1 /**************************************************************************
3 * Copyright © 2009 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 uint32_t vram_placement_flags = TTM_PL_FLAG_VRAM |
36 static uint32_t vram_ne_placement_flags = TTM_PL_FLAG_VRAM |
40 static uint32_t sys_placement_flags = TTM_PL_FLAG_SYSTEM |
43 static uint32_t gmr_placement_flags = VMW_PL_FLAG_GMR |
46 static uint32_t gmr_ne_placement_flags = VMW_PL_FLAG_GMR |
50 struct ttm_placement vmw_vram_placement = {
54 .placement = &vram_placement_flags,
55 .num_busy_placement = 1,
56 .busy_placement = &vram_placement_flags
59 static uint32_t vram_gmr_placement_flags[] = {
60 TTM_PL_FLAG_VRAM | TTM_PL_FLAG_CACHED,
61 VMW_PL_FLAG_GMR | TTM_PL_FLAG_CACHED
64 static uint32_t gmr_vram_placement_flags[] = {
65 VMW_PL_FLAG_GMR | TTM_PL_FLAG_CACHED,
66 TTM_PL_FLAG_VRAM | TTM_PL_FLAG_CACHED
69 struct ttm_placement vmw_vram_gmr_placement = {
73 .placement = vram_gmr_placement_flags,
74 .num_busy_placement = 1,
75 .busy_placement = &gmr_placement_flags
78 static uint32_t vram_gmr_ne_placement_flags[] = {
79 TTM_PL_FLAG_VRAM | TTM_PL_FLAG_CACHED | TTM_PL_FLAG_NO_EVICT,
80 VMW_PL_FLAG_GMR | TTM_PL_FLAG_CACHED | TTM_PL_FLAG_NO_EVICT
83 struct ttm_placement vmw_vram_gmr_ne_placement = {
87 .placement = vram_gmr_ne_placement_flags,
88 .num_busy_placement = 1,
89 .busy_placement = &gmr_ne_placement_flags
92 struct ttm_placement vmw_vram_sys_placement = {
96 .placement = &vram_placement_flags,
97 .num_busy_placement = 1,
98 .busy_placement = &sys_placement_flags
101 struct ttm_placement vmw_vram_ne_placement = {
105 .placement = &vram_ne_placement_flags,
106 .num_busy_placement = 1,
107 .busy_placement = &vram_ne_placement_flags
110 struct ttm_placement vmw_sys_placement = {
114 .placement = &sys_placement_flags,
115 .num_busy_placement = 1,
116 .busy_placement = &sys_placement_flags
119 static uint32_t evictable_placement_flags[] = {
120 TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED,
121 TTM_PL_FLAG_VRAM | TTM_PL_FLAG_CACHED,
122 VMW_PL_FLAG_GMR | TTM_PL_FLAG_CACHED
125 struct ttm_placement vmw_evictable_placement = {
129 .placement = evictable_placement_flags,
130 .num_busy_placement = 1,
131 .busy_placement = &sys_placement_flags
134 struct ttm_placement vmw_srf_placement = {
138 .num_busy_placement = 2,
139 .placement = &gmr_placement_flags,
140 .busy_placement = gmr_vram_placement_flags
144 struct ttm_dma_tt dma_ttm;
145 struct vmw_private *dev_priv;
148 struct vmw_sg_table vsgt;
149 uint64_t sg_alloc_size;
153 const size_t vmw_tt_size = sizeof(struct vmw_ttm_tt);
156 * Helper functions to advance a struct vmw_piter iterator.
158 * @viter: Pointer to the iterator.
160 * These functions return false if past the end of the list,
161 * true otherwise. Functions are selected depending on the current
164 static bool __vmw_piter_non_sg_next(struct vmw_piter *viter)
166 return ++(viter->i) < viter->num_pages;
169 static bool __vmw_piter_sg_next(struct vmw_piter *viter)
171 return __sg_page_iter_next(&viter->iter);
176 * Helper functions to return a pointer to the current page.
178 * @viter: Pointer to the iterator
180 * These functions return a pointer to the page currently
181 * pointed to by @viter. Functions are selected depending on the
182 * current mapping mode.
