1 /**************************************************************************
2 *
3 * Copyright (c) 2006-2009 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 #define pr_fmt(fmt) "[TTM] " fmt
32
33 #include <drm/ttm/ttm_module.h>
34 #include <drm/ttm/ttm_bo_driver.h>
35 #include <drm/ttm/ttm_placement.h>
36 #include <linux/jiffies.h>
37 #include <linux/slab.h>
38 #include <linux/sched.h>
39 #include <linux/mm.h>
40 #include <linux/file.h>
41 #include <linux/module.h>
42 #include <linux/atomic.h>
43 #include <linux/reservation.h>
44
45 #define TTM_ASSERT_LOCKED(param)
46 #define TTM_DEBUG(fmt, arg...)
47 #define TTM_BO_HASH_ORDER 13
48
49 static void ttm_bo_global_kobj_release(struct kobject *kobj);
50
51 static struct attribute ttm_bo_count = {
52 .name = "bo_count",
53 .mode = S_IRUGO
54 };
55
56 /* default destructor */
ttm_bo_default_destroy(struct ttm_buffer_object * bo)57 static void ttm_bo_default_destroy(struct ttm_buffer_object *bo)
58 {
59 kfree(bo);
60 }
61
ttm_mem_type_from_place(const struct ttm_place * place,uint32_t * mem_type)62 static inline int ttm_mem_type_from_place(const struct ttm_place *place,
63 uint32_t *mem_type)
64 {
65 int pos;
66
67 pos = ffs(place->flags & TTM_PL_MASK_MEM);
68 if (unlikely(!pos))
69 return -EINVAL;
70
71 *mem_type = pos - 1;
72 return 0;
73 }
74
ttm_mem_type_debug(struct ttm_bo_device * bdev,int mem_type)75 static void ttm_mem_type_debug(struct ttm_bo_device *bdev, int mem_type)
76 {
77 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
78 struct drm_printer p = drm_debug_printer(TTM_PFX);
79
80 pr_err(" has_type: %d\n", man->has_type);
81 pr_err(" use_type: %d\n", man->use_type);
82 pr_err(" flags: 0x%08X\n", man->flags);
83 pr_err(" gpu_offset: 0x%08lX\n", man->gpu_offset);
84 pr_err(" size: %ju\n", man->size);
85 pr_err(" available_caching: 0x%08X\n", man->available_caching);
86 pr_err(" default_caching: 0x%08X\n", man->default_caching);
87 if (mem_type != TTM_PL_SYSTEM)
88 (*man->func->debug)(man, &p);
89 }
90
ttm_bo_mem_space_debug(struct ttm_buffer_object * bo,struct ttm_placement * placement)91 static void ttm_bo_mem_space_debug(struct ttm_buffer_object *bo,
92 struct ttm_placement *placement)
93 {
94 int i, ret, mem_type;
95
96 pr_err("No space for %p (%lu pages, %luK, %luM)\n",
97 bo, bo->mem.num_pages, bo->mem.size >> 10,
98 bo->mem.size >> 20);
99 for (i = 0; i < placement->num_placement; i++) {
100 ret = ttm_mem_type_from_place(&placement->placement[i],
101 &mem_type);
102 if (ret)
103 return;
104 pr_err(" placement[%d]=0x%08X (%d)\n",
105 i, placement->placement[i].flags, mem_type);
106 ttm_mem_type_debug(bo->bdev, mem_type);
107 }
108 }
109
ttm_bo_global_show(struct kobject * kobj,struct attribute * attr,char * buffer)110 static ssize_t ttm_bo_global_show(struct kobject *kobj,
111 struct attribute *attr,
112 char *buffer)
113 {
114 struct ttm_bo_global *glob =
115 container_of(kobj, struct ttm_bo_global, kobj);
116
117 return snprintf(buffer, PAGE_SIZE, "%d\n",
118 atomic_read(&glob->bo_count));
119 }
120
121 static struct attribute *ttm_bo_global_attrs[] = {
122 &ttm_bo_count,
123 NULL
124 };
125
126 static const struct sysfs_ops ttm_bo_global_ops = {
127 .show = &ttm_bo_global_show
128 };
129
130 static struct kobj_type ttm_bo_glob_kobj_type = {
131 .release = &ttm_bo_global_kobj_release,
132 .sysfs_ops = &ttm_bo_global_ops,
133 .default_attrs = ttm_bo_global_attrs
134 };
135
136
ttm_bo_type_flags(unsigned type)137 static inline uint32_t ttm_bo_type_flags(unsigned type)
138 {
139 return 1 << (type);
140 }
141
ttm_bo_release_list(struct kref * list_kref)142 static void ttm_bo_release_list(struct kref *list_kref)
143 {
144 struct ttm_buffer_object *bo =
145 container_of(list_kref, struct ttm_buffer_object, list_kref);
146 struct ttm_bo_device *bdev = bo->bdev;
147 size_t acc_size = bo->acc_size;
148
149 BUG_ON(kref_read(&bo->list_kref));
150 BUG_ON(kref_read(&bo->kref));
151 BUG_ON(atomic_read(&bo->cpu_writers));
152 BUG_ON(bo->mem.mm_node != NULL);
153 BUG_ON(!list_empty(&bo->lru));
154 BUG_ON(!list_empty(&bo->ddestroy));
155 ttm_tt_destroy(bo->ttm);
156 atomic_dec(&bo->bdev->glob->bo_count);
157 dma_fence_put(bo->moving);
158 reservation_object_fini(&bo->ttm_resv);
159 mutex_destroy(&bo->wu_mutex);
160 bo->destroy(bo);
161 ttm_mem_global_free(bdev->glob->mem_glob, acc_size);
162 }
163
ttm_bo_add_to_lru(struct ttm_buffer_object * bo)164 void ttm_bo_add_to_lru(struct ttm_buffer_object *bo)
165 {
166 struct ttm_bo_device *bdev = bo->bdev;
167 struct ttm_mem_type_manager *man;
168
169 lockdep_assert_held(&bo->resv->lock.base);
170
171 if (!(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
172
173 #ifdef __DragonFly__
174 if (WARN_ON(!list_empty(&bo->lru)))
175 return;
176 #endif
177
178 man = &bdev->man[bo->mem.mem_type];
179 list_add_tail(&bo->lru, &man->lru[bo->priority]);
180 kref_get(&bo->list_kref);
181
182 if (bo->ttm && !(bo->ttm->page_flags &
183 (TTM_PAGE_FLAG_SG | TTM_PAGE_FLAG_SWAPPED))) {
184 list_add_tail(&bo->swap,
185 &bdev->glob->swap_lru[bo->priority]);
186 kref_get(&bo->list_kref);
187 }
188 }
189 }
190 EXPORT_SYMBOL(ttm_bo_add_to_lru);
191
ttm_bo_ref_bug(struct kref * list_kref)192 static void ttm_bo_ref_bug(struct kref *list_kref)
193 {
194 BUG();
195 }
196
ttm_bo_del_from_lru(struct ttm_buffer_object * bo)197 void ttm_bo_del_from_lru(struct ttm_buffer_object *bo)
198 {
199 if (!list_empty(&bo->swap)) {
200 list_del_init(&bo->swap);
201 kref_put(&bo->list_kref, ttm_bo_ref_bug);
202 }
203 if (!list_empty(&bo->lru)) {
204 list_del_init(&bo->lru);
205 kref_put(&bo->list_kref, ttm_bo_ref_bug);
206 }
207
208 /*
209 * TODO: Add a driver hook to delete from
210 * driver-specific LRU's here.
