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 struct list_head node;
36 struct drm_mm_node mm_node;
37};
38
39static inline struct amdgpu_vram_mgr *
40to_vram_mgr(struct ttm_resource_manager *man)
41{
42 return container_of(man, struct amdgpu_vram_mgr, manager);
43}
44
45static inline struct amdgpu_device *
46to_amdgpu_device(struct amdgpu_vram_mgr *mgr)
47{
48 return container_of(mgr, struct amdgpu_device, mman.vram_mgr);
49}
50
51/**
52 * DOC: mem_info_vram_total
53 *
54 * The amdgpu driver provides a sysfs API for reporting current total VRAM
55 * available on the device
56 * The file mem_info_vram_total is used for this and returns the total
57 * amount of VRAM in bytes
58 */
59static ssize_t amdgpu_mem_info_vram_total_show(struct device *dev,
60 struct device_attribute *attr, char *buf)
61{
62 struct drm_device *ddev = dev_get_drvdata(dev);
63 struct amdgpu_device *adev = drm_to_adev(ddev);
64
65 return sysfs_emit(buf, "%llu\n", adev->gmc.real_vram_size);
66}
67
68/**
69 * DOC: mem_info_vis_vram_total
70 *
71 * The amdgpu driver provides a sysfs API for reporting current total
72 * visible VRAM available on the device
73 * The file mem_info_vis_vram_total is used for this and returns the total
74 * amount of visible VRAM in bytes
75 */
76static ssize_t amdgpu_mem_info_vis_vram_total_show(struct device *dev,
77 struct device_attribute *attr, char *buf)
78{
79 struct drm_device *ddev = dev_get_drvdata(dev);
80 struct amdgpu_device *adev = drm_to_adev(ddev);
81
82 return sysfs_emit(buf, "%llu\n", adev->gmc.visible_vram_size);
83}
84
85/**
86 * DOC: mem_info_vram_used
87 *
88 * The amdgpu driver provides a sysfs API for reporting current total VRAM
89 * available on the device
90 * The file mem_info_vram_used is used for this and returns the total
91 * amount of currently used VRAM in bytes
92 */
93static ssize_t amdgpu_mem_info_vram_used_show(struct device *dev,
94 struct device_attribute *attr,
95 char *buf)
96{
97 struct drm_device *ddev = dev_get_drvdata(dev);
98 struct amdgpu_device *adev = drm_to_adev(ddev);
99 struct ttm_resource_manager *man;
100
101 man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
102 return sysfs_emit(buf, "%llu\n", amdgpu_vram_mgr_usage(man));
103}
104
105/**
106 * DOC: mem_info_vis_vram_used
107 *
108 * The amdgpu driver provides a sysfs API for reporting current total of
109 * used visible VRAM
110 * The file mem_info_vis_vram_used is used for this and returns the total
111 * amount of currently used visible VRAM in bytes
112 */
113static ssize_t amdgpu_mem_info_vis_vram_used_show(struct device *dev,
114 struct device_attribute *attr,
115 char *buf)
116{
117 struct drm_device *ddev = dev_get_drvdata(dev);
118 struct amdgpu_device *adev = drm_to_adev(ddev);
119 struct ttm_resource_manager *man;
120
121 man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
122 return sysfs_emit(buf, "%llu\n", amdgpu_vram_mgr_vis_usage(man));
123}
124
125/**
126 * DOC: mem_info_vram_vendor
127 *
128 * The amdgpu driver provides a sysfs API for reporting the vendor of the
129 * installed VRAM
130 * The file mem_info_vram_vendor is used for this and returns the name of the
131 * vendor.
132 */
133static ssize_t amdgpu_mem_info_vram_vendor(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
140 switch (adev->gmc.vram_vendor) {
141 case SAMSUNG:
142 return sysfs_emit(buf, "samsung\n");
143 case INFINEON:
144 return sysfs_emit(buf, "infineon\n");
145 case ELPIDA:
146 return sysfs_emit(buf, "elpida\n");
147 case ETRON:
148 return sysfs_emit(buf, "etron\n");
149 case NANYA:
150 return sysfs_emit(buf, "nanya\n");
151 case HYNIX:
152 return sysfs_emit(buf, "hynix\n");
153 case MOSEL:
154 return sysfs_emit(buf, "mosel\n");
155 case WINBOND:
156 return sysfs_emit(buf, "winbond\n");
157 case ESMT:
158 return sysfs_emit(buf, "esmt\n");
159 case MICRON:
160 return sysfs_emit(buf, "micron\n");
161 default:
162 return sysfs_emit(buf, "unknown\n");
163 }
164}
165
166static DEVICE_ATTR(mem_info_vram_total, S_IRUGO,
167 amdgpu_mem_info_vram_total_show, NULL);
168static DEVICE_ATTR(mem_info_vis_vram_total, S_IRUGO,
169 amdgpu_mem_info_vis_vram_total_show,NULL);
170static DEVICE_ATTR(mem_info_vram_used, S_IRUGO,
171 amdgpu_mem_info_vram_used_show, NULL);
172static DEVICE_ATTR(mem_info_vis_vram_used, S_IRUGO,
173 amdgpu_mem_info_vis_vram_used_show, NULL);
174static DEVICE_ATTR(mem_info_vram_vendor, S_IRUGO,
175 amdgpu_mem_info_vram_vendor, NULL);
176
177static struct attribute *amdgpu_vram_mgr_attributes[] = {
178 &dev_attr_mem_info_vram_total.attr,
179 &dev_attr_mem_info_vis_vram_total.attr,
180 &dev_attr_mem_info_vram_used.attr,
181 &dev_attr_mem_info_vis_vram_used.attr,
182 &dev_attr_mem_info_vram_vendor.attr,
183 NULL
184};
185
186const struct attribute_group amdgpu_vram_mgr_attr_group = {
187 .attrs = amdgpu_vram_mgr_attributes
188};
189
190/**
191 * amdgpu_vram_mgr_vis_size - Calculate visible node size
192 *
193 * @adev: amdgpu_device pointer
194 * @node: MM node structure
195 *
196 * Calculate how many bytes of the MM node are inside visible VRAM
197 */
198static u64 amdgpu_vram_mgr_vis_size(struct amdgpu_device *adev,
199 struct drm_mm_node *node)
200{
201 uint64_t start = node->start << PAGE_SHIFT;
202 uint64_t end = (node->size + node->start) << PAGE_SHIFT;
203
204 if (start >= adev->gmc.visible_vram_size)
205 return 0;
206
207 return (end > adev->gmc.visible_vram_size ?
