25fd1bd949d8c952347ab00a6cfb4fa19dee1711
[cascardo/linux.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 #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 #include <linux/slab.h>
58 #include <linux/sysctl.h>
59
60 #include <asm/page.h>
61 #include <asm/pgalloc.h>
62 #include <asm/pgtable.h>
63 #include <asm/tlb.h>
64
65 #include <asm/xen/hypervisor.h>
66 #include <asm/xen/hypercall.h>
67
68 #include <xen/xen.h>
69 #include <xen/interface/xen.h>
70 #include <xen/interface/memory.h>
71 #include <xen/balloon.h>
72 #include <xen/features.h>
73 #include <xen/page.h>
74
75 static int xen_hotplug_unpopulated;
76
77 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
78
79 static int zero;
80 static int one = 1;
81
82 static struct ctl_table balloon_table[] = {
83         {
84                 .procname       = "hotplug_unpopulated",
85                 .data           = &xen_hotplug_unpopulated,
86                 .maxlen         = sizeof(int),
87                 .mode           = 0644,
88                 .proc_handler   = proc_dointvec_minmax,
89                 .extra1         = &zero,
90                 .extra2         = &one,
91         },
92         { }
93 };
94
95 static struct ctl_table balloon_root[] = {
96         {
97                 .procname       = "balloon",
98                 .mode           = 0555,
99                 .child          = balloon_table,
100         },
101         { }
102 };
103
104 static struct ctl_table xen_root[] = {
105         {
106                 .procname       = "xen",
107                 .mode           = 0555,
108                 .child          = balloon_root,
109         },
110         { }
111 };
112
113 #endif
114
115 /*
116  * balloon_process() state:
117  *
118  * BP_DONE: done or nothing to do,
119  * BP_WAIT: wait to be rescheduled,
120  * BP_EAGAIN: error, go to sleep,
121  * BP_ECANCELED: error, balloon operation canceled.
122  */
123
124 enum bp_state {
125         BP_DONE,
126         BP_WAIT,
127         BP_EAGAIN,
128         BP_ECANCELED
129 };
130
131
132 static DEFINE_MUTEX(balloon_mutex);
133
134 struct balloon_stats balloon_stats;
135 EXPORT_SYMBOL_GPL(balloon_stats);
136
137 /* We increase/decrease in batches which fit in a page */
138 static xen_pfn_t frame_list[PAGE_SIZE / sizeof(unsigned long)];
139
140
141 /* List of ballooned pages, threaded through the mem_map array. */
142 static LIST_HEAD(ballooned_pages);
143 static DECLARE_WAIT_QUEUE_HEAD(balloon_wq);
144
145 /* Main work function, always executed in process context. */
146 static void balloon_process(struct work_struct *work);
147 static DECLARE_DELAYED_WORK(balloon_worker, balloon_process);
148
149 /* When ballooning out (allocating memory to return to Xen) we don't really
150    want the kernel to try too hard since that can trigger the oom killer. */
151 #define GFP_BALLOON \
152         (GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC)
153
154 static void scrub_page(struct page *page)
155 {
156 #ifdef CONFIG_XEN_SCRUB_PAGES
157         clear_highpage(page);
158 #endif
159 }
160
161 /* balloon_append: add the given page to the balloon. */
162 static void __balloon_append(struct page *page)
163 {
164         /* Lowmem is re-populated first, so highmem pages go at list tail. */
165         if (PageHighMem(page)) {
166                 list_add_tail(&page->lru, &ballooned_pages);
167                 balloon_stats.balloon_high++;
168         } else {
169                 list_add(&page->lru, &ballooned_pages);
170                 balloon_stats.balloon_low++;
171         }
172         wake_up(&balloon_wq);
173 }
174
175 static void balloon_append(struct page *page)
176 {
177         __balloon_append(page);
178         adjust_managed_page_count(page, -1);
179 }
180
181 /* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
182 static struct page *balloon_retrieve(bool require_lowmem)
183 {
184         struct page *page;
185
186         if (list_empty(&ballooned_pages))
187                 return NULL;
188
189         page = list_entry(ballooned_pages.next, struct page, lru);
