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>
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>
74 #include <xen/mem-reservation.h>
76 static int xen_hotplug_unpopulated;
78 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
83 static struct ctl_table balloon_table[] = {
85 .procname = "hotplug_unpopulated",
86 .data = &xen_hotplug_unpopulated,
87 .maxlen = sizeof(int),
89 .proc_handler = proc_dointvec_minmax,
96 static struct ctl_table balloon_root[] = {
98 .procname = "balloon",
100 .child = balloon_table,
105 static struct ctl_table xen_root[] = {
109 .child = balloon_root,
117 * Use one extent per PAGE_SIZE to avoid to break down the page into
120 #define EXTENT_ORDER (fls(XEN_PFN_PER_PAGE) - 1)
123 * balloon_process() state:
125 * BP_DONE: done or nothing to do,
126 * BP_WAIT: wait to be rescheduled,
127 * BP_EAGAIN: error, go to sleep,
128 * BP_ECANCELED: error, balloon operation canceled.
139 static DEFINE_MUTEX(balloon_mutex);
141 struct balloon_stats balloon_stats;
142 EXPORT_SYMBOL_GPL(balloon_stats);
144 /* We increase/decrease in batches which fit in a page */
145 static xen_pfn_t frame_list[PAGE_SIZE / sizeof(xen_pfn_t)];
148 /* List of ballooned pages, threaded through the mem_map array. */
149 static LIST_HEAD(ballooned_pages);
150 static DECLARE_WAIT_QUEUE_HEAD(balloon_wq);
152 /* Main work function, always executed in process context. */
153 static void balloon_process(struct work_struct *work);
154 static DECLARE_DELAYED_WORK(balloon_worker, balloon_process);
156 /* When ballooning out (allocating memory to return to Xen) we don't really
157 want the kernel to try too hard since that can trigger the oom killer. */
158 #define GFP_BALLOON \
159 (GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC)
161 /* balloon_append: add the given page to the balloon. */
162 static void __balloon_append(struct page *page)
164 /* Lowmem is re-populated first, so highmem pages go at list tail. */
165 if (PageHighMem(page)) {
166 list_add_tail(&page->lru, &ballooned_pages);
167 balloon_stats.balloon_high++;
169 list_add(&page->lru, &ballooned_pages);
170 balloon_stats.balloon_low++;
172 wake_up(&balloon_wq);
175 static void balloon_append(struct page *page)
177 __balloon_append(page);
180 /* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
181 static struct page *balloon_retrieve(bool require_lowmem)
185 if (list_empty(&ballooned_pages))
188 page = list_entry(ballooned_pages.next, struct page, lru);
189 if (require_lowmem && PageHighMem(page))
191 list_del(&page->lru);
193 if (PageHighMem(page))
194 balloon_stats.balloon_high--;
196 balloon_stats.balloon_low--;
201 static struct page *balloon_next_page(struct page *page)
203 struct list_head *next = page->lru.next;
204 if (next == &ballooned_pages)
206 return list_entry(next, struct page, lru);
209 static enum bp_state update_schedule(enum bp_state state)
211 if (state == BP_WAIT)
214 if (state == BP_ECANCELED)
217 if (state == BP_DONE) {
218 balloon_stats.schedule_delay = 1;
219 balloon_stats.retry_count = 1;
223 ++balloon_stats.retry_count;
225 if (balloon_stats.max_retry_count != RETRY_UNLIMITED &&
226 balloon_stats.retry_count > balloon_stats.max_retry_count) {
227 balloon_stats.schedule_delay = 1;
228 balloon_stats.retry_count = 1;
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;
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);
255 * Host memory not allocated to dom0. We can use this range for hotplug-based
258 * It's a type-less resource. Setting IORESOURCE_MEM will make resource
259 * management algorithms (arch_remove_reservations()) look into guest e820,
260 * which we don't want.
