Merge branch 'slub/urgent' of git://git.kernel.org/pub/scm/linux/kernel/git/penberg...
[pandora-kernel.git] / 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 #include <linux/gfp.h>
47
48 #include <asm/page.h>
49 #include <asm/pgalloc.h>
50 #include <asm/pgtable.h>
51 #include <asm/uaccess.h>
52 #include <asm/tlb.h>
53
54 #include <asm/xen/hypervisor.h>
55 #include <asm/xen/hypercall.h>
56
57 #include <xen/xen.h>
58 #include <xen/interface/xen.h>
59 #include <xen/interface/memory.h>
60 #include <xen/xenbus.h>
61 #include <xen/features.h>
62 #include <xen/page.h>
63
64 #define PAGES2KB(_p) ((_p)<<(PAGE_SHIFT-10))
65
66 #define BALLOON_CLASS_NAME "xen_memory"
67
68 struct balloon_stats {
69         /* We aim for 'current allocation' == 'target allocation'. */
70         unsigned long current_pages;
71         unsigned long target_pages;
72         /*
73          * Drivers may alter the memory reservation independently, but they
74          * must inform the balloon driver so we avoid hitting the hard limit.
75          */
76         unsigned long driver_pages;
77         /* Number of pages in high- and low-memory balloons. */
78         unsigned long balloon_low;
79         unsigned long balloon_high;
80 };
81
82 static DEFINE_MUTEX(balloon_mutex);
83
84 static struct sys_device balloon_sysdev;
85
86 static int register_balloon(struct sys_device *sysdev);
87
88 /*
89  * Protects atomic reservation decrease/increase against concurrent increases.
90  * Also protects non-atomic updates of current_pages and driver_pages, and
91  * balloon lists.
92  */
93 static DEFINE_SPINLOCK(balloon_lock);
94
95 static struct balloon_stats balloon_stats;
96
97 /* We increase/decrease in batches which fit in a page */
98 static unsigned long frame_list[PAGE_SIZE / sizeof(unsigned long)];
99
100 #ifdef CONFIG_HIGHMEM
101 #define inc_totalhigh_pages() (totalhigh_pages++)
102 #define dec_totalhigh_pages() (totalhigh_pages--)
103 #else
104 #define inc_totalhigh_pages() do {} while(0)
105 #define dec_totalhigh_pages() do {} while(0)
106 #endif
107
108 /* List of ballooned pages, threaded through the mem_map array. */
109 static LIST_HEAD(ballooned_pages);
110
111 /* Main work function, always executed in process context. */
112 static void balloon_process(struct work_struct *work);
113 static DECLARE_WORK(balloon_worker, balloon_process);
114 static struct timer_list balloon_timer;
115
116 /* When ballooning out (allocating memory to return to Xen) we don't really
117    want the kernel to try too hard since that can trigger the oom killer. */
118 #define GFP_BALLOON \
119         (GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC)
120
121 static void scrub_page(struct page *page)
122 {
123 #ifdef CONFIG_XEN_SCRUB_PAGES
124         clear_highpage(page);
125 #endif
126 }
127
128 /* balloon_append: add the given page to the balloon. */
129 static void balloon_append(struct page *page)
130 {
131         /* Lowmem is re-populated first, so highmem pages go at list tail. */
132         if (PageHighMem(page)) {
133                 list_add_tail(&page->lru, &ballooned_pages);
134                 balloon_stats.balloon_high++;
135                 dec_totalhigh_pages();
136         } else {
137                 list_add(&page->lru, &ballooned_pages);
138                 balloon_stats.balloon_low++;
139         }
140
141         totalram_pages--;
142 }
143
144 /* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
145 static struct page *balloon_retrieve(void)
146 {
147         struct page *page;
148
149         if (list_empty(&ballooned_pages))
150                 return NULL;
151
152         page = list_entry(ballooned_pages.next, struct page, lru);
153         list_del(&page->lru);
154
155         if (PageHighMem(page)) {
156                 balloon_stats.balloon_high--;
157                 inc_totalhigh_pages();
158         }
159         else
160                 balloon_stats.balloon_low--;
161
162         totalram_pages++;
163
164         return page;
165 }
166
167 static struct page *balloon_first_page(void)
168 {
169         if (list_empty(&ballooned_pages))
170                 return NULL;
171         return list_entry(ballooned_pages.next, struct page, lru);
172 }
173
174 static struct page *balloon_next_page(struct page *page)
175 {
176         struct list_head *next = page->lru.next;
177         if (next == &ballooned_pages)
178                 return NULL;
179         return list_entry(next, struct page, lru);
180 }
181
182 static void balloon_alarm(unsigned long unused)
183 {
184         schedule_work(&balloon_worker);
185 }
186
187 static unsigned long current_target(void)
188 {
189         unsigned long target = balloon_stats.target_pages;
190
191         target = min(target,
