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 "amdgpu.h"
27#include "amdgpu_vm.h"
28#include "amdgpu_atomfirmware.h"
29#include "atom.h"
30
31struct amdgpu_vram_mgr {
32 struct drm_mm mm;
33 spinlock_t lock;
34 atomic64_t usage;
35 atomic64_t vis_usage;
36};
37
38/**
39 * DOC: mem_info_vram_total
40 *
41 * The amdgpu driver provides a sysfs API for reporting current total VRAM
42 * available on the device
43 * The file mem_info_vram_total is used for this and returns the total
44 * amount of VRAM in bytes
45 */
46static ssize_t amdgpu_mem_info_vram_total_show(struct device *dev,
47 struct device_attribute *attr, char *buf)
48{
49 struct drm_device *ddev = dev_get_drvdata(dev);
50 struct amdgpu_device *adev = ddev->dev_private;
51
52 return snprintf(buf, PAGE_SIZE, "%llu\n", adev->gmc.real_vram_size);
53}
54
55/**
56 * DOC: mem_info_vis_vram_total
57 *
58 * The amdgpu driver provides a sysfs API for reporting current total
59 * visible VRAM available on the device
60 * The file mem_info_vis_vram_total is used for this and returns the total
61 * amount of visible VRAM in bytes
62 */
63static ssize_t amdgpu_mem_info_vis_vram_total_show(struct device *dev,
64 struct device_attribute *attr, char *buf)
65{
66 struct drm_device *ddev = dev_get_drvdata(dev);
67 struct amdgpu_device *adev = ddev->dev_private;
68
69 return snprintf(buf, PAGE_SIZE, "%llu\n", adev->gmc.visible_vram_size);
70}
71
72/**
73 * DOC: mem_info_vram_used
74 *
75 * The amdgpu driver provides a sysfs API for reporting current total VRAM
76 * available on the device
77 * The file mem_info_vram_used is used for this and returns the total
78 * amount of currently used VRAM in bytes
79 */
80static ssize_t amdgpu_mem_info_vram_used_show(struct device *dev,
81 struct device_attribute *attr, char *buf)
82{
83 struct drm_device *ddev = dev_get_drvdata(dev);
84 struct amdgpu_device *adev = ddev->dev_private;
85
86 return snprintf(buf, PAGE_SIZE, "%llu\n",
87 amdgpu_vram_mgr_usage(&adev->mman.bdev.man[TTM_PL_VRAM]));
88}
89
90/**
91 * DOC: mem_info_vis_vram_used
92 *
93 * The amdgpu driver provides a sysfs API for reporting current total of
94 * used visible VRAM
95 * The file mem_info_vis_vram_used is used for this and returns the total
96 * amount of currently used visible VRAM in bytes
97 */
98static ssize_t amdgpu_mem_info_vis_vram_used_show(struct device *dev,
99 struct device_attribute *attr, char *buf)
100{
101 struct drm_device *ddev = dev_get_drvdata(dev);
102 struct amdgpu_device *adev = ddev->dev_private;
103
104 return snprintf(buf, PAGE_SIZE, "%llu\n",
105 amdgpu_vram_mgr_vis_usage(&adev->mman.bdev.man[TTM_PL_VRAM]));
106}
107
108static ssize_t amdgpu_mem_info_vram_vendor(struct device *dev,
109 struct device_attribute *attr,
110 char *buf)
111{
112 struct drm_device *ddev = dev_get_drvdata(dev);
113 struct amdgpu_device *adev = ddev->dev_private;
114
115 switch (adev->gmc.vram_vendor) {
116 case SAMSUNG:
117 return snprintf(buf, PAGE_SIZE, "samsung\n");
118 case INFINEON:
119 return snprintf(buf, PAGE_SIZE, "infineon\n");
120 case ELPIDA:
121 return snprintf(buf, PAGE_SIZE, "elpida\n");
122 case ETRON:
123 return snprintf(buf, PAGE_SIZE, "etron\n");
124 case NANYA:
