xref: /openbmc/linux/drivers/xen/balloon.c (revision 01a6e126)
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/bootmem.h>
48 #include <linux/pagemap.h>
49 #include <linux/highmem.h>
50 #include <linux/mutex.h>
51 #include <linux/list.h>
52 #include <linux/gfp.h>
53 #include <linux/notifier.h>
54 #include <linux/memory.h>
55 #include <linux/memory_hotplug.h>
56 #include <linux/percpu-defs.h>
57 #include <linux/slab.h>
58 #include <linux/sysctl.h>
59 
60 #include <asm/page.h>
61 #include <asm/pgalloc.h>
62 #include <asm/pgtable.h>
63 #include <asm/tlb.h>
64 
65 #include <asm/xen/hypervisor.h>
66 #include <asm/xen/hypercall.h>
67 
68 #include <xen/xen.h>
69 #include <xen/interface/xen.h>
70 #include <xen/interface/memory.h>
71 #include <xen/balloon.h>
72 #include <xen/features.h>
73 #include <xen/page.h>
74 
75 static int xen_hotplug_unpopulated;
76 
77 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
78 
79 static int zero;
80 static int one = 1;
81 
82 static struct ctl_table balloon_table[] = {
83 	{
84 		.procname	= "hotplug_unpopulated",
85 		.data		= &xen_hotplug_unpopulated,
86 		.maxlen		= sizeof(int),
87 		.mode		= 0644,
88 		.proc_handler	= proc_dointvec_minmax,
89 		.extra1         = &zero,
90 		.extra2         = &one,
91 	},
92 	{ }
93 };
94 
95 static struct ctl_table balloon_root[] = {
96 	{
97 		.procname	= "balloon",
98 		.mode		= 0555,
99 		.child		= balloon_table,
100 	},
101 	{ }
102 };
103 
104 static struct ctl_table xen_root[] = {
105 	{
106 		.procname	= "xen",
107 		.mode		= 0555,
108 		.child		= balloon_root,
109 	},
110 	{ }
111 };
112 
113 #endif
114 
115 /*
116  * Use one extent per PAGE_SIZE to avoid to break down the page into
117  * multiple frame.
118  */
119 #define EXTENT_ORDER (fls(XEN_PFN_PER_PAGE) - 1)
120 
121 /*
122  * balloon_process() state:
123  *
124  * BP_DONE: done or nothing to do,
125  * BP_WAIT: wait to be rescheduled,
126  * BP_EAGAIN: error, go to sleep,
127  * BP_ECANCELED: error, balloon operation canceled.
128  */
129 
130 enum bp_state {
131 	BP_DONE,
132 	BP_WAIT,
133 	BP_EAGAIN,
134 	BP_ECANCELED
135 };
136 
137 
138 static DEFINE_MUTEX(balloon_mutex);
139 
140 struct balloon_stats balloon_stats;
141 EXPORT_SYMBOL_GPL(balloon_stats);
142 
143 /* We increase/decrease in batches which fit in a page */
144 static xen_pfn_t frame_list[PAGE_SIZE / sizeof(xen_pfn_t)];
145 
146 
147 /* List of ballooned pages, threaded through the mem_map array. */
148 static LIST_HEAD(ballooned_pages);
149 static DECLARE_WAIT_QUEUE_HEAD(balloon_wq);
150 
151 /* Main work function, always executed in process context. */
152 static void balloon_process(struct work_struct *work);
153 static DECLARE_DELAYED_WORK(balloon_worker, balloon_process);
154 
155 /* When ballooning out (allocating memory to return to Xen) we don't really
156    want the kernel to try too hard since that can trigger the oom killer. */
157 #define GFP_BALLOON \
158 	(GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC)
159 
160 static void scrub_page(struct page *page)
161 {
162 #ifdef CONFIG_XEN_SCRUB_PAGES
163 	clear_highpage(page);
164 #endif
165 }
166 
167 /* balloon_append: add the given page to the balloon. */
168 static void __balloon_append(struct page *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 static void balloon_append(struct page *page)
