xref: /netbsd-src/sys/external/bsd/drm2/dist/drm/vmwgfx/vmwgfx_ttm_buffer.c (revision 41ec02673d281bbb3d38e6c78504ce6e30c228c1)
1 /*	$NetBSD: vmwgfx_ttm_buffer.c,v 1.2 2021/12/18 23:45:45 riastradh Exp $	*/
2 
3 // SPDX-License-Identifier: GPL-2.0 OR MIT
4 /**************************************************************************
5  *
6  * Copyright 2009-2015 VMware, Inc., Palo Alto, CA., USA
7  *
8  * Permission is hereby granted, free of charge, to any person obtaining a
9  * copy of this software and associated documentation files (the
10  * "Software"), to deal in the Software without restriction, including
11  * without limitation the rights to use, copy, modify, merge, publish,
12  * distribute, sub license, and/or sell copies of the Software, and to
13  * permit persons to whom the Software is furnished to do so, subject to
14  * the following conditions:
15  *
16  * The above copyright notice and this permission notice (including the
17  * next paragraph) shall be included in all copies or substantial portions
18  * of the Software.
19  *
20  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
21  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
22  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
23  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
24  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
25  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
26  * USE OR OTHER DEALINGS IN THE SOFTWARE.
27  *
28  **************************************************************************/
29 
30 #include <sys/cdefs.h>
31 __KERNEL_RCSID(0, "$NetBSD: vmwgfx_ttm_buffer.c,v 1.2 2021/12/18 23:45:45 riastradh Exp $");
32 
33 #include "vmwgfx_drv.h"
34 #include <drm/ttm/ttm_bo_driver.h>
35 #include <drm/ttm/ttm_placement.h>
36 #include <drm/ttm/ttm_page_alloc.h>
37 
38 static const struct ttm_place vram_placement_flags = {
39 	.fpfn = 0,
40 	.lpfn = 0,
41 	.flags = TTM_PL_FLAG_VRAM | TTM_PL_FLAG_CACHED
42 };
43 
44 static const struct ttm_place vram_ne_placement_flags = {
45 	.fpfn = 0,
46 	.lpfn = 0,
47 	.flags = TTM_PL_FLAG_VRAM | TTM_PL_FLAG_CACHED | TTM_PL_FLAG_NO_EVICT
48 };
49 
50 static const struct ttm_place sys_placement_flags = {
51 	.fpfn = 0,
52 	.lpfn = 0,
53 	.flags = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED
54 };
55 
56 static const struct ttm_place sys_ne_placement_flags = {
57 	.fpfn = 0,
58 	.lpfn = 0,
59 	.flags = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED | TTM_PL_FLAG_NO_EVICT
60 };
61 
62 static const struct ttm_place gmr_placement_flags = {
63 	.fpfn = 0,
64 	.lpfn = 0,
65 	.flags = VMW_PL_FLAG_GMR | TTM_PL_FLAG_CACHED
66 };
67 
68 static const struct ttm_place gmr_ne_placement_flags = {
69 	.fpfn = 0,
70 	.lpfn = 0,
71 	.flags = VMW_PL_FLAG_GMR | TTM_PL_FLAG_CACHED | TTM_PL_FLAG_NO_EVICT
72 };
73 
74 static const struct ttm_place mob_placement_flags = {
75 	.fpfn = 0,
76 	.lpfn = 0,
77 	.flags = VMW_PL_FLAG_MOB | TTM_PL_FLAG_CACHED
78 };
79 
80 static const struct ttm_place mob_ne_placement_flags = {
81 	.fpfn = 0,
82 	.lpfn = 0,
83 	.flags = VMW_PL_FLAG_MOB | TTM_PL_FLAG_CACHED | TTM_PL_FLAG_NO_EVICT
84 };
85 
86 struct ttm_placement vmw_vram_placement = {
87 	.num_placement = 1,
88 	.placement = &vram_placement_flags,
89 	.num_busy_placement = 1,
90 	.busy_placement = &vram_placement_flags
91 };
92 
93 static const struct ttm_place vram_gmr_placement_flags[] = {
94 	{
95 		.fpfn = 0,
96 		.lpfn = 0,
97 		.flags = TTM_PL_FLAG_VRAM | TTM_PL_FLAG_CACHED
98 	}, {
99 		.fpfn = 0,
100 		.lpfn = 0,
101 		.flags = VMW_PL_FLAG_GMR | TTM_PL_FLAG_CACHED
102 	}
103 };
104 
105 static const struct ttm_place gmr_vram_placement_flags[] = {
106 	{
107 		.fpfn = 0,
108 		.lpfn = 0,
109 		.flags = VMW_PL_FLAG_GMR | TTM_PL_FLAG_CACHED
110 	}, {
111 		.fpfn = 0,
112 		.lpfn = 0,
113 		.flags = TTM_PL_FLAG_VRAM | TTM_PL_FLAG_CACHED
114 	}
115 };
116 
