2 * Copyright 2020 Advanced Micro Devices, Inc.
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
22 * Authors: Christian König
25 #include <linux/iosys-map.h>
26 #include <linux/io-mapping.h>
27 #include <linux/scatterlist.h>
29 #include <drm/ttm/ttm_resource.h>
30 #include <drm/ttm/ttm_bo_driver.h>
33 * ttm_lru_bulk_move_init - initialize a bulk move structure
34 * @bulk: the structure to init
36 * For now just memset the structure to zero.
38 void ttm_lru_bulk_move_init(struct ttm_lru_bulk_move *bulk)
40 memset(bulk, 0, sizeof(*bulk));
42 EXPORT_SYMBOL(ttm_lru_bulk_move_init);
45 * ttm_lru_bulk_move_tail - bulk move range of resources to the LRU tail.
47 * @bulk: bulk move structure
49 * Bulk move BOs to the LRU tail, only valid to use when driver makes sure that
50 * resource order never changes. Should be called with &ttm_device.lru_lock held.
52 void ttm_lru_bulk_move_tail(struct ttm_lru_bulk_move *bulk)
56 for (i = 0; i < TTM_NUM_MEM_TYPES; ++i) {
57 for (j = 0; j < TTM_MAX_BO_PRIORITY; ++j) {
58 struct ttm_lru_bulk_move_pos *pos = &bulk->pos[i][j];
59 struct ttm_resource_manager *man;
64 lockdep_assert_held(&pos->first->bo->bdev->lru_lock);
65 dma_resv_assert_held(pos->first->bo->base.resv);
66 dma_resv_assert_held(pos->last->bo->base.resv);
68 man = ttm_manager_type(pos->first->bo->bdev, i);
69 list_bulk_move_tail(&man->lru[j], &pos->first->lru,
74 EXPORT_SYMBOL(ttm_lru_bulk_move_tail);
76 /* Return the bulk move pos object for this resource */
77 static struct ttm_lru_bulk_move_pos *
78 ttm_lru_bulk_move_pos(struct ttm_lru_bulk_move *bulk, struct ttm_resource *res)
80 return &bulk->pos[res->mem_type][res->bo->priority];
83 /* Move the resource to the tail of the bulk move range */
84 static void ttm_lru_bulk_move_pos_tail(struct ttm_lru_bulk_move_pos *pos,
85 struct ttm_resource *res)
87 if (pos->last != res) {
88 list_move(&res->lru, &pos->last->lru);
93 /* Add the resource to a bulk_move cursor */
94 static void ttm_lru_bulk_move_add(struct ttm_lru_bulk_move *bulk,
95 struct ttm_resource *res)
97 struct ttm_lru_bulk_move_pos *pos = ttm_lru_bulk_move_pos(bulk, res);
103 ttm_lru_bulk_move_pos_tail(pos, res);
107 /* Remove the resource from a bulk_move range */
108 static void ttm_lru_bulk_move_del(struct ttm_lru_bulk_move *bulk,
109 struct ttm_resource *res)
111 struct ttm_lru_bulk_move_pos *pos = ttm_lru_bulk_move_pos(bulk, res);
113 if (unlikely(pos->first == res && pos->last == res)) {
116 } else if (pos->first == res) {
117 pos->first = list_next_entry(res, lru);
118 } else if (pos->last == res) {
119 pos->last = list_prev_entry(res, lru);
121 list_move(&res->lru, &pos->last->lru);
125 /* Add the resource to a bulk move if the BO is configured for it */
126 void ttm_resource_add_bulk_move(struct ttm_resource *res,
127 struct ttm_buffer_object *bo)
129 if (bo->bulk_move && !bo->pin_count)
130 ttm_lru_bulk_move_add(bo->bulk_move, res);
133 /* Remove the resource from a bulk move if the BO is configured for it */
134 void ttm_resource_del_bulk_move(struct ttm_resource *res,
135 struct ttm_buffer_object *bo)
137 if (bo->bulk_move && !bo->pin_count)
138 ttm_lru_bulk_move_del(bo->bulk_move, res);
141 /* Move a resource to the LRU or bulk tail */
142 void ttm_resource_move_to_lru_tail(struct ttm_resource *res)
144 struct ttm_buffer_object *bo = res->bo;
145 struct ttm_device *bdev = bo->bdev;
147 lockdep_assert_held(&bo->bdev->lru_lock);
150 list_move_tail(&res->lru, &bdev->pinned);
152 } else if (bo->bulk_move) {
153 struct ttm_lru_bulk_move_pos *pos =
154 ttm_lru_bulk_move_pos(bo->bulk_move, res);
156 ttm_lru_bulk_move_pos_tail(pos, res);
158 struct ttm_resource_manager *man;
160 man = ttm_manager_type(bdev, res->mem_type);
161 list_move_tail(&res->lru, &man->lru[bo->priority]);
166 * ttm_resource_init - resource object constructure
167 * @bo: buffer object this resources is allocated for
168 * @place: placement of the resource
169 * @res: the resource object to inistilize
171 * Initialize a new resource object. Counterpart of ttm_resource_fini().
