Loading...
1/*
2 * VMware Balloon driver.
3 *
4 * Copyright (C) 2000-2010, VMware, Inc. All Rights Reserved.
5 *
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.
9 *
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
14 * details.
15 *
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.
19 *
20 * Maintained by: Dmitry Torokhov <dtor@vmware.com>
21 */
22
23/*
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.
31 */
32
33//#define DEBUG
34#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
35
36#include <linux/types.h>
37#include <linux/kernel.h>
38#include <linux/mm.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>
45
46MODULE_AUTHOR("VMware, Inc.");
47MODULE_DESCRIPTION("VMware Memory Control (Balloon) Driver");
48MODULE_VERSION("1.2.1.3-k");
49MODULE_ALIAS("dmi:*:svnVMware*:*");
50MODULE_ALIAS("vmware_vmmemctl");
51MODULE_LICENSE("GPL");
52
53/*
54 * Various constants controlling rate of inflaint/deflating balloon,
55 * measured in pages.
56 */
57
58/*
59 * Rate of allocating memory when there is no memory pressure
60 * (driver performs non-sleeping allocations).
61 */
62#define VMW_BALLOON_NOSLEEP_ALLOC_MAX 16384U
63
64/*
65 * Rates of memory allocaton when guest experiences memory pressure
66 * (driver performs sleeping allocations).
67 */
68#define VMW_BALLOON_RATE_ALLOC_MIN 512U
69#define VMW_BALLOON_RATE_ALLOC_MAX 2048U
70#define VMW_BALLOON_RATE_ALLOC_INC 16U
71
72/*
73 * Rates for releasing pages while deflating balloon.
74 */
75#define VMW_BALLOON_RATE_FREE_MIN 512U
76#define VMW_BALLOON_RATE_FREE_MAX 16384U
77#define VMW_BALLOON_RATE_FREE_INC 16U
78
79/*
80 * When guest is under memory pressure, use a reduced page allocation
81 * rate for next several cycles.
82 */
83#define VMW_BALLOON_SLOW_CYCLES 4
84
85/*
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.
89 */
90#define VMW_PAGE_ALLOC_NOSLEEP (__GFP_HIGHMEM|__GFP_NOWARN)
91
92/*
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
97 * low-memory pools.
98 */
99#define VMW_PAGE_ALLOC_CANSLEEP (GFP_HIGHUSER)
100
101/* Maximum number of page allocations without yielding processor */
102#define VMW_BALLOON_YIELD_THRESHOLD 1024
103
104/* Maximum number of refused pages we accumulate during inflation cycle */
105#define VMW_BALLOON_MAX_REFUSED 16
106
107/*
108 * Hypervisor communication port definitions.
109 */
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 */
114
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
120
121/* error codes */
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
132
133#define VMWARE_BALLOON_CMD(cmd, data, result) \
134({ \
135 unsigned long __stat, __dummy1, __dummy2; \
136 __asm__ __volatile__ ("inl (%%dx)" : \
137 "=a"(__stat), \
138 "=c"(__dummy1), \
139 "=d"(__dummy2), \
140 "=b"(result) : \
141 "0"(VMW_BALLOON_HV_MAGIC), \
142 "1"(VMW_BALLOON_CMD_##cmd), \
143 "2"(VMW_BALLOON_HV_PORT), \
144 "3"(data) : \
145 "memory"); \
146 result &= -1UL; \
147 __stat & -1UL; \
148})
149
150#ifdef CONFIG_DEBUG_FS
151struct vmballoon_stats {
152 unsigned int timer;
153
154 /* allocation statustics */
155 unsigned int alloc;
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;
161 unsigned int free;
162
163 /* monitor operations */
164 unsigned int lock;
165 unsigned int lock_fail;
166 unsigned int unlock;
167 unsigned int unlock_fail;
168 unsigned int target;
169 unsigned int target_fail;
170 unsigned int start;
171 unsigned int start_fail;
172 unsigned int guest_type;
173 unsigned int guest_type_fail;
174};
175
176#define STATS_INC(stat) (stat)++
177#else
178#define STATS_INC(stat)
179#endif
180
181struct vmballoon {
182
183 /* list of reserved physical pages */
184 struct list_head pages;
185
186 /* transient list of non-balloonable pages */
187 struct list_head refused_pages;
188 unsigned int n_refused_pages;
189
190 /* balloon size in pages */
191 unsigned int size;
192 unsigned int target;
193
194 /* reset flag */
195 bool reset_required;
196
197 /* adjustment rates (pages per second) */
198 unsigned int rate_alloc;
199 unsigned int rate_free;
200
201 /* slowdown page allocations for next few cycles */
202 unsigned int slow_allocation_cycles;
203
204#ifdef CONFIG_DEBUG_FS
205 /* statistics */
206 struct vmballoon_stats stats;
207
208 /* debugfs file exporting statistics */
209 struct dentry *dbg_entry;
210#endif
211
212 struct sysinfo sysinfo;
213
214 struct delayed_work dwork;
215};
216
217static struct vmballoon balloon;
218
219/*
220 * Send "start" command to the host, communicating supported version
221 * of the protocol.
222 */
223static bool vmballoon_send_start(struct vmballoon *b)
224{
225 unsigned long status, dummy;
226
227 STATS_INC(b->stats.start);
228
229 status = VMWARE_BALLOON_CMD(START, VMW_BALLOON_PROTOCOL_VERSION, dummy);
230 if (status == VMW_BALLOON_SUCCESS)
231 return true;
232
233 pr_debug("%s - failed, hv returns %ld\n", __func__, status);
234 STATS_INC(b->stats.start_fail);
235 return false;
236}
237
238static bool vmballoon_check_status(struct vmballoon *b, unsigned long status)
239{
240 switch (status) {
241 case VMW_BALLOON_SUCCESS:
242 return true;
243
244 case VMW_BALLOON_ERROR_RESET:
245 b->reset_required = true;
246 /* fall through */
247
248 default:
249 return false;
250 }
251}
252
253/*
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.
258 */
259static bool vmballoon_send_guest_id(struct vmballoon *b)
260{
261 unsigned long status, dummy;
262
263 status = VMWARE_BALLOON_CMD(GUEST_ID, VMW_BALLOON_GUEST_ID, dummy);
264
265 STATS_INC(b->stats.guest_type);
266
267 if (vmballoon_check_status(b, status))
268 return true;
269
270 pr_debug("%s - failed, hv returns %ld\n", __func__, status);
271 STATS_INC(b->stats.guest_type_fail);
272 return false;
273}
274
275/*
276 * Retrieve desired balloon size from the host.
277 */
278static bool vmballoon_send_get_target(struct vmballoon *b, u32 *new_target)
279{
280 unsigned long status;
281 unsigned long target;
282 unsigned long limit;
283 u32 limit32;
284
285 /*
286 * si_meminfo() is cheap. Moreover, we want to provide dynamic
287 * max balloon size later. So let us call si_meminfo() every
288 * iteration.
289 */
290 si_meminfo(&b->sysinfo);
291 limit = b->sysinfo.totalram;
292
293 /* Ensure limit fits in 32-bits */
294 limit32 = (u32)limit;
295 if (limit != limit32)
296 return false;
297
298 /* update stats */
299 STATS_INC(b->stats.target);
300
301 status = VMWARE_BALLOON_CMD(GET_TARGET, limit, target);
302 if (vmballoon_check_status(b, status)) {
303 *new_target = target;
304 return true;
305 }
306
307 pr_debug("%s - failed, hv returns %ld\n", __func__, status);
308 STATS_INC(b->stats.target_fail);
309 return false;
310}
311
312/*
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.
316 */
317static bool vmballoon_send_lock_page(struct vmballoon *b, unsigned long pfn,
318 unsigned int *hv_status)
319{
320 unsigned long status, dummy;
321 u32 pfn32;
322
323 pfn32 = (u32)pfn;
324 if (pfn32 != pfn)
325 return false;
326
327 STATS_INC(b->stats.lock);
328
329 *hv_status = status = VMWARE_BALLOON_CMD(LOCK, pfn, dummy);
330 if (vmballoon_check_status(b, status))
331 return true;
332
333 pr_debug("%s - ppn %lx, hv returns %ld\n", __func__, pfn, status);
334 STATS_INC(b->stats.lock_fail);
335 return false;
336}
337
338/*
339 * Notify the host that guest intends to release given page back into
340 * the pool of available (to the guest) pages.
