1 // SPDX-License-Identifier: GPL-2.0 OR MIT
2 /**************************************************************************
4 * Copyright 2009-2015 VMware, Inc., Palo Alto, CA., USA
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:
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
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.
26 **************************************************************************/
28 #include "vmwgfx_drv.h"
29 #include <drm/ttm/ttm_bo_driver.h>
30 #include <drm/ttm/ttm_placement.h>
31 #include <drm/ttm/ttm_page_alloc.h>
33 static const struct ttm_place vram_placement_flags = {
36 .mem_type = TTM_PL_VRAM,
40 static const struct ttm_place sys_placement_flags = {
43 .mem_type = TTM_PL_SYSTEM,
47 static const struct ttm_place gmr_placement_flags = {
50 .mem_type = VMW_PL_GMR,
54 static const struct ttm_place mob_placement_flags = {
57 .mem_type = VMW_PL_MOB,
61 struct ttm_placement vmw_vram_placement = {
63 .placement = &vram_placement_flags,
64 .num_busy_placement = 1,
65 .busy_placement = &vram_placement_flags
68 static const struct ttm_place vram_gmr_placement_flags[] = {
72 .mem_type = TTM_PL_VRAM,
77 .mem_type = VMW_PL_GMR,
82 static const struct ttm_place gmr_vram_placement_flags[] = {
86 .mem_type = VMW_PL_GMR,
91 .mem_type = TTM_PL_VRAM,
96 struct ttm_placement vmw_vram_gmr_placement = {
98 .placement = vram_gmr_placement_flags,
99 .num_busy_placement = 1,
100 .busy_placement = &gmr_placement_flags
103 struct ttm_placement vmw_vram_sys_placement = {
105 .placement = &vram_placement_flags,
106 .num_busy_placement = 1,
107 .busy_placement = &sys_placement_flags
110 struct ttm_placement vmw_sys_placement = {
112 .placement = &sys_placement_flags,
113 .num_busy_placement = 1,
114 .busy_placement = &sys_placement_flags
117 static const struct ttm_place evictable_placement_flags[] = {
121 .mem_type = TTM_PL_SYSTEM,
126 .mem_type = TTM_PL_VRAM,
131 .mem_type = VMW_PL_GMR,
136 .mem_type = VMW_PL_MOB,
141 static const struct ttm_place nonfixed_placement_flags[] = {
145 .mem_type = TTM_PL_SYSTEM,
150 .mem_type = VMW_PL_GMR,
155 .mem_type = VMW_PL_MOB,
160 struct ttm_placement vmw_evictable_placement = {
162 .placement = evictable_placement_flags,
163 .num_busy_placement = 1,
164 .busy_placement = &sys_placement_flags
167 struct ttm_placement vmw_srf_placement = {
169 .num_busy_placement = 2,
170 .placement = &gmr_placement_flags,
171 .busy_placement = gmr_vram_placement_flags
174 struct ttm_placement vmw_mob_placement = {
176 .num_busy_placement = 1,
177 .placement = &mob_placement_flags,
178 .busy_placement = &mob_placement_flags
181 struct ttm_placement vmw_nonfixed_placement = {
183 .placement = nonfixed_placement_flags,
184 .num_busy_placement = 1,
185 .busy_placement = &sys_placement_flags
189 struct ttm_tt dma_ttm;
190 struct vmw_private *dev_priv;
195 struct vmw_sg_table vsgt;
196 uint64_t sg_alloc_size;
201 const size_t vmw_tt_size = sizeof(struct vmw_ttm_tt);
204 * Helper functions to advance a struct vmw_piter iterator.
206 * @viter: Pointer to the iterator.
208 * These functions return false if past the end of the list,
209 * true otherwise. Functions are selected depending on the current
212 static bool __vmw_piter_non_sg_next(struct vmw_piter *viter)
214 return ++(viter->i) < viter->num_pages;
217 static bool __vmw_piter_sg_next(struct vmw_piter *viter)
219 bool ret = __vmw_piter_non_sg_next(viter);
221 return __sg_page_iter_dma_next(&viter->iter) && ret;
226 * Helper functions to return a pointer to the current page.
228 * @viter: Pointer to the iterator
230 * These functions return a pointer to the page currently
231 * pointed to by @viter. Functions are selected depending on the
232 * current mapping mode.
234 static struct page *__vmw_piter_non_sg_page(struct vmw_piter *viter)
236 return viter->pages[viter->i];
240 * Helper functions to return the DMA address of the current page.
