]> Pileus Git - ~andy/linux/blob - drivers/xen/balloon.c
xen/balloon: don't alloc page while non-preemptible
[~andy/linux] / 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 #define pr_fmt(fmt) "xen:" KBUILD_MODNAME ": " fmt
40
41 #include <linux/cpu.h>
42 #include <linux/kernel.h>
43 #include <linux/sched.h>
44 #include <linux/errno.h>
45 #include <linux/module.h>
46 #include <linux/mm.h>
47 #include <linux/bootmem.h>
48 #include <linux/pagemap.h>
49 #include <linux/highmem.h>
50 #include <linux/mutex.h>
51 #include <linux/list.h>
52 #include <linux/gfp.h>
53 #include <linux/notifier.h>
54 #include <linux/memory.h>
55 #include <linux/memory_hotplug.h>
56 #include <linux/percpu-defs.h>
57
58 #include <asm/page.h>
59 #include <asm/pgalloc.h>
60 #include <asm/pgtable.h>
61 #include <asm/tlb.h>
62
63 #include <asm/xen/hypervisor.h>
64 #include <asm/xen/hypercall.h>
65
66 #include <xen/xen.h>
67 #include <xen/interface/xen.h>
68 #include <xen/interface/memory.h>
69 #include <xen/balloon.h>
70 #include <xen/features.h>
71 #include <xen/page.h>
72
73 /*
74  * balloon_process() state:
75  *
76  * BP_DONE: done or nothing to do,
77  * BP_EAGAIN: error, go to sleep,
78  * BP_ECANCELED: error, balloon operation canceled.
79  */
80
81 enum bp_state {
82         BP_DONE,
83         BP_EAGAIN,
84         BP_ECANCELED
85 };
86
87
88 static DEFINE_MUTEX(balloon_mutex);
89
90 struct balloon_stats balloon_stats;
91 EXPORT_SYMBOL_GPL(balloon_stats);
92
93 /* We increase/decrease in batches which fit in a page */
94 static xen_pfn_t frame_list[PAGE_SIZE / sizeof(unsigned long)];
95 static DEFINE_PER_CPU(struct page *, balloon_scratch_page);
96
97
98 /* List of ballooned pages, threaded through the mem_map array. */
99 static LIST_HEAD(ballooned_pages);
100
101 /* Main work function, always executed in process context. */
102 static void balloon_process(struct work_struct *work);
103 static DECLARE_DELAYED_WORK(balloon_worker, balloon_process);
104
105 /* When ballooning out (allocating memory to return to Xen) we don't really
106    want the kernel to try too hard since that can trigger the oom killer. */
107 #define GFP_BALLOON \
108         (GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC)
109
110 static void scrub_page(struct page *page)
111 {
112 #ifdef CONFIG_XEN_SCRUB_PAGES
113         clear_highpage(page);
114 #endif
115 }
116
117 /* balloon_append: add the given page to the balloon. */
118 static void __balloon_append(struct page *page)
119 {
120         /* Lowmem is re-populated first, so highmem pages go at list tail. */
121         if (PageHighMem(page)) {
122                 list_add_tail(&page->lru, &ballooned_pages);
123                 balloon_stats.balloon_high++;
124         } else {
125                 list_add(&page->lru, &ballooned_pages);
126                 balloon_stats.balloon_low++;
127         }
128 }
129
130 static void balloon_append(struct page *page)
131 {
132         __balloon_append(page);
133         adjust_managed_page_count(page, -1);
134 }
135
136 /* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
137 static struct page *balloon_retrieve(bool prefer_highmem)
138 {
139         struct page *page;
140
141         if (list_empty(&ballooned_pages))
142                 return NULL;
143
144         if (prefer_highmem)
145                 page = list_entry(ballooned_pages.prev, struct page, lru);
146         else
147                 page = list_entry(ballooned_pages.next, struct page, lru);
148         list_del(&page->lru);
149
150         if (PageHighMem(page))
151                 balloon_stats.balloon_high--;
152         else
153                 balloon_stats.balloon_low--;
154
155         adjust_managed_page_count(page, 1);
156
157         return page;
158 }
159
160 static struct page *balloon_first_page(void)
161 {
162         if (list_empty(&ballooned_pages))
163                 return NULL;
164         return list_entry(ballooned_pages.next, struct page, lru);
165 }
166
167 static struct page *balloon_next_page(struct page *page)
168 {
169         struct list_head *next = page->lru.next;
170         if (next == &ballooned_pages)
171                 return NULL;
172         return list_entry(next, struct page, lru);
173 }
174
175 static enum bp_state update_schedule(enum bp_state state)
