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/errno.h>
45 #include <linux/module.h>
47 #include <linux/bootmem.h>
48 #include <linux/pagemap.h>
49 #include <linux/highmem.h>
50 #include <linux/mutex.h>
51 #include <linux/list.h>
52 #include <linux/gfp.h>
53 #include <linux/notifier.h>
54 #include <linux/memory.h>
55 #include <linux/memory_hotplug.h>
56 #include <linux/percpu-defs.h>
57 #include <linux/slab.h>
58 #include <linux/sysctl.h>
61 #include <asm/pgalloc.h>
62 #include <asm/pgtable.h>
65 #include <asm/xen/hypervisor.h>
66 #include <asm/xen/hypercall.h>
69 #include <xen/interface/xen.h>
70 #include <xen/interface/memory.h>
71 #include <xen/balloon.h>
72 #include <xen/features.h>
75 static int xen_hotplug_unpopulated;
77 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
82 static struct ctl_table balloon_table[] = {
84 .procname = "hotplug_unpopulated",
85 .data = &xen_hotplug_unpopulated,
86 .maxlen = sizeof(int),
88 .proc_handler = proc_dointvec_minmax,
95 static struct ctl_table balloon_root[] = {
97 .procname = "balloon",
99 .child = balloon_table,
104 static struct ctl_table xen_root[] = {
108 .child = balloon_root,
116 * Use one extent per PAGE_SIZE to avoid to break down the page into
119 #define EXTENT_ORDER (fls(XEN_PFN_PER_PAGE) - 1)
122 * balloon_process() state:
124 * BP_DONE: done or nothing to do,
125 * BP_WAIT: wait to be rescheduled,
126 * BP_EAGAIN: error, go to sleep,
127 * BP_ECANCELED: error, balloon operation canceled.
138 static DEFINE_MUTEX(balloon_mutex);
140 struct balloon_stats balloon_stats;
141 EXPORT_SYMBOL_GPL(balloon_stats);
143 /* We increase/decrease in batches which fit in a page */
144 static xen_pfn_t frame_list[PAGE_SIZE / sizeof(xen_pfn_t)];
147 /* List of ballooned pages, threaded through the mem_map array. */
148 static LIST_HEAD(ballooned_pages);
149 static DECLARE_WAIT_QUEUE_HEAD(balloon_wq);
151 /* Main work function, always executed in process context. */
152 static void balloon_process(struct work_struct *work);
153 static DECLARE_DELAYED_WORK(balloon_worker, balloon_process);
155 /* When ballooning out (allocating memory to return to Xen) we don't really
156 want the kernel to try too hard since that can trigger the oom killer. */
157 #define GFP_BALLOON \
158 (GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC)
160 static void scrub_page(struct page *page)
162 #ifdef CONFIG_XEN_SCRUB_PAGES
163 clear_highpage(page);
167 /* balloon_append: add the given page to the balloon. */
168 static void __balloon_append(struct page *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 static void balloon_append(struct page *page)
183 __balloon_append(page);
184 adjust_managed_page_count(page, -1);
187 /* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
188 static struct page *balloon_retrieve(bool require_lowmem)
192 if (list_empty(&ballooned_pages))
195 page = list_entry(ballooned_pages.next, struct page, lru);
196 if (require_lowmem && PageHighMem(page))
198 list_del(&page->lru);
200 if (PageHighMem(page))
201 balloon_stats.balloon_high--;
203 balloon_stats.balloon_low--;
205 adjust_managed_page_count(page, 1);
210 static struct page *balloon_next_page(struct page *page)
212 struct list_head *next = page->lru.next;
213 if (next == &ballooned_pages)
215 return list_entry(next, struct page, lru);
218 static enum bp_state update_schedule(enum bp_state state)
220 if (state == BP_WAIT)
223 if (state == BP_ECANCELED)
226 if (state == BP_DONE) {
227 balloon_stats.schedule_delay = 1;
228 balloon_stats.retry_count = 1;
232 ++balloon_stats.retry_count;
234 if (balloon_stats.max_retry_count != RETRY_UNLIMITED &&
235 balloon_stats.retry_count > balloon_stats.max_retry_count) {
236 balloon_stats.schedule_delay = 1;
237 balloon_stats.retry_count = 1;
241 balloon_stats.schedule_delay <<= 1;
243 if (balloon_stats.schedule_delay > balloon_stats.max_schedule_delay)
244 balloon_stats.schedule_delay = balloon_stats.max_schedule_delay;
249 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