117 struct ttm_placement vmw_vram_gmr_placement = {
118 	.num_placement = 2,
119 	.placement = vram_gmr_placement_flags,
120 	.num_busy_placement = 1,
121 	.busy_placement = &gmr_placement_flags
122 };
123 
124 static const struct ttm_place vram_gmr_ne_placement_flags[] = {
125 	{
126 		.fpfn = 0,
127 		.lpfn = 0,
128 		.flags = TTM_PL_FLAG_VRAM | TTM_PL_FLAG_CACHED |
129 			 TTM_PL_FLAG_NO_EVICT
130 	}, {
131 		.fpfn = 0,
132 		.lpfn = 0,
133 		.flags = VMW_PL_FLAG_GMR | TTM_PL_FLAG_CACHED |
134 			 TTM_PL_FLAG_NO_EVICT
135 	}
136 };
137 
138 struct ttm_placement vmw_vram_gmr_ne_placement = {
139 	.num_placement = 2,
140 	.placement = vram_gmr_ne_placement_flags,
141 	.num_busy_placement = 1,
142 	.busy_placement = &gmr_ne_placement_flags
143 };
144 
145 struct ttm_placement vmw_vram_sys_placement = {
146 	.num_placement = 1,
147 	.placement = &vram_placement_flags,
148 	.num_busy_placement = 1,
149 	.busy_placement = &sys_placement_flags
150 };
151 
152 struct ttm_placement vmw_vram_ne_placement = {
153 	.num_placement = 1,
154 	.placement = &vram_ne_placement_flags,
155 	.num_busy_placement = 1,
156 	.busy_placement = &vram_ne_placement_flags
157 };
158 
159 struct ttm_placement vmw_sys_placement = {
160 	.num_placement = 1,
161 	.placement = &sys_placement_flags,
162 	.num_busy_placement = 1,
163 	.busy_placement = &sys_placement_flags
164 };
165 
166 struct ttm_placement vmw_sys_ne_placement = {
167 	.num_placement = 1,
168 	.placement = &sys_ne_placement_flags,
169 	.num_busy_placement = 1,
170 	.busy_placement = &sys_ne_placement_flags
171 };
172 
173 static const struct ttm_place evictable_placement_flags[] = {
174 	{
175 		.fpfn = 0,
176 		.lpfn = 0,
177 		.flags = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED
178 	}, {
179 		.fpfn = 0,
180 		.lpfn = 0,
181 		.flags = TTM_PL_FLAG_VRAM | TTM_PL_FLAG_CACHED
182 	}, {
183 		.fpfn = 0,
184 		.lpfn = 0,
185 		.flags = VMW_PL_FLAG_GMR | TTM_PL_FLAG_CACHED
186 	}, {
187 		.fpfn = 0,
188 		.lpfn = 0,
189 		.flags = VMW_PL_FLAG_MOB | TTM_PL_FLAG_CACHED
190 	}
191 };
192 
193 static const struct ttm_place nonfixed_placement_flags[] = {
194 	{
195 		.fpfn = 0,
196 		.lpfn = 0,
197 		.flags = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED
198 	}, {
199 		.fpfn = 0,
200 		.lpfn = 0,
201 		.flags = VMW_PL_FLAG_GMR | TTM_PL_FLAG_CACHED
202 	}, {
203 		.fpfn = 0,
204 		.lpfn = 0,
205 		.flags = VMW_PL_FLAG_MOB | TTM_PL_FLAG_CACHED
206 	}
207 };
208 
209 struct ttm_placement vmw_evictable_placement = {
210 	.num_placement = 4,
211 	.placement = evictable_placement_flags,
212 	.num_busy_placement = 1,
213 	.busy_placement = &sys_placement_flags
214 };
215 
216 struct ttm_placement vmw_srf_placement = {
217 	.num_placement = 1,
218 	.num_busy_placement = 2,
219 	.placement = &gmr_placement_flags,
220 	.busy_placement = gmr_vram_placement_flags
221 };
222 
223 struct ttm_placement vmw_mob_placement = {
224 	.num_placement = 1,
225 	.num_busy_placement = 1,
226 	.placement = &mob_placement_flags,
227 	.busy_placement = &mob_placement_flags
228 };
229 
230 struct ttm_placement vmw_mob_ne_placement = {
231 	.num_placement = 1,
232 	.num_busy_placement = 1,
233 	.placement = &mob_ne_placement_flags,
234 	.busy_placement = &mob_ne_placement_flags
235 };
236 
237 struct ttm_placement vmw_nonfixed_placement = {
238 	.num_placement = 3,
239 	.placement = nonfixed_placement_flags,
240 	.num_busy_placement = 1,
241 	.busy_placement = &sys_placement_flags
242 };
243 
244 struct vmw_ttm_tt {
245 	struct ttm_dma_tt dma_ttm;
246 	struct vmw_private *dev_priv;
247 	int gmr_id;
248 	struct vmw_mob *mob;
249 	int mem_type;
250 	struct sg_table sgt;
251 	struct vmw_sg_table vsgt;
252 	uint64_t sg_alloc_size;
253 	bool mapped;
254 };
255 
256 const size_t vmw_tt_size = sizeof(struct vmw_ttm_tt);
257 
258 /**
259  * Helper functions to advance a struct vmw_piter iterator.
