Linux Audio

Check our new training course

Loading...
v4.6
  1/**************************************************************************
  2 *
  3 * Copyright © 2009-2015 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#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>
 32
 33static struct ttm_place vram_placement_flags = {
 34	.fpfn = 0,
 35	.lpfn = 0,
 36	.flags = TTM_PL_FLAG_VRAM | TTM_PL_FLAG_CACHED
 37};
 38
 39static struct ttm_place vram_ne_placement_flags = {
 40	.fpfn = 0,
 41	.lpfn = 0,
 42	.flags = TTM_PL_FLAG_VRAM | TTM_PL_FLAG_CACHED | TTM_PL_FLAG_NO_EVICT
 43};
 44
 45static struct ttm_place sys_placement_flags = {
 46	.fpfn = 0,
 47	.lpfn = 0,
 48	.flags = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED
 49};
 50
 51static struct ttm_place sys_ne_placement_flags = {
 52	.fpfn = 0,
 53	.lpfn = 0,
 54	.flags = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED | TTM_PL_FLAG_NO_EVICT
 55};
 56
 57static struct ttm_place gmr_placement_flags = {
 58	.fpfn = 0,
 59	.lpfn = 0,
 60	.flags = VMW_PL_FLAG_GMR | TTM_PL_FLAG_CACHED
 61};
 62
 63static struct ttm_place gmr_ne_placement_flags = {
 64	.fpfn = 0,
 65	.lpfn = 0,
 66	.flags = VMW_PL_FLAG_GMR | TTM_PL_FLAG_CACHED | TTM_PL_FLAG_NO_EVICT
 67};
 68
 69static struct ttm_place mob_placement_flags = {
 70	.fpfn = 0,
 71	.lpfn = 0,
 72	.flags = VMW_PL_FLAG_MOB | TTM_PL_FLAG_CACHED
 73};
 74
 75static struct ttm_place mob_ne_placement_flags = {
 76	.fpfn = 0,
 77	.lpfn = 0,
 78	.flags = VMW_PL_FLAG_MOB | TTM_PL_FLAG_CACHED | TTM_PL_FLAG_NO_EVICT
 79};
 80
 81struct ttm_placement vmw_vram_placement = {
 82	.num_placement = 1,
 83	.placement = &vram_placement_flags,
 84	.num_busy_placement = 1,
 85	.busy_placement = &vram_placement_flags
 86};
 87
 88static struct ttm_place vram_gmr_placement_flags[] = {
 89	{
 90		.fpfn = 0,
 91		.lpfn = 0,
 92		.flags = TTM_PL_FLAG_VRAM | TTM_PL_FLAG_CACHED
 93	}, {
 94		.fpfn = 0,
 95		.lpfn = 0,
 96		.flags = VMW_PL_FLAG_GMR | TTM_PL_FLAG_CACHED
 97	}
 98};
 99
100static struct ttm_place gmr_vram_placement_flags[] = {
101	{
102		.fpfn = 0,
103		.lpfn = 0,
104		.flags = VMW_PL_FLAG_GMR | TTM_PL_FLAG_CACHED
105	}, {
106		.fpfn = 0,
107		.lpfn = 0,
108		.flags = TTM_PL_FLAG_VRAM | TTM_PL_FLAG_CACHED
109	}
110};
111
112struct ttm_placement vmw_vram_gmr_placement = {
 
 
113	.num_placement = 2,
114	.placement = vram_gmr_placement_flags,
115	.num_busy_placement = 1,
116	.busy_placement = &gmr_placement_flags
117};
118
119static struct ttm_place vram_gmr_ne_placement_flags[] = {
120	{
121		.fpfn = 0,
122		.lpfn = 0,
123		.flags = TTM_PL_FLAG_VRAM | TTM_PL_FLAG_CACHED |
124			 TTM_PL_FLAG_NO_EVICT
125	}, {
126		.fpfn = 0,
127		.lpfn = 0,
128		.flags = VMW_PL_FLAG_GMR | TTM_PL_FLAG_CACHED |
129			 TTM_PL_FLAG_NO_EVICT
130	}
131};
132
133struct ttm_placement vmw_vram_gmr_ne_placement = {
 
 
134	.num_placement = 2,
135	.placement = vram_gmr_ne_placement_flags,
136	.num_busy_placement = 1,
137	.busy_placement = &gmr_ne_placement_flags
138};
139
140struct ttm_placement vmw_vram_sys_placement = {
 
 
141	.num_placement = 1,
142	.placement = &vram_placement_flags,
143	.num_busy_placement = 1,
144	.busy_placement = &sys_placement_flags
145};
146
147struct ttm_placement vmw_vram_ne_placement = {
 
 
148	.num_placement = 1,
149	.placement = &vram_ne_placement_flags,
150	.num_busy_placement = 1,
151	.busy_placement = &vram_ne_placement_flags
152};
153
154struct ttm_placement vmw_sys_placement = {
 
 
155	.num_placement = 1,
156	.placement = &sys_placement_flags,
157	.num_busy_placement = 1,
158	.busy_placement = &sys_placement_flags
159};
160
161struct ttm_placement vmw_sys_ne_placement = {
 
 
162	.num_placement = 1,
163	.placement = &sys_ne_placement_flags,
164	.num_busy_placement = 1,
165	.busy_placement = &sys_ne_placement_flags
166};
167
168static struct ttm_place evictable_placement_flags[] = {
169	{
170		.fpfn = 0,
171		.lpfn = 0,
172		.flags = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED
173	}, {
174		.fpfn = 0,
175		.lpfn = 0,
176		.flags = TTM_PL_FLAG_VRAM | TTM_PL_FLAG_CACHED
177	}, {
178		.fpfn = 0,
179		.lpfn = 0,
180		.flags = VMW_PL_FLAG_GMR | TTM_PL_FLAG_CACHED
181	}, {
182		.fpfn = 0,
183		.lpfn = 0,
184		.flags = VMW_PL_FLAG_MOB | TTM_PL_FLAG_CACHED
185	}
186};
187
188struct ttm_placement vmw_evictable_placement = {
 
 
189	.num_placement = 4,
190	.placement = evictable_placement_flags,
191	.num_busy_placement = 1,
192	.busy_placement = &sys_placement_flags
193};
194
195struct ttm_placement vmw_srf_placement = {
 
 
196	.num_placement = 1,
197	.num_busy_placement = 2,
198	.placement = &gmr_placement_flags,
199	.busy_placement = gmr_vram_placement_flags
200};
201
202struct ttm_placement vmw_mob_placement = {
 