250 static void release_memory_resource(struct resource *resource)
256 * No need to reset region to identity mapped since we now
257 * know that no I/O can be in this region
259 release_resource(resource);
263 static struct resource *additional_memory_resource(phys_addr_t size)
265 struct resource *res;
268 res = kzalloc(sizeof(*res), GFP_KERNEL);
272 res->name = "System RAM";
273 res->flags = IORESOURCE_MEM | IORESOURCE_BUSY;
275 ret = allocate_resource(&iomem_resource, res,
277 PAGES_PER_SECTION * PAGE_SIZE, NULL, NULL);
279 pr_err("Cannot allocate new System RAM resource\n");
284 #ifdef CONFIG_SPARSEMEM
286 unsigned long limit = 1UL << (MAX_PHYSMEM_BITS - PAGE_SHIFT);
287 unsigned long pfn = res->start >> PAGE_SHIFT;
290 pr_err("New System RAM resource outside addressable RAM (%lu > %lu)\n",
292 release_memory_resource(res);
301 static enum bp_state reserve_additional_memory(void)
304 struct resource *resource;
306 unsigned long balloon_hotplug;
308 credit = balloon_stats.target_pages + balloon_stats.target_unpopulated
309 - balloon_stats.total_pages;
312 * Already hotplugged enough pages? Wait for them to be
318 balloon_hotplug = round_up(credit, PAGES_PER_SECTION);
320 resource = additional_memory_resource(balloon_hotplug * PAGE_SIZE);
324 nid = memory_add_physaddr_to_nid(resource->start);
326 #ifdef CONFIG_XEN_HAVE_PVMMU
328 * We don't support PV MMU when Linux and Xen is using
329 * different page granularity.
331 BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
334 * add_memory() will build page tables for the new memory so
335 * the p2m must contain invalid entries so the correct
336 * non-present PTEs will be written.
338 * If a failure occurs, the original (identity) p2m entries
339 * are not restored since this region is now known not to
340 * conflict with any devices.
342 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
343 unsigned long pfn, i;
345 pfn = PFN_DOWN(resource->start);
346 for (i = 0; i < balloon_hotplug; i++) {
347 if (!set_phys_to_machine(pfn + i, INVALID_P2M_ENTRY)) {
348 pr_warn("set_phys_to_machine() failed, no memory added\n");
355 rc = add_memory_resource(nid, resource);
357 pr_warn("Cannot add additional memory (%i)\n", rc);
361 balloon_stats.total_pages += balloon_hotplug;
365 release_memory_resource(resource);
369 static void xen_online_page(struct page *page)
371 __online_page_set_limits(page);
373 mutex_lock(&balloon_mutex);
375 __balloon_append(page);
377 mutex_unlock(&balloon_mutex);
380 static int xen_memory_notifier(struct notifier_block *nb, unsigned long val, void *v)
382 if (val == MEM_ONLINE)
383 schedule_delayed_work(&balloon_worker, 0);
388 static struct notifier_block xen_memory_nb = {
389 .notifier_call = xen_memory_notifier,
393 static enum bp_state reserve_additional_memory(void)
395 balloon_stats.target_pages = balloon_stats.current_pages;
398 #endif /* CONFIG_XEN_BALLOON_MEMORY_HOTPLUG */
400 static long current_credit(void)
402 return balloon_stats.target_pages - balloon_stats.current_pages;
405 static bool balloon_is_inflated(void)
407 return balloon_stats.balloon_low || balloon_stats.balloon_high;
410 static enum bp_state increase_reservation(unsigned long nr_pages)
415 struct xen_memory_reservation reservation = {
417 .extent_order = EXTENT_ORDER,
421 if (nr_pages > ARRAY_SIZE(frame_list))
422 nr_pages = ARRAY_SIZE(frame_list);
424 page = list_first_entry_or_null(&ballooned_pages, struct page, lru);
425 for (i = 0; i < nr_pages; i++) {
431 /* XENMEM_populate_physmap requires a PFN based on Xen
434 frame_list[i] = page_to_xen_pfn(page);
435 page = balloon_next_page(page);
438 set_xen_guest_handle(reservation.extent_start, frame_list);
439 reservation.nr_extents = nr_pages;
440 rc = HYPERVISOR_memory_op(XENMEM_populate_physmap, &reservation);
444 for (i = 0; i < rc; i++) {
445 page = balloon_retrieve(false);
446 BUG_ON(page == NULL);
448 #ifdef CONFIG_XEN_HAVE_PVMMU
450 * We don't support PV MMU when Linux and Xen is using
451 * different page granularity.
