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/memblock.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
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_process() 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.
136 static DEFINE_MUTEX(balloon_mutex);
138 struct balloon_stats balloon_stats;
139 EXPORT_SYMBOL_GPL(balloon_stats);
141 /* We increase/decrease in batches which fit in a page */
142 static xen_pfn_t frame_list[PAGE_SIZE / sizeof(xen_pfn_t)];
145 /* List of ballooned pages, threaded through the mem_map array. */
146 static LIST_HEAD(ballooned_pages);
147 static DECLARE_WAIT_QUEUE_HEAD(balloon_wq);
149 /* Main work function, always executed in process context. */
150 static void balloon_process(struct work_struct *work);
151 static DECLARE_DELAYED_WORK(balloon_worker, balloon_process);
153 /* When ballooning out (allocating memory to return to Xen) we don't really
154 want the kernel to try too hard since that can trigger the oom killer. */
155 #define GFP_BALLOON \
156 (GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC)
158 /* balloon_append: add the given page to the balloon. */
159 static void __balloon_append(struct page *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 static void balloon_append(struct page *page)
174 __balloon_append(page);
177 /* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
178 static struct page *balloon_retrieve(bool require_lowmem)
182 if (list_empty(&ballooned_pages))
185 page = list_entry(ballooned_pages.next, struct page, lru);
186 if (require_lowmem && PageHighMem(page))
188 list_del(&page->lru);
190 if (PageHighMem(page))
191 balloon_stats.balloon_high--;
193 balloon_stats.balloon_low--;
198 static struct page *balloon_next_page(struct page *page)
200 struct list_head *next = page->lru.next;
201 if (next == &ballooned_pages)
203 return list_entry(next, struct page, lru);
206 static enum bp_state update_schedule(enum bp_state state)
208 if (state == BP_WAIT)
211 if (state == BP_ECANCELED)
214 if (state == BP_DONE) {
215 balloon_stats.schedule_delay = 1;
216 balloon_stats.retry_count = 1;
220 ++balloon_stats.retry_count;
222 if (balloon_stats.max_retry_count != RETRY_UNLIMITED &&
223 balloon_stats.retry_count > balloon_stats.max_retry_count) {
224 balloon_stats.schedule_delay = 1;
225 balloon_stats.retry_count = 1;
229 balloon_stats.schedule_delay <<= 1;
231 if (balloon_stats.schedule_delay > balloon_stats.max_schedule_delay)
232 balloon_stats.schedule_delay = balloon_stats.max_schedule_delay;
237 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
238 static void release_memory_resource(struct resource *resource)
244 * No need to reset region to identity mapped since we now
245 * know that no I/O can be in this region
247 release_resource(resource);
251 static struct resource *additional_memory_resource(phys_addr_t size)
253 struct resource *res;
256 res = kzalloc(sizeof(*res), GFP_KERNEL);
260 res->name = "System RAM";
261 res->flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
263 ret = allocate_resource(&iomem_resource, res,
265 PAGES_PER_SECTION * PAGE_SIZE, NULL, NULL);
267 pr_err("Cannot allocate new System RAM resource\n");
272 #ifdef CONFIG_SPARSEMEM
274 unsigned long limit = 1UL << (MAX_PHYSMEM_BITS - PAGE_SHIFT);
275 unsigned long pfn = res->start >> PAGE_SHIFT;
278 pr_err("New System RAM resource outside addressable RAM (%lu > %lu)\n",
280 release_memory_resource(res);
289 static enum bp_state reserve_additional_memory(void)
292 struct resource *resource;
294 unsigned long balloon_hotplug;
296 credit = balloon_stats.target_pages + balloon_stats.target_unpopulated
297 - balloon_stats.total_pages;
300 * Already hotplugged enough pages? Wait for them to be
306 balloon_hotplug = round_up(credit, PAGES_PER_SECTION);
308 resource = additional_memory_resource(balloon_hotplug * PAGE_SIZE);
312 nid = memory_add_physaddr_to_nid(resource->start);
314 #ifdef CONFIG_XEN_HAVE_PVMMU
316 * We don't support PV MMU when Linux and Xen is using
317 * different page granularity.
319 BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
322 * add_memory() will build page tables for the new memory so
323 * the p2m must contain invalid entries so the correct
324 * non-present PTEs will be written.
326 * If a failure occurs, the original (identity) p2m entries
327 * are not restored since this region is now known not to
328 * conflict with any devices.
