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("1.2.1.1-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 statustics */
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;
218 static struct workqueue_struct *vmballoon_wq;
221 * Send "start" command to the host, communicating supported version
224 static bool vmballoon_send_start(struct vmballoon *b)
226 unsigned long status, dummy;
228 STATS_INC(b->stats.start);
230 status = VMWARE_BALLOON_CMD(START, VMW_BALLOON_PROTOCOL_VERSION, dummy);
231 if (status == VMW_BALLOON_SUCCESS)
234 pr_debug("%s - failed, hv returns %ld\n", __func__, status);
235 STATS_INC(b->stats.start_fail);
239 static bool vmballoon_check_status(struct vmballoon *b, unsigned long status)
242 case VMW_BALLOON_SUCCESS:
245 case VMW_BALLOON_ERROR_RESET:
246 b->reset_required = true;
255 * Communicate guest type to the host so that it can adjust ballooning
256 * algorithm to the one most appropriate for the guest. This command
257 * is normally issued after sending "start" command and is part of
258 * standard reset sequence.
260 static bool vmballoon_send_guest_id(struct vmballoon *b)
262 unsigned long status, dummy;
264 status = VMWARE_BALLOON_CMD(GUEST_ID, VMW_BALLOON_GUEST_ID, dummy);
266 STATS_INC(b->stats.guest_type);
268 if (vmballoon_check_status(b, status))
271 pr_debug("%s - failed, hv returns %ld\n", __func__, status);
272 STATS_INC(b->stats.guest_type_fail);
277 * Retrieve desired balloon size from the host.
279 static bool vmballoon_send_get_target(struct vmballoon *b, u32 *new_target)
281 unsigned long status;
282 unsigned long target;
287 * si_meminfo() is cheap. Moreover, we want to provide dynamic
288 * max balloon size later. So let us call si_meminfo() every
291 si_meminfo(&b->sysinfo);
292 limit = b->sysinfo.totalram;
294 /* Ensure limit fits in 32-bits */
295 limit32 = (u32)limit;
296 if (limit != limit32)
300 STATS_INC(b->stats.target);
302 status = VMWARE_BALLOON_CMD(GET_TARGET, limit, target);
303 if (vmballoon_check_status(b, status)) {
304 *new_target = target;
308 pr_debug("%s - failed, hv returns %ld\n", __func__, status);
309 STATS_INC(b->stats.target_fail);
314 * Notify the host about allocated page so that host can use it without
315 * fear that guest will need it. Host may reject some pages, we need to
316 * check the return value and maybe submit a different page.
318 static bool vmballoon_send_lock_page(struct vmballoon *b, unsigned long pfn)
320 unsigned long status, dummy;
327 STATS_INC(b->stats.lock);
329 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)
417 STATS_INC(b->stats.alloc);
419 STATS_INC(b->stats.sleep_alloc);
421 flags = can_sleep ? VMW_PAGE_ALLOC_CANSLEEP : VMW_PAGE_ALLOC_NOSLEEP;
422 page = alloc_page(flags);
425 STATS_INC(b->stats.alloc_fail);
427 STATS_INC(b->stats.sleep_alloc_fail);
432 locked = vmballoon_send_lock_page(b, page_to_pfn(page));
434 STATS_INC(b->stats.refused_alloc);
436 if (b->reset_required) {
442 * Place page on the list of non-balloonable pages
443 * and retry allocation, unless we already accumulated
444 * too many of them, in which case take a breather.
446 list_add(&page->lru, &b->refused_pages);
447 if (++b->n_refused_pages >= VMW_BALLOON_MAX_REFUSED)
452 /* track allocated page */
453 list_add(&page->lru, &b->pages);
455 /* update balloon size */
462 * Release the page allocated for the balloon. Note that we first notify
463 * the host so it can make sure the page will be available for the guest
466 static int vmballoon_release_page(struct vmballoon *b, struct page *page)
468 if (!vmballoon_send_unlock_page(b, page_to_pfn(page)))
471 list_del(&page->lru);
473 /* deallocate page */
475 STATS_INC(b->stats.free);
477 /* update balloon size */
484 * Release pages that were allocated while attempting to inflate the
485 * balloon but were refused by the host for one reason or another.
