Linux Audio

Check our new training course

Loading...
v4.6
  1/*
  2 * Copyright 2009 Jerome Glisse.
  3 * All Rights Reserved.
  4 *
  5 * Permission is hereby granted, free of charge, to any person obtaining a
  6 * copy of this software and associated documentation files (the
  7 * "Software"), to deal in the Software without restriction, including
  8 * without limitation the rights to use, copy, modify, merge, publish,
  9 * distribute, sub license, and/or sell copies of the Software, and to
 10 * permit persons to whom the Software is furnished to do so, subject to
 11 * the following conditions:
 12 *
 13 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 14 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 15 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
 16 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
 17 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
 18 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
 19 * USE OR OTHER DEALINGS IN THE SOFTWARE.
 20 *
 21 * The above copyright notice and this permission notice (including the
 22 * next paragraph) shall be included in all copies or substantial portions
 23 * of the Software.
 24 *
 25 */
 26/*
 27 * Authors:
 28 *    Jerome Glisse <glisse@freedesktop.org>
 29 *    Thomas Hellstrom <thomas-at-tungstengraphics-dot-com>
 30 *    Dave Airlie
 31 */
 
 
 32#include <linux/list.h>
 33#include <linux/slab.h>
 34#include <drm/drmP.h>
 35#include <drm/radeon_drm.h>
 36#include <drm/drm_cache.h>
 
 
 
 37#include "radeon.h"
 38#include "radeon_trace.h"
 
 39
 40
 41int radeon_ttm_init(struct radeon_device *rdev);
 42void radeon_ttm_fini(struct radeon_device *rdev);
 43static void radeon_bo_clear_surface_reg(struct radeon_bo *bo);
 44
 45/*
 46 * To exclude mutual BO access we rely on bo_reserve exclusion, as all
 47 * function are calling it.
 48 */
 49
 50static void radeon_update_memory_usage(struct radeon_bo *bo,
 51				       unsigned mem_type, int sign)
 52{
 53	struct radeon_device *rdev = bo->rdev;
 54	u64 size = (u64)bo->tbo.num_pages << PAGE_SHIFT;
 55
 56	switch (mem_type) {
 57	case TTM_PL_TT:
 58		if (sign > 0)
 59			atomic64_add(size, &rdev->gtt_usage);
 60		else
 61			atomic64_sub(size, &rdev->gtt_usage);
 62		break;
 63	case TTM_PL_VRAM:
 64		if (sign > 0)
 65			atomic64_add(size, &rdev->vram_usage);
 66		else
 67			atomic64_sub(size, &rdev->vram_usage);
 68		break;
 69	}
 70}
 71
 72static void radeon_ttm_bo_destroy(struct ttm_buffer_object *tbo)
 73{
 74	struct radeon_bo *bo;
 75
 76	bo = container_of(tbo, struct radeon_bo, tbo);
 77
 78	radeon_update_memory_usage(bo, bo->tbo.mem.mem_type, -1);
 79
 80	mutex_lock(&bo->rdev->gem.mutex);
 81	list_del_init(&bo->list);
 82	mutex_unlock(&bo->rdev->gem.mutex);
 83	radeon_bo_clear_surface_reg(bo);
 84	WARN_ON(!list_empty(&bo->va));
 85	drm_gem_object_release(&bo->gem_base);
 
 
 86	kfree(bo);
 87}
 88
 89bool radeon_ttm_bo_is_radeon_bo(struct ttm_buffer_object *bo)
 90{
 91	if (bo->destroy == &radeon_ttm_bo_destroy)
 92		return true;
 93	return false;
 94}
 95
 96void radeon_ttm_placement_from_domain(struct radeon_bo *rbo, u32 domain)
 97{
 98	u32 c = 0, i;
 99
100	rbo->placement.placement = rbo->placements;
101	rbo->placement.busy_placement = rbo->placements;
102	if (domain & RADEON_GEM_DOMAIN_VRAM) {
103		/* Try placing BOs which don't need CPU access outside of the
104		 * CPU accessible part of VRAM
105		 */
106		if ((rbo->flags & RADEON_GEM_NO_CPU_ACCESS) &&
107		    rbo->rdev->mc.visible_vram_size < rbo->rdev->mc.real_vram_size) {
108			rbo->placements[c].fpfn =
109				rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
110			rbo->placements[c++].flags = TTM_PL_FLAG_WC |
111						     TTM_PL_FLAG_UNCACHED |
112						     TTM_PL_FLAG_VRAM;
113		}
114
115		rbo->placements[c].fpfn = 0;
116		rbo->placements[c++].flags = TTM_PL_FLAG_WC |
117					     TTM_PL_FLAG_UNCACHED |
118					     TTM_PL_FLAG_VRAM;
119	}
120
121	if (domain & RADEON_GEM_DOMAIN_GTT) {
122		if (rbo->flags & RADEON_GEM_GTT_UC) {
123			rbo->placements[c].fpfn = 0;
124			rbo->placements[c++].flags = TTM_PL_FLAG_UNCACHED |
125				TTM_PL_FLAG_TT;
126
127		} else if ((rbo->flags & RADEON_GEM_GTT_WC) ||
128			   (rbo->rdev->flags & RADEON_IS_AGP)) {
129			rbo->placements[c].fpfn = 0;
130			rbo->placements[c++].flags = TTM_PL_FLAG_WC |
131				TTM_PL_FLAG_UNCACHED |
132				TTM_PL_FLAG_TT;
133		} else {
134			rbo->placements[c].fpfn = 0;
135			rbo->placements[c++].flags = TTM_PL_FLAG_CACHED |
136						     TTM_PL_FLAG_TT;
137		}
138	}
139
140	if (domain & RADEON_GEM_DOMAIN_CPU) {
141		if (rbo->flags & RADEON_GEM_GTT_UC) {
142			rbo->placements[c].fpfn = 0;
143			rbo->placements[c++].flags = TTM_PL_FLAG_UNCACHED |
144				TTM_PL_FLAG_SYSTEM;
145
146		} else if ((rbo->flags & RADEON_GEM_GTT_WC) ||
147		    rbo->rdev->flags & RADEON_IS_AGP) {
148			rbo->placements[c].fpfn = 0;
149			rbo->placements[c++].flags = TTM_PL_FLAG_WC |
150				TTM_PL_FLAG_UNCACHED |
151				TTM_PL_FLAG_SYSTEM;
152		} else {
153			rbo->placements[c].fpfn = 0;
154			rbo->placements[c++].flags = TTM_PL_FLAG_CACHED |
155						     TTM_PL_FLAG_SYSTEM;
156		}
157	}
158	if (!c) {
159		rbo->placements[c].fpfn = 0;
160		rbo->placements[c++].flags = TTM_PL_MASK_CACHING |
161					     TTM_PL_FLAG_SYSTEM;
162	}
163
164	rbo->placement.num_placement = c;
165	rbo->placement.num_busy_placement = c;
166
167	for (i = 0; i < c; ++i) {
168		if ((rbo->flags & RADEON_GEM_CPU_ACCESS) &&
169		    (rbo->placements[i].flags & TTM_PL_FLAG_VRAM) &&
170		    !rbo->placements[i].fpfn)
171			rbo->placements[i].lpfn =
172				rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
173		else
174			rbo->placements[i].lpfn = 0;
175	}
176}
177
178int radeon_bo_create(struct radeon_device *rdev,
179		     unsigned long size, int byte_align, bool kernel,
180		     u32 domain, u32 flags, struct sg_table *sg,
181		     struct reservation_object *resv,
182		     struct radeon_bo **bo_ptr)
183{
184	struct radeon_bo *bo;
185	enum ttm_bo_type type;
186	unsigned long page_align = roundup(byte_align, PAGE_SIZE) >> PAGE_SHIFT;
187	size_t acc_size;
188	int r;
189
190	size = ALIGN(size, PAGE_SIZE);
191
192	if (kernel) {
193		type = ttm_bo_type_kernel;
194	} else if (sg) {
195		type = ttm_bo_type_sg;
196	} else {
197		type = ttm_bo_type_device;
198	}
199	*bo_ptr = NULL;
200
201	acc_size = ttm_bo_dma_acc_size(&rdev->mman.bdev, size,
202				       sizeof(struct radeon_bo));
203
204	bo = kzalloc(sizeof(struct radeon_bo), GFP_KERNEL);
205	if (bo == NULL)
206		return -ENOMEM;
207	r = drm_gem_object_init(rdev->ddev, &bo->gem_base, size);
208	if (unlikely(r)) {
209		kfree(bo);
210		return r;
211	}
212	bo->rdev = rdev;
213	bo->surface_reg = -1;
214	INIT_LIST_HEAD(&bo->list);
215	INIT_LIST_HEAD(&bo->va);
216	bo->initial_domain = domain & (RADEON_GEM_DOMAIN_VRAM |
217				       RADEON_GEM_DOMAIN_GTT |
218				       RADEON_GEM_DOMAIN_CPU);
219
220	bo->flags = flags;
221	/* PCI GART is always snooped */
222	if (!(rdev->flags & RADEON_IS_PCIE))
223		bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
224
225	/* Write-combined CPU mappings of GTT cause GPU hangs with RV6xx
226	 * See https://bugs.freedesktop.org/show_bug.cgi?id=91268
227	 */
228	if (rdev->family >= CHIP_RV610 && rdev->family <= CHIP_RV635)
229		bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
230
231#ifdef CONFIG_X86_32
232	/* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
233	 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
234	 */
235	bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
236#elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
237	/* Don't try to enable write-combining when it can't work, or things
238	 * may be slow
239	 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
240	 */
241
242#warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
243	 thanks to write-combining
 
