Loading...
1/*
2 * Copyright 2016 Advanced Micro Devices, Inc.
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
13 *
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
21 *
22 * Authors: Christian König
23 */
24
25#include <linux/dma-mapping.h>
26#include <drm/ttm/ttm_range_manager.h>
27
28#include "amdgpu.h"
29#include "amdgpu_vm.h"
30#include "amdgpu_res_cursor.h"
31#include "amdgpu_atomfirmware.h"
32#include "atom.h"
33
34#define AMDGPU_MAX_SG_SEGMENT_SIZE (2UL << 30)
35
36struct amdgpu_vram_reservation {
37 u64 start;
38 u64 size;
39 struct list_head allocated;
40 struct list_head blocks;
41};
42
43static inline struct amdgpu_vram_mgr *
44to_vram_mgr(struct ttm_resource_manager *man)
45{
46 return container_of(man, struct amdgpu_vram_mgr, manager);
47}
48
49static inline struct amdgpu_device *
50to_amdgpu_device(struct amdgpu_vram_mgr *mgr)
51{
52 return container_of(mgr, struct amdgpu_device, mman.vram_mgr);
53}
54
55static inline struct drm_buddy_block *
56amdgpu_vram_mgr_first_block(struct list_head *list)
57{
58 return list_first_entry_or_null(list, struct drm_buddy_block, link);
59}
60
61static inline bool amdgpu_is_vram_mgr_blocks_contiguous(struct list_head *head)
62{
63 struct drm_buddy_block *block;
64 u64 start, size;
65
66 block = amdgpu_vram_mgr_first_block(head);
67 if (!block)
68 return false;
69
70 while (head != block->link.next) {
71 start = amdgpu_vram_mgr_block_start(block);
72 size = amdgpu_vram_mgr_block_size(block);
73
74 block = list_entry(block->link.next, struct drm_buddy_block, link);
75 if (start + size != amdgpu_vram_mgr_block_start(block))
76 return false;
77 }
78
79 return true;
80}
81
82static inline u64 amdgpu_vram_mgr_blocks_size(struct list_head *head)
83{
84 struct drm_buddy_block *block;
85 u64 size = 0;
86
87 list_for_each_entry(block, head, link)
88 size += amdgpu_vram_mgr_block_size(block);
89
90 return size;
91}
92
93/**
94 * DOC: mem_info_vram_total
95 *
96 * The amdgpu driver provides a sysfs API for reporting current total VRAM
97 * available on the device
98 * The file mem_info_vram_total is used for this and returns the total
99 * amount of VRAM in bytes
100 */
101static ssize_t amdgpu_mem_info_vram_total_show(struct device *dev,
102 struct device_attribute *attr, char *buf)
103{
104 struct drm_device *ddev = dev_get_drvdata(dev);
105 struct amdgpu_device *adev = drm_to_adev(ddev);
106
107 return sysfs_emit(buf, "%llu\n", adev->gmc.real_vram_size);
108}
109
110/**
111 * DOC: mem_info_vis_vram_total
112 *
113 * The amdgpu driver provides a sysfs API for reporting current total
114 * visible VRAM available on the device
115 * The file mem_info_vis_vram_total is used for this and returns the total
116 * amount of visible VRAM in bytes
117 */
118static ssize_t amdgpu_mem_info_vis_vram_total_show(struct device *dev,
119 struct device_attribute *attr, char *buf)
120{
121 struct drm_device *ddev = dev_get_drvdata(dev);
122 struct amdgpu_device *adev = drm_to_adev(ddev);
123
124 return sysfs_emit(buf, "%llu\n", adev->gmc.visible_vram_size);
125}
126
127/**
128 * DOC: mem_info_vram_used
129 *
130 * The amdgpu driver provides a sysfs API for reporting current total VRAM
131 * available on the device
132 * The file mem_info_vram_used is used for this and returns the total
133 * amount of currently used VRAM in bytes
134 */
135static ssize_t amdgpu_mem_info_vram_used_show(struct device *dev,
136 struct device_attribute *attr,
137 char *buf)
138{
139 struct drm_device *ddev = dev_get_drvdata(dev);
140 struct amdgpu_device *adev = drm_to_adev(ddev);
141 struct ttm_resource_manager *man = &adev->mman.vram_mgr.manager;
142
143 return sysfs_emit(buf, "%llu\n", ttm_resource_manager_usage(man));
144}
145
146/**
147 * DOC: mem_info_vis_vram_used
148 *
149 * The amdgpu driver provides a sysfs API for reporting current total of
150 * used visible VRAM
151 * The file mem_info_vis_vram_used is used for this and returns the total
152 * amount of currently used visible VRAM in bytes
153 */
154static ssize_t amdgpu_mem_info_vis_vram_used_show(struct device *dev,
155 struct device_attribute *attr,
156 char *buf)
157{
158 struct drm_device *ddev = dev_get_drvdata(dev);
159 struct amdgpu_device *adev = drm_to_adev(ddev);
160
161 return sysfs_emit(buf, "%llu\n",
162 amdgpu_vram_mgr_vis_usage(&adev->mman.vram_mgr));
163}
164
165/**
166 * DOC: mem_info_vram_vendor
167 *
168 * The amdgpu driver provides a sysfs API for reporting the vendor of the
169 * installed VRAM
170 * The file mem_info_vram_vendor is used for this and returns the name of the
171 * vendor.
172 */
173static ssize_t amdgpu_mem_info_vram_vendor(struct device *dev,
174 struct device_attribute *attr,
175 char *buf)
176{
177 struct drm_device *ddev = dev_get_drvdata(dev);
178 struct amdgpu_device *adev = drm_to_adev(ddev);
179
180 switch (adev->gmc.vram_vendor) {
181 case SAMSUNG:
182 return sysfs_emit(buf, "samsung\n");
183 case INFINEON:
184 return sysfs_emit(buf, "infineon\n");
185 case ELPIDA:
186 return sysfs_emit(buf, "elpida\n");
187 case ETRON:
188 return sysfs_emit(buf, "etron\n");
189 case NANYA:
190 return sysfs_emit(buf, "nanya\n");
191 case HYNIX:
192 return sysfs_emit(buf, "hynix\n");
193 case MOSEL:
194 return sysfs_emit(buf, "mosel\n");
195 case WINBOND:
196 return sysfs_emit(buf, "winbond\n");
197 case ESMT:
198 return sysfs_emit(buf, "esmt\n");
199 case MICRON:
200 return sysfs_emit(buf, "micron\n");
201 default:
202 return sysfs_emit(buf, "unknown\n");
203 }
204}
205
206static DEVICE_ATTR(mem_info_vram_total, S_IRUGO,
207 amdgpu_mem_info_vram_total_show, NULL);
208static DEVICE_ATTR(mem_info_vis_vram_total, S_IRUGO,
209 amdgpu_mem_info_vis_vram_total_show,NULL);
210static DEVICE_ATTR(mem_info_vram_used, S_IRUGO,
211 amdgpu_mem_info_vram_used_show, NULL);
212static DEVICE_ATTR(mem_info_vis_vram_used, S_IRUGO,
213 amdgpu_mem_info_vis_vram_used_show, NULL);
214static DEVICE_ATTR(mem_info_vram_vendor, S_IRUGO,
215 amdgpu_mem_info_vram_vendor, NULL);
216
217static struct attribute *amdgpu_vram_mgr_attributes[] = {
218 &dev_attr_mem_info_vram_total.attr,
219 &dev_attr_mem_info_vis_vram_total.attr,
220 &dev_attr_mem_info_vram_used.attr,
221 &dev_attr_mem_info_vis_vram_used.attr,
222 &dev_attr_mem_info_vram_vendor.attr,
223 NULL
224};
225
226static umode_t amdgpu_vram_attrs_is_visible(struct kobject *kobj,
227 struct attribute *attr, int i)
228{
229 struct device *dev = kobj_to_dev(kobj);
230 struct drm_device *ddev = dev_get_drvdata(dev);
231 struct amdgpu_device *adev = drm_to_adev(ddev);
232
233 if (attr == &dev_attr_mem_info_vram_vendor.attr &&
234 !adev->gmc.vram_vendor)
235 return 0;
236
237 return attr->mode;
238}
239
240const struct attribute_group amdgpu_vram_mgr_attr_group = {
241 .attrs = amdgpu_vram_mgr_attributes,
242 .is_visible = amdgpu_vram_attrs_is_visible
243};
244
245/**
246 * amdgpu_vram_mgr_vis_size - Calculate visible block size
247 *
248 * @adev: amdgpu_device pointer
249 * @block: DRM BUDDY block structure
250 *
251 * Calculate how many bytes of the DRM BUDDY block are inside visible VRAM
252 */
253static u64 amdgpu_vram_mgr_vis_size(struct amdgpu_device *adev,
254 struct drm_buddy_block *block)
255{
256 u64 start = amdgpu_vram_mgr_block_start(block);
257 u64 end = start + amdgpu_vram_mgr_block_size(block);
258
259 if (start >= adev->gmc.visible_vram_size)
260 return 0;
261
262 return (end > adev->gmc.visible_vram_size ?
