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>
60 #include <asm/pgalloc.h>
61 #include <asm/pgtable.h>
64 #include <asm/xen/hypervisor.h>
65 #include <asm/xen/hypercall.h>
68 #include <xen/interface/xen.h>
69 #include <xen/interface/memory.h>
70 #include <xen/balloon.h>
71 #include <xen/features.h>
75 * balloon_process() state:
77 * BP_DONE: done or nothing to do,
78 * BP_WAIT: wait to be rescheduled,
79 * BP_EAGAIN: error, go to sleep,
80 * BP_ECANCELED: error, balloon operation canceled.
91 static DEFINE_MUTEX(balloon_mutex);
93 struct balloon_stats balloon_stats;
94 EXPORT_SYMBOL_GPL(balloon_stats);
96 /* We increase/decrease in batches which fit in a page */
97 static xen_pfn_t frame_list[PAGE_SIZE / sizeof(unsigned long)];
100 /* List of ballooned pages, threaded through the mem_map array. */
101 static LIST_HEAD(ballooned_pages);
103 /* Main work function, always executed in process context. */
104 static void balloon_process(struct work_struct *work);
105 static DECLARE_DELAYED_WORK(balloon_worker, balloon_process);
107 /* When ballooning out (allocating memory to return to Xen) we don't really
108 want the kernel to try too hard since that can trigger the oom killer. */
109 #define GFP_BALLOON \
110 (GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC)
112 static void scrub_page(struct page *page)
114 #ifdef CONFIG_XEN_SCRUB_PAGES
115 clear_highpage(page);
119 /* balloon_append: add the given page to the balloon. */
120 static void __balloon_append(struct page *page)
122 /* Lowmem is re-populated first, so highmem pages go at list tail. */
123 if (PageHighMem(page)) {
124 list_add_tail(&page->lru, &ballooned_pages);
125 balloon_stats.balloon_high++;
127 list_add(&page->lru, &ballooned_pages);
128 balloon_stats.balloon_low++;
132 static void balloon_append(struct page *page)
134 __balloon_append(page);
135 adjust_managed_page_count(page, -1);
138 /* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
139 static struct page *balloon_retrieve(bool require_lowmem)
143 if (list_empty(&ballooned_pages))
146 page = list_entry(ballooned_pages.next, struct page, lru);
147 if (require_lowmem && PageHighMem(page))
149 list_del(&page->lru);
151 if (PageHighMem(page))
152 balloon_stats.balloon_high--;
154 balloon_stats.balloon_low--;
156 adjust_managed_page_count(page, 1);
161 static struct page *balloon_next_page(struct page *page)
163 struct list_head *next = page->lru.next;
164 if (next == &ballooned_pages)
166 return list_entry(next, struct page, lru);
169 static enum bp_state update_schedule(enum bp_state state)
171 if (state == BP_WAIT)
174 if (state == BP_ECANCELED)
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 struct resource *additional_memory_resource(phys_addr_t size)
203 struct resource *res;
206 res = kzalloc(sizeof(*res), GFP_KERNEL);
210 res->name = "System RAM";
211 res->flags = IORESOURCE_MEM | IORESOURCE_BUSY;
213 ret = allocate_resource(&iomem_resource, res,
215 PAGES_PER_SECTION * PAGE_SIZE, NULL, NULL);
217 pr_err("Cannot allocate new System RAM resource\n");
225 static void release_memory_resource(struct resource *resource)
231 * No need to reset region to identity mapped since we now
232 * know that no I/O can be in this region
234 release_resource(resource);
238 static enum bp_state reserve_additional_memory(void)
241 struct resource *resource;
243 unsigned long balloon_hotplug;
245 credit = balloon_stats.target_pages - balloon_stats.total_pages;
248 * Already hotplugged enough pages? Wait for them to be
254 balloon_hotplug = round_up(credit, PAGES_PER_SECTION);
256 resource = additional_memory_resource(balloon_hotplug * PAGE_SIZE);
260 nid = memory_add_physaddr_to_nid(resource->start);
262 #ifdef CONFIG_XEN_HAVE_PVMMU
264 * add_memory() will build page tables for the new memory so
265 * the p2m must contain invalid entries so the correct
266 * non-present PTEs will be written.
268 * If a failure occurs, the original (identity) p2m entries
269 * are not restored since this region is now known not to
270 * conflict with any devices.
