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#define pr_fmt(fmt) "[TTM] " fmt
29
30#include "ttm/ttm_memory.h"
31#include "ttm/ttm_module.h"
32#include "ttm/ttm_page_alloc.h"
33#include <linux/spinlock.h>
34#include <linux/sched.h>
35#include <linux/wait.h>
36#include <linux/mm.h>
37#include <linux/module.h>
38#include <linux/slab.h>
39
40#define TTM_MEMORY_ALLOC_RETRIES 4
41
42struct ttm_mem_zone {
43	struct kobject kobj;
44	struct ttm_mem_global *glob;
45	const char *name;
46	uint64_t zone_mem;
47	uint64_t emer_mem;
48	uint64_t max_mem;
49	uint64_t swap_limit;
50	uint64_t used_mem;
51};
52
53static struct attribute ttm_mem_sys = {
54	.name = "zone_memory",
55	.mode = S_IRUGO
56};
57static struct attribute ttm_mem_emer = {
58	.name = "emergency_memory",
59	.mode = S_IRUGO | S_IWUSR
60};
61static struct attribute ttm_mem_max = {
62	.name = "available_memory",
63	.mode = S_IRUGO | S_IWUSR
64};
65static struct attribute ttm_mem_swap = {
66	.name = "swap_limit",
67	.mode = S_IRUGO | S_IWUSR
68};
69static struct attribute ttm_mem_used = {
70	.name = "used_memory",
71	.mode = S_IRUGO
72};
73
74static void ttm_mem_zone_kobj_release(struct kobject *kobj)
75{
76	struct ttm_mem_zone *zone =
77		container_of(kobj, struct ttm_mem_zone, kobj);
78
79	pr_info("Zone %7s: Used memory at exit: %llu kiB\n",
80		zone->name, (unsigned long long)zone->used_mem >> 10);
81	kfree(zone);
82}
83
84static ssize_t ttm_mem_zone_show(struct kobject *kobj,
85				 struct attribute *attr,
86				 char *buffer)
87{
88	struct ttm_mem_zone *zone =
89		container_of(kobj, struct ttm_mem_zone, kobj);
90	uint64_t val = 0;
91
92	spin_lock(&zone->glob->lock);
93	if (attr == &ttm_mem_sys)
94		val = zone->zone_mem;
95	else if (attr == &ttm_mem_emer)
96		val = zone->emer_mem;
97	else if (attr == &ttm_mem_max)
98		val = zone->max_mem;
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);
104
105	return snprintf(buffer, PAGE_SIZE, "%llu\n",
106			(unsigned long long) val >> 10);
107}
108
109static void ttm_check_swapping(struct ttm_mem_global *glob);
110
111static ssize_t ttm_mem_zone_store(struct kobject *kobj,
112				  struct attribute *attr,
113				  const char *buffer,
114				  size_t size)
115{
116	struct ttm_mem_zone *zone =
117		container_of(kobj, struct ttm_mem_zone, kobj);
118	int chars;
119	unsigned long val;
120	uint64_t val64;
121
122	chars = sscanf(buffer, "%lu", &val);
123	if (chars == 0)
124		return size;
125
126	val64 = val;
127	val64 <<= 10;
128
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);
143
144	ttm_check_swapping(zone->glob);
145
146	return size;
147}
148
149static struct attribute *ttm_mem_zone_attrs[] = {
150	&ttm_mem_sys,
151	&ttm_mem_emer,
152	&ttm_mem_max,
153	&ttm_mem_swap,
154	&ttm_mem_used,
155	NULL
156};
157
158static const struct sysfs_ops ttm_mem_zone_ops = {
159	.show = &ttm_mem_zone_show,
160	.store = &ttm_mem_zone_store
161};
162
163static 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,
167};
168
169static void ttm_mem_global_kobj_release(struct kobject *kobj)
170{
171	struct ttm_mem_global *glob =
172		container_of(kobj, struct ttm_mem_global, kobj);
173
174	kfree(glob);
175}
176
177static struct kobj_type ttm_mem_glob_kobj_type = {
178	.release = &ttm_mem_global_kobj_release,
179};
180
181static bool ttm_zones_above_swap_target(struct ttm_mem_global *glob,
182					bool from_wq, uint64_t extra)
183{
184	unsigned int i;
185	struct ttm_mem_zone *zone;
186	uint64_t target;
187
188	for (i = 0; i < glob->num_zones; ++i) {
189		zone = glob->zones[i];
190
191		if (from_wq)
192			target = zone->swap_limit;
193		else if (capable(CAP_SYS_ADMIN))
194			target = zone->emer_mem;
195		else
196			target = zone->max_mem;
197
198		target = (extra > target) ? 0ULL : target;
199
200		if (zone->used_mem > target)
201			return true;
202	}
203	return false;
204}
205
206/**
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.
