Linux Audio

Check our new training course

Loading...
v6.13.7
  1/******************************************************************************
  2 * Xen balloon driver - enables returning/claiming memory to/from Xen.
  3 *
  4 * Copyright (c) 2003, B Dragovic
  5 * Copyright (c) 2003-2004, M Williamson, K Fraser
  6 * Copyright (c) 2005 Dan M. Smith, IBM Corporation
  7 * Copyright (c) 2010 Daniel Kiper
  8 *
  9 * Memory hotplug support was written by Daniel Kiper. Work on
 10 * it was sponsored by Google under Google Summer of Code 2010
 11 * program. Jeremy Fitzhardinge from Citrix was the mentor for
 12 * this project.
 13 *
 14 * This program is free software; you can redistribute it and/or
 15 * modify it under the terms of the GNU General Public License version 2
 16 * as published by the Free Software Foundation; or, when distributed
 17 * separately from the Linux kernel or incorporated into other
 18 * software packages, subject to the following license:
 19 *
 20 * Permission is hereby granted, free of charge, to any person obtaining a copy
 21 * of this source file (the "Software"), to deal in the Software without
 22 * restriction, including without limitation the rights to use, copy, modify,
 23 * merge, publish, distribute, sublicense, and/or sell copies of the Software,
 24 * and to permit persons to whom the Software is furnished to do so, subject to
 25 * the following conditions:
 26 *
 27 * The above copyright notice and this permission notice shall be included in
 28 * all copies or substantial portions of the Software.
 29 *
 30 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 31 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 32 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
 33 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 34 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
 35 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
 36 * IN THE SOFTWARE.
 37 */
 38
 39#define pr_fmt(fmt) "xen:" KBUILD_MODNAME ": " fmt
 40
 41#include <linux/cpu.h>
 42#include <linux/kernel.h>
 43#include <linux/sched.h>
 44#include <linux/cred.h>
 45#include <linux/errno.h>
 46#include <linux/freezer.h>
 47#include <linux/kthread.h>
 48#include <linux/mm.h>
 49#include <linux/memblock.h>
 50#include <linux/pagemap.h>
 51#include <linux/highmem.h>
 52#include <linux/mutex.h>
 53#include <linux/list.h>
 54#include <linux/gfp.h>
 55#include <linux/notifier.h>
 56#include <linux/memory.h>
 57#include <linux/memory_hotplug.h>
 58#include <linux/percpu-defs.h>
 59#include <linux/slab.h>
 60#include <linux/sysctl.h>
 61#include <linux/moduleparam.h>
 62#include <linux/jiffies.h>
 63
 64#include <asm/page.h>
 65#include <asm/tlb.h>
 66
 67#include <asm/xen/hypervisor.h>
 68#include <asm/xen/hypercall.h>
 69
 70#include <xen/xen.h>
 71#include <xen/interface/xen.h>
 72#include <xen/interface/memory.h>
 73#include <xen/balloon.h>
 74#include <xen/features.h>
 75#include <xen/page.h>
 76#include <xen/mem-reservation.h>
 77
 78#undef MODULE_PARAM_PREFIX
 79#define MODULE_PARAM_PREFIX "xen."
 80
 81static uint __read_mostly balloon_boot_timeout = 180;
 82module_param(balloon_boot_timeout, uint, 0444);
 83
 84#ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
 85static int xen_hotplug_unpopulated;
 86
 87static struct ctl_table balloon_table[] = {
 88	{
 89		.procname	= "hotplug_unpopulated",
 90		.data		= &xen_hotplug_unpopulated,
 91		.maxlen		= sizeof(int),
 92		.mode		= 0644,
 93		.proc_handler	= proc_dointvec_minmax,
 94		.extra1         = SYSCTL_ZERO,
 95		.extra2         = SYSCTL_ONE,
 96	},
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 97};
 98
 99#else
100#define xen_hotplug_unpopulated 0
101#endif
102
103/*
104 * Use one extent per PAGE_SIZE to avoid to break down the page into
105 * multiple frame.
106 */
107#define EXTENT_ORDER (fls(XEN_PFN_PER_PAGE) - 1)
108
109/*
110 * balloon_thread() state:
111 *
112 * BP_DONE: done or nothing to do,
113 * BP_WAIT: wait to be rescheduled,
114 * BP_EAGAIN: error, go to sleep,
115 * BP_ECANCELED: error, balloon operation canceled.
116 */
117
118static enum bp_state {
119	BP_DONE,
120	BP_WAIT,
121	BP_EAGAIN,
122	BP_ECANCELED
123} balloon_state = BP_DONE;
124
125/* Main waiting point for xen-balloon thread. */
126static DECLARE_WAIT_QUEUE_HEAD(balloon_thread_wq);
127
128static DEFINE_MUTEX(balloon_mutex);
129
130struct balloon_stats balloon_stats;
131EXPORT_SYMBOL_GPL(balloon_stats);
132
133/* We increase/decrease in batches which fit in a page */
134static xen_pfn_t frame_list[PAGE_SIZE / sizeof(xen_pfn_t)];
135
136
137/* List of ballooned pages, threaded through the mem_map array. */
138static LIST_HEAD(ballooned_pages);
139static DECLARE_WAIT_QUEUE_HEAD(balloon_wq);
140
141/* When ballooning out (allocating memory to return to Xen) we don't really
142   want the kernel to try too hard since that can trigger the oom killer. */
143#define GFP_BALLOON \
144	(GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC)
145
146/* balloon_append: add the given page to the balloon. */
147static void balloon_append(struct page *page)
148{
149	if (!PageOffline(page))
150		__SetPageOffline(page);
151
152	/* Lowmem is re-populated first, so highmem pages go at list tail. */
153	if (PageHighMem(page)) {
154		list_add_tail(&page->lru, &ballooned_pages);
155		balloon_stats.balloon_high++;
156	} else {
157		list_add(&page->lru, &ballooned_pages);
158		balloon_stats.balloon_low++;
159	}
160	wake_up(&balloon_wq);
161}
162
163/* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
164static struct page *balloon_retrieve(bool require_lowmem)
165{
166	struct page *page;
167
168	if (list_empty(&ballooned_pages))
169		return NULL;
170
171	page = list_entry(ballooned_pages.next, struct page, lru);
172	if (require_lowmem && PageHighMem(page))
173		return NULL;
174	list_del(&page->lru);
175
176	if (PageHighMem(page))
177		balloon_stats.balloon_high--;
178	else
179		balloon_stats.balloon_low--;
180
181	__ClearPageOffline(page);
182	return page;
183}
184
185static struct page *balloon_next_page(struct page *page)
186{
187	struct list_head *next = page->lru.next;
188	if (next == &ballooned_pages)
189		return NULL;
190	return list_entry(next, struct page, lru);
191}
192
193static void update_schedule(void)
194{