260  *
261  * @viter: Pointer to the iterator.
262  *
263  * These functions return false if past the end of the list,
264  * true otherwise. Functions are selected depending on the current
265  * DMA mapping mode.
266  */
__vmw_piter_non_sg_next(struct vmw_piter * viter)267 static bool __vmw_piter_non_sg_next(struct vmw_piter *viter)
268 {
269 	return ++(viter->i) < viter->num_pages;
270 }
271 
__vmw_piter_sg_next(struct vmw_piter * viter)272 static bool __vmw_piter_sg_next(struct vmw_piter *viter)
273 {
274 	bool ret = __vmw_piter_non_sg_next(viter);
275 
276 	return __sg_page_iter_dma_next(&viter->iter) && ret;
277 }
278 
279 
280 /**
281  * Helper functions to return a pointer to the current page.
282  *
283  * @viter: Pointer to the iterator
284  *
285  * These functions return a pointer to the page currently
286  * pointed to by @viter. Functions are selected depending on the
287  * current mapping mode.
288  */
__vmw_piter_non_sg_page(struct vmw_piter * viter)289 static struct page *__vmw_piter_non_sg_page(struct vmw_piter *viter)
290 {
291 	return viter->pages[viter->i];
292 }
293 
294 /**
295  * Helper functions to return the DMA address of the current page.
296  *
297  * @viter: Pointer to the iterator
298  *
299  * These functions return the DMA address of the page currently
300  * pointed to by @viter. Functions are selected depending on the
301  * current mapping mode.
302  */
__vmw_piter_phys_addr(struct vmw_piter * viter)303 static dma_addr_t __vmw_piter_phys_addr(struct vmw_piter *viter)
304 {
305 	return page_to_phys(viter->pages[viter->i]);
306 }
307 
__vmw_piter_dma_addr(struct vmw_piter * viter)308 static dma_addr_t __vmw_piter_dma_addr(struct vmw_piter *viter)
309 {
310 	return viter->addrs[viter->i];
311 }
312 
__vmw_piter_sg_addr(struct vmw_piter * viter)313 static dma_addr_t __vmw_piter_sg_addr(struct vmw_piter *viter)
314 {
315 	return sg_page_iter_dma_address(&viter->iter);
316 }
317 
318 
319 /**
320  * vmw_piter_start - Initialize a struct vmw_piter.
321  *
322  * @viter: Pointer to the iterator to initialize
323  * @vsgt: Pointer to a struct vmw_sg_table to initialize from
324  *
325  * Note that we're following the convention of __sg_page_iter_start, so that
326  * the iterator doesn't point to a valid page after initialization; it has
327  * to be advanced one step first.
328  */
vmw_piter_start(struct vmw_piter * viter,const struct vmw_sg_table * vsgt,unsigned long p_offset)329 void vmw_piter_start(struct vmw_piter *viter, const struct vmw_sg_table *vsgt,
330 		     unsigned long p_offset)
331 {
332 	viter->i = p_offset - 1;
333 	viter->num_pages = vsgt->num_pages;
334 	viter->page = &__vmw_piter_non_sg_page;
335 	viter->pages = vsgt->pages;
336 	switch (vsgt->mode) {
337 	case vmw_dma_phys:
338 		viter->next = &__vmw_piter_non_sg_next;
339 		viter->dma_address = &__vmw_piter_phys_addr;
340 		break;
341 	case vmw_dma_alloc_coherent:
342 		viter->next = &__vmw_piter_non_sg_next;
343 		viter->dma_address = &__vmw_piter_dma_addr;
344 		viter->addrs = vsgt->addrs;
345 		break;
346 	case vmw_dma_map_populate:
347 	case vmw_dma_map_bind:
348 		viter->next = &__vmw_piter_sg_next;
349 		viter->dma_address = &__vmw_piter_sg_addr;
350 		__sg_page_iter_start(&viter->iter.base, vsgt->sgt->sgl,
351 				     vsgt->sgt->orig_nents, p_offset);
352 		break;
353 	default:
354 		BUG();
355 	}
356 }
357 
358 /**
359  * vmw_ttm_unmap_from_dma - unmap  device addresses previsouly mapped for
360  * TTM pages
361  *
362  * @vmw_tt: Pointer to a struct vmw_ttm_backend
363  *
364  * Used to free dma mappings previously mapped by vmw_ttm_map_for_dma.