242 * @viter: Pointer to the iterator
244 * These functions return the DMA address of the page currently
245 * pointed to by @viter. Functions are selected depending on the
246 * current mapping mode.
248 static dma_addr_t __vmw_piter_phys_addr(struct vmw_piter *viter)
250 return page_to_phys(viter->pages[viter->i]);
253 static dma_addr_t __vmw_piter_dma_addr(struct vmw_piter *viter)
255 return viter->addrs[viter->i];
258 static dma_addr_t __vmw_piter_sg_addr(struct vmw_piter *viter)
260 return sg_page_iter_dma_address(&viter->iter);
265 * vmw_piter_start - Initialize a struct vmw_piter.
267 * @viter: Pointer to the iterator to initialize
268 * @vsgt: Pointer to a struct vmw_sg_table to initialize from
270 * Note that we're following the convention of __sg_page_iter_start, so that
271 * the iterator doesn't point to a valid page after initialization; it has
272 * to be advanced one step first.
274 void vmw_piter_start(struct vmw_piter *viter, const struct vmw_sg_table *vsgt,
275 unsigned long p_offset)
277 viter->i = p_offset - 1;
278 viter->num_pages = vsgt->num_pages;
279 viter->page = &__vmw_piter_non_sg_page;
280 viter->pages = vsgt->pages;
281 switch (vsgt->mode) {
283 viter->next = &__vmw_piter_non_sg_next;
284 viter->dma_address = &__vmw_piter_phys_addr;
286 case vmw_dma_alloc_coherent:
287 viter->next = &__vmw_piter_non_sg_next;
288 viter->dma_address = &__vmw_piter_dma_addr;
289 viter->addrs = vsgt->addrs;
291 case vmw_dma_map_populate:
292 case vmw_dma_map_bind:
293 viter->next = &__vmw_piter_sg_next;
294 viter->dma_address = &__vmw_piter_sg_addr;
295 __sg_page_iter_start(&viter->iter.base, vsgt->sgt->sgl,
296 vsgt->sgt->orig_nents, p_offset);
304 * vmw_ttm_unmap_from_dma - unmap device addresses previsouly mapped for
307 * @vmw_tt: Pointer to a struct vmw_ttm_backend
309 * Used to free dma mappings previously mapped by vmw_ttm_map_for_dma.
311 static void vmw_ttm_unmap_from_dma(struct vmw_ttm_tt *vmw_tt)
313 struct device *dev = vmw_tt->dev_priv->dev->dev;
315 dma_unmap_sgtable(dev, &vmw_tt->sgt, DMA_BIDIRECTIONAL, 0);
316 vmw_tt->sgt.nents = vmw_tt->sgt.orig_nents;
320 * vmw_ttm_map_for_dma - map TTM pages to get device addresses
322 * @vmw_tt: Pointer to a struct vmw_ttm_backend
324 * This function is used to get device addresses from the kernel DMA layer.
325 * However, it's violating the DMA API in that when this operation has been
326 * performed, it's illegal for the CPU to write to the pages without first
327 * unmapping the DMA mappings, or calling dma_sync_sg_for_cpu(). It is
328 * therefore only legal to call this function if we know that the function
329 * dma_sync_sg_for_cpu() is a NOP, and dma_sync_sg_for_device() is at most
330 * a CPU write buffer flush.
332 static int vmw_ttm_map_for_dma(struct vmw_ttm_tt *vmw_tt)
334 struct device *dev = vmw_tt->dev_priv->dev->dev;
336 return dma_map_sgtable(dev, &vmw_tt->sgt, DMA_BIDIRECTIONAL, 0);
340 * vmw_ttm_map_dma - Make sure TTM pages are visible to the device
342 * @vmw_tt: Pointer to a struct vmw_ttm_tt
344 * Select the correct function for and make sure the TTM pages are
345 * visible to the device. Allocate storage for the device mappings.
346 * If a mapping has already been performed, indicated by the storage
347 * pointer being non NULL, the function returns success.