176 {
177         if (state == BP_DONE) {
178                 balloon_stats.schedule_delay = 1;
179                 balloon_stats.retry_count = 1;
180                 return BP_DONE;
181         }
182
183         ++balloon_stats.retry_count;
184
185         if (balloon_stats.max_retry_count != RETRY_UNLIMITED &&
186                         balloon_stats.retry_count > balloon_stats.max_retry_count) {
187                 balloon_stats.schedule_delay = 1;
188                 balloon_stats.retry_count = 1;
189                 return BP_ECANCELED;
190         }
191
192         balloon_stats.schedule_delay <<= 1;
193
194         if (balloon_stats.schedule_delay > balloon_stats.max_schedule_delay)
195                 balloon_stats.schedule_delay = balloon_stats.max_schedule_delay;
196
197         return BP_EAGAIN;
198 }
199
200 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
201 static long current_credit(void)
202 {
203         return balloon_stats.target_pages - balloon_stats.current_pages -
204                 balloon_stats.hotplug_pages;
205 }
206
207 static bool balloon_is_inflated(void)
208 {
209         if (balloon_stats.balloon_low || balloon_stats.balloon_high ||
210                         balloon_stats.balloon_hotplug)
211                 return true;
212         else
213                 return false;
214 }
215
216 /*
217  * reserve_additional_memory() adds memory region of size >= credit above
218  * max_pfn. New region is section aligned and size is modified to be multiple
219  * of section size. Those features allow optimal use of address space and
220  * establish proper alignment when this function is called first time after
221  * boot (last section not fully populated at boot time contains unused memory
222  * pages with PG_reserved bit not set; online_pages_range() does not allow page
223  * onlining in whole range if first onlined page does not have PG_reserved
224  * bit set). Real size of added memory is established at page onlining stage.
225  */
226
227 static enum bp_state reserve_additional_memory(long credit)
228 {
229         int nid, rc;
230         u64 hotplug_start_paddr;
231         unsigned long balloon_hotplug = credit;
232
233         hotplug_start_paddr = PFN_PHYS(SECTION_ALIGN_UP(max_pfn));
234         balloon_hotplug = round_up(balloon_hotplug, PAGES_PER_SECTION);
235         nid = memory_add_physaddr_to_nid(hotplug_start_paddr);
236
237         rc = add_memory(nid, hotplug_start_paddr, balloon_hotplug << PAGE_SHIFT);
238
239         if (rc) {
240                 pr_info("%s: add_memory() failed: %i\n", __func__, rc);
241                 return BP_EAGAIN;
242         }
243
244         balloon_hotplug -= credit;
245
246         balloon_stats.hotplug_pages += credit;
247         balloon_stats.balloon_hotplug = balloon_hotplug;
248
249         return BP_DONE;
250 }
251
252 static void xen_online_page(struct page *page)
253 {
254         __online_page_set_limits(page);
255
256         mutex_lock(&balloon_mutex);
257
258         __balloon_append(page);
259
260         if (balloon_stats.hotplug_pages)
261                 --balloon_stats.hotplug_pages;
262         else
263                 --balloon_stats.balloon_hotplug;
264
265         mutex_unlock(&balloon_mutex);
266 }
267
268 static int xen_memory_notifier(struct notifier_block *nb, unsigned long val, void *v)
269 {
270         if (val == MEM_ONLINE)
271                 schedule_delayed_work(&balloon_worker, 0);
272
273         return NOTIFY_OK;
274 }
275
276 static struct notifier_block xen_memory_nb = {
277         .notifier_call = xen_memory_notifier,
278         .priority = 0
279 };
280 #else
281 static long current_credit(void)
282 {
283         unsigned long target = balloon_stats.target_pages;
284
285         target = min(target,
286                      balloon_stats.current_pages +
287                      balloon_stats.balloon_low +
288                      balloon_stats.balloon_high);
289
290         return target - balloon_stats.current_pages;
291 }
292
293 static bool balloon_is_inflated(void)
294 {
295         if (balloon_stats.balloon_low || balloon_stats.balloon_high)
296                 return true;
297         else
298                 return false;
299 }
300