263 adev->gmc.visible_vram_size : end) - start;
264}
265
266/**
267 * amdgpu_vram_mgr_bo_visible_size - CPU visible BO size
268 *
269 * @bo: &amdgpu_bo buffer object (must be in VRAM)
270 *
271 * Returns:
272 * How much of the given &amdgpu_bo buffer object lies in CPU visible VRAM.
273 */
274u64 amdgpu_vram_mgr_bo_visible_size(struct amdgpu_bo *bo)
275{
276 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
277 struct ttm_resource *res = bo->tbo.resource;
278 struct amdgpu_vram_mgr_resource *vres = to_amdgpu_vram_mgr_resource(res);
279 struct drm_buddy_block *block;
280 u64 usage = 0;
281
282 if (amdgpu_gmc_vram_full_visible(&adev->gmc))
283 return amdgpu_bo_size(bo);
284
285 if (res->start >= adev->gmc.visible_vram_size >> PAGE_SHIFT)
286 return 0;
287
288 list_for_each_entry(block, &vres->blocks, link)
289 usage += amdgpu_vram_mgr_vis_size(adev, block);
290
291 return usage;
292}
293
294/* Commit the reservation of VRAM pages */
295static void amdgpu_vram_mgr_do_reserve(struct ttm_resource_manager *man)
296{
297 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
298 struct amdgpu_device *adev = to_amdgpu_device(mgr);
299 struct drm_buddy *mm = &mgr->mm;
300 struct amdgpu_vram_reservation *rsv, *temp;
301 struct drm_buddy_block *block;
302 uint64_t vis_usage;
303
304 list_for_each_entry_safe(rsv, temp, &mgr->reservations_pending, blocks) {
305 if (drm_buddy_alloc_blocks(mm, rsv->start, rsv->start + rsv->size,
306 rsv->size, mm->chunk_size, &rsv->allocated,
307 DRM_BUDDY_RANGE_ALLOCATION))
308 continue;
309
310 block = amdgpu_vram_mgr_first_block(&rsv->allocated);
311 if (!block)
312 continue;
313
314 dev_dbg(adev->dev, "Reservation 0x%llx - %lld, Succeeded\n",
315 rsv->start, rsv->size);
316
317 vis_usage = amdgpu_vram_mgr_vis_size(adev, block);
318 atomic64_add(vis_usage, &mgr->vis_usage);
319 spin_lock(&man->bdev->lru_lock);
320 man->usage += rsv->size;
321 spin_unlock(&man->bdev->lru_lock);
322 list_move(&rsv->blocks, &mgr->reserved_pages);
323 }
324}
325
326/**
327 * amdgpu_vram_mgr_reserve_range - Reserve a range from VRAM
328 *
329 * @mgr: amdgpu_vram_mgr pointer
330 * @start: start address of the range in VRAM
331 * @size: size of the range
332 *
333 * Reserve memory from start address with the specified size in VRAM
334 */
335int amdgpu_vram_mgr_reserve_range(struct amdgpu_vram_mgr *mgr,
336 uint64_t start, uint64_t size)
337{
338 struct amdgpu_vram_reservation *rsv;
339
340 rsv = kzalloc(sizeof(*rsv), GFP_KERNEL);
341 if (!rsv)
342 return -ENOMEM;
343
344 INIT_LIST_HEAD(&rsv->allocated);
345 INIT_LIST_HEAD(&rsv->blocks);
346
347 rsv->start = start;
348 rsv->size = size;
349
350 mutex_lock(&mgr->lock);
351 list_add_tail(&rsv->blocks, &mgr->reservations_pending);
352 amdgpu_vram_mgr_do_reserve(&mgr->manager);
353 mutex_unlock(&mgr->lock);
354
355 return 0;
356}
357
358/**
359 * amdgpu_vram_mgr_query_page_status - query the reservation status
360 *
361 * @mgr: amdgpu_vram_mgr pointer
362 * @start: start address of a page in VRAM
363 *
364 * Returns:
365 * -EBUSY: the page is still hold and in pending list
366 * 0: the page has been reserved
367 * -ENOENT: the input page is not a reservation
368 */
369int amdgpu_vram_mgr_query_page_status(struct amdgpu_vram_mgr *mgr,
370 uint64_t start)
371{
372 struct amdgpu_vram_reservation *rsv;
373 int ret;
374
375 mutex_lock(&mgr->lock);
376
377 list_for_each_entry(rsv, &mgr->reservations_pending, blocks) {
378 if (rsv->start <= start &&
379 (start < (rsv->start + rsv->size))) {
380 ret = -EBUSY;
381 goto out;
382 }
383 }
384
385 list_for_each_entry(rsv, &mgr->reserved_pages, blocks) {
386 if (rsv->start <= start &&
387 (start < (rsv->start + rsv->size))) {
388 ret = 0;
389 goto out;
390 }
391 }
392
393 ret = -ENOENT;
394out:
395 mutex_unlock(&mgr->lock);
396 return ret;
397}
398
399static void amdgpu_dummy_vram_mgr_debug(struct ttm_resource_manager *man,
400 struct drm_printer *printer)
401{
402 DRM_DEBUG_DRIVER("Dummy vram mgr debug\n");
403}
404
405static bool amdgpu_dummy_vram_mgr_compatible(struct ttm_resource_manager *man,
406 struct ttm_resource *res,
407 const struct ttm_place *place,
408 size_t size)
409{
410 DRM_DEBUG_DRIVER("Dummy vram mgr compatible\n");
411 return false;
412}
413
414static bool amdgpu_dummy_vram_mgr_intersects(struct ttm_resource_manager *man,
415 struct ttm_resource *res,
416 const struct ttm_place *place,
417 size_t size)
418{
419 DRM_DEBUG_DRIVER("Dummy vram mgr intersects\n");
420 return true;
421}
422
423static void amdgpu_dummy_vram_mgr_del(struct ttm_resource_manager *man,
424 struct ttm_resource *res)
425{
426 DRM_DEBUG_DRIVER("Dummy vram mgr deleted\n");
427}
428
429static int amdgpu_dummy_vram_mgr_new(struct ttm_resource_manager *man,
430 struct ttm_buffer_object *tbo,
431 const struct ttm_place *place,
432 struct ttm_resource **res)
433{
434 DRM_DEBUG_DRIVER("Dummy vram mgr new\n");
435 return -ENOSPC;
436}
437
438/**
439 * amdgpu_vram_mgr_new - allocate new ranges
440 *
441 * @man: TTM memory type manager
442 * @tbo: TTM BO we need this range for
443 * @place: placement flags and restrictions
444 * @res: the resulting mem object
445 *
446 * Allocate VRAM for the given BO.
