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 static void release_memory_resource(struct resource *resource);
157 /* When ballooning out (allocating memory to return to Xen) we don't really
158 want the kernel to try too hard since that can trigger the oom killer. */
159 #define GFP_BALLOON \
160 (GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC)
162 static void scrub_page(struct page *page)
164 #ifdef CONFIG_XEN_SCRUB_PAGES
165 clear_highpage(page);
169 /* balloon_append: add the given page to the balloon. */
170 static void __balloon_append(struct page *page)
172 /* Lowmem is re-populated first, so highmem pages go at list tail. */
173 if (PageHighMem(page)) {
174 list_add_tail(&page->lru, &ballooned_pages);
175 balloon_stats.balloon_high++;
177 list_add(&page->lru, &ballooned_pages);
178 balloon_stats.balloon_low++;
180 wake_up(&balloon_wq);
183 static void balloon_append(struct page *page)
185 __balloon_append(page);
186 adjust_managed_page_count(page, -1);
189 /* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
190 static struct page *balloon_retrieve(bool require_lowmem)
194 if (list_empty(&ballooned_pages))
197 page = list_entry(ballooned_pages.next, struct page, lru);
198 if (require_lowmem && PageHighMem(page))
200 list_del(&page->lru);
202 if (PageHighMem(page))
203 balloon_stats.balloon_high--;
205 balloon_stats.balloon_low--;
207 adjust_managed_page_count(page, 1);
212 static struct page *balloon_next_page(struct page *page)
214 struct list_head *next = page->lru.next;
215 if (next == &ballooned_pages)
217 return list_entry(next, struct page, lru);
220 static enum bp_state update_schedule(enum bp_state state)
222 if (state == BP_WAIT)
225 if (state == BP_ECANCELED)
228 if (state == BP_DONE) {
229 balloon_stats.schedule_delay = 1;
230 balloon_stats.retry_count = 1;
234 ++balloon_stats.retry_count;
236 if (balloon_stats.max_retry_count != RETRY_UNLIMITED &&
237 balloon_stats.retry_count > balloon_stats.max_retry_count) {
238 balloon_stats.schedule_delay = 1;
239 balloon_stats.retry_count = 1;
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;
251 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
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_MEM | 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");
273 #ifdef CONFIG_SPARSEMEM
275 unsigned long limit = 1UL << (MAX_PHYSMEM_BITS - PAGE_SHIFT);
276 unsigned long pfn = res->start >> PAGE_SHIFT;
279 pr_err("New System RAM resource outside addressable RAM (%lu > %lu)\n",
281 release_memory_resource(res);
290 static void release_memory_resource(struct resource *resource)
296 * No need to reset region to identity mapped since we now
297 * know that no I/O can be in this region
299 release_resource(resource);
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");
357 rc = add_memory_resource(nid, resource);
359 pr_warn("Cannot add additional memory (%i)\n", rc);
363 balloon_stats.total_pages += balloon_hotplug;
367 release_memory_resource(resource);
371 static void xen_online_page(struct page *page)
373 __online_page_set_limits(page);
375 mutex_lock(&balloon_mutex);
377 __balloon_append(page);
379 mutex_unlock(&balloon_mutex);
382 static int xen_memory_notifier(struct notifier_block *nb, unsigned long val, void *v)
384 if (val == MEM_ONLINE)
385 schedule_delayed_work(&balloon_worker, 0);
390 static struct notifier_block xen_memory_nb = {
391 .notifier_call = xen_memory_notifier,
395 static enum bp_state reserve_additional_memory(void)
397 balloon_stats.target_pages = balloon_stats.current_pages;
400 #endif /* CONFIG_XEN_BALLOON_MEMORY_HOTPLUG */
402 static long current_credit(void)
404 return balloon_stats.target_pages - balloon_stats.current_pages;
407 static bool balloon_is_inflated(void)
409 return balloon_stats.balloon_low || balloon_stats.balloon_high;
412 static enum bp_state increase_reservation(unsigned long nr_pages)
417 struct xen_memory_reservation reservation = {
419 .extent_order = EXTENT_ORDER,
423 if (nr_pages > ARRAY_SIZE(frame_list))
424 nr_pages = ARRAY_SIZE(frame_list);
426 page = list_first_entry_or_null(&ballooned_pages, struct page, lru);
427 for (i = 0; i < nr_pages; i++) {
433 /* XENMEM_populate_physmap requires a PFN based on Xen
436 frame_list[i] = page_to_xen_pfn(page);
437 page = balloon_next_page(page);
440 set_xen_guest_handle(reservation.extent_start, frame_list);
441 reservation.nr_extents = nr_pages;
442 rc = HYPERVISOR_memory_op(XENMEM_populate_physmap, &reservation);
446 for (i = 0; i < rc; i++) {
447 page = balloon_retrieve(false);
448 BUG_ON(page == NULL);
450 #ifdef CONFIG_XEN_HAVE_PVMMU
452 * We don't support PV MMU when Linux and Xen is using
453 * different page granularity.