301 static enum bp_state reserve_additional_memory(long credit)
302 {
303         balloon_stats.target_pages = balloon_stats.current_pages;
304         return BP_DONE;
305 }
306 #endif /* CONFIG_XEN_BALLOON_MEMORY_HOTPLUG */
307
308 static enum bp_state increase_reservation(unsigned long nr_pages)
309 {
310         int rc;
311         unsigned long  pfn, i;
312         struct page   *page;
313         struct xen_memory_reservation reservation = {
314                 .address_bits = 0,
315                 .extent_order = 0,
316                 .domid        = DOMID_SELF
317         };
318
319 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
320         if (!balloon_stats.balloon_low && !balloon_stats.balloon_high) {
321                 nr_pages = min(nr_pages, balloon_stats.balloon_hotplug);
322                 balloon_stats.hotplug_pages += nr_pages;
323                 balloon_stats.balloon_hotplug -= nr_pages;
324                 return BP_DONE;
325         }
326 #endif
327
328         if (nr_pages > ARRAY_SIZE(frame_list))
329                 nr_pages = ARRAY_SIZE(frame_list);
330
331         page = balloon_first_page();
332         for (i = 0; i < nr_pages; i++) {
333                 if (!page) {
334                         nr_pages = i;
335                         break;
336                 }
337                 frame_list[i] = page_to_pfn(page);
338                 page = balloon_next_page(page);
339         }
340
341         set_xen_guest_handle(reservation.extent_start, frame_list);
342         reservation.nr_extents = nr_pages;
343         rc = HYPERVISOR_memory_op(XENMEM_populate_physmap, &reservation);
344         if (rc <= 0)
345                 return BP_EAGAIN;
346
347         for (i = 0; i < rc; i++) {
348                 page = balloon_retrieve(false);
349                 BUG_ON(page == NULL);
350
351                 pfn = page_to_pfn(page);
352
353                 set_phys_to_machine(pfn, frame_list[i]);
354
355 #ifdef CONFIG_XEN_HAVE_PVMMU
356                 /* Link back into the page tables if not highmem. */
357                 if (xen_pv_domain() && !PageHighMem(page)) {
358                         int ret;
359                         ret = HYPERVISOR_update_va_mapping(
360                                 (unsigned long)__va(pfn << PAGE_SHIFT),
361                                 mfn_pte(frame_list[i], PAGE_KERNEL),
362                                 0);
363                         BUG_ON(ret);
364                 }
365 #endif
366
367                 /* Relinquish the page back to the allocator. */
368                 __free_reserved_page(page);
369         }
370
371         balloon_stats.current_pages += rc;
372
373         return BP_DONE;
374 }
375
376 static enum bp_state decrease_reservation(unsigned long nr_pages, gfp_t gfp)
377 {
378         enum bp_state state = BP_DONE;
379         unsigned long  pfn, i;
380         struct page   *page;
381         struct page   *scratch_page;
382         int ret;
383         struct xen_memory_reservation reservation = {
384                 .address_bits = 0,
385                 .extent_order = 0,
386                 .domid        = DOMID_SELF
387         };
388
389 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
390         if (balloon_stats.hotplug_pages) {
391                 nr_pages = min(nr_pages, balloon_stats.hotplug_pages);
392                 balloon_stats.hotplug_pages -= nr_pages;
393                 balloon_stats.balloon_hotplug += nr_pages;
394                 return BP_DONE;
395         }
396 #endif
397
398         if (nr_pages > ARRAY_SIZE(frame_list))
399                 nr_pages = ARRAY_SIZE(frame_list);
400
401         for (i = 0; i < nr_pages; i++) {
402                 page = alloc_page(gfp);
403                 if (page == NULL) {
404                         nr_pages = i;
405                         state = BP_EAGAIN;
406                         break;
407                 }
408
409                 pfn = page_to_pfn(page);
410                 frame_list[i] = pfn_to_mfn(pfn);
411
412                 scrub_page(page);
413
414                 /*
415                  * Ballooned out frames are effectively replaced with
416                  * a scratch frame.  Ensure direct mappings and the
417                  * p2m are consistent.