244
245	if (bo->flags & RADEON_GEM_GTT_WC)
246		DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
247			      "better performance thanks to write-combining\n");
248	bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
249#else
250	/* For architectures that don't support WC memory,
251	 * mask out the WC flag from the BO
252	 */
253	if (!drm_arch_can_wc_memory())
254		bo->flags &= ~RADEON_GEM_GTT_WC;
255#endif
256
257	radeon_ttm_placement_from_domain(bo, domain);
258	/* Kernel allocation are uninterruptible */
259	down_read(&rdev->pm.mclk_lock);
260	r = ttm_bo_init(&rdev->mman.bdev, &bo->tbo, size, type,
261			&bo->placement, page_align, !kernel, NULL,
262			acc_size, sg, resv, &radeon_ttm_bo_destroy);
263	up_read(&rdev->pm.mclk_lock);
264	if (unlikely(r != 0)) {
265		return r;
266	}
267	*bo_ptr = bo;
268
269	trace_radeon_bo_create(bo);
270
271	return 0;
272}
273
274int radeon_bo_kmap(struct radeon_bo *bo, void **ptr)
275{
276	bool is_iomem;
277	int r;
278
279	if (bo->kptr) {
280		if (ptr) {
281			*ptr = bo->kptr;
282		}
283		return 0;
284	}
285	r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages, &bo->kmap);
286	if (r) {
287		return r;
288	}
289	bo->kptr = ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
290	if (ptr) {
291		*ptr = bo->kptr;
292	}
293	radeon_bo_check_tiling(bo, 0, 0);
294	return 0;
295}
296
297void radeon_bo_kunmap(struct radeon_bo *bo)
298{
299	if (bo->kptr == NULL)
300		return;
301	bo->kptr = NULL;
302	radeon_bo_check_tiling(bo, 0, 0);
303	ttm_bo_kunmap(&bo->kmap);
304}
305
306struct radeon_bo *radeon_bo_ref(struct radeon_bo *bo)
307{
308	if (bo == NULL)
309		return NULL;
310
311	ttm_bo_reference(&bo->tbo);
312	return bo;
313}
314
315void radeon_bo_unref(struct radeon_bo **bo)
316{
317	struct ttm_buffer_object *tbo;
318	struct radeon_device *rdev;
319
320	if ((*bo) == NULL)
321		return;
322	rdev = (*bo)->rdev;
323	tbo = &((*bo)->tbo);
324	ttm_bo_unref(&tbo);
325	if (tbo == NULL)
326		*bo = NULL;
327}
328
329int radeon_bo_pin_restricted(struct radeon_bo *bo, u32 domain, u64 max_offset,
330			     u64 *gpu_addr)
331{
 
332	int r, i;
333
334	if (radeon_ttm_tt_has_userptr(bo->tbo.ttm))
335		return -EPERM;
336
337	if (bo->pin_count) {
338		bo->pin_count++;
339		if (gpu_addr)
340			*gpu_addr = radeon_bo_gpu_offset(bo);
341
342		if (max_offset != 0) {
343			u64 domain_start;
344
345			if (domain == RADEON_GEM_DOMAIN_VRAM)
346				domain_start = bo->rdev->mc.vram_start;
347			else
348				domain_start = bo->rdev->mc.gtt_start;
349			WARN_ON_ONCE(max_offset <
350				     (radeon_bo_gpu_offset(bo) - domain_start));
351		}
352
353		return 0;
354	}
 
 
 
 
 
355	radeon_ttm_placement_from_domain(bo, domain);
356	for (i = 0; i < bo->placement.num_placement; i++) {
357		/* force to pin into visible video ram */
358		if ((bo->placements[i].flags & TTM_PL_FLAG_VRAM) &&
359		    !(bo->flags & RADEON_GEM_NO_CPU_ACCESS) &&
360		    (!max_offset || max_offset > bo->rdev->mc.visible_vram_size))
361			bo->placements[i].lpfn =
362				bo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
363		else
364			bo->placements[i].lpfn = max_offset >> PAGE_SHIFT;
365
366		bo->placements[i].flags |= TTM_PL_FLAG_NO_EVICT;
367	}
368
369	r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
370	if (likely(r == 0)) {
371		bo->pin_count = 1;
372		if (gpu_addr != NULL)
373			*gpu_addr = radeon_bo_gpu_offset(bo);
374		if (domain == RADEON_GEM_DOMAIN_VRAM)
375			bo->rdev->vram_pin_size += radeon_bo_size(bo);
376		else
377			bo->rdev->gart_pin_size += radeon_bo_size(bo);
378	} else {
379		dev_err(bo->rdev->dev, "%p pin failed\n", bo);
380	}
381	return r;
382}
383
384int radeon_bo_pin(struct radeon_bo *bo, u32 domain, u64 *gpu_addr)
385{
386	return radeon_bo_pin_restricted(bo, domain, 0, gpu_addr);
387}
388
389int radeon_bo_unpin(struct radeon_bo *bo)
390{
391	int r, i;
392
393	if (!bo->pin_count) {
394		dev_warn(bo->rdev->dev, "%p unpin not necessary\n", bo);
395		return 0;
396	}
397	bo->pin_count--;
398	if (bo->pin_count)
399		return 0;
400	for (i = 0; i < bo->placement.num_placement; i++) {
401		bo->placements[i].lpfn = 0;
402		bo->placements[i].flags &= ~TTM_PL_FLAG_NO_EVICT;
403	}
404	r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
405	if (likely(r == 0)) {
406		if (bo->tbo.mem.mem_type == TTM_PL_VRAM)
407			bo->rdev->vram_pin_size -= radeon_bo_size(bo);
408		else
409			bo->rdev->gart_pin_size -= radeon_bo_size(bo);
410	} else {
411		dev_err(bo->rdev->dev, "%p validate failed for unpin\n", bo);
412	}
413	return r;
414}
415
416int radeon_bo_evict_vram(struct radeon_device *rdev)
417{
 
