2 * VMware Balloon driver.
4 * Copyright (C) 2000-2010, VMware, Inc. All Rights Reserved.
6 * This program is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License as published by the
8 * Free Software Foundation; version 2 of the License and no later version.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
13 * NON INFRINGEMENT. See the GNU General Public License for more
16 * You should have received a copy of the GNU General Public License
17 * along with this program; if not, write to the Free Software
18 * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
20 * Maintained by: Dmitry Torokhov <dtor@vmware.com>
24 * This is VMware physical memory management driver for Linux. The driver
25 * acts like a "balloon" that can be inflated to reclaim physical pages by
26 * reserving them in the guest and invalidating them in the monitor,
27 * freeing up the underlying machine pages so they can be allocated to
28 * other guests. The balloon can also be deflated to allow the guest to
29 * use more physical memory. Higher level policies can control the sizes
30 * of balloons in VMs in order to manage physical memory resources.
34 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
36 #include <linux/types.h>
37 #include <linux/kernel.h>
39 #include <linux/sched.h>
40 #include <linux/module.h>
41 #include <linux/workqueue.h>
42 #include <linux/debugfs.h>
43 #include <linux/seq_file.h>
44 #include <asm/hypervisor.h>
46 MODULE_AUTHOR("VMware, Inc.");
47 MODULE_DESCRIPTION("VMware Memory Control (Balloon) Driver");
48 MODULE_VERSION("1.2.1.3-k");
49 MODULE_ALIAS("dmi:*:svnVMware*:*");
50 MODULE_ALIAS("vmware_vmmemctl");
51 MODULE_LICENSE("GPL");
54 * Various constants controlling rate of inflaint/deflating balloon,
59 * Rate of allocating memory when there is no memory pressure
60 * (driver performs non-sleeping allocations).
62 #define VMW_BALLOON_NOSLEEP_ALLOC_MAX 16384U
65 * Rates of memory allocaton when guest experiences memory pressure
66 * (driver performs sleeping allocations).
68 #define VMW_BALLOON_RATE_ALLOC_MIN 512U
69 #define VMW_BALLOON_RATE_ALLOC_MAX 2048U
70 #define VMW_BALLOON_RATE_ALLOC_INC 16U
73 * Rates for releasing pages while deflating balloon.
75 #define VMW_BALLOON_RATE_FREE_MIN 512U
76 #define VMW_BALLOON_RATE_FREE_MAX 16384U
77 #define VMW_BALLOON_RATE_FREE_INC 16U
80 * When guest is under memory pressure, use a reduced page allocation
81 * rate for next several cycles.
83 #define VMW_BALLOON_SLOW_CYCLES 4
86 * Use __GFP_HIGHMEM to allow pages from HIGHMEM zone. We don't
87 * allow wait (__GFP_WAIT) for NOSLEEP page allocations. Use
88 * __GFP_NOWARN, to suppress page allocation failure warnings.
90 #define VMW_PAGE_ALLOC_NOSLEEP (__GFP_HIGHMEM|__GFP_NOWARN)
93 * Use GFP_HIGHUSER when executing in a separate kernel thread
94 * context and allocation can sleep. This is less stressful to
95 * the guest memory system, since it allows the thread to block
96 * while memory is reclaimed, and won't take pages from emergency
99 #define VMW_PAGE_ALLOC_CANSLEEP (GFP_HIGHUSER)
101 /* Maximum number of page allocations without yielding processor */
102 #define VMW_BALLOON_YIELD_THRESHOLD 1024
104 /* Maximum number of refused pages we accumulate during inflation cycle */
105 #define VMW_BALLOON_MAX_REFUSED 16
108 * Hypervisor communication port definitions.
