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/memblock.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>
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
80 static struct ctl_table balloon_table[] = {
82 .procname = "hotplug_unpopulated",
83 .data = &xen_hotplug_unpopulated,
84 .maxlen = sizeof(int),
86 .proc_handler = proc_dointvec_minmax,
87 .extra1 = SYSCTL_ZERO,
93 static struct ctl_table balloon_root[] = {
95 .procname = "balloon",
97 .child = balloon_table,
102 static struct ctl_table xen_root[] = {
106 .child = balloon_root,
114 * Use one extent per PAGE_SIZE to avoid to break down the page into
117 #define EXTENT_ORDER (fls(XEN_PFN_PER_PAGE) - 1)
120 * balloon_thread() state:
122 * BP_DONE: done or nothing to do,
123 * BP_WAIT: wait to be rescheduled,
124 * BP_EAGAIN: error, go to sleep,
125 * BP_ECANCELED: error, balloon operation canceled.
135 /* Main waiting point for xen-balloon thread. */
136 static DECLARE_WAIT_QUEUE_HEAD(balloon_thread_wq);
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 /* When ballooning out (allocating memory to return to Xen) we don't really
152 want the kernel to try too hard since that can trigger the oom killer. */
153 #define GFP_BALLOON \
154 (GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC)
156 /* balloon_append: add the given page to the balloon. */
157 static void balloon_append(struct page *page)
159 __SetPageOffline(page);
161 /* Lowmem is re-populated first, so highmem pages go at list tail. */
162 if (PageHighMem(page)) {
163 list_add_tail(&page->lru, &ballooned_pages);
164 balloon_stats.balloon_high++;
166 list_add(&page->lru, &ballooned_pages);
167 balloon_stats.balloon_low++;
169 wake_up(&balloon_wq);
172 /* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
173 static struct page *balloon_retrieve(bool require_lowmem)
177 if (list_empty(&ballooned_pages))
180 page = list_entry(ballooned_pages.next, struct page, lru);
181 if (require_lowmem && PageHighMem(page))
183 list_del(&page->lru);
185 if (PageHighMem(page))
186 balloon_stats.balloon_high--;
188 balloon_stats.balloon_low--;
190 __ClearPageOffline(page);
194 static struct page *balloon_next_page(struct page *page)
196 struct list_head *next = page->lru.next;
197 if (next == &ballooned_pages)
199 return list_entry(next, struct page, lru);
202 static enum bp_state update_schedule(enum bp_state state)
204 if (state == BP_WAIT)
207 if (state == BP_ECANCELED)
210 if (state == BP_DONE) {
211 balloon_stats.schedule_delay = 1;
212 balloon_stats.retry_count = 1;
216 ++balloon_stats.retry_count;
218 if (balloon_stats.max_retry_count != RETRY_UNLIMITED &&
219 balloon_stats.retry_count > balloon_stats.max_retry_count) {
220 balloon_stats.schedule_delay = 1;
221 balloon_stats.retry_count = 1;
225 balloon_stats.schedule_delay <<= 1;
227 if (balloon_stats.schedule_delay > balloon_stats.max_schedule_delay)
228 balloon_stats.schedule_delay = balloon_stats.max_schedule_delay;
233 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
234 static void release_memory_resource(struct resource *resource)
240 * No need to reset region to identity mapped since we now
241 * know that no I/O can be in this region
243 release_resource(resource);
247 static struct resource *additional_memory_resource(phys_addr_t size)
249 struct resource *res;
252 res = kzalloc(sizeof(*res), GFP_KERNEL);
256 res->name = "System RAM";
257 res->flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
259 ret = allocate_resource(&iomem_resource, res,
261 PAGES_PER_SECTION * PAGE_SIZE, NULL, NULL);
263 pr_err("Cannot allocate new System RAM resource\n");
271 static enum bp_state reserve_additional_memory(void)
274 struct resource *resource;
276 unsigned long balloon_hotplug;
278 credit = balloon_stats.target_pages + balloon_stats.target_unpopulated
279 - balloon_stats.total_pages;
282 * Already hotplugged enough pages? Wait for them to be
288 balloon_hotplug = round_up(credit, PAGES_PER_SECTION);
290 resource = additional_memory_resource(balloon_hotplug * PAGE_SIZE);
294 nid = memory_add_physaddr_to_nid(resource->start);
296 #ifdef CONFIG_XEN_HAVE_PVMMU
298 * We don't support PV MMU when Linux and Xen is using
299 * different page granularity.
