Merge commit 'v2.6.28-rc4' into x86/apic
[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/xen/hypervisor.h>
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 <xen/interface/memory.h>
55 #include <xen/xenbus.h>
56 #include <xen/features.h>
57 #include <xen/page.h>
58
59 #define PAGES2KB(_p) ((_p)<<(PAGE_SHIFT-10))
60
61 #define BALLOON_CLASS_NAME "xen_memory"
62
63 struct balloon_stats {
64         /* We aim for 'current allocation' == 'target allocation'. */
65         unsigned long current_pages;
66         unsigned long target_pages;
67         /* We may hit the hard limit in Xen. If we do then we remember it. */
68         unsigned long hard_limit;
69         /*
70          * Drivers may alter the memory reservation independently, but they
71          * must inform the balloon driver so we avoid hitting the hard limit.
72          */
73         unsigned long driver_pages;
74         /* Number of pages in high- and low-memory balloons. */
75         unsigned long balloon_low;
76         unsigned long balloon_high;
77 };
78
79 static DEFINE_MUTEX(balloon_mutex);
80
81 static struct sys_device balloon_sysdev;
82
83 static int register_balloon(struct sys_device *sysdev);
84
85 /*
86  * Protects atomic reservation decrease/increase against concurrent increases.
87  * Also protects non-atomic updates of current_pages and driver_pages, and
88  * balloon lists.
89  */
90 static DEFINE_SPINLOCK(balloon_lock);
91
92 static struct balloon_stats balloon_stats;
93
94 /* We increase/decrease in batches which fit in a page */
95 static unsigned long frame_list[PAGE_SIZE / sizeof(unsigned long)];
96
97 /* VM /proc information for memory */
98 extern unsigned long totalram_pages;
99
100 #ifdef CONFIG_HIGHMEM
101 extern unsigned long totalhigh_pages;
102 #define inc_totalhigh_pages() (totalhigh_pages++)
103 #define dec_totalhigh_pages() (totalhigh_pages--)
104 #else
105 #define inc_totalhigh_pages() do {} while(0)
106 #define dec_totalhigh_pages() do {} while(0)
107 #endif
108
109 /* List of ballooned pages, threaded through the mem_map array. */
110 static LIST_HEAD(ballooned_pages);
111
112 /* Main work function, always executed in process context. */
113 static void balloon_process(struct work_struct *work);
114 static DECLARE_WORK(balloon_worker, balloon_process);
115 static struct timer_list balloon_timer;
116
117 /* When ballooning out (allocating memory to return to Xen) we don't really
118    want the kernel to try too hard since that can trigger the oom killer. */
119 #define GFP_BALLOON \
120         (GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC)
121
122 static void scrub_page(struct page *page)
123 {
124 #ifdef CONFIG_XEN_SCRUB_PAGES
125         if (PageHighMem(page)) {
126                 void *v = kmap(page);
127                 clear_page(v);
128                 kunmap(v);
129         } else {
130                 void *v = page_address(page);
131                 clear_page(v);
132         }
133 #endif
134 }
135
136 /* balloon_append: add the given page to the balloon. */
137 static void balloon_append(struct page *page)
138 {
139         /* Lowmem is re-populated first, so highmem pages go at list tail. */
140         if (PageHighMem(page)) {
141                 list_add_tail(&page->lru, &ballooned_pages);
142                 balloon_stats.balloon_high++;
143                 dec_totalhigh_pages();
144         } else {
145                 list_add(&page->lru, &ballooned_pages);
146                 balloon_stats.balloon_low++;
147         }
148 }
149
150 /* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
151 static struct page *balloon_retrieve(void)
152 {
153         struct page *page;
154
155         if (list_empty(&ballooned_pages))
156                 return NULL;
157
158         page = list_entry(ballooned_pages.next, struct page, lru);
159         list_del(&page->lru);
160
161         if (PageHighMem(page)) {
162                 balloon_stats.balloon_high--;
163                 inc_totalhigh_pages();
164         }
165         else
166                 balloon_stats.balloon_low--;
167
168         return page;
169 }
170
171 static struct page *balloon_first_page(void)
172 {
173         if (list_empty(&ballooned_pages))
174                 return NULL;
175         return list_entry(ballooned_pages.next, struct page, lru);
176 }
177
178 static struct page *balloon_next_page(struct page *page)
179 {
180         struct list_head *next = page->lru.next;
181         if (next == &ballooned_pages)
182                 return NULL;
183         return list_entry(next, struct page, lru);
184 }
185
186 static void balloon_alarm(unsigned long unused)
187 {
188         schedule_work(&balloon_worker);
189 }
190
191 static unsigned long current_target(void)
192 {
193         unsigned long target = min(balloon_stats.target_pages, balloon_stats.hard_limit);
194
195         target = min(target,
196                      balloon_stats.current_pages +
197                      balloon_stats.balloon_low +
198                      balloon_stats.balloon_high);
