1 /*	$NetBSD: radeon_object.c,v 1.8 2021/12/18 23:45:43 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_object.c,v 1.8 2021/12/18 23:45:43 riastradh Exp $");
37 
38 #include <linux/io.h>
39 #include <linux/list.h>
40 #include <linux/slab.h>
41 
42 #include <drm/drm_cache.h>
43 #include <drm/drm_prime.h>
44 #include <drm/radeon_drm.h>
45 
46 #include "radeon.h"
47 #include "radeon_trace.h"
48 
49 #include <linux/nbsd-namespace.h>
50 
51 int radeon_ttm_init(struct radeon_device *rdev);
52 void radeon_ttm_fini(struct radeon_device *rdev);
53 static void radeon_bo_clear_surface_reg(struct radeon_bo *bo);
54 
55 /*
56  * To exclude mutual BO access we rely on bo_reserve exclusion, as all
57  * function are calling it.
58  */
59 
radeon_update_memory_usage(struct radeon_bo * bo,unsigned mem_type,int sign)60 static void radeon_update_memory_usage(struct radeon_bo *bo,
61 				       unsigned mem_type, int sign)
62 {
63 	struct radeon_device *rdev = bo->rdev;
64 	u64 size = (u64)bo->tbo.num_pages << PAGE_SHIFT;
65 
66 	switch (mem_type) {
67 	case TTM_PL_TT:
68 		if (sign > 0)
69 			atomic64_add(size, &rdev->gtt_usage);
70 		else
71 			atomic64_sub(size, &rdev->gtt_usage);
72 		break;
73 	case TTM_PL_VRAM:
74 		if (sign > 0)
75 			atomic64_add(size, &rdev->vram_usage);
76 		else
77 			atomic64_sub(size, &rdev->vram_usage);
78 		break;
79 	}
80 }
81 
radeon_ttm_bo_destroy(struct ttm_buffer_object * tbo)82 static void radeon_ttm_bo_destroy(struct ttm_buffer_object *tbo)
83 {
84 	struct radeon_bo *bo;
85 
86 	bo = container_of(tbo, struct radeon_bo, tbo);
87 
88 	radeon_update_memory_usage(bo, bo->tbo.mem.mem_type, -1);
89 
90 	mutex_lock(&bo->rdev->gem.mutex);
91 	list_del_init(&bo->list);
92 	mutex_unlock(&bo->rdev->gem.mutex);
93 	radeon_bo_clear_surface_reg(bo);
94 	WARN_ON_ONCE(!list_empty(&bo->va));
95 	if (bo->tbo.base.import_attach)
96 		drm_prime_gem_destroy(&bo->tbo.base, bo->tbo.sg);
97 	drm_gem_object_release(&bo->tbo.base);
98 	kfree(bo);
99 }
100 
radeon_ttm_bo_is_radeon_bo(struct ttm_buffer_object * bo)101 bool radeon_ttm_bo_is_radeon_bo(struct ttm_buffer_object *bo)
102 {
103 	if (bo->destroy == &radeon_ttm_bo_destroy)
104 		return true;
105 	return false;
106 }
107 
radeon_ttm_placement_from_domain(struct radeon_bo * rbo,u32 domain)108 void radeon_ttm_placement_from_domain(struct radeon_bo *rbo, u32 domain)
109 {
110 	u32 c = 0, i;
111 
112 	rbo->placement.placement = rbo->placements;
113 	rbo->placement.busy_placement = rbo->placements;
114 	if (domain & RADEON_GEM_DOMAIN_VRAM) {
115 		/* Try placing BOs which don't need CPU access outside of the
116 		 * CPU accessible part of VRAM
117 		 */
118 		if ((rbo->flags & RADEON_GEM_NO_CPU_ACCESS) &&
119 		    rbo->rdev->mc.visible_vram_size < rbo->rdev->mc.real_vram_size) {
120 			rbo->placements[c].fpfn =
121 				rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
122 			rbo->placements[c++].flags = TTM_PL_FLAG_WC |
123 						     TTM_PL_FLAG_UNCACHED |
124 						     TTM_PL_FLAG_VRAM;
125 		}
126 
127 		rbo->placements[c].fpfn = 0;
128 		rbo->placements[c++].flags = TTM_PL_FLAG_WC |
129 					     TTM_PL_FLAG_UNCACHED |
130 					     TTM_PL_FLAG_VRAM;
131 	}
132 
133 	if (domain & RADEON_GEM_DOMAIN_GTT) {
134 		if (rbo->flags & RADEON_GEM_GTT_UC) {
135 			rbo->placements[c].fpfn = 0;
136 			rbo->placements[c++].flags = TTM_PL_FLAG_UNCACHED |
137 				TTM_PL_FLAG_TT;