341 */
342static bool vmballoon_send_unlock_page(struct vmballoon *b, unsigned long pfn)
343{
344 unsigned long status, dummy;
345 u32 pfn32;
346
347 pfn32 = (u32)pfn;
348 if (pfn32 != pfn)
349 return false;
350
351 STATS_INC(b->stats.unlock);
352
353 status = VMWARE_BALLOON_CMD(UNLOCK, pfn, dummy);
354 if (vmballoon_check_status(b, status))
355 return true;
356
357 pr_debug("%s - ppn %lx, hv returns %ld\n", __func__, pfn, status);
358 STATS_INC(b->stats.unlock_fail);
359 return false;
360}
361
362/*
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
366 * being released.
367 */
368static void vmballoon_pop(struct vmballoon *b)
369{
370 struct page *page, *next;
371 unsigned int count = 0;
372
373 list_for_each_entry_safe(page, next, &b->pages, lru) {
374 list_del(&page->lru);
375 __free_page(page);
376 STATS_INC(b->stats.free);
377 b->size--;
378
379 if (++count >= b->rate_free) {
380 count = 0;
381 cond_resched();
382 }
383 }
384}
385
386/*
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.
390 */
391static void vmballoon_reset(struct vmballoon *b)
392{
393 /* free all pages, skipping monitor unlock */
394 vmballoon_pop(b);
395
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");
400 }
401}
402
403/*
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).
408 */
409static int vmballoon_reserve_page(struct vmballoon *b, bool can_sleep)
410{
411 struct page *page;
412 gfp_t flags;
413 unsigned int hv_status;
414 bool locked = false;
415
416 do {
417 if (!can_sleep)
418 STATS_INC(b->stats.alloc);
419 else
420 STATS_INC(b->stats.sleep_alloc);
421
422 flags = can_sleep ? VMW_PAGE_ALLOC_CANSLEEP : VMW_PAGE_ALLOC_NOSLEEP;
423 page = alloc_page(flags);
424 if (!page) {
425 if (!can_sleep)
426 STATS_INC(b->stats.alloc_fail);
427 else
428 STATS_INC(b->stats.sleep_alloc_fail);
429 return -ENOMEM;
430 }
431
432 /* inform monitor */
433 locked = vmballoon_send_lock_page(b, page_to_pfn(page), &hv_status);
434 if (!locked) {
435 STATS_INC(b->stats.refused_alloc);
436
437 if (hv_status == VMW_BALLOON_ERROR_RESET ||
438 hv_status == VMW_BALLOON_ERROR_PPN_NOTNEEDED) {
439 __free_page(page);
440 return -EIO;
441 }
442
443 /*
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.
447 */
448 list_add(&page->lru, &b->refused_pages);
449 if (++b->n_refused_pages >= VMW_BALLOON_MAX_REFUSED)
450 return -EIO;
451 }
452 } while (!locked);
453
454 /* track allocated page */
455 list_add(&page->lru, &b->pages);
456
457 /* update balloon size */
458 b->size++;
459
460 return 0;
461}
462
463/*
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
466 * to use, if needed.
467 */
468static int vmballoon_release_page(struct vmballoon *b, struct page *page)
469{
470 if (!vmballoon_send_unlock_page(b, page_to_pfn(page)))
471 return -EIO;
472
473 list_del(&page->lru);
474
475 /* deallocate page */
476 __free_page(page);
477 STATS_INC(b->stats.free);
478
479 /* update balloon size */
480 b->size--;
481
482 return 0;
483}
484
485/*
486 * Release pages that were allocated while attempting to inflate the
487 * balloon but were refused by the host for one reason or another.
488 */
489static void vmballoon_release_refused_pages(struct vmballoon *b)
490{
491 struct page *page, *next;
492
493 list_for_each_entry_safe(page, next, &b->refused_pages, lru) {
494 list_del(&page->lru);
495 __free_page(page);
496 STATS_INC(b->stats.refused_free);
497 }
498
499 b->n_refused_pages = 0;
500}
501
502/*
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
505 * system.
506 */
507static void vmballoon_inflate(struct vmballoon *b)
508{
509 unsigned int goal;
510 unsigned int rate;
511 unsigned int i;
512 unsigned int allocations = 0;
513 int error = 0;
514 bool alloc_can_sleep = false;
515
516 pr_debug("%s - size: %d, target %d\n", __func__, b->size, b->target);
517
518 /*
519 * First try NOSLEEP page allocations to inflate balloon.
520 *
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.
531 */
532
533 goal = b->target - b->size;
534 /*
535 * Start with no sleep allocation rate which may be higher
536 * than sleeping allocation rate.
537 */
538 rate = b->slow_allocation_cycles ?
539 b->rate_alloc : VMW_BALLOON_NOSLEEP_ALLOC_MAX;
540
541 pr_debug("%s - goal: %d, no-sleep rate: %d, sleep rate: %d\n",
542 __func__, goal, rate, b->rate_alloc);
543
544 for (i = 0; i < goal; i++) {
545
546 error = vmballoon_reserve_page(b, alloc_can_sleep);
547 if (error) {
548 if (error != -ENOMEM) {
549 /*
550 * Not a page allocation failure, stop this
551 * cycle. Maybe we'll get new target from
552 * the host soon.
553 */
554 break;
555 }
556
557 if (alloc_can_sleep) {
558 /*
559 * CANSLEEP page allocation failed, so guest
560 * is under severe memory pressure. Quickly
561 * decrease allocation rate.
562 */
563 b->rate_alloc = max(b->rate_alloc / 2,
564 VMW_BALLOON_RATE_ALLOC_MIN);
565 break;
566 }
567
568 /*
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.
575 */
576 b->slow_allocation_cycles = VMW_BALLOON_SLOW_CYCLES;
577
578 if (i >= b->rate_alloc)
579 break;
580
581 alloc_can_sleep = true;
582 /* Lower rate for sleeping allocations. */
583 rate = b->rate_alloc;
584 }
585
586 if (++allocations > VMW_BALLOON_YIELD_THRESHOLD) {
587 cond_resched();
588 allocations = 0;
589 }
590
591 if (i >= rate) {
592 /* We allocated enough pages, let's take a break. */
593 break;
594 }
595 }
596
597 /*
598 * We reached our goal without failures so try increasing
599 * allocation rate.
600 */
601 if (error == 0 && i >= b->rate_alloc) {
602 unsigned int mult = i / b->rate_alloc;
603
604 b->rate_alloc =
605 min(b->rate_alloc + mult * VMW_BALLOON_RATE_ALLOC_INC,
606 VMW_BALLOON_RATE_ALLOC_MAX);
607 }
608
609 vmballoon_release_refused_pages(b);
610}
611
612/*
613 * Decrease the size of the balloon allowing guest to use more memory.
614 */
615static void vmballoon_deflate(struct vmballoon *b)
616{
617 struct page *page, *next;
618 unsigned int i = 0;
619 unsigned int goal;
620 int error;
621
622 pr_debug("%s - size: %d, target %d\n", __func__, b->size, b->target);
623
624 /* limit deallocation rate */
625 goal = min(b->size - b->target, b->rate_free);
626
627 pr_debug("%s - goal: %d, rate: %d\n", __func__, goal, b->rate_free);
628
629 /* free pages to reach target */
630 list_for_each_entry_safe(page, next, &b->pages, lru) {
631 error = vmballoon_release_page(b, page);
632 if (error) {
633 /* quickly decrease rate in case of error */
634 b->rate_free = max(b->rate_free / 2,
635 VMW_BALLOON_RATE_FREE_MIN);
636 return;
637 }
638
639 if (++i >= goal)
640 break;
641 }
642
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);
646}
647
648/*
649 * Balloon work function: reset protocol, if needed, get the new size and
650 * adjust balloon as needed. Repeat in 1 sec.
