Loading...
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/cred.h>
45#include <linux/errno.h>
46#include <linux/mm.h>
47#include <linux/memblock.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/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#include <xen/mem-reservation.h>
73
74static int xen_hotplug_unpopulated;
75
76#ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
77
78static struct ctl_table balloon_table[] = {
79 {
80 .procname = "hotplug_unpopulated",
81 .data = &xen_hotplug_unpopulated,
82 .maxlen = sizeof(int),
83 .mode = 0644,
84 .proc_handler = proc_dointvec_minmax,
85 .extra1 = SYSCTL_ZERO,
86 .extra2 = SYSCTL_ONE,
87 },
88 { }
89};
90
91static struct ctl_table balloon_root[] = {
92 {
93 .procname = "balloon",
94 .mode = 0555,
95 .child = balloon_table,
96 },
97 { }
98};
99
100static struct ctl_table xen_root[] = {
101 {
102 .procname = "xen",
103 .mode = 0555,
104 .child = balloon_root,
105 },
106 { }
107};
108
109#endif
110
111/*
112 * Use one extent per PAGE_SIZE to avoid to break down the page into
113 * multiple frame.
114 */
115#define EXTENT_ORDER (fls(XEN_PFN_PER_PAGE) - 1)
116
117/*
118 * balloon_process() state:
119 *
120 * BP_DONE: done or nothing to do,
121 * BP_WAIT: wait to be rescheduled,
122 * BP_EAGAIN: error, go to sleep,
123 * BP_ECANCELED: error, balloon operation canceled.
124 */
125
126enum bp_state {
127 BP_DONE,
128 BP_WAIT,
129 BP_EAGAIN,
130 BP_ECANCELED
131};
132
133
134static DEFINE_MUTEX(balloon_mutex);
135
136struct balloon_stats balloon_stats;
137EXPORT_SYMBOL_GPL(balloon_stats);
138
139/* We increase/decrease in batches which fit in a page */
140static xen_pfn_t frame_list[PAGE_SIZE / sizeof(xen_pfn_t)];
141
142
143/* List of ballooned pages, threaded through the mem_map array. */
144static LIST_HEAD(ballooned_pages);
145static DECLARE_WAIT_QUEUE_HEAD(balloon_wq);
146
147/* Main work function, always executed in process context. */
148static void balloon_process(struct work_struct *work);
149static DECLARE_DELAYED_WORK(balloon_worker, balloon_process);
150
151/* When ballooning out (allocating memory to return to Xen) we don't really
152 want the kernel to try too hard since that can trigger the oom killer. */
153#define GFP_BALLOON \
154 (GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC)
155
156/* balloon_append: add the given page to the balloon. */
157static void balloon_append(struct page *page)
158{
159 __SetPageOffline(page);
160
161 /* Lowmem is re-populated first, so highmem pages go at list tail. */
162 if (PageHighMem(page)) {
163 list_add_tail(&page->lru, &ballooned_pages);
164 balloon_stats.balloon_high++;
165 } else {
166 list_add(&page->lru, &ballooned_pages);
167 balloon_stats.balloon_low++;
168 }
169 wake_up(&balloon_wq);
170}
171
172/* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
173static struct page *balloon_retrieve(bool require_lowmem)
174{
175 struct page *page;
176
177 if (list_empty(&ballooned_pages))
178 return NULL;
179
180 page = list_entry(ballooned_pages.next, struct page, lru);
181 if (require_lowmem && PageHighMem(page))
182 return NULL;
183 list_del(&page->lru);
184
185 if (PageHighMem(page))
186 balloon_stats.balloon_high--;
187 else
188 balloon_stats.balloon_low--;
189
190 __ClearPageOffline(page);
191 return page;
192}
193
194static struct page *balloon_next_page(struct page *page)
195{
196 struct list_head *next = page->lru.next;
197 if (next == &ballooned_pages)
198 return NULL;
199 return list_entry(next, struct page, lru);
200}
201
202static enum bp_state update_schedule(enum bp_state state)
203{
204 if (state == BP_WAIT)
205 return BP_WAIT;
206
207 if (state == BP_ECANCELED)
208 return BP_ECANCELED;
209
210 if (state == BP_DONE) {
211 balloon_stats.schedule_delay = 1;
212 balloon_stats.retry_count = 1;
213 return BP_DONE;
214 }
215
216 ++balloon_stats.retry_count;
217
218 if (balloon_stats.max_retry_count != RETRY_UNLIMITED &&
219 balloon_stats.retry_count > balloon_stats.max_retry_count) {
220 balloon_stats.schedule_delay = 1;
221 balloon_stats.retry_count = 1;
222 return BP_ECANCELED;
223 }
224
225 balloon_stats.schedule_delay <<= 1;
226
227 if (balloon_stats.schedule_delay > balloon_stats.max_schedule_delay)
228 balloon_stats.schedule_delay = balloon_stats.max_schedule_delay;
229
230 return BP_EAGAIN;
231}
232
233#ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
234static void release_memory_resource(struct resource *resource)
235{
236 if (!resource)
237 return;
238
239 /*
240 * No need to reset region to identity mapped since we now
241 * know that no I/O can be in this region
242 */
243 release_resource(resource);
244 kfree(resource);
245}
246
247static struct resource *additional_memory_resource(phys_addr_t size)
248{
249 struct resource *res;
250 int ret;
251
252 res = kzalloc(sizeof(*res), GFP_KERNEL);
253 if (!res)
254 return NULL;
255
256 res->name = "System RAM";
257 res->flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
258
259 ret = allocate_resource(&iomem_resource, res,
260 size, 0, -1,
261 PAGES_PER_SECTION * PAGE_SIZE, NULL, NULL);
262 if (ret < 0) {
263 pr_err("Cannot allocate new System RAM resource\n");
264 kfree(res);
265 return NULL;
266 }
267
268 return res;
269}
270
271static enum bp_state reserve_additional_memory(void)
272{
273 long credit;
274 struct resource *resource;
275 int nid, rc;
276 unsigned long balloon_hotplug;
277
278 credit = balloon_stats.target_pages + balloon_stats.target_unpopulated
279 - balloon_stats.total_pages;
280
281 /*
282 * Already hotplugged enough pages? Wait for them to be
283 * onlined.
