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>
32 static const struct ttm_place vram_placement_flags
= {
35 .mem_type
= TTM_PL_VRAM
,
39 static const struct ttm_place sys_placement_flags
= {
42 .mem_type
= TTM_PL_SYSTEM
,
46 static const struct ttm_place gmr_placement_flags
= {
49 .mem_type
= VMW_PL_GMR
,
53 static const struct ttm_place mob_placement_flags
= {
56 .mem_type
= VMW_PL_MOB
,
60 struct ttm_placement vmw_vram_placement
= {
62 .placement
= &vram_placement_flags
,
63 .num_busy_placement
= 1,
64 .busy_placement
= &vram_placement_flags
67 static const struct ttm_place vram_gmr_placement_flags
[] = {
71 .mem_type
= TTM_PL_VRAM
,
76 .mem_type
= VMW_PL_GMR
,
81 static const struct ttm_place gmr_vram_placement_flags
[] = {
85 .mem_type
= VMW_PL_GMR
,
90 .mem_type
= TTM_PL_VRAM
,
95 static const struct ttm_place vmw_sys_placement_flags
= {
98 .mem_type
= VMW_PL_SYSTEM
,
102 struct ttm_placement vmw_vram_gmr_placement
= {
104 .placement
= vram_gmr_placement_flags
,
105 .num_busy_placement
= 1,
106 .busy_placement
= &gmr_placement_flags
109 struct ttm_placement vmw_vram_sys_placement
= {
111 .placement
= &vram_placement_flags
,
112 .num_busy_placement
= 1,
113 .busy_placement
= &sys_placement_flags
116 struct ttm_placement vmw_sys_placement
= {
118 .placement
= &sys_placement_flags
,
119 .num_busy_placement
= 1,
120 .busy_placement
= &sys_placement_flags
123 struct ttm_placement vmw_pt_sys_placement
= {
125 .placement
= &vmw_sys_placement_flags
,
126 .num_busy_placement
= 1,
127 .busy_placement
= &vmw_sys_placement_flags
130 static const struct ttm_place nonfixed_placement_flags
[] = {
134 .mem_type
= TTM_PL_SYSTEM
,
139 .mem_type
= VMW_PL_GMR
,
144 .mem_type
= VMW_PL_MOB
,
149 struct ttm_placement vmw_srf_placement
= {
151 .num_busy_placement
= 2,
152 .placement
= &gmr_placement_flags
,
153 .busy_placement
= gmr_vram_placement_flags
156 struct ttm_placement vmw_mob_placement
= {
158 .num_busy_placement
= 1,
159 .placement
= &mob_placement_flags
,
160 .busy_placement
= &mob_placement_flags
163 struct ttm_placement vmw_nonfixed_placement
= {
165 .placement
= nonfixed_placement_flags
,
166 .num_busy_placement
= 1,
167 .busy_placement
= &sys_placement_flags
171 struct ttm_tt dma_ttm
;
172 struct vmw_private
*dev_priv
;
177 struct vmw_sg_table vsgt
;
178 uint64_t sg_alloc_size
;
183 const size_t vmw_tt_size
= sizeof(struct vmw_ttm_tt
);
186 * __vmw_piter_non_sg_next: Helper functions to advance
187 * a struct vmw_piter iterator.
189 * @viter: Pointer to the iterator.
191 * These functions return false if past the end of the list,
192 * true otherwise. Functions are selected depending on the current
195 static bool __vmw_piter_non_sg_next(struct vmw_piter
*viter
)
197 return ++(viter
->i
) < viter
->num_pages
;
200 static bool __vmw_piter_sg_next(struct vmw_piter
*viter
)
202 bool ret
= __vmw_piter_non_sg_next(viter
);
204 return __sg_page_iter_dma_next(&viter
->iter
) && ret
;
208 static dma_addr_t
__vmw_piter_dma_addr(struct vmw_piter
*viter
)
210 return viter
->addrs
[viter
->i
];
213 static dma_addr_t
__vmw_piter_sg_addr(struct vmw_piter
*viter
)
215 return sg_page_iter_dma_address(&viter
->iter
);
220 * vmw_piter_start - Initialize a struct vmw_piter.
222 * @viter: Pointer to the iterator to initialize
223 * @vsgt: Pointer to a struct vmw_sg_table to initialize from
224 * @p_offset: Pointer offset used to update current array position
226 * Note that we're following the convention of __sg_page_iter_start, so that
227 * the iterator doesn't point to a valid page after initialization; it has
228 * to be advanced one step first.
