1 /*	$NetBSD: radeon_ttm.c,v 1.26 2022/07/20 01:22:38 riastradh Exp $	*/
2 
3 /*
4  * Copyright 2009 Jerome Glisse.
5  * All Rights Reserved.
6  *
7  * Permission is hereby granted, free of charge, to any person obtaining a
8  * copy of this software and associated documentation files (the
9  * "Software"), to deal in the Software without restriction, including
10  * without limitation the rights to use, copy, modify, merge, publish,
11  * distribute, sub license, and/or sell copies of the Software, and to
12  * permit persons to whom the Software is furnished to do so, subject to
13  * the following conditions:
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
18  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
19  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
20  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
21  * USE OR OTHER DEALINGS IN THE SOFTWARE.
22  *
23  * The above copyright notice and this permission notice (including the
24  * next paragraph) shall be included in all copies or substantial portions
25  * of the Software.
26  *
27  */
28 /*
29  * Authors:
30  *    Jerome Glisse <glisse@freedesktop.org>
31  *    Thomas Hellstrom <thomas-at-tungstengraphics-dot-com>
32  *    Dave Airlie
33  */
34 
35 #include <sys/cdefs.h>
36 __KERNEL_RCSID(0, "$NetBSD: radeon_ttm.c,v 1.26 2022/07/20 01:22:38 riastradh Exp $");
37 
38 #include <linux/dma-mapping.h>
39 #include <linux/pagemap.h>
40 #include <linux/pci.h>
41 #include <linux/seq_file.h>
42 #include <linux/slab.h>
43 #include <linux/swap.h>
44 #include <linux/swiotlb.h>
45 
46 #include <drm/drm_agpsupport.h>
47 #include <drm/drm_debugfs.h>
48 #include <drm/drm_device.h>
49 #include <drm/drm_file.h>
50 #include <drm/drm_prime.h>
51 #include <drm/radeon_drm.h>
52 #include <drm/ttm/ttm_bo_api.h>
53 #include <drm/ttm/ttm_bo_driver.h>
54 #include <drm/ttm/ttm_module.h>
55 #include <drm/ttm/ttm_page_alloc.h>
56 #include <drm/ttm/ttm_placement.h>
57 
58 #include "radeon_reg.h"
59 #include "radeon.h"
60 
61 #ifdef __NetBSD__
62 #include <uvm/uvm_extern.h>
63 #include <uvm/uvm_fault.h>
64 #include <uvm/uvm_param.h>
65 #include <drm/bus_dma_hacks.h>
66 #endif
67 
68 static int radeon_ttm_debugfs_init(struct radeon_device *rdev);
69 static void radeon_ttm_debugfs_fini(struct radeon_device *rdev);
70 
radeon_get_rdev(struct ttm_bo_device * bdev)71 static struct radeon_device *radeon_get_rdev(struct ttm_bo_device *bdev)
72 {
73 	struct radeon_mman *mman;
74 	struct radeon_device *rdev;
75 
76 	mman = container_of(bdev, struct radeon_mman, bdev);
77 	rdev = container_of(mman, struct radeon_device, mman);
78 	return rdev;
79 }
80 
radeon_invalidate_caches(struct ttm_bo_device * bdev,uint32_t flags)81 static int radeon_invalidate_caches(struct ttm_bo_device *bdev, uint32_t flags)
82 {
83 	return 0;
84 }
85 
radeon_init_mem_type(struct ttm_bo_device * bdev,uint32_t type,struct ttm_mem_type_manager * man)86 static int radeon_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
87 				struct ttm_mem_type_manager *man)
88 {
89 	struct radeon_device *rdev;
90 
91 	rdev = radeon_get_rdev(bdev);
92 
93 	switch (type) {
94 	case TTM_PL_SYSTEM:
95 		/* System memory */
96 		man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
97 		man->available_caching = TTM_PL_MASK_CACHING;
98 		man->default_caching = TTM_PL_FLAG_CACHED;
99 		break;
100 	case TTM_PL_TT:
101 		man->func = &ttm_bo_manager_func;
102 		man->gpu_offset = rdev->mc.gtt_start;
103 		man->available_caching = TTM_PL_MASK_CACHING;
104 		man->default_caching = TTM_PL_FLAG_CACHED;
105 		man->flags = TTM_MEMTYPE_FLAG_MAPPABLE | TTM_MEMTYPE_FLAG_CMA;
106 #if IS_ENABLED(CONFIG_AGP)
107 		if (rdev->flags & RADEON_IS_AGP) {
108 			if (!rdev->ddev->agp) {
109 				DRM_ERROR("AGP is not enabled for memory type %u\n",
110 					  (unsigned)type);
111 				return -EINVAL;
112 			}
113 			if (!rdev->ddev->agp->cant_use_aperture)
114 				man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
115 			man->available_caching = TTM_PL_FLAG_UNCACHED |
116 						 TTM_PL_FLAG_WC;
117 			man->default_caching = TTM_PL_FLAG_WC;
118 		}
119 #endif
120 		break;
121 	case TTM_PL_VRAM:
122 		/* "On-card" video ram */
123 		man->func = &ttm_bo_manager_func;
124 		man->gpu_offset = rdev->mc.vram_start;
125 		man->flags = TTM_MEMTYPE_FLAG_FIXED |
126 			     TTM_MEMTYPE_FLAG_MAPPABLE;
127 		man->available_caching = TTM_PL_FLAG_UNCACHED | TTM_PL_FLAG_WC;
128 		man->default_caching = TTM_PL_FLAG_WC;
129 		break;
130 	default:
131 		DRM_ERROR("Unsupported memory type %u\n", (unsigned)type);
132 		return -EINVAL;
133 	}
134 	return 0;
135 }
136 
radeon_evict_flags(struct ttm_buffer_object * bo,struct ttm_placement * placement)137 static void radeon_evict_flags(struct ttm_buffer_object *bo,
138 				struct ttm_placement *placement)
139 {
140 	static const struct ttm_place placements = {
141 		.fpfn = 0,
142 		.lpfn = 0,
143 		.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM
144 	};
145 
146 	struct radeon_bo *rbo;
147 
148 	if (!radeon_ttm_bo_is_radeon_bo(bo)) {
149 		placement->placement = &placements;
150 		placement->busy_placement = &placements;
151 		placement->num_placement = 1;
152 		placement->num_busy_placement = 1;
153 		return;
154 	}
155 	rbo = container_of(bo, struct radeon_bo, tbo);
156 	switch (bo->mem.mem_type) {
157 	case TTM_PL_VRAM:
158 		if (rbo->rdev->ring[radeon_copy_ring_index(rbo->rdev)].ready == false)
159 			radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_CPU);
160 		else if (rbo->rdev->mc.visible_vram_size < rbo->rdev->mc.real_vram_size &&
161 			 bo->mem.start < (rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT)) {
