dm snapshot: separate out exception store interface
[linux-2.6] / drivers / xen / balloon.c
1 /******************************************************************************
2  * balloon.c
3  *
4  * Xen balloon driver - enables returning/claiming memory to/from Xen.
5  *
6  * Copyright (c) 2003, B Dragovic
7  * Copyright (c) 2003-2004, M Williamson, K Fraser
8  * Copyright (c) 2005 Dan M. Smith, IBM Corporation
9  *
10  * This program is free software; you can redistribute it and/or
11  * modify it under the terms of the GNU General Public License version 2
12  * as published by the Free Software Foundation; or, when distributed
13  * separately from the Linux kernel or incorporated into other
14  * software packages, subject to the following license:
15  *
16  * Permission is hereby granted, free of charge, to any person obtaining a copy
17  * of this source file (the "Software"), to deal in the Software without
18  * restriction, including without limitation the rights to use, copy, modify,
19  * merge, publish, distribute, sublicense, and/or sell copies of the Software,
20  * and to permit persons to whom the Software is furnished to do so, subject to
21  * the following conditions:
22  *
23  * The above copyright notice and this permission notice shall be included in
24  * all copies or substantial portions of the Software.
25  *
26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
27  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
28  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
29  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
30  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
31  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
32  * IN THE SOFTWARE.
33  */
34
35 #include <linux/kernel.h>
36 #include <linux/module.h>
37 #include <linux/sched.h>
38 #include <linux/errno.h>
39 #include <linux/mm.h>
40 #include <linux/bootmem.h>
41 #include <linux/pagemap.h>
42 #include <linux/highmem.h>
43 #include <linux/mutex.h>
44 #include <linux/list.h>
45 #include <linux/sysdev.h>
46
47 #include <asm/page.h>
48 #include <asm/pgalloc.h>
49 #include <asm/pgtable.h>
50 #include <asm/uaccess.h>
51 #include <asm/tlb.h>
52
53 #include <asm/xen/hypervisor.h>
54 #include <asm/xen/hypercall.h>
55 #include <xen/interface/xen.h>
56 #include <xen/interface/memory.h>
57 #include <xen/xenbus.h>
58 #include <xen/features.h>
59 #include <xen/page.h>
60
61 #define PAGES2KB(_p) ((_p)<<(PAGE_SHIFT-10))
62
63 #define BALLOON_CLASS_NAME "xen_memory"
64
65 struct balloon_stats {
66         /* We aim for 'current allocation' == 'target allocation'. */
67         unsigned long current_pages;
68         unsigned long target_pages;
69         /* We may hit the hard limit in Xen. If we do then we remember it. */
70         unsigned long hard_limit;
71         /*
72          * Drivers may alter the memory reservation independently, but they
73          * must inform the balloon driver so we avoid hitting the hard limit.
74          */
75         unsigned long driver_pages;
76         /* Number of pages in high- and low-memory balloons. */
77         unsigned long balloon_low;
78         unsigned long balloon_high;
79 };
80
81 static DEFINE_MUTEX(balloon_mutex);
82
83 static struct sys_device balloon_sysdev;
84
85 static int register_balloon(struct sys_device *sysdev);
86
87 /*
88  * Protects atomic reservation decrease/increase against concurrent increases.
89  * Also protects non-atomic updates of current_pages and driver_pages, and
90  * balloon lists.
