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>
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
);
186 /* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
187 static struct page
*balloon_retrieve(bool require_lowmem
)
191 if (list_empty(&ballooned_pages
))
194 page
= list_entry(ballooned_pages
.next
, struct page
, lru
);
195 if (require_lowmem
&& PageHighMem(page
))
197 list_del(&page
->lru
);
199 if (PageHighMem(page
))
200 balloon_stats
.balloon_high
--;
202 balloon_stats
.balloon_low
--;
207 static struct page
*balloon_next_page(struct page
*page
)
209 struct list_head
*next
= page
->lru
.next
;
210 if (next
== &ballooned_pages
)
212 return list_entry(next
, struct page
, lru
);
215 static enum bp_state
update_schedule(enum bp_state state
)
217 if (state
== BP_WAIT
)
220 if (state
== BP_ECANCELED
)
223 if (state
== BP_DONE
) {
224 balloon_stats
.schedule_delay
= 1;
225 balloon_stats
.retry_count
= 1;
229 ++balloon_stats
.retry_count
;
231 if (balloon_stats
.max_retry_count
!= RETRY_UNLIMITED
&&
232 balloon_stats
.retry_count
> balloon_stats
.max_retry_count
) {
233 balloon_stats
.schedule_delay
= 1;
234 balloon_stats
.retry_count
= 1;
238 balloon_stats
.schedule_delay
<<= 1;
240 if (balloon_stats
.schedule_delay
> balloon_stats
.max_schedule_delay
)
241 balloon_stats
.schedule_delay
= balloon_stats
.max_schedule_delay
;
246 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
247 static void release_memory_resource(struct resource
*resource
)
253 * No need to reset region to identity mapped since we now
254 * know that no I/O can be in this region
256 release_resource(resource
);
261 * Host memory not allocated to dom0. We can use this range for hotplug-based
264 * It's a type-less resource. Setting IORESOURCE_MEM will make resource
265 * management algorithms (arch_remove_reservations()) look into guest e820,
266 * which we don't want.
268 static struct resource hostmem_resource
= {
272 void __attribute__((weak
)) __init
arch_xen_balloon_init(struct resource
*res
)
275 static struct resource
*additional_memory_resource(phys_addr_t size
)
277 struct resource
*res
, *res_hostmem
;
280 res
= kzalloc(sizeof(*res
), GFP_KERNEL
);
284 res
->name
= "System RAM";
285 res
->flags
= IORESOURCE_SYSTEM_RAM
| IORESOURCE_BUSY
;
287 res_hostmem
= kzalloc(sizeof(*res
), GFP_KERNEL
);
289 /* Try to grab a range from hostmem */
290 res_hostmem
->name
= "Host memory";
291 ret
= allocate_resource(&hostmem_resource
, res_hostmem
,
293 PAGES_PER_SECTION
* PAGE_SIZE
, NULL
, NULL
);
298 * Insert this resource into iomem. Because hostmem_resource
299 * tracks portion of guest e820 marked as UNUSABLE noone else
300 * should try to use it.
302 res
->start
= res_hostmem
->start
;
303 res
->end
= res_hostmem
->end
;
304 ret
= insert_resource(&iomem_resource
, res
);
306 pr_err("Can't insert iomem_resource [%llx - %llx]\n",
307 res
->start
, res
->end
);
308 release_memory_resource(res_hostmem
);
310 res
->start
= res
->end
= 0;
315 ret
= allocate_resource(&iomem_resource
, res
,
317 PAGES_PER_SECTION
* PAGE_SIZE
, NULL
, NULL
);
319 pr_err("Cannot allocate new System RAM resource\n");
325 #ifdef CONFIG_SPARSEMEM
327 unsigned long limit
= 1UL << (MAX_PHYSMEM_BITS
- PAGE_SHIFT
);
328 unsigned long pfn
= res
->start
>> PAGE_SHIFT
;
331 pr_err("New System RAM resource outside addressable RAM (%lu > %lu)\n",
333 release_memory_resource(res
);
334 release_memory_resource(res_hostmem
);
343 static enum bp_state
reserve_additional_memory(void)
346 struct resource
*resource
;
348 unsigned long balloon_hotplug
;
350 credit
= balloon_stats
.target_pages
+ balloon_stats
.target_unpopulated
351 - balloon_stats
.total_pages
;
354 * Already hotplugged enough pages? Wait for them to be
360 balloon_hotplug
= round_up(credit
, PAGES_PER_SECTION
);
362 resource
= additional_memory_resource(balloon_hotplug
* PAGE_SIZE
);
366 nid
= memory_add_physaddr_to_nid(resource
->start
);
368 #ifdef CONFIG_XEN_HAVE_PVMMU
370 * We don't support PV MMU when Linux and Xen is using
371 * different page granularity.