208 adev->gmc.visible_vram_size : end) - start;
209}
210
211/**
212 * amdgpu_vram_mgr_bo_visible_size - CPU visible BO size
213 *
214 * @bo: &amdgpu_bo buffer object (must be in VRAM)
215 *
216 * Returns:
217 * How much of the given &amdgpu_bo buffer object lies in CPU visible VRAM.
218 */
219u64 amdgpu_vram_mgr_bo_visible_size(struct amdgpu_bo *bo)
220{
221 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
222 struct ttm_resource *res = bo->tbo.resource;
223 unsigned pages = res->num_pages;
224 struct drm_mm_node *mm;
225 u64 usage;
226
227 if (amdgpu_gmc_vram_full_visible(&adev->gmc))
228 return amdgpu_bo_size(bo);
229
230 if (res->start >= adev->gmc.visible_vram_size >> PAGE_SHIFT)
231 return 0;
232
233 mm = &container_of(res, struct ttm_range_mgr_node, base)->mm_nodes[0];
234 for (usage = 0; pages; pages -= mm->size, mm++)
235 usage += amdgpu_vram_mgr_vis_size(adev, mm);
236
237 return usage;
238}
239
240/* Commit the reservation of VRAM pages */
241static void amdgpu_vram_mgr_do_reserve(struct ttm_resource_manager *man)
242{
243 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
244 struct amdgpu_device *adev = to_amdgpu_device(mgr);
245 struct drm_mm *mm = &mgr->mm;
246 struct amdgpu_vram_reservation *rsv, *temp;
247 uint64_t vis_usage;
248
249 list_for_each_entry_safe(rsv, temp, &mgr->reservations_pending, node) {
250 if (drm_mm_reserve_node(mm, &rsv->mm_node))
251 continue;
252
253 dev_dbg(adev->dev, "Reservation 0x%llx - %lld, Succeeded\n",
254 rsv->mm_node.start, rsv->mm_node.size);
255
256 vis_usage = amdgpu_vram_mgr_vis_size(adev, &rsv->mm_node);
257 atomic64_add(vis_usage, &mgr->vis_usage);
258 atomic64_add(rsv->mm_node.size << PAGE_SHIFT, &mgr->usage);
259 list_move(&rsv->node, &mgr->reserved_pages);
260 }
261}
262
263/**
264 * amdgpu_vram_mgr_reserve_range - Reserve a range from VRAM
265 *
266 * @man: TTM memory type manager
267 * @start: start address of the range in VRAM
268 * @size: size of the range
269 *
270 * Reserve memory from start addess with the specified size in VRAM
271 */
272int amdgpu_vram_mgr_reserve_range(struct ttm_resource_manager *man,
273 uint64_t start, uint64_t size)
274{
275 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
276 struct amdgpu_vram_reservation *rsv;
277
278 rsv = kzalloc(sizeof(*rsv), GFP_KERNEL);
279 if (!rsv)
280 return -ENOMEM;
281
282 INIT_LIST_HEAD(&rsv->node);
283 rsv->mm_node.start = start >> PAGE_SHIFT;
284 rsv->mm_node.size = size >> PAGE_SHIFT;
285
286 spin_lock(&mgr->lock);
287 list_add_tail(&mgr->reservations_pending, &rsv->node);
288 amdgpu_vram_mgr_do_reserve(man);
289 spin_unlock(&mgr->lock);
290
291 return 0;
292}
293
294/**
295 * amdgpu_vram_mgr_query_page_status - query the reservation status
296 *
297 * @man: TTM memory type manager
298 * @start: start address of a page in VRAM
299 *
300 * Returns:
301 * -EBUSY: the page is still hold and in pending list
302 * 0: the page has been reserved
303 * -ENOENT: the input page is not a reservation
304 */
305int amdgpu_vram_mgr_query_page_status(struct ttm_resource_manager *man,
306 uint64_t start)
307{
308 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
309 struct amdgpu_vram_reservation *rsv;
310 int ret;
311
312 spin_lock(&mgr->lock);
313
314 list_for_each_entry(rsv, &mgr->reservations_pending, node) {
315 if ((rsv->mm_node.start <= start) &&
316 (start < (rsv->mm_node.start + rsv->mm_node.size))) {
317 ret = -EBUSY;
318 goto out;
319 }
320 }
321
322 list_for_each_entry(rsv, &mgr->reserved_pages, node) {
323 if ((rsv->mm_node.start <= start) &&
324 (start < (rsv->mm_node.start + rsv->mm_node.size))) {
325 ret = 0;
326 goto out;
327 }
328 }
329
330 ret = -ENOENT;
331out:
332 spin_unlock(&mgr->lock);
333 return ret;
334}
335
336/**
337 * amdgpu_vram_mgr_virt_start - update virtual start address
338 *
339 * @mem: ttm_resource to update
340 * @node: just allocated node
341 *
342 * Calculate a virtual BO start address to easily check if everything is CPU
343 * accessible.
