340 void *old_iomap;
341 void *new_iomap;
342 int ret;
343 unsigned long i;
344 unsigned long page;
345 unsigned long add = 0;
346 int dir;
347
348 ret = ttm_bo_wait(bo, interruptible, no_wait_gpu);
349 if (ret)
350 return ret;
351
352 ret = ttm_mem_reg_ioremap(bdev, old_mem, &old_iomap);
353 if (ret)
354 return ret;
355 ret = ttm_mem_reg_ioremap(bdev, new_mem, &new_iomap);
356 if (ret)
357 goto out;
358
359 /*
360 * Single TTM move. NOP.
361 */
362 if (old_iomap == NULL && new_iomap == NULL)
363 goto out2;
364
365 /*
366 * Don't move nonexistent data. Clear destination instead.
367 */
368 if (old_iomap == NULL &&
369 (ttm == NULL || (ttm->state == tt_unpopulated &&
370 !(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)))) {
371 memset_io(new_iomap, 0, new_mem->num_pages*PAGE_SIZE);
372 goto out2;
373 }
374
375 /*
376 * TTM might be null for moves within the same region.
377 */
378 if (ttm && ttm->state == tt_unpopulated) {
379 ret = ttm->bdev->driver->ttm_tt_populate(ttm);
380 if (ret)
381 goto out1;
382 }
383
384 add = 0;
385 dir = 1;
386
387 if ((old_mem->mem_type == new_mem->mem_type) &&
388 (new_mem->start < old_mem->start + old_mem->size)) {
389 dir = -1;
390 add = new_mem->num_pages - 1;
391 }
392
393 for (i = 0; i < new_mem->num_pages; ++i) {
394 page = i * dir + add;
395 if (old_iomap == NULL) {
396 pgprot_t prot = ttm_io_prot(old_mem->placement,
397 PAGE_KERNEL);
398 ret = ttm_copy_ttm_io_page(ttm, new_iomap, page,
399 prot);
400 } else if (new_iomap == NULL) {
401 pgprot_t prot = ttm_io_prot(new_mem->placement,
402 PAGE_KERNEL);
403 ret = ttm_copy_io_ttm_page(ttm, old_iomap, page,
404 prot);
405 } else
406 ret = ttm_copy_io_page(new_iomap, old_iomap, page);
407 if (ret)
408 goto out1;
409 }
410 mb();
411out2:
412 old_copy = *old_mem;
413 *old_mem = *new_mem;
414 new_mem->mm_node = NULL;
415
416 if (man->flags & TTM_MEMTYPE_FLAG_FIXED) {
417 ttm_tt_destroy(ttm);
418 bo->ttm = NULL;
419 }
420
421out1:
422 ttm_mem_reg_iounmap(bdev, old_mem, new_iomap);
423out:
424 ttm_mem_reg_iounmap(bdev, &old_copy, old_iomap);
425
426 /*
427 * On error, keep the mm node!
428 */
429 if (!ret)
430 ttm_bo_mem_put(bo, &old_copy);
431 return ret;
432}
433EXPORT_SYMBOL(ttm_bo_move_memcpy);
434
435static void ttm_transfered_destroy(struct ttm_buffer_object *bo)
436{
437 kfree(bo);
438}
439
440/**
441 * ttm_buffer_object_transfer
442 *
443 * @bo: A pointer to a struct ttm_buffer_object.
444 * @new_obj: A pointer to a pointer to a newly created ttm_buffer_object,
445 * holding the data of @bo with the old placement.
446 *
447 * This is a utility function that may be called after an accelerated move
448 * has been scheduled. A new buffer object is created as a placeholder for
449 * the old data while it's being copied. When that buffer object is idle,
450 * it can be destroyed, releasing the space of the old placement.
451 * Returns:
452 * !0: Failure.
453 */
454
455static int ttm_buffer_object_transfer(struct ttm_buffer_object *bo,
456 struct ttm_buffer_object **new_obj)
457{
458 struct ttm_buffer_object *fbo;
459 int ret;
460
461 fbo = kmalloc(sizeof(*fbo), GFP_KERNEL);
462 if (!fbo)
463 return -ENOMEM;
464
465 *fbo = *bo;
466
467 /**
468 * Fix up members that we shouldn't copy directly:
469 * TODO: Explicit member copy would probably be better here.