195	if (balloon_state == BP_WAIT || balloon_state == BP_ECANCELED)
196		return;
197
198	if (balloon_state == BP_DONE) {
199		balloon_stats.schedule_delay = 1;
200		balloon_stats.retry_count = 1;
201		return;
202	}
203
204	++balloon_stats.retry_count;
205
206	if (balloon_stats.max_retry_count != RETRY_UNLIMITED &&
207			balloon_stats.retry_count > balloon_stats.max_retry_count) {
208		balloon_stats.schedule_delay = 1;
209		balloon_stats.retry_count = 1;
210		balloon_state = BP_ECANCELED;
211		return;
212	}
213
214	balloon_stats.schedule_delay <<= 1;
215
216	if (balloon_stats.schedule_delay > balloon_stats.max_schedule_delay)
217		balloon_stats.schedule_delay = balloon_stats.max_schedule_delay;
218
219	balloon_state = BP_EAGAIN;
220}
221
222#ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
223static void release_memory_resource(struct resource *resource)
224{
225	if (!resource)
226		return;
227
228	/*
229	 * No need to reset region to identity mapped since we now
230	 * know that no I/O can be in this region
231	 */
232	release_resource(resource);
233	kfree(resource);
234}
235
236static struct resource *additional_memory_resource(phys_addr_t size)
237{
238	struct resource *res;
239	int ret;
240
241	res = kzalloc(sizeof(*res), GFP_KERNEL);
242	if (!res)
243		return NULL;
244
245	res->name = "System RAM";
246	res->flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
247
248	ret = allocate_resource(&iomem_resource, res,
249				size, 0, -1,
250				PAGES_PER_SECTION * PAGE_SIZE, NULL, NULL);
251	if (ret < 0) {
252		pr_err("Cannot allocate new System RAM resource\n");
253		kfree(res);
254		return NULL;
255	}
256
257	return res;
258}
259
260static enum bp_state reserve_additional_memory(void)
261{
262	long credit;
263	struct resource *resource;
264	int nid, rc;
265	unsigned long balloon_hotplug;
266
267	credit = balloon_stats.target_pages + balloon_stats.target_unpopulated
268		- balloon_stats.total_pages;
269
270	/*
271	 * Already hotplugged enough pages?  Wait for them to be
272	 * onlined.
273	 */
274	if (credit <= 0)
275		return BP_WAIT;
276
277	balloon_hotplug = round_up(credit, PAGES_PER_SECTION);
278
279	resource = additional_memory_resource(balloon_hotplug * PAGE_SIZE);
280	if (!resource)
281		goto err;
282
283	nid = memory_add_physaddr_to_nid(resource->start);
284
285#ifdef CONFIG_XEN_HAVE_PVMMU
286	/*
287	 * We don't support PV MMU when Linux and Xen is using
288	 * different page granularity.
289	 */
290	BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
291
292        /*
293         * add_memory() will build page tables for the new memory so
294         * the p2m must contain invalid entries so the correct
295         * non-present PTEs will be written.
296         *
297         * If a failure occurs, the original (identity) p2m entries
298         * are not restored since this region is now known not to
299         * conflict with any devices.
300         */ 
301	if (!xen_feature(XENFEAT_auto_translated_physmap)) {
302		unsigned long pfn, i;
303
304		pfn = PFN_DOWN(resource->start);
305		for (i = 0; i < balloon_hotplug; i++) {
306			if (!set_phys_to_machine(pfn + i, INVALID_P2M_ENTRY)) {
307				pr_warn("set_phys_to_machine() failed, no memory added\n");
308				goto err;
309			}
310                }
311	}
312#endif
313
314	/*
315	 * add_memory_resource() will call online_pages() which in its turn
316	 * will call xen_online_page() callback causing deadlock if we don't
317	 * release balloon_mutex here. Unlocking here is safe because the
318	 * callers drop the mutex before trying again.
319	 */
320	mutex_unlock(&balloon_mutex);
321	/* add_memory_resource() requires the device_hotplug lock */
322	lock_device_hotplug();
323	rc = add_memory_resource(nid, resource, MHP_MERGE_RESOURCE);
324	unlock_device_hotplug();
325	mutex_lock(&balloon_mutex);
326
327	if (rc) {
328		pr_warn("Cannot add additional memory (%i)\n", rc);
329		goto err;
330	}
331
332	balloon_stats.total_pages += balloon_hotplug;
333
334	return BP_WAIT;
335  err:
336	release_memory_resource(resource);
337	return BP_ECANCELED;
338}
339
340static void xen_online_page(struct page *page, unsigned int order)
341{
342	unsigned long i, size = (1 << order);
343	unsigned long start_pfn = page_to_pfn(page);
344	struct page *p;
345
346	pr_debug("Online %lu pages starting at pfn 0x%lx\n", size, start_pfn);
347	mutex_lock(&balloon_mutex);
348	for (i = 0; i < size; i++) {
349		p = pfn_to_page(start_pfn + i);
350		balloon_append(p);
351	}
352	mutex_unlock(&balloon_mutex);
353}
354
355static int xen_memory_notifier(struct notifier_block *nb, unsigned long val, void *v)
356{
357	if (val == MEM_ONLINE)
358		wake_up(&balloon_thread_wq);
359
360	return NOTIFY_OK;
361}
362
363static struct notifier_block xen_memory_nb = {
364	.notifier_call = xen_memory_notifier,
365	.priority = 0
366};
367#else
368static enum bp_state reserve_additional_memory(void)
369{
370	balloon_stats.target_pages = balloon_stats.current_pages +
371				     balloon_stats.target_unpopulated;
372	return BP_ECANCELED;
373}
374#endif /* CONFIG_XEN_BALLOON_MEMORY_HOTPLUG */
375
376static long current_credit(void)
377{
378	return balloon_stats.target_pages - balloon_stats.current_pages;
379}
380
381static bool balloon_is_inflated(void)
382{
383	return balloon_stats.balloon_low || balloon_stats.balloon_high;
384}
385
386static enum bp_state increase_reservation(unsigned long nr_pages)
387{
388	int rc;
389	unsigned long i;
390	struct page   *page;
391
392	if (nr_pages > ARRAY_SIZE(frame_list))
393		nr_pages = ARRAY_SIZE(frame_list);
394
395	page = list_first_entry_or_null(&ballooned_pages, struct page, lru);
396	for (i = 0; i < nr_pages; i++) {
397		if (!page) {
398			nr_pages = i;
399			break;
400		}
401
402		frame_list[i] = page_to_xen_pfn(page);
403		page = balloon_next_page(page);
404	}
405
406	rc = xenmem_reservation_increase(nr_pages, frame_list);
407	if (rc <= 0)
408		return BP_EAGAIN;
409
410	for (i = 0; i < rc; i++) {
411		page = balloon_retrieve(false);
412		BUG_ON(page == NULL);
413
414		xenmem_reservation_va_mapping_update(1, &page, &frame_list[i]);
415
416		/*
417		 * Relinquish the page back to the allocator. Note that
418		 * some pages, including ones added via xen_online_page(), might
419		 * not be marked reserved; free_reserved_page() will handle that.
