1 /**************************************************************************
3 * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sub license, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
19 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
20 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
21 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
22 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
23 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
24 * USE OR OTHER DEALINGS IN THE SOFTWARE.
26 **************************************************************************/
28 #define pr_fmt(fmt) "[TTM] " fmt
30 #include <drm/ttm/ttm_memory.h>
31 #include <drm/ttm/ttm_module.h>
32 #include <drm/ttm/ttm_page_alloc.h>
33 #include <linux/spinlock.h>
34 #include <linux/sched.h>
35 #include <linux/wait.h>
37 #include <linux/module.h>
38 #include <linux/slab.h>
40 #define TTM_MEMORY_ALLOC_RETRIES 4
44 struct ttm_mem_global *glob;
53 static struct attribute ttm_mem_sys = {
54 .name = "zone_memory",
57 static struct attribute ttm_mem_emer = {
58 .name = "emergency_memory",
59 .mode = S_IRUGO | S_IWUSR
61 static struct attribute ttm_mem_max = {
62 .name = "available_memory",
63 .mode = S_IRUGO | S_IWUSR
65 static struct attribute ttm_mem_swap = {
67 .mode = S_IRUGO | S_IWUSR
69 static struct attribute ttm_mem_used = {
70 .name = "used_memory",
74 static void ttm_mem_zone_kobj_release(struct kobject *kobj)
76 struct ttm_mem_zone *zone =
77 container_of(kobj, struct ttm_mem_zone, kobj);
79 pr_info("Zone %7s: Used memory at exit: %llu kiB\n",
80 zone->name, (unsigned long long)zone->used_mem >> 10);
84 static ssize_t ttm_mem_zone_show(struct kobject *kobj,
85 struct attribute *attr,
88 struct ttm_mem_zone *zone =
89 container_of(kobj, struct ttm_mem_zone, kobj);
92 spin_lock(&zone->glob->lock);
93 if (attr == &ttm_mem_sys)
95 else if (attr == &ttm_mem_emer)
97 else if (attr == &ttm_mem_max)
99 else if (attr == &ttm_mem_swap)
100 val = zone->swap_limit;
101 else if (attr == &ttm_mem_used)
102 val = zone->used_mem;
103 spin_unlock(&zone->glob->lock);
105 return snprintf(buffer, PAGE_SIZE, "%llu\n",
106 (unsigned long long) val >> 10);
109 static void ttm_check_swapping(struct ttm_mem_global *glob);
111 static ssize_t ttm_mem_zone_store(struct kobject *kobj,
112 struct attribute *attr,
116 struct ttm_mem_zone *zone =
117 container_of(kobj, struct ttm_mem_zone, kobj);
122 chars = sscanf(buffer, "%lu", &val);
129 spin_lock(&zone->glob->lock);
130 if (val64 > zone->zone_mem)
131 val64 = zone->zone_mem;
132 if (attr == &ttm_mem_emer) {
133 zone->emer_mem = val64;
134 if (zone->max_mem > val64)
135 zone->max_mem = val64;
136 } else if (attr == &ttm_mem_max) {
137 zone->max_mem = val64;
138 if (zone->emer_mem < val64)
139 zone->emer_mem = val64;
140 } else if (attr == &ttm_mem_swap)
141 zone->swap_limit = val64;
142 spin_unlock(&zone->glob->lock);
144 ttm_check_swapping(zone->glob);
149 static struct attribute *ttm_mem_zone_attrs[] = {
158 static const struct sysfs_ops ttm_mem_zone_ops = {
159 .show = &ttm_mem_zone_show,
160 .store = &ttm_mem_zone_store
163 static struct kobj_type ttm_mem_zone_kobj_type = {
164 .release = &ttm_mem_zone_kobj_release,
165 .sysfs_ops = &ttm_mem_zone_ops,
166 .default_attrs = ttm_mem_zone_attrs,
169 static void ttm_mem_global_kobj_release(struct kobject *kobj)
171 struct ttm_mem_global *glob =
172 container_of(kobj, struct ttm_mem_global, kobj);
177 static struct kobj_type ttm_mem_glob_kobj_type = {
178 .release = &ttm_mem_global_kobj_release,
181 static bool ttm_zones_above_swap_target(struct ttm_mem_global *glob,
182 bool from_wq, uint64_t extra)
185 struct ttm_mem_zone *zone;
188 for (i = 0; i < glob->num_zones; ++i) {
189 zone = glob->zones[i];
192 target = zone->swap_limit;
193 else if (capable(CAP_SYS_ADMIN))
194 target = zone->emer_mem;
196 target = zone->max_mem;
198 target = (extra > target) ? 0ULL : target;
200 if (zone->used_mem > target)
207 * At this point we only support a single shrink callback.
208 * Extend this if needed, perhaps using a linked list of callbacks.
209 * Note that this function is reentrant:
210 * many threads may try to swap out at any given time.
