]> git.openfabrics.org - ~shefty/rdma-dev.git/blob - drivers/xen/balloon.c
Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/mason/linux...
[~shefty/rdma-dev.git] / drivers / xen / balloon.c
1 /******************************************************************************
2  * Xen balloon driver - enables returning/claiming memory to/from Xen.
3  *
4  * Copyright (c) 2003, B Dragovic
5  * Copyright (c) 2003-2004, M Williamson, K Fraser
6  * Copyright (c) 2005 Dan M. Smith, IBM Corporation
7  * Copyright (c) 2010 Daniel Kiper
8  *
9  * Memory hotplug support was written by Daniel Kiper. Work on
10  * it was sponsored by Google under Google Summer of Code 2010
11  * program. Jeremy Fitzhardinge from Citrix was the mentor for
12  * this project.
13  *
14  * This program is free software; you can redistribute it and/or
15  * modify it under the terms of the GNU General Public License version 2
16  * as published by the Free Software Foundation; or, when distributed
17  * separately from the Linux kernel or incorporated into other
18  * software packages, subject to the following license:
19  *
20  * Permission is hereby granted, free of charge, to any person obtaining a copy
21  * of this source file (the "Software"), to deal in the Software without
22  * restriction, including without limitation the rights to use, copy, modify,
23  * merge, publish, distribute, sublicense, and/or sell copies of the Software,
24  * and to permit persons to whom the Software is furnished to do so, subject to
25  * the following conditions:
26  *
27  * The above copyright notice and this permission notice shall be included in
28  * all copies or substantial portions of the Software.
29  *
30  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
31  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
32  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
33  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
34  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
35  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
36  * IN THE SOFTWARE.
37  */
38
39 #include <linux/kernel.h>
40 #include <linux/sched.h>
41 #include <linux/errno.h>
42 #include <linux/module.h>
43 #include <linux/mm.h>
44 #include <linux/bootmem.h>
45 #include <linux/pagemap.h>
46 #include <linux/highmem.h>
47 #include <linux/mutex.h>
48 #include <linux/list.h>
49 #include <linux/gfp.h>
50 #include <linux/notifier.h>
51 #include <linux/memory.h>
52 #include <linux/memory_hotplug.h>
53
54 #include <asm/page.h>
55 #include <asm/pgalloc.h>
56 #include <asm/pgtable.h>
57 #include <asm/tlb.h>
58 #include <asm/e820.h>
59
60 #include <asm/xen/hypervisor.h>
61 #include <asm/xen/hypercall.h>
62
63 #include <xen/xen.h>
64 #include <xen/interface/xen.h>
65 #include <xen/interface/memory.h>
66 #include <xen/balloon.h>
67 #include <xen/features.h>
68 #include <xen/page.h>
69
70 /*
71  * balloon_process() state:
72  *
73  * BP_DONE: done or nothing to do,
74  * BP_EAGAIN: error, go to sleep,
75  * BP_ECANCELED: error, balloon operation canceled.