330 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
331 unsigned long pfn, i;
333 pfn = PFN_DOWN(resource->start);
334 for (i = 0; i < balloon_hotplug; i++) {
335 if (!set_phys_to_machine(pfn + i, INVALID_P2M_ENTRY)) {
336 pr_warn("set_phys_to_machine() failed, no memory added\n");
344 * add_memory_resource() will call online_pages() which in its turn
345 * will call xen_online_page() callback causing deadlock if we don't
346 * release balloon_mutex here. Unlocking here is safe because the
347 * callers drop the mutex before trying again.
349 mutex_unlock(&balloon_mutex);
350 /* add_memory_resource() requires the device_hotplug lock */
351 lock_device_hotplug();
352 rc = add_memory_resource(nid, resource);
353 unlock_device_hotplug();
354 mutex_lock(&balloon_mutex);
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, unsigned int order)
371 unsigned long i, size = (1 << order);
372 unsigned long start_pfn = page_to_pfn(page);
375 pr_debug("Online %lu pages starting at pfn 0x%lx\n", size, start_pfn);
376 mutex_lock(&balloon_mutex);
377 for (i = 0; i < size; i++) {
378 p = pfn_to_page(start_pfn + i);
379 __online_page_set_limits(p);
383 mutex_unlock(&balloon_mutex);
386 static int xen_memory_notifier(struct notifier_block *nb, unsigned long val, void *v)
388 if (val == MEM_ONLINE)
389 schedule_delayed_work(&balloon_worker, 0);
394 static struct notifier_block xen_memory_nb = {
395 .notifier_call = xen_memory_notifier,
399 static enum bp_state reserve_additional_memory(void)
401 balloon_stats.target_pages = balloon_stats.current_pages;
404 #endif /* CONFIG_XEN_BALLOON_MEMORY_HOTPLUG */
406 static long current_credit(void)
408 return balloon_stats.target_pages - balloon_stats.current_pages;
411 static bool balloon_is_inflated(void)
413 return balloon_stats.balloon_low || balloon_stats.balloon_high;
416 static enum bp_state increase_reservation(unsigned long nr_pages)
422 if (nr_pages > ARRAY_SIZE(frame_list))
423 nr_pages = ARRAY_SIZE(frame_list);
425 page = list_first_entry_or_null(&ballooned_pages, struct page, lru);
426 for (i = 0; i < nr_pages; i++) {
432 frame_list[i] = page_to_xen_pfn(page);
433 page = balloon_next_page(page);
436 rc = xenmem_reservation_increase(nr_pages, frame_list);
440 for (i = 0; i < rc; i++) {
441 page = balloon_retrieve(false);
442 BUG_ON(page == NULL);
444 xenmem_reservation_va_mapping_update(1, &page, &frame_list[i]);
446 /* Relinquish the page back to the allocator. */
447 __ClearPageOffline(page);
448 free_reserved_page(page);
451 balloon_stats.current_pages += rc;
456 static enum bp_state decrease_reservation(unsigned long nr_pages, gfp_t gfp)
458 enum bp_state state = BP_DONE;
460 struct page *page, *tmp;
464 if (nr_pages > ARRAY_SIZE(frame_list))
465 nr_pages = ARRAY_SIZE(frame_list);
467 for (i = 0; i < nr_pages; i++) {
468 page = alloc_page(gfp);
474 __SetPageOffline(page);
475 adjust_managed_page_count(page, -1);
476 xenmem_reservation_scrub_page(page);
477 list_add(&page->lru, &pages);
481 * Ensure that ballooned highmem pages don't have kmaps.
483 * Do this before changing the p2m as kmap_flush_unused()
484 * reads PTEs to obtain pages (and hence needs the original
490 * Setup the frame, update direct mapping, invalidate P2M,
491 * and add to balloon.
494 list_for_each_entry_safe(page, tmp, &pages, lru) {
495 frame_list[i++] = xen_page_to_gfn(page);
497 xenmem_reservation_va_mapping_reset(1, &page);
499 list_del(&page->lru);
501 balloon_append(page);
506 ret = xenmem_reservation_decrease(nr_pages, frame_list);
507 BUG_ON(ret != nr_pages);
509 balloon_stats.current_pages -= nr_pages;
515 * As this is a work item it is guaranteed to run as a single instance only.
516 * We may of course race updates of the target counts (which are protected
517 * by the balloon lock), or with changes to the Xen hard limit, but we will
518 * recover from these in time.