125 return snprintf(buf, PAGE_SIZE, "nanya\n");
126 case HYNIX:
127 return snprintf(buf, PAGE_SIZE, "hynix\n");
128 case MOSEL:
129 return snprintf(buf, PAGE_SIZE, "mosel\n");
130 case WINBOND:
131 return snprintf(buf, PAGE_SIZE, "winbond\n");
132 case ESMT:
133 return snprintf(buf, PAGE_SIZE, "esmt\n");
134 case MICRON:
135 return snprintf(buf, PAGE_SIZE, "micron\n");
136 default:
137 return snprintf(buf, PAGE_SIZE, "unknown\n");
138 }
139}
140
141static DEVICE_ATTR(mem_info_vram_total, S_IRUGO,
142 amdgpu_mem_info_vram_total_show, NULL);
143static DEVICE_ATTR(mem_info_vis_vram_total, S_IRUGO,
144 amdgpu_mem_info_vis_vram_total_show,NULL);
145static DEVICE_ATTR(mem_info_vram_used, S_IRUGO,
146 amdgpu_mem_info_vram_used_show, NULL);
147static DEVICE_ATTR(mem_info_vis_vram_used, S_IRUGO,
148 amdgpu_mem_info_vis_vram_used_show, NULL);
149static DEVICE_ATTR(mem_info_vram_vendor, S_IRUGO,
150 amdgpu_mem_info_vram_vendor, NULL);
151
152static const struct attribute *amdgpu_vram_mgr_attributes[] = {
153 &dev_attr_mem_info_vram_total.attr,
154 &dev_attr_mem_info_vis_vram_total.attr,
155 &dev_attr_mem_info_vram_used.attr,
156 &dev_attr_mem_info_vis_vram_used.attr,
157 &dev_attr_mem_info_vram_vendor.attr,
158 NULL
159};
160
161/**
162 * amdgpu_vram_mgr_init - init VRAM manager and DRM MM
163 *
164 * @man: TTM memory type manager
165 * @p_size: maximum size of VRAM
166 *
167 * Allocate and initialize the VRAM manager.
168 */
169static int amdgpu_vram_mgr_init(struct ttm_mem_type_manager *man,
170 unsigned long p_size)
171{
172 struct amdgpu_device *adev = amdgpu_ttm_adev(man->bdev);
173 struct amdgpu_vram_mgr *mgr;
174 int ret;
175
176 mgr = kzalloc(sizeof(*mgr), GFP_KERNEL);
177 if (!mgr)
178 return -ENOMEM;
179
180 drm_mm_init(&mgr->mm, 0, p_size);
181 spin_lock_init(&mgr->lock);
182 man->priv = mgr;
183
184 /* Add the two VRAM-related sysfs files */
185 ret = sysfs_create_files(&adev->dev->kobj, amdgpu_vram_mgr_attributes);
186 if (ret)
187 DRM_ERROR("Failed to register sysfs\n");
188
189 return 0;
190}
191
192/**
193 * amdgpu_vram_mgr_fini - free and destroy VRAM manager
194 *
195 * @man: TTM memory type manager
196 *
197 * Destroy and free the VRAM manager, returns -EBUSY if ranges are still
198 * allocated inside it.
199 */
200static int amdgpu_vram_mgr_fini(struct ttm_mem_type_manager *man)
201{
202 struct amdgpu_device *adev = amdgpu_ttm_adev(man->bdev);
203 struct amdgpu_vram_mgr *mgr = man->priv;
204
205 spin_lock(&mgr->lock);
206 drm_mm_takedown(&mgr->mm);
207 spin_unlock(&mgr->lock);
208 kfree(mgr);
209 man->priv = NULL;
210 sysfs_remove_files(&adev->dev->kobj, amdgpu_vram_mgr_attributes);
211 return 0;
212}
213
214/**
215 * amdgpu_vram_mgr_vis_size - Calculate visible node size
216 *
217 * @adev: amdgpu device structure
218 * @node: MM node structure
219 *
220 * Calculate how many bytes of the MM node are inside visible VRAM
221 */
222static u64 amdgpu_vram_mgr_vis_size(struct amdgpu_device *adev,
223 struct drm_mm_node *node)
224{
225 uint64_t start = node->start << PAGE_SHIFT;
226 uint64_t end = (node->size + node->start) << PAGE_SHIFT;
227
228 if (start >= adev->gmc.visible_vram_size)
229 return 0;
230
231 return (end > adev->gmc.visible_vram_size ?
