ttm_memory.c revision 2334b75ffbef6b8932f09ec4418b65ddb764ae99 
1/**************************************************************************
2 *
3 * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA
4 * All Rights Reserved.
5 *
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:
13 *
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
16 * of the Software.
17 *
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.
25 *
26 **************************************************************************/
27
28#include "ttm/ttm_memory.h"
29#include "ttm/ttm_module.h"
30#include "ttm/ttm_page_alloc.h"
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
38#define TTM_MEMORY_ALLOC_RETRIES 4
39
40struct ttm_mem_zone {
41	struct kobject kobj;
42	struct ttm_mem_global *glob;
43	const char *name;
44	uint64_t zone_mem;
45	uint64_t emer_mem;
46	uint64_t max_mem;
47	uint64_t swap_limit;
48	uint64_t used_mem;
49};
50
51static struct attribute ttm_mem_sys = {
52	.name = "zone_memory",
53	.mode = S_IRUGO
54};
55static struct attribute ttm_mem_emer = {
56	.name = "emergency_memory",
57	.mode = S_IRUGO | S_IWUSR
58};
59static struct attribute ttm_mem_max = {
60	.name = "available_memory",
61	.mode = S_IRUGO | S_IWUSR
62};
63static struct attribute ttm_mem_swap = {
64	.name = "swap_limit",
65	.mode = S_IRUGO | S_IWUSR
66};
67static struct attribute ttm_mem_used = {
68	.name = "used_memory",
69	.mode = S_IRUGO
70};
71
72static void ttm_mem_zone_kobj_release(struct kobject *kobj)
73{
74	struct ttm_mem_zone *zone =
75		container_of(kobj, struct ttm_mem_zone, kobj);
76
77	printk(KERN_INFO TTM_PFX
78	       "Zone %7s: Used memory at exit: %llu kiB.\n",
79	       zone->name, (unsigned long long) zone->used_mem >> 10);
80	kfree(zone);
81}
82
83static ssize_t ttm_mem_zone_show(struct kobject *kobj,
84				 struct attribute *attr,
85				 char *buffer)
86{
87	struct ttm_mem_zone *zone =
88		container_of(kobj, struct ttm_mem_zone, kobj);
89	uint64_t val = 0;
90
91	spin_lock(&zone->glob->lock);
92	if (attr == &ttm_mem_sys)
93		val = zone->zone_mem;
94	else if (attr == &ttm_mem_emer)
95		val = zone->emer_mem;
96	else if (attr == &ttm_mem_max)
97		val = zone->max_mem;
98	else if (attr == &ttm_mem_swap)
99		val = zone->swap_limit;
100	else if (attr == &ttm_mem_used)
101		val = zone->used_mem;
102	spin_unlock(&zone->glob->lock);
103
104	return snprintf(buffer, PAGE_SIZE, "%llu\n",
105			(unsigned long long) val >> 10);
106}
107
108static void ttm_check_swapping(struct ttm_mem_global *glob);
109
110static ssize_t ttm_mem_zone_store(struct kobject *kobj,
111				  struct attribute *attr,
112				  const char *buffer,
113				  size_t size)
114{
115	struct ttm_mem_zone *zone =
116		container_of(kobj, struct ttm_mem_zone, kobj);
117	int chars;
118	unsigned long val;
119	uint64_t val64;
120
121	chars = sscanf(buffer, "%lu", &val);
122	if (chars == 0)
123		return size;
124
125	val64 = val;
126	val64 <<= 10;
127
128	spin_lock(&zone->glob->lock);
129	if (val64 > zone->zone_mem)
130		val64 = zone->zone_mem;
131	if (attr == &ttm_mem_emer) {
132		zone->emer_mem = val64;
133		if (zone->max_mem > val64)
134			zone->max_mem = val64;
135	} else if (attr == &ttm_mem_max) {
136		zone->max_mem = val64;
137		if (zone->emer_mem < val64)
138			zone->emer_mem = val64;
139	} else if (attr == &ttm_mem_swap)
140		zone->swap_limit = val64;
141	spin_unlock(&zone->glob->lock);
142
143	ttm_check_swapping(zone->glob);
144
145	return size;
146}
147
148static struct attribute *ttm_mem_zone_attrs[] = {
149	&ttm_mem_sys,
150	&ttm_mem_emer,
151	&ttm_mem_max,
152	&ttm_mem_swap,
153	&ttm_mem_used,
154	NULL
155};
156
157static const struct sysfs_ops ttm_mem_zone_ops = {
158	.show = &ttm_mem_zone_show,
159	.store = &ttm_mem_zone_store
160};
161
162static struct kobj_type ttm_mem_zone_kobj_type = {
163	.release = &ttm_mem_zone_kobj_release,
164	.sysfs_ops = &ttm_mem_zone_ops,
165	.default_attrs = ttm_mem_zone_attrs,
166};
167
168static void ttm_mem_global_kobj_release(struct kobject *kobj)
169{
170	struct ttm_mem_global *glob =
171		container_of(kobj, struct ttm_mem_global, kobj);
172
173	kfree(glob);
174}
175
176static struct kobj_type ttm_mem_glob_kobj_type = {
177	.release = &ttm_mem_global_kobj_release,
178};
179
180static bool ttm_zones_above_swap_target(struct ttm_mem_global *glob,
181					bool from_wq, uint64_t extra)
182{
183	unsigned int i;
184	struct ttm_mem_zone *zone;
185	uint64_t target;
186
187	for (i = 0; i < glob->num_zones; ++i) {
188		zone = glob->zones[i];
189
190		if (from_wq)
191			target = zone->swap_limit;
192		else if (capable(CAP_SYS_ADMIN))
193			target = zone->emer_mem;
194		else
195			target = zone->max_mem;
196
197		target = (extra > target) ? 0ULL : target;
198
199		if (zone->used_mem > target)
200			return true;
201	}
202	return false;
203}
204
205/**
206 * At this point we only support a single shrink callback.