192                      balloon_stats.current_pages +
193                      balloon_stats.balloon_low +
194                      balloon_stats.balloon_high);
195
196         return target;
197 }
198
199 static int increase_reservation(unsigned long nr_pages)
200 {
201         unsigned long  pfn, i, flags;
202         struct page   *page;
203         long           rc;
204         struct xen_memory_reservation reservation = {
205                 .address_bits = 0,
206                 .extent_order = 0,
207                 .domid        = DOMID_SELF
208         };
209
210         if (nr_pages > ARRAY_SIZE(frame_list))
211                 nr_pages = ARRAY_SIZE(frame_list);
212
213         spin_lock_irqsave(&balloon_lock, flags);
214
215         page = balloon_first_page();
216         for (i = 0; i < nr_pages; i++) {
217                 BUG_ON(page == NULL);
218                 frame_list[i] = page_to_pfn(page);
219                 page = balloon_next_page(page);
220         }
221
222         set_xen_guest_handle(reservation.extent_start, frame_list);
223         reservation.nr_extents = nr_pages;
224         rc = HYPERVISOR_memory_op(XENMEM_populate_physmap, &reservation);
225         if (rc < 0)
226                 goto out;
227
228         for (i = 0; i < rc; i++) {
229                 page = balloon_retrieve();
230                 BUG_ON(page == NULL);
231
232                 pfn = page_to_pfn(page);
233                 BUG_ON(!xen_feature(XENFEAT_auto_translated_physmap) &&
234                        phys_to_machine_mapping_valid(pfn));
235
236                 set_phys_to_machine(pfn, frame_list[i]);
237
238                 /* Link back into the page tables if not highmem. */
239                 if (pfn < max_low_pfn) {
240                         int ret;
241                         ret = HYPERVISOR_update_va_mapping(
242                                 (unsigned long)__va(pfn << PAGE_SHIFT),
243                                 mfn_pte(frame_list[i], PAGE_KERNEL),
244                                 0);
245                         BUG_ON(ret);
246                 }
247
248                 /* Relinquish the page back to the allocator. */
249                 ClearPageReserved(page);
250                 init_page_count(page);
251                 __free_page(page);
252         }
253
254         balloon_stats.current_pages += rc;
255
256  out:
257         spin_unlock_irqrestore(&balloon_lock, flags);
258
259         return rc < 0 ? rc : rc != nr_pages;
260 }
261
262 static int decrease_reservation(unsigned long nr_pages)
263 {
264         unsigned long  pfn, i, flags;
265         struct page   *page;
266         int            need_sleep = 0;
267         int ret;
268         struct xen_memory_reservation reservation = {
269                 .address_bits = 0,
270                 .extent_order = 0,
271                 .domid        = DOMID_SELF
272         };
273
274         if (nr_pages > ARRAY_SIZE(frame_list))
275                 nr_pages = ARRAY_SIZE(frame_list);
276
277         for (i = 0; i < nr_pages; i++) {
278                 if ((page = alloc_page(GFP_BALLOON)) == NULL) {
279                         nr_pages = i;
280                         need_sleep = 1;
281                         break;
282                 }
283
284                 pfn = page_to_pfn(page);
285                 frame_list[i] = pfn_to_mfn(pfn);
286
287                 scrub_page(page);
288
289                 if (!PageHighMem(page)) {
290                         ret = HYPERVISOR_update_va_mapping(
291                                 (unsigned long)__va(pfn << PAGE_SHIFT),
292                                 __pte_ma(0), 0);
293                         BUG_ON(ret);
294                 }
295
296         }
297
298         /* Ensure that ballooned highmem pages don't have kmaps. */
299         kmap_flush_unused();
300         flush_tlb_all();
301
302         spin_lock_irqsave(&balloon_lock, flags);
303
304         /* No more mappings: invalidate P2M and add to balloon. */
305         for (i = 0; i < nr_pages; i++) {
306                 pfn = mfn_to_pfn(frame_list[i]);
307                 set_phys_to_machine(pfn, INVALID_P2M_ENTRY);
308                 balloon_append(pfn_to_page(pfn));
309         }
310
311         set_xen_guest_handle(reservation.extent_start, frame_list);
312         reservation.nr_extents   = nr_pages;
313         ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation, &reservation);
314         BUG_ON(ret != nr_pages);
315
316         balloon_stats.current_pages -= nr_pages;
317
318         spin_unlock_irqrestore(&balloon_lock, flags);
319
320         return need_sleep;
321 }
322
323 /*
324  * We avoid multiple worker processes conflicting via the balloon mutex.