162 			unsigned fpfn = rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
163 			int i;
164 
165 			/* Try evicting to the CPU inaccessible part of VRAM
166 			 * first, but only set GTT as busy placement, so this
167 			 * BO will be evicted to GTT rather than causing other
168 			 * BOs to be evicted from VRAM
169 			 */
170 			radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_VRAM |
171 							 RADEON_GEM_DOMAIN_GTT);
172 			rbo->placement.num_busy_placement = 0;
173 			for (i = 0; i < rbo->placement.num_placement; i++) {
174 				if (rbo->placements[i].flags & TTM_PL_FLAG_VRAM) {
175 					if (rbo->placements[i].fpfn < fpfn)
176 						rbo->placements[i].fpfn = fpfn;
177 				} else {
178 					rbo->placement.busy_placement =
179 						&rbo->placements[i];
180 					rbo->placement.num_busy_placement = 1;
181 				}
182 			}
183 		} else
184 			radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_GTT);
185 		break;
186 	case TTM_PL_TT:
187 	default:
188 		radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_CPU);
189 	}
190 	*placement = rbo->placement;
191 }
192 
radeon_verify_access(struct ttm_buffer_object * bo,struct file * filp)193 static int radeon_verify_access(struct ttm_buffer_object *bo, struct file *filp)
194 {
195 	struct radeon_bo *rbo = container_of(bo, struct radeon_bo, tbo);
196 
197 	if (radeon_ttm_tt_has_userptr(bo->ttm))
198 		return -EPERM;
199 #ifdef __NetBSD__
200 	struct drm_file *drm_file = filp->f_data;
201 	return drm_vma_node_verify_access(&rbo->tbo.base.vma_node, drm_file);
202 #else
203 	return drm_vma_node_verify_access(&rbo->tbo.base.vma_node,
204 					  filp->private_data);
205 #endif
206 }
207 
radeon_move_null(struct ttm_buffer_object * bo,struct ttm_mem_reg * new_mem)208 static void radeon_move_null(struct ttm_buffer_object *bo,
209 			     struct ttm_mem_reg *new_mem)
210 {
211 	struct ttm_mem_reg *old_mem = &bo->mem;
212 
213 	BUG_ON(old_mem->mm_node != NULL);
214 	*old_mem = *new_mem;
215 	new_mem->mm_node = NULL;
216 }
217 
radeon_move_blit(struct ttm_buffer_object * bo,bool evict,bool no_wait_gpu,struct ttm_mem_reg * new_mem,struct ttm_mem_reg * old_mem)218 static int radeon_move_blit(struct ttm_buffer_object *bo,
219 			bool evict, bool no_wait_gpu,
220 			struct ttm_mem_reg *new_mem,
221 			struct ttm_mem_reg *old_mem)
222 {
223 	struct radeon_device *rdev;
224 	uint64_t old_start, new_start;
225 	struct radeon_fence *fence;
226 	unsigned num_pages;
227 	int r, ridx;
228 
229 	rdev = radeon_get_rdev(bo->bdev);
230 	ridx = radeon_copy_ring_index(rdev);
231 	old_start = (u64)old_mem->start << PAGE_SHIFT;
232 	new_start = (u64)new_mem->start << PAGE_SHIFT;
233 
234 	switch (old_mem->mem_type) {
235 	case TTM_PL_VRAM:
236 		old_start += rdev->mc.vram_start;
237 		break;
238 	case TTM_PL_TT:
239 		old_start += rdev->mc.gtt_start;
240 		break;
241 	default:
242 		DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
243 		return -EINVAL;
244 	}
245 	switch (new_mem->mem_type) {
246 	case TTM_PL_VRAM:
247 		new_start += rdev->mc.vram_start;
248 		break;
249 	case TTM_PL_TT:
250 		new_start += rdev->mc.gtt_start;
251 		break;
252 	default:
253 		DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
254 		return -EINVAL;
255 	}
256 	if (!rdev->ring[ridx].ready) {
257 		DRM_ERROR("Trying to move memory with ring turned off.\n");
258 		return -EINVAL;
259 	}
260 
261 	BUILD_BUG_ON((PAGE_SIZE % RADEON_GPU_PAGE_SIZE) != 0);
262 
263 	num_pages = new_mem->num_pages * (PAGE_SIZE / RADEON_GPU_PAGE_SIZE);
264 	fence = radeon_copy(rdev, old_start, new_start, num_pages, bo->base.resv);
265 	if (IS_ERR(fence))
266 		return PTR_ERR(fence);
267 
268 	r = ttm_bo_move_accel_cleanup(bo, &fence->base, evict, new_mem);
269 	radeon_fence_unref(&fence);
270 	return r;
271 }
272 
radeon_move_vram_ram(struct ttm_buffer_object * bo,bool evict,bool interruptible,bool no_wait_gpu,struct ttm_mem_reg * new_mem)273 static int radeon_move_vram_ram(struct ttm_buffer_object *bo,
274 				bool evict, bool interruptible,
275 				bool no_wait_gpu,
276 				struct ttm_mem_reg *new_mem)
277 {
278 	struct ttm_operation_ctx ctx = { interruptible, no_wait_gpu };
279 	struct ttm_mem_reg *old_mem = &bo->mem;
280 	struct ttm_mem_reg tmp_mem;
281 	struct ttm_place placements;
282 	struct ttm_placement placement;
283 	int r;
284 
285 	tmp_mem = *new_mem;
286 	tmp_mem.mm_node = NULL;
287 	placement.num_placement = 1;
288 	placement.placement = &placements;
289 	placement.num_busy_placement = 1;
290 	placement.busy_placement = &placements;
291 	placements.fpfn = 0;
292 	placements.lpfn = 0;
293 	placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
294 	r = ttm_bo_mem_space(bo, &placement, &tmp_mem, &ctx);
295 	if (unlikely(r)) {
296 		return r;
297 	}
298 
299 	r = ttm_tt_set_placement_caching(bo->ttm, tmp_mem.placement);
300 	if (unlikely(r)) {
301 		goto out_cleanup;
302 	}
303 
304 	r = ttm_tt_bind(bo->ttm, &tmp_mem, &ctx);
305 	if (unlikely(r)) {
306 		goto out_cleanup;
307 	}
308 	r = radeon_move_blit(bo, true, no_wait_gpu, &tmp_mem, old_mem);
309 	if (unlikely(r)) {
310 		goto out_cleanup;
311 	}
312 	r = ttm_bo_move_ttm(bo, &ctx, new_mem);
313 out_cleanup:
314 	ttm_bo_mem_put(bo, &tmp_mem);
315 	return r;
316 }
317 
radeon_move_ram_vram(struct ttm_buffer_object * bo,bool evict,bool interruptible,bool no_wait_gpu,struct ttm_mem_reg * new_mem)318 static int radeon_move_ram_vram(struct ttm_buffer_object *bo,
319 				bool evict, bool interruptible,
320 				bool no_wait_gpu,
321 				struct ttm_mem_reg *new_mem)
322 {
323 	struct ttm_operation_ctx ctx = { interruptible, no_wait_gpu };
324 	struct ttm_mem_reg *old_mem = &bo->mem;