138 
139 		} else if ((rbo->flags & RADEON_GEM_GTT_WC) ||
140 			   (rbo->rdev->flags & RADEON_IS_AGP)) {
141 			rbo->placements[c].fpfn = 0;
142 			rbo->placements[c++].flags = TTM_PL_FLAG_WC |
143 				TTM_PL_FLAG_UNCACHED |
144 				TTM_PL_FLAG_TT;
145 		} else {
146 			rbo->placements[c].fpfn = 0;
147 			rbo->placements[c++].flags = TTM_PL_FLAG_CACHED |
148 						     TTM_PL_FLAG_TT;
149 		}
150 	}
151 
152 	if (domain & RADEON_GEM_DOMAIN_CPU) {
153 		if (rbo->flags & RADEON_GEM_GTT_UC) {
154 			rbo->placements[c].fpfn = 0;
155 			rbo->placements[c++].flags = TTM_PL_FLAG_UNCACHED |
156 				TTM_PL_FLAG_SYSTEM;
157 
158 		} else if ((rbo->flags & RADEON_GEM_GTT_WC) ||
159 		    rbo->rdev->flags & RADEON_IS_AGP) {
160 			rbo->placements[c].fpfn = 0;
161 			rbo->placements[c++].flags = TTM_PL_FLAG_WC |
162 				TTM_PL_FLAG_UNCACHED |
163 				TTM_PL_FLAG_SYSTEM;
164 		} else {
165 			rbo->placements[c].fpfn = 0;
166 			rbo->placements[c++].flags = TTM_PL_FLAG_CACHED |
167 						     TTM_PL_FLAG_SYSTEM;
168 		}
169 	}
170 	if (!c) {
171 		rbo->placements[c].fpfn = 0;
172 		rbo->placements[c++].flags = TTM_PL_MASK_CACHING |
173 					     TTM_PL_FLAG_SYSTEM;
174 	}
175 
176 	rbo->placement.num_placement = c;
177 	rbo->placement.num_busy_placement = c;
178 
179 	for (i = 0; i < c; ++i) {
180 		if ((rbo->flags & RADEON_GEM_CPU_ACCESS) &&
181 		    (rbo->placements[i].flags & TTM_PL_FLAG_VRAM) &&
182 		    !rbo->placements[i].fpfn)
183 			rbo->placements[i].lpfn =
184 				rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
185 		else
186 			rbo->placements[i].lpfn = 0;
187 	}
188 }
189 
radeon_bo_create(struct radeon_device * rdev,unsigned long size,int byte_align,bool kernel,u32 domain,u32 flags,struct sg_table * sg,struct dma_resv * resv,struct radeon_bo ** bo_ptr)190 int radeon_bo_create(struct radeon_device *rdev,
191 		     unsigned long size, int byte_align, bool kernel,
192 		     u32 domain, u32 flags, struct sg_table *sg,
193 		     struct dma_resv *resv,
194 		     struct radeon_bo **bo_ptr)
195 {
196 	struct radeon_bo *bo;
197 	enum ttm_bo_type type;
198 	unsigned long page_align = roundup(byte_align, PAGE_SIZE) >> PAGE_SHIFT;
199 	size_t acc_size;
200 	int r;
201 
202 	size = ALIGN(size, PAGE_SIZE);
203 
204 	if (kernel) {
205 		type = ttm_bo_type_kernel;
206 	} else if (sg) {
207 		type = ttm_bo_type_sg;
208 	} else {
209 		type = ttm_bo_type_device;
210 	}
211 	*bo_ptr = NULL;
212 
213 	acc_size = ttm_bo_dma_acc_size(&rdev->mman.bdev, size,
214 				       sizeof(struct radeon_bo));
215 
216 	bo = kzalloc(sizeof(struct radeon_bo), GFP_KERNEL);
217 	if (bo == NULL)
218 		return -ENOMEM;
219 	drm_gem_private_object_init(rdev->ddev, &bo->tbo.base, size);
220 	bo->rdev = rdev;
221 	bo->surface_reg = -1;
222 	INIT_LIST_HEAD(&bo->list);
223 	INIT_LIST_HEAD(&bo->va);
224 	bo->initial_domain = domain & (RADEON_GEM_DOMAIN_VRAM |
225 				       RADEON_GEM_DOMAIN_GTT |
226 				       RADEON_GEM_DOMAIN_CPU);
227 
228 	bo->flags = flags;
229 	/* PCI GART is always snooped */
230 	if (!(rdev->flags & RADEON_IS_PCIE))
231 		bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
232 
233 	/* Write-combined CPU mappings of GTT cause GPU hangs with RV6xx
234 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=91268
235 	 */
236 	if (rdev->family >= CHIP_RV610 && rdev->family <= CHIP_RV635)
237 		bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