232 adev->gmc.visible_vram_size : end) - start;
233}
234
235/**
236 * amdgpu_vram_mgr_bo_visible_size - CPU visible BO size
237 *
238 * @bo: &amdgpu_bo buffer object (must be in VRAM)
239 *
240 * Returns:
241 * How much of the given &amdgpu_bo buffer object lies in CPU visible VRAM.
242 */
243u64 amdgpu_vram_mgr_bo_visible_size(struct amdgpu_bo *bo)
244{
245 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
246 struct ttm_mem_reg *mem = &bo->tbo.mem;
247 struct drm_mm_node *nodes = mem->mm_node;
248 unsigned pages = mem->num_pages;
249 u64 usage;
250
251 if (amdgpu_gmc_vram_full_visible(&adev->gmc))
252 return amdgpu_bo_size(bo);
253
254 if (mem->start >= adev->gmc.visible_vram_size >> PAGE_SHIFT)
255 return 0;
256
257 for (usage = 0; nodes && pages; pages -= nodes->size, nodes++)
258 usage += amdgpu_vram_mgr_vis_size(adev, nodes);
259
260 return usage;
261}
262
263/**
264 * amdgpu_vram_mgr_virt_start - update virtual start address
265 *
266 * @mem: ttm_mem_reg to update
267 * @node: just allocated node
268 *
269 * Calculate a virtual BO start address to easily check if everything is CPU
270 * accessible.
271 */
272static void amdgpu_vram_mgr_virt_start(struct ttm_mem_reg *mem,
273 struct drm_mm_node *node)
274{
275 unsigned long start;
276
277 start = node->start + node->size;
278 if (start > mem->num_pages)
279 start -= mem->num_pages;
280 else
281 start = 0;
282 mem->start = max(mem->start, start);
283}
284
285/**
286 * amdgpu_vram_mgr_new - allocate new ranges
287 *
288 * @man: TTM memory type manager
289 * @tbo: TTM BO we need this range for
290 * @place: placement flags and restrictions
291 * @mem: the resulting mem object
292 *
293 * Allocate VRAM for the given BO.
294 */
295static int amdgpu_vram_mgr_new(struct ttm_mem_type_manager *man,
296 struct ttm_buffer_object *tbo,
297 const struct ttm_place *place,
298 struct ttm_mem_reg *mem)
299{
300 struct amdgpu_device *adev = amdgpu_ttm_adev(man->bdev);
301 struct amdgpu_vram_mgr *mgr = man->priv;
302 struct drm_mm *mm = &mgr->mm;
303 struct drm_mm_node *nodes;
304 enum drm_mm_insert_mode mode;
305 unsigned long lpfn, num_nodes, pages_per_node, pages_left;
306 uint64_t vis_usage = 0, mem_bytes, max_bytes;
307 unsigned i;
308 int r;
309
310 lpfn = place->lpfn;
311 if (!lpfn)
312 lpfn = man->size;
313
314 max_bytes = adev->gmc.mc_vram_size;
315 if (tbo->type != ttm_bo_type_kernel)
316 max_bytes -= AMDGPU_VM_RESERVED_VRAM;
317
318 /* bail out quickly if there's likely not enough VRAM for this BO */
319 mem_bytes = (u64)mem->num_pages << PAGE_SHIFT;
320 if (atomic64_add_return(mem_bytes, &mgr->usage) > max_bytes) {
321 atomic64_sub(mem_bytes, &mgr->usage);
322 return -ENOSPC;
323 }
324
325 if (place->flags & TTM_PL_FLAG_CONTIGUOUS) {
326 pages_per_node = ~0ul;
327 num_nodes = 1;
328 } else {
329#ifdef CONFIG_TRANSPARENT_HUGEPAGE
330 pages_per_node = HPAGE_PMD_NR;
331#else
332 /* default to 2MB */
333 pages_per_node = (2UL << (20UL - PAGE_SHIFT));
334#endif
