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
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}