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