325 	struct ttm_mem_reg tmp_mem;
326 	struct ttm_placement placement;
327 	struct ttm_place placements;
328 	int r;
329 
330 	tmp_mem = *new_mem;
331 	tmp_mem.mm_node = NULL;
332 	placement.num_placement = 1;
333 	placement.placement = &placements;
334 	placement.num_busy_placement = 1;
335 	placement.busy_placement = &placements;
336 	placements.fpfn = 0;
337 	placements.lpfn = 0;
338 	placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
339 	r = ttm_bo_mem_space(bo, &placement, &tmp_mem, &ctx);
340 	if (unlikely(r)) {
341 		return r;
342 	}
343 	r = ttm_bo_move_ttm(bo, &ctx, &tmp_mem);
344 	if (unlikely(r)) {
345 		goto out_cleanup;
346 	}
347 	r = radeon_move_blit(bo, true, no_wait_gpu, new_mem, old_mem);
348 	if (unlikely(r)) {
349 		goto out_cleanup;
350 	}
351 out_cleanup:
352 	ttm_bo_mem_put(bo, &tmp_mem);
353 	return r;
354 }
355 
radeon_bo_move(struct ttm_buffer_object * bo,bool evict,struct ttm_operation_ctx * ctx,struct ttm_mem_reg * new_mem)356 static int radeon_bo_move(struct ttm_buffer_object *bo, bool evict,
357 			  struct ttm_operation_ctx *ctx,
358 			  struct ttm_mem_reg *new_mem)
359 {
360 	struct radeon_device *rdev;
361 	struct radeon_bo *rbo;
362 	struct ttm_mem_reg *old_mem = &bo->mem;
363 	int r;
364 
365 	r = ttm_bo_wait(bo, ctx->interruptible, ctx->no_wait_gpu);
366 	if (r)
367 		return r;
368 
369 	/* Can't move a pinned BO */
370 	rbo = container_of(bo, struct radeon_bo, tbo);
371 	if (WARN_ON_ONCE(rbo->pin_count > 0))
372 		return -EINVAL;
373 
374 	rdev = radeon_get_rdev(bo->bdev);
375 	if (old_mem->mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
376 		radeon_move_null(bo, new_mem);
377 		return 0;
378 	}
379 	if ((old_mem->mem_type == TTM_PL_TT &&
380 	     new_mem->mem_type == TTM_PL_SYSTEM) ||
381 	    (old_mem->mem_type == TTM_PL_SYSTEM &&
382 	     new_mem->mem_type == TTM_PL_TT)) {
383 		/* bind is enough */
384 		radeon_move_null(bo, new_mem);
385 		return 0;
386 	}
387 	if (!rdev->ring[radeon_copy_ring_index(rdev)].ready ||
388 	    rdev->asic->copy.copy == NULL) {
389 		/* use memcpy */
390 		goto memcpy;
391 	}
392 
393 	if (old_mem->mem_type == TTM_PL_VRAM &&
394 	    new_mem->mem_type == TTM_PL_SYSTEM) {
395 		r = radeon_move_vram_ram(bo, evict, ctx->interruptible,
396 					ctx->no_wait_gpu, new_mem);
397 	} else if (old_mem->mem_type == TTM_PL_SYSTEM &&
398 		   new_mem->mem_type == TTM_PL_VRAM) {
399 		r = radeon_move_ram_vram(bo, evict, ctx->interruptible,
400 					    ctx->no_wait_gpu, new_mem);
401 	} else {
402 		r = radeon_move_blit(bo, evict, ctx->no_wait_gpu,
403 				     new_mem, old_mem);
404 	}
405 
406 	if (r) {
407 memcpy:
408 		r = ttm_bo_move_memcpy(bo, ctx, new_mem);
409 		if (r) {
410 			return r;
411 		}
412 	}
413 
414 	/* update statistics */
415 	atomic64_add((u64)bo->num_pages << PAGE_SHIFT, &rdev->num_bytes_moved);
416 	return 0;
417 }
418 
radeon_ttm_io_mem_reserve(struct ttm_bo_device * bdev,struct ttm_mem_reg * mem)419 static int radeon_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
420 {
421 	struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
422 	struct radeon_device *rdev = radeon_get_rdev(bdev);
423 
424 	mem->bus.addr = NULL;
425 	mem->bus.offset = 0;
426 	mem->bus.size = mem->num_pages << PAGE_SHIFT;
427 	mem->bus.base = 0;
428 	mem->bus.is_iomem = false;
429 	if (!(man->flags & TTM_MEMTYPE_FLAG_MAPPABLE))
430 		return -EINVAL;
431 	switch (mem->mem_type) {
432 	case TTM_PL_SYSTEM:
433 		/* system memory */
434 		return 0;
435 	case TTM_PL_TT:
436 #if IS_ENABLED(CONFIG_AGP)
437 		if (rdev->flags & RADEON_IS_AGP) {
438 			/* RADEON_IS_AGP is set only if AGP is active */
439 			mem->bus.offset = mem->start << PAGE_SHIFT;
440 			mem->bus.base = rdev->mc.agp_base;
441 			mem->bus.is_iomem = !rdev->ddev->agp->cant_use_aperture;
442 			KASSERTMSG((mem->bus.base & (PAGE_SIZE - 1)) == 0,
443 			    "agp aperture is not page-aligned: %" PRIx64 "",
444 			    (uint64_t)mem->bus.base);
445 			KASSERT((mem->bus.offset & (PAGE_SIZE - 1)) == 0);
446 		}
447 #endif
448 		break;
449 	case TTM_PL_VRAM:
450 		mem->bus.offset = mem->start << PAGE_SHIFT;
451 		/* check if it's visible */
452 		if ((mem->bus.offset + mem->bus.size) > rdev->mc.visible_vram_size)
453 			return -EINVAL;
454 		mem->bus.base = rdev->mc.aper_base;
455 		mem->bus.is_iomem = true;
456 #ifndef __NetBSD__		/* alpha hose handled through bus_space(9) */
457 #ifdef __alpha__
458 		/*
459 		 * Alpha: use bus.addr to hold the ioremap() return,
460 		 * so we can modify bus.base below.
461 		 */
462 		if (mem->placement & TTM_PL_FLAG_WC)
463 			mem->bus.addr =
464 				ioremap_wc(mem->bus.base + mem->bus.offset,
465 					   mem->bus.size);
466 		else
467 			mem->bus.addr =
468 				ioremap(mem->bus.base + mem->bus.offset,
469 						mem->bus.size);
470 		if (!mem->bus.addr)
471 			return -ENOMEM;
472 
473 		/*
474 		 * Alpha: Use just the bus offset plus
475 		 * the hose/domain memory base for bus.base.
476 		 * It then can be used to build PTEs for VRAM
477 		 * access, as done in ttm_bo_vm_fault().
478 		 */
479 		mem->bus.base = (mem->bus.base & 0x0ffffffffUL) +
480 			rdev->ddev->hose->dense_mem_base;
481 #endif
482 #endif
483 		KASSERTMSG((mem->bus.base & (PAGE_SIZE - 1)) == 0,
484 		    "mc aperture is not page-aligned: %" PRIx64 "",
485 		    (uint64_t)mem->bus.base);
486 		KASSERT((mem->bus.offset & (PAGE_SIZE - 1)) == 0);
487 		break;
488 	default:
489 		return -EINVAL;
490 	}
491 	return 0;
492 }
493 
radeon_ttm_io_mem_free(struct ttm_bo_device * bdev,struct ttm_mem_reg * mem)494 static void radeon_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
495 {
496 }
497 
498 /*
499  * TTM backend functions.