184 static struct page *__vmw_piter_non_sg_page(struct vmw_piter *viter)
186 return viter->pages[viter->i];
189 static struct page *__vmw_piter_sg_page(struct vmw_piter *viter)
191 return sg_page_iter_page(&viter->iter);
196 * Helper functions to return the DMA address of the current page.
198 * @viter: Pointer to the iterator
200 * These functions return the DMA address of the page currently
201 * pointed to by @viter. Functions are selected depending on the
202 * current mapping mode.
204 static dma_addr_t __vmw_piter_phys_addr(struct vmw_piter *viter)
206 return page_to_phys(viter->pages[viter->i]);
209 static dma_addr_t __vmw_piter_dma_addr(struct vmw_piter *viter)
211 return viter->addrs[viter->i];
214 static dma_addr_t __vmw_piter_sg_addr(struct vmw_piter *viter)
216 return sg_page_iter_dma_address(&viter->iter);
221 * vmw_piter_start - Initialize a struct vmw_piter.
223 * @viter: Pointer to the iterator to initialize
224 * @vsgt: Pointer to a struct vmw_sg_table to initialize from
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 switch (vsgt->mode) {
237 viter->next = &__vmw_piter_non_sg_next;
238 viter->dma_address = &__vmw_piter_phys_addr;
239 viter->page = &__vmw_piter_non_sg_page;
240 viter->pages = vsgt->pages;
242 case vmw_dma_alloc_coherent:
243 viter->next = &__vmw_piter_non_sg_next;
244 viter->dma_address = &__vmw_piter_dma_addr;
245 viter->page = &__vmw_piter_non_sg_page;
246 viter->addrs = vsgt->addrs;
248 case vmw_dma_map_populate:
249 case vmw_dma_map_bind:
250 viter->next = &__vmw_piter_sg_next;
251 viter->dma_address = &__vmw_piter_sg_addr;
252 viter->page = &__vmw_piter_sg_page;
253 __sg_page_iter_start(&viter->iter, vsgt->sgt->sgl,
254 vsgt->sgt->orig_nents, p_offset);
262 * vmw_ttm_unmap_from_dma - unmap device addresses previsouly mapped for
265 * @vmw_tt: Pointer to a struct vmw_ttm_backend
267 * Used to free dma mappings previously mapped by vmw_ttm_map_for_dma.
269 static void vmw_ttm_unmap_from_dma(struct vmw_ttm_tt *vmw_tt)
271 struct device *dev = vmw_tt->dev_priv->dev->dev;
273 dma_unmap_sg(dev, vmw_tt->sgt.sgl, vmw_tt->sgt.nents,
275 vmw_tt->sgt.nents = vmw_tt->sgt.orig_nents;
279 * vmw_ttm_map_for_dma - map TTM pages to get device addresses
281 * @vmw_tt: Pointer to a struct vmw_ttm_backend
283 * This function is used to get device addresses from the kernel DMA layer.
284 * However, it's violating the DMA API in that when this operation has been
285 * performed, it's illegal for the CPU to write to the pages without first
286 * unmapping the DMA mappings, or calling dma_sync_sg_for_cpu(). It is
287 * therefore only legal to call this function if we know that the function
288 * dma_sync_sg_for_cpu() is a NOP, and dma_sync_sg_for_device() is at most
289 * a CPU write buffer flush.
291 static int vmw_ttm_map_for_dma(struct vmw_ttm_tt *vmw_tt)
293 struct device *dev = vmw_tt->dev_priv->dev->dev;
296 ret = dma_map_sg(dev, vmw_tt->sgt.sgl, vmw_tt->sgt.orig_nents,
298 if (unlikely(ret == 0))
301 vmw_tt->sgt.nents = ret;
307 * vmw_ttm_map_dma - Make sure TTM pages are visible to the device
309 * @vmw_tt: Pointer to a struct vmw_ttm_tt
311 * Select the correct function for and make sure the TTM pages are
312 * visible to the device. Allocate storage for the device mappings.
313 * If a mapping has already been performed, indicated by the storage
314 * pointer being non NULL, the function returns success.