213 static void ttm_shrink(struct ttm_mem_global *glob, bool from_wq,
214 uint64_t extra, struct ttm_operation_ctx *ctx)
218 spin_lock(&glob->lock);
220 while (ttm_zones_above_swap_target(glob, from_wq, extra)) {
221 spin_unlock(&glob->lock);
222 ret = ttm_bo_swapout(glob->bo_glob, ctx);
223 spin_lock(&glob->lock);
224 if (unlikely(ret != 0))
228 spin_unlock(&glob->lock);
231 static void ttm_shrink_work(struct work_struct *work)
233 struct ttm_operation_ctx ctx = {
234 .interruptible = false,
237 struct ttm_mem_global *glob =
238 container_of(work, struct ttm_mem_global, work);
240 ttm_shrink(glob, true, 0ULL, &ctx);
243 static int ttm_mem_init_kernel_zone(struct ttm_mem_global *glob,
244 const struct sysinfo *si)
246 struct ttm_mem_zone *zone = kzalloc(sizeof(*zone), GFP_KERNEL);
253 mem = si->totalram - si->totalhigh;
256 zone->name = "kernel";
257 zone->zone_mem = mem;
258 zone->max_mem = mem >> 1;
259 zone->emer_mem = (mem >> 1) + (mem >> 2);
260 zone->swap_limit = zone->max_mem - (mem >> 3);
263 glob->zone_kernel = zone;
264 ret = kobject_init_and_add(
265 &zone->kobj, &ttm_mem_zone_kobj_type, &glob->kobj, zone->name);
266 if (unlikely(ret != 0)) {
267 kobject_put(&zone->kobj);
270 glob->zones[glob->num_zones++] = zone;
274 #ifdef CONFIG_HIGHMEM
275 static int ttm_mem_init_highmem_zone(struct ttm_mem_global *glob,
276 const struct sysinfo *si)
278 struct ttm_mem_zone *zone;
282 if (si->totalhigh == 0)
285 zone = kzalloc(sizeof(*zone), GFP_KERNEL);
292 zone->name = "highmem";
293 zone->zone_mem = mem;
294 zone->max_mem = mem >> 1;
295 zone->emer_mem = (mem >> 1) + (mem >> 2);
296 zone->swap_limit = zone->max_mem - (mem >> 3);
299 glob->zone_highmem = zone;
300 ret = kobject_init_and_add(
301 &zone->kobj, &ttm_mem_zone_kobj_type, &glob->kobj, "%s",
303 if (unlikely(ret != 0)) {
304 kobject_put(&zone->kobj);
307 glob->zones[glob->num_zones++] = zone;
311 static int ttm_mem_init_dma32_zone(struct ttm_mem_global *glob,
312 const struct sysinfo *si)
314 struct ttm_mem_zone *zone = kzalloc(sizeof(*zone), GFP_KERNEL);
325 * No special dma32 zone needed.
328 if (mem <= ((uint64_t) 1ULL << 32)) {
334 * Limit max dma32 memory to 4GB for now
335 * until we can figure out how big this
339 mem = ((uint64_t) 1ULL << 32);
340 zone->name = "dma32";
341 zone->zone_mem = mem;
342 zone->max_mem = mem >> 1;
343 zone->emer_mem = (mem >> 1) + (mem >> 2);
344 zone->swap_limit = zone->max_mem - (mem >> 3);
347 glob->zone_dma32 = zone;
348 ret = kobject_init_and_add(
349 &zone->kobj, &ttm_mem_zone_kobj_type, &glob->kobj, zone->name);
350 if (unlikely(ret != 0)) {
351 kobject_put(&zone->kobj);
354 glob->zones[glob->num_zones++] = zone;
359 int ttm_mem_global_init(struct ttm_mem_global *glob)
364 struct ttm_mem_zone *zone;
366 spin_lock_init(&glob->lock);
367 glob->swap_queue = create_singlethread_workqueue("ttm_swap");
368 INIT_WORK(&glob->work, ttm_shrink_work);
369 ret = kobject_init_and_add(
370 &glob->kobj, &ttm_mem_glob_kobj_type, ttm_get_kobj(), "memory_accounting");
371 if (unlikely(ret != 0)) {
372 kobject_put(&glob->kobj);
378 ret = ttm_mem_init_kernel_zone(glob, &si);
379 if (unlikely(ret != 0))
381 #ifdef CONFIG_HIGHMEM
382 ret = ttm_mem_init_highmem_zone(glob, &si);
383 if (unlikely(ret != 0))
386 ret = ttm_mem_init_dma32_zone(glob, &si);
387 if (unlikely(ret != 0))
390 for (i = 0; i < glob->num_zones; ++i) {
391 zone = glob->zones[i];
392 pr_info("Zone %7s: Available graphics memory: %llu kiB\n",
393 zone->name, (unsigned long long)zone->max_mem >> 10);
395 ttm_page_alloc_init(glob, glob->zone_kernel->max_mem/(2*PAGE_SIZE));
396 ttm_dma_page_alloc_init(glob, glob->zone_kernel->max_mem/(2*PAGE_SIZE));
399 ttm_mem_global_release(glob);
402 EXPORT_SYMBOL(ttm_mem_global_init);
404 void ttm_mem_global_release(struct ttm_mem_global *glob)
407 struct ttm_mem_zone *zone;
409 /* let the page allocator first stop the shrink work. */
