Loading...
Note: File does not exist in v6.13.7.
1/* SPDX-License-Identifier: GPL-2.0 OR MIT */
2/**************************************************************************
3 *
4 * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA
5 * All Rights Reserved.
6 *
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the
9 * "Software"), to deal in the Software without restriction, including
10 * without limitation the rights to use, copy, modify, merge, publish,
11 * distribute, sub license, and/or sell copies of the Software, and to
12 * permit persons to whom the Software is furnished to do so, subject to
13 * the following conditions:
14 *
15 * The above copyright notice and this permission notice (including the
16 * next paragraph) shall be included in all copies or substantial portions
17 * of the Software.
18 *
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
22 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
23 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
24 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
25 * USE OR OTHER DEALINGS IN THE SOFTWARE.
26 *
27 **************************************************************************/
28
29#define pr_fmt(fmt) "[TTM] " fmt
30
31#include <drm/ttm/ttm_memory.h>
32#include <drm/ttm/ttm_module.h>
33#include <drm/ttm/ttm_page_alloc.h>
34#include <linux/spinlock.h>
35#include <linux/sched.h>
36#include <linux/wait.h>
37#include <linux/mm.h>
38#include <linux/module.h>
39#include <linux/slab.h>
40#include <linux/swap.h>
41
42#define TTM_MEMORY_ALLOC_RETRIES 4
43
44struct ttm_mem_global ttm_mem_glob;
45EXPORT_SYMBOL(ttm_mem_glob);
46
47struct ttm_mem_zone {
48 struct kobject kobj;
49 struct ttm_mem_global *glob;
50 const char *name;
51 uint64_t zone_mem;
52 uint64_t emer_mem;
53 uint64_t max_mem;
54 uint64_t swap_limit;
55 uint64_t used_mem;
56};
57
58static struct attribute ttm_mem_sys = {
59 .name = "zone_memory",
60 .mode = S_IRUGO
61};
62static struct attribute ttm_mem_emer = {
63 .name = "emergency_memory",
64 .mode = S_IRUGO | S_IWUSR
65};
66static struct attribute ttm_mem_max = {
67 .name = "available_memory",
68 .mode = S_IRUGO | S_IWUSR
69};
70static struct attribute ttm_mem_swap = {
71 .name = "swap_limit",
72 .mode = S_IRUGO | S_IWUSR
73};
74static struct attribute ttm_mem_used = {
75 .name = "used_memory",
76 .mode = S_IRUGO
77};
78
79static void ttm_mem_zone_kobj_release(struct kobject *kobj)
80{
81 struct ttm_mem_zone *zone =
82 container_of(kobj, struct ttm_mem_zone, kobj);
83
84 pr_info("Zone %7s: Used memory at exit: %llu KiB\n",
85 zone->name, (unsigned long long)zone->used_mem >> 10);
86 kfree(zone);
87}
88
89static ssize_t ttm_mem_zone_show(struct kobject *kobj,
90 struct attribute *attr,
91 char *buffer)
92{
93 struct ttm_mem_zone *zone =
94 container_of(kobj, struct ttm_mem_zone, kobj);
95 uint64_t val = 0;
96
97 spin_lock(&zone->glob->lock);
98 if (attr == &ttm_mem_sys)
99 val = zone->zone_mem;
100 else if (attr == &ttm_mem_emer)
101 val = zone->emer_mem;
102 else if (attr == &ttm_mem_max)
103 val = zone->max_mem;
104 else if (attr == &ttm_mem_swap)
105 val = zone->swap_limit;
106 else if (attr == &ttm_mem_used)
107 val = zone->used_mem;
108 spin_unlock(&zone->glob->lock);
109
110 return snprintf(buffer, PAGE_SIZE, "%llu\n",
111 (unsigned long long) val >> 10);
112}
113
114static void ttm_check_swapping(struct ttm_mem_global *glob);
115
116static ssize_t ttm_mem_zone_store(struct kobject *kobj,
117 struct attribute *attr,
118 const char *buffer,
119 size_t size)
120{
121 struct ttm_mem_zone *zone =
122 container_of(kobj, struct ttm_mem_zone, kobj);