211 */
212 }
213
ttm_bo_del_sub_from_lru(struct ttm_buffer_object * bo)214 void ttm_bo_del_sub_from_lru(struct ttm_buffer_object *bo)
215 {
216 struct ttm_bo_global *glob = bo->bdev->glob;
217
218 lockmgr(&glob->lru_lock, LK_EXCLUSIVE);
219 ttm_bo_del_from_lru(bo);
220 lockmgr(&glob->lru_lock, LK_RELEASE);
221 }
222 EXPORT_SYMBOL(ttm_bo_del_sub_from_lru);
223
ttm_bo_move_to_lru_tail(struct ttm_buffer_object * bo)224 void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo)
225 {
226 lockdep_assert_held(&bo->resv->lock.base);
227
228 ttm_bo_del_from_lru(bo);
229 ttm_bo_add_to_lru(bo);
230 }
231 EXPORT_SYMBOL(ttm_bo_move_to_lru_tail);
232
ttm_bo_handle_move_mem(struct ttm_buffer_object * bo,struct ttm_mem_reg * mem,bool evict,struct ttm_operation_ctx * ctx)233 static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
234 struct ttm_mem_reg *mem, bool evict,
235 struct ttm_operation_ctx *ctx)
236 {
237 struct ttm_bo_device *bdev = bo->bdev;
238 bool old_is_pci = ttm_mem_reg_is_pci(bdev, &bo->mem);
239 bool new_is_pci = ttm_mem_reg_is_pci(bdev, mem);
240 struct ttm_mem_type_manager *old_man = &bdev->man[bo->mem.mem_type];
241 struct ttm_mem_type_manager *new_man = &bdev->man[mem->mem_type];
242 int ret = 0;
243
244 if (old_is_pci || new_is_pci ||
245 ((mem->placement & bo->mem.placement & TTM_PL_MASK_CACHING) == 0)) {
246 ret = ttm_mem_io_lock(old_man, true);
247 if (unlikely(ret != 0))
248 goto out_err;
249 ttm_bo_unmap_virtual_locked(bo);
250 ttm_mem_io_unlock(old_man);
251 }
252
253 /*
254 * Create and bind a ttm if required.
255 */
256
257 if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
258 if (bo->ttm == NULL) {
259 bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED);
260 ret = ttm_tt_create(bo, zero);
261 if (ret)
262 goto out_err;
263 }
264
265 ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement);
266 if (ret)
267 goto out_err;
268
269 if (mem->mem_type != TTM_PL_SYSTEM) {
270 ret = ttm_tt_bind(bo->ttm, mem, ctx);
271 if (ret)
272 goto out_err;
273 }
274
275 if (bo->mem.mem_type == TTM_PL_SYSTEM) {
276 if (bdev->driver->move_notify)
277 bdev->driver->move_notify(bo, evict, mem);
278 bo->mem = *mem;
279 mem->mm_node = NULL;
280 goto moved;
281 }
282 }
283
284 if (bdev->driver->move_notify)
285 bdev->driver->move_notify(bo, evict, mem);
286
287 if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
288 !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED))
289 ret = ttm_bo_move_ttm(bo, ctx, mem);
290 else if (bdev->driver->move)
291 ret = bdev->driver->move(bo, evict, ctx, mem);
292 else
293 ret = ttm_bo_move_memcpy(bo, ctx, mem);
294
295 if (ret) {
296 if (bdev->driver->move_notify) {
297 swap(*mem, bo->mem);
298 bdev->driver->move_notify(bo, false, mem);
299 swap(*mem, bo->mem);
300 }
301
302 goto out_err;
303 }
304
305 moved:
306 if (bo->evicted) {
307 if (bdev->driver->invalidate_caches) {
308 ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement);
309 if (ret)
310 pr_err("Can not flush read caches\n");
311 }
312 bo->evicted = false;
313 }
314
315 if (bo->mem.mm_node)
316 bo->offset = (bo->mem.start << PAGE_SHIFT) +
317 bdev->man[bo->mem.mem_type].gpu_offset;
318 else
319 bo->offset = 0;
320
321 ctx->bytes_moved += bo->num_pages << PAGE_SHIFT;
322 return 0;
323
324 out_err:
325 new_man = &bdev->man[bo->mem.mem_type];
326 if (new_man->flags & TTM_MEMTYPE_FLAG_FIXED) {
327 ttm_tt_destroy(bo->ttm);
328 bo->ttm = NULL;
329 }
330
331 return ret;
332 }
333
334 /**
335 * Call bo::reserved.
336 * Will release GPU memory type usage on destruction.
337 * This is the place to put in driver specific hooks to release
338 * driver private resources.
339 * Will release the bo::reserved lock.
340 */
341
ttm_bo_cleanup_memtype_use(struct ttm_buffer_object * bo)342 static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
343 {
344 if (bo->bdev->driver->move_notify)
345 bo->bdev->driver->move_notify(bo, false, NULL);
346
347 ttm_tt_destroy(bo->ttm);
348 bo->ttm = NULL;
349 ttm_bo_mem_put(bo, &bo->mem);
350 }
351
ttm_bo_individualize_resv(struct ttm_buffer_object * bo)352 static int ttm_bo_individualize_resv(struct ttm_buffer_object *bo)
353 {
354 int r;
355
356 if (bo->resv == &bo->ttm_resv)
357 return 0;
358
359 BUG_ON(!reservation_object_trylock(&bo->ttm_resv));
360
361 r = reservation_object_copy_fences(&bo->ttm_resv, bo->resv);
362 if (r)
363 reservation_object_unlock(&bo->ttm_resv);
364
365 return r;
366 }
367
ttm_bo_flush_all_fences(struct ttm_buffer_object * bo)368 static void ttm_bo_flush_all_fences(struct ttm_buffer_object *bo)
369 {
370 struct reservation_object_list *fobj;
371 struct dma_fence *fence;
372 int i;
373
374 fobj = reservation_object_get_list(&bo->ttm_resv);
375 fence = reservation_object_get_excl(&bo->ttm_resv);
376 if (fence && !fence->ops->signaled)
377 dma_fence_enable_sw_signaling(fence);
378
379 for (i = 0; fobj && i < fobj->shared_count; ++i) {
380 fence = rcu_dereference_protected(fobj->shared[i],
381 reservation_object_held(bo->resv));
382
383 if (!fence->ops->signaled)
384 dma_fence_enable_sw_signaling(fence);
385 }
386 }
387
ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object * bo)388 static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo)
389 {
390 struct ttm_bo_device *bdev = bo->bdev;
391 struct ttm_bo_global *glob = bdev->glob;
392 int ret;
393
394 ret = ttm_bo_individualize_resv(bo);
395 if (ret) {
396 /* Last resort, if we fail to allocate memory for the
397 * fences block for the BO to become idle
398 */
399 reservation_object_wait_timeout_rcu(bo->resv, true, false,
400 30 * HZ);
401 lockmgr(&glob->lru_lock, LK_EXCLUSIVE);
402 goto error;
403 }
404
405 lockmgr(&glob->lru_lock, LK_EXCLUSIVE);
406 ret = __ttm_bo_reserve(bo, false, true, NULL);
407 if (!ret) {
408 if (reservation_object_test_signaled_rcu(&bo->ttm_resv, true)) {
409 ttm_bo_del_from_lru(bo);
410 lockmgr(&glob->lru_lock, LK_RELEASE);
411 if (bo->resv != &bo->ttm_resv)
412 reservation_object_unlock(&bo->ttm_resv);
413
414 ttm_bo_cleanup_memtype_use(bo);
415 __ttm_bo_unreserve(bo);
416 return;
417 }
418
419 ttm_bo_flush_all_fences(bo);
420
421 /*
422 * Make NO_EVICT bos immediately available to
423 * shrinkers, now that they are queued for
424 * destruction.
425 */
426 if (bo->mem.placement & TTM_PL_FLAG_NO_EVICT) {
427 bo->mem.placement &= ~TTM_PL_FLAG_NO_EVICT;
428 ttm_bo_add_to_lru(bo);
429 }
430
431 __ttm_bo_unreserve(bo);
432 }
433 if (bo->resv != &bo->ttm_resv)
434 reservation_object_unlock(&bo->ttm_resv);
435
436 error:
437 kref_get(&bo->list_kref);
438 list_add_tail(&bo->ddestroy, &bdev->ddestroy);
439 lockmgr(&glob->lru_lock, LK_RELEASE);
440
441 schedule_delayed_work(&bdev->wq,
442 ((HZ / 100) < 1) ? 1 : HZ / 100);
443 }
444
445 /**
446 * function ttm_bo_cleanup_refs
447 * If bo idle, remove from delayed- and lru lists, and unref.
448 * If not idle, do nothing.
449 *
450 * Must be called with lru_lock and reservation held, this function
451 * will drop the lru lock and optionally the reservation lock before returning.
452 *
453 * @interruptible Any sleeps should occur interruptibly.
454 * @no_wait_gpu Never wait for gpu. Return -EBUSY instead.
455 * @unlock_resv Unlock the reservation lock as well.