344 */
345static void amdgpu_vram_mgr_virt_start(struct ttm_resource *mem,
346 struct drm_mm_node *node)
347{
348 unsigned long start;
349
350 start = node->start + node->size;
351 if (start > mem->num_pages)
352 start -= mem->num_pages;
353 else
354 start = 0;
355 mem->start = max(mem->start, start);
356}
357
358/**
359 * amdgpu_vram_mgr_new - allocate new ranges
360 *
361 * @man: TTM memory type manager
362 * @tbo: TTM BO we need this range for
363 * @place: placement flags and restrictions
364 * @mem: the resulting mem object
365 *
366 * Allocate VRAM for the given BO.
367 */
368static int amdgpu_vram_mgr_new(struct ttm_resource_manager *man,
369 struct ttm_buffer_object *tbo,
370 const struct ttm_place *place,
371 struct ttm_resource **res)
372{
373 unsigned long lpfn, num_nodes, pages_per_node, pages_left, pages;
374 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
375 struct amdgpu_device *adev = to_amdgpu_device(mgr);
376 uint64_t vis_usage = 0, mem_bytes, max_bytes;
377 struct ttm_range_mgr_node *node;
378 struct drm_mm *mm = &mgr->mm;
379 enum drm_mm_insert_mode mode;
380 unsigned i;
381 int r;
382
383 lpfn = place->lpfn;
384 if (!lpfn)
385 lpfn = man->size;
386
387 max_bytes = adev->gmc.mc_vram_size;
388 if (tbo->type != ttm_bo_type_kernel)
389 max_bytes -= AMDGPU_VM_RESERVED_VRAM;
390
391 /* bail out quickly if there's likely not enough VRAM for this BO */
392 mem_bytes = tbo->base.size;
393 if (atomic64_add_return(mem_bytes, &mgr->usage) > max_bytes) {
394 r = -ENOSPC;
395 goto error_sub;
396 }
397
398 if (place->flags & TTM_PL_FLAG_CONTIGUOUS) {
399 pages_per_node = ~0ul;
400 num_nodes = 1;
401 } else {
402#ifdef CONFIG_TRANSPARENT_HUGEPAGE
403 pages_per_node = HPAGE_PMD_NR;
404#else
405 /* default to 2MB */
406 pages_per_node = 2UL << (20UL - PAGE_SHIFT);
407#endif
408 pages_per_node = max_t(uint32_t, pages_per_node,
409 tbo->page_alignment);
410 num_nodes = DIV_ROUND_UP_ULL(PFN_UP(mem_bytes), pages_per_node);
411 }
412
413 node = kvmalloc(struct_size(node, mm_nodes, num_nodes),
414 GFP_KERNEL | __GFP_ZERO);
415 if (!node) {
416 r = -ENOMEM;
417 goto error_sub;
418 }
419
420 ttm_resource_init(tbo, place, &node->base);
421
422 mode = DRM_MM_INSERT_BEST;
423 if (place->flags & TTM_PL_FLAG_TOPDOWN)
424 mode = DRM_MM_INSERT_HIGH;
425
426 pages_left = node->base.num_pages;
427
428 /* Limit maximum size to 2GB due to SG table limitations */
429 pages = min(pages_left, 2UL << (30 - PAGE_SHIFT));
430
431 i = 0;
432 spin_lock(&mgr->lock);
433 while (pages_left) {
434 uint32_t alignment = tbo->page_alignment;
435
436 if (pages >= pages_per_node)
437 alignment = pages_per_node;
438
439 r = drm_mm_insert_node_in_range(mm, &node->mm_nodes[i], pages,
440 alignment, 0, place->fpfn,
441 lpfn, mode);
442 if (unlikely(r)) {
443 if (pages > pages_per_node) {
444 if (is_power_of_2(pages))
445 pages = pages / 2;
446 else
447 pages = rounddown_pow_of_two(pages);
448 continue;
449 }
450 goto error_free;
451 }
452
453 vis_usage += amdgpu_vram_mgr_vis_size(adev, &node->mm_nodes[i]);
454 amdgpu_vram_mgr_virt_start(&node->base, &node->mm_nodes[i]);
455 pages_left -= pages;
456 ++i;
457
458 if (pages > pages_left)
459 pages = pages_left;
460 }
461 spin_unlock(&mgr->lock);
462
463 if (i == 1)
464 node->base.placement |= TTM_PL_FLAG_CONTIGUOUS;
465
466 if (adev->gmc.xgmi.connected_to_cpu)
467 node->base.bus.caching = ttm_cached;
468 else
469 node->base.bus.caching = ttm_write_combined;
470
471 atomic64_add(vis_usage, &mgr->vis_usage);
472 *res = &node->base;
473 return 0;
474
475error_free:
476 while (i--)
477 drm_mm_remove_node(&node->mm_nodes[i]);
478 spin_unlock(&mgr->lock);
479 kvfree(node);
480
481error_sub:
482 atomic64_sub(mem_bytes, &mgr->usage);
483 return r;
484}
485
486/**
487 * amdgpu_vram_mgr_del - free ranges
488 *
489 * @man: TTM memory type manager
490 * @mem: TTM memory object
491 *
492 * Free the allocated VRAM again.