487 static void vmballoon_release_refused_pages(struct vmballoon *b)
489 struct page *page, *next;
491 list_for_each_entry_safe(page, next, &b->refused_pages, lru) {
492 list_del(&page->lru);
494 STATS_INC(b->stats.refused_free);
497 b->n_refused_pages = 0;
501 * Inflate the balloon towards its target size. Note that we try to limit
502 * the rate of allocation to make sure we are not choking the rest of the
505 static void vmballoon_inflate(struct vmballoon *b)
510 unsigned int allocations = 0;
512 bool alloc_can_sleep = false;
514 pr_debug("%s - size: %d, target %d\n", __func__, b->size, b->target);
517 * First try NOSLEEP page allocations to inflate balloon.
519 * If we do not throttle nosleep allocations, we can drain all
520 * free pages in the guest quickly (if the balloon target is high).
521 * As a side-effect, draining free pages helps to inform (force)
522 * the guest to start swapping if balloon target is not met yet,
523 * which is a desired behavior. However, balloon driver can consume
524 * all available CPU cycles if too many pages are allocated in a
525 * second. Therefore, we throttle nosleep allocations even when
526 * the guest is not under memory pressure. OTOH, if we have already
527 * predicted that the guest is under memory pressure, then we
528 * slowdown page allocations considerably.
531 goal = b->target - b->size;
533 * Start with no sleep allocation rate which may be higher
534 * than sleeping allocation rate.
536 rate = b->slow_allocation_cycles ?
537 b->rate_alloc : VMW_BALLOON_NOSLEEP_ALLOC_MAX;
539 pr_debug("%s - goal: %d, no-sleep rate: %d, sleep rate: %d\n",
540 __func__, goal, rate, b->rate_alloc);
542 for (i = 0; i < goal; i++) {
544 error = vmballoon_reserve_page(b, alloc_can_sleep);
546 if (error != -ENOMEM) {
548 * Not a page allocation failure, stop this
549 * cycle. Maybe we'll get new target from
555 if (alloc_can_sleep) {
557 * CANSLEEP page allocation failed, so guest
558 * is under severe memory pressure. Quickly
559 * decrease allocation rate.
561 b->rate_alloc = max(b->rate_alloc / 2,
562 VMW_BALLOON_RATE_ALLOC_MIN);
567 * NOSLEEP page allocation failed, so the guest is
568 * under memory pressure. Let us slow down page
569 * allocations for next few cycles so that the guest
570 * gets out of memory pressure. Also, if we already
571 * allocated b->rate_alloc pages, let's pause,
572 * otherwise switch to sleeping allocations.
574 b->slow_allocation_cycles = VMW_BALLOON_SLOW_CYCLES;
576 if (i >= b->rate_alloc)
579 alloc_can_sleep = true;
580 /* Lower rate for sleeping allocations. */
581 rate = b->rate_alloc;
584 if (++allocations > VMW_BALLOON_YIELD_THRESHOLD) {
590 /* We allocated enough pages, let's take a break. */
596 * We reached our goal without failures so try increasing
599 if (error == 0 && i >= b->rate_alloc) {
600 unsigned int mult = i / b->rate_alloc;
603 min(b->rate_alloc + mult * VMW_BALLOON_RATE_ALLOC_INC,
604 VMW_BALLOON_RATE_ALLOC_MAX);
607 vmballoon_release_refused_pages(b);
611 * Decrease the size of the balloon allowing guest to use more memory.
613 static void vmballoon_deflate(struct vmballoon *b)
615 struct page *page, *next;
620 pr_debug("%s - size: %d, target %d\n", __func__, b->size, b->target);
622 /* limit deallocation rate */
623 goal = min(b->size - b->target, b->rate_free);
625 pr_debug("%s - goal: %d, rate: %d\n", __func__, goal, b->rate_free);
627 /* free pages to reach target */
628 list_for_each_entry_safe(page, next, &b->pages, lru) {
629 error = vmballoon_release_page(b, page);
631 /* quickly decrease rate in case of error */
632 b->rate_free = max(b->rate_free / 2,
633 VMW_BALLOON_RATE_FREE_MIN);
641 /* slowly increase rate if there were no errors */
642 b->rate_free = min(b->rate_free + VMW_BALLOON_RATE_FREE_INC,
643 VMW_BALLOON_RATE_FREE_MAX);
647 * Balloon work function: reset protocol, if needed, get the new size and
648 * adjust balloon as needed. Repeat in 1 sec.