301 BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
304 * add_memory() will build page tables for the new memory so
305 * the p2m must contain invalid entries so the correct
306 * non-present PTEs will be written.
308 * If a failure occurs, the original (identity) p2m entries
309 * are not restored since this region is now known not to
310 * conflict with any devices.
312 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
313 unsigned long pfn, i;
315 pfn = PFN_DOWN(resource->start);
316 for (i = 0; i < balloon_hotplug; i++) {
317 if (!set_phys_to_machine(pfn + i, INVALID_P2M_ENTRY)) {
318 pr_warn("set_phys_to_machine() failed, no memory added\n");
326 * add_memory_resource() will call online_pages() which in its turn
327 * will call xen_online_page() callback causing deadlock if we don't
328 * release balloon_mutex here. Unlocking here is safe because the
329 * callers drop the mutex before trying again.
331 mutex_unlock(&balloon_mutex);
332 /* add_memory_resource() requires the device_hotplug lock */
333 lock_device_hotplug();
334 rc = add_memory_resource(nid, resource, MHP_MERGE_RESOURCE);
335 unlock_device_hotplug();
336 mutex_lock(&balloon_mutex);
339 pr_warn("Cannot add additional memory (%i)\n", rc);
343 balloon_stats.total_pages += balloon_hotplug;
347 release_memory_resource(resource);
351 static void xen_online_page(struct page *page, unsigned int order)
353 unsigned long i, size = (1 << order);
354 unsigned long start_pfn = page_to_pfn(page);
357 pr_debug("Online %lu pages starting at pfn 0x%lx\n", size, start_pfn);
358 mutex_lock(&balloon_mutex);
359 for (i = 0; i < size; i++) {
360 p = pfn_to_page(start_pfn + i);
363 mutex_unlock(&balloon_mutex);
366 static int xen_memory_notifier(struct notifier_block *nb, unsigned long val, void *v)
368 if (val == MEM_ONLINE)
369 wake_up(&balloon_thread_wq);
374 static struct notifier_block xen_memory_nb = {
375 .notifier_call = xen_memory_notifier,
379 static enum bp_state reserve_additional_memory(void)
381 balloon_stats.target_pages = balloon_stats.current_pages +
382 balloon_stats.target_unpopulated;
385 #endif /* CONFIG_XEN_BALLOON_MEMORY_HOTPLUG */
387 static long current_credit(void)
389 return balloon_stats.target_pages - balloon_stats.current_pages;
392 static bool balloon_is_inflated(void)
394 return balloon_stats.balloon_low || balloon_stats.balloon_high;
397 static enum bp_state increase_reservation(unsigned long nr_pages)
403 if (nr_pages > ARRAY_SIZE(frame_list))
404 nr_pages = ARRAY_SIZE(frame_list);
406 page = list_first_entry_or_null(&ballooned_pages, struct page, lru);
407 for (i = 0; i < nr_pages; i++) {
413 frame_list[i] = page_to_xen_pfn(page);
414 page = balloon_next_page(page);
417 rc = xenmem_reservation_increase(nr_pages, frame_list);
421 for (i = 0; i < rc; i++) {
422 page = balloon_retrieve(false);
423 BUG_ON(page == NULL);
425 xenmem_reservation_va_mapping_update(1, &page, &frame_list[i]);
427 /* Relinquish the page back to the allocator. */
428 free_reserved_page(page);
431 balloon_stats.current_pages += rc;
436 static enum bp_state decrease_reservation(unsigned long nr_pages, gfp_t gfp)
438 enum bp_state state = BP_DONE;
440 struct page *page, *tmp;
444 if (nr_pages > ARRAY_SIZE(frame_list))
445 nr_pages = ARRAY_SIZE(frame_list);
447 for (i = 0; i < nr_pages; i++) {
448 page = alloc_page(gfp);
454 adjust_managed_page_count(page, -1);
455 xenmem_reservation_scrub_page(page);
456 list_add(&page->lru, &pages);
460 * Ensure that ballooned highmem pages don't have kmaps.