470 */
471
472 INIT_LIST_HEAD(&fbo->ddestroy);
473 INIT_LIST_HEAD(&fbo->lru);
474 INIT_LIST_HEAD(&fbo->swap);
475 INIT_LIST_HEAD(&fbo->io_reserve_lru);
476 fbo->moving = NULL;
477 drm_vma_node_reset(&fbo->vma_node);
478 atomic_set(&fbo->cpu_writers, 0);
479
480 kref_init(&fbo->list_kref);
481 kref_init(&fbo->kref);
482 fbo->destroy = &ttm_transfered_destroy;
483 fbo->acc_size = 0;
484 fbo->resv = &fbo->ttm_resv;
485 reservation_object_init(fbo->resv);
486 ret = ww_mutex_trylock(&fbo->resv->lock);
487 WARN_ON(!ret);
488
489 *new_obj = fbo;
490 return 0;
491}
492
493pgprot_t ttm_io_prot(uint32_t caching_flags, pgprot_t tmp)
494{
495 /* Cached mappings need no adjustment */
496 if (caching_flags & TTM_PL_FLAG_CACHED)
497 return tmp;
498
499#if defined(__i386__) || defined(__x86_64__)
500 if (caching_flags & TTM_PL_FLAG_WC)
501 tmp = pgprot_writecombine(tmp);
502 else if (boot_cpu_data.x86 > 3)
503 tmp = pgprot_noncached(tmp);
504#endif
505#if defined(__ia64__) || defined(__arm__) || defined(__aarch64__) || \
506 defined(__powerpc__)
507 if (caching_flags & TTM_PL_FLAG_WC)
508 tmp = pgprot_writecombine(tmp);
509 else
510 tmp = pgprot_noncached(tmp);
511#endif
512#if defined(__sparc__) || defined(__mips__)
513 tmp = pgprot_noncached(tmp);
514#endif
515 return tmp;
516}
517EXPORT_SYMBOL(ttm_io_prot);
518
519static int ttm_bo_ioremap(struct ttm_buffer_object *bo,
520 unsigned long offset,
521 unsigned long size,
522 struct ttm_bo_kmap_obj *map)
523{
524 struct ttm_mem_reg *mem = &bo->mem;
525
526 if (bo->mem.bus.addr) {
527 map->bo_kmap_type = ttm_bo_map_premapped;
528 map->virtual = (void *)(((u8 *)bo->mem.bus.addr) + offset);
529 } else {
530 map->bo_kmap_type = ttm_bo_map_iomap;
531 if (mem->placement & TTM_PL_FLAG_WC)
532 map->virtual = ioremap_wc(bo->mem.bus.base + bo->mem.bus.offset + offset,
533 size);
534 else
535 map->virtual = ioremap_nocache(bo->mem.bus.base + bo->mem.bus.offset + offset,
536 size);
537 }
538 return (!map->virtual) ? -ENOMEM : 0;
539}
540
541static int ttm_bo_kmap_ttm(struct ttm_buffer_object *bo,
542 unsigned long start_page,
543 unsigned long num_pages,
544 struct ttm_bo_kmap_obj *map)
545{
546 struct ttm_mem_reg *mem = &bo->mem; pgprot_t prot;
547 struct ttm_tt *ttm = bo->ttm;
548 int ret;
549
550 BUG_ON(!ttm);
551
552 if (ttm->state == tt_unpopulated) {
553 ret = ttm->bdev->driver->ttm_tt_populate(ttm);
554 if (ret)
555 return ret;
556 }
557
558 if (num_pages == 1 && (mem->placement & TTM_PL_FLAG_CACHED)) {
559 /*
560 * We're mapping a single page, and the desired
561 * page protection is consistent with the bo.
562 */
563
564 map->bo_kmap_type = ttm_bo_map_kmap;
565 map->page = ttm->pages[start_page];
566 map->virtual = kmap(map->page);
567 } else {
568 /*
569 * We need to use vmap to get the desired page protection
570 * or to make the buffer object look contiguous.
571 */
572 prot = 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 struct dma_fence *fence,
648 bool evict,
649 struct ttm_mem_reg *new_mem)
650{
651 struct ttm_bo_device *bdev = bo->bdev;
652 struct ttm_mem_type_manager *man = &bdev->man[new_mem->mem_type];
653 struct ttm_mem_reg *old_mem = &bo->mem;
654 int ret;
655 struct ttm_buffer_object *ghost_obj;
656
657 reservation_object_add_excl_fence(bo->resv, fence);
658 if (evict) {
659 ret = ttm_bo_wait(bo, false, false);
660 if (ret)
661 return ret;
662
663 if (man->flags & TTM_MEMTYPE_FLAG_FIXED) {
664 ttm_tt_destroy(bo->ttm);
665 bo->ttm = NULL;
666 }
667 ttm_bo_free_old_node(bo);
668 } else {
669 /**
670 * This should help pipeline ordinary buffer moves.