493 */
494static void amdgpu_vram_mgr_del(struct ttm_resource_manager *man,
495 struct ttm_resource *res)
496{
497 struct ttm_range_mgr_node *node = to_ttm_range_mgr_node(res);
498 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
499 struct amdgpu_device *adev = to_amdgpu_device(mgr);
500 uint64_t usage = 0, vis_usage = 0;
501 unsigned i, pages;
502
503 spin_lock(&mgr->lock);
504 for (i = 0, pages = res->num_pages; pages;
505 pages -= node->mm_nodes[i].size, ++i) {
506 struct drm_mm_node *mm = &node->mm_nodes[i];
507
508 drm_mm_remove_node(mm);
509 usage += mm->size << PAGE_SHIFT;
510 vis_usage += amdgpu_vram_mgr_vis_size(adev, mm);
511 }
512 amdgpu_vram_mgr_do_reserve(man);
513 spin_unlock(&mgr->lock);
514
515 atomic64_sub(usage, &mgr->usage);
516 atomic64_sub(vis_usage, &mgr->vis_usage);
517
518 kvfree(node);
519}
520
521/**
522 * amdgpu_vram_mgr_alloc_sgt - allocate and fill a sg table
523 *
524 * @adev: amdgpu device pointer
525 * @mem: TTM memory object
526 * @offset: byte offset from the base of VRAM BO
527 * @length: number of bytes to export in sg_table
528 * @dev: the other device
529 * @dir: dma direction
530 * @sgt: resulting sg table
531 *
532 * Allocate and fill a sg table from a VRAM allocation.
533 */
534int amdgpu_vram_mgr_alloc_sgt(struct amdgpu_device *adev,
535 struct ttm_resource *res,
536 u64 offset, u64 length,
537 struct device *dev,
538 enum dma_data_direction dir,
539 struct sg_table **sgt)
540{
541 struct amdgpu_res_cursor cursor;
542 struct scatterlist *sg;
543 int num_entries = 0;
544 int i, r;
545
546 *sgt = kmalloc(sizeof(**sgt), GFP_KERNEL);
547 if (!*sgt)
548 return -ENOMEM;
549
550 /* Determine the number of DRM_MM nodes to export */
551 amdgpu_res_first(res, offset, length, &cursor);
552 while (cursor.remaining) {
553 num_entries++;
554 amdgpu_res_next(&cursor, cursor.size);
555 }
556
557 r = sg_alloc_table(*sgt, num_entries, GFP_KERNEL);
558 if (r)
559 goto error_free;
560
561 /* Initialize scatterlist nodes of sg_table */
562 for_each_sgtable_sg((*sgt), sg, i)
563 sg->length = 0;
564
565 /*
566 * Walk down DRM_MM nodes to populate scatterlist nodes
567 * @note: Use iterator api to get first the DRM_MM node
568 * and the number of bytes from it. Access the following
569 * DRM_MM node(s) if more buffer needs to exported
570 */
571 amdgpu_res_first(res, offset, length, &cursor);
572 for_each_sgtable_sg((*sgt), sg, i) {
573 phys_addr_t phys = cursor.start + adev->gmc.aper_base;
574 size_t size = cursor.size;
575 dma_addr_t addr;
576
577 addr = dma_map_resource(dev, phys, size, dir,
578 DMA_ATTR_SKIP_CPU_SYNC);
579 r = dma_mapping_error(dev, addr);
580 if (r)
581 goto error_unmap;
582
583 sg_set_page(sg, NULL, size, 0);
584 sg_dma_address(sg) = addr;
585 sg_dma_len(sg) = size;
586
587 amdgpu_res_next(&cursor, cursor.size);
588 }
589
590 return 0;
591
592error_unmap:
593 for_each_sgtable_sg((*sgt), sg, i) {
594 if (!sg->length)
595 continue;
596
597 dma_unmap_resource(dev, sg->dma_address,
598 sg->length, dir,
599 DMA_ATTR_SKIP_CPU_SYNC);
600 }
601 sg_free_table(*sgt);
602
603error_free:
604 kfree(*sgt);
605 return r;
606}
607
608/**
609 * amdgpu_vram_mgr_free_sgt - allocate and fill a sg table
610 *
611 * @dev: device pointer
612 * @dir: data direction of resource to unmap
613 * @sgt: sg table to free
614 *
615 * Free a previously allocate sg table.
616 */
617void amdgpu_vram_mgr_free_sgt(struct device *dev,
618 enum dma_data_direction dir,
619 struct sg_table *sgt)
620{
621 struct scatterlist *sg;
622 int i;
623
624 for_each_sgtable_sg(sgt, sg, i)
625 dma_unmap_resource(dev, sg->dma_address,
626 sg->length, dir,
627 DMA_ATTR_SKIP_CPU_SYNC);
628 sg_free_table(sgt);
629 kfree(sgt);
630}
631
632/**
633 * amdgpu_vram_mgr_usage - how many bytes are used in this domain
634 *
635 * @man: TTM memory type manager
636 *
637 * Returns how many bytes are used in this domain.