335 pages_per_node = max((uint32_t)pages_per_node, mem->page_alignment);
336 num_nodes = DIV_ROUND_UP(mem->num_pages, pages_per_node);
337 }
338
339 nodes = kvmalloc_array((uint32_t)num_nodes, sizeof(*nodes),
340 GFP_KERNEL | __GFP_ZERO);
341 if (!nodes) {
342 atomic64_sub(mem_bytes, &mgr->usage);
343 return -ENOMEM;
344 }
345
346 mode = DRM_MM_INSERT_BEST;
347 if (place->flags & TTM_PL_FLAG_TOPDOWN)
348 mode = DRM_MM_INSERT_HIGH;
349
350 mem->start = 0;
351 pages_left = mem->num_pages;
352
353 spin_lock(&mgr->lock);
354 for (i = 0; pages_left >= pages_per_node; ++i) {
355 unsigned long pages = rounddown_pow_of_two(pages_left);
356
357 r = drm_mm_insert_node_in_range(mm, &nodes[i], pages,
358 pages_per_node, 0,
359 place->fpfn, lpfn,
360 mode);
361 if (unlikely(r))
362 break;
363
364 vis_usage += amdgpu_vram_mgr_vis_size(adev, &nodes[i]);
365 amdgpu_vram_mgr_virt_start(mem, &nodes[i]);
366 pages_left -= pages;
367 }
368
369 for (; pages_left; ++i) {
370 unsigned long pages = min(pages_left, pages_per_node);
371 uint32_t alignment = mem->page_alignment;
372
373 if (pages == pages_per_node)
374 alignment = pages_per_node;
375
376 r = drm_mm_insert_node_in_range(mm, &nodes[i],
377 pages, alignment, 0,
378 place->fpfn, lpfn,
379 mode);
380 if (unlikely(r))
381 goto error;
382
383 vis_usage += amdgpu_vram_mgr_vis_size(adev, &nodes[i]);
384 amdgpu_vram_mgr_virt_start(mem, &nodes[i]);
385 pages_left -= pages;
386 }
387 spin_unlock(&mgr->lock);
388
389 atomic64_add(vis_usage, &mgr->vis_usage);
390
391 mem->mm_node = nodes;
392
393 return 0;
394
395error:
396 while (i--)
397 drm_mm_remove_node(&nodes[i]);
398 spin_unlock(&mgr->lock);
399 atomic64_sub(mem->num_pages << PAGE_SHIFT, &mgr->usage);
400
401 kvfree(nodes);
402 return r;
403}
404
405/**
406 * amdgpu_vram_mgr_del - free ranges
407 *
408 * @man: TTM memory type manager
409 * @mem: TTM memory object
410 *
411 * Free the allocated VRAM again.
412 */
413static void amdgpu_vram_mgr_del(struct ttm_mem_type_manager *man,
414 struct ttm_mem_reg *mem)
415{
416 struct amdgpu_device *adev = amdgpu_ttm_adev(man->bdev);
417 struct amdgpu_vram_mgr *mgr = man->priv;
418 struct drm_mm_node *nodes = mem->mm_node;
419 uint64_t usage = 0, vis_usage = 0;
420 unsigned pages = mem->num_pages;
421
422 if (!mem->mm_node)
423 return;
424
425 spin_lock(&mgr->lock);
426 while (pages) {
427 pages -= nodes->size;
428 drm_mm_remove_node(nodes);
429 usage += nodes->size << PAGE_SHIFT;
430 vis_usage += amdgpu_vram_mgr_vis_size(adev, nodes);
431 ++nodes;
432 }
433 spin_unlock(&mgr->lock);
434
435 atomic64_sub(usage, &mgr->usage);
436 atomic64_sub(vis_usage, &mgr->vis_usage);
437
438 kvfree(mem->mm_node);
439 mem->mm_node = NULL;
440}
441
442/**
443 * amdgpu_vram_mgr_alloc_sgt - allocate and fill a sg table
444 *
445 * @adev: amdgpu device pointer
446 * @mem: TTM memory object
447 * @dev: the other device
448 * @dir: dma direction
449 * @sgt: resulting sg table
450 *
451 * Allocate and fill a sg table from a VRAM allocation.