349 static int vmw_ttm_map_dma(struct vmw_ttm_tt *vmw_tt)
351 struct vmw_private *dev_priv = vmw_tt->dev_priv;
352 struct ttm_mem_global *glob = vmw_mem_glob(dev_priv);
353 struct vmw_sg_table *vsgt = &vmw_tt->vsgt;
354 struct ttm_operation_ctx ctx = {
355 .interruptible = true,
358 struct vmw_piter iter;
361 static size_t sgl_size;
362 static size_t sgt_size;
363 struct scatterlist *sg;
368 vsgt->mode = dev_priv->map_mode;
369 vsgt->pages = vmw_tt->dma_ttm.pages;
370 vsgt->num_pages = vmw_tt->dma_ttm.num_pages;
371 vsgt->addrs = vmw_tt->dma_ttm.dma_address;
372 vsgt->sgt = &vmw_tt->sgt;
374 switch (dev_priv->map_mode) {
375 case vmw_dma_map_bind:
376 case vmw_dma_map_populate:
377 if (unlikely(!sgl_size)) {
378 sgl_size = ttm_round_pot(sizeof(struct scatterlist));
379 sgt_size = ttm_round_pot(sizeof(struct sg_table));
381 vmw_tt->sg_alloc_size = sgt_size + sgl_size * vsgt->num_pages;
382 ret = ttm_mem_global_alloc(glob, vmw_tt->sg_alloc_size, &ctx);
383 if (unlikely(ret != 0))
386 sg = __sg_alloc_table_from_pages(&vmw_tt->sgt, vsgt->pages,
388 (unsigned long) vsgt->num_pages << PAGE_SHIFT,
389 dma_get_max_seg_size(dev_priv->dev->dev),
390 NULL, 0, GFP_KERNEL);
393 goto out_sg_alloc_fail;
396 if (vsgt->num_pages > vmw_tt->sgt.orig_nents) {
397 uint64_t over_alloc =
398 sgl_size * (vsgt->num_pages -
399 vmw_tt->sgt.orig_nents);
401 ttm_mem_global_free(glob, over_alloc);
402 vmw_tt->sg_alloc_size -= over_alloc;
405 ret = vmw_ttm_map_for_dma(vmw_tt);
406 if (unlikely(ret != 0))
414 old = ~((dma_addr_t) 0);
415 vmw_tt->vsgt.num_regions = 0;
416 for (vmw_piter_start(&iter, vsgt, 0); vmw_piter_next(&iter);) {
417 dma_addr_t cur = vmw_piter_dma_addr(&iter);
419 if (cur != old + PAGE_SIZE)
420 vmw_tt->vsgt.num_regions++;
424 vmw_tt->mapped = true;
428 sg_free_table(vmw_tt->vsgt.sgt);
429 vmw_tt->vsgt.sgt = NULL;
431 ttm_mem_global_free(glob, vmw_tt->sg_alloc_size);
436 * vmw_ttm_unmap_dma - Tear down any TTM page device mappings
438 * @vmw_tt: Pointer to a struct vmw_ttm_tt
440 * Tear down any previously set up device DMA mappings and free
441 * any storage space allocated for them. If there are no mappings set up,
442 * this function is a NOP.
444 static void vmw_ttm_unmap_dma(struct vmw_ttm_tt *vmw_tt)
446 struct vmw_private *dev_priv = vmw_tt->dev_priv;
448 if (!vmw_tt->vsgt.sgt)
451 switch (dev_priv->map_mode) {
452 case vmw_dma_map_bind:
453 case vmw_dma_map_populate:
454 vmw_ttm_unmap_from_dma(vmw_tt);
455 sg_free_table(vmw_tt->vsgt.sgt);
456 vmw_tt->vsgt.sgt = NULL;
457 ttm_mem_global_free(vmw_mem_glob(dev_priv),
458 vmw_tt->sg_alloc_size);
463 vmw_tt->mapped = false;
467 * vmw_bo_sg_table - Return a struct vmw_sg_table object for a
470 * @bo: Pointer to a struct ttm_buffer_object
472 * Returns a pointer to a struct vmw_sg_table object. The object should
473 * not be freed after use.
474 * Note that for the device addresses to be valid, the buffer object must
475 * either be reserved or pinned.