91  */
92 static DEFINE_SPINLOCK(balloon_lock);
93
94 static struct balloon_stats balloon_stats;
95
96 /* We increase/decrease in batches which fit in a page */
97 static unsigned long frame_list[PAGE_SIZE / sizeof(unsigned long)];
98
99 /* VM /proc information for memory */
100 extern unsigned long totalram_pages;
101
102 #ifdef CONFIG_HIGHMEM
103 extern unsigned long totalhigh_pages;
104 #define inc_totalhigh_pages() (totalhigh_pages++)
105 #define dec_totalhigh_pages() (totalhigh_pages--)
106 #else
107 #define inc_totalhigh_pages() do {} while(0)
108 #define dec_totalhigh_pages() do {} while(0)
109 #endif
110
111 /* List of ballooned pages, threaded through the mem_map array. */
112 static LIST_HEAD(ballooned_pages);
113
114 /* Main work function, always executed in process context. */
115 static void balloon_process(struct work_struct *work);
116 static DECLARE_WORK(balloon_worker, balloon_process);
117 static struct timer_list balloon_timer;
118
119 /* When ballooning out (allocating memory to return to Xen) we don't really
120    want the kernel to try too hard since that can trigger the oom killer. */
121 #define GFP_BALLOON \
122         (GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC)
123
124 static void scrub_page(struct page *page)
125 {
126 #ifdef CONFIG_XEN_SCRUB_PAGES
127         clear_highpage(page);
128 #endif
129 }
130
131 /* balloon_append: add the given page to the balloon. */
132 static void balloon_append(struct page *page)
133 {
134         /* Lowmem is re-populated first, so highmem pages go at list tail. */
135         if (PageHighMem(page)) {
136                 list_add_tail(&page->lru, &ballooned_pages);
137                 balloon_stats.balloon_high++;
138                 dec_totalhigh_pages();
139         } else {
140                 list_add(&page->lru, &ballooned_pages);
141                 balloon_stats.balloon_low++;
142         }
143 }
144
145 /* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
146 static struct page *balloon_retrieve(void)
147 {
148         struct page *page;
149
150         if (list_empty(&ballooned_pages))
151                 return NULL;
152
153         page = list_entry(ballooned_pages.next, struct page, lru);
154         list_del(&page->lru);
155
156         if (PageHighMem(page)) {
157                 balloon_stats.balloon_high--;
158                 inc_totalhigh_pages();
159         }
160         else
161                 balloon_stats.balloon_low--;
162
163         return page;
164 }
165
166 static struct page *balloon_first_page(void)
167 {
168         if (list_empty(&ballooned_pages))
169                 return NULL;
170         return list_entry(ballooned_pages.next, struct page, lru);
171 }
172
173 static struct page *balloon_next_page(struct page *page)
174 {
175         struct list_head *next = page->lru.next;
176         if (next == &ballooned_pages)
177                 return NULL;
178         return list_entry(next, struct page, lru);
179 }
180
181 static void balloon_alarm(unsigned long unused)
182 {
183         schedule_work(&balloon_worker);
184 }
185
186 static unsigned long current_target(void)
187 {
188         unsigned long target = min(balloon_stats.target_pages, balloon_stats.hard_limit);
189
190         target = min(target,
191                      balloon_stats.current_pages +
192                      balloon_stats.balloon_low +
193                      balloon_stats.balloon_high);
194
195         return target;
196 }
197
198 static int increase_reservation(unsigned long nr_pages)
199 {
200         unsigned long  pfn, i, flags;
201         struct page   *page;
202         long           rc;
203         struct xen_memory_reservation reservation = {
204                 .address_bits = 0,
205                 .extent_order = 0,
206                 .domid        = DOMID_SELF
207         };
208
209         if (nr_pages > ARRAY_SIZE(frame_list))
210                 nr_pages = ARRAY_SIZE(frame_list);
211
212         spin_lock_irqsave(&balloon_lock, flags);
213
214         page = balloon_first_page();
215         for (i = 0; i < nr_pages; i++) {
216                 BUG_ON(page == NULL);
217                 frame_list[i] = page_to_pfn(page);;
218                 page = balloon_next_page(page);
219         }
220
221         set_xen_guest_handle(reservation.extent_start, frame_list);
