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>
59 #include <asm/pgalloc.h>
60 #include <asm/pgtable.h>
63 #include <asm/xen/hypervisor.h>
64 #include <asm/xen/hypercall.h>
67 #include <xen/interface/xen.h>
68 #include <xen/interface/memory.h>
69 #include <xen/balloon.h>
70 #include <xen/features.h>
74 * balloon_process() state:
76 * BP_DONE: done or nothing to do,
77 * BP_EAGAIN: error, go to sleep,
78 * BP_ECANCELED: error, balloon operation canceled.
88 static DEFINE_MUTEX(balloon_mutex);
90 struct balloon_stats balloon_stats;
91 EXPORT_SYMBOL_GPL(balloon_stats);
93 /* We increase/decrease in batches which fit in a page */
94 static xen_pfn_t frame_list[PAGE_SIZE / sizeof(unsigned long)];
95 static DEFINE_PER_CPU(struct page *, balloon_scratch_page);
98 /* List of ballooned pages, threaded through the mem_map array. */
99 static LIST_HEAD(ballooned_pages);
101 /* Main work function, always executed in process context. */
102 static void balloon_process(struct work_struct *work);
103 static DECLARE_DELAYED_WORK(balloon_worker, balloon_process);
105 /* When ballooning out (allocating memory to return to Xen) we don't really
106 want the kernel to try too hard since that can trigger the oom killer. */
107 #define GFP_BALLOON \
108 (GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC)
110 static void scrub_page(struct page *page)
112 #ifdef CONFIG_XEN_SCRUB_PAGES
113 clear_highpage(page);
117 /* balloon_append: add the given page to the balloon. */
118 static void __balloon_append(struct page *page)
120 /* Lowmem is re-populated first, so highmem pages go at list tail. */
121 if (PageHighMem(page)) {
122 list_add_tail(&page->lru, &ballooned_pages);
123 balloon_stats.balloon_high++;
125 list_add(&page->lru, &ballooned_pages);
126 balloon_stats.balloon_low++;
130 static void balloon_append(struct page *page)
132 __balloon_append(page);
133 adjust_managed_page_count(page, -1);
136 /* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
137 static struct page *balloon_retrieve(bool prefer_highmem)
141 if (list_empty(&ballooned_pages))
145 page = list_entry(ballooned_pages.prev, struct page, lru);
147 page = list_entry(ballooned_pages.next, struct page, lru);
148 list_del(&page->lru);
150 if (PageHighMem(page))
151 balloon_stats.balloon_high--;
153 balloon_stats.balloon_low--;
155 adjust_managed_page_count(page, 1);
160 static struct page *balloon_first_page(void)
162 if (list_empty(&ballooned_pages))
164 return list_entry(ballooned_pages.next, struct page, lru);
167 static struct page *balloon_next_page(struct page *page)
169 struct list_head *next = page->lru.next;
170 if (next == &ballooned_pages)
172 return list_entry(next, struct page, lru);
175 static enum bp_state update_schedule(enum bp_state state)
177 if (state == BP_DONE) {
178 balloon_stats.schedule_delay = 1;
179 balloon_stats.retry_count = 1;
183 ++balloon_stats.retry_count;
185 if (balloon_stats.max_retry_count != RETRY_UNLIMITED &&
186 balloon_stats.retry_count > balloon_stats.max_retry_count) {
187 balloon_stats.schedule_delay = 1;
188 balloon_stats.retry_count = 1;
192 balloon_stats.schedule_delay <<= 1;
194 if (balloon_stats.schedule_delay > balloon_stats.max_schedule_delay)
195 balloon_stats.schedule_delay = balloon_stats.max_schedule_delay;
200 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
201 static long current_credit(void)
203 return balloon_stats.target_pages - balloon_stats.current_pages -
204 balloon_stats.hotplug_pages;
207 static bool balloon_is_inflated(void)
209 if (balloon_stats.balloon_low || balloon_stats.balloon_high ||
210 balloon_stats.balloon_hotplug)
217 * reserve_additional_memory() adds memory region of size >= credit above
218 * max_pfn. New region is section aligned and size is modified to be multiple
219 * of section size. Those features allow optimal use of address space and
220 * establish proper alignment when this function is called first time after
221 * boot (last section not fully populated at boot time contains unused memory
222 * pages with PG_reserved bit not set; online_pages_range() does not allow page
223 * onlining in whole range if first onlined page does not have PG_reserved
224 * bit set). Real size of added memory is established at page onlining stage.