 
203	.num_placement = 1,
204	.num_busy_placement = 1,
205	.placement = &mob_placement_flags,
206	.busy_placement = &mob_placement_flags
207};
208
209struct ttm_placement vmw_mob_ne_placement = {
210	.num_placement = 1,
211	.num_busy_placement = 1,
212	.placement = &mob_ne_placement_flags,
213	.busy_placement = &mob_ne_placement_flags
214};
215
216struct vmw_ttm_tt {
217	struct ttm_dma_tt dma_ttm;
218	struct vmw_private *dev_priv;
219	int gmr_id;
220	struct vmw_mob *mob;
221	int mem_type;
222	struct sg_table sgt;
223	struct vmw_sg_table vsgt;
224	uint64_t sg_alloc_size;
225	bool mapped;
226};
227
228const size_t vmw_tt_size = sizeof(struct vmw_ttm_tt);
229
230/**
231 * Helper functions to advance a struct vmw_piter iterator.
232 *
233 * @viter: Pointer to the iterator.
234 *
235 * These functions return false if past the end of the list,
236 * true otherwise. Functions are selected depending on the current
237 * DMA mapping mode.
238 */
239static bool __vmw_piter_non_sg_next(struct vmw_piter *viter)
240{
241	return ++(viter->i) < viter->num_pages;
242}
243
244static bool __vmw_piter_sg_next(struct vmw_piter *viter)
245{
246	return __sg_page_iter_next(&viter->iter);
247}
248
249
250/**
251 * Helper functions to return a pointer to the current page.
252 *
253 * @viter: Pointer to the iterator
254 *
255 * These functions return a pointer to the page currently
256 * pointed to by @viter. Functions are selected depending on the
257 * current mapping mode.
258 */
259static struct page *__vmw_piter_non_sg_page(struct vmw_piter *viter)
260{
261	return viter->pages[viter->i];
262}
263
264static struct page *__vmw_piter_sg_page(struct vmw_piter *viter)
265{
266	return sg_page_iter_page(&viter->iter);
267}
268
269
270/**
271 * Helper functions to return the DMA address of the current page.
272 *
273 * @viter: Pointer to the iterator
274 *
275 * These functions return the DMA address of the page currently
276 * pointed to by @viter. Functions are selected depending on the
277 * current mapping mode.
278 */
279static dma_addr_t __vmw_piter_phys_addr(struct vmw_piter *viter)
280{
281	return page_to_phys(viter->pages[viter->i]);
282}
283
284static dma_addr_t __vmw_piter_dma_addr(struct vmw_piter *viter)
285{
286	return viter->addrs[viter->i];
287}
288
289static dma_addr_t __vmw_piter_sg_addr(struct vmw_piter *viter)
290{
291	return sg_page_iter_dma_address(&viter->iter);
292}
293
294
295/**
296 * vmw_piter_start - Initialize a struct vmw_piter.
297 *
298 * @viter: Pointer to the iterator to initialize
299 * @vsgt: Pointer to a struct vmw_sg_table to initialize from
300 *
301 * Note that we're following the convention of __sg_page_iter_start, so that
302 * the iterator doesn't point to a valid page after initialization; it has
303 * to be advanced one step first.
304 */
305void vmw_piter_start(struct vmw_piter *viter, const struct vmw_sg_table *vsgt,
306		     unsigned long p_offset)
307{
308	viter->i = p_offset - 1;
309	viter->num_pages = vsgt->num_pages;
310	switch (vsgt->mode) {
311	case vmw_dma_phys:
312		viter->next = &__vmw_piter_non_sg_next;
313		viter->dma_address = &__vmw_piter_phys_addr;
314		viter->page = &__vmw_piter_non_sg_page;
315		viter->pages = vsgt->pages;
316		break;
317	case vmw_dma_alloc_coherent:
318		viter->next = &__vmw_piter_non_sg_next;
319		viter->dma_address = &__vmw_piter_dma_addr;
320		viter->page = &__vmw_piter_non_sg_page;
321		viter->addrs = vsgt->addrs;
322		viter->pages = vsgt->pages;
323		break;
324	case vmw_dma_map_populate:
325	case vmw_dma_map_bind:
326		viter->next = &__vmw_piter_sg_next;
327		viter->dma_address = &__vmw_piter_sg_addr;
328		viter->page = &__vmw_piter_sg_page;
329		__sg_page_iter_start(&viter->iter, vsgt->sgt->sgl,
330				     vsgt->sgt->orig_nents, p_offset);
331		break;
332	default:
333		BUG();
334	}
335}
336
337/**
338 * vmw_ttm_unmap_from_dma - unmap  device addresses previsouly mapped for
339 * TTM pages
340 *
341 * @vmw_tt: Pointer to a struct vmw_ttm_backend
342 *
343 * Used to free dma mappings previously mapped by vmw_ttm_map_for_dma.
344 */
345static void vmw_ttm_unmap_from_dma(struct vmw_ttm_tt *vmw_tt)
346{
347	struct device *dev = vmw_tt->dev_priv->dev->dev;
348
349	dma_unmap_sg(dev, vmw_tt->sgt.sgl, vmw_tt->sgt.nents,
350		DMA_BIDIRECTIONAL);
351	vmw_tt->sgt.nents = vmw_tt->sgt.orig_nents;
352}
353
354/**
355 * vmw_ttm_map_for_dma - map TTM pages to get device addresses
356 *
357 * @vmw_tt: Pointer to a struct vmw_ttm_backend
358 *
359 * This function is used to get device addresses from the kernel DMA layer.
360 * However, it's violating the DMA API in that when this operation has been
361 * performed, it's illegal for the CPU to write to the pages without first
362 * unmapping the DMA mappings, or calling dma_sync_sg_for_cpu(). It is
363 * therefore only legal to call this function if we know that the function
364 * dma_sync_sg_for_cpu() is a NOP, and dma_sync_sg_for_device() is at most
365 * a CPU write buffer flush.
366 */
367static int vmw_ttm_map_for_dma(struct vmw_ttm_tt *vmw_tt)
368{
369	struct device *dev = vmw_tt->dev_priv->dev->dev;
370	int ret;
371
372	ret = dma_map_sg(dev, vmw_tt->sgt.sgl, vmw_tt->sgt.orig_nents,
373			 DMA_BIDIRECTIONAL);
374	if (unlikely(ret == 0))
375		return -ENOMEM;
376
377	vmw_tt->sgt.nents = ret;
378
379	return 0;
380}
381
382/**
383 * vmw_ttm_map_dma - Make sure TTM pages are visible to the device
384 *
385 * @vmw_tt: Pointer to a struct vmw_ttm_tt
386 *
387 * Select the correct function for and make sure the TTM pages are
388 * visible to the device. Allocate storage for the device mappings.
389 * If a mapping has already been performed, indicated by the storage
390 * pointer being non NULL, the function returns success.
391 */
392static int vmw_ttm_map_dma(struct vmw_ttm_tt *vmw_tt)
393{
394	struct vmw_private *dev_priv = vmw_tt->dev_priv;
395	struct ttm_mem_global *glob = vmw_mem_glob(dev_priv);
396	struct vmw_sg_table *vsgt = &vmw_tt->vsgt;
397	struct vmw_piter iter;
398	dma_addr_t old;
399	int ret = 0;
400	static size_t sgl_size;
401	static size_t sgt_size;
402
403	if (vmw_tt->mapped)
404		return 0;
405
406	vsgt->mode = dev_priv->map_mode;
407	vsgt->pages = vmw_tt->dma_ttm.ttm.pages;
408	vsgt->num_pages = vmw_tt->dma_ttm.ttm.num_pages;
409	vsgt->addrs = vmw_tt->dma_ttm.dma_address;
410	vsgt->sgt = &vmw_tt->sgt;
411
412	switch (dev_priv->map_mode) {
413	case vmw_dma_map_bind:
414	case vmw_dma_map_populate:
415		if (unlikely(!sgl_size)) {
416			sgl_size = ttm_round_pot(sizeof(struct scatterlist));
417			sgt_size = ttm_round_pot(sizeof(struct sg_table));
418		}
419		vmw_tt->sg_alloc_size = sgt_size + sgl_size * vsgt->num_pages;
420		ret = ttm_mem_global_alloc(glob, vmw_tt->sg_alloc_size, false,
421					   true);
422		if (unlikely(ret != 0))
423			return ret;
424
425		ret = sg_alloc_table_from_pages(&vmw_tt->sgt, vsgt->pages,
426						vsgt->num_pages, 0,
427						(unsigned long)
428						vsgt->num_pages << PAGE_SHIFT,
429						GFP_KERNEL);
430		if (unlikely(ret != 0))
431			goto out_sg_alloc_fail;
432
433		if (vsgt->num_pages > vmw_tt->sgt.nents) {
434			uint64_t over_alloc =
435				sgl_size * (vsgt->num_pages -
436					    vmw_tt->sgt.nents);
437
438			ttm_mem_global_free(glob, over_alloc);
439			vmw_tt->sg_alloc_size -= over_alloc;
440		}
441
442		ret = vmw_ttm_map_for_dma(vmw_tt);
443		if (unlikely(ret != 0))
444			goto out_map_fail;
445
446		break;
447	default:
448		break;
449	}
450
451	old = ~((dma_addr_t) 0);
452	vmw_tt->vsgt.num_regions = 0;
453	for (vmw_piter_start(&iter, vsgt, 0); vmw_piter_next(&iter);) {
454		dma_addr_t cur = vmw_piter_dma_addr(&iter);
455
456		if (cur != old + PAGE_SIZE)
457			vmw_tt->vsgt.num_regions++;
458		old = cur;
459	}
460
461	vmw_tt->mapped = true;
462	return 0;
463
464out_map_fail:
465	sg_free_table(vmw_tt->vsgt.sgt);
466	vmw_tt->vsgt.sgt = NULL;
467out_sg_alloc_fail:
468	ttm_mem_global_free(glob, vmw_tt->sg_alloc_size);
469	return ret;
470}
471
472/**
473 * vmw_ttm_unmap_dma - Tear down any TTM page device mappings
474 *
475 * @vmw_tt: Pointer to a struct vmw_ttm_tt
476 *
477 * Tear down any previously set up device DMA mappings and free
478 * any storage space allocated for them. If there are no mappings set up,
479 * this function is a NOP.
480 */
481static void vmw_ttm_unmap_dma(struct vmw_ttm_tt *vmw_tt)
482{
483	struct vmw_private *dev_priv = vmw_tt->dev_priv;
484
485	if (!vmw_tt->vsgt.sgt)
486		return;
487
488	switch (dev_priv->map_mode) {
489	case vmw_dma_map_bind:
490	case vmw_dma_map_populate:
491		vmw_ttm_unmap_from_dma(vmw_tt);
492		sg_free_table(vmw_tt->vsgt.sgt);
493		vmw_tt->vsgt.sgt = NULL;
494		ttm_mem_global_free(vmw_mem_glob(dev_priv),
495				    vmw_tt->sg_alloc_size);
496		break;
497	default:
498		break;
499	}
500	vmw_tt->mapped = false;
501}
502
503
504/**
505 * vmw_bo_map_dma - Make sure buffer object pages are visible to the device
506 *
507 * @bo: Pointer to a struct ttm_buffer_object
508 *
509 * Wrapper around vmw_ttm_map_dma, that takes a TTM buffer object pointer
510 * instead of a pointer to a struct vmw_ttm_backend as argument.
511 * Note that the buffer object must be either pinned or reserved before