650 static void vmballoon_work(struct work_struct *work)
652 struct delayed_work *dwork = to_delayed_work(work);
653 struct vmballoon *b = container_of(dwork, struct vmballoon, dwork);
656 STATS_INC(b->stats.timer);
658 if (b->reset_required)
661 if (b->slow_allocation_cycles > 0)
662 b->slow_allocation_cycles--;
664 if (vmballoon_send_get_target(b, &target)) {
665 /* update target, adjust size */
668 if (b->size < target)
669 vmballoon_inflate(b);
670 else if (b->size > target)
671 vmballoon_deflate(b);
674 queue_delayed_work(vmballoon_wq, dwork, round_jiffies_relative(HZ));
680 #ifdef CONFIG_DEBUG_FS
682 static int vmballoon_debug_show(struct seq_file *f, void *offset)
684 struct vmballoon *b = f->private;
685 struct vmballoon_stats *stats = &b->stats;
687 /* format size info */
689 "target: %8d pages\n"
690 "current: %8d pages\n",
693 /* format rate info */
695 "rateNoSleepAlloc: %8d pages/sec\n"
696 "rateSleepAlloc: %8d pages/sec\n"
697 "rateFree: %8d pages/sec\n",
698 VMW_BALLOON_NOSLEEP_ALLOC_MAX,
699 b->rate_alloc, b->rate_free);
704 "start: %8u (%4u failed)\n"
705 "guestType: %8u (%4u failed)\n"
706 "lock: %8u (%4u failed)\n"
707 "unlock: %8u (%4u failed)\n"
708 "target: %8u (%4u failed)\n"
709 "primNoSleepAlloc: %8u (%4u failed)\n"
710 "primCanSleepAlloc: %8u (%4u failed)\n"
715 stats->start, stats->start_fail,
716 stats->guest_type, stats->guest_type_fail,
717 stats->lock, stats->lock_fail,
718 stats->unlock, stats->unlock_fail,
719 stats->target, stats->target_fail,
720 stats->alloc, stats->alloc_fail,
721 stats->sleep_alloc, stats->sleep_alloc_fail,
723 stats->refused_alloc, stats->refused_free);
728 static int vmballoon_debug_open(struct inode *inode, struct file *file)
730 return single_open(file, vmballoon_debug_show, inode->i_private);
733 static const struct file_operations vmballoon_debug_fops = {
734 .owner = THIS_MODULE,
735 .open = vmballoon_debug_open,
738 .release = single_release,
741 static int __init vmballoon_debugfs_init(struct vmballoon *b)
745 b->dbg_entry = debugfs_create_file("vmmemctl", S_IRUGO, NULL, b,
746 &vmballoon_debug_fops);
747 if (IS_ERR(b->dbg_entry)) {
748 error = PTR_ERR(b->dbg_entry);
749 pr_err("failed to create debugfs entry, error: %d\n", error);
756 static void __exit vmballoon_debugfs_exit(struct vmballoon *b)
758 debugfs_remove(b->dbg_entry);
763 static inline int vmballoon_debugfs_init(struct vmballoon *b)
768 static inline void vmballoon_debugfs_exit(struct vmballoon *b)
772 #endif /* CONFIG_DEBUG_FS */
774 static int __init vmballoon_init(void)
779 * Check if we are running on VMware's hypervisor and bail out
782 if (x86_hyper != &x86_hyper_vmware)
785 vmballoon_wq = create_freezeable_workqueue("vmmemctl");
787 pr_err("failed to create workqueue\n");
791 INIT_LIST_HEAD(&balloon.pages);
792 INIT_LIST_HEAD(&balloon.refused_pages);
794 /* initialize rates */
795 balloon.rate_alloc = VMW_BALLOON_RATE_ALLOC_MAX;
796 balloon.rate_free = VMW_BALLOON_RATE_FREE_MAX;
798 INIT_DELAYED_WORK(&balloon.dwork, vmballoon_work);
803 if (!vmballoon_send_start(&balloon)) {
804 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");
815 error = vmballoon_debugfs_init(&balloon);
819 queue_delayed_work(vmballoon_wq, &balloon.dwork, 0);
824 destroy_workqueue(vmballoon_wq);
827 module_init(vmballoon_init);
829 static void __exit vmballoon_exit(void)
831 cancel_delayed_work_sync(&balloon.dwork);
832 destroy_workqueue(vmballoon_wq);
834 vmballoon_debugfs_exit(&balloon);
837 * Deallocate all reserved memory, and reset connection with monitor.
838 * Reset connection before deallocating memory to avoid potential for
839 * additional spurious resets from guest touching deallocated pages.
841 vmballoon_send_start(&balloon);
842 vmballoon_pop(&balloon);
844 module_exit(vmballoon_exit);