Loading...
Note: File does not exist in v4.6.
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 <linux/spinlock.h>
32#include <linux/sched.h>
33#include <linux/wait.h>
34#include <linux/mm.h>
35#include <linux/module.h>
36#include <linux/slab.h>
37#include <linux/swap.h>
38
39#include <drm/drm_device.h>
40#include <drm/drm_file.h>
41#include <drm/ttm/ttm_device.h>
42
43#include "ttm_memory.h"
44
45#define TTM_MEMORY_ALLOC_RETRIES 4
46
47struct ttm_mem_global ttm_mem_glob;
48EXPORT_SYMBOL(ttm_mem_glob);
49
50struct ttm_mem_zone {
51 struct kobject kobj;
52 struct ttm_mem_global *glob;
53 const char *name;
54 uint64_t zone_mem;
55 uint64_t emer_mem;
56 uint64_t max_mem;
57 uint64_t swap_limit;
58 uint64_t used_mem;
59};
60
61static struct attribute ttm_mem_sys = {
62 .name = "zone_memory",
63 .mode = S_IRUGO
64};
65static struct attribute ttm_mem_emer = {
66 .name = "emergency_memory",
67 .mode = S_IRUGO | S_IWUSR
68};
69static struct attribute ttm_mem_max = {
70 .name = "available_memory",
71 .mode = S_IRUGO | S_IWUSR
72};
73static struct attribute ttm_mem_swap = {
74 .name = "swap_limit",
75 .mode = S_IRUGO | S_IWUSR
76};
77static struct attribute ttm_mem_used = {
78 .name = "used_memory",
79 .mode = S_IRUGO
80};
81
82static void ttm_mem_zone_kobj_release(struct kobject *kobj)
83{
84 struct ttm_mem_zone *zone =
85 container_of(kobj, struct ttm_mem_zone, kobj);
86
87 pr_info("Zone %7s: Used memory at exit: %llu KiB\n",
88 zone->name, (unsigned long long)zone->used_mem >> 10);
89 kfree(zone);
90}
91
92static ssize_t ttm_mem_zone_show(struct kobject *kobj,
93 struct attribute *attr,
94 char *buffer)
95{
96 struct ttm_mem_zone *zone =
97 container_of(kobj, struct ttm_mem_zone, kobj);
98 uint64_t val = 0;
99
100 spin_lock(&zone->glob->lock);
101 if (attr == &ttm_mem_sys)
102 val = zone->zone_mem;
103 else if (attr == &ttm_mem_emer)
104 val = zone->emer_mem;
105 else if (attr == &ttm_mem_max)
106 val = zone->max_mem;
107 else if (attr == &ttm_mem_swap)
108 val = zone->swap_limit;
109 else if (attr == &ttm_mem_used)
110 val = zone->used_mem;
111 spin_unlock(&zone->glob->lock);
112
113 return snprintf(buffer, PAGE_SIZE, "%llu\n",
114 (unsigned long long) val >> 10);
115}
116
117static void ttm_check_swapping(struct ttm_mem_global *glob);
118
119static ssize_t ttm_mem_zone_store(struct kobject *kobj,
120 struct attribute *attr,
121 const char *buffer,
122 size_t size)
123{
124 struct ttm_mem_zone *zone =
125 container_of(kobj, struct ttm_mem_zone, kobj);
126 int chars;
127 unsigned long val;
128 uint64_t val64;
129
130 chars = sscanf(buffer, "%lu", &val);
131 if (chars == 0)
132 return size;
133
134 val64 = val;
135 val64 <<= 10;
136
137 spin_lock(&zone->glob->lock);
138 if (val64 > zone->zone_mem)
139 val64 = zone->zone_mem;
140 if (attr == &ttm_mem_emer) {
141 zone->emer_mem = val64;
142 if (zone->max_mem > val64)
143 zone->max_mem = val64;
144 } else if (attr == &ttm_mem_max) {
145 zone->max_mem = val64;
146 if (zone->emer_mem < val64)
147 zone->emer_mem = val64;
148 } else if (attr == &ttm_mem_swap)
149 zone->swap_limit = val64;
150 spin_unlock(&zone->glob->lock);
151
152 ttm_check_swapping(zone->glob);
153
154 return size;
155}
156
157static struct attribute *ttm_mem_zone_attrs[] = {
158 &ttm_mem_sys,
159 &ttm_mem_emer,
160 &ttm_mem_max,
161 &ttm_mem_swap,
162 &ttm_mem_used,
163 NULL
164};
165
166static const struct sysfs_ops ttm_mem_zone_ops = {
167 .show = &ttm_mem_zone_show,
168 .store = &ttm_mem_zone_store
169};
170
171static struct kobj_type ttm_mem_zone_kobj_type = {
172 .release = &ttm_mem_zone_kobj_release,
173 .sysfs_ops = &ttm_mem_zone_ops,
174 .default_attrs = ttm_mem_zone_attrs,
175};
176
177static struct attribute ttm_mem_global_lower_mem_limit = {
178 .name = "lower_mem_limit",
179 .mode = S_IRUGO | S_IWUSR
180};
181