123 int chars;
124 unsigned long val;
125 uint64_t val64;
126
127 chars = sscanf(buffer, "%lu", &val);
128 if (chars == 0)
129 return size;
130
131 val64 = val;
132 val64 <<= 10;
133
134 spin_lock(&zone->glob->lock);
135 if (val64 > zone->zone_mem)
136 val64 = zone->zone_mem;
137 if (attr == &ttm_mem_emer) {
138 zone->emer_mem = val64;
139 if (zone->max_mem > val64)
140 zone->max_mem = val64;
141 } else if (attr == &ttm_mem_max) {
142 zone->max_mem = val64;
143 if (zone->emer_mem < val64)
144 zone->emer_mem = val64;
145 } else if (attr == &ttm_mem_swap)
146 zone->swap_limit = val64;
147 spin_unlock(&zone->glob->lock);
148
149 ttm_check_swapping(zone->glob);
150
151 return size;
152}
153
154static struct attribute *ttm_mem_zone_attrs[] = {
155 &ttm_mem_sys,
156 &ttm_mem_emer,
157 &ttm_mem_max,
158 &ttm_mem_swap,
159 &ttm_mem_used,
160 NULL
161};
162
163static const struct sysfs_ops ttm_mem_zone_ops = {
164 .show = &ttm_mem_zone_show,
165 .store = &ttm_mem_zone_store
166};
167
168static struct kobj_type ttm_mem_zone_kobj_type = {
169 .release = &ttm_mem_zone_kobj_release,
170 .sysfs_ops = &ttm_mem_zone_ops,
171 .default_attrs = ttm_mem_zone_attrs,
172};
173
174static struct attribute ttm_mem_global_lower_mem_limit = {
175 .name = "lower_mem_limit",
176 .mode = S_IRUGO | S_IWUSR
177};
178
179static ssize_t ttm_mem_global_show(struct kobject *kobj,
180 struct attribute *attr,
181 char *buffer)
182{
183 struct ttm_mem_global *glob =
184 container_of(kobj, struct ttm_mem_global, kobj);
185 uint64_t val = 0;
186
187 spin_lock(&glob->lock);
188 val = glob->lower_mem_limit;
189 spin_unlock(&glob->lock);
190 /* convert from number of pages to KB */
191 val <<= (PAGE_SHIFT - 10);
192 return snprintf(buffer, PAGE_SIZE, "%llu\n",
193 (unsigned long long) val);
194}
195
196static ssize_t ttm_mem_global_store(struct kobject *kobj,
197 struct attribute *attr,
198 const char *buffer,
199 size_t size)
200{
201 int chars;
202 uint64_t val64;
203 unsigned long val;
204 struct ttm_mem_global *glob =
205 container_of(kobj, struct ttm_mem_global, kobj);
206
207 chars = sscanf(buffer, "%lu", &val);
208 if (chars == 0)
209 return size;
210
211 val64 = val;
212 /* convert from KB to number of pages */
213 val64 >>= (PAGE_SHIFT - 10);
214
215 spin_lock(&glob->lock);
216 glob->lower_mem_limit = val64;
217 spin_unlock(&glob->lock);
218
219 return size;
220}
221
222static struct attribute *ttm_mem_global_attrs[] = {
223 &ttm_mem_global_lower_mem_limit,
224 NULL
225};
226
227static const struct sysfs_ops ttm_mem_global_ops = {
228 .show = &ttm_mem_global_show,
229 .store = &ttm_mem_global_store,
230};
231
232static struct kobj_type ttm_mem_glob_kobj_type = {
233 .sysfs_ops = &ttm_mem_global_ops,
234 .default_attrs = ttm_mem_global_attrs,
235};
236
237static bool ttm_zones_above_swap_target(struct ttm_mem_global *glob,
238 bool from_wq, uint64_t extra)
239{
240 unsigned int i;
241 struct ttm_mem_zone *zone;
242 uint64_t target;
243
244 for (i = 0; i < glob->num_zones; ++i) {
245 zone = glob->zones[i];
246
247 if (from_wq)
248 target = zone->swap_limit;
249 else if (capable(CAP_SYS_ADMIN))
250 target = zone->emer_mem;
251 else
252 target = zone->max_mem;
253
254 target = (extra > target) ? 0ULL : target;
255
256 if (zone->used_mem > target)
257 return true;
258 }
259 return false;
260}
261
262/**
263 * At this point we only support a single shrink callback.
264 * Extend this if needed, perhaps using a linked list of callbacks.
265 * Note that this function is reentrant:
266 * many threads may try to swap out at any given time.