500  */
501 struct radeon_ttm_tt {
502 	struct ttm_dma_tt		ttm;
503 	struct radeon_device		*rdev;
504 	u64				offset;
505 
506 	uint64_t			userptr;
507 #ifdef __NetBSD__
508 	struct vmspace			*usermm;
509 #else
510 	struct mm_struct		*usermm;
511 #endif
512 	uint32_t			userflags;
513 };
514 
515 /* prepare the sg table with the user pages */
radeon_ttm_tt_pin_userptr(struct ttm_tt * ttm)516 static int radeon_ttm_tt_pin_userptr(struct ttm_tt *ttm)
517 {
518 	struct radeon_device *rdev = radeon_get_rdev(ttm->bdev);
519 	struct radeon_ttm_tt *gtt = (void *)ttm;
520 #ifndef __NetBSD__
521 	unsigned pinned = 0, nents;
522 #endif
523 	int r;
524 
525 	int write = !(gtt->userflags & RADEON_GEM_USERPTR_READONLY);
526 #ifndef __NetBSD__
527 	enum dma_data_direction direction = write ?
528 		DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
529 #endif
530 
531 #ifdef __NetBSD__
532 	if (curproc->p_vmspace != gtt->usermm)
533 		return -EPERM;
534 #else
535 	if (current->mm != gtt->usermm)
536 		return -EPERM;
537 #endif
538 
539 	if (gtt->userflags & RADEON_GEM_USERPTR_ANONONLY) {
540 		/* check that we only pin down anonymous memory
541 		   to prevent problems with writeback */
542 		unsigned long end = gtt->userptr + ttm->num_pages * PAGE_SIZE;
543 #ifdef __NetBSD__
544 		/* XXX ???  TOCTOU, anyone?  */
545 		/* XXX should do range_test */
546 		struct vm_map_entry *entry;
547 		bool ok;
548 		vm_map_lock_read(&gtt->usermm->vm_map);
549 		ok = uvm_map_lookup_entry(&gtt->usermm->vm_map,
550 		    (vaddr_t)gtt->userptr, &entry);
551 		if (ok)
552 			ok = !UVM_ET_ISOBJ(entry) && end <= entry->end;
553 		vm_map_unlock_read(&gtt->usermm->vm_map);
554 		if (!ok)
555 			return -EPERM;
556 #else
557 		struct vm_area_struct *vma;
558 		vma = find_vma(gtt->usermm, gtt->userptr);
559 		if (!vma || vma->vm_file || vma->vm_end < end)
560 			return -EPERM;
561 #endif
562 	}
563 
564 #ifdef __NetBSD__
565 	struct iovec iov = {
566 		.iov_base = (void *)(vaddr_t)gtt->userptr,
567 		.iov_len = ttm->num_pages << PAGE_SHIFT,
568 	};
569 	struct uio uio = {
570 		.uio_iov = &iov,
571 		.uio_iovcnt = 1,
572 		.uio_offset = 0,
573 		.uio_resid = ttm->num_pages << PAGE_SHIFT,
574 		.uio_rw = (write ? UIO_READ : UIO_WRITE), /* XXX ??? */
575 		.uio_vmspace = gtt->usermm,
576 	};
577 	unsigned long i;
578 
579 	/* Wire the relevant part of the user's address space.  */
580 	/* XXX What happens if user does munmap?  */
581 	/* XXX errno NetBSD->Linux */
582 	r = -uvm_vslock(gtt->usermm, (void *)(vaddr_t)gtt->userptr,
583 	    ttm->num_pages << PAGE_SHIFT,
584 	    (write ? VM_PROT_WRITE : VM_PROT_READ)); /* XXX ??? */
585 	if (r)
586 		goto fail0;
587 
588 	/* Load it up for DMA.  */
589 	/* XXX errno NetBSD->Linux */
590 	r = -bus_dmamap_load_uio(rdev->ddev->dmat, gtt->ttm.dma_address, &uio,
591 	    BUS_DMA_WAITOK);
592 	if (r)
593 		goto fail1;
594 
595 	/* Get each of the pages as ttm requests.  */
596 	for (i = 0; i < ttm->num_pages; i++) {
597 		vaddr_t va = (vaddr_t)gtt->userptr + (i << PAGE_SHIFT);
598 		paddr_t pa;
599 		struct vm_page *vmp;
600 
601 		if (!pmap_extract(gtt->usermm->vm_map.pmap, va, &pa)) {
602 			r = -EFAULT;
603 			goto fail2;
604 		}
605 		vmp = PHYS_TO_VM_PAGE(pa);
606 		ttm->pages[i] = container_of(vmp, struct page, p_vmp);
607 	}
608 
609 	/* Success!  */
610 	return 0;
611 
612 fail2:	while (i --> 0)
613 		ttm->pages[i] = NULL; /* paranoia */
614 	bus_dmamap_unload(rdev->ddev->dmat, gtt->ttm.dma_address);
615 fail1:	uvm_vsunlock(gtt->usermm, (void *)(vaddr_t)gtt->userptr,
616 	    ttm->num_pages << PAGE_SHIFT);
617 fail0:	return r;
618 #else
619 	do {
620 		unsigned num_pages = ttm->num_pages - pinned;
621 		uint64_t userptr = gtt->userptr + pinned * PAGE_SIZE;
622 		struct page **pages = ttm->pages + pinned;
623 
624 		r = get_user_pages(userptr, num_pages, write ? FOLL_WRITE : 0,
625 				   pages, NULL);
626 		if (r < 0)
627 			goto release_pages;
628 
629 		pinned += r;
630 
631 	} while (pinned < ttm->num_pages);
632 
633 	r = sg_alloc_table_from_pages(ttm->sg, ttm->pages, ttm->num_pages, 0,
634 				      ttm->num_pages << PAGE_SHIFT,
635 				      GFP_KERNEL);
636 	if (r)
637 		goto release_sg;
638 
639 	r = -ENOMEM;
640 	nents = dma_map_sg(rdev->dev, ttm->sg->sgl, ttm->sg->nents, direction);
641 	if (nents != ttm->sg->nents)
642 		goto release_sg;
643 
644 	drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
645 					 gtt->ttm.dma_address, ttm->num_pages);
646 
647 	return 0;
648 
649 release_sg:
650 	kfree(ttm->sg);
651 
652 release_pages:
653 	release_pages(ttm->pages, pinned);
654 	return r;
655 #endif
656 }
657 
radeon_ttm_tt_unpin_userptr(struct ttm_tt * ttm)658 static void radeon_ttm_tt_unpin_userptr(struct ttm_tt *ttm)
659 {
660 #ifdef __NetBSD__
661 	struct radeon_device *rdev = radeon_get_rdev(ttm->bdev);
662 	struct radeon_ttm_tt *gtt = (void *)ttm;
663 
664 	bus_dmamap_unload(rdev->ddev->dmat, gtt->ttm.dma_address);
665 	uvm_vsunlock(gtt->usermm, (void *)(vaddr_t)gtt->userptr,
666 	    ttm->num_pages << PAGE_SHIFT);
667 #else
668 	struct radeon_device *rdev = radeon_get_rdev(ttm->bdev);
669 	struct radeon_ttm_tt *gtt = (void *)ttm;
670 	struct sg_page_iter sg_iter;
671 
672 	int write = !(gtt->userflags & RADEON_GEM_USERPTR_READONLY);
673 	enum dma_data_direction direction = write ?