477 const struct vmw_sg_table *vmw_bo_sg_table(struct ttm_buffer_object *bo)
479 struct vmw_ttm_tt *vmw_tt =
480 container_of(bo->ttm, struct vmw_ttm_tt, dma_ttm);
482 return &vmw_tt->vsgt;
486 static int vmw_ttm_bind(struct ttm_bo_device *bdev,
487 struct ttm_tt *ttm, struct ttm_resource *bo_mem)
489 struct vmw_ttm_tt *vmw_be =
490 container_of(ttm, struct vmw_ttm_tt, dma_ttm);
499 ret = vmw_ttm_map_dma(vmw_be);
500 if (unlikely(ret != 0))
503 vmw_be->gmr_id = bo_mem->start;
504 vmw_be->mem_type = bo_mem->mem_type;
506 switch (bo_mem->mem_type) {
508 ret = vmw_gmr_bind(vmw_be->dev_priv, &vmw_be->vsgt,
509 ttm->num_pages, vmw_be->gmr_id);
512 if (unlikely(vmw_be->mob == NULL)) {
514 vmw_mob_create(ttm->num_pages);
515 if (unlikely(vmw_be->mob == NULL))
519 ret = vmw_mob_bind(vmw_be->dev_priv, vmw_be->mob,
520 &vmw_be->vsgt, ttm->num_pages,
526 vmw_be->bound = true;
530 static void vmw_ttm_unbind(struct ttm_bo_device *bdev,
533 struct vmw_ttm_tt *vmw_be =
534 container_of(ttm, struct vmw_ttm_tt, dma_ttm);
539 switch (vmw_be->mem_type) {
541 vmw_gmr_unbind(vmw_be->dev_priv, vmw_be->gmr_id);
544 vmw_mob_unbind(vmw_be->dev_priv, vmw_be->mob);
550 if (vmw_be->dev_priv->map_mode == vmw_dma_map_bind)
551 vmw_ttm_unmap_dma(vmw_be);
552 vmw_be->bound = false;
556 static void vmw_ttm_destroy(struct ttm_bo_device *bdev, struct ttm_tt *ttm)
558 struct vmw_ttm_tt *vmw_be =
559 container_of(ttm, struct vmw_ttm_tt, dma_ttm);
561 vmw_ttm_unbind(bdev, ttm);
562 ttm_tt_destroy_common(bdev, ttm);
563 vmw_ttm_unmap_dma(vmw_be);
564 if (vmw_be->dev_priv->map_mode == vmw_dma_alloc_coherent)
565 ttm_tt_fini(&vmw_be->dma_ttm);
570 vmw_mob_destroy(vmw_be->mob);
576 static int vmw_ttm_populate(struct ttm_bo_device *bdev,
577 struct ttm_tt *ttm, struct ttm_operation_ctx *ctx)
579 struct vmw_ttm_tt *vmw_tt =
580 container_of(ttm, struct vmw_ttm_tt, dma_ttm);
581 struct vmw_private *dev_priv = vmw_tt->dev_priv;
582 struct ttm_mem_global *glob = vmw_mem_glob(dev_priv);
585 if (ttm_tt_is_populated(ttm))
588 if (dev_priv->map_mode == vmw_dma_alloc_coherent) {
590 ttm_round_pot(ttm->num_pages * sizeof(dma_addr_t));
591 ret = ttm_mem_global_alloc(glob, size, ctx);
592 if (unlikely(ret != 0))
595 ret = ttm_dma_populate(&vmw_tt->dma_ttm, dev_priv->dev->dev,
597 if (unlikely(ret != 0))
598 ttm_mem_global_free(glob, size);
600 ret = ttm_pool_populate(ttm, ctx);
605 static void vmw_ttm_unpopulate(struct ttm_bo_device *bdev,
608 struct vmw_ttm_tt *vmw_tt = container_of(ttm, struct vmw_ttm_tt,
610 struct vmw_private *dev_priv = vmw_tt->dev_priv;
611 struct ttm_mem_global *glob = vmw_mem_glob(dev_priv);
615 vmw_mob_destroy(vmw_tt->mob);
619 vmw_ttm_unmap_dma(vmw_tt);
620 if (dev_priv->map_mode == vmw_dma_alloc_coherent) {
622 ttm_round_pot(ttm->num_pages * sizeof(dma_addr_t));
624 ttm_dma_unpopulate(&vmw_tt->dma_ttm, dev_priv->dev->dev);
625 ttm_mem_global_free(glob, size);
627 ttm_pool_unpopulate(ttm);
630 static struct ttm_tt *vmw_ttm_tt_create(struct ttm_buffer_object *bo,
633 struct vmw_ttm_tt *vmw_be;
636 vmw_be = kzalloc(sizeof(*vmw_be), GFP_KERNEL);
640 vmw_be->dev_priv = container_of(bo->bdev, struct vmw_private, bdev);
643 if (vmw_be->dev_priv->map_mode == vmw_dma_alloc_coherent)
644 ret = ttm_dma_tt_init(&vmw_be->dma_ttm, bo, page_flags,
647 ret = ttm_tt_init(&vmw_be->dma_ttm, bo, page_flags,
649 if (unlikely(ret != 0))
652 return &vmw_be->dma_ttm;
658 static void vmw_evict_flags(struct ttm_buffer_object *bo,
659 struct ttm_placement *placement)
661 *placement = vmw_sys_placement;
664 static int vmw_verify_access(struct ttm_buffer_object *bo, struct file *filp)
666 struct ttm_object_file *tfile =
667 vmw_fpriv((struct drm_file *)filp->private_data)->tfile;
669 return vmw_user_bo_verify_access(bo, tfile);
672 static int vmw_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_resource *mem)
674 struct vmw_private *dev_priv = container_of(bdev, struct vmw_private, bdev);
676 switch (mem->mem_type) {
682 mem->bus.offset = (mem->start << PAGE_SHIFT) +
683 dev_priv->vram_start;
684 mem->bus.is_iomem = true;
685 mem->bus.caching = ttm_cached;
694 * vmw_move_notify - TTM move_notify_callback
696 * @bo: The TTM buffer object about to move.