76  */
77
78 enum bp_state {
79         BP_DONE,
80         BP_EAGAIN,
81         BP_ECANCELED
82 };
83
84
85 static DEFINE_MUTEX(balloon_mutex);
86
87 struct balloon_stats balloon_stats;
88 EXPORT_SYMBOL_GPL(balloon_stats);
89
90 /* We increase/decrease in batches which fit in a page */
91 static unsigned long frame_list[PAGE_SIZE / sizeof(unsigned long)];
92
93 #ifdef CONFIG_HIGHMEM
94 #define inc_totalhigh_pages() (totalhigh_pages++)
95 #define dec_totalhigh_pages() (totalhigh_pages--)
96 #else
97 #define inc_totalhigh_pages() do {} while(0)
98 #define dec_totalhigh_pages() do {} while(0)
99 #endif
100
101 /* List of ballooned pages, threaded through the mem_map array. */
102 static LIST_HEAD(ballooned_pages);
103
104 /* Main work function, always executed in process context. */
105 static void balloon_process(struct work_struct *work);
106 static DECLARE_DELAYED_WORK(balloon_worker, balloon_process);
107
108 /* When ballooning out (allocating memory to return to Xen) we don't really
109    want the kernel to try too hard since that can trigger the oom killer. */
110 #define GFP_BALLOON \
111         (GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC)
112
113 static void scrub_page(struct page *page)
114 {
115 #ifdef CONFIG_XEN_SCRUB_PAGES
116         clear_highpage(page);
117 #endif
118 }
119
120 /* balloon_append: add the given page to the balloon. */
121 static void __balloon_append(struct page *page)
122 {
123         /* Lowmem is re-populated first, so highmem pages go at list tail. */
124         if (PageHighMem(page)) {
125                 list_add_tail(&page->lru, &ballooned_pages);
126                 balloon_stats.balloon_high++;
127         } else {
128                 list_add(&page->lru, &ballooned_pages);
129                 balloon_stats.balloon_low++;
130         }
131 }
132
133 static void balloon_append(struct page *page)
134 {
135         __balloon_append(page);
136         if (PageHighMem(page))
137                 dec_totalhigh_pages();
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(bool prefer_highmem)
143 {
144         struct page *page;
145
146         if (list_empty(&ballooned_pages))
147                 return NULL;
148
149         if (prefer_highmem)
150                 page = list_entry(ballooned_pages.prev, struct page, lru);
151         else
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 enum bp_state update_schedule(enum bp_state state)
183 {
184         if (state == BP_DONE) {
185                 balloon_stats.schedule_delay = 1;
186                 balloon_stats.retry_count = 1;
187                 return BP_DONE;
188         }
189
190         ++balloon_stats.retry_count;
191
192         if (balloon_stats.max_retry_count != RETRY_UNLIMITED &&
193                         balloon_stats.retry_count > balloon_stats.max_retry_count) {
194                 balloon_stats.schedule_delay = 1;
195                 balloon_stats.retry_count = 1;
196                 return BP_ECANCELED;
197         }
198
199         balloon_stats.schedule_delay <<= 1;
200
201         if (balloon_stats.schedule_delay > balloon_stats.max_schedule_delay)
202                 balloon_stats.schedule_delay = balloon_stats.max_schedule_delay;
203
204         return BP_EAGAIN;
205 }
206
207 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
208 static long current_credit(void)
209 {
210         return balloon_stats.target_pages - balloon_stats.current_pages -
211                 balloon_stats.hotplug_pages;
212 }
213
214 static bool balloon_is_inflated(void)
215 {
216         if (balloon_stats.balloon_low || balloon_stats.balloon_high ||
217                         balloon_stats.balloon_hotplug)
218                 return true;
219         else
220                 return false;
221 }
222
223 /*
224  * reserve_additional_memory() adds memory region of size >= credit above
225  * max_pfn. New region is section aligned and size is modified to be multiple
226  * of section size. Those features allow optimal use of address space and
227  * establish proper alignment when this function is called first time after
228  * boot (last section not fully populated at boot time contains unused memory
229  * pages with PG_reserved bit not set; online_pages_range() does not allow page
230  * onlining in whole range if first onlined page does not have PG_reserved
231  * bit set). Real size of added memory is established at page onlining stage.