262 static struct resource hostmem_resource = {
266 void __attribute__((weak)) __init arch_xen_balloon_init(struct resource *res)
269 static struct resource *additional_memory_resource(phys_addr_t size)
271 struct resource *res, *res_hostmem;
274 res = kzalloc(sizeof(*res), GFP_KERNEL);
278 res->name = "System RAM";
279 res->flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
281 res_hostmem = kzalloc(sizeof(*res), GFP_KERNEL);
283 /* Try to grab a range from hostmem */
284 res_hostmem->name = "Host memory";
285 ret = allocate_resource(&hostmem_resource, res_hostmem,
287 PAGES_PER_SECTION * PAGE_SIZE, NULL, NULL);
292 * Insert this resource into iomem. Because hostmem_resource
293 * tracks portion of guest e820 marked as UNUSABLE noone else
294 * should try to use it.
296 res->start = res_hostmem->start;
297 res->end = res_hostmem->end;
298 ret = insert_resource(&iomem_resource, res);
300 pr_err("Can't insert iomem_resource [%llx - %llx]\n",
301 res->start, res->end);
302 release_memory_resource(res_hostmem);
304 res->start = res->end = 0;
309 ret = allocate_resource(&iomem_resource, res,
311 PAGES_PER_SECTION * PAGE_SIZE, NULL, NULL);
313 pr_err("Cannot allocate new System RAM resource\n");
319 #ifdef CONFIG_SPARSEMEM
321 unsigned long limit = 1UL << (MAX_PHYSMEM_BITS - PAGE_SHIFT);
322 unsigned long pfn = res->start >> PAGE_SHIFT;
325 pr_err("New System RAM resource outside addressable RAM (%lu > %lu)\n",
327 release_memory_resource(res);
328 release_memory_resource(res_hostmem);
337 static enum bp_state reserve_additional_memory(void)
340 struct resource *resource;
342 unsigned long balloon_hotplug;
344 credit = balloon_stats.target_pages + balloon_stats.target_unpopulated
345 - balloon_stats.total_pages;
348 * Already hotplugged enough pages? Wait for them to be
354 balloon_hotplug = round_up(credit, PAGES_PER_SECTION);
356 resource = additional_memory_resource(balloon_hotplug * PAGE_SIZE);
360 nid = memory_add_physaddr_to_nid(resource->start);
362 #ifdef CONFIG_XEN_HAVE_PVMMU
364 * We don't support PV MMU when Linux and Xen is using
365 * different page granularity.
367 BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
370 * add_memory() will build page tables for the new memory so
371 * the p2m must contain invalid entries so the correct
372 * non-present PTEs will be written.
374 * If a failure occurs, the original (identity) p2m entries
375 * are not restored since this region is now known not to
376 * conflict with any devices.
378 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
379 unsigned long pfn, i;
381 pfn = PFN_DOWN(resource->start);
382 for (i = 0; i < balloon_hotplug; i++) {
383 if (!set_phys_to_machine(pfn + i, INVALID_P2M_ENTRY)) {
384 pr_warn("set_phys_to_machine() failed, no memory added\n");
392 * add_memory_resource() will call online_pages() which in its turn
393 * will call xen_online_page() callback causing deadlock if we don't
394 * release balloon_mutex here. Unlocking here is safe because the
395 * callers drop the mutex before trying again.