207 * Extend this if needed, perhaps using a linked list of callbacks.
208 * Note that this function is reentrant:
209 * many threads may try to swap out at any given time.
210 */
211
212static void ttm_shrink(struct ttm_mem_global *glob, bool from_wq,
213		       uint64_t extra)
214{
215	int ret;
216	struct ttm_mem_shrink *shrink;
217
218	spin_lock(&glob->lock);
219	if (glob->shrink == NULL)
220		goto out;
221
222	while (ttm_zones_above_swap_target(glob, from_wq, extra)) {
223		shrink = glob->shrink;
224		spin_unlock(&glob->lock);
225		ret = shrink->do_shrink(shrink);
226		spin_lock(&glob->lock);
227		if (unlikely(ret != 0))
228			goto out;
229	}
230out:
231	spin_unlock(&glob->lock);
232}
233
234
235
236static void ttm_shrink_work(struct work_struct *work)
237{
238	struct ttm_mem_global *glob =
239	    container_of(work, struct ttm_mem_global, work);
240
241	ttm_shrink(glob, true, 0ULL);
242}
243
244static int ttm_mem_init_kernel_zone(struct ttm_mem_global *glob,
245				    const struct sysinfo *si)
246{
247	struct ttm_mem_zone *zone = kzalloc(sizeof(*zone), GFP_KERNEL);
248	uint64_t mem;
249	int ret;
250
251	if (unlikely(!zone))
252		return -ENOMEM;
253
254	mem = si->totalram - si->totalhigh;
255	mem *= si->mem_unit;
256
257	zone->name = "kernel";
258	zone->zone_mem = mem;
259	zone->max_mem = mem >> 1;
260	zone->emer_mem = (mem >> 1) + (mem >> 2);
261	zone->swap_limit = zone->max_mem - (mem >> 3);
262	zone->used_mem = 0;
263	zone->glob = glob;
264	glob->zone_kernel = zone;
265	ret = kobject_init_and_add(
266		&zone->kobj, &ttm_mem_zone_kobj_type, &glob->kobj, zone->name);
267	if (unlikely(ret != 0)) {
268		kobject_put(&zone->kobj);
269		return ret;
270	}
271	glob->zones[glob->num_zones++] = zone;
272	return 0;
273}
274
275#ifdef CONFIG_HIGHMEM
276static int ttm_mem_init_highmem_zone(struct ttm_mem_global *glob,
277				     const struct sysinfo *si)
278{
279	struct ttm_mem_zone *zone;
280	uint64_t mem;
281	int ret;
282
283	if (si->totalhigh == 0)
284		return 0;
285
286	zone = kzalloc(sizeof(*zone), GFP_KERNEL);
287	if (unlikely(!zone))
288		return -ENOMEM;
289
290	mem = si->totalram;
291	mem *= si->mem_unit;
292
293	zone->name = "highmem";
294	zone->zone_mem = mem;
295	zone->max_mem = mem >> 1;
296	zone->emer_mem = (mem >> 1) + (mem >> 2);
297	zone->swap_limit = zone->max_mem - (mem >> 3);
298	zone->used_mem = 0;
299	zone->glob = glob;
300	glob->zone_highmem = zone;
301	ret = kobject_init_and_add(
302		&zone->kobj, &ttm_mem_zone_kobj_type, &glob->kobj, zone->name);
303	if (unlikely(ret != 0)) {
304		kobject_put(&zone->kobj);
305		return ret;
306	}
307	glob->zones[glob->num_zones++] = zone;
308	return 0;
309}
310#else
311static int ttm_mem_init_dma32_zone(struct ttm_mem_global *glob,
312				   const struct sysinfo *si)
313{
314	struct ttm_mem_zone *zone = kzalloc(sizeof(*zone), GFP_KERNEL);
315	uint64_t mem;
316	int ret;
317
318	if (unlikely(!zone))
319		return -ENOMEM;
320
321	mem = si->totalram;
322	mem *= si->mem_unit;
323
324	/**
325	 * No special dma32 zone needed.