232  */
233
234 static enum bp_state reserve_additional_memory(long credit)
235 {
236         int nid, rc;
237         u64 hotplug_start_paddr;
238         unsigned long balloon_hotplug = credit;
239
240         hotplug_start_paddr = PFN_PHYS(SECTION_ALIGN_UP(max_pfn));
241         balloon_hotplug = round_up(balloon_hotplug, PAGES_PER_SECTION);
242         nid = memory_add_physaddr_to_nid(hotplug_start_paddr);
243
244         rc = add_memory(nid, hotplug_start_paddr, balloon_hotplug << PAGE_SHIFT);
245
246         if (rc) {
247                 pr_info("xen_balloon: %s: add_memory() failed: %i\n", __func__, rc);
248                 return BP_EAGAIN;
249         }
250
251         balloon_hotplug -= credit;
252
253         balloon_stats.hotplug_pages += credit;
254         balloon_stats.balloon_hotplug = balloon_hotplug;
255
256         return BP_DONE;
257 }
258
259 static void xen_online_page(struct page *page)
260 {
261         __online_page_set_limits(page);
262
263         mutex_lock(&balloon_mutex);
264
265         __balloon_append(page);
266
267         if (balloon_stats.hotplug_pages)
268                 --balloon_stats.hotplug_pages;
269         else
270                 --balloon_stats.balloon_hotplug;
271
272         mutex_unlock(&balloon_mutex);
273 }
274
275 static int xen_memory_notifier(struct notifier_block *nb, unsigned long val, void *v)
276 {
277         if (val == MEM_ONLINE)
278                 schedule_delayed_work(&balloon_worker, 0);
279
280         return NOTIFY_OK;
281 }
282
283 static struct notifier_block xen_memory_nb = {
284         .notifier_call = xen_memory_notifier,
285         .priority = 0
286 };
287 #else
288 static long current_credit(void)
289 {
290         unsigned long target = balloon_stats.target_pages;
291
292         target = min(target,
293                      balloon_stats.current_pages +
294                      balloon_stats.balloon_low +
295                      balloon_stats.balloon_high);
296
297         return target - balloon_stats.current_pages;
298 }
299
300 static bool balloon_is_inflated(void)
301 {
302         if (balloon_stats.balloon_low || balloon_stats.balloon_high)
303                 return true;
304         else
305                 return false;
306 }
307
308 static enum bp_state reserve_additional_memory(long credit)
309 {
310         balloon_stats.target_pages = balloon_stats.current_pages;
311         return BP_DONE;
312 }
313 #endif /* CONFIG_XEN_BALLOON_MEMORY_HOTPLUG */
314
315 static enum bp_state increase_reservation(unsigned long nr_pages)
316 {
317         int rc;
318         unsigned long  pfn, i;
319         struct page   *page;
320         struct xen_memory_reservation reservation = {
321                 .address_bits = 0,
322                 .extent_order = 0,
323                 .domid        = DOMID_SELF
324         };
325
326 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
327         if (!balloon_stats.balloon_low && !balloon_stats.balloon_high) {
328                 nr_pages = min(nr_pages, balloon_stats.balloon_hotplug);
329                 balloon_stats.hotplug_pages += nr_pages;
330                 balloon_stats.balloon_hotplug -= nr_pages;
331                 return BP_DONE;
332         }
333 #endif
334
335         if (nr_pages > ARRAY_SIZE(frame_list))
336                 nr_pages = ARRAY_SIZE(frame_list);
337
338         page = balloon_first_page();
339         for (i = 0; i < nr_pages; i++) {
340                 if (!page) {
341                         nr_pages = i;
342                         break;
343                 }
344                 frame_list[i] = page_to_pfn(page);
345                 page = balloon_next_page(page);
346         }
347
348         set_xen_guest_handle(reservation.extent_start, frame_list);
349         reservation.nr_extents = nr_pages;
350         rc = HYPERVISOR_memory_op(XENMEM_populate_physmap, &reservation);
351         if (rc <= 0)
352                 return BP_EAGAIN;
353
354         for (i = 0; i < rc; i++) {
355                 page = balloon_retrieve(false);
356                 BUG_ON(page == NULL);
357
358                 pfn = page_to_pfn(page);
359                 BUG_ON(!xen_feature(XENFEAT_auto_translated_physmap) &&
360                        phys_to_machine_mapping_valid(pfn));
361
362                 set_phys_to_machine(pfn, frame_list[i]);
363
364                 /* Link back into the page tables if not highmem. */
365                 if (xen_pv_domain() && !PageHighMem(page)) {
366                         int ret;
367                         ret = HYPERVISOR_update_va_mapping(
368                                 (unsigned long)__va(pfn << PAGE_SHIFT),