182 {
183 	__balloon_append(page);
184 }
185 
186 /* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
187 static struct page *balloon_retrieve(bool require_lowmem)
188 {
189 	struct page *page;
190 
191 	if (list_empty(&ballooned_pages))
192 		return NULL;
193 
194 	page = list_entry(ballooned_pages.next, struct page, lru);
195 	if (require_lowmem && PageHighMem(page))
196 		return NULL;
197 	list_del(&page->lru);
198 
199 	if (PageHighMem(page))
200 		balloon_stats.balloon_high--;
201 	else
202 		balloon_stats.balloon_low--;
203 
204 	return page;
205 }
206 
207 static struct page *balloon_next_page(struct page *page)
208 {
209 	struct list_head *next = page->lru.next;
210 	if (next == &ballooned_pages)
211 		return NULL;
212 	return list_entry(next, struct page, lru);
213 }
214 
215 static enum bp_state update_schedule(enum bp_state state)
216 {
217 	if (state == BP_WAIT)
218 		return BP_WAIT;
219 
220 	if (state == BP_ECANCELED)
221 		return BP_ECANCELED;
222 
223 	if (state == BP_DONE) {
224 		balloon_stats.schedule_delay = 1;
225 		balloon_stats.retry_count = 1;
226 		return BP_DONE;
227 	}
228 
229 	++balloon_stats.retry_count;
230 
231 	if (balloon_stats.max_retry_count != RETRY_UNLIMITED &&
232 			balloon_stats.retry_count > balloon_stats.max_retry_count) {
233 		balloon_stats.schedule_delay = 1;
234 		balloon_stats.retry_count = 1;
235 		return BP_ECANCELED;
236 	}
237 
238 	balloon_stats.schedule_delay <<= 1;
239 
240 	if (balloon_stats.schedule_delay > balloon_stats.max_schedule_delay)
241 		balloon_stats.schedule_delay = balloon_stats.max_schedule_delay;
242 
243 	return BP_EAGAIN;
244 }
245 
246 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
247 static void release_memory_resource(struct resource *resource)
248 {
249 	if (!resource)
250 		return;
251 
252 	/*
253 	 * No need to reset region to identity mapped since we now
254 	 * know that no I/O can be in this region
255 	 */
256 	release_resource(resource);
257 	kfree(resource);
258 }
259 
260 /*
261  * Host memory not allocated to dom0. We can use this range for hotplug-based
262  * ballooning.
263  *
264  * It's a type-less resource. Setting IORESOURCE_MEM will make resource
265  * management algorithms (arch_remove_reservations()) look into guest e820,
266  * which we don't want.
267  */
268 static struct resource hostmem_resource = {
269 	.name   = "Host RAM",
270 };
271 
272 void __attribute__((weak)) __init arch_xen_balloon_init(struct resource *res)
273 {}
274 
275 static struct resource *additional_memory_resource(phys_addr_t size)
276 {
277 	struct resource *res, *res_hostmem;
278 	int ret = -ENOMEM;
279 
280 	res = kzalloc(sizeof(*res), GFP_KERNEL);
281 	if (!res)
282 		return NULL;
283 
284 	res->name = "System RAM";
285 	res->flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
286 
287 	res_hostmem = kzalloc(sizeof(*res), GFP_KERNEL);
288 	if (res_hostmem) {
289 		/* Try to grab a range from hostmem */
290 		res_hostmem->name = "Host memory";
291 		ret = allocate_resource(&hostmem_resource, res_hostmem,
292 					size, 0, -1,
293 					PAGES_PER_SECTION * PAGE_SIZE, NULL, NULL);
294 	}
295 
296 	if (!ret) {
297 		/*
298 		 * Insert this resource into iomem. Because hostmem_resource
299 		 * tracks portion of guest e820 marked as UNUSABLE noone else
300 		 * should try to use it.