 
 
418	/* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
419	if (0 && (rdev->flags & RADEON_IS_IGP)) {
 
420		if (rdev->mc.igp_sideport_enabled == false)
421			/* Useless to evict on IGP chips */
422			return 0;
423	}
424	return ttm_bo_evict_mm(&rdev->mman.bdev, TTM_PL_VRAM);
 
 
 
 
425}
426
427void radeon_bo_force_delete(struct radeon_device *rdev)
428{
429	struct radeon_bo *bo, *n;
430
431	if (list_empty(&rdev->gem.objects)) {
432		return;
433	}
434	dev_err(rdev->dev, "Userspace still has active objects !\n");
435	list_for_each_entry_safe(bo, n, &rdev->gem.objects, list) {
436		dev_err(rdev->dev, "%p %p %lu %lu force free\n",
437			&bo->gem_base, bo, (unsigned long)bo->gem_base.size,
438			*((unsigned long *)&bo->gem_base.refcount));
439		mutex_lock(&bo->rdev->gem.mutex);
440		list_del_init(&bo->list);
441		mutex_unlock(&bo->rdev->gem.mutex);
442		/* this should unref the ttm bo */
443		drm_gem_object_unreference_unlocked(&bo->gem_base);
444	}
445}
446
447int radeon_bo_init(struct radeon_device *rdev)
448{
 
 
 
 
449	/* Add an MTRR for the VRAM */
450	if (!rdev->fastfb_working) {
451		rdev->mc.vram_mtrr = arch_phys_wc_add(rdev->mc.aper_base,
452						      rdev->mc.aper_size);
453	}
454	DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
455		rdev->mc.mc_vram_size >> 20,
456		(unsigned long long)rdev->mc.aper_size >> 20);
457	DRM_INFO("RAM width %dbits %cDR\n",
458			rdev->mc.vram_width, rdev->mc.vram_is_ddr ? 'D' : 'S');
459	return radeon_ttm_init(rdev);
460}
461
462void radeon_bo_fini(struct radeon_device *rdev)
463{
464	radeon_ttm_fini(rdev);
465	arch_phys_wc_del(rdev->mc.vram_mtrr);
 
466}
467
468/* Returns how many bytes TTM can move per IB.
469 */
470static u64 radeon_bo_get_threshold_for_moves(struct radeon_device *rdev)
471{
472	u64 real_vram_size = rdev->mc.real_vram_size;
473	u64 vram_usage = atomic64_read(&rdev->vram_usage);
474
475	/* This function is based on the current VRAM usage.
476	 *
477	 * - If all of VRAM is free, allow relocating the number of bytes that
478	 *   is equal to 1/4 of the size of VRAM for this IB.
479
480	 * - If more than one half of VRAM is occupied, only allow relocating
481	 *   1 MB of data for this IB.
482	 *
483	 * - From 0 to one half of used VRAM, the threshold decreases
484	 *   linearly.
485	 *         __________________
486	 * 1/4 of -|\               |
487	 * VRAM    | \              |
488	 *         |  \             |
489	 *         |   \            |
490	 *         |    \           |
491	 *         |     \          |
492	 *         |      \         |
493	 *         |       \________|1 MB
494	 *         |----------------|
495	 *    VRAM 0 %             100 %
496	 *         used            used
497	 *
498	 * Note: It's a threshold, not a limit. The threshold must be crossed
499	 * for buffer relocations to stop, so any buffer of an arbitrary size
500	 * can be moved as long as the threshold isn't crossed before
501	 * the relocation takes place. We don't want to disable buffer
502	 * relocations completely.
503	 *
504	 * The idea is that buffers should be placed in VRAM at creation time
505	 * and TTM should only do a minimum number of relocations during
506	 * command submission. In practice, you need to submit at least
507	 * a dozen IBs to move all buffers to VRAM if they are in GTT.
508	 *
509	 * Also, things can get pretty crazy under memory pressure and actual
510	 * VRAM usage can change a lot, so playing safe even at 50% does
511	 * consistently increase performance.
512	 */
513
514	u64 half_vram = real_vram_size >> 1;
515	u64 half_free_vram = vram_usage >= half_vram ? 0 : half_vram - vram_usage;
516	u64 bytes_moved_threshold = half_free_vram >> 1;
517	return max(bytes_moved_threshold, 1024*1024ull);
518}
519
520int radeon_bo_list_validate(struct radeon_device *rdev,
521			    struct ww_acquire_ctx *ticket,
522			    struct list_head *head, int ring)
523{
 
