Linux Audio

Check our new training course

Loading...
v4.6
  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
154static void release_memory_resource(struct resource *resource);
155
156/* When ballooning out (allocating memory to return to Xen) we don't really
157   want the kernel to try too hard since that can trigger the oom killer. */
158#define GFP_BALLOON \
159	(GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC)
160
161static void scrub_page(struct page *page)
162{
163#ifdef CONFIG_XEN_SCRUB_PAGES
164	clear_highpage(page);
165#endif
166}
167
168/* balloon_append: add the given page to the balloon. */
169static void __balloon_append(struct page *page)
170{
171	/* Lowmem is re-populated first, so highmem pages go at list tail. */
172	if (PageHighMem(page)) {
173		list_add_tail(&page->lru, &ballooned_pages);
174		balloon_stats.balloon_high++;
175	} else {
176		list_add(&page->lru, &ballooned_pages);
177		balloon_stats.balloon_low++;
178	}
179	wake_up(&balloon_wq);
180}
181
182static void balloon_append(struct page *page)
183{
184	__balloon_append(page);
185	adjust_managed_page_count(page, -1);
186}
187
188/* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
189static struct page *balloon_retrieve(bool require_lowmem)
190{
191	struct page *page;
192
193	if (list_empty(&ballooned_pages))
194		return NULL;
195
196	page = list_entry(ballooned_pages.next, struct page, lru);
197	if (require_lowmem && PageHighMem(page))
198		return NULL;
199	list_del(&page->lru);
200
201	if (PageHighMem(page))
202		balloon_stats.balloon_high--;
203	else
204		balloon_stats.balloon_low--;
205
206	adjust_managed_page_count(page, 1);
207
208	return page;
209}
210
211static struct page *balloon_next_page(struct page *page)
212{
213	struct list_head *next = page->lru.next;
214	if (next == &ballooned_pages)
215		return NULL;
216	return list_entry(next, struct page, lru);
217}
218
219static enum bp_state update_schedule(enum bp_state state)
220{
221	if (state == BP_WAIT)
222		return BP_WAIT;
223
224	if (state == BP_ECANCELED)
225		return BP_ECANCELED;
226
227	if (state == BP_DONE) {
228		balloon_stats.schedule_delay = 1;
229		balloon_stats.retry_count = 1;
230		return BP_DONE;
231	}
232
233	++balloon_stats.retry_count;
234
235	if (balloon_stats.max_retry_count != RETRY_UNLIMITED &&
236			balloon_stats.retry_count > balloon_stats.max_retry_count) {
237		balloon_stats.schedule_delay = 1;
238		balloon_stats.retry_count = 1;
239		return BP_ECANCELED;
240	}
241
242	balloon_stats.schedule_delay <<= 1;
243
244	if (balloon_stats.schedule_delay > balloon_stats.max_schedule_delay)
245		balloon_stats.schedule_delay = balloon_stats.max_schedule_delay;
246
247	return BP_EAGAIN;
248}
249
250#ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
 
 
 
 
 
 
 
 
 
 
 
 
 
