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/**************************************************************************
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->pages[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->pages[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 if (ttm->state == tt_unpopulated) {
346 ret = ttm->bdev->driver->ttm_tt_populate(ttm);
347 if (ret)
348 goto out1;
349 }
350
351 add = 0;
352 dir = 1;
353
354 if ((old_mem->mem_type == new_mem->mem_type) &&
355 (new_mem->start < old_mem->start + old_mem->size)) {
356 dir = -1;
357 add = new_mem->num_pages - 1;
358 }
359
360 for (i = 0; i < new_mem->num_pages; ++i) {
361 page = i * dir + add;
362 if (old_iomap == NULL) {
363 pgprot_t prot = ttm_io_prot(old_mem->placement,
364 PAGE_KERNEL);
365 ret = ttm_copy_ttm_io_page(ttm, new_iomap, page,
366 prot);
367 } else if (new_iomap == NULL) {
368 pgprot_t prot = ttm_io_prot(new_mem->placement,
369 PAGE_KERNEL);
370 ret = ttm_copy_io_ttm_page(ttm, old_iomap, page,
371 prot);
372 } else
373 ret = ttm_copy_io_page(new_iomap, old_iomap, page);
374 if (ret)
375 goto out1;
376 }
377 mb();
378out2:
379 old_copy = *old_mem;
380 *old_mem = *new_mem;
381 new_mem->mm_node = NULL;
382
383 if ((man->flags & TTM_MEMTYPE_FLAG_FIXED) && (ttm != NULL)) {
384 ttm_tt_unbind(ttm);
385 ttm_tt_destroy(ttm);
386 bo->ttm = NULL;
387 }
388
389out1:
390 ttm_mem_reg_iounmap(bdev, old_mem, new_iomap);
391out:
392 ttm_mem_reg_iounmap(bdev, &old_copy, old_iomap);
393 ttm_bo_mem_put(bo, &old_copy);
394 return ret;
395}
396EXPORT_SYMBOL(ttm_bo_move_memcpy);
397
398static void ttm_transfered_destroy(struct ttm_buffer_object *bo)
399{
400 kfree(bo);
401}
402
403/**
404 * ttm_buffer_object_transfer
405 *
406 * @bo: A pointer to a struct ttm_buffer_object.
407 * @new_obj: A pointer to a pointer to a newly created ttm_buffer_object,
408 * holding the data of @bo with the old placement.
409 *
410 * This is a utility function that may be called after an accelerated move
411 * has been scheduled. A new buffer object is created as a placeholder for
412 * the old data while it's being copied. When that buffer object is idle,
413 * it can be destroyed, releasing the space of the old placement.
414 * Returns:
415 * !0: Failure.
416 */
417
418static int ttm_buffer_object_transfer(struct ttm_buffer_object *bo,
419 struct ttm_buffer_object **new_obj)
420{
421 struct ttm_buffer_object *fbo;
422 struct ttm_bo_device *bdev = bo->bdev;
423 struct ttm_bo_driver *driver = bdev->driver;
424
425 fbo = kzalloc(sizeof(*fbo), GFP_KERNEL);
426 if (!fbo)
427 return -ENOMEM;
428
429 *fbo = *bo;
430
431 /**
432 * Fix up members that we shouldn't copy directly:
433 * TODO: Explicit member copy would probably be better here.
