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/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>
74 #include <xen/mem-reservation.h>
76 static int xen_hotplug_unpopulated
;
78 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
83 static struct ctl_table balloon_table
[] = {
85 .procname
= "hotplug_unpopulated",
86 .data
= &xen_hotplug_unpopulated
,
87 .maxlen
= sizeof(int),
89 .proc_handler
= proc_dointvec_minmax
,
96 static struct ctl_table balloon_root
[] = {
98 .procname
= "balloon",
100 .child
= balloon_table
,
105 static struct ctl_table xen_root
[] = {
109 .child
= balloon_root
,
117 * Use one extent per PAGE_SIZE to avoid to break down the page into
120 #define EXTENT_ORDER (fls(XEN_PFN_PER_PAGE) - 1)
123 * balloon_process() state:
125 * BP_DONE: done or nothing to do,
126 * BP_WAIT: wait to be rescheduled,
127 * BP_EAGAIN: error, go to sleep,
128 * BP_ECANCELED: error, balloon operation canceled.
139 static DEFINE_MUTEX(balloon_mutex
);
141 struct balloon_stats balloon_stats
;
142 EXPORT_SYMBOL_GPL(balloon_stats
);
144 /* We increase/decrease in batches which fit in a page */
145 static xen_pfn_t frame_list
[PAGE_SIZE
/ sizeof(xen_pfn_t
)];
148 /* List of ballooned pages, threaded through the mem_map array. */
149 static LIST_HEAD(ballooned_pages
);
150 static DECLARE_WAIT_QUEUE_HEAD(balloon_wq
);
152 /* Main work function, always executed in process context. */
153 static void balloon_process(struct work_struct
*work
);
154 static DECLARE_DELAYED_WORK(balloon_worker
, balloon_process
);
156 /* When ballooning out (allocating memory to return to Xen) we don't really
157 want the kernel to try too hard since that can trigger the oom killer. */
158 #define GFP_BALLOON \
159 (GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC)
161 /* balloon_append: add the given page to the balloon. */
162 static void __balloon_append(struct page
*page
)
164 /* Lowmem is re-populated first, so highmem pages go at list tail. */
165 if (PageHighMem(page
)) {
166 list_add_tail(&page
->lru
, &ballooned_pages
);
167 balloon_stats
.balloon_high
++;
169 list_add(&page
->lru
, &ballooned_pages
);
170 balloon_stats
.balloon_low
++;
172 wake_up(&balloon_wq
);
175 static void balloon_append(struct page
*page
)
177 __balloon_append(page
);
180 /* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
181 static struct page
*balloon_retrieve(bool require_lowmem
)
185 if (list_empty(&ballooned_pages
))
188 page
= list_entry(ballooned_pages
.next
, struct page
, lru
);
189 if (require_lowmem
&& PageHighMem(page
))
191 list_del(&page
->lru
);
193 if (PageHighMem(page
))
194 balloon_stats
.balloon_high
--;
196 balloon_stats
.balloon_low
--;
201 static struct page
*balloon_next_page(struct page
*page
)
203 struct list_head
*next
= page
->lru
.next
;
204 if (next
== &ballooned_pages
)
206 return list_entry(next
, struct page
, lru
);
209 static enum bp_state
update_schedule(enum bp_state state
)
211 if (state
== BP_WAIT
)
214 if (state
== BP_ECANCELED
)
217 if (state
== BP_DONE
) {
218 balloon_stats
.schedule_delay
= 1;
219 balloon_stats
.retry_count
= 1;
223 ++balloon_stats
.retry_count
;
225 if (balloon_stats
.max_retry_count
!= RETRY_UNLIMITED
&&
226 balloon_stats
.retry_count
> balloon_stats
.max_retry_count
) {
227 balloon_stats
.schedule_delay
= 1;
228 balloon_stats
.retry_count
= 1;
232 balloon_stats
.schedule_delay
<<= 1;
234 if (balloon_stats
.schedule_delay
> balloon_stats
.max_schedule_delay
)
235 balloon_stats
.schedule_delay
= balloon_stats
.max_schedule_delay
;
240 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
