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/bootmem.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>
78 #undef MODULE_PARAM_PREFIX
79 #define MODULE_PARAM_PREFIX "xen."
81 static uint __read_mostly balloon_boot_timeout = 180;
82 module_param(balloon_boot_timeout, uint, 0444);
84 static int xen_hotplug_unpopulated;
86 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
91 static struct ctl_table balloon_table[] = {
93 .procname = "hotplug_unpopulated",
94 .data = &xen_hotplug_unpopulated,
95 .maxlen = sizeof(int),
97 .proc_handler = proc_dointvec_minmax,
104 static struct ctl_table balloon_root[] = {
106 .procname = "balloon",
108 .child = balloon_table,
113 static struct ctl_table xen_root[] = {
117 .child = balloon_root,
125 * Use one extent per PAGE_SIZE to avoid to break down the page into
128 #define EXTENT_ORDER (fls(XEN_PFN_PER_PAGE) - 1)
131 * balloon_thread() state:
133 * BP_DONE: done or nothing to do,
134 * BP_WAIT: wait to be rescheduled,
135 * BP_EAGAIN: error, go to sleep,
136 * BP_ECANCELED: error, balloon operation canceled.
139 static enum bp_state {
144 } balloon_state = BP_DONE;
146 /* Main waiting point for xen-balloon thread. */
147 static DECLARE_WAIT_QUEUE_HEAD(balloon_thread_wq);
149 static DEFINE_MUTEX(balloon_mutex);
151 struct balloon_stats balloon_stats;
152 EXPORT_SYMBOL_GPL(balloon_stats);
154 /* We increase/decrease in batches which fit in a page */
155 static xen_pfn_t frame_list[PAGE_SIZE / sizeof(xen_pfn_t)];
158 /* List of ballooned pages, threaded through the mem_map array. */
159 static LIST_HEAD(ballooned_pages);
160 static DECLARE_WAIT_QUEUE_HEAD(balloon_wq);
162 /* When ballooning out (allocating memory to return to Xen) we don't really
163 want the kernel to try too hard since that can trigger the oom killer. */
164 #define GFP_BALLOON \
165 (GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC)
167 static void scrub_page(struct page *page)
169 #ifdef CONFIG_XEN_SCRUB_PAGES
170 clear_highpage(page);
174 /* balloon_append: add the given page to the balloon. */
175 static void __balloon_append(struct page *page)
177 /* Lowmem is re-populated first, so highmem pages go at list tail. */
178 if (PageHighMem(page)) {
179 list_add_tail(&page->lru, &ballooned_pages);
180 balloon_stats.balloon_high++;
182 list_add(&page->lru, &ballooned_pages);
183 balloon_stats.balloon_low++;
185 wake_up(&balloon_wq);
188 static void balloon_append(struct page *page)
190 __balloon_append(page);
193 /* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
194 static struct page *balloon_retrieve(bool require_lowmem)
198 if (list_empty(&ballooned_pages))
201 page = list_entry(ballooned_pages.next, struct page, lru);
202 if (require_lowmem && PageHighMem(page))
204 list_del(&page->lru);
206 if (PageHighMem(page))
207 balloon_stats.balloon_high--;
209 balloon_stats.balloon_low--;
214 static struct page *balloon_next_page(struct page *page)
216 struct list_head *next = page->lru.next;
217 if (next == &ballooned_pages)
219 return list_entry(next, struct page, lru);
222 static void update_schedule(void)
224 if (balloon_state == BP_WAIT || balloon_state == BP_ECANCELED)
227 if (balloon_state == BP_DONE) {
228 balloon_stats.schedule_delay = 1;
229 balloon_stats.retry_count = 1;
233 ++balloon_stats.retry_count;
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 balloon_state = BP_ECANCELED;
243 balloon_stats.schedule_delay <<= 1;
245 if (balloon_stats.schedule_delay > balloon_stats.max_schedule_delay)
246 balloon_stats.schedule_delay = balloon_stats.max_schedule_delay;
248 balloon_state = BP_EAGAIN;
251 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
252 static void release_memory_resource(struct resource *resource)
258 * No need to reset region to identity mapped since we now
259 * know that no I/O can be in this region
261 release_resource(resource);
265 static struct resource *additional_memory_resource(phys_addr_t size)
267 struct resource *res;
270 res = kzalloc(sizeof(*res), GFP_KERNEL);
274 res->name = "System RAM";
275 res->flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
277 ret = allocate_resource(&iomem_resource, res,
279 PAGES_PER_SECTION * PAGE_SIZE, NULL, NULL);
281 pr_err("Cannot allocate new System RAM resource\n");
286 #ifdef CONFIG_SPARSEMEM
288 unsigned long limit = 1UL << (MAX_PHYSMEM_BITS - PAGE_SHIFT);
289 unsigned long pfn = res->start >> PAGE_SHIFT;
292 pr_err("New System RAM resource outside addressable RAM (%lu > %lu)\n",
294 release_memory_resource(res);
303 static enum bp_state reserve_additional_memory(void)
306 struct resource *resource;
308 unsigned long balloon_hotplug;
310 credit = balloon_stats.target_pages + balloon_stats.target_unpopulated
311 - balloon_stats.total_pages;
314 * Already hotplugged enough pages? Wait for them to be
320 balloon_hotplug = round_up(credit, PAGES_PER_SECTION);
322 resource = additional_memory_resource(balloon_hotplug * PAGE_SIZE);
326 nid = memory_add_physaddr_to_nid(resource->start);
328 #ifdef CONFIG_XEN_HAVE_PVMMU
330 * We don't support PV MMU when Linux and Xen is using
331 * different page granularity.
