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 #include <linux/kernel.h>
40 #include <linux/sched.h>
41 #include <linux/errno.h>
43 #include <linux/bootmem.h>
44 #include <linux/pagemap.h>
45 #include <linux/highmem.h>
46 #include <linux/mutex.h>
47 #include <linux/list.h>
48 #include <linux/gfp.h>
49 #include <linux/notifier.h>
50 #include <linux/memory.h>
51 #include <linux/memory_hotplug.h>
54 #include <asm/pgalloc.h>
55 #include <asm/pgtable.h>
59 #include <asm/xen/hypervisor.h>
60 #include <asm/xen/hypercall.h>
63 #include <xen/interface/xen.h>
64 #include <xen/interface/memory.h>
65 #include <xen/balloon.h>
66 #include <xen/features.h>
70 * balloon_process() state:
72 * BP_DONE: done or nothing to do,
73 * BP_EAGAIN: error, go to sleep,
74 * BP_ECANCELED: error, balloon operation canceled.
84 static DEFINE_MUTEX(balloon_mutex);
86 struct balloon_stats balloon_stats;
87 EXPORT_SYMBOL_GPL(balloon_stats);
89 /* We increase/decrease in batches which fit in a page */
90 static unsigned long frame_list[PAGE_SIZE / sizeof(unsigned long)];
93 #define inc_totalhigh_pages() (totalhigh_pages++)
94 #define dec_totalhigh_pages() (totalhigh_pages--)
96 #define inc_totalhigh_pages() do {} while (0)
97 #define dec_totalhigh_pages() do {} while (0)
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 if (PageHighMem(page))
136 dec_totalhigh_pages();
140 /* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
141 static struct page *balloon_retrieve(bool prefer_highmem)
145 if (list_empty(&ballooned_pages))
149 page = list_entry(ballooned_pages.prev, struct page, lru);
151 page = list_entry(ballooned_pages.next, struct page, lru);
152 list_del(&page->lru);
154 if (PageHighMem(page)) {
155 balloon_stats.balloon_high--;
156 inc_totalhigh_pages();
158 balloon_stats.balloon_low--;
165 static struct page *balloon_first_page(void)
167 if (list_empty(&ballooned_pages))
169 return list_entry(ballooned_pages.next, struct page, lru);
172 static struct page *balloon_next_page(struct page *page)
174 struct list_head *next = page->lru.next;
175 if (next == &ballooned_pages)
177 return list_entry(next, struct page, lru);
180 static enum bp_state update_schedule(enum bp_state state)
182 if (state == BP_DONE) {
183 balloon_stats.schedule_delay = 1;
184 balloon_stats.retry_count = 1;
188 ++balloon_stats.retry_count;
190 if (balloon_stats.max_retry_count != RETRY_UNLIMITED &&
191 balloon_stats.retry_count > balloon_stats.max_retry_count) {
192 balloon_stats.schedule_delay = 1;
193 balloon_stats.retry_count = 1;
197 balloon_stats.schedule_delay <<= 1;
199 if (balloon_stats.schedule_delay > balloon_stats.max_schedule_delay)
200 balloon_stats.schedule_delay = balloon_stats.max_schedule_delay;
205 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
206 static long current_credit(void)
208 return balloon_stats.target_pages - balloon_stats.current_pages -
209 balloon_stats.hotplug_pages;
212 static bool balloon_is_inflated(void)
214 if (balloon_stats.balloon_low || balloon_stats.balloon_high ||
215 balloon_stats.balloon_hotplug)
222 * reserve_additional_memory() adds memory region of size >= credit above
223 * max_pfn. New region is section aligned and size is modified to be multiple
224 * of section size. Those features allow optimal use of address space and
225 * establish proper alignment when this function is called first time after
226 * boot (last section not fully populated at boot time contains unused memory
227 * pages with PG_reserved bit not set; online_pages_range() does not allow page
228 * onlining in whole range if first onlined page does not have PG_reserved
229 * bit set). Real size of added memory is established at page onlining stage.