373 BUILD_BUG_ON(XEN_PAGE_SIZE
!= PAGE_SIZE
);
376 * add_memory() will build page tables for the new memory so
377 * the p2m must contain invalid entries so the correct
378 * non-present PTEs will be written.
380 * If a failure occurs, the original (identity) p2m entries
381 * are not restored since this region is now known not to
382 * conflict with any devices.
384 if (!xen_feature(XENFEAT_auto_translated_physmap
)) {
385 unsigned long pfn
, i
;
387 pfn
= PFN_DOWN(resource
->start
);
388 for (i
= 0; i
< balloon_hotplug
; i
++) {
389 if (!set_phys_to_machine(pfn
+ i
, INVALID_P2M_ENTRY
)) {
390 pr_warn("set_phys_to_machine() failed, no memory added\n");
398 * add_memory_resource() will call online_pages() which in its turn
399 * will call xen_online_page() callback causing deadlock if we don't
400 * release balloon_mutex here. Unlocking here is safe because the
401 * callers drop the mutex before trying again.
403 mutex_unlock(&balloon_mutex
);
404 rc
= add_memory_resource(nid
, resource
, memhp_auto_online
);
405 mutex_lock(&balloon_mutex
);
408 pr_warn("Cannot add additional memory (%i)\n", rc
);
412 balloon_stats
.total_pages
+= balloon_hotplug
;
416 release_memory_resource(resource
);
420 static void xen_online_page(struct page
*page
)
422 __online_page_set_limits(page
);
424 mutex_lock(&balloon_mutex
);
426 __balloon_append(page
);
428 mutex_unlock(&balloon_mutex
);
431 static int xen_memory_notifier(struct notifier_block
*nb
, unsigned long val
, void *v
)
433 if (val
== MEM_ONLINE
)
434 schedule_delayed_work(&balloon_worker
, 0);
439 static struct notifier_block xen_memory_nb
= {
440 .notifier_call
= xen_memory_notifier
,
444 static enum bp_state
reserve_additional_memory(void)
446 balloon_stats
.target_pages
= balloon_stats
.current_pages
;
449 #endif /* CONFIG_XEN_BALLOON_MEMORY_HOTPLUG */
451 static long current_credit(void)
453 return balloon_stats
.target_pages
- balloon_stats
.current_pages
;
456 static bool balloon_is_inflated(void)
458 return balloon_stats
.balloon_low
|| balloon_stats
.balloon_high
;
461 static enum bp_state
increase_reservation(unsigned long nr_pages
)
466 struct xen_memory_reservation reservation
= {
468 .extent_order
= EXTENT_ORDER
,
472 if (nr_pages
> ARRAY_SIZE(frame_list
))
473 nr_pages
= ARRAY_SIZE(frame_list
);
475 page
= list_first_entry_or_null(&ballooned_pages
, struct page
, lru
);
476 for (i
= 0; i
< nr_pages
; i
++) {
482 /* XENMEM_populate_physmap requires a PFN based on Xen
485 frame_list
[i
] = page_to_xen_pfn(page
);
486 page
= balloon_next_page(page
);
489 set_xen_guest_handle(reservation
.extent_start
, frame_list
);
490 reservation
.nr_extents
= nr_pages
;
491 rc
= HYPERVISOR_memory_op(XENMEM_populate_physmap
, &reservation
);
495 for (i
= 0; i
< rc
; i
++) {
496 page
= balloon_retrieve(false);
497 BUG_ON(page
== NULL
);
499 #ifdef CONFIG_XEN_HAVE_PVMMU
501 * We don't support PV MMU when Linux and Xen is using
502 * different page granularity.