199
200         return target;
201 }
202
203 static int increase_reservation(unsigned long nr_pages)
204 {
205         unsigned long  pfn, i, flags;
206         struct page   *page;
207         long           rc;
208         struct xen_memory_reservation reservation = {
209                 .address_bits = 0,
210                 .extent_order = 0,
211                 .domid        = DOMID_SELF
212         };
213
214         if (nr_pages > ARRAY_SIZE(frame_list))
215                 nr_pages = ARRAY_SIZE(frame_list);
216
217         spin_lock_irqsave(&balloon_lock, flags);
218
219         page = balloon_first_page();
220         for (i = 0; i < nr_pages; i++) {
221                 BUG_ON(page == NULL);
222                 frame_list[i] = page_to_pfn(page);;
223                 page = balloon_next_page(page);
224         }
225
226         set_xen_guest_handle(reservation.extent_start, frame_list);
227         reservation.nr_extents = nr_pages;
228         rc = HYPERVISOR_memory_op(XENMEM_populate_physmap, &reservation);
229         if (rc < nr_pages) {
230                 if (rc > 0) {
231                         int ret;
232
233                         /* We hit the Xen hard limit: reprobe. */
234                         reservation.nr_extents = rc;
235                         ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation,
236                                                    &reservation);
237                         BUG_ON(ret != rc);
238                 }
239                 if (rc >= 0)
240                         balloon_stats.hard_limit = (balloon_stats.current_pages + rc -
241                                                     balloon_stats.driver_pages);
242                 goto out;
243         }
244
245         for (i = 0; i < nr_pages; i++) {
246                 page = balloon_retrieve();
247                 BUG_ON(page == NULL);
248
249                 pfn = page_to_pfn(page);
250                 BUG_ON(!xen_feature(XENFEAT_auto_translated_physmap) &&
251                        phys_to_machine_mapping_valid(pfn));
252
253                 set_phys_to_machine(pfn, frame_list[i]);
254
255                 /* Link back into the page tables if not highmem. */
256                 if (pfn < max_low_pfn) {
257                         int ret;
258                         ret = HYPERVISOR_update_va_mapping(
259                                 (unsigned long)__va(pfn << PAGE_SHIFT),
260                                 mfn_pte(frame_list[i], PAGE_KERNEL),
261                                 0);
262                         BUG_ON(ret);
263                 }
264
265                 /* Relinquish the page back to the allocator. */
266                 ClearPageReserved(page);
267                 init_page_count(page);
268                 __free_page(page);
269         }
270
271         balloon_stats.current_pages += nr_pages;
272         totalram_pages = balloon_stats.current_pages;
273
274  out:
275         spin_unlock_irqrestore(&balloon_lock, flags);
276
277         return 0;
278 }
279
280 static int decrease_reservation(unsigned long nr_pages)
281 {
282         unsigned long  pfn, i, flags;
283         struct page   *page;
284         int            need_sleep = 0;
285         int ret;
286         struct xen_memory_reservation reservation = {
287                 .address_bits = 0,
288                 .extent_order = 0,
289                 .domid        = DOMID_SELF
290         };
291
292         if (nr_pages > ARRAY_SIZE(frame_list))
293                 nr_pages = ARRAY_SIZE(frame_list);
294
295         for (i = 0; i < nr_pages; i++) {
296                 if ((page = alloc_page(GFP_BALLOON)) == NULL) {
297                         nr_pages = i;
298                         need_sleep = 1;
299                         break;
300                 }
301
302                 pfn = page_to_pfn(page);
303                 frame_list[i] = pfn_to_mfn(pfn);
304
305                 scrub_page(page);
306         }
307
308         /* Ensure that ballooned highmem pages don't have kmaps. */
309         kmap_flush_unused();
310         flush_tlb_all();
311
312         spin_lock_irqsave(&balloon_lock, flags);
313
314         /* No more mappings: invalidate P2M and add to balloon. */
315         for (i = 0; i < nr_pages; i++) {
316                 pfn = mfn_to_pfn(frame_list[i]);
317                 set_phys_to_machine(pfn, INVALID_P2M_ENTRY);
318                 balloon_append(pfn_to_page(pfn));
319         }
320
321         set_xen_guest_handle(reservation.extent_start, frame_list);
322         reservation.nr_extents   = nr_pages;
323         ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation, &reservation);
324         BUG_ON(ret != nr_pages);
325
326         balloon_stats.current_pages -= nr_pages;
327         totalram_pages = balloon_stats.current_pages;
328
329         spin_unlock_irqrestore(&balloon_lock, flags);
330
331         return need_sleep;
332 }
333
334 /*
335  * We avoid multiple worker processes conflicting via the balloon mutex.