651 */
652static void vmballoon_work(struct work_struct *work)
653{
654 struct delayed_work *dwork = to_delayed_work(work);
655 struct vmballoon *b = container_of(dwork, struct vmballoon, dwork);
656 unsigned int target;
657
658 STATS_INC(b->stats.timer);
659
660 if (b->reset_required)
661 vmballoon_reset(b);
662
663 if (b->slow_allocation_cycles > 0)
664 b->slow_allocation_cycles--;
665
666 if (vmballoon_send_get_target(b, &target)) {
667 /* update target, adjust size */
668 b->target = target;
669
670 if (b->size < target)
671 vmballoon_inflate(b);
672 else if (b->size > target)
673 vmballoon_deflate(b);
674 }
675
676 /*
677 * We are using a freezable workqueue so that balloon operations are
678 * stopped while the system transitions to/from sleep/hibernation.
679 */
680 queue_delayed_work(system_freezable_wq,
681 dwork, round_jiffies_relative(HZ));
682}
683
684/*
685 * DEBUGFS Interface
686 */
687#ifdef CONFIG_DEBUG_FS
688
689static int vmballoon_debug_show(struct seq_file *f, void *offset)
690{
691 struct vmballoon *b = f->private;
692 struct vmballoon_stats *stats = &b->stats;
693
694 /* format size info */
695 seq_printf(f,
696 "target: %8d pages\n"
697 "current: %8d pages\n",
698 b->target, b->size);
699
700 /* format rate info */
701 seq_printf(f,
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);
707
708 seq_printf(f,
709 "\n"
710 "timer: %8u\n"
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"
718 "primFree: %8u\n"
719 "errAlloc: %8u\n"
720 "errFree: %8u\n",
721 stats->timer,
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,
729 stats->free,
730 stats->refused_alloc, stats->refused_free);
731
732 return 0;
733}
734
735static int vmballoon_debug_open(struct inode *inode, struct file *file)
736{
737 return single_open(file, vmballoon_debug_show, inode->i_private);
738}
739
740static const struct file_operations vmballoon_debug_fops = {
741 .owner = THIS_MODULE,
742 .open = vmballoon_debug_open,
743 .read = seq_read,
744 .llseek = seq_lseek,
745 .release = single_release,
746};
747
748static int __init vmballoon_debugfs_init(struct vmballoon *b)
749{
750 int error;
751
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);
757 return error;
758 }
759
760 return 0;
761}
762
763static void __exit vmballoon_debugfs_exit(struct vmballoon *b)
764{
765 debugfs_remove(b->dbg_entry);
766}
767
768#else
769
770static inline int vmballoon_debugfs_init(struct vmballoon *b)
771{
772 return 0;
773}
774
775static inline void vmballoon_debugfs_exit(struct vmballoon *b)
776{
777}
778
779#endif /* CONFIG_DEBUG_FS */
780
781static int __init vmballoon_init(void)
782{
783 int error;
784
785 /*
786 * Check if we are running on VMware's hypervisor and bail out
787 * if we are not.
788 */
789 if (x86_hyper != &x86_hyper_vmware)
790 return -ENODEV;
791
792 INIT_LIST_HEAD(&balloon.pages);
793 INIT_LIST_HEAD(&balloon.refused_pages);
794
795 /* initialize rates */
796 balloon.rate_alloc = VMW_BALLOON_RATE_ALLOC_MAX;
797 balloon.rate_free = VMW_BALLOON_RATE_FREE_MAX;
798
799 INIT_DELAYED_WORK(&balloon.dwork, vmballoon_work);
800
801 /*
802 * Start balloon.
803 */
804 if (!vmballoon_send_start(&balloon)) {
805 pr_err("failed to send start command to the host\n");
806 return -EIO;
807 }
808
809 if (!vmballoon_send_guest_id(&balloon)) {
810 pr_err("failed to send guest ID to the host\n");
811 return -EIO;
812 }
813
814 error = vmballoon_debugfs_init(&balloon);
815 if (error)
816 return error;
817
818 queue_delayed_work(system_freezable_wq, &balloon.dwork, 0);
819
820 return 0;
821}
822module_init(vmballoon_init);
823
824static void __exit vmballoon_exit(void)
825{
826 cancel_delayed_work_sync(&balloon.dwork);
827
828 vmballoon_debugfs_exit(&balloon);
829
830 /*
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.
834 */
835 vmballoon_send_start(&balloon);
836 vmballoon_pop(&balloon);
837}
838module_exit(vmballoon_exit);
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * VMware Balloon driver.
4 *
5 * Copyright (C) 2000-2018, VMware, Inc. All Rights Reserved.
6 *
7 * This is VMware physical memory management driver for Linux. The driver
8 * acts like a "balloon" that can be inflated to reclaim physical pages by
9 * reserving them in the guest and invalidating them in the monitor,
10 * freeing up the underlying machine pages so they can be allocated to
11 * other guests. The balloon can also be deflated to allow the guest to
12 * use more physical memory. Higher level policies can control the sizes
13 * of balloons in VMs in order to manage physical memory resources.
14 */
15
16//#define DEBUG
17#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
18
19#include <linux/types.h>
20#include <linux/io.h>
21#include <linux/kernel.h>
22#include <linux/mm.h>
23#include <linux/vmalloc.h>
24#include <linux/sched.h>
25#include <linux/module.h>
26#include <linux/workqueue.h>
27#include <linux/debugfs.h>
28#include <linux/seq_file.h>
29#include <linux/rwsem.h>
30#include <linux/slab.h>
31#include <linux/spinlock.h>
32#include <linux/mount.h>
33#include <linux/pseudo_fs.h>
34#include <linux/balloon_compaction.h>
35#include <linux/vmw_vmci_defs.h>
36#include <linux/vmw_vmci_api.h>
37#include <asm/hypervisor.h>
38
39MODULE_AUTHOR("VMware, Inc.");
40MODULE_DESCRIPTION("VMware Memory Control (Balloon) Driver");
41MODULE_ALIAS("dmi:*:svnVMware*:*");
42MODULE_ALIAS("vmware_vmmemctl");
43MODULE_LICENSE("GPL");
44
45static bool __read_mostly vmwballoon_shrinker_enable;
46module_param(vmwballoon_shrinker_enable, bool, 0444);
47MODULE_PARM_DESC(vmwballoon_shrinker_enable,
48 "Enable non-cooperative out-of-memory protection. Disabled by default as it may degrade performance.");
49
50/* Delay in seconds after shrink before inflation. */
51#define VMBALLOON_SHRINK_DELAY (5)
52
53/* Maximum number of refused pages we accumulate during inflation cycle */
54#define VMW_BALLOON_MAX_REFUSED 16
55
56/* Magic number for the balloon mount-point */
57#define BALLOON_VMW_MAGIC 0x0ba11007
58
59/*
60 * Hypervisor communication port definitions.
61 */
62#define VMW_BALLOON_HV_PORT 0x5670
63#define VMW_BALLOON_HV_MAGIC 0x456c6d6f
64#define VMW_BALLOON_GUEST_ID 1 /* Linux */
65
66enum vmwballoon_capabilities {
67 /*
68 * Bit 0 is reserved and not associated to any capability.
69 */
70 VMW_BALLOON_BASIC_CMDS = (1 << 1),
71 VMW_BALLOON_BATCHED_CMDS = (1 << 2),
72 VMW_BALLOON_BATCHED_2M_CMDS = (1 << 3),
73 VMW_BALLOON_SIGNALLED_WAKEUP_CMD = (1 << 4),
74 VMW_BALLOON_64_BIT_TARGET = (1 << 5)
75};
76
77#define VMW_BALLOON_CAPABILITIES_COMMON (VMW_BALLOON_BASIC_CMDS \
78 | VMW_BALLOON_BATCHED_CMDS \
79 | VMW_BALLOON_BATCHED_2M_CMDS \
80 | VMW_BALLOON_SIGNALLED_WAKEUP_CMD)
81
82#define VMW_BALLOON_2M_ORDER (PMD_SHIFT - PAGE_SHIFT)
83
84/*
85 * 64-bit targets are only supported in 64-bit
86 */
87#ifdef CONFIG_64BIT
88#define VMW_BALLOON_CAPABILITIES (VMW_BALLOON_CAPABILITIES_COMMON \
89 | VMW_BALLOON_64_BIT_TARGET)
90#else
91#define VMW_BALLOON_CAPABILITIES VMW_BALLOON_CAPABILITIES_COMMON
92#endif
93
94enum vmballoon_page_size_type {
95 VMW_BALLOON_4K_PAGE,
96 VMW_BALLOON_2M_PAGE,
97 VMW_BALLOON_LAST_SIZE = VMW_BALLOON_2M_PAGE
98};
99
100#define VMW_BALLOON_NUM_PAGE_SIZES (VMW_BALLOON_LAST_SIZE + 1)
101
102static const char * const vmballoon_page_size_names[] = {
103 [VMW_BALLOON_4K_PAGE] = "4k",
104 [VMW_BALLOON_2M_PAGE] = "2M"
105};
106
107enum vmballoon_op {
108 VMW_BALLOON_INFLATE,
109 VMW_BALLOON_DEFLATE
110};
111
112enum vmballoon_op_stat_type {
113 VMW_BALLOON_OP_STAT,
114 VMW_BALLOON_OP_FAIL_STAT
115};
116
117#define VMW_BALLOON_OP_STAT_TYPES (VMW_BALLOON_OP_FAIL_STAT + 1)
118
119/**
120 * enum vmballoon_cmd_type - backdoor commands.