316 static int vmw_ttm_map_dma(struct vmw_ttm_tt *vmw_tt)
318 struct vmw_private *dev_priv = vmw_tt->dev_priv;
319 struct ttm_mem_global *glob = vmw_mem_glob(dev_priv);
320 struct vmw_sg_table *vsgt = &vmw_tt->vsgt;
321 struct vmw_piter iter;
324 static size_t sgl_size;
325 static size_t sgt_size;
330 vsgt->mode = dev_priv->map_mode;
331 vsgt->pages = vmw_tt->dma_ttm.ttm.pages;
332 vsgt->num_pages = vmw_tt->dma_ttm.ttm.num_pages;
333 vsgt->addrs = vmw_tt->dma_ttm.dma_address;
334 vsgt->sgt = &vmw_tt->sgt;
336 switch (dev_priv->map_mode) {
337 case vmw_dma_map_bind:
338 case vmw_dma_map_populate:
339 if (unlikely(!sgl_size)) {
340 sgl_size = ttm_round_pot(sizeof(struct scatterlist));
341 sgt_size = ttm_round_pot(sizeof(struct sg_table));
343 vmw_tt->sg_alloc_size = sgt_size + sgl_size * vsgt->num_pages;
344 ret = ttm_mem_global_alloc(glob, vmw_tt->sg_alloc_size, false,
346 if (unlikely(ret != 0))
349 ret = sg_alloc_table_from_pages(&vmw_tt->sgt, vsgt->pages,
352 vsgt->num_pages << PAGE_SHIFT,
354 if (unlikely(ret != 0))
355 goto out_sg_alloc_fail;
357 if (vsgt->num_pages > vmw_tt->sgt.nents) {
358 uint64_t over_alloc =
359 sgl_size * (vsgt->num_pages -
362 ttm_mem_global_free(glob, over_alloc);
363 vmw_tt->sg_alloc_size -= over_alloc;
366 ret = vmw_ttm_map_for_dma(vmw_tt);
367 if (unlikely(ret != 0))
375 old = ~((dma_addr_t) 0);
376 vmw_tt->vsgt.num_regions = 0;
377 for (vmw_piter_start(&iter, vsgt, 0); vmw_piter_next(&iter);) {
378 dma_addr_t cur = vmw_piter_dma_addr(&iter);
380 if (cur != old + PAGE_SIZE)
381 vmw_tt->vsgt.num_regions++;
385 vmw_tt->mapped = true;
389 sg_free_table(vmw_tt->vsgt.sgt);
390 vmw_tt->vsgt.sgt = NULL;
392 ttm_mem_global_free(glob, vmw_tt->sg_alloc_size);
397 * vmw_ttm_unmap_dma - Tear down any TTM page device mappings
399 * @vmw_tt: Pointer to a struct vmw_ttm_tt
401 * Tear down any previously set up device DMA mappings and free
402 * any storage space allocated for them. If there are no mappings set up,
403 * this function is a NOP.
405 static void vmw_ttm_unmap_dma(struct vmw_ttm_tt *vmw_tt)
407 struct vmw_private *dev_priv = vmw_tt->dev_priv;
409 if (!vmw_tt->vsgt.sgt)
412 switch (dev_priv->map_mode) {
413 case vmw_dma_map_bind:
414 case vmw_dma_map_populate:
415 vmw_ttm_unmap_from_dma(vmw_tt);
416 sg_free_table(vmw_tt->vsgt.sgt);
417 vmw_tt->vsgt.sgt = NULL;
418 ttm_mem_global_free(vmw_mem_glob(dev_priv),
419 vmw_tt->sg_alloc_size);
424 vmw_tt->mapped = false;
427 static int vmw_ttm_bind(struct ttm_tt *ttm, struct ttm_mem_reg *bo_mem)
429 struct vmw_ttm_tt *vmw_be =
430 container_of(ttm, struct vmw_ttm_tt, dma_ttm.ttm);
433 ret = vmw_ttm_map_dma(vmw_be);
434 if (unlikely(ret != 0))
437 vmw_be->gmr_id = bo_mem->start;
439 return vmw_gmr_bind(vmw_be->dev_priv, &vmw_be->vsgt,
440 ttm->num_pages, vmw_be->gmr_id);
443 static int vmw_ttm_unbind(struct ttm_tt *ttm)
445 struct vmw_ttm_tt *vmw_be =
446 container_of(ttm, struct vmw_ttm_tt, dma_ttm.ttm);
448 vmw_gmr_unbind(vmw_be->dev_priv, vmw_be->gmr_id);