369                                 mfn_pte(frame_list[i], PAGE_KERNEL),
370                                 0);
371                         BUG_ON(ret);
372                 }
373
374                 /* Relinquish the page back to the allocator. */
375                 ClearPageReserved(page);
376                 init_page_count(page);
377                 __free_page(page);
378         }
379
380         balloon_stats.current_pages += rc;
381
382         return BP_DONE;
383 }
384
385 static enum bp_state decrease_reservation(unsigned long nr_pages, gfp_t gfp)
386 {
387         enum bp_state state = BP_DONE;
388         unsigned long  pfn, i;
389         struct page   *page;
390         int ret;
391         struct xen_memory_reservation reservation = {
392                 .address_bits = 0,
393                 .extent_order = 0,
394                 .domid        = DOMID_SELF
395         };
396
397 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
398         if (balloon_stats.hotplug_pages) {
399                 nr_pages = min(nr_pages, balloon_stats.hotplug_pages);
400                 balloon_stats.hotplug_pages -= nr_pages;
401                 balloon_stats.balloon_hotplug += nr_pages;
402                 return BP_DONE;
403         }
404 #endif
405
406         if (nr_pages > ARRAY_SIZE(frame_list))
407                 nr_pages = ARRAY_SIZE(frame_list);
408
409         for (i = 0; i < nr_pages; i++) {
410                 if ((page = alloc_page(gfp)) == NULL) {
411                         nr_pages = i;
412                         state = BP_EAGAIN;
413                         break;
414                 }
415
416                 pfn = page_to_pfn(page);
417                 frame_list[i] = pfn_to_mfn(pfn);
418
419                 scrub_page(page);
420
421                 if (xen_pv_domain() && !PageHighMem(page)) {
422                         ret = HYPERVISOR_update_va_mapping(
423                                 (unsigned long)__va(pfn << PAGE_SHIFT),
424                                 __pte_ma(0), 0);
425                         BUG_ON(ret);
426                 }
427
428         }
429
430         /* Ensure that ballooned highmem pages don't have kmaps. */
431         kmap_flush_unused();
432         flush_tlb_all();
433
434         /* No more mappings: invalidate P2M and add to balloon. */
435         for (i = 0; i < nr_pages; i++) {
436                 pfn = mfn_to_pfn(frame_list[i]);
437                 __set_phys_to_machine(pfn, INVALID_P2M_ENTRY);
438                 balloon_append(pfn_to_page(pfn));
439         }
440
441         set_xen_guest_handle(reservation.extent_start, frame_list);
442         reservation.nr_extents   = nr_pages;
443         ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation, &reservation);
444         BUG_ON(ret != nr_pages);
445
446         balloon_stats.current_pages -= nr_pages;
447
448         return state;
449 }
450
451 /*
452  * We avoid multiple worker processes conflicting via the balloon mutex.
453  * We may of course race updates of the target counts (which are protected
454  * by the balloon lock), or with changes to the Xen hard limit, but we will
455  * recover from these in time.
456  */
457 static void balloon_process(struct work_struct *work)
458 {
459         enum bp_state state = BP_DONE;
460         long credit;
461
462         mutex_lock(&balloon_mutex);
463
464         do {
465                 credit = current_credit();
466
467                 if (credit > 0) {
468                         if (balloon_is_inflated())
469                                 state = increase_reservation(credit);
470                         else
471                                 state = reserve_additional_memory(credit);
472                 }
473
474                 if (credit < 0)
475                         state = decrease_reservation(-credit, GFP_BALLOON);
476
477                 state = update_schedule(state);
478
479 #ifndef CONFIG_PREEMPT
480                 if (need_resched())
481                         schedule();
482 #endif
483         } while (credit && state == BP_DONE);
484
485         /* Schedule more work if there is some still to be done. */
486         if (state == BP_EAGAIN)
487                 schedule_delayed_work(&balloon_worker, balloon_stats.schedule_delay * HZ);
488
489         mutex_unlock(&balloon_mutex);
490 }
491
492 /* Resets the Xen limit, sets new target, and kicks off processing. */
493 void balloon_set_new_target(unsigned long target)
494 {
495         /* No need for lock. Not read-modify-write updates. */
496         balloon_stats.target_pages = target;
497         schedule_delayed_work(&balloon_worker, 0);