336  * We may of course race updates of the target counts (which are protected
337  * by the balloon lock), or with changes to the Xen hard limit, but we will
338  * recover from these in time.
339  */
340 static void balloon_process(struct work_struct *work)
341 {
342         int need_sleep = 0;
343         long credit;
344
345         mutex_lock(&balloon_mutex);
346
347         do {
348                 credit = current_target() - balloon_stats.current_pages;
349                 if (credit > 0)
350                         need_sleep = (increase_reservation(credit) != 0);
351                 if (credit < 0)
352                         need_sleep = (decrease_reservation(-credit) != 0);
353
354 #ifndef CONFIG_PREEMPT
355                 if (need_resched())
356                         schedule();
357 #endif
358         } while ((credit != 0) && !need_sleep);
359
360         /* Schedule more work if there is some still to be done. */
361         if (current_target() != balloon_stats.current_pages)
362                 mod_timer(&balloon_timer, jiffies + HZ);
363
364         mutex_unlock(&balloon_mutex);
365 }
366
367 /* Resets the Xen limit, sets new target, and kicks off processing. */
368 static void balloon_set_new_target(unsigned long target)
369 {
370         /* No need for lock. Not read-modify-write updates. */
371         balloon_stats.hard_limit   = ~0UL;
372         balloon_stats.target_pages = target;
373         schedule_work(&balloon_worker);
374 }
375
376 static struct xenbus_watch target_watch =
377 {
378         .node = "memory/target"
379 };
380
381 /* React to a change in the target key */
382 static void watch_target(struct xenbus_watch *watch,
383                          const char **vec, unsigned int len)
384 {
385         unsigned long long new_target;
386         int err;
387
388         err = xenbus_scanf(XBT_NIL, "memory", "target", "%llu", &new_target);
389         if (err != 1) {
390                 /* This is ok (for domain0 at least) - so just return */
391                 return;
392         }
393
394         /* The given memory/target value is in KiB, so it needs converting to
395          * pages. PAGE_SHIFT converts bytes to pages, hence PAGE_SHIFT - 10.
396          */
397         balloon_set_new_target(new_target >> (PAGE_SHIFT - 10));
398 }
399
400 static int balloon_init_watcher(struct notifier_block *notifier,
401                                 unsigned long event,
402                                 void *data)
403 {
404         int err;
405
406         err = register_xenbus_watch(&target_watch);
407         if (err)
408                 printk(KERN_ERR "Failed to set balloon watcher\n");
409
410         return NOTIFY_DONE;
411 }
412
413 static struct notifier_block xenstore_notifier;
414
415 static int __init balloon_init(void)
416 {
417         unsigned long pfn;
418         struct page *page;
419
420         if (!xen_pv_domain())
421                 return -ENODEV;
422
423         pr_info("xen_balloon: Initialising balloon driver.\n");
424
425         balloon_stats.current_pages = min(xen_start_info->nr_pages, max_pfn);
426         totalram_pages   = balloon_stats.current_pages;
427         balloon_stats.target_pages  = balloon_stats.current_pages;
428         balloon_stats.balloon_low   = 0;
429         balloon_stats.balloon_high  = 0;
430         balloon_stats.driver_pages  = 0UL;
431         balloon_stats.hard_limit    = ~0UL;
432
433         init_timer(&balloon_timer);
434         balloon_timer.data = 0;
435         balloon_timer.function = balloon_alarm;
436
437         register_balloon(&balloon_sysdev);
438
439         /* Initialise the balloon with excess memory space. */
440         for (pfn = xen_start_info->nr_pages; pfn < max_pfn; pfn++) {
441                 page = pfn_to_page(pfn);
442                 if (!PageReserved(page))
443                         balloon_append(page);
444         }
445
446         target_watch.callback = watch_target;
447         xenstore_notifier.notifier_call = balloon_init_watcher;