121 *
122 * Availability of the commands is as followed:
123 *
124 * %VMW_BALLOON_CMD_START, %VMW_BALLOON_CMD_GET_TARGET and
125 * %VMW_BALLOON_CMD_GUEST_ID are always available.
126 *
127 * If the host reports %VMW_BALLOON_BASIC_CMDS are supported then
128 * %VMW_BALLOON_CMD_LOCK and %VMW_BALLOON_CMD_UNLOCK commands are available.
129 *
130 * If the host reports %VMW_BALLOON_BATCHED_CMDS are supported then
131 * %VMW_BALLOON_CMD_BATCHED_LOCK and VMW_BALLOON_CMD_BATCHED_UNLOCK commands
132 * are available.
133 *
134 * If the host reports %VMW_BALLOON_BATCHED_2M_CMDS are supported then
135 * %VMW_BALLOON_CMD_BATCHED_2M_LOCK and %VMW_BALLOON_CMD_BATCHED_2M_UNLOCK
136 * are supported.
137 *
138 * If the host reports VMW_BALLOON_SIGNALLED_WAKEUP_CMD is supported then
139 * VMW_BALLOON_CMD_VMCI_DOORBELL_SET command is supported.
140 *
141 * @VMW_BALLOON_CMD_START: Communicating supported version with the hypervisor.
142 * @VMW_BALLOON_CMD_GET_TARGET: Gets the balloon target size.
143 * @VMW_BALLOON_CMD_LOCK: Informs the hypervisor about a ballooned page.
144 * @VMW_BALLOON_CMD_UNLOCK: Informs the hypervisor about a page that is about
145 * to be deflated from the balloon.
146 * @VMW_BALLOON_CMD_GUEST_ID: Informs the hypervisor about the type of OS that
147 * runs in the VM.
148 * @VMW_BALLOON_CMD_BATCHED_LOCK: Inform the hypervisor about a batch of
149 * ballooned pages (up to 512).
150 * @VMW_BALLOON_CMD_BATCHED_UNLOCK: Inform the hypervisor about a batch of
151 * pages that are about to be deflated from the
152 * balloon (up to 512).
153 * @VMW_BALLOON_CMD_BATCHED_2M_LOCK: Similar to @VMW_BALLOON_CMD_BATCHED_LOCK
154 * for 2MB pages.
155 * @VMW_BALLOON_CMD_BATCHED_2M_UNLOCK: Similar to
156 * @VMW_BALLOON_CMD_BATCHED_UNLOCK for 2MB
157 * pages.
158 * @VMW_BALLOON_CMD_VMCI_DOORBELL_SET: A command to set doorbell notification
159 * that would be invoked when the balloon
160 * size changes.
161 * @VMW_BALLOON_CMD_LAST: Value of the last command.
162 */
163enum vmballoon_cmd_type {
164 VMW_BALLOON_CMD_START,
165 VMW_BALLOON_CMD_GET_TARGET,
166 VMW_BALLOON_CMD_LOCK,
167 VMW_BALLOON_CMD_UNLOCK,
168 VMW_BALLOON_CMD_GUEST_ID,
169 /* No command 5 */
170 VMW_BALLOON_CMD_BATCHED_LOCK = 6,
171 VMW_BALLOON_CMD_BATCHED_UNLOCK,
172 VMW_BALLOON_CMD_BATCHED_2M_LOCK,
173 VMW_BALLOON_CMD_BATCHED_2M_UNLOCK,
174 VMW_BALLOON_CMD_VMCI_DOORBELL_SET,
175 VMW_BALLOON_CMD_LAST = VMW_BALLOON_CMD_VMCI_DOORBELL_SET,
176};
177
178#define VMW_BALLOON_CMD_NUM (VMW_BALLOON_CMD_LAST + 1)
179
180enum vmballoon_error_codes {
181 VMW_BALLOON_SUCCESS,
182 VMW_BALLOON_ERROR_CMD_INVALID,
183 VMW_BALLOON_ERROR_PPN_INVALID,
184 VMW_BALLOON_ERROR_PPN_LOCKED,
185 VMW_BALLOON_ERROR_PPN_UNLOCKED,
186 VMW_BALLOON_ERROR_PPN_PINNED,
187 VMW_BALLOON_ERROR_PPN_NOTNEEDED,
188 VMW_BALLOON_ERROR_RESET,
189 VMW_BALLOON_ERROR_BUSY
190};
191
192#define VMW_BALLOON_SUCCESS_WITH_CAPABILITIES (0x03000000)
193
194#define VMW_BALLOON_CMD_WITH_TARGET_MASK \
195 ((1UL << VMW_BALLOON_CMD_GET_TARGET) | \
196 (1UL << VMW_BALLOON_CMD_LOCK) | \
197 (1UL << VMW_BALLOON_CMD_UNLOCK) | \
198 (1UL << VMW_BALLOON_CMD_BATCHED_LOCK) | \
199 (1UL << VMW_BALLOON_CMD_BATCHED_UNLOCK) | \
200 (1UL << VMW_BALLOON_CMD_BATCHED_2M_LOCK) | \
201 (1UL << VMW_BALLOON_CMD_BATCHED_2M_UNLOCK))
202
203static const char * const vmballoon_cmd_names[] = {
204 [VMW_BALLOON_CMD_START] = "start",
205 [VMW_BALLOON_CMD_GET_TARGET] = "target",
206 [VMW_BALLOON_CMD_LOCK] = "lock",
207 [VMW_BALLOON_CMD_UNLOCK] = "unlock",
208 [VMW_BALLOON_CMD_GUEST_ID] = "guestType",
209 [VMW_BALLOON_CMD_BATCHED_LOCK] = "batchLock",
210 [VMW_BALLOON_CMD_BATCHED_UNLOCK] = "batchUnlock",
211 [VMW_BALLOON_CMD_BATCHED_2M_LOCK] = "2m-lock",
212 [VMW_BALLOON_CMD_BATCHED_2M_UNLOCK] = "2m-unlock",
213 [VMW_BALLOON_CMD_VMCI_DOORBELL_SET] = "doorbellSet"
214};
215
216enum vmballoon_stat_page {
217 VMW_BALLOON_PAGE_STAT_ALLOC,
218 VMW_BALLOON_PAGE_STAT_ALLOC_FAIL,
219 VMW_BALLOON_PAGE_STAT_REFUSED_ALLOC,
220 VMW_BALLOON_PAGE_STAT_REFUSED_FREE,
221 VMW_BALLOON_PAGE_STAT_FREE,
222 VMW_BALLOON_PAGE_STAT_LAST = VMW_BALLOON_PAGE_STAT_FREE
223};
224
225#define VMW_BALLOON_PAGE_STAT_NUM (VMW_BALLOON_PAGE_STAT_LAST + 1)
226
227enum vmballoon_stat_general {
228 VMW_BALLOON_STAT_TIMER,
229 VMW_BALLOON_STAT_DOORBELL,
230 VMW_BALLOON_STAT_RESET,
231 VMW_BALLOON_STAT_SHRINK,
232 VMW_BALLOON_STAT_SHRINK_FREE,
233 VMW_BALLOON_STAT_LAST = VMW_BALLOON_STAT_SHRINK_FREE
234};
235
236#define VMW_BALLOON_STAT_NUM (VMW_BALLOON_STAT_LAST + 1)
237
238static DEFINE_STATIC_KEY_TRUE(vmw_balloon_batching);
239static DEFINE_STATIC_KEY_FALSE(balloon_stat_enabled);
240
241struct vmballoon_ctl {
242 struct list_head pages;
243 struct list_head refused_pages;
244 struct list_head prealloc_pages;
245 unsigned int n_refused_pages;
246 unsigned int n_pages;
247 enum vmballoon_page_size_type page_size;
248 enum vmballoon_op op;
249};
250
251/**
252 * struct vmballoon_batch_entry - a batch entry for lock or unlock.