190         if (require_lowmem && PageHighMem(page))
191                 return NULL;
192         list_del(&page->lru);
193
194         if (PageHighMem(page))
195                 balloon_stats.balloon_high--;
196         else
197                 balloon_stats.balloon_low--;
198
199         adjust_managed_page_count(page, 1);
200
201         return page;
202 }
203
204 static struct page *balloon_next_page(struct page *page)
205 {
206         struct list_head *next = page->lru.next;
207         if (next == &ballooned_pages)
208                 return NULL;
209         return list_entry(next, struct page, lru);
210 }
211
212 static enum bp_state update_schedule(enum bp_state state)
213 {
214         if (state == BP_WAIT)
215                 return BP_WAIT;
216
217         if (state == BP_ECANCELED)
218                 return BP_ECANCELED;
219
220         if (state == BP_DONE) {
221                 balloon_stats.schedule_delay = 1;
222                 balloon_stats.retry_count = 1;
223                 return BP_DONE;
224         }
225
226         ++balloon_stats.retry_count;
227
228         if (balloon_stats.max_retry_count != RETRY_UNLIMITED &&
229                         balloon_stats.retry_count > balloon_stats.max_retry_count) {
230                 balloon_stats.schedule_delay = 1;
231                 balloon_stats.retry_count = 1;
232                 return BP_ECANCELED;
233         }
234
235         balloon_stats.schedule_delay <<= 1;
236
237         if (balloon_stats.schedule_delay > balloon_stats.max_schedule_delay)
238                 balloon_stats.schedule_delay = balloon_stats.max_schedule_delay;
239
240         return BP_EAGAIN;
241 }
242
243 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
244 static struct resource *additional_memory_resource(phys_addr_t size)
245 {
246         struct resource *res;
247         int ret;
248
249         res = kzalloc(sizeof(*res), GFP_KERNEL);
250         if (!res)
251                 return NULL;
252
253         res->name = "System RAM";
254         res->flags = IORESOURCE_MEM | IORESOURCE_BUSY;
255
256         ret = allocate_resource(&iomem_resource, res,
257                                 size, 0, -1,
258                                 PAGES_PER_SECTION * PAGE_SIZE, NULL, NULL);
259         if (ret < 0) {
260                 pr_err("Cannot allocate new System RAM resource\n");
261                 kfree(res);
262                 return NULL;
263         }
264
265         return res;
266 }
267
268 static void release_memory_resource(struct resource *resource)
269 {
270         if (!resource)
271                 return;
272
273         /*
274          * No need to reset region to identity mapped since we now
275          * know that no I/O can be in this region
276          */
277         release_resource(resource);
278         kfree(resource);
279 }
280
281 static enum bp_state reserve_additional_memory(void)
282 {
283         long credit;
284         struct resource *resource;
285         int nid, rc;
286         unsigned long balloon_hotplug;
287
288         credit = balloon_stats.target_pages + balloon_stats.target_unpopulated
289                 - balloon_stats.total_pages;
290
291         /*
292          * Already hotplugged enough pages?  Wait for them to be
293          * onlined.
294          */
295         if (credit <= 0)
296                 return BP_WAIT;
297
298         balloon_hotplug = round_up(credit, PAGES_PER_SECTION);
299
300         resource = additional_memory_resource(balloon_hotplug * PAGE_SIZE);
301         if (!resource)
302                 goto err;
303
304         nid = memory_add_physaddr_to_nid(resource->start);
305
306 #ifdef CONFIG_XEN_HAVE_PVMMU
307         /*
308          * add_memory() will build page tables for the new memory so
309          * the p2m must contain invalid entries so the correct
310          * non-present PTEs will be written.
311          *
312          * If a failure occurs, the original (identity) p2m entries
313          * are not restored since this region is now known not to
314          * conflict with any devices.