241 static void release_memory_resource(struct resource
*resource
)
247 * No need to reset region to identity mapped since we now
248 * know that no I/O can be in this region
250 release_resource(resource
);
254 static struct resource
*additional_memory_resource(phys_addr_t size
)
256 struct resource
*res
;
259 res
= kzalloc(sizeof(*res
), GFP_KERNEL
);
263 res
->name
= "System RAM";
264 res
->flags
= IORESOURCE_SYSTEM_RAM
| IORESOURCE_BUSY
;
266 ret
= allocate_resource(&iomem_resource
, res
,
268 PAGES_PER_SECTION
* PAGE_SIZE
, NULL
, NULL
);
270 pr_err("Cannot allocate new System RAM resource\n");
275 #ifdef CONFIG_SPARSEMEM
277 unsigned long limit
= 1UL << (MAX_PHYSMEM_BITS
- PAGE_SHIFT
);
278 unsigned long pfn
= res
->start
>> PAGE_SHIFT
;
281 pr_err("New System RAM resource outside addressable RAM (%lu > %lu)\n",
283 release_memory_resource(res
);
292 static enum bp_state
reserve_additional_memory(void)
295 struct resource
*resource
;
297 unsigned long balloon_hotplug
;
299 credit
= balloon_stats
.target_pages
+ balloon_stats
.target_unpopulated
300 - balloon_stats
.total_pages
;
303 * Already hotplugged enough pages? Wait for them to be
309 balloon_hotplug
= round_up(credit
, PAGES_PER_SECTION
);
311 resource
= additional_memory_resource(balloon_hotplug
* PAGE_SIZE
);
315 nid
= memory_add_physaddr_to_nid(resource
->start
);
317 #ifdef CONFIG_XEN_HAVE_PVMMU
319 * We don't support PV MMU when Linux and Xen is using
320 * different page granularity.
322 BUILD_BUG_ON(XEN_PAGE_SIZE
!= PAGE_SIZE
);
325 * add_memory() will build page tables for the new memory so
326 * the p2m must contain invalid entries so the correct
327 * non-present PTEs will be written.
329 * If a failure occurs, the original (identity) p2m entries
330 * are not restored since this region is now known not to
331 * conflict with any devices.
333 if (!xen_feature(XENFEAT_auto_translated_physmap
)) {
334 unsigned long pfn
, i
;
336 pfn
= PFN_DOWN(resource
->start
);
337 for (i
= 0; i
< balloon_hotplug
; i
++) {
338 if (!set_phys_to_machine(pfn
+ i
, INVALID_P2M_ENTRY
)) {
339 pr_warn("set_phys_to_machine() failed, no memory added\n");
347 * add_memory_resource() will call online_pages() which in its turn
348 * will call xen_online_page() callback causing deadlock if we don't
349 * release balloon_mutex here. Unlocking here is safe because the
350 * callers drop the mutex before trying again.
352 mutex_unlock(&balloon_mutex
);
353 rc
= add_memory_resource(nid
, resource
, memhp_auto_online
);
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
)
371 __online_page_set_limits(page
);
373 mutex_lock(&balloon_mutex
);
375 __balloon_append(page
);
377 mutex_unlock(&balloon_mutex
);
380 static int xen_memory_notifier(struct notifier_block
*nb
, unsigned long val
, void *v
)
382 if (val
== MEM_ONLINE
)
383 schedule_delayed_work(&balloon_worker
, 0);
388 static struct notifier_block xen_memory_nb
= {
389 .notifier_call
= xen_memory_notifier
,
393 static enum bp_state
reserve_additional_memory(void)
395 balloon_stats
.target_pages
= balloon_stats
.current_pages
;
398 #endif /* CONFIG_XEN_BALLOON_MEMORY_HOTPLUG */
400 static long current_credit(void)
402 return balloon_stats
.target_pages
- balloon_stats
.current_pages
;
405 static bool balloon_is_inflated(void)
407 return balloon_stats
.balloon_low
|| balloon_stats
.balloon_high
;
410 static enum bp_state
increase_reservation(unsigned long nr_pages
)
416 if (nr_pages
> ARRAY_SIZE(frame_list
))
417 nr_pages
= ARRAY_SIZE(frame_list
);
419 page
= list_first_entry_or_null(&ballooned_pages
, struct page
, lru
);
420 for (i
= 0; i
< nr_pages
; i
++) {