251static struct resource *additional_memory_resource(phys_addr_t size)
252{
253	struct resource *res;
254	int ret;
255
256	res = kzalloc(sizeof(*res), GFP_KERNEL);
257	if (!res)
258		return NULL;
259
260	res->name = "System RAM";
261	res->flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
262
263	ret = allocate_resource(&iomem_resource, res,
264				size, 0, -1,
265				PAGES_PER_SECTION * PAGE_SIZE, NULL, NULL);
266	if (ret < 0) {
267		pr_err("Cannot allocate new System RAM resource\n");
268		kfree(res);
269		return NULL;
270	}
271
272#ifdef CONFIG_SPARSEMEM
273	{
274		unsigned long limit = 1UL << (MAX_PHYSMEM_BITS - PAGE_SHIFT);
275		unsigned long pfn = res->start >> PAGE_SHIFT;
276
277		if (pfn > limit) {
278			pr_err("New System RAM resource outside addressable RAM (%lu > %lu)\n",
279			       pfn, limit);
280			release_memory_resource(res);
281			return NULL;
282		}
283	}
284#endif
285
286	return res;
287}
288
289static void release_memory_resource(struct resource *resource)
290{
291	if (!resource)
292		return;
293
294	/*
295	 * No need to reset region to identity mapped since we now
296	 * know that no I/O can be in this region
297	 */
298	release_resource(resource);
299	kfree(resource);
300}
301
302static enum bp_state reserve_additional_memory(void)
303{
304	long credit;
305	struct resource *resource;
306	int nid, rc;
307	unsigned long balloon_hotplug;
308
309	credit = balloon_stats.target_pages + balloon_stats.target_unpopulated
310		- balloon_stats.total_pages;
311
312	/*
313	 * Already hotplugged enough pages?  Wait for them to be
314	 * onlined.
315	 */
316	if (credit <= 0)
317		return BP_WAIT;
318
319	balloon_hotplug = round_up(credit, PAGES_PER_SECTION);
320
321	resource = additional_memory_resource(balloon_hotplug * PAGE_SIZE);
322	if (!resource)
323		goto err;
324
325	nid = memory_add_physaddr_to_nid(resource->start);
326
327#ifdef CONFIG_XEN_HAVE_PVMMU
328	/*
329	 * We don't support PV MMU when Linux and Xen is using
330	 * different page granularity.
331	 */
332	BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
333
334        /*
335         * add_memory() will build page tables for the new memory so
336         * the p2m must contain invalid entries so the correct
337         * non-present PTEs will be written.
338         *
339         * If a failure occurs, the original (identity) p2m entries
340         * are not restored since this region is now known not to
341         * conflict with any devices.
342         */ 
343	if (!xen_feature(XENFEAT_auto_translated_physmap)) {
344		unsigned long pfn, i;
345
346		pfn = PFN_DOWN(resource->start);
347		for (i = 0; i < balloon_hotplug; i++) {
348			if (!set_phys_to_machine(pfn + i, INVALID_P2M_ENTRY)) {
349				pr_warn("set_phys_to_machine() failed, no memory added\n");
350				goto err;
351			}
352                }
353	}
354#endif
355
356	/*
357	 * add_memory_resource() will call online_pages() which in its turn
358	 * will call xen_online_page() callback causing deadlock if we don't
359	 * release balloon_mutex here. Unlocking here is safe because the
360	 * callers drop the mutex before trying again.
361	 */
362	mutex_unlock(&balloon_mutex);
363	rc = add_memory_resource(nid, resource, memhp_auto_online);
364	mutex_lock(&balloon_mutex);
365
366	if (rc) {
367		pr_warn("Cannot add additional memory (%i)\n", rc);
368		goto err;
369	}
370
371	balloon_stats.total_pages += balloon_hotplug;
372
373	return BP_WAIT;
374  err:
375	release_memory_resource(resource);