397 mutex_unlock(&balloon_mutex);
398 rc = add_memory_resource(nid, resource, memhp_auto_online);
399 mutex_lock(&balloon_mutex);
402 pr_warn("Cannot add additional memory (%i)\n", rc);
406 balloon_stats.total_pages += balloon_hotplug;
410 release_memory_resource(resource);
414 static void xen_online_page(struct page *page)
416 __online_page_set_limits(page);
418 mutex_lock(&balloon_mutex);
420 __balloon_append(page);
422 mutex_unlock(&balloon_mutex);
425 static int xen_memory_notifier(struct notifier_block *nb, unsigned long val, void *v)
427 if (val == MEM_ONLINE)
428 schedule_delayed_work(&balloon_worker, 0);
433 static struct notifier_block xen_memory_nb = {
434 .notifier_call = xen_memory_notifier,
438 static enum bp_state reserve_additional_memory(void)
440 balloon_stats.target_pages = balloon_stats.current_pages;
443 #endif /* CONFIG_XEN_BALLOON_MEMORY_HOTPLUG */
445 static long current_credit(void)
447 return balloon_stats.target_pages - balloon_stats.current_pages;
450 static bool balloon_is_inflated(void)
452 return balloon_stats.balloon_low || balloon_stats.balloon_high;
455 static enum bp_state increase_reservation(unsigned long nr_pages)
461 if (nr_pages > ARRAY_SIZE(frame_list))
462 nr_pages = ARRAY_SIZE(frame_list);
464 page = list_first_entry_or_null(&ballooned_pages, struct page, lru);
465 for (i = 0; i < nr_pages; i++) {
471 frame_list[i] = page_to_xen_pfn(page);
472 page = balloon_next_page(page);
475 rc = xenmem_reservation_increase(nr_pages, frame_list);
479 for (i = 0; i < rc; i++) {
480 page = balloon_retrieve(false);
481 BUG_ON(page == NULL);
483 xenmem_reservation_va_mapping_update(1, &page, &frame_list[i]);
485 /* Relinquish the page back to the allocator. */
486 free_reserved_page(page);
489 balloon_stats.current_pages += rc;
494 static enum bp_state decrease_reservation(unsigned long nr_pages, gfp_t gfp)
496 enum bp_state state = BP_DONE;
498 struct page *page, *tmp;
502 if (nr_pages > ARRAY_SIZE(frame_list))
503 nr_pages = ARRAY_SIZE(frame_list);
505 for (i = 0; i < nr_pages; i++) {
506 page = alloc_page(gfp);
512 adjust_managed_page_count(page, -1);
513 xenmem_reservation_scrub_page(page);
514 list_add(&page->lru, &pages);
518 * Ensure that ballooned highmem pages don't have kmaps.
520 * Do this before changing the p2m as kmap_flush_unused()
521 * reads PTEs to obtain pages (and hence needs the original
527 * Setup the frame, update direct mapping, invalidate P2M,
528 * and add to balloon.
531 list_for_each_entry_safe(page, tmp, &pages, lru) {
532 frame_list[i++] = xen_page_to_gfn(page);
534 xenmem_reservation_va_mapping_reset(1, &page);
536 list_del(&page->lru);
538 balloon_append(page);
543 ret = xenmem_reservation_decrease(nr_pages, frame_list);
544 BUG_ON(ret != nr_pages);
546 balloon_stats.current_pages -= nr_pages;
552 * As this is a work item it is guaranteed to run as a single instance only.
553 * We may of course race updates of the target counts (which are protected
554 * by the balloon lock), or with changes to the Xen hard limit, but we will
555 * recover from these in time.
557 static void balloon_process(struct work_struct *work)
559 enum bp_state state = BP_DONE;
564 mutex_lock(&balloon_mutex);
566 credit = current_credit();
569 if (balloon_is_inflated())
570 state = increase_reservation(credit);
572 state = reserve_additional_memory();
576 state = decrease_reservation(-credit, GFP_BALLOON);
578 state = update_schedule(state);
580 mutex_unlock(&balloon_mutex);
584 } while (credit && state == BP_DONE);
586 /* Schedule more work if there is some still to be done. */
587 if (state == BP_EAGAIN)
588 schedule_delayed_work(&balloon_worker, balloon_stats.schedule_delay * HZ);
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 schedule_delayed_work(&balloon_worker, 0);
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) {
607 mutex_unlock(&balloon_mutex);
608 wait_event(balloon_wq,
609 !list_empty(&ballooned_pages));
610 mutex_lock(&balloon_mutex);
615 st = decrease_reservation(nr_pages, GFP_USER);
623 * alloc_xenballooned_pages - get pages that have been ballooned out
624 * @nr_pages: Number of pages to get
625 * @pages: pages returned
626 * @return 0 on success, error otherwise
628 int alloc_xenballooned_pages(int nr_pages, struct page **pages)
634 mutex_lock(&balloon_mutex);
636 balloon_stats.target_unpopulated += nr_pages;
638 while (pgno < nr_pages) {
639 page = balloon_retrieve(true);
641 pages[pgno++] = page;
642 #ifdef CONFIG_XEN_HAVE_PVMMU
644 * We don't support PV MMU when Linux and Xen is using
645 * different page granularity.