453 BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
455 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
456 unsigned long pfn = page_to_pfn(page);
458 set_phys_to_machine(pfn, frame_list[i]);
460 /* Link back into the page tables if not highmem. */
461 if (!PageHighMem(page)) {
463 ret = HYPERVISOR_update_va_mapping(
464 (unsigned long)__va(pfn << PAGE_SHIFT),
465 mfn_pte(frame_list[i], PAGE_KERNEL),
472 /* Relinquish the page back to the allocator. */
473 __free_reserved_page(page);
476 balloon_stats.current_pages += rc;
481 static enum bp_state decrease_reservation(unsigned long nr_pages, gfp_t gfp)
483 enum bp_state state = BP_DONE;
485 struct page *page, *tmp;
487 struct xen_memory_reservation reservation = {
489 .extent_order = EXTENT_ORDER,
494 if (nr_pages > ARRAY_SIZE(frame_list))
495 nr_pages = ARRAY_SIZE(frame_list);
497 for (i = 0; i < nr_pages; i++) {
498 page = alloc_page(gfp);
505 list_add(&page->lru, &pages);
509 * Ensure that ballooned highmem pages don't have kmaps.
511 * Do this before changing the p2m as kmap_flush_unused()
512 * reads PTEs to obtain pages (and hence needs the original
518 * Setup the frame, update direct mapping, invalidate P2M,
519 * and add to balloon.
522 list_for_each_entry_safe(page, tmp, &pages, lru) {
523 /* XENMEM_decrease_reservation requires a GFN */
524 frame_list[i++] = xen_page_to_gfn(page);
526 #ifdef CONFIG_XEN_HAVE_PVMMU
528 * We don't support PV MMU when Linux and Xen is using
529 * different page granularity.
531 BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
533 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
534 unsigned long pfn = page_to_pfn(page);
536 if (!PageHighMem(page)) {
537 ret = HYPERVISOR_update_va_mapping(
538 (unsigned long)__va(pfn << PAGE_SHIFT),
542 __set_phys_to_machine(pfn, INVALID_P2M_ENTRY);
545 list_del(&page->lru);
547 balloon_append(page);
552 set_xen_guest_handle(reservation.extent_start, frame_list);
553 reservation.nr_extents = nr_pages;
554 ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation, &reservation);
555 BUG_ON(ret != nr_pages);
557 balloon_stats.current_pages -= nr_pages;
563 * As this is a work item it is guaranteed to run as a single instance only.
564 * We may of course race updates of the target counts (which are protected
565 * by the balloon lock), or with changes to the Xen hard limit, but we will
566 * recover from these in time.