110 #define VMW_BALLOON_HV_PORT 0x5670
111 #define VMW_BALLOON_HV_MAGIC 0x456c6d6f
112 #define VMW_BALLOON_PROTOCOL_VERSION 2
113 #define VMW_BALLOON_GUEST_ID 1 /* Linux */
115 #define VMW_BALLOON_CMD_START 0
116 #define VMW_BALLOON_CMD_GET_TARGET 1
117 #define VMW_BALLOON_CMD_LOCK 2
118 #define VMW_BALLOON_CMD_UNLOCK 3
119 #define VMW_BALLOON_CMD_GUEST_ID 4
122 #define VMW_BALLOON_SUCCESS 0
123 #define VMW_BALLOON_FAILURE -1
124 #define VMW_BALLOON_ERROR_CMD_INVALID 1
125 #define VMW_BALLOON_ERROR_PPN_INVALID 2
126 #define VMW_BALLOON_ERROR_PPN_LOCKED 3
127 #define VMW_BALLOON_ERROR_PPN_UNLOCKED 4
128 #define VMW_BALLOON_ERROR_PPN_PINNED 5
129 #define VMW_BALLOON_ERROR_PPN_NOTNEEDED 6
130 #define VMW_BALLOON_ERROR_RESET 7
131 #define VMW_BALLOON_ERROR_BUSY 8
133 #define VMWARE_BALLOON_CMD(cmd, data, result) \
135 unsigned long __stat, __dummy1, __dummy2; \
136 __asm__ __volatile__ ("inl (%%dx)" : \
141 "0"(VMW_BALLOON_HV_MAGIC), \
142 "1"(VMW_BALLOON_CMD_##cmd), \
143 "2"(VMW_BALLOON_HV_PORT), \
150 #ifdef CONFIG_DEBUG_FS
151 struct vmballoon_stats
{
154 /* allocation statistics */
156 unsigned int alloc_fail
;
157 unsigned int sleep_alloc
;
158 unsigned int sleep_alloc_fail
;
159 unsigned int refused_alloc
;
160 unsigned int refused_free
;
163 /* monitor operations */
165 unsigned int lock_fail
;
167 unsigned int unlock_fail
;
169 unsigned int target_fail
;
171 unsigned int start_fail
;
172 unsigned int guest_type
;
173 unsigned int guest_type_fail
;
176 #define STATS_INC(stat) (stat)++
178 #define STATS_INC(stat)
183 /* list of reserved physical pages */
184 struct list_head pages
;
186 /* transient list of non-balloonable pages */
187 struct list_head refused_pages
;
188 unsigned int n_refused_pages
;
190 /* balloon size in pages */
197 /* adjustment rates (pages per second) */
198 unsigned int rate_alloc
;
199 unsigned int rate_free
;
201 /* slowdown page allocations for next few cycles */
202 unsigned int slow_allocation_cycles
;
204 #ifdef CONFIG_DEBUG_FS
206 struct vmballoon_stats stats
;
208 /* debugfs file exporting statistics */
209 struct dentry
*dbg_entry
;
212 struct sysinfo sysinfo
;
214 struct delayed_work dwork
;
217 static struct vmballoon balloon
;
220 * Send "start" command to the host, communicating supported version
223 static bool vmballoon_send_start(struct vmballoon
*b
)
225 unsigned long status
, dummy
;
227 STATS_INC(b
->stats
.start
);
229 status
= VMWARE_BALLOON_CMD(START
, VMW_BALLOON_PROTOCOL_VERSION
, dummy
);
230 if (status
== VMW_BALLOON_SUCCESS
)
233 pr_debug("%s - failed, hv returns %ld\n", __func__
, status
);
234 STATS_INC(b
->stats
.start_fail
);
238 static bool vmballoon_check_status(struct vmballoon
*b
, unsigned long status
)
241 case VMW_BALLOON_SUCCESS
:
244 case VMW_BALLOON_ERROR_RESET
:
245 b
->reset_required
= true;
254 * Communicate guest type to the host so that it can adjust ballooning
255 * algorithm to the one most appropriate for the guest. This command
256 * is normally issued after sending "start" command and is part of
257 * standard reset sequence.