697 * @mem: The struct ttm_resource indicating to what memory
698 * region the move is taking place.
700 * Calls move_notify for all subsystems needing it.
701 * (currently only resources).
703 static void vmw_move_notify(struct ttm_buffer_object *bo,
705 struct ttm_resource *mem)
709 vmw_bo_move_notify(bo, mem);
710 vmw_query_move_notify(bo, mem);
715 * vmw_swap_notify - TTM move_notify_callback
717 * @bo: The TTM buffer object about to be swapped out.
719 static void vmw_swap_notify(struct ttm_buffer_object *bo)
721 vmw_bo_swap_notify(bo);
722 (void) ttm_bo_wait(bo, false, false);
725 static int vmw_move(struct ttm_buffer_object *bo,
727 struct ttm_operation_ctx *ctx,
728 struct ttm_resource *new_mem)
730 struct ttm_resource_manager *old_man = ttm_manager_type(bo->bdev, bo->mem.mem_type);
731 struct ttm_resource_manager *new_man = ttm_manager_type(bo->bdev, new_mem->mem_type);
734 if (new_man->use_tt && new_mem->mem_type != TTM_PL_SYSTEM) {
735 ret = vmw_ttm_bind(bo->bdev, bo->ttm, new_mem);
740 vmw_move_notify(bo, evict, new_mem);
742 if (old_man->use_tt && new_man->use_tt) {
743 if (bo->mem.mem_type == TTM_PL_SYSTEM) {
744 ttm_bo_assign_mem(bo, new_mem);
747 ret = ttm_bo_wait_ctx(bo, ctx);
751 vmw_ttm_unbind(bo->bdev, bo->ttm);
752 ttm_resource_free(bo, &bo->mem);
753 ttm_bo_assign_mem(bo, new_mem);
756 ret = ttm_bo_move_memcpy(bo, ctx, new_mem);
762 swap(*new_mem, bo->mem);
763 vmw_move_notify(bo, false, new_mem);
764 swap(*new_mem, bo->mem);
769 vmw_delete_mem_notify(struct ttm_buffer_object *bo)
771 vmw_move_notify(bo, false, NULL);
774 struct ttm_bo_driver vmw_bo_driver = {
775 .ttm_tt_create = &vmw_ttm_tt_create,
776 .ttm_tt_populate = &vmw_ttm_populate,
777 .ttm_tt_unpopulate = &vmw_ttm_unpopulate,
778 .ttm_tt_destroy = &vmw_ttm_destroy,
779 .eviction_valuable = ttm_bo_eviction_valuable,
780 .evict_flags = vmw_evict_flags,
782 .verify_access = vmw_verify_access,
783 .delete_mem_notify = vmw_delete_mem_notify,
784 .swap_notify = vmw_swap_notify,
785 .io_mem_reserve = &vmw_ttm_io_mem_reserve,
788 int vmw_bo_create_and_populate(struct vmw_private *dev_priv,
789 unsigned long bo_size,
790 struct ttm_buffer_object **bo_p)
792 struct ttm_operation_ctx ctx = {
793 .interruptible = false,
796 struct ttm_buffer_object *bo;
799 ret = vmw_bo_create_kernel(dev_priv, bo_size,
802 if (unlikely(ret != 0))
805 ret = ttm_bo_reserve(bo, false, true, NULL);
807 ret = vmw_ttm_populate(bo->bdev, bo->ttm, &ctx);
808 if (likely(ret == 0)) {
809 struct vmw_ttm_tt *vmw_tt =
810 container_of(bo->ttm, struct vmw_ttm_tt, dma_ttm);
811 ret = vmw_ttm_map_dma(vmw_tt);
814 ttm_bo_unreserve(bo);
816 if (likely(ret == 0))