434 */
435
436 init_waitqueue_head(&fbo->event_queue);
437 INIT_LIST_HEAD(&fbo->ddestroy);
438 INIT_LIST_HEAD(&fbo->lru);
439 INIT_LIST_HEAD(&fbo->swap);
440 INIT_LIST_HEAD(&fbo->io_reserve_lru);
441 fbo->vm_node = NULL;
442 atomic_set(&fbo->cpu_writers, 0);
443
444 fbo->sync_obj = driver->sync_obj_ref(bo->sync_obj);
445 kref_init(&fbo->list_kref);
446 kref_init(&fbo->kref);
447 fbo->destroy = &ttm_transfered_destroy;
448 fbo->acc_size = 0;
449
450 *new_obj = fbo;
451 return 0;
452}
453
454pgprot_t ttm_io_prot(uint32_t caching_flags, pgprot_t tmp)
455{
456#if defined(__i386__) || defined(__x86_64__)
457 if (caching_flags & TTM_PL_FLAG_WC)
458 tmp = pgprot_writecombine(tmp);
459 else if (boot_cpu_data.x86 > 3)
460 tmp = pgprot_noncached(tmp);
461
462#elif defined(__powerpc__)
463 if (!(caching_flags & TTM_PL_FLAG_CACHED)) {
464 pgprot_val(tmp) |= _PAGE_NO_CACHE;
465 if (caching_flags & TTM_PL_FLAG_UNCACHED)
466 pgprot_val(tmp) |= _PAGE_GUARDED;
467 }
468#endif
469#if defined(__ia64__)
470 if (caching_flags & TTM_PL_FLAG_WC)
471 tmp = pgprot_writecombine(tmp);
472 else
473 tmp = pgprot_noncached(tmp);
474#endif
475#if defined(__sparc__)
476 if (!(caching_flags & TTM_PL_FLAG_CACHED))
477 tmp = pgprot_noncached(tmp);
478#endif
479 return tmp;
480}
481EXPORT_SYMBOL(ttm_io_prot);
482
483static int ttm_bo_ioremap(struct ttm_buffer_object *bo,
484 unsigned long offset,
485 unsigned long size,
486 struct ttm_bo_kmap_obj *map)
487{
488 struct ttm_mem_reg *mem = &bo->mem;
489
490 if (bo->mem.bus.addr) {
491 map->bo_kmap_type = ttm_bo_map_premapped;
492 map->virtual = (void *)(((u8 *)bo->mem.bus.addr) + offset);
493 } else {
494 map->bo_kmap_type = ttm_bo_map_iomap;
495 if (mem->placement & TTM_PL_FLAG_WC)
496 map->virtual = ioremap_wc(bo->mem.bus.base + bo->mem.bus.offset + offset,
497 size);
498 else
499 map->virtual = ioremap_nocache(bo->mem.bus.base + bo->mem.bus.offset + offset,
500 size);
501 }
502 return (!map->virtual) ? -ENOMEM : 0;
503}
504
505static int ttm_bo_kmap_ttm(struct ttm_buffer_object *bo,
506 unsigned long start_page,
507 unsigned long num_pages,
508 struct ttm_bo_kmap_obj *map)
509{
510 struct ttm_mem_reg *mem = &bo->mem; pgprot_t prot;
511 struct ttm_tt *ttm = bo->ttm;
512 int ret;
513
514 BUG_ON(!ttm);
515
516 if (ttm->state == tt_unpopulated) {
517 ret = ttm->bdev->driver->ttm_tt_populate(ttm);
518 if (ret)
519 return ret;
520 }
521
522 if (num_pages == 1 && (mem->placement & TTM_PL_FLAG_CACHED)) {
523 /*
524 * We're mapping a single page, and the desired
525 * page protection is consistent with the bo.
526 */
527
528 map->bo_kmap_type = ttm_bo_map_kmap;
529 map->page = ttm->pages[start_page];
530 map->virtual = kmap(map->page);
531 } else {
532 /*
533 * We need to use vmap to get the desired page protection
534 * or to make the buffer object look contiguous.
535 */
536 prot = (mem->placement & TTM_PL_FLAG_CACHED) ?
