Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/sparc-2.6
[linux-2.6.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
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
56 #include <xen/xen.h>
57 #include <xen/interface/xen.h>
58 #include <xen/interface/memory.h>
59 #include <xen/xenbus.h>
60 #include <xen/features.h>
61 #include <xen/page.h>
62
63 #define PAGES2KB(_p) ((_p)<<(PAGE_SHIFT-10))
64
65 #define BALLOON_CLASS_NAME "xen_memory"
66
67 struct balloon_stats {
68         /* We aim for 'current allocation' == 'target allocation'. */
69         unsigned long current_pages;
70         unsigned long target_pages;
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 #ifdef CONFIG_HIGHMEM
100 #define inc_totalhigh_pages() (totalhigh_pages++)
101 #define dec_totalhigh_pages() (totalhigh_pages--)
102 #else
103 #define inc_totalhigh_pages() do {} while(0)
104 #define dec_totalhigh_pages() do {} while(0)
105 #endif
106
107 /* List of ballooned pages, threaded through the mem_map array. */
108 static LIST_HEAD(ballooned_pages);
109
110 /* Main work function, always executed in process context. */
111 static void balloon_process(struct work_struct *work);
112 static DECLARE_WORK(balloon_worker, balloon_process);
113 static struct timer_list balloon_timer;
114
115 /* When ballooning out (allocating memory to return to Xen) we don't really
116    want the kernel to try too hard since that can trigger the oom killer. */
117 #define GFP_BALLOON \
118         (GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC)
119
120 static void scrub_page(struct page *page)
121 {
122 #ifdef CONFIG_XEN_SCRUB_PAGES
123         clear_highpage(page);
124 #endif
125 }
126
127 /* balloon_append: add the given page to the balloon. */
128 static void balloon_append(struct page *page)
129 {
130         /* Lowmem is re-populated first, so highmem pages go at list tail. */
131         if (PageHighMem(page)) {
132                 list_add_tail(&page->lru, &ballooned_pages);
133                 balloon_stats.balloon_high++;
134                 dec_totalhigh_pages();
135         } else {
136                 list_add(&page->lru, &ballooned_pages);
137                 balloon_stats.balloon_low++;
138         }
139
140         totalram_pages--;
141 }
142
143 /* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
144 static struct page *balloon_retrieve(void)
145 {
146         struct page *page;
147
148         if (list_empty(&ballooned_pages))
149                 return NULL;
150
151         page = list_entry(ballooned_pages.next, struct page, lru);
152         list_del(&page->lru);
153
154         if (PageHighMem(page)) {
155                 balloon_stats.balloon_high--;
156                 inc_totalhigh_pages();
157         }
158         else
159                 balloon_stats.balloon_low--;
160
161         totalram_pages++;
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 = balloon_stats.target_pages;
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 < 0)
225                 goto out;
226
227         for (i = 0; i < rc; i++) {
228                 page = balloon_retrieve();
229                 BUG_ON(page == NULL);
230
231                 pfn = page_to_pfn(page);
232                 BUG_ON(!xen_feature(XENFEAT_auto_translated_physmap) &&
233                        phys_to_machine_mapping_valid(pfn));
234
235                 set_phys_to_machine(pfn, frame_list[i]);
236
237                 /* Link back into the page tables if not highmem. */
238                 if (pfn < max_low_pfn) {
239                         int ret;
240                         ret = HYPERVISOR_update_va_mapping(
241                                 (unsigned long)__va(pfn << PAGE_SHIFT),
242                                 mfn_pte(frame_list[i], PAGE_KERNEL),
243                                 0);
244                         BUG_ON(ret);
245                 }
246
247                 /* Relinquish the page back to the allocator. */
248                 ClearPageReserved(page);
249                 init_page_count(page);
250                 __free_page(page);
251         }
252
253         balloon_stats.current_pages += rc;
254
255  out:
256         spin_unlock_irqrestore(&balloon_lock, flags);
257
258         return rc < 0 ? rc : rc != nr_pages;
259 }
260
261 static int decrease_reservation(unsigned long nr_pages)
262 {
263         unsigned long  pfn, i, flags;
264         struct page   *page;
265         int            need_sleep = 0;
266         int ret;
267         struct xen_memory_reservation reservation = {
268                 .address_bits = 0,
269                 .extent_order = 0,
270                 .domid        = DOMID_SELF
271         };
272
273         if (nr_pages > ARRAY_SIZE(frame_list))
274                 nr_pages = ARRAY_SIZE(frame_list);
275
276         for (i = 0; i < nr_pages; i++) {
277                 if ((page = alloc_page(GFP_BALLOON)) == NULL) {
278                         nr_pages = i;
279                         need_sleep = 1;
280                         break;
281                 }
282
283                 pfn = page_to_pfn(page);
284                 frame_list[i] = pfn_to_mfn(pfn);
285
286                 scrub_page(page);
287
288                 if (!PageHighMem(page)) {
289                         ret = HYPERVISOR_update_va_mapping(
290                                 (unsigned long)__va(pfn << PAGE_SHIFT),
291                                 __pte_ma(0), 0);
292                         BUG_ON(ret);
293                 }
294
295         }
296
297         /* Ensure that ballooned highmem pages don't have kmaps. */
298         kmap_flush_unused();
299         flush_tlb_all();
300
301         spin_lock_irqsave(&balloon_lock, flags);
302
303         /* No more mappings: invalidate P2M and add to balloon. */
304         for (i = 0; i < nr_pages; i++) {
305                 pfn = mfn_to_pfn(frame_list[i]);
306                 set_phys_to_machine(pfn, INVALID_P2M_ENTRY);
307                 balloon_append(pfn_to_page(pfn));
308         }
309
310         set_xen_guest_handle(reservation.extent_start, frame_list);
311         reservation.nr_extents   = nr_pages;
312         ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation, &reservation);
313         BUG_ON(ret != nr_pages);
314
315         balloon_stats.current_pages -= nr_pages;
316
317         spin_unlock_irqrestore(&balloon_lock, flags);
318
319         return need_sleep;
320 }
321
322 /*
323  * We avoid multiple worker processes conflicting via the balloon mutex.