333 BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
336 * add_memory() will build page tables for the new memory so
337 * the p2m must contain invalid entries so the correct
338 * non-present PTEs will be written.
340 * If a failure occurs, the original (identity) p2m entries
341 * are not restored since this region is now known not to
342 * conflict with any devices.
344 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
345 unsigned long pfn, i;
347 pfn = PFN_DOWN(resource->start);
348 for (i = 0; i < balloon_hotplug; i++) {
349 if (!set_phys_to_machine(pfn + i, INVALID_P2M_ENTRY)) {
350 pr_warn("set_phys_to_machine() failed, no memory added\n");
358 * add_memory_resource() will call online_pages() which in its turn
359 * will call xen_online_page() callback causing deadlock if we don't
360 * release balloon_mutex here. Unlocking here is safe because the
361 * callers drop the mutex before trying again.
363 mutex_unlock(&balloon_mutex);
364 /* add_memory_resource() requires the device_hotplug lock */
365 lock_device_hotplug();
366 rc = add_memory_resource(nid, resource, memhp_auto_online);
367 unlock_device_hotplug();
368 mutex_lock(&balloon_mutex);
371 pr_warn("Cannot add additional memory (%i)\n", rc);
375 balloon_stats.total_pages += balloon_hotplug;
379 release_memory_resource(resource);
383 static void xen_online_page(struct page *page)
385 __online_page_set_limits(page);
387 mutex_lock(&balloon_mutex);
389 __balloon_append(page);
391 mutex_unlock(&balloon_mutex);
394 static int xen_memory_notifier(struct notifier_block *nb, unsigned long val, void *v)
396 if (val == MEM_ONLINE)
397 wake_up(&balloon_thread_wq);
402 static struct notifier_block xen_memory_nb = {
403 .notifier_call = xen_memory_notifier,
407 static enum bp_state reserve_additional_memory(void)
409 balloon_stats.target_pages = balloon_stats.current_pages +
410 balloon_stats.target_unpopulated;
413 #endif /* CONFIG_XEN_BALLOON_MEMORY_HOTPLUG */
415 static long current_credit(void)
417 return balloon_stats.target_pages - balloon_stats.current_pages;
420 static bool balloon_is_inflated(void)
422 return balloon_stats.balloon_low || balloon_stats.balloon_high;
425 static enum bp_state increase_reservation(unsigned long nr_pages)
430 struct xen_memory_reservation reservation = {
432 .extent_order = EXTENT_ORDER,
436 if (nr_pages > ARRAY_SIZE(frame_list))
437 nr_pages = ARRAY_SIZE(frame_list);
439 page = list_first_entry_or_null(&ballooned_pages, struct page, lru);
440 for (i = 0; i < nr_pages; i++) {
446 /* XENMEM_populate_physmap requires a PFN based on Xen
449 frame_list[i] = page_to_xen_pfn(page);
450 page = balloon_next_page(page);
453 set_xen_guest_handle(reservation.extent_start, frame_list);
454 reservation.nr_extents = nr_pages;
455 rc = HYPERVISOR_memory_op(XENMEM_populate_physmap, &reservation);
459 for (i = 0; i < rc; i++) {
460 page = balloon_retrieve(false);
461 BUG_ON(page == NULL);
463 #ifdef CONFIG_XEN_HAVE_PVMMU
465 * We don't support PV MMU when Linux and Xen is using
466 * different page granularity.