420		 */
421		free_reserved_page(page);
422	}
423
424	balloon_stats.current_pages += rc;
425
426	return BP_DONE;
427}
428
429static enum bp_state decrease_reservation(unsigned long nr_pages, gfp_t gfp)
430{
431	enum bp_state state = BP_DONE;
432	unsigned long i;
433	struct page *page, *tmp;
434	int ret;
435	LIST_HEAD(pages);
436
437	if (nr_pages > ARRAY_SIZE(frame_list))
438		nr_pages = ARRAY_SIZE(frame_list);
439
440	for (i = 0; i < nr_pages; i++) {
441		page = alloc_page(gfp);
442		if (page == NULL) {
443			nr_pages = i;
444			state = BP_EAGAIN;
445			break;
446		}
447		adjust_managed_page_count(page, -1);
448		xenmem_reservation_scrub_page(page);
449		list_add(&page->lru, &pages);
450	}
451
452	/*
453	 * Ensure that ballooned highmem pages don't have kmaps.
454	 *
455	 * Do this before changing the p2m as kmap_flush_unused()
456	 * reads PTEs to obtain pages (and hence needs the original
457	 * p2m entry).
458	 */
459	kmap_flush_unused();
460
461	/*
462	 * Setup the frame, update direct mapping, invalidate P2M,
463	 * and add to balloon.
464	 */
465	i = 0;
466	list_for_each_entry_safe(page, tmp, &pages, lru) {
467		frame_list[i++] = xen_page_to_gfn(page);
468
469		xenmem_reservation_va_mapping_reset(1, &page);
470
471		list_del(&page->lru);
472
473		balloon_append(page);
474	}
475
476	flush_tlb_all();
477
478	ret = xenmem_reservation_decrease(nr_pages, frame_list);
479	BUG_ON(ret != nr_pages);
480
481	balloon_stats.current_pages -= nr_pages;
482
483	return state;
484}
485
486/*
487 * Stop waiting if either state is BP_DONE and ballooning action is
488 * needed, or if the credit has changed while state is not BP_DONE.
489 */
490static bool balloon_thread_cond(long credit)
491{
492	if (balloon_state == BP_DONE)
493		credit = 0;
494
495	return current_credit() != credit || kthread_should_stop();
496}
497
498/*
499 * As this is a kthread it is guaranteed to run as a single instance only.
500 * We may of course race updates of the target counts (which are protected
501 * by the balloon lock), or with changes to the Xen hard limit, but we will
502 * recover from these in time.
503 */
504static int balloon_thread(void *unused)
505{
506	long credit;
507	unsigned long timeout;
508
509	set_freezable();
510	for (;;) {
511		switch (balloon_state) {
512		case BP_DONE:
513		case BP_ECANCELED:
514			timeout = 3600 * HZ;
515			break;
516		case BP_EAGAIN:
517			timeout = balloon_stats.schedule_delay * HZ;
518			break;
519		case BP_WAIT:
520			timeout = HZ;
521			break;
522		}
523
524		credit = current_credit();
525
526		wait_event_freezable_timeout(balloon_thread_wq,
527			balloon_thread_cond(credit), timeout);
528
529		if (kthread_should_stop())
530			return 0;
531
532		mutex_lock(&balloon_mutex);
533
534		credit = current_credit();
535
536		if (credit > 0) {
537			if (balloon_is_inflated())
538				balloon_state = increase_reservation(credit);
539			else
540				balloon_state = reserve_additional_memory();
541		}
542
543		if (credit < 0) {
544			long n_pages;
545
546			n_pages = min(-credit, si_mem_available());
547			balloon_state = decrease_reservation(n_pages,
548							     GFP_BALLOON);
549			if (balloon_state == BP_DONE && n_pages != -credit &&
550			    n_pages < totalreserve_pages)
551				balloon_state = BP_EAGAIN;
552		}
553
554		update_schedule();
555
556		mutex_unlock(&balloon_mutex);
557
558		cond_resched();
559	}
560}
561
562/* Resets the Xen limit, sets new target, and kicks off processing. */
563void balloon_set_new_target(unsigned long target)
564{
565	/* No need for lock. Not read-modify-write updates. */
566	balloon_stats.target_pages = target;
567	wake_up(&balloon_thread_wq);
568}
569EXPORT_SYMBOL_GPL(balloon_set_new_target);
570
571static int add_ballooned_pages(unsigned int nr_pages)
572{
573	enum bp_state st;
574
575	if (xen_hotplug_unpopulated) {
576		st = reserve_additional_memory();
577		if (st != BP_ECANCELED) {
578			int rc;
579
580			mutex_unlock(&balloon_mutex);
581			rc = wait_event_interruptible(balloon_wq,
582				   !list_empty(&ballooned_pages));
583			mutex_lock(&balloon_mutex);
584			return rc ? -ENOMEM : 0;
585		}
586	}
587
588	if (si_mem_available() < nr_pages)
589		return -ENOMEM;
590
591	st = decrease_reservation(nr_pages, GFP_USER);
592	if (st != BP_DONE)
593		return -ENOMEM;
594
595	return 0;
596}
597
598/**
599 * xen_alloc_ballooned_pages - get pages that have been ballooned out
600 * @nr_pages: Number of pages to get
601 * @pages: pages returned
602 * @return 0 on success, error otherwise
603 */
604int xen_alloc_ballooned_pages(unsigned int nr_pages, struct page **pages)
605{
606	unsigned int pgno = 0;
607	struct page *page;
608	int ret;
609
610	mutex_lock(&balloon_mutex);
611
612	balloon_stats.target_unpopulated += nr_pages;
613
614	while (pgno < nr_pages) {
615		page = balloon_retrieve(true);
616		if (page) {
617			pages[pgno++] = page;
618#ifdef CONFIG_XEN_HAVE_PVMMU
619			/*
620			 * We don't support PV MMU when Linux and Xen is using
621			 * different page granularity.