182static ssize_t ttm_mem_global_show(struct kobject *kobj,
183 struct attribute *attr,
184 char *buffer)
185{
186 struct ttm_mem_global *glob =
187 container_of(kobj, struct ttm_mem_global, kobj);
188 uint64_t val = 0;
189
190 spin_lock(&glob->lock);
191 val = glob->lower_mem_limit;
192 spin_unlock(&glob->lock);
193 /* convert from number of pages to KB */
194 val <<= (PAGE_SHIFT - 10);
195 return snprintf(buffer, PAGE_SIZE, "%llu\n",
196 (unsigned long long) val);
197}
198
199static ssize_t ttm_mem_global_store(struct kobject *kobj,
200 struct attribute *attr,
201 const char *buffer,
202 size_t size)
203{
204 int chars;
205 uint64_t val64;
206 unsigned long val;
207 struct ttm_mem_global *glob =
208 container_of(kobj, struct ttm_mem_global, kobj);
209
210 chars = sscanf(buffer, "%lu", &val);
211 if (chars == 0)
212 return size;
213
214 val64 = val;
215 /* convert from KB to number of pages */
216 val64 >>= (PAGE_SHIFT - 10);
217
218 spin_lock(&glob->lock);
219 glob->lower_mem_limit = val64;
220 spin_unlock(&glob->lock);
221
222 return size;
223}
224
225static struct attribute *ttm_mem_global_attrs[] = {
226 &ttm_mem_global_lower_mem_limit,
227 NULL
228};
229
230static const struct sysfs_ops ttm_mem_global_ops = {
231 .show = &ttm_mem_global_show,
232 .store = &ttm_mem_global_store,
233};
234
235static struct kobj_type ttm_mem_glob_kobj_type = {
236 .sysfs_ops = &ttm_mem_global_ops,
237 .default_attrs = ttm_mem_global_attrs,
238};
239
240static bool ttm_zones_above_swap_target(struct ttm_mem_global *glob,
241 bool from_wq, uint64_t extra)
242{
243 unsigned int i;
244 struct ttm_mem_zone *zone;
245 uint64_t target;
246
247 for (i = 0; i < glob->num_zones; ++i) {
248 zone = glob->zones[i];
249
250 if (from_wq)
251 target = zone->swap_limit;
252 else if (capable(CAP_SYS_ADMIN))
253 target = zone->emer_mem;
254 else
255 target = zone->max_mem;
256
257 target = (extra > target) ? 0ULL : target;
258
259 if (zone->used_mem > target)
260 return true;
261 }
262 return false;
263}
264
265/*
266 * At this point we only support a single shrink callback.
267 * Extend this if needed, perhaps using a linked list of callbacks.
268 * Note that this function is reentrant:
269 * many threads may try to swap out at any given time.
270 */
271
272static void ttm_shrink(struct ttm_mem_global *glob, bool from_wq,
273 uint64_t extra, struct ttm_operation_ctx *ctx)
274{
275 int ret;
276
277 spin_lock(&glob->lock);
278
279 while (ttm_zones_above_swap_target(glob, from_wq, extra)) {
280 spin_unlock(&glob->lock);
281 ret = ttm_global_swapout(ctx, GFP_KERNEL);
282 spin_lock(&glob->lock);
283 if (unlikely(ret <= 0))
284 break;
285 }
286
287 spin_unlock(&glob->lock);
288}
289
290static void ttm_shrink_work(struct work_struct *work)
291{
292 struct ttm_operation_ctx ctx = {
293 .interruptible = false,
294 .no_wait_gpu = false
295 };
296 struct ttm_mem_global *glob =
297 container_of(work, struct ttm_mem_global, work);
298
299 ttm_shrink(glob, true, 0ULL, &ctx);
300}
301
302static int ttm_mem_init_kernel_zone(struct ttm_mem_global *glob,
303 const struct sysinfo *si)
304{
305 struct ttm_mem_zone *zone = kzalloc(sizeof(*zone), GFP_KERNEL);
306 uint64_t mem;
307 int ret;
308
309 if (unlikely(!zone))
310 return -ENOMEM;
311
312 mem = si->totalram - si->totalhigh;
313 mem *= si->mem_unit;
314
315 zone->name = "kernel";
316 zone->zone_mem = mem;
317 zone->max_mem = mem >> 1;
318 zone->emer_mem = (mem >> 1) + (mem >> 2);
319 zone->swap_limit = zone->max_mem - (mem >> 3);
320 zone->used_mem = 0;
321 zone->glob = glob;
322 glob->zone_kernel = zone;
323 ret = kobject_init_and_add(
324 &zone->kobj, &ttm_mem_zone_kobj_type, &glob->kobj, zone->name);
325 if (unlikely(ret != 0)) {
326 kobject_put(&zone->kobj);
327 return ret;
328 }
329 glob->zones[glob->num_zones++] = zone;
330 return 0;
331}
332
