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>
64 #include <asm/pgalloc.h>
65 #include <asm/pgtable.h>
68 #include <asm/xen/hypervisor.h>
69 #include <asm/xen/hypercall.h>
72 #include <xen/interface/xen.h>
73 #include <xen/interface/memory.h>
74 #include <xen/balloon.h>
75 #include <xen/features.h>
77 #include <xen/mem-reservation.h>
79 #undef MODULE_PARAM_PREFIX
80 #define MODULE_PARAM_PREFIX "xen."
82 static uint __read_mostly balloon_boot_timeout = 180;
83 module_param(balloon_boot_timeout, uint, 0444);
85 static int xen_hotplug_unpopulated;
87 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
89 static struct ctl_table balloon_table[] = {
91 .procname = "hotplug_unpopulated",
92 .data = &xen_hotplug_unpopulated,
93 .maxlen = sizeof(int),
95 .proc_handler = proc_dointvec_minmax,
96 .extra1 = SYSCTL_ZERO,
102 static struct ctl_table balloon_root[] = {
104 .procname = "balloon",
106 .child = balloon_table,
111 static struct ctl_table xen_root[] = {
115 .child = balloon_root,
123 * Use one extent per PAGE_SIZE to avoid to break down the page into
126 #define EXTENT_ORDER (fls(XEN_PFN_PER_PAGE) - 1)
129 * balloon_thread() state:
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.
137 static enum bp_state {
142 } balloon_state = BP_DONE;
144 /* Main waiting point for xen-balloon thread. */
145 static DECLARE_WAIT_QUEUE_HEAD(balloon_thread_wq);
147 static DEFINE_MUTEX(balloon_mutex);
149 struct balloon_stats balloon_stats;
150 EXPORT_SYMBOL_GPL(balloon_stats);
152 /* We increase/decrease in batches which fit in a page */
153 static xen_pfn_t frame_list[PAGE_SIZE / sizeof(xen_pfn_t)];
156 /* List of ballooned pages, threaded through the mem_map array. */
157 static LIST_HEAD(ballooned_pages);
158 static DECLARE_WAIT_QUEUE_HEAD(balloon_wq);
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)
165 /* balloon_append: add the given page to the balloon. */
166 static void balloon_append(struct page *page)
168 __SetPageOffline(page);
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++;
175 list_add(&page->lru, &ballooned_pages);
176 balloon_stats.balloon_low++;
178 wake_up(&balloon_wq);
181 /* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
182 static struct page *balloon_retrieve(bool require_lowmem)
186 if (list_empty(&ballooned_pages))
189 page = list_entry(ballooned_pages.next, struct page, lru);
190 if (require_lowmem && PageHighMem(page))
192 list_del(&page->lru);
194 if (PageHighMem(page))
195 balloon_stats.balloon_high--;
197 balloon_stats.balloon_low--;
199 __ClearPageOffline(page);
203 static struct page *balloon_next_page(struct page *page)
205 struct list_head *next = page->lru.next;
206 if (next == &ballooned_pages)
208 return list_entry(next, struct page, lru);
211 static void update_schedule(void)
213 if (balloon_state == BP_WAIT || balloon_state == BP_ECANCELED)
216 if (balloon_state == BP_DONE) {
217 balloon_stats.schedule_delay = 1;
218 balloon_stats.retry_count = 1;
222 ++balloon_stats.retry_count;
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;
232 balloon_stats.schedule_delay <<= 1;
234 if (balloon_stats.schedule_delay > balloon_stats.max_schedule_delay)
235 balloon_stats.schedule_delay = balloon_stats.max_schedule_delay;
237 balloon_state = BP_EAGAIN;
240 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
241 static void release_memory_resource(struct resource *resource)
247 * No need to reset region to identity mapped since we now
248 * know that no I/O can be in this region
250 release_resource(resource);
254 static struct resource *additional_memory_resource(phys_addr_t size)
256 struct resource *res;
259 res = kzalloc(sizeof(*res), GFP_KERNEL);
263 res->name = "System RAM";
264 res->flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
266 ret = allocate_resource(&iomem_resource, res,
268 PAGES_PER_SECTION * PAGE_SIZE, NULL, NULL);
270 pr_err("Cannot allocate new System RAM resource\n");
275 #ifdef CONFIG_SPARSEMEM
277 unsigned long limit = 1UL << (MAX_PHYSMEM_BITS - PAGE_SHIFT);
278 unsigned long pfn = res->start >> PAGE_SHIFT;
281 pr_err("New System RAM resource outside addressable RAM (%lu > %lu)\n",
283 release_memory_resource(res);
292 static enum bp_state reserve_additional_memory(void)
295 struct resource *resource;
297 unsigned long balloon_hotplug;
299 credit = balloon_stats.target_pages + balloon_stats.target_unpopulated
300 - balloon_stats.total_pages;
303 * Already hotplugged enough pages? Wait for them to be
309 balloon_hotplug = round_up(credit, PAGES_PER_SECTION);
311 resource = additional_memory_resource(balloon_hotplug * PAGE_SIZE);
315 nid = memory_add_physaddr_to_nid(resource->start);
317 #ifdef CONFIG_XEN_HAVE_PVMMU
319 * We don't support PV MMU when Linux and Xen is using
320 * different page granularity.