365  */
vmw_ttm_unmap_from_dma(struct vmw_ttm_tt * vmw_tt)366 static void vmw_ttm_unmap_from_dma(struct vmw_ttm_tt *vmw_tt)
367 {
368 	struct device *dev = vmw_tt->dev_priv->dev->dev;
369 
370 	dma_unmap_sg(dev, vmw_tt->sgt.sgl, vmw_tt->sgt.nents,
371 		DMA_BIDIRECTIONAL);
372 	vmw_tt->sgt.nents = vmw_tt->sgt.orig_nents;
373 }
374 
375 /**
376  * vmw_ttm_map_for_dma - map TTM pages to get device addresses
377  *
378  * @vmw_tt: Pointer to a struct vmw_ttm_backend
379  *
380  * This function is used to get device addresses from the kernel DMA layer.
381  * However, it's violating the DMA API in that when this operation has been
382  * performed, it's illegal for the CPU to write to the pages without first
383  * unmapping the DMA mappings, or calling dma_sync_sg_for_cpu(). It is
384  * therefore only legal to call this function if we know that the function
385  * dma_sync_sg_for_cpu() is a NOP, and dma_sync_sg_for_device() is at most
386  * a CPU write buffer flush.
387  */
vmw_ttm_map_for_dma(struct vmw_ttm_tt * vmw_tt)388 static int vmw_ttm_map_for_dma(struct vmw_ttm_tt *vmw_tt)
389 {
390 	struct device *dev = vmw_tt->dev_priv->dev->dev;
391 	int ret;
392 
393 	ret = dma_map_sg(dev, vmw_tt->sgt.sgl, vmw_tt->sgt.orig_nents,
394 			 DMA_BIDIRECTIONAL);
395 	if (unlikely(ret == 0))
396 		return -ENOMEM;
397 
398 	vmw_tt->sgt.nents = ret;
399 
400 	return 0;
401 }
402 
403 /**
404  * vmw_ttm_map_dma - Make sure TTM pages are visible to the device
405  *
406  * @vmw_tt: Pointer to a struct vmw_ttm_tt
407  *
408  * Select the correct function for and make sure the TTM pages are
409  * visible to the device. Allocate storage for the device mappings.
410  * If a mapping has already been performed, indicated by the storage
411  * pointer being non NULL, the function returns success.
412  */
vmw_ttm_map_dma(struct vmw_ttm_tt * vmw_tt)413 static int vmw_ttm_map_dma(struct vmw_ttm_tt *vmw_tt)
414 {
415 	struct vmw_private *dev_priv = vmw_tt->dev_priv;
416 	struct ttm_mem_global *glob = vmw_mem_glob(dev_priv);
417 	struct vmw_sg_table *vsgt = &vmw_tt->vsgt;
418 	struct ttm_operation_ctx ctx = {
419 		.interruptible = true,
420 		.no_wait_gpu = false
421 	};
422 	struct vmw_piter iter;
423 	dma_addr_t old;
424 	int ret = 0;
425 	static size_t sgl_size;
426 	static size_t sgt_size;
427 
428 	if (vmw_tt->mapped)
429 		return 0;
430 
431 	vsgt->mode = dev_priv->map_mode;
432 	vsgt->pages = vmw_tt->dma_ttm.ttm.pages;
433 	vsgt->num_pages = vmw_tt->dma_ttm.ttm.num_pages;
434 	vsgt->addrs = vmw_tt->dma_ttm.dma_address;
435 	vsgt->sgt = &vmw_tt->sgt;
436 
437 	switch (dev_priv->map_mode) {
438 	case vmw_dma_map_bind:
439 	case vmw_dma_map_populate:
440 		if (unlikely(!sgl_size)) {
441 			sgl_size = ttm_round_pot(sizeof(struct scatterlist));
442 			sgt_size = ttm_round_pot(sizeof(struct sg_table));
443 		}
444 		vmw_tt->sg_alloc_size = sgt_size + sgl_size * vsgt->num_pages;
445 		ret = ttm_mem_global_alloc(glob, vmw_tt->sg_alloc_size, &ctx);
446 		if (unlikely(ret != 0))
447 			return ret;
448 
449 		ret = __sg_alloc_table_from_pages
450 			(&vmw_tt->sgt, vsgt->pages, vsgt->num_pages, 0,
451 			 (unsigned long) vsgt->num_pages << PAGE_SHIFT,
452 			 dma_get_max_seg_size(dev_priv->dev->dev),
453 			 GFP_KERNEL);
454 		if (unlikely(ret != 0))
455 			goto out_sg_alloc_fail;
456 
