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/mm.h>
46#include <linux/bootmem.h>
47#include <linux/pagemap.h>
48#include <linux/highmem.h>
49#include <linux/mutex.h>
50#include <linux/list.h>
51#include <linux/gfp.h>
52#include <linux/notifier.h>
53#include <linux/memory.h>
54#include <linux/memory_hotplug.h>
55#include <linux/percpu-defs.h>
56#include <linux/slab.h>
57#include <linux/sysctl.h>
58
59#include <asm/page.h>
60#include <asm/pgalloc.h>
61#include <asm/pgtable.h>
62#include <asm/tlb.h>
63
64#include <asm/xen/hypervisor.h>
65#include <asm/xen/hypercall.h>
66
67#include <xen/xen.h>
68#include <xen/interface/xen.h>
69#include <xen/interface/memory.h>
70#include <xen/balloon.h>
71#include <xen/features.h>
72#include <xen/page.h>
73
74static int xen_hotplug_unpopulated;
75
76#ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
77
78static int zero;
79static int one = 1;
80
81static struct ctl_table balloon_table[] = {
82 {
83 .procname = "hotplug_unpopulated",
84 .data = &xen_hotplug_unpopulated,
85 .maxlen = sizeof(int),
86 .mode = 0644,
87 .proc_handler = proc_dointvec_minmax,
88 .extra1 = &zero,
89 .extra2 = &one,
90 },
91 { }
92};
93
94static struct ctl_table balloon_root[] = {
95 {
96 .procname = "balloon",
97 .mode = 0555,
98 .child = balloon_table,
99 },
100 { }
101};
102
103static struct ctl_table xen_root[] = {
104 {
105 .procname = "xen",
106 .mode = 0555,
107 .child = balloon_root,
108 },
109 { }
110};
111
112#endif
113
114/*
115 * Use one extent per PAGE_SIZE to avoid to break down the page into
116 * multiple frame.
117 */
118#define EXTENT_ORDER (fls(XEN_PFN_PER_PAGE) - 1)
119
120/*
121 * balloon_process() state:
122 *
123 * BP_DONE: done or nothing to do,
124 * BP_WAIT: wait to be rescheduled,
125 * BP_EAGAIN: error, go to sleep,
126 * BP_ECANCELED: error, balloon operation canceled.
127 */
128
129enum bp_state {
130 BP_DONE,
131 BP_WAIT,
132 BP_EAGAIN,
133 BP_ECANCELED
134};
135
136
137static DEFINE_MUTEX(balloon_mutex);
138
139struct balloon_stats balloon_stats;
140EXPORT_SYMBOL_GPL(balloon_stats);
141
142/* We increase/decrease in batches which fit in a page */
143static xen_pfn_t frame_list[PAGE_SIZE / sizeof(xen_pfn_t)];
144
145
146/* List of ballooned pages, threaded through the mem_map array. */
147static LIST_HEAD(ballooned_pages);
148static DECLARE_WAIT_QUEUE_HEAD(balloon_wq);
149
150/* Main work function, always executed in process context. */
151static void balloon_process(struct work_struct *work);
152static DECLARE_DELAYED_WORK(balloon_worker, balloon_process);
153
154/* When ballooning out (allocating memory to return to Xen) we don't really
155 want the kernel to try too hard since that can trigger the oom killer. */
156#define GFP_BALLOON \
157 (GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC)
158
159static void scrub_page(struct page *page)
160{
161#ifdef CONFIG_XEN_SCRUB_PAGES
162 clear_highpage(page);
163#endif
164}
165
166/* balloon_append: add the given page to the balloon. */
167static void __balloon_append(struct page *page)
168{
169 /* Lowmem is re-populated first, so highmem pages go at list tail. */
170 if (PageHighMem(page)) {
171 list_add_tail(&page->lru, &ballooned_pages);
172 balloon_stats.balloon_high++;
173 } else {
174 list_add(&page->lru, &ballooned_pages);
175 balloon_stats.balloon_low++;
176 }
177 wake_up(&balloon_wq);
178}
179
180static void balloon_append(struct page *page)
181{
182 __balloon_append(page);
183}
184
185/* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
186static struct page *balloon_retrieve(bool require_lowmem)
187{
188 struct page *page;
189
190 if (list_empty(&ballooned_pages))
191 return NULL;
192
193 page = list_entry(ballooned_pages.next, struct page, lru);
194 if (require_lowmem && PageHighMem(page))
195 return NULL;
196 list_del(&page->lru);
197
198 if (PageHighMem(page))
199 balloon_stats.balloon_high--;
200 else
201 balloon_stats.balloon_low--;
202
203 return page;
204}
205
206static struct page *balloon_next_page(struct page *page)