173 void ttm_resource_init(struct ttm_buffer_object *bo,
174 const struct ttm_place *place,
175 struct ttm_resource *res)
177 struct ttm_resource_manager *man;
180 res->num_pages = PFN_UP(bo->base.size);
181 res->mem_type = place->mem_type;
182 res->placement = place->flags;
183 res->bus.addr = NULL;
185 res->bus.is_iomem = false;
186 res->bus.caching = ttm_cached;
189 man = ttm_manager_type(bo->bdev, place->mem_type);
190 spin_lock(&bo->bdev->lru_lock);
192 list_add_tail(&res->lru, &bo->bdev->pinned);
194 list_add_tail(&res->lru, &man->lru[bo->priority]);
195 man->usage += res->num_pages << PAGE_SHIFT;
196 spin_unlock(&bo->bdev->lru_lock);
198 EXPORT_SYMBOL(ttm_resource_init);
201 * ttm_resource_fini - resource destructor
202 * @man: the resource manager this resource belongs to
203 * @res: the resource to clean up
205 * Should be used by resource manager backends to clean up the TTM resource
206 * objects before freeing the underlying structure. Makes sure the resource is
207 * removed from the LRU before destruction.
208 * Counterpart of ttm_resource_init().
210 void ttm_resource_fini(struct ttm_resource_manager *man,
211 struct ttm_resource *res)
213 struct ttm_device *bdev = man->bdev;
215 spin_lock(&bdev->lru_lock);
216 list_del_init(&res->lru);
217 man->usage -= res->num_pages << PAGE_SHIFT;
218 spin_unlock(&bdev->lru_lock);
220 EXPORT_SYMBOL(ttm_resource_fini);
222 int ttm_resource_alloc(struct ttm_buffer_object *bo,
223 const struct ttm_place *place,
224 struct ttm_resource **res_ptr)
226 struct ttm_resource_manager *man =
227 ttm_manager_type(bo->bdev, place->mem_type);
230 ret = man->func->alloc(man, bo, place, res_ptr);
234 spin_lock(&bo->bdev->lru_lock);
235 ttm_resource_add_bulk_move(*res_ptr, bo);
236 spin_unlock(&bo->bdev->lru_lock);
240 void ttm_resource_free(struct ttm_buffer_object *bo, struct ttm_resource **res)
242 struct ttm_resource_manager *man;
247 spin_lock(&bo->bdev->lru_lock);
248 ttm_resource_del_bulk_move(*res, bo);
249 spin_unlock(&bo->bdev->lru_lock);
250 man = ttm_manager_type(bo->bdev, (*res)->mem_type);
251 man->func->free(man, *res);
254 EXPORT_SYMBOL(ttm_resource_free);
256 static bool ttm_resource_places_compat(struct ttm_resource *res,
257 const struct ttm_place *places,
258 unsigned num_placement)
262 if (res->placement & TTM_PL_FLAG_TEMPORARY)
265 for (i = 0; i < num_placement; i++) {
266 const struct ttm_place *heap = &places[i];
268 if (res->start < heap->fpfn || (heap->lpfn &&
269 (res->start + res->num_pages) > heap->lpfn))
272 if ((res->mem_type == heap->mem_type) &&
273 (!(heap->flags & TTM_PL_FLAG_CONTIGUOUS) ||
274 (res->placement & TTM_PL_FLAG_CONTIGUOUS)))
281 * ttm_resource_compat - check if resource is compatible with placement
283 * @res: the resource to check
284 * @placement: the placement to check against
286 * Returns true if the placement is compatible.