324  * We may of course race updates of the target counts (which are protected
325  * by the balloon lock), or with changes to the Xen hard limit, but we will
326  * recover from these in time.
327  */
328 static void balloon_process(struct work_struct *work)
329 {
330         int need_sleep = 0;
331         long credit;
332
333         mutex_lock(&balloon_mutex);
334
335         do {
336                 credit = current_target() - balloon_stats.current_pages;
337                 if (credit > 0)
338                         need_sleep = (increase_reservation(credit) != 0);
339                 if (credit < 0)
340                         need_sleep = (decrease_reservation(-credit) != 0);
341
342 #ifndef CONFIG_PREEMPT
343                 if (need_resched())
344                         schedule();
345 #endif
346         } while ((credit != 0) && !need_sleep);
347
348         /* Schedule more work if there is some still to be done. */
349         if (current_target() != balloon_stats.current_pages)
350                 mod_timer(&balloon_timer, jiffies + HZ);
351
352         mutex_unlock(&balloon_mutex);
353 }
354
355 /* Resets the Xen limit, sets new target, and kicks off processing. */
356 static void balloon_set_new_target(unsigned long target)
357 {
358         /* No need for lock. Not read-modify-write updates. */
359         balloon_stats.target_pages = target;
360         schedule_work(&balloon_worker);
361 }
362
363 static struct xenbus_watch target_watch =
364 {
365         .node = "memory/target"
366 };
367
368 /* React to a change in the target key */
369 static void watch_target(struct xenbus_watch *watch,
370                          const char **vec, unsigned int len)
371 {
372         unsigned long long new_target;
373         int err;
374
375         err = xenbus_scanf(XBT_NIL, "memory", "target", "%llu", &new_target);
376         if (err != 1) {
377                 /* This is ok (for domain0 at least) - so just return */
378                 return;
379         }
380
381         /* The given memory/target value is in KiB, so it needs converting to
382          * pages. PAGE_SHIFT converts bytes to pages, hence PAGE_SHIFT - 10.
383          */
384         balloon_set_new_target(new_target >> (PAGE_SHIFT - 10));
385 }
386
387 static int balloon_init_watcher(struct notifier_block *notifier,
388                                 unsigned long event,
389                                 void *data)
390 {
391         int err;
392
393         err = register_xenbus_watch(&target_watch);
394         if (err)
395                 printk(KERN_ERR "Failed to set balloon watcher\n");
396
397         return NOTIFY_DONE;
398 }
399
400 static struct notifier_block xenstore_notifier;
401
402 static int __init balloon_init(void)
403 {
404         unsigned long pfn;
405         struct page *page;
406
407         if (!xen_pv_domain())
408                 return -ENODEV;
409
410         pr_info("xen_balloon: Initialising balloon driver.\n");
411
412         balloon_stats.current_pages = min(xen_start_info->nr_pages, max_pfn);
413         balloon_stats.target_pages  = balloon_stats.current_pages;
414         balloon_stats.balloon_low   = 0;
415         balloon_stats.balloon_high  = 0;
416         balloon_stats.driver_pages  = 0UL;
417
418         init_timer(&balloon_timer);
419         balloon_timer.data = 0;
420         balloon_timer.function = balloon_alarm;
421
422         register_balloon(&balloon_sysdev);
423
424         /* Initialise the balloon with excess memory space. */
425         for (pfn = xen_start_info->nr_pages; pfn < max_pfn; pfn++) {
426                 page = pfn_to_page(pfn);
427                 if (!PageReserved(page))
428                         balloon_append(page);
429         }
430
431         target_watch.callback = watch_target;
432         xenstore_notifier.notifier_call = balloon_init_watcher;
433
434         register_xenstore_notifier(&xenstore_notifier);
435
436         return 0;
437 }
438
439 subsys_initcall(balloon_init);
440
441 static void balloon_exit(void)
442 {
443     /* XXX - release balloon here */
444     return;
445 }
446
447 module_exit(balloon_exit);
448
449 #define BALLOON_SHOW(name, format, args...)                             \
450         static ssize_t show_##name(struct sys_device *dev,              \
