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>
59 #include <asm/pgalloc.h>
60 #include <asm/pgtable.h>
63 #include <asm/xen/hypervisor.h>
64 #include <asm/xen/hypercall.h>
67 #include <xen/interface/xen.h>
68 #include <xen/interface/memory.h>
69 #include <xen/balloon.h>
70 #include <xen/features.h>
74 * balloon_process() state:
76 * BP_DONE: done or nothing to do,
77 * BP_EAGAIN: error, go to sleep,
78 * BP_ECANCELED: error, balloon operation canceled.
88 static DEFINE_MUTEX(balloon_mutex
);
90 struct balloon_stats balloon_stats
;
91 EXPORT_SYMBOL_GPL(balloon_stats
);
93 /* We increase/decrease in batches which fit in a page */
94 static xen_pfn_t frame_list
[PAGE_SIZE
/ sizeof(unsigned long)];
95 static DEFINE_PER_CPU(struct page
*, balloon_scratch_page
);
98 /* List of ballooned pages, threaded through the mem_map array. */
99 static LIST_HEAD(ballooned_pages
);
101 /* Main work function, always executed in process context. */
102 static void balloon_process(struct work_struct
*work
);
103 static DECLARE_DELAYED_WORK(balloon_worker
, balloon_process
);
105 /* When ballooning out (allocating memory to return to Xen) we don't really
106 want the kernel to try too hard since that can trigger the oom killer. */
107 #define GFP_BALLOON \
108 (GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC)
110 static void scrub_page(struct page
*page
)
112 #ifdef CONFIG_XEN_SCRUB_PAGES
113 clear_highpage(page
);
117 /* balloon_append: add the given page to the balloon. */
118 static void __balloon_append(struct page
*page
)
120 /* Lowmem is re-populated first, so highmem pages go at list tail. */
121 if (PageHighMem(page
)) {
122 list_add_tail(&page
->lru
, &ballooned_pages
);
123 balloon_stats
.balloon_high
++;
125 list_add(&page
->lru
, &ballooned_pages
);
126 balloon_stats
.balloon_low
++;
130 static void balloon_append(struct page
*page
)
132 __balloon_append(page
);
133 adjust_managed_page_count(page
, -1);
136 /* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
137 static struct page
*balloon_retrieve(bool prefer_highmem
)
141 if (list_empty(&ballooned_pages
))
145 page
= list_entry(ballooned_pages
.prev
, struct page
, lru
);
147 page
= list_entry(ballooned_pages
.next
, struct page
, lru
);
148 list_del(&page
->lru
);
150 if (PageHighMem(page
))
151 balloon_stats
.balloon_high
--;
153 balloon_stats
.balloon_low
--;
155 adjust_managed_page_count(page
, 1);
160 static struct page
*balloon_next_page(struct page
*page
)
162 struct list_head
*next
= page
->lru
.next
;
163 if (next
== &ballooned_pages
)
165 return list_entry(next
, struct page
, lru
);
168 static enum bp_state
update_schedule(enum bp_state state
)
170 if (state
== BP_DONE
) {
171 balloon_stats
.schedule_delay
= 1;
172 balloon_stats
.retry_count
= 1;
176 ++balloon_stats
.retry_count
;
178 if (balloon_stats
.max_retry_count
!= RETRY_UNLIMITED
&&
179 balloon_stats
.retry_count
> balloon_stats
.max_retry_count
) {
180 balloon_stats
.schedule_delay
= 1;
181 balloon_stats
.retry_count
= 1;
185 balloon_stats
.schedule_delay
<<= 1;
187 if (balloon_stats
.schedule_delay
> balloon_stats
.max_schedule_delay
)
188 balloon_stats
.schedule_delay
= balloon_stats
.max_schedule_delay
;
193 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
194 static long current_credit(void)
196 return balloon_stats
.target_pages
- balloon_stats
.current_pages
-
197 balloon_stats
.hotplug_pages
;
200 static bool balloon_is_inflated(void)
202 if (balloon_stats
.balloon_low
|| balloon_stats
.balloon_high
||
203 balloon_stats
.balloon_hotplug
)
210 * reserve_additional_memory() adds memory region of size >= credit above
211 * max_pfn. New region is section aligned and size is modified to be multiple
212 * of section size. Those features allow optimal use of address space and
213 * establish proper alignment when this function is called first time after
214 * boot (last section not fully populated at boot time contains unused memory
215 * pages with PG_reserved bit not set; online_pages_range() does not allow page
216 * onlining in whole range if first onlined page does not have PG_reserved
217 * bit set). Real size of added memory is established at page onlining stage.