238 
239 #ifdef CONFIG_X86_32
240 	/* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
241 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
242 	 */
243 	bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
244 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
245 	/* Don't try to enable write-combining when it can't work, or things
246 	 * may be slow
247 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
248 	 */
249 #ifndef CONFIG_COMPILE_TEST
250 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
251 	 thanks to write-combining
252 #endif
253 
254 	if (bo->flags & RADEON_GEM_GTT_WC)
255 		DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
256 			      "better performance thanks to write-combining\n");
257 	bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
258 #else
259 	/* For architectures that don't support WC memory,
260 	 * mask out the WC flag from the BO
261 	 */
262 	if (!drm_arch_can_wc_memory())
263 		bo->flags &= ~RADEON_GEM_GTT_WC;
264 #endif
265 
266 	radeon_ttm_placement_from_domain(bo, domain);
267 	/* Kernel allocation are uninterruptible */
268 	down_read(&rdev->pm.mclk_lock);
269 	r = ttm_bo_init(&rdev->mman.bdev, &bo->tbo, size, type,
270 			&bo->placement, page_align, !kernel, acc_size,
271 			sg, resv, &radeon_ttm_bo_destroy);
272 	up_read(&rdev->pm.mclk_lock);
273 	if (unlikely(r != 0)) {
274 		return r;
275 	}
276 	*bo_ptr = bo;
277 
278 	trace_radeon_bo_create(bo);
279 
280 	return 0;
281 }
282 
radeon_bo_kmap(struct radeon_bo * bo,void ** ptr)283 int radeon_bo_kmap(struct radeon_bo *bo, void **ptr)
284 {
285 	bool is_iomem;
286 	int r;
287 
288 	if (bo->kptr) {
289 		if (ptr) {
290 			*ptr = bo->kptr;
291 		}
292 		return 0;
293 	}
294 	r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages, &bo->kmap);
295 	if (r) {
296 		return r;
297 	}
298 	bo->kptr = ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
299 	if (ptr) {
300 		*ptr = bo->kptr;
301 	}
302 	radeon_bo_check_tiling(bo, 0, 0);
303 	return 0;
304 }
305 
radeon_bo_kunmap(struct radeon_bo * bo)306 void radeon_bo_kunmap(struct radeon_bo *bo)
307 {
308 	if (bo->kptr == NULL)
309 		return;
310 	bo->kptr = NULL;
311 	radeon_bo_check_tiling(bo, 0, 0);
312 	ttm_bo_kunmap(&bo->kmap);
313 }
314 
radeon_bo_ref(struct radeon_bo * bo)315 struct radeon_bo *radeon_bo_ref(struct radeon_bo *bo)
316 {
317 	if (bo == NULL)
318 		return NULL;
319 
320 	ttm_bo_get(&bo->tbo);
321 	return bo;
322 }
323 
radeon_bo_unref(struct radeon_bo ** bo)324 void radeon_bo_unref(struct radeon_bo **bo)
325 {
326 	struct ttm_buffer_object *tbo;
327 
328 	if ((*bo) == NULL)
329 		return;
330 	tbo = &((*bo)->tbo);
331 	ttm_bo_put(tbo);
332 	*bo = NULL;
333 }
334 
radeon_bo_pin_restricted(struct radeon_bo * bo,u32 domain,u64 max_offset,u64 * gpu_addr)335 int radeon_bo_pin_restricted(struct radeon_bo *bo, u32 domain, u64 max_offset,
336 			     u64 *gpu_addr)
337 {
338 	struct ttm_operation_ctx ctx = { false, false };
339 	int r, i;
340 
341 	if (radeon_ttm_tt_has_userptr(bo->tbo.ttm))
342 		return -EPERM;
343 
344 	if (bo->pin_count) {
345 		bo->pin_count++;
346 		if (gpu_addr)
347 			*gpu_addr = radeon_bo_gpu_offset(bo);
348 
349 		if (max_offset != 0) {
350 			u64 domain_start;
351 
352 			if (domain == RADEON_GEM_DOMAIN_VRAM)
353 				domain_start = bo->rdev->mc.vram_start;
354 			else
355 				domain_start = bo->rdev->mc.gtt_start;