272 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
273 unsigned long pfn, i;
275 pfn = PFN_DOWN(resource->start);
276 for (i = 0; i < balloon_hotplug; i++) {
277 if (!set_phys_to_machine(pfn + i, INVALID_P2M_ENTRY)) {
278 pr_warn("set_phys_to_machine() failed, no memory added\n");
285 rc = add_memory_resource(nid, resource);
287 pr_warn("Cannot add additional memory (%i)\n", rc);
291 balloon_stats.total_pages += balloon_hotplug;
295 release_memory_resource(resource);
299 static void xen_online_page(struct page *page)
301 __online_page_set_limits(page);
303 mutex_lock(&balloon_mutex);
305 __balloon_append(page);
307 mutex_unlock(&balloon_mutex);
310 static int xen_memory_notifier(struct notifier_block *nb, unsigned long val, void *v)
312 if (val == MEM_ONLINE)
313 schedule_delayed_work(&balloon_worker, 0);
318 static struct notifier_block xen_memory_nb = {
319 .notifier_call = xen_memory_notifier,
323 static enum bp_state reserve_additional_memory(void)
325 balloon_stats.target_pages = balloon_stats.current_pages;
328 #endif /* CONFIG_XEN_BALLOON_MEMORY_HOTPLUG */
330 static long current_credit(void)
332 return balloon_stats.target_pages - balloon_stats.current_pages;
335 static bool balloon_is_inflated(void)
337 return balloon_stats.balloon_low || balloon_stats.balloon_high;
340 static enum bp_state increase_reservation(unsigned long nr_pages)
343 unsigned long pfn, i;
345 struct xen_memory_reservation reservation = {
351 if (nr_pages > ARRAY_SIZE(frame_list))
352 nr_pages = ARRAY_SIZE(frame_list);
354 page = list_first_entry_or_null(&ballooned_pages, struct page, lru);
355 for (i = 0; i < nr_pages; i++) {
360 frame_list[i] = page_to_pfn(page);
361 page = balloon_next_page(page);
364 set_xen_guest_handle(reservation.extent_start, frame_list);
365 reservation.nr_extents = nr_pages;
366 rc = HYPERVISOR_memory_op(XENMEM_populate_physmap, &reservation);
370 for (i = 0; i < rc; i++) {
371 page = balloon_retrieve(false);
372 BUG_ON(page == NULL);
374 pfn = page_to_pfn(page);
376 #ifdef CONFIG_XEN_HAVE_PVMMU
377 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
378 set_phys_to_machine(pfn, frame_list[i]);
380 /* Link back into the page tables if not highmem. */
381 if (!PageHighMem(page)) {
383 ret = HYPERVISOR_update_va_mapping(
384 (unsigned long)__va(pfn << PAGE_SHIFT),
385 mfn_pte(frame_list[i], PAGE_KERNEL),
392 /* Relinquish the page back to the allocator. */
393 __free_reserved_page(page);
396 balloon_stats.current_pages += rc;
401 static enum bp_state decrease_reservation(unsigned long nr_pages, gfp_t gfp)
403 enum bp_state state = BP_DONE;
404 unsigned long pfn, i;
407 struct xen_memory_reservation reservation = {
413 if (nr_pages > ARRAY_SIZE(frame_list))
414 nr_pages = ARRAY_SIZE(frame_list);
416 for (i = 0; i < nr_pages; i++) {
417 page = alloc_page(gfp);
425 frame_list[i] = page_to_pfn(page);
429 * Ensure that ballooned highmem pages don't have kmaps.
431 * Do this before changing the p2m as kmap_flush_unused()
432 * reads PTEs to obtain pages (and hence needs the original
437 /* Update direct mapping, invalidate P2M, and add to balloon. */
438 for (i = 0; i < nr_pages; i++) {
440 frame_list[i] = pfn_to_gfn(pfn);
441 page = pfn_to_page(pfn);
443 #ifdef CONFIG_XEN_HAVE_PVMMU
444 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
445 if (!PageHighMem(page)) {
446 ret = HYPERVISOR_update_va_mapping(
447 (unsigned long)__va(pfn << PAGE_SHIFT),
451 __set_phys_to_machine(pfn, INVALID_P2M_ENTRY);
455 balloon_append(page);
460 set_xen_guest_handle(reservation.extent_start, frame_list);
461 reservation.nr_extents = nr_pages;
462 ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation, &reservation);
463 BUG_ON(ret != nr_pages);
465 balloon_stats.current_pages -= nr_pages;
471 * As this is a work item it is guaranteed to run as a single instance only.
472 * We may of course race updates of the target counts (which are protected
473 * by the balloon lock), or with changes to the Xen hard limit, but we will
474 * recover from these in time.
