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/cred.h>
45#include <linux/errno.h>
46#include <linux/freezer.h>
47#include <linux/kthread.h>
48#include <linux/mm.h>
49#include <linux/memblock.h>
50#include <linux/pagemap.h>
51#include <linux/highmem.h>
52#include <linux/mutex.h>
53#include <linux/list.h>
54#include <linux/gfp.h>
55#include <linux/notifier.h>
56#include <linux/memory.h>
57#include <linux/memory_hotplug.h>
58#include <linux/percpu-defs.h>
59#include <linux/slab.h>
60#include <linux/sysctl.h>
61#include <linux/moduleparam.h>
62#include <linux/jiffies.h>
63
64#include <asm/page.h>
65#include <asm/tlb.h>
66
67#include <asm/xen/hypervisor.h>
68#include <asm/xen/hypercall.h>
69
70#include <xen/xen.h>
71#include <xen/interface/xen.h>
72#include <xen/interface/memory.h>
73#include <xen/balloon.h>
74#include <xen/features.h>
75#include <xen/page.h>
76#include <xen/mem-reservation.h>
77
78#undef MODULE_PARAM_PREFIX
79#define MODULE_PARAM_PREFIX "xen."
80
81static uint __read_mostly balloon_boot_timeout = 180;
82module_param(balloon_boot_timeout, uint, 0444);
83
84#ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
85static int xen_hotplug_unpopulated;
86
87static struct ctl_table balloon_table[] = {
88 {
89 .procname = "hotplug_unpopulated",
90 .data = &xen_hotplug_unpopulated,
91 .maxlen = sizeof(int),
92 .mode = 0644,
93 .proc_handler = proc_dointvec_minmax,
94 .extra1 = SYSCTL_ZERO,
95 .extra2 = SYSCTL_ONE,
96 },
97 { }
98};
99
100static struct ctl_table balloon_root[] = {
101 {
102 .procname = "balloon",
103 .mode = 0555,
104 .child = balloon_table,
105 },
106 { }
107};
108
109static struct ctl_table xen_root[] = {
110 {
111 .procname = "xen",
112 .mode = 0555,
113 .child = balloon_root,
114 },
115 { }
116};
117
118#else
119#define xen_hotplug_unpopulated 0
120#endif
121
122/*
123 * Use one extent per PAGE_SIZE to avoid to break down the page into
124 * multiple frame.
125 */
126#define EXTENT_ORDER (fls(XEN_PFN_PER_PAGE) - 1)
127
128/*
129 * balloon_thread() state:
130 *
131 * BP_DONE: done or nothing to do,
132 * BP_WAIT: wait to be rescheduled,
133 * BP_EAGAIN: error, go to sleep,
134 * BP_ECANCELED: error, balloon operation canceled.
135 */
136
137static enum bp_state {
138 BP_DONE,
139 BP_WAIT,
140 BP_EAGAIN,
141 BP_ECANCELED
142} balloon_state = BP_DONE;
143
144/* Main waiting point for xen-balloon thread. */
145static DECLARE_WAIT_QUEUE_HEAD(balloon_thread_wq);
146
147static DEFINE_MUTEX(balloon_mutex);
148
149struct balloon_stats balloon_stats;
150EXPORT_SYMBOL_GPL(balloon_stats);
151
152/* We increase/decrease in batches which fit in a page */
153static xen_pfn_t frame_list[PAGE_SIZE / sizeof(xen_pfn_t)];
154
155
156/* List of ballooned pages, threaded through the mem_map array. */
157static LIST_HEAD(ballooned_pages);
158static DECLARE_WAIT_QUEUE_HEAD(balloon_wq);
159
160/* When ballooning out (allocating memory to return to Xen) we don't really
161 want the kernel to try too hard since that can trigger the oom killer. */
162#define GFP_BALLOON \
163 (GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC)
164
165/* balloon_append: add the given page to the balloon. */
166static void balloon_append(struct page *page)
167{
168 __SetPageOffline(page);
169
170 /* Lowmem is re-populated first, so highmem pages go at list tail. */
171 if (PageHighMem(page)) {
172 list_add_tail(&page->lru, &ballooned_pages);
173 balloon_stats.balloon_high++;
174 } else {
175 list_add(&page->lru, &ballooned_pages);
176 balloon_stats.balloon_low++;
177 }
178 wake_up(&balloon_wq);
179}
180
181/* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
182static 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 __ClearPageOffline(page);
