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>
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>
75 static int xen_hotplug_unpopulated
;
77 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
82 static struct ctl_table balloon_table
[] = {
84 .procname
= "hotplug_unpopulated",
85 .data
= &xen_hotplug_unpopulated
,
86 .maxlen
= sizeof(int),
88 .proc_handler
= proc_dointvec_minmax
,
95 static struct ctl_table balloon_root
[] = {
97 .procname
= "balloon",
99 .child
= balloon_table
,
104 static struct ctl_table xen_root
[] = {
108 .child
= balloon_root
,
116 * Use one extent per PAGE_SIZE to avoid to break down the page into
119 #define EXTENT_ORDER (fls(XEN_PFN_PER_PAGE) - 1)
122 * balloon_process() state:
124 * BP_DONE: done or nothing to do,
125 * BP_WAIT: wait to be rescheduled,
126 * BP_EAGAIN: error, go to sleep,
127 * BP_ECANCELED: error, balloon operation canceled.
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 /* Main work function, always executed in process context. */
152 static void balloon_process(struct work_struct
*work
);
153 static DECLARE_DELAYED_WORK(balloon_worker
, balloon_process
);
155 /* When ballooning out (allocating memory to return to Xen) we don't really
156 want the kernel to try too hard since that can trigger the oom killer. */
157 #define GFP_BALLOON \
158 (GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC)
160 static void scrub_page(struct page
*page
)
162 #ifdef CONFIG_XEN_SCRUB_PAGES
163 clear_highpage(page
);
167 /* balloon_append: add the given page to the balloon. */
168 static void __balloon_append(struct page
*page
)
170 /* Lowmem is re-populated first, so highmem pages go at list tail. */
171 if (PageHighMem(page
)) {
172 list_add_tail(&page
->lru
, &ballooned_pages
);
173 balloon_stats
.balloon_high
++;
175 list_add(&page
->lru
, &ballooned_pages
);
176 balloon_stats
.balloon_low
++;
178 wake_up(&balloon_wq
);
181 static void balloon_append(struct page
*page
)
183 __balloon_append(page
);
184 adjust_managed_page_count(page
, -1);
187 /* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
188 static struct page
*balloon_retrieve(bool require_lowmem
)
192 if (list_empty(&ballooned_pages
))
195 page
= list_entry(ballooned_pages
.next
, struct page
, lru
);
196 if (require_lowmem
&& PageHighMem(page
))
198 list_del(&page
->lru
);
200 if (PageHighMem(page
))
201 balloon_stats
.balloon_high
--;
203 balloon_stats
.balloon_low
--;
205 adjust_managed_page_count(page
, 1);
210 static struct page
*balloon_next_page(struct page
*page
)
212 struct list_head
*next
= page
->lru
.next
;
213 if (next
== &ballooned_pages
)
215 return list_entry(next
, struct page
, lru
);
218 static enum bp_state
update_schedule(enum bp_state state
)
220 if (state
== BP_WAIT
)
223 if (state
== BP_ECANCELED
)
226 if (state
== BP_DONE
) {
227 balloon_stats
.schedule_delay
= 1;
228 balloon_stats
.retry_count
= 1;
232 ++balloon_stats
.retry_count
;
234 if (balloon_stats
.max_retry_count
!= RETRY_UNLIMITED
&&
235 balloon_stats
.retry_count
> balloon_stats
.max_retry_count
) {
236 balloon_stats
.schedule_delay
= 1;
237 balloon_stats
.retry_count
= 1;
241 balloon_stats
.schedule_delay
<<= 1;
243 if (balloon_stats
.schedule_delay
> balloon_stats
.max_schedule_delay
)
244 balloon_stats
.schedule_delay
= balloon_stats
.max_schedule_delay
;
249 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
250 static struct resource
*additional_memory_resource(phys_addr_t size
)
252 struct resource
*res
;
255 res
= kzalloc(sizeof(*res
), GFP_KERNEL
);
259 res
->name
= "System RAM";
260 res
->flags
= IORESOURCE_MEM
| IORESOURCE_BUSY
;
262 ret
= allocate_resource(&iomem_resource
, res
,
264 PAGES_PER_SECTION
* PAGE_SIZE
, NULL
, NULL
);
266 pr_err("Cannot allocate new System RAM resource\n");
274 static void release_memory_resource(struct resource
*resource
)
280 * No need to reset region to identity mapped since we now
281 * know that no I/O can be in this region
283 release_resource(resource
);
287 static enum bp_state
reserve_additional_memory(void)
290 struct resource
*resource
;
292 unsigned long balloon_hotplug
;
294 credit
= balloon_stats
.target_pages
+ balloon_stats
.target_unpopulated
295 - balloon_stats
.total_pages
;
298 * Already hotplugged enough pages? Wait for them to be
304 balloon_hotplug
= round_up(credit
, PAGES_PER_SECTION
);
306 resource
= additional_memory_resource(balloon_hotplug
* PAGE_SIZE
);
310 nid
= memory_add_physaddr_to_nid(resource
->start
);
312 #ifdef CONFIG_XEN_HAVE_PVMMU
314 * We don't support PV MMU when Linux and Xen is using
315 * different page granularity.