450 if (vmw_be->dev_priv->map_mode == vmw_dma_map_bind)
451 vmw_ttm_unmap_dma(vmw_be);
456 static void vmw_ttm_destroy(struct ttm_tt *ttm)
458 struct vmw_ttm_tt *vmw_be =
459 container_of(ttm, struct vmw_ttm_tt, dma_ttm.ttm);
461 vmw_ttm_unmap_dma(vmw_be);
462 if (vmw_be->dev_priv->map_mode == vmw_dma_alloc_coherent)
463 ttm_dma_tt_fini(&vmw_be->dma_ttm);
469 static int vmw_ttm_populate(struct ttm_tt *ttm)
471 struct vmw_ttm_tt *vmw_tt =
472 container_of(ttm, struct vmw_ttm_tt, dma_ttm.ttm);
473 struct vmw_private *dev_priv = vmw_tt->dev_priv;
474 struct ttm_mem_global *glob = vmw_mem_glob(dev_priv);
477 if (ttm->state != tt_unpopulated)
480 if (dev_priv->map_mode == vmw_dma_alloc_coherent) {
482 ttm_round_pot(ttm->num_pages * sizeof(dma_addr_t));
483 ret = ttm_mem_global_alloc(glob, size, false, true);
484 if (unlikely(ret != 0))
487 ret = ttm_dma_populate(&vmw_tt->dma_ttm, dev_priv->dev->dev);
488 if (unlikely(ret != 0))
489 ttm_mem_global_free(glob, size);
491 ret = ttm_pool_populate(ttm);
496 static void vmw_ttm_unpopulate(struct ttm_tt *ttm)
498 struct vmw_ttm_tt *vmw_tt = container_of(ttm, struct vmw_ttm_tt,
500 struct vmw_private *dev_priv = vmw_tt->dev_priv;
501 struct ttm_mem_global *glob = vmw_mem_glob(dev_priv);
503 vmw_ttm_unmap_dma(vmw_tt);
504 if (dev_priv->map_mode == vmw_dma_alloc_coherent) {
506 ttm_round_pot(ttm->num_pages * sizeof(dma_addr_t));
508 ttm_dma_unpopulate(&vmw_tt->dma_ttm, dev_priv->dev->dev);
509 ttm_mem_global_free(glob, size);
511 ttm_pool_unpopulate(ttm);
514 static struct ttm_backend_func vmw_ttm_func = {
515 .bind = vmw_ttm_bind,
516 .unbind = vmw_ttm_unbind,
517 .destroy = vmw_ttm_destroy,
520 struct ttm_tt *vmw_ttm_tt_create(struct ttm_bo_device *bdev,
521 unsigned long size, uint32_t page_flags,
522 struct page *dummy_read_page)
524 struct vmw_ttm_tt *vmw_be;
527 vmw_be = kzalloc(sizeof(*vmw_be), GFP_KERNEL);
531 vmw_be->dma_ttm.ttm.func = &vmw_ttm_func;
532 vmw_be->dev_priv = container_of(bdev, struct vmw_private, bdev);
534 if (vmw_be->dev_priv->map_mode == vmw_dma_alloc_coherent)
535 ret = ttm_dma_tt_init(&vmw_be->dma_ttm, bdev, size, page_flags,
538 ret = ttm_tt_init(&vmw_be->dma_ttm.ttm, bdev, size, page_flags,
540 if (unlikely(ret != 0))
543 return &vmw_be->dma_ttm.ttm;
549 int vmw_invalidate_caches(struct ttm_bo_device *bdev, uint32_t flags)
554 int vmw_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
555 struct ttm_mem_type_manager *man)
561 man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
562 man->available_caching = TTM_PL_FLAG_CACHED;
563 man->default_caching = TTM_PL_FLAG_CACHED;
566 /* "On-card" video ram */
567 man->func = &ttm_bo_manager_func;
569 man->flags = TTM_MEMTYPE_FLAG_FIXED | TTM_MEMTYPE_FLAG_MAPPABLE;
570 man->available_caching = TTM_PL_FLAG_CACHED;
571 man->default_caching = TTM_PL_FLAG_CACHED;
575 * "Guest Memory Regions" is an aperture like feature with
576 * one slot per bo. There is an upper limit of the number of
577 * slots as well as the bo size.