200 return page;
201}
202
203static struct page *balloon_next_page(struct page *page)
204{
205 struct list_head *next = page->lru.next;
206 if (next == &ballooned_pages)
207 return NULL;
208 return list_entry(next, struct page, lru);
209}
210
211static void update_schedule(void)
212{
213 if (balloon_state == BP_WAIT || balloon_state == BP_ECANCELED)
214 return;
215
216 if (balloon_state == BP_DONE) {
217 balloon_stats.schedule_delay = 1;
218 balloon_stats.retry_count = 1;
219 return;
220 }
221
222 ++balloon_stats.retry_count;
223
224 if (balloon_stats.max_retry_count != RETRY_UNLIMITED &&
225 balloon_stats.retry_count > balloon_stats.max_retry_count) {
226 balloon_stats.schedule_delay = 1;
227 balloon_stats.retry_count = 1;
228 balloon_state = BP_ECANCELED;
229 return;
230 }
231
232 balloon_stats.schedule_delay <<= 1;
233
234 if (balloon_stats.schedule_delay > balloon_stats.max_schedule_delay)
235 balloon_stats.schedule_delay = balloon_stats.max_schedule_delay;
236
237 balloon_state = BP_EAGAIN;
238}
239
240#ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
241static void release_memory_resource(struct resource *resource)
242{
243 if (!resource)
244 return;
245
246 /*
247 * No need to reset region to identity mapped since we now
248 * know that no I/O can be in this region
249 */
250 release_resource(resource);
251 kfree(resource);
252}
253
254static struct resource *additional_memory_resource(phys_addr_t size)
255{
256 struct resource *res;
257 int ret;
258
259 res = kzalloc(sizeof(*res), GFP_KERNEL);
260 if (!res)
261 return NULL;
262
263 res->name = "System RAM";
264 res->flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
265
266 ret = allocate_resource(&iomem_resource, res,
267 size, 0, -1,
268 PAGES_PER_SECTION * PAGE_SIZE, NULL, NULL);
269 if (ret < 0) {
270 pr_err("Cannot allocate new System RAM resource\n");
271 kfree(res);
272 return NULL;
273 }
274
275 return res;
276}
277
278static enum bp_state reserve_additional_memory(void)
279{
280 long credit;
281 struct resource *resource;
282 int nid, rc;
283 unsigned long balloon_hotplug;
284
285 credit = balloon_stats.target_pages + balloon_stats.target_unpopulated
286 - balloon_stats.total_pages;
287
288 /*
289 * Already hotplugged enough pages? Wait for them to be
290 * onlined.
291 */
292 if (credit <= 0)
293 return BP_WAIT;
294
295 balloon_hotplug = round_up(credit, PAGES_PER_SECTION);
296
297 resource = additional_memory_resource(balloon_hotplug * PAGE_SIZE);
298 if (!resource)
299 goto err;
300
301 nid = memory_add_physaddr_to_nid(resource->start);
302
303#ifdef CONFIG_XEN_HAVE_PVMMU
304 /*
305 * We don't support PV MMU when Linux and Xen is using
306 * different page granularity.
307 */
308 BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
309
310 /*
311 * add_memory() will build page tables for the new memory so
312 * the p2m must contain invalid entries so the correct
313 * non-present PTEs will be written.
314 *
315 * If a failure occurs, the original (identity) p2m entries
316 * are not restored since this region is now known not to
317 * conflict with any devices.
318 */
319 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
320 unsigned long pfn, i;
321
322 pfn = PFN_DOWN(resource->start);
323 for (i = 0; i < balloon_hotplug; i++) {
324 if (!set_phys_to_machine(pfn + i, INVALID_P2M_ENTRY)) {
325 pr_warn("set_phys_to_machine() failed, no memory added\n");
326 goto err;
327 }
328 }
329 }
330#endif
331
332 /*
333 * add_memory_resource() will call online_pages() which in its turn
334 * will call xen_online_page() callback causing deadlock if we don't
335 * release balloon_mutex here. Unlocking here is safe because the
336 * callers drop the mutex before trying again.