284 */
285 if (credit <= 0)
286 return BP_WAIT;
287
288 balloon_hotplug = round_up(credit, PAGES_PER_SECTION);
289
290 resource = additional_memory_resource(balloon_hotplug * PAGE_SIZE);
291 if (!resource)
292 goto err;
293
294 nid = memory_add_physaddr_to_nid(resource->start);
295
296#ifdef CONFIG_XEN_HAVE_PVMMU
297 /*
298 * We don't support PV MMU when Linux and Xen is using
299 * different page granularity.
300 */
301 BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
302
303 /*
304 * add_memory() will build page tables for the new memory so
305 * the p2m must contain invalid entries so the correct
306 * non-present PTEs will be written.
307 *
308 * If a failure occurs, the original (identity) p2m entries
309 * are not restored since this region is now known not to
310 * conflict with any devices.
311 */
312 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
313 unsigned long pfn, i;
314
315 pfn = PFN_DOWN(resource->start);
316 for (i = 0; i < balloon_hotplug; i++) {
317 if (!set_phys_to_machine(pfn + i, INVALID_P2M_ENTRY)) {
318 pr_warn("set_phys_to_machine() failed, no memory added\n");
319 goto err;
320 }
321 }
322 }
323#endif
324
325 /*
326 * add_memory_resource() will call online_pages() which in its turn
327 * will call xen_online_page() callback causing deadlock if we don't
328 * release balloon_mutex here. Unlocking here is safe because the
329 * callers drop the mutex before trying again.
330 */
331 mutex_unlock(&balloon_mutex);
332 /* add_memory_resource() requires the device_hotplug lock */
333 lock_device_hotplug();
334 rc = add_memory_resource(nid, resource);
335 unlock_device_hotplug();
336 mutex_lock(&balloon_mutex);
337
338 if (rc) {
339 pr_warn("Cannot add additional memory (%i)\n", rc);
340 goto err;
341 }
342
343 balloon_stats.total_pages += balloon_hotplug;
344
345 return BP_WAIT;
346 err:
347 release_memory_resource(resource);
348 return BP_ECANCELED;
349}
350
351static void xen_online_page(struct page *page, unsigned int order)
352{
353 unsigned long i, size = (1 << order);
354 unsigned long start_pfn = page_to_pfn(page);
355 struct page *p;
356
357 pr_debug("Online %lu pages starting at pfn 0x%lx\n", size, start_pfn);
358 mutex_lock(&balloon_mutex);
359 for (i = 0; i < size; i++) {
360 p = pfn_to_page(start_pfn + i);
361 balloon_append(p);
362 }
363 mutex_unlock(&balloon_mutex);
364}
365
366static int xen_memory_notifier(struct notifier_block *nb, unsigned long val, void *v)
367{
368 if (val == MEM_ONLINE)
369 schedule_delayed_work(&balloon_worker, 0);
370
371 return NOTIFY_OK;
372}
373
374static struct notifier_block xen_memory_nb = {
375 .notifier_call = xen_memory_notifier,
376 .priority = 0
377};
378#else
379static enum bp_state reserve_additional_memory(void)
380{
381 balloon_stats.target_pages = balloon_stats.current_pages +
382 balloon_stats.target_unpopulated;
383 return BP_ECANCELED;
384}
385#endif /* CONFIG_XEN_BALLOON_MEMORY_HOTPLUG */
386
387static long current_credit(void)
388{
389 return balloon_stats.target_pages - balloon_stats.current_pages;
390}
391
392static bool balloon_is_inflated(void)
393{
394 return balloon_stats.balloon_low || balloon_stats.balloon_high;
395}
396
397static enum bp_state increase_reservation(unsigned long nr_pages)
398{
399 int rc;
400 unsigned long i;
401 struct page *page;
402
403 if (nr_pages > ARRAY_SIZE(frame_list))
404 nr_pages = ARRAY_SIZE(frame_list);
405
406 page = list_first_entry_or_null(&ballooned_pages, struct page, lru);
407 for (i = 0; i < nr_pages; i++) {
408 if (!page) {
409 nr_pages = i;
410 break;
411 }
412
413 frame_list[i] = page_to_xen_pfn(page);
414 page = balloon_next_page(page);
415 }
416
417 rc = xenmem_reservation_increase(nr_pages, frame_list);
418 if (rc <= 0)
419 return BP_EAGAIN;
420
421 for (i = 0; i < rc; i++) {
422 page = balloon_retrieve(false);
423 BUG_ON(page == NULL);
424
425 xenmem_reservation_va_mapping_update(1, &page, &frame_list[i]);
426
427 /* Relinquish the page back to the allocator. */
428 free_reserved_page(page);
429 }
430
431 balloon_stats.current_pages += rc;
432
433 return BP_DONE;
434}
435