468 BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
470 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
471 unsigned long pfn = page_to_pfn(page);
473 set_phys_to_machine(pfn, frame_list[i]);
475 /* Link back into the page tables if not highmem. */
476 if (!PageHighMem(page)) {
478 ret = HYPERVISOR_update_va_mapping(
479 (unsigned long)__va(pfn << PAGE_SHIFT),
480 mfn_pte(frame_list[i], PAGE_KERNEL),
487 /* Relinquish the page back to the allocator. */
488 free_reserved_page(page);
491 balloon_stats.current_pages += rc;
496 static enum bp_state decrease_reservation(unsigned long nr_pages, gfp_t gfp)
498 enum bp_state state = BP_DONE;
500 struct page *page, *tmp;
502 struct xen_memory_reservation reservation = {
504 .extent_order = EXTENT_ORDER,
509 if (nr_pages > ARRAY_SIZE(frame_list))
510 nr_pages = ARRAY_SIZE(frame_list);
512 for (i = 0; i < nr_pages; i++) {
513 page = alloc_page(gfp);
519 adjust_managed_page_count(page, -1);
521 list_add(&page->lru, &pages);
525 * Ensure that ballooned highmem pages don't have kmaps.
527 * Do this before changing the p2m as kmap_flush_unused()
528 * reads PTEs to obtain pages (and hence needs the original
534 * Setup the frame, update direct mapping, invalidate P2M,
535 * and add to balloon.
538 list_for_each_entry_safe(page, tmp, &pages, lru) {
539 /* XENMEM_decrease_reservation requires a GFN */
540 frame_list[i++] = xen_page_to_gfn(page);
542 #ifdef CONFIG_XEN_HAVE_PVMMU
544 * We don't support PV MMU when Linux and Xen is using
545 * different page granularity.
547 BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
549 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
550 unsigned long pfn = page_to_pfn(page);
552 if (!PageHighMem(page)) {
553 ret = HYPERVISOR_update_va_mapping(
554 (unsigned long)__va(pfn << PAGE_SHIFT),
558 __set_phys_to_machine(pfn, INVALID_P2M_ENTRY);
561 list_del(&page->lru);
563 balloon_append(page);
568 set_xen_guest_handle(reservation.extent_start, frame_list);
569 reservation.nr_extents = nr_pages;
570 ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation, &reservation);
571 BUG_ON(ret != nr_pages);
573 balloon_stats.current_pages -= nr_pages;
579 * Stop waiting if either state is BP_DONE and ballooning action is
580 * needed, or if the credit has changed while state is not BP_DONE.
582 static bool balloon_thread_cond(long credit)
584 if (balloon_state == BP_DONE)
587 return current_credit() != credit || kthread_should_stop();
591 * As this is a kthread it is guaranteed to run as a single instance only.
592 * We may of course race updates of the target counts (which are protected
593 * by the balloon lock), or with changes to the Xen hard limit, but we will
594 * recover from these in time.
596 static int balloon_thread(void *unused)
599 unsigned long timeout;
603 switch (balloon_state) {
609 timeout = balloon_stats.schedule_delay * HZ;
616 credit = current_credit();
618 wait_event_freezable_timeout(balloon_thread_wq,
619 balloon_thread_cond(credit), timeout);
621 if (kthread_should_stop())
624 mutex_lock(&balloon_mutex);
626 credit = current_credit();
629 if (balloon_is_inflated())
630 balloon_state = increase_reservation(credit);
632 balloon_state = reserve_additional_memory();
638 n_pages = min(-credit, si_mem_available());
639 balloon_state = decrease_reservation(n_pages,
641 if (balloon_state == BP_DONE && n_pages != -credit &&
642 n_pages < totalreserve_pages)
643 balloon_state = BP_EAGAIN;
648 mutex_unlock(&balloon_mutex);
654 /* Resets the Xen limit, sets new target, and kicks off processing. */
655 void balloon_set_new_target(unsigned long target)
657 /* No need for lock. Not read-modify-write updates. */
658 balloon_stats.target_pages = target;
659 wake_up(&balloon_thread_wq);
661 EXPORT_SYMBOL_GPL(balloon_set_new_target);
663 static int add_ballooned_pages(int nr_pages)
667 if (xen_hotplug_unpopulated) {
668 st = reserve_additional_memory();
669 if (st != BP_ECANCELED) {
672 mutex_unlock(&balloon_mutex);
673 rc = wait_event_interruptible(balloon_wq,
674 !list_empty(&ballooned_pages));
675 mutex_lock(&balloon_mutex);
676 return rc ? -ENOMEM : 0;
680 if (si_mem_available() < nr_pages)
683 st = decrease_reservation(nr_pages, GFP_USER);
691 * alloc_xenballooned_pages - get pages that have been ballooned out
692 * @nr_pages: Number of pages to get
693 * @pages: pages returned
694 * @return 0 on success, error otherwise
696 int alloc_xenballooned_pages(int nr_pages, struct page **pages)
702 mutex_lock(&balloon_mutex);
704 balloon_stats.target_unpopulated += nr_pages;
706 while (pgno < nr_pages) {
707 page = balloon_retrieve(true);
709 pages[pgno++] = page;
710 #ifdef CONFIG_XEN_HAVE_PVMMU
712 * We don't support PV MMU when Linux and Xen is using
713 * different page granularity.