337 */
338 mutex_unlock(&balloon_mutex);
339 /* add_memory_resource() requires the device_hotplug lock */
340 lock_device_hotplug();
341 rc = add_memory_resource(nid, resource, MHP_MERGE_RESOURCE);
342 unlock_device_hotplug();
343 mutex_lock(&balloon_mutex);
344
345 if (rc) {
346 pr_warn("Cannot add additional memory (%i)\n", rc);
347 goto err;
348 }
349
350 balloon_stats.total_pages += balloon_hotplug;
351
352 return BP_WAIT;
353 err:
354 release_memory_resource(resource);
355 return BP_ECANCELED;
356}
357
358static void xen_online_page(struct page *page, unsigned int order)
359{
360 unsigned long i, size = (1 << order);
361 unsigned long start_pfn = page_to_pfn(page);
362 struct page *p;
363
364 pr_debug("Online %lu pages starting at pfn 0x%lx\n", size, start_pfn);
365 mutex_lock(&balloon_mutex);
366 for (i = 0; i < size; i++) {
367 p = pfn_to_page(start_pfn + i);
368 balloon_append(p);
369 }
370 mutex_unlock(&balloon_mutex);
371}
372
373static int xen_memory_notifier(struct notifier_block *nb, unsigned long val, void *v)
374{
375 if (val == MEM_ONLINE)
376 wake_up(&balloon_thread_wq);
377
378 return NOTIFY_OK;
379}
380
381static struct notifier_block xen_memory_nb = {
382 .notifier_call = xen_memory_notifier,
383 .priority = 0
384};
385#else
386static enum bp_state reserve_additional_memory(void)
387{
388 balloon_stats.target_pages = balloon_stats.current_pages +
389 balloon_stats.target_unpopulated;
390 return BP_ECANCELED;
391}
392#endif /* CONFIG_XEN_BALLOON_MEMORY_HOTPLUG */
393
394static long current_credit(void)
395{
396 return balloon_stats.target_pages - balloon_stats.current_pages;
397}
398
399static bool balloon_is_inflated(void)
400{
401 return balloon_stats.balloon_low || balloon_stats.balloon_high;
402}
403
404static enum bp_state increase_reservation(unsigned long nr_pages)
405{
406 int rc;
407 unsigned long i;
408 struct page *page;
409
410 if (nr_pages > ARRAY_SIZE(frame_list))
411 nr_pages = ARRAY_SIZE(frame_list);
412
413 page = list_first_entry_or_null(&ballooned_pages, struct page, lru);
414 for (i = 0; i < nr_pages; i++) {
415 if (!page) {
416 nr_pages = i;
417 break;
418 }
419
420 frame_list[i] = page_to_xen_pfn(page);
421 page = balloon_next_page(page);
422 }
423
424 rc = xenmem_reservation_increase(nr_pages, frame_list);
425 if (rc <= 0)
426 return BP_EAGAIN;
427
428 for (i = 0; i < rc; i++) {
429 page = balloon_retrieve(false);
430 BUG_ON(page == NULL);
431
432 xenmem_reservation_va_mapping_update(1, &page, &frame_list[i]);
433
434 /* Relinquish the page back to the allocator. */
435 free_reserved_page(page);
436 }
437
438 balloon_stats.current_pages += rc;
439
440 return BP_DONE;
441}
442
443static enum bp_state decrease_reservation(unsigned long nr_pages, gfp_t gfp)
444{
445 enum bp_state state = BP_DONE;
446 unsigned long i;
447 struct page *page, *tmp;
448 int ret;
449 LIST_HEAD(pages);
450
451 if (nr_pages > ARRAY_SIZE(frame_list))
452 nr_pages = ARRAY_SIZE(frame_list);
453
454 for (i = 0; i < nr_pages; i++) {
455 page = alloc_page(gfp);
456 if (page == NULL) {
457 nr_pages = i;
458 state = BP_EAGAIN;
459 break;
460 }
461 adjust_managed_page_count(page, -1);
462 xenmem_reservation_scrub_page(page);
463 list_add(&page->lru, &pages);
464 }
465
466 /*
467 * Ensure that ballooned highmem pages don't have kmaps.
468 *
469 * Do this before changing the p2m as kmap_flush_unused()
470 * reads PTEs to obtain pages (and hence needs the original
471 * p2m entry).
472 */
473 kmap_flush_unused();
474
475 /*
476 * Setup the frame, update direct mapping, invalidate P2M,
477 * and add to balloon.
478 */
479 i = 0;
480 list_for_each_entry_safe(page, tmp, &pages, lru) {
481 frame_list[i++] = xen_page_to_gfn(page);
482
483 xenmem_reservation_va_mapping_reset(1, &page);
484
485 list_del(&page->lru);
486
487 balloon_append(page);
488 }
489
490 flush_tlb_all();
491
492 ret = xenmem_reservation_decrease(nr_pages, frame_list);
493 BUG_ON(ret != nr_pages);
494
495 balloon_stats.current_pages -= nr_pages;
496
497 return state;
498}
499
500/*
501 * Stop waiting if either state is BP_DONE and ballooning action is
502 * needed, or if the credit has changed while state is not BP_DONE.