456 */
457
ttm_bo_cleanup_refs(struct ttm_buffer_object * bo,bool interruptible,bool no_wait_gpu,bool unlock_resv)458 static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo,
459 bool interruptible, bool no_wait_gpu,
460 bool unlock_resv)
461 {
462 struct ttm_bo_global *glob = bo->bdev->glob;
463 struct reservation_object *resv;
464 int ret;
465
466 if (unlikely(list_empty(&bo->ddestroy)))
467 resv = bo->resv;
468 else
469 resv = &bo->ttm_resv;
470
471 if (reservation_object_test_signaled_rcu(resv, true))
472 ret = 0;
473 else
474 ret = -EBUSY;
475
476 if (ret && !no_wait_gpu) {
477 long lret;
478
479 if (unlock_resv)
480 ww_mutex_unlock(&bo->resv->lock);
481 lockmgr(&glob->lru_lock, LK_RELEASE);
482
483 lret = reservation_object_wait_timeout_rcu(resv, true,
484 interruptible,
485 30 * HZ);
486
487 if (lret < 0)
488 return lret;
489 else if (lret == 0)
490 return -EBUSY;
491
492 lockmgr(&glob->lru_lock, LK_EXCLUSIVE);
493 if (unlock_resv && __ttm_bo_reserve(bo, false, true, NULL)) {
494 /*
495 * We raced, and lost, someone else holds the reservation now,
496 * and is probably busy in ttm_bo_cleanup_memtype_use.
497 *
498 * Even if it's not the case, because we finished waiting any
499 * delayed destruction would succeed, so just return success
500 * here.
501 */
502 lockmgr(&glob->lru_lock, LK_RELEASE);
503 return 0;
504 }
505 ret = 0;
506 }
507
508 if (ret || unlikely(list_empty(&bo->ddestroy))) {
509 if (unlock_resv)
510 __ttm_bo_unreserve(bo);
511 lockmgr(&glob->lru_lock, LK_RELEASE);
512 return ret;
513 }
514
515 ttm_bo_del_from_lru(bo);
516 list_del_init(&bo->ddestroy);
517 kref_put(&bo->list_kref, ttm_bo_ref_bug);
518
519 lockmgr(&glob->lru_lock, LK_RELEASE);
520 ttm_bo_cleanup_memtype_use(bo);
521
522 if (unlock_resv)
523 __ttm_bo_unreserve(bo);
524
525 return 0;
526 }
527
528 /**
529 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
530 * encountered buffers.
531 */
ttm_bo_delayed_delete(struct ttm_bo_device * bdev,bool remove_all)532 static bool ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all)
533 {
534 struct ttm_bo_global *glob = bdev->glob;
535 struct list_head removed;
536 bool empty;
537
538 INIT_LIST_HEAD(&removed);
539
540 lockmgr(&glob->lru_lock, LK_EXCLUSIVE);
541 while (!list_empty(&bdev->ddestroy)) {
542 struct ttm_buffer_object *bo;
543
544 bo = list_first_entry(&bdev->ddestroy, struct ttm_buffer_object,
545 ddestroy);
546 kref_get(&bo->list_kref);
547 list_move_tail(&bo->ddestroy, &removed);
548
549 if (remove_all || bo->resv != &bo->ttm_resv) {
550 lockmgr(&glob->lru_lock, LK_RELEASE);
551 reservation_object_lock(bo->resv, NULL);
552
553 lockmgr(&glob->lru_lock, LK_EXCLUSIVE);
554 ttm_bo_cleanup_refs(bo, false, !remove_all, true);
555
556 } else if (reservation_object_trylock(bo->resv)) {
557 ttm_bo_cleanup_refs(bo, false, !remove_all, true);
558 } else {
559 lockmgr(&glob->lru_lock, LK_RELEASE);
560 }
561
562 kref_put(&bo->list_kref, ttm_bo_release_list);
563 lockmgr(&glob->lru_lock, LK_EXCLUSIVE);
564 }
565 list_splice_tail(&removed, &bdev->ddestroy);
566 empty = list_empty(&bdev->ddestroy);
567 lockmgr(&glob->lru_lock, LK_RELEASE);
568
569 return empty;
570 }
571
ttm_bo_delayed_workqueue(struct work_struct * work)572 static void ttm_bo_delayed_workqueue(struct work_struct *work)
573 {
574 struct ttm_bo_device *bdev =
575 container_of(work, struct ttm_bo_device, wq.work);
576
577 if (!ttm_bo_delayed_delete(bdev, false))
578 schedule_delayed_work(&bdev->wq,
579 ((HZ / 100) < 1) ? 1 : HZ / 100);
580 }
581
ttm_bo_release(struct kref * kref)582 static void ttm_bo_release(struct kref *kref)
583 {
584 struct ttm_buffer_object *bo =
585 container_of(kref, struct ttm_buffer_object, kref);
586 struct ttm_bo_device *bdev = bo->bdev;
587 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
588
589 drm_vma_offset_remove(&bdev->vma_manager, &bo->vma_node);
590 ttm_mem_io_lock(man, false);
591 ttm_mem_io_free_vm(bo);
592 ttm_mem_io_unlock(man);
593 ttm_bo_cleanup_refs_or_queue(bo);
594 kref_put(&bo->list_kref, ttm_bo_release_list);
595 }
596
ttm_bo_put(struct ttm_buffer_object * bo)597 void ttm_bo_put(struct ttm_buffer_object *bo)
598 {
599 kref_put(&bo->kref, ttm_bo_release);
600 }
601 EXPORT_SYMBOL(ttm_bo_put);
602
ttm_bo_unref(struct ttm_buffer_object ** p_bo)603 void ttm_bo_unref(struct ttm_buffer_object **p_bo)
604 {
605 struct ttm_buffer_object *bo = *p_bo;
606
607 *p_bo = NULL;
608 ttm_bo_put(bo);
609 }
610 EXPORT_SYMBOL(ttm_bo_unref);
611
ttm_bo_lock_delayed_workqueue(struct ttm_bo_device * bdev)612 int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev)
613 {
614 return cancel_delayed_work_sync(&bdev->wq);
615 }
616 EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue);
617
ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device * bdev,int resched)618 void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched)
619 {
620 if (resched)
621 schedule_delayed_work(&bdev->wq,
622 ((HZ / 100) < 1) ? 1 : HZ / 100);
623 }
624 EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue);
625
ttm_bo_evict(struct ttm_buffer_object * bo,struct ttm_operation_ctx * ctx)626 static int ttm_bo_evict(struct ttm_buffer_object *bo,
627 struct ttm_operation_ctx *ctx)
628 {
629 struct ttm_bo_device *bdev = bo->bdev;
630 struct ttm_mem_reg evict_mem;
631 struct ttm_placement placement;
632 int ret = 0;
633
634 lockdep_assert_held(&bo->resv->lock.base);
635
636 placement.num_placement = 0;
637 placement.num_busy_placement = 0;
638 bdev->driver->evict_flags(bo, &placement);
639
640 if (!placement.num_placement && !placement.num_busy_placement) {
641 ret = ttm_bo_pipeline_gutting(bo);
642 if (ret)
643 return ret;
644
645 return ttm_tt_create(bo, false);
646 }
647
648 evict_mem = bo->mem;
649 evict_mem.mm_node = NULL;
650 evict_mem.bus.io_reserved_vm = false;
651 evict_mem.bus.io_reserved_count = 0;
652
653 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, ctx);
654 if (ret) {
655 if (ret != -ERESTARTSYS) {
656 pr_err("Failed to find memory space for buffer 0x%p eviction\n",
657 bo);
658 ttm_bo_mem_space_debug(bo, &placement);
659 }
660 goto out;
661 }
662
663 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, ctx);
664 if (unlikely(ret)) {
665 if (ret != -ERESTARTSYS)
666 pr_err("Buffer eviction failed\n");
667 ttm_bo_mem_put(bo, &evict_mem);
668 goto out;
669 }
670 bo->evicted = true;
671 out:
672 return ret;
673 }
674
ttm_bo_eviction_valuable(struct ttm_buffer_object * bo,const struct ttm_place * place)675 bool ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
676 const struct ttm_place *place)
677 {
678 /* Don't evict this BO if it's outside of the
679 * requested placement range
680 */
681 if (place->fpfn >= (bo->mem.start + bo->mem.num_pages) ||
682 (place->lpfn && place->lpfn <= bo->mem.start))
683 return false;
684
685 return true;
686 }
687 EXPORT_SYMBOL(ttm_bo_eviction_valuable);
688
689 /**
690 * Check the target bo is allowable to be evicted or swapout, including cases:
691 *
692 * a. if share same reservation object with ctx->resv, have assumption
693 * reservation objects should already be locked, so not lock again and
694 * return true directly when either the opreation allow_reserved_eviction
695 * or the target bo already is in delayed free list;
696 *
697 * b. Otherwise, trylock it.