325  * We may of course race updates of the target counts (which are protected
326  * by the balloon lock), or with changes to the Xen hard limit, but we will
327  * recover from these in time.
328  */
329 static void balloon_process(struct work_struct *work)
330 {
331         int need_sleep = 0;
332         long credit;
333
334         mutex_lock(&balloon_mutex);
335
336         do {
337                 credit = current_target() - balloon_stats.current_pages;
338                 if (credit > 0)
339                         need_sleep = (increase_reservation(credit) != 0);
340                 if (credit < 0)
341                         need_sleep = (decrease_reservation(-credit) != 0);
342
343 #ifndef CONFIG_PREEMPT
344                 if (need_resched())
345                         schedule();
346 #endif
347         } while ((credit != 0) && !need_sleep);
348
349         /* Schedule more work if there is some still to be done. */
350         if (current_target() != balloon_stats.current_pages)
351                 mod_timer(&balloon_timer, jiffies + HZ);
352
353         mutex_unlock(&balloon_mutex);
354 }
355
356 /* Resets the Xen limit, sets new target, and kicks off processing. */
357 static void balloon_set_new_target(unsigned long target)
358 {
359         /* No need for lock. Not read-modify-write updates. */
360         balloon_stats.target_pages = target;
361         schedule_work(&balloon_worker);
362 }
363
364 static struct xenbus_watch target_watch =
365 {
366         .node = "memory/target"
367 };
368
369 /* React to a change in the target key */
370 static void watch_target(struct xenbus_watch *watch,
371                          const char **vec, unsigned int len)
372 {
373         unsigned long long new_target;
374         int err;
375
376         err = xenbus_scanf(XBT_NIL, "memory", "target", "%llu", &new_target);
377         if (err != 1) {
378                 /* This is ok (for domain0 at least) - so just return */
379                 return;
380         }
381
382         /* The given memory/target value is in KiB, so it needs converting to
383          * pages. PAGE_SHIFT converts bytes to pages, hence PAGE_SHIFT - 10.
384          */
385         balloon_set_new_target(new_target >> (PAGE_SHIFT - 10));
386 }
387
388 static int balloon_init_watcher(struct notifier_block *notifier,
389                                 unsigned long event,
390                                 void *data)
391 {
392         int err;
393
394         err = register_xenbus_watch(&target_watch);
395         if (err)
396                 printk(KERN_ERR "Failed to set balloon watcher\n");
397
398         return NOTIFY_DONE;
399 }
400
401 static struct notifier_block xenstore_notifier;
402
403 static int __init balloon_init(void)
404 {
405         unsigned long pfn;
406         struct page *page;
407
408         if (!xen_pv_domain())
409                 return -ENODEV;
410
411         pr_info("xen_balloon: Initialising balloon driver.\n");
412
413         balloon_stats.current_pages = min(xen_start_info->nr_pages, max_pfn);
414         balloon_stats.target_pages  = balloon_stats.current_pages;
415         balloon_stats.balloon_low   = 0;
416         balloon_stats.balloon_high  = 0;
417         balloon_stats.driver_pages  = 0UL;
418
419         init_timer(&balloon_timer);
420         balloon_timer.data = 0;
421         balloon_timer.function = balloon_alarm;
422
423         register_balloon(&balloon_sysdev);
424
425         /* Initialise the balloon with excess memory space. */
426         for (pfn = xen_start_info->nr_pages; pfn < max_pfn; pfn++) {
427                 page = pfn_to_page(pfn);
428                 if (!PageReserved(page))
429                         balloon_append(page);
430         }
431
432         target_watch.callback = watch_target;
433         xenstore_notifier.notifier_call = balloon_init_watcher;
434
435         register_xenstore_notifier(&xenstore_notifier);
436
437         return 0;
438 }
439
440 subsys_initcall(balloon_init);
441
442 static void balloon_exit(void)
443 {
444     /* XXX - release balloon here */
445     return;
446 }
447
448 module_exit(balloon_exit);
449
450 #define BALLOON_SHOW(name, format, args...)                             \
451         static ssize_t show_##name(struct sys_device *dev,              \