503 */
504static bool balloon_thread_cond(long credit)
505{
506 if (balloon_state == BP_DONE)
507 credit = 0;
508
509 return current_credit() != credit || kthread_should_stop();
510}
511
512/*
513 * As this is a kthread it is guaranteed to run as a single instance only.
514 * We may of course race updates of the target counts (which are protected
515 * by the balloon lock), or with changes to the Xen hard limit, but we will
516 * recover from these in time.
517 */
518static int balloon_thread(void *unused)
519{
520 long credit;
521 unsigned long timeout;
522
523 set_freezable();
524 for (;;) {
525 switch (balloon_state) {
526 case BP_DONE:
527 case BP_ECANCELED:
528 timeout = 3600 * HZ;
529 break;
530 case BP_EAGAIN:
531 timeout = balloon_stats.schedule_delay * HZ;
532 break;
533 case BP_WAIT:
534 timeout = HZ;
535 break;
536 }
537
538 credit = current_credit();
539
540 wait_event_freezable_timeout(balloon_thread_wq,
541 balloon_thread_cond(credit), timeout);
542
543 if (kthread_should_stop())
544 return 0;
545
546 mutex_lock(&balloon_mutex);
547
548 credit = current_credit();
549
550 if (credit > 0) {
551 if (balloon_is_inflated())
552 balloon_state = increase_reservation(credit);
553 else
554 balloon_state = reserve_additional_memory();
555 }
556
557 if (credit < 0) {
558 long n_pages;
559
560 n_pages = min(-credit, si_mem_available());
561 balloon_state = decrease_reservation(n_pages,
562 GFP_BALLOON);
563 if (balloon_state == BP_DONE && n_pages != -credit &&
564 n_pages < totalreserve_pages)
565 balloon_state = BP_EAGAIN;
566 }
567
568 update_schedule();
569
570 mutex_unlock(&balloon_mutex);
571
572 cond_resched();
573 }
574}
575
576/* Resets the Xen limit, sets new target, and kicks off processing. */
577void balloon_set_new_target(unsigned long target)
578{
579 /* No need for lock. Not read-modify-write updates. */
580 balloon_stats.target_pages = target;
581 wake_up(&balloon_thread_wq);
582}
583EXPORT_SYMBOL_GPL(balloon_set_new_target);
584
585static int add_ballooned_pages(unsigned int nr_pages)
586{
587 enum bp_state st;
588
589 if (xen_hotplug_unpopulated) {
590 st = reserve_additional_memory();
591 if (st != BP_ECANCELED) {
592 int rc;
593
594 mutex_unlock(&balloon_mutex);
595 rc = wait_event_interruptible(balloon_wq,
596 !list_empty(&ballooned_pages));
597 mutex_lock(&balloon_mutex);
598 return rc ? -ENOMEM : 0;
599 }
600 }
601
602 if (si_mem_available() < nr_pages)
603 return -ENOMEM;
604
605 st = decrease_reservation(nr_pages, GFP_USER);
606 if (st != BP_DONE)
607 return -ENOMEM;
608
609 return 0;
610}
611
612/**
613 * xen_alloc_ballooned_pages - get pages that have been ballooned out
614 * @nr_pages: Number of pages to get
615 * @pages: pages returned
616 * @return 0 on success, error otherwise
617 */
618int xen_alloc_ballooned_pages(unsigned int nr_pages, struct page **pages)
619{
620 unsigned int pgno = 0;
621 struct page *page;
622 int ret;
623
624 mutex_lock(&balloon_mutex);
625
626 balloon_stats.target_unpopulated += nr_pages;
627
628 while (pgno < nr_pages) {
629 page = balloon_retrieve(true);
630 if (page) {
631 pages[pgno++] = page;
632#ifdef CONFIG_XEN_HAVE_PVMMU
633 /*
634 * We don't support PV MMU when Linux and Xen is using
635 * different page granularity.
636 */
637 BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
638
639 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
640 ret = xen_alloc_p2m_entry(page_to_pfn(page));
641 if (ret < 0)
642 goto out_undo;
643 }
644#endif
645 } else {
646 ret = add_ballooned_pages(nr_pages - pgno);
647 if (ret < 0)
648 goto out_undo;
649 }
650 }
651 mutex_unlock(&balloon_mutex);
652 return 0;
653 out_undo:
654 mutex_unlock(&balloon_mutex);
655 xen_free_ballooned_pages(pgno, pages);
656 /*
657 * NB: xen_free_ballooned_pages will only subtract pgno pages, but since
658 * target_unpopulated is incremented with nr_pages at the start we need
659 * to remove the remaining ones also, or accounting will be screwed.