455 BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
457 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
458 unsigned long pfn = page_to_pfn(page);
460 set_phys_to_machine(pfn, frame_list[i]);
462 /* Link back into the page tables if not highmem. */
463 if (!PageHighMem(page)) {
465 ret = HYPERVISOR_update_va_mapping(
466 (unsigned long)__va(pfn << PAGE_SHIFT),
467 mfn_pte(frame_list[i], PAGE_KERNEL),
474 /* Relinquish the page back to the allocator. */
475 __free_reserved_page(page);
478 balloon_stats.current_pages += rc;
483 static enum bp_state decrease_reservation(unsigned long nr_pages, gfp_t gfp)
485 enum bp_state state = BP_DONE;
487 struct page *page, *tmp;
489 struct xen_memory_reservation reservation = {
491 .extent_order = EXTENT_ORDER,
496 if (nr_pages > ARRAY_SIZE(frame_list))
497 nr_pages = ARRAY_SIZE(frame_list);
499 for (i = 0; i < nr_pages; i++) {
500 page = alloc_page(gfp);
507 list_add(&page->lru, &pages);
511 * Ensure that ballooned highmem pages don't have kmaps.
513 * Do this before changing the p2m as kmap_flush_unused()
514 * reads PTEs to obtain pages (and hence needs the original
520 * Setup the frame, update direct mapping, invalidate P2M,
521 * and add to balloon.
524 list_for_each_entry_safe(page, tmp, &pages, lru) {
525 /* XENMEM_decrease_reservation requires a GFN */
526 frame_list[i++] = xen_page_to_gfn(page);
528 #ifdef CONFIG_XEN_HAVE_PVMMU
530 * We don't support PV MMU when Linux and Xen is using
531 * different page granularity.
533 BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
535 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
536 unsigned long pfn = page_to_pfn(page);
538 if (!PageHighMem(page)) {
539 ret = HYPERVISOR_update_va_mapping(
540 (unsigned long)__va(pfn << PAGE_SHIFT),
544 __set_phys_to_machine(pfn, INVALID_P2M_ENTRY);
547 list_del(&page->lru);
549 balloon_append(page);
554 set_xen_guest_handle(reservation.extent_start, frame_list);
555 reservation.nr_extents = nr_pages;
556 ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation, &reservation);
557 BUG_ON(ret != nr_pages);
559 balloon_stats.current_pages -= nr_pages;
565 * As this is a work item it is guaranteed to run as a single instance only.
566 * We may of course race updates of the target counts (which are protected
567 * by the balloon lock), or with changes to the Xen hard limit, but we will
568 * recover from these in time.