622			 */
623			BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
624
625			if (!xen_feature(XENFEAT_auto_translated_physmap)) {
626				ret = xen_alloc_p2m_entry(page_to_pfn(page));
627				if (ret < 0)
628					goto out_undo;
629			}
630#endif
631		} else {
632			ret = add_ballooned_pages(nr_pages - pgno);
633			if (ret < 0)
634				goto out_undo;
635		}
636	}
637	mutex_unlock(&balloon_mutex);
638	return 0;
639 out_undo:
640	mutex_unlock(&balloon_mutex);
641	xen_free_ballooned_pages(pgno, pages);
642	/*
643	 * NB: xen_free_ballooned_pages will only subtract pgno pages, but since
644	 * target_unpopulated is incremented with nr_pages at the start we need
645	 * to remove the remaining ones also, or accounting will be screwed.
646	 */
647	balloon_stats.target_unpopulated -= nr_pages - pgno;
648	return ret;
649}
650EXPORT_SYMBOL(xen_alloc_ballooned_pages);
651
652/**
653 * xen_free_ballooned_pages - return pages retrieved with get_ballooned_pages
654 * @nr_pages: Number of pages
655 * @pages: pages to return
656 */
657void xen_free_ballooned_pages(unsigned int nr_pages, struct page **pages)
658{
659	unsigned int i;
660
661	mutex_lock(&balloon_mutex);
662
663	for (i = 0; i < nr_pages; i++) {
664		if (pages[i])
665			balloon_append(pages[i]);
666	}
667
668	balloon_stats.target_unpopulated -= nr_pages;
669
670	/* The balloon may be too large now. Shrink it if needed. */
671	if (current_credit())
672		wake_up(&balloon_thread_wq);
673
674	mutex_unlock(&balloon_mutex);
675}
676EXPORT_SYMBOL(xen_free_ballooned_pages);
677
678static void __init balloon_add_regions(void)
679{
 
680	unsigned long start_pfn, pages;
681	unsigned long pfn, extra_pfn_end;
682	unsigned int i;
683
684	for (i = 0; i < XEN_EXTRA_MEM_MAX_REGIONS; i++) {
685		pages = xen_extra_mem[i].n_pfns;
686		if (!pages)
687			continue;
688
689		start_pfn = xen_extra_mem[i].start_pfn;
690
691		/*
692		 * If the amount of usable memory has been limited (e.g., with
693		 * the 'mem' command line parameter), don't add pages beyond
694		 * this limit.
695		 */
696		extra_pfn_end = min(max_pfn, start_pfn + pages);
697
698		for (pfn = start_pfn; pfn < extra_pfn_end; pfn++)
699			balloon_append(pfn_to_page(pfn));
700
701		balloon_stats.total_pages += extra_pfn_end - start_pfn;
702	}
 
703}
704
705static int __init balloon_init(void)
706{
707	struct task_struct *task;
708
709	if (!xen_domain())
710		return -ENODEV;
711
712	pr_info("Initialising balloon driver\n");
713
714#ifdef CONFIG_XEN_PV
715	balloon_stats.current_pages = xen_pv_domain()
716		? min(xen_start_info->nr_pages - xen_released_pages, max_pfn)
717		: get_num_physpages();
718#else
719	balloon_stats.current_pages = get_num_physpages();
720#endif
721	balloon_stats.target_pages  = balloon_stats.current_pages;
722	balloon_stats.balloon_low   = 0;
723	balloon_stats.balloon_high  = 0;
724	balloon_stats.total_pages   = balloon_stats.current_pages;
725
726	balloon_stats.schedule_delay = 1;
727	balloon_stats.max_schedule_delay = 32;
728	balloon_stats.retry_count = 1;
729	balloon_stats.max_retry_count = 4;
730
731#ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
732	set_online_page_callback(&xen_online_page);
733	register_memory_notifier(&xen_memory_nb);
734	register_sysctl_init("xen/balloon", balloon_table);
735#endif
736
737	balloon_add_regions();
738
739	task = kthread_run(balloon_thread, NULL, "xen-balloon");
740	if (IS_ERR(task)) {
741		pr_err("xen-balloon thread could not be started, ballooning will not work!\n");
742		return PTR_ERR(task);
743	}
744
745	/* Init the xen-balloon driver. */
746	xen_balloon_init();
747
748	return 0;
749}
750subsys_initcall(balloon_init);
751
752static int __init balloon_wait_finish(void)
753{
754	long credit, last_credit = 0;
755	unsigned long last_changed = 0;
756
757	if (!xen_domain())
758		return -ENODEV;
759
760	/* PV guests don't need to wait. */
761	if (xen_pv_domain() || !current_credit())
762		return 0;
763
764	pr_notice("Waiting for initial ballooning down having finished.\n");
765
766	while ((credit = current_credit()) < 0) {
767		if (credit != last_credit) {
768			last_changed = jiffies;
769			last_credit = credit;
770		}
771		if (balloon_state == BP_ECANCELED) {
772			pr_warn_once("Initial ballooning failed, %ld pages need to be freed.\n",
773				     -credit);
774			if (time_is_before_eq_jiffies(last_changed + HZ * balloon_boot_timeout))
775				panic("Initial ballooning failed!\n");
776		}
777
778		schedule_timeout_interruptible(HZ / 10);
779	}
780
781	pr_notice("Initial ballooning down finished.\n");
782
783	return 0;
784}
785late_initcall_sync(balloon_wait_finish);
v6.2
  1/******************************************************************************
  2 * Xen balloon driver - enables returning/claiming memory to/from Xen.