317 BUILD_BUG_ON(XEN_PAGE_SIZE
!= PAGE_SIZE
);
320 * add_memory() will build page tables for the new memory so
321 * the p2m must contain invalid entries so the correct
322 * non-present PTEs will be written.
324 * If a failure occurs, the original (identity) p2m entries
325 * are not restored since this region is now known not to
326 * conflict with any devices.
328 if (!xen_feature(XENFEAT_auto_translated_physmap
)) {
329 unsigned long pfn
, i
;
331 pfn
= PFN_DOWN(resource
->start
);
332 for (i
= 0; i
< balloon_hotplug
; i
++) {
333 if (!set_phys_to_machine(pfn
+ i
, INVALID_P2M_ENTRY
)) {
334 pr_warn("set_phys_to_machine() failed, no memory added\n");
341 rc
= add_memory_resource(nid
, resource
);
343 pr_warn("Cannot add additional memory (%i)\n", rc
);
347 balloon_stats
.total_pages
+= balloon_hotplug
;
351 release_memory_resource(resource
);
355 static void xen_online_page(struct page
*page
)
357 __online_page_set_limits(page
);
359 mutex_lock(&balloon_mutex
);
361 __balloon_append(page
);
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 schedule_delayed_work(&balloon_worker
, 0);
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
;
384 #endif /* CONFIG_XEN_BALLOON_MEMORY_HOTPLUG */
386 static long current_credit(void)
388 return balloon_stats
.target_pages
- balloon_stats
.current_pages
;
391 static bool balloon_is_inflated(void)
393 return balloon_stats
.balloon_low
|| balloon_stats
.balloon_high
;
396 static enum bp_state
increase_reservation(unsigned long nr_pages
)
401 struct xen_memory_reservation reservation
= {
403 .extent_order
= EXTENT_ORDER
,
407 if (nr_pages
> ARRAY_SIZE(frame_list
))
408 nr_pages
= ARRAY_SIZE(frame_list
);
410 page
= list_first_entry_or_null(&ballooned_pages
, struct page
, lru
);
411 for (i
= 0; i
< nr_pages
; i
++) {
417 /* XENMEM_populate_physmap requires a PFN based on Xen
420 frame_list
[i
] = page_to_xen_pfn(page
);
421 page
= balloon_next_page(page
);
424 set_xen_guest_handle(reservation
.extent_start
, frame_list
);
425 reservation
.nr_extents
= nr_pages
;
426 rc
= HYPERVISOR_memory_op(XENMEM_populate_physmap
, &reservation
);
430 for (i
= 0; i
< rc
; i
++) {
431 page
= balloon_retrieve(false);
432 BUG_ON(page
== NULL
);
434 #ifdef CONFIG_XEN_HAVE_PVMMU
436 * We don't support PV MMU when Linux and Xen is using
437 * different page granularity.