698 */
ttm_bo_evict_swapout_allowable(struct ttm_buffer_object * bo,struct ttm_operation_ctx * ctx,bool * locked)699 static bool ttm_bo_evict_swapout_allowable(struct ttm_buffer_object *bo,
700 struct ttm_operation_ctx *ctx, bool *locked)
701 {
702 bool ret = false;
703
704 *locked = false;
705 if (bo->resv == ctx->resv) {
706 lockdep_assert_held(&bo->resv->lock.base);
707 if (ctx->flags & TTM_OPT_FLAG_ALLOW_RES_EVICT
708 || !list_empty(&bo->ddestroy))
709 ret = true;
710 } else {
711 *locked = ww_mutex_trylock(&bo->resv->lock);
712 ret = *locked;
713 }
714
715 return ret;
716 }
717
ttm_mem_evict_first(struct ttm_bo_device * bdev,uint32_t mem_type,const struct ttm_place * place,struct ttm_operation_ctx * ctx)718 static int ttm_mem_evict_first(struct ttm_bo_device *bdev,
719 uint32_t mem_type,
720 const struct ttm_place *place,
721 struct ttm_operation_ctx *ctx)
722 {
723 struct ttm_bo_global *glob = bdev->glob;
724 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
725 struct ttm_buffer_object *bo = NULL;
726 bool locked = false;
727 unsigned i;
728 int ret = -EBUSY;
729
730 lockmgr(&glob->lru_lock, LK_EXCLUSIVE);
731 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
732 list_for_each_entry(bo, &man->lru[i], lru) {
733 if (!ttm_bo_evict_swapout_allowable(bo, ctx, &locked))
734 continue;
735
736 if (place && !bdev->driver->eviction_valuable(bo,
737 place)) {
738 if (locked)
739 __ttm_bo_unreserve(bo);
740 continue;
741 }
742 break;
743 }
744
745 /* If the inner loop terminated early, we have our candidate */
746 if (&bo->lru != &man->lru[i])
747 break;
748
749 bo = NULL;
750 }
751
752 if (!bo) {
753 lockmgr(&glob->lru_lock, LK_RELEASE);
754 return -EBUSY;
755 }
756
757 kref_get(&bo->list_kref);
758
759 if (!list_empty(&bo->ddestroy)) {
760 ret = ttm_bo_cleanup_refs(bo, ctx->interruptible,
761 ctx->no_wait_gpu, locked);
762 kref_put(&bo->list_kref, ttm_bo_release_list);
763 return ret;
764 }
765
766 ttm_bo_del_from_lru(bo);
767 lockmgr(&glob->lru_lock, LK_RELEASE);
768
769 ret = ttm_bo_evict(bo, ctx);
770 if (locked) {
771 ttm_bo_unreserve(bo);
772 } else {
773 lockmgr(&glob->lru_lock, LK_EXCLUSIVE);
774 ttm_bo_add_to_lru(bo);
775 lockmgr(&glob->lru_lock, LK_RELEASE);
776 }
777
778 kref_put(&bo->list_kref, ttm_bo_release_list);
779 return ret;
780 }
781
ttm_bo_mem_put(struct ttm_buffer_object * bo,struct ttm_mem_reg * mem)782 void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem)
783 {
784 struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type];
785
786 if (mem->mm_node)
787 (*man->func->put_node)(man, mem);
788 }
789 EXPORT_SYMBOL(ttm_bo_mem_put);
790
791 /**
792 * Add the last move fence to the BO and reserve a new shared slot.
793 */
ttm_bo_add_move_fence(struct ttm_buffer_object * bo,struct ttm_mem_type_manager * man,struct ttm_mem_reg * mem)794 static int ttm_bo_add_move_fence(struct ttm_buffer_object *bo,
795 struct ttm_mem_type_manager *man,
796 struct ttm_mem_reg *mem)
797 {
798 struct dma_fence *fence;
799 int ret;
800
801 lockmgr(&man->move_lock, LK_EXCLUSIVE);
802 fence = dma_fence_get(man->move);
803 lockmgr(&man->move_lock, LK_RELEASE);
804
805 if (fence) {
806 reservation_object_add_shared_fence(bo->resv, fence);
807
808 ret = reservation_object_reserve_shared(bo->resv);
809 if (unlikely(ret))
810 return ret;
811
812 dma_fence_put(bo->moving);
813 bo->moving = fence;
814 }
815
816 return 0;
817 }
818
819 /**
820 * Repeatedly evict memory from the LRU for @mem_type until we create enough
821 * space, or we've evicted everything and there isn't enough space.
822 */
ttm_bo_mem_force_space(struct ttm_buffer_object * bo,uint32_t mem_type,const struct ttm_place * place,struct ttm_mem_reg * mem,struct ttm_operation_ctx * ctx)823 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
824 uint32_t mem_type,
825 const struct ttm_place *place,
826 struct ttm_mem_reg *mem,
827 struct ttm_operation_ctx *ctx)
828 {
829 struct ttm_bo_device *bdev = bo->bdev;
830 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
831 int ret;
832
833 do {
834 ret = (*man->func->get_node)(man, bo, place, mem);
835 if (unlikely(ret != 0))
836 return ret;
837 if (mem->mm_node)
838 break;
839 ret = ttm_mem_evict_first(bdev, mem_type, place, ctx);
840 if (unlikely(ret != 0))
841 return ret;
842 } while (1);
843 mem->mem_type = mem_type;
844 return ttm_bo_add_move_fence(bo, man, mem);
845 }
846
ttm_bo_select_caching(struct ttm_mem_type_manager * man,uint32_t cur_placement,uint32_t proposed_placement)847 static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man,
848 uint32_t cur_placement,
849 uint32_t proposed_placement)
850 {
851 uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING;
852 uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING;
853
854 /**
855 * Keep current caching if possible.
856 */
857
858 if ((cur_placement & caching) != 0)
859 result |= (cur_placement & caching);
860 else if ((man->default_caching & caching) != 0)
861 result |= man->default_caching;
862 else if ((TTM_PL_FLAG_CACHED & caching) != 0)
863 result |= TTM_PL_FLAG_CACHED;
864 else if ((TTM_PL_FLAG_WC & caching) != 0)
865 result |= TTM_PL_FLAG_WC;
866 else if ((TTM_PL_FLAG_UNCACHED & caching) != 0)
867 result |= TTM_PL_FLAG_UNCACHED;
868
869 return result;
870 }
871
ttm_bo_mt_compatible(struct ttm_mem_type_manager * man,uint32_t mem_type,const struct ttm_place * place,uint32_t * masked_placement)872 static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man,
873 uint32_t mem_type,
874 const struct ttm_place *place,
875 uint32_t *masked_placement)
876 {
877 uint32_t cur_flags = ttm_bo_type_flags(mem_type);
878
879 if ((cur_flags & place->flags & TTM_PL_MASK_MEM) == 0)
880 return false;
881
882 if ((place->flags & man->available_caching) == 0)
883 return false;
884
885 cur_flags |= (place->flags & man->available_caching);
886
887 *masked_placement = cur_flags;
888 return true;
889 }
890
891 /**
892 * Creates space for memory region @mem according to its type.
893 *
894 * This function first searches for free space in compatible memory types in
895 * the priority order defined by the driver. If free space isn't found, then
896 * ttm_bo_mem_force_space is attempted in priority order to evict and find
897 * space.