512 * calling this function.
513 */
514int vmw_bo_map_dma(struct ttm_buffer_object *bo)
515{
516	struct vmw_ttm_tt *vmw_tt =
517		container_of(bo->ttm, struct vmw_ttm_tt, dma_ttm.ttm);
518
519	return vmw_ttm_map_dma(vmw_tt);
520}
521
522
523/**
524 * vmw_bo_unmap_dma - Make sure buffer object pages are visible to the device
525 *
526 * @bo: Pointer to a struct ttm_buffer_object
527 *
528 * Wrapper around vmw_ttm_unmap_dma, that takes a TTM buffer object pointer
529 * instead of a pointer to a struct vmw_ttm_backend as argument.
530 */
531void vmw_bo_unmap_dma(struct ttm_buffer_object *bo)
532{
533	struct vmw_ttm_tt *vmw_tt =
534		container_of(bo->ttm, struct vmw_ttm_tt, dma_ttm.ttm);
535
536	vmw_ttm_unmap_dma(vmw_tt);
537}
538
539
540/**
541 * vmw_bo_sg_table - Return a struct vmw_sg_table object for a
542 * TTM buffer object
543 *
544 * @bo: Pointer to a struct ttm_buffer_object
545 *
546 * Returns a pointer to a struct vmw_sg_table object. The object should
547 * not be freed after use.
548 * Note that for the device addresses to be valid, the buffer object must
549 * either be reserved or pinned.
550 */
551const struct vmw_sg_table *vmw_bo_sg_table(struct ttm_buffer_object *bo)
552{
553	struct vmw_ttm_tt *vmw_tt =
554		container_of(bo->ttm, struct vmw_ttm_tt, dma_ttm.ttm);
555
556	return &vmw_tt->vsgt;
557}
558
559
560static int vmw_ttm_bind(struct ttm_tt *ttm, struct ttm_mem_reg *bo_mem)
561{
562	struct vmw_ttm_tt *vmw_be =
563		container_of(ttm, struct vmw_ttm_tt, dma_ttm.ttm);
564	int ret;
565
566	ret = vmw_ttm_map_dma(vmw_be);
567	if (unlikely(ret != 0))
568		return ret;
569
570	vmw_be->gmr_id = bo_mem->start;
571	vmw_be->mem_type = bo_mem->mem_type;
572
573	switch (bo_mem->mem_type) {
574	case VMW_PL_GMR:
575		return vmw_gmr_bind(vmw_be->dev_priv, &vmw_be->vsgt,
576				    ttm->num_pages, vmw_be->gmr_id);
577	case VMW_PL_MOB:
578		if (unlikely(vmw_be->mob == NULL)) {
579			vmw_be->mob =
580				vmw_mob_create(ttm->num_pages);
581			if (unlikely(vmw_be->mob == NULL))
582				return -ENOMEM;
583		}
584
585		return vmw_mob_bind(vmw_be->dev_priv, vmw_be->mob,
586				    &vmw_be->vsgt, ttm->num_pages,
587				    vmw_be->gmr_id);
588	default:
589		BUG();
590	}
591	return 0;
592}
593
594static int vmw_ttm_unbind(struct ttm_tt *ttm)
595{
596	struct vmw_ttm_tt *vmw_be =
597		container_of(ttm, struct vmw_ttm_tt, dma_ttm.ttm);
598
599	switch (vmw_be->mem_type) {
600	case VMW_PL_GMR:
601		vmw_gmr_unbind(vmw_be->dev_priv, vmw_be->gmr_id);
602		break;
603	case VMW_PL_MOB:
604		vmw_mob_unbind(vmw_be->dev_priv, vmw_be->mob);
605		break;
606	default:
607		BUG();
608	}
609
610	if (vmw_be->dev_priv->map_mode == vmw_dma_map_bind)
611		vmw_ttm_unmap_dma(vmw_be);
612
613	return 0;
614}
615
616
617static void vmw_ttm_destroy(struct ttm_tt *ttm)
618{
619	struct vmw_ttm_tt *vmw_be =
620		container_of(ttm, struct vmw_ttm_tt, dma_ttm.ttm);
621
622	vmw_ttm_unmap_dma(vmw_be);
623	if (vmw_be->dev_priv->map_mode == vmw_dma_alloc_coherent)
624		ttm_dma_tt_fini(&vmw_be->dma_ttm);
625	else
626		ttm_tt_fini(ttm);
627
628	if (vmw_be->mob)
629		vmw_mob_destroy(vmw_be->mob);
630
631	kfree(vmw_be);
632}
633
634
635static int vmw_ttm_populate(struct ttm_tt *ttm)
636{
637	struct vmw_ttm_tt *vmw_tt =
638		container_of(ttm, struct vmw_ttm_tt, dma_ttm.ttm);
639	struct vmw_private *dev_priv = vmw_tt->dev_priv;
640	struct ttm_mem_global *glob = vmw_mem_glob(dev_priv);
641	int ret;
642
643	if (ttm->state != tt_unpopulated)
644		return 0;
645
646	if (dev_priv->map_mode == vmw_dma_alloc_coherent) {
647		size_t size =
648			ttm_round_pot(ttm->num_pages * sizeof(dma_addr_t));
649		ret = ttm_mem_global_alloc(glob, size, false, true);
650		if (unlikely(ret != 0))
651			return ret;
652
653		ret = ttm_dma_populate(&vmw_tt->dma_ttm, dev_priv->dev->dev);
654		if (unlikely(ret != 0))
655			ttm_mem_global_free(glob, size);
656	} else
657		ret = ttm_pool_populate(ttm);
658
659	return ret;
660}
661
662static void vmw_ttm_unpopulate(struct ttm_tt *ttm)
663{
664	struct vmw_ttm_tt *vmw_tt = container_of(ttm, struct vmw_ttm_tt,
665						 dma_ttm.ttm);
666	struct vmw_private *dev_priv = vmw_tt->dev_priv;
667	struct ttm_mem_global *glob = vmw_mem_glob(dev_priv);
668
669
670	if (vmw_tt->mob) {
671		vmw_mob_destroy(vmw_tt->mob);
672		vmw_tt->mob = NULL;
673	}
674
675	vmw_ttm_unmap_dma(vmw_tt);
676	if (dev_priv->map_mode == vmw_dma_alloc_coherent) {
677		size_t size =
678			ttm_round_pot(ttm->num_pages * sizeof(dma_addr_t));
679
680		ttm_dma_unpopulate(&vmw_tt->dma_ttm, dev_priv->dev->dev);
681		ttm_mem_global_free(glob, size);
682	} else
683		ttm_pool_unpopulate(ttm);
684}
685
686static struct ttm_backend_func vmw_ttm_func = {
687	.bind = vmw_ttm_bind,
688	.unbind = vmw_ttm_unbind,
689	.destroy = vmw_ttm_destroy,
690};
691
692static struct ttm_tt *vmw_ttm_tt_create(struct ttm_bo_device *bdev,
693				 unsigned long size, uint32_t page_flags,
694				 struct page *dummy_read_page)
695{
696	struct vmw_ttm_tt *vmw_be;
697	int ret;
698
699	vmw_be = kzalloc(sizeof(*vmw_be), GFP_KERNEL);
700	if (!vmw_be)
701		return NULL;
702
703	vmw_be->dma_ttm.ttm.func = &vmw_ttm_func;
704	vmw_be->dev_priv = container_of(bdev, struct vmw_private, bdev);
705	vmw_be->mob = NULL;
706
707	if (vmw_be->dev_priv->map_mode == vmw_dma_alloc_coherent)
708		ret = ttm_dma_tt_init(&vmw_be->dma_ttm, bdev, size, page_flags,
709				      dummy_read_page);
710	else
711		ret = ttm_tt_init(&vmw_be->dma_ttm.ttm, bdev, size, page_flags,
712				  dummy_read_page);
713	if (unlikely(ret != 0))
714		goto out_no_init;
715
716	return &vmw_be->dma_ttm.ttm;
717out_no_init:
718	kfree(vmw_be);
719	return NULL;
720}
721
722static int vmw_invalidate_caches(struct ttm_bo_device *bdev, uint32_t flags)
723{
724	return 0;
725}
726
727static int vmw_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
728		      struct ttm_mem_type_manager *man)
729{
730	switch (type) {
731	case TTM_PL_SYSTEM:
732		/* System memory */
733
734		man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
735		man->available_caching = TTM_PL_FLAG_CACHED;
736		man->default_caching = TTM_PL_FLAG_CACHED;
737		break;
738	case TTM_PL_VRAM:
739		/* "On-card" video ram */
740		man->func = &ttm_bo_manager_func;
741		man->gpu_offset = 0;
742		man->flags = TTM_MEMTYPE_FLAG_FIXED | TTM_MEMTYPE_FLAG_MAPPABLE;
743		man->available_caching = TTM_PL_FLAG_CACHED;
744		man->default_caching = TTM_PL_FLAG_CACHED;
745		break;
746	case VMW_PL_GMR:
747	case VMW_PL_MOB:
748		/*
749		 * "Guest Memory Regions" is an aperture like feature with
750		 *  one slot per bo. There is an upper limit of the number of
751		 *  slots as well as the bo size.
752		 */
753		man->func = &vmw_gmrid_manager_func;
754		man->gpu_offset = 0;
755		man->flags = TTM_MEMTYPE_FLAG_CMA | TTM_MEMTYPE_FLAG_MAPPABLE;
756		man->available_caching = TTM_PL_FLAG_CACHED;
757		man->default_caching = TTM_PL_FLAG_CACHED;
758		break;
759	default:
760		DRM_ERROR("Unsupported memory type %u\n", (unsigned)type);
761		return -EINVAL;
762	}
763	return 0;
764}
765
766static void vmw_evict_flags(struct ttm_buffer_object *bo,
767		     struct ttm_placement *placement)
768{
769	*placement = vmw_sys_placement;
770}
771
772static int vmw_verify_access(struct ttm_buffer_object *bo, struct file *filp)
773{
774	struct ttm_object_file *tfile =
775		vmw_fpriv((struct drm_file *)filp->private_data)->tfile;
776
777	return vmw_user_dmabuf_verify_access(bo, tfile);
778}
779
780static int vmw_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
781{
782	struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
783	struct vmw_private *dev_priv = container_of(bdev, struct vmw_private, bdev);
784
785	mem->bus.addr = NULL;
786	mem->bus.is_iomem = false;
787	mem->bus.offset = 0;
788	mem->bus.size = mem->num_pages << PAGE_SHIFT;
789	mem->bus.base = 0;
790	if (!(man->flags & TTM_MEMTYPE_FLAG_MAPPABLE))
791		return -EINVAL;
792	switch (mem->mem_type) {
793	case TTM_PL_SYSTEM:
794	case VMW_PL_GMR:
795	case VMW_PL_MOB:
796		return 0;
797	case TTM_PL_VRAM:
798		mem->bus.offset = mem->start << PAGE_SHIFT;
799		mem->bus.base = dev_priv->vram_start;
800		mem->bus.is_iomem = true;
801		break;
802	default:
803		return -EINVAL;
804	}
805	return 0;
806}
807
808static void vmw_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
809{
810}
811
812static int vmw_ttm_fault_reserve_notify(struct ttm_buffer_object *bo)
813{
814	return 0;
815}
816
817/**
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
818 * vmw_move_notify - TTM move_notify_callback
819 *
820 * @bo: The TTM buffer object about to move.
821 * @mem: The struct ttm_mem_reg indicating to what memory
822 *       region the move is taking place.
823 *
824 * Calls move_notify for all subsystems needing it.
825 * (currently only resources).
826 */
827static void vmw_move_notify(struct ttm_buffer_object *bo,
828			    struct ttm_mem_reg *mem)
829{
830	vmw_resource_move_notify(bo, mem);
831	vmw_query_move_notify(bo, mem);
832}
833
834
835/**
836 * vmw_swap_notify - TTM move_notify_callback
837 *
838 * @bo: The TTM buffer object about to be swapped out.
839 */
840static void vmw_swap_notify(struct ttm_buffer_object *bo)
841{
 