436static enum bp_state decrease_reservation(unsigned long nr_pages, gfp_t gfp)
437{
438 enum bp_state state = BP_DONE;
439 unsigned long i;
440 struct page *page, *tmp;
441 int ret;
442 LIST_HEAD(pages);
443
444 if (nr_pages > ARRAY_SIZE(frame_list))
445 nr_pages = ARRAY_SIZE(frame_list);
446
447 for (i = 0; i < nr_pages; i++) {
448 page = alloc_page(gfp);
449 if (page == NULL) {
450 nr_pages = i;
451 state = BP_EAGAIN;
452 break;
453 }
454 adjust_managed_page_count(page, -1);
455 xenmem_reservation_scrub_page(page);
456 list_add(&page->lru, &pages);
457 }
458
459 /*
460 * Ensure that ballooned highmem pages don't have kmaps.
461 *
462 * Do this before changing the p2m as kmap_flush_unused()
463 * reads PTEs to obtain pages (and hence needs the original
464 * p2m entry).
465 */
466 kmap_flush_unused();
467
468 /*
469 * Setup the frame, update direct mapping, invalidate P2M,
470 * and add to balloon.
471 */
472 i = 0;
473 list_for_each_entry_safe(page, tmp, &pages, lru) {
474 frame_list[i++] = xen_page_to_gfn(page);
475
476 xenmem_reservation_va_mapping_reset(1, &page);
477
478 list_del(&page->lru);
479
480 balloon_append(page);
481 }
482
483 flush_tlb_all();
484
485 ret = xenmem_reservation_decrease(nr_pages, frame_list);
486 BUG_ON(ret != nr_pages);
487
488 balloon_stats.current_pages -= nr_pages;
489
490 return state;
491}
492
493/*
494 * As this is a work item it is guaranteed to run as a single instance only.
495 * We may of course race updates of the target counts (which are protected
496 * by the balloon lock), or with changes to the Xen hard limit, but we will
497 * recover from these in time.
498 */
499static void balloon_process(struct work_struct *work)
500{
501 enum bp_state state = BP_DONE;
502 long credit;
503
504
505 do {
506 mutex_lock(&balloon_mutex);
507
508 credit = current_credit();
509
510 if (credit > 0) {
511 if (balloon_is_inflated())
512 state = increase_reservation(credit);
513 else
514 state = reserve_additional_memory();
515 }
516
517 if (credit < 0) {
518 long n_pages;
519
520 n_pages = min(-credit, si_mem_available());
521 state = decrease_reservation(n_pages, GFP_BALLOON);
522 if (state == BP_DONE && n_pages != -credit &&
523 n_pages < totalreserve_pages)
524 state = BP_EAGAIN;
525 }
526
527 state = update_schedule(state);
528
529 mutex_unlock(&balloon_mutex);
530
531 cond_resched();
532
533 } while (credit && state == BP_DONE);
534
535 /* Schedule more work if there is some still to be done. */
536 if (state == BP_EAGAIN)
537 schedule_delayed_work(&balloon_worker, balloon_stats.schedule_delay * HZ);
538}
539
540/* Resets the Xen limit, sets new target, and kicks off processing. */
541void balloon_set_new_target(unsigned long target)
542{
543 /* No need for lock. Not read-modify-write updates. */
544 balloon_stats.target_pages = target;
545 schedule_delayed_work(&balloon_worker, 0);
546}
547EXPORT_SYMBOL_GPL(balloon_set_new_target);
548
549static int add_ballooned_pages(int nr_pages)
550{
551 enum bp_state st;
552
553 if (xen_hotplug_unpopulated) {
554 st = reserve_additional_memory();
555 if (st != BP_ECANCELED) {
556 int rc;
557
558 mutex_unlock(&balloon_mutex);
559 rc = wait_event_interruptible(balloon_wq,
560 !list_empty(&ballooned_pages));
561 mutex_lock(&balloon_mutex);
562 return rc ? -ENOMEM : 0;
563 }
564 }
565
566 if (si_mem_available() < nr_pages)
567 return -ENOMEM;
568
569 st = decrease_reservation(nr_pages, GFP_USER);
570 if (st != BP_DONE)
571 return -ENOMEM;
572
573 return 0;
574}
575
576/**
577 * alloc_xenballooned_pages - get pages that have been ballooned out
578 * @nr_pages: Number of pages to get
579 * @pages: pages returned
580 * @return 0 on success, error otherwise
581 */
582int alloc_xenballooned_pages(int nr_pages, struct page **pages)
583{
584 int pgno = 0;
585 struct page *page;
586 int ret;
587
588 mutex_lock(&balloon_mutex);
589
590 balloon_stats.target_unpopulated += nr_pages;
591
592 while (pgno < nr_pages) {
593 page = balloon_retrieve(true);
594 if (page) {
595 pages[pgno++] = page;
596#ifdef CONFIG_XEN_HAVE_PVMMU
597 /*
598 * We don't support PV MMU when Linux and Xen is using
599 * different page granularity.