898 */
ttm_bo_mem_space(struct ttm_buffer_object * bo,struct ttm_placement * placement,struct ttm_mem_reg * mem,struct ttm_operation_ctx * ctx)899 int ttm_bo_mem_space(struct ttm_buffer_object *bo,
900 struct ttm_placement *placement,
901 struct ttm_mem_reg *mem,
902 struct ttm_operation_ctx *ctx)
903 {
904 struct ttm_bo_device *bdev = bo->bdev;
905 struct ttm_mem_type_manager *man;
906 uint32_t mem_type = TTM_PL_SYSTEM;
907 uint32_t cur_flags = 0;
908 bool type_found = false;
909 bool type_ok = false;
910 bool has_erestartsys = false;
911 int i, ret;
912
913 ret = reservation_object_reserve_shared(bo->resv);
914 if (unlikely(ret))
915 return ret;
916
917 mem->mm_node = NULL;
918 for (i = 0; i < placement->num_placement; ++i) {
919 const struct ttm_place *place = &placement->placement[i];
920
921 ret = ttm_mem_type_from_place(place, &mem_type);
922 if (ret)
923 return ret;
924 man = &bdev->man[mem_type];
925 if (!man->has_type || !man->use_type)
926 continue;
927
928 type_ok = ttm_bo_mt_compatible(man, mem_type, place,
929 &cur_flags);
930
931 if (!type_ok)
932 continue;
933
934 type_found = true;
935 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
936 cur_flags);
937 /*
938 * Use the access and other non-mapping-related flag bits from
939 * the memory placement flags to the current flags
940 */
941 ttm_flag_masked(&cur_flags, place->flags,
942 ~TTM_PL_MASK_MEMTYPE);
943
944 if (mem_type == TTM_PL_SYSTEM)
945 break;
946
947 ret = (*man->func->get_node)(man, bo, place, mem);
948 if (unlikely(ret))
949 return ret;
950
951 if (mem->mm_node) {
952 ret = ttm_bo_add_move_fence(bo, man, mem);
953 if (unlikely(ret)) {
954 (*man->func->put_node)(man, mem);
955 return ret;
956 }
957 break;
958 }
959 }
960
961 if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) {
962 mem->mem_type = mem_type;
963 mem->placement = cur_flags;
964 return 0;
965 }
966
967 for (i = 0; i < placement->num_busy_placement; ++i) {
968 const struct ttm_place *place = &placement->busy_placement[i];
969
970 ret = ttm_mem_type_from_place(place, &mem_type);
971 if (ret)
972 return ret;
973 man = &bdev->man[mem_type];
974 if (!man->has_type || !man->use_type)
975 continue;
976 if (!ttm_bo_mt_compatible(man, mem_type, place, &cur_flags))
977 continue;
978
979 type_found = true;
980 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
981 cur_flags);
982 /*
983 * Use the access and other non-mapping-related flag bits from
984 * the memory placement flags to the current flags
985 */
986 ttm_flag_masked(&cur_flags, place->flags,
987 ~TTM_PL_MASK_MEMTYPE);
988
989 if (mem_type == TTM_PL_SYSTEM) {
990 mem->mem_type = mem_type;
991 mem->placement = cur_flags;
992 mem->mm_node = NULL;
993 return 0;
994 }
995
996 ret = ttm_bo_mem_force_space(bo, mem_type, place, mem, ctx);
997 if (ret == 0 && mem->mm_node) {
998 mem->placement = cur_flags;
999 return 0;
1000 }
1001 if (ret == -ERESTARTSYS)
1002 has_erestartsys = true;
1003 }
1004
1005 if (!type_found) {
1006 pr_err(TTM_PFX "No compatible memory type found\n");
1007 return -EINVAL;
1008 }
1009
1010 return (has_erestartsys) ? -ERESTARTSYS : -ENOMEM;
1011 }
1012 EXPORT_SYMBOL(ttm_bo_mem_space);
1013
ttm_bo_move_buffer(struct ttm_buffer_object * bo,struct ttm_placement * placement,struct ttm_operation_ctx * ctx)1014 static int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
1015 struct ttm_placement *placement,
1016 struct ttm_operation_ctx *ctx)
1017 {
1018 int ret = 0;
1019 struct ttm_mem_reg mem;
1020
1021 lockdep_assert_held(&bo->resv->lock.base);
1022
1023 mem.num_pages = bo->num_pages;
1024 mem.size = mem.num_pages << PAGE_SHIFT;
1025 mem.page_alignment = bo->mem.page_alignment;
1026 mem.bus.io_reserved_vm = false;
1027 mem.bus.io_reserved_count = 0;
1028 /*
1029 * Determine where to move the buffer.
1030 */
1031 ret = ttm_bo_mem_space(bo, placement, &mem, ctx);
1032 if (ret)
1033 goto out_unlock;
1034 ret = ttm_bo_handle_move_mem(bo, &mem, false, ctx);
1035 out_unlock:
1036 if (ret && mem.mm_node)
1037 ttm_bo_mem_put(bo, &mem);
1038 return ret;
1039 }
1040
ttm_bo_places_compat(const struct ttm_place * places,unsigned num_placement,struct ttm_mem_reg * mem,uint32_t * new_flags)1041 static bool ttm_bo_places_compat(const struct ttm_place *places,
1042 unsigned num_placement,
1043 struct ttm_mem_reg *mem,
1044 uint32_t *new_flags)
1045 {
1046 unsigned i;
1047
1048 for (i = 0; i < num_placement; i++) {
1049 const struct ttm_place *heap = &places[i];
1050
1051 if (mem->mm_node && (mem->start < heap->fpfn ||
1052 (heap->lpfn != 0 && (mem->start + mem->num_pages) > heap->lpfn)))
1053 continue;
1054
1055 *new_flags = heap->flags;
1056 if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) &&
1057 (*new_flags & mem->placement & TTM_PL_MASK_MEM) &&
1058 (!(*new_flags & TTM_PL_FLAG_CONTIGUOUS) ||
1059 (mem->placement & TTM_PL_FLAG_CONTIGUOUS)))
1060 return true;
1061 }
1062 return false;
1063 }
1064
ttm_bo_mem_compat(struct ttm_placement * placement,struct ttm_mem_reg * mem,uint32_t * new_flags)1065 bool ttm_bo_mem_compat(struct ttm_placement *placement,
1066 struct ttm_mem_reg *mem,
1067 uint32_t *new_flags)
1068 {
1069 if (ttm_bo_places_compat(placement->placement, placement->num_placement,
1070 mem, new_flags))
1071 return true;
1072
1073 if ((placement->busy_placement != placement->placement ||
1074 placement->num_busy_placement > placement->num_placement) &&
1075 ttm_bo_places_compat(placement->busy_placement,
1076 placement->num_busy_placement,
1077 mem, new_flags))
1078 return true;
1079
1080 return false;
1081 }
1082 EXPORT_SYMBOL(ttm_bo_mem_compat);
1083
ttm_bo_validate(struct ttm_buffer_object * bo,struct ttm_placement * placement,struct ttm_operation_ctx * ctx)1084 int ttm_bo_validate(struct ttm_buffer_object *bo,
1085 struct ttm_placement *placement,
1086 struct ttm_operation_ctx *ctx)
1087 {
1088 int ret;
1089 uint32_t new_flags;
1090
1091 lockdep_assert_held(&bo->resv->lock.base);
1092 /*
1093 * Check whether we need to move buffer.
1094 */
1095 if (!ttm_bo_mem_compat(placement, &bo->mem, &new_flags)) {
1096 ret = ttm_bo_move_buffer(bo, placement, ctx);
1097 if (ret)
1098 return ret;
1099 } else {
1100 /*
1101 * Use the access and other non-mapping-related flag bits from
1102 * the compatible memory placement flags to the active flags
1103 */
1104 ttm_flag_masked(&bo->mem.placement, new_flags,
1105 ~TTM_PL_MASK_MEMTYPE);
1106 }
1107 /*
1108 * We might need to add a TTM.