671 *
672 * Hang old buffer memory on a new buffer object,
673 * and leave it to be released when the GPU
674 * operation has completed.
675 */
676
677 dma_fence_put(bo->moving);
678 bo->moving = dma_fence_get(fence);
679
680 ret = ttm_buffer_object_transfer(bo, &ghost_obj);
681 if (ret)
682 return ret;
683
684 reservation_object_add_excl_fence(ghost_obj->resv, fence);
685
686 /**
687 * If we're not moving to fixed memory, the TTM object
688 * needs to stay alive. Otherwhise hang it on the ghost
689 * bo to be unbound and destroyed.
690 */
691
692 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED))
693 ghost_obj->ttm = NULL;
694 else
695 bo->ttm = NULL;
696
697 ttm_bo_unreserve(ghost_obj);
698 ttm_bo_unref(&ghost_obj);
699 }
700
701 *old_mem = *new_mem;
702 new_mem->mm_node = NULL;
703
704 return 0;
705}
706EXPORT_SYMBOL(ttm_bo_move_accel_cleanup);
707
708int ttm_bo_pipeline_move(struct ttm_buffer_object *bo,
709 struct dma_fence *fence, bool evict,
710 struct ttm_mem_reg *new_mem)
711{
712 struct ttm_bo_device *bdev = bo->bdev;
713 struct ttm_mem_reg *old_mem = &bo->mem;
714
715 struct ttm_mem_type_manager *from = &bdev->man[old_mem->mem_type];
716 struct ttm_mem_type_manager *to = &bdev->man[new_mem->mem_type];
717
718 int ret;
719
720 reservation_object_add_excl_fence(bo->resv, fence);
721
722 if (!evict) {
723 struct ttm_buffer_object *ghost_obj;
724
725 /**
726 * This should help pipeline ordinary buffer moves.
727 *
728 * Hang old buffer memory on a new buffer object,
729 * and leave it to be released when the GPU
730 * operation has completed.
731 */
732
733 dma_fence_put(bo->moving);
734 bo->moving = dma_fence_get(fence);
735
736 ret = ttm_buffer_object_transfer(bo, &ghost_obj);
737 if (ret)
738 return ret;
739
740 reservation_object_add_excl_fence(ghost_obj->resv, fence);
741
742 /**
743 * If we're not moving to fixed memory, the TTM object
744 * needs to stay alive. Otherwhise hang it on the ghost
745 * bo to be unbound and destroyed.
746 */
747
748 if (!(to->flags & TTM_MEMTYPE_FLAG_FIXED))
749 ghost_obj->ttm = NULL;
750 else
751 bo->ttm = NULL;
752
753 ttm_bo_unreserve(ghost_obj);
754 ttm_bo_unref(&ghost_obj);
755
756 } else if (from->flags & TTM_MEMTYPE_FLAG_FIXED) {
757
758 /**
759 * BO doesn't have a TTM we need to bind/unbind. Just remember
760 * this eviction and free up the allocation
761 */
762
763 spin_lock(&from->move_lock);
764 if (!from->move || dma_fence_is_later(fence, from->move)) {
765 dma_fence_put(from->move);
766 from->move = dma_fence_get(fence);
767 }
768 spin_unlock(&from->move_lock);
769
770 ttm_bo_free_old_node(bo);
771
772 dma_fence_put(bo->moving);
773 bo->moving = dma_fence_get(fence);
774
775 } else {
776 /**
777 * Last resort, wait for the move to be completed.
778 *
779 * Should never happen in pratice.
780 */
781
782 ret = ttm_bo_wait(bo, false, false);
783 if (ret)
784 return ret;
785
786 if (to->flags & TTM_MEMTYPE_FLAG_FIXED) {
787 ttm_tt_destroy(bo->ttm);
788 bo->ttm = NULL;
789 }
790 ttm_bo_free_old_node(bo);
791 }
792
793 *old_mem = *new_mem;
794 new_mem->mm_node = NULL;
795
796 return 0;
797}
798EXPORT_SYMBOL(ttm_bo_pipeline_move);