647 BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
649 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
650 ret = xen_alloc_p2m_entry(page_to_pfn(page));
656 ret = add_ballooned_pages(nr_pages - pgno);
661 mutex_unlock(&balloon_mutex);
664 mutex_unlock(&balloon_mutex);
665 free_xenballooned_pages(pgno, pages);
668 EXPORT_SYMBOL(alloc_xenballooned_pages);
671 * free_xenballooned_pages - return pages retrieved with get_ballooned_pages
672 * @nr_pages: Number of pages
673 * @pages: pages to return
675 void free_xenballooned_pages(int nr_pages, struct page **pages)
679 mutex_lock(&balloon_mutex);
681 for (i = 0; i < nr_pages; i++) {
683 balloon_append(pages[i]);
686 balloon_stats.target_unpopulated -= nr_pages;
688 /* The balloon may be too large now. Shrink it if needed. */
689 if (current_credit())
690 schedule_delayed_work(&balloon_worker, 0);
692 mutex_unlock(&balloon_mutex);
694 EXPORT_SYMBOL(free_xenballooned_pages);
697 static void __init balloon_add_region(unsigned long start_pfn,
700 unsigned long pfn, extra_pfn_end;
704 * If the amount of usable memory has been limited (e.g., with
705 * the 'mem' command line parameter), don't add pages beyond
708 extra_pfn_end = min(max_pfn, start_pfn + pages);
710 for (pfn = start_pfn; pfn < extra_pfn_end; pfn++) {
711 page = pfn_to_page(pfn);
712 /* totalram_pages and totalhigh_pages do not
713 include the boot-time balloon extension, so
714 don't subtract from it. */
715 __balloon_append(page);
718 balloon_stats.total_pages += extra_pfn_end - start_pfn;
722 static int __init balloon_init(void)
727 pr_info("Initialising balloon driver\n");
730 balloon_stats.current_pages = xen_pv_domain()
731 ? min(xen_start_info->nr_pages - xen_released_pages, max_pfn)
732 : get_num_physpages();
734 balloon_stats.current_pages = get_num_physpages();
736 balloon_stats.target_pages = balloon_stats.current_pages;
737 balloon_stats.balloon_low = 0;
738 balloon_stats.balloon_high = 0;
739 balloon_stats.total_pages = balloon_stats.current_pages;
741 balloon_stats.schedule_delay = 1;
742 balloon_stats.max_schedule_delay = 32;
743 balloon_stats.retry_count = 1;
744 balloon_stats.max_retry_count = RETRY_UNLIMITED;
746 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
747 set_online_page_callback(&xen_online_page);
748 register_memory_notifier(&xen_memory_nb);
749 register_sysctl_table(xen_root);
751 arch_xen_balloon_init(&hostmem_resource);
759 * Initialize the balloon with pages from the extra memory
760 * regions (see arch/x86/xen/setup.c).
762 for (i = 0; i < XEN_EXTRA_MEM_MAX_REGIONS; i++)
763 if (xen_extra_mem[i].n_pfns)
764 balloon_add_region(xen_extra_mem[i].start_pfn,
765 xen_extra_mem[i].n_pfns);
769 /* Init the xen-balloon driver. */
774 subsys_initcall(balloon_init);