600 */
601 BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
602
603 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
604 ret = xen_alloc_p2m_entry(page_to_pfn(page));
605 if (ret < 0)
606 goto out_undo;
607 }
608#endif
609 } else {
610 ret = add_ballooned_pages(nr_pages - pgno);
611 if (ret < 0)
612 goto out_undo;
613 }
614 }
615 mutex_unlock(&balloon_mutex);
616 return 0;
617 out_undo:
618 mutex_unlock(&balloon_mutex);
619 free_xenballooned_pages(pgno, pages);
620 /*
621 * NB: free_xenballooned_pages will only subtract pgno pages, but since
622 * target_unpopulated is incremented with nr_pages at the start we need
623 * to remove the remaining ones also, or accounting will be screwed.
624 */
625 balloon_stats.target_unpopulated -= nr_pages - pgno;
626 return ret;
627}
628EXPORT_SYMBOL(alloc_xenballooned_pages);
629
630/**
631 * free_xenballooned_pages - return pages retrieved with get_ballooned_pages
632 * @nr_pages: Number of pages
633 * @pages: pages to return
634 */
635void free_xenballooned_pages(int nr_pages, struct page **pages)
636{
637 int i;
638
639 mutex_lock(&balloon_mutex);
640
641 for (i = 0; i < nr_pages; i++) {
642 if (pages[i])
643 balloon_append(pages[i]);
644 }
645
646 balloon_stats.target_unpopulated -= nr_pages;
647
648 /* The balloon may be too large now. Shrink it if needed. */
649 if (current_credit())
650 schedule_delayed_work(&balloon_worker, 0);
651
652 mutex_unlock(&balloon_mutex);
653}
654EXPORT_SYMBOL(free_xenballooned_pages);
655
656#if defined(CONFIG_XEN_PV) && !defined(CONFIG_XEN_UNPOPULATED_ALLOC)
657static void __init balloon_add_region(unsigned long start_pfn,
658 unsigned long pages)
659{
660 unsigned long pfn, extra_pfn_end;
661
662 /*
663 * If the amount of usable memory has been limited (e.g., with
664 * the 'mem' command line parameter), don't add pages beyond
665 * this limit.
666 */
667 extra_pfn_end = min(max_pfn, start_pfn + pages);
668
669 for (pfn = start_pfn; pfn < extra_pfn_end; pfn++) {
670 /* totalram_pages and totalhigh_pages do not
671 include the boot-time balloon extension, so
672 don't subtract from it. */
673 balloon_append(pfn_to_page(pfn));
674 }
675
676 balloon_stats.total_pages += extra_pfn_end - start_pfn;
677}
678#endif
679
680static int __init balloon_init(void)
681{
682 if (!xen_domain())
683 return -ENODEV;
684
685 pr_info("Initialising balloon driver\n");
686
687#ifdef CONFIG_XEN_PV
688 balloon_stats.current_pages = xen_pv_domain()
689 ? min(xen_start_info->nr_pages - xen_released_pages, max_pfn)
690 : get_num_physpages();
691#else
692 balloon_stats.current_pages = get_num_physpages();
693#endif
694 balloon_stats.target_pages = balloon_stats.current_pages;
695 balloon_stats.balloon_low = 0;
696 balloon_stats.balloon_high = 0;
697 balloon_stats.total_pages = balloon_stats.current_pages;
698
699 balloon_stats.schedule_delay = 1;
700 balloon_stats.max_schedule_delay = 32;
701 balloon_stats.retry_count = 1;
702 balloon_stats.max_retry_count = 4;
703
704#ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
705 set_online_page_callback(&xen_online_page);
706 register_memory_notifier(&xen_memory_nb);
707 register_sysctl_table(xen_root);
708#endif
709
710#if defined(CONFIG_XEN_PV) && !defined(CONFIG_XEN_UNPOPULATED_ALLOC)
711 {
712 int i;
713
714 /*
715 * Initialize the balloon with pages from the extra memory
716 * regions (see arch/x86/xen/setup.c).