520 static void balloon_process(struct work_struct *work)
522 enum bp_state state = BP_DONE;
527 mutex_lock(&balloon_mutex);
529 credit = current_credit();
532 if (balloon_is_inflated())
533 state = increase_reservation(credit);
535 state = reserve_additional_memory();
541 n_pages = min(-credit, si_mem_available());
542 state = decrease_reservation(n_pages, GFP_BALLOON);
543 if (state == BP_DONE && n_pages != -credit &&
544 n_pages < totalreserve_pages)
548 state = update_schedule(state);
550 mutex_unlock(&balloon_mutex);
554 } while (credit && state == BP_DONE);
556 /* Schedule more work if there is some still to be done. */
557 if (state == BP_EAGAIN)
558 schedule_delayed_work(&balloon_worker, balloon_stats.schedule_delay * HZ);
561 /* Resets the Xen limit, sets new target, and kicks off processing. */
562 void balloon_set_new_target(unsigned long target)
564 /* No need for lock. Not read-modify-write updates. */
565 balloon_stats.target_pages = target;
566 schedule_delayed_work(&balloon_worker, 0);
568 EXPORT_SYMBOL_GPL(balloon_set_new_target);
570 static int add_ballooned_pages(int nr_pages)
574 if (xen_hotplug_unpopulated) {
575 st = reserve_additional_memory();
576 if (st != BP_ECANCELED) {
577 mutex_unlock(&balloon_mutex);
578 wait_event(balloon_wq,
579 !list_empty(&ballooned_pages));
580 mutex_lock(&balloon_mutex);
585 if (si_mem_available() < nr_pages)
588 st = decrease_reservation(nr_pages, GFP_USER);
596 * alloc_xenballooned_pages - get pages that have been ballooned out
597 * @nr_pages: Number of pages to get
598 * @pages: pages returned
599 * @return 0 on success, error otherwise
601 int alloc_xenballooned_pages(int nr_pages, struct page **pages)
607 mutex_lock(&balloon_mutex);
609 balloon_stats.target_unpopulated += nr_pages;
611 while (pgno < nr_pages) {
612 page = balloon_retrieve(true);
614 __ClearPageOffline(page);
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);
642 EXPORT_SYMBOL(alloc_xenballooned_pages);
645 * free_xenballooned_pages - return pages retrieved with get_ballooned_pages
646 * @nr_pages: Number of pages
647 * @pages: pages to return
649 void free_xenballooned_pages(int nr_pages, struct page **pages)
653 mutex_lock(&balloon_mutex);
655 for (i = 0; i < nr_pages; i++) {
657 __SetPageOffline(pages[i]);
658 balloon_append(pages[i]);
662 balloon_stats.target_unpopulated -= nr_pages;
664 /* The balloon may be too large now. Shrink it if needed. */
665 if (current_credit())
666 schedule_delayed_work(&balloon_worker, 0);
668 mutex_unlock(&balloon_mutex);
670 EXPORT_SYMBOL(free_xenballooned_pages);
673 static void __init balloon_add_region(unsigned long start_pfn,
676 unsigned long pfn, extra_pfn_end;
680 * If the amount of usable memory has been limited (e.g., with
681 * the 'mem' command line parameter), don't add pages beyond
684 extra_pfn_end = min(max_pfn, start_pfn + pages);
686 for (pfn = start_pfn; pfn < extra_pfn_end; pfn++) {
687 page = pfn_to_page(pfn);
688 /* totalram_pages and totalhigh_pages do not
689 include the boot-time balloon extension, so
690 don't subtract from it. */
691 __balloon_append(page);
694 balloon_stats.total_pages += extra_pfn_end - start_pfn;
698 static int __init balloon_init(void)
703 pr_info("Initialising balloon driver\n");
706 balloon_stats.current_pages = xen_pv_domain()
707 ? min(xen_start_info->nr_pages - xen_released_pages, max_pfn)
708 : get_num_physpages();
710 balloon_stats.current_pages = get_num_physpages();
712 balloon_stats.target_pages = balloon_stats.current_pages;
713 balloon_stats.balloon_low = 0;
714 balloon_stats.balloon_high = 0;
715 balloon_stats.total_pages = balloon_stats.current_pages;
717 balloon_stats.schedule_delay = 1;
718 balloon_stats.max_schedule_delay = 32;
719 balloon_stats.retry_count = 1;
720 balloon_stats.max_retry_count = 4;
722 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
723 set_online_page_callback(&xen_online_page);
724 register_memory_notifier(&xen_memory_nb);
725 register_sysctl_table(xen_root);
733 * Initialize the balloon with pages from the extra memory
734 * regions (see arch/x86/xen/setup.c).
736 for (i = 0; i < XEN_EXTRA_MEM_MAX_REGIONS; i++)
737 if (xen_extra_mem[i].n_pfns)
738 balloon_add_region(xen_extra_mem[i].start_pfn,
739 xen_extra_mem[i].n_pfns);
743 /* Init the xen-balloon driver. */
748 subsys_initcall(balloon_init);