288 bool ttm_resource_compat(struct ttm_resource *res,
289 struct ttm_placement *placement)
291 if (ttm_resource_places_compat(res, placement->placement,
292 placement->num_placement))
295 if ((placement->busy_placement != placement->placement ||
296 placement->num_busy_placement > placement->num_placement) &&
297 ttm_resource_places_compat(res, placement->busy_placement,
298 placement->num_busy_placement))
303 EXPORT_SYMBOL(ttm_resource_compat);
305 void ttm_resource_set_bo(struct ttm_resource *res,
306 struct ttm_buffer_object *bo)
308 spin_lock(&bo->bdev->lru_lock);
310 spin_unlock(&bo->bdev->lru_lock);
314 * ttm_resource_manager_init
316 * @man: memory manager object to init
317 * @bdev: ttm device this manager belongs to
318 * @size: size of managed resources in arbitrary units
320 * Initialise core parts of a manager object.
322 void ttm_resource_manager_init(struct ttm_resource_manager *man,
323 struct ttm_device *bdev,
328 spin_lock_init(&man->move_lock);
333 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
334 INIT_LIST_HEAD(&man->lru[i]);
337 EXPORT_SYMBOL(ttm_resource_manager_init);
340 * ttm_resource_manager_evict_all
342 * @bdev - device to use
343 * @man - manager to use
345 * Evict all the objects out of a memory manager until it is empty.
346 * Part of memory manager cleanup sequence.
348 int ttm_resource_manager_evict_all(struct ttm_device *bdev,
349 struct ttm_resource_manager *man)
351 struct ttm_operation_ctx ctx = {
352 .interruptible = false,
353 .no_wait_gpu = false,
356 struct dma_fence *fence;
361 * Can't use standard list traversal since we're unlocking.
364 spin_lock(&bdev->lru_lock);
365 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
366 while (!list_empty(&man->lru[i])) {
367 spin_unlock(&bdev->lru_lock);
368 ret = ttm_mem_evict_first(bdev, man, NULL, &ctx,
372 spin_lock(&bdev->lru_lock);
375 spin_unlock(&bdev->lru_lock);
377 spin_lock(&man->move_lock);
378 fence = dma_fence_get(man->move);
379 spin_unlock(&man->move_lock);
382 ret = dma_fence_wait(fence, false);
383 dma_fence_put(fence);
390 EXPORT_SYMBOL(ttm_resource_manager_evict_all);
393 * ttm_resource_manager_usage
395 * @man: A memory manager object.
397 * Return how many resources are currently used.
399 uint64_t ttm_resource_manager_usage(struct ttm_resource_manager *man)
403 spin_lock(&man->bdev->lru_lock);
405 spin_unlock(&man->bdev->lru_lock);
408 EXPORT_SYMBOL(ttm_resource_manager_usage);
411 * ttm_resource_manager_debug
413 * @man: manager type to dump.
414 * @p: printer to use for debug.
416 void ttm_resource_manager_debug(struct ttm_resource_manager *man,
417 struct drm_printer *p)
419 drm_printf(p, " use_type: %d\n", man->use_type);
420 drm_printf(p, " use_tt: %d\n", man->use_tt);
421 drm_printf(p, " size: %llu\n", man->size);
422 drm_printf(p, " usage: %llu\n", ttm_resource_manager_usage(man));
423 if (man->func->debug)
424 man->func->debug(man, p);
426 EXPORT_SYMBOL(ttm_resource_manager_debug);
429 * ttm_resource_manager_first
431 * @man: resource manager to iterate over
432 * @cursor: cursor to record the position
434 * Returns the first resource from the resource manager.