259 static bool vmballoon_send_guest_id(struct vmballoon
*b
)
261 unsigned long status
, dummy
;
263 status
= VMWARE_BALLOON_CMD(GUEST_ID
, VMW_BALLOON_GUEST_ID
, dummy
);
265 STATS_INC(b
->stats
.guest_type
);
267 if (vmballoon_check_status(b
, status
))
270 pr_debug("%s - failed, hv returns %ld\n", __func__
, status
);
271 STATS_INC(b
->stats
.guest_type_fail
);
276 * Retrieve desired balloon size from the host.
278 static bool vmballoon_send_get_target(struct vmballoon
*b
, u32
*new_target
)
280 unsigned long status
;
281 unsigned long target
;
286 * si_meminfo() is cheap. Moreover, we want to provide dynamic
287 * max balloon size later. So let us call si_meminfo() every
290 si_meminfo(&b
->sysinfo
);
291 limit
= b
->sysinfo
.totalram
;
293 /* Ensure limit fits in 32-bits */
294 limit32
= (u32
)limit
;
295 if (limit
!= limit32
)
299 STATS_INC(b
->stats
.target
);
301 status
= VMWARE_BALLOON_CMD(GET_TARGET
, limit
, target
);
302 if (vmballoon_check_status(b
, status
)) {
303 *new_target
= target
;
307 pr_debug("%s - failed, hv returns %ld\n", __func__
, status
);
308 STATS_INC(b
->stats
.target_fail
);
313 * Notify the host about allocated page so that host can use it without
314 * fear that guest will need it. Host may reject some pages, we need to
315 * check the return value and maybe submit a different page.
317 static bool vmballoon_send_lock_page(struct vmballoon
*b
, unsigned long pfn
,
318 unsigned int *hv_status
)
320 unsigned long status
, dummy
;
327 STATS_INC(b
->stats
.lock
);
329 *hv_status
= status
= VMWARE_BALLOON_CMD(LOCK
, pfn
, dummy
);
330 if (vmballoon_check_status(b
, status
))
333 pr_debug("%s - ppn %lx, hv returns %ld\n", __func__
, pfn
, status
);
334 STATS_INC(b
->stats
.lock_fail
);
339 * Notify the host that guest intends to release given page back into
340 * the pool of available (to the guest) pages.
342 static bool vmballoon_send_unlock_page(struct vmballoon
*b
, unsigned long pfn
)
344 unsigned long status
, dummy
;
351 STATS_INC(b
->stats
.unlock
);
353 status
= VMWARE_BALLOON_CMD(UNLOCK
, pfn
, dummy
);
354 if (vmballoon_check_status(b
, status
))
357 pr_debug("%s - ppn %lx, hv returns %ld\n", __func__
, pfn
, status
);
358 STATS_INC(b
->stats
.unlock_fail
);
363 * Quickly release all pages allocated for the balloon. This function is
364 * called when host decides to "reset" balloon for one reason or another.
365 * Unlike normal "deflate" we do not (shall not) notify host of the pages
368 static void vmballoon_pop(struct vmballoon
*b
)
370 struct page
*page
, *next
;
371 unsigned int count
= 0;
373 list_for_each_entry_safe(page
, next
, &b
->pages
, lru
) {
374 list_del(&page
->lru
);
376 STATS_INC(b
->stats
.free
);
379 if (++count
>= b
->rate_free
) {
387 * Perform standard reset sequence by popping the balloon (in case it
388 * is not empty) and then restarting protocol. This operation normally
389 * happens when host responds with VMW_BALLOON_ERROR_RESET to a command.
391 static void vmballoon_reset(struct vmballoon
*b
)
393 /* free all pages, skipping monitor unlock */
396 if (vmballoon_send_start(b
)) {
397 b
->reset_required
= false;
398 if (!vmballoon_send_guest_id(b
))
399 pr_err("failed to send guest ID to the host\n");
404 * Allocate (or reserve) a page for the balloon and notify the host. If host
405 * refuses the page put it on "refuse" list and allocate another one until host
406 * is satisfied. "Refused" pages are released at the end of inflation cycle
407 * (when we allocate b->rate_alloc pages).