356 			WARN_ON_ONCE(max_offset <
357 				     (radeon_bo_gpu_offset(bo) - domain_start));
358 		}
359 
360 		return 0;
361 	}
362 	if (bo->prime_shared_count && domain == RADEON_GEM_DOMAIN_VRAM) {
363 		/* A BO shared as a dma-buf cannot be sensibly migrated to VRAM */
364 		return -EINVAL;
365 	}
366 
367 	radeon_ttm_placement_from_domain(bo, domain);
368 	for (i = 0; i < bo->placement.num_placement; i++) {
369 		/* force to pin into visible video ram */
370 		if ((bo->placements[i].flags & TTM_PL_FLAG_VRAM) &&
371 		    !(bo->flags & RADEON_GEM_NO_CPU_ACCESS) &&
372 		    (!max_offset || max_offset > bo->rdev->mc.visible_vram_size))
373 			bo->placements[i].lpfn =
374 				bo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
375 		else
376 			bo->placements[i].lpfn = max_offset >> PAGE_SHIFT;
377 
378 		bo->placements[i].flags |= TTM_PL_FLAG_NO_EVICT;
379 	}
380 
381 	r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
382 	if (likely(r == 0)) {
383 		bo->pin_count = 1;
384 		if (gpu_addr != NULL)
385 			*gpu_addr = radeon_bo_gpu_offset(bo);
386 		if (domain == RADEON_GEM_DOMAIN_VRAM)
387 			bo->rdev->vram_pin_size += radeon_bo_size(bo);
388 		else
389 			bo->rdev->gart_pin_size += radeon_bo_size(bo);
390 	} else {
391 		dev_err(bo->rdev->dev, "%p pin failed\n", bo);
392 	}
393 	return r;
394 }
395 
radeon_bo_pin(struct radeon_bo * bo,u32 domain,u64 * gpu_addr)396 int radeon_bo_pin(struct radeon_bo *bo, u32 domain, u64 *gpu_addr)
397 {
398 	return radeon_bo_pin_restricted(bo, domain, 0, gpu_addr);
399 }
400 
radeon_bo_unpin(struct radeon_bo * bo)401 int radeon_bo_unpin(struct radeon_bo *bo)
402 {
403 	struct ttm_operation_ctx ctx = { false, false };
404 	int r, i;
405 
406 	if (!bo->pin_count) {
407 		dev_warn(bo->rdev->dev, "%p unpin not necessary\n", bo);
408 		return 0;
409 	}
410 	bo->pin_count--;
411 	if (bo->pin_count)
412 		return 0;
413 	for (i = 0; i < bo->placement.num_placement; i++) {
414 		bo->placements[i].lpfn = 0;
415 		bo->placements[i].flags &= ~TTM_PL_FLAG_NO_EVICT;
416 	}
417 	r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
418 	if (likely(r == 0)) {
419 		if (bo->tbo.mem.mem_type == TTM_PL_VRAM)
420 			bo->rdev->vram_pin_size -= radeon_bo_size(bo);
421 		else
422 			bo->rdev->gart_pin_size -= radeon_bo_size(bo);
423 	} else {
424 		dev_err(bo->rdev->dev, "%p validate failed for unpin\n", bo);
425 	}
426 	return r;
427 }
428 
radeon_bo_evict_vram(struct radeon_device * rdev)429 int radeon_bo_evict_vram(struct radeon_device *rdev)
430 {
431 	/* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
432 #ifndef CONFIG_HIBERNATION
433 	if (rdev->flags & RADEON_IS_IGP) {
434 		if (rdev->mc.igp_sideport_enabled == false)
435 			/* Useless to evict on IGP chips */
436 			return 0;
437 	}
438 #endif
439 	return ttm_bo_evict_mm(&rdev->mman.bdev, TTM_PL_VRAM);
440 }
441 
radeon_bo_force_delete(struct radeon_device * rdev)442 void radeon_bo_force_delete(struct radeon_device *rdev)
443 {
444 	struct radeon_bo *bo, *n;
445 
446 	if (list_empty(&rdev->gem.objects)) {
447 		return;
448 	}
449 	dev_err(rdev->dev, "Userspace still has active objects !\n");
450 	list_for_each_entry_safe(bo, n, &rdev->gem.objects, list) {
451 		dev_err(rdev->dev, "%p %p %lu %lu force free\n",
452 			&bo->tbo.base, bo, (unsigned long)bo->tbo.base.size,
453 			*((unsigned long *)&bo->tbo.base.refcount));