301 		 */
302 		res->start = res_hostmem->start;
303 		res->end = res_hostmem->end;
304 		ret = insert_resource(&iomem_resource, res);
305 		if (ret < 0) {
306 			pr_err("Can't insert iomem_resource [%llx - %llx]\n",
307 				res->start, res->end);
308 			release_memory_resource(res_hostmem);
309 			res_hostmem = NULL;
310 			res->start = res->end = 0;
311 		}
312 	}
313 
314 	if (ret) {
315 		ret = allocate_resource(&iomem_resource, res,
316 					size, 0, -1,
317 					PAGES_PER_SECTION * PAGE_SIZE, NULL, NULL);
318 		if (ret < 0) {
319 			pr_err("Cannot allocate new System RAM resource\n");
320 			kfree(res);
321 			return NULL;
322 		}
323 	}
324 
325 #ifdef CONFIG_SPARSEMEM
326 	{
327 		unsigned long limit = 1UL << (MAX_PHYSMEM_BITS - PAGE_SHIFT);
328 		unsigned long pfn = res->start >> PAGE_SHIFT;
329 
330 		if (pfn > limit) {
331 			pr_err("New System RAM resource outside addressable RAM (%lu > %lu)\n",
332 			       pfn, limit);
333 			release_memory_resource(res);
334 			release_memory_resource(res_hostmem);
335 			return NULL;
336 		}
337 	}
338 #endif
339 
340 	return res;
341 }
342 
343 static enum bp_state reserve_additional_memory(void)
344 {
345 	long credit;
346 	struct resource *resource;
347 	int nid, rc;
348 	unsigned long balloon_hotplug;
349 
350 	credit = balloon_stats.target_pages + balloon_stats.target_unpopulated
351 		- balloon_stats.total_pages;
352 
353 	/*
354 	 * Already hotplugged enough pages?  Wait for them to be
355 	 * onlined.
356 	 */
357 	if (credit <= 0)
358 		return BP_WAIT;
359 
360 	balloon_hotplug = round_up(credit, PAGES_PER_SECTION);
361 
362 	resource = additional_memory_resource(balloon_hotplug * PAGE_SIZE);
363 	if (!resource)
364 		goto err;
365 
366 	nid = memory_add_physaddr_to_nid(resource->start);
367 
368 #ifdef CONFIG_XEN_HAVE_PVMMU
369 	/*
370 	 * We don't support PV MMU when Linux and Xen is using
371 	 * different page granularity.
372 	 */
373 	BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
374 
375         /*
376          * add_memory() will build page tables for the new memory so
377          * the p2m must contain invalid entries so the correct
378          * non-present PTEs will be written.
379          *
380          * If a failure occurs, the original (identity) p2m entries
381          * are not restored since this region is now known not to
382          * conflict with any devices.
383          */
384 	if (!xen_feature(XENFEAT_auto_translated_physmap)) {
385 		unsigned long pfn, i;
386 
387 		pfn = PFN_DOWN(resource->start);
388 		for (i = 0; i < balloon_hotplug; i++) {
389 			if (!set_phys_to_machine(pfn + i, INVALID_P2M_ENTRY)) {
390 				pr_warn("set_phys_to_machine() failed, no memory added\n");
391 				goto err;
392 			}
393                 }
394 	}
395 #endif
396 
397 	/*
398 	 * add_memory_resource() will call online_pages() which in its turn
399 	 * will call xen_online_page() callback causing deadlock if we don't
400 	 * release balloon_mutex here. Unlocking here is safe because the
401 	 * callers drop the mutex before trying again.
402 	 */
403 	mutex_unlock(&balloon_mutex);
404 	rc = add_memory_resource(nid, resource, memhp_auto_online);
405 	mutex_lock(&balloon_mutex);
406 
407 	if (rc) {
408 		pr_warn("Cannot add additional memory (%i)\n", rc);
409 		goto err;
410 	}
411 
412 	balloon_stats.total_pages += balloon_hotplug;
413 
414 	return BP_WAIT;
415   err:
416 	release_memory_resource(resource);
417 	return BP_ECANCELED;
418 }
419 
420 static void xen_online_page(struct page *page)
421 {
422 	__online_page_set_limits(page);
423 
424 	mutex_lock(&balloon_mutex);
425 
426 	__balloon_append(page);
427 
428 	mutex_unlock(&balloon_mutex);
429 }
430 
431 static int xen_memory_notifier(struct notifier_block *nb, unsigned long val, void *v)
432 {
433 	if (val == MEM_ONLINE)
434 		schedule_delayed_work(&balloon_worker, 0);
435 
436 	return NOTIFY_OK;
437 }
438 
439 static struct notifier_block xen_memory_nb = {
440 	.notifier_call = xen_memory_notifier,
441 	.priority = 0
442 };
443 #else
444 static enum bp_state reserve_additional_memory(void)
445 {
446 	balloon_stats.target_pages = balloon_stats.current_pages;
447 	return BP_ECANCELED;
448 }
449 #endif /* CONFIG_XEN_BALLOON_MEMORY_HOTPLUG */
450 
451 static long current_credit(void)
452 {
453 	return balloon_stats.target_pages - balloon_stats.current_pages;
454 }
455 
456 static bool balloon_is_inflated(void)
457 {
458 	return balloon_stats.balloon_low || balloon_stats.balloon_high;
459 }
460 
461 static enum bp_state increase_reservation(unsigned long nr_pages)
462 {
463 	int rc;
464 	unsigned long i;
465 	struct page   *page;
466 	struct xen_memory_reservation reservation = {
467 		.address_bits = 0,
468 		.extent_order = EXTENT_ORDER,
469 		.domid        = DOMID_SELF
470 	};
471 
472 	if (nr_pages > ARRAY_SIZE(frame_list))
473 		nr_pages = ARRAY_SIZE(frame_list);
474 
475 	page = list_first_entry_or_null(&ballooned_pages, struct page, lru);
476 	for (i = 0; i < nr_pages; i++) {
477 		if (!page) {
478 			nr_pages = i;
479 			break;
480 		}
481 
482 		/* XENMEM_populate_physmap requires a PFN based on Xen
483 		 * granularity.
