1 /******************************************************************************
2 * Xen balloon driver - enables returning/claiming memory to/from Xen.
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
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
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:
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:
27 * The above copyright notice and this permission notice shall be included in
28 * all copies or substantial portions of the Software.
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
39 #define pr_fmt(fmt) "xen:" KBUILD_MODNAME ": " fmt
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>
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>
66 #include <asm/xen/hypervisor.h>
67 #include <asm/xen/hypercall.h>
70 #include <xen/interface/xen.h>
71 #include <xen/interface/memory.h>
72 #include <xen/balloon.h>
73 #include <xen/features.h>
75 #include <xen/mem-reservation.h>
77 #undef MODULE_PARAM_PREFIX
78 #define MODULE_PARAM_PREFIX "xen."
80 static uint __read_mostly balloon_boot_timeout = 180;
81 module_param(balloon_boot_timeout, uint, 0444);
83 static int xen_hotplug_unpopulated;
85 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
87 static struct ctl_table balloon_table[] = {
89 .procname = "hotplug_unpopulated",
90 .data = &xen_hotplug_unpopulated,
91 .maxlen = sizeof(int),
93 .proc_handler = proc_dointvec_minmax,
94 .extra1 = SYSCTL_ZERO,
100 static struct ctl_table balloon_root[] = {
102 .procname = "balloon",
104 .child = balloon_table,
109 static struct ctl_table xen_root[] = {
113 .child = balloon_root,
121 * Use one extent per PAGE_SIZE to avoid to break down the page into
124 #define EXTENT_ORDER (fls(XEN_PFN_PER_PAGE) - 1)
127 * balloon_thread() state:
129 * BP_DONE: done or nothing to do,
130 * BP_WAIT: wait to be rescheduled,
131 * BP_EAGAIN: error, go to sleep,
132 * BP_ECANCELED: error, balloon operation canceled.
135 static enum bp_state {
140 } balloon_state = BP_DONE;
142 /* Main waiting point for xen-balloon thread. */
143 static DECLARE_WAIT_QUEUE_HEAD(balloon_thread_wq);
145 static DEFINE_MUTEX(balloon_mutex);
147 struct balloon_stats balloon_stats;
148 EXPORT_SYMBOL_GPL(balloon_stats);
150 /* We increase/decrease in batches which fit in a page */
151 static xen_pfn_t frame_list[PAGE_SIZE / sizeof(xen_pfn_t)];
154 /* List of ballooned pages, threaded through the mem_map array. */
155 static LIST_HEAD(ballooned_pages);
156 static DECLARE_WAIT_QUEUE_HEAD(balloon_wq);
158 /* When ballooning out (allocating memory to return to Xen) we don't really
159 want the kernel to try too hard since that can trigger the oom killer. */
160 #define GFP_BALLOON \
161 (GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC)
163 /* balloon_append: add the given page to the balloon. */
164 static void balloon_append(struct page *page)
166 __SetPageOffline(page);
168 /* Lowmem is re-populated first, so highmem pages go at list tail. */
169 if (PageHighMem(page)) {
170 list_add_tail(&page->lru, &ballooned_pages);
171 balloon_stats.balloon_high++;
173 list_add(&page->lru, &ballooned_pages);
174 balloon_stats.balloon_low++;
176 wake_up(&balloon_wq);
179 /* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
180 static struct page *balloon_retrieve(bool require_lowmem)
184 if (list_empty(&ballooned_pages))
187 page = list_entry(ballooned_pages.next, struct page, lru);
188 if (require_lowmem && PageHighMem(page))
190 list_del(&page->lru);
192 if (PageHighMem(page))
193 balloon_stats.balloon_high--;
195 balloon_stats.balloon_low--;
197 __ClearPageOffline(page);
201 static struct page *balloon_next_page(struct page *page)
203 struct list_head *next = page->lru.next;
204 if (next == &ballooned_pages)
206 return list_entry(next, struct page, lru);
209 static void update_schedule(void)
211 if (balloon_state == BP_WAIT || balloon_state == BP_ECANCELED)
214 if (balloon_state == BP_DONE) {
215 balloon_stats.schedule_delay = 1;
216 balloon_stats.retry_count = 1;
220 ++balloon_stats.retry_count;
222 if (balloon_stats.max_retry_count != RETRY_UNLIMITED &&
223 balloon_stats.retry_count > balloon_stats.max_retry_count) {
224 balloon_stats.schedule_delay = 1;
225 balloon_stats.retry_count = 1;
226 balloon_state = BP_ECANCELED;
230 balloon_stats.schedule_delay <<= 1;
232 if (balloon_stats.schedule_delay > balloon_stats.max_schedule_delay)
233 balloon_stats.schedule_delay = balloon_stats.max_schedule_delay;
235 balloon_state = BP_EAGAIN;
238 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
239 static void release_memory_resource(struct resource *resource)