454 		mutex_lock(&bo->rdev->gem.mutex);
455 		list_del_init(&bo->list);
456 		mutex_unlock(&bo->rdev->gem.mutex);
457 		/* this should unref the ttm bo */
458 		drm_gem_object_put_unlocked(&bo->tbo.base);
459 	}
460 }
461 
radeon_bo_init(struct radeon_device * rdev)462 int radeon_bo_init(struct radeon_device *rdev)
463 {
464 	/* reserve PAT memory space to WC for VRAM */
465 	arch_io_reserve_memtype_wc(rdev->mc.aper_base,
466 				   rdev->mc.aper_size);
467 
468 	/* Add an MTRR for the VRAM */
469 	if (!rdev->fastfb_working) {
470 		rdev->mc.vram_mtrr = arch_phys_wc_add(rdev->mc.aper_base,
471 						      rdev->mc.aper_size);
472 	}
473 #ifdef __NetBSD__
474 	if (rdev->mc.aper_base)
475 		pmap_pv_track(rdev->mc.aper_base, rdev->mc.aper_size);
476 #endif
477 	DRM_INFO("Detected VRAM RAM=%"PRIx64"M, BAR=%lluM\n",
478 		rdev->mc.mc_vram_size >> 20,
479 		(unsigned long long)rdev->mc.aper_size >> 20);
480 	DRM_INFO("RAM width %dbits %cDR\n",
481 			rdev->mc.vram_width, rdev->mc.vram_is_ddr ? 'D' : 'S');
482 	return radeon_ttm_init(rdev);
483 }
484 
radeon_bo_fini(struct radeon_device * rdev)485 void radeon_bo_fini(struct radeon_device *rdev)
486 {
487 	radeon_ttm_fini(rdev);
488 #ifdef __NetBSD__
489 	if (rdev->mc.aper_base)
490 		pmap_pv_untrack(rdev->mc.aper_base, rdev->mc.aper_size);
491 #endif
492 	arch_phys_wc_del(rdev->mc.vram_mtrr);
493 	arch_io_free_memtype_wc(rdev->mc.aper_base, rdev->mc.aper_size);
494 }
495 
496 /* Returns how many bytes TTM can move per IB.
497  */
radeon_bo_get_threshold_for_moves(struct radeon_device * rdev)498 static u64 radeon_bo_get_threshold_for_moves(struct radeon_device *rdev)
499 {
500 	u64 real_vram_size = rdev->mc.real_vram_size;
501 	u64 vram_usage = atomic64_read(&rdev->vram_usage);
502 
503 	/* This function is based on the current VRAM usage.
504 	 *
505 	 * - If all of VRAM is free, allow relocating the number of bytes that
506 	 *   is equal to 1/4 of the size of VRAM for this IB.
507 
508 	 * - If more than one half of VRAM is occupied, only allow relocating
509 	 *   1 MB of data for this IB.
510 	 *
511 	 * - From 0 to one half of used VRAM, the threshold decreases
512 	 *   linearly.
513 	 *         __________________
514 	 * 1/4 of -|\               |
515 	 * VRAM    | \              |
516 	 *         |  \             |
517 	 *         |   \            |
518 	 *         |    \           |
519 	 *         |     \          |
520 	 *         |      \         |
521 	 *         |       \________|1 MB
522 	 *         |----------------|
523 	 *    VRAM 0 %             100 %
524 	 *         used            used
525 	 *
526 	 * Note: It's a threshold, not a limit. The threshold must be crossed
527 	 * for buffer relocations to stop, so any buffer of an arbitrary size
528 	 * can be moved as long as the threshold isn't crossed before
529 	 * the relocation takes place. We don't want to disable buffer
530 	 * relocations completely.
531 	 *
532 	 * The idea is that buffers should be placed in VRAM at creation time
533 	 * and TTM should only do a minimum number of relocations during
534 	 * command submission. In practice, you need to submit at least
535 	 * a dozen IBs to move all buffers to VRAM if they are in GTT.
536 	 *
537 	 * Also, things can get pretty crazy under memory pressure and actual
538 	 * VRAM usage can change a lot, so playing safe even at 50% does
539 	 * consistently increase performance.