253 *
254 * @status: the status of the operation, which is written by the hypervisor.
255 * @reserved: reserved for future use. Must be set to zero.
256 * @pfn: the physical frame number of the page to be locked or unlocked.
257 */
258struct vmballoon_batch_entry {
259 u64 status : 5;
260 u64 reserved : PAGE_SHIFT - 5;
261 u64 pfn : 52;
262} __packed;
263
264struct vmballoon {
265 /**
266 * @max_page_size: maximum supported page size for ballooning.
267 *
268 * Protected by @conf_sem
269 */
270 enum vmballoon_page_size_type max_page_size;
271
272 /**
273 * @size: balloon actual size in basic page size (frames).
274 *
275 * While we currently do not support size which is bigger than 32-bit,
276 * in preparation for future support, use 64-bits.
277 */
278 atomic64_t size;
279
280 /**
281 * @target: balloon target size in basic page size (frames).
282 *
283 * We do not protect the target under the assumption that setting the
284 * value is always done through a single write. If this assumption ever
285 * breaks, we would have to use X_ONCE for accesses, and suffer the less
286 * optimized code. Although we may read stale target value if multiple
287 * accesses happen at once, the performance impact should be minor.
288 */
289 unsigned long target;
290
291 /**
292 * @reset_required: reset flag
293 *
294 * Setting this flag may introduce races, but the code is expected to
295 * handle them gracefully. In the worst case, another operation will
296 * fail as reset did not take place. Clearing the flag is done while
297 * holding @conf_sem for write.
298 */
299 bool reset_required;
300
301 /**
302 * @capabilities: hypervisor balloon capabilities.
303 *
304 * Protected by @conf_sem.
305 */
306 unsigned long capabilities;
307
308 /**
309 * @batch_page: pointer to communication batch page.
310 *
311 * When batching is used, batch_page points to a page, which holds up to
312 * %VMW_BALLOON_BATCH_MAX_PAGES entries for locking or unlocking.
313 */
314 struct vmballoon_batch_entry *batch_page;
315
316 /**
317 * @batch_max_pages: maximum pages that can be locked/unlocked.
318 *
319 * Indicates the number of pages that the hypervisor can lock or unlock
320 * at once, according to whether batching is enabled. If batching is
321 * disabled, only a single page can be locked/unlock on each operation.
322 *
323 * Protected by @conf_sem.
324 */
325 unsigned int batch_max_pages;
326
327 /**
328 * @page: page to be locked/unlocked by the hypervisor
329 *
330 * @page is only used when batching is disabled and a single page is
331 * reclaimed on each iteration.
332 *
333 * Protected by @comm_lock.
334 */
335 struct page *page;
336
337 /**
338 * @shrink_timeout: timeout until the next inflation.
339 *
340 * After an shrink event, indicates the time in jiffies after which
341 * inflation is allowed again. Can be written concurrently with reads,
342 * so must use READ_ONCE/WRITE_ONCE when accessing.
343 */
344 unsigned long shrink_timeout;
345
346 /* statistics */
347 struct vmballoon_stats *stats;
348
349 /**
350 * @b_dev_info: balloon device information descriptor.
351 */
352 struct balloon_dev_info b_dev_info;
353
354 struct delayed_work dwork;
355
356 /**
357 * @huge_pages - list of the inflated 2MB pages.
358 *
359 * Protected by @b_dev_info.pages_lock .
360 */
361 struct list_head huge_pages;
362
363 /**
364 * @vmci_doorbell.
365 *
366 * Protected by @conf_sem.
367 */
368 struct vmci_handle vmci_doorbell;
369
370 /**
371 * @conf_sem: semaphore to protect the configuration and the statistics.
372 */
373 struct rw_semaphore conf_sem;
374
375 /**
376 * @comm_lock: lock to protect the communication with the host.
377 *
378 * Lock ordering: @conf_sem -> @comm_lock .
379 */
380 spinlock_t comm_lock;
381
382 /**
383 * @shrinker: shrinker interface that is used to avoid over-inflation.
384 */
385 struct shrinker shrinker;
386
387 /**
388 * @shrinker_registered: whether the shrinker was registered.
389 *
390 * The shrinker interface does not handle gracefully the removal of
391 * shrinker that was not registered before. This indication allows to
392 * simplify the unregistration process.
393 */
394 bool shrinker_registered;
395};
396
397static struct vmballoon balloon;
398
399struct vmballoon_stats {
400 /* timer / doorbell operations */
401 atomic64_t general_stat[VMW_BALLOON_STAT_NUM];
402
403 /* allocation statistics for huge and small pages */
404 atomic64_t
405 page_stat[VMW_BALLOON_PAGE_STAT_NUM][VMW_BALLOON_NUM_PAGE_SIZES];
406
407 /* Monitor operations: total operations, and failures */
408 atomic64_t ops[VMW_BALLOON_CMD_NUM][VMW_BALLOON_OP_STAT_TYPES];
409};
410
411static inline bool is_vmballoon_stats_on(void)
412{
413 return IS_ENABLED(CONFIG_DEBUG_FS) &&
414 static_branch_unlikely(&balloon_stat_enabled);
415}
416
417static inline void vmballoon_stats_op_inc(struct vmballoon *b, unsigned int op,
418 enum vmballoon_op_stat_type type)
419{
420 if (is_vmballoon_stats_on())
421 atomic64_inc(&b->stats->ops[op][type]);
422}
423
424static inline void vmballoon_stats_gen_inc(struct vmballoon *b,
425 enum vmballoon_stat_general stat)
426{
427 if (is_vmballoon_stats_on())
428 atomic64_inc(&b->stats->general_stat[stat]);
429}
430
431static inline void vmballoon_stats_gen_add(struct vmballoon *b,
432 enum vmballoon_stat_general stat,
433 unsigned int val)
434{
435 if (is_vmballoon_stats_on())
436 atomic64_add(val, &b->stats->general_stat[stat]);
437}
438
439static inline void vmballoon_stats_page_inc(struct vmballoon *b,
440 enum vmballoon_stat_page stat,
441 enum vmballoon_page_size_type size)
442{
443 if (is_vmballoon_stats_on())
444 atomic64_inc(&b->stats->page_stat[stat][size]);
445}
446
447static inline void vmballoon_stats_page_add(struct vmballoon *b,
448 enum vmballoon_stat_page stat,
449 enum vmballoon_page_size_type size,
450 unsigned int val)
451{
452 if (is_vmballoon_stats_on())
453 atomic64_add(val, &b->stats->page_stat[stat][size]);
454}
455
456static inline unsigned long
457__vmballoon_cmd(struct vmballoon *b, unsigned long cmd, unsigned long arg1,
458 unsigned long arg2, unsigned long *result)
459{
460 unsigned long status, dummy1, dummy2, dummy3, local_result;
461
462 vmballoon_stats_op_inc(b, cmd, VMW_BALLOON_OP_STAT);
463
464 asm volatile ("inl %%dx" :
465 "=a"(status),
466 "=c"(dummy1),
467 "=d"(dummy2),
468 "=b"(local_result),
469 "=S"(dummy3) :
470 "0"(VMW_BALLOON_HV_MAGIC),
471 "1"(cmd),
472 "2"(VMW_BALLOON_HV_PORT),
473 "3"(arg1),
474 "4"(arg2) :
475 "memory");
476
477 /* update the result if needed */
478 if (result)
479 *result = (cmd == VMW_BALLOON_CMD_START) ? dummy1 :
480 local_result;
481
482 /* update target when applicable */
483 if (status == VMW_BALLOON_SUCCESS &&
484 ((1ul << cmd) & VMW_BALLOON_CMD_WITH_TARGET_MASK))
485 WRITE_ONCE(b->target, local_result);
486
487 if (status != VMW_BALLOON_SUCCESS &&
488 status != VMW_BALLOON_SUCCESS_WITH_CAPABILITIES) {
489 vmballoon_stats_op_inc(b, cmd, VMW_BALLOON_OP_FAIL_STAT);
490 pr_debug("%s: %s [0x%lx,0x%lx) failed, returned %ld\n",
491 __func__, vmballoon_cmd_names[cmd], arg1, arg2,
492 status);
493 }
494
495 /* mark reset required accordingly */
496 if (status == VMW_BALLOON_ERROR_RESET)
497 b->reset_required = true;
498
499 return status;
500}
501
502static __always_inline unsigned long
503vmballoon_cmd(struct vmballoon *b, unsigned long cmd, unsigned long arg1,
504 unsigned long arg2)
505{
506 unsigned long dummy;
507
508 return __vmballoon_cmd(b, cmd, arg1, arg2, &dummy);
509}
510
511/*
512 * Send "start" command to the host, communicating supported version
513 * of the protocol.