  3 *
  4 * Copyright (c) 2003, B Dragovic
  5 * Copyright (c) 2003-2004, M Williamson, K Fraser
  6 * Copyright (c) 2005 Dan M. Smith, IBM Corporation
  7 * Copyright (c) 2010 Daniel Kiper
  8 *
  9 * Memory hotplug support was written by Daniel Kiper. Work on
 10 * it was sponsored by Google under Google Summer of Code 2010
 11 * program. Jeremy Fitzhardinge from Citrix was the mentor for
 12 * this project.
 13 *
 14 * This program is free software; you can redistribute it and/or
 15 * modify it under the terms of the GNU General Public License version 2
 16 * as published by the Free Software Foundation; or, when distributed
 17 * separately from the Linux kernel or incorporated into other
 18 * software packages, subject to the following license:
 19 *
 20 * Permission is hereby granted, free of charge, to any person obtaining a copy
 21 * of this source file (the "Software"), to deal in the Software without
 22 * restriction, including without limitation the rights to use, copy, modify,
 23 * merge, publish, distribute, sublicense, and/or sell copies of the Software,
 24 * and to permit persons to whom the Software is furnished to do so, subject to
 25 * the following conditions:
 26 *
 27 * The above copyright notice and this permission notice shall be included in
 28 * all copies or substantial portions of the Software.
 29 *
 30 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 31 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 32 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
 33 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 34 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
 35 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
 36 * IN THE SOFTWARE.
 37 */
 38
 39#define pr_fmt(fmt) "xen:" KBUILD_MODNAME ": " fmt
 40
 41#include <linux/cpu.h>
 42#include <linux/kernel.h>
 43#include <linux/sched.h>
 44#include <linux/cred.h>
 45#include <linux/errno.h>
 46#include <linux/freezer.h>
 47#include <linux/kthread.h>
 48#include <linux/mm.h>
 49#include <linux/memblock.h>
 50#include <linux/pagemap.h>
 51#include <linux/highmem.h>
 52#include <linux/mutex.h>
 53#include <linux/list.h>
 54#include <linux/gfp.h>
 55#include <linux/notifier.h>
 56#include <linux/memory.h>
 57#include <linux/memory_hotplug.h>
 58#include <linux/percpu-defs.h>
 59#include <linux/slab.h>
 60#include <linux/sysctl.h>
 61#include <linux/moduleparam.h>
 62#include <linux/jiffies.h>
 63
 64#include <asm/page.h>
 65#include <asm/tlb.h>
 66
 67#include <asm/xen/hypervisor.h>
 68#include <asm/xen/hypercall.h>
 69
 70#include <xen/xen.h>
 71#include <xen/interface/xen.h>
 72#include <xen/interface/memory.h>
 73#include <xen/balloon.h>
 74#include <xen/features.h>
 75#include <xen/page.h>
 76#include <xen/mem-reservation.h>
 77
 78#undef MODULE_PARAM_PREFIX
 79#define MODULE_PARAM_PREFIX "xen."
 80
 81static uint __read_mostly balloon_boot_timeout = 180;
 82module_param(balloon_boot_timeout, uint, 0444);
 83
 84#ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
 85static int xen_hotplug_unpopulated;
 86
 87static struct ctl_table balloon_table[] = {
 88	{
 89		.procname	= "hotplug_unpopulated",
 90		.data		= &xen_hotplug_unpopulated,
 91		.maxlen		= sizeof(int),
 92		.mode		= 0644,
 93		.proc_handler	= proc_dointvec_minmax,
 94		.extra1         = SYSCTL_ZERO,
 95		.extra2         = SYSCTL_ONE,
 96	},
 97	{ }
 98};
 99
100static struct ctl_table balloon_root[] = {
101	{
102		.procname	= "balloon",
103		.mode		= 0555,
104		.child		= balloon_table,
105	},
106	{ }
107};
108
109static struct ctl_table xen_root[] = {
110	{
111		.procname	= "xen",
112		.mode		= 0555,
113		.child		= balloon_root,
114	},
115	{ }
116};
117
118#else
119#define xen_hotplug_unpopulated 0
120#endif
121
122/*
123 * Use one extent per PAGE_SIZE to avoid to break down the page into
124 * multiple frame.
125 */
126#define EXTENT_ORDER (fls(XEN_PFN_PER_PAGE) - 1)
127
128/*
129 * balloon_thread() state:
130 *
131 * BP_DONE: done or nothing to do,
132 * BP_WAIT: wait to be rescheduled,
133 * BP_EAGAIN: error, go to sleep,
134 * BP_ECANCELED: error, balloon operation canceled.