447 */
448static int amdgpu_vram_mgr_new(struct ttm_resource_manager *man,
449 struct ttm_buffer_object *tbo,
450 const struct ttm_place *place,
451 struct ttm_resource **res)
452{
453 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
454 struct amdgpu_device *adev = to_amdgpu_device(mgr);
455 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo);
456 u64 vis_usage = 0, max_bytes, min_block_size;
457 struct amdgpu_vram_mgr_resource *vres;
458 u64 size, remaining_size, lpfn, fpfn;
459 unsigned int adjust_dcc_size = 0;
460 struct drm_buddy *mm = &mgr->mm;
461 struct drm_buddy_block *block;
462 unsigned long pages_per_block;
463 int r;
464
465 lpfn = (u64)place->lpfn << PAGE_SHIFT;
466 if (!lpfn)
467 lpfn = man->size;
468
469 fpfn = (u64)place->fpfn << PAGE_SHIFT;
470
471 max_bytes = adev->gmc.mc_vram_size;
472 if (tbo->type != ttm_bo_type_kernel)
473 max_bytes -= AMDGPU_VM_RESERVED_VRAM;
474
475 if (bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS) {
476 pages_per_block = ~0ul;
477 } else {
478#ifdef CONFIG_TRANSPARENT_HUGEPAGE
479 pages_per_block = HPAGE_PMD_NR;
480#else
481 /* default to 2MB */
482 pages_per_block = 2UL << (20UL - PAGE_SHIFT);
483#endif
484 pages_per_block = max_t(u32, pages_per_block,
485 tbo->page_alignment);
486 }
487
488 vres = kzalloc(sizeof(*vres), GFP_KERNEL);
489 if (!vres)
490 return -ENOMEM;
491
492 ttm_resource_init(tbo, place, &vres->base);
493
494 /* bail out quickly if there's likely not enough VRAM for this BO */
495 if (ttm_resource_manager_usage(man) > max_bytes) {
496 r = -ENOSPC;
497 goto error_fini;
498 }
499
500 INIT_LIST_HEAD(&vres->blocks);
501
502 if (place->flags & TTM_PL_FLAG_TOPDOWN)
503 vres->flags |= DRM_BUDDY_TOPDOWN_ALLOCATION;
504
505 if (bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)
506 vres->flags |= DRM_BUDDY_CONTIGUOUS_ALLOCATION;
507
508 if (bo->flags & AMDGPU_GEM_CREATE_VRAM_CLEARED)
509 vres->flags |= DRM_BUDDY_CLEAR_ALLOCATION;
510
511 if (fpfn || lpfn != mgr->mm.size)
512 /* Allocate blocks in desired range */
513 vres->flags |= DRM_BUDDY_RANGE_ALLOCATION;
514
515 if (bo->flags & AMDGPU_GEM_CREATE_GFX12_DCC &&
516 adev->gmc.gmc_funcs->get_dcc_alignment)
517 adjust_dcc_size = amdgpu_gmc_get_dcc_alignment(adev);
518
519 remaining_size = (u64)vres->base.size;
520 if (bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS && adjust_dcc_size) {
521 unsigned int dcc_size;
522
523 dcc_size = roundup_pow_of_two(vres->base.size + adjust_dcc_size);
524 remaining_size = (u64)dcc_size;
525
526 vres->flags |= DRM_BUDDY_TRIM_DISABLE;
527 }
528
529 mutex_lock(&mgr->lock);
530 while (remaining_size) {
531 if (tbo->page_alignment)
532 min_block_size = (u64)tbo->page_alignment << PAGE_SHIFT;
533 else
534 min_block_size = mgr->default_page_size;
535
536 size = remaining_size;
537
538 if (bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS && adjust_dcc_size)
539 min_block_size = size;
540 else if ((size >= (u64)pages_per_block << PAGE_SHIFT) &&
541 !(size & (((u64)pages_per_block << PAGE_SHIFT) - 1)))
542 min_block_size = (u64)pages_per_block << PAGE_SHIFT;
543
544 BUG_ON(min_block_size < mm->chunk_size);
545
546 r = drm_buddy_alloc_blocks(mm, fpfn,
547 lpfn,
548 size,
549 min_block_size,
550 &vres->blocks,
551 vres->flags);
552
553 if (unlikely(r == -ENOSPC) && pages_per_block == ~0ul &&
554 !(place->flags & TTM_PL_FLAG_CONTIGUOUS)) {
555 vres->flags &= ~DRM_BUDDY_CONTIGUOUS_ALLOCATION;
556 pages_per_block = max_t(u32, 2UL << (20UL - PAGE_SHIFT),
557 tbo->page_alignment);
558
559 continue;
560 }
561
562 if (unlikely(r))
563 goto error_free_blocks;
564
565 if (size > remaining_size)
566 remaining_size = 0;
567 else
568 remaining_size -= size;
569 }
570
571 if (bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS && adjust_dcc_size) {
572 struct drm_buddy_block *dcc_block;
573 unsigned long dcc_start;
574 u64 trim_start;
575
576 dcc_block = amdgpu_vram_mgr_first_block(&vres->blocks);
577 /* Adjust the start address for DCC buffers only */
578 dcc_start =
579 roundup((unsigned long)amdgpu_vram_mgr_block_start(dcc_block),
580 adjust_dcc_size);
581 trim_start = (u64)dcc_start;
582 drm_buddy_block_trim(mm, &trim_start,
583 (u64)vres->base.size,
584 &vres->blocks);
585 }
586 mutex_unlock(&mgr->lock);
587
588 vres->base.start = 0;
589 size = max_t(u64, amdgpu_vram_mgr_blocks_size(&vres->blocks),
590 vres->base.size);
591 list_for_each_entry(block, &vres->blocks, link) {
592 unsigned long start;
593
594 start = amdgpu_vram_mgr_block_start(block) +
595 amdgpu_vram_mgr_block_size(block);
596 start >>= PAGE_SHIFT;
597
598 if (start > PFN_UP(size))
599 start -= PFN_UP(size);
600 else
601 start = 0;
602 vres->base.start = max(vres->base.start, start);
603
604 vis_usage += amdgpu_vram_mgr_vis_size(adev, block);
605 }
606
607 if (amdgpu_is_vram_mgr_blocks_contiguous(&vres->blocks))
608 vres->base.placement |= TTM_PL_FLAG_CONTIGUOUS;
609
610 if (adev->gmc.xgmi.connected_to_cpu)
611 vres->base.bus.caching = ttm_cached;
612 else
613 vres->base.bus.caching = ttm_write_combined;
614
615 atomic64_add(vis_usage, &mgr->vis_usage);
616 *res = &vres->base;
617 return 0;
618
619error_free_blocks:
620 drm_buddy_free_list(mm, &vres->blocks, 0);
621 mutex_unlock(&mgr->lock);
622error_fini:
623 ttm_resource_fini(man, &vres->base);
624 kfree(vres);
625
626 return r;
627}
628
629/**
630 * amdgpu_vram_mgr_del - free ranges
631 *
632 * @man: TTM memory type manager
633 * @res: TTM memory object
634 *
635 * Free the allocated VRAM again.
636 */
637static void amdgpu_vram_mgr_del(struct ttm_resource_manager *man,
638 struct ttm_resource *res)
639{
640 struct amdgpu_vram_mgr_resource *vres = to_amdgpu_vram_mgr_resource(res);
641 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
642 struct amdgpu_device *adev = to_amdgpu_device(mgr);
643 struct drm_buddy *mm = &mgr->mm;
644 struct drm_buddy_block *block;
645 uint64_t vis_usage = 0;
646
647 mutex_lock(&mgr->lock);
648 list_for_each_entry(block, &vres->blocks, link)
649 vis_usage += amdgpu_vram_mgr_vis_size(adev, block);
650
651 amdgpu_vram_mgr_do_reserve(man);
652
653 drm_buddy_free_list(mm, &vres->blocks, vres->flags);
654 mutex_unlock(&mgr->lock);
655
656 atomic64_sub(vis_usage, &mgr->vis_usage);
657
658 ttm_resource_fini(man, res);
659 kfree(vres);
660}
661
662/**
663 * amdgpu_vram_mgr_alloc_sgt - allocate and fill a sg table
664 *
665 * @adev: amdgpu device pointer
666 * @res: TTM memory object
667 * @offset: byte offset from the base of VRAM BO
668 * @length: number of bytes to export in sg_table
669 * @dev: the other device
670 * @dir: dma direction
671 * @sgt: resulting sg table
672 *
673 * Allocate and fill a sg table from a VRAM allocation.