579 man->func = &vmw_gmrid_manager_func;
581 man->flags = TTM_MEMTYPE_FLAG_CMA | TTM_MEMTYPE_FLAG_MAPPABLE;
582 man->available_caching = TTM_PL_FLAG_CACHED;
583 man->default_caching = TTM_PL_FLAG_CACHED;
586 DRM_ERROR("Unsupported memory type %u\n", (unsigned)type);
592 void vmw_evict_flags(struct ttm_buffer_object *bo,
593 struct ttm_placement *placement)
595 *placement = vmw_sys_placement;
598 static int vmw_verify_access(struct ttm_buffer_object *bo, struct file *filp)
600 struct ttm_object_file *tfile =
601 vmw_fpriv((struct drm_file *)filp->private_data)->tfile;
603 return vmw_user_dmabuf_verify_access(bo, tfile);
606 static int vmw_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
608 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
609 struct vmw_private *dev_priv = container_of(bdev, struct vmw_private, bdev);
611 mem->bus.addr = NULL;
612 mem->bus.is_iomem = false;
614 mem->bus.size = mem->num_pages << PAGE_SHIFT;
616 if (!(man->flags & TTM_MEMTYPE_FLAG_MAPPABLE))
618 switch (mem->mem_type) {
623 mem->bus.offset = mem->start << PAGE_SHIFT;
624 mem->bus.base = dev_priv->vram_start;
625 mem->bus.is_iomem = true;
633 static void vmw_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
637 static int vmw_ttm_fault_reserve_notify(struct ttm_buffer_object *bo)
643 * FIXME: We're using the old vmware polling method to sync.
644 * Do this with fences instead.
647 static void *vmw_sync_obj_ref(void *sync_obj)
651 vmw_fence_obj_reference((struct vmw_fence_obj *) sync_obj);
654 static void vmw_sync_obj_unref(void **sync_obj)
656 vmw_fence_obj_unreference((struct vmw_fence_obj **) sync_obj);
659 static int vmw_sync_obj_flush(void *sync_obj)
661 vmw_fence_obj_flush((struct vmw_fence_obj *) sync_obj);
665 static bool vmw_sync_obj_signaled(void *sync_obj)
667 return vmw_fence_obj_signaled((struct vmw_fence_obj *) sync_obj,
668 DRM_VMW_FENCE_FLAG_EXEC);
672 static int vmw_sync_obj_wait(void *sync_obj, bool lazy, bool interruptible)
674 return vmw_fence_obj_wait((struct vmw_fence_obj *) sync_obj,
675 DRM_VMW_FENCE_FLAG_EXEC,
677 VMW_FENCE_WAIT_TIMEOUT);
680 struct ttm_bo_driver vmw_bo_driver = {
681 .ttm_tt_create = &vmw_ttm_tt_create,
682 .ttm_tt_populate = &vmw_ttm_populate,
683 .ttm_tt_unpopulate = &vmw_ttm_unpopulate,
684 .invalidate_caches = vmw_invalidate_caches,
685 .init_mem_type = vmw_init_mem_type,
686 .evict_flags = vmw_evict_flags,
688 .verify_access = vmw_verify_access,
689 .sync_obj_signaled = vmw_sync_obj_signaled,
690 .sync_obj_wait = vmw_sync_obj_wait,
691 .sync_obj_flush = vmw_sync_obj_flush,
692 .sync_obj_unref = vmw_sync_obj_unref,
693 .sync_obj_ref = vmw_sync_obj_ref,
696 .fault_reserve_notify = &vmw_ttm_fault_reserve_notify,
697 .io_mem_reserve = &vmw_ttm_io_mem_reserve,
698 .io_mem_free = &vmw_ttm_io_mem_free,