570 static void balloon_process(struct work_struct *work)
572 enum bp_state state = BP_DONE;
577 mutex_lock(&balloon_mutex);
579 credit = current_credit();
582 if (balloon_is_inflated())
583 state = increase_reservation(credit);
585 state = reserve_additional_memory();
589 state = decrease_reservation(-credit, GFP_BALLOON);
591 state = update_schedule(state);
593 mutex_unlock(&balloon_mutex);
597 } while (credit && state == BP_DONE);
599 /* Schedule more work if there is some still to be done. */
600 if (state == BP_EAGAIN)
601 schedule_delayed_work(&balloon_worker, balloon_stats.schedule_delay * HZ);
604 /* Resets the Xen limit, sets new target, and kicks off processing. */
605 void balloon_set_new_target(unsigned long target)
607 /* No need for lock. Not read-modify-write updates. */
608 balloon_stats.target_pages = target;
609 schedule_delayed_work(&balloon_worker, 0);
611 EXPORT_SYMBOL_GPL(balloon_set_new_target);
613 static int add_ballooned_pages(int nr_pages)
617 if (xen_hotplug_unpopulated) {
618 st = reserve_additional_memory();
619 if (st != BP_ECANCELED) {
620 mutex_unlock(&balloon_mutex);
621 wait_event(balloon_wq,
622 !list_empty(&ballooned_pages));
623 mutex_lock(&balloon_mutex);
628 st = decrease_reservation(nr_pages, GFP_USER);
636 * alloc_xenballooned_pages - get pages that have been ballooned out
637 * @nr_pages: Number of pages to get
638 * @pages: pages returned
639 * @return 0 on success, error otherwise
641 int alloc_xenballooned_pages(int nr_pages, struct page **pages)
647 mutex_lock(&balloon_mutex);
649 balloon_stats.target_unpopulated += nr_pages;
651 while (pgno < nr_pages) {
652 page = balloon_retrieve(true);
654 pages[pgno++] = page;
655 #ifdef CONFIG_XEN_HAVE_PVMMU
657 * We don't support PV MMU when Linux and Xen is using
658 * different page granularity.
660 BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
662 ret = xen_alloc_p2m_entry(page_to_pfn(page));
667 ret = add_ballooned_pages(nr_pages - pgno);
672 mutex_unlock(&balloon_mutex);
675 mutex_unlock(&balloon_mutex);
676 free_xenballooned_pages(pgno, pages);
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 schedule_delayed_work(&balloon_worker, 0);
703 mutex_unlock(&balloon_mutex);
705 EXPORT_SYMBOL(free_xenballooned_pages);
707 static void __init balloon_add_region(unsigned long start_pfn,
710 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 page = pfn_to_page(pfn);
722 /* totalram_pages and totalhigh_pages do not
723 include the boot-time balloon extension, so
724 don't subtract from it. */
725 __balloon_append(page);
728 balloon_stats.total_pages += extra_pfn_end - start_pfn;
731 static int __init balloon_init(void)
738 pr_info("Initialising balloon driver\n");
740 balloon_stats.current_pages = xen_pv_domain()
741 ? min(xen_start_info->nr_pages - xen_released_pages, max_pfn)
742 : get_num_physpages();
743 balloon_stats.target_pages = balloon_stats.current_pages;
744 balloon_stats.balloon_low = 0;
745 balloon_stats.balloon_high = 0;
746 balloon_stats.total_pages = balloon_stats.current_pages;
748 balloon_stats.schedule_delay = 1;
749 balloon_stats.max_schedule_delay = 32;
750 balloon_stats.retry_count = 1;
751 balloon_stats.max_retry_count = RETRY_UNLIMITED;
753 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
754 set_online_page_callback(&xen_online_page);
755 register_memory_notifier(&xen_memory_nb);
756 register_sysctl_table(xen_root);
760 * Initialize the balloon with pages from the extra memory
761 * regions (see arch/x86/xen/setup.c).
763 for (i = 0; i < XEN_EXTRA_MEM_MAX_REGIONS; i++)
764 if (xen_extra_mem[i].n_pfns)
765 balloon_add_region(xen_extra_mem[i].start_pfn,
766 xen_extra_mem[i].n_pfns);
771 subsys_initcall(balloon_init);
773 MODULE_LICENSE("GPL");