 
 
842	ttm_bo_wait(bo, false, false, false);
 
843}
844
845
846struct ttm_bo_driver vmw_bo_driver = {
847	.ttm_tt_create = &vmw_ttm_tt_create,
848	.ttm_tt_populate = &vmw_ttm_populate,
849	.ttm_tt_unpopulate = &vmw_ttm_unpopulate,
850	.invalidate_caches = vmw_invalidate_caches,
851	.init_mem_type = vmw_init_mem_type,
852	.evict_flags = vmw_evict_flags,
853	.move = NULL,
854	.verify_access = vmw_verify_access,
 
 
 
 
 
855	.move_notify = vmw_move_notify,
856	.swap_notify = vmw_swap_notify,
857	.fault_reserve_notify = &vmw_ttm_fault_reserve_notify,
858	.io_mem_reserve = &vmw_ttm_io_mem_reserve,
859	.io_mem_free = &vmw_ttm_io_mem_free,
860};
v3.15
  1/**************************************************************************
  2 *
  3 * Copyright © 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#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>
 32
 33static uint32_t vram_placement_flags = TTM_PL_FLAG_VRAM |
 34	TTM_PL_FLAG_CACHED;
 
 
 
 35
 36static uint32_t vram_ne_placement_flags = TTM_PL_FLAG_VRAM |
 37	TTM_PL_FLAG_CACHED |
 38	TTM_PL_FLAG_NO_EVICT;
 
 
 39
 40static uint32_t sys_placement_flags = TTM_PL_FLAG_SYSTEM |
 41	TTM_PL_FLAG_CACHED;
 
 
 
 42
 43static uint32_t sys_ne_placement_flags = TTM_PL_FLAG_SYSTEM |
 44	TTM_PL_FLAG_CACHED |
 45	TTM_PL_FLAG_NO_EVICT;
 
 
 46
 47static uint32_t gmr_placement_flags = VMW_PL_FLAG_GMR |
 48	TTM_PL_FLAG_CACHED;
 
 
 
 49
 50static uint32_t gmr_ne_placement_flags = VMW_PL_FLAG_GMR |
 51	TTM_PL_FLAG_CACHED |
 52	TTM_PL_FLAG_NO_EVICT;
 
 
 53
 54static uint32_t mob_placement_flags = VMW_PL_FLAG_MOB |
 55	TTM_PL_FLAG_CACHED;
 
 
 
 56
 57struct ttm_placement vmw_vram_placement = {
 58	.fpfn = 0,
 59	.lpfn = 0,
 
 
 
 
 60	.num_placement = 1,
 61	.placement = &vram_placement_flags,
 62	.num_busy_placement = 1,
 63	.busy_placement = &vram_placement_flags
 64};
 65
 66static uint32_t vram_gmr_placement_flags[] = {
 67	TTM_PL_FLAG_VRAM | TTM_PL_FLAG_CACHED,
 68	VMW_PL_FLAG_GMR | TTM_PL_FLAG_CACHED
 
 
 
 
 
 
 
 69};
 70
 71static uint32_t gmr_vram_placement_flags[] = {
 72	VMW_PL_FLAG_GMR | TTM_PL_FLAG_CACHED,
 73	TTM_PL_FLAG_VRAM | TTM_PL_FLAG_CACHED
 
 
 
 
 
 
 
 74};
 75
 76struct ttm_placement vmw_vram_gmr_placement = {
 77	.fpfn = 0,
 78	.lpfn = 0,
 79	.num_placement = 2,
 80	.placement = vram_gmr_placement_flags,
 81	.num_busy_placement = 1,
 82	.busy_placement = &gmr_placement_flags
 83};
 84
 85static uint32_t vram_gmr_ne_placement_flags[] = {
 86	TTM_PL_FLAG_VRAM | TTM_PL_FLAG_CACHED | TTM_PL_FLAG_NO_EVICT,
 87	VMW_PL_FLAG_GMR | TTM_PL_FLAG_CACHED | TTM_PL_FLAG_NO_EVICT
 
 
 
 
 
 
 
 
 
 88};
 89
 90struct ttm_placement vmw_vram_gmr_ne_placement = {
 91	.fpfn = 0,
 92	.lpfn = 0,
 93	.num_placement = 2,
 94	.placement = vram_gmr_ne_placement_flags,
 95	.num_busy_placement = 1,
 96	.busy_placement = &gmr_ne_placement_flags
 97};
 98
 99struct ttm_placement vmw_vram_sys_placement = {
100	.fpfn = 0,
101	.lpfn = 0,
102	.num_placement = 1,
103	.placement = &vram_placement_flags,
104	.num_busy_placement = 1,
105	.busy_placement = &sys_placement_flags
106};
107
108struct ttm_placement vmw_vram_ne_placement = {
109	.fpfn = 0,
110	.lpfn = 0,
111	.num_placement = 1,
112	.placement = &vram_ne_placement_flags,
113	.num_busy_placement = 1,
114	.busy_placement = &vram_ne_placement_flags
115};
116
117struct ttm_placement vmw_sys_placement = {
118	.fpfn = 0,
119	.lpfn = 0,
120	.num_placement = 1,
121	.placement = &sys_placement_flags,
122	.num_busy_placement = 1,
123	.busy_placement = &sys_placement_flags
124};
125
126struct ttm_placement vmw_sys_ne_placement = {
127	.fpfn = 0,
128	.lpfn = 0,
129	.num_placement = 1,
130	.placement = &sys_ne_placement_flags,
131	.num_busy_placement = 1,
132	.busy_placement = &sys_ne_placement_flags
133};
134
135static uint32_t evictable_placement_flags[] = {
136	TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED,
137	TTM_PL_FLAG_VRAM | TTM_PL_FLAG_CACHED,
138	VMW_PL_FLAG_GMR | TTM_PL_FLAG_CACHED,
139	VMW_PL_FLAG_MOB | TTM_PL_FLAG_CACHED
 
 
 
 
 
 
 
 
 
 
 
 
 
140};
141
142struct ttm_placement vmw_evictable_placement = {
143	.fpfn = 0,
144	.lpfn = 0,
145	.num_placement = 4,
146	.placement = evictable_placement_flags,
147	.num_busy_placement = 1,
148	.busy_placement = &sys_placement_flags
149};
150
151struct ttm_placement vmw_srf_placement = {
152	.fpfn = 0,
153	.lpfn = 0,
154	.num_placement = 1,
155	.num_busy_placement = 2,
156	.placement = &gmr_placement_flags,
157	.busy_placement = gmr_vram_placement_flags
158};
159
160struct ttm_placement vmw_mob_placement = {
161	.fpfn = 0,
162	.lpfn = 0,
163	.num_placement = 1,
164	.num_busy_placement = 1,
165	.placement = &mob_placement_flags,
166	.busy_placement = &mob_placement_flags
167};
168
 
 
 
 
 
 
 