333#ifdef CONFIG_HIGHMEM
334static int ttm_mem_init_highmem_zone(struct ttm_mem_global *glob,
335 const struct sysinfo *si)
336{
337 struct ttm_mem_zone *zone;
338 uint64_t mem;
339 int ret;
340
341 if (si->totalhigh == 0)
342 return 0;
343
344 zone = kzalloc(sizeof(*zone), GFP_KERNEL);
345 if (unlikely(!zone))
346 return -ENOMEM;
347
348 mem = si->totalram;
349 mem *= si->mem_unit;
350
351 zone->name = "highmem";
352 zone->zone_mem = mem;
353 zone->max_mem = mem >> 1;
354 zone->emer_mem = (mem >> 1) + (mem >> 2);
355 zone->swap_limit = zone->max_mem - (mem >> 3);
356 zone->used_mem = 0;
357 zone->glob = glob;
358 glob->zone_highmem = zone;
359 ret = kobject_init_and_add(
360 &zone->kobj, &ttm_mem_zone_kobj_type, &glob->kobj, "%s",
361 zone->name);
362 if (unlikely(ret != 0)) {
363 kobject_put(&zone->kobj);
364 return ret;
365 }
366 glob->zones[glob->num_zones++] = zone;
367 return 0;
368}
369#else
370static int ttm_mem_init_dma32_zone(struct ttm_mem_global *glob,
371 const struct sysinfo *si)
372{
373 struct ttm_mem_zone *zone = kzalloc(sizeof(*zone), GFP_KERNEL);
374 uint64_t mem;
375 int ret;
376
377 if (unlikely(!zone))
378 return -ENOMEM;
379
380 mem = si->totalram;
381 mem *= si->mem_unit;
382
383 /**
384 * No special dma32 zone needed.
385 */
386
387 if (mem <= ((uint64_t) 1ULL << 32)) {
388 kfree(zone);
389 return 0;
390 }
391
392 /*
393 * Limit max dma32 memory to 4GB for now
394 * until we can figure out how big this
395 * zone really is.
396 */
397
398 mem = ((uint64_t) 1ULL << 32);
399 zone->name = "dma32";
400 zone->zone_mem = mem;
401 zone->max_mem = mem >> 1;
402 zone->emer_mem = (mem >> 1) + (mem >> 2);
403 zone->swap_limit = zone->max_mem - (mem >> 3);
404 zone->used_mem = 0;
405 zone->glob = glob;
406 glob->zone_dma32 = zone;
407 ret = kobject_init_and_add(
408 &zone->kobj, &ttm_mem_zone_kobj_type, &glob->kobj, zone->name);
409 if (unlikely(ret != 0)) {
410 kobject_put(&zone->kobj);
411 return ret;
412 }
413 glob->zones[glob->num_zones++] = zone;
414 return 0;
415}
416#endif
417
418int ttm_mem_global_init(struct ttm_mem_global *glob, struct device *dev)
419{
420 struct sysinfo si;
421 int ret;
422 int i;
423 struct ttm_mem_zone *zone;
424
425 spin_lock_init(&glob->lock);
426 glob->swap_queue = create_singlethread_workqueue("ttm_swap");
427 INIT_WORK(&glob->work, ttm_shrink_work);
428
429 ret = kobject_init_and_add(&glob->kobj, &ttm_mem_glob_kobj_type,
430 &dev->kobj, "memory_accounting");
431 if (unlikely(ret != 0)) {
432 kobject_put(&glob->kobj);
433 return ret;
434 }
435
436 si_meminfo(&si);
437
438 spin_lock(&glob->lock);
439 /* set it as 0 by default to keep original behavior of OOM */
440 glob->lower_mem_limit = 0;
441 spin_unlock(&glob->lock);
442
443 ret = ttm_mem_init_kernel_zone(glob, &si);
444 if (unlikely(ret != 0))
445 goto out_no_zone;
446#ifdef CONFIG_HIGHMEM
447 ret = ttm_mem_init_highmem_zone(glob, &si);
448 if (unlikely(ret != 0))
449 goto out_no_zone;
450#else
451 ret = ttm_mem_init_dma32_zone(glob, &si);
452 if (unlikely(ret != 0))
453 goto out_no_zone;
454#endif
455 for (i = 0; i < glob->num_zones; ++i) {
456 zone = glob->zones[i];
457 pr_info("Zone %7s: Available graphics memory: %llu KiB\n",
458 zone->name, (unsigned long long)zone->max_mem >> 10);
459 }
460 return 0;
461out_no_zone:
462 ttm_mem_global_release(glob);
463 return ret;
464}
465
466void ttm_mem_global_release(struct ttm_mem_global *glob)
467{
468 struct ttm_mem_zone *zone;
469 unsigned int i;
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 /* We allow over commit during suspend */
548 if (ctx->force_alloc)
549 return false;
550
551 available = get_nr_swap_pages() + si_mem_available();
552 available -= num_pages;
553 if (available < glob->lower_mem_limit)
554 return true;
555
556 return false;
557}
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);