674 		DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
675 
676 	/* double check that we don't free the table twice */
677 	if (!ttm->sg->sgl)
678 		return;
679 
680 	/* free the sg table and pages again */
681 	dma_unmap_sg(rdev->dev, ttm->sg->sgl, ttm->sg->nents, direction);
682 
683 	for_each_sg_page(ttm->sg->sgl, &sg_iter, ttm->sg->nents, 0) {
684 		struct page *page = sg_page_iter_page(&sg_iter);
685 		if (!(gtt->userflags & RADEON_GEM_USERPTR_READONLY))
686 			set_page_dirty(page);
687 
688 		mark_page_accessed(page);
689 		put_page(page);
690 	}
691 
692 	sg_free_table(ttm->sg);
693 #endif
694 }
695 
radeon_ttm_backend_bind(struct ttm_tt * ttm,struct ttm_mem_reg * bo_mem)696 static int radeon_ttm_backend_bind(struct ttm_tt *ttm,
697 				   struct ttm_mem_reg *bo_mem)
698 {
699 	struct radeon_ttm_tt *gtt = (void*)ttm;
700 	uint32_t flags = RADEON_GART_PAGE_VALID | RADEON_GART_PAGE_READ |
701 		RADEON_GART_PAGE_WRITE;
702 	int r;
703 
704 	if (gtt->userptr) {
705 		radeon_ttm_tt_pin_userptr(ttm);
706 		flags &= ~RADEON_GART_PAGE_WRITE;
707 	}
708 
709 	gtt->offset = (unsigned long)(bo_mem->start << PAGE_SHIFT);
710 	if (!ttm->num_pages) {
711 		WARN(1, "nothing to bind %lu pages for mreg %p back %p!\n",
712 		     ttm->num_pages, bo_mem, ttm);
713 	}
714 	if (ttm->caching_state == tt_cached)
715 		flags |= RADEON_GART_PAGE_SNOOP;
716 	r = radeon_gart_bind(gtt->rdev, gtt->offset, ttm->num_pages,
717 			     ttm->pages, gtt->ttm.dma_address, flags);
718 	if (r) {
719 		DRM_ERROR("failed to bind %lu pages at 0x%08X\n",
720 			  ttm->num_pages, (unsigned)gtt->offset);
721 		return r;
722 	}
723 	return 0;
724 }
725 
radeon_ttm_backend_unbind(struct ttm_tt * ttm)726 static int radeon_ttm_backend_unbind(struct ttm_tt *ttm)
727 {
728 	struct radeon_ttm_tt *gtt = (void *)ttm;
729 
730 	radeon_gart_unbind(gtt->rdev, gtt->offset, ttm->num_pages);
731 
732 	if (gtt->userptr)
733 		radeon_ttm_tt_unpin_userptr(ttm);
734 
735 	return 0;
736 }
737 
radeon_ttm_backend_destroy(struct ttm_tt * ttm)738 static void radeon_ttm_backend_destroy(struct ttm_tt *ttm)
739 {
740 	struct radeon_ttm_tt *gtt = (void *)ttm;
741 
742 	ttm_dma_tt_fini(&gtt->ttm);
743 	kfree(gtt);
744 }
745 
746 static struct ttm_backend_func radeon_backend_func = {
747 	.bind = &radeon_ttm_backend_bind,
748 	.unbind = &radeon_ttm_backend_unbind,
749 	.destroy = &radeon_ttm_backend_destroy,
750 };
751 
radeon_ttm_tt_create(struct ttm_buffer_object * bo,uint32_t page_flags)752 static struct ttm_tt *radeon_ttm_tt_create(struct ttm_buffer_object *bo,
753 					   uint32_t page_flags)
754 {
755 	struct radeon_device *rdev;
756 	struct radeon_ttm_tt *gtt;
757 
758 	rdev = radeon_get_rdev(bo->bdev);
759 #if IS_ENABLED(CONFIG_AGP)
760 	if (rdev->flags & RADEON_IS_AGP) {
761 		return ttm_agp_tt_create(bo, rdev->ddev->agp->bridge,
762 					 page_flags);
763 	}
764 #endif
765 
766 	gtt = kzalloc(sizeof(struct radeon_ttm_tt), GFP_KERNEL);
767 	if (gtt == NULL) {
768 		return NULL;
769 	}
770 	gtt->ttm.ttm.func = &radeon_backend_func;
771 	gtt->rdev = rdev;
772 	if (ttm_dma_tt_init(&gtt->ttm, bo, page_flags)) {
773 		kfree(gtt);
774 		return NULL;
775 	}
776 	return &gtt->ttm.ttm;
777 }
778 
radeon_ttm_tt_to_gtt(struct ttm_tt * ttm)779 static struct radeon_ttm_tt *radeon_ttm_tt_to_gtt(struct ttm_tt *ttm)
780 {
781 	if (!ttm || ttm->func != &radeon_backend_func)
782 		return NULL;
783 	return (struct radeon_ttm_tt *)ttm;
784 }
785 
radeon_ttm_tt_populate(struct ttm_tt * ttm,struct ttm_operation_ctx * ctx)786 static int radeon_ttm_tt_populate(struct ttm_tt *ttm,
787 			struct ttm_operation_ctx *ctx)
788 {
789 	struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm);
790 #if !defined(__NetBSD__) || IS_ENABLED(CONFIG_AGP)
791 	struct radeon_device *rdev;
792 #endif
793 	bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
794 
795 	if (gtt && gtt->userptr) {
796 #ifdef __NetBSD__
797 		ttm->sg = NULL;
798 #else
799 		ttm->sg = kzalloc(sizeof(struct sg_table), GFP_KERNEL);
800 		if (!ttm->sg)
801 			return -ENOMEM;
802 #endif
803 
804 		ttm->page_flags |= TTM_PAGE_FLAG_SG;
805 		ttm->state = tt_unbound;
806 		return 0;
807 	}
808 
809 	if (slave && ttm->sg) {
810 #ifdef __NetBSD__
811 		int r = drm_prime_bus_dmamap_load_sgt(ttm->bdev->dmat,
812 		    gtt->ttm.dma_address, ttm->sg);
813 		if (r)
814 			return r;
815 #else
816 		drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
817 						 gtt->ttm.dma_address, ttm->num_pages);
818 #endif
819 		ttm->state = tt_unbound;
820 		return 0;
821 	}
822 
823 #if !defined(__NetBSD__) || IS_ENABLED(CONFIG_AGP)
824 	rdev = radeon_get_rdev(ttm->bdev);
825 #endif
826 #if IS_ENABLED(CONFIG_AGP)
827 	if (rdev->flags & RADEON_IS_AGP) {
828 		return ttm_agp_tt_populate(ttm, ctx);
829 	}
830 #endif
831 
832 #ifdef __NetBSD__
833 	/* XXX errno NetBSD->Linux */
834 	return ttm_bus_dma_populate(&gtt->ttm);
835 #else
836 
837 #ifdef CONFIG_SWIOTLB
838 	if (rdev->need_swiotlb && swiotlb_nr_tbl()) {
839 		return ttm_dma_populate(&gtt->ttm, rdev->dev, ctx);
840 	}
841 #endif
842 
843 	return ttm_populate_and_map_pages(rdev->dev, &gtt->ttm, ctx);