524	struct radeon_bo_list *lobj;
525	struct list_head duplicates;
526	int r;
527	u64 bytes_moved = 0, initial_bytes_moved;
528	u64 bytes_moved_threshold = radeon_bo_get_threshold_for_moves(rdev);
529
530	INIT_LIST_HEAD(&duplicates);
531	r = ttm_eu_reserve_buffers(ticket, head, true, &duplicates);
532	if (unlikely(r != 0)) {
533		return r;
534	}
535
536	list_for_each_entry(lobj, head, tv.head) {
537		struct radeon_bo *bo = lobj->robj;
538		if (!bo->pin_count) {
539			u32 domain = lobj->prefered_domains;
540			u32 allowed = lobj->allowed_domains;
541			u32 current_domain =
542				radeon_mem_type_to_domain(bo->tbo.mem.mem_type);
543
544			/* Check if this buffer will be moved and don't move it
545			 * if we have moved too many buffers for this IB already.
546			 *
547			 * Note that this allows moving at least one buffer of
548			 * any size, because it doesn't take the current "bo"
549			 * into account. We don't want to disallow buffer moves
550			 * completely.
551			 */
552			if ((allowed & current_domain) != 0 &&
553			    (domain & current_domain) == 0 && /* will be moved */
554			    bytes_moved > bytes_moved_threshold) {
555				/* don't move it */
556				domain = current_domain;
557			}
558
559		retry:
560			radeon_ttm_placement_from_domain(bo, domain);
561			if (ring == R600_RING_TYPE_UVD_INDEX)
562				radeon_uvd_force_into_uvd_segment(bo, allowed);
563
564			initial_bytes_moved = atomic64_read(&rdev->num_bytes_moved);
565			r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
566			bytes_moved += atomic64_read(&rdev->num_bytes_moved) -
567				       initial_bytes_moved;
568
569			if (unlikely(r)) {
570				if (r != -ERESTARTSYS &&
571				    domain != lobj->allowed_domains) {
572					domain = lobj->allowed_domains;
573					goto retry;
574				}
575				ttm_eu_backoff_reservation(ticket, head);
576				return r;
577			}
578		}
579		lobj->gpu_offset = radeon_bo_gpu_offset(bo);
580		lobj->tiling_flags = bo->tiling_flags;
581	}
582
583	list_for_each_entry(lobj, &duplicates, tv.head) {
584		lobj->gpu_offset = radeon_bo_gpu_offset(lobj->robj);
585		lobj->tiling_flags = lobj->robj->tiling_flags;
586	}
587
588	return 0;
589}
590
591int radeon_bo_get_surface_reg(struct radeon_bo *bo)
592{
593	struct radeon_device *rdev = bo->rdev;
594	struct radeon_surface_reg *reg;
595	struct radeon_bo *old_object;
596	int steal;
597	int i;
598
599	lockdep_assert_held(&bo->tbo.resv->lock.base);
600
601	if (!bo->tiling_flags)
602		return 0;
603
604	if (bo->surface_reg >= 0) {
605		reg = &rdev->surface_regs[bo->surface_reg];
606		i = bo->surface_reg;
607		goto out;
608	}
609
610	steal = -1;
611	for (i = 0; i < RADEON_GEM_MAX_SURFACES; i++) {
612
613		reg = &rdev->surface_regs[i];
614		if (!reg->bo)
615			break;
616
617		old_object = reg->bo;
618		if (old_object->pin_count == 0)
619			steal = i;
620	}
621
622	/* if we are all out */
623	if (i == RADEON_GEM_MAX_SURFACES) {
624		if (steal == -1)
625			return -ENOMEM;
626		/* find someone with a surface reg and nuke their BO */
627		reg = &rdev->surface_regs[steal];
628		old_object = reg->bo;
629		/* blow away the mapping */
630		DRM_DEBUG("stealing surface reg %d from %p\n", steal, old_object);
631		ttm_bo_unmap_virtual(&old_object->tbo);
632		old_object->surface_reg = -1;
633		i = steal;
634	}
635
636	bo->surface_reg = i;
637	reg->bo = bo;
638
639out:
640	radeon_set_surface_reg(rdev, i, bo->tiling_flags, bo->pitch,
641			       bo->tbo.mem.start << PAGE_SHIFT,
642			       bo->tbo.num_pages << PAGE_SHIFT);
643	return 0;
644}
645
646static void radeon_bo_clear_surface_reg(struct radeon_bo *bo)
647{
648	struct radeon_device *rdev = bo->rdev;
649	struct radeon_surface_reg *reg;
650
651	if (bo->surface_reg == -1)
652		return;
653
654	reg = &rdev->surface_regs[bo->surface_reg];
655	radeon_clear_surface_reg(rdev, bo->surface_reg);
656
657	reg->bo = NULL;
658	bo->surface_reg = -1;
659}
660
661int radeon_bo_set_tiling_flags(struct radeon_bo *bo,
662				uint32_t tiling_flags, uint32_t pitch)
663{
664	struct radeon_device *rdev = bo->rdev;
665	int r;
666
667	if (rdev->family >= CHIP_CEDAR) {
668		unsigned bankw, bankh, mtaspect, tilesplit, stilesplit;
669
670		bankw = (tiling_flags >> RADEON_TILING_EG_BANKW_SHIFT) & RADEON_TILING_EG_BANKW_MASK;
671		bankh = (tiling_flags >> RADEON_TILING_EG_BANKH_SHIFT) & RADEON_TILING_EG_BANKH_MASK;
672		mtaspect = (tiling_flags >> RADEON_TILING_EG_MACRO_TILE_ASPECT_SHIFT) & RADEON_TILING_EG_MACRO_TILE_ASPECT_MASK;
673		tilesplit = (tiling_flags >> RADEON_TILING_EG_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_TILE_SPLIT_MASK;
674		stilesplit = (tiling_flags >> RADEON_TILING_EG_STENCIL_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_STENCIL_TILE_SPLIT_MASK;
675		switch (bankw) {
676		case 0:
677		case 1:
678		case 2:
679		case 4:
680		case 8:
681			break;
682		default:
683			return -EINVAL;
684		}
685		switch (bankh) {
686		case 0:
687		case 1:
688		case 2:
689		case 4:
690		case 8:
691			break;
692		default:
693			return -EINVAL;
694		}
695		switch (mtaspect) {
696		case 0:
697		case 1:
698		case 2:
699		case 4:
700		case 8:
701			break;
702		default:
703			return -EINVAL;
704		}
705		if (tilesplit > 6) {
706			return -EINVAL;
707		}
708		if (stilesplit > 6) {
709			return -EINVAL;
710		}
711	}
712	r = radeon_bo_reserve(bo, false);
713	if (unlikely(r != 0))
714		return r;
715	bo->tiling_flags = tiling_flags;
716	bo->pitch = pitch;
717	radeon_bo_unreserve(bo);
718	return 0;
719}
720
721void radeon_bo_get_tiling_flags(struct radeon_bo *bo,
722				uint32_t *tiling_flags,
723				uint32_t *pitch)
724{
725	lockdep_assert_held(&bo->tbo.resv->lock.base);
726
727	if (tiling_flags)
728		*tiling_flags = bo->tiling_flags;
729	if (pitch)
730		*pitch = bo->pitch;
731}
732
733int radeon_bo_check_tiling(struct radeon_bo *bo, bool has_moved,
734				bool force_drop)
735{
736	if (!force_drop)
737		lockdep_assert_held(&bo->tbo.resv->lock.base);
738
739	if (!(bo->tiling_flags & RADEON_TILING_SURFACE))
740		return 0;
741
742	if (force_drop) {
743		radeon_bo_clear_surface_reg(bo);
744		return 0;
745	}
746
747	if (bo->tbo.mem.mem_type != TTM_PL_VRAM) {
748		if (!has_moved)
749			return 0;
750
751		if (bo->surface_reg >= 0)
752			radeon_bo_clear_surface_reg(bo);
753		return 0;
754	}
755
756	if ((bo->surface_reg >= 0) && !has_moved)
757		return 0;
758
759	return radeon_bo_get_surface_reg(bo);
760}
761
762void radeon_bo_move_notify(struct ttm_buffer_object *bo,
763			   struct ttm_mem_reg *new_mem)
 
764{
765	struct radeon_bo *rbo;
766
 
 
 
 
767	if (!radeon_ttm_bo_is_radeon_bo(bo))
768		return;
769
770	rbo = container_of(bo, struct radeon_bo, tbo);
771	radeon_bo_check_tiling(rbo, 0, 1);
772	radeon_vm_bo_invalidate(rbo->rdev, rbo);
773
774	/* update statistics */
775	if (!new_mem)
776		return;
777
778	radeon_update_memory_usage(rbo, bo->mem.mem_type, -1);
779	radeon_update_memory_usage(rbo, new_mem->mem_type, 1);
780}
781
782int radeon_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
783{
 
784	struct radeon_device *rdev;
785	struct radeon_bo *rbo;
786	unsigned long offset, size, lpfn;
787	int i, r;
788
789	if (!radeon_ttm_bo_is_radeon_bo(bo))
790		return 0;
791	rbo = container_of(bo, struct radeon_bo, tbo);
792	radeon_bo_check_tiling(rbo, 0, 0);
793	rdev = rbo->rdev;
794	if (bo->mem.mem_type != TTM_PL_VRAM)
795		return 0;
796
797	size = bo->mem.num_pages << PAGE_SHIFT;
798	offset = bo->mem.start << PAGE_SHIFT;
799	if ((offset + size) <= rdev->mc.visible_vram_size)
800		return 0;
801
802	/* Can't move a pinned BO to visible VRAM */
803	if (rbo->pin_count > 0)
804		return -EINVAL;
805
806	/* hurrah the memory is not visible ! */
807	radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_VRAM);
808	lpfn =	rdev->mc.visible_vram_size >> PAGE_SHIFT;
809	for (i = 0; i < rbo->placement.num_placement; i++) {
810		/* Force into visible VRAM */
811		if ((rbo->placements[i].flags & TTM_PL_FLAG_VRAM) &&
812		    (!rbo->placements[i].lpfn || rbo->placements[i].lpfn > lpfn))
813			rbo->placements[i].lpfn = lpfn;
814	}
815	r = ttm_bo_validate(bo, &rbo->placement, false, false);
816	if (unlikely(r == -ENOMEM)) {
817		radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_GTT);
818		return ttm_bo_validate(bo, &rbo->placement, false, false);
819	} else if (unlikely(r != 0)) {
820		return r;
 
 
 
821	}
822
823	offset = bo->mem.start << PAGE_SHIFT;
824	/* this should never happen */
825	if ((offset + size) > rdev->mc.visible_vram_size)
826		return -EINVAL;
827
 