326	 */
327
328	if (mem <= ((uint64_t) 1ULL << 32)) {
329		kfree(zone);
330		return 0;
331	}
332
333	/*
334	 * Limit max dma32 memory to 4GB for now
335	 * until we can figure out how big this
336	 * zone really is.
337	 */
338
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);
345	zone->used_mem = 0;
346	zone->glob = glob;
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);
352		return ret;
353	}
354	glob->zones[glob->num_zones++] = zone;
355	return 0;
356}
357#endif
358
359int ttm_mem_global_init(struct ttm_mem_global *glob)
360{
361	struct sysinfo si;
362	int ret;
363	int i;
364	struct ttm_mem_zone *zone;
365
366	spin_lock_init(&glob->lock);
367	glob->swap_queue = create_singlethread_workqueue("ttm_swap");
368	INIT_WORK(&glob->work, ttm_shrink_work);
369	init_waitqueue_head(&glob->queue);
370	ret = kobject_init_and_add(
371		&glob->kobj, &ttm_mem_glob_kobj_type, ttm_get_kobj(), "memory_accounting");
372	if (unlikely(ret != 0)) {
373		kobject_put(&glob->kobj);
374		return ret;
375	}
376
377	si_meminfo(&si);
378
379	ret = ttm_mem_init_kernel_zone(glob, &si);
380	if (unlikely(ret != 0))
381		goto out_no_zone;
382#ifdef CONFIG_HIGHMEM
383	ret = ttm_mem_init_highmem_zone(glob, &si);
384	if (unlikely(ret != 0))
385		goto out_no_zone;
386#else
387	ret = ttm_mem_init_dma32_zone(glob, &si);
388	if (unlikely(ret != 0))
389		goto out_no_zone;
390#endif
391	for (i = 0; i < glob->num_zones; ++i) {
392		zone = glob->zones[i];
393		printk(KERN_INFO TTM_PFX
394		       "Zone %7s: Available graphics memory: %llu kiB.\n",
395		       zone->name, (unsigned long long) zone->max_mem >> 10);
396	}
397	ttm_page_alloc_init(glob, glob->zone_kernel->max_mem/(2*PAGE_SIZE));
398	ttm_dma_page_alloc_init(glob, glob->zone_kernel->max_mem/(2*PAGE_SIZE));
399	return 0;
400out_no_zone:
401	ttm_mem_global_release(glob);
402	return ret;
403}
404EXPORT_SYMBOL(ttm_mem_global_init);
405
406void ttm_mem_global_release(struct ttm_mem_global *glob)
407{
408	unsigned int i;
409	struct ttm_mem_zone *zone;
410
411	/* let the page allocator first stop the shrink work. */
412	ttm_page_alloc_fini();
413	ttm_dma_page_alloc_fini();
414
415	flush_workqueue(glob->swap_queue);
416	destroy_workqueue(glob->swap_queue);
417	glob->swap_queue = NULL;
418	for (i = 0; i < glob->num_zones; ++i) {
419		zone = glob->zones[i];
420		kobject_del(&zone->kobj);
421		kobject_put(&zone->kobj);
422			}
423	kobject_del(&glob->kobj);
424	kobject_put(&glob->kobj);
425}
426EXPORT_SYMBOL(ttm_mem_global_release);
427
428static void ttm_check_swapping(struct ttm_mem_global *glob)
429{
430	bool needs_swapping = false;
431	unsigned int i;
432	struct ttm_mem_zone *zone;
433
434	spin_lock(&glob->lock);
435	for (i = 0; i < glob->num_zones; ++i) {
436		zone = glob->zones[i];
437		if (zone->used_mem > zone->swap_limit) {
438			needs_swapping = true;
439			break;
440		}
441	}
442
443	spin_unlock(&glob->lock);
444
445	if (unlikely(needs_swapping))
446		(void)queue_work(glob->swap_queue, &glob->work);
447
448}
449
450static void ttm_mem_global_free_zone(struct ttm_mem_global *glob,
451				     struct ttm_mem_zone *single_zone,
452				     uint64_t amount)
453{
454	unsigned int i;
455	struct ttm_mem_zone *zone;
456
457	spin_lock(&glob->lock);
458	for (i = 0; i < glob->num_zones; ++i) {
459		zone = glob->zones[i];
460		if (single_zone && zone != single_zone)
461			continue;
462		zone->used_mem -= amount;
463	}
464	spin_unlock(&glob->lock);
465}
466
467void ttm_mem_global_free(struct ttm_mem_global *glob,
468			 uint64_t amount)
469{
470	return ttm_mem_global_free_zone(glob, NULL, amount);
471}
472EXPORT_SYMBOL(ttm_mem_global_free);
473
474static int ttm_mem_global_reserve(struct ttm_mem_global *glob,
475				  struct ttm_mem_zone *single_zone,
476				  uint64_t amount, bool reserve)
477{
478	uint64_t limit;
479	int ret = -ENOMEM;
480	unsigned int i;
481	struct ttm_mem_zone *zone;
482
483	spin_lock(&glob->lock);
484	for (i = 0; i < glob->num_zones; ++i) {
485		zone = glob->zones[i];
486		if (single_zone && zone != single_zone)
487			continue;
488
489		limit = (capable(CAP_SYS_ADMIN)) ?