638 */
639uint64_t amdgpu_vram_mgr_usage(struct ttm_resource_manager *man)
640{
641 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
642
643 return atomic64_read(&mgr->usage);
644}
645
646/**
647 * amdgpu_vram_mgr_vis_usage - how many bytes are used in the visible part
648 *
649 * @man: TTM memory type manager
650 *
651 * Returns how many bytes are used in the visible part of VRAM
652 */
653uint64_t amdgpu_vram_mgr_vis_usage(struct ttm_resource_manager *man)
654{
655 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
656
657 return atomic64_read(&mgr->vis_usage);
658}
659
660/**
661 * amdgpu_vram_mgr_debug - dump VRAM table
662 *
663 * @man: TTM memory type manager
664 * @printer: DRM printer to use
665 *
666 * Dump the table content using printk.
667 */
668static void amdgpu_vram_mgr_debug(struct ttm_resource_manager *man,
669 struct drm_printer *printer)
670{
671 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
672
673 spin_lock(&mgr->lock);
674 drm_mm_print(&mgr->mm, printer);
675 spin_unlock(&mgr->lock);
676
677 drm_printf(printer, "man size:%llu pages, ram usage:%lluMB, vis usage:%lluMB\n",
678 man->size, amdgpu_vram_mgr_usage(man) >> 20,
679 amdgpu_vram_mgr_vis_usage(man) >> 20);
680}
681
682static const struct ttm_resource_manager_func amdgpu_vram_mgr_func = {
683 .alloc = amdgpu_vram_mgr_new,
684 .free = amdgpu_vram_mgr_del,
685 .debug = amdgpu_vram_mgr_debug
686};
687
688/**
689 * amdgpu_vram_mgr_init - init VRAM manager and DRM MM
690 *
691 * @adev: amdgpu_device pointer
692 *
693 * Allocate and initialize the VRAM manager.
694 */
695int amdgpu_vram_mgr_init(struct amdgpu_device *adev)
696{
697 struct amdgpu_vram_mgr *mgr = &adev->mman.vram_mgr;
698 struct ttm_resource_manager *man = &mgr->manager;
699
700 ttm_resource_manager_init(man, adev->gmc.real_vram_size >> PAGE_SHIFT);
701
702 man->func = &amdgpu_vram_mgr_func;
703
704 drm_mm_init(&mgr->mm, 0, man->size);
705 spin_lock_init(&mgr->lock);
706 INIT_LIST_HEAD(&mgr->reservations_pending);
707 INIT_LIST_HEAD(&mgr->reserved_pages);
708
709 ttm_set_driver_manager(&adev->mman.bdev, TTM_PL_VRAM, &mgr->manager);
710 ttm_resource_manager_set_used(man, true);
711 return 0;
712}
713
714/**
715 * amdgpu_vram_mgr_fini - free and destroy VRAM manager
716 *
717 * @adev: amdgpu_device pointer
718 *
719 * Destroy and free the VRAM manager, returns -EBUSY if ranges are still
720 * allocated inside it.
721 */
722void amdgpu_vram_mgr_fini(struct amdgpu_device *adev)
723{
724 struct amdgpu_vram_mgr *mgr = &adev->mman.vram_mgr;
725 struct ttm_resource_manager *man = &mgr->manager;
726 int ret;
727 struct amdgpu_vram_reservation *rsv, *temp;
728
729 ttm_resource_manager_set_used(man, false);
730
731 ret = ttm_resource_manager_evict_all(&adev->mman.bdev, man);
732 if (ret)
733 return;
734
735 spin_lock(&mgr->lock);
736 list_for_each_entry_safe(rsv, temp, &mgr->reservations_pending, node)
737 kfree(rsv);
738
739 list_for_each_entry_safe(rsv, temp, &mgr->reserved_pages, node) {
740 drm_mm_remove_node(&rsv->mm_node);
741 kfree(rsv);
742 }
743 drm_mm_takedown(&mgr->mm);
744 spin_unlock(&mgr->lock);
745
746 ttm_resource_manager_cleanup(man);
747 ttm_set_driver_manager(&adev->mman.bdev, TTM_PL_VRAM, NULL);
748}
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 "amdgpu.h"
26
27struct amdgpu_vram_mgr {
28 struct drm_mm mm;
29 spinlock_t lock;
30 atomic64_t usage;
31 atomic64_t vis_usage;
32};
33
34/**
35 * DOC: mem_info_vram_total
36 *
37 * The amdgpu driver provides a sysfs API for reporting current total VRAM
38 * available on the device
39 * The file mem_info_vram_total is used for this and returns the total
40 * amount of VRAM in bytes
41 */
42static ssize_t amdgpu_mem_info_vram_total_show(struct device *dev,
43 struct device_attribute *attr, char *buf)
44{
45 struct drm_device *ddev = dev_get_drvdata(dev);
46 struct amdgpu_device *adev = ddev->dev_private;
47