232 static enum bp_state reserve_additional_memory(long credit)
235 u64 hotplug_start_paddr;
236 unsigned long balloon_hotplug = credit;
238 hotplug_start_paddr = PFN_PHYS(SECTION_ALIGN_UP(max_pfn));
239 balloon_hotplug = round_up(balloon_hotplug, PAGES_PER_SECTION);
240 nid = memory_add_physaddr_to_nid(hotplug_start_paddr);
242 rc = add_memory(nid, hotplug_start_paddr, balloon_hotplug << PAGE_SHIFT);
245 pr_info("xen_balloon: %s: add_memory() failed: %i\n", __func__, rc);
249 balloon_hotplug -= credit;
251 balloon_stats.hotplug_pages += credit;
252 balloon_stats.balloon_hotplug = balloon_hotplug;
257 static void xen_online_page(struct page *page)
259 __online_page_set_limits(page);
261 mutex_lock(&balloon_mutex);
263 __balloon_append(page);
265 if (balloon_stats.hotplug_pages)
266 --balloon_stats.hotplug_pages;
268 --balloon_stats.balloon_hotplug;
270 mutex_unlock(&balloon_mutex);
273 static int xen_memory_notifier(struct notifier_block *nb, unsigned long val, void *v)
275 if (val == MEM_ONLINE)
276 schedule_delayed_work(&balloon_worker, 0);
281 static struct notifier_block xen_memory_nb = {
282 .notifier_call = xen_memory_notifier,
286 static long current_credit(void)
288 unsigned long target = balloon_stats.target_pages;
291 balloon_stats.current_pages +
292 balloon_stats.balloon_low +
293 balloon_stats.balloon_high);
295 return target - balloon_stats.current_pages;
298 static bool balloon_is_inflated(void)
300 if (balloon_stats.balloon_low || balloon_stats.balloon_high)
306 static enum bp_state reserve_additional_memory(long credit)
308 balloon_stats.target_pages = balloon_stats.current_pages;
311 #endif /* CONFIG_XEN_BALLOON_MEMORY_HOTPLUG */
313 static enum bp_state increase_reservation(unsigned long nr_pages)
316 unsigned long pfn, i;
318 struct xen_memory_reservation reservation = {
324 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
325 if (!balloon_stats.balloon_low && !balloon_stats.balloon_high) {
326 nr_pages = min(nr_pages, balloon_stats.balloon_hotplug);
327 balloon_stats.hotplug_pages += nr_pages;
328 balloon_stats.balloon_hotplug -= nr_pages;
333 if (nr_pages > ARRAY_SIZE(frame_list))
334 nr_pages = ARRAY_SIZE(frame_list);
336 page = balloon_first_page();
337 for (i = 0; i < nr_pages; i++) {
342 frame_list[i] = page_to_pfn(page);
343 page = balloon_next_page(page);
346 set_xen_guest_handle(reservation.extent_start, frame_list);
347 reservation.nr_extents = nr_pages;
348 rc = HYPERVISOR_memory_op(XENMEM_populate_physmap, &reservation);
352 for (i = 0; i < rc; i++) {
353 page = balloon_retrieve(false);
354 BUG_ON(page == NULL);
356 pfn = page_to_pfn(page);
357 BUG_ON(!xen_feature(XENFEAT_auto_translated_physmap) &&
358 phys_to_machine_mapping_valid(pfn));
360 set_phys_to_machine(pfn, frame_list[i]);
362 /* Link back into the page tables if not highmem. */
363 if (xen_pv_domain() && !PageHighMem(page)) {
365 ret = HYPERVISOR_update_va_mapping(
366 (unsigned long)__va(pfn << PAGE_SHIFT),
367 mfn_pte(frame_list[i], PAGE_KERNEL),
372 /* Relinquish the page back to the allocator. */
373 ClearPageReserved(page);
374 init_page_count(page);
378 balloon_stats.current_pages += rc;
383 static enum bp_state decrease_reservation(unsigned long nr_pages, gfp_t gfp)
385 enum bp_state state = BP_DONE;
386 unsigned long pfn, i;
389 struct xen_memory_reservation reservation = {
395 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
396 if (balloon_stats.hotplug_pages) {
397 nr_pages = min(nr_pages, balloon_stats.hotplug_pages);
398 balloon_stats.hotplug_pages -= nr_pages;
399 balloon_stats.balloon_hotplug += nr_pages;
404 if (nr_pages > ARRAY_SIZE(frame_list))
405 nr_pages = ARRAY_SIZE(frame_list);
407 for (i = 0; i < nr_pages; i++) {
408 if ((page = alloc_page(gfp)) == NULL) {
414 pfn = page_to_pfn(page);
415 frame_list[i] = pfn_to_mfn(pfn);
419 if (xen_pv_domain() && !PageHighMem(page)) {
420 ret = HYPERVISOR_update_va_mapping(
421 (unsigned long)__va(pfn << PAGE_SHIFT),
428 /* Ensure that ballooned highmem pages don't have kmaps. */
432 /* No more mappings: invalidate P2M and add to balloon. */
433 for (i = 0; i < nr_pages; i++) {
434 pfn = mfn_to_pfn(frame_list[i]);
435 __set_phys_to_machine(pfn, INVALID_P2M_ENTRY);
436 balloon_append(pfn_to_page(pfn));
439 set_xen_guest_handle(reservation.extent_start, frame_list);
440 reservation.nr_extents = nr_pages;
441 ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation, &reservation);
442 BUG_ON(ret != nr_pages);
444 balloon_stats.current_pages -= nr_pages;
450 * We avoid multiple worker processes conflicting via the balloon mutex.