1109 */
1110 if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
1111 ret = ttm_tt_create(bo, true);
1112 if (ret)
1113 return ret;
1114 }
1115 return 0;
1116 }
1117 EXPORT_SYMBOL(ttm_bo_validate);
1118
ttm_bo_init_reserved(struct ttm_bo_device * bdev,struct ttm_buffer_object * bo,unsigned long size,enum ttm_bo_type type,struct ttm_placement * placement,uint32_t page_alignment,struct ttm_operation_ctx * ctx,size_t acc_size,struct sg_table * sg,struct reservation_object * resv,void (* destroy)(struct ttm_buffer_object *))1119 int ttm_bo_init_reserved(struct ttm_bo_device *bdev,
1120 struct ttm_buffer_object *bo,
1121 unsigned long size,
1122 enum ttm_bo_type type,
1123 struct ttm_placement *placement,
1124 uint32_t page_alignment,
1125 struct ttm_operation_ctx *ctx,
1126 size_t acc_size,
1127 struct sg_table *sg,
1128 struct reservation_object *resv,
1129 void (*destroy) (struct ttm_buffer_object *))
1130 {
1131 int ret = 0;
1132 unsigned long num_pages;
1133 struct ttm_mem_global *mem_glob = bdev->glob->mem_glob;
1134 bool locked;
1135
1136 ret = ttm_mem_global_alloc(mem_glob, acc_size, ctx);
1137 if (ret) {
1138 pr_err("Out of kernel memory\n");
1139 if (destroy)
1140 (*destroy)(bo);
1141 else
1142 kfree(bo);
1143 return -ENOMEM;
1144 }
1145
1146 num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1147 if (num_pages == 0) {
1148 pr_err("Illegal buffer object size\n");
1149 if (destroy)
1150 (*destroy)(bo);
1151 else
1152 kfree(bo);
1153 ttm_mem_global_free(mem_glob, acc_size);
1154 return -EINVAL;
1155 }
1156 bo->destroy = destroy ? destroy : ttm_bo_default_destroy;
1157
1158 kref_init(&bo->kref);
1159 kref_init(&bo->list_kref);
1160 atomic_set(&bo->cpu_writers, 0);
1161 INIT_LIST_HEAD(&bo->lru);
1162 INIT_LIST_HEAD(&bo->ddestroy);
1163 INIT_LIST_HEAD(&bo->swap);
1164 INIT_LIST_HEAD(&bo->io_reserve_lru);
1165 lockinit(&bo->wu_mutex, "ttmbwm", 0, LK_CANRECURSE);
1166 bo->bdev = bdev;
1167 bo->type = type;
1168 bo->num_pages = num_pages;
1169 bo->mem.size = num_pages << PAGE_SHIFT;
1170 bo->mem.mem_type = TTM_PL_SYSTEM;
1171 bo->mem.num_pages = bo->num_pages;
1172 bo->mem.mm_node = NULL;
1173 bo->mem.page_alignment = page_alignment;
1174 bo->mem.bus.io_reserved_vm = false;
1175 bo->mem.bus.io_reserved_count = 0;
1176 bo->moving = NULL;
1177 bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED);
1178 bo->acc_size = acc_size;
1179 bo->sg = sg;
1180 if (resv) {
1181 bo->resv = resv;
1182 lockdep_assert_held(&bo->resv->lock.base);
1183 } else {
1184 bo->resv = &bo->ttm_resv;
1185 }
1186 reservation_object_init(&bo->ttm_resv);
1187 atomic_inc(&bo->bdev->glob->bo_count);
1188 drm_vma_node_reset(&bo->vma_node);
1189 bo->priority = 0;
1190
1191 /*
1192 * For ttm_bo_type_device buffers, allocate
1193 * address space from the device.
1194 */
1195 if (bo->type == ttm_bo_type_device ||
1196 bo->type == ttm_bo_type_sg)
1197 ret = drm_vma_offset_add(&bdev->vma_manager, &bo->vma_node,
1198 bo->mem.num_pages);
1199
1200 /* passed reservation objects should already be locked,
1201 * since otherwise lockdep will be angered in radeon.
1202 */
1203 if (!resv) {
1204 locked = ww_mutex_trylock(&bo->resv->lock);
1205 WARN_ON(!locked);
1206 }
1207
1208 if (likely(!ret))
1209 ret = ttm_bo_validate(bo, placement, ctx);
1210
1211 if (unlikely(ret)) {
1212 if (!resv)
1213 ttm_bo_unreserve(bo);
1214
1215 ttm_bo_put(bo);
1216 return ret;
1217 }
1218
1219 if (resv && !(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
1220 lockmgr(&bdev->glob->lru_lock, LK_EXCLUSIVE);
1221 ttm_bo_add_to_lru(bo);
1222 lockmgr(&bdev->glob->lru_lock, LK_RELEASE);
1223 }
1224
1225 return ret;
1226 }
1227 EXPORT_SYMBOL(ttm_bo_init_reserved);
1228
ttm_bo_init(struct ttm_bo_device * bdev,struct ttm_buffer_object * bo,unsigned long size,enum ttm_bo_type type,struct ttm_placement * placement,uint32_t page_alignment,bool interruptible,size_t acc_size,struct sg_table * sg,struct reservation_object * resv,void (* destroy)(struct ttm_buffer_object *))1229 int ttm_bo_init(struct ttm_bo_device *bdev,
1230 struct ttm_buffer_object *bo,
1231 unsigned long size,
1232 enum ttm_bo_type type,
1233 struct ttm_placement *placement,
1234 uint32_t page_alignment,
1235 bool interruptible,
1236 size_t acc_size,
1237 struct sg_table *sg,
1238 struct reservation_object *resv,
1239 void (*destroy) (struct ttm_buffer_object *))
1240 {
1241 struct ttm_operation_ctx ctx = { interruptible, false };
1242 int ret;
1243
1244 ret = ttm_bo_init_reserved(bdev, bo, size, type, placement,
1245 page_alignment, &ctx, acc_size,
1246 sg, resv, destroy);
1247 if (ret)
1248 return ret;
1249
1250 if (!resv)
1251 ttm_bo_unreserve(bo);
1252
1253 return 0;
1254 }
1255 EXPORT_SYMBOL(ttm_bo_init);
1256
ttm_bo_acc_size(struct ttm_bo_device * bdev,unsigned long bo_size,unsigned struct_size)1257 size_t ttm_bo_acc_size(struct ttm_bo_device *bdev,
1258 unsigned long bo_size,
1259 unsigned struct_size)
1260 {
1261 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1262 size_t size = 0;
1263
1264 size += ttm_round_pot(struct_size);
1265 size += ttm_round_pot(npages * sizeof(void *));
1266 size += ttm_round_pot(sizeof(struct ttm_tt));
1267 return size;
1268 }
1269 EXPORT_SYMBOL(ttm_bo_acc_size);
1270
ttm_bo_dma_acc_size(struct ttm_bo_device * bdev,unsigned long bo_size,unsigned struct_size)1271 size_t ttm_bo_dma_acc_size(struct ttm_bo_device *bdev,
1272 unsigned long bo_size,
1273 unsigned struct_size)
1274 {
1275 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1276 size_t size = 0;
1277
1278 size += ttm_round_pot(struct_size);
1279 size += ttm_round_pot(npages * (2*sizeof(void *) + sizeof(dma_addr_t)));
1280 size += ttm_round_pot(sizeof(struct ttm_dma_tt));
1281 return size;
1282 }
1283 EXPORT_SYMBOL(ttm_bo_dma_acc_size);
1284
ttm_bo_create(struct ttm_bo_device * bdev,unsigned long size,enum ttm_bo_type type,struct ttm_placement * placement,uint32_t page_alignment,bool interruptible,struct ttm_buffer_object ** p_bo)1285 int ttm_bo_create(struct ttm_bo_device *bdev,
1286 unsigned long size,
1287 enum ttm_bo_type type,
1288 struct ttm_placement *placement,
1289 uint32_t page_alignment,
1290 bool interruptible,
1291 struct ttm_buffer_object **p_bo)
1292 {
1293 struct ttm_buffer_object *bo;
1294 size_t acc_size;
1295 int ret;
1296
1297 bo = kzalloc(sizeof(*bo), GFP_KERNEL);
1298 if (unlikely(bo == NULL))
1299 return -ENOMEM;
1300
1301 acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object));
1302 ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment,
1303 interruptible, acc_size,
1304 NULL, NULL, NULL);
1305 if (likely(ret == 0))
1306 *p_bo = bo;
1307
1308 return ret;
1309 }
1310 EXPORT_SYMBOL(ttm_bo_create);
1311
ttm_bo_force_list_clean(struct ttm_bo_device * bdev,unsigned mem_type)1312 static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev,
1313 unsigned mem_type)
1314 {
1315 struct ttm_operation_ctx ctx = {
1316 .interruptible = false,
1317 .no_wait_gpu = false,
1318 .flags = TTM_OPT_FLAG_FORCE_ALLOC
1319 };
1320 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1321 struct ttm_bo_global *glob = bdev->glob;
1322 struct dma_fence *fence;
1323 int ret;
1324 unsigned i;
1325
1326 /*
1327 * Can't use standard list traversal since we're unlocking.