462 * Do this before changing the p2m as kmap_flush_unused()
463 * reads PTEs to obtain pages (and hence needs the original
469 * Setup the frame, update direct mapping, invalidate P2M,
470 * and add to balloon.
473 list_for_each_entry_safe(page, tmp, &pages, lru) {
474 frame_list[i++] = xen_page_to_gfn(page);
476 xenmem_reservation_va_mapping_reset(1, &page);
478 list_del(&page->lru);
480 balloon_append(page);
485 ret = xenmem_reservation_decrease(nr_pages, frame_list);
486 BUG_ON(ret != nr_pages);
488 balloon_stats.current_pages -= nr_pages;
494 * Stop waiting if either state is not BP_EAGAIN and ballooning action is
495 * needed, or if the credit has changed while state is BP_EAGAIN.
497 static bool balloon_thread_cond(enum bp_state state, long credit)
499 if (state != BP_EAGAIN)
502 return current_credit() != credit || kthread_should_stop();
506 * As this is a kthread it is guaranteed to run as a single instance only.
507 * We may of course race updates of the target counts (which are protected
508 * by the balloon lock), or with changes to the Xen hard limit, but we will
509 * recover from these in time.
511 static int balloon_thread(void *unused)
513 enum bp_state state = BP_DONE;
515 unsigned long timeout;
519 if (state == BP_EAGAIN)
520 timeout = balloon_stats.schedule_delay * HZ;
523 credit = current_credit();
525 wait_event_freezable_timeout(balloon_thread_wq,
526 balloon_thread_cond(state, credit), timeout);
528 if (kthread_should_stop())
531 mutex_lock(&balloon_mutex);
533 credit = current_credit();
536 if (balloon_is_inflated())
537 state = increase_reservation(credit);
539 state = reserve_additional_memory();
545 n_pages = min(-credit, si_mem_available());
546 state = decrease_reservation(n_pages, GFP_BALLOON);
547 if (state == BP_DONE && n_pages != -credit &&
548 n_pages < totalreserve_pages)
552 state = update_schedule(state);
554 mutex_unlock(&balloon_mutex);
560 /* Resets the Xen limit, sets new target, and kicks off processing. */
561 void balloon_set_new_target(unsigned long target)
563 /* No need for lock. Not read-modify-write updates. */
564 balloon_stats.target_pages = target;
565 wake_up(&balloon_thread_wq);
567 EXPORT_SYMBOL_GPL(balloon_set_new_target);
569 static int add_ballooned_pages(int nr_pages)
573 if (xen_hotplug_unpopulated) {
574 st = reserve_additional_memory();
575 if (st != BP_ECANCELED) {
578 mutex_unlock(&balloon_mutex);
579 rc = wait_event_interruptible(balloon_wq,
580 !list_empty(&ballooned_pages));
581 mutex_lock(&balloon_mutex);
582 return rc ? -ENOMEM : 0;
586 if (si_mem_available() < nr_pages)
589 st = decrease_reservation(nr_pages, GFP_USER);
597 * alloc_xenballooned_pages - get pages that have been ballooned out
598 * @nr_pages: Number of pages to get
599 * @pages: pages returned
600 * @return 0 on success, error otherwise
602 int alloc_xenballooned_pages(int nr_pages, struct page **pages)
608 mutex_lock(&balloon_mutex);
610 balloon_stats.target_unpopulated += nr_pages;
612 while (pgno < nr_pages) {
613 page = balloon_retrieve(true);
615 pages[pgno++] = page;
616 #ifdef CONFIG_XEN_HAVE_PVMMU
618 * We don't support PV MMU when Linux and Xen is using
619 * different page granularity.