448
449         register_xenstore_notifier(&xenstore_notifier);
450
451         return 0;
452 }
453
454 subsys_initcall(balloon_init);
455
456 static void balloon_exit(void)
457 {
458     /* XXX - release balloon here */
459     return;
460 }
461
462 module_exit(balloon_exit);
463
464 #define BALLOON_SHOW(name, format, args...)                             \
465         static ssize_t show_##name(struct sys_device *dev,              \
466                                    struct sysdev_attribute *attr,       \
467                                    char *buf)                           \
468         {                                                               \
469                 return sprintf(buf, format, ##args);                    \
470         }                                                               \
471         static SYSDEV_ATTR(name, S_IRUGO, show_##name, NULL)
472
473 BALLOON_SHOW(current_kb, "%lu\n", PAGES2KB(balloon_stats.current_pages));
474 BALLOON_SHOW(low_kb, "%lu\n", PAGES2KB(balloon_stats.balloon_low));
475 BALLOON_SHOW(high_kb, "%lu\n", PAGES2KB(balloon_stats.balloon_high));
476 BALLOON_SHOW(hard_limit_kb,
477              (balloon_stats.hard_limit!=~0UL) ? "%lu\n" : "???\n",
478              (balloon_stats.hard_limit!=~0UL) ? PAGES2KB(balloon_stats.hard_limit) : 0);
479 BALLOON_SHOW(driver_kb, "%lu\n", PAGES2KB(balloon_stats.driver_pages));
480
481 static ssize_t show_target_kb(struct sys_device *dev, struct sysdev_attribute *attr,
482                               char *buf)
483 {
484         return sprintf(buf, "%lu\n", PAGES2KB(balloon_stats.target_pages));
485 }
486
487 static ssize_t store_target_kb(struct sys_device *dev,
488                                struct sysdev_attribute *attr,
489                                const char *buf,
490                                size_t count)
491 {
492         char *endchar;
493         unsigned long long target_bytes;
494
495         if (!capable(CAP_SYS_ADMIN))
496                 return -EPERM;
497
498         target_bytes = memparse(buf, &endchar);
499
500         balloon_set_new_target(target_bytes >> PAGE_SHIFT);
501
502         return count;
503 }
504
505 static SYSDEV_ATTR(target_kb, S_IRUGO | S_IWUSR,
506                    show_target_kb, store_target_kb);
507
508 static struct sysdev_attribute *balloon_attrs[] = {
509         &attr_target_kb,
510 };
511
512 static struct attribute *balloon_info_attrs[] = {
513         &attr_current_kb.attr,
514         &attr_low_kb.attr,
515         &attr_high_kb.attr,
516         &attr_hard_limit_kb.attr,
517         &attr_driver_kb.attr,
518         NULL
519 };
520
521 static struct attribute_group balloon_info_group = {
522         .name = "info",
523         .attrs = balloon_info_attrs,
524 };
525
526 static struct sysdev_class balloon_sysdev_class = {
527         .name = BALLOON_CLASS_NAME,
528 };
529
530 static int register_balloon(struct sys_device *sysdev)
531 {
532         int i, error;
533
534         error = sysdev_class_register(&balloon_sysdev_class);
535         if (error)
536                 return error;
537
538         sysdev->id = 0;
539         sysdev->cls = &balloon_sysdev_class;
540
541         error = sysdev_register(sysdev);
542         if (error) {
543                 sysdev_class_unregister(&balloon_sysdev_class);
544                 return error;
545         }
546
547         for (i = 0; i < ARRAY_SIZE(balloon_attrs); i++) {
548                 error = sysdev_create_file(sysdev, balloon_attrs[i]);
549                 if (error)
550                         goto fail;
551         }
552
553         error = sysfs_create_group(&sysdev->kobj, &balloon_info_group);
554         if (error)
555                 goto fail;
556
557         return 0;
558
559  fail:
560         while (--i >= 0)
561                 sysdev_remove_file(sysdev, balloon_attrs[i]);
562         sysdev_unregister(sysdev);
563         sysdev_class_unregister(&balloon_sysdev_class);
564         return error;
565 }
566
567 MODULE_LICENSE("GPL");