230 void vmw_piter_start(struct vmw_piter
*viter
, const struct vmw_sg_table
*vsgt
,
231 unsigned long p_offset
)
233 viter
->i
= p_offset
- 1;
234 viter
->num_pages
= vsgt
->num_pages
;
235 viter
->pages
= vsgt
->pages
;
236 switch (vsgt
->mode
) {
237 case vmw_dma_alloc_coherent
:
238 viter
->next
= &__vmw_piter_non_sg_next
;
239 viter
->dma_address
= &__vmw_piter_dma_addr
;
240 viter
->addrs
= vsgt
->addrs
;
242 case vmw_dma_map_populate
:
243 case vmw_dma_map_bind
:
244 viter
->next
= &__vmw_piter_sg_next
;
245 viter
->dma_address
= &__vmw_piter_sg_addr
;
246 __sg_page_iter_start(&viter
->iter
.base
, vsgt
->sgt
->sgl
,
247 vsgt
->sgt
->orig_nents
, p_offset
);
255 * vmw_ttm_unmap_from_dma - unmap device addresses previsouly mapped for
258 * @vmw_tt: Pointer to a struct vmw_ttm_backend
260 * Used to free dma mappings previously mapped by vmw_ttm_map_for_dma.
262 static void vmw_ttm_unmap_from_dma(struct vmw_ttm_tt
*vmw_tt
)
264 struct device
*dev
= vmw_tt
->dev_priv
->drm
.dev
;
266 dma_unmap_sgtable(dev
, &vmw_tt
->sgt
, DMA_BIDIRECTIONAL
, 0);
267 vmw_tt
->sgt
.nents
= vmw_tt
->sgt
.orig_nents
;
271 * vmw_ttm_map_for_dma - map TTM pages to get device addresses
273 * @vmw_tt: Pointer to a struct vmw_ttm_backend
275 * This function is used to get device addresses from the kernel DMA layer.
276 * However, it's violating the DMA API in that when this operation has been
277 * performed, it's illegal for the CPU to write to the pages without first
278 * unmapping the DMA mappings, or calling dma_sync_sg_for_cpu(). It is
279 * therefore only legal to call this function if we know that the function
280 * dma_sync_sg_for_cpu() is a NOP, and dma_sync_sg_for_device() is at most
281 * a CPU write buffer flush.
283 static int vmw_ttm_map_for_dma(struct vmw_ttm_tt
*vmw_tt
)
285 struct device
*dev
= vmw_tt
->dev_priv
->drm
.dev
;
287 return dma_map_sgtable(dev
, &vmw_tt
->sgt
, DMA_BIDIRECTIONAL
, 0);
291 * vmw_ttm_map_dma - Make sure TTM pages are visible to the device
293 * @vmw_tt: Pointer to a struct vmw_ttm_tt
295 * Select the correct function for and make sure the TTM pages are
296 * visible to the device. Allocate storage for the device mappings.
297 * If a mapping has already been performed, indicated by the storage
298 * pointer being non NULL, the function returns success.
300 static int vmw_ttm_map_dma(struct vmw_ttm_tt
*vmw_tt
)
302 struct vmw_private
*dev_priv
= vmw_tt
->dev_priv
;
303 struct ttm_mem_global
*glob
= vmw_mem_glob(dev_priv
);
304 struct vmw_sg_table
*vsgt
= &vmw_tt
->vsgt
;
305 struct ttm_operation_ctx ctx
= {
306 .interruptible
= true,
309 struct vmw_piter iter
;
312 static size_t sgl_size
;
313 static size_t sgt_size