436 struct ttm_resource *
437 ttm_resource_manager_first(struct ttm_resource_manager *man,
438 struct ttm_resource_cursor *cursor)
440 struct ttm_resource *res;
442 lockdep_assert_held(&man->bdev->lru_lock);
444 for (cursor->priority = 0; cursor->priority < TTM_MAX_BO_PRIORITY;
446 list_for_each_entry(res, &man->lru[cursor->priority], lru)
453 * ttm_resource_manager_next
455 * @man: resource manager to iterate over
456 * @cursor: cursor to record the position
457 * @res: the current resource pointer
459 * Returns the next resource from the resource manager.
461 struct ttm_resource *
462 ttm_resource_manager_next(struct ttm_resource_manager *man,
463 struct ttm_resource_cursor *cursor,
464 struct ttm_resource *res)
466 lockdep_assert_held(&man->bdev->lru_lock);
468 list_for_each_entry_continue(res, &man->lru[cursor->priority], lru)
471 for (++cursor->priority; cursor->priority < TTM_MAX_BO_PRIORITY;
473 list_for_each_entry(res, &man->lru[cursor->priority], lru)
479 static void ttm_kmap_iter_iomap_map_local(struct ttm_kmap_iter *iter,
480 struct iosys_map *dmap,
483 struct ttm_kmap_iter_iomap *iter_io =
484 container_of(iter, typeof(*iter_io), base);
488 while (i >= iter_io->cache.end) {
489 iter_io->cache.sg = iter_io->cache.sg ?
490 sg_next(iter_io->cache.sg) : iter_io->st->sgl;
491 iter_io->cache.i = iter_io->cache.end;
492 iter_io->cache.end += sg_dma_len(iter_io->cache.sg) >>
494 iter_io->cache.offs = sg_dma_address(iter_io->cache.sg) -
498 if (i < iter_io->cache.i) {
499 iter_io->cache.end = 0;
500 iter_io->cache.sg = NULL;
504 addr = io_mapping_map_local_wc(iter_io->iomap, iter_io->cache.offs +
505 (((resource_size_t)i - iter_io->cache.i)
507 iosys_map_set_vaddr_iomem(dmap, addr);
510 static void ttm_kmap_iter_iomap_unmap_local(struct ttm_kmap_iter *iter,
511 struct iosys_map *map)
513 io_mapping_unmap_local(map->vaddr_iomem);
516 static const struct ttm_kmap_iter_ops ttm_kmap_iter_io_ops = {
517 .map_local = ttm_kmap_iter_iomap_map_local,
518 .unmap_local = ttm_kmap_iter_iomap_unmap_local,
523 * ttm_kmap_iter_iomap_init - Initialize a struct ttm_kmap_iter_iomap
524 * @iter_io: The struct ttm_kmap_iter_iomap to initialize.
525 * @iomap: The struct io_mapping representing the underlying linear io_memory.
526 * @st: sg_table into @iomap, representing the memory of the struct
528 * @start: Offset that needs to be subtracted from @st to make
529 * sg_dma_address(st->sgl) - @start == 0 for @iomap start.
531 * Return: Pointer to the embedded struct ttm_kmap_iter.
533 struct ttm_kmap_iter *
534 ttm_kmap_iter_iomap_init(struct ttm_kmap_iter_iomap *iter_io,
535 struct io_mapping *iomap,
537 resource_size_t start)
539 iter_io->base.ops = &ttm_kmap_iter_io_ops;
540 iter_io->iomap = iomap;
542 iter_io->start = start;
543 memset(&iter_io->cache, 0, sizeof(iter_io->cache));
545 return &iter_io->base;
547 EXPORT_SYMBOL(ttm_kmap_iter_iomap_init);
550 * DOC: Linear io iterator
552 * This code should die in the not too near future. Best would be if we could
553 * make io-mapping use memremap for all io memory, and have memremap
554 * implement a kmap_local functionality. We could then strip a huge amount of
555 * code. These linear io iterators are implemented to mimic old functionality,
556 * and they don't use kmap_local semantics at all internally. Rather ioremap or
557 * friends, and at least on 32-bit they add global TLB flushes and points
561 static void ttm_kmap_iter_linear_io_map_local(struct ttm_kmap_iter *iter,
562 struct iosys_map *dmap,
565 struct ttm_kmap_iter_linear_io *iter_io =
566 container_of(iter, typeof(*iter_io), base);
568 *dmap = iter_io->dmap;
569 iosys_map_incr(dmap, i * PAGE_SIZE);
572 static const struct ttm_kmap_iter_ops ttm_kmap_iter_linear_io_ops = {
573 .map_local = ttm_kmap_iter_linear_io_map_local,
578 * ttm_kmap_iter_linear_io_init - Initialize an iterator for linear io memory
579 * @iter_io: The iterator to initialize
580 * @bdev: The TTM device
581 * @mem: The ttm resource representing the iomap.