540 	 */
541 
542 	u64 half_vram = real_vram_size >> 1;
543 	u64 half_free_vram = vram_usage >= half_vram ? 0 : half_vram - vram_usage;
544 	u64 bytes_moved_threshold = half_free_vram >> 1;
545 	return max(bytes_moved_threshold, 1024*1024ull);
546 }
547 
radeon_bo_list_validate(struct radeon_device * rdev,struct ww_acquire_ctx * ticket,struct list_head * head,int ring)548 int radeon_bo_list_validate(struct radeon_device *rdev,
549 			    struct ww_acquire_ctx *ticket,
550 			    struct list_head *head, int ring)
551 {
552 	struct ttm_operation_ctx ctx = { true, false };
553 	struct radeon_bo_list *lobj;
554 	struct list_head duplicates;
555 	int r;
556 	u64 bytes_moved = 0, initial_bytes_moved;
557 	u64 bytes_moved_threshold = radeon_bo_get_threshold_for_moves(rdev);
558 
559 	INIT_LIST_HEAD(&duplicates);
560 	r = ttm_eu_reserve_buffers(ticket, head, true, &duplicates);
561 	if (unlikely(r != 0)) {
562 		return r;
563 	}
564 
565 	list_for_each_entry(lobj, head, tv.head) {
566 		struct radeon_bo *bo = lobj->robj;
567 		if (!bo->pin_count) {
568 			u32 domain = lobj->preferred_domains;
569 			u32 allowed = lobj->allowed_domains;
570 			u32 current_domain =
571 				radeon_mem_type_to_domain(bo->tbo.mem.mem_type);
572 
573 			/* Check if this buffer will be moved and don't move it
574 			 * if we have moved too many buffers for this IB already.
575 			 *
576 			 * Note that this allows moving at least one buffer of
577 			 * any size, because it doesn't take the current "bo"
578 			 * into account. We don't want to disallow buffer moves
579 			 * completely.
580 			 */
581 			if ((allowed & current_domain) != 0 &&
582 			    (domain & current_domain) == 0 && /* will be moved */
583 			    bytes_moved > bytes_moved_threshold) {
584 				/* don't move it */
585 				domain = current_domain;
586 			}
587 
588 		retry:
589 			radeon_ttm_placement_from_domain(bo, domain);
590 			if (ring == R600_RING_TYPE_UVD_INDEX)
591 				radeon_uvd_force_into_uvd_segment(bo, allowed);
592 
593 			initial_bytes_moved = atomic64_read(&rdev->num_bytes_moved);
594 			r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
595 			bytes_moved += atomic64_read(&rdev->num_bytes_moved) -
596 				       initial_bytes_moved;
597 
598 			if (unlikely(r)) {
599 				if (r != -ERESTARTSYS &&
600 				    domain != lobj->allowed_domains) {
601 					domain = lobj->allowed_domains;
602 					goto retry;
603 				}
604 				ttm_eu_backoff_reservation(ticket, head);
605 				return r;
606 			}
607 		}
608 		lobj->gpu_offset = radeon_bo_gpu_offset(bo);
609 		lobj->tiling_flags = bo->tiling_flags;
610 	}
611 
612 	list_for_each_entry(lobj, &duplicates, tv.head) {
613 		lobj->gpu_offset = radeon_bo_gpu_offset(lobj->robj);
614 		lobj->tiling_flags = lobj->robj->tiling_flags;
615 	}
616 
617 	return 0;
618 }
619 
radeon_bo_get_surface_reg(struct radeon_bo * bo)620 int radeon_bo_get_surface_reg(struct radeon_bo *bo)
621 {
622 	struct radeon_device *rdev = bo->rdev;
623 	struct radeon_surface_reg *reg;
624 	struct radeon_bo *old_object;
625 	int steal;
626 	int i;
627 
628 	dma_resv_assert_held(bo->tbo.base.resv);
629 
630 	if (!bo->tiling_flags)
631 		return 0;
632 
633 	if (bo->surface_reg >= 0) {
634 		reg = &rdev->surface_regs[bo->surface_reg];
635 		i = bo->surface_reg;
636 		goto out;
637 	}
638 
639 	steal = -1;
640 	for (i = 0; i < RADEON_GEM_MAX_SURFACES; i++) {
641 
642 		reg = &rdev->surface_regs[i];
643 		if (!reg->bo)
644 			break;
645 
646 		old_object = reg->bo;
647 		if (old_object->pin_count == 0)
648 			steal = i;
649 	}
650 
651 	/* if we are all out */
652 	if (i == RADEON_GEM_MAX_SURFACES) {