844 #endif
845 }
846 
radeon_ttm_tt_unpopulate(struct ttm_tt * ttm)847 static void radeon_ttm_tt_unpopulate(struct ttm_tt *ttm)
848 {
849 #if !defined(__NetBSD__) || IS_ENABLED(CONFIG_AGP)
850 	struct radeon_device *rdev;
851 #endif
852 	struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm);
853 	bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
854 
855 #ifdef __NetBSD__
856 	if (slave && ttm->sg) {
857 		bus_dmamap_unload(ttm->bdev->dmat, gtt->ttm.dma_address);
858 	}
859 #endif
860 	if (gtt && gtt->userptr) {
861 		kfree(ttm->sg);
862 		ttm->page_flags &= ~TTM_PAGE_FLAG_SG;
863 		return;
864 	}
865 
866 	if (slave)
867 		return;
868 
869 #if !defined(__NetBSD__) || IS_ENABLED(CONFIG_AGP)
870 	rdev = radeon_get_rdev(ttm->bdev);
871 #endif
872 #if IS_ENABLED(CONFIG_AGP)
873 	if (rdev->flags & RADEON_IS_AGP) {
874 		ttm_agp_tt_unpopulate(ttm);
875 		return;
876 	}
877 #endif
878 
879 #ifdef __NetBSD__
880 	ttm_bus_dma_unpopulate(&gtt->ttm);
881 	return;
882 #else
883 
884 #ifdef CONFIG_SWIOTLB
885 	if (rdev->need_swiotlb && swiotlb_nr_tbl()) {
886 		ttm_dma_unpopulate(&gtt->ttm, rdev->dev);
887 		return;
888 	}
889 #endif
890 
891 	ttm_unmap_and_unpopulate_pages(rdev->dev, &gtt->ttm);
892 #endif
893 }
894 
895 #ifdef __NetBSD__
radeon_ttm_tt_swapout(struct ttm_tt * ttm)896 static void radeon_ttm_tt_swapout(struct ttm_tt *ttm)
897 {
898 	struct radeon_ttm_tt *gtt = container_of(ttm, struct radeon_ttm_tt,
899 	    ttm.ttm);
900 	struct ttm_dma_tt *ttm_dma = &gtt->ttm;
901 
902 	ttm_bus_dma_swapout(ttm_dma);
903 }
904 
905 static int	radeon_ttm_fault(struct uvm_faultinfo *, vaddr_t,
906 		    struct vm_page **, int, int, vm_prot_t, int);
907 
908 static const struct uvm_pagerops radeon_uvm_ops = {
909 	.pgo_reference = &ttm_bo_uvm_reference,
910 	.pgo_detach = &ttm_bo_uvm_detach,
911 	.pgo_fault = &radeon_ttm_fault,
912 };
913 #endif
914 
radeon_ttm_tt_set_userptr(struct ttm_tt * ttm,uint64_t addr,uint32_t flags)915 int radeon_ttm_tt_set_userptr(struct ttm_tt *ttm, uint64_t addr,
916 			      uint32_t flags)
917 {
918 	struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm);
919 
920 	if (gtt == NULL)
921 		return -EINVAL;
922 
923 	gtt->userptr = addr;
924 #ifdef __NetBSD__
925 	gtt->usermm = curproc->p_vmspace;
926 #else
927 	gtt->usermm = current->mm;
928 #endif
929 	gtt->userflags = flags;
930 	return 0;
931 }
932 
radeon_ttm_tt_has_userptr(struct ttm_tt * ttm)933 bool radeon_ttm_tt_has_userptr(struct ttm_tt *ttm)
934 {
935 	struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm);
936 
937 	if (gtt == NULL)
938 		return false;
939 
940 	return !!gtt->userptr;
941 }
942 
radeon_ttm_tt_is_readonly(struct ttm_tt * ttm)943 bool radeon_ttm_tt_is_readonly(struct ttm_tt *ttm)
944 {
945 	struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm);
946 
947 	if (gtt == NULL)
948 		return false;
949 
950 	return !!(gtt->userflags & RADEON_GEM_USERPTR_READONLY);
951 }
952 
953 static struct ttm_bo_driver radeon_bo_driver = {
954 	.ttm_tt_create = &radeon_ttm_tt_create,
955 	.ttm_tt_populate = &radeon_ttm_tt_populate,
956 	.ttm_tt_unpopulate = &radeon_ttm_tt_unpopulate,
957 #ifdef __NetBSD__
958 	.ttm_tt_swapout = &radeon_ttm_tt_swapout,
959 	.ttm_uvm_ops = &radeon_uvm_ops,
960 #endif
961 	.invalidate_caches = &radeon_invalidate_caches,
962 	.init_mem_type = &radeon_init_mem_type,
963 	.eviction_valuable = ttm_bo_eviction_valuable,
964 	.evict_flags = &radeon_evict_flags,
965 	.move = &radeon_bo_move,
966 	.verify_access = &radeon_verify_access,
967 	.move_notify = &radeon_bo_move_notify,
968 	.fault_reserve_notify = &radeon_bo_fault_reserve_notify,
969 	.io_mem_reserve = &radeon_ttm_io_mem_reserve,
970 	.io_mem_free = &radeon_ttm_io_mem_free,
971 };
972 
radeon_ttm_init(struct radeon_device * rdev)973 int radeon_ttm_init(struct radeon_device *rdev)
974 {
975 	int r;
976 
977 	/* No others user of address space so set it to 0 */
978 	r = ttm_bo_device_init(&rdev->mman.bdev,
979 			       &radeon_bo_driver,
980 #ifdef __NetBSD__
981 			       rdev->ddev->bst,
982 			       rdev->ddev->dmat,
983 #else
984 			       rdev->ddev->anon_inode->i_mapping,
985 #endif
986 			       rdev->ddev->vma_offset_manager,
987 			       dma_addressing_limited(pci_dev_dev(rdev->pdev)));
988 	if (r) {
989 		DRM_ERROR("failed initializing buffer object driver(%d).\n", r);
990 		return r;
991 	}
992 	rdev->mman.initialized = true;
993 	r = ttm_bo_init_mm(&rdev->mman.bdev, TTM_PL_VRAM,
994 				rdev->mc.real_vram_size >> PAGE_SHIFT);
995 	if (r) {
996 		DRM_ERROR("Failed initializing VRAM heap.\n");
997 		return r;
998 	}
999 	/* Change the size here instead of the init above so only lpfn is affected */
1000 	radeon_ttm_set_active_vram_size(rdev, rdev->mc.visible_vram_size);
1001 
1002 	r = radeon_bo_create(rdev, 256 * 1024, PAGE_SIZE, true,
1003 			     RADEON_GEM_DOMAIN_VRAM, 0, NULL,
1004 			     NULL, &rdev->stolen_vga_memory);
1005 	if (r) {
1006 		return r;
1007 	}
1008 	r = radeon_bo_reserve(rdev->stolen_vga_memory, false);
1009 	if (r)
1010 		return r;
1011 	r = radeon_bo_pin(rdev->stolen_vga_memory, RADEON_GEM_DOMAIN_VRAM, NULL);