484 		 */
485 		frame_list[i] = page_to_xen_pfn(page);
486 		page = balloon_next_page(page);
487 	}
488 
489 	set_xen_guest_handle(reservation.extent_start, frame_list);
490 	reservation.nr_extents = nr_pages;
491 	rc = HYPERVISOR_memory_op(XENMEM_populate_physmap, &reservation);
492 	if (rc <= 0)
493 		return BP_EAGAIN;
494 
495 	for (i = 0; i < rc; i++) {
496 		page = balloon_retrieve(false);
497 		BUG_ON(page == NULL);
498 
499 #ifdef CONFIG_XEN_HAVE_PVMMU
500 		/*
501 		 * We don't support PV MMU when Linux and Xen is using
502 		 * different page granularity.
503 		 */
504 		BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
505 
506 		if (!xen_feature(XENFEAT_auto_translated_physmap)) {
507 			unsigned long pfn = page_to_pfn(page);
508 
509 			set_phys_to_machine(pfn, frame_list[i]);
510 
511 			/* Link back into the page tables if not highmem. */
512 			if (!PageHighMem(page)) {
513 				int ret;
514 				ret = HYPERVISOR_update_va_mapping(
515 						(unsigned long)__va(pfn << PAGE_SHIFT),
516 						mfn_pte(frame_list[i], PAGE_KERNEL),
517 						0);
518 				BUG_ON(ret);
519 			}
520 		}
521 #endif
522 
523 		/* Relinquish the page back to the allocator. */
524 		free_reserved_page(page);
525 	}
526 
527 	balloon_stats.current_pages += rc;
528 
529 	return BP_DONE;
530 }
531 
532 static enum bp_state decrease_reservation(unsigned long nr_pages, gfp_t gfp)
533 {
534 	enum bp_state state = BP_DONE;
535 	unsigned long i;
536 	struct page *page, *tmp;
537 	int ret;
538 	struct xen_memory_reservation reservation = {
539 		.address_bits = 0,
540 		.extent_order = EXTENT_ORDER,
541 		.domid        = DOMID_SELF
542 	};
543 	LIST_HEAD(pages);
544 
545 	if (nr_pages > ARRAY_SIZE(frame_list))
546 		nr_pages = ARRAY_SIZE(frame_list);
547 
548 	for (i = 0; i < nr_pages; i++) {
549 		page = alloc_page(gfp);
550 		if (page == NULL) {
551 			nr_pages = i;
552 			state = BP_EAGAIN;
553 			break;
554 		}
555 		adjust_managed_page_count(page, -1);
556 		scrub_page(page);
557 		list_add(&page->lru, &pages);
558 	}
559 
560 	/*
561 	 * Ensure that ballooned highmem pages don't have kmaps.
562 	 *
563 	 * Do this before changing the p2m as kmap_flush_unused()
564 	 * reads PTEs to obtain pages (and hence needs the original
565 	 * p2m entry).
566 	 */
567 	kmap_flush_unused();
568 
569 	/*
570 	 * Setup the frame, update direct mapping, invalidate P2M,
571 	 * and add to balloon.
572 	 */
573 	i = 0;
574 	list_for_each_entry_safe(page, tmp, &pages, lru) {
575 		/* XENMEM_decrease_reservation requires a GFN */
576 		frame_list[i++] = xen_page_to_gfn(page);
577 
578 #ifdef CONFIG_XEN_HAVE_PVMMU
579 		/*
580 		 * We don't support PV MMU when Linux and Xen is using
581 		 * different page granularity.