621 BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
623 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
624 ret = xen_alloc_p2m_entry(page_to_pfn(page));
630 ret = add_ballooned_pages(nr_pages - pgno);
635 mutex_unlock(&balloon_mutex);
638 mutex_unlock(&balloon_mutex);
639 free_xenballooned_pages(pgno, pages);
641 * NB: free_xenballooned_pages will only subtract pgno pages, but since
642 * target_unpopulated is incremented with nr_pages at the start we need
643 * to remove the remaining ones also, or accounting will be screwed.
645 balloon_stats.target_unpopulated -= nr_pages - pgno;
648 EXPORT_SYMBOL(alloc_xenballooned_pages);
651 * free_xenballooned_pages - return pages retrieved with get_ballooned_pages
652 * @nr_pages: Number of pages
653 * @pages: pages to return
655 void free_xenballooned_pages(int nr_pages, struct page **pages)
659 mutex_lock(&balloon_mutex);
661 for (i = 0; i < nr_pages; i++) {
663 balloon_append(pages[i]);
666 balloon_stats.target_unpopulated -= nr_pages;
668 /* The balloon may be too large now. Shrink it if needed. */
669 if (current_credit())
670 wake_up(&balloon_thread_wq);
672 mutex_unlock(&balloon_mutex);
674 EXPORT_SYMBOL(free_xenballooned_pages);
676 #if defined(CONFIG_XEN_PV) && !defined(CONFIG_XEN_UNPOPULATED_ALLOC)
677 static void __init balloon_add_region(unsigned long start_pfn,
680 unsigned long pfn, extra_pfn_end;
683 * If the amount of usable memory has been limited (e.g., with
684 * the 'mem' command line parameter), don't add pages beyond
687 extra_pfn_end = min(max_pfn, start_pfn + pages);
689 for (pfn = start_pfn; pfn < extra_pfn_end; pfn++) {
690 /* totalram_pages and totalhigh_pages do not
691 include the boot-time balloon extension, so
692 don't subtract from it. */
693 balloon_append(pfn_to_page(pfn));
696 balloon_stats.total_pages += extra_pfn_end - start_pfn;
700 static int __init balloon_init(void)
702 struct task_struct *task;
707 pr_info("Initialising balloon driver\n");
710 balloon_stats.current_pages = xen_pv_domain()
711 ? min(xen_start_info->nr_pages - xen_released_pages, max_pfn)
712 : get_num_physpages();
714 balloon_stats.current_pages = get_num_physpages();
716 balloon_stats.target_pages = balloon_stats.current_pages;
717 balloon_stats.balloon_low = 0;
718 balloon_stats.balloon_high = 0;
719 balloon_stats.total_pages = balloon_stats.current_pages;
721 balloon_stats.schedule_delay = 1;
722 balloon_stats.max_schedule_delay = 32;
723 balloon_stats.retry_count = 1;
724 balloon_stats.max_retry_count = 4;
726 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
727 set_online_page_callback(&xen_online_page);
728 register_memory_notifier(&xen_memory_nb);
729 register_sysctl_table(xen_root);
732 #if defined(CONFIG_XEN_PV) && !defined(CONFIG_XEN_UNPOPULATED_ALLOC)
737 * Initialize the balloon with pages from the extra memory
738 * regions (see arch/x86/xen/setup.c).
740 for (i = 0; i < XEN_EXTRA_MEM_MAX_REGIONS; i++)
741 if (xen_extra_mem[i].n_pfns)
742 balloon_add_region(xen_extra_mem[i].start_pfn,
743 xen_extra_mem[i].n_pfns);
747 task = kthread_run(balloon_thread, NULL, "xen-balloon");
749 pr_err("xen-balloon thread could not be started, ballooning will not work!\n");
750 return PTR_ERR(task);
753 /* Init the xen-balloon driver. */
758 subsys_initcall(balloon_init);