211 */
212
213static void ttm_shrink(struct ttm_mem_global *glob, bool from_wq,
214		       uint64_t extra)
215{
216	int ret;
217	struct ttm_mem_shrink *shrink;
218
219	spin_lock(&glob->lock);
220	if (glob->shrink == NULL)
221		goto out;
222
223	while (ttm_zones_above_swap_target(glob, from_wq, extra)) {
224		shrink = glob->shrink;
225		spin_unlock(&glob->lock);
226		ret = shrink->do_shrink(shrink);
227		spin_lock(&glob->lock);
228		if (unlikely(ret != 0))
229			goto out;
230	}
231out:
232	spin_unlock(&glob->lock);
233}
234
235
236
237static void ttm_shrink_work(struct work_struct *work)
238{
239	struct ttm_mem_global *glob =
240	    container_of(work, struct ttm_mem_global, work);
241
242	ttm_shrink(glob, true, 0ULL);
243}
244
245static int ttm_mem_init_kernel_zone(struct ttm_mem_global *glob,
246				    const struct sysinfo *si)
247{
248	struct ttm_mem_zone *zone = kzalloc(sizeof(*zone), GFP_KERNEL);
249	uint64_t mem;
250	int ret;
251
252	if (unlikely(!zone))
253		return -ENOMEM;
254
255	mem = si->totalram - si->totalhigh;
256	mem *= si->mem_unit;
257
258	zone->name = "kernel";
259	zone->zone_mem = mem;
260	zone->max_mem = mem >> 1;
261	zone->emer_mem = (mem >> 1) + (mem >> 2);
262	zone->swap_limit = zone->max_mem - (mem >> 3);
263	zone->used_mem = 0;
264	zone->glob = glob;
265	glob->zone_kernel = zone;
266	ret = kobject_init_and_add(
267		&zone->kobj, &ttm_mem_zone_kobj_type, &glob->kobj, zone->name);
268	if (unlikely(ret != 0)) {
269		kobject_put(&zone->kobj);
270		return ret;
271	}
272	glob->zones[glob->num_zones++] = zone;
273	return 0;
274}
275
276#ifdef CONFIG_HIGHMEM
277static int ttm_mem_init_highmem_zone(struct ttm_mem_global *glob,
278				     const struct sysinfo *si)
279{
280	struct ttm_mem_zone *zone;
281	uint64_t mem;
282	int ret;
283
284	if (si->totalhigh == 0)
285		return 0;
286
287	zone = kzalloc(sizeof(*zone), GFP_KERNEL);
288	if (unlikely(!zone))
289		return -ENOMEM;
290
291	mem = si->totalram;
292	mem *= si->mem_unit;
293
294	zone->name = "highmem";
295	zone->zone_mem = mem;
296	zone->max_mem = mem >> 1;
297	zone->emer_mem = (mem >> 1) + (mem >> 2);
298	zone->swap_limit = zone->max_mem - (mem >> 3);
299	zone->used_mem = 0;
300	zone->glob = glob;
301	glob->zone_highmem = zone;
302	ret = kobject_init_and_add(
303		&zone->kobj, &ttm_mem_zone_kobj_type, &glob->kobj, zone->name);
304	if (unlikely(ret != 0)) {
305		kobject_put(&zone->kobj);
306		return ret;
307	}
308	glob->zones[glob->num_zones++] = zone;
309	return 0;
310}
311#else
312static int ttm_mem_init_dma32_zone(struct ttm_mem_global *glob,
313				   const struct sysinfo *si)
314{
315	struct ttm_mem_zone *zone = kzalloc(sizeof(*zone), GFP_KERNEL);
316	uint64_t mem;
317	int ret;
318
319	if (unlikely(!zone))
320		return -ENOMEM;
321
322	mem = si->totalram;
323	mem *= si->mem_unit;
324
325	/**
326	 * No special dma32 zone needed.
327	 */
328
329	if (mem <= ((uint64_t) 1ULL << 32)) {
330		kfree(zone);
331		return 0;
332	}
333
334	/*
335	 * Limit max dma32 memory to 4GB for now
336	 * until we can figure out how big this
337	 * zone really is.
338	 */
339
340	mem = ((uint64_t) 1ULL << 32);
341	zone->name = "dma32";
342	zone->zone_mem = mem;
343	zone->max_mem = mem >> 1;
344	zone->emer_mem = (mem >> 1) + (mem >> 2);
345	zone->swap_limit = zone->max_mem - (mem >> 3);
346	zone->used_mem = 0;
347	zone->glob = glob;
348	glob->zone_dma32 = zone;
349	ret = kobject_init_and_add(
350		&zone->kobj, &ttm_mem_zone_kobj_type, &glob->kobj, zone->name);
351	if (unlikely(ret != 0)) {
352		kobject_put(&zone->kobj);
353		return ret;
354	}
355	glob->zones[glob->num_zones++] = zone;
356	return 0;
357}
358#endif
359
360int ttm_mem_global_init(struct ttm_mem_global *glob)
361{
362	struct sysinfo si;
363	int ret;
364	int i;
365	struct ttm_mem_zone *zone;
366
367	spin_lock_init(&glob->lock);
368	glob->swap_queue = create_singlethread_workqueue("ttm_swap");
369	INIT_WORK(&glob->work, ttm_shrink_work);
370	init_waitqueue_head(&glob->queue);
371	ret = kobject_init_and_add(
372		&glob->kobj, &ttm_mem_glob_kobj_type, ttm_get_kobj(), "memory_accounting");
373	if (unlikely(ret != 0)) {
374		kobject_put(&glob->kobj);
375		return ret;
376	}
377
378	si_meminfo(&si);
379
380	ret = ttm_mem_init_kernel_zone(glob, &si);
381	if (unlikely(ret != 0))
382		goto out_no_zone;
383#ifdef CONFIG_HIGHMEM
384	ret = ttm_mem_init_highmem_zone(glob, &si);
385	if (unlikely(ret != 0))
386		goto out_no_zone;
387#else
388	ret = ttm_mem_init_dma32_zone(glob, &si);
389	if (unlikely(ret != 0))
390		goto out_no_zone;
391#endif
392	for (i = 0; i < glob->num_zones; ++i) {
393		zone = glob->zones[i];
394		pr_info("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