222         reservation.nr_extents = nr_pages;
223         rc = HYPERVISOR_memory_op(XENMEM_populate_physmap, &reservation);
224         if (rc < nr_pages) {
225                 if (rc > 0) {
226                         int ret;
227
228                         /* We hit the Xen hard limit: reprobe. */
229                         reservation.nr_extents = rc;
230                         ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation,
231                                                    &reservation);
232                         BUG_ON(ret != rc);
233                 }
234                 if (rc >= 0)
235                         balloon_stats.hard_limit = (balloon_stats.current_pages + rc -
236                                                     balloon_stats.driver_pages);
237                 goto out;
238         }
239
240         for (i = 0; i < nr_pages; i++) {
241                 page = balloon_retrieve();
242                 BUG_ON(page == NULL);
243
244                 pfn = page_to_pfn(page);
245                 BUG_ON(!xen_feature(XENFEAT_auto_translated_physmap) &&
246                        phys_to_machine_mapping_valid(pfn));
247
248                 set_phys_to_machine(pfn, frame_list[i]);
249
250                 /* Link back into the page tables if not highmem. */
251                 if (pfn < max_low_pfn) {
252                         int ret;
253                         ret = HYPERVISOR_update_va_mapping(
254                                 (unsigned long)__va(pfn << PAGE_SHIFT),
255                                 mfn_pte(frame_list[i], PAGE_KERNEL),
256                                 0);
257                         BUG_ON(ret);
258                 }
259
260                 /* Relinquish the page back to the allocator. */
261                 ClearPageReserved(page);
262                 init_page_count(page);
263                 __free_page(page);
264         }
265
266         balloon_stats.current_pages += nr_pages;
267         totalram_pages = balloon_stats.current_pages;
268
269  out:
270         spin_unlock_irqrestore(&balloon_lock, flags);
271
272         return 0;
273 }
274
275 static int decrease_reservation(unsigned long nr_pages)
276 {
277         unsigned long  pfn, i, flags;
278         struct page   *page;
279         int            need_sleep = 0;
280         int ret;
281         struct xen_memory_reservation reservation = {
282                 .address_bits = 0,
283                 .extent_order = 0,
284                 .domid        = DOMID_SELF
285         };
286
287         if (nr_pages > ARRAY_SIZE(frame_list))
288                 nr_pages = ARRAY_SIZE(frame_list);
289
290         for (i = 0; i < nr_pages; i++) {
291                 if ((page = alloc_page(GFP_BALLOON)) == NULL) {
292                         nr_pages = i;
293                         need_sleep = 1;
294                         break;
295                 }
296
297                 pfn = page_to_pfn(page);
298                 frame_list[i] = pfn_to_mfn(pfn);
299
300                 scrub_page(page);
301         }
302
303         /* Ensure that ballooned highmem pages don't have kmaps. */
304         kmap_flush_unused();
305         flush_tlb_all();
306
307         spin_lock_irqsave(&balloon_lock, flags);
308
309         /* No more mappings: invalidate P2M and add to balloon. */
310         for (i = 0; i < nr_pages; i++) {
311                 pfn = mfn_to_pfn(frame_list[i]);
312                 set_phys_to_machine(pfn, INVALID_P2M_ENTRY);
313                 balloon_append(pfn_to_page(pfn));
314         }
315
316         set_xen_guest_handle(reservation.extent_start, frame_list);
317         reservation.nr_extents   = nr_pages;
318         ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation, &reservation);
319         BUG_ON(ret != nr_pages);
320
321         balloon_stats.current_pages -= nr_pages;
322         totalram_pages = balloon_stats.current_pages;
323
324         spin_unlock_irqrestore(&balloon_lock, flags);
325
326         return need_sleep;
327 }
328
329 /*
330  * We avoid multiple worker processes conflicting via the balloon mutex.
331  * We may of course race updates of the target counts (which are protected
332  * by the balloon lock), or with changes to the Xen hard limit, but we will
333  * recover from these in time.