169struct vmw_ttm_tt {
170	struct ttm_dma_tt dma_ttm;
171	struct vmw_private *dev_priv;
172	int gmr_id;
173	struct vmw_mob *mob;
174	int mem_type;
175	struct sg_table sgt;
176	struct vmw_sg_table vsgt;
177	uint64_t sg_alloc_size;
178	bool mapped;
179};
180
181const size_t vmw_tt_size = sizeof(struct vmw_ttm_tt);
182
183/**
184 * Helper functions to advance a struct vmw_piter iterator.
185 *
186 * @viter: Pointer to the iterator.
187 *
188 * These functions return false if past the end of the list,
189 * true otherwise. Functions are selected depending on the current
190 * DMA mapping mode.
191 */
192static bool __vmw_piter_non_sg_next(struct vmw_piter *viter)
193{
194	return ++(viter->i) < viter->num_pages;
195}
196
197static bool __vmw_piter_sg_next(struct vmw_piter *viter)
198{
199	return __sg_page_iter_next(&viter->iter);
200}
201
202
203/**
204 * Helper functions to return a pointer to the current page.
205 *
206 * @viter: Pointer to the iterator
207 *
208 * These functions return a pointer to the page currently
209 * pointed to by @viter. Functions are selected depending on the
210 * current mapping mode.
211 */
212static struct page *__vmw_piter_non_sg_page(struct vmw_piter *viter)
213{
214	return viter->pages[viter->i];
215}
216
217static struct page *__vmw_piter_sg_page(struct vmw_piter *viter)
218{
219	return sg_page_iter_page(&viter->iter);
220}
221
222
223/**
224 * Helper functions to return the DMA address of the current page.
225 *
226 * @viter: Pointer to the iterator
227 *
228 * These functions return the DMA address of the page currently
229 * pointed to by @viter. Functions are selected depending on the
230 * current mapping mode.
231 */
232static dma_addr_t __vmw_piter_phys_addr(struct vmw_piter *viter)
233{
234	return page_to_phys(viter->pages[viter->i]);
235}
236
237static dma_addr_t __vmw_piter_dma_addr(struct vmw_piter *viter)
238{
239	return viter->addrs[viter->i];
240}
241
242static dma_addr_t __vmw_piter_sg_addr(struct vmw_piter *viter)
243{
244	return sg_page_iter_dma_address(&viter->iter);
245}
246
247
248/**
249 * vmw_piter_start - Initialize a struct vmw_piter.
250 *
251 * @viter: Pointer to the iterator to initialize
252 * @vsgt: Pointer to a struct vmw_sg_table to initialize from
253 *
254 * Note that we're following the convention of __sg_page_iter_start, so that
255 * the iterator doesn't point to a valid page after initialization; it has
256 * to be advanced one step first.
257 */
258void vmw_piter_start(struct vmw_piter *viter, const struct vmw_sg_table *vsgt,
259		     unsigned long p_offset)
260{
261	viter->i = p_offset - 1;
262	viter->num_pages = vsgt->num_pages;
263	switch (vsgt->mode) {
264	case vmw_dma_phys:
265		viter->next = &__vmw_piter_non_sg_next;
266		viter->dma_address = &__vmw_piter_phys_addr;
267		viter->page = &__vmw_piter_non_sg_page;
268		viter->pages = vsgt->pages;
269		break;
270	case vmw_dma_alloc_coherent:
271		viter->next = &__vmw_piter_non_sg_next;
272		viter->dma_address = &__vmw_piter_dma_addr;
273		viter->page = &__vmw_piter_non_sg_page;
274		viter->addrs = vsgt->addrs;
275		viter->pages = vsgt->pages;
276		break;
277	case vmw_dma_map_populate:
278	case vmw_dma_map_bind:
279		viter->next = &__vmw_piter_sg_next;
280		viter->dma_address = &__vmw_piter_sg_addr;
281		viter->page = &__vmw_piter_sg_page;
282		__sg_page_iter_start(&viter->iter, vsgt->sgt->sgl,
283				     vsgt->sgt->orig_nents, p_offset);
284		break;
285	default:
286		BUG();
287	}
288}
289
290/**
291 * vmw_ttm_unmap_from_dma - unmap  device addresses previsouly mapped for
292 * TTM pages
293 *
294 * @vmw_tt: Pointer to a struct vmw_ttm_backend
295 *
296 * Used to free dma mappings previously mapped by vmw_ttm_map_for_dma.
297 */
298static void vmw_ttm_unmap_from_dma(struct vmw_ttm_tt *vmw_tt)
299{
300	struct device *dev = vmw_tt->dev_priv->dev->dev;
301
302	dma_unmap_sg(dev, vmw_tt->sgt.sgl, vmw_tt->sgt.nents,
303		DMA_BIDIRECTIONAL);
304	vmw_tt->sgt.nents = vmw_tt->sgt.orig_nents;
305}
306
307/**
308 * vmw_ttm_map_for_dma - map TTM pages to get device addresses
309 *
310 * @vmw_tt: Pointer to a struct vmw_ttm_backend
311 *
312 * This function is used to get device addresses from the kernel DMA layer.
313 * However, it's violating the DMA API in that when this operation has been
314 * performed, it's illegal for the CPU to write to the pages without first
315 * unmapping the DMA mappings, or calling dma_sync_sg_for_cpu(). It is
316 * therefore only legal to call this function if we know that the function
317 * dma_sync_sg_for_cpu() is a NOP, and dma_sync_sg_for_device() is at most
318 * a CPU write buffer flush.
319 */
320static int vmw_ttm_map_for_dma(struct vmw_ttm_tt *vmw_tt)
321{
322	struct device *dev = vmw_tt->dev_priv->dev->dev;
323	int ret;
324
325	ret = dma_map_sg(dev, vmw_tt->sgt.sgl, vmw_tt->sgt.orig_nents,
326			 DMA_BIDIRECTIONAL);
327	if (unlikely(ret == 0))
328		return -ENOMEM;
329
330	vmw_tt->sgt.nents = ret;
331
332	return 0;
333}
334
335/**
336 * vmw_ttm_map_dma - Make sure TTM pages are visible to the device
337 *
338 * @vmw_tt: Pointer to a struct vmw_ttm_tt
339 *
340 * Select the correct function for and make sure the TTM pages are
341 * visible to the device. Allocate storage for the device mappings.
342 * If a mapping has already been performed, indicated by the storage
343 * pointer being non NULL, the function returns success.
344 */
345static int vmw_ttm_map_dma(struct vmw_ttm_tt *vmw_tt)
346{
347	struct vmw_private *dev_priv = vmw_tt->dev_priv;
348	struct ttm_mem_global *glob = vmw_mem_glob(dev_priv);
349	struct vmw_sg_table *vsgt = &vmw_tt->vsgt;
350	struct vmw_piter iter;
351	dma_addr_t old;
352	int ret = 0;
353	static size_t sgl_size;
354	static size_t sgt_size;
355
356	if (vmw_tt->mapped)
357		return 0;
358
359	vsgt->mode = dev_priv->map_mode;
360	vsgt->pages = vmw_tt->dma_ttm.ttm.pages;
361	vsgt->num_pages = vmw_tt->dma_ttm.ttm.num_pages;
362	vsgt->addrs = vmw_tt->dma_ttm.dma_address;
363	vsgt->sgt = &vmw_tt->sgt;
364
365	switch (dev_priv->map_mode) {
366	case vmw_dma_map_bind:
367	case vmw_dma_map_populate:
368		if (unlikely(!sgl_size)) {
369			sgl_size = ttm_round_pot(sizeof(struct scatterlist));
370			sgt_size = ttm_round_pot(sizeof(struct sg_table));