451                                    struct sysdev_attribute *attr,       \
452                                    char *buf)                           \
453         {                                                               \
454                 return sprintf(buf, format, ##args);                    \
455         }                                                               \
456         static SYSDEV_ATTR(name, S_IRUGO, show_##name, NULL)
457
458 BALLOON_SHOW(current_kb, "%lu\n", PAGES2KB(balloon_stats.current_pages));
459 BALLOON_SHOW(low_kb, "%lu\n", PAGES2KB(balloon_stats.balloon_low));
460 BALLOON_SHOW(high_kb, "%lu\n", PAGES2KB(balloon_stats.balloon_high));
461 BALLOON_SHOW(driver_kb, "%lu\n", PAGES2KB(balloon_stats.driver_pages));
462
463 static ssize_t show_target_kb(struct sys_device *dev, struct sysdev_attribute *attr,
464                               char *buf)
465 {
466         return sprintf(buf, "%lu\n", PAGES2KB(balloon_stats.target_pages));
467 }
468
469 static ssize_t store_target_kb(struct sys_device *dev,
470                                struct sysdev_attribute *attr,
471                                const char *buf,
472                                size_t count)
473 {
474         char *endchar;
475         unsigned long long target_bytes;
476
477         if (!capable(CAP_SYS_ADMIN))
478                 return -EPERM;
479
480         target_bytes = simple_strtoull(buf, &endchar, 0) * 1024;
481
482         balloon_set_new_target(target_bytes >> PAGE_SHIFT);
483
484         return count;
485 }
486
487 static SYSDEV_ATTR(target_kb, S_IRUGO | S_IWUSR,
488                    show_target_kb, store_target_kb);
489
490
491 static ssize_t show_target(struct sys_device *dev, struct sysdev_attribute *attr,
492                               char *buf)
493 {
494         return sprintf(buf, "%llu\n",
495                        (unsigned long long)balloon_stats.target_pages
496                        << PAGE_SHIFT);
497 }
498
499 static ssize_t store_target(struct sys_device *dev,
500                             struct sysdev_attribute *attr,
501                             const char *buf,
502                             size_t count)
503 {
504         char *endchar;
505         unsigned long long target_bytes;
506
507         if (!capable(CAP_SYS_ADMIN))
508                 return -EPERM;
509
510         target_bytes = memparse(buf, &endchar);
511
512         balloon_set_new_target(target_bytes >> PAGE_SHIFT);
513
514         return count;
515 }
516
517 static SYSDEV_ATTR(target, S_IRUGO | S_IWUSR,
518                    show_target, store_target);
519
520
521 static struct sysdev_attribute *balloon_attrs[] = {
522         &attr_target_kb,
523         &attr_target,
524 };
525
526 static struct attribute *balloon_info_attrs[] = {
527         &attr_current_kb.attr,
528         &attr_low_kb.attr,
529         &attr_high_kb.attr,
530         &attr_driver_kb.attr,
531         NULL
532 };
533
534 static struct attribute_group balloon_info_group = {
535         .name = "info",
536         .attrs = balloon_info_attrs,
537 };
538
539 static struct sysdev_class balloon_sysdev_class = {
540         .name = BALLOON_CLASS_NAME,
541 };
542
543 static int register_balloon(struct sys_device *sysdev)
544 {
545         int i, error;
546
547         error = sysdev_class_register(&balloon_sysdev_class);
548         if (error)
549                 return error;
550
551         sysdev->id = 0;
552         sysdev->cls = &balloon_sysdev_class;
553
554         error = sysdev_register(sysdev);
555         if (error) {
556                 sysdev_class_unregister(&balloon_sysdev_class);
557                 return error;
558         }
559
560         for (i = 0; i < ARRAY_SIZE(balloon_attrs); i++) {
561                 error = sysdev_create_file(sysdev, balloon_attrs[i]);
562                 if (error)
563                         goto fail;
564         }
565
566         error = sysfs_create_group(&sysdev->kobj, &balloon_info_group);
567         if (error)
568                 goto fail;
569
570         return 0;
571
572  fail:
573         while (--i >= 0)
574                 sysdev_remove_file(sysdev, balloon_attrs[i]);
575         sysdev_unregister(sysdev);
576         sysdev_class_unregister(&balloon_sysdev_class);
577         return error;
578 }
579
580 MODULE_LICENSE("GPL");