498 }
499 EXPORT_SYMBOL_GPL(balloon_set_new_target);
500
501 /**
502  * alloc_xenballooned_pages - get pages that have been ballooned out
503  * @nr_pages: Number of pages to get
504  * @pages: pages returned
505  * @highmem: highmem or lowmem pages
506  * @return 0 on success, error otherwise
507  */
508 int alloc_xenballooned_pages(int nr_pages, struct page **pages, bool highmem)
509 {
510         int pgno = 0;
511         struct page* page;
512         mutex_lock(&balloon_mutex);
513         while (pgno < nr_pages) {
514                 page = balloon_retrieve(highmem);
515                 if (page && PageHighMem(page) == highmem) {
516                         pages[pgno++] = page;
517                 } else {
518                         enum bp_state st;
519                         if (page)
520                                 balloon_append(page);
521                         st = decrease_reservation(nr_pages - pgno,
522                                         highmem ? GFP_HIGHUSER : GFP_USER);
523                         if (st != BP_DONE)
524                                 goto out_undo;
525                 }
526         }
527         mutex_unlock(&balloon_mutex);
528         return 0;
529  out_undo:
530         while (pgno)
531                 balloon_append(pages[--pgno]);
532         /* Free the memory back to the kernel soon */
533         schedule_delayed_work(&balloon_worker, 0);
534         mutex_unlock(&balloon_mutex);
535         return -ENOMEM;
536 }
537 EXPORT_SYMBOL(alloc_xenballooned_pages);
538
539 /**
540  * free_xenballooned_pages - return pages retrieved with get_ballooned_pages
541  * @nr_pages: Number of pages
542  * @pages: pages to return
543  */
544 void free_xenballooned_pages(int nr_pages, struct page** pages)
545 {
546         int i;
547
548         mutex_lock(&balloon_mutex);
549
550         for (i = 0; i < nr_pages; i++) {
551                 if (pages[i])
552                         balloon_append(pages[i]);
553         }
554
555         /* The balloon may be too large now. Shrink it if needed. */
556         if (current_credit())
557                 schedule_delayed_work(&balloon_worker, 0);
558
559         mutex_unlock(&balloon_mutex);
560 }
561 EXPORT_SYMBOL(free_xenballooned_pages);
562
563 static void __init balloon_add_region(unsigned long start_pfn,
564                                       unsigned long pages)
565 {
566         unsigned long pfn, extra_pfn_end;
567         struct page *page;
568
569         /*
570          * If the amount of usable memory has been limited (e.g., with
571          * the 'mem' command line parameter), don't add pages beyond
572          * this limit.
573          */
574         extra_pfn_end = min(max_pfn, start_pfn + pages);
575
576         for (pfn = start_pfn; pfn < extra_pfn_end; pfn++) {
577                 page = pfn_to_page(pfn);
578                 /* totalram_pages and totalhigh_pages do not
579                    include the boot-time balloon extension, so
580                    don't subtract from it. */
581                 __balloon_append(page);
582         }
583 }
584
585 static int __init balloon_init(void)
586 {
587         int i;
588
589         if (!xen_domain())
590                 return -ENODEV;
591
592         pr_info("xen/balloon: Initialising balloon driver.\n");
593
594         balloon_stats.current_pages = xen_pv_domain()
595                 ? min(xen_start_info->nr_pages - xen_released_pages, max_pfn)
596                 : max_pfn;
597         balloon_stats.target_pages  = balloon_stats.current_pages;
598         balloon_stats.balloon_low   = 0;
599         balloon_stats.balloon_high  = 0;
600
601         balloon_stats.schedule_delay = 1;
602         balloon_stats.max_schedule_delay = 32;
603         balloon_stats.retry_count = 1;
604         balloon_stats.max_retry_count = RETRY_UNLIMITED;
605
606 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
607         balloon_stats.hotplug_pages = 0;
608         balloon_stats.balloon_hotplug = 0;
609
610         set_online_page_callback(&xen_online_page);
611         register_memory_notifier(&xen_memory_nb);
612 #endif
613
614         /*
615          * Initialize the balloon with pages from the extra memory
616          * regions (see arch/x86/xen/setup.c).
617          */
618         for (i = 0; i < XEN_EXTRA_MEM_MAX_REGIONS; i++)
619                 if (xen_extra_mem[i].size)
620                         balloon_add_region(PFN_UP(xen_extra_mem[i].start),
621                                            PFN_DOWN(xen_extra_mem[i].size));
622
623         return 0;
624 }
625
626 subsys_initcall(balloon_init);
627
628 MODULE_LICENSE("GPL");