227 static enum bp_state reserve_additional_memory(long credit)
230 u64 hotplug_start_paddr;
231 unsigned long balloon_hotplug = credit;
233 hotplug_start_paddr = PFN_PHYS(SECTION_ALIGN_UP(max_pfn));
234 balloon_hotplug = round_up(balloon_hotplug, PAGES_PER_SECTION);
235 nid = memory_add_physaddr_to_nid(hotplug_start_paddr);
237 rc = add_memory(nid, hotplug_start_paddr, balloon_hotplug << PAGE_SHIFT);
240 pr_info("%s: add_memory() failed: %i\n", __func__, rc);
244 balloon_hotplug -= credit;
246 balloon_stats.hotplug_pages += credit;
247 balloon_stats.balloon_hotplug = balloon_hotplug;
252 static void xen_online_page(struct page *page)
254 __online_page_set_limits(page);
256 mutex_lock(&balloon_mutex);
258 __balloon_append(page);
260 if (balloon_stats.hotplug_pages)
261 --balloon_stats.hotplug_pages;
263 --balloon_stats.balloon_hotplug;
265 mutex_unlock(&balloon_mutex);
268 static int xen_memory_notifier(struct notifier_block *nb, unsigned long val, void *v)
270 if (val == MEM_ONLINE)
271 schedule_delayed_work(&balloon_worker, 0);
276 static struct notifier_block xen_memory_nb = {
277 .notifier_call = xen_memory_notifier,
281 static long current_credit(void)
283 unsigned long target = balloon_stats.target_pages;
286 balloon_stats.current_pages +
287 balloon_stats.balloon_low +
288 balloon_stats.balloon_high);
290 return target - balloon_stats.current_pages;
293 static bool balloon_is_inflated(void)
295 if (balloon_stats.balloon_low || balloon_stats.balloon_high)
301 static enum bp_state reserve_additional_memory(long credit)
303 balloon_stats.target_pages = balloon_stats.current_pages;
306 #endif /* CONFIG_XEN_BALLOON_MEMORY_HOTPLUG */
308 static enum bp_state increase_reservation(unsigned long nr_pages)
311 unsigned long pfn, i;
313 struct xen_memory_reservation reservation = {
319 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
320 if (!balloon_stats.balloon_low && !balloon_stats.balloon_high) {
321 nr_pages = min(nr_pages, balloon_stats.balloon_hotplug);
322 balloon_stats.hotplug_pages += nr_pages;
323 balloon_stats.balloon_hotplug -= nr_pages;
328 if (nr_pages > ARRAY_SIZE(frame_list))
329 nr_pages = ARRAY_SIZE(frame_list);
331 page = balloon_first_page();
332 for (i = 0; i < nr_pages; i++) {
337 frame_list[i] = page_to_pfn(page);
338 page = balloon_next_page(page);
341 set_xen_guest_handle(reservation.extent_start, frame_list);
342 reservation.nr_extents = nr_pages;
343 rc = HYPERVISOR_memory_op(XENMEM_populate_physmap, &reservation);
347 for (i = 0; i < rc; i++) {
348 page = balloon_retrieve(false);
349 BUG_ON(page == NULL);
351 pfn = page_to_pfn(page);
353 set_phys_to_machine(pfn, frame_list[i]);
355 #ifdef CONFIG_XEN_HAVE_PVMMU
356 /* Link back into the page tables if not highmem. */
357 if (xen_pv_domain() && !PageHighMem(page)) {
359 ret = HYPERVISOR_update_va_mapping(
360 (unsigned long)__va(pfn << PAGE_SHIFT),
361 mfn_pte(frame_list[i], PAGE_KERNEL),
367 /* Relinquish the page back to the allocator. */
368 __free_reserved_page(page);
371 balloon_stats.current_pages += rc;
376 static enum bp_state decrease_reservation(unsigned long nr_pages, gfp_t gfp)
378 enum bp_state state = BP_DONE;
379 unsigned long pfn, i;
381 struct page *scratch_page;
383 struct xen_memory_reservation reservation = {
389 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