376	return BP_ECANCELED;
377}
378
379static void xen_online_page(struct page *page)
380{
381	__online_page_set_limits(page);
382
383	mutex_lock(&balloon_mutex);
384
385	__balloon_append(page);
386
387	mutex_unlock(&balloon_mutex);
388}
389
390static int xen_memory_notifier(struct notifier_block *nb, unsigned long val, void *v)
391{
392	if (val == MEM_ONLINE)
393		schedule_delayed_work(&balloon_worker, 0);
394
395	return NOTIFY_OK;
396}
397
398static struct notifier_block xen_memory_nb = {
399	.notifier_call = xen_memory_notifier,
400	.priority = 0
401};
402#else
403static enum bp_state reserve_additional_memory(void)
404{
405	balloon_stats.target_pages = balloon_stats.current_pages;
406	return BP_ECANCELED;
407}
408#endif /* CONFIG_XEN_BALLOON_MEMORY_HOTPLUG */
409
410static long current_credit(void)
411{
412	return balloon_stats.target_pages - balloon_stats.current_pages;
413}
414
415static bool balloon_is_inflated(void)
416{
417	return balloon_stats.balloon_low || balloon_stats.balloon_high;
418}
419
420static enum bp_state increase_reservation(unsigned long nr_pages)
421{
422	int rc;
423	unsigned long i;
424	struct page   *page;
425	struct xen_memory_reservation reservation = {
426		.address_bits = 0,
427		.extent_order = EXTENT_ORDER,
428		.domid        = DOMID_SELF
429	};
430
431	if (nr_pages > ARRAY_SIZE(frame_list))
432		nr_pages = ARRAY_SIZE(frame_list);
433
434	page = list_first_entry_or_null(&ballooned_pages, struct page, lru);
435	for (i = 0; i < nr_pages; i++) {
436		if (!page) {
437			nr_pages = i;
438			break;
439		}
440
441		/* XENMEM_populate_physmap requires a PFN based on Xen
442		 * granularity.
443		 */
444		frame_list[i] = page_to_xen_pfn(page);
445		page = balloon_next_page(page);
446	}
447
448	set_xen_guest_handle(reservation.extent_start, frame_list);
449	reservation.nr_extents = nr_pages;
450	rc = HYPERVISOR_memory_op(XENMEM_populate_physmap, &reservation);
451	if (rc <= 0)
452		return BP_EAGAIN;
453
454	for (i = 0; i < rc; i++) {
455		page = balloon_retrieve(false);
456		BUG_ON(page == NULL);
457
458#ifdef CONFIG_XEN_HAVE_PVMMU
459		/*
460		 * We don't support PV MMU when Linux and Xen is using
461		 * different page granularity.
462		 */
463		BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
464
465		if (!xen_feature(XENFEAT_auto_translated_physmap)) {
466			unsigned long pfn = page_to_pfn(page);
467
468			set_phys_to_machine(pfn, frame_list[i]);
469
470			/* Link back into the page tables if not highmem. */
471			if (!PageHighMem(page)) {
472				int ret;
473				ret = HYPERVISOR_update_va_mapping(
474						(unsigned long)__va(pfn << PAGE_SHIFT),
475						mfn_pte(frame_list[i], PAGE_KERNEL),
476						0);
477				BUG_ON(ret);
478			}
479		}
480#endif
481
482		/* Relinquish the page back to the allocator. */
483		__free_reserved_page(page);
484	}
485
486	balloon_stats.current_pages += rc;
487
488	return BP_DONE;
489}
490
491static enum bp_state decrease_reservation(unsigned long nr_pages, gfp_t gfp)
492{
493	enum bp_state state = BP_DONE;
494	unsigned long i;
495	struct page *page, *tmp;
496	int ret;
497	struct xen_memory_reservation reservation = {
498		.address_bits = 0,
499		.extent_order = EXTENT_ORDER,
500		.domid        = DOMID_SELF
501	};
502	LIST_HEAD(pages);
503
504	if (nr_pages > ARRAY_SIZE(frame_list))
505		nr_pages = ARRAY_SIZE(frame_list);
506
507	for (i = 0; i < nr_pages; i++) {
508		page = alloc_page(gfp);
509		if (page == NULL) {
510			nr_pages = i;
511			state = BP_EAGAIN;
512			break;
513		}
 
