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