390 if (balloon_stats.hotplug_pages) {
391 nr_pages = min(nr_pages, balloon_stats.hotplug_pages);
392 balloon_stats.hotplug_pages -= nr_pages;
393 balloon_stats.balloon_hotplug += nr_pages;
398 if (nr_pages > ARRAY_SIZE(frame_list))
399 nr_pages = ARRAY_SIZE(frame_list);
401 scratch_page = get_balloon_scratch_page();
403 for (i = 0; i < nr_pages; i++) {
404 page = alloc_page(gfp);
411 pfn = page_to_pfn(page);
412 frame_list[i] = pfn_to_mfn(pfn);
416 #ifdef CONFIG_XEN_HAVE_PVMMU
417 if (xen_pv_domain() && !PageHighMem(page)) {
418 ret = HYPERVISOR_update_va_mapping(
419 (unsigned long)__va(pfn << PAGE_SHIFT),
420 pfn_pte(page_to_pfn(scratch_page),
427 /* Ensure that ballooned highmem pages don't have kmaps. */
431 /* No more mappings: invalidate P2M and add to balloon. */
432 for (i = 0; i < nr_pages; i++) {
433 pfn = mfn_to_pfn(frame_list[i]);
434 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
436 p = page_to_pfn(scratch_page);
437 __set_phys_to_machine(pfn, pfn_to_mfn(p));
439 balloon_append(pfn_to_page(pfn));
442 put_balloon_scratch_page();
444 set_xen_guest_handle(reservation.extent_start, frame_list);
445 reservation.nr_extents = nr_pages;
446 ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation, &reservation);
447 BUG_ON(ret != nr_pages);
449 balloon_stats.current_pages -= nr_pages;
455 * We avoid multiple worker processes conflicting via the balloon mutex.
456 * We may of course race updates of the target counts (which are protected
457 * by the balloon lock), or with changes to the Xen hard limit, but we will
458 * recover from these in time.
460 static void balloon_process(struct work_struct *work)
462 enum bp_state state = BP_DONE;
465 mutex_lock(&balloon_mutex);
468 credit = current_credit();
471 if (balloon_is_inflated())
472 state = increase_reservation(credit);
474 state = reserve_additional_memory(credit);
478 state = decrease_reservation(-credit, GFP_BALLOON);
480 state = update_schedule(state);
482 #ifndef CONFIG_PREEMPT
486 } while (credit && state == BP_DONE);
488 /* Schedule more work if there is some still to be done. */
489 if (state == BP_EAGAIN)
490 schedule_delayed_work(&balloon_worker, balloon_stats.schedule_delay * HZ);
492 mutex_unlock(&balloon_mutex);
495 struct page *get_balloon_scratch_page(void)
497 struct page *ret = get_cpu_var(balloon_scratch_page);
502 void put_balloon_scratch_page(void)
504 put_cpu_var(balloon_scratch_page);
507 /* Resets the Xen limit, sets new target, and kicks off processing. */
508 void balloon_set_new_target(unsigned long target)
510 /* No need for lock. Not read-modify-write updates. */
511 balloon_stats.target_pages = target;
512 schedule_delayed_work(&balloon_worker, 0);
514 EXPORT_SYMBOL_GPL(balloon_set_new_target);
517 * alloc_xenballooned_pages - get pages that have been ballooned out
518 * @nr_pages: Number of pages to get
519 * @pages: pages returned
520 * @highmem: allow highmem pages
521 * @return 0 on success, error otherwise
523 int alloc_xenballooned_pages(int nr_pages, struct page **pages, bool highmem)
527 mutex_lock(&balloon_mutex);
528 while (pgno < nr_pages) {
529 page = balloon_retrieve(highmem);
530 if (page && (highmem || !PageHighMem(page))) {
531 pages[pgno++] = page;
535 balloon_append(page);
536 st = decrease_reservation(nr_pages - pgno,