717 */
718 for (i = 0; i < XEN_EXTRA_MEM_MAX_REGIONS; i++)
719 if (xen_extra_mem[i].n_pfns)
720 balloon_add_region(xen_extra_mem[i].start_pfn,
721 xen_extra_mem[i].n_pfns);
722 }
723#endif
724
725 /* Init the xen-balloon driver. */
726 xen_balloon_init();
727
728 return 0;
729}
730subsys_initcall(balloon_init);
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
81enum bp_state {
82 BP_DONE,
83 BP_EAGAIN,
84 BP_ECANCELED
85};
86
87
88static DEFINE_MUTEX(balloon_mutex);
89
90struct balloon_stats balloon_stats;
91EXPORT_SYMBOL_GPL(balloon_stats);
92
93/* We increase/decrease in batches which fit in a page */
94static xen_pfn_t frame_list[PAGE_SIZE / sizeof(unsigned long)];
95static DEFINE_PER_CPU(struct page *, balloon_scratch_page);
96
97
98/* List of ballooned pages, threaded through the mem_map array. */
99static LIST_HEAD(ballooned_pages);
100
101/* Main work function, always executed in process context. */
102static void balloon_process(struct work_struct *work);
103static 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
110static 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. */
118static 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
130static 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. */
137static 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
160static struct page *balloon_next_page(struct page *page)
161{
162 struct list_head *next = page->lru.next;
163 if (next == &ballooned_pages)
164 return NULL;
165 return list_entry(next, struct page, lru);
166}
167
168static enum bp_state update_schedule(enum bp_state state)
169{
170 if (state == BP_DONE) {
171 balloon_stats.schedule_delay = 1;
172 balloon_stats.retry_count = 1;
173 return BP_DONE;
174 }
175
176 ++balloon_stats.retry_count;
177
178 if (balloon_stats.max_retry_count != RETRY_UNLIMITED &&
179 balloon_stats.retry_count > balloon_stats.max_retry_count) {
180 balloon_stats.schedule_delay = 1;
181 balloon_stats.retry_count = 1;
182 return BP_ECANCELED;
183 }
184
185 balloon_stats.schedule_delay <<= 1;
186
187 if (balloon_stats.schedule_delay > balloon_stats.max_schedule_delay)
188 balloon_stats.schedule_delay = balloon_stats.max_schedule_delay;
189
190 return BP_EAGAIN;
191}
192
193#ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
194static long current_credit(void)
195{
196 return balloon_stats.target_pages - balloon_stats.current_pages -
197 balloon_stats.hotplug_pages;
198}
199
200static bool balloon_is_inflated(void)
201{
202 if (balloon_stats.balloon_low || balloon_stats.balloon_high ||
203 balloon_stats.balloon_hotplug)
204 return true;
205 else
206 return false;
207}
208
209/*
210 * reserve_additional_memory() adds memory region of size >= credit above
211 * max_pfn. New region is section aligned and size is modified to be multiple
212 * of section size. Those features allow optimal use of address space and
213 * establish proper alignment when this function is called first time after
214 * boot (last section not fully populated at boot time contains unused memory
215 * pages with PG_reserved bit not set; online_pages_range() does not allow page
216 * onlining in whole range if first onlined page does not have PG_reserved
217 * bit set). Real size of added memory is established at page onlining stage.