418                  */
419                 scratch_page = get_balloon_scratch_page();
420 #ifdef CONFIG_XEN_HAVE_PVMMU
421                 if (xen_pv_domain() && !PageHighMem(page)) {
422                         ret = HYPERVISOR_update_va_mapping(
423                                 (unsigned long)__va(pfn << PAGE_SHIFT),
424                                 pfn_pte(page_to_pfn(scratch_page),
425                                         PAGE_KERNEL_RO), 0);
426                         BUG_ON(ret);
427                 }
428 #endif
429                 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
430                         unsigned long p;
431                         p = page_to_pfn(scratch_page);
432                         __set_phys_to_machine(pfn, pfn_to_mfn(p));
433                 }
434                 put_balloon_scratch_page();
435
436                 balloon_append(pfn_to_page(pfn));
437         }
438
439         /* Ensure that ballooned highmem pages don't have kmaps. */
440         kmap_flush_unused();
441         flush_tlb_all();
442
443         set_xen_guest_handle(reservation.extent_start, frame_list);
444         reservation.nr_extents   = nr_pages;
445         ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation, &reservation);
446         BUG_ON(ret != nr_pages);
447
448         balloon_stats.current_pages -= nr_pages;
449
450         return state;
451 }
452
453 /*
454  * We avoid multiple worker processes conflicting via the balloon mutex.
455  * We may of course race updates of the target counts (which are protected
456  * by the balloon lock), or with changes to the Xen hard limit, but we will
457  * recover from these in time.
458  */
459 static void balloon_process(struct work_struct *work)
460 {
461         enum bp_state state = BP_DONE;
462         long credit;
463
464         mutex_lock(&balloon_mutex);
465
466         do {
467                 credit = current_credit();
468
469                 if (credit > 0) {
470                         if (balloon_is_inflated())
471                                 state = increase_reservation(credit);
472                         else
473                                 state = reserve_additional_memory(credit);
474                 }
475
476                 if (credit < 0)
477                         state = decrease_reservation(-credit, GFP_BALLOON);
478
479                 state = update_schedule(state);
480
481 #ifndef CONFIG_PREEMPT
482                 if (need_resched())
483                         schedule();
484 #endif
485         } while (credit && state == BP_DONE);
486
487         /* Schedule more work if there is some still to be done. */
488         if (state == BP_EAGAIN)
489                 schedule_delayed_work(&balloon_worker, balloon_stats.schedule_delay * HZ);
490
491         mutex_unlock(&balloon_mutex);
492 }
493
494 struct page *get_balloon_scratch_page(void)
495 {
496         struct page *ret = get_cpu_var(balloon_scratch_page);
497         BUG_ON(ret == NULL);
498         return ret;
499 }
500
501 void put_balloon_scratch_page(void)
502 {
503         put_cpu_var(balloon_scratch_page);
504 }
505
506 /* Resets the Xen limit, sets new target, and kicks off processing. */
507 void balloon_set_new_target(unsigned long target)
508 {
509         /* No need for lock. Not read-modify-write updates. */
510         balloon_stats.target_pages = target;
511         schedule_delayed_work(&balloon_worker, 0);
512 }
513 EXPORT_SYMBOL_GPL(balloon_set_new_target);
514
515 /**
516  * alloc_xenballooned_pages - get pages that have been ballooned out
517  * @nr_pages: Number of pages to get
518  * @pages: pages returned
519  * @highmem: allow highmem pages
520  * @return 0 on success, error otherwise
521  */
522 int alloc_xenballooned_pages(int nr_pages, struct page **pages, bool highmem)
523 {
524         int pgno = 0;
525         struct page *page;
526         mutex_lock(&balloon_mutex);
527         while (pgno < nr_pages) {
528                 page = balloon_retrieve(highmem);
529                 if (page && (highmem || !PageHighMem(page))) {
530                         pages[pgno++] = page;
531                 } else {
532                         enum bp_state st;
533                         if (page)
534                                 balloon_append(page);
535                         st = decrease_reservation(nr_pages - pgno,
536                                         highmem ? GFP_HIGHUSER : GFP_USER);
537                         if (st != BP_DONE)
538                                 goto out_undo;
539                 }
540         }
541         mutex_unlock(&balloon_mutex);
542         return 0;
543  out_undo:
544         while (pgno)
545                 balloon_append(pages[--pgno]);
546         /* Free the memory back to the kernel soon */
547         schedule_delayed_work(&balloon_worker, 0);
548         mutex_unlock(&balloon_mutex);
549         return -ENOMEM;
550 }
551 EXPORT_SYMBOL(alloc_xenballooned_pages);
552
553 /**
554  * free_xenballooned_pages - return pages retrieved with get_ballooned_pages
555  * @nr_pages: Number of pages
556  * @pages: pages to return
557  */
558 void free_xenballooned_pages(int nr_pages, struct page **pages)
559 {
560         int i;
561
562         mutex_lock(&balloon_mutex);
563
564         for (i = 0; i < nr_pages; i++) {
565                 if (pages[i])
566                         balloon_append(pages[i]);
567         }
568
569         /* The balloon may be too large now. Shrink it if needed. */
570         if (current_credit())
571                 schedule_delayed_work(&balloon_worker, 0);
572
573         mutex_unlock(&balloon_mutex);
574 }
575 EXPORT_SYMBOL(free_xenballooned_pages);
576
577 static void __init balloon_add_region(unsigned long start_pfn,
578                                       unsigned long pages)
579 {
580         unsigned long pfn, extra_pfn_end;
581         struct page *page;
582
583         /*
584          * If the amount of usable memory has been limited (e.g., with
585          * the 'mem' command line parameter), don't add pages beyond
586          * this limit.