457 		if (vsgt->num_pages > vmw_tt->sgt.nents) {
458 			uint64_t over_alloc =
459 				sgl_size * (vsgt->num_pages -
460 					    vmw_tt->sgt.nents);
461 
462 			ttm_mem_global_free(glob, over_alloc);
463 			vmw_tt->sg_alloc_size -= over_alloc;
464 		}
465 
466 		ret = vmw_ttm_map_for_dma(vmw_tt);
467 		if (unlikely(ret != 0))
468 			goto out_map_fail;
469 
470 		break;
471 	default:
472 		break;
473 	}
474 
475 	old = ~((dma_addr_t) 0);
476 	vmw_tt->vsgt.num_regions = 0;
477 	for (vmw_piter_start(&iter, vsgt, 0); vmw_piter_next(&iter);) {
478 		dma_addr_t cur = vmw_piter_dma_addr(&iter);
479 
480 		if (cur != old + PAGE_SIZE)
481 			vmw_tt->vsgt.num_regions++;
482 		old = cur;
483 	}
484 
485 	vmw_tt->mapped = true;
486 	return 0;
487 
488 out_map_fail:
489 	sg_free_table(vmw_tt->vsgt.sgt);
490 	vmw_tt->vsgt.sgt = NULL;
491 out_sg_alloc_fail:
492 	ttm_mem_global_free(glob, vmw_tt->sg_alloc_size);
493 	return ret;
494 }
495 
496 /**
497  * vmw_ttm_unmap_dma - Tear down any TTM page device mappings
498  *
499  * @vmw_tt: Pointer to a struct vmw_ttm_tt
500  *
501  * Tear down any previously set up device DMA mappings and free
502  * any storage space allocated for them. If there are no mappings set up,
503  * this function is a NOP.
504  */
vmw_ttm_unmap_dma(struct vmw_ttm_tt * vmw_tt)505 static void vmw_ttm_unmap_dma(struct vmw_ttm_tt *vmw_tt)
506 {
507 	struct vmw_private *dev_priv = vmw_tt->dev_priv;
508 
509 	if (!vmw_tt->vsgt.sgt)
510 		return;
511 
512 	switch (dev_priv->map_mode) {
513 	case vmw_dma_map_bind:
514 	case vmw_dma_map_populate:
515 		vmw_ttm_unmap_from_dma(vmw_tt);
516 		sg_free_table(vmw_tt->vsgt.sgt);
517 		vmw_tt->vsgt.sgt = NULL;
518 		ttm_mem_global_free(vmw_mem_glob(dev_priv),
519 				    vmw_tt->sg_alloc_size);
520 		break;
521 	default:
522 		break;
523 	}
524 	vmw_tt->mapped = false;
525 }
526 
527 
528 /**
529  * vmw_bo_map_dma - Make sure buffer object pages are visible to the device
530  *
531  * @bo: Pointer to a struct ttm_buffer_object
532  *
533  * Wrapper around vmw_ttm_map_dma, that takes a TTM buffer object pointer
534  * instead of a pointer to a struct vmw_ttm_backend as argument.
535  * Note that the buffer object must be either pinned or reserved before
536  * calling this function.
537  */
vmw_bo_map_dma(struct ttm_buffer_object * bo)538 int vmw_bo_map_dma(struct ttm_buffer_object *bo)
539 {
540 	struct vmw_ttm_tt *vmw_tt =
541 		container_of(bo->ttm, struct vmw_ttm_tt, dma_ttm.ttm);
542 
543 	return vmw_ttm_map_dma(vmw_tt);
544 }
545 
546 
547 /**
548  * vmw_bo_unmap_dma - Make sure buffer object pages are visible to the device
549  *
550  * @bo: Pointer to a struct ttm_buffer_object
551  *
552  * Wrapper around vmw_ttm_unmap_dma, that takes a TTM buffer object pointer
553  * instead of a pointer to a struct vmw_ttm_backend as argument.
554  */
vmw_bo_unmap_dma(struct ttm_buffer_object * bo)555 void vmw_bo_unmap_dma(struct ttm_buffer_object *bo)
556 {
557 	struct vmw_ttm_tt *vmw_tt =
558 		container_of(bo->ttm, struct vmw_ttm_tt, dma_ttm.ttm);
559 
560 	vmw_ttm_unmap_dma(vmw_tt);
561 }
562 
563 
564 /**
565  * vmw_bo_sg_table - Return a struct vmw_sg_table object for a
566  * TTM buffer object
567  *
568  * @bo: Pointer to a struct ttm_buffer_object
569  *
570  * Returns a pointer to a struct vmw_sg_table object. The object should
571  * not be freed after use.