409 static int vmballoon_reserve_page(struct vmballoon
*b
, bool can_sleep
)
413 unsigned int hv_status
;
415 flags
= can_sleep
? VMW_PAGE_ALLOC_CANSLEEP
: VMW_PAGE_ALLOC_NOSLEEP
;
419 STATS_INC(b
->stats
.alloc
);
421 STATS_INC(b
->stats
.sleep_alloc
);
423 page
= alloc_page(flags
);
426 STATS_INC(b
->stats
.alloc_fail
);
428 STATS_INC(b
->stats
.sleep_alloc_fail
);
433 locked
= vmballoon_send_lock_page(b
, page_to_pfn(page
), &hv_status
);
435 STATS_INC(b
->stats
.refused_alloc
);
437 if (hv_status
== VMW_BALLOON_ERROR_RESET
||
438 hv_status
== VMW_BALLOON_ERROR_PPN_NOTNEEDED
) {
444 * Place page on the list of non-balloonable pages
445 * and retry allocation, unless we already accumulated
446 * too many of them, in which case take a breather.
448 list_add(&page
->lru
, &b
->refused_pages
);
449 if (++b
->n_refused_pages
>= VMW_BALLOON_MAX_REFUSED
)
454 /* track allocated page */
455 list_add(&page
->lru
, &b
->pages
);
457 /* update balloon size */
464 * Release the page allocated for the balloon. Note that we first notify
465 * the host so it can make sure the page will be available for the guest
468 static int vmballoon_release_page(struct vmballoon
*b
, struct page
*page
)
470 if (!vmballoon_send_unlock_page(b
, page_to_pfn(page
)))
473 list_del(&page
->lru
);
475 /* deallocate page */
477 STATS_INC(b
->stats
.free
);
479 /* update balloon size */
486 * Release pages that were allocated while attempting to inflate the
487 * balloon but were refused by the host for one reason or another.
489 static void vmballoon_release_refused_pages(struct vmballoon
*b
)
491 struct page
*page
, *next
;
493 list_for_each_entry_safe(page
, next
, &b
->refused_pages
, lru
) {
494 list_del(&page
->lru
);
496 STATS_INC(b
->stats
.refused_free
);
499 b
->n_refused_pages
= 0;
503 * Inflate the balloon towards its target size. Note that we try to limit
504 * the rate of allocation to make sure we are not choking the rest of the
507 static void vmballoon_inflate(struct vmballoon
*b
)
512 unsigned int allocations
= 0;
514 bool alloc_can_sleep
= false;
516 pr_debug("%s - size: %d, target %d\n", __func__
, b
->size
, b
->target
);
519 * First try NOSLEEP page allocations to inflate balloon.
521 * If we do not throttle nosleep allocations, we can drain all
522 * free pages in the guest quickly (if the balloon target is high).
523 * As a side-effect, draining free pages helps to inform (force)
524 * the guest to start swapping if balloon target is not met yet,
525 * which is a desired behavior. However, balloon driver can consume
526 * all available CPU cycles if too many pages are allocated in a
527 * second. Therefore, we throttle nosleep allocations even when
528 * the guest is not under memory pressure. OTOH, if we have already
529 * predicted that the guest is under memory pressure, then we
530 * slowdown page allocations considerably.
533 goal
= b
->target
- b
->size
;
535 * Start with no sleep allocation rate which may be higher
536 * than sleeping allocation rate.
538 rate
= b
->slow_allocation_cycles
?
539 b
->rate_alloc
: VMW_BALLOON_NOSLEEP_ALLOC_MAX
;
541 pr_debug("%s - goal: %d, no-sleep rate: %d, sleep rate: %d\n",
542 __func__
, goal
, rate
, b
->rate_alloc
);
544 for (i
= 0; i
< goal
; i
++) {
546 error
= vmballoon_reserve_page(b
, alloc_can_sleep
);
548 if (error
!= -ENOMEM
) {
550 * Not a page allocation failure, stop this
551 * cycle. Maybe we'll get new target from
557 if (alloc_can_sleep
) {
559 * CANSLEEP page allocation failed, so guest
560 * is under severe memory pressure. Quickly
561 * decrease allocation rate.