1012 	radeon_bo_unreserve(rdev->stolen_vga_memory);
1013 	if (r) {
1014 		radeon_bo_unref(&rdev->stolen_vga_memory);
1015 		return r;
1016 	}
1017 	DRM_INFO("radeon: %uM of VRAM memory ready\n",
1018 		 (unsigned) (rdev->mc.real_vram_size / (1024 * 1024)));
1019 	r = ttm_bo_init_mm(&rdev->mman.bdev, TTM_PL_TT,
1020 				rdev->mc.gtt_size >> PAGE_SHIFT);
1021 	if (r) {
1022 		DRM_ERROR("Failed initializing GTT heap.\n");
1023 		return r;
1024 	}
1025 	DRM_INFO("radeon: %uM of GTT memory ready.\n",
1026 		 (unsigned)(rdev->mc.gtt_size / (1024 * 1024)));
1027 
1028 	r = radeon_ttm_debugfs_init(rdev);
1029 	if (r) {
1030 		DRM_ERROR("Failed to init debugfs\n");
1031 		return r;
1032 	}
1033 	return 0;
1034 }
1035 
radeon_ttm_fini(struct radeon_device * rdev)1036 void radeon_ttm_fini(struct radeon_device *rdev)
1037 {
1038 	int r;
1039 
1040 	if (!rdev->mman.initialized)
1041 		return;
1042 	radeon_ttm_debugfs_fini(rdev);
1043 	if (rdev->stolen_vga_memory) {
1044 		r = radeon_bo_reserve(rdev->stolen_vga_memory, false);
1045 		if (r == 0) {
1046 			radeon_bo_unpin(rdev->stolen_vga_memory);
1047 			radeon_bo_unreserve(rdev->stolen_vga_memory);
1048 		}
1049 		radeon_bo_unref(&rdev->stolen_vga_memory);
1050 	}
1051 	ttm_bo_clean_mm(&rdev->mman.bdev, TTM_PL_VRAM);
1052 	ttm_bo_clean_mm(&rdev->mman.bdev, TTM_PL_TT);
1053 	ttm_bo_device_release(&rdev->mman.bdev);
1054 	radeon_gart_fini(rdev);
1055 	rdev->mman.initialized = false;
1056 	DRM_INFO("radeon: ttm finalized\n");
1057 }
1058 
1059 /* this should only be called at bootup or when userspace
1060  * isn't running */
radeon_ttm_set_active_vram_size(struct radeon_device * rdev,u64 size)1061 void radeon_ttm_set_active_vram_size(struct radeon_device *rdev, u64 size)
1062 {
1063 	struct ttm_mem_type_manager *man;
1064 
1065 	if (!rdev->mman.initialized)
1066 		return;
1067 
1068 	man = &rdev->mman.bdev.man[TTM_PL_VRAM];
1069 	/* this just adjusts TTM size idea, which sets lpfn to the correct value */
1070 	man->size = size >> PAGE_SHIFT;
1071 }
1072 
1073 #ifdef __NetBSD__
1074 
1075 static int
radeon_ttm_fault(struct uvm_faultinfo * ufi,vaddr_t vaddr,struct vm_page ** pps,int npages,int centeridx,vm_prot_t access_type,int flags)1076 radeon_ttm_fault(struct uvm_faultinfo *ufi, vaddr_t vaddr,
1077     struct vm_page **pps, int npages, int centeridx, vm_prot_t access_type,
1078     int flags)
1079 {
1080 	struct uvm_object *const uobj = ufi->entry->object.uvm_obj;
1081 	struct ttm_buffer_object *const bo = container_of(uobj,
1082 	    struct ttm_buffer_object, uvmobj);
1083 	struct radeon_device *const rdev = radeon_get_rdev(bo->bdev);
1084 	int error;
1085 
1086 	KASSERT(rdev != NULL);
1087 	down_read(&rdev->pm.mclk_lock);
1088 	error = ttm_bo_uvm_fault(ufi, vaddr, pps, npages, centeridx,
1089 	    access_type, flags);
1090 	up_read(&rdev->pm.mclk_lock);
1091 
1092 	return error;
1093 }
1094 
1095 int
radeon_mmap_object(struct drm_device * dev,off_t offset,size_t size,vm_prot_t prot,struct uvm_object ** uobjp,voff_t * uoffsetp,struct file * file)1096 radeon_mmap_object(struct drm_device *dev, off_t offset, size_t size,
1097     vm_prot_t prot, struct uvm_object **uobjp, voff_t *uoffsetp,
1098     struct file *file)
1099 {
1100 	struct radeon_device *rdev = dev->dev_private;
1101 
1102 	KASSERT(0 == (offset & (PAGE_SIZE - 1)));
1103 
1104 	if (__predict_false(rdev == NULL))	/* XXX How?? */
1105 		return -EINVAL;
1106 
1107 	return ttm_bo_mmap_object(&rdev->mman.bdev, offset, size, prot,
1108 	    uobjp, uoffsetp, file);
1109 }
1110 
1111 #else
1112 
radeon_ttm_fault(struct vm_fault * vmf)1113 static vm_fault_t radeon_ttm_fault(struct vm_fault *vmf)
1114 {
1115 	struct ttm_buffer_object *bo;
1116 	struct radeon_device *rdev;
1117 	vm_fault_t ret;
1118 
1119 	bo = (struct ttm_buffer_object *)vmf->vma->vm_private_data;
1120 	if (bo == NULL)
1121 		return VM_FAULT_NOPAGE;
1122 
1123 	rdev = radeon_get_rdev(bo->bdev);
1124 	down_read(&rdev->pm.mclk_lock);
1125 	ret = ttm_bo_vm_fault(vmf);
1126 	up_read(&rdev->pm.mclk_lock);
1127 	return ret;
1128 }
1129 
1130 static struct vm_operations_struct radeon_ttm_vm_ops = {
1131 	.fault = radeon_ttm_fault,
1132 	.open = ttm_bo_vm_open,
1133 	.close = ttm_bo_vm_close,
1134 	.access = ttm_bo_vm_access
1135 };
1136 
radeon_mmap(struct file * filp,struct vm_area_struct * vma)1137 int radeon_mmap(struct file *filp, struct vm_area_struct *vma)
1138 {
1139 	int r;
1140 	struct drm_file *file_priv = filp->private_data;
1141 	struct radeon_device *rdev = file_priv->minor->dev->dev_private;
1142 
1143 	if (rdev == NULL)
1144 		return -EINVAL;
1145 
1146 	r = ttm_bo_mmap(filp, vma, &rdev->mman.bdev);
1147 	if (unlikely(r != 0))
1148 		return r;
1149 
1150 	vma->vm_ops = &radeon_ttm_vm_ops;
1151 	return 0;
1152 }
1153 
1154 #endif	/* __NetBSD__ */
1155 
1156 #if defined(CONFIG_DEBUG_FS)
1157 
radeon_mm_dump_table(struct seq_file * m,void * data)1158 static int radeon_mm_dump_table(struct seq_file *m, void *data)
1159 {
1160 	struct drm_info_node *node = (struct drm_info_node *)m->private;
1161 	unsigned ttm_pl = *(int*)node->info_ent->data;