587          */
588         extra_pfn_end = min(max_pfn, start_pfn + pages);
589
590         for (pfn = start_pfn; pfn < extra_pfn_end; pfn++) {
591                 page = pfn_to_page(pfn);
592                 /* totalram_pages and totalhigh_pages do not
593                    include the boot-time balloon extension, so
594                    don't subtract from it. */
595                 __balloon_append(page);
596         }
597 }
598
599 static int __cpuinit balloon_cpu_notify(struct notifier_block *self,
600                                     unsigned long action, void *hcpu)
601 {
602         int cpu = (long)hcpu;
603         switch (action) {
604         case CPU_UP_PREPARE:
605                 if (per_cpu(balloon_scratch_page, cpu) != NULL)
606                         break;
607                 per_cpu(balloon_scratch_page, cpu) = alloc_page(GFP_KERNEL);
608                 if (per_cpu(balloon_scratch_page, cpu) == NULL) {
609                         pr_warn("Failed to allocate balloon_scratch_page for cpu %d\n", cpu);
610                         return NOTIFY_BAD;
611                 }
612                 break;
613         default:
614                 break;
615         }
616         return NOTIFY_OK;
617 }
618
619 static struct notifier_block balloon_cpu_notifier __cpuinitdata = {
620         .notifier_call  = balloon_cpu_notify,
621 };
622
623 static int __init balloon_init(void)
624 {
625         int i, cpu;
626
627         if (!xen_domain())
628                 return -ENODEV;
629
630         for_each_online_cpu(cpu)
631         {
632                 per_cpu(balloon_scratch_page, cpu) = alloc_page(GFP_KERNEL);
633                 if (per_cpu(balloon_scratch_page, cpu) == NULL) {
634                         pr_warn("Failed to allocate balloon_scratch_page for cpu %d\n", cpu);
635                         return -ENOMEM;
636                 }
637         }
638         register_cpu_notifier(&balloon_cpu_notifier);
639
640         pr_info("Initialising balloon driver\n");
641
642         balloon_stats.current_pages = xen_pv_domain()
643                 ? min(xen_start_info->nr_pages - xen_released_pages, max_pfn)
644                 : max_pfn;
645         balloon_stats.target_pages  = balloon_stats.current_pages;
646         balloon_stats.balloon_low   = 0;
647         balloon_stats.balloon_high  = 0;
648
649         balloon_stats.schedule_delay = 1;
650         balloon_stats.max_schedule_delay = 32;
651         balloon_stats.retry_count = 1;
652         balloon_stats.max_retry_count = RETRY_UNLIMITED;
653
654 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
655         balloon_stats.hotplug_pages = 0;
656         balloon_stats.balloon_hotplug = 0;
657
658         set_online_page_callback(&xen_online_page);
659         register_memory_notifier(&xen_memory_nb);
660 #endif
661
662         /*
663          * Initialize the balloon with pages from the extra memory
664          * regions (see arch/x86/xen/setup.c).
665          */
666         for (i = 0; i < XEN_EXTRA_MEM_MAX_REGIONS; i++)
667                 if (xen_extra_mem[i].size)
668                         balloon_add_region(PFN_UP(xen_extra_mem[i].start),
669                                            PFN_DOWN(xen_extra_mem[i].size));
670
671         return 0;
672 }
673
674 subsys_initcall(balloon_init);
675
676 static int __init balloon_clear(void)
677 {
678         int cpu;
679
680         for_each_possible_cpu(cpu)
681                 per_cpu(balloon_scratch_page, cpu) = NULL;
682
683         return 0;
684 }
685 early_initcall(balloon_clear);
686
687 MODULE_LICENSE("GPL");