563 b
->rate_alloc
= max(b
->rate_alloc
/ 2,
564 VMW_BALLOON_RATE_ALLOC_MIN
);
569 * NOSLEEP page allocation failed, so the guest is
570 * under memory pressure. Let us slow down page
571 * allocations for next few cycles so that the guest
572 * gets out of memory pressure. Also, if we already
573 * allocated b->rate_alloc pages, let's pause,
574 * otherwise switch to sleeping allocations.
576 b
->slow_allocation_cycles
= VMW_BALLOON_SLOW_CYCLES
;
578 if (i
>= b
->rate_alloc
)
581 alloc_can_sleep
= true;
582 /* Lower rate for sleeping allocations. */
583 rate
= b
->rate_alloc
;
586 if (++allocations
> VMW_BALLOON_YIELD_THRESHOLD
) {
592 /* We allocated enough pages, let's take a break. */
598 * We reached our goal without failures so try increasing
601 if (error
== 0 && i
>= b
->rate_alloc
) {
602 unsigned int mult
= i
/ b
->rate_alloc
;
605 min(b
->rate_alloc
+ mult
* VMW_BALLOON_RATE_ALLOC_INC
,
606 VMW_BALLOON_RATE_ALLOC_MAX
);
609 vmballoon_release_refused_pages(b
);
613 * Decrease the size of the balloon allowing guest to use more memory.
615 static void vmballoon_deflate(struct vmballoon
*b
)
617 struct page
*page
, *next
;
622 pr_debug("%s - size: %d, target %d\n", __func__
, b
->size
, b
->target
);
624 /* limit deallocation rate */
625 goal
= min(b
->size
- b
->target
, b
->rate_free
);
627 pr_debug("%s - goal: %d, rate: %d\n", __func__
, goal
, b
->rate_free
);
629 /* free pages to reach target */
630 list_for_each_entry_safe(page
, next
, &b
->pages
, lru
) {
631 error
= vmballoon_release_page(b
, page
);
633 /* quickly decrease rate in case of error */
634 b
->rate_free
= max(b
->rate_free
/ 2,
635 VMW_BALLOON_RATE_FREE_MIN
);
643 /* slowly increase rate if there were no errors */
644 b
->rate_free
= min(b
->rate_free
+ VMW_BALLOON_RATE_FREE_INC
,
645 VMW_BALLOON_RATE_FREE_MAX
);
649 * Balloon work function: reset protocol, if needed, get the new size and
650 * adjust balloon as needed. Repeat in 1 sec.
652 static void vmballoon_work(struct work_struct
*work
)
654 struct delayed_work
*dwork
= to_delayed_work(work
);
655 struct vmballoon
*b
= container_of(dwork
, struct vmballoon
, dwork
);
658 STATS_INC(b
->stats
.timer
);
660 if (b
->reset_required
)
663 if (b
->slow_allocation_cycles
> 0)
664 b
->slow_allocation_cycles
--;
666 if (vmballoon_send_get_target(b
, &target
)) {
667 /* update target, adjust size */
670 if (b
->size
< target
)
671 vmballoon_inflate(b
);
672 else if (b
->size
> target
)
673 vmballoon_deflate(b
);
677 * We are using a freezable workqueue so that balloon operations are
678 * stopped while the system transitions to/from sleep/hibernation.