504 BUILD_BUG_ON(XEN_PAGE_SIZE
!= PAGE_SIZE
);
506 if (!xen_feature(XENFEAT_auto_translated_physmap
)) {
507 unsigned long pfn
= page_to_pfn(page
);
509 set_phys_to_machine(pfn
, frame_list
[i
]);
511 /* Link back into the page tables if not highmem. */
512 if (!PageHighMem(page
)) {
514 ret
= HYPERVISOR_update_va_mapping(
515 (unsigned long)__va(pfn
<< PAGE_SHIFT
),
516 mfn_pte(frame_list
[i
], PAGE_KERNEL
),
523 /* Relinquish the page back to the allocator. */
524 free_reserved_page(page
);
527 balloon_stats
.current_pages
+= rc
;
532 static enum bp_state
decrease_reservation(unsigned long nr_pages
, gfp_t gfp
)
534 enum bp_state state
= BP_DONE
;
536 struct page
*page
, *tmp
;
538 struct xen_memory_reservation reservation
= {
540 .extent_order
= EXTENT_ORDER
,
545 if (nr_pages
> ARRAY_SIZE(frame_list
))
546 nr_pages
= ARRAY_SIZE(frame_list
);
548 for (i
= 0; i
< nr_pages
; i
++) {
549 page
= alloc_page(gfp
);
555 adjust_managed_page_count(page
, -1);
557 list_add(&page
->lru
, &pages
);
561 * Ensure that ballooned highmem pages don't have kmaps.
563 * Do this before changing the p2m as kmap_flush_unused()
564 * reads PTEs to obtain pages (and hence needs the original
570 * Setup the frame, update direct mapping, invalidate P2M,
571 * and add to balloon.
574 list_for_each_entry_safe(page
, tmp
, &pages
, lru
) {
575 /* XENMEM_decrease_reservation requires a GFN */
576 frame_list
[i
++] = xen_page_to_gfn(page
);
578 #ifdef CONFIG_XEN_HAVE_PVMMU
580 * We don't support PV MMU when Linux and Xen is using
581 * different page granularity.
583 BUILD_BUG_ON(XEN_PAGE_SIZE
!= PAGE_SIZE
);
585 if (!xen_feature(XENFEAT_auto_translated_physmap
)) {
586 unsigned long pfn
= page_to_pfn(page
);
588 if (!PageHighMem(page
)) {
589 ret
= HYPERVISOR_update_va_mapping(
590 (unsigned long)__va(pfn
<< PAGE_SHIFT
),
594 __set_phys_to_machine(pfn
, INVALID_P2M_ENTRY
);
597 list_del(&page
->lru
);
599 balloon_append(page
);
604 set_xen_guest_handle(reservation
.extent_start
, frame_list
);
605 reservation
.nr_extents
= nr_pages
;
606 ret
= HYPERVISOR_memory_op(XENMEM_decrease_reservation
, &reservation
);
607 BUG_ON(ret
!= nr_pages
);
609 balloon_stats
.current_pages
-= nr_pages
;
615 * As this is a work item it is guaranteed to run as a single instance only.
616 * We may of course race updates of the target counts (which are protected
617 * by the balloon lock), or with changes to the Xen hard limit, but we will
618 * recover from these in time.
620 static void balloon_process(struct work_struct
*work
)
622 enum bp_state state
= BP_DONE
;
627 mutex_lock(&balloon_mutex
);
629 credit
= current_credit();
632 if (balloon_is_inflated())
633 state
= increase_reservation(credit
);
635 state
= reserve_additional_memory();
639 state
= decrease_reservation(-credit
, GFP_BALLOON
);
641 state
= update_schedule(state
);
643 mutex_unlock(&balloon_mutex
);
647 } while (credit
&& state
== BP_DONE
);
649 /* Schedule more work if there is some still to be done. */
650 if (state
== BP_EAGAIN
)
651 schedule_delayed_work(&balloon_worker
, balloon_stats
.schedule_delay
* HZ
);
654 /* Resets the Xen limit, sets new target, and kicks off processing. */
655 void balloon_set_new_target(unsigned long target
)
657 /* No need for lock. Not read-modify-write updates. */
658 balloon_stats
.target_pages
= target
;
659 schedule_delayed_work(&balloon_worker
, 0);
661 EXPORT_SYMBOL_GPL(balloon_set_new_target
);
663 static int add_ballooned_pages(int nr_pages
)
667 if (xen_hotplug_unpopulated
) {
668 st
= reserve_additional_memory();
669 if (st
!= BP_ECANCELED
) {
670 mutex_unlock(&balloon_mutex
);
671 wait_event(balloon_wq
,
672 !list_empty(&ballooned_pages
));
673 mutex_lock(&balloon_mutex
);
678 st
= decrease_reservation(nr_pages
, GFP_USER
);
686 * alloc_xenballooned_pages - get pages that have been ballooned out
687 * @nr_pages: Number of pages to get
688 * @pages: pages returned
689 * @return 0 on success, error otherwise
691 int alloc_xenballooned_pages(int nr_pages
, struct page
**pages
)
697 mutex_lock(&balloon_mutex
);
699 balloon_stats
.target_unpopulated
+= nr_pages
;
701 while (pgno
< nr_pages
) {
702 page
= balloon_retrieve(true);
704 pages
[pgno
++] = page
;
705 #ifdef CONFIG_XEN_HAVE_PVMMU
707 * We don't support PV MMU when Linux and Xen is using
708 * different page granularity.