439 BUILD_BUG_ON(XEN_PAGE_SIZE
!= PAGE_SIZE
);
441 if (!xen_feature(XENFEAT_auto_translated_physmap
)) {
442 unsigned long pfn
= page_to_pfn(page
);
444 set_phys_to_machine(pfn
, frame_list
[i
]);
446 /* Link back into the page tables if not highmem. */
447 if (!PageHighMem(page
)) {
449 ret
= HYPERVISOR_update_va_mapping(
450 (unsigned long)__va(pfn
<< PAGE_SHIFT
),
451 mfn_pte(frame_list
[i
], PAGE_KERNEL
),
458 /* Relinquish the page back to the allocator. */
459 __free_reserved_page(page
);
462 balloon_stats
.current_pages
+= rc
;
467 static enum bp_state
decrease_reservation(unsigned long nr_pages
, gfp_t gfp
)
469 enum bp_state state
= BP_DONE
;
471 struct page
*page
, *tmp
;
473 struct xen_memory_reservation reservation
= {
475 .extent_order
= EXTENT_ORDER
,
480 if (nr_pages
> ARRAY_SIZE(frame_list
))
481 nr_pages
= ARRAY_SIZE(frame_list
);
483 for (i
= 0; i
< nr_pages
; i
++) {
484 page
= alloc_page(gfp
);
491 list_add(&page
->lru
, &pages
);
495 * Ensure that ballooned highmem pages don't have kmaps.
497 * Do this before changing the p2m as kmap_flush_unused()
498 * reads PTEs to obtain pages (and hence needs the original
504 * Setup the frame, update direct mapping, invalidate P2M,
505 * and add to balloon.
508 list_for_each_entry_safe(page
, tmp
, &pages
, lru
) {
509 /* XENMEM_decrease_reservation requires a GFN */
510 frame_list
[i
++] = xen_page_to_gfn(page
);
512 #ifdef CONFIG_XEN_HAVE_PVMMU
514 * We don't support PV MMU when Linux and Xen is using
515 * different page granularity.
517 BUILD_BUG_ON(XEN_PAGE_SIZE
!= PAGE_SIZE
);
519 if (!xen_feature(XENFEAT_auto_translated_physmap
)) {
520 unsigned long pfn
= page_to_pfn(page
);
522 if (!PageHighMem(page
)) {
523 ret
= HYPERVISOR_update_va_mapping(
524 (unsigned long)__va(pfn
<< PAGE_SHIFT
),
528 __set_phys_to_machine(pfn
, INVALID_P2M_ENTRY
);
531 list_del(&page
->lru
);
533 balloon_append(page
);
538 set_xen_guest_handle(reservation
.extent_start
, frame_list
);
539 reservation
.nr_extents
= nr_pages
;
540 ret
= HYPERVISOR_memory_op(XENMEM_decrease_reservation
, &reservation
);
541 BUG_ON(ret
!= nr_pages
);
543 balloon_stats
.current_pages
-= nr_pages
;
549 * As this is a work item it is guaranteed to run as a single instance only.
550 * We may of course race updates of the target counts (which are protected
551 * by the balloon lock), or with changes to the Xen hard limit, but we will
552 * recover from these in time.
554 static void balloon_process(struct work_struct
*work
)
556 enum bp_state state
= BP_DONE
;
561 mutex_lock(&balloon_mutex
);
563 credit
= current_credit();
566 if (balloon_is_inflated())
567 state
= increase_reservation(credit
);
569 state
= reserve_additional_memory();
573 state
= decrease_reservation(-credit
, GFP_BALLOON
);
575 state
= update_schedule(state
);
577 mutex_unlock(&balloon_mutex
);
581 } while (credit
&& state
== BP_DONE
);
583 /* Schedule more work if there is some still to be done. */
584 if (state
== BP_EAGAIN
)
585 schedule_delayed_work(&balloon_worker
, balloon_stats
.schedule_delay
* HZ
);
588 /* Resets the Xen limit, sets new target, and kicks off processing. */
589 void balloon_set_new_target(unsigned long target
)
591 /* No need for lock. Not read-modify-write updates. */
592 balloon_stats
.target_pages
= target
;
593 schedule_delayed_work(&balloon_worker
, 0);
595 EXPORT_SYMBOL_GPL(balloon_set_new_target
);
597 static int add_ballooned_pages(int nr_pages
)
601 if (xen_hotplug_unpopulated
) {
602 st
= reserve_additional_memory();
603 if (st
!= BP_ECANCELED
) {
604 mutex_unlock(&balloon_mutex
);
605 wait_event(balloon_wq
,
606 !list_empty(&ballooned_pages
));
607 mutex_lock(&balloon_mutex
);
612 st
= decrease_reservation(nr_pages
, GFP_USER
);
620 * alloc_xenballooned_pages - get pages that have been ballooned out
621 * @nr_pages: Number of pages to get
622 * @pages: pages returned
623 * @return 0 on success, error otherwise
625 int alloc_xenballooned_pages(int nr_pages
, struct page
**pages
)
631 mutex_lock(&balloon_mutex
);
633 balloon_stats
.target_unpopulated
+= nr_pages
;
635 while (pgno
< nr_pages
) {
636 page
= balloon_retrieve(true);
638 pages
[pgno
++] = page
;
639 #ifdef CONFIG_XEN_HAVE_PVMMU
641 * We don't support PV MMU when Linux and Xen is using
642 * different page granularity.