582 		 */
583 		BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
584 
585 		if (!xen_feature(XENFEAT_auto_translated_physmap)) {
586 			unsigned long pfn = page_to_pfn(page);
587 
588 			if (!PageHighMem(page)) {
589 				ret = HYPERVISOR_update_va_mapping(
590 						(unsigned long)__va(pfn << PAGE_SHIFT),
591 						__pte_ma(0), 0);
592 				BUG_ON(ret);
593 			}
594 			__set_phys_to_machine(pfn, INVALID_P2M_ENTRY);
595 		}
596 #endif
597 		list_del(&page->lru);
598 
599 		balloon_append(page);
600 	}
601 
602 	flush_tlb_all();
603 
604 	set_xen_guest_handle(reservation.extent_start, frame_list);
605 	reservation.nr_extents   = nr_pages;
606 	ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation, &reservation);
607 	BUG_ON(ret != nr_pages);
608 
609 	balloon_stats.current_pages -= nr_pages;
610 
611 	return state;
612 }
613 
614 /*
615  * As this is a work item it is guaranteed to run as a single instance only.
616  * We may of course race updates of the target counts (which are protected
617  * by the balloon lock), or with changes to the Xen hard limit, but we will
618  * recover from these in time.
619  */
620 static void balloon_process(struct work_struct *work)
621 {
622 	enum bp_state state = BP_DONE;
623 	long credit;
624 
625 
626 	do {
627 		mutex_lock(&balloon_mutex);
628 
629 		credit = current_credit();
630 
631 		if (credit > 0) {
632 			if (balloon_is_inflated())
633 				state = increase_reservation(credit);
634 			else
635 				state = reserve_additional_memory();
636 		}
637 
638 		if (credit < 0)
639 			state = decrease_reservation(-credit, GFP_BALLOON);
640 
641 		state = update_schedule(state);
642 
643 		mutex_unlock(&balloon_mutex);
644 
645 		cond_resched();
646 
647 	} while (credit && state == BP_DONE);
648 
649 	/* Schedule more work if there is some still to be done. */
650 	if (state == BP_EAGAIN)
651 		schedule_delayed_work(&balloon_worker, balloon_stats.schedule_delay * HZ);
652 }
653 
654 /* Resets the Xen limit, sets new target, and kicks off processing. */
655 void balloon_set_new_target(unsigned long target)
656 {
657 	/* No need for lock. Not read-modify-write updates. */
658 	balloon_stats.target_pages = target;
659 	schedule_delayed_work(&balloon_worker, 0);
660 }
661 EXPORT_SYMBOL_GPL(balloon_set_new_target);
662 
663 static int add_ballooned_pages(int nr_pages)
664 {
665 	enum bp_state st;
666 
667 	if (xen_hotplug_unpopulated) {
668 		st = reserve_additional_memory();
669 		if (st != BP_ECANCELED) {
670 			mutex_unlock(&balloon_mutex);
671 			wait_event(balloon_wq,
672 				   !list_empty(&ballooned_pages));
673 			mutex_lock(&balloon_mutex);
674 			return 0;
675 		}
676 	}
677 
678 	st = decrease_reservation(nr_pages, GFP_USER);
679 	if (st != BP_DONE)
680 		return -ENOMEM;
681 
682 	return 0;
683 }
684 
685 /**
686  * alloc_xenballooned_pages - get pages that have been ballooned out
687  * @nr_pages: Number of pages to get
688  * @pages: pages returned
689  * @return 0 on success, error otherwise
690  */
691 int alloc_xenballooned_pages(int nr_pages, struct page **pages)
692 {
693 	int pgno = 0;
694 	struct page *page;
695 	int ret;
696 
697 	mutex_lock(&balloon_mutex);
698 
699 	balloon_stats.target_unpopulated += nr_pages;
700 
701 	while (pgno < nr_pages) {
702 		page = balloon_retrieve(true);
703 		if (page) {
704 			pages[pgno++] = page;
705 #ifdef CONFIG_XEN_HAVE_PVMMU
706 			/*
707 			 * We don't support PV MMU when Linux and Xen is using
708 			 * different page granularity.