710 BUILD_BUG_ON(XEN_PAGE_SIZE
!= PAGE_SIZE
);
712 if (!xen_feature(XENFEAT_auto_translated_physmap
)) {
713 ret
= xen_alloc_p2m_entry(page_to_pfn(page
));
719 ret
= add_ballooned_pages(nr_pages
- pgno
);
724 mutex_unlock(&balloon_mutex
);
727 mutex_unlock(&balloon_mutex
);
728 free_xenballooned_pages(pgno
, pages
);
731 EXPORT_SYMBOL(alloc_xenballooned_pages
);
734 * free_xenballooned_pages - return pages retrieved with get_ballooned_pages
735 * @nr_pages: Number of pages
736 * @pages: pages to return
738 void free_xenballooned_pages(int nr_pages
, struct page
**pages
)
742 mutex_lock(&balloon_mutex
);
744 for (i
= 0; i
< nr_pages
; i
++) {
746 balloon_append(pages
[i
]);
749 balloon_stats
.target_unpopulated
-= nr_pages
;
751 /* The balloon may be too large now. Shrink it if needed. */
752 if (current_credit())
753 schedule_delayed_work(&balloon_worker
, 0);
755 mutex_unlock(&balloon_mutex
);
757 EXPORT_SYMBOL(free_xenballooned_pages
);
760 static void __init
balloon_add_region(unsigned long start_pfn
,
763 unsigned long pfn
, extra_pfn_end
;
767 * If the amount of usable memory has been limited (e.g., with
768 * the 'mem' command line parameter), don't add pages beyond
771 extra_pfn_end
= min(max_pfn
, start_pfn
+ pages
);
773 for (pfn
= start_pfn
; pfn
< extra_pfn_end
; pfn
++) {
774 page
= pfn_to_page(pfn
);
775 /* totalram_pages and totalhigh_pages do not
776 include the boot-time balloon extension, so
777 don't subtract from it. */
778 __balloon_append(page
);
781 balloon_stats
.total_pages
+= extra_pfn_end
- start_pfn
;
785 static int __init
balloon_init(void)
790 pr_info("Initialising balloon driver\n");
793 balloon_stats
.current_pages
= xen_pv_domain()
794 ? min(xen_start_info
->nr_pages
- xen_released_pages
, max_pfn
)
795 : get_num_physpages();
797 balloon_stats
.current_pages
= get_num_physpages();
799 balloon_stats
.target_pages
= balloon_stats
.current_pages
;
800 balloon_stats
.balloon_low
= 0;
801 balloon_stats
.balloon_high
= 0;
802 balloon_stats
.total_pages
= balloon_stats
.current_pages
;
804 balloon_stats
.schedule_delay
= 1;
805 balloon_stats
.max_schedule_delay
= 32;
806 balloon_stats
.retry_count
= 1;
807 balloon_stats
.max_retry_count
= RETRY_UNLIMITED
;
809 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
810 set_online_page_callback(&xen_online_page
);
811 register_memory_notifier(&xen_memory_nb
);
812 register_sysctl_table(xen_root
);
814 arch_xen_balloon_init(&hostmem_resource
);
822 * Initialize the balloon with pages from the extra memory
823 * regions (see arch/x86/xen/setup.c).
825 for (i
= 0; i
< XEN_EXTRA_MEM_MAX_REGIONS
; i
++)
826 if (xen_extra_mem
[i
].n_pfns
)
827 balloon_add_region(xen_extra_mem
[i
].start_pfn
,
828 xen_extra_mem
[i
].n_pfns
);
832 /* Init the xen-balloon driver. */
837 subsys_initcall(balloon_init
);