572  * Note that for the device addresses to be valid, the buffer object must
573  * either be reserved or pinned.
574  */
vmw_bo_sg_table(struct ttm_buffer_object * bo)575 const struct vmw_sg_table *vmw_bo_sg_table(struct ttm_buffer_object *bo)
576 {
577 	struct vmw_ttm_tt *vmw_tt =
578 		container_of(bo->ttm, struct vmw_ttm_tt, dma_ttm.ttm);
579 
580 	return &vmw_tt->vsgt;
581 }
582 
583 
vmw_ttm_bind(struct ttm_tt * ttm,struct ttm_mem_reg * bo_mem)584 static int vmw_ttm_bind(struct ttm_tt *ttm, struct ttm_mem_reg *bo_mem)
585 {
586 	struct vmw_ttm_tt *vmw_be =
587 		container_of(ttm, struct vmw_ttm_tt, dma_ttm.ttm);
588 	int ret;
589 
590 	ret = vmw_ttm_map_dma(vmw_be);
591 	if (unlikely(ret != 0))
592 		return ret;
593 
594 	vmw_be->gmr_id = bo_mem->start;
595 	vmw_be->mem_type = bo_mem->mem_type;
596 
597 	switch (bo_mem->mem_type) {
598 	case VMW_PL_GMR:
599 		return vmw_gmr_bind(vmw_be->dev_priv, &vmw_be->vsgt,
600 				    ttm->num_pages, vmw_be->gmr_id);
601 	case VMW_PL_MOB:
602 		if (unlikely(vmw_be->mob == NULL)) {
603 			vmw_be->mob =
604 				vmw_mob_create(ttm->num_pages);
605 			if (unlikely(vmw_be->mob == NULL))
606 				return -ENOMEM;
607 		}
608 
609 		return vmw_mob_bind(vmw_be->dev_priv, vmw_be->mob,
610 				    &vmw_be->vsgt, ttm->num_pages,
611 				    vmw_be->gmr_id);
612 	default:
613 		BUG();
614 	}
615 	return 0;
616 }
617 
vmw_ttm_unbind(struct ttm_tt * ttm)618 static int vmw_ttm_unbind(struct ttm_tt *ttm)
619 {
620 	struct vmw_ttm_tt *vmw_be =
621 		container_of(ttm, struct vmw_ttm_tt, dma_ttm.ttm);
622 
623 	switch (vmw_be->mem_type) {
624 	case VMW_PL_GMR:
625 		vmw_gmr_unbind(vmw_be->dev_priv, vmw_be->gmr_id);
626 		break;
627 	case VMW_PL_MOB:
628 		vmw_mob_unbind(vmw_be->dev_priv, vmw_be->mob);
629 		break;
630 	default:
631 		BUG();
632 	}
633 
634 	if (vmw_be->dev_priv->map_mode == vmw_dma_map_bind)
635 		vmw_ttm_unmap_dma(vmw_be);
636 
637 	return 0;
638 }
639 
640 
vmw_ttm_destroy(struct ttm_tt * ttm)641 static void vmw_ttm_destroy(struct ttm_tt *ttm)
642 {
643 	struct vmw_ttm_tt *vmw_be =
644 		container_of(ttm, struct vmw_ttm_tt, dma_ttm.ttm);
645 
646 	vmw_ttm_unmap_dma(vmw_be);
647 	if (vmw_be->dev_priv->map_mode == vmw_dma_alloc_coherent)
648 		ttm_dma_tt_fini(&vmw_be->dma_ttm);
649 	else
650 		ttm_tt_fini(ttm);
651 
652 	if (vmw_be->mob)
653 		vmw_mob_destroy(vmw_be->mob);
654 
655 	kfree(vmw_be);
656 }
657 
658 
vmw_ttm_populate(struct ttm_tt * ttm,struct ttm_operation_ctx * ctx)659 static int vmw_ttm_populate(struct ttm_tt *ttm, struct ttm_operation_ctx *ctx)
660 {
661 	struct vmw_ttm_tt *vmw_tt =
662 		container_of(ttm, struct vmw_ttm_tt, dma_ttm.ttm);
663 	struct vmw_private *dev_priv = vmw_tt->dev_priv;
664 	struct ttm_mem_global *glob = vmw_mem_glob(dev_priv);
665 	int ret;
666 
667 	if (ttm->state != tt_unpopulated)
668 		return 0;
669 
670 	if (dev_priv->map_mode == vmw_dma_alloc_coherent) {
671 		size_t size =
672 			ttm_round_pot(ttm->num_pages * sizeof(dma_addr_t));
673 		ret = ttm_mem_global_alloc(glob, size, ctx);
674 		if (unlikely(ret != 0))
675 			return ret;
676 
677 		ret = ttm_dma_populate(&vmw_tt->dma_ttm, dev_priv->dev->dev,