680 queue_delayed_work(system_freezable_wq
,
681 dwork
, round_jiffies_relative(HZ
));
687 #ifdef CONFIG_DEBUG_FS
689 static int vmballoon_debug_show(struct seq_file
*f
, void *offset
)
691 struct vmballoon
*b
= f
->private;
692 struct vmballoon_stats
*stats
= &b
->stats
;
694 /* format size info */
696 "target: %8d pages\n"
697 "current: %8d pages\n",
700 /* format rate info */
702 "rateNoSleepAlloc: %8d pages/sec\n"
703 "rateSleepAlloc: %8d pages/sec\n"
704 "rateFree: %8d pages/sec\n",
705 VMW_BALLOON_NOSLEEP_ALLOC_MAX
,
706 b
->rate_alloc
, b
->rate_free
);
711 "start: %8u (%4u failed)\n"
712 "guestType: %8u (%4u failed)\n"
713 "lock: %8u (%4u failed)\n"
714 "unlock: %8u (%4u failed)\n"
715 "target: %8u (%4u failed)\n"
716 "primNoSleepAlloc: %8u (%4u failed)\n"
717 "primCanSleepAlloc: %8u (%4u failed)\n"
722 stats
->start
, stats
->start_fail
,
723 stats
->guest_type
, stats
->guest_type_fail
,
724 stats
->lock
, stats
->lock_fail
,
725 stats
->unlock
, stats
->unlock_fail
,
726 stats
->target
, stats
->target_fail
,
727 stats
->alloc
, stats
->alloc_fail
,
728 stats
->sleep_alloc
, stats
->sleep_alloc_fail
,
730 stats
->refused_alloc
, stats
->refused_free
);
735 static int vmballoon_debug_open(struct inode
*inode
, struct file
*file
)
737 return single_open(file
, vmballoon_debug_show
, inode
->i_private
);
740 static const struct file_operations vmballoon_debug_fops
= {
741 .owner
= THIS_MODULE
,
742 .open
= vmballoon_debug_open
,
745 .release
= single_release
,
748 static int __init
vmballoon_debugfs_init(struct vmballoon
*b
)
752 b
->dbg_entry
= debugfs_create_file("vmmemctl", S_IRUGO
, NULL
, b
,
753 &vmballoon_debug_fops
);
754 if (IS_ERR(b
->dbg_entry
)) {
755 error
= PTR_ERR(b
->dbg_entry
);
756 pr_err("failed to create debugfs entry, error: %d\n", error
);
763 static void __exit
vmballoon_debugfs_exit(struct vmballoon
*b
)
765 debugfs_remove(b
->dbg_entry
);
770 static inline int vmballoon_debugfs_init(struct vmballoon
*b
)
775 static inline void vmballoon_debugfs_exit(struct vmballoon
*b
)
779 #endif /* CONFIG_DEBUG_FS */
781 static int __init
vmballoon_init(void)
786 * Check if we are running on VMware's hypervisor and bail out
789 if (x86_hyper
!= &x86_hyper_vmware
)
792 INIT_LIST_HEAD(&balloon
.pages
);
793 INIT_LIST_HEAD(&balloon
.refused_pages
);
795 /* initialize rates */
796 balloon
.rate_alloc
= VMW_BALLOON_RATE_ALLOC_MAX
;
797 balloon
.rate_free
= VMW_BALLOON_RATE_FREE_MAX
;
799 INIT_DELAYED_WORK(&balloon
.dwork
, vmballoon_work
);
804 if (!vmballoon_send_start(&balloon
)) {
805 pr_err("failed to send start command to the host\n");
809 if (!vmballoon_send_guest_id(&balloon
)) {
810 pr_err("failed to send guest ID to the host\n");
814 error
= vmballoon_debugfs_init(&balloon
);
818 queue_delayed_work(system_freezable_wq
, &balloon
.dwork
, 0);
822 module_init(vmballoon_init
);
824 static void __exit
vmballoon_exit(void)
826 cancel_delayed_work_sync(&balloon
.dwork
);
828 vmballoon_debugfs_exit(&balloon
);
831 * Deallocate all reserved memory, and reset connection with monitor.
832 * Reset connection before deallocating memory to avoid potential for
833 * additional spurious resets from guest touching deallocated pages.
835 vmballoon_send_start(&balloon
);
836 vmballoon_pop(&balloon
);
838 module_exit(vmballoon_exit
);