674 */
675int amdgpu_vram_mgr_alloc_sgt(struct amdgpu_device *adev,
676 struct ttm_resource *res,
677 u64 offset, u64 length,
678 struct device *dev,
679 enum dma_data_direction dir,
680 struct sg_table **sgt)
681{
682 struct amdgpu_res_cursor cursor;
683 struct scatterlist *sg;
684 int num_entries = 0;
685 int i, r;
686
687 *sgt = kmalloc(sizeof(**sgt), GFP_KERNEL);
688 if (!*sgt)
689 return -ENOMEM;
690
691 /* Determine the number of DRM_BUDDY blocks to export */
692 amdgpu_res_first(res, offset, length, &cursor);
693 while (cursor.remaining) {
694 num_entries++;
695 amdgpu_res_next(&cursor, min(cursor.size, AMDGPU_MAX_SG_SEGMENT_SIZE));
696 }
697
698 r = sg_alloc_table(*sgt, num_entries, GFP_KERNEL);
699 if (r)
700 goto error_free;
701
702 /* Initialize scatterlist nodes of sg_table */
703 for_each_sgtable_sg((*sgt), sg, i)
704 sg->length = 0;
705
706 /*
707 * Walk down DRM_BUDDY blocks to populate scatterlist nodes
708 * @note: Use iterator api to get first the DRM_BUDDY block
709 * and the number of bytes from it. Access the following
710 * DRM_BUDDY block(s) if more buffer needs to exported
711 */
712 amdgpu_res_first(res, offset, length, &cursor);
713 for_each_sgtable_sg((*sgt), sg, i) {
714 phys_addr_t phys = cursor.start + adev->gmc.aper_base;
715 unsigned long size = min(cursor.size, AMDGPU_MAX_SG_SEGMENT_SIZE);
716 dma_addr_t addr;
717
718 addr = dma_map_resource(dev, phys, size, dir,
719 DMA_ATTR_SKIP_CPU_SYNC);
720 r = dma_mapping_error(dev, addr);
721 if (r)
722 goto error_unmap;
723
724 sg_set_page(sg, NULL, size, 0);
725 sg_dma_address(sg) = addr;
726 sg_dma_len(sg) = size;
727
728 amdgpu_res_next(&cursor, size);
729 }
730
731 return 0;
732
733error_unmap:
734 for_each_sgtable_sg((*sgt), sg, i) {
735 if (!sg->length)
736 continue;
737
738 dma_unmap_resource(dev, sg->dma_address,
739 sg->length, dir,
740 DMA_ATTR_SKIP_CPU_SYNC);
741 }
742 sg_free_table(*sgt);
743
744error_free:
745 kfree(*sgt);
746 return r;
747}
748
749/**
750 * amdgpu_vram_mgr_free_sgt - allocate and fill a sg table
751 *
752 * @dev: device pointer
753 * @dir: data direction of resource to unmap
754 * @sgt: sg table to free
755 *
756 * Free a previously allocate sg table.
757 */
758void amdgpu_vram_mgr_free_sgt(struct device *dev,
759 enum dma_data_direction dir,
760 struct sg_table *sgt)
761{
762 struct scatterlist *sg;
763 int i;
764
765 for_each_sgtable_sg(sgt, sg, i)
766 dma_unmap_resource(dev, sg->dma_address,
767 sg->length, dir,
768 DMA_ATTR_SKIP_CPU_SYNC);
769 sg_free_table(sgt);
770 kfree(sgt);
771}
772
773/**
774 * amdgpu_vram_mgr_vis_usage - how many bytes are used in the visible part
775 *
776 * @mgr: amdgpu_vram_mgr pointer
777 *
778 * Returns how many bytes are used in the visible part of VRAM
779 */
780uint64_t amdgpu_vram_mgr_vis_usage(struct amdgpu_vram_mgr *mgr)
781{
782 return atomic64_read(&mgr->vis_usage);
783}
784
785/**
786 * amdgpu_vram_mgr_intersects - test each drm buddy block for intersection
787 *
788 * @man: TTM memory type manager
789 * @res: The resource to test
790 * @place: The place to test against
791 * @size: Size of the new allocation
792 *
793 * Test each drm buddy block for intersection for eviction decision.
794 */
795static bool amdgpu_vram_mgr_intersects(struct ttm_resource_manager *man,
796 struct ttm_resource *res,
797 const struct ttm_place *place,
798 size_t size)
799{
800 struct amdgpu_vram_mgr_resource *mgr = to_amdgpu_vram_mgr_resource(res);
801 struct drm_buddy_block *block;
802
803 /* Check each drm buddy block individually */
804 list_for_each_entry(block, &mgr->blocks, link) {
805 unsigned long fpfn =
806 amdgpu_vram_mgr_block_start(block) >> PAGE_SHIFT;
807 unsigned long lpfn = fpfn +
808 (amdgpu_vram_mgr_block_size(block) >> PAGE_SHIFT);
809
810 if (place->fpfn < lpfn &&
811 (!place->lpfn || place->lpfn > fpfn))
812 return true;
813 }
814
815 return false;
816}
817
818/**
819 * amdgpu_vram_mgr_compatible - test each drm buddy block for compatibility
820 *
821 * @man: TTM memory type manager
822 * @res: The resource to test
823 * @place: The place to test against
824 * @size: Size of the new allocation
825 *
826 * Test each drm buddy block for placement compatibility.
827 */
828static bool amdgpu_vram_mgr_compatible(struct ttm_resource_manager *man,
829 struct ttm_resource *res,
830 const struct ttm_place *place,
831 size_t size)
832{
833 struct amdgpu_vram_mgr_resource *mgr = to_amdgpu_vram_mgr_resource(res);
834 struct drm_buddy_block *block;
835
836 /* Check each drm buddy block individually */
837 list_for_each_entry(block, &mgr->blocks, link) {
838 unsigned long fpfn =
839 amdgpu_vram_mgr_block_start(block) >> PAGE_SHIFT;
840 unsigned long lpfn = fpfn +
841 (amdgpu_vram_mgr_block_size(block) >> PAGE_SHIFT);
842
843 if (fpfn < place->fpfn ||
844 (place->lpfn && lpfn > place->lpfn))
845 return false;
846 }
847
848 return true;
849}
850
851/**
852 * amdgpu_vram_mgr_debug - dump VRAM table
853 *
854 * @man: TTM memory type manager
855 * @printer: DRM printer to use
856 *
857 * Dump the table content using printk.
858 */
859static void amdgpu_vram_mgr_debug(struct ttm_resource_manager *man,
860 struct drm_printer *printer)
861{
862 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
863 struct drm_buddy *mm = &mgr->mm;
864 struct amdgpu_vram_reservation *rsv;
865
866 drm_printf(printer, " vis usage:%llu\n",
867 amdgpu_vram_mgr_vis_usage(mgr));
868
869 mutex_lock(&mgr->lock);
870 drm_printf(printer, "default_page_size: %lluKiB\n",
871 mgr->default_page_size >> 10);
872
873 drm_buddy_print(mm, printer);
874
875 drm_printf(printer, "reserved:\n");
876 list_for_each_entry(rsv, &mgr->reserved_pages, blocks)
877 drm_printf(printer, "%#018llx-%#018llx: %llu\n",
878 rsv->start, rsv->start + rsv->size, rsv->size);
879 mutex_unlock(&mgr->lock);
880}
881
882static const struct ttm_resource_manager_func amdgpu_dummy_vram_mgr_func = {
883 .alloc = amdgpu_dummy_vram_mgr_new,
884 .free = amdgpu_dummy_vram_mgr_del,
885 .intersects = amdgpu_dummy_vram_mgr_intersects,
886 .compatible = amdgpu_dummy_vram_mgr_compatible,
887 .debug = amdgpu_dummy_vram_mgr_debug
888};
889
890static const struct ttm_resource_manager_func amdgpu_vram_mgr_func = {
891 .alloc = amdgpu_vram_mgr_new,
892 .free = amdgpu_vram_mgr_del,
893 .intersects = amdgpu_vram_mgr_intersects,
894 .compatible = amdgpu_vram_mgr_compatible,
895 .debug = amdgpu_vram_mgr_debug
896};
897
898/**
899 * amdgpu_vram_mgr_init - init VRAM manager and DRM MM
900 *
901 * @adev: amdgpu_device pointer
902 *
903 * Allocate and initialize the VRAM manager.
904 */
905int amdgpu_vram_mgr_init(struct amdgpu_device *adev)
906{
907 struct amdgpu_vram_mgr *mgr = &adev->mman.vram_mgr;
908 struct ttm_resource_manager *man = &mgr->manager;
909 int err;
910
911 ttm_resource_manager_init(man, &adev->mman.bdev,
912 adev->gmc.real_vram_size);
913
914 mutex_init(&mgr->lock);
915 INIT_LIST_HEAD(&mgr->reservations_pending);
916 INIT_LIST_HEAD(&mgr->reserved_pages);
917 mgr->default_page_size = PAGE_SIZE;
918
919 if (!adev->gmc.is_app_apu) {
920 man->func = &amdgpu_vram_mgr_func;
921
922 err = drm_buddy_init(&mgr->mm, man->size, PAGE_SIZE);
923 if (err)
924 return err;
925 } else {
926 man->func = &amdgpu_dummy_vram_mgr_func;
927 DRM_INFO("Setup dummy vram mgr\n");
928 }
929
930 ttm_set_driver_manager(&adev->mman.bdev, TTM_PL_VRAM, &mgr->manager);
931 ttm_resource_manager_set_used(man, true);
932 return 0;
933}
934
935/**
936 * amdgpu_vram_mgr_fini - free and destroy VRAM manager
937 *
938 * @adev: amdgpu_device pointer
939 *
940 * Destroy and free the VRAM manager, returns -EBUSY if ranges are still
941 * allocated inside it.