715 BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
717 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
718 ret = xen_alloc_p2m_entry(page_to_pfn(page));
724 ret = add_ballooned_pages(nr_pages - pgno);
729 mutex_unlock(&balloon_mutex);
732 mutex_unlock(&balloon_mutex);
733 free_xenballooned_pages(pgno, pages);
735 * NB: free_xenballooned_pages will only subtract pgno pages, but since
736 * target_unpopulated is incremented with nr_pages at the start we need
737 * to remove the remaining ones also, or accounting will be screwed.
739 balloon_stats.target_unpopulated -= nr_pages - pgno;
742 EXPORT_SYMBOL(alloc_xenballooned_pages);
745 * free_xenballooned_pages - return pages retrieved with get_ballooned_pages
746 * @nr_pages: Number of pages
747 * @pages: pages to return
749 void free_xenballooned_pages(int nr_pages, struct page **pages)
753 mutex_lock(&balloon_mutex);
755 for (i = 0; i < nr_pages; i++) {
757 balloon_append(pages[i]);
760 balloon_stats.target_unpopulated -= nr_pages;
762 /* The balloon may be too large now. Shrink it if needed. */
763 if (current_credit())
764 wake_up(&balloon_thread_wq);
766 mutex_unlock(&balloon_mutex);
768 EXPORT_SYMBOL(free_xenballooned_pages);
771 static void __init balloon_add_region(unsigned long start_pfn,
774 unsigned long pfn, extra_pfn_end;
778 * If the amount of usable memory has been limited (e.g., with
779 * the 'mem' command line parameter), don't add pages beyond
782 extra_pfn_end = min(max_pfn, start_pfn + pages);
784 for (pfn = start_pfn; pfn < extra_pfn_end; pfn++) {
785 page = pfn_to_page(pfn);
786 /* totalram_pages and totalhigh_pages do not
787 include the boot-time balloon extension, so
788 don't subtract from it. */
789 __balloon_append(page);
792 balloon_stats.total_pages += extra_pfn_end - start_pfn;
796 static int __init balloon_init(void)
798 struct task_struct *task;
803 pr_info("Initialising balloon driver\n");
806 balloon_stats.current_pages = xen_pv_domain()
807 ? min(xen_start_info->nr_pages - xen_released_pages, max_pfn)
808 : get_num_physpages();
810 balloon_stats.current_pages = get_num_physpages();
812 balloon_stats.target_pages = balloon_stats.current_pages;
813 balloon_stats.balloon_low = 0;
814 balloon_stats.balloon_high = 0;
815 balloon_stats.total_pages = balloon_stats.current_pages;
817 balloon_stats.schedule_delay = 1;
818 balloon_stats.max_schedule_delay = 32;
819 balloon_stats.retry_count = 1;
820 balloon_stats.max_retry_count = 4;
822 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
823 set_online_page_callback(&xen_online_page);
824 register_memory_notifier(&xen_memory_nb);
825 register_sysctl_table(xen_root);
833 * Initialize the balloon with pages from the extra memory
834 * regions (see arch/x86/xen/setup.c).
836 for (i = 0; i < XEN_EXTRA_MEM_MAX_REGIONS; i++)
837 if (xen_extra_mem[i].n_pfns)
838 balloon_add_region(xen_extra_mem[i].start_pfn,
839 xen_extra_mem[i].n_pfns);
843 task = kthread_run(balloon_thread, NULL, "xen-balloon");
845 pr_err("xen-balloon thread could not be started, ballooning will not work!\n");
846 return PTR_ERR(task);
849 /* Init the xen-balloon driver. */
854 subsys_initcall(balloon_init);
856 static int __init balloon_wait_finish(void)
858 long credit, last_credit = 0;
859 unsigned long last_changed = 0;
864 /* PV guests don't need to wait. */
865 if (xen_pv_domain() || !current_credit())
868 pr_notice("Waiting for initial ballooning down having finished.\n");
870 while ((credit = current_credit()) < 0) {
871 if (credit != last_credit) {
872 last_changed = jiffies;
873 last_credit = credit;
875 if (balloon_state == BP_ECANCELED) {
876 pr_warn_once("Initial ballooning failed, %ld pages need to be freed.\n",
878 if (jiffies - last_changed >= HZ * balloon_boot_timeout)
879 panic("Initial ballooning failed!\n");
882 schedule_timeout_interruptible(HZ / 10);
885 pr_notice("Initial ballooning down finished.\n");
889 late_initcall_sync(balloon_wait_finish);