452 */
453int amdgpu_vram_mgr_alloc_sgt(struct amdgpu_device *adev,
454 struct ttm_mem_reg *mem,
455 struct device *dev,
456 enum dma_data_direction dir,
457 struct sg_table **sgt)
458{
459 struct drm_mm_node *node;
460 struct scatterlist *sg;
461 int num_entries = 0;
462 unsigned int pages;
463 int i, r;
464
465 *sgt = kmalloc(sizeof(**sgt), GFP_KERNEL);
466 if (!*sgt)
467 return -ENOMEM;
468
469 for (pages = mem->num_pages, node = mem->mm_node;
470 pages; pages -= node->size, ++node)
471 ++num_entries;
472
473 r = sg_alloc_table(*sgt, num_entries, GFP_KERNEL);
474 if (r)
475 goto error_free;
476
477 for_each_sgtable_sg((*sgt), sg, i)
478 sg->length = 0;
479
480 node = mem->mm_node;
481 for_each_sgtable_sg((*sgt), sg, i) {
482 phys_addr_t phys = (node->start << PAGE_SHIFT) +
483 adev->gmc.aper_base;
484 size_t size = node->size << PAGE_SHIFT;
485 dma_addr_t addr;
486
487 ++node;
488 addr = dma_map_resource(dev, phys, size, dir,
489 DMA_ATTR_SKIP_CPU_SYNC);
490 r = dma_mapping_error(dev, addr);
491 if (r)
492 goto error_unmap;
493
494 sg_set_page(sg, NULL, size, 0);
495 sg_dma_address(sg) = addr;
496 sg_dma_len(sg) = size;
497 }
498 return 0;
499
500error_unmap:
501 for_each_sgtable_sg((*sgt), sg, i) {
502 if (!sg->length)
503 continue;
504
505 dma_unmap_resource(dev, sg->dma_address,
506 sg->length, dir,
507 DMA_ATTR_SKIP_CPU_SYNC);
508 }
509 sg_free_table(*sgt);
510
511error_free:
512 kfree(*sgt);
513 return r;
514}
515
516/**
517 * amdgpu_vram_mgr_alloc_sgt - allocate and fill a sg table
518 *
519 * @adev: amdgpu device pointer
520 * @sgt: sg table to free
521 *
522 * Free a previously allocate sg table.
523 */
524void amdgpu_vram_mgr_free_sgt(struct amdgpu_device *adev,
525 struct device *dev,
526 enum dma_data_direction dir,
527 struct sg_table *sgt)
528{
529 struct scatterlist *sg;
530 int i;
531
532 for_each_sgtable_sg(sgt, sg, i)
533 dma_unmap_resource(dev, sg->dma_address,
534 sg->length, dir,
535 DMA_ATTR_SKIP_CPU_SYNC);
536 sg_free_table(sgt);
537 kfree(sgt);
538}
539
540/**
541 * amdgpu_vram_mgr_usage - how many bytes are used in this domain
542 *
543 * @man: TTM memory type manager
544 *
545 * Returns how many bytes are used in this domain.
546 */
547uint64_t amdgpu_vram_mgr_usage(struct ttm_mem_type_manager *man)
548{
549 struct amdgpu_vram_mgr *mgr = man->priv;
550
551 return atomic64_read(&mgr->usage);
552}
553
554/**
555 * amdgpu_vram_mgr_vis_usage - how many bytes are used in the visible part
556 *
557 * @man: TTM memory type manager
558 *
559 * Returns how many bytes are used in the visible part of VRAM
560 */
561uint64_t amdgpu_vram_mgr_vis_usage(struct ttm_mem_type_manager *man)
562{
563 struct amdgpu_vram_mgr *mgr = man->priv;
564
565 return atomic64_read(&mgr->vis_usage);
566}
567
568/**
569 * amdgpu_vram_mgr_debug - dump VRAM table
570 *
571 * @man: TTM memory type manager
572 * @printer: DRM printer to use
573 *
574 * Dump the table content using printk.
575 */
576static void amdgpu_vram_mgr_debug(struct ttm_mem_type_manager *man,
577 struct drm_printer *printer)
578{
579 struct amdgpu_vram_mgr *mgr = man->priv;
580
581 spin_lock(&mgr->lock);
582 drm_mm_print(&mgr->mm, printer);
583 spin_unlock(&mgr->lock);
584
585 drm_printf(printer, "man size:%llu pages, ram usage:%lluMB, vis usage:%lluMB\n",
586 man->size, amdgpu_vram_mgr_usage(man) >> 20,
587 amdgpu_vram_mgr_vis_usage(man) >> 20);
588}
589
590const struct ttm_mem_type_manager_func amdgpu_vram_mgr_func = {
591 .init = amdgpu_vram_mgr_init,
592 .takedown = amdgpu_vram_mgr_fini,
593 .get_node = amdgpu_vram_mgr_new,
594 .put_node = amdgpu_vram_mgr_del,
595 .debug = amdgpu_vram_mgr_debug
596};