322 BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
325 * add_memory() will build page tables for the new memory so
326 * the p2m must contain invalid entries so the correct
327 * non-present PTEs will be written.
329 * If a failure occurs, the original (identity) p2m entries
330 * are not restored since this region is now known not to
331 * conflict with any devices.
333 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
334 unsigned long pfn, i;
336 pfn = PFN_DOWN(resource->start);
337 for (i = 0; i < balloon_hotplug; i++) {
338 if (!set_phys_to_machine(pfn + i, INVALID_P2M_ENTRY)) {
339 pr_warn("set_phys_to_machine() failed, no memory added\n");
347 * add_memory_resource() will call online_pages() which in its turn
348 * will call xen_online_page() callback causing deadlock if we don't
349 * release balloon_mutex here. Unlocking here is safe because the
350 * callers drop the mutex before trying again.
352 mutex_unlock(&balloon_mutex);
353 /* add_memory_resource() requires the device_hotplug lock */
354 lock_device_hotplug();
355 rc = add_memory_resource(nid, resource);
356 unlock_device_hotplug();
357 mutex_lock(&balloon_mutex);
360 pr_warn("Cannot add additional memory (%i)\n", rc);
364 balloon_stats.total_pages += balloon_hotplug;
368 release_memory_resource(resource);
372 static void xen_online_page(struct page *page, unsigned int order)
374 unsigned long i, size = (1 << order);
375 unsigned long start_pfn = page_to_pfn(page);
378 pr_debug("Online %lu pages starting at pfn 0x%lx\n", size, start_pfn);
379 mutex_lock(&balloon_mutex);
380 for (i = 0; i < size; i++) {
381 p = pfn_to_page(start_pfn + i);
382 __online_page_set_limits(p);
385 mutex_unlock(&balloon_mutex);
388 static int xen_memory_notifier(struct notifier_block *nb, unsigned long val, void *v)
390 if (val == MEM_ONLINE)
391 wake_up(&balloon_thread_wq);
396 static struct notifier_block xen_memory_nb = {
397 .notifier_call = xen_memory_notifier,
401 static enum bp_state reserve_additional_memory(void)
403 balloon_stats.target_pages = balloon_stats.current_pages +
404 balloon_stats.target_unpopulated;
407 #endif /* CONFIG_XEN_BALLOON_MEMORY_HOTPLUG */
409 static long current_credit(void)
411 return balloon_stats.target_pages - balloon_stats.current_pages;
414 static bool balloon_is_inflated(void)
416 return balloon_stats.balloon_low || balloon_stats.balloon_high;
419 static enum bp_state increase_reservation(unsigned long nr_pages)
425 if (nr_pages > ARRAY_SIZE(frame_list))
426 nr_pages = ARRAY_SIZE(frame_list);
428 page = list_first_entry_or_null(&ballooned_pages, struct page, lru);
429 for (i = 0; i < nr_pages; i++) {
435 frame_list[i] = page_to_xen_pfn(page);
436 page = balloon_next_page(page);
439 rc = xenmem_reservation_increase(nr_pages, frame_list);
443 for (i = 0; i < rc; i++) {
444 page = balloon_retrieve(false);
445 BUG_ON(page == NULL);
447 xenmem_reservation_va_mapping_update(1, &page, &frame_list[i]);
449 /* Relinquish the page back to the allocator. */
450 free_reserved_page(page);
453 balloon_stats.current_pages += rc;
458 static enum bp_state decrease_reservation(unsigned long nr_pages, gfp_t gfp)
460 enum bp_state state = BP_DONE;
462 struct page *page, *tmp;
466 if (nr_pages > ARRAY_SIZE(frame_list))
467 nr_pages = ARRAY_SIZE(frame_list);
469 for (i = 0; i < nr_pages; i++) {
470 page = alloc_page(gfp);
476 adjust_managed_page_count(page, -1);
477 xenmem_reservation_scrub_page(page);
478 list_add(&page->lru, &pages);
482 * Ensure that ballooned highmem pages don't have kmaps.