828	return 0;
829}
830
831int radeon_bo_wait(struct radeon_bo *bo, u32 *mem_type, bool no_wait)
832{
833	int r;
834
835	r = ttm_bo_reserve(&bo->tbo, true, no_wait, false, NULL);
836	if (unlikely(r != 0))
837		return r;
838	if (mem_type)
839		*mem_type = bo->tbo.mem.mem_type;
840
841	r = ttm_bo_wait(&bo->tbo, true, true, no_wait);
842	ttm_bo_unreserve(&bo->tbo);
843	return r;
844}
845
846/**
847 * radeon_bo_fence - add fence to buffer object
848 *
849 * @bo: buffer object in question
850 * @fence: fence to add
851 * @shared: true if fence should be added shared
852 *
853 */
854void radeon_bo_fence(struct radeon_bo *bo, struct radeon_fence *fence,
855		     bool shared)
856{
857	struct reservation_object *resv = bo->tbo.resv;
858
859	if (shared)
860		reservation_object_add_shared_fence(resv, &fence->base);
861	else
862		reservation_object_add_excl_fence(resv, &fence->base);
863}
v5.14.15
  1/*
  2 * Copyright 2009 Jerome Glisse.
  3 * All Rights Reserved.
  4 *
  5 * Permission is hereby granted, free of charge, to any person obtaining a
  6 * copy of this software and associated documentation files (the
  7 * "Software"), to deal in the Software without restriction, including
  8 * without limitation the rights to use, copy, modify, merge, publish,
  9 * distribute, sub license, and/or sell copies of the Software, and to
 10 * permit persons to whom the Software is furnished to do so, subject to
 11 * the following conditions:
 12 *
 13 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 14 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 15 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
 16 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
 17 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
 18 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
 19 * USE OR OTHER DEALINGS IN THE SOFTWARE.
 20 *
 21 * The above copyright notice and this permission notice (including the
 22 * next paragraph) shall be included in all copies or substantial portions
 23 * of the Software.
 24 *
 25 */
 26/*
 27 * Authors:
 28 *    Jerome Glisse <glisse@freedesktop.org>
 29 *    Thomas Hellstrom <thomas-at-tungstengraphics-dot-com>
 30 *    Dave Airlie
 31 */
 32
 33#include <linux/io.h>
 34#include <linux/list.h>
 35#include <linux/slab.h>
 36
 
 37#include <drm/drm_cache.h>
 38#include <drm/drm_prime.h>
 39#include <drm/radeon_drm.h>
 40
 41#include "radeon.h"
 42#include "radeon_trace.h"
 43#include "radeon_ttm.h"
 44
 
 
 
 45static void radeon_bo_clear_surface_reg(struct radeon_bo *bo);
 46
 47/*
 48 * To exclude mutual BO access we rely on bo_reserve exclusion, as all
 49 * function are calling it.
 50 */
 51
 52static void radeon_update_memory_usage(struct ttm_buffer_object *bo,
 53				       unsigned int mem_type, int sign)
 54{
 55	struct radeon_device *rdev = radeon_get_rdev(bo->bdev);
 
 56
 57	switch (mem_type) {
 58	case TTM_PL_TT:
 59		if (sign > 0)
 60			atomic64_add(bo->base.size, &rdev->gtt_usage);
 61		else
 62			atomic64_sub(bo->base.size, &rdev->gtt_usage);
 63		break;
 64	case TTM_PL_VRAM:
 65		if (sign > 0)
 66			atomic64_add(bo->base.size, &rdev->vram_usage);
 67		else
 68			atomic64_sub(bo->base.size, &rdev->vram_usage);
 69		break;
 70	}
 71}
 72
 73static void radeon_ttm_bo_destroy(struct ttm_buffer_object *tbo)
 74{
 75	struct radeon_bo *bo;
 76
 77	bo = container_of(tbo, struct radeon_bo, tbo);
 78
 
 
 79	mutex_lock(&bo->rdev->gem.mutex);
 80	list_del_init(&bo->list);
 81	mutex_unlock(&bo->rdev->gem.mutex);
 82	radeon_bo_clear_surface_reg(bo);
 83	WARN_ON_ONCE(!list_empty(&bo->va));
 84	if (bo->tbo.base.import_attach)
 85		drm_prime_gem_destroy(&bo->tbo.base, bo->tbo.sg);
 86	drm_gem_object_release(&bo->tbo.base);
 87	kfree(bo);
 88}
 89
 90bool radeon_ttm_bo_is_radeon_bo(struct ttm_buffer_object *bo)
 91{
 92	if (bo->destroy == &radeon_ttm_bo_destroy)
 93		return true;
 94	return false;
 95}
 96
 97void radeon_ttm_placement_from_domain(struct radeon_bo *rbo, u32 domain)
 98{
 99	u32 c = 0, i;
100
101	rbo->placement.placement = rbo->placements;
102	rbo->placement.busy_placement = rbo->placements;
103	if (domain & RADEON_GEM_DOMAIN_VRAM) {
104		/* Try placing BOs which don't need CPU access outside of the
105		 * CPU accessible part of VRAM
106		 */
107		if ((rbo->flags & RADEON_GEM_NO_CPU_ACCESS) &&
108		    rbo->rdev->mc.visible_vram_size < rbo->rdev->mc.real_vram_size) {
109			rbo->placements[c].fpfn =
110				rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
111			rbo->placements[c].mem_type = TTM_PL_VRAM;
112			rbo->placements[c++].flags = 0;
 
113		}
114
115		rbo->placements[c].fpfn = 0;
116		rbo->placements[c].mem_type = TTM_PL_VRAM;
117		rbo->placements[c++].flags = 0;
 
118	}
119
120	if (domain & RADEON_GEM_DOMAIN_GTT) {
121		rbo->placements[c].fpfn = 0;
122		rbo->placements[c].mem_type = TTM_PL_TT;
123		rbo->placements[c++].flags = 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
124	}
125
126	if (domain & RADEON_GEM_DOMAIN_CPU) {
127		rbo->placements[c].fpfn = 0;
128		rbo->placements[c].mem_type = TTM_PL_SYSTEM;
129		rbo->placements[c++].flags = 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
130	}
131	if (!c) {
132		rbo->placements[c].fpfn = 0;
133		rbo->placements[c].mem_type = TTM_PL_SYSTEM;
134		rbo->placements[c++].flags = 0;
135	}
136
137	rbo->placement.num_placement = c;
138	rbo->placement.num_busy_placement = c;
139
140	for (i = 0; i < c; ++i) {
141		if ((rbo->flags & RADEON_GEM_CPU_ACCESS) &&
142		    (rbo->placements[i].mem_type == TTM_PL_VRAM) &&
143		    !rbo->placements[i].fpfn)
144			rbo->placements[i].lpfn =
145				rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
146		else
147			rbo->placements[i].lpfn = 0;
148	}
149}
150
151int radeon_bo_create(struct radeon_device *rdev,
152		     unsigned long size, int byte_align, bool kernel,
153		     u32 domain, u32 flags, struct sg_table *sg,
154		     struct dma_resv *resv,
155		     struct radeon_bo **bo_ptr)
156{
157	struct radeon_bo *bo;
158	enum ttm_bo_type type;
159	unsigned long page_align = roundup(byte_align, PAGE_SIZE) >> PAGE_SHIFT;
 
160	int r;
161
162	size = ALIGN(size, PAGE_SIZE);
163
164	if (kernel) {
165		type = ttm_bo_type_kernel;
166	} else if (sg) {
167		type = ttm_bo_type_sg;
168	} else {
169		type = ttm_bo_type_device;
170	}
171	*bo_ptr = NULL;
172
 
 
 
173	bo = kzalloc(sizeof(struct radeon_bo), GFP_KERNEL);
174	if (bo == NULL)
175		return -ENOMEM;
176	drm_gem_private_object_init(rdev->ddev, &bo->tbo.base, size);
 
 
 