207{
208 struct list_head *next = page->lru.next;
209 if (next == &ballooned_pages)
210 return NULL;
211 return list_entry(next, struct page, lru);
212}
213
214static enum bp_state update_schedule(enum bp_state state)
215{
216 if (state == BP_WAIT)
217 return BP_WAIT;
218
219 if (state == BP_ECANCELED)
220 return BP_ECANCELED;
221
222 if (state == BP_DONE) {
223 balloon_stats.schedule_delay = 1;
224 balloon_stats.retry_count = 1;
225 return BP_DONE;
226 }
227
228 ++balloon_stats.retry_count;
229
230 if (balloon_stats.max_retry_count != RETRY_UNLIMITED &&
231 balloon_stats.retry_count > balloon_stats.max_retry_count) {
232 balloon_stats.schedule_delay = 1;
233 balloon_stats.retry_count = 1;
234 return BP_ECANCELED;
235 }
236
237 balloon_stats.schedule_delay <<= 1;
238
239 if (balloon_stats.schedule_delay > balloon_stats.max_schedule_delay)
240 balloon_stats.schedule_delay = balloon_stats.max_schedule_delay;
241
242 return BP_EAGAIN;
243}
244
245#ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
246static void release_memory_resource(struct resource *resource)
247{
248 if (!resource)
249 return;
250
251 /*
252 * No need to reset region to identity mapped since we now
253 * know that no I/O can be in this region
254 */
255 release_resource(resource);
256 kfree(resource);
257}
258
259static struct resource *additional_memory_resource(phys_addr_t size)
260{
261 struct resource *res;
262 int ret;
263
264 res = kzalloc(sizeof(*res), GFP_KERNEL);
265 if (!res)
266 return NULL;
267
268 res->name = "System RAM";
269 res->flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
270
271 ret = allocate_resource(&iomem_resource, res,
272 size, 0, -1,
273 PAGES_PER_SECTION * PAGE_SIZE, NULL, NULL);
274 if (ret < 0) {
275 pr_err("Cannot allocate new System RAM resource\n");
276 kfree(res);
277 return NULL;
278 }
279
280#ifdef CONFIG_SPARSEMEM
281 {
282 unsigned long limit = 1UL << (MAX_PHYSMEM_BITS - PAGE_SHIFT);
283 unsigned long pfn = res->start >> PAGE_SHIFT;
284
285 if (pfn > limit) {
286 pr_err("New System RAM resource outside addressable RAM (%lu > %lu)\n",
287 pfn, limit);
288 release_memory_resource(res);
289 return NULL;
290 }
291 }
292#endif
293
294 return res;
295}
296
297static enum bp_state reserve_additional_memory(void)
298{
299 long credit;
300 struct resource *resource;
301 int nid, rc;
302 unsigned long balloon_hotplug;
303
304 credit = balloon_stats.target_pages + balloon_stats.target_unpopulated
305 - balloon_stats.total_pages;
306
307 /*
308 * Already hotplugged enough pages? Wait for them to be
309 * onlined.
310 */
311 if (credit <= 0)
312 return BP_WAIT;
313
314 balloon_hotplug = round_up(credit, PAGES_PER_SECTION);
315
316 resource = additional_memory_resource(balloon_hotplug * PAGE_SIZE);
317 if (!resource)
318 goto err;
319
320 nid = memory_add_physaddr_to_nid(resource->start);
321
322#ifdef CONFIG_XEN_HAVE_PVMMU
323 /*
324 * We don't support PV MMU when Linux and Xen is using
325 * different page granularity.
326 */
327 BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
328
329 /*
330 * add_memory() will build page tables for the new memory so
331 * the p2m must contain invalid entries so the correct
332 * non-present PTEs will be written.
333 *
334 * If a failure occurs, the original (identity) p2m entries
335 * are not restored since this region is now known not to
336 * conflict with any devices.
337 */
338 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
339 unsigned long pfn, i;
340
341 pfn = PFN_DOWN(resource->start);
342 for (i = 0; i < balloon_hotplug; i++) {
343 if (!set_phys_to_machine(pfn + i, INVALID_P2M_ENTRY)) {
344 pr_warn("set_phys_to_machine() failed, no memory added\n");
345 goto err;
346 }
347 }
348 }
349#endif
350
351 /*
352 * add_memory_resource() will call online_pages() which in its turn
353 * will call xen_online_page() callback causing deadlock if we don't
354 * release balloon_mutex here. Unlocking here is safe because the
355 * callers drop the mutex before trying again.