135 */
136
137static enum bp_state {
138	BP_DONE,
139	BP_WAIT,
140	BP_EAGAIN,
141	BP_ECANCELED
142} balloon_state = BP_DONE;
143
144/* Main waiting point for xen-balloon thread. */
145static DECLARE_WAIT_QUEUE_HEAD(balloon_thread_wq);
146
147static DEFINE_MUTEX(balloon_mutex);
148
149struct balloon_stats balloon_stats;
150EXPORT_SYMBOL_GPL(balloon_stats);
151
152/* We increase/decrease in batches which fit in a page */
153static xen_pfn_t frame_list[PAGE_SIZE / sizeof(xen_pfn_t)];
154
155
156/* List of ballooned pages, threaded through the mem_map array. */
157static LIST_HEAD(ballooned_pages);
158static DECLARE_WAIT_QUEUE_HEAD(balloon_wq);
159
160/* When ballooning out (allocating memory to return to Xen) we don't really
161   want the kernel to try too hard since that can trigger the oom killer. */
162#define GFP_BALLOON \
163	(GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC)
164
165/* balloon_append: add the given page to the balloon. */
166static void balloon_append(struct page *page)
167{
168	__SetPageOffline(page);
 
169
170	/* Lowmem is re-populated first, so highmem pages go at list tail. */
171	if (PageHighMem(page)) {
172		list_add_tail(&page->lru, &ballooned_pages);
173		balloon_stats.balloon_high++;
174	} else {
175		list_add(&page->lru, &ballooned_pages);
176		balloon_stats.balloon_low++;
177	}
178	wake_up(&balloon_wq);
179}
180
181/* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
182static struct page *balloon_retrieve(bool require_lowmem)
183{
184	struct page *page;
185
186	if (list_empty(&ballooned_pages))
187		return NULL;
188
189	page = list_entry(ballooned_pages.next, struct page, lru);
190	if (require_lowmem && PageHighMem(page))
191		return NULL;
192	list_del(&page->lru);
193
194	if (PageHighMem(page))
195		balloon_stats.balloon_high--;
196	else
197		balloon_stats.balloon_low--;
198
199	__ClearPageOffline(page);
200	return page;
201}
202
203static struct page *balloon_next_page(struct page *page)
204{
205	struct list_head *next = page->lru.next;
206	if (next == &ballooned_pages)
207		return NULL;
208	return list_entry(next, struct page, lru);
209}
210
211static void update_schedule(void)
212{
213	if (balloon_state == BP_WAIT || balloon_state == BP_ECANCELED)
214		return;
215
216	if (balloon_state == BP_DONE) {
217		balloon_stats.schedule_delay = 1;
218		balloon_stats.retry_count = 1;
219		return;
220	}
221
222	++balloon_stats.retry_count;
223
224	if (balloon_stats.max_retry_count != RETRY_UNLIMITED &&
225			balloon_stats.retry_count > balloon_stats.max_retry_count) {
226		balloon_stats.schedule_delay = 1;
227		balloon_stats.retry_count = 1;
228		balloon_state = BP_ECANCELED;
229		return;
230	}
231
232	balloon_stats.schedule_delay <<= 1;
233
234	if (balloon_stats.schedule_delay > balloon_stats.max_schedule_delay)
235		balloon_stats.schedule_delay = balloon_stats.max_schedule_delay;
236
237	balloon_state = BP_EAGAIN;
238}
239
240#ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
241static void release_memory_resource(struct resource *resource)
242{
243	if (!resource)
244		return;
245
246	/*
247	 * No need to reset region to identity mapped since we now
248	 * know that no I/O can be in this region
249	 */
250	release_resource(resource);
251	kfree(resource);
252}
253
254static struct resource *additional_memory_resource(phys_addr_t size)
255{
256	struct resource *res;
257	int ret;
258
259	res = kzalloc(sizeof(*res), GFP_KERNEL);
260	if (!res)
261		return NULL;
262
263	res->name = "System RAM";
264	res->flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
265
266	ret = allocate_resource(&iomem_resource, res,
267				size, 0, -1,
268				PAGES_PER_SECTION * PAGE_SIZE, NULL, NULL);
269	if (ret < 0) {
270		pr_err("Cannot allocate new System RAM resource\n");
271		kfree(res);
272		return NULL;
273	}
274
275	return res;
276}
277
278static enum bp_state reserve_additional_memory(void)
279{
280	long credit;
281	struct resource *resource;
282	int nid, rc;
283	unsigned long balloon_hotplug;
284
285	credit = balloon_stats.target_pages + balloon_stats.target_unpopulated
286		- balloon_stats.total_pages;
287
288	/*
289	 * Already hotplugged enough pages?  Wait for them to be
290	 * onlined.
291	 */
292	if (credit <= 0)
293		return BP_WAIT;
294
295	balloon_hotplug = round_up(credit, PAGES_PER_SECTION);
296
297	resource = additional_memory_resource(balloon_hotplug * PAGE_SIZE);
298	if (!resource)
299		goto err;
300
301	nid = memory_add_physaddr_to_nid(resource->start);
302
303#ifdef CONFIG_XEN_HAVE_PVMMU
304	/*
305	 * We don't support PV MMU when Linux and Xen is using
306	 * different page granularity.
307	 */
308	BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
309
310        /*
311         * add_memory() will build page tables for the new memory so
312         * the p2m must contain invalid entries so the correct
313         * non-present PTEs will be written.
314         *
315         * If a failure occurs, the original (identity) p2m entries
316         * are not restored since this region is now known not to
317         * conflict with any devices.
318         */ 
319	if (!xen_feature(XENFEAT_auto_translated_physmap)) {
320		unsigned long pfn, i;
321
322		pfn = PFN_DOWN(resource->start);
323		for (i = 0; i < balloon_hotplug; i++) {
324			if (!set_phys_to_machine(pfn + i, INVALID_P2M_ENTRY)) {
325				pr_warn("set_phys_to_machine() failed, no memory added\n");
326				goto err;
327			}
328                }
329	}
330#endif
331
332	/*
333	 * add_memory_resource() will call online_pages() which in its turn
334	 * will call xen_online_page() callback causing deadlock if we don't
335	 * release balloon_mutex here. Unlocking here is safe because the
336	 * callers drop the mutex before trying again.