315          */ 
316         if (!xen_feature(XENFEAT_auto_translated_physmap)) {
317                 unsigned long pfn, i;
318
319                 pfn = PFN_DOWN(resource->start);
320                 for (i = 0; i < balloon_hotplug; i++) {
321                         if (!set_phys_to_machine(pfn + i, INVALID_P2M_ENTRY)) {
322                                 pr_warn("set_phys_to_machine() failed, no memory added\n");
323                                 goto err;
324                         }
325                 }
326         }
327 #endif
328
329         rc = add_memory_resource(nid, resource);
330         if (rc) {
331                 pr_warn("Cannot add additional memory (%i)\n", rc);
332                 goto err;
333         }
334
335         balloon_stats.total_pages += balloon_hotplug;
336
337         return BP_WAIT;
338   err:
339         release_memory_resource(resource);
340         return BP_ECANCELED;
341 }
342
343 static void xen_online_page(struct page *page)
344 {
345         __online_page_set_limits(page);
346
347         mutex_lock(&balloon_mutex);
348
349         __balloon_append(page);
350
351         mutex_unlock(&balloon_mutex);
352 }
353
354 static int xen_memory_notifier(struct notifier_block *nb, unsigned long val, void *v)
355 {
356         if (val == MEM_ONLINE)
357                 schedule_delayed_work(&balloon_worker, 0);
358
359         return NOTIFY_OK;
360 }
361
362 static struct notifier_block xen_memory_nb = {
363         .notifier_call = xen_memory_notifier,
364         .priority = 0
365 };
366 #else
367 static enum bp_state reserve_additional_memory(void)
368 {
369         balloon_stats.target_pages = balloon_stats.current_pages;
370         return BP_ECANCELED;
371 }
372 #endif /* CONFIG_XEN_BALLOON_MEMORY_HOTPLUG */
373
374 static long current_credit(void)
375 {
376         return balloon_stats.target_pages - balloon_stats.current_pages;
377 }
378
379 static bool balloon_is_inflated(void)
380 {
381         return balloon_stats.balloon_low || balloon_stats.balloon_high;
382 }
383
384 static enum bp_state increase_reservation(unsigned long nr_pages)
385 {
386         int rc;
387         unsigned long  pfn, i;
388         struct page   *page;
389         struct xen_memory_reservation reservation = {
390                 .address_bits = 0,
391                 .extent_order = 0,
392                 .domid        = DOMID_SELF
393         };
394
395         if (nr_pages > ARRAY_SIZE(frame_list))
396                 nr_pages = ARRAY_SIZE(frame_list);
397
398         page = list_first_entry_or_null(&ballooned_pages, struct page, lru);
399         for (i = 0; i < nr_pages; i++) {
400                 if (!page) {
401                         nr_pages = i;
402                         break;
403                 }
404                 frame_list[i] = page_to_pfn(page);
405                 page = balloon_next_page(page);
406         }
407
408         set_xen_guest_handle(reservation.extent_start, frame_list);
409         reservation.nr_extents = nr_pages;
410         rc = HYPERVISOR_memory_op(XENMEM_populate_physmap, &reservation);
411         if (rc <= 0)
412                 return BP_EAGAIN;
413
414         for (i = 0; i < rc; i++) {
415                 page = balloon_retrieve(false);
416                 BUG_ON(page == NULL);
417
418                 pfn = page_to_pfn(page);
419
420 #ifdef CONFIG_XEN_HAVE_PVMMU
421                 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
422                         set_phys_to_machine(pfn, frame_list[i]);
423
424                         /* Link back into the page tables if not highmem. */
425                         if (!PageHighMem(page)) {
426                                 int ret;
427                                 ret = HYPERVISOR_update_va_mapping(
428                                                 (unsigned long)__va(pfn << PAGE_SHIFT),
429                                                 mfn_pte(frame_list[i], PAGE_KERNEL),
430                                                 0);
431                                 BUG_ON(ret);
432                         }
433                 }
434 #endif
435
436                 /* Relinquish the page back to the allocator. */
437                 __free_reserved_page(page);
438         }
439
440         balloon_stats.current_pages += rc;
441
442         return BP_DONE;
443 }
444
445 static enum bp_state decrease_reservation(unsigned long nr_pages, gfp_t gfp)
446 {
447         enum bp_state state = BP_DONE;
448         unsigned long  pfn, i;
449         struct page   *page;
450         int ret;
451         struct xen_memory_reservation reservation = {
452                 .address_bits = 0,
453                 .extent_order = 0,
454                 .domid        = DOMID_SELF
455         };
456
457         if (nr_pages > ARRAY_SIZE(frame_list))
458                 nr_pages = ARRAY_SIZE(frame_list);
459
460         for (i = 0; i < nr_pages; i++) {
461                 page = alloc_page(gfp);
462                 if (page == NULL) {
463                         nr_pages = i;
464                         state = BP_EAGAIN;
465                         break;
466                 }
467                 scrub_page(page);
468
469                 frame_list[i] = page_to_pfn(page);
470         }
471
472         /*
473          * Ensure that ballooned highmem pages don't have kmaps.