;
318 vsgt
->mode
= dev_priv
->map_mode
;
319 vsgt
->pages
= vmw_tt
->dma_ttm
.pages
;
320 vsgt
->num_pages
= vmw_tt
->dma_ttm
.num_pages
;
321 vsgt
->addrs
= vmw_tt
->dma_ttm
.dma_address
;
322 vsgt
->sgt
= &vmw_tt
->sgt
;
324 switch (dev_priv
->map_mode
) {
325 case vmw_dma_map_bind
:
326 case vmw_dma_map_populate
:
327 if (unlikely(!sgl_size
)) {
328 sgl_size
= ttm_round_pot(sizeof(struct scatterlist
));
329 sgt_size
= ttm_round_pot(sizeof(struct sg_table
));
331 vmw_tt
->sg_alloc_size
= sgt_size
+ sgl_size
* vsgt
->num_pages
;
332 ret
= ttm_mem_global_alloc(glob
, vmw_tt
->sg_alloc_size
, &ctx
);
333 if (unlikely(ret
!= 0))
336 ret
= sg_alloc_table_from_pages_segment(
337 &vmw_tt
->sgt
, vsgt
->pages
, vsgt
->num_pages
, 0,
338 (unsigned long)vsgt
->num_pages
<< PAGE_SHIFT
,
339 dma_get_max_seg_size(dev_priv
->drm
.dev
), GFP_KERNEL
);
341 goto out_sg_alloc_fail
;
343 if (vsgt
->num_pages
> vmw_tt
->sgt
.orig_nents
) {
344 uint64_t over_alloc
=
345 sgl_size
* (vsgt
->num_pages
-
346 vmw_tt
->sgt
.orig_nents
);
348 ttm_mem_global_free(glob
, over_alloc
);
349 vmw_tt
->sg_alloc_size
-= over_alloc
;
352 ret
= vmw_ttm_map_for_dma(vmw_tt
);
353 if (unlikely(ret
!= 0))
361 old
= ~((dma_addr_t
) 0);
362 vmw_tt
->vsgt
.num_regions
= 0;
363 for (vmw_piter_start(&iter
, vsgt
, 0); vmw_piter_next(&iter
);) {
364 dma_addr_t cur
= vmw_piter_dma_addr(&iter
);
366 if (cur
!= old
+ PAGE_SIZE
)
367 vmw_tt
->vsgt
.num_regions
++;
371 vmw_tt
->mapped
= true;
375 sg_free_table(vmw_tt
->vsgt
.sgt
);
376 vmw_tt
->vsgt
.sgt
= NULL
;
378 ttm_mem_global_free(glob
, vmw_tt
->sg_alloc_size
);
383 * vmw_ttm_unmap_dma - Tear down any TTM page device mappings
385 * @vmw_tt: Pointer to a struct vmw_ttm_tt
387 * Tear down any previously set up device DMA mappings and free
388 * any storage space allocated for them. If there are no mappings set up,
389 * this function is a NOP.
391 static void vmw_ttm_unmap_dma(struct vmw_ttm_tt
*vmw_tt
)
393 struct vmw_private
*dev_priv
= vmw_tt
->dev_priv
;
395 if (!vmw_tt
->vsgt
.sgt
)
398 switch (dev_priv
->map_mode
) {
399 case vmw_dma_map_bind
:
400 case vmw_dma_map_populate
:
401 vmw_ttm_unmap_from_dma(vmw_tt
);
402 sg_free_table(vmw_tt
->vsgt
.sgt
);
403 vmw_tt
->vsgt
.sgt
= NULL
;
404 ttm_mem_global_free(vmw_mem_glob(dev_priv
),
405 vmw_tt
->sg_alloc_size
);
410 vmw_tt
->mapped
= false;
414 * vmw_bo_sg_table - Return a struct vmw_sg_table object for a
417 * @bo: Pointer to a struct ttm_buffer_object
419 * Returns a pointer to a struct vmw_sg_table object. The object should
420 * not be freed after use.
421 * Note that for the device addresses to be valid, the buffer object must
422 * either be reserved or pinned.