653 		if (steal == -1)
654 			return -ENOMEM;
655 		/* find someone with a surface reg and nuke their BO */
656 		reg = &rdev->surface_regs[steal];
657 		old_object = reg->bo;
658 		/* blow away the mapping */
659 		DRM_DEBUG("stealing surface reg %d from %p\n", steal, old_object);
660 		ttm_bo_unmap_virtual(&old_object->tbo);
661 		old_object->surface_reg = -1;
662 		i = steal;
663 	}
664 
665 	bo->surface_reg = i;
666 	reg->bo = bo;
667 
668 out:
669 	radeon_set_surface_reg(rdev, i, bo->tiling_flags, bo->pitch,
670 			       bo->tbo.mem.start << PAGE_SHIFT,
671 			       bo->tbo.num_pages << PAGE_SHIFT);
672 	return 0;
673 }
674 
radeon_bo_clear_surface_reg(struct radeon_bo * bo)675 static void radeon_bo_clear_surface_reg(struct radeon_bo *bo)
676 {
677 	struct radeon_device *rdev = bo->rdev;
678 	struct radeon_surface_reg *reg;
679 
680 	if (bo->surface_reg == -1)
681 		return;
682 
683 	reg = &rdev->surface_regs[bo->surface_reg];
684 	radeon_clear_surface_reg(rdev, bo->surface_reg);
685 
686 	reg->bo = NULL;
687 	bo->surface_reg = -1;
688 }
689 
radeon_bo_set_tiling_flags(struct radeon_bo * bo,uint32_t tiling_flags,uint32_t pitch)690 int radeon_bo_set_tiling_flags(struct radeon_bo *bo,
691 				uint32_t tiling_flags, uint32_t pitch)
692 {
693 	struct radeon_device *rdev = bo->rdev;
694 	int r;
695 
696 	if (rdev->family >= CHIP_CEDAR) {
697 		unsigned bankw, bankh, mtaspect, tilesplit, stilesplit;
698 
699 		bankw = (tiling_flags >> RADEON_TILING_EG_BANKW_SHIFT) & RADEON_TILING_EG_BANKW_MASK;
700 		bankh = (tiling_flags >> RADEON_TILING_EG_BANKH_SHIFT) & RADEON_TILING_EG_BANKH_MASK;
701 		mtaspect = (tiling_flags >> RADEON_TILING_EG_MACRO_TILE_ASPECT_SHIFT) & RADEON_TILING_EG_MACRO_TILE_ASPECT_MASK;
702 		tilesplit = (tiling_flags >> RADEON_TILING_EG_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_TILE_SPLIT_MASK;
703 		stilesplit = (tiling_flags >> RADEON_TILING_EG_STENCIL_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_STENCIL_TILE_SPLIT_MASK;
704 		switch (bankw) {
705 		case 0:
706 		case 1:
707 		case 2:
708 		case 4:
709 		case 8:
710 			break;
711 		default:
712 			return -EINVAL;
713 		}
714 		switch (bankh) {
715 		case 0:
716 		case 1:
717 		case 2:
718 		case 4:
719 		case 8:
720 			break;
721 		default:
722 			return -EINVAL;
723 		}
724 		switch (mtaspect) {
725 		case 0:
726 		case 1:
727 		case 2:
728 		case 4:
729 		case 8:
730 			break;
731 		default:
732 			return -EINVAL;
733 		}
734 		if (tilesplit > 6) {
735 			return -EINVAL;
736 		}
737 		if (stilesplit > 6) {
738 			return -EINVAL;
739 		}
740 	}
741 	r = radeon_bo_reserve(bo, false);
742 	if (unlikely(r != 0))
743 		return r;
744 	bo->tiling_flags = tiling_flags;
745 	bo->pitch = pitch;
746 	radeon_bo_unreserve(bo);
747 	return 0;
748 }
749 
radeon_bo_get_tiling_flags(struct radeon_bo * bo,uint32_t * tiling_flags,uint32_t * pitch)750 void radeon_bo_get_tiling_flags(struct radeon_bo *bo,
751 				uint32_t *tiling_flags,
752 				uint32_t *pitch)
753 {
754 	dma_resv_assert_held(bo->tbo.base.resv);
755 
756 	if (tiling_flags)
757 		*tiling_flags = bo->tiling_flags;
758 	if (pitch)
759 		*pitch = bo->pitch;
760 }
761 
radeon_bo_check_tiling(struct radeon_bo * bo,bool has_moved,bool force_drop)762 int radeon_bo_check_tiling(struct radeon_bo *bo, bool has_moved,
763 				bool force_drop)
764 {
765 	if (!force_drop)
766 		dma_resv_assert_held(bo->tbo.base.resv);
767 
768 	if (!(bo->tiling_flags & RADEON_TILING_SURFACE))
769 		return 0;
770 
771 	if (force_drop) {
772 		radeon_bo_clear_surface_reg(bo);
773 		return 0;
774 	}
775 
776 	if (bo->tbo.mem.mem_type != TTM_PL_VRAM) {
777 		if (!has_moved)