371		}
372		vmw_tt->sg_alloc_size = sgt_size + sgl_size * vsgt->num_pages;
373		ret = ttm_mem_global_alloc(glob, vmw_tt->sg_alloc_size, false,
374					   true);
375		if (unlikely(ret != 0))
376			return ret;
377
378		ret = sg_alloc_table_from_pages(&vmw_tt->sgt, vsgt->pages,
379						vsgt->num_pages, 0,
380						(unsigned long)
381						vsgt->num_pages << PAGE_SHIFT,
382						GFP_KERNEL);
383		if (unlikely(ret != 0))
384			goto out_sg_alloc_fail;
385
386		if (vsgt->num_pages > vmw_tt->sgt.nents) {
387			uint64_t over_alloc =
388				sgl_size * (vsgt->num_pages -
389					    vmw_tt->sgt.nents);
390
391			ttm_mem_global_free(glob, over_alloc);
392			vmw_tt->sg_alloc_size -= over_alloc;
393		}
394
395		ret = vmw_ttm_map_for_dma(vmw_tt);
396		if (unlikely(ret != 0))
397			goto out_map_fail;
398
399		break;
400	default:
401		break;
402	}
403
404	old = ~((dma_addr_t) 0);
405	vmw_tt->vsgt.num_regions = 0;
406	for (vmw_piter_start(&iter, vsgt, 0); vmw_piter_next(&iter);) {
407		dma_addr_t cur = vmw_piter_dma_addr(&iter);
408
409		if (cur != old + PAGE_SIZE)
410			vmw_tt->vsgt.num_regions++;
411		old = cur;
412	}
413
414	vmw_tt->mapped = true;
415	return 0;
416
417out_map_fail:
418	sg_free_table(vmw_tt->vsgt.sgt);
419	vmw_tt->vsgt.sgt = NULL;
420out_sg_alloc_fail:
421	ttm_mem_global_free(glob, vmw_tt->sg_alloc_size);
422	return ret;
423}
424
425/**
426 * vmw_ttm_unmap_dma - Tear down any TTM page device mappings
427 *
428 * @vmw_tt: Pointer to a struct vmw_ttm_tt
429 *
430 * Tear down any previously set up device DMA mappings and free
431 * any storage space allocated for them. If there are no mappings set up,
432 * this function is a NOP.
433 */
434static void vmw_ttm_unmap_dma(struct vmw_ttm_tt *vmw_tt)
435{
436	struct vmw_private *dev_priv = vmw_tt->dev_priv;
437
438	if (!vmw_tt->vsgt.sgt)
439		return;
440
441	switch (dev_priv->map_mode) {
442	case vmw_dma_map_bind:
443	case vmw_dma_map_populate:
444		vmw_ttm_unmap_from_dma(vmw_tt);
445		sg_free_table(vmw_tt->vsgt.sgt);
446		vmw_tt->vsgt.sgt = NULL;
447		ttm_mem_global_free(vmw_mem_glob(dev_priv),
448				    vmw_tt->sg_alloc_size);
449		break;
450	default:
451		break;
452	}
453	vmw_tt->mapped = false;
454}
455
456
457/**
458 * vmw_bo_map_dma - Make sure buffer object pages are visible to the device
459 *
460 * @bo: Pointer to a struct ttm_buffer_object
461 *
462 * Wrapper around vmw_ttm_map_dma, that takes a TTM buffer object pointer
463 * instead of a pointer to a struct vmw_ttm_backend as argument.
464 * Note that the buffer object must be either pinned or reserved before
465 * calling this function.
466 */
467int vmw_bo_map_dma(struct ttm_buffer_object *bo)
468{
469	struct vmw_ttm_tt *vmw_tt =
470		container_of(bo->ttm, struct vmw_ttm_tt, dma_ttm.ttm);
471
472	return vmw_ttm_map_dma(vmw_tt);
473}
474
475
476/**
477 * vmw_bo_unmap_dma - Make sure buffer object pages are visible to the device
478 *
479 * @bo: Pointer to a struct ttm_buffer_object
480 *
481 * Wrapper around vmw_ttm_unmap_dma, that takes a TTM buffer object pointer
482 * instead of a pointer to a struct vmw_ttm_backend as argument.
483 */
484void vmw_bo_unmap_dma(struct ttm_buffer_object *bo)
485{
486	struct vmw_ttm_tt *vmw_tt =
487		container_of(bo->ttm, struct vmw_ttm_tt, dma_ttm.ttm);
488
489	vmw_ttm_unmap_dma(vmw_tt);
490}
491
492
493/**
494 * vmw_bo_sg_table - Return a struct vmw_sg_table object for a
495 * TTM buffer object
496 *
497 * @bo: Pointer to a struct ttm_buffer_object
498 *
499 * Returns a pointer to a struct vmw_sg_table object. The object should
500 * not be freed after use.
501 * Note that for the device addresses to be valid, the buffer object must
502 * either be reserved or pinned.
503 */
504const struct vmw_sg_table *vmw_bo_sg_table(struct ttm_buffer_object *bo)
505{
506	struct vmw_ttm_tt *vmw_tt =
507		container_of(bo->ttm, struct vmw_ttm_tt, dma_ttm.ttm);
508
509	return &vmw_tt->vsgt;
510}
511
512
513static int vmw_ttm_bind(struct ttm_tt *ttm, struct ttm_mem_reg *bo_mem)
514{
515	struct vmw_ttm_tt *vmw_be =
516		container_of(ttm, struct vmw_ttm_tt, dma_ttm.ttm);
517	int ret;
518
519	ret = vmw_ttm_map_dma(vmw_be);
520	if (unlikely(ret != 0))
521		return ret;
522
523	vmw_be->gmr_id = bo_mem->start;
524	vmw_be->mem_type = bo_mem->mem_type;
525
526	switch (bo_mem->mem_type) {
527	case VMW_PL_GMR:
528		return vmw_gmr_bind(vmw_be->dev_priv, &vmw_be->vsgt,
529				    ttm->num_pages, vmw_be->gmr_id);
530	case VMW_PL_MOB:
531		if (unlikely(vmw_be->mob == NULL)) {
532			vmw_be->mob =
533				vmw_mob_create(ttm->num_pages);
534			if (unlikely(vmw_be->mob == NULL))
535				return -ENOMEM;
536		}
537
538		return vmw_mob_bind(vmw_be->dev_priv, vmw_be->mob,
539				    &vmw_be->vsgt, ttm->num_pages,
540				    vmw_be->gmr_id);
541	default:
542		BUG();
543	}
544	return 0;
545}
546
547static int vmw_ttm_unbind(struct ttm_tt *ttm)
548{
549	struct vmw_ttm_tt *vmw_be =
550		container_of(ttm, struct vmw_ttm_tt, dma_ttm.ttm);
551
552	switch (vmw_be->mem_type) {
553	case VMW_PL_GMR:
554		vmw_gmr_unbind(vmw_be->dev_priv, vmw_be->gmr_id);
555		break;
556	case VMW_PL_MOB:
557		vmw_mob_unbind(vmw_be->dev_priv, vmw_be->mob);
558		break;
559	default:
560		BUG();
561	}
562
563	if (vmw_be->dev_priv->map_mode == vmw_dma_map_bind)
564		vmw_ttm_unmap_dma(vmw_be);
565
566	return 0;
567}
568
569
570static void vmw_ttm_destroy(struct ttm_tt *ttm)
571{
572	struct vmw_ttm_tt *vmw_be =
573		container_of(ttm, struct vmw_ttm_tt, dma_ttm.ttm);
574
575	vmw_ttm_unmap_dma(vmw_be);
576	if (vmw_be->dev_priv->map_mode == vmw_dma_alloc_coherent)
577		ttm_dma_tt_fini(&vmw_be->dma_ttm);
578	else
579		ttm_tt_fini(ttm);
580
581	if (vmw_be->mob)
582		vmw_mob_destroy(vmw_be->mob);
583
584	kfree(vmw_be);
585}
586
587
588static int vmw_ttm_populate(struct ttm_tt *ttm)
589{
590	struct vmw_ttm_tt *vmw_tt =
591		container_of(ttm, struct vmw_ttm_tt, dma_ttm.ttm);
592	struct vmw_private *dev_priv = vmw_tt->dev_priv;
593	struct ttm_mem_global *glob = vmw_mem_glob(dev_priv);
594	int ret;
595
596	if (ttm->state != tt_unpopulated)
597		return 0;
598
599	if (dev_priv->map_mode == vmw_dma_alloc_coherent) {
600		size_t size =
601			ttm_round_pot(ttm->num_pages * sizeof(dma_addr_t));