 
177	bo->rdev = rdev;
178	bo->surface_reg = -1;
179	INIT_LIST_HEAD(&bo->list);
180	INIT_LIST_HEAD(&bo->va);
181	bo->initial_domain = domain & (RADEON_GEM_DOMAIN_VRAM |
182				       RADEON_GEM_DOMAIN_GTT |
183				       RADEON_GEM_DOMAIN_CPU);
184
185	bo->flags = flags;
186	/* PCI GART is always snooped */
187	if (!(rdev->flags & RADEON_IS_PCIE))
188		bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
189
190	/* Write-combined CPU mappings of GTT cause GPU hangs with RV6xx
191	 * See https://bugs.freedesktop.org/show_bug.cgi?id=91268
192	 */
193	if (rdev->family >= CHIP_RV610 && rdev->family <= CHIP_RV635)
194		bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
195
196#ifdef CONFIG_X86_32
197	/* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
198	 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
199	 */
200	bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
201#elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
202	/* Don't try to enable write-combining when it can't work, or things
203	 * may be slow
204	 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
205	 */
206#ifndef CONFIG_COMPILE_TEST
207#warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
208	 thanks to write-combining
209#endif
210
211	if (bo->flags & RADEON_GEM_GTT_WC)
212		DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
213			      "better performance thanks to write-combining\n");
214	bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
215#else
216	/* For architectures that don't support WC memory,
217	 * mask out the WC flag from the BO
218	 */
219	if (!drm_arch_can_wc_memory())
220		bo->flags &= ~RADEON_GEM_GTT_WC;
221#endif
222
223	radeon_ttm_placement_from_domain(bo, domain);
224	/* Kernel allocation are uninterruptible */
225	down_read(&rdev->pm.mclk_lock);
226	r = ttm_bo_init(&rdev->mman.bdev, &bo->tbo, size, type,
227			&bo->placement, page_align, !kernel, sg, resv,
228			&radeon_ttm_bo_destroy);
229	up_read(&rdev->pm.mclk_lock);
230	if (unlikely(r != 0)) {
231		return r;
232	}
233	*bo_ptr = bo;
234
235	trace_radeon_bo_create(bo);
236
237	return 0;
238}
239
240int radeon_bo_kmap(struct radeon_bo *bo, void **ptr)
241{
242	bool is_iomem;
243	int r;
244
245	if (bo->kptr) {
246		if (ptr) {
247			*ptr = bo->kptr;
248		}
249		return 0;
250	}
251	r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.resource->num_pages, &bo->kmap);
252	if (r) {
253		return r;
254	}
255	bo->kptr = ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
256	if (ptr) {
257		*ptr = bo->kptr;
258	}
259	radeon_bo_check_tiling(bo, 0, 0);
260	return 0;
261}
262
263void radeon_bo_kunmap(struct radeon_bo *bo)
264{
265	if (bo->kptr == NULL)
266		return;
267	bo->kptr = NULL;
268	radeon_bo_check_tiling(bo, 0, 0);
269	ttm_bo_kunmap(&bo->kmap);
270}
271
272struct radeon_bo *radeon_bo_ref(struct radeon_bo *bo)
273{
274	if (bo == NULL)
275		return NULL;
276
277	ttm_bo_get(&bo->tbo);
278	return bo;
279}
280
281void radeon_bo_unref(struct radeon_bo **bo)
282{
283	struct ttm_buffer_object *tbo;
 
284
285	if ((*bo) == NULL)
286		return;
 
287	tbo = &((*bo)->tbo);
288	ttm_bo_put(tbo);
289	*bo = NULL;
 
290}
291
292int radeon_bo_pin_restricted(struct radeon_bo *bo, u32 domain, u64 max_offset,
293			     u64 *gpu_addr)
294{
295	struct ttm_operation_ctx ctx = { false, false };
296	int r, i;
297
298	if (radeon_ttm_tt_has_userptr(bo->rdev, bo->tbo.ttm))
299		return -EPERM;
300
301	if (bo->tbo.pin_count) {
302		ttm_bo_pin(&bo->tbo);
303		if (gpu_addr)
304			*gpu_addr = radeon_bo_gpu_offset(bo);
305
306		if (max_offset != 0) {
307			u64 domain_start;
308
309			if (domain == RADEON_GEM_DOMAIN_VRAM)
310				domain_start = bo->rdev->mc.vram_start;
311			else
312				domain_start = bo->rdev->mc.gtt_start;
313			WARN_ON_ONCE(max_offset <
314				     (radeon_bo_gpu_offset(bo) - domain_start));
315		}
316
317		return 0;
318	}
319	if (bo->prime_shared_count && domain == RADEON_GEM_DOMAIN_VRAM) {
320		/* A BO shared as a dma-buf cannot be sensibly migrated to VRAM */
321		return -EINVAL;
322	}
323
324	radeon_ttm_placement_from_domain(bo, domain);
325	for (i = 0; i < bo->placement.num_placement; i++) {
326		/* force to pin into visible video ram */
327		if ((bo->placements[i].mem_type == TTM_PL_VRAM) &&
328		    !(bo->flags & RADEON_GEM_NO_CPU_ACCESS) &&
329		    (!max_offset || max_offset > bo->rdev->mc.visible_vram_size))
330			bo->placements[i].lpfn =
331				bo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
332		else
333			bo->placements[i].lpfn = max_offset >> PAGE_SHIFT;
 
 
334	}
335
336	r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
337	if (likely(r == 0)) {
338		ttm_bo_pin(&bo->tbo);
339		if (gpu_addr != NULL)
340			*gpu_addr = radeon_bo_gpu_offset(bo);
341		if (domain == RADEON_GEM_DOMAIN_VRAM)
342			bo->rdev->vram_pin_size += radeon_bo_size(bo);
343		else
344			bo->rdev->gart_pin_size += radeon_bo_size(bo);
345	} else {
346		dev_err(bo->rdev->dev, "%p pin failed\n", bo);
347	}
348	return r;
349}
350
351int radeon_bo_pin(struct radeon_bo *bo, u32 domain, u64 *gpu_addr)
352{
353	return radeon_bo_pin_restricted(bo, domain, 0, gpu_addr);
354}
355
356void radeon_bo_unpin(struct radeon_bo *bo)
357{
358	ttm_bo_unpin(&bo->tbo);
359	if (!bo->tbo.pin_count) {
360		if (bo->tbo.resource->mem_type == TTM_PL_VRAM)
 
 
 
 
 
 
 
 
 
 
 
 
 
361			bo->rdev->vram_pin_size -= radeon_bo_size(bo);
362		else
363			bo->rdev->gart_pin_size -= radeon_bo_size(bo);
 