218 */
219
220static enum bp_state reserve_additional_memory(long credit)
221{
222 int nid, rc;
223 u64 hotplug_start_paddr;
224 unsigned long balloon_hotplug = credit;
225
226 hotplug_start_paddr = PFN_PHYS(SECTION_ALIGN_UP(max_pfn));
227 balloon_hotplug = round_up(balloon_hotplug, PAGES_PER_SECTION);
228 nid = memory_add_physaddr_to_nid(hotplug_start_paddr);
229
230 rc = add_memory(nid, hotplug_start_paddr, balloon_hotplug << PAGE_SHIFT);
231
232 if (rc) {
233 pr_info("%s: add_memory() failed: %i\n", __func__, rc);
234 return BP_EAGAIN;
235 }
236
237 balloon_hotplug -= credit;
238
239 balloon_stats.hotplug_pages += credit;
240 balloon_stats.balloon_hotplug = balloon_hotplug;
241
242 return BP_DONE;
243}
244
245static void xen_online_page(struct page *page)
246{
247 __online_page_set_limits(page);
248
249 mutex_lock(&balloon_mutex);
250
251 __balloon_append(page);
252
253 if (balloon_stats.hotplug_pages)
254 --balloon_stats.hotplug_pages;
255 else
256 --balloon_stats.balloon_hotplug;
257
258 mutex_unlock(&balloon_mutex);
259}
260
261static int xen_memory_notifier(struct notifier_block *nb, unsigned long val, void *v)
262{
263 if (val == MEM_ONLINE)
264 schedule_delayed_work(&balloon_worker, 0);
265
266 return NOTIFY_OK;
267}
268
269static struct notifier_block xen_memory_nb = {
270 .notifier_call = xen_memory_notifier,
271 .priority = 0
272};
273#else
274static long current_credit(void)
275{
276 unsigned long target = balloon_stats.target_pages;
277
278 target = min(target,
279 balloon_stats.current_pages +
280 balloon_stats.balloon_low +
281 balloon_stats.balloon_high);
282
283 return target - balloon_stats.current_pages;
284}
285
286static bool balloon_is_inflated(void)
287{
288 if (balloon_stats.balloon_low || balloon_stats.balloon_high)
289 return true;
290 else
291 return false;
292}
293
294static enum bp_state reserve_additional_memory(long credit)
295{
296 balloon_stats.target_pages = balloon_stats.current_pages;
297 return BP_DONE;
298}
299#endif /* CONFIG_XEN_BALLOON_MEMORY_HOTPLUG */
300
301static enum bp_state increase_reservation(unsigned long nr_pages)
302{
303 int rc;
304 unsigned long pfn, i;
305 struct page *page;
306 struct xen_memory_reservation reservation = {
307 .address_bits = 0,
308 .extent_order = 0,
309 .domid = DOMID_SELF
310 };
311
312#ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
313 if (!balloon_stats.balloon_low && !balloon_stats.balloon_high) {
314 nr_pages = min(nr_pages, balloon_stats.balloon_hotplug);
315 balloon_stats.hotplug_pages += nr_pages;
316 balloon_stats.balloon_hotplug -= nr_pages;
317 return BP_DONE;
318 }
319#endif
320
321 if (nr_pages > ARRAY_SIZE(frame_list))
322 nr_pages = ARRAY_SIZE(frame_list);
323
324 page = list_first_entry_or_null(&ballooned_pages, struct page, lru);
325 for (i = 0; i < nr_pages; i++) {
326 if (!page) {
327 nr_pages = i;
328 break;
329 }
330 frame_list[i] = page_to_pfn(page);
331 page = balloon_next_page(page);
332 }
333
334 set_xen_guest_handle(reservation.extent_start, frame_list);
335 reservation.nr_extents = nr_pages;
336 rc = HYPERVISOR_memory_op(XENMEM_populate_physmap, &reservation);
337 if (rc <= 0)
338 return BP_EAGAIN;
339
340 for (i = 0; i < rc; i++) {
341 page = balloon_retrieve(false);
342 BUG_ON(page == NULL);
343
344 pfn = page_to_pfn(page);
345
346#ifdef CONFIG_XEN_HAVE_PVMMU
347 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
348 set_phys_to_machine(pfn, frame_list[i]);
349
350 /* Link back into the page tables if not highmem. */
351 if (!PageHighMem(page)) {
352 int ret;
353 ret = HYPERVISOR_update_va_mapping(
354 (unsigned long)__va(pfn << PAGE_SHIFT),
355 mfn_pte(frame_list[i], PAGE_KERNEL),
356 0);
357 BUG_ON(ret);
358 }
359 }
360#endif
361
362 /* Relinquish the page back to the allocator. */
363 __free_reserved_page(page);
364 }
365
366 balloon_stats.current_pages += rc;
367
368 return BP_DONE;
369}
370
371static enum bp_state decrease_reservation(unsigned long nr_pages, gfp_t gfp)
372{
373 enum bp_state state = BP_DONE;
374 unsigned long pfn, i;
375 struct page *page;
376 int ret;
377 struct xen_memory_reservation reservation = {
378 .address_bits = 0,
379 .extent_order = 0,
380 .domid = DOMID_SELF
381 };
382
383#ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
384 if (balloon_stats.hotplug_pages) {
385 nr_pages = min(nr_pages, balloon_stats.hotplug_pages);
386 balloon_stats.hotplug_pages -= nr_pages;
387 balloon_stats.balloon_hotplug += nr_pages;
388 return BP_DONE;
389 }
390#endif
391
392 if (nr_pages > ARRAY_SIZE(frame_list))
393 nr_pages = ARRAY_SIZE(frame_list);
394
395 for (i = 0; i < nr_pages; i++) {
396 page = alloc_page(gfp);
397 if (page == NULL) {
398 nr_pages = i;
399 state = BP_EAGAIN;
400 break;
401 }
402 scrub_page(page);
403
404 frame_list[i] = page_to_pfn(page);
405 }
406
407 /*
408 * Ensure that ballooned highmem pages don't have kmaps.