942 */
943void amdgpu_vram_mgr_fini(struct amdgpu_device *adev)
944{
945 struct amdgpu_vram_mgr *mgr = &adev->mman.vram_mgr;
946 struct ttm_resource_manager *man = &mgr->manager;
947 int ret;
948 struct amdgpu_vram_reservation *rsv, *temp;
949
950 ttm_resource_manager_set_used(man, false);
951
952 ret = ttm_resource_manager_evict_all(&adev->mman.bdev, man);
953 if (ret)
954 return;
955
956 mutex_lock(&mgr->lock);
957 list_for_each_entry_safe(rsv, temp, &mgr->reservations_pending, blocks)
958 kfree(rsv);
959
960 list_for_each_entry_safe(rsv, temp, &mgr->reserved_pages, blocks) {
961 drm_buddy_free_list(&mgr->mm, &rsv->allocated, 0);
962 kfree(rsv);
963 }
964 if (!adev->gmc.is_app_apu)
965 drm_buddy_fini(&mgr->mm);
966 mutex_unlock(&mgr->lock);
967
968 ttm_resource_manager_cleanup(man);
969 ttm_set_driver_manager(&adev->mman.bdev, TTM_PL_VRAM, NULL);
970}
1/*
2 * Copyright 2016 Advanced Micro Devices, Inc.
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
13 *
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
21 *
22 * Authors: Christian König
23 */
24
25#include <linux/dma-mapping.h>
26#include <drm/ttm/ttm_range_manager.h>
27
28#include "amdgpu.h"
29#include "amdgpu_vm.h"
30#include "amdgpu_res_cursor.h"
31#include "amdgpu_atomfirmware.h"
32#include "atom.h"
33
34struct amdgpu_vram_reservation {
35 u64 start;
36 u64 size;
37 struct list_head allocated;
38 struct list_head blocks;
39};
40
41static inline struct amdgpu_vram_mgr *
42to_vram_mgr(struct ttm_resource_manager *man)
43{
44 return container_of(man, struct amdgpu_vram_mgr, manager);
45}
46
47static inline struct amdgpu_device *
48to_amdgpu_device(struct amdgpu_vram_mgr *mgr)
49{
50 return container_of(mgr, struct amdgpu_device, mman.vram_mgr);
51}
52
53static inline struct drm_buddy_block *
54amdgpu_vram_mgr_first_block(struct list_head *list)
55{
56 return list_first_entry_or_null(list, struct drm_buddy_block, link);
57}
58
59static inline bool amdgpu_is_vram_mgr_blocks_contiguous(struct list_head *head)
60{
61 struct drm_buddy_block *block;
62 u64 start, size;
63
64 block = amdgpu_vram_mgr_first_block(head);
65 if (!block)
66 return false;
67
68 while (head != block->link.next) {
69 start = amdgpu_vram_mgr_block_start(block);
70 size = amdgpu_vram_mgr_block_size(block);
71
72 block = list_entry(block->link.next, struct drm_buddy_block, link);
73 if (start + size != amdgpu_vram_mgr_block_start(block))
74 return false;
75 }
76
77 return true;
78}
79
80
81
82/**
83 * DOC: mem_info_vram_total
84 *
85 * The amdgpu driver provides a sysfs API for reporting current total VRAM
86 * available on the device
87 * The file mem_info_vram_total is used for this and returns the total
88 * amount of VRAM in bytes
89 */
90static ssize_t amdgpu_mem_info_vram_total_show(struct device *dev,
91 struct device_attribute *attr, char *buf)
92{
93 struct drm_device *ddev = dev_get_drvdata(dev);
94 struct amdgpu_device *adev = drm_to_adev(ddev);
95
96 return sysfs_emit(buf, "%llu\n", adev->gmc.real_vram_size);
97}
98
99/**
100 * DOC: mem_info_vis_vram_total
101 *
102 * The amdgpu driver provides a sysfs API for reporting current total
103 * visible VRAM available on the device
104 * The file mem_info_vis_vram_total is used for this and returns the total
105 * amount of visible VRAM in bytes
106 */
107static ssize_t amdgpu_mem_info_vis_vram_total_show(struct device *dev,
108 struct device_attribute *attr, char *buf)
109{
110 struct drm_device *ddev = dev_get_drvdata(dev);
111 struct amdgpu_device *adev = drm_to_adev(ddev);
112
113 return sysfs_emit(buf, "%llu\n", adev->gmc.visible_vram_size);
114}
115
116/**
117 * DOC: mem_info_vram_used
118 *
119 * The amdgpu driver provides a sysfs API for reporting current total VRAM
120 * available on the device
121 * The file mem_info_vram_used is used for this and returns the total
122 * amount of currently used VRAM in bytes
123 */
124static ssize_t amdgpu_mem_info_vram_used_show(struct device *dev,
125 struct device_attribute *attr,
126 char *buf)
127{
128 struct drm_device *ddev = dev_get_drvdata(dev);
129 struct amdgpu_device *adev = drm_to_adev(ddev);
130 struct ttm_resource_manager *man = &adev->mman.vram_mgr.manager;
131
132 return sysfs_emit(buf, "%llu\n", ttm_resource_manager_usage(man));
133}
134
135/**
136 * DOC: mem_info_vis_vram_used
137 *
138 * The amdgpu driver provides a sysfs API for reporting current total of
139 * used visible VRAM
140 * The file mem_info_vis_vram_used is used for this and returns the total
141 * amount of currently used visible VRAM in bytes
142 */
143static ssize_t amdgpu_mem_info_vis_vram_used_show(struct device *dev,
144 struct device_attribute *attr,
145 char *buf)
146{
147 struct drm_device *ddev = dev_get_drvdata(dev);
148 struct amdgpu_device *adev = drm_to_adev(ddev);
149
150 return sysfs_emit(buf, "%llu\n",
151 amdgpu_vram_mgr_vis_usage(&adev->mman.vram_mgr));
152}
153
154/**
155 * DOC: mem_info_vram_vendor
156 *
157 * The amdgpu driver provides a sysfs API for reporting the vendor of the
158 * installed VRAM
159 * The file mem_info_vram_vendor is used for this and returns the name of the
160 * vendor.