709 			 */
710 			BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
711 
712 			if (!xen_feature(XENFEAT_auto_translated_physmap)) {
713 				ret = xen_alloc_p2m_entry(page_to_pfn(page));
714 				if (ret < 0)
715 					goto out_undo;
716 			}
717 #endif
718 		} else {
719 			ret = add_ballooned_pages(nr_pages - pgno);
720 			if (ret < 0)
721 				goto out_undo;
722 		}
723 	}
724 	mutex_unlock(&balloon_mutex);
725 	return 0;
726  out_undo:
727 	mutex_unlock(&balloon_mutex);
728 	free_xenballooned_pages(pgno, pages);
729 	return ret;
730 }
731 EXPORT_SYMBOL(alloc_xenballooned_pages);
732 
733 /**
734  * free_xenballooned_pages - return pages retrieved with get_ballooned_pages
735  * @nr_pages: Number of pages
736  * @pages: pages to return
737  */
738 void free_xenballooned_pages(int nr_pages, struct page **pages)
739 {
740 	int i;
741 
742 	mutex_lock(&balloon_mutex);
743 
744 	for (i = 0; i < nr_pages; i++) {
745 		if (pages[i])
746 			balloon_append(pages[i]);
747 	}
748 
749 	balloon_stats.target_unpopulated -= nr_pages;
750 
751 	/* The balloon may be too large now. Shrink it if needed. */
752 	if (current_credit())
753 		schedule_delayed_work(&balloon_worker, 0);
754 
755 	mutex_unlock(&balloon_mutex);
756 }
757 EXPORT_SYMBOL(free_xenballooned_pages);
758 
759 #ifdef CONFIG_XEN_PV
760 static void __init balloon_add_region(unsigned long start_pfn,
761 				      unsigned long pages)
762 {
763 	unsigned long pfn, extra_pfn_end;
764 	struct page *page;
765 
766 	/*
767 	 * If the amount of usable memory has been limited (e.g., with
768 	 * the 'mem' command line parameter), don't add pages beyond
769 	 * this limit.
770 	 */
771 	extra_pfn_end = min(max_pfn, start_pfn + pages);
772 
773 	for (pfn = start_pfn; pfn < extra_pfn_end; pfn++) {
774 		page = pfn_to_page(pfn);
775 		/* totalram_pages and totalhigh_pages do not
776 		   include the boot-time balloon extension, so
777 		   don't subtract from it. */
778 		__balloon_append(page);
779 	}
780 
781 	balloon_stats.total_pages += extra_pfn_end - start_pfn;
782 }
783 #endif
784 
785 static int __init balloon_init(void)
786 {
787 	if (!xen_domain())
788 		return -ENODEV;
789 
790 	pr_info("Initialising balloon driver\n");
791 
792 #ifdef CONFIG_XEN_PV
793 	balloon_stats.current_pages = xen_pv_domain()
794 		? min(xen_start_info->nr_pages - xen_released_pages, max_pfn)
795 		: get_num_physpages();
796 #else
797 	balloon_stats.current_pages = get_num_physpages();
798 #endif
799 	balloon_stats.target_pages  = balloon_stats.current_pages;
800 	balloon_stats.balloon_low   = 0;
801 	balloon_stats.balloon_high  = 0;
802 	balloon_stats.total_pages   = balloon_stats.current_pages;
803 
804 	balloon_stats.schedule_delay = 1;
805 	balloon_stats.max_schedule_delay = 32;
806 	balloon_stats.retry_count = 1;
807 	balloon_stats.max_retry_count = RETRY_UNLIMITED;
808 
809 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
810 	set_online_page_callback(&xen_online_page);
811 	register_memory_notifier(&xen_memory_nb);
812 	register_sysctl_table(xen_root);
813 
814 	arch_xen_balloon_init(&hostmem_resource);
815 #endif
816 
817 #ifdef CONFIG_XEN_PV
818 	{
819 		int i;
820 
821 		/*
822 		 * Initialize the balloon with pages from the extra memory
823 		 * regions (see arch/x86/xen/setup.c).
824 		 */
825 		for (i = 0; i < XEN_EXTRA_MEM_MAX_REGIONS; i++)
826 			if (xen_extra_mem[i].n_pfns)
827 				balloon_add_region(xen_extra_mem[i].start_pfn,
828 						   xen_extra_mem[i].n_pfns);
829 	}
830 #endif
831 
832 	/* Init the xen-balloon driver. */
833 	xen_balloon_init();
834 
835 	return 0;
836 }
837 subsys_initcall(balloon_init);
838