220 static enum bp_state
reserve_additional_memory(long credit
)
223 u64 hotplug_start_paddr
;
224 unsigned long balloon_hotplug
= credit
;
226 hotplug_start_paddr
= PFN_PHYS(SECTION_ALIGN_UP(max_pfn
));
227 balloon_hotplug
= round_up(balloon_hotplug
, PAGES_PER_SECTION
);
228 nid
= memory_add_physaddr_to_nid(hotplug_start_paddr
);
230 rc
= add_memory(nid
, hotplug_start_paddr
, balloon_hotplug
<< PAGE_SHIFT
);
233 pr_info("%s: add_memory() failed: %i\n", __func__
, rc
);
237 balloon_hotplug
-= credit
;
239 balloon_stats
.hotplug_pages
+= credit
;
240 balloon_stats
.balloon_hotplug
= balloon_hotplug
;
245 static void xen_online_page(struct page
*page
)
247 __online_page_set_limits(page
);
249 mutex_lock(&balloon_mutex
);
251 __balloon_append(page
);
253 if (balloon_stats
.hotplug_pages
)
254 --balloon_stats
.hotplug_pages
;
256 --balloon_stats
.balloon_hotplug
;
258 mutex_unlock(&balloon_mutex
);
261 static int xen_memory_notifier(struct notifier_block
*nb
, unsigned long val
, void *v
)
263 if (val
== MEM_ONLINE
)
264 schedule_delayed_work(&balloon_worker
, 0);
269 static struct notifier_block xen_memory_nb
= {
270 .notifier_call
= xen_memory_notifier
,
274 static long current_credit(void)
276 unsigned long target
= balloon_stats
.target_pages
;
279 balloon_stats
.current_pages
+
280 balloon_stats
.balloon_low
+
281 balloon_stats
.balloon_high
);
283 return target
- balloon_stats
.current_pages
;
286 static bool balloon_is_inflated(void)
288 if (balloon_stats
.balloon_low
|| balloon_stats
.balloon_high
)
294 static enum bp_state
reserve_additional_memory(long credit
)
296 balloon_stats
.target_pages
= balloon_stats
.current_pages
;
299 #endif /* CONFIG_XEN_BALLOON_MEMORY_HOTPLUG */
301 static enum bp_state
increase_reservation(unsigned long nr_pages
)
304 unsigned long pfn
, i
;
306 struct xen_memory_reservation reservation
= {
312 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
313 if (!balloon_stats
.balloon_low
&& !balloon_stats
.balloon_high
) {
314 nr_pages
= min(nr_pages
, balloon_stats
.balloon_hotplug
);
315 balloon_stats
.hotplug_pages
+= nr_pages
;
316 balloon_stats
.balloon_hotplug
-= nr_pages
;
321 if (nr_pages
> ARRAY_SIZE(frame_list
))
322 nr_pages
= ARRAY_SIZE(frame_list
);
324 page
= list_first_entry_or_null(&ballooned_pages
, struct page
, lru
);
325 for (i
= 0; i
< nr_pages
; i
++) {
330 frame_list
[i
] = page_to_pfn(page
);
331 page
= balloon_next_page(page
);
334 set_xen_guest_handle(reservation
.extent_start
, frame_list
);
335 reservation
.nr_extents
= nr_pages
;
336 rc
= HYPERVISOR_memory_op(XENMEM_populate_physmap
, &reservation
);
340 for (i
= 0; i
< rc
; i
++) {
341 page
= balloon_retrieve(false);
342 BUG_ON(page
== NULL
);
344 pfn
= page_to_pfn(page
);
346 #ifdef CONFIG_XEN_HAVE_PVMMU
347 if (!xen_feature(XENFEAT_auto_translated_physmap
)) {
348 set_phys_to_machine(pfn
, frame_list
[i
]);
350 /* Link back into the page tables if not highmem. */
351 if (!PageHighMem(page
)) {
353 ret
= HYPERVISOR_update_va_mapping(
354 (unsigned long)__va(pfn
<< PAGE_SHIFT
),
355 mfn_pte(frame_list
[i
], PAGE_KERNEL
),
362 /* Relinquish the page back to the allocator. */
363 __free_reserved_page(page
);
366 balloon_stats
.current_pages
+= rc
;
371 static enum bp_state
decrease_reservation(unsigned long nr_pages
, gfp_t gfp
)
373 enum bp_state state
= BP_DONE
;
374 unsigned long pfn
, i
;
377 struct xen_memory_reservation reservation
= {
383 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
384 if (balloon_stats
.hotplug_pages
) {
385 nr_pages
= min(nr_pages
, balloon_stats
.hotplug_pages
);
386 balloon_stats
.hotplug_pages
-= nr_pages
;
387 balloon_stats
.balloon_hotplug
+= nr_pages
;
392 if (nr_pages
> ARRAY_SIZE(frame_list
))
393 nr_pages
= ARRAY_SIZE(frame_list
);
395 for (i
= 0; i
< nr_pages
; i
++) {
396 page
= alloc_page(gfp
);
404 frame_list
[i
] = page_to_pfn(page
);
408 * Ensure that ballooned highmem pages don't have kmaps.