409 *
410 * Do this before changing the p2m as kmap_flush_unused()
411 * reads PTEs to obtain pages (and hence needs the original
412 * p2m entry).
413 */
414 kmap_flush_unused();
415
416 /* Update direct mapping, invalidate P2M, and add to balloon. */
417 for (i = 0; i < nr_pages; i++) {
418 pfn = frame_list[i];
419 frame_list[i] = pfn_to_mfn(pfn);
420 page = pfn_to_page(pfn);
421
422#ifdef CONFIG_XEN_HAVE_PVMMU
423 /*
424 * Ballooned out frames are effectively replaced with
425 * a scratch frame. Ensure direct mappings and the
426 * p2m are consistent.
427 */
428 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
429 unsigned long p;
430 struct page *scratch_page = get_balloon_scratch_page();
431
432 if (!PageHighMem(page)) {
433 ret = HYPERVISOR_update_va_mapping(
434 (unsigned long)__va(pfn << PAGE_SHIFT),
435 pfn_pte(page_to_pfn(scratch_page),
436 PAGE_KERNEL_RO), 0);
437 BUG_ON(ret);
438 }
439 p = page_to_pfn(scratch_page);
440 __set_phys_to_machine(pfn, pfn_to_mfn(p));
441
442 put_balloon_scratch_page();
443 }
444#endif
445
446 balloon_append(page);
447 }
448
449 flush_tlb_all();
450
451 set_xen_guest_handle(reservation.extent_start, frame_list);
452 reservation.nr_extents = nr_pages;
453 ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation, &reservation);
454 BUG_ON(ret != nr_pages);
455
456 balloon_stats.current_pages -= nr_pages;
457
458 return state;
459}
460
461/*
462 * We avoid multiple worker processes conflicting via the balloon mutex.
463 * We may of course race updates of the target counts (which are protected
464 * by the balloon lock), or with changes to the Xen hard limit, but we will
465 * recover from these in time.
466 */
467static void balloon_process(struct work_struct *work)
468{
469 enum bp_state state = BP_DONE;
470 long credit;
471
472 mutex_lock(&balloon_mutex);
473
474 do {
475 credit = current_credit();
476
477 if (credit > 0) {
478 if (balloon_is_inflated())
479 state = increase_reservation(credit);
480 else
481 state = reserve_additional_memory(credit);
482 }
483
484 if (credit < 0)
485 state = decrease_reservation(-credit, GFP_BALLOON);
486
487 state = update_schedule(state);
488
489#ifndef CONFIG_PREEMPT
490 if (need_resched())
491 schedule();
492#endif
493 } while (credit && state == BP_DONE);
494
495 /* Schedule more work if there is some still to be done. */
496 if (state == BP_EAGAIN)
497 schedule_delayed_work(&balloon_worker, balloon_stats.schedule_delay * HZ);
498
499 mutex_unlock(&balloon_mutex);
500}
501
502struct page *get_balloon_scratch_page(void)
503{
504 struct page *ret = get_cpu_var(balloon_scratch_page);
505 BUG_ON(ret == NULL);
506 return ret;
507}
508
509void put_balloon_scratch_page(void)
510{
511 put_cpu_var(balloon_scratch_page);
512}
513
514/* Resets the Xen limit, sets new target, and kicks off processing. */
515void balloon_set_new_target(unsigned long target)
516{
517 /* No need for lock. Not read-modify-write updates. */
518 balloon_stats.target_pages = target;
519 schedule_delayed_work(&balloon_worker, 0);
520}
521EXPORT_SYMBOL_GPL(balloon_set_new_target);
522
523/**
524 * alloc_xenballooned_pages - get pages that have been ballooned out
525 * @nr_pages: Number of pages to get
526 * @pages: pages returned
527 * @highmem: allow highmem pages
528 * @return 0 on success, error otherwise
529 */
530int alloc_xenballooned_pages(int nr_pages, struct page **pages, bool highmem)
531{
532 int pgno = 0;
533 struct page *page;
534 mutex_lock(&balloon_mutex);
535 while (pgno < nr_pages) {
536 page = balloon_retrieve(highmem);
537 if (page && (highmem || !PageHighMem(page))) {
538 pages[pgno++] = page;
539 } else {
540 enum bp_state st;
541 if (page)
542 balloon_append(page);
543 st = decrease_reservation(nr_pages - pgno,
544 highmem ? GFP_HIGHUSER : GFP_USER);
545 if (st != BP_DONE)
546 goto out_undo;
547 }
548 }
549 mutex_unlock(&balloon_mutex);
550 return 0;
551 out_undo:
552 while (pgno)
553 balloon_append(pages[--pgno]);
554 /* Free the memory back to the kernel soon */
555 schedule_delayed_work(&balloon_worker, 0);
556 mutex_unlock(&balloon_mutex);
557 return -ENOMEM;
558}
559EXPORT_SYMBOL(alloc_xenballooned_pages);
560
561/**
562 * free_xenballooned_pages - return pages retrieved with get_ballooned_pages
563 * @nr_pages: Number of pages
564 * @pages: pages to return
565 */
566void free_xenballooned_pages(int nr_pages, struct page **pages)
567{
568 int i;
569
570 mutex_lock(&balloon_mutex);
571
572 for (i = 0; i < nr_pages; i++) {
573 if (pages[i])
574 balloon_append(pages[i]);
575 }
576
577 /* The balloon may be too large now. Shrink it if needed. */
578 if (current_credit())
579 schedule_delayed_work(&balloon_worker, 0);
580
581 mutex_unlock(&balloon_mutex);
582}
583EXPORT_SYMBOL(free_xenballooned_pages);
584
585static void __init balloon_add_region(unsigned long start_pfn,
586 unsigned long pages)
587{
588 unsigned long pfn, extra_pfn_end;
589 struct page *page;
590
591 /*
592 * If the amount of usable memory has been limited (e.g., with
593 * the 'mem' command line parameter), don't add pages beyond
594 * this limit.