337	 */
338	mutex_unlock(&balloon_mutex);
339	/* add_memory_resource() requires the device_hotplug lock */
340	lock_device_hotplug();
341	rc = add_memory_resource(nid, resource, MHP_MERGE_RESOURCE);
342	unlock_device_hotplug();
343	mutex_lock(&balloon_mutex);
344
345	if (rc) {
346		pr_warn("Cannot add additional memory (%i)\n", rc);
347		goto err;
348	}
349
350	balloon_stats.total_pages += balloon_hotplug;
351
352	return BP_WAIT;
353  err:
354	release_memory_resource(resource);
355	return BP_ECANCELED;
356}
357
358static void xen_online_page(struct page *page, unsigned int order)
359{
360	unsigned long i, size = (1 << order);
361	unsigned long start_pfn = page_to_pfn(page);
362	struct page *p;
363
364	pr_debug("Online %lu pages starting at pfn 0x%lx\n", size, start_pfn);
365	mutex_lock(&balloon_mutex);
366	for (i = 0; i < size; i++) {
367		p = pfn_to_page(start_pfn + i);
368		balloon_append(p);
369	}
370	mutex_unlock(&balloon_mutex);
371}
372
373static int xen_memory_notifier(struct notifier_block *nb, unsigned long val, void *v)
374{
375	if (val == MEM_ONLINE)
376		wake_up(&balloon_thread_wq);
377
378	return NOTIFY_OK;
379}
380
381static struct notifier_block xen_memory_nb = {
382	.notifier_call = xen_memory_notifier,
383	.priority = 0
384};
385#else
386static enum bp_state reserve_additional_memory(void)
387{
388	balloon_stats.target_pages = balloon_stats.current_pages +
389				     balloon_stats.target_unpopulated;
390	return BP_ECANCELED;
391}
392#endif /* CONFIG_XEN_BALLOON_MEMORY_HOTPLUG */
393
394static long current_credit(void)
395{
396	return balloon_stats.target_pages - balloon_stats.current_pages;
397}
398
399static bool balloon_is_inflated(void)
400{
401	return balloon_stats.balloon_low || balloon_stats.balloon_high;
402}
403
404static enum bp_state increase_reservation(unsigned long nr_pages)
405{
406	int rc;
407	unsigned long i;
408	struct page   *page;
409
410	if (nr_pages > ARRAY_SIZE(frame_list))
411		nr_pages = ARRAY_SIZE(frame_list);
412
413	page = list_first_entry_or_null(&ballooned_pages, struct page, lru);
414	for (i = 0; i < nr_pages; i++) {
415		if (!page) {
416			nr_pages = i;
417			break;
418		}
419
420		frame_list[i] = page_to_xen_pfn(page);
421		page = balloon_next_page(page);
422	}
423
424	rc = xenmem_reservation_increase(nr_pages, frame_list);
425	if (rc <= 0)
426		return BP_EAGAIN;
427
428	for (i = 0; i < rc; i++) {
429		page = balloon_retrieve(false);
430		BUG_ON(page == NULL);
431
432		xenmem_reservation_va_mapping_update(1, &page, &frame_list[i]);
433
434		/* Relinquish the page back to the allocator. */
 
 
 
 
435		free_reserved_page(page);
436	}
437
438	balloon_stats.current_pages += rc;
439
440	return BP_DONE;
441}
442
443static enum bp_state decrease_reservation(unsigned long nr_pages, gfp_t gfp)
444{
445	enum bp_state state = BP_DONE;
446	unsigned long i;
447	struct page *page, *tmp;
448	int ret;
449	LIST_HEAD(pages);
450
451	if (nr_pages > ARRAY_SIZE(frame_list))
452		nr_pages = ARRAY_SIZE(frame_list);
453
454	for (i = 0; i < nr_pages; i++) {
455		page = alloc_page(gfp);
456		if (page == NULL) {
457			nr_pages = i;
458			state = BP_EAGAIN;
459			break;
460		}
461		adjust_managed_page_count(page, -1);
462		xenmem_reservation_scrub_page(page);
463		list_add(&page->lru, &pages);
464	}
465
466	/*
467	 * Ensure that ballooned highmem pages don't have kmaps.
468	 *
469	 * Do this before changing the p2m as kmap_flush_unused()
470	 * reads PTEs to obtain pages (and hence needs the original
471	 * p2m entry).
472	 */
473	kmap_flush_unused();
474
475	/*
476	 * Setup the frame, update direct mapping, invalidate P2M,
477	 * and add to balloon.
478	 */
479	i = 0;
480	list_for_each_entry_safe(page, tmp, &pages, lru) {
481		frame_list[i++] = xen_page_to_gfn(page);
482
483		xenmem_reservation_va_mapping_reset(1, &page);
484
485		list_del(&page->lru);
486
487		balloon_append(page);
488	}
489
490	flush_tlb_all();
491
492	ret = xenmem_reservation_decrease(nr_pages, frame_list);
493	BUG_ON(ret != nr_pages);
494
495	balloon_stats.current_pages -= nr_pages;
496
497	return state;
498}
499
500/*
501 * Stop waiting if either state is BP_DONE and ballooning action is
502 * needed, or if the credit has changed while state is not BP_DONE.
503 */
504static bool balloon_thread_cond(long credit)
505{
506	if (balloon_state == BP_DONE)
507		credit = 0;
508
509	return current_credit() != credit || kthread_should_stop();
510}
511
512/*
513 * As this is a kthread it is guaranteed to run as a single instance only.
514 * We may of course race updates of the target counts (which are protected
515 * by the balloon lock), or with changes to the Xen hard limit, but we will
516 * recover from these in time.
517 */
518static int balloon_thread(void *unused)
519{
520	long credit;
521	unsigned long timeout;
522
523	set_freezable();
524	for (;;) {
525		switch (balloon_state) {
526		case BP_DONE:
527		case BP_ECANCELED:
528			timeout = 3600 * HZ;
529			break;
530		case BP_EAGAIN:
531			timeout = balloon_stats.schedule_delay * HZ;
532			break;
533		case BP_WAIT:
534			timeout = HZ;
535			break;
536		}
537
538		credit = current_credit();
539
540		wait_event_freezable_timeout(balloon_thread_wq,
541			balloon_thread_cond(credit), timeout);
542
543		if (kthread_should_stop())
544			return 0;
545
546		mutex_lock(&balloon_mutex);
547
548		credit = current_credit();
549
550		if (credit > 0) {
551			if (balloon_is_inflated())
552				balloon_state = increase_reservation(credit);
553			else
554				balloon_state = reserve_additional_memory();
555		}
556
557		if (credit < 0) {
558			long n_pages;
559
560			n_pages = min(-credit, si_mem_available());
561			balloon_state = decrease_reservation(n_pages,
562							     GFP_BALLOON);
563			if (balloon_state == BP_DONE && n_pages != -credit &&
564			    n_pages < totalreserve_pages)
565				balloon_state = BP_EAGAIN;
566		}
567
568		update_schedule();
569
570		mutex_unlock(&balloon_mutex);
571
572		cond_resched();
573	}
574}
575
576/* Resets the Xen limit, sets new target, and kicks off processing. */
577void balloon_set_new_target(unsigned long target)
578{
579	/* No need for lock. Not read-modify-write updates. */
580	balloon_stats.target_pages = target;
581	wake_up(&balloon_thread_wq);
582}
583EXPORT_SYMBOL_GPL(balloon_set_new_target);
584
585static int add_ballooned_pages(unsigned int nr_pages)
586{
587	enum bp_state st;
588
589	if (xen_hotplug_unpopulated) {
590		st = reserve_additional_memory();
591		if (st != BP_ECANCELED) {
592			int rc;
593
594			mutex_unlock(&balloon_mutex);
595			rc = wait_event_interruptible(balloon_wq,
596				   !list_empty(&ballooned_pages));
597			mutex_lock(&balloon_mutex);
598			return rc ? -ENOMEM : 0;
599		}
600	}
601
602	if (si_mem_available() < nr_pages)
603		return -ENOMEM;
604
605	st = decrease_reservation(nr_pages, GFP_USER);
606	if (st != BP_DONE)
607		return -ENOMEM;
608
609	return 0;
610}
611
612/**
613 * xen_alloc_ballooned_pages - get pages that have been ballooned out
614 * @nr_pages: Number of pages to get
615 * @pages: pages returned
616 * @return 0 on success, error otherwise
617 */
618int xen_alloc_ballooned_pages(unsigned int nr_pages, struct page **pages)
619{
620	unsigned int pgno = 0;
621	struct page *page;
622	int ret;
623
624	mutex_lock(&balloon_mutex);
625
626	balloon_stats.target_unpopulated += nr_pages;
627
628	while (pgno < nr_pages) {
629		page = balloon_retrieve(true);
630		if (page) {
631			pages[pgno++] = page;
632#ifdef CONFIG_XEN_HAVE_PVMMU
633			/*
634			 * We don't support PV MMU when Linux and Xen is using
635			 * different page granularity.