48 return snprintf(buf, PAGE_SIZE, "%llu\n", adev->gmc.real_vram_size);
49}
50
51/**
52 * DOC: mem_info_vis_vram_total
53 *
54 * The amdgpu driver provides a sysfs API for reporting current total
55 * visible VRAM available on the device
56 * The file mem_info_vis_vram_total is used for this and returns the total
57 * amount of visible VRAM in bytes
58 */
59static ssize_t amdgpu_mem_info_vis_vram_total_show(struct device *dev,
60 struct device_attribute *attr, char *buf)
61{
62 struct drm_device *ddev = dev_get_drvdata(dev);
63 struct amdgpu_device *adev = ddev->dev_private;
64
65 return snprintf(buf, PAGE_SIZE, "%llu\n", adev->gmc.visible_vram_size);
66}
67
68/**
69 * DOC: mem_info_vram_used
70 *
71 * The amdgpu driver provides a sysfs API for reporting current total VRAM
72 * available on the device
73 * The file mem_info_vram_used is used for this and returns the total
74 * amount of currently used VRAM in bytes
75 */
76static ssize_t amdgpu_mem_info_vram_used_show(struct device *dev,
77 struct device_attribute *attr, char *buf)
78{
79 struct drm_device *ddev = dev_get_drvdata(dev);
80 struct amdgpu_device *adev = ddev->dev_private;
81
82 return snprintf(buf, PAGE_SIZE, "%llu\n",
83 amdgpu_vram_mgr_usage(&adev->mman.bdev.man[TTM_PL_VRAM]));
84}
85
86/**
87 * DOC: mem_info_vis_vram_used
88 *
89 * The amdgpu driver provides a sysfs API for reporting current total of
90 * used visible VRAM
91 * The file mem_info_vis_vram_used is used for this and returns the total
92 * amount of currently used visible VRAM in bytes
93 */
94static ssize_t amdgpu_mem_info_vis_vram_used_show(struct device *dev,
95 struct device_attribute *attr, char *buf)
96{
97 struct drm_device *ddev = dev_get_drvdata(dev);
98 struct amdgpu_device *adev = ddev->dev_private;
99
100 return snprintf(buf, PAGE_SIZE, "%llu\n",
101 amdgpu_vram_mgr_vis_usage(&adev->mman.bdev.man[TTM_PL_VRAM]));
102}
103
104static DEVICE_ATTR(mem_info_vram_total, S_IRUGO,
105 amdgpu_mem_info_vram_total_show, NULL);
106static DEVICE_ATTR(mem_info_vis_vram_total, S_IRUGO,
107 amdgpu_mem_info_vis_vram_total_show,NULL);
108static DEVICE_ATTR(mem_info_vram_used, S_IRUGO,
109 amdgpu_mem_info_vram_used_show, NULL);
110static DEVICE_ATTR(mem_info_vis_vram_used, S_IRUGO,
111 amdgpu_mem_info_vis_vram_used_show, NULL);
112
113/**
114 * amdgpu_vram_mgr_init - init VRAM manager and DRM MM
115 *
116 * @man: TTM memory type manager
117 * @p_size: maximum size of VRAM
118 *
119 * Allocate and initialize the VRAM manager.
120 */
121static int amdgpu_vram_mgr_init(struct ttm_mem_type_manager *man,
122 unsigned long p_size)
123{
124 struct amdgpu_device *adev = amdgpu_ttm_adev(man->bdev);
125 struct amdgpu_vram_mgr *mgr;
126 int ret;
127
128 mgr = kzalloc(sizeof(*mgr), GFP_KERNEL);
129 if (!mgr)
130 return -ENOMEM;
131
132 drm_mm_init(&mgr->mm, 0, p_size);
133 spin_lock_init(&mgr->lock);
134 man->priv = mgr;
135
136 /* Add the two VRAM-related sysfs files */
137 ret = device_create_file(adev->dev, &dev_attr_mem_info_vram_total);
138 if (ret) {
139 DRM_ERROR("Failed to create device file mem_info_vram_total\n");
140 return ret;
141 }
142 ret = device_create_file(adev->dev, &dev_attr_mem_info_vis_vram_total);
143 if (ret) {
144 DRM_ERROR("Failed to create device file mem_info_vis_vram_total\n");
145 return ret;
146 }
147 ret = device_create_file(adev->dev, &dev_attr_mem_info_vram_used);
148 if (ret) {
149 DRM_ERROR("Failed to create device file mem_info_vram_used\n");
150 return ret;
151 }
152 ret = device_create_file(adev->dev, &dev_attr_mem_info_vis_vram_used);
153 if (ret) {
154 DRM_ERROR("Failed to create device file mem_info_vis_vram_used\n");
155 return ret;
156 }
157
158 return 0;
159}
160
161/**
162 * amdgpu_vram_mgr_fini - free and destroy VRAM manager
163 *
164 * @man: TTM memory type manager
165 *
166 * Destroy and free the VRAM manager, returns -EBUSY if ranges are still
167 * allocated inside it.