410 * Do this before changing the p2m as kmap_flush_unused()
411 * reads PTEs to obtain pages (and hence needs the original
416 /* Update direct mapping, invalidate P2M, and add to balloon. */
417 for (i
= 0; i
< nr_pages
; i
++) {
419 frame_list
[i
] = pfn_to_mfn(pfn
);
420 page
= pfn_to_page(pfn
);
422 #ifdef CONFIG_XEN_HAVE_PVMMU
424 * Ballooned out frames are effectively replaced with
425 * a scratch frame. Ensure direct mappings and the
426 * p2m are consistent.
428 if (!xen_feature(XENFEAT_auto_translated_physmap
)) {
430 struct page
*scratch_page
= get_balloon_scratch_page();
432 if (!PageHighMem(page
)) {
433 ret
= HYPERVISOR_update_va_mapping(
434 (unsigned long)__va(pfn
<< PAGE_SHIFT
),
435 pfn_pte(page_to_pfn(scratch_page
),
439 p
= page_to_pfn(scratch_page
);
440 __set_phys_to_machine(pfn
, pfn_to_mfn(p
));
442 put_balloon_scratch_page();
446 balloon_append(page
);
451 set_xen_guest_handle(reservation
.extent_start
, frame_list
);
452 reservation
.nr_extents
= nr_pages
;
453 ret
= HYPERVISOR_memory_op(XENMEM_decrease_reservation
, &reservation
);
454 BUG_ON(ret
!= nr_pages
);
456 balloon_stats
.current_pages
-= nr_pages
;
462 * We avoid multiple worker processes conflicting via the balloon mutex.
463 * We may of course race updates of the target counts (which are protected
464 * by the balloon lock), or with changes to the Xen hard limit, but we will
465 * recover from these in time.
467 static void balloon_process(struct work_struct
*work
)
469 enum bp_state state
= BP_DONE
;
472 mutex_lock(&balloon_mutex
);
475 credit
= current_credit();
478 if (balloon_is_inflated())
479 state
= increase_reservation(credit
);
481 state
= reserve_additional_memory(credit
);
485 state
= decrease_reservation(-credit
, GFP_BALLOON
);
487 state
= update_schedule(state
);
489 #ifndef CONFIG_PREEMPT
493 } while (credit
&& state
== BP_DONE
);
495 /* Schedule more work if there is some still to be done. */
496 if (state
== BP_EAGAIN
)
497 schedule_delayed_work(&balloon_worker
, balloon_stats
.schedule_delay
* HZ
);
499 mutex_unlock(&balloon_mutex
);
502 struct page
*get_balloon_scratch_page(void)
504 struct page
*ret
= get_cpu_var(balloon_scratch_page
);
509 void put_balloon_scratch_page(void)
511 put_cpu_var(balloon_scratch_page
);
514 /* Resets the Xen limit, sets new target, and kicks off processing. */
515 void balloon_set_new_target(unsigned long target
)
517 /* No need for lock. Not read-modify-write updates. */
518 balloon_stats
.target_pages
= target
;
519 schedule_delayed_work(&balloon_worker
, 0);
521 EXPORT_SYMBOL_GPL(balloon_set_new_target
);
524 * alloc_xenballooned_pages - get pages that have been ballooned out
525 * @nr_pages: Number of pages to get
526 * @pages: pages returned
527 * @highmem: allow highmem pages
528 * @return 0 on success, error otherwise
530 int alloc_xenballooned_pages(int nr_pages
, struct page
**pages
, bool highmem
)
534 mutex_lock(&balloon_mutex
);
535 while (pgno
< nr_pages
) {
536 page
= balloon_retrieve(highmem
);
537 if (page
&& (highmem
|| !PageHighMem(page
))) {
538 pages
[pgno
++] = page
;
542 balloon_append(page
);
543 st
= decrease_reservation(nr_pages
- pgno
,
544 highmem
? GFP_HIGHUSER
: GFP_USER
);
549 mutex_unlock(&balloon_mutex
);
553 balloon_append(pages
[--pgno
]);
554 /* Free the memory back to the kernel soon */