356 */
357 mutex_unlock(&balloon_mutex);
358 rc = add_memory_resource(nid, resource, memhp_auto_online);
359 mutex_lock(&balloon_mutex);
360
361 if (rc) {
362 pr_warn("Cannot add additional memory (%i)\n", rc);
363 goto err;
364 }
365
366 balloon_stats.total_pages += balloon_hotplug;
367
368 return BP_WAIT;
369 err:
370 release_memory_resource(resource);
371 return BP_ECANCELED;
372}
373
374static void xen_online_page(struct page *page)
375{
376 __online_page_set_limits(page);
377
378 mutex_lock(&balloon_mutex);
379
380 __balloon_append(page);
381
382 mutex_unlock(&balloon_mutex);
383}
384
385static int xen_memory_notifier(struct notifier_block *nb, unsigned long val, void *v)
386{
387 if (val == MEM_ONLINE)
388 schedule_delayed_work(&balloon_worker, 0);
389
390 return NOTIFY_OK;
391}
392
393static struct notifier_block xen_memory_nb = {
394 .notifier_call = xen_memory_notifier,
395 .priority = 0
396};
397#else
398static enum bp_state reserve_additional_memory(void)
399{
400 balloon_stats.target_pages = balloon_stats.current_pages;
401 return BP_ECANCELED;
402}
403#endif /* CONFIG_XEN_BALLOON_MEMORY_HOTPLUG */
404
405static long current_credit(void)
406{
407 return balloon_stats.target_pages - balloon_stats.current_pages;
408}
409
410static bool balloon_is_inflated(void)
411{
412 return balloon_stats.balloon_low || balloon_stats.balloon_high;
413}
414
415static enum bp_state increase_reservation(unsigned long nr_pages)
416{
417 int rc;
418 unsigned long i;
419 struct page *page;
420 struct xen_memory_reservation reservation = {
421 .address_bits = 0,
422 .extent_order = EXTENT_ORDER,
423 .domid = DOMID_SELF
424 };
425
426 if (nr_pages > ARRAY_SIZE(frame_list))
427 nr_pages = ARRAY_SIZE(frame_list);
428
429 page = list_first_entry_or_null(&ballooned_pages, struct page, lru);
430 for (i = 0; i < nr_pages; i++) {
431 if (!page) {
432 nr_pages = i;
433 break;
434 }
435
436 /* XENMEM_populate_physmap requires a PFN based on Xen
437 * granularity.
438 */
439 frame_list[i] = page_to_xen_pfn(page);
440 page = balloon_next_page(page);
441 }
442
443 set_xen_guest_handle(reservation.extent_start, frame_list);
444 reservation.nr_extents = nr_pages;
445 rc = HYPERVISOR_memory_op(XENMEM_populate_physmap, &reservation);
446 if (rc <= 0)
447 return BP_EAGAIN;
448
449 for (i = 0; i < rc; i++) {
450 page = balloon_retrieve(false);
451 BUG_ON(page == NULL);
452
453#ifdef CONFIG_XEN_HAVE_PVMMU
454 /*
455 * We don't support PV MMU when Linux and Xen is using
456 * different page granularity.
457 */
458 BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
459
460 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
461 unsigned long pfn = page_to_pfn(page);
462
463 set_phys_to_machine(pfn, frame_list[i]);
464
465 /* Link back into the page tables if not highmem. */
466 if (!PageHighMem(page)) {
467 int ret;
468 ret = HYPERVISOR_update_va_mapping(
469 (unsigned long)__va(pfn << PAGE_SHIFT),
470 mfn_pte(frame_list[i], PAGE_KERNEL),
471 0);
472 BUG_ON(ret);
473 }
474 }
475#endif
476
477 /* Relinquish the page back to the allocator. */
478 free_reserved_page(page);
479 }
480
481 balloon_stats.current_pages += rc;
482
483 return BP_DONE;
484}
485
486static enum bp_state decrease_reservation(unsigned long nr_pages, gfp_t gfp)
487{
488 enum bp_state state = BP_DONE;
489 unsigned long i;
490 struct page *page, *tmp;
491 int ret;
492 struct xen_memory_reservation reservation = {
493 .address_bits = 0,
494 .extent_order = EXTENT_ORDER,
495 .domid = DOMID_SELF
496 };
497 LIST_HEAD(pages);
498
499 if (nr_pages > ARRAY_SIZE(frame_list))
500 nr_pages = ARRAY_SIZE(frame_list);
501
502 for (i = 0; i < nr_pages; i++) {
503 page = alloc_page(gfp);
504 if (page == NULL) {
505 nr_pages = i;
506 state = BP_EAGAIN;
507 break;
508 }
509 adjust_managed_page_count(page, -1);
510 scrub_page(page);
511 list_add(&page->lru, &pages);
512 }
513
514 /*
515 * Ensure that ballooned highmem pages don't have kmaps.