644 BUILD_BUG_ON(XEN_PAGE_SIZE
!= PAGE_SIZE
);
646 ret
= xen_alloc_p2m_entry(page_to_pfn(page
));
651 ret
= add_ballooned_pages(nr_pages
- pgno
);
656 mutex_unlock(&balloon_mutex
);
659 mutex_unlock(&balloon_mutex
);
660 free_xenballooned_pages(pgno
, pages
);
663 EXPORT_SYMBOL(alloc_xenballooned_pages
);
666 * free_xenballooned_pages - return pages retrieved with get_ballooned_pages
667 * @nr_pages: Number of pages
668 * @pages: pages to return
670 void free_xenballooned_pages(int nr_pages
, struct page
**pages
)
674 mutex_lock(&balloon_mutex
);
676 for (i
= 0; i
< nr_pages
; i
++) {
678 balloon_append(pages
[i
]);
681 balloon_stats
.target_unpopulated
-= nr_pages
;
683 /* The balloon may be too large now. Shrink it if needed. */
684 if (current_credit())
685 schedule_delayed_work(&balloon_worker
, 0);
687 mutex_unlock(&balloon_mutex
);
689 EXPORT_SYMBOL(free_xenballooned_pages
);
691 static void __init
balloon_add_region(unsigned long start_pfn
,
694 unsigned long pfn
, extra_pfn_end
;
698 * If the amount of usable memory has been limited (e.g., with
699 * the 'mem' command line parameter), don't add pages beyond
702 extra_pfn_end
= min(max_pfn
, start_pfn
+ pages
);
704 for (pfn
= start_pfn
; pfn
< extra_pfn_end
; pfn
++) {
705 page
= pfn_to_page(pfn
);
706 /* totalram_pages and totalhigh_pages do not
707 include the boot-time balloon extension, so
708 don't subtract from it. */
709 __balloon_append(page
);
712 balloon_stats
.total_pages
+= extra_pfn_end
- start_pfn
;
715 static int __init
balloon_init(void)
722 pr_info("Initialising balloon driver\n");
724 balloon_stats
.current_pages
= xen_pv_domain()
725 ? min(xen_start_info
->nr_pages
- xen_released_pages
, max_pfn
)
726 : get_num_physpages();
727 balloon_stats
.target_pages
= balloon_stats
.current_pages
;
728 balloon_stats
.balloon_low
= 0;
729 balloon_stats
.balloon_high
= 0;
730 balloon_stats
.total_pages
= balloon_stats
.current_pages
;
732 balloon_stats
.schedule_delay
= 1;
733 balloon_stats
.max_schedule_delay
= 32;
734 balloon_stats
.retry_count
= 1;
735 balloon_stats
.max_retry_count
= RETRY_UNLIMITED
;
737 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
738 set_online_page_callback(&xen_online_page
);
739 register_memory_notifier(&xen_memory_nb
);
740 register_sysctl_table(xen_root
);
744 * Initialize the balloon with pages from the extra memory
745 * regions (see arch/x86/xen/setup.c).
747 for (i
= 0; i
< XEN_EXTRA_MEM_MAX_REGIONS
; i
++)
748 if (xen_extra_mem
[i
].n_pfns
)
749 balloon_add_region(xen_extra_mem
[i
].start_pfn
,
750 xen_extra_mem
[i
].n_pfns
);
755 subsys_initcall(balloon_init
);
757 MODULE_LICENSE("GPL");