168 */
169static int amdgpu_vram_mgr_fini(struct ttm_mem_type_manager *man)
170{
171 struct amdgpu_device *adev = amdgpu_ttm_adev(man->bdev);
172 struct amdgpu_vram_mgr *mgr = man->priv;
173
174 spin_lock(&mgr->lock);
175 drm_mm_takedown(&mgr->mm);
176 spin_unlock(&mgr->lock);
177 kfree(mgr);
178 man->priv = NULL;
179 device_remove_file(adev->dev, &dev_attr_mem_info_vram_total);
180 device_remove_file(adev->dev, &dev_attr_mem_info_vis_vram_total);
181 device_remove_file(adev->dev, &dev_attr_mem_info_vram_used);
182 device_remove_file(adev->dev, &dev_attr_mem_info_vis_vram_used);
183 return 0;
184}
185
186/**
187 * amdgpu_vram_mgr_vis_size - Calculate visible node size
188 *
189 * @adev: amdgpu device structure
190 * @node: MM node structure
191 *
192 * Calculate how many bytes of the MM node are inside visible VRAM
193 */
194static u64 amdgpu_vram_mgr_vis_size(struct amdgpu_device *adev,
195 struct drm_mm_node *node)
196{
197 uint64_t start = node->start << PAGE_SHIFT;
198 uint64_t end = (node->size + node->start) << PAGE_SHIFT;
199
200 if (start >= adev->gmc.visible_vram_size)
201 return 0;
202
203 return (end > adev->gmc.visible_vram_size ?
204 adev->gmc.visible_vram_size : end) - start;
205}
206
207/**
208 * amdgpu_vram_mgr_bo_visible_size - CPU visible BO size
209 *
210 * @bo: &amdgpu_bo buffer object (must be in VRAM)
211 *
212 * Returns:
213 * How much of the given &amdgpu_bo buffer object lies in CPU visible VRAM.
214 */
215u64 amdgpu_vram_mgr_bo_visible_size(struct amdgpu_bo *bo)
216{
217 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
218 struct ttm_mem_reg *mem = &bo->tbo.mem;
219 struct drm_mm_node *nodes = mem->mm_node;
220 unsigned pages = mem->num_pages;
221 u64 usage;
222
223 if (amdgpu_gmc_vram_full_visible(&adev->gmc))
224 return amdgpu_bo_size(bo);
225
226 if (mem->start >= adev->gmc.visible_vram_size >> PAGE_SHIFT)
227 return 0;
228
229 for (usage = 0; nodes && pages; pages -= nodes->size, nodes++)
230 usage += amdgpu_vram_mgr_vis_size(adev, nodes);
231
232 return usage;
233}
234
235/**
236 * amdgpu_vram_mgr_virt_start - update virtual start address
237 *
238 * @mem: ttm_mem_reg to update
239 * @node: just allocated node
240 *
241 * Calculate a virtual BO start address to easily check if everything is CPU
242 * accessible.
243 */
244static void amdgpu_vram_mgr_virt_start(struct ttm_mem_reg *mem,
245 struct drm_mm_node *node)
246{
247 unsigned long start;
248
249 start = node->start + node->size;
250 if (start > mem->num_pages)
251 start -= mem->num_pages;
252 else
253 start = 0;
254 mem->start = max(mem->start, start);
255}
256
257/**
258 * amdgpu_vram_mgr_new - allocate new ranges
259 *
260 * @man: TTM memory type manager
261 * @tbo: TTM BO we need this range for
262 * @place: placement flags and restrictions
263 * @mem: the resulting mem object
264 *
265 * Allocate VRAM for the given BO.
266 */
267static int amdgpu_vram_mgr_new(struct ttm_mem_type_manager *man,
268 struct ttm_buffer_object *tbo,
269 const struct ttm_place *place,
270 struct ttm_mem_reg *mem)
271{
272 struct amdgpu_device *adev = amdgpu_ttm_adev(man->bdev);
273 struct amdgpu_vram_mgr *mgr = man->priv;
274 struct drm_mm *mm = &mgr->mm;
275 struct drm_mm_node *nodes;
276 enum drm_mm_insert_mode mode;
277 unsigned long lpfn, num_nodes, pages_per_node, pages_left;
278 uint64_t vis_usage = 0, mem_bytes;
279 unsigned i;
280 int r;
281
282 lpfn = place->lpfn;
283 if (!lpfn)
284 lpfn = man->size;
285
286 /* bail out quickly if there's likely not enough VRAM for this BO */
287 mem_bytes = (u64)mem->num_pages << PAGE_SHIFT;
288 if (atomic64_add_return(mem_bytes, &mgr->usage) > adev->gmc.mc_vram_size) {
289 atomic64_sub(mem_bytes, &mgr->usage);
290 mem->mm_node = NULL;
291 return 0;
292 }
293
294 if (place->flags & TTM_PL_FLAG_CONTIGUOUS) {
295 pages_per_node = ~0ul;
296 num_nodes = 1;
297 } else {
298#ifdef CONFIG_TRANSPARENT_HUGEPAGE
299 pages_per_node = HPAGE_PMD_NR;
300#else
301 /* default to 2MB */
302 pages_per_node = (2UL << (20UL - PAGE_SHIFT));
303#endif
304 pages_per_node = max((uint32_t)pages_per_node, mem->page_alignment);
305 num_nodes = DIV_ROUND_UP(mem->num_pages, pages_per_node);
306 }
307
308 nodes = kvmalloc_array((uint32_t)num_nodes, sizeof(*nodes),