514		scrub_page(page);
515		list_add(&page->lru, &pages);
516	}
517
518	/*
519	 * Ensure that ballooned highmem pages don't have kmaps.
520	 *
521	 * Do this before changing the p2m as kmap_flush_unused()
522	 * reads PTEs to obtain pages (and hence needs the original
523	 * p2m entry).
524	 */
525	kmap_flush_unused();
526
527	/*
528	 * Setup the frame, update direct mapping, invalidate P2M,
529	 * and add to balloon.
530	 */
531	i = 0;
532	list_for_each_entry_safe(page, tmp, &pages, lru) {
533		/* XENMEM_decrease_reservation requires a GFN */
534		frame_list[i++] = xen_page_to_gfn(page);
535
536#ifdef CONFIG_XEN_HAVE_PVMMU
537		/*
538		 * We don't support PV MMU when Linux and Xen is using
539		 * different page granularity.
540		 */
541		BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
542
543		if (!xen_feature(XENFEAT_auto_translated_physmap)) {
544			unsigned long pfn = page_to_pfn(page);
545
546			if (!PageHighMem(page)) {
547				ret = HYPERVISOR_update_va_mapping(
548						(unsigned long)__va(pfn << PAGE_SHIFT),
549						__pte_ma(0), 0);
550				BUG_ON(ret);
551			}
552			__set_phys_to_machine(pfn, INVALID_P2M_ENTRY);
553		}
554#endif
555		list_del(&page->lru);
556
557		balloon_append(page);
558	}
559
560	flush_tlb_all();
561
562	set_xen_guest_handle(reservation.extent_start, frame_list);
563	reservation.nr_extents   = nr_pages;
564	ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation, &reservation);
565	BUG_ON(ret != nr_pages);
566
567	balloon_stats.current_pages -= nr_pages;
568
569	return state;
570}
571
572/*
573 * As this is a work item it is guaranteed to run as a single instance only.
574 * We may of course race updates of the target counts (which are protected
575 * by the balloon lock), or with changes to the Xen hard limit, but we will
576 * recover from these in time.
577 */
578static void balloon_process(struct work_struct *work)
579{
580	enum bp_state state = BP_DONE;
581	long credit;
582
583
584	do {
585		mutex_lock(&balloon_mutex);
586
587		credit = current_credit();
588
589		if (credit > 0) {
590			if (balloon_is_inflated())
591				state = increase_reservation(credit);
592			else
593				state = reserve_additional_memory();
594		}
595
596		if (credit < 0)
597			state = decrease_reservation(-credit, GFP_BALLOON);
598
599		state = update_schedule(state);
600
601		mutex_unlock(&balloon_mutex);
602
603		cond_resched();
604
605	} while (credit && state == BP_DONE);
606
607	/* Schedule more work if there is some still to be done. */
608	if (state == BP_EAGAIN)
609		schedule_delayed_work(&balloon_worker, balloon_stats.schedule_delay * HZ);
610}
611
612/* Resets the Xen limit, sets new target, and kicks off processing. */
613void balloon_set_new_target(unsigned long target)
614{
615	/* No need for lock. Not read-modify-write updates. */
616	balloon_stats.target_pages = target;
617	schedule_delayed_work(&balloon_worker, 0);
618}
619EXPORT_SYMBOL_GPL(balloon_set_new_target);
620
621static int add_ballooned_pages(int nr_pages)
622{
623	enum bp_state st;
624
625	if (xen_hotplug_unpopulated) {
626		st = reserve_additional_memory();
627		if (st != BP_ECANCELED) {
628			mutex_unlock(&balloon_mutex);
629			wait_event(balloon_wq,
630				   !list_empty(&ballooned_pages));
631			mutex_lock(&balloon_mutex);
632			return 0;
633		}
634	}
635
636	st = decrease_reservation(nr_pages, GFP_USER);
637	if (st != BP_DONE)
638		return -ENOMEM;
639
640	return 0;
641}
642
643/**
644 * alloc_xenballooned_pages - get pages that have been ballooned out
645 * @nr_pages: Number of pages to get
646 * @pages: pages returned
647 * @return 0 on success, error otherwise
648 */