568 static void balloon_process(struct work_struct *work)
570 enum bp_state state = BP_DONE;
575 mutex_lock(&balloon_mutex);
577 credit = current_credit();
580 if (balloon_is_inflated())
581 state = increase_reservation(credit);
583 state = reserve_additional_memory();
587 state = decrease_reservation(-credit, GFP_BALLOON);
589 state = update_schedule(state);
591 mutex_unlock(&balloon_mutex);
595 } while (credit && state == BP_DONE);
597 /* Schedule more work if there is some still to be done. */
598 if (state == BP_EAGAIN)
599 schedule_delayed_work(&balloon_worker, balloon_stats.schedule_delay * HZ);
602 /* Resets the Xen limit, sets new target, and kicks off processing. */
603 void balloon_set_new_target(unsigned long target)
605 /* No need for lock. Not read-modify-write updates. */
606 balloon_stats.target_pages = target;
607 schedule_delayed_work(&balloon_worker, 0);
609 EXPORT_SYMBOL_GPL(balloon_set_new_target);
611 static int add_ballooned_pages(int nr_pages)
615 if (xen_hotplug_unpopulated) {
616 st = reserve_additional_memory();
617 if (st != BP_ECANCELED) {
618 mutex_unlock(&balloon_mutex);
619 wait_event(balloon_wq,
620 !list_empty(&ballooned_pages));
621 mutex_lock(&balloon_mutex);
626 st = decrease_reservation(nr_pages, GFP_USER);
634 * alloc_xenballooned_pages - get pages that have been ballooned out
635 * @nr_pages: Number of pages to get
636 * @pages: pages returned
637 * @return 0 on success, error otherwise
639 int alloc_xenballooned_pages(int nr_pages, struct page **pages)
645 mutex_lock(&balloon_mutex);
647 balloon_stats.target_unpopulated += nr_pages;
649 while (pgno < nr_pages) {
650 page = balloon_retrieve(true);
652 pages[pgno++] = page;
653 #ifdef CONFIG_XEN_HAVE_PVMMU
655 * We don't support PV MMU when Linux and Xen is using
656 * different page granularity.
658 BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
660 ret = xen_alloc_p2m_entry(page_to_pfn(page));
665 ret = add_ballooned_pages(nr_pages - pgno);
670 mutex_unlock(&balloon_mutex);
673 mutex_unlock(&balloon_mutex);
674 free_xenballooned_pages(pgno, pages);
677 EXPORT_SYMBOL(alloc_xenballooned_pages);
680 * free_xenballooned_pages - return pages retrieved with get_ballooned_pages
681 * @nr_pages: Number of pages
682 * @pages: pages to return
684 void free_xenballooned_pages(int nr_pages, struct page **pages)
688 mutex_lock(&balloon_mutex);
690 for (i = 0; i < nr_pages; i++) {
692 balloon_append(pages[i]);
695 balloon_stats.target_unpopulated -= nr_pages;
697 /* The balloon may be too large now. Shrink it if needed. */
698 if (current_credit())
699 schedule_delayed_work(&balloon_worker, 0);
701 mutex_unlock(&balloon_mutex);
703 EXPORT_SYMBOL(free_xenballooned_pages);
705 static void __init balloon_add_region(unsigned long start_pfn,
708 unsigned long pfn, extra_pfn_end;
712 * If the amount of usable memory has been limited (e.g., with
713 * the 'mem' command line parameter), don't add pages beyond
716 extra_pfn_end = min(max_pfn, start_pfn + pages);
718 for (pfn = start_pfn; pfn < extra_pfn_end; pfn++) {
719 page = pfn_to_page(pfn);
720 /* totalram_pages and totalhigh_pages do not
721 include the boot-time balloon extension, so
722 don't subtract from it. */
723 __balloon_append(page);
726 balloon_stats.total_pages += extra_pfn_end - start_pfn;
729 static int __init balloon_init(void)
736 pr_info("Initialising balloon driver\n");
738 balloon_stats.current_pages = xen_pv_domain()
739 ? min(xen_start_info->nr_pages - xen_released_pages, max_pfn)
740 : get_num_physpages();
741 balloon_stats.target_pages = balloon_stats.current_pages;
742 balloon_stats.balloon_low = 0;
743 balloon_stats.balloon_high = 0;
744 balloon_stats.total_pages = balloon_stats.current_pages;
746 balloon_stats.schedule_delay = 1;
747 balloon_stats.max_schedule_delay = 32;
748 balloon_stats.retry_count = 1;
749 balloon_stats.max_retry_count = RETRY_UNLIMITED;
751 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
752 set_online_page_callback(&xen_online_page);
753 register_memory_notifier(&xen_memory_nb);
754 register_sysctl_table(xen_root);
758 * Initialize the balloon with pages from the extra memory
759 * regions (see arch/x86/xen/setup.c).
761 for (i = 0; i < XEN_EXTRA_MEM_MAX_REGIONS; i++)
762 if (xen_extra_mem[i].n_pfns)
763 balloon_add_region(xen_extra_mem[i].start_pfn,
764 xen_extra_mem[i].n_pfns);
769 subsys_initcall(balloon_init);
771 MODULE_LICENSE("GPL");