 
364	}
 
365}
366
367int radeon_bo_evict_vram(struct radeon_device *rdev)
368{
369	struct ttm_device *bdev = &rdev->mman.bdev;
370	struct ttm_resource_manager *man;
371
372	/* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
373#ifndef CONFIG_HIBERNATION
374	if (rdev->flags & RADEON_IS_IGP) {
375		if (rdev->mc.igp_sideport_enabled == false)
376			/* Useless to evict on IGP chips */
377			return 0;
378	}
379#endif
380	man = ttm_manager_type(bdev, TTM_PL_VRAM);
381	if (!man)
382		return 0;
383	return ttm_resource_manager_evict_all(bdev, man);
384}
385
386void radeon_bo_force_delete(struct radeon_device *rdev)
387{
388	struct radeon_bo *bo, *n;
389
390	if (list_empty(&rdev->gem.objects)) {
391		return;
392	}
393	dev_err(rdev->dev, "Userspace still has active objects !\n");
394	list_for_each_entry_safe(bo, n, &rdev->gem.objects, list) {
395		dev_err(rdev->dev, "%p %p %lu %lu force free\n",
396			&bo->tbo.base, bo, (unsigned long)bo->tbo.base.size,
397			*((unsigned long *)&bo->tbo.base.refcount));
398		mutex_lock(&bo->rdev->gem.mutex);
399		list_del_init(&bo->list);
400		mutex_unlock(&bo->rdev->gem.mutex);
401		/* this should unref the ttm bo */
402		drm_gem_object_put(&bo->tbo.base);
403	}
404}
405
406int radeon_bo_init(struct radeon_device *rdev)
407{
408	/* reserve PAT memory space to WC for VRAM */
409	arch_io_reserve_memtype_wc(rdev->mc.aper_base,
410				   rdev->mc.aper_size);
411
412	/* Add an MTRR for the VRAM */
413	if (!rdev->fastfb_working) {
414		rdev->mc.vram_mtrr = arch_phys_wc_add(rdev->mc.aper_base,
415						      rdev->mc.aper_size);
416	}
417	DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
418		rdev->mc.mc_vram_size >> 20,
419		(unsigned long long)rdev->mc.aper_size >> 20);
420	DRM_INFO("RAM width %dbits %cDR\n",
421			rdev->mc.vram_width, rdev->mc.vram_is_ddr ? 'D' : 'S');
422	return radeon_ttm_init(rdev);
423}
424
425void radeon_bo_fini(struct radeon_device *rdev)
426{
427	radeon_ttm_fini(rdev);
428	arch_phys_wc_del(rdev->mc.vram_mtrr);
429	arch_io_free_memtype_wc(rdev->mc.aper_base, rdev->mc.aper_size);
430}
431
432/* Returns how many bytes TTM can move per IB.
433 */
434static u64 radeon_bo_get_threshold_for_moves(struct radeon_device *rdev)
435{
436	u64 real_vram_size = rdev->mc.real_vram_size;
437	u64 vram_usage = atomic64_read(&rdev->vram_usage);
438
439	/* This function is based on the current VRAM usage.
440	 *
441	 * - If all of VRAM is free, allow relocating the number of bytes that
442	 *   is equal to 1/4 of the size of VRAM for this IB.
443
444	 * - If more than one half of VRAM is occupied, only allow relocating
445	 *   1 MB of data for this IB.
446	 *
447	 * - From 0 to one half of used VRAM, the threshold decreases
448	 *   linearly.
449	 *         __________________
450	 * 1/4 of -|\               |
451	 * VRAM    | \              |
452	 *         |  \             |
453	 *         |   \            |
454	 *         |    \           |
455	 *         |     \          |
456	 *         |      \         |
457	 *         |       \________|1 MB
458	 *         |----------------|
459	 *    VRAM 0 %             100 %
460	 *         used            used
461	 *
462	 * Note: It's a threshold, not a limit. The threshold must be crossed
463	 * for buffer relocations to stop, so any buffer of an arbitrary size
464	 * can be moved as long as the threshold isn't crossed before
465	 * the relocation takes place. We don't want to disable buffer
466	 * relocations completely.
467	 *
468	 * The idea is that buffers should be placed in VRAM at creation time
469	 * and TTM should only do a minimum number of relocations during
470	 * command submission. In practice, you need to submit at least
471	 * a dozen IBs to move all buffers to VRAM if they are in GTT.
472	 *
473	 * Also, things can get pretty crazy under memory pressure and actual
474	 * VRAM usage can change a lot, so playing safe even at 50% does
475	 * consistently increase performance.
476	 */
477
478	u64 half_vram = real_vram_size >> 1;
479	u64 half_free_vram = vram_usage >= half_vram ? 0 : half_vram - vram_usage;
480	u64 bytes_moved_threshold = half_free_vram >> 1;
481	return max(bytes_moved_threshold, 1024*1024ull);
482}
483
484int radeon_bo_list_validate(struct radeon_device *rdev,
485			    struct ww_acquire_ctx *ticket,
486			    struct list_head *head, int ring)
487{
488	struct ttm_operation_ctx ctx = { true, false };
489	struct radeon_bo_list *lobj;
490	struct list_head duplicates;
491	int r;
492	u64 bytes_moved = 0, initial_bytes_moved;
493	u64 bytes_moved_threshold = radeon_bo_get_threshold_for_moves(rdev);
494
495	INIT_LIST_HEAD(&duplicates);
496	r = ttm_eu_reserve_buffers(ticket, head, true, &duplicates);
497	if (unlikely(r != 0)) {
498		return r;
499	}
500
501	list_for_each_entry(lobj, head, tv.head) {
502		struct radeon_bo *bo = lobj->robj;
503		if (!bo->tbo.pin_count) {
504			u32 domain = lobj->preferred_domains;
505			u32 allowed = lobj->allowed_domains;
506			u32 current_domain =
507				radeon_mem_type_to_domain(bo->tbo.resource->mem_type);
508
509			/* Check if this buffer will be moved and don't move it
510			 * if we have moved too many buffers for this IB already.
511			 *
512			 * Note that this allows moving at least one buffer of
513			 * any size, because it doesn't take the current "bo"
514			 * into account. We don't want to disallow buffer moves
515			 * completely.
516			 */
517			if ((allowed & current_domain) != 0 &&
518			    (domain & current_domain) == 0 && /* will be moved */
519			    bytes_moved > bytes_moved_threshold) {
520				/* don't move it */
521				domain = current_domain;
522			}
523
524		retry:
525			radeon_ttm_placement_from_domain(bo, domain);
526			if (ring == R600_RING_TYPE_UVD_INDEX)
527				radeon_uvd_force_into_uvd_segment(bo, allowed);
528
529			initial_bytes_moved = atomic64_read(&rdev->num_bytes_moved);
530			r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
531			bytes_moved += atomic64_read(&rdev->num_bytes_moved) -
532				       initial_bytes_moved;
533
534			if (unlikely(r)) {
535				if (r != -ERESTARTSYS &&
536				    domain != lobj->allowed_domains) {
537					domain = lobj->allowed_domains;
538					goto retry;
539				}
540				ttm_eu_backoff_reservation(ticket, head);
541				return r;
542			}
543		}
544		lobj->gpu_offset = radeon_bo_gpu_offset(bo);
545		lobj->tiling_flags = bo->tiling_flags;
546	}
547
548	list_for_each_entry(lobj, &duplicates, tv.head) {
549		lobj->gpu_offset = radeon_bo_gpu_offset(lobj->robj);
550		lobj->tiling_flags = lobj->robj->tiling_flags;
551	}
552
553	return 0;
554}
555
556int radeon_bo_get_surface_reg(struct radeon_bo *bo)
557{
558	struct radeon_device *rdev = bo->rdev;
559	struct radeon_surface_reg *reg;
560	struct radeon_bo *old_object;
561	int steal;
562	int i;
563
564	dma_resv_assert_held(bo->tbo.base.resv);
565
566	if (!bo->tiling_flags)
567		return 0;
568
569	if (bo->surface_reg >= 0) {
570		reg = &rdev->surface_regs[bo->surface_reg];
571		i = bo->surface_reg;
572		goto out;
573	}
574
575	steal = -1;
576	for (i = 0; i < RADEON_GEM_MAX_SURFACES; i++) {
577
578		reg = &rdev->surface_regs[i];
579		if (!reg->bo)
580			break;
581
582		old_object = reg->bo;
583		if (old_object->tbo.pin_count == 0)
584			steal = i;
585	}
586
587	/* if we are all out */
588	if (i == RADEON_GEM_MAX_SURFACES) {
589		if (steal == -1)
590			return -ENOMEM;
591		/* find someone with a surface reg and nuke their BO */
592		reg = &rdev->surface_regs[steal];
593		old_object = reg->bo;
594		/* blow away the mapping */
595		DRM_DEBUG("stealing surface reg %d from %p\n", steal, old_object);
596		ttm_bo_unmap_virtual(&old_object->tbo);
597		old_object->surface_reg = -1;
598		i = steal;
599	}
600
601	bo->surface_reg = i;
602	reg->bo = bo;
603
604out:
605	radeon_set_surface_reg(rdev, i, bo->tiling_flags, bo->pitch,
606			       bo->tbo.resource->start << PAGE_SHIFT,
607			       bo->tbo.base.size);
608	return 0;
609}
610
611static void radeon_bo_clear_surface_reg(struct radeon_bo *bo)
612{
613	struct radeon_device *rdev = bo->rdev;
614	struct radeon_surface_reg *reg;
615
616	if (bo->surface_reg == -1)
617		return;
618
619	reg = &rdev->surface_regs[bo->surface_reg];
620	radeon_clear_surface_reg(rdev, bo->surface_reg);
621
622	reg->bo = NULL;
623	bo->surface_reg = -1;
624}
625
626int radeon_bo_set_tiling_flags(struct radeon_bo *bo,
627				uint32_t tiling_flags, uint32_t pitch)
628{
629	struct radeon_device *rdev = bo->rdev;
630	int r;
631
632	if (rdev->family >= CHIP_CEDAR) {
633		unsigned bankw, bankh, mtaspect, tilesplit, stilesplit;
634
635		bankw = (tiling_flags >> RADEON_TILING_EG_BANKW_SHIFT) & RADEON_TILING_EG_BANKW_MASK;
636		bankh = (tiling_flags >> RADEON_TILING_EG_BANKH_SHIFT) & RADEON_TILING_EG_BANKH_MASK;
637		mtaspect = (tiling_flags >> RADEON_TILING_EG_MACRO_TILE_ASPECT_SHIFT) & RADEON_TILING_EG_MACRO_TILE_ASPECT_MASK;
638		tilesplit = (tiling_flags >> RADEON_TILING_EG_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_TILE_SPLIT_MASK;
639		stilesplit = (tiling_flags >> RADEON_TILING_EG_STENCIL_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_STENCIL_TILE_SPLIT_MASK;
640		switch (bankw) {
641		case 0:
642		case 1:
643		case 2:
644		case 4:
645		case 8:
646			break;
647		default:
648			return -EINVAL;
649		}
650		switch (bankh) {
651		case 0:
652		case 1:
653		case 2:
654		case 4:
655		case 8:
656			break;
657		default:
658			return -EINVAL;
659		}
660		switch (mtaspect) {
661		case 0:
662		case 1:
663		case 2:
664		case 4:
665		case 8:
666			break;
667		default:
668			return -EINVAL;
669		}
670		if (tilesplit > 6) {
671			return -EINVAL;
672		}
673		if (stilesplit > 6) {
674			return -EINVAL;
675		}
676	}
677	r = radeon_bo_reserve(bo, false);
678	if (unlikely(r != 0))
679		return r;
680	bo->tiling_flags = tiling_flags;
681	bo->pitch = pitch;
682	radeon_bo_unreserve(bo);
683	return 0;
684}
685
686void radeon_bo_get_tiling_flags(struct radeon_bo *bo,
687				uint32_t *tiling_flags,
688				uint32_t *pitch)
689{
690	dma_resv_assert_held(bo->tbo.base.resv);
691
692	if (tiling_flags)
693		*tiling_flags = bo->tiling_flags;
694	if (pitch)
695		*pitch = bo->pitch;
696}
697
698int radeon_bo_check_tiling(struct radeon_bo *bo, bool has_moved,
699				bool force_drop)
700{
701	if (!force_drop)
702		dma_resv_assert_held(bo->tbo.base.resv);
703
704	if (!(bo->tiling_flags & RADEON_TILING_SURFACE))
705		return 0;
706
707	if (force_drop) {
708		radeon_bo_clear_surface_reg(bo);
709		return 0;
710	}
711
712	if (bo->tbo.resource->mem_type != TTM_PL_VRAM) {
713		if (!has_moved)
714			return 0;
715
716		if (bo->surface_reg >= 0)
717			radeon_bo_clear_surface_reg(bo);
718		return 0;
719	}
720
721	if ((bo->surface_reg >= 0) && !has_moved)
722		return 0;
723
724	return radeon_bo_get_surface_reg(bo);
725}
726
727void radeon_bo_move_notify(struct ttm_buffer_object *bo,
728			   unsigned int old_type,
729			   struct ttm_resource *new_mem)
730{
731	struct radeon_bo *rbo;
732
733	radeon_update_memory_usage(bo, old_type, -1);
734	if (new_mem)
735		radeon_update_memory_usage(bo, new_mem->mem_type, 1);
736
737	if (!radeon_ttm_bo_is_radeon_bo(bo))
738		return;
739
740	rbo = container_of(bo, struct radeon_bo, tbo);
741	radeon_bo_check_tiling(rbo, 0, 1);
742	radeon_vm_bo_invalidate(rbo->rdev, rbo);
 