484 * Do this before changing the p2m as kmap_flush_unused()
485 * reads PTEs to obtain pages (and hence needs the original
491 * Setup the frame, update direct mapping, invalidate P2M,
492 * and add to balloon.
495 list_for_each_entry_safe(page, tmp, &pages, lru) {
496 frame_list[i++] = xen_page_to_gfn(page);
498 xenmem_reservation_va_mapping_reset(1, &page);
500 list_del(&page->lru);
502 balloon_append(page);
507 ret = xenmem_reservation_decrease(nr_pages, frame_list);
508 BUG_ON(ret != nr_pages);
510 balloon_stats.current_pages -= nr_pages;
516 * Stop waiting if either state is BP_DONE and ballooning action is
517 * needed, or if the credit has changed while state is not BP_DONE.
519 static bool balloon_thread_cond(long credit)
521 if (balloon_state == BP_DONE)
524 return current_credit() != credit || kthread_should_stop();
528 * As this is a kthread it is guaranteed to run as a single instance only.
529 * We may of course race updates of the target counts (which are protected
530 * by the balloon lock), or with changes to the Xen hard limit, but we will
531 * recover from these in time.
533 static int balloon_thread(void *unused)
536 unsigned long timeout;
540 switch (balloon_state) {
546 timeout = balloon_stats.schedule_delay * HZ;
553 credit = current_credit();
555 wait_event_freezable_timeout(balloon_thread_wq,
556 balloon_thread_cond(credit), timeout);
558 if (kthread_should_stop())
561 mutex_lock(&balloon_mutex);
563 credit = current_credit();
566 if (balloon_is_inflated())
567 balloon_state = increase_reservation(credit);
569 balloon_state = reserve_additional_memory();
575 n_pages = min(-credit, si_mem_available());
576 balloon_state = decrease_reservation(n_pages,
578 if (balloon_state == BP_DONE && n_pages != -credit &&
579 n_pages < totalreserve_pages)
580 balloon_state = BP_EAGAIN;
585 mutex_unlock(&balloon_mutex);
591 /* Resets the Xen limit, sets new target, and kicks off processing. */
592 void balloon_set_new_target(unsigned long target)
594 /* No need for lock. Not read-modify-write updates. */
595 balloon_stats.target_pages = target;
596 wake_up(&balloon_thread_wq);
598 EXPORT_SYMBOL_GPL(balloon_set_new_target);
600 static int add_ballooned_pages(int nr_pages)
604 if (xen_hotplug_unpopulated) {
605 st = reserve_additional_memory();
606 if (st != BP_ECANCELED) {
609 mutex_unlock(&balloon_mutex);
610 rc = wait_event_interruptible(balloon_wq,
611 !list_empty(&ballooned_pages));
612 mutex_lock(&balloon_mutex);
613 return rc ? -ENOMEM : 0;
617 if (si_mem_available() < nr_pages)
620 st = decrease_reservation(nr_pages, GFP_USER);
628 * alloc_xenballooned_pages - get pages that have been ballooned out
629 * @nr_pages: Number of pages to get
630 * @pages: pages returned
631 * @return 0 on success, error otherwise
633 int alloc_xenballooned_pages(int nr_pages, struct page **pages)
639 mutex_lock(&balloon_mutex);
641 balloon_stats.target_unpopulated += nr_pages;
643 while (pgno < nr_pages) {
644 page = balloon_retrieve(true);
646 pages[pgno++] = page;
647 #ifdef CONFIG_XEN_HAVE_PVMMU
649 * We don't support PV MMU when Linux and Xen is using
650 * different page granularity.