1328 */
1329
1330 lockmgr(&glob->lru_lock, LK_EXCLUSIVE);
1331 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
1332 while (!list_empty(&man->lru[i])) {
1333 lockmgr(&glob->lru_lock, LK_RELEASE);
1334 ret = ttm_mem_evict_first(bdev, mem_type, NULL, &ctx);
1335 if (ret)
1336 return ret;
1337 lockmgr(&glob->lru_lock, LK_EXCLUSIVE);
1338 }
1339 }
1340 lockmgr(&glob->lru_lock, LK_RELEASE);
1341
1342 lockmgr(&man->move_lock, LK_EXCLUSIVE);
1343 fence = dma_fence_get(man->move);
1344 lockmgr(&man->move_lock, LK_RELEASE);
1345
1346 if (fence) {
1347 ret = dma_fence_wait(fence, false);
1348 dma_fence_put(fence);
1349 if (ret)
1350 return ret;
1351 }
1352
1353 return 0;
1354 }
1355
ttm_bo_clean_mm(struct ttm_bo_device * bdev,unsigned mem_type)1356 int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1357 {
1358 struct ttm_mem_type_manager *man;
1359 int ret = -EINVAL;
1360
1361 if (mem_type >= TTM_NUM_MEM_TYPES) {
1362 pr_err("Illegal memory type %d\n", mem_type);
1363 return ret;
1364 }
1365 man = &bdev->man[mem_type];
1366
1367 if (!man->has_type) {
1368 pr_err("Trying to take down uninitialized memory manager type %u\n",
1369 mem_type);
1370 return ret;
1371 }
1372
1373 man->use_type = false;
1374 man->has_type = false;
1375
1376 ret = 0;
1377 if (mem_type > 0) {
1378 ret = ttm_bo_force_list_clean(bdev, mem_type);
1379 if (ret) {
1380 pr_err("Cleanup eviction failed\n");
1381 return ret;
1382 }
1383
1384 ret = (*man->func->takedown)(man);
1385 }
1386
1387 dma_fence_put(man->move);
1388 man->move = NULL;
1389
1390 return ret;
1391 }
1392 EXPORT_SYMBOL(ttm_bo_clean_mm);
1393
ttm_bo_evict_mm(struct ttm_bo_device * bdev,unsigned mem_type)1394 int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1395 {
1396 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1397
1398 if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) {
1399 pr_err("Illegal memory manager memory type %u\n", mem_type);
1400 return -EINVAL;
1401 }
1402
1403 if (!man->has_type) {
1404 pr_err("Memory type %u has not been initialized\n", mem_type);
1405 return 0;
1406 }
1407
1408 return ttm_bo_force_list_clean(bdev, mem_type);
1409 }
1410 EXPORT_SYMBOL(ttm_bo_evict_mm);
1411
ttm_bo_init_mm(struct ttm_bo_device * bdev,unsigned type,unsigned long p_size)1412 int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type,
1413 unsigned long p_size)
1414 {
1415 int ret;
1416 struct ttm_mem_type_manager *man;
1417 unsigned i;
1418
1419 BUG_ON(type >= TTM_NUM_MEM_TYPES);
1420 man = &bdev->man[type];
1421 BUG_ON(man->has_type);
1422 man->io_reserve_fastpath = true;
1423 man->use_io_reserve_lru = false;
1424 lockinit(&man->io_reserve_mutex, "ttmior", 0, 0);
1425 lockinit(&man->move_lock, "ttmml", 0, 0);
1426 INIT_LIST_HEAD(&man->io_reserve_lru);
1427
1428 ret = bdev->driver->init_mem_type(bdev, type, man);
1429 if (ret)
1430 return ret;
1431 man->bdev = bdev;
1432
1433 if (type != TTM_PL_SYSTEM) {
1434 ret = (*man->func->init)(man, p_size);
1435 if (ret)
1436 return ret;
1437 }
1438 man->has_type = true;
1439 man->use_type = true;
1440 man->size = p_size;
1441
1442 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
1443 INIT_LIST_HEAD(&man->lru[i]);
1444 man->move = NULL;
1445
1446 return 0;
1447 }
1448 EXPORT_SYMBOL(ttm_bo_init_mm);
1449
ttm_bo_global_kobj_release(struct kobject * kobj)1450 static void ttm_bo_global_kobj_release(struct kobject *kobj)
1451 {
1452 struct ttm_bo_global *glob =
1453 container_of(kobj, struct ttm_bo_global, kobj);
1454
1455 __free_page(glob->dummy_read_page);
1456 }
1457
ttm_bo_global_release(struct drm_global_reference * ref)1458 void ttm_bo_global_release(struct drm_global_reference *ref)
1459 {
1460 struct ttm_bo_global *glob = ref->object;
1461
1462 kobject_del(&glob->kobj);
1463 kobject_put(&glob->kobj);
1464 }
1465 EXPORT_SYMBOL(ttm_bo_global_release);
1466
ttm_bo_global_init(struct drm_global_reference * ref)1467 int ttm_bo_global_init(struct drm_global_reference *ref)
1468 {
1469 struct ttm_bo_global_ref *bo_ref =
1470 container_of(ref, struct ttm_bo_global_ref, ref);
1471 struct ttm_bo_global *glob = ref->object;
1472 int ret;
1473 unsigned i;
1474
1475 lockinit(&glob->device_list_mutex, "ttmdlm", 0, 0);
1476 lockinit(&glob->lru_lock, "ttmlru", 0, 0);
1477 glob->mem_glob = bo_ref->mem_glob;
1478 glob->mem_glob->bo_glob = glob;
1479 glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32);
1480
1481 if (unlikely(glob->dummy_read_page == NULL)) {
1482 ret = -ENOMEM;
1483 goto out_no_drp;
1484 }
1485
1486 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
1487 INIT_LIST_HEAD(&glob->swap_lru[i]);
1488 INIT_LIST_HEAD(&glob->device_list);
1489 atomic_set(&glob->bo_count, 0);
1490
1491 ret = kobject_init_and_add(
1492 &glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects");
1493 if (unlikely(ret != 0))
1494 kobject_put(&glob->kobj);
1495 return ret;
1496 out_no_drp:
1497 kfree(glob);
1498 return ret;
1499 }
1500 EXPORT_SYMBOL(ttm_bo_global_init);
1501
1502
ttm_bo_device_release(struct ttm_bo_device * bdev)1503 int ttm_bo_device_release(struct ttm_bo_device *bdev)
1504 {
1505 int ret = 0;
1506 unsigned i = TTM_NUM_MEM_TYPES;
1507 struct ttm_mem_type_manager *man;
1508 struct ttm_bo_global *glob = bdev->glob;
1509
1510 while (i--) {
1511 man = &bdev->man[i];
1512 if (man->has_type) {
1513 man->use_type = false;
1514 if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) {
1515 ret = -EBUSY;
1516 pr_err("DRM memory manager type %d is not clean\n",
1517 i);
1518 }
1519 man->has_type = false;
1520 }
1521 }
1522
1523 mutex_lock(&glob->device_list_mutex);
1524 list_del(&bdev->device_list);
1525 mutex_unlock(&glob->device_list_mutex);
1526
1527 cancel_delayed_work_sync(&bdev->wq);
1528
1529 if (ttm_bo_delayed_delete(bdev, true))
1530 pr_debug("Delayed destroy list was clean\n");
1531
1532 lockmgr(&glob->lru_lock, LK_EXCLUSIVE);
1533 if (list_empty(&bdev->ddestroy))
1534 TTM_DEBUG("Delayed destroy list was clean\n");
1535 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
1536 if (list_empty(&bdev->man[0].lru[0]))
1537 pr_debug("Swap list %d was clean\n", i);
1538 lockmgr(&glob->lru_lock, LK_RELEASE);
1539
1540 drm_vma_offset_manager_destroy(&bdev->vma_manager);
1541
1542 return ret;
1543 }
1544 EXPORT_SYMBOL(ttm_bo_device_release);
1545
ttm_bo_device_init(struct ttm_bo_device * bdev,struct ttm_bo_global * glob,struct ttm_bo_driver * driver,struct address_space * mapping,uint64_t file_page_offset,bool need_dma32)1546 int ttm_bo_device_init(struct ttm_bo_device *bdev,
1547 struct ttm_bo_global *glob,
1548 struct ttm_bo_driver *driver,
1549 struct address_space *mapping,
1550 uint64_t file_page_offset,
1551 bool need_dma32)
1552 {
1553 int ret = -EINVAL;
1554
1555 bdev->driver = driver;
1556
1557 memset(bdev->man, 0, sizeof(bdev->man));
1558
1559 /*
1560 * Initialize the system memory buffer type.
1561 * Other types need to be driver / IOCTL initialized.
1562 */
1563 ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0);
1564 if (unlikely(ret != 0))
1565 goto out_no_sys;
1566
1567 drm_vma_offset_manager_init(&bdev->vma_manager, file_page_offset,
1568 0x10000000);
1569 INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue);
1570 INIT_LIST_HEAD(&bdev->ddestroy);
1571 /*
1572 * XXX DRAGONFLY - dev_mapping NULL atm, find other XXX DRAGONFLY
1573 * lines and fix when it no longer is in later API change.
1574 */
1575 bdev->dev_mapping = mapping;
1576 bdev->glob = glob;
1577 bdev->need_dma32 = need_dma32;
1578 mutex_lock(&glob->device_list_mutex);
1579 list_add_tail(&bdev->device_list, &glob->device_list);
1580 mutex_unlock(&glob->device_list_mutex);
1581
1582 return 0;
1583 out_no_sys:
1584 return ret;
1585 }
1586 EXPORT_SYMBOL(ttm_bo_device_init);
1587
1588 /*
1589 * buffer object vm functions.