1162 	struct drm_device *dev = node->minor->dev;
1163 	struct radeon_device *rdev = dev->dev_private;
1164 	struct ttm_mem_type_manager *man = &rdev->mman.bdev.man[ttm_pl];
1165 	struct drm_printer p = drm_seq_file_printer(m);
1166 
1167 	man->func->debug(man, &p);
1168 	return 0;
1169 }
1170 
1171 
1172 static int ttm_pl_vram = TTM_PL_VRAM;
1173 static int ttm_pl_tt = TTM_PL_TT;
1174 
1175 static struct drm_info_list radeon_ttm_debugfs_list[] = {
1176 	{"radeon_vram_mm", radeon_mm_dump_table, 0, &ttm_pl_vram},
1177 	{"radeon_gtt_mm", radeon_mm_dump_table, 0, &ttm_pl_tt},
1178 	{"ttm_page_pool", ttm_page_alloc_debugfs, 0, NULL},
1179 #ifdef CONFIG_SWIOTLB
1180 	{"ttm_dma_page_pool", ttm_dma_page_alloc_debugfs, 0, NULL}
1181 #endif
1182 };
1183 
radeon_ttm_vram_open(struct inode * inode,struct file * filep)1184 static int radeon_ttm_vram_open(struct inode *inode, struct file *filep)
1185 {
1186 	struct radeon_device *rdev = inode->i_private;
1187 	i_size_write(inode, rdev->mc.mc_vram_size);
1188 	filep->private_data = inode->i_private;
1189 	return 0;
1190 }
1191 
radeon_ttm_vram_read(struct file * f,char __user * buf,size_t size,loff_t * pos)1192 static ssize_t radeon_ttm_vram_read(struct file *f, char __user *buf,
1193 				    size_t size, loff_t *pos)
1194 {
1195 	struct radeon_device *rdev = f->private_data;
1196 	ssize_t result = 0;
1197 	int r;
1198 
1199 	if (size & 0x3 || *pos & 0x3)
1200 		return -EINVAL;
1201 
1202 	while (size) {
1203 		unsigned long flags;
1204 		uint32_t value;
1205 
1206 		if (*pos >= rdev->mc.mc_vram_size)
1207 			return result;
1208 
1209 		spin_lock_irqsave(&rdev->mmio_idx_lock, flags);
1210 		WREG32(RADEON_MM_INDEX, ((uint32_t)*pos) | 0x80000000);
1211 		if (rdev->family >= CHIP_CEDAR)
1212 			WREG32(EVERGREEN_MM_INDEX_HI, *pos >> 31);
1213 		value = RREG32(RADEON_MM_DATA);
1214 		spin_unlock_irqrestore(&rdev->mmio_idx_lock, flags);
1215 
1216 		r = put_user(value, (uint32_t *)buf);
1217 		if (r)
1218 			return r;
1219 
1220 		result += 4;
1221 		buf += 4;
1222 		*pos += 4;
1223 		size -= 4;
1224 	}
1225 
1226 	return result;
1227 }
1228 
1229 static const struct file_operations radeon_ttm_vram_fops = {
1230 	.owner = THIS_MODULE,
1231 	.open = radeon_ttm_vram_open,
1232 	.read = radeon_ttm_vram_read,
1233 	.llseek = default_llseek
1234 };
1235 
radeon_ttm_gtt_open(struct inode * inode,struct file * filep)1236 static int radeon_ttm_gtt_open(struct inode *inode, struct file *filep)
1237 {
1238 	struct radeon_device *rdev = inode->i_private;
1239 	i_size_write(inode, rdev->mc.gtt_size);
1240 	filep->private_data = inode->i_private;
1241 	return 0;
1242 }
1243 
radeon_ttm_gtt_read(struct file * f,char __user * buf,size_t size,loff_t * pos)1244 static ssize_t radeon_ttm_gtt_read(struct file *f, char __user *buf,
1245 				   size_t size, loff_t *pos)
1246 {
1247 	struct radeon_device *rdev = f->private_data;
1248 	ssize_t result = 0;
1249 	int r;
1250 
1251 	while (size) {
1252 		loff_t p = *pos / PAGE_SIZE;
1253 		unsigned off = *pos & ~PAGE_MASK;
1254 		size_t cur_size = min_t(size_t, size, PAGE_SIZE - off);
1255 		struct page *page;
1256 		void *ptr;
1257 
1258 		if (p >= rdev->gart.num_cpu_pages)
1259 			return result;
1260 
1261 		page = rdev->gart.pages[p];
1262 		if (page) {
1263 			ptr = kmap(page);
1264 			ptr += off;
1265 
1266 			r = copy_to_user(buf, ptr, cur_size);
1267 			kunmap(rdev->gart.pages[p]);
1268 		} else
1269 			r = clear_user(buf, cur_size);
1270 
1271 		if (r)
1272 			return -EFAULT;
1273 
1274 		result += cur_size;
1275 		buf += cur_size;
1276 		*pos += cur_size;
1277 		size -= cur_size;
1278 	}
1279 
1280 	return result;
1281 }
1282 
1283 static const struct file_operations radeon_ttm_gtt_fops = {
1284 	.owner = THIS_MODULE,
1285 	.open = radeon_ttm_gtt_open,
1286 	.read = radeon_ttm_gtt_read,
1287 	.llseek = default_llseek
1288 };
1289 
1290 #endif
1291 
radeon_ttm_debugfs_init(struct radeon_device * rdev)1292 static int radeon_ttm_debugfs_init(struct radeon_device *rdev)
1293 {
1294 #if defined(CONFIG_DEBUG_FS)
1295 	unsigned count;
1296 
1297 	struct drm_minor *minor = rdev->ddev->primary;
1298 	struct dentry *root = minor->debugfs_root;
1299 
1300 	rdev->mman.vram = debugfs_create_file("radeon_vram", S_IFREG | S_IRUGO,
1301 					      root, rdev,
1302 					      &radeon_ttm_vram_fops);
1303 
1304 	rdev->mman.gtt = debugfs_create_file("radeon_gtt", S_IFREG | S_IRUGO,
1305 					     root, rdev, &radeon_ttm_gtt_fops);
1306 
1307 	count = ARRAY_SIZE(radeon_ttm_debugfs_list);
1308 
1309 #ifdef CONFIG_SWIOTLB
1310 	if (!(rdev->need_swiotlb && swiotlb_nr_tbl()))
1311 		--count;
1312 #endif
1313 
1314 	return radeon_debugfs_add_files(rdev, radeon_ttm_debugfs_list, count);
1315 #else
1316 
1317 	return 0;
1318 #endif
1319 }
1320 
radeon_ttm_debugfs_fini(struct radeon_device * rdev)1321 static void radeon_ttm_debugfs_fini(struct radeon_device *rdev)
1322 {
1323 #if defined(CONFIG_DEBUG_FS)
1324 
1325 	debugfs_remove(rdev->mman.vram);
1326 	rdev->mman.vram = NULL;
1327 
1328 	debugfs_remove(rdev->mman.gtt);
1329 	rdev->mman.gtt = NULL;
1330 #endif
1331 }
1332