245 * No need to reset region to identity mapped since we now
246 * know that no I/O can be in this region
248 release_resource(resource);
252 static struct resource *additional_memory_resource(phys_addr_t size)
254 struct resource *res;
257 res = kzalloc(sizeof(*res), GFP_KERNEL);
261 res->name = "System RAM";
262 res->flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
264 ret = allocate_resource(&iomem_resource, res,
266 PAGES_PER_SECTION * PAGE_SIZE, NULL, NULL);
268 pr_err("Cannot allocate new System RAM resource\n");
276 static enum bp_state reserve_additional_memory(void)
279 struct resource *resource;
281 unsigned long balloon_hotplug;
283 credit = balloon_stats.target_pages + balloon_stats.target_unpopulated
284 - balloon_stats.total_pages;
287 * Already hotplugged enough pages? Wait for them to be
293 balloon_hotplug = round_up(credit, PAGES_PER_SECTION);
295 resource = additional_memory_resource(balloon_hotplug * PAGE_SIZE);
299 nid = memory_add_physaddr_to_nid(resource->start);
301 #ifdef CONFIG_XEN_HAVE_PVMMU
303 * We don't support PV MMU when Linux and Xen is using
304 * different page granularity.
306 BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
309 * add_memory() will build page tables for the new memory so
310 * the p2m must contain invalid entries so the correct
311 * non-present PTEs will be written.
313 * If a failure occurs, the original (identity) p2m entries
314 * are not restored since this region is now known not to
315 * conflict with any devices.
317 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
318 unsigned long pfn, i;
320 pfn = PFN_DOWN(resource->start);
321 for (i = 0; i < balloon_hotplug; i++) {
322 if (!set_phys_to_machine(pfn + i, INVALID_P2M_ENTRY)) {
323 pr_warn("set_phys_to_machine() failed, no memory added\n");
331 * add_memory_resource() will call online_pages() which in its turn
332 * will call xen_online_page() callback causing deadlock if we don't
333 * release balloon_mutex here. Unlocking here is safe because the
334 * callers drop the mutex before trying again.
336 mutex_unlock(&balloon_mutex);
337 /* add_memory_resource() requires the device_hotplug lock */
338 lock_device_hotplug();
339 rc = add_memory_resource(nid, resource, MEMHP_MERGE_RESOURCE);
340 unlock_device_hotplug();
341 mutex_lock(&balloon_mutex);
344 pr_warn("Cannot add additional memory (%i)\n", rc);
348 balloon_stats.total_pages += balloon_hotplug;
352 release_memory_resource(resource);
356 static void xen_online_page(struct page *page, unsigned int order)
358 unsigned long i, size = (1 << order);
359 unsigned long start_pfn = page_to_pfn(page);
362 pr_debug("Online %lu pages starting at pfn 0x%lx\n", size, start_pfn);
363 mutex_lock(&balloon_mutex);
364 for (i = 0; i < size; i++) {
365 p = pfn_to_page(start_pfn + i);
368 mutex_unlock(&balloon_mutex);
371 static int xen_memory_notifier(struct notifier_block *nb, unsigned long val, void *v)
373 if (val == MEM_ONLINE)
374 wake_up(&balloon_thread_wq);
379 static struct notifier_block xen_memory_nb = {
380 .notifier_call = xen_memory_notifier,
384 static enum bp_state reserve_additional_memory(void)
386 balloon_stats.target_pages = balloon_stats.current_pages +
387 balloon_stats.target_unpopulated;
390 #endif /* CONFIG_XEN_BALLOON_MEMORY_HOTPLUG */
392 static long current_credit(void)
394 return balloon_stats.target_pages - balloon_stats.current_pages;
397 static bool balloon_is_inflated(void)
399 return balloon_stats.balloon_low || balloon_stats.balloon_high;
402 static enum bp_state increase_reservation(unsigned long nr_pages)
408 if (nr_pages > ARRAY_SIZE(frame_list))
409 nr_pages = ARRAY_SIZE(frame_list);
411 page = list_first_entry_or_null(&ballooned_pages, struct page, lru);
412 for (i = 0; i < nr_pages; i++) {
418 frame_list[i] = page_to_xen_pfn(page);
419 page = balloon_next_page(page);
422 rc = xenmem_reservation_increase(nr_pages, frame_list);
426 for (i = 0; i < rc; i++) {
427 page = balloon_retrieve(false);
428 BUG_ON(page == NULL);
430 xenmem_reservation_va_mapping_update(1, &page, &frame_list[i]);
432 /* Relinquish the page back to the allocator. */
433 free_reserved_page(page);
436 balloon_stats.current_pages += rc;
441 static enum bp_state decrease_reservation(unsigned long nr_pages, gfp_t gfp)
443 enum bp_state state = BP_DONE;
445 struct page *page, *tmp;
449 if (nr_pages > ARRAY_SIZE(frame_list))
450 nr_pages = ARRAY_SIZE(frame_list);
452 for (i = 0; i < nr_pages; i++) {
453 page = alloc_page(gfp);
459 adjust_managed_page_count(page, -1);
460 xenmem_reservation_scrub_page(page);
461 list_add(&page->lru, &pages);
465 * Ensure that ballooned highmem pages don't have kmaps.