451 * We may of course race updates of the target counts (which are protected
452 * by the balloon lock), or with changes to the Xen hard limit, but we will
453 * recover from these in time.
455 static void balloon_process(struct work_struct *work)
457 enum bp_state state = BP_DONE;
460 mutex_lock(&balloon_mutex);
463 credit = current_credit();
466 if (balloon_is_inflated())
467 state = increase_reservation(credit);
469 state = reserve_additional_memory(credit);
473 state = decrease_reservation(-credit, GFP_BALLOON);
475 state = update_schedule(state);
477 #ifndef CONFIG_PREEMPT
481 } while (credit && state == BP_DONE);
483 /* Schedule more work if there is some still to be done. */
484 if (state == BP_EAGAIN)
485 schedule_delayed_work(&balloon_worker, balloon_stats.schedule_delay * HZ);
487 mutex_unlock(&balloon_mutex);
490 /* Resets the Xen limit, sets new target, and kicks off processing. */
491 void balloon_set_new_target(unsigned long target)
493 /* No need for lock. Not read-modify-write updates. */
494 balloon_stats.target_pages = target;
495 schedule_delayed_work(&balloon_worker, 0);
497 EXPORT_SYMBOL_GPL(balloon_set_new_target);
500 * alloc_xenballooned_pages - get pages that have been ballooned out
501 * @nr_pages: Number of pages to get
502 * @pages: pages returned
503 * @return 0 on success, error otherwise
505 int alloc_xenballooned_pages(int nr_pages, struct page **pages)
509 mutex_lock(&balloon_mutex);
510 while (pgno < nr_pages) {
511 page = balloon_retrieve(true);
513 pages[pgno++] = page;
516 st = decrease_reservation(nr_pages - pgno, GFP_HIGHUSER);
521 mutex_unlock(&balloon_mutex);
525 balloon_append(pages[--pgno]);
526 /* Free the memory back to the kernel soon */
527 schedule_delayed_work(&balloon_worker, 0);
528 mutex_unlock(&balloon_mutex);
531 EXPORT_SYMBOL(alloc_xenballooned_pages);
534 * free_xenballooned_pages - return pages retrieved with get_ballooned_pages
535 * @nr_pages: Number of pages
536 * @pages: pages to return
538 void free_xenballooned_pages(int nr_pages, struct page **pages)
542 mutex_lock(&balloon_mutex);
544 for (i = 0; i < nr_pages; i++) {
546 balloon_append(pages[i]);
549 /* The balloon may be too large now. Shrink it if needed. */
550 if (current_credit())
551 schedule_delayed_work(&balloon_worker, 0);
553 mutex_unlock(&balloon_mutex);
555 EXPORT_SYMBOL(free_xenballooned_pages);
557 static int __init balloon_init(void)
559 unsigned long pfn, extra_pfn_end;
565 pr_info("xen/balloon: Initialising balloon driver.\n");
567 balloon_stats.current_pages = xen_pv_domain() ? min(xen_start_info->nr_pages, max_pfn) : max_pfn;
568 balloon_stats.target_pages = balloon_stats.current_pages;
569 balloon_stats.balloon_low = 0;
570 balloon_stats.balloon_high = 0;
572 balloon_stats.schedule_delay = 1;
573 balloon_stats.max_schedule_delay = 32;
574 balloon_stats.retry_count = 1;
575 balloon_stats.max_retry_count = RETRY_UNLIMITED;
577 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
578 balloon_stats.hotplug_pages = 0;
579 balloon_stats.balloon_hotplug = 0;
581 set_online_page_callback(&xen_online_page);
582 register_memory_notifier(&xen_memory_nb);
586 * Initialise the balloon with excess memory space. We need
587 * to make sure we don't add memory which doesn't exist or
588 * logically exist. The E820 map can be trimmed to be smaller
589 * than the amount of physical memory due to the mem= command
590 * line parameter. And if this is a 32-bit non-HIGHMEM kernel
591 * on a system with memory which requires highmem to access,
592 * don't try to use it.
594 extra_pfn_end = min(min(max_pfn, e820_end_of_ram_pfn()),
595 (unsigned long)PFN_DOWN(xen_extra_mem_start + xen_extra_mem_size));
596 for (pfn = PFN_UP(xen_extra_mem_start);
599 page = pfn_to_page(pfn);
600 /* totalram_pages and totalhigh_pages do not include the boot-time
601 balloon extension, so don't subtract from it. */
602 __balloon_append(page);
608 subsys_initcall(balloon_init);
610 MODULE_LICENSE("GPL");