476 static void balloon_process(struct work_struct *work)
478 enum bp_state state = BP_DONE;
483 mutex_lock(&balloon_mutex);
485 credit = current_credit();
488 if (balloon_is_inflated())
489 state = increase_reservation(credit);
491 state = reserve_additional_memory();
495 state = decrease_reservation(-credit, GFP_BALLOON);
497 state = update_schedule(state);
499 mutex_unlock(&balloon_mutex);
503 } while (credit && state == BP_DONE);
505 /* Schedule more work if there is some still to be done. */
506 if (state == BP_EAGAIN)
507 schedule_delayed_work(&balloon_worker, balloon_stats.schedule_delay * HZ);
510 /* Resets the Xen limit, sets new target, and kicks off processing. */
511 void balloon_set_new_target(unsigned long target)
513 /* No need for lock. Not read-modify-write updates. */
514 balloon_stats.target_pages = target;
515 schedule_delayed_work(&balloon_worker, 0);
517 EXPORT_SYMBOL_GPL(balloon_set_new_target);
520 * alloc_xenballooned_pages - get pages that have been ballooned out
521 * @nr_pages: Number of pages to get
522 * @pages: pages returned
523 * @return 0 on success, error otherwise
525 int alloc_xenballooned_pages(int nr_pages, struct page **pages)
529 mutex_lock(&balloon_mutex);
530 while (pgno < nr_pages) {
531 page = balloon_retrieve(true);
533 pages[pgno++] = page;
536 st = decrease_reservation(nr_pages - pgno, GFP_USER);
541 mutex_unlock(&balloon_mutex);
545 balloon_append(pages[--pgno]);
546 /* Free the memory back to the kernel soon */
547 schedule_delayed_work(&balloon_worker, 0);
548 mutex_unlock(&balloon_mutex);
551 EXPORT_SYMBOL(alloc_xenballooned_pages);
554 * free_xenballooned_pages - return pages retrieved with get_ballooned_pages
555 * @nr_pages: Number of pages
556 * @pages: pages to return
558 void free_xenballooned_pages(int nr_pages, struct page **pages)
562 mutex_lock(&balloon_mutex);
564 for (i = 0; i < nr_pages; i++) {
566 balloon_append(pages[i]);
569 /* The balloon may be too large now. Shrink it if needed. */
570 if (current_credit())
571 schedule_delayed_work(&balloon_worker, 0);
573 mutex_unlock(&balloon_mutex);
575 EXPORT_SYMBOL(free_xenballooned_pages);
577 static void __init balloon_add_region(unsigned long start_pfn,
580 unsigned long pfn, extra_pfn_end;
584 * If the amount of usable memory has been limited (e.g., with
585 * the 'mem' command line parameter), don't add pages beyond
588 extra_pfn_end = min(max_pfn, start_pfn + pages);
590 for (pfn = start_pfn; pfn < extra_pfn_end; pfn++) {
591 page = pfn_to_page(pfn);
592 /* totalram_pages and totalhigh_pages do not
593 include the boot-time balloon extension, so
594 don't subtract from it. */
595 __balloon_append(page);
598 balloon_stats.total_pages += extra_pfn_end - start_pfn;
601 static int __init balloon_init(void)
608 pr_info("Initialising balloon driver\n");
610 balloon_stats.current_pages = xen_pv_domain()
611 ? min(xen_start_info->nr_pages - xen_released_pages, max_pfn)
612 : get_num_physpages();
613 balloon_stats.target_pages = balloon_stats.current_pages;
614 balloon_stats.balloon_low = 0;
615 balloon_stats.balloon_high = 0;
616 balloon_stats.total_pages = balloon_stats.current_pages;
618 balloon_stats.schedule_delay = 1;
619 balloon_stats.max_schedule_delay = 32;
620 balloon_stats.retry_count = 1;
621 balloon_stats.max_retry_count = RETRY_UNLIMITED;
623 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
624 set_online_page_callback(&xen_online_page);
625 register_memory_notifier(&xen_memory_nb);
629 * Initialize the balloon with pages from the extra memory
630 * regions (see arch/x86/xen/setup.c).
632 for (i = 0; i < XEN_EXTRA_MEM_MAX_REGIONS; i++)
633 if (xen_extra_mem[i].n_pfns)
634 balloon_add_region(xen_extra_mem[i].start_pfn,
635 xen_extra_mem[i].n_pfns);
640 subsys_initcall(balloon_init);
642 MODULE_LICENSE("GPL");