161 */
162static ssize_t amdgpu_mem_info_vram_vendor(struct device *dev,
163 struct device_attribute *attr,
164 char *buf)
165{
166 struct drm_device *ddev = dev_get_drvdata(dev);
167 struct amdgpu_device *adev = drm_to_adev(ddev);
168
169 switch (adev->gmc.vram_vendor) {
170 case SAMSUNG:
171 return sysfs_emit(buf, "samsung\n");
172 case INFINEON:
173 return sysfs_emit(buf, "infineon\n");
174 case ELPIDA:
175 return sysfs_emit(buf, "elpida\n");
176 case ETRON:
177 return sysfs_emit(buf, "etron\n");
178 case NANYA:
179 return sysfs_emit(buf, "nanya\n");
180 case HYNIX:
181 return sysfs_emit(buf, "hynix\n");
182 case MOSEL:
183 return sysfs_emit(buf, "mosel\n");
184 case WINBOND:
185 return sysfs_emit(buf, "winbond\n");
186 case ESMT:
187 return sysfs_emit(buf, "esmt\n");
188 case MICRON:
189 return sysfs_emit(buf, "micron\n");
190 default:
191 return sysfs_emit(buf, "unknown\n");
192 }
193}
194
195static DEVICE_ATTR(mem_info_vram_total, S_IRUGO,
196 amdgpu_mem_info_vram_total_show, NULL);
197static DEVICE_ATTR(mem_info_vis_vram_total, S_IRUGO,
198 amdgpu_mem_info_vis_vram_total_show,NULL);
199static DEVICE_ATTR(mem_info_vram_used, S_IRUGO,
200 amdgpu_mem_info_vram_used_show, NULL);
201static DEVICE_ATTR(mem_info_vis_vram_used, S_IRUGO,
202 amdgpu_mem_info_vis_vram_used_show, NULL);
203static DEVICE_ATTR(mem_info_vram_vendor, S_IRUGO,
204 amdgpu_mem_info_vram_vendor, NULL);
205
206static struct attribute *amdgpu_vram_mgr_attributes[] = {
207 &dev_attr_mem_info_vram_total.attr,
208 &dev_attr_mem_info_vis_vram_total.attr,
209 &dev_attr_mem_info_vram_used.attr,
210 &dev_attr_mem_info_vis_vram_used.attr,
211 &dev_attr_mem_info_vram_vendor.attr,
212 NULL
213};
214
215const struct attribute_group amdgpu_vram_mgr_attr_group = {
216 .attrs = amdgpu_vram_mgr_attributes
217};
218
219/**
220 * amdgpu_vram_mgr_vis_size - Calculate visible block size
221 *
222 * @adev: amdgpu_device pointer
223 * @block: DRM BUDDY block structure
224 *
225 * Calculate how many bytes of the DRM BUDDY block are inside visible VRAM
226 */
227static u64 amdgpu_vram_mgr_vis_size(struct amdgpu_device *adev,
228 struct drm_buddy_block *block)
229{
230 u64 start = amdgpu_vram_mgr_block_start(block);
231 u64 end = start + amdgpu_vram_mgr_block_size(block);
232
233 if (start >= adev->gmc.visible_vram_size)
234 return 0;
235
236 return (end > adev->gmc.visible_vram_size ?
237 adev->gmc.visible_vram_size : end) - start;
238}
239
240/**
241 * amdgpu_vram_mgr_bo_visible_size - CPU visible BO size
242 *
243 * @bo: &amdgpu_bo buffer object (must be in VRAM)
244 *
245 * Returns:
246 * How much of the given &amdgpu_bo buffer object lies in CPU visible VRAM.
247 */
248u64 amdgpu_vram_mgr_bo_visible_size(struct amdgpu_bo *bo)
249{
250 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
251 struct ttm_resource *res = bo->tbo.resource;
252 struct amdgpu_vram_mgr_resource *vres = to_amdgpu_vram_mgr_resource(res);
253 struct drm_buddy_block *block;
254 u64 usage = 0;
255
256 if (amdgpu_gmc_vram_full_visible(&adev->gmc))
257 return amdgpu_bo_size(bo);
258
259 if (res->start >= adev->gmc.visible_vram_size >> PAGE_SHIFT)
260 return 0;
261
262 list_for_each_entry(block, &vres->blocks, link)
263 usage += amdgpu_vram_mgr_vis_size(adev, block);
264
265 return usage;
266}
267
268/* Commit the reservation of VRAM pages */
269static void amdgpu_vram_mgr_do_reserve(struct ttm_resource_manager *man)
270{
271 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
272 struct amdgpu_device *adev = to_amdgpu_device(mgr);
273 struct drm_buddy *mm = &mgr->mm;
274 struct amdgpu_vram_reservation *rsv, *temp;
275 struct drm_buddy_block *block;
276 uint64_t vis_usage;
277
278 list_for_each_entry_safe(rsv, temp, &mgr->reservations_pending, blocks) {
279 if (drm_buddy_alloc_blocks(mm, rsv->start, rsv->start + rsv->size,
280 rsv->size, mm->chunk_size, &rsv->allocated,
281 DRM_BUDDY_RANGE_ALLOCATION))
282 continue;
283
284 block = amdgpu_vram_mgr_first_block(&rsv->allocated);
285 if (!block)
286 continue;
287
288 dev_dbg(adev->dev, "Reservation 0x%llx - %lld, Succeeded\n",
289 rsv->start, rsv->size);
290
291 vis_usage = amdgpu_vram_mgr_vis_size(adev, block);
292 atomic64_add(vis_usage, &mgr->vis_usage);
293 spin_lock(&man->bdev->lru_lock);
294 man->usage += rsv->size;
295 spin_unlock(&man->bdev->lru_lock);
296 list_move(&rsv->blocks, &mgr->reserved_pages);
297 }
298}
299
300/**
301 * amdgpu_vram_mgr_reserve_range - Reserve a range from VRAM
302 *
303 * @mgr: amdgpu_vram_mgr pointer
304 * @start: start address of the range in VRAM
305 * @size: size of the range
306 *
307 * Reserve memory from start address with the specified size in VRAM
308 */
309int amdgpu_vram_mgr_reserve_range(struct amdgpu_vram_mgr *mgr,
310 uint64_t start, uint64_t size)
311{
312 struct amdgpu_vram_reservation *rsv;
313
314 rsv = kzalloc(sizeof(*rsv), GFP_KERNEL);
315 if (!rsv)
316 return -ENOMEM;
317
318 INIT_LIST_HEAD(&rsv->allocated);
319 INIT_LIST_HEAD(&rsv->blocks);
320
321 rsv->start = start;
322 rsv->size = size;
323
324 mutex_lock(&mgr->lock);
325 list_add_tail(&rsv->blocks, &mgr->reservations_pending);
326 amdgpu_vram_mgr_do_reserve(&mgr->manager);
327 mutex_unlock(&mgr->lock);
328
329 return 0;
330}
331
332/**
333 * amdgpu_vram_mgr_query_page_status - query the reservation status
334 *
335 * @mgr: amdgpu_vram_mgr pointer
336 * @start: start address of a page in VRAM
337 *
338 * Returns:
339 * -EBUSY: the page is still hold and in pending list
340 * 0: the page has been reserved
341 * -ENOENT: the input page is not a reservation
342 */
343int amdgpu_vram_mgr_query_page_status(struct amdgpu_vram_mgr *mgr,
344 uint64_t start)
345{
346 struct amdgpu_vram_reservation *rsv;
347 int ret;
348
349 mutex_lock(&mgr->lock);
350
351 list_for_each_entry(rsv, &mgr->reservations_pending, blocks) {
352 if (rsv->start <= start &&
353 (start < (rsv->start + rsv->size))) {
354 ret = -EBUSY;
355 goto out;
356 }
357 }
358
359 list_for_each_entry(rsv, &mgr->reserved_pages, blocks) {
360 if (rsv->start <= start &&
361 (start < (rsv->start + rsv->size))) {
362 ret = 0;
363 goto out;
364 }
365 }
366
367 ret = -ENOENT;
368out:
369 mutex_unlock(&mgr->lock);
370 return ret;
371}
372
373/**
374 * amdgpu_vram_mgr_new - allocate new ranges
375 *
376 * @man: TTM memory type manager
377 * @tbo: TTM BO we need this range for
378 * @place: placement flags and restrictions
379 * @res: the resulting mem object
380 *
381 * Allocate VRAM for the given BO.