678 					ctx);
679 		if (unlikely(ret != 0))
680 			ttm_mem_global_free(glob, size);
681 	} else
682 		ret = ttm_pool_populate(ttm, ctx);
683 
684 	return ret;
685 }
686 
vmw_ttm_unpopulate(struct ttm_tt * ttm)687 static void vmw_ttm_unpopulate(struct ttm_tt *ttm)
688 {
689 	struct vmw_ttm_tt *vmw_tt = container_of(ttm, struct vmw_ttm_tt,
690 						 dma_ttm.ttm);
691 	struct vmw_private *dev_priv = vmw_tt->dev_priv;
692 	struct ttm_mem_global *glob = vmw_mem_glob(dev_priv);
693 
694 
695 	if (vmw_tt->mob) {
696 		vmw_mob_destroy(vmw_tt->mob);
697 		vmw_tt->mob = NULL;
698 	}
699 
700 	vmw_ttm_unmap_dma(vmw_tt);
701 	if (dev_priv->map_mode == vmw_dma_alloc_coherent) {
702 		size_t size =
703 			ttm_round_pot(ttm->num_pages * sizeof(dma_addr_t));
704 
705 		ttm_dma_unpopulate(&vmw_tt->dma_ttm, dev_priv->dev->dev);
706 		ttm_mem_global_free(glob, size);
707 	} else
708 		ttm_pool_unpopulate(ttm);
709 }
710 
711 static struct ttm_backend_func vmw_ttm_func = {
712 	.bind = vmw_ttm_bind,
713 	.unbind = vmw_ttm_unbind,
714 	.destroy = vmw_ttm_destroy,
715 };
716 
vmw_ttm_tt_create(struct ttm_buffer_object * bo,uint32_t page_flags)717 static struct ttm_tt *vmw_ttm_tt_create(struct ttm_buffer_object *bo,
718 					uint32_t page_flags)
719 {
720 	struct vmw_ttm_tt *vmw_be;
721 	int ret;
722 
723 	vmw_be = kzalloc(sizeof(*vmw_be), GFP_KERNEL);
724 	if (!vmw_be)
725 		return NULL;
726 
727 	vmw_be->dma_ttm.ttm.func = &vmw_ttm_func;
728 	vmw_be->dev_priv = container_of(bo->bdev, struct vmw_private, bdev);
729 	vmw_be->mob = NULL;
730 
731 	if (vmw_be->dev_priv->map_mode == vmw_dma_alloc_coherent)
732 		ret = ttm_dma_tt_init(&vmw_be->dma_ttm, bo, page_flags);
733 	else
734 		ret = ttm_tt_init(&vmw_be->dma_ttm.ttm, bo, page_flags);
735 	if (unlikely(ret != 0))
736 		goto out_no_init;
737 
738 	return &vmw_be->dma_ttm.ttm;
739 out_no_init:
740 	kfree(vmw_be);
741 	return NULL;
742 }
743 
vmw_invalidate_caches(struct ttm_bo_device * bdev,uint32_t flags)744 static int vmw_invalidate_caches(struct ttm_bo_device *bdev, uint32_t flags)
745 {
746 	return 0;
747 }
748 
vmw_init_mem_type(struct ttm_bo_device * bdev,uint32_t type,struct ttm_mem_type_manager * man)749 static int vmw_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
750 		      struct ttm_mem_type_manager *man)
751 {
752 	switch (type) {
753 	case TTM_PL_SYSTEM:
754 		/* System memory */
755 
756 		man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
757 		man->available_caching = TTM_PL_FLAG_CACHED;
758 		man->default_caching = TTM_PL_FLAG_CACHED;
759 		break;
760 	case TTM_PL_VRAM:
761 		/* "On-card" video ram */
762 		man->func = &ttm_bo_manager_func;
763 		man->gpu_offset = 0;
764 		man->flags = TTM_MEMTYPE_FLAG_FIXED | TTM_MEMTYPE_FLAG_MAPPABLE;
765 		man->available_caching = TTM_PL_FLAG_CACHED;
766 		man->default_caching = TTM_PL_FLAG_CACHED;
767 		break;
768 	case VMW_PL_GMR:
769 	case VMW_PL_MOB:
770 		/*
771 		 * "Guest Memory Regions" is an aperture like feature with
772 		 *  one slot per bo. There is an upper limit of the number of
773 		 *  slots as well as the bo size.