452                                    struct sysdev_attribute *attr,       \
453                                    char *buf)                           \
454         {                                                               \
455                 return sprintf(buf, format, ##args);                    \
456         }                                                               \
457         static SYSDEV_ATTR(name, S_IRUGO, show_##name, NULL)
458
459 BALLOON_SHOW(current_kb, "%lu\n", PAGES2KB(balloon_stats.current_pages));
460 BALLOON_SHOW(low_kb, "%lu\n", PAGES2KB(balloon_stats.balloon_low));
461 BALLOON_SHOW(high_kb, "%lu\n", PAGES2KB(balloon_stats.balloon_high));
462 BALLOON_SHOW(driver_kb, "%lu\n", PAGES2KB(balloon_stats.driver_pages));
463
464 static ssize_t show_target_kb(struct sys_device *dev, struct sysdev_attribute *attr,
465                               char *buf)
466 {
467         return sprintf(buf, "%lu\n", PAGES2KB(balloon_stats.target_pages));
468 }
469
470 static ssize_t store_target_kb(struct sys_device *dev,
471                                struct sysdev_attribute *attr,
472                                const char *buf,
473                                size_t count)
474 {
475         char *endchar;
476         unsigned long long target_bytes;
477
478         if (!capable(CAP_SYS_ADMIN))
479                 return -EPERM;
480
481         target_bytes = simple_strtoull(buf, &endchar, 0) * 1024;
482
483         balloon_set_new_target(target_bytes >> PAGE_SHIFT);
484
485         return count;
486 }
487
488 static SYSDEV_ATTR(target_kb, S_IRUGO | S_IWUSR,
489                    show_target_kb, store_target_kb);
490
491
492 static ssize_t show_target(struct sys_device *dev, struct sysdev_attribute *attr,
493                               char *buf)
494 {
495         return sprintf(buf, "%llu\n",
496                        (unsigned long long)balloon_stats.target_pages
497                        << PAGE_SHIFT);
498 }
499
500 static ssize_t store_target(struct sys_device *dev,
501                             struct sysdev_attribute *attr,
502                             const char *buf,
503                             size_t count)
504 {
505         char *endchar;
506         unsigned long long target_bytes;
507
508         if (!capable(CAP_SYS_ADMIN))
509                 return -EPERM;
510
511         target_bytes = memparse(buf, &endchar);
512
513         balloon_set_new_target(target_bytes >> PAGE_SHIFT);
514
515         return count;
516 }
517
518 static SYSDEV_ATTR(target, S_IRUGO | S_IWUSR,
519                    show_target, store_target);
520
521
522 static struct sysdev_attribute *balloon_attrs[] = {
523         &attr_target_kb,
524         &attr_target,
525 };
526
527 static struct attribute *balloon_info_attrs[] = {
528         &attr_current_kb.attr,
529         &attr_low_kb.attr,
530         &attr_high_kb.attr,
531         &attr_driver_kb.attr,
532         NULL
533 };
534
535 static struct attribute_group balloon_info_group = {
536         .name = "info",
537         .attrs = balloon_info_attrs,
538 };
539
540 static struct sysdev_class balloon_sysdev_class = {
541         .name = BALLOON_CLASS_NAME,
542 };
543
544 static int register_balloon(struct sys_device *sysdev)
545 {
546         int i, error;
547
548         error = sysdev_class_register(&balloon_sysdev_class);
549         if (error)
550                 return error;
551
552         sysdev->id = 0;
553         sysdev->cls = &balloon_sysdev_class;
554
555         error = sysdev_register(sysdev);
556         if (error) {
557                 sysdev_class_unregister(&balloon_sysdev_class);
558                 return error;
559         }
560
561         for (i = 0; i < ARRAY_SIZE(balloon_attrs); i++) {
562                 error = sysdev_create_file(sysdev, balloon_attrs[i]);
563                 if (error)
564                         goto fail;
565         }
566
567         error = sysfs_create_group(&sysdev->kobj, &balloon_info_group);
568         if (error)
569                 goto fail;
570
571         return 0;
572
573  fail:
574         while (--i >= 0)
575                 sysdev_remove_file(sysdev, balloon_attrs[i]);
576         sysdev_unregister(sysdev);
577         sysdev_class_unregister(&balloon_sysdev_class);
578         return error;
579 }
580
581 MODULE_LICENSE("GPL");