516 *
517 * Do this before changing the p2m as kmap_flush_unused()
518 * reads PTEs to obtain pages (and hence needs the original
519 * p2m entry).
520 */
521 kmap_flush_unused();
522
523 /*
524 * Setup the frame, update direct mapping, invalidate P2M,
525 * and add to balloon.
526 */
527 i = 0;
528 list_for_each_entry_safe(page, tmp, &pages, lru) {
529 /* XENMEM_decrease_reservation requires a GFN */
530 frame_list[i++] = xen_page_to_gfn(page);
531
532#ifdef CONFIG_XEN_HAVE_PVMMU
533 /*
534 * We don't support PV MMU when Linux and Xen is using
535 * different page granularity.
536 */
537 BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
538
539 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
540 unsigned long pfn = page_to_pfn(page);
541
542 if (!PageHighMem(page)) {
543 ret = HYPERVISOR_update_va_mapping(
544 (unsigned long)__va(pfn << PAGE_SHIFT),
545 __pte_ma(0), 0);
546 BUG_ON(ret);
547 }
548 __set_phys_to_machine(pfn, INVALID_P2M_ENTRY);
549 }
550#endif
551 list_del(&page->lru);
552
553 balloon_append(page);
554 }
555
556 flush_tlb_all();
557
558 set_xen_guest_handle(reservation.extent_start, frame_list);
559 reservation.nr_extents = nr_pages;
560 ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation, &reservation);
561 BUG_ON(ret != nr_pages);
562
563 balloon_stats.current_pages -= nr_pages;
564
565 return state;
566}
567
568/*
569 * As this is a work item it is guaranteed to run as a single instance only.
570 * We may of course race updates of the target counts (which are protected
571 * by the balloon lock), or with changes to the Xen hard limit, but we will
572 * recover from these in time.
573 */
574static void balloon_process(struct work_struct *work)
575{
576 enum bp_state state = BP_DONE;
577 long credit;
578
579
580 do {
581 mutex_lock(&balloon_mutex);
582
583 credit = current_credit();
584
585 if (credit > 0) {
586 if (balloon_is_inflated())
587 state = increase_reservation(credit);
588 else
589 state = reserve_additional_memory();
590 }
591
592 if (credit < 0)
593 state = decrease_reservation(-credit, GFP_BALLOON);
594
595 state = update_schedule(state);
596
597 mutex_unlock(&balloon_mutex);
598
599 cond_resched();
600
601 } while (credit && state == BP_DONE);
602
603 /* Schedule more work if there is some still to be done. */
604 if (state == BP_EAGAIN)
605 schedule_delayed_work(&balloon_worker, balloon_stats.schedule_delay * HZ);
606}
607
608/* Resets the Xen limit, sets new target, and kicks off processing. */
609void balloon_set_new_target(unsigned long target)
610{
611 /* No need for lock. Not read-modify-write updates. */
612 balloon_stats.target_pages = target;
613 schedule_delayed_work(&balloon_worker, 0);
614}
615EXPORT_SYMBOL_GPL(balloon_set_new_target);
616
617static int add_ballooned_pages(int nr_pages)
618{
619 enum bp_state st;
620
621 if (xen_hotplug_unpopulated) {
622 st = reserve_additional_memory();
623 if (st != BP_ECANCELED) {
624 mutex_unlock(&balloon_mutex);
625 wait_event(balloon_wq,
626 !list_empty(&ballooned_pages));
627 mutex_lock(&balloon_mutex);
628 return 0;
629 }
630 }
631
632 st = decrease_reservation(nr_pages, GFP_USER);
633 if (st != BP_DONE)
634 return -ENOMEM;
635
636 return 0;
637}
638
639/**
640 * alloc_xenballooned_pages - get pages that have been ballooned out
641 * @nr_pages: Number of pages to get
642 * @pages: pages returned
643 * @return 0 on success, error otherwise
644 */
645int alloc_xenballooned_pages(int nr_pages, struct page **pages)
646{
647 int pgno = 0;
648 struct page *page;
649 int ret;
650
651 mutex_lock(&balloon_mutex);
652
653 balloon_stats.target_unpopulated += nr_pages;
654
655 while (pgno < nr_pages) {
656 page = balloon_retrieve(true);
657 if (page) {
658 pages[pgno++] = page;
659#ifdef CONFIG_XEN_HAVE_PVMMU
660 /*
661 * We don't support PV MMU when Linux and Xen is using
662 * different page granularity.