267 */
268
269static void ttm_shrink(struct ttm_mem_global *glob, bool from_wq,
270 uint64_t extra, struct ttm_operation_ctx *ctx)
271{
272 int ret;
273
274 spin_lock(&glob->lock);
275
276 while (ttm_zones_above_swap_target(glob, from_wq, extra)) {
277 spin_unlock(&glob->lock);
278 ret = ttm_bo_swapout(glob->bo_glob, ctx);
279 spin_lock(&glob->lock);
280 if (unlikely(ret != 0))
281 break;
282 }
283
284 spin_unlock(&glob->lock);
285}
286
287static void ttm_shrink_work(struct work_struct *work)
288{
289 struct ttm_operation_ctx ctx = {
290 .interruptible = false,
291 .no_wait_gpu = false
292 };
293 struct ttm_mem_global *glob =
294 container_of(work, struct ttm_mem_global, work);
295
296 ttm_shrink(glob, true, 0ULL, &ctx);
297}
298
299static int ttm_mem_init_kernel_zone(struct ttm_mem_global *glob,
300 const struct sysinfo *si)
301{
302 struct ttm_mem_zone *zone = kzalloc(sizeof(*zone), GFP_KERNEL);
303 uint64_t mem;
304 int ret;
305
306 if (unlikely(!zone))
307 return -ENOMEM;
308
309 mem = si->totalram - si->totalhigh;
310 mem *= si->mem_unit;
311
312 zone->name = "kernel";
313 zone->zone_mem = mem;
314 zone->max_mem = mem >> 1;
315 zone->emer_mem = (mem >> 1) + (mem >> 2);
316 zone->swap_limit = zone->max_mem - (mem >> 3);
317 zone->used_mem = 0;
318 zone->glob = glob;
319 glob->zone_kernel = zone;
320 ret = kobject_init_and_add(
321 &zone->kobj, &ttm_mem_zone_kobj_type, &glob->kobj, zone->name);
322 if (unlikely(ret != 0)) {
323 kobject_put(&zone->kobj);
324 return ret;
325 }
326 glob->zones[glob->num_zones++] = zone;
327 return 0;
328}
329
330#ifdef CONFIG_HIGHMEM
331static int ttm_mem_init_highmem_zone(struct ttm_mem_global *glob,
332 const struct sysinfo *si)
333{
334 struct ttm_mem_zone *zone;
335 uint64_t mem;
336 int ret;
337
338 if (si->totalhigh == 0)
339 return 0;
340
341 zone = kzalloc(sizeof(*zone), GFP_KERNEL);
342 if (unlikely(!zone))
343 return -ENOMEM;
344
345 mem = si->totalram;
346 mem *= si->mem_unit;
347
348 zone->name = "highmem";
349 zone->zone_mem = mem;
350 zone->max_mem = mem >> 1;
351 zone->emer_mem = (mem >> 1) + (mem >> 2);
352 zone->swap_limit = zone->max_mem - (mem >> 3);
353 zone->used_mem = 0;
354 zone->glob = glob;
355 glob->zone_highmem = zone;
356 ret = kobject_init_and_add(
357 &zone->kobj, &ttm_mem_zone_kobj_type, &glob->kobj, "%s",
358 zone->name);
359 if (unlikely(ret != 0)) {
360 kobject_put(&zone->kobj);
361 return ret;
362 }
363 glob->zones[glob->num_zones++] = zone;
364 return 0;
365}
366#else
367static int ttm_mem_init_dma32_zone(struct ttm_mem_global *glob,
368 const struct sysinfo *si)
369{
370 struct ttm_mem_zone *zone = kzalloc(sizeof(*zone), GFP_KERNEL);
371 uint64_t mem;
372 int ret;
373
374 if (unlikely(!zone))
375 return -ENOMEM;
376
377 mem = si->totalram;
378 mem *= si->mem_unit;
379
380 /**
381 * No special dma32 zone needed.
382 */
383
384 if (mem <= ((uint64_t) 1ULL << 32)) {
385 kfree(zone);
386 return 0;
387 }
388
389 /*
390 * Limit max dma32 memory to 4GB for now
391 * until we can figure out how big this
392 * zone really is.