595 */
596 extra_pfn_end = min(max_pfn, start_pfn + pages);
597
598 for (pfn = start_pfn; pfn < extra_pfn_end; pfn++) {
599 page = pfn_to_page(pfn);
600 /* totalram_pages and totalhigh_pages do not
601 include the boot-time balloon extension, so
602 don't subtract from it. */
603 __balloon_append(page);
604 }
605}
606
607static int alloc_balloon_scratch_page(int cpu)
608{
609 if (per_cpu(balloon_scratch_page, cpu) != NULL)
610 return 0;
611
612 per_cpu(balloon_scratch_page, cpu) = alloc_page(GFP_KERNEL);
613 if (per_cpu(balloon_scratch_page, cpu) == NULL) {
614 pr_warn("Failed to allocate balloon_scratch_page for cpu %d\n", cpu);
615 return -ENOMEM;
616 }
617
618 return 0;
619}
620
621
622static int balloon_cpu_notify(struct notifier_block *self,
623 unsigned long action, void *hcpu)
624{
625 int cpu = (long)hcpu;
626 switch (action) {
627 case CPU_UP_PREPARE:
628 if (alloc_balloon_scratch_page(cpu))
629 return NOTIFY_BAD;
630 break;
631 default:
632 break;
633 }
634 return NOTIFY_OK;
635}
636
637static struct notifier_block balloon_cpu_notifier = {
638 .notifier_call = balloon_cpu_notify,
639};
640
641static int __init balloon_init(void)
642{
643 int i, cpu;
644
645 if (!xen_domain())
646 return -ENODEV;
647
648 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
649 register_cpu_notifier(&balloon_cpu_notifier);
650
651 get_online_cpus();
652 for_each_online_cpu(cpu) {
653 if (alloc_balloon_scratch_page(cpu)) {
654 put_online_cpus();
655 unregister_cpu_notifier(&balloon_cpu_notifier);
656 return -ENOMEM;
657 }
658 }
659 put_online_cpus();
660 }
661
662 pr_info("Initialising balloon driver\n");
663
664 balloon_stats.current_pages = xen_pv_domain()
665 ? min(xen_start_info->nr_pages - xen_released_pages, max_pfn)
666 : get_num_physpages();
667 balloon_stats.target_pages = balloon_stats.current_pages;
668 balloon_stats.balloon_low = 0;
669 balloon_stats.balloon_high = 0;
670
671 balloon_stats.schedule_delay = 1;
672 balloon_stats.max_schedule_delay = 32;
673 balloon_stats.retry_count = 1;
674 balloon_stats.max_retry_count = RETRY_UNLIMITED;
675
676#ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
677 balloon_stats.hotplug_pages = 0;
678 balloon_stats.balloon_hotplug = 0;
679
680 set_online_page_callback(&xen_online_page);
681 register_memory_notifier(&xen_memory_nb);
682#endif
683
684 /*
685 * Initialize the balloon with pages from the extra memory
686 * regions (see arch/x86/xen/setup.c).
687 */
688 for (i = 0; i < XEN_EXTRA_MEM_MAX_REGIONS; i++)
689 if (xen_extra_mem[i].size)
690 balloon_add_region(PFN_UP(xen_extra_mem[i].start),
691 PFN_DOWN(xen_extra_mem[i].size));
692
693 return 0;
694}
695
696subsys_initcall(balloon_init);
697
698static int __init balloon_clear(void)
699{
700 int cpu;
701
702 for_each_possible_cpu(cpu)
703 per_cpu(balloon_scratch_page, cpu) = NULL;
704
705 return 0;
706}
707early_initcall(balloon_clear);
708
709MODULE_LICENSE("GPL");