663 */
664 BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
665
666 ret = xen_alloc_p2m_entry(page_to_pfn(page));
667 if (ret < 0)
668 goto out_undo;
669#endif
670 } else {
671 ret = add_ballooned_pages(nr_pages - pgno);
672 if (ret < 0)
673 goto out_undo;
674 }
675 }
676 mutex_unlock(&balloon_mutex);
677 return 0;
678 out_undo:
679 mutex_unlock(&balloon_mutex);
680 free_xenballooned_pages(pgno, pages);
681 return ret;
682}
683EXPORT_SYMBOL(alloc_xenballooned_pages);
684
685/**
686 * free_xenballooned_pages - return pages retrieved with get_ballooned_pages
687 * @nr_pages: Number of pages
688 * @pages: pages to return
689 */
690void free_xenballooned_pages(int nr_pages, struct page **pages)
691{
692 int i;
693
694 mutex_lock(&balloon_mutex);
695
696 for (i = 0; i < nr_pages; i++) {
697 if (pages[i])
698 balloon_append(pages[i]);
699 }
700
701 balloon_stats.target_unpopulated -= nr_pages;
702
703 /* The balloon may be too large now. Shrink it if needed. */
704 if (current_credit())
705 schedule_delayed_work(&balloon_worker, 0);
706
707 mutex_unlock(&balloon_mutex);
708}
709EXPORT_SYMBOL(free_xenballooned_pages);
710
711static void __init balloon_add_region(unsigned long start_pfn,
712 unsigned long pages)
713{
714 unsigned long pfn, extra_pfn_end;
715 struct page *page;
716
717 /*
718 * If the amount of usable memory has been limited (e.g., with
719 * the 'mem' command line parameter), don't add pages beyond
720 * this limit.
721 */
722 extra_pfn_end = min(max_pfn, start_pfn + pages);
723
724 for (pfn = start_pfn; pfn < extra_pfn_end; pfn++) {
725 page = pfn_to_page(pfn);
726 /* totalram_pages and totalhigh_pages do not
727 include the boot-time balloon extension, so
728 don't subtract from it. */
729 __balloon_append(page);
730 }
731
732 balloon_stats.total_pages += extra_pfn_end - start_pfn;
733}
734
735static int __init balloon_init(void)
736{
737 int i;
738
739 if (!xen_domain())
740 return -ENODEV;
741
742 pr_info("Initialising balloon driver\n");
743
744 balloon_stats.current_pages = xen_pv_domain()
745 ? min(xen_start_info->nr_pages - xen_released_pages, max_pfn)
746 : get_num_physpages();
747 balloon_stats.target_pages = balloon_stats.current_pages;
748 balloon_stats.balloon_low = 0;
749 balloon_stats.balloon_high = 0;
750 balloon_stats.total_pages = balloon_stats.current_pages;
751
752 balloon_stats.schedule_delay = 1;
753 balloon_stats.max_schedule_delay = 32;
754 balloon_stats.retry_count = 1;
755 balloon_stats.max_retry_count = RETRY_UNLIMITED;
756
757#ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
758 set_online_page_callback(&xen_online_page);
759 register_memory_notifier(&xen_memory_nb);
760 register_sysctl_table(xen_root);
761#endif
762
763 /*
764 * Initialize the balloon with pages from the extra memory
765 * regions (see arch/x86/xen/setup.c).
766 */
767 for (i = 0; i < XEN_EXTRA_MEM_MAX_REGIONS; i++)
768 if (xen_extra_mem[i].n_pfns)
769 balloon_add_region(xen_extra_mem[i].start_pfn,
770 xen_extra_mem[i].n_pfns);
771
772 return 0;
773}
774subsys_initcall(balloon_init);