660 */
661 balloon_stats.target_unpopulated -= nr_pages - pgno;
662 return ret;
663}
664EXPORT_SYMBOL(xen_alloc_ballooned_pages);
665
666/**
667 * xen_free_ballooned_pages - return pages retrieved with get_ballooned_pages
668 * @nr_pages: Number of pages
669 * @pages: pages to return
670 */
671void xen_free_ballooned_pages(unsigned int nr_pages, struct page **pages)
672{
673 unsigned int i;
674
675 mutex_lock(&balloon_mutex);
676
677 for (i = 0; i < nr_pages; i++) {
678 if (pages[i])
679 balloon_append(pages[i]);
680 }
681
682 balloon_stats.target_unpopulated -= nr_pages;
683
684 /* The balloon may be too large now. Shrink it if needed. */
685 if (current_credit())
686 wake_up(&balloon_thread_wq);
687
688 mutex_unlock(&balloon_mutex);
689}
690EXPORT_SYMBOL(xen_free_ballooned_pages);
691
692static void __init balloon_add_regions(void)
693{
694#if defined(CONFIG_XEN_PV)
695 unsigned long start_pfn, pages;
696 unsigned long pfn, extra_pfn_end;
697 unsigned int i;
698
699 for (i = 0; i < XEN_EXTRA_MEM_MAX_REGIONS; i++) {
700 pages = xen_extra_mem[i].n_pfns;
701 if (!pages)
702 continue;
703
704 start_pfn = xen_extra_mem[i].start_pfn;
705
706 /*
707 * If the amount of usable memory has been limited (e.g., with
708 * the 'mem' command line parameter), don't add pages beyond
709 * this limit.
710 */
711 extra_pfn_end = min(max_pfn, start_pfn + pages);
712
713 for (pfn = start_pfn; pfn < extra_pfn_end; pfn++)
714 balloon_append(pfn_to_page(pfn));
715
716 balloon_stats.total_pages += extra_pfn_end - start_pfn;
717 }
718#endif
719}
720
721static int __init balloon_init(void)
722{
723 struct task_struct *task;
724
725 if (!xen_domain())
726 return -ENODEV;
727
728 pr_info("Initialising balloon driver\n");
729
730#ifdef CONFIG_XEN_PV
731 balloon_stats.current_pages = xen_pv_domain()
732 ? min(xen_start_info->nr_pages - xen_released_pages, max_pfn)
733 : get_num_physpages();
734#else
735 balloon_stats.current_pages = get_num_physpages();
736#endif
737 balloon_stats.target_pages = balloon_stats.current_pages;
738 balloon_stats.balloon_low = 0;
739 balloon_stats.balloon_high = 0;
740 balloon_stats.total_pages = balloon_stats.current_pages;
741
742 balloon_stats.schedule_delay = 1;
743 balloon_stats.max_schedule_delay = 32;
744 balloon_stats.retry_count = 1;
745 balloon_stats.max_retry_count = 4;
746
747#ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
748 set_online_page_callback(&xen_online_page);
749 register_memory_notifier(&xen_memory_nb);
750 register_sysctl_table(xen_root);
751#endif
752
753 balloon_add_regions();
754
755 task = kthread_run(balloon_thread, NULL, "xen-balloon");
756 if (IS_ERR(task)) {
757 pr_err("xen-balloon thread could not be started, ballooning will not work!\n");
758 return PTR_ERR(task);
759 }
760
761 /* Init the xen-balloon driver. */
762 xen_balloon_init();
763
764 return 0;
765}
766subsys_initcall(balloon_init);
767
768static int __init balloon_wait_finish(void)
769{
770 long credit, last_credit = 0;
771 unsigned long last_changed = 0;
772
773 if (!xen_domain())
774 return -ENODEV;
775
776 /* PV guests don't need to wait. */
777 if (xen_pv_domain() || !current_credit())
778 return 0;
779
780 pr_notice("Waiting for initial ballooning down having finished.\n");
781
782 while ((credit = current_credit()) < 0) {
783 if (credit != last_credit) {
784 last_changed = jiffies;
785 last_credit = credit;
786 }
787 if (balloon_state == BP_ECANCELED) {
788 pr_warn_once("Initial ballooning failed, %ld pages need to be freed.\n",
789 -credit);
790 if (time_is_before_eq_jiffies(last_changed + HZ * balloon_boot_timeout))
791 panic("Initial ballooning failed!\n");
792 }
793
794 schedule_timeout_interruptible(HZ / 10);
795 }
796
797 pr_notice("Initial ballooning down finished.\n");
798
799 return 0;
800}
801late_initcall_sync(balloon_wait_finish);