334  */
335 static void balloon_process(struct work_struct *work)
336 {
337         int need_sleep = 0;
338         long credit;
339
340         mutex_lock(&balloon_mutex);
341
342         do {
343                 credit = current_target() - balloon_stats.current_pages;
344                 if (credit > 0)
345                         need_sleep = (increase_reservation(credit) != 0);
346                 if (credit < 0)
347                         need_sleep = (decrease_reservation(-credit) != 0);
348
349 #ifndef CONFIG_PREEMPT
350                 if (need_resched())
351                         schedule();
352 #endif
353         } while ((credit != 0) && !need_sleep);
354
355         /* Schedule more work if there is some still to be done. */
356         if (current_target() != balloon_stats.current_pages)
357                 mod_timer(&balloon_timer, jiffies + HZ);
358
359         mutex_unlock(&balloon_mutex);
360 }
361
362 /* Resets the Xen limit, sets new target, and kicks off processing. */
363 static void balloon_set_new_target(unsigned long target)
364 {
365         /* No need for lock. Not read-modify-write updates. */
366         balloon_stats.hard_limit   = ~0UL;
367         balloon_stats.target_pages = target;
368         schedule_work(&balloon_worker);
369 }
370
371 static struct xenbus_watch target_watch =
372 {
373         .node = "memory/target"
374 };
375
376 /* React to a change in the target key */
377 static void watch_target(struct xenbus_watch *watch,
378                          const char **vec, unsigned int len)
379 {
380         unsigned long long new_target;
381         int err;
382
383         err = xenbus_scanf(XBT_NIL, "memory", "target", "%llu", &new_target);
384         if (err != 1) {
385                 /* This is ok (for domain0 at least) - so just return */
386                 return;
387         }
388
389         /* The given memory/target value is in KiB, so it needs converting to
390          * pages. PAGE_SHIFT converts bytes to pages, hence PAGE_SHIFT - 10.
391          */
392         balloon_set_new_target(new_target >> (PAGE_SHIFT - 10));
393 }
394
395 static int balloon_init_watcher(struct notifier_block *notifier,
396                                 unsigned long event,
397                                 void *data)
398 {
399         int err;
400
401         err = register_xenbus_watch(&target_watch);
402         if (err)
403                 printk(KERN_ERR "Failed to set balloon watcher\n");
404
405         return NOTIFY_DONE;
406 }
407
408 static struct notifier_block xenstore_notifier;
409
410 static int __init balloon_init(void)
411 {
412         unsigned long pfn;
413         struct page *page;
414
415         if (!xen_pv_domain())
416                 return -ENODEV;
417
418         pr_info("xen_balloon: Initialising balloon driver.\n");
419
420         balloon_stats.current_pages = min(xen_start_info->nr_pages, max_pfn);
421         totalram_pages   = balloon_stats.current_pages;
422         balloon_stats.target_pages  = balloon_stats.current_pages;
423         balloon_stats.balloon_low   = 0;
424         balloon_stats.balloon_high  = 0;
425         balloon_stats.driver_pages  = 0UL;
426         balloon_stats.hard_limit    = ~0UL;
427
428         init_timer(&balloon_timer);
429         balloon_timer.data = 0;
430         balloon_timer.function = balloon_alarm;
431
432         register_balloon(&balloon_sysdev);
433
434         /* Initialise the balloon with excess memory space. */
435         for (pfn = xen_start_info->nr_pages; pfn < max_pfn; pfn++) {
436                 page = pfn_to_page(pfn);
437                 if (!PageReserved(page))
438                         balloon_append(page);
439         }
440
441         target_watch.callback = watch_target;
442         xenstore_notifier.notifier_call = balloon_init_watcher;
443
444         register_xenstore_notifier(&xenstore_notifier);
445
446         return 0;