514 */
515static int vmballoon_send_start(struct vmballoon *b, unsigned long req_caps)
516{
517 unsigned long status, capabilities;
518
519 status = __vmballoon_cmd(b, VMW_BALLOON_CMD_START, req_caps, 0,
520 &capabilities);
521
522 switch (status) {
523 case VMW_BALLOON_SUCCESS_WITH_CAPABILITIES:
524 b->capabilities = capabilities;
525 break;
526 case VMW_BALLOON_SUCCESS:
527 b->capabilities = VMW_BALLOON_BASIC_CMDS;
528 break;
529 default:
530 return -EIO;
531 }
532
533 /*
534 * 2MB pages are only supported with batching. If batching is for some
535 * reason disabled, do not use 2MB pages, since otherwise the legacy
536 * mechanism is used with 2MB pages, causing a failure.
537 */
538 b->max_page_size = VMW_BALLOON_4K_PAGE;
539 if ((b->capabilities & VMW_BALLOON_BATCHED_2M_CMDS) &&
540 (b->capabilities & VMW_BALLOON_BATCHED_CMDS))
541 b->max_page_size = VMW_BALLOON_2M_PAGE;
542
543
544 return 0;
545}
546
547/**
548 * vmballoon_send_guest_id - communicate guest type to the host.
549 *
550 * @b: pointer to the balloon.
551 *
552 * Communicate guest type to the host so that it can adjust ballooning
553 * algorithm to the one most appropriate for the guest. This command
554 * is normally issued after sending "start" command and is part of
555 * standard reset sequence.
556 *
557 * Return: zero on success or appropriate error code.
558 */
559static int vmballoon_send_guest_id(struct vmballoon *b)
560{
561 unsigned long status;
562
563 status = vmballoon_cmd(b, VMW_BALLOON_CMD_GUEST_ID,
564 VMW_BALLOON_GUEST_ID, 0);
565
566 return status == VMW_BALLOON_SUCCESS ? 0 : -EIO;
567}
568
569/**
570 * vmballoon_page_order() - return the order of the page
571 * @page_size: the size of the page.
572 *
573 * Return: the allocation order.
574 */
575static inline
576unsigned int vmballoon_page_order(enum vmballoon_page_size_type page_size)
577{
578 return page_size == VMW_BALLOON_2M_PAGE ? VMW_BALLOON_2M_ORDER : 0;
579}
580
581/**
582 * vmballoon_page_in_frames() - returns the number of frames in a page.
583 * @page_size: the size of the page.
584 *
585 * Return: the number of 4k frames.
586 */
587static inline unsigned int
588vmballoon_page_in_frames(enum vmballoon_page_size_type page_size)
589{
590 return 1 << vmballoon_page_order(page_size);
591}
592
593/**
594 * vmballoon_mark_page_offline() - mark a page as offline
595 * @page: pointer for the page.
596 * @page_size: the size of the page.
597 */
598static void
599vmballoon_mark_page_offline(struct page *page,
600 enum vmballoon_page_size_type page_size)
601{
602 int i;
603
604 for (i = 0; i < vmballoon_page_in_frames(page_size); i++)
605 __SetPageOffline(page + i);
606}
607
608/**
609 * vmballoon_mark_page_online() - mark a page as online
610 * @page: pointer for the page.
611 * @page_size: the size of the page.
612 */
613static void
614vmballoon_mark_page_online(struct page *page,
615 enum vmballoon_page_size_type page_size)
616{
617 int i;
618
619 for (i = 0; i < vmballoon_page_in_frames(page_size); i++)
620 __ClearPageOffline(page + i);
621}
622
623/**
624 * vmballoon_send_get_target() - Retrieve desired balloon size from the host.
625 *
626 * @b: pointer to the balloon.
627 *
628 * Return: zero on success, EINVAL if limit does not fit in 32-bit, as required
629 * by the host-guest protocol and EIO if an error occurred in communicating with
630 * the host.
631 */
632static int vmballoon_send_get_target(struct vmballoon *b)
633{
634 unsigned long status;
635 unsigned long limit;
636
637 limit = totalram_pages();
638
639 /* Ensure limit fits in 32-bits if 64-bit targets are not supported */
640 if (!(b->capabilities & VMW_BALLOON_64_BIT_TARGET) &&
641 limit != (u32)limit)
642 return -EINVAL;
643
644 status = vmballoon_cmd(b, VMW_BALLOON_CMD_GET_TARGET, limit, 0);
645
646 return status == VMW_BALLOON_SUCCESS ? 0 : -EIO;
647}
648
649/**
650 * vmballoon_alloc_page_list - allocates a list of pages.
651 *
652 * @b: pointer to the balloon.
653 * @ctl: pointer for the %struct vmballoon_ctl, which defines the operation.
654 * @req_n_pages: the number of requested pages.
655 *
656 * Tries to allocate @req_n_pages. Add them to the list of balloon pages in
657 * @ctl.pages and updates @ctl.n_pages to reflect the number of pages.
658 *
659 * Return: zero on success or error code otherwise.
660 */
661static int vmballoon_alloc_page_list(struct vmballoon *b,
662 struct vmballoon_ctl *ctl,
663 unsigned int req_n_pages)
664{
665 struct page *page;
666 unsigned int i;
667
668 for (i = 0; i < req_n_pages; i++) {
669 /*
670 * First check if we happen to have pages that were allocated
671 * before. This happens when 2MB page rejected during inflation
672 * by the hypervisor, and then split into 4KB pages.
673 */
674 if (!list_empty(&ctl->prealloc_pages)) {
675 page = list_first_entry(&ctl->prealloc_pages,
676 struct page, lru);
677 list_del(&page->lru);
678 } else {
679 if (ctl->page_size == VMW_BALLOON_2M_PAGE)
680 page = alloc_pages(__GFP_HIGHMEM|__GFP_NOWARN|
681 __GFP_NOMEMALLOC, VMW_BALLOON_2M_ORDER);
682 else
683 page = balloon_page_alloc();
684
685 vmballoon_stats_page_inc(b, VMW_BALLOON_PAGE_STAT_ALLOC,
686 ctl->page_size);
687 }
688
689 if (page) {
690 /* Success. Add the page to the list and continue. */
691 list_add(&page->lru, &ctl->pages);
692 continue;
693 }
694
695 /* Allocation failed. Update statistics and stop. */
696 vmballoon_stats_page_inc(b, VMW_BALLOON_PAGE_STAT_ALLOC_FAIL,
697 ctl->page_size);
698 break;
699 }
700
701 ctl->n_pages = i;
702
703 return req_n_pages == ctl->n_pages ? 0 : -ENOMEM;
704}
705
706/**
707 * vmballoon_handle_one_result - Handle lock/unlock result for a single page.
708 *
709 * @b: pointer for %struct vmballoon.
710 * @page: pointer for the page whose result should be handled.
711 * @page_size: size of the page.
712 * @status: status of the operation as provided by the hypervisor.