537 highmem ? GFP_HIGHUSER : GFP_USER);
542 mutex_unlock(&balloon_mutex);
546 balloon_append(pages[--pgno]);
547 /* Free the memory back to the kernel soon */
548 schedule_delayed_work(&balloon_worker, 0);
549 mutex_unlock(&balloon_mutex);
552 EXPORT_SYMBOL(alloc_xenballooned_pages);
555 * free_xenballooned_pages - return pages retrieved with get_ballooned_pages
556 * @nr_pages: Number of pages
557 * @pages: pages to return
559 void free_xenballooned_pages(int nr_pages, struct page **pages)
563 mutex_lock(&balloon_mutex);
565 for (i = 0; i < nr_pages; i++) {
567 balloon_append(pages[i]);
570 /* The balloon may be too large now. Shrink it if needed. */
571 if (current_credit())
572 schedule_delayed_work(&balloon_worker, 0);
574 mutex_unlock(&balloon_mutex);
576 EXPORT_SYMBOL(free_xenballooned_pages);
578 static void __init balloon_add_region(unsigned long start_pfn,
581 unsigned long pfn, extra_pfn_end;
585 * If the amount of usable memory has been limited (e.g., with
586 * the 'mem' command line parameter), don't add pages beyond
589 extra_pfn_end = min(max_pfn, start_pfn + pages);
591 for (pfn = start_pfn; pfn < extra_pfn_end; pfn++) {
592 page = pfn_to_page(pfn);
593 /* totalram_pages and totalhigh_pages do not
594 include the boot-time balloon extension, so
595 don't subtract from it. */
596 __balloon_append(page);
600 static int __cpuinit balloon_cpu_notify(struct notifier_block *self,
601 unsigned long action, void *hcpu)
603 int cpu = (long)hcpu;
606 if (per_cpu(balloon_scratch_page, cpu) != NULL)
608 per_cpu(balloon_scratch_page, cpu) = alloc_page(GFP_KERNEL);
609 if (per_cpu(balloon_scratch_page, cpu) == NULL) {
610 pr_warn("Failed to allocate balloon_scratch_page for cpu %d\n", cpu);
620 static struct notifier_block balloon_cpu_notifier __cpuinitdata = {
621 .notifier_call = balloon_cpu_notify,
624 static int __init balloon_init(void)
631 for_each_online_cpu(cpu)
633 per_cpu(balloon_scratch_page, cpu) = alloc_page(GFP_KERNEL);
634 if (per_cpu(balloon_scratch_page, cpu) == NULL) {
635 pr_warn("Failed to allocate balloon_scratch_page for cpu %d\n", cpu);
639 register_cpu_notifier(&balloon_cpu_notifier);
641 pr_info("Initialising balloon driver\n");
643 balloon_stats.current_pages = xen_pv_domain()
644 ? min(xen_start_info->nr_pages - xen_released_pages, max_pfn)
646 balloon_stats.target_pages = balloon_stats.current_pages;
647 balloon_stats.balloon_low = 0;
648 balloon_stats.balloon_high = 0;
650 balloon_stats.schedule_delay = 1;
651 balloon_stats.max_schedule_delay = 32;
652 balloon_stats.retry_count = 1;
653 balloon_stats.max_retry_count = RETRY_UNLIMITED;
655 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
656 balloon_stats.hotplug_pages = 0;
657 balloon_stats.balloon_hotplug = 0;
659 set_online_page_callback(&xen_online_page);
660 register_memory_notifier(&xen_memory_nb);
664 * Initialize the balloon with pages from the extra memory
665 * regions (see arch/x86/xen/setup.c).
667 for (i = 0; i < XEN_EXTRA_MEM_MAX_REGIONS; i++)
668 if (xen_extra_mem[i].size)
669 balloon_add_region(PFN_UP(xen_extra_mem[i].start),
670 PFN_DOWN(xen_extra_mem[i].size));
675 subsys_initcall(balloon_init);
677 static int __init balloon_clear(void)
681 for_each_possible_cpu(cpu)
682 per_cpu(balloon_scratch_page, cpu) = NULL;
686 early_initcall(balloon_clear);
688 MODULE_LICENSE("GPL");