382 */
383static int amdgpu_vram_mgr_new(struct ttm_resource_manager *man,
384 struct ttm_buffer_object *tbo,
385 const struct ttm_place *place,
386 struct ttm_resource **res)
387{
388 u64 vis_usage = 0, max_bytes, cur_size, min_block_size;
389 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
390 struct amdgpu_device *adev = to_amdgpu_device(mgr);
391 struct amdgpu_vram_mgr_resource *vres;
392 u64 size, remaining_size, lpfn, fpfn;
393 struct drm_buddy *mm = &mgr->mm;
394 struct drm_buddy_block *block;
395 unsigned long pages_per_block;
396 int r;
397
398 lpfn = (u64)place->lpfn << PAGE_SHIFT;
399 if (!lpfn)
400 lpfn = man->size;
401
402 fpfn = (u64)place->fpfn << PAGE_SHIFT;
403
404 max_bytes = adev->gmc.mc_vram_size;
405 if (tbo->type != ttm_bo_type_kernel)
406 max_bytes -= AMDGPU_VM_RESERVED_VRAM;
407
408 if (place->flags & TTM_PL_FLAG_CONTIGUOUS) {
409 pages_per_block = ~0ul;
410 } else {
411#ifdef CONFIG_TRANSPARENT_HUGEPAGE
412 pages_per_block = HPAGE_PMD_NR;
413#else
414 /* default to 2MB */
415 pages_per_block = 2UL << (20UL - PAGE_SHIFT);
416#endif
417 pages_per_block = max_t(uint32_t, pages_per_block,
418 tbo->page_alignment);
419 }
420
421 vres = kzalloc(sizeof(*vres), GFP_KERNEL);
422 if (!vres)
423 return -ENOMEM;
424
425 ttm_resource_init(tbo, place, &vres->base);
426
427 /* bail out quickly if there's likely not enough VRAM for this BO */
428 if (ttm_resource_manager_usage(man) > max_bytes) {
429 r = -ENOSPC;
430 goto error_fini;
431 }
432
433 INIT_LIST_HEAD(&vres->blocks);
434
435 if (place->flags & TTM_PL_FLAG_TOPDOWN)
436 vres->flags |= DRM_BUDDY_TOPDOWN_ALLOCATION;
437
438 if (fpfn || lpfn != mgr->mm.size)
439 /* Allocate blocks in desired range */
440 vres->flags |= DRM_BUDDY_RANGE_ALLOCATION;
441
442 remaining_size = (u64)vres->base.size;
443
444 mutex_lock(&mgr->lock);
445 while (remaining_size) {
446 if (tbo->page_alignment)
447 min_block_size = (u64)tbo->page_alignment << PAGE_SHIFT;
448 else
449 min_block_size = mgr->default_page_size;
450
451 BUG_ON(min_block_size < mm->chunk_size);
452
453 /* Limit maximum size to 2GiB due to SG table limitations */
454 size = min(remaining_size, 2ULL << 30);
455
456 if (size >= (u64)pages_per_block << PAGE_SHIFT)
457 min_block_size = (u64)pages_per_block << PAGE_SHIFT;
458
459 cur_size = size;
460
461 if (fpfn + size != (u64)place->lpfn << PAGE_SHIFT) {
462 /*
463 * Except for actual range allocation, modify the size and
464 * min_block_size conforming to continuous flag enablement
465 */
466 if (place->flags & TTM_PL_FLAG_CONTIGUOUS) {
467 size = roundup_pow_of_two(size);
468 min_block_size = size;
469 /*
470 * Modify the size value if size is not
471 * aligned with min_block_size
472 */
473 } else if (!IS_ALIGNED(size, min_block_size)) {
474 size = round_up(size, min_block_size);
475 }
476 }
477
478 r = drm_buddy_alloc_blocks(mm, fpfn,
479 lpfn,
480 size,
481 min_block_size,
482 &vres->blocks,
483 vres->flags);
484 if (unlikely(r))
485 goto error_free_blocks;
486
487 if (size > remaining_size)
488 remaining_size = 0;
489 else
490 remaining_size -= size;
491 }
492 mutex_unlock(&mgr->lock);
493
494 if (cur_size != size) {
495 struct drm_buddy_block *block;
496 struct list_head *trim_list;
497 u64 original_size;
498 LIST_HEAD(temp);
499
500 trim_list = &vres->blocks;
501 original_size = (u64)vres->base.size;
502
503 /*
504 * If size value is rounded up to min_block_size, trim the last
505 * block to the required size
506 */
507 if (!list_is_singular(&vres->blocks)) {
508 block = list_last_entry(&vres->blocks, typeof(*block), link);
509 list_move_tail(&block->link, &temp);
510 trim_list = &temp;
511 /*
512 * Compute the original_size value by subtracting the
513 * last block size with (aligned size - original size)
514 */
515 original_size = amdgpu_vram_mgr_block_size(block) - (size - cur_size);
516 }
517
518 mutex_lock(&mgr->lock);
519 drm_buddy_block_trim(mm,
520 original_size,
521 trim_list);
522 mutex_unlock(&mgr->lock);
523
524 if (!list_empty(&temp))
525 list_splice_tail(trim_list, &vres->blocks);
526 }
527
528 vres->base.start = 0;
529 list_for_each_entry(block, &vres->blocks, link) {
530 unsigned long start;
531
532 start = amdgpu_vram_mgr_block_start(block) +
533 amdgpu_vram_mgr_block_size(block);
534 start >>= PAGE_SHIFT;
535
536 if (start > PFN_UP(vres->base.size))
537 start -= PFN_UP(vres->base.size);
538 else
539 start = 0;
540 vres->base.start = max(vres->base.start, start);
541
542 vis_usage += amdgpu_vram_mgr_vis_size(adev, block);
543 }
544
545 if (amdgpu_is_vram_mgr_blocks_contiguous(&vres->blocks))
546 vres->base.placement |= TTM_PL_FLAG_CONTIGUOUS;
547
548 if (adev->gmc.xgmi.connected_to_cpu)
549 vres->base.bus.caching = ttm_cached;
550 else
551 vres->base.bus.caching = ttm_write_combined;
552
553 atomic64_add(vis_usage, &mgr->vis_usage);
554 *res = &vres->base;
555 return 0;
556
557error_free_blocks:
558 drm_buddy_free_list(mm, &vres->blocks);
559 mutex_unlock(&mgr->lock);
560error_fini:
561 ttm_resource_fini(man, &vres->base);
562 kfree(vres);
563
564 return r;
565}
566
567/**
568 * amdgpu_vram_mgr_del - free ranges
569 *
570 * @man: TTM memory type manager
571 * @res: TTM memory object
572 *
573 * Free the allocated VRAM again.
574 */
575static void amdgpu_vram_mgr_del(struct ttm_resource_manager *man,
576 struct ttm_resource *res)
577{
578 struct amdgpu_vram_mgr_resource *vres = to_amdgpu_vram_mgr_resource(res);
579 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
580 struct amdgpu_device *adev = to_amdgpu_device(mgr);
581 struct drm_buddy *mm = &mgr->mm;
582 struct drm_buddy_block *block;
583 uint64_t vis_usage = 0;
584
585 mutex_lock(&mgr->lock);
586 list_for_each_entry(block, &vres->blocks, link)
587 vis_usage += amdgpu_vram_mgr_vis_size(adev, block);
588
589 amdgpu_vram_mgr_do_reserve(man);
590
591 drm_buddy_free_list(mm, &vres->blocks);
592 mutex_unlock(&mgr->lock);
593
594 atomic64_sub(vis_usage, &mgr->vis_usage);
595
596 ttm_resource_fini(man, res);
597 kfree(vres);
598}
599
600/**
601 * amdgpu_vram_mgr_alloc_sgt - allocate and fill a sg table
602 *
603 * @adev: amdgpu device pointer
604 * @res: TTM memory object
605 * @offset: byte offset from the base of VRAM BO
606 * @length: number of bytes to export in sg_table
607 * @dev: the other device
608 * @dir: dma direction
609 * @sgt: resulting sg table
610 *
611 * Allocate and fill a sg table from a VRAM allocation.
612 */
613int amdgpu_vram_mgr_alloc_sgt(struct amdgpu_device *adev,
614 struct ttm_resource *res,
615 u64 offset, u64 length,
616 struct device *dev,
617 enum dma_data_direction dir,
618 struct sg_table **sgt)
619{
620 struct amdgpu_res_cursor cursor;
621 struct scatterlist *sg;
622 int num_entries = 0;
623 int i, r;
624
625 *sgt = kmalloc(sizeof(**sgt), GFP_KERNEL);
626 if (!*sgt)
627 return -ENOMEM;
628
629 /* Determine the number of DRM_BUDDY blocks to export */
630 amdgpu_res_first(res, offset, length, &cursor);
631 while (cursor.remaining) {
632 num_entries++;
633 amdgpu_res_next(&cursor, cursor.size);
634 }
635
636 r = sg_alloc_table(*sgt, num_entries, GFP_KERNEL);
637 if (r)
638 goto error_free;
639
640 /* Initialize scatterlist nodes of sg_table */
641 for_each_sgtable_sg((*sgt), sg, i)
642 sg->length = 0;
643
644 /*
645 * Walk down DRM_BUDDY blocks to populate scatterlist nodes
646 * @note: Use iterator api to get first the DRM_BUDDY block
647 * and the number of bytes from it. Access the following
648 * DRM_BUDDY block(s) if more buffer needs to exported
649 */
650 amdgpu_res_first(res, offset, length, &cursor);
651 for_each_sgtable_sg((*sgt), sg, i) {
652 phys_addr_t phys = cursor.start + adev->gmc.aper_base;
653 size_t size = cursor.size;
654 dma_addr_t addr;
655
656 addr = dma_map_resource(dev, phys, size, dir,
657 DMA_ATTR_SKIP_CPU_SYNC);
658 r = dma_mapping_error(dev, addr);
659 if (r)
660 goto error_unmap;
661
662 sg_set_page(sg, NULL, size, 0);
663 sg_dma_address(sg) = addr;
664 sg_dma_len(sg) = size;
665
666 amdgpu_res_next(&cursor, cursor.size);
667 }
668
669 return 0;
670
671error_unmap:
672 for_each_sgtable_sg((*sgt), sg, i) {
673 if (!sg->length)
674 continue;
675
676 dma_unmap_resource(dev, sg->dma_address,
677 sg->length, dir,
678 DMA_ATTR_SKIP_CPU_SYNC);
679 }
680 sg_free_table(*sgt);
681
682error_free:
683 kfree(*sgt);
684 return r;
685}
686
687/**
688 * amdgpu_vram_mgr_free_sgt - allocate and fill a sg table
689 *
690 * @dev: device pointer
691 * @dir: data direction of resource to unmap
692 * @sgt: sg table to free
693 *
694 * Free a previously allocate sg table.