 
 
 
 
 
 
743}
744
745vm_fault_t radeon_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
746{
747	struct ttm_operation_ctx ctx = { false, false };
748	struct radeon_device *rdev;
749	struct radeon_bo *rbo;
750	unsigned long offset, size, lpfn;
751	int i, r;
752
753	if (!radeon_ttm_bo_is_radeon_bo(bo))
754		return 0;
755	rbo = container_of(bo, struct radeon_bo, tbo);
756	radeon_bo_check_tiling(rbo, 0, 0);
757	rdev = rbo->rdev;
758	if (bo->resource->mem_type != TTM_PL_VRAM)
759		return 0;
760
761	size = bo->resource->num_pages << PAGE_SHIFT;
762	offset = bo->resource->start << PAGE_SHIFT;
763	if ((offset + size) <= rdev->mc.visible_vram_size)
764		return 0;
765
766	/* Can't move a pinned BO to visible VRAM */
767	if (rbo->tbo.pin_count > 0)
768		return VM_FAULT_SIGBUS;
769
770	/* hurrah the memory is not visible ! */
771	radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_VRAM);
772	lpfn =	rdev->mc.visible_vram_size >> PAGE_SHIFT;
773	for (i = 0; i < rbo->placement.num_placement; i++) {
774		/* Force into visible VRAM */
775		if ((rbo->placements[i].mem_type == TTM_PL_VRAM) &&
776		    (!rbo->placements[i].lpfn || rbo->placements[i].lpfn > lpfn))
777			rbo->placements[i].lpfn = lpfn;
778	}
779	r = ttm_bo_validate(bo, &rbo->placement, &ctx);
780	if (unlikely(r == -ENOMEM)) {
781		radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_GTT);
782		r = ttm_bo_validate(bo, &rbo->placement, &ctx);
783	} else if (likely(!r)) {
784		offset = bo->resource->start << PAGE_SHIFT;
785		/* this should never happen */
786		if ((offset + size) > rdev->mc.visible_vram_size)
787			return VM_FAULT_SIGBUS;
788	}
789
790	if (unlikely(r == -EBUSY || r == -ERESTARTSYS))
791		return VM_FAULT_NOPAGE;
792	else if (unlikely(r))
793		return VM_FAULT_SIGBUS;
794
795	ttm_bo_move_to_lru_tail_unlocked(bo);
796	return 0;
797}
798
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
799/**
800 * radeon_bo_fence - add fence to buffer object
801 *
802 * @bo: buffer object in question
803 * @fence: fence to add
804 * @shared: true if fence should be added shared
805 *
806 */
807void radeon_bo_fence(struct radeon_bo *bo, struct radeon_fence *fence,
808		     bool shared)
809{
810	struct dma_resv *resv = bo->tbo.base.resv;
811
812	if (shared)
813		dma_resv_add_shared_fence(resv, &fence->base);
814	else
815		dma_resv_add_excl_fence(resv, &fence->base);
816}