602		ret = ttm_mem_global_alloc(glob, size, false, true);
603		if (unlikely(ret != 0))
604			return ret;
605
606		ret = ttm_dma_populate(&vmw_tt->dma_ttm, dev_priv->dev->dev);
607		if (unlikely(ret != 0))
608			ttm_mem_global_free(glob, size);
609	} else
610		ret = ttm_pool_populate(ttm);
611
612	return ret;
613}
614
615static void vmw_ttm_unpopulate(struct ttm_tt *ttm)
616{
617	struct vmw_ttm_tt *vmw_tt = container_of(ttm, struct vmw_ttm_tt,
618						 dma_ttm.ttm);
619	struct vmw_private *dev_priv = vmw_tt->dev_priv;
620	struct ttm_mem_global *glob = vmw_mem_glob(dev_priv);
621
622
623	if (vmw_tt->mob) {
624		vmw_mob_destroy(vmw_tt->mob);
625		vmw_tt->mob = NULL;
626	}
627
628	vmw_ttm_unmap_dma(vmw_tt);
629	if (dev_priv->map_mode == vmw_dma_alloc_coherent) {
630		size_t size =
631			ttm_round_pot(ttm->num_pages * sizeof(dma_addr_t));
632
633		ttm_dma_unpopulate(&vmw_tt->dma_ttm, dev_priv->dev->dev);
634		ttm_mem_global_free(glob, size);
635	} else
636		ttm_pool_unpopulate(ttm);
637}
638
639static struct ttm_backend_func vmw_ttm_func = {
640	.bind = vmw_ttm_bind,
641	.unbind = vmw_ttm_unbind,
642	.destroy = vmw_ttm_destroy,
643};
644
645static struct ttm_tt *vmw_ttm_tt_create(struct ttm_bo_device *bdev,
646				 unsigned long size, uint32_t page_flags,
647				 struct page *dummy_read_page)
648{
649	struct vmw_ttm_tt *vmw_be;
650	int ret;
651
652	vmw_be = kzalloc(sizeof(*vmw_be), GFP_KERNEL);
653	if (!vmw_be)
654		return NULL;
655
656	vmw_be->dma_ttm.ttm.func = &vmw_ttm_func;
657	vmw_be->dev_priv = container_of(bdev, struct vmw_private, bdev);
658	vmw_be->mob = NULL;
659
660	if (vmw_be->dev_priv->map_mode == vmw_dma_alloc_coherent)
661		ret = ttm_dma_tt_init(&vmw_be->dma_ttm, bdev, size, page_flags,
662				      dummy_read_page);
663	else
664		ret = ttm_tt_init(&vmw_be->dma_ttm.ttm, bdev, size, page_flags,
665				  dummy_read_page);
666	if (unlikely(ret != 0))
667		goto out_no_init;
668
669	return &vmw_be->dma_ttm.ttm;
670out_no_init:
671	kfree(vmw_be);
672	return NULL;
673}
674
675static int vmw_invalidate_caches(struct ttm_bo_device *bdev, uint32_t flags)
676{
677	return 0;
678}
679
680static int vmw_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
681		      struct ttm_mem_type_manager *man)
682{
683	switch (type) {
684	case TTM_PL_SYSTEM:
685		/* System memory */
686
687		man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
688		man->available_caching = TTM_PL_FLAG_CACHED;
689		man->default_caching = TTM_PL_FLAG_CACHED;
690		break;
691	case TTM_PL_VRAM:
692		/* "On-card" video ram */
693		man->func = &ttm_bo_manager_func;
694		man->gpu_offset = 0;
695		man->flags = TTM_MEMTYPE_FLAG_FIXED | TTM_MEMTYPE_FLAG_MAPPABLE;
696		man->available_caching = TTM_PL_FLAG_CACHED;
697		man->default_caching = TTM_PL_FLAG_CACHED;
698		break;
699	case VMW_PL_GMR:
700	case VMW_PL_MOB:
701		/*
702		 * "Guest Memory Regions" is an aperture like feature with
703		 *  one slot per bo. There is an upper limit of the number of
704		 *  slots as well as the bo size.
705		 */
706		man->func = &vmw_gmrid_manager_func;
707		man->gpu_offset = 0;
708		man->flags = TTM_MEMTYPE_FLAG_CMA | TTM_MEMTYPE_FLAG_MAPPABLE;
709		man->available_caching = TTM_PL_FLAG_CACHED;
710		man->default_caching = TTM_PL_FLAG_CACHED;
711		break;
712	default:
713		DRM_ERROR("Unsupported memory type %u\n", (unsigned)type);
714		return -EINVAL;
715	}
716	return 0;
717}
718
719static void vmw_evict_flags(struct ttm_buffer_object *bo,
720		     struct ttm_placement *placement)
721{
722	*placement = vmw_sys_placement;
723}
724
725static int vmw_verify_access(struct ttm_buffer_object *bo, struct file *filp)
726{
727	struct ttm_object_file *tfile =
728		vmw_fpriv((struct drm_file *)filp->private_data)->tfile;
729
730	return vmw_user_dmabuf_verify_access(bo, tfile);
731}
732
733static int vmw_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
734{
735	struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
736	struct vmw_private *dev_priv = container_of(bdev, struct vmw_private, bdev);
737
738	mem->bus.addr = NULL;
739	mem->bus.is_iomem = false;
740	mem->bus.offset = 0;
741	mem->bus.size = mem->num_pages << PAGE_SHIFT;
742	mem->bus.base = 0;
743	if (!(man->flags & TTM_MEMTYPE_FLAG_MAPPABLE))
744		return -EINVAL;
745	switch (mem->mem_type) {
746	case TTM_PL_SYSTEM:
747	case VMW_PL_GMR:
748	case VMW_PL_MOB:
749		return 0;
750	case TTM_PL_VRAM:
751		mem->bus.offset = mem->start << PAGE_SHIFT;
752		mem->bus.base = dev_priv->vram_start;
753		mem->bus.is_iomem = true;
754		break;
755	default:
756		return -EINVAL;
757	}
758	return 0;
759}
760
761static void vmw_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
762{
763}
764
765static int vmw_ttm_fault_reserve_notify(struct ttm_buffer_object *bo)
766{
767	return 0;
768}
769
770/**
771 * FIXME: We're using the old vmware polling method to sync.
772 * Do this with fences instead.
773 */
774
775static void *vmw_sync_obj_ref(void *sync_obj)
776{
777
778	return (void *)
779		vmw_fence_obj_reference((struct vmw_fence_obj *) sync_obj);
780}
781
782static void vmw_sync_obj_unref(void **sync_obj)
783{
784	vmw_fence_obj_unreference((struct vmw_fence_obj **) sync_obj);
785}
786
787static int vmw_sync_obj_flush(void *sync_obj)
788{
789	vmw_fence_obj_flush((struct vmw_fence_obj *) sync_obj);
790	return 0;
791}
792
793static bool vmw_sync_obj_signaled(void *sync_obj)
794{
795	return	vmw_fence_obj_signaled((struct vmw_fence_obj *) sync_obj,
796				       DRM_VMW_FENCE_FLAG_EXEC);
797
798}
799
800static int vmw_sync_obj_wait(void *sync_obj, bool lazy, bool interruptible)
801{
802	return vmw_fence_obj_wait((struct vmw_fence_obj *) sync_obj,
803				  DRM_VMW_FENCE_FLAG_EXEC,
804				  lazy, interruptible,
805				  VMW_FENCE_WAIT_TIMEOUT);
806}
807
808/**
809 * vmw_move_notify - TTM move_notify_callback
810 *
811 * @bo:             The TTM buffer object about to move.
812 * @mem:            The truct ttm_mem_reg indicating to what memory
813 *                  region the move is taking place.
814 *
815 * Calls move_notify for all subsystems needing it.
816 * (currently only resources).
817 */
818static void vmw_move_notify(struct ttm_buffer_object *bo,
819			    struct ttm_mem_reg *mem)
820{
821	vmw_resource_move_notify(bo, mem);
 