695 */
696void amdgpu_vram_mgr_free_sgt(struct device *dev,
697 enum dma_data_direction dir,
698 struct sg_table *sgt)
699{
700 struct scatterlist *sg;
701 int i;
702
703 for_each_sgtable_sg(sgt, sg, i)
704 dma_unmap_resource(dev, sg->dma_address,
705 sg->length, dir,
706 DMA_ATTR_SKIP_CPU_SYNC);
707 sg_free_table(sgt);
708 kfree(sgt);
709}
710
711/**
712 * amdgpu_vram_mgr_vis_usage - how many bytes are used in the visible part
713 *
714 * @mgr: amdgpu_vram_mgr pointer
715 *
716 * Returns how many bytes are used in the visible part of VRAM
717 */
718uint64_t amdgpu_vram_mgr_vis_usage(struct amdgpu_vram_mgr *mgr)
719{
720 return atomic64_read(&mgr->vis_usage);
721}
722
723/**
724 * amdgpu_vram_mgr_intersects - test each drm buddy block for intersection
725 *
726 * @man: TTM memory type manager
727 * @res: The resource to test
728 * @place: The place to test against
729 * @size: Size of the new allocation
730 *
731 * Test each drm buddy block for intersection for eviction decision.
732 */
733static bool amdgpu_vram_mgr_intersects(struct ttm_resource_manager *man,
734 struct ttm_resource *res,
735 const struct ttm_place *place,
736 size_t size)
737{
738 struct amdgpu_vram_mgr_resource *mgr = to_amdgpu_vram_mgr_resource(res);
739 struct drm_buddy_block *block;
740
741 /* Check each drm buddy block individually */
742 list_for_each_entry(block, &mgr->blocks, link) {
743 unsigned long fpfn =
744 amdgpu_vram_mgr_block_start(block) >> PAGE_SHIFT;
745 unsigned long lpfn = fpfn +
746 (amdgpu_vram_mgr_block_size(block) >> PAGE_SHIFT);
747
748 if (place->fpfn < lpfn &&
749 (!place->lpfn || place->lpfn > fpfn))
750 return true;
751 }
752
753 return false;
754}
755
756/**
757 * amdgpu_vram_mgr_compatible - test each drm buddy block for compatibility
758 *
759 * @man: TTM memory type manager
760 * @res: The resource to test
761 * @place: The place to test against
762 * @size: Size of the new allocation
763 *
764 * Test each drm buddy block for placement compatibility.
765 */
766static bool amdgpu_vram_mgr_compatible(struct ttm_resource_manager *man,
767 struct ttm_resource *res,
768 const struct ttm_place *place,
769 size_t size)
770{
771 struct amdgpu_vram_mgr_resource *mgr = to_amdgpu_vram_mgr_resource(res);
772 struct drm_buddy_block *block;
773
774 /* Check each drm buddy block individually */
775 list_for_each_entry(block, &mgr->blocks, link) {
776 unsigned long fpfn =
777 amdgpu_vram_mgr_block_start(block) >> PAGE_SHIFT;
778 unsigned long lpfn = fpfn +
779 (amdgpu_vram_mgr_block_size(block) >> PAGE_SHIFT);
780
781 if (fpfn < place->fpfn ||
782 (place->lpfn && lpfn > place->lpfn))
783 return false;
784 }
785
786 return true;
787}
788
789/**
790 * amdgpu_vram_mgr_debug - dump VRAM table
791 *
792 * @man: TTM memory type manager
793 * @printer: DRM printer to use
794 *
795 * Dump the table content using printk.
796 */
797static void amdgpu_vram_mgr_debug(struct ttm_resource_manager *man,
798 struct drm_printer *printer)
799{
800 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
801 struct drm_buddy *mm = &mgr->mm;
802 struct drm_buddy_block *block;
803
804 drm_printf(printer, " vis usage:%llu\n",
805 amdgpu_vram_mgr_vis_usage(mgr));
806
807 mutex_lock(&mgr->lock);
808 drm_printf(printer, "default_page_size: %lluKiB\n",
809 mgr->default_page_size >> 10);
810
811 drm_buddy_print(mm, printer);
812
813 drm_printf(printer, "reserved:\n");
814 list_for_each_entry(block, &mgr->reserved_pages, link)
815 drm_buddy_block_print(mm, block, printer);
816 mutex_unlock(&mgr->lock);
817}
818
819static const struct ttm_resource_manager_func amdgpu_vram_mgr_func = {
820 .alloc = amdgpu_vram_mgr_new,
821 .free = amdgpu_vram_mgr_del,
822 .intersects = amdgpu_vram_mgr_intersects,
823 .compatible = amdgpu_vram_mgr_compatible,
824 .debug = amdgpu_vram_mgr_debug
825};
826
827/**
828 * amdgpu_vram_mgr_init - init VRAM manager and DRM MM
829 *
830 * @adev: amdgpu_device pointer
831 *
832 * Allocate and initialize the VRAM manager.
833 */
834int amdgpu_vram_mgr_init(struct amdgpu_device *adev)
835{
836 struct amdgpu_vram_mgr *mgr = &adev->mman.vram_mgr;
837 struct ttm_resource_manager *man = &mgr->manager;
838 int err;
839
840 ttm_resource_manager_init(man, &adev->mman.bdev,
841 adev->gmc.real_vram_size);
842
843 man->func = &amdgpu_vram_mgr_func;
844
845 err = drm_buddy_init(&mgr->mm, man->size, PAGE_SIZE);
846 if (err)
847 return err;
848
849 mutex_init(&mgr->lock);
850 INIT_LIST_HEAD(&mgr->reservations_pending);
851 INIT_LIST_HEAD(&mgr->reserved_pages);
852 mgr->default_page_size = PAGE_SIZE;
853
854 ttm_set_driver_manager(&adev->mman.bdev, TTM_PL_VRAM, &mgr->manager);
855 ttm_resource_manager_set_used(man, true);
856 return 0;
857}
858
859/**
860 * amdgpu_vram_mgr_fini - free and destroy VRAM manager
861 *
862 * @adev: amdgpu_device pointer
863 *
864 * Destroy and free the VRAM manager, returns -EBUSY if ranges are still
865 * allocated inside it.
866 */
867void amdgpu_vram_mgr_fini(struct amdgpu_device *adev)
868{
869 struct amdgpu_vram_mgr *mgr = &adev->mman.vram_mgr;
870 struct ttm_resource_manager *man = &mgr->manager;
871 int ret;
872 struct amdgpu_vram_reservation *rsv, *temp;
873
874 ttm_resource_manager_set_used(man, false);
875
876 ret = ttm_resource_manager_evict_all(&adev->mman.bdev, man);
877 if (ret)
878 return;
879
880 mutex_lock(&mgr->lock);
881 list_for_each_entry_safe(rsv, temp, &mgr->reservations_pending, blocks)
882 kfree(rsv);
883
884 list_for_each_entry_safe(rsv, temp, &mgr->reserved_pages, blocks) {
885 drm_buddy_free_list(&mgr->mm, &rsv->allocated);
886 kfree(rsv);
887 }
888 drm_buddy_fini(&mgr->mm);
889 mutex_unlock(&mgr->lock);
890
891 ttm_resource_manager_cleanup(man);
892 ttm_set_driver_manager(&adev->mman.bdev, TTM_PL_VRAM, NULL);
893}