713 */
714static int vmballoon_handle_one_result(struct vmballoon *b, struct page *page,
715 enum vmballoon_page_size_type page_size,
716 unsigned long status)
717{
718 /* On success do nothing. The page is already on the balloon list. */
719 if (likely(status == VMW_BALLOON_SUCCESS))
720 return 0;
721
722 pr_debug("%s: failed comm pfn %lx status %lu page_size %s\n", __func__,
723 page_to_pfn(page), status,
724 vmballoon_page_size_names[page_size]);
725
726 /* Error occurred */
727 vmballoon_stats_page_inc(b, VMW_BALLOON_PAGE_STAT_REFUSED_ALLOC,
728 page_size);
729
730 return -EIO;
731}
732
733/**
734 * vmballoon_status_page - returns the status of (un)lock operation
735 *
736 * @b: pointer to the balloon.
737 * @idx: index for the page for which the operation is performed.
738 * @p: pointer to where the page struct is returned.
739 *
740 * Following a lock or unlock operation, returns the status of the operation for
741 * an individual page. Provides the page that the operation was performed on on
742 * the @page argument.
743 *
744 * Returns: The status of a lock or unlock operation for an individual page.
745 */
746static unsigned long vmballoon_status_page(struct vmballoon *b, int idx,
747 struct page **p)
748{
749 if (static_branch_likely(&vmw_balloon_batching)) {
750 /* batching mode */
751 *p = pfn_to_page(b->batch_page[idx].pfn);
752 return b->batch_page[idx].status;
753 }
754
755 /* non-batching mode */
756 *p = b->page;
757
758 /*
759 * If a failure occurs, the indication will be provided in the status
760 * of the entire operation, which is considered before the individual
761 * page status. So for non-batching mode, the indication is always of
762 * success.
763 */
764 return VMW_BALLOON_SUCCESS;
765}
766
767/**
768 * vmballoon_lock_op - notifies the host about inflated/deflated pages.
769 * @b: pointer to the balloon.
770 * @num_pages: number of inflated/deflated pages.
771 * @page_size: size of the page.
772 * @op: the type of operation (lock or unlock).
773 *
774 * Notify the host about page(s) that were ballooned (or removed from the
775 * balloon) so that host can use it without fear that guest will need it (or
776 * stop using them since the VM does). Host may reject some pages, we need to
777 * check the return value and maybe submit a different page. The pages that are
778 * inflated/deflated are pointed by @b->page.
779 *
780 * Return: result as provided by the hypervisor.
781 */
782static unsigned long vmballoon_lock_op(struct vmballoon *b,
783 unsigned int num_pages,
784 enum vmballoon_page_size_type page_size,
785 enum vmballoon_op op)
786{
787 unsigned long cmd, pfn;
788
789 lockdep_assert_held(&b->comm_lock);
790
791 if (static_branch_likely(&vmw_balloon_batching)) {
792 if (op == VMW_BALLOON_INFLATE)
793 cmd = page_size == VMW_BALLOON_2M_PAGE ?
794 VMW_BALLOON_CMD_BATCHED_2M_LOCK :
795 VMW_BALLOON_CMD_BATCHED_LOCK;
796 else
797 cmd = page_size == VMW_BALLOON_2M_PAGE ?
798 VMW_BALLOON_CMD_BATCHED_2M_UNLOCK :
799 VMW_BALLOON_CMD_BATCHED_UNLOCK;
800
801 pfn = PHYS_PFN(virt_to_phys(b->batch_page));
802 } else {
803 cmd = op == VMW_BALLOON_INFLATE ? VMW_BALLOON_CMD_LOCK :
804 VMW_BALLOON_CMD_UNLOCK;
805 pfn = page_to_pfn(b->page);
806
807 /* In non-batching mode, PFNs must fit in 32-bit */
808 if (unlikely(pfn != (u32)pfn))
809 return VMW_BALLOON_ERROR_PPN_INVALID;
810 }
811
812 return vmballoon_cmd(b, cmd, pfn, num_pages);
813}
814
815/**
816 * vmballoon_add_page - adds a page towards lock/unlock operation.
817 *
818 * @b: pointer to the balloon.
819 * @idx: index of the page to be ballooned in this batch.
820 * @p: pointer to the page that is about to be ballooned.
821 *
822 * Adds the page to be ballooned. Must be called while holding @comm_lock.
823 */
824static void vmballoon_add_page(struct vmballoon *b, unsigned int idx,
825 struct page *p)
826{
827 lockdep_assert_held(&b->comm_lock);
828
829 if (static_branch_likely(&vmw_balloon_batching))
830 b->batch_page[idx] = (struct vmballoon_batch_entry)
831 { .pfn = page_to_pfn(p) };
832 else
833 b->page = p;
834}
835
836/**
837 * vmballoon_lock - lock or unlock a batch of pages.
838 *
839 * @b: pointer to the balloon.
840 * @ctl: pointer for the %struct vmballoon_ctl, which defines the operation.
841 *
842 * Notifies the host of about ballooned pages (after inflation or deflation,
843 * according to @ctl). If the host rejects the page put it on the
844 * @ctl refuse list. These refused page are then released when moving to the
845 * next size of pages.
846 *
847 * Note that we neither free any @page here nor put them back on the ballooned
848 * pages list. Instead we queue it for later processing. We do that for several
849 * reasons. First, we do not want to free the page under the lock. Second, it
850 * allows us to unify the handling of lock and unlock. In the inflate case, the
851 * caller will check if there are too many refused pages and release them.
852 * Although it is not identical to the past behavior, it should not affect
853 * performance.
854 */
855static int vmballoon_lock(struct vmballoon *b, struct vmballoon_ctl *ctl)
856{
857 unsigned long batch_status;
858 struct page *page;
859 unsigned int i, num_pages;
860
861 num_pages = ctl->n_pages;
862 if (num_pages == 0)
863 return 0;
864
865 /* communication with the host is done under the communication lock */
866 spin_lock(&b->comm_lock);
867
868 i = 0;
869 list_for_each_entry(page, &ctl->pages, lru)
870 vmballoon_add_page(b, i++, page);
871
872 batch_status = vmballoon_lock_op(b, ctl->n_pages, ctl->page_size,
873 ctl->op);
874
875 /*
876 * Iterate over the pages in the provided list. Since we are changing
877 * @ctl->n_pages we are saving the original value in @num_pages and
878 * use this value to bound the loop.
879 */
880 for (i = 0; i < num_pages; i++) {
881 unsigned long status;
882
883 status = vmballoon_status_page(b, i, &page);
884
885 /*
886 * Failure of the whole batch overrides a single operation
887 * results.
888 */
889 if (batch_status != VMW_BALLOON_SUCCESS)
890 status = batch_status;
891
892 /* Continue if no error happened */
893 if (!vmballoon_handle_one_result(b, page, ctl->page_size,
894 status))
895 continue;
896
897 /*
898 * Error happened. Move the pages to the refused list and update
899 * the pages number.
900 */
901 list_move(&page->lru, &ctl->refused_pages);
902 ctl->n_pages--;
903 ctl->n_refused_pages++;
904 }
905
906 spin_unlock(&b->comm_lock);
907
908 return batch_status == VMW_BALLOON_SUCCESS ? 0 : -EIO;
909}
910
911/**
912 * vmballoon_release_page_list() - Releases a page list
913 *
914 * @page_list: list of pages to release.
915 * @n_pages: pointer to the number of pages.
916 * @page_size: whether the pages in the list are 2MB (or else 4KB).
917 *
918 * Releases the list of pages and zeros the number of pages.
919 */
920static void vmballoon_release_page_list(struct list_head *page_list,
921 int *n_pages,
922 enum vmballoon_page_size_type page_size)
923{
924 struct page *page, *tmp;
925
926 list_for_each_entry_safe(page, tmp, page_list, lru) {
927 list_del(&page->lru);
928 __free_pages(page, vmballoon_page_order(page_size));
929 }
930
931 if (n_pages)
932 *n_pages = 0;
933}
934
935
936/*
937 * Release pages that were allocated while attempting to inflate the
938 * balloon but were refused by the host for one reason or another.
939 */
940static void vmballoon_release_refused_pages(struct vmballoon *b,
941 struct vmballoon_ctl *ctl)
942{
943 vmballoon_stats_page_inc(b, VMW_BALLOON_PAGE_STAT_REFUSED_FREE,
944 ctl->page_size);
945
946 vmballoon_release_page_list(&ctl->refused_pages, &ctl->n_refused_pages,
947 ctl->page_size);
948}
949
950/**
951 * vmballoon_change - retrieve the required balloon change
952 *
953 * @b: pointer for the balloon.