636			 */
637			BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
638
639			if (!xen_feature(XENFEAT_auto_translated_physmap)) {
640				ret = xen_alloc_p2m_entry(page_to_pfn(page));
641				if (ret < 0)
642					goto out_undo;
643			}
644#endif
645		} else {
646			ret = add_ballooned_pages(nr_pages - pgno);
647			if (ret < 0)
648				goto out_undo;
649		}
650	}
651	mutex_unlock(&balloon_mutex);
652	return 0;
653 out_undo:
654	mutex_unlock(&balloon_mutex);
655	xen_free_ballooned_pages(pgno, pages);
656	/*
657	 * NB: xen_free_ballooned_pages will only subtract pgno pages, but since
658	 * target_unpopulated is incremented with nr_pages at the start we need
659	 * to remove the remaining ones also, or accounting will be screwed.
660	 */
661	balloon_stats.target_unpopulated -= nr_pages - pgno;
662	return ret;
663}
664EXPORT_SYMBOL(xen_alloc_ballooned_pages);
665
666/**
667 * xen_free_ballooned_pages - return pages retrieved with get_ballooned_pages
668 * @nr_pages: Number of pages
669 * @pages: pages to return
670 */
671void xen_free_ballooned_pages(unsigned int nr_pages, struct page **pages)
672{
673	unsigned int i;
674
675	mutex_lock(&balloon_mutex);
676
677	for (i = 0; i < nr_pages; i++) {
678		if (pages[i])
679			balloon_append(pages[i]);
680	}
681
682	balloon_stats.target_unpopulated -= nr_pages;
683
684	/* The balloon may be too large now. Shrink it if needed. */
685	if (current_credit())
686		wake_up(&balloon_thread_wq);
687
688	mutex_unlock(&balloon_mutex);
689}
690EXPORT_SYMBOL(xen_free_ballooned_pages);
691
692static void __init balloon_add_regions(void)
693{
694#if defined(CONFIG_XEN_PV)
695	unsigned long start_pfn, pages;
696	unsigned long pfn, extra_pfn_end;
697	unsigned int i;
698
699	for (i = 0; i < XEN_EXTRA_MEM_MAX_REGIONS; i++) {
700		pages = xen_extra_mem[i].n_pfns;
701		if (!pages)
702			continue;
703
704		start_pfn = xen_extra_mem[i].start_pfn;
705
706		/*
707		 * If the amount of usable memory has been limited (e.g., with
708		 * the 'mem' command line parameter), don't add pages beyond
709		 * this limit.
710		 */
711		extra_pfn_end = min(max_pfn, start_pfn + pages);
712
713		for (pfn = start_pfn; pfn < extra_pfn_end; pfn++)
714			balloon_append(pfn_to_page(pfn));
715
716		balloon_stats.total_pages += extra_pfn_end - start_pfn;
717	}
718#endif
719}
720
721static int __init balloon_init(void)
722{
723	struct task_struct *task;
724
725	if (!xen_domain())
726		return -ENODEV;
727
728	pr_info("Initialising balloon driver\n");
729
730#ifdef CONFIG_XEN_PV
731	balloon_stats.current_pages = xen_pv_domain()
732		? min(xen_start_info->nr_pages - xen_released_pages, max_pfn)
733		: get_num_physpages();
734#else
735	balloon_stats.current_pages = get_num_physpages();
736#endif
737	balloon_stats.target_pages  = balloon_stats.current_pages;
738	balloon_stats.balloon_low   = 0;
739	balloon_stats.balloon_high  = 0;
740	balloon_stats.total_pages   = balloon_stats.current_pages;
741
742	balloon_stats.schedule_delay = 1;
743	balloon_stats.max_schedule_delay = 32;
744	balloon_stats.retry_count = 1;
745	balloon_stats.max_retry_count = 4;
746
747#ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
748	set_online_page_callback(&xen_online_page);
749	register_memory_notifier(&xen_memory_nb);
750	register_sysctl_table(xen_root);
751#endif
752
753	balloon_add_regions();
754
755	task = kthread_run(balloon_thread, NULL, "xen-balloon");
756	if (IS_ERR(task)) {
757		pr_err("xen-balloon thread could not be started, ballooning will not work!\n");
758		return PTR_ERR(task);
759	}
760
761	/* Init the xen-balloon driver. */
762	xen_balloon_init();
763
764	return 0;
765}
766subsys_initcall(balloon_init);
767
768static int __init balloon_wait_finish(void)
769{
770	long credit, last_credit = 0;
771	unsigned long last_changed = 0;
772
773	if (!xen_domain())
774		return -ENODEV;
775
776	/* PV guests don't need to wait. */
777	if (xen_pv_domain() || !current_credit())
778		return 0;
779
780	pr_notice("Waiting for initial ballooning down having finished.\n");
781
782	while ((credit = current_credit()) < 0) {
783		if (credit != last_credit) {
784			last_changed = jiffies;
785			last_credit = credit;
786		}
787		if (balloon_state == BP_ECANCELED) {
788			pr_warn_once("Initial ballooning failed, %ld pages need to be freed.\n",
789				     -credit);
790			if (time_is_before_eq_jiffies(last_changed + HZ * balloon_boot_timeout))
791				panic("Initial ballooning failed!\n");
792		}
793
794		schedule_timeout_interruptible(HZ / 10);
795	}
796
797	pr_notice("Initial ballooning down finished.\n");
798
799	return 0;
800}
801late_initcall_sync(balloon_wait_finish);