410 ttm_page_alloc_fini();
411 ttm_dma_page_alloc_fini();
413 flush_workqueue(glob->swap_queue);
414 destroy_workqueue(glob->swap_queue);
415 glob->swap_queue = NULL;
416 for (i = 0; i < glob->num_zones; ++i) {
417 zone = glob->zones[i];
418 kobject_del(&zone->kobj);
419 kobject_put(&zone->kobj);
421 kobject_del(&glob->kobj);
422 kobject_put(&glob->kobj);
424 EXPORT_SYMBOL(ttm_mem_global_release);
426 static void ttm_check_swapping(struct ttm_mem_global *glob)
428 bool needs_swapping = false;
430 struct ttm_mem_zone *zone;
432 spin_lock(&glob->lock);
433 for (i = 0; i < glob->num_zones; ++i) {
434 zone = glob->zones[i];
435 if (zone->used_mem > zone->swap_limit) {
436 needs_swapping = true;
441 spin_unlock(&glob->lock);
443 if (unlikely(needs_swapping))
444 (void)queue_work(glob->swap_queue, &glob->work);
448 static void ttm_mem_global_free_zone(struct ttm_mem_global *glob,
449 struct ttm_mem_zone *single_zone,
453 struct ttm_mem_zone *zone;
455 spin_lock(&glob->lock);
456 for (i = 0; i < glob->num_zones; ++i) {
457 zone = glob->zones[i];
458 if (single_zone && zone != single_zone)
460 zone->used_mem -= amount;
462 spin_unlock(&glob->lock);
465 void ttm_mem_global_free(struct ttm_mem_global *glob,
468 return ttm_mem_global_free_zone(glob, NULL, amount);
470 EXPORT_SYMBOL(ttm_mem_global_free);
472 static int ttm_mem_global_reserve(struct ttm_mem_global *glob,
473 struct ttm_mem_zone *single_zone,
474 uint64_t amount, bool reserve)
479 struct ttm_mem_zone *zone;
481 spin_lock(&glob->lock);
482 for (i = 0; i < glob->num_zones; ++i) {
483 zone = glob->zones[i];
484 if (single_zone && zone != single_zone)
487 limit = (capable(CAP_SYS_ADMIN)) ?
488 zone->emer_mem : zone->max_mem;
490 if (zone->used_mem > limit)
495 for (i = 0; i < glob->num_zones; ++i) {
496 zone = glob->zones[i];
497 if (single_zone && zone != single_zone)
499 zone->used_mem += amount;
505 spin_unlock(&glob->lock);
506 ttm_check_swapping(glob);
512 static int ttm_mem_global_alloc_zone(struct ttm_mem_global *glob,
513 struct ttm_mem_zone *single_zone,
515 struct ttm_operation_ctx *ctx)
517 int count = TTM_MEMORY_ALLOC_RETRIES;
519 while (unlikely(ttm_mem_global_reserve(glob,
523 if (ctx->no_wait_gpu)
525 if (unlikely(count-- == 0))
527 ttm_shrink(glob, false, memory + (memory >> 2) + 16, ctx);
533 int ttm_mem_global_alloc(struct ttm_mem_global *glob, uint64_t memory,
534 struct ttm_operation_ctx *ctx)
537 * Normal allocations of kernel memory are registered in
541 return ttm_mem_global_alloc_zone(glob, NULL, memory, ctx);
543 EXPORT_SYMBOL(ttm_mem_global_alloc);
545 int ttm_mem_global_alloc_page(struct ttm_mem_global *glob,
546 struct page *page, uint64_t size,
547 struct ttm_operation_ctx *ctx)
549 struct ttm_mem_zone *zone = NULL;
552 * Page allocations may be registed in a single zone
553 * only if highmem or !dma32.
556 #ifdef CONFIG_HIGHMEM
557 if (PageHighMem(page) && glob->zone_highmem != NULL)
558 zone = glob->zone_highmem;
560 if (glob->zone_dma32 && page_to_pfn(page) > 0x00100000UL)
561 zone = glob->zone_kernel;
563 return ttm_mem_global_alloc_zone(glob, zone, size, ctx);
566 void ttm_mem_global_free_page(struct ttm_mem_global *glob, struct page *page,
569 struct ttm_mem_zone *zone = NULL;
571 #ifdef CONFIG_HIGHMEM
572 if (PageHighMem(page) && glob->zone_highmem != NULL)
573 zone = glob->zone_highmem;
575 if (glob->zone_dma32 && page_to_pfn(page) > 0x00100000UL)
576 zone = glob->zone_kernel;
578 ttm_mem_global_free_zone(glob, zone, size);
581 size_t ttm_round_pot(size_t size)
583 if ((size & (size - 1)) == 0)
585 else if (size > PAGE_SIZE)
586 return PAGE_ALIGN(size);
590 while (tmp_size < size)
597 EXPORT_SYMBOL(ttm_round_pot);
599 uint64_t ttm_get_kernel_zone_memory_size(struct ttm_mem_global *glob)
601 return glob->zone_kernel->max_mem;
603 EXPORT_SYMBOL(ttm_get_kernel_zone_memory_size);