1590 */
1591
ttm_mem_reg_is_pci(struct ttm_bo_device * bdev,struct ttm_mem_reg * mem)1592 bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1593 {
1594 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
1595
1596 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
1597 if (mem->mem_type == TTM_PL_SYSTEM)
1598 return false;
1599
1600 if (man->flags & TTM_MEMTYPE_FLAG_CMA)
1601 return false;
1602
1603 if (mem->placement & TTM_PL_FLAG_CACHED)
1604 return false;
1605 }
1606 return true;
1607 }
1608
1609 #ifdef __DragonFly__
1610
1611 /*
1612 * XXX DRAGONFLY - device_mapping not yet implemented so
1613 * file_mapping is basically always NULL. We have to properly
1614 * release the mmap, etc.
1615 */
1616 void ttm_bo_release_mmap(struct ttm_buffer_object *bo);
1617
1618 #endif
1619
ttm_bo_unmap_virtual_locked(struct ttm_buffer_object * bo)1620 void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo)
1621 {
1622 // drm_vma_node_unmap(&bo->vma_node, bdev->dev_mapping);
1623 ttm_bo_release_mmap(bo);
1624 ttm_mem_io_free_vm(bo);
1625 }
1626
ttm_bo_unmap_virtual(struct ttm_buffer_object * bo)1627 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1628 {
1629 struct ttm_bo_device *bdev = bo->bdev;
1630 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
1631
1632 ttm_mem_io_lock(man, false);
1633 ttm_bo_unmap_virtual_locked(bo);
1634 ttm_mem_io_unlock(man);
1635 }
1636
1637
1638 EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1639
ttm_bo_wait(struct ttm_buffer_object * bo,bool interruptible,bool no_wait)1640 int ttm_bo_wait(struct ttm_buffer_object *bo,
1641 bool interruptible, bool no_wait)
1642 {
1643 long timeout = 15 * HZ;
1644
1645 if (no_wait) {
1646 if (reservation_object_test_signaled_rcu(bo->resv, true))
1647 return 0;
1648 else
1649 return -EBUSY;
1650 }
1651
1652 timeout = reservation_object_wait_timeout_rcu(bo->resv, true,
1653 interruptible, timeout);
1654 if (timeout < 0)
1655 return timeout;
1656
1657 if (timeout == 0)
1658 return -EBUSY;
1659
1660 reservation_object_add_excl_fence(bo->resv, NULL);
1661 return 0;
1662 }
1663 EXPORT_SYMBOL(ttm_bo_wait);
1664
ttm_bo_synccpu_write_grab(struct ttm_buffer_object * bo,bool no_wait)1665 int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait)
1666 {
1667 int ret = 0;
1668
1669 /*
1670 * Using ttm_bo_reserve makes sure the lru lists are updated.
1671 */
1672
1673 ret = ttm_bo_reserve(bo, true, no_wait, NULL);
1674 if (unlikely(ret != 0))
1675 return ret;
1676 ret = ttm_bo_wait(bo, true, no_wait);
1677 if (likely(ret == 0))
1678 atomic_inc(&bo->cpu_writers);
1679 ttm_bo_unreserve(bo);
1680 return ret;
1681 }
1682 EXPORT_SYMBOL(ttm_bo_synccpu_write_grab);
1683
ttm_bo_synccpu_write_release(struct ttm_buffer_object * bo)1684 void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo)
1685 {
1686 atomic_dec(&bo->cpu_writers);
1687 }
1688 EXPORT_SYMBOL(ttm_bo_synccpu_write_release);
1689
1690 /**
1691 * A buffer object shrink method that tries to swap out the first
1692 * buffer object on the bo_global::swap_lru list.
1693 */
ttm_bo_swapout(struct ttm_bo_global * glob,struct ttm_operation_ctx * ctx)1694 int ttm_bo_swapout(struct ttm_bo_global *glob, struct ttm_operation_ctx *ctx)
1695 {
1696 struct ttm_buffer_object *bo;
1697 int ret = -EBUSY;
1698 bool locked;
1699 unsigned i;
1700
1701 lockmgr(&glob->lru_lock, LK_EXCLUSIVE);
1702 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
1703 list_for_each_entry(bo, &glob->swap_lru[i], swap) {
1704 if (ttm_bo_evict_swapout_allowable(bo, ctx, &locked)) {
1705 ret = 0;
1706 break;
1707 }
1708 }
1709 if (!ret)
1710 break;
1711 }
1712
1713 if (ret) {
1714 lockmgr(&glob->lru_lock, LK_RELEASE);
1715 return ret;
1716 }
1717
1718 kref_get(&bo->list_kref);
1719
1720 if (!list_empty(&bo->ddestroy)) {
1721 ret = ttm_bo_cleanup_refs(bo, false, false, locked);
1722 kref_put(&bo->list_kref, ttm_bo_release_list);
1723 return ret;
1724 }
1725
1726 ttm_bo_del_from_lru(bo);
1727 lockmgr(&glob->lru_lock, LK_RELEASE);
1728
1729 /**
1730 * Move to system cached
1731 */
1732
1733 if (bo->mem.mem_type != TTM_PL_SYSTEM ||
1734 bo->ttm->caching_state != tt_cached) {
1735 struct ttm_operation_ctx ctx = { false, false };
1736 struct ttm_mem_reg evict_mem;
1737
1738 evict_mem = bo->mem;
1739 evict_mem.mm_node = NULL;
1740 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED;
1741 evict_mem.mem_type = TTM_PL_SYSTEM;
1742
1743 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, &ctx);
1744 if (unlikely(ret != 0))
1745 goto out;
1746 }
1747
1748 /**
1749 * Make sure BO is idle.
1750 */
1751
1752 ret = ttm_bo_wait(bo, false, false);
1753 if (unlikely(ret != 0))
1754 goto out;
1755
1756 ttm_bo_unmap_virtual(bo);
1757
1758 /**
1759 * Swap out. Buffer will be swapped in again as soon as
1760 * anyone tries to access a ttm page.
1761 */
1762
1763 if (bo->bdev->driver->swap_notify)
1764 bo->bdev->driver->swap_notify(bo);
1765
1766 ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage);
1767 out:
1768
1769 /**
1770 *
1771 * Unreserve without putting on LRU to avoid swapping out an
1772 * already swapped buffer.
1773 */
1774 if (locked)
1775 __ttm_bo_unreserve(bo);
1776 kref_put(&bo->list_kref, ttm_bo_release_list);
1777 return ret;
1778 }
1779 EXPORT_SYMBOL(ttm_bo_swapout);
1780
ttm_bo_swapout_all(struct ttm_bo_device * bdev)1781 void ttm_bo_swapout_all(struct ttm_bo_device *bdev)
1782 {
1783 struct ttm_operation_ctx ctx = {
1784 .interruptible = false,
1785 .no_wait_gpu = false
1786 };
1787
1788 while (ttm_bo_swapout(bdev->glob, &ctx) == 0)
1789 ;
1790 }
1791 EXPORT_SYMBOL(ttm_bo_swapout_all);
1792
1793 /**
1794 * ttm_bo_wait_unreserved - interruptible wait for a buffer object to become
1795 * unreserved
1796 *
1797 * @bo: Pointer to buffer
1798 */
ttm_bo_wait_unreserved(struct ttm_buffer_object * bo)1799 int ttm_bo_wait_unreserved(struct ttm_buffer_object *bo)
1800 {
1801 int ret;
1802
1803 /*
1804 * In the absense of a wait_unlocked API,
1805 * Use the bo::wu_mutex to avoid triggering livelocks due to
1806 * concurrent use of this function. Note that this use of
1807 * bo::wu_mutex can go away if we change locking order to
1808 * mmap_sem -> bo::reserve.
1809 */
1810 ret = mutex_lock_interruptible(&bo->wu_mutex);
1811 if (unlikely(ret != 0))
1812 return -ERESTARTSYS;
1813 if (!ww_mutex_is_locked(&bo->resv->lock))
1814 goto out_unlock;
1815 ret = __ttm_bo_reserve(bo, true, false, NULL);
1816 if (unlikely(ret != 0))
1817 goto out_unlock;
1818 __ttm_bo_unreserve(bo);
1819
1820 out_unlock:
1821 mutex_unlock(&bo->wu_mutex);
1822 return ret;
1823 }
1824