424 const struct vmw_sg_table
*vmw_bo_sg_table(struct ttm_buffer_object
*bo
)
426 struct vmw_ttm_tt
*vmw_tt
=
427 container_of(bo
->ttm
, struct vmw_ttm_tt
, dma_ttm
);
429 return &vmw_tt
->vsgt
;
433 static int vmw_ttm_bind(struct ttm_device
*bdev
,
434 struct ttm_tt
*ttm
, struct ttm_resource
*bo_mem
)
436 struct vmw_ttm_tt
*vmw_be
=
437 container_of(ttm
, struct vmw_ttm_tt
, dma_ttm
);
446 ret
= vmw_ttm_map_dma(vmw_be
);
447 if (unlikely(ret
!= 0))
450 vmw_be
->gmr_id
= bo_mem
->start
;
451 vmw_be
->mem_type
= bo_mem
->mem_type
;
453 switch (bo_mem
->mem_type
) {
455 ret
= vmw_gmr_bind(vmw_be
->dev_priv
, &vmw_be
->vsgt
,
456 ttm
->num_pages
, vmw_be
->gmr_id
);
459 if (unlikely(vmw_be
->mob
== NULL
)) {
461 vmw_mob_create(ttm
->num_pages
);
462 if (unlikely(vmw_be
->mob
== NULL
))
466 ret
= vmw_mob_bind(vmw_be
->dev_priv
, vmw_be
->mob
,
467 &vmw_be
->vsgt
, ttm
->num_pages
,
471 /* Nothing to be done for a system bind */
476 vmw_be
->bound
= true;
480 static void vmw_ttm_unbind(struct ttm_device
*bdev
,
483 struct vmw_ttm_tt
*vmw_be
=
484 container_of(ttm
, struct vmw_ttm_tt
, dma_ttm
);
489 switch (vmw_be
->mem_type
) {
491 vmw_gmr_unbind(vmw_be
->dev_priv
, vmw_be
->gmr_id
);
494 vmw_mob_unbind(vmw_be
->dev_priv
, vmw_be
->mob
);
502 if (vmw_be
->dev_priv
->map_mode
== vmw_dma_map_bind
)
503 vmw_ttm_unmap_dma(vmw_be
);
504 vmw_be
->bound
= false;
508 static void vmw_ttm_destroy(struct ttm_device
*bdev
, struct ttm_tt
*ttm
)
510 struct vmw_ttm_tt
*vmw_be
=
511 container_of(ttm
, struct vmw_ttm_tt
, dma_ttm
);
513 vmw_ttm_unbind(bdev
, ttm
);
514 ttm_tt_destroy_common(bdev
, ttm
);
515 vmw_ttm_unmap_dma(vmw_be
);
516 if (vmw_be
->dev_priv
->map_mode
== vmw_dma_alloc_coherent
)
517 ttm_tt_fini(&vmw_be
->dma_ttm
);
522 vmw_mob_destroy(vmw_be
->mob
);
528 static int vmw_ttm_populate(struct ttm_device
*bdev
,
529 struct ttm_tt
*ttm
, struct ttm_operation_ctx
*ctx
)
534 /* TODO: maybe completely drop this ? */
535 if (ttm_tt_is_populated(ttm
))
538 ret
= ttm_pool_alloc(&bdev
->pool
, ttm
, ctx
);
542 for (i
= 0; i
< ttm
->num_pages
; ++i
) {
543 ret
= ttm_mem_global_alloc_page(&ttm_mem_glob
, ttm
->pages
[i
],
552 ttm_mem_global_free_page(&ttm_mem_glob
, ttm
->pages
[i
],
554 ttm_pool_free(&bdev
->pool
, ttm
);
558 static void vmw_ttm_unpopulate(struct ttm_device
*bdev
,
561 struct vmw_ttm_tt
*vmw_tt
= container_of(ttm
, struct vmw_ttm_tt
,
566 vmw_mob_destroy(vmw_tt
->mob
);
570 vmw_ttm_unmap_dma(vmw_tt
);
572 for (i
= 0; i
< ttm
->num_pages
; ++i
)
573 ttm_mem_global_free_page(&ttm_mem_glob
, ttm
->pages
[i
],
576 ttm_pool_free(&bdev
->pool
, ttm
);
579 static struct ttm_tt
*vmw_ttm_tt_create(struct ttm_buffer_object
*bo
,
582 struct vmw_ttm_tt
*vmw_be
;
585 vmw_be
= kzalloc(sizeof(*vmw_be
), GFP_KERNEL
);
589 vmw_be
->dev_priv
= container_of(bo
->bdev
, struct vmw_private
, bdev
);
592 if (vmw_be
->dev_priv
->map_mode
== vmw_dma_alloc_coherent
)
593 ret
= ttm_sg_tt_init(&vmw_be
->dma_ttm
, bo
, page_flags
,
596 ret
= ttm_tt_init(&vmw_be
->dma_ttm
, bo
, page_flags
,
598 if (unlikely(ret
!= 0))
601 return &vmw_be
->dma_ttm
;
607 static void vmw_evict_flags(struct ttm_buffer_object
*bo
,
608 struct ttm_placement
*placement
)
610 *placement
= vmw_sys_placement
;
613 static int vmw_ttm_io_mem_reserve(struct ttm_device
*bdev
, struct ttm_resource
*mem
)
615 struct vmw_private
*dev_priv
= container_of(bdev
, struct vmw_private
, bdev
);
617 switch (mem
->mem_type
) {
624 mem
->bus
.offset
= (mem
->start
<< PAGE_SHIFT
) +
625 dev_priv
->vram_start
;
626 mem
->bus
.is_iomem
= true;
627 mem
->bus
.caching
= ttm_cached
;
636 * vmw_move_notify - TTM move_notify_callback
638 * @bo: The TTM buffer object about to move.