555 schedule_delayed_work(&balloon_worker
, 0);
556 mutex_unlock(&balloon_mutex
);
559 EXPORT_SYMBOL(alloc_xenballooned_pages
);
562 * free_xenballooned_pages - return pages retrieved with get_ballooned_pages
563 * @nr_pages: Number of pages
564 * @pages: pages to return
566 void free_xenballooned_pages(int nr_pages
, struct page
**pages
)
570 mutex_lock(&balloon_mutex
);
572 for (i
= 0; i
< nr_pages
; i
++) {
574 balloon_append(pages
[i
]);
577 /* The balloon may be too large now. Shrink it if needed. */
578 if (current_credit())
579 schedule_delayed_work(&balloon_worker
, 0);
581 mutex_unlock(&balloon_mutex
);
583 EXPORT_SYMBOL(free_xenballooned_pages
);
585 static void __init
balloon_add_region(unsigned long start_pfn
,
588 unsigned long pfn
, extra_pfn_end
;
592 * If the amount of usable memory has been limited (e.g., with
593 * the 'mem' command line parameter), don't add pages beyond
596 extra_pfn_end
= min(max_pfn
, start_pfn
+ pages
);
598 for (pfn
= start_pfn
; pfn
< extra_pfn_end
; pfn
++) {
599 page
= pfn_to_page(pfn
);
600 /* totalram_pages and totalhigh_pages do not
601 include the boot-time balloon extension, so
602 don't subtract from it. */
603 __balloon_append(page
);
607 static int alloc_balloon_scratch_page(int cpu
)
609 if (per_cpu(balloon_scratch_page
, cpu
) != NULL
)
612 per_cpu(balloon_scratch_page
, cpu
) = alloc_page(GFP_KERNEL
);
613 if (per_cpu(balloon_scratch_page
, cpu
) == NULL
) {
614 pr_warn("Failed to allocate balloon_scratch_page for cpu %d\n", cpu
);
622 static int balloon_cpu_notify(struct notifier_block
*self
,
623 unsigned long action
, void *hcpu
)
625 int cpu
= (long)hcpu
;
628 if (alloc_balloon_scratch_page(cpu
))
637 static struct notifier_block balloon_cpu_notifier
= {
638 .notifier_call
= balloon_cpu_notify
,
641 static int __init
balloon_init(void)
648 if (!xen_feature(XENFEAT_auto_translated_physmap
)) {
649 register_cpu_notifier(&balloon_cpu_notifier
);
652 for_each_online_cpu(cpu
) {
653 if (alloc_balloon_scratch_page(cpu
)) {
655 unregister_cpu_notifier(&balloon_cpu_notifier
);
662 pr_info("Initialising balloon driver\n");
664 balloon_stats
.current_pages
= xen_pv_domain()
665 ? min(xen_start_info
->nr_pages
- xen_released_pages
, max_pfn
)
666 : get_num_physpages();
667 balloon_stats
.target_pages
= balloon_stats
.current_pages
;
668 balloon_stats
.balloon_low
= 0;
669 balloon_stats
.balloon_high
= 0;
671 balloon_stats
.schedule_delay
= 1;
672 balloon_stats
.max_schedule_delay
= 32;
673 balloon_stats
.retry_count
= 1;
674 balloon_stats
.max_retry_count
= RETRY_UNLIMITED
;
676 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
677 balloon_stats
.hotplug_pages
= 0;
678 balloon_stats
.balloon_hotplug
= 0;
680 set_online_page_callback(&xen_online_page
);
681 register_memory_notifier(&xen_memory_nb
);
685 * Initialize the balloon with pages from the extra memory
686 * regions (see arch/x86/xen/setup.c).
688 for (i
= 0; i
< XEN_EXTRA_MEM_MAX_REGIONS
; i
++)
689 if (xen_extra_mem
[i
].size
)
690 balloon_add_region(PFN_UP(xen_extra_mem
[i
].start
),
691 PFN_DOWN(xen_extra_mem
[i
].size
));
696 subsys_initcall(balloon_init
);
698 static int __init
balloon_clear(void)
702 for_each_possible_cpu(cpu
)
703 per_cpu(balloon_scratch_page
, cpu
) = NULL
;
707 early_initcall(balloon_clear
);
709 MODULE_LICENSE("GPL");