393 */
394
395 mem = ((uint64_t) 1ULL << 32);
396 zone->name = "dma32";
397 zone->zone_mem = mem;
398 zone->max_mem = mem >> 1;
399 zone->emer_mem = (mem >> 1) + (mem >> 2);
400 zone->swap_limit = zone->max_mem - (mem >> 3);
401 zone->used_mem = 0;
402 zone->glob = glob;
403 glob->zone_dma32 = zone;
404 ret = kobject_init_and_add(
405 &zone->kobj, &ttm_mem_zone_kobj_type, &glob->kobj, zone->name);
406 if (unlikely(ret != 0)) {
407 kobject_put(&zone->kobj);
408 return ret;
409 }
410 glob->zones[glob->num_zones++] = zone;
411 return 0;
412}
413#endif
414
415int ttm_mem_global_init(struct ttm_mem_global *glob)
416{
417 struct sysinfo si;
418 int ret;
419 int i;
420 struct ttm_mem_zone *zone;
421
422 spin_lock_init(&glob->lock);
423 glob->swap_queue = create_singlethread_workqueue("ttm_swap");
424 INIT_WORK(&glob->work, ttm_shrink_work);
425 ret = kobject_init_and_add(
426 &glob->kobj, &ttm_mem_glob_kobj_type, ttm_get_kobj(), "memory_accounting");
427 if (unlikely(ret != 0)) {
428 kobject_put(&glob->kobj);
429 return ret;
430 }
431
432 si_meminfo(&si);
433
434 /* set it as 0 by default to keep original behavior of OOM */
435 glob->lower_mem_limit = 0;
436
437 ret = ttm_mem_init_kernel_zone(glob, &si);
438 if (unlikely(ret != 0))
439 goto out_no_zone;
440#ifdef CONFIG_HIGHMEM
441 ret = ttm_mem_init_highmem_zone(glob, &si);
442 if (unlikely(ret != 0))
443 goto out_no_zone;
444#else
445 ret = ttm_mem_init_dma32_zone(glob, &si);
446 if (unlikely(ret != 0))
447 goto out_no_zone;
448#endif
449 for (i = 0; i < glob->num_zones; ++i) {
450 zone = glob->zones[i];
451 pr_info("Zone %7s: Available graphics memory: %llu KiB\n",
452 zone->name, (unsigned long long)zone->max_mem >> 10);
453 }
454 ttm_page_alloc_init(glob, glob->zone_kernel->max_mem/(2*PAGE_SIZE));
455 ttm_dma_page_alloc_init(glob, glob->zone_kernel->max_mem/(2*PAGE_SIZE));
456 return 0;
457out_no_zone:
458 ttm_mem_global_release(glob);
459 return ret;
460}
461
462void ttm_mem_global_release(struct ttm_mem_global *glob)
463{
464 struct ttm_mem_zone *zone;
465 unsigned int i;
466
467 /* let the page allocator first stop the shrink work. */
468 ttm_page_alloc_fini();
469 ttm_dma_page_alloc_fini();
470
471 flush_workqueue(glob->swap_queue);
472 destroy_workqueue(glob->swap_queue);
473 glob->swap_queue = NULL;
474 for (i = 0; i < glob->num_zones; ++i) {
475 zone = glob->zones[i];
476 kobject_del(&zone->kobj);
477 kobject_put(&zone->kobj);
478 }
479 kobject_del(&glob->kobj);
480 kobject_put(&glob->kobj);
481 memset(glob, 0, sizeof(*glob));
482}
483
484static void ttm_check_swapping(struct ttm_mem_global *glob)
485{
486 bool needs_swapping = false;
487 unsigned int i;
488 struct ttm_mem_zone *zone;
489
490 spin_lock(&glob->lock);
491 for (i = 0; i < glob->num_zones; ++i) {
492 zone = glob->zones[i];
493 if (zone->used_mem > zone->swap_limit) {
494 needs_swapping = true;
495 break;
496 }
497 }
498
499 spin_unlock(&glob->lock);
500
501 if (unlikely(needs_swapping))
502 (void)queue_work(glob->swap_queue, &glob->work);
503
504}
505
506static void ttm_mem_global_free_zone(struct ttm_mem_global *glob,
507 struct ttm_mem_zone *single_zone,
508 uint64_t amount)
509{
510 unsigned int i;
511 struct ttm_mem_zone *zone;
512
513 spin_lock(&glob->lock);
514 for (i = 0; i < glob->num_zones; ++i) {
515 zone = glob->zones[i];
516 if (single_zone && zone != single_zone)
517 continue;
518 zone->used_mem -= amount;
519 }
520 spin_unlock(&glob->lock);
521}
522
523void ttm_mem_global_free(struct ttm_mem_global *glob,
524 uint64_t amount)
525{
526 return ttm_mem_global_free_zone(glob, glob->zone_kernel, amount);
527}
528EXPORT_SYMBOL(ttm_mem_global_free);
529
530/*
531 * check if the available mem is under lower memory limit
532 *
533 * a. if no swap disk at all or free swap space is under swap_mem_limit
534 * but available system mem is bigger than sys_mem_limit, allow TTM
535 * allocation;
536 *
537 * b. if the available system mem is less than sys_mem_limit but free
538 * swap disk is bigger than swap_mem_limit, allow TTM allocation.