774 		 */
775 		man->func = &vmw_gmrid_manager_func;
776 		man->gpu_offset = 0;
777 		man->flags = TTM_MEMTYPE_FLAG_CMA | TTM_MEMTYPE_FLAG_MAPPABLE;
778 		man->available_caching = TTM_PL_FLAG_CACHED;
779 		man->default_caching = TTM_PL_FLAG_CACHED;
780 		break;
781 	default:
782 		DRM_ERROR("Unsupported memory type %u\n", (unsigned)type);
783 		return -EINVAL;
784 	}
785 	return 0;
786 }
787 
vmw_evict_flags(struct ttm_buffer_object * bo,struct ttm_placement * placement)788 static void vmw_evict_flags(struct ttm_buffer_object *bo,
789 		     struct ttm_placement *placement)
790 {
791 	*placement = vmw_sys_placement;
792 }
793 
vmw_verify_access(struct ttm_buffer_object * bo,struct file * filp)794 static int vmw_verify_access(struct ttm_buffer_object *bo, struct file *filp)
795 {
796 	struct ttm_object_file *tfile =
797 		vmw_fpriv((struct drm_file *)filp->private_data)->tfile;
798 
799 	return vmw_user_bo_verify_access(bo, tfile);
800 }
801 
vmw_ttm_io_mem_reserve(struct ttm_bo_device * bdev,struct ttm_mem_reg * mem)802 static int vmw_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
803 {
804 	struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
805 	struct vmw_private *dev_priv = container_of(bdev, struct vmw_private, bdev);
806 
807 	mem->bus.addr = NULL;
808 	mem->bus.is_iomem = false;
809 	mem->bus.offset = 0;
810 	mem->bus.size = mem->num_pages << PAGE_SHIFT;
811 	mem->bus.base = 0;
812 	if (!(man->flags & TTM_MEMTYPE_FLAG_MAPPABLE))
813 		return -EINVAL;
814 	switch (mem->mem_type) {
815 	case TTM_PL_SYSTEM:
816 	case VMW_PL_GMR:
817 	case VMW_PL_MOB:
818 		return 0;
819 	case TTM_PL_VRAM:
820 		mem->bus.offset = mem->start << PAGE_SHIFT;
821 		mem->bus.base = dev_priv->vram_start;
822 		mem->bus.is_iomem = true;
823 		break;
824 	default:
825 		return -EINVAL;
826 	}
827 	return 0;
828 }
829 
vmw_ttm_io_mem_free(struct ttm_bo_device * bdev,struct ttm_mem_reg * mem)830 static void vmw_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
831 {
832 }
833 
vmw_ttm_fault_reserve_notify(struct ttm_buffer_object * bo)834 static int vmw_ttm_fault_reserve_notify(struct ttm_buffer_object *bo)
835 {
836 	return 0;
837 }
838 
839 /**
840  * vmw_move_notify - TTM move_notify_callback
841  *
842  * @bo: The TTM buffer object about to move.
843  * @mem: The struct ttm_mem_reg indicating to what memory
844  *       region the move is taking place.
845  *
846  * Calls move_notify for all subsystems needing it.
847  * (currently only resources).
848  */
vmw_move_notify(struct ttm_buffer_object * bo,bool evict,struct ttm_mem_reg * mem)849 static void vmw_move_notify(struct ttm_buffer_object *bo,
850 			    bool evict,
851 			    struct ttm_mem_reg *mem)
852 {
853 	vmw_bo_move_notify(bo, mem);
854 	vmw_query_move_notify(bo, mem);
855 }
856 
857 
858 /**
859  * vmw_swap_notify - TTM move_notify_callback
860  *
861  * @bo: The TTM buffer object about to be swapped out.
862  */
vmw_swap_notify(struct ttm_buffer_object * bo)863 static void vmw_swap_notify(struct ttm_buffer_object *bo)
864 {
865 	vmw_bo_swap_notify(bo);
866 	(void) ttm_bo_wait(bo, false, false);
867 }
868 
869 
870 struct ttm_bo_driver vmw_bo_driver = {
871 	.ttm_tt_create = &vmw_ttm_tt_create,
872 	.ttm_tt_populate = &vmw_ttm_populate,
873 	.ttm_tt_unpopulate = &vmw_ttm_unpopulate,
874 	.invalidate_caches = vmw_invalidate_caches,
875 	.init_mem_type = vmw_init_mem_type,
876 	.eviction_valuable = ttm_bo_eviction_valuable,
877 	.evict_flags = vmw_evict_flags,
878 	.move = NULL,
879 	.verify_access = vmw_verify_access,
880 	.move_notify = vmw_move_notify,
881 	.swap_notify = vmw_swap_notify,
882 	.fault_reserve_notify = &vmw_ttm_fault_reserve_notify,
883 	.io_mem_reserve = &vmw_ttm_io_mem_reserve,
884 	.io_mem_free = &vmw_ttm_io_mem_free,
885 };
886