649int alloc_xenballooned_pages(int nr_pages, struct page **pages)
650{
651	int pgno = 0;
652	struct page *page;
653	int ret;
654
655	mutex_lock(&balloon_mutex);
656
657	balloon_stats.target_unpopulated += nr_pages;
658
659	while (pgno < nr_pages) {
660		page = balloon_retrieve(true);
661		if (page) {
662			pages[pgno++] = page;
663#ifdef CONFIG_XEN_HAVE_PVMMU
664			/*
665			 * We don't support PV MMU when Linux and Xen is using
666			 * different page granularity.
667			 */
668			BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
669
670			ret = xen_alloc_p2m_entry(page_to_pfn(page));
671			if (ret < 0)
672				goto out_undo;
673#endif
674		} else {
675			ret = add_ballooned_pages(nr_pages - pgno);
676			if (ret < 0)
677				goto out_undo;
678		}
679	}
680	mutex_unlock(&balloon_mutex);
681	return 0;
682 out_undo:
683	mutex_unlock(&balloon_mutex);
684	free_xenballooned_pages(pgno, pages);
685	return ret;
686}
687EXPORT_SYMBOL(alloc_xenballooned_pages);
688
689/**
690 * free_xenballooned_pages - return pages retrieved with get_ballooned_pages
691 * @nr_pages: Number of pages
692 * @pages: pages to return
693 */
694void free_xenballooned_pages(int nr_pages, struct page **pages)
695{
696	int i;
697
698	mutex_lock(&balloon_mutex);
699
700	for (i = 0; i < nr_pages; i++) {
701		if (pages[i])
702			balloon_append(pages[i]);
703	}
704
705	balloon_stats.target_unpopulated -= nr_pages;
706
707	/* The balloon may be too large now. Shrink it if needed. */
708	if (current_credit())
709		schedule_delayed_work(&balloon_worker, 0);
710
711	mutex_unlock(&balloon_mutex);
712}
713EXPORT_SYMBOL(free_xenballooned_pages);
714
715static void __init balloon_add_region(unsigned long start_pfn,
716				      unsigned long pages)
717{
718	unsigned long pfn, extra_pfn_end;
719	struct page *page;
720
721	/*
722	 * If the amount of usable memory has been limited (e.g., with
723	 * the 'mem' command line parameter), don't add pages beyond
724	 * this limit.
725	 */
726	extra_pfn_end = min(max_pfn, start_pfn + pages);
727
728	for (pfn = start_pfn; pfn < extra_pfn_end; pfn++) {
729		page = pfn_to_page(pfn);
730		/* totalram_pages and totalhigh_pages do not
731		   include the boot-time balloon extension, so
732		   don't subtract from it. */
733		__balloon_append(page);
734	}
735
736	balloon_stats.total_pages += extra_pfn_end - start_pfn;
737}
738
739static int __init balloon_init(void)
740{
741	int i;
742
743	if (!xen_domain())
744		return -ENODEV;
745
746	pr_info("Initialising balloon driver\n");
747
748	balloon_stats.current_pages = xen_pv_domain()
749		? min(xen_start_info->nr_pages - xen_released_pages, max_pfn)
750		: get_num_physpages();
751	balloon_stats.target_pages  = balloon_stats.current_pages;
752	balloon_stats.balloon_low   = 0;
753	balloon_stats.balloon_high  = 0;
754	balloon_stats.total_pages   = balloon_stats.current_pages;
755
756	balloon_stats.schedule_delay = 1;
757	balloon_stats.max_schedule_delay = 32;
758	balloon_stats.retry_count = 1;
759	balloon_stats.max_retry_count = RETRY_UNLIMITED;
760
761#ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
762	set_online_page_callback(&xen_online_page);
763	register_memory_notifier(&xen_memory_nb);
764	register_sysctl_table(xen_root);
765#endif
766
767	/*
768	 * Initialize the balloon with pages from the extra memory
769	 * regions (see arch/x86/xen/setup.c).
770	 */
771	for (i = 0; i < XEN_EXTRA_MEM_MAX_REGIONS; i++)
772		if (xen_extra_mem[i].n_pfns)
773			balloon_add_region(xen_extra_mem[i].start_pfn,
774					   xen_extra_mem[i].n_pfns);
775
776	return 0;
777}
778subsys_initcall(balloon_init);
v4.10.11
  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);