309 GFP_KERNEL | __GFP_ZERO);
310 if (!nodes) {
311 atomic64_sub(mem_bytes, &mgr->usage);
312 return -ENOMEM;
313 }
314
315 mode = DRM_MM_INSERT_BEST;
316 if (place->flags & TTM_PL_FLAG_TOPDOWN)
317 mode = DRM_MM_INSERT_HIGH;
318
319 mem->start = 0;
320 pages_left = mem->num_pages;
321
322 spin_lock(&mgr->lock);
323 for (i = 0; pages_left >= pages_per_node; ++i) {
324 unsigned long pages = rounddown_pow_of_two(pages_left);
325
326 r = drm_mm_insert_node_in_range(mm, &nodes[i], pages,
327 pages_per_node, 0,
328 place->fpfn, lpfn,
329 mode);
330 if (unlikely(r))
331 break;
332
333 vis_usage += amdgpu_vram_mgr_vis_size(adev, &nodes[i]);
334 amdgpu_vram_mgr_virt_start(mem, &nodes[i]);
335 pages_left -= pages;
336 }
337
338 for (; pages_left; ++i) {
339 unsigned long pages = min(pages_left, pages_per_node);
340 uint32_t alignment = mem->page_alignment;
341
342 if (pages == pages_per_node)
343 alignment = pages_per_node;
344
345 r = drm_mm_insert_node_in_range(mm, &nodes[i],
346 pages, alignment, 0,
347 place->fpfn, lpfn,
348 mode);
349 if (unlikely(r))
350 goto error;
351
352 vis_usage += amdgpu_vram_mgr_vis_size(adev, &nodes[i]);
353 amdgpu_vram_mgr_virt_start(mem, &nodes[i]);
354 pages_left -= pages;
355 }
356 spin_unlock(&mgr->lock);
357
358 atomic64_add(vis_usage, &mgr->vis_usage);
359
360 mem->mm_node = nodes;
361
362 return 0;
363
364error:
365 while (i--)
366 drm_mm_remove_node(&nodes[i]);
367 spin_unlock(&mgr->lock);
368 atomic64_sub(mem->num_pages << PAGE_SHIFT, &mgr->usage);
369
370 kvfree(nodes);
371 return r == -ENOSPC ? 0 : r;
372}
373
374/**
375 * amdgpu_vram_mgr_del - free ranges
376 *
377 * @man: TTM memory type manager
378 * @tbo: TTM BO we need this range for
379 * @place: placement flags and restrictions
380 * @mem: TTM memory object
381 *
382 * Free the allocated VRAM again.
383 */
384static void amdgpu_vram_mgr_del(struct ttm_mem_type_manager *man,
385 struct ttm_mem_reg *mem)
386{
387 struct amdgpu_device *adev = amdgpu_ttm_adev(man->bdev);
388 struct amdgpu_vram_mgr *mgr = man->priv;
389 struct drm_mm_node *nodes = mem->mm_node;
390 uint64_t usage = 0, vis_usage = 0;
391 unsigned pages = mem->num_pages;
392
393 if (!mem->mm_node)
394 return;
395
396 spin_lock(&mgr->lock);
397 while (pages) {
398 pages -= nodes->size;
399 drm_mm_remove_node(nodes);
400 usage += nodes->size << PAGE_SHIFT;
401 vis_usage += amdgpu_vram_mgr_vis_size(adev, nodes);
402 ++nodes;
403 }
404 spin_unlock(&mgr->lock);
405
406 atomic64_sub(usage, &mgr->usage);
407 atomic64_sub(vis_usage, &mgr->vis_usage);
408
409 kvfree(mem->mm_node);
410 mem->mm_node = NULL;
411}
412
413/**
414 * amdgpu_vram_mgr_usage - how many bytes are used in this domain
415 *
416 * @man: TTM memory type manager
417 *
418 * Returns how many bytes are used in this domain.
419 */
420uint64_t amdgpu_vram_mgr_usage(struct ttm_mem_type_manager *man)
421{
422 struct amdgpu_vram_mgr *mgr = man->priv;
423
424 return atomic64_read(&mgr->usage);
425}
426
427/**
428 * amdgpu_vram_mgr_vis_usage - how many bytes are used in the visible part
429 *
430 * @man: TTM memory type manager
431 *
432 * Returns how many bytes are used in the visible part of VRAM
433 */
434uint64_t amdgpu_vram_mgr_vis_usage(struct ttm_mem_type_manager *man)
435{
436 struct amdgpu_vram_mgr *mgr = man->priv;
437
438 return atomic64_read(&mgr->vis_usage);
439}
440
441/**
442 * amdgpu_vram_mgr_debug - dump VRAM table
443 *
444 * @man: TTM memory type manager
445 * @printer: DRM printer to use
446 *
447 * Dump the table content using printk.
448 */
449static void amdgpu_vram_mgr_debug(struct ttm_mem_type_manager *man,
450 struct drm_printer *printer)
451{
452 struct amdgpu_vram_mgr *mgr = man->priv;
453
454 spin_lock(&mgr->lock);
455 drm_mm_print(&mgr->mm, printer);
456 spin_unlock(&mgr->lock);
457
458 drm_printf(printer, "man size:%llu pages, ram usage:%lluMB, vis usage:%lluMB\n",
459 man->size, amdgpu_vram_mgr_usage(man) >> 20,
460 amdgpu_vram_mgr_vis_usage(man) >> 20);
461}
462
463const struct ttm_mem_type_manager_func amdgpu_vram_mgr_func = {
464 .init = amdgpu_vram_mgr_init,
465 .takedown = amdgpu_vram_mgr_fini,
466 .get_node = amdgpu_vram_mgr_new,
467 .put_node = amdgpu_vram_mgr_del,
468 .debug = amdgpu_vram_mgr_debug
469};