426 frame_list
[i
] = page_to_xen_pfn(page
);
427 page
= balloon_next_page(page
);
430 rc
= xenmem_reservation_increase(nr_pages
, frame_list
);
434 for (i
= 0; i
< rc
; i
++) {
435 page
= balloon_retrieve(false);
436 BUG_ON(page
== NULL
);
438 xenmem_reservation_va_mapping_update(1, &page
, &frame_list
[i
]);
440 /* Relinquish the page back to the allocator. */
441 free_reserved_page(page
);
444 balloon_stats
.current_pages
+= rc
;
449 static enum bp_state
decrease_reservation(unsigned long nr_pages
, gfp_t gfp
)
451 enum bp_state state
= BP_DONE
;
453 struct page
*page
, *tmp
;
457 if (nr_pages
> ARRAY_SIZE(frame_list
))
458 nr_pages
= ARRAY_SIZE(frame_list
);
460 for (i
= 0; i
< nr_pages
; i
++) {
461 page
= alloc_page(gfp
);
467 adjust_managed_page_count(page
, -1);
468 xenmem_reservation_scrub_page(page
);
469 list_add(&page
->lru
, &pages
);
473 * Ensure that ballooned highmem pages don't have kmaps.
475 * Do this before changing the p2m as kmap_flush_unused()
476 * reads PTEs to obtain pages (and hence needs the original
482 * Setup the frame, update direct mapping, invalidate P2M,
483 * and add to balloon.
486 list_for_each_entry_safe(page
, tmp
, &pages
, lru
) {
487 frame_list
[i
++] = xen_page_to_gfn(page
);
489 xenmem_reservation_va_mapping_reset(1, &page
);
491 list_del(&page
->lru
);
493 balloon_append(page
);
498 ret
= xenmem_reservation_decrease(nr_pages
, frame_list
);
499 BUG_ON(ret
!= nr_pages
);
501 balloon_stats
.current_pages
-= nr_pages
;
507 * As this is a work item it is guaranteed to run as a single instance only.
508 * We may of course race updates of the target counts (which are protected
509 * by the balloon lock), or with changes to the Xen hard limit, but we will
510 * recover from these in time.
512 static void balloon_process(struct work_struct
*work
)
514 enum bp_state state
= BP_DONE
;
519 mutex_lock(&balloon_mutex
);
521 credit
= current_credit();
524 if (balloon_is_inflated())
525 state
= increase_reservation(credit
);
527 state
= reserve_additional_memory();
531 state
= decrease_reservation(-credit
, GFP_BALLOON
);
533 state
= update_schedule(state
);
535 mutex_unlock(&balloon_mutex
);
539 } while (credit
&& state
== BP_DONE
);
541 /* Schedule more work if there is some still to be done. */
542 if (state
== BP_EAGAIN
)
543 schedule_delayed_work(&balloon_worker
, balloon_stats
.schedule_delay
* HZ
);
546 /* Resets the Xen limit, sets new target, and kicks off processing. */
547 void balloon_set_new_target(unsigned long target
)
549 /* No need for lock. Not read-modify-write updates. */
550 balloon_stats
.target_pages
= target
;
551 schedule_delayed_work(&balloon_worker
, 0);
553 EXPORT_SYMBOL_GPL(balloon_set_new_target
);
555 static int add_ballooned_pages(int nr_pages
)
559 if (xen_hotplug_unpopulated
) {
560 st
= reserve_additional_memory();
561 if (st
!= BP_ECANCELED
) {
562 mutex_unlock(&balloon_mutex
);
563 wait_event(balloon_wq
,
564 !list_empty(&ballooned_pages
));
565 mutex_lock(&balloon_mutex
);
570 st
= decrease_reservation(nr_pages
, GFP_USER
);
578 * alloc_xenballooned_pages - get pages that have been ballooned out
579 * @nr_pages: Number of pages to get
580 * @pages: pages returned
581 * @return 0 on success, error otherwise
583 int alloc_xenballooned_pages(int nr_pages
, struct page
**pages
)
589 mutex_lock(&balloon_mutex
);
591 balloon_stats
.target_unpopulated
+= nr_pages
;
593 while (pgno
< nr_pages
) {
594 page
= balloon_retrieve(true);
596 pages
[pgno
++] = page
;
597 #ifdef CONFIG_XEN_HAVE_PVMMU
599 * We don't support PV MMU when Linux and Xen is using
600 * different page granularity.