474          *
475          * Do this before changing the p2m as kmap_flush_unused()
476          * reads PTEs to obtain pages (and hence needs the original
477          * p2m entry).
478          */
479         kmap_flush_unused();
480
481         /* Update direct mapping, invalidate P2M, and add to balloon. */
482         for (i = 0; i < nr_pages; i++) {
483                 pfn = frame_list[i];
484                 frame_list[i] = pfn_to_gfn(pfn);
485                 page = pfn_to_page(pfn);
486
487 #ifdef CONFIG_XEN_HAVE_PVMMU
488                 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
489                         if (!PageHighMem(page)) {
490                                 ret = HYPERVISOR_update_va_mapping(
491                                                 (unsigned long)__va(pfn << PAGE_SHIFT),
492                                                 __pte_ma(0), 0);
493                                 BUG_ON(ret);
494                         }
495                         __set_phys_to_machine(pfn, INVALID_P2M_ENTRY);
496                 }
497 #endif
498
499                 balloon_append(page);
500         }
501
502         flush_tlb_all();
503
504         set_xen_guest_handle(reservation.extent_start, frame_list);
505         reservation.nr_extents   = nr_pages;
506         ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation, &reservation);
507         BUG_ON(ret != nr_pages);
508
509         balloon_stats.current_pages -= nr_pages;
510
511         return state;
512 }
513
514 /*
515  * As this is a work item it is guaranteed to run as a single instance only.
516  * We may of course race updates of the target counts (which are protected
517  * by the balloon lock), or with changes to the Xen hard limit, but we will
518  * recover from these in time.
519  */
520 static void balloon_process(struct work_struct *work)
521 {
522         enum bp_state state = BP_DONE;
523         long credit;
524
525
526         do {
527                 mutex_lock(&balloon_mutex);
528
529                 credit = current_credit();
530
531                 if (credit > 0) {
532                         if (balloon_is_inflated())
533                                 state = increase_reservation(credit);
534                         else
535                                 state = reserve_additional_memory();
536                 }
537
538                 if (credit < 0)
539                         state = decrease_reservation(-credit, GFP_BALLOON);
540
541                 state = update_schedule(state);
542
543                 mutex_unlock(&balloon_mutex);
544
545                 cond_resched();
546
547         } while (credit && state == BP_DONE);
548
549         /* Schedule more work if there is some still to be done. */
550         if (state == BP_EAGAIN)
551                 schedule_delayed_work(&balloon_worker, balloon_stats.schedule_delay * HZ);
552 }
553
554 /* Resets the Xen limit, sets new target, and kicks off processing. */
555 void balloon_set_new_target(unsigned long target)
556 {
557         /* No need for lock. Not read-modify-write updates. */
558         balloon_stats.target_pages = target;
559         schedule_delayed_work(&balloon_worker, 0);
560 }
561 EXPORT_SYMBOL_GPL(balloon_set_new_target);
562
563 static int add_ballooned_pages(int nr_pages)
564 {
565         enum bp_state st;
566
567         if (xen_hotplug_unpopulated) {
568                 st = reserve_additional_memory();
569                 if (st != BP_ECANCELED) {
570                         mutex_unlock(&balloon_mutex);
571                         wait_event(balloon_wq,
572                                    !list_empty(&ballooned_pages));
573                         mutex_lock(&balloon_mutex);
574                         return 0;
575                 }
576         }
577
578         st = decrease_reservation(nr_pages, GFP_USER);
579         if (st != BP_DONE)
580                 return -ENOMEM;
581
582         return 0;
583 }
584
585 /**
586  * alloc_xenballooned_pages - get pages that have been ballooned out
587  * @nr_pages: Number of pages to get
588  * @pages: pages returned
589  * @return 0 on success, error otherwise
590  */
