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