583 * This function is for internal TTM use only. It sets up a memcpy kmap iterator
584 * pointing at a linear chunk of io memory.
586 * Return: A pointer to the embedded struct ttm_kmap_iter or error pointer on
589 struct ttm_kmap_iter *
590 ttm_kmap_iter_linear_io_init(struct ttm_kmap_iter_linear_io *iter_io,
591 struct ttm_device *bdev,
592 struct ttm_resource *mem)
596 ret = ttm_mem_io_reserve(bdev, mem);
599 if (!mem->bus.is_iomem) {
605 iosys_map_set_vaddr(&iter_io->dmap, mem->bus.addr);
606 iter_io->needs_unmap = false;
608 size_t bus_size = (size_t)mem->num_pages << PAGE_SHIFT;
610 iter_io->needs_unmap = true;
611 memset(&iter_io->dmap, 0, sizeof(iter_io->dmap));
612 if (mem->bus.caching == ttm_write_combined)
613 iosys_map_set_vaddr_iomem(&iter_io->dmap,
614 ioremap_wc(mem->bus.offset,
616 else if (mem->bus.caching == ttm_cached)
617 iosys_map_set_vaddr(&iter_io->dmap,
618 memremap(mem->bus.offset, bus_size,
623 /* If uncached requested or if mapping cached or wc failed */
624 if (iosys_map_is_null(&iter_io->dmap))
625 iosys_map_set_vaddr_iomem(&iter_io->dmap,
626 ioremap(mem->bus.offset,
629 if (iosys_map_is_null(&iter_io->dmap)) {
635 iter_io->base.ops = &ttm_kmap_iter_linear_io_ops;
636 return &iter_io->base;
639 ttm_mem_io_free(bdev, mem);
645 * ttm_kmap_iter_linear_io_fini - Clean up an iterator for linear io memory
646 * @iter_io: The iterator to initialize
647 * @bdev: The TTM device
648 * @mem: The ttm resource representing the iomap.
650 * This function is for internal TTM use only. It cleans up a memcpy kmap
651 * iterator initialized by ttm_kmap_iter_linear_io_init.
654 ttm_kmap_iter_linear_io_fini(struct ttm_kmap_iter_linear_io *iter_io,
655 struct ttm_device *bdev,
656 struct ttm_resource *mem)
658 if (iter_io->needs_unmap && iosys_map_is_set(&iter_io->dmap)) {
659 if (iter_io->dmap.is_iomem)
660 iounmap(iter_io->dmap.vaddr_iomem);
662 memunmap(iter_io->dmap.vaddr);
665 ttm_mem_io_free(bdev, mem);
668 #if defined(CONFIG_DEBUG_FS)
670 static int ttm_resource_manager_show(struct seq_file *m, void *unused)
672 struct ttm_resource_manager *man =
673 (struct ttm_resource_manager *)m->private;
674 struct drm_printer p = drm_seq_file_printer(m);
675 ttm_resource_manager_debug(man, &p);
678 DEFINE_SHOW_ATTRIBUTE(ttm_resource_manager);
683 * ttm_resource_manager_create_debugfs - Create debugfs entry for specified
685 * @man: The TTM resource manager for which the debugfs stats file be creates
686 * @parent: debugfs directory in which the file will reside
687 * @name: The filename to create.
689 * This function setups up a debugfs file that can be used to look
690 * at debug statistics of the specified ttm_resource_manager.
692 void ttm_resource_manager_create_debugfs(struct ttm_resource_manager *man,
693 struct dentry * parent,
696 #if defined(CONFIG_DEBUG_FS)
697 debugfs_create_file(name, 0444, parent, man, &ttm_resource_manager_fops);
700 EXPORT_SYMBOL(ttm_resource_manager_create_debugfs);