539 */
540bool
541ttm_check_under_lowerlimit(struct ttm_mem_global *glob,
542 uint64_t num_pages,
543 struct ttm_operation_ctx *ctx)
544{
545 int64_t available;
546
547 if (ctx->flags & TTM_OPT_FLAG_FORCE_ALLOC)
548 return false;
549
550 available = get_nr_swap_pages() + si_mem_available();
551 available -= num_pages;
552 if (available < glob->lower_mem_limit)
553 return true;
554
555 return false;
556}
557EXPORT_SYMBOL(ttm_check_under_lowerlimit);
558
559static int ttm_mem_global_reserve(struct ttm_mem_global *glob,
560 struct ttm_mem_zone *single_zone,
561 uint64_t amount, bool reserve)
562{
563 uint64_t limit;
564 int ret = -ENOMEM;
565 unsigned int i;
566 struct ttm_mem_zone *zone;
567
568 spin_lock(&glob->lock);
569 for (i = 0; i < glob->num_zones; ++i) {
570 zone = glob->zones[i];
571 if (single_zone && zone != single_zone)
572 continue;
573
574 limit = (capable(CAP_SYS_ADMIN)) ?
575 zone->emer_mem : zone->max_mem;
576
577 if (zone->used_mem > limit)
578 goto out_unlock;
579 }
580
581 if (reserve) {
582 for (i = 0; i < glob->num_zones; ++i) {
583 zone = glob->zones[i];
584 if (single_zone && zone != single_zone)
585 continue;
586 zone->used_mem += amount;
587 }
588 }
589
590 ret = 0;
591out_unlock:
592 spin_unlock(&glob->lock);
593 ttm_check_swapping(glob);
594
595 return ret;
596}
597
598
599static int ttm_mem_global_alloc_zone(struct ttm_mem_global *glob,
600 struct ttm_mem_zone *single_zone,
601 uint64_t memory,
602 struct ttm_operation_ctx *ctx)
603{
604 int count = TTM_MEMORY_ALLOC_RETRIES;
605
606 while (unlikely(ttm_mem_global_reserve(glob,
607 single_zone,
608 memory, true)
609 != 0)) {
610 if (ctx->no_wait_gpu)
611 return -ENOMEM;
612 if (unlikely(count-- == 0))
613 return -ENOMEM;
614 ttm_shrink(glob, false, memory + (memory >> 2) + 16, ctx);
615 }
616
617 return 0;
618}
619
620int ttm_mem_global_alloc(struct ttm_mem_global *glob, uint64_t memory,
621 struct ttm_operation_ctx *ctx)
622{
623 /**
624 * Normal allocations of kernel memory are registered in
625 * the kernel zone.
626 */
627
628 return ttm_mem_global_alloc_zone(glob, glob->zone_kernel, memory, ctx);
629}
630EXPORT_SYMBOL(ttm_mem_global_alloc);
631
632int ttm_mem_global_alloc_page(struct ttm_mem_global *glob,
633 struct page *page, uint64_t size,
634 struct ttm_operation_ctx *ctx)
635{
636 struct ttm_mem_zone *zone = NULL;
637
638 /**
639 * Page allocations may be registed in a single zone
640 * only if highmem or !dma32.
641 */
642
643#ifdef CONFIG_HIGHMEM
644 if (PageHighMem(page) && glob->zone_highmem != NULL)
645 zone = glob->zone_highmem;
646#else
647 if (glob->zone_dma32 && page_to_pfn(page) > 0x00100000UL)
648 zone = glob->zone_kernel;
649#endif
650 return ttm_mem_global_alloc_zone(glob, zone, size, ctx);
651}
652
653void ttm_mem_global_free_page(struct ttm_mem_global *glob, struct page *page,
654 uint64_t size)
655{
656 struct ttm_mem_zone *zone = NULL;
657
658#ifdef CONFIG_HIGHMEM
659 if (PageHighMem(page) && glob->zone_highmem != NULL)
660 zone = glob->zone_highmem;
661#else
662 if (glob->zone_dma32 && page_to_pfn(page) > 0x00100000UL)
663 zone = glob->zone_kernel;
664#endif
665 ttm_mem_global_free_zone(glob, zone, size);
666}
667
668size_t ttm_round_pot(size_t size)
669{
670 if ((size & (size - 1)) == 0)
671 return size;
672 else if (size > PAGE_SIZE)
673 return PAGE_ALIGN(size);
674 else {
675 size_t tmp_size = 4;
676
677 while (tmp_size < size)
678 tmp_size <<= 1;
679
680 return tmp_size;
681 }
682 return 0;
683}
684EXPORT_SYMBOL(ttm_round_pot);
685
686uint64_t ttm_get_kernel_zone_memory_size(struct ttm_mem_global *glob)
687{
688 return glob->zone_kernel->max_mem;
689}
690EXPORT_SYMBOL(ttm_get_kernel_zone_memory_size);