822}
823
824
825/**
826 * vmw_swap_notify - TTM move_notify_callback
827 *
828 * @bo:             The TTM buffer object about to be swapped out.
829 */
830static void vmw_swap_notify(struct ttm_buffer_object *bo)
831{
832	struct ttm_bo_device *bdev = bo->bdev;
833
834	spin_lock(&bdev->fence_lock);
835	ttm_bo_wait(bo, false, false, false);
836	spin_unlock(&bdev->fence_lock);
837}
838
839
840struct ttm_bo_driver vmw_bo_driver = {
841	.ttm_tt_create = &vmw_ttm_tt_create,
842	.ttm_tt_populate = &vmw_ttm_populate,
843	.ttm_tt_unpopulate = &vmw_ttm_unpopulate,
844	.invalidate_caches = vmw_invalidate_caches,
845	.init_mem_type = vmw_init_mem_type,
846	.evict_flags = vmw_evict_flags,
847	.move = NULL,
848	.verify_access = vmw_verify_access,
849	.sync_obj_signaled = vmw_sync_obj_signaled,
850	.sync_obj_wait = vmw_sync_obj_wait,
851	.sync_obj_flush = vmw_sync_obj_flush,
852	.sync_obj_unref = vmw_sync_obj_unref,
853	.sync_obj_ref = vmw_sync_obj_ref,
854	.move_notify = vmw_move_notify,
855	.swap_notify = vmw_swap_notify,
856	.fault_reserve_notify = &vmw_ttm_fault_reserve_notify,
857	.io_mem_reserve = &vmw_ttm_io_mem_reserve,
858	.io_mem_free = &vmw_ttm_io_mem_free,
859};