537 PAGE_KERNEL :
538 ttm_io_prot(mem->placement, PAGE_KERNEL);
539 map->bo_kmap_type = ttm_bo_map_vmap;
540 map->virtual = vmap(ttm->pages + start_page, num_pages,
541 0, prot);
542 }
543 return (!map->virtual) ? -ENOMEM : 0;
544}
545
546int ttm_bo_kmap(struct ttm_buffer_object *bo,
547 unsigned long start_page, unsigned long num_pages,
548 struct ttm_bo_kmap_obj *map)
549{
550 struct ttm_mem_type_manager *man =
551 &bo->bdev->man[bo->mem.mem_type];
552 unsigned long offset, size;
553 int ret;
554
555 BUG_ON(!list_empty(&bo->swap));
556 map->virtual = NULL;
557 map->bo = bo;
558 if (num_pages > bo->num_pages)
559 return -EINVAL;
560 if (start_page > bo->num_pages)
561 return -EINVAL;
562#if 0
563 if (num_pages > 1 && !DRM_SUSER(DRM_CURPROC))
564 return -EPERM;
565#endif
566 (void) ttm_mem_io_lock(man, false);
567 ret = ttm_mem_io_reserve(bo->bdev, &bo->mem);
568 ttm_mem_io_unlock(man);
569 if (ret)
570 return ret;
571 if (!bo->mem.bus.is_iomem) {
572 return ttm_bo_kmap_ttm(bo, start_page, num_pages, map);
573 } else {
574 offset = start_page << PAGE_SHIFT;
575 size = num_pages << PAGE_SHIFT;
576 return ttm_bo_ioremap(bo, offset, size, map);
577 }
578}
579EXPORT_SYMBOL(ttm_bo_kmap);
580
581void ttm_bo_kunmap(struct ttm_bo_kmap_obj *map)
582{
583 struct ttm_buffer_object *bo = map->bo;
584 struct ttm_mem_type_manager *man =
585 &bo->bdev->man[bo->mem.mem_type];
586
587 if (!map->virtual)
588 return;
589 switch (map->bo_kmap_type) {
590 case ttm_bo_map_iomap:
591 iounmap(map->virtual);
592 break;
593 case ttm_bo_map_vmap:
594 vunmap(map->virtual);
595 break;
596 case ttm_bo_map_kmap:
597 kunmap(map->page);
598 break;
599 case ttm_bo_map_premapped:
600 break;
601 default:
602 BUG();
603 }
604 (void) ttm_mem_io_lock(man, false);
605 ttm_mem_io_free(map->bo->bdev, &map->bo->mem);
606 ttm_mem_io_unlock(man);
607 map->virtual = NULL;
608 map->page = NULL;
609}
610EXPORT_SYMBOL(ttm_bo_kunmap);
611
612int ttm_bo_move_accel_cleanup(struct ttm_buffer_object *bo,
613 void *sync_obj,
614 void *sync_obj_arg,
615 bool evict, bool no_wait_reserve,
616 bool no_wait_gpu,
617 struct ttm_mem_reg *new_mem)
618{
619 struct ttm_bo_device *bdev = bo->bdev;
620 struct ttm_bo_driver *driver = bdev->driver;
621 struct ttm_mem_type_manager *man = &bdev->man[new_mem->mem_type];
622 struct ttm_mem_reg *old_mem = &bo->mem;
623 int ret;
624 struct ttm_buffer_object *ghost_obj;
625 void *tmp_obj = NULL;
626
627 spin_lock(&bdev->fence_lock);
628 if (bo->sync_obj) {
629 tmp_obj = bo->sync_obj;
630 bo->sync_obj = NULL;
631 }
632 bo->sync_obj = driver->sync_obj_ref(sync_obj);
633 bo->sync_obj_arg = sync_obj_arg;
634 if (evict) {
635 ret = ttm_bo_wait(bo, false, false, false);
636 spin_unlock(&bdev->fence_lock);
637 if (tmp_obj)
638 driver->sync_obj_unref(&tmp_obj);
639 if (ret)
640 return ret;
641
642 if ((man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
643 (bo->ttm != NULL)) {
644 ttm_tt_unbind(bo->ttm);
645 ttm_tt_destroy(bo->ttm);
646 bo->ttm = NULL;
647 }
648 ttm_bo_free_old_node(bo);
649 } else {
650 /**
651 * This should help pipeline ordinary buffer moves.
652 *
653 * Hang old buffer memory on a new buffer object,
654 * and leave it to be released when the GPU
655 * operation has completed.
656 */
657
658 set_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
659 spin_unlock(&bdev->fence_lock);
660 if (tmp_obj)
661 driver->sync_obj_unref(&tmp_obj);
662
663 ret = ttm_buffer_object_transfer(bo, &ghost_obj);
664 if (ret)
665 return ret;
666
667 /**
668 * If we're not moving to fixed memory, the TTM object
669 * needs to stay alive. Otherwhise hang it on the ghost
670 * bo to be unbound and destroyed.
671 */
672
673 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED))
674 ghost_obj->ttm = NULL;
675 else
676 bo->ttm = NULL;
677
678 ttm_bo_unreserve(ghost_obj);
679 ttm_bo_unref(&ghost_obj);
680 }
681
682 *old_mem = *new_mem;
683 new_mem->mm_node = NULL;
684
685 return 0;
686}
687EXPORT_SYMBOL(ttm_bo_move_accel_cleanup);