467 * Do this before changing the p2m as kmap_flush_unused()
468 * reads PTEs to obtain pages (and hence needs the original
474 * Setup the frame, update direct mapping, invalidate P2M,
475 * and add to balloon.
478 list_for_each_entry_safe(page, tmp, &pages, lru) {
479 frame_list[i++] = xen_page_to_gfn(page);
481 xenmem_reservation_va_mapping_reset(1, &page);
483 list_del(&page->lru);
485 balloon_append(page);
490 ret = xenmem_reservation_decrease(nr_pages, frame_list);
491 BUG_ON(ret != nr_pages);
493 balloon_stats.current_pages -= nr_pages;
499 * Stop waiting if either state is BP_DONE and ballooning action is
500 * needed, or if the credit has changed while state is not BP_DONE.
502 static bool balloon_thread_cond(long credit)
504 if (balloon_state == BP_DONE)
507 return current_credit() != credit || kthread_should_stop();
511 * As this is a kthread it is guaranteed to run as a single instance only.
512 * We may of course race updates of the target counts (which are protected
513 * by the balloon lock), or with changes to the Xen hard limit, but we will
514 * recover from these in time.
516 static int balloon_thread(void *unused)
519 unsigned long timeout;
523 switch (balloon_state) {
529 timeout = balloon_stats.schedule_delay * HZ;
536 credit = current_credit();
538 wait_event_freezable_timeout(balloon_thread_wq,
539 balloon_thread_cond(credit), timeout);
541 if (kthread_should_stop())
544 mutex_lock(&balloon_mutex);
546 credit = current_credit();
549 if (balloon_is_inflated())
550 balloon_state = increase_reservation(credit);
552 balloon_state = reserve_additional_memory();
558 n_pages = min(-credit, si_mem_available());
559 balloon_state = decrease_reservation(n_pages,
561 if (balloon_state == BP_DONE && n_pages != -credit &&
562 n_pages < totalreserve_pages)
563 balloon_state = BP_EAGAIN;
568 mutex_unlock(&balloon_mutex);
574 /* Resets the Xen limit, sets new target, and kicks off processing. */
575 void balloon_set_new_target(unsigned long target)
577 /* No need for lock. Not read-modify-write updates. */
578 balloon_stats.target_pages = target;
579 wake_up(&balloon_thread_wq);
581 EXPORT_SYMBOL_GPL(balloon_set_new_target);
583 static int add_ballooned_pages(int nr_pages)
587 if (xen_hotplug_unpopulated) {
588 st = reserve_additional_memory();
589 if (st != BP_ECANCELED) {
592 mutex_unlock(&balloon_mutex);
593 rc = wait_event_interruptible(balloon_wq,
594 !list_empty(&ballooned_pages));
595 mutex_lock(&balloon_mutex);
596 return rc ? -ENOMEM : 0;
600 if (si_mem_available() < nr_pages)
603 st = decrease_reservation(nr_pages, GFP_USER);
611 * alloc_xenballooned_pages - get pages that have been ballooned out
612 * @nr_pages: Number of pages to get
613 * @pages: pages returned
614 * @return 0 on success, error otherwise
616 int alloc_xenballooned_pages(int nr_pages, struct page **pages)
622 mutex_lock(&balloon_mutex);
624 balloon_stats.target_unpopulated += nr_pages;
626 while (pgno < nr_pages) {
627 page = balloon_retrieve(true);
629 pages[pgno++] = page;
630 #ifdef CONFIG_XEN_HAVE_PVMMU
632 * We don't support PV MMU when Linux and Xen is using
633 * different page granularity.