639 * @old_mem: The old memory where we move from
640 * @new_mem: The struct ttm_resource indicating to what memory
641 * region the move is taking place.
643 * Calls move_notify for all subsystems needing it.
644 * (currently only resources).
646 static void vmw_move_notify(struct ttm_buffer_object
*bo
,
647 struct ttm_resource
*old_mem
,
648 struct ttm_resource
*new_mem
)
650 vmw_bo_move_notify(bo
, new_mem
);
651 vmw_query_move_notify(bo
, old_mem
, new_mem
);
656 * vmw_swap_notify - TTM move_notify_callback
658 * @bo: The TTM buffer object about to be swapped out.
660 static void vmw_swap_notify(struct ttm_buffer_object
*bo
)
662 vmw_bo_swap_notify(bo
);
663 (void) ttm_bo_wait(bo
, false, false);
666 static bool vmw_memtype_is_system(uint32_t mem_type
)
668 return mem_type
== TTM_PL_SYSTEM
|| mem_type
== VMW_PL_SYSTEM
;
671 static int vmw_move(struct ttm_buffer_object
*bo
,
673 struct ttm_operation_ctx
*ctx
,
674 struct ttm_resource
*new_mem
,
675 struct ttm_place
*hop
)
677 struct ttm_resource_manager
*old_man
= ttm_manager_type(bo
->bdev
, bo
->resource
->mem_type
);
678 struct ttm_resource_manager
*new_man
= ttm_manager_type(bo
->bdev
, new_mem
->mem_type
);
681 if (new_man
->use_tt
&& !vmw_memtype_is_system(new_mem
->mem_type
)) {
682 ret
= vmw_ttm_bind(bo
->bdev
, bo
->ttm
, new_mem
);
687 vmw_move_notify(bo
, bo
->resource
, new_mem
);
689 if (old_man
->use_tt
&& new_man
->use_tt
) {
690 if (vmw_memtype_is_system(bo
->resource
->mem_type
)) {
691 ttm_bo_move_null(bo
, new_mem
);
694 ret
= ttm_bo_wait_ctx(bo
, ctx
);
698 vmw_ttm_unbind(bo
->bdev
, bo
->ttm
);
699 ttm_resource_free(bo
, &bo
->resource
);
700 ttm_bo_assign_mem(bo
, new_mem
);
703 ret
= ttm_bo_move_memcpy(bo
, ctx
, new_mem
);
709 vmw_move_notify(bo
, new_mem
, bo
->resource
);
713 struct ttm_device_funcs vmw_bo_driver
= {
714 .ttm_tt_create
= &vmw_ttm_tt_create
,
715 .ttm_tt_populate
= &vmw_ttm_populate
,
716 .ttm_tt_unpopulate
= &vmw_ttm_unpopulate
,
717 .ttm_tt_destroy
= &vmw_ttm_destroy
,
718 .eviction_valuable
= ttm_bo_eviction_valuable
,
719 .evict_flags
= vmw_evict_flags
,
721 .swap_notify
= vmw_swap_notify
,
722 .io_mem_reserve
= &vmw_ttm_io_mem_reserve
,
725 int vmw_bo_create_and_populate(struct vmw_private
*dev_priv
,
726 unsigned long bo_size
,
727 struct ttm_buffer_object
**bo_p
)
729 struct ttm_operation_ctx ctx
= {
730 .interruptible
= false,
733 struct ttm_buffer_object
*bo
;
736 ret
= vmw_bo_create_kernel(dev_priv
, bo_size
,
737 &vmw_pt_sys_placement
,
739 if (unlikely(ret
!= 0))
742 ret
= ttm_bo_reserve(bo
, false, true, NULL
);
744 ret
= vmw_ttm_populate(bo
->bdev
, bo
->ttm
, &ctx
);
745 if (likely(ret
== 0)) {
746 struct vmw_ttm_tt
*vmw_tt
=
747 container_of(bo
->ttm
, struct vmw_ttm_tt
, dma_ttm
);
748 ret
= vmw_ttm_map_dma(vmw_tt
);
751 ttm_bo_unreserve(bo
);
753 if (likely(ret
== 0))