490			zone->emer_mem : zone->max_mem;
491
492		if (zone->used_mem > limit)
493			goto out_unlock;
494	}
495
496	if (reserve) {
497		for (i = 0; i < glob->num_zones; ++i) {
498			zone = glob->zones[i];
499			if (single_zone && zone != single_zone)
500				continue;
501			zone->used_mem += amount;
502		}
503	}
504
505	ret = 0;
506out_unlock:
507	spin_unlock(&glob->lock);
508	ttm_check_swapping(glob);
509
510	return ret;
511}
512
513
514static int ttm_mem_global_alloc_zone(struct ttm_mem_global *glob,
515				     struct ttm_mem_zone *single_zone,
516				     uint64_t memory,
517				     bool no_wait, bool interruptible)
518{
519	int count = TTM_MEMORY_ALLOC_RETRIES;
520
521	while (unlikely(ttm_mem_global_reserve(glob,
522					       single_zone,
523					       memory, true)
524			!= 0)) {
525		if (no_wait)
526			return -ENOMEM;
527		if (unlikely(count-- == 0))
528			return -ENOMEM;
529		ttm_shrink(glob, false, memory + (memory >> 2) + 16);
530	}
531
532	return 0;
533}
534
535int ttm_mem_global_alloc(struct ttm_mem_global *glob, uint64_t memory,
536			 bool no_wait, bool interruptible)
537{
538	/**
539	 * Normal allocations of kernel memory are registered in
540	 * all zones.
541	 */
542
543	return ttm_mem_global_alloc_zone(glob, NULL, memory, no_wait,
544					 interruptible);
545}
546EXPORT_SYMBOL(ttm_mem_global_alloc);
547
548int ttm_mem_global_alloc_page(struct ttm_mem_global *glob,
549			      struct page *page,
550			      bool no_wait, bool interruptible)
551{
552
553	struct ttm_mem_zone *zone = NULL;
554
555	/**
556	 * Page allocations may be registed in a single zone
557	 * only if highmem or !dma32.
558	 */
559
560#ifdef CONFIG_HIGHMEM
561	if (PageHighMem(page) && glob->zone_highmem != NULL)
562		zone = glob->zone_highmem;
563#else
564	if (glob->zone_dma32 && page_to_pfn(page) > 0x00100000UL)
565		zone = glob->zone_kernel;
566#endif
567	return ttm_mem_global_alloc_zone(glob, zone, PAGE_SIZE, no_wait,
568					 interruptible);
569}
570
571void ttm_mem_global_free_page(struct ttm_mem_global *glob, struct page *page)
572{
573	struct ttm_mem_zone *zone = NULL;
574
575#ifdef CONFIG_HIGHMEM
576	if (PageHighMem(page) && glob->zone_highmem != NULL)
577		zone = glob->zone_highmem;
578#else
579	if (glob->zone_dma32 && page_to_pfn(page) > 0x00100000UL)
580		zone = glob->zone_kernel;
581#endif
582	ttm_mem_global_free_zone(glob, zone, PAGE_SIZE);
583}
584
585
586size_t ttm_round_pot(size_t size)
587{
588	if ((size & (size - 1)) == 0)
589		return size;
590	else if (size > PAGE_SIZE)
591		return PAGE_ALIGN(size);
592	else {
593		size_t tmp_size = 4;
594
595		while (tmp_size < size)
596			tmp_size <<= 1;
597
598		return tmp_size;
599	}
600	return 0;
601}
602EXPORT_SYMBOL(ttm_round_pot);
603