602 BUILD_BUG_ON(XEN_PAGE_SIZE
!= PAGE_SIZE
);
604 if (!xen_feature(XENFEAT_auto_translated_physmap
)) {
605 ret
= xen_alloc_p2m_entry(page_to_pfn(page
));
611 ret
= add_ballooned_pages(nr_pages
- pgno
);
616 mutex_unlock(&balloon_mutex
);
619 mutex_unlock(&balloon_mutex
);
620 free_xenballooned_pages(pgno
, pages
);
623 EXPORT_SYMBOL(alloc_xenballooned_pages
);
626 * free_xenballooned_pages - return pages retrieved with get_ballooned_pages
627 * @nr_pages: Number of pages
628 * @pages: pages to return
630 void free_xenballooned_pages(int nr_pages
, struct page
**pages
)
634 mutex_lock(&balloon_mutex
);
636 for (i
= 0; i
< nr_pages
; i
++) {
638 balloon_append(pages
[i
]);
641 balloon_stats
.target_unpopulated
-= nr_pages
;
643 /* The balloon may be too large now. Shrink it if needed. */
644 if (current_credit())
645 schedule_delayed_work(&balloon_worker
, 0);
647 mutex_unlock(&balloon_mutex
);
649 EXPORT_SYMBOL(free_xenballooned_pages
);
652 static void __init
balloon_add_region(unsigned long start_pfn
,
655 unsigned long pfn
, extra_pfn_end
;
659 * If the amount of usable memory has been limited (e.g., with
660 * the 'mem' command line parameter), don't add pages beyond
663 extra_pfn_end
= min(max_pfn
, start_pfn
+ pages
);
665 for (pfn
= start_pfn
; pfn
< extra_pfn_end
; pfn
++) {
666 page
= pfn_to_page(pfn
);
667 /* totalram_pages and totalhigh_pages do not
668 include the boot-time balloon extension, so
669 don't subtract from it. */
670 __balloon_append(page
);
673 balloon_stats
.total_pages
+= extra_pfn_end
- start_pfn
;
677 static int __init
balloon_init(void)
682 pr_info("Initialising balloon driver\n");
685 balloon_stats
.current_pages
= xen_pv_domain()
686 ? min(xen_start_info
->nr_pages
- xen_released_pages
, max_pfn
)
687 : get_num_physpages();
689 balloon_stats
.current_pages
= get_num_physpages();
691 balloon_stats
.target_pages
= balloon_stats
.current_pages
;
692 balloon_stats
.balloon_low
= 0;
693 balloon_stats
.balloon_high
= 0;
694 balloon_stats
.total_pages
= balloon_stats
.current_pages
;
696 balloon_stats
.schedule_delay
= 1;
697 balloon_stats
.max_schedule_delay
= 32;
698 balloon_stats
.retry_count
= 1;
699 balloon_stats
.max_retry_count
= RETRY_UNLIMITED
;
701 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
702 set_online_page_callback(&xen_online_page
);
703 register_memory_notifier(&xen_memory_nb
);
704 register_sysctl_table(xen_root
);
712 * Initialize the balloon with pages from the extra memory
713 * regions (see arch/x86/xen/setup.c).
715 for (i
= 0; i
< XEN_EXTRA_MEM_MAX_REGIONS
; i
++)
716 if (xen_extra_mem
[i
].n_pfns
)
717 balloon_add_region(xen_extra_mem
[i
].start_pfn
,
718 xen_extra_mem
[i
].n_pfns
);
722 /* Init the xen-balloon driver. */
727 subsys_initcall(balloon_init
);