778 			return 0;
779 
780 		if (bo->surface_reg >= 0)
781 			radeon_bo_clear_surface_reg(bo);
782 		return 0;
783 	}
784 
785 	if ((bo->surface_reg >= 0) && !has_moved)
786 		return 0;
787 
788 	return radeon_bo_get_surface_reg(bo);
789 }
790 
radeon_bo_move_notify(struct ttm_buffer_object * bo,bool evict,struct ttm_mem_reg * new_mem)791 void radeon_bo_move_notify(struct ttm_buffer_object *bo,
792 			   bool evict,
793 			   struct ttm_mem_reg *new_mem)
794 {
795 	struct radeon_bo *rbo;
796 
797 	if (!radeon_ttm_bo_is_radeon_bo(bo))
798 		return;
799 
800 	rbo = container_of(bo, struct radeon_bo, tbo);
801 	radeon_bo_check_tiling(rbo, 0, 1);
802 	radeon_vm_bo_invalidate(rbo->rdev, rbo);
803 
804 	/* update statistics */
805 	if (!new_mem)
806 		return;
807 
808 	radeon_update_memory_usage(rbo, bo->mem.mem_type, -1);
809 	radeon_update_memory_usage(rbo, new_mem->mem_type, 1);
810 }
811 
radeon_bo_fault_reserve_notify(struct ttm_buffer_object * bo)812 int radeon_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
813 {
814 	struct ttm_operation_ctx ctx = { false, false };
815 	struct radeon_device *rdev;
816 	struct radeon_bo *rbo;
817 	unsigned long offset, size, lpfn;
818 	int i, r;
819 
820 	if (!radeon_ttm_bo_is_radeon_bo(bo))
821 		return 0;
822 	rbo = container_of(bo, struct radeon_bo, tbo);
823 	radeon_bo_check_tiling(rbo, 0, 0);
824 	rdev = rbo->rdev;
825 	if (bo->mem.mem_type != TTM_PL_VRAM)
826 		return 0;
827 
828 	size = bo->mem.num_pages << PAGE_SHIFT;
829 	offset = bo->mem.start << PAGE_SHIFT;
830 	if ((offset + size) <= rdev->mc.visible_vram_size)
831 		return 0;
832 
833 	/* Can't move a pinned BO to visible VRAM */
834 	if (rbo->pin_count > 0)
835 		return -EINVAL;
836 
837 	/* hurrah the memory is not visible ! */
838 	radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_VRAM);
839 	lpfn =	rdev->mc.visible_vram_size >> PAGE_SHIFT;
840 	for (i = 0; i < rbo->placement.num_placement; i++) {
841 		/* Force into visible VRAM */
842 		if ((rbo->placements[i].flags & TTM_PL_FLAG_VRAM) &&
843 		    (!rbo->placements[i].lpfn || rbo->placements[i].lpfn > lpfn))
844 			rbo->placements[i].lpfn = lpfn;
845 	}
846 	r = ttm_bo_validate(bo, &rbo->placement, &ctx);
847 	if (unlikely(r == -ENOMEM)) {
848 		radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_GTT);
849 		return ttm_bo_validate(bo, &rbo->placement, &ctx);
850 	} else if (unlikely(r != 0)) {
851 		return r;
852 	}
853 
854 	offset = bo->mem.start << PAGE_SHIFT;
855 	/* this should never happen */
856 	if ((offset + size) > rdev->mc.visible_vram_size)
857 		return -EINVAL;
858 
859 	return 0;
860 }
861 
radeon_bo_wait(struct radeon_bo * bo,u32 * mem_type,bool no_wait)862 int radeon_bo_wait(struct radeon_bo *bo, u32 *mem_type, bool no_wait)
863 {
864 	int r;
865 
866 	r = ttm_bo_reserve(&bo->tbo, true, no_wait, NULL);
867 	if (unlikely(r != 0))
868 		return r;
869 	if (mem_type)
870 		*mem_type = bo->tbo.mem.mem_type;
871 
872 	r = ttm_bo_wait(&bo->tbo, true, no_wait);
873 	ttm_bo_unreserve(&bo->tbo);
874 	return r;
875 }
876 
877 /**
878  * radeon_bo_fence - add fence to buffer object
879  *
880  * @bo: buffer object in question
881  * @fence: fence to add
882  * @shared: true if fence should be added shared
883  *
884  */
radeon_bo_fence(struct radeon_bo * bo,struct radeon_fence * fence,bool shared)885 void radeon_bo_fence(struct radeon_bo *bo, struct radeon_fence *fence,
886 		     bool shared)
887 {
888 	struct dma_resv *resv = bo->tbo.base.resv;
889 
890 	if (shared)
891 		dma_resv_add_shared_fence(resv, &fence->base);
892 	else
893 		dma_resv_add_excl_fence(resv, &fence->base);
894 }
895