591 int alloc_xenballooned_pages(int nr_pages, struct page **pages)
592 {
593         int pgno = 0;
594         struct page *page;
595         int ret;
596
597         mutex_lock(&balloon_mutex);
598
599         balloon_stats.target_unpopulated += nr_pages;
600
601         while (pgno < nr_pages) {
602                 page = balloon_retrieve(true);
603                 if (page) {
604                         pages[pgno++] = page;
605                 } else {
606                         ret = add_ballooned_pages(nr_pages - pgno);
607                         if (ret < 0)
608                                 goto out_undo;
609                 }
610         }
611         mutex_unlock(&balloon_mutex);
612         return 0;
613  out_undo:
614         mutex_unlock(&balloon_mutex);
615         free_xenballooned_pages(pgno, pages);
616         return ret;
617 }
618 EXPORT_SYMBOL(alloc_xenballooned_pages);
619
620 /**
621  * free_xenballooned_pages - return pages retrieved with get_ballooned_pages
622  * @nr_pages: Number of pages
623  * @pages: pages to return
624  */
625 void free_xenballooned_pages(int nr_pages, struct page **pages)
626 {
627         int i;
628
629         mutex_lock(&balloon_mutex);
630
631         for (i = 0; i < nr_pages; i++) {
632                 if (pages[i])
633                         balloon_append(pages[i]);
634         }
635
636         balloon_stats.target_unpopulated -= nr_pages;
637
638         /* The balloon may be too large now. Shrink it if needed. */
639         if (current_credit())
640                 schedule_delayed_work(&balloon_worker, 0);
641
642         mutex_unlock(&balloon_mutex);
643 }
644 EXPORT_SYMBOL(free_xenballooned_pages);
645
646 static void __init balloon_add_region(unsigned long start_pfn,
647                                       unsigned long pages)
648 {
649         unsigned long pfn, extra_pfn_end;
650         struct page *page;
651
652         /*
653          * If the amount of usable memory has been limited (e.g., with
654          * the 'mem' command line parameter), don't add pages beyond
655          * this limit.
656          */
657         extra_pfn_end = min(max_pfn, start_pfn + pages);
658
659         for (pfn = start_pfn; pfn < extra_pfn_end; pfn++) {
660                 page = pfn_to_page(pfn);
661                 /* totalram_pages and totalhigh_pages do not
662                    include the boot-time balloon extension, so
663                    don't subtract from it. */
664                 __balloon_append(page);
665         }
666
667         balloon_stats.total_pages += extra_pfn_end - start_pfn;
668 }
669
670 static int __init balloon_init(void)
671 {
672         int i;
673
674         if (!xen_domain())
675                 return -ENODEV;
676
677         pr_info("Initialising balloon driver\n");
678
679         balloon_stats.current_pages = xen_pv_domain()
680                 ? min(xen_start_info->nr_pages - xen_released_pages, max_pfn)
681                 : get_num_physpages();
682         balloon_stats.target_pages  = balloon_stats.current_pages;
683         balloon_stats.balloon_low   = 0;
684         balloon_stats.balloon_high  = 0;
685         balloon_stats.total_pages   = balloon_stats.current_pages;
686
687         balloon_stats.schedule_delay = 1;
688         balloon_stats.max_schedule_delay = 32;
689         balloon_stats.retry_count = 1;
690         balloon_stats.max_retry_count = RETRY_UNLIMITED;
691
692 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
693         set_online_page_callback(&xen_online_page);
694         register_memory_notifier(&xen_memory_nb);
695         register_sysctl_table(xen_root);
696 #endif
697
698         /*
699          * Initialize the balloon with pages from the extra memory
700          * regions (see arch/x86/xen/setup.c).
701          */
702         for (i = 0; i < XEN_EXTRA_MEM_MAX_REGIONS; i++)
703                 if (xen_extra_mem[i].n_pfns)
704                         balloon_add_region(xen_extra_mem[i].start_pfn,
705                                            xen_extra_mem[i].n_pfns);
706
707         return 0;
708 }
709
710 subsys_initcall(balloon_init);
711
712 MODULE_LICENSE("GPL");