635 BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
637 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
638 ret = xen_alloc_p2m_entry(page_to_pfn(page));
644 ret = add_ballooned_pages(nr_pages - pgno);
649 mutex_unlock(&balloon_mutex);
652 mutex_unlock(&balloon_mutex);
653 free_xenballooned_pages(pgno, pages);
655 * NB: free_xenballooned_pages will only subtract pgno pages, but since
656 * target_unpopulated is incremented with nr_pages at the start we need
657 * to remove the remaining ones also, or accounting will be screwed.
659 balloon_stats.target_unpopulated -= nr_pages - pgno;
662 EXPORT_SYMBOL(alloc_xenballooned_pages);
665 * free_xenballooned_pages - return pages retrieved with get_ballooned_pages
666 * @nr_pages: Number of pages
667 * @pages: pages to return
669 void free_xenballooned_pages(int nr_pages, struct page **pages)
673 mutex_lock(&balloon_mutex);
675 for (i = 0; i < nr_pages; i++) {
677 balloon_append(pages[i]);
680 balloon_stats.target_unpopulated -= nr_pages;
682 /* The balloon may be too large now. Shrink it if needed. */
683 if (current_credit())
684 wake_up(&balloon_thread_wq);
686 mutex_unlock(&balloon_mutex);
688 EXPORT_SYMBOL(free_xenballooned_pages);
690 #if defined(CONFIG_XEN_PV) && !defined(CONFIG_XEN_UNPOPULATED_ALLOC)
691 static void __init balloon_add_region(unsigned long start_pfn,
694 unsigned long pfn, extra_pfn_end;
697 * If the amount of usable memory has been limited (e.g., with
698 * the 'mem' command line parameter), don't add pages beyond
701 extra_pfn_end = min(max_pfn, start_pfn + pages);
703 for (pfn = start_pfn; pfn < extra_pfn_end; pfn++) {
704 /* totalram_pages and totalhigh_pages do not
705 include the boot-time balloon extension, so
706 don't subtract from it. */
707 balloon_append(pfn_to_page(pfn));
710 balloon_stats.total_pages += extra_pfn_end - start_pfn;
714 static int __init balloon_init(void)
716 struct task_struct *task;
721 pr_info("Initialising balloon driver\n");
724 balloon_stats.current_pages = xen_pv_domain()
725 ? min(xen_start_info->nr_pages - xen_released_pages, max_pfn)
726 : get_num_physpages();
728 balloon_stats.current_pages = get_num_physpages();
730 balloon_stats.target_pages = balloon_stats.current_pages;
731 balloon_stats.balloon_low = 0;
732 balloon_stats.balloon_high = 0;
733 balloon_stats.total_pages = balloon_stats.current_pages;
735 balloon_stats.schedule_delay = 1;
736 balloon_stats.max_schedule_delay = 32;
737 balloon_stats.retry_count = 1;
738 balloon_stats.max_retry_count = 4;
740 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
741 set_online_page_callback(&xen_online_page);
742 register_memory_notifier(&xen_memory_nb);
743 register_sysctl_table(xen_root);
746 #if defined(CONFIG_XEN_PV) && !defined(CONFIG_XEN_UNPOPULATED_ALLOC)
751 * Initialize the balloon with pages from the extra memory
752 * regions (see arch/x86/xen/setup.c).
754 for (i = 0; i < XEN_EXTRA_MEM_MAX_REGIONS; i++)
755 if (xen_extra_mem[i].n_pfns)
756 balloon_add_region(xen_extra_mem[i].start_pfn,
757 xen_extra_mem[i].n_pfns);
761 task = kthread_run(balloon_thread, NULL, "xen-balloon");
763 pr_err("xen-balloon thread could not be started, ballooning will not work!\n");
764 return PTR_ERR(task);
767 /* Init the xen-balloon driver. */
772 subsys_initcall(balloon_init);
774 static int __init balloon_wait_finish(void)
776 long credit, last_credit = 0;
777 unsigned long last_changed = 0;
782 /* PV guests don't need to wait. */
783 if (xen_pv_domain() || !current_credit())
786 pr_notice("Waiting for initial ballooning down having finished.\n");
788 while ((credit = current_credit()) < 0) {
789 if (credit != last_credit) {
790 last_changed = jiffies;
791 last_credit = credit;
793 if (balloon_state == BP_ECANCELED) {
794 pr_warn_once("Initial ballooning failed, %ld pages need to be freed.\n",
796 if (jiffies - last_changed >= HZ * balloon_boot_timeout)
797 panic("Initial ballooning failed!\n");
800 schedule_timeout_interruptible(HZ / 10);
803 pr_notice("Initial ballooning down finished.\n");
807 late_initcall_sync(balloon_wait_finish);