652 BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
654 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
655 ret = xen_alloc_p2m_entry(page_to_pfn(page));
661 ret = add_ballooned_pages(nr_pages - pgno);
666 mutex_unlock(&balloon_mutex);
669 mutex_unlock(&balloon_mutex);
670 free_xenballooned_pages(pgno, pages);
672 * NB: free_xenballooned_pages will only subtract pgno pages, but since
673 * target_unpopulated is incremented with nr_pages at the start we need
674 * to remove the remaining ones also, or accounting will be screwed.
676 balloon_stats.target_unpopulated -= nr_pages - pgno;
679 EXPORT_SYMBOL(alloc_xenballooned_pages);
682 * free_xenballooned_pages - return pages retrieved with get_ballooned_pages
683 * @nr_pages: Number of pages
684 * @pages: pages to return
686 void free_xenballooned_pages(int nr_pages, struct page **pages)
690 mutex_lock(&balloon_mutex);
692 for (i = 0; i < nr_pages; i++) {
694 balloon_append(pages[i]);
697 balloon_stats.target_unpopulated -= nr_pages;
699 /* The balloon may be too large now. Shrink it if needed. */
700 if (current_credit())
701 wake_up(&balloon_thread_wq);
703 mutex_unlock(&balloon_mutex);
705 EXPORT_SYMBOL(free_xenballooned_pages);
708 static void __init balloon_add_region(unsigned long start_pfn,
711 unsigned long pfn, extra_pfn_end;
714 * If the amount of usable memory has been limited (e.g., with
715 * the 'mem' command line parameter), don't add pages beyond
718 extra_pfn_end = min(max_pfn, start_pfn + pages);
720 for (pfn = start_pfn; pfn < extra_pfn_end; pfn++) {
721 /* totalram_pages and totalhigh_pages do not
722 include the boot-time balloon extension, so
723 don't subtract from it. */
724 balloon_append(pfn_to_page(pfn));
727 balloon_stats.total_pages += extra_pfn_end - start_pfn;
731 static int __init balloon_init(void)
733 struct task_struct *task;
738 pr_info("Initialising balloon driver\n");
741 balloon_stats.current_pages = xen_pv_domain()
742 ? min(xen_start_info->nr_pages - xen_released_pages, max_pfn)
743 : get_num_physpages();
745 balloon_stats.current_pages = 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;
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 = 4;
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);
768 * Initialize the balloon with pages from the extra memory
769 * regions (see arch/x86/xen/setup.c).
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);
778 task = kthread_run(balloon_thread, NULL, "xen-balloon");
780 pr_err("xen-balloon thread could not be started, ballooning will not work!\n");
781 return PTR_ERR(task);
784 /* Init the xen-balloon driver. */
789 subsys_initcall(balloon_init);
791 static int __init balloon_wait_finish(void)
793 long credit, last_credit = 0;
794 unsigned long last_changed = 0;
799 /* PV guests don't need to wait. */
800 if (xen_pv_domain() || !current_credit())
803 pr_notice("Waiting for initial ballooning down having finished.\n");
805 while ((credit = current_credit()) < 0) {
806 if (credit != last_credit) {
807 last_changed = jiffies;
808 last_credit = credit;
810 if (balloon_state == BP_ECANCELED) {
811 pr_warn_once("Initial ballooning failed, %ld pages need to be freed.\n",
813 if (jiffies - last_changed >= HZ * balloon_boot_timeout)
814 panic("Initial ballooning failed!\n");
817 schedule_timeout_interruptible(HZ / 10);
820 pr_notice("Initial ballooning down finished.\n");
824 late_initcall_sync(balloon_wait_finish);