447 }
448
449 subsys_initcall(balloon_init);
450
451 static void balloon_exit(void)
452 {
453     /* XXX - release balloon here */
454     return;
455 }
456
457 module_exit(balloon_exit);
458
459 #define BALLOON_SHOW(name, format, args...)                             \
460         static ssize_t show_##name(struct sys_device *dev,              \
461                                    struct sysdev_attribute *attr,       \
462                                    char *buf)                           \
463         {                                                               \
464                 return sprintf(buf, format, ##args);                    \
465         }                                                               \
466         static SYSDEV_ATTR(name, S_IRUGO, show_##name, NULL)
467
468 BALLOON_SHOW(current_kb, "%lu\n", PAGES2KB(balloon_stats.current_pages));
469 BALLOON_SHOW(low_kb, "%lu\n", PAGES2KB(balloon_stats.balloon_low));
470 BALLOON_SHOW(high_kb, "%lu\n", PAGES2KB(balloon_stats.balloon_high));
471 BALLOON_SHOW(hard_limit_kb,
472              (balloon_stats.hard_limit!=~0UL) ? "%lu\n" : "???\n",
473              (balloon_stats.hard_limit!=~0UL) ? PAGES2KB(balloon_stats.hard_limit) : 0);
474 BALLOON_SHOW(driver_kb, "%lu\n", PAGES2KB(balloon_stats.driver_pages));
475
476 static ssize_t show_target_kb(struct sys_device *dev, struct sysdev_attribute *attr,
477                               char *buf)
478 {
479         return sprintf(buf, "%lu\n", PAGES2KB(balloon_stats.target_pages));
480 }
481
482 static ssize_t store_target_kb(struct sys_device *dev,
483                                struct sysdev_attribute *attr,
484                                const char *buf,
485                                size_t count)
486 {
487         char *endchar;
488         unsigned long long target_bytes;
489
490         if (!capable(CAP_SYS_ADMIN))
491                 return -EPERM;
492
493         target_bytes = memparse(buf, &endchar);
494
495         balloon_set_new_target(target_bytes >> PAGE_SHIFT);
496
497         return count;
498 }
499
500 static SYSDEV_ATTR(target_kb, S_IRUGO | S_IWUSR,
501                    show_target_kb, store_target_kb);
502
503 static struct sysdev_attribute *balloon_attrs[] = {
504         &attr_target_kb,
505 };
506
507 static struct attribute *balloon_info_attrs[] = {
508         &attr_current_kb.attr,
509         &attr_low_kb.attr,
510         &attr_high_kb.attr,
511         &attr_hard_limit_kb.attr,
512         &attr_driver_kb.attr,
513         NULL
514 };
515
516 static struct attribute_group balloon_info_group = {
517         .name = "info",
518         .attrs = balloon_info_attrs,
519 };
520
521 static struct sysdev_class balloon_sysdev_class = {
522         .name = BALLOON_CLASS_NAME,
523 };
524
525 static int register_balloon(struct sys_device *sysdev)
526 {
527         int i, error;
528
529         error = sysdev_class_register(&balloon_sysdev_class);
530         if (error)
531                 return error;
532
533         sysdev->id = 0;
534         sysdev->cls = &balloon_sysdev_class;
535
536         error = sysdev_register(sysdev);
537         if (error) {
538                 sysdev_class_unregister(&balloon_sysdev_class);
539                 return error;
540         }
541
542         for (i = 0; i < ARRAY_SIZE(balloon_attrs); i++) {
543                 error = sysdev_create_file(sysdev, balloon_attrs[i]);
544                 if (error)
545                         goto fail;
546         }
547
548         error = sysfs_create_group(&sysdev->kobj, &balloon_info_group);
549         if (error)
550                 goto fail;
551
552         return 0;
553
554  fail:
555         while (--i >= 0)
556                 sysdev_remove_file(sysdev, balloon_attrs[i]);
557         sysdev_unregister(sysdev);
558         sysdev_class_unregister(&balloon_sysdev_class);
559         return error;
560 }
561
562 MODULE_LICENSE("GPL");