954 *
955 * Return: the required change for the balloon size. A positive number
956 * indicates inflation, a negative number indicates a deflation.
957 */
958static int64_t vmballoon_change(struct vmballoon *b)
959{
960 int64_t size, target;
961
962 size = atomic64_read(&b->size);
963 target = READ_ONCE(b->target);
964
965 /*
966 * We must cast first because of int sizes
967 * Otherwise we might get huge positives instead of negatives
968 */
969
970 if (b->reset_required)
971 return 0;
972
973 /* consider a 2MB slack on deflate, unless the balloon is emptied */
974 if (target < size && target != 0 &&
975 size - target < vmballoon_page_in_frames(VMW_BALLOON_2M_PAGE))
976 return 0;
977
978 /* If an out-of-memory recently occurred, inflation is disallowed. */
979 if (target > size && time_before(jiffies, READ_ONCE(b->shrink_timeout)))
980 return 0;
981
982 return target - size;
983}
984
985/**
986 * vmballoon_enqueue_page_list() - Enqueues list of pages after inflation.
987 *
988 * @b: pointer to balloon.
989 * @pages: list of pages to enqueue.
990 * @n_pages: pointer to number of pages in list. The value is zeroed.
991 * @page_size: whether the pages are 2MB or 4KB pages.
992 *
993 * Enqueues the provides list of pages in the ballooned page list, clears the
994 * list and zeroes the number of pages that was provided.
995 */
996static void vmballoon_enqueue_page_list(struct vmballoon *b,
997 struct list_head *pages,
998 unsigned int *n_pages,
999 enum vmballoon_page_size_type page_size)
1000{
1001 unsigned long flags;
1002 struct page *page;
1003
1004 if (page_size == VMW_BALLOON_4K_PAGE) {
1005 balloon_page_list_enqueue(&b->b_dev_info, pages);
1006 } else {
1007 /*
1008 * Keep the huge pages in a local list which is not available
1009 * for the balloon compaction mechanism.
1010 */
1011 spin_lock_irqsave(&b->b_dev_info.pages_lock, flags);
1012
1013 list_for_each_entry(page, pages, lru) {
1014 vmballoon_mark_page_offline(page, VMW_BALLOON_2M_PAGE);
1015 }
1016
1017 list_splice_init(pages, &b->huge_pages);
1018 __count_vm_events(BALLOON_INFLATE, *n_pages *
1019 vmballoon_page_in_frames(VMW_BALLOON_2M_PAGE));
1020 spin_unlock_irqrestore(&b->b_dev_info.pages_lock, flags);
1021 }
1022
1023 *n_pages = 0;
1024}
1025
1026/**
1027 * vmballoon_dequeue_page_list() - Dequeues page lists for deflation.
1028 *
1029 * @b: pointer to balloon.
1030 * @pages: list of pages to enqueue.
1031 * @n_pages: pointer to number of pages in list. The value is zeroed.
1032 * @page_size: whether the pages are 2MB or 4KB pages.
1033 * @n_req_pages: the number of requested pages.
1034 *
1035 * Dequeues the number of requested pages from the balloon for deflation. The
1036 * number of dequeued pages may be lower, if not enough pages in the requested
1037 * size are available.
1038 */
1039static void vmballoon_dequeue_page_list(struct vmballoon *b,
1040 struct list_head *pages,
1041 unsigned int *n_pages,
1042 enum vmballoon_page_size_type page_size,
1043 unsigned int n_req_pages)
1044{
1045 struct page *page, *tmp;
1046 unsigned int i = 0;
1047 unsigned long flags;
1048
1049 /* In the case of 4k pages, use the compaction infrastructure */
1050 if (page_size == VMW_BALLOON_4K_PAGE) {
1051 *n_pages = balloon_page_list_dequeue(&b->b_dev_info, pages,
1052 n_req_pages);
1053 return;
1054 }
1055
1056 /* 2MB pages */
1057 spin_lock_irqsave(&b->b_dev_info.pages_lock, flags);
1058 list_for_each_entry_safe(page, tmp, &b->huge_pages, lru) {
1059 vmballoon_mark_page_online(page, VMW_BALLOON_2M_PAGE);
1060
1061 list_move(&page->lru, pages);
1062 if (++i == n_req_pages)
1063 break;
1064 }
1065
1066 __count_vm_events(BALLOON_DEFLATE,
1067 i * vmballoon_page_in_frames(VMW_BALLOON_2M_PAGE));
1068 spin_unlock_irqrestore(&b->b_dev_info.pages_lock, flags);
1069 *n_pages = i;
1070}
1071
1072/**
1073 * vmballoon_split_refused_pages() - Split the 2MB refused pages to 4k.
1074 *
1075 * If inflation of 2MB pages was denied by the hypervisor, it is likely to be
1076 * due to one or few 4KB pages. These 2MB pages may keep being allocated and
1077 * then being refused. To prevent this case, this function splits the refused
1078 * pages into 4KB pages and adds them into @prealloc_pages list.
1079 *
1080 * @ctl: pointer for the %struct vmballoon_ctl, which defines the operation.
1081 */
1082static void vmballoon_split_refused_pages(struct vmballoon_ctl *ctl)
1083{
1084 struct page *page, *tmp;
1085 unsigned int i, order;
1086
1087 order = vmballoon_page_order(ctl->page_size);
1088
1089 list_for_each_entry_safe(page, tmp, &ctl->refused_pages, lru) {
1090 list_del(&page->lru);
1091 split_page(page, order);
1092 for (i = 0; i < (1 << order); i++)
1093 list_add(&page[i].lru, &ctl->prealloc_pages);
1094 }
1095 ctl->n_refused_pages = 0;
1096}
1097
1098/**
1099 * vmballoon_inflate() - Inflate the balloon towards its target size.
1100 *
1101 * @b: pointer to the balloon.
1102 */
1103static void vmballoon_inflate(struct vmballoon *b)
1104{
1105 int64_t to_inflate_frames;
1106 struct vmballoon_ctl ctl = {
1107 .pages = LIST_HEAD_INIT(ctl.pages),
1108 .refused_pages = LIST_HEAD_INIT(ctl.refused_pages),
1109 .prealloc_pages = LIST_HEAD_INIT(ctl.prealloc_pages),
1110 .page_size = b->max_page_size,
1111 .op = VMW_BALLOON_INFLATE
1112 };
1113
1114 while ((to_inflate_frames = vmballoon_change(b)) > 0) {
1115 unsigned int to_inflate_pages, page_in_frames;
1116 int alloc_error, lock_error = 0;
1117
1118 VM_BUG_ON(!list_empty(&ctl.pages));
1119 VM_BUG_ON(ctl.n_pages != 0);
1120
1121 page_in_frames = vmballoon_page_in_frames(ctl.page_size);
1122
1123 to_inflate_pages = min_t(unsigned long, b->batch_max_pages,
1124 DIV_ROUND_UP_ULL(to_inflate_frames,
1125 page_in_frames));
1126
1127 /* Start by allocating */
1128 alloc_error = vmballoon_alloc_page_list(b, &ctl,
1129 to_inflate_pages);
1130
1131 /* Actually lock the pages by telling the hypervisor */
1132 lock_error = vmballoon_lock(b, &ctl);
1133
1134 /*
1135 * If an error indicates that something serious went wrong,
1136 * stop the inflation.
1137 */
1138 if (lock_error)
1139 break;
1140
1141 /* Update the balloon size */
1142 atomic64_add(ctl.n_pages * page_in_frames, &b->size);
1143
1144 vmballoon_enqueue_page_list(b, &ctl.pages, &ctl.n_pages,
1145 ctl.page_size);
1146
1147 /*
1148 * If allocation failed or the number of refused pages exceeds
1149 * the maximum allowed, move to the next page size.
1150 */
1151 if (alloc_error ||
1152 ctl.n_refused_pages >= VMW_BALLOON_MAX_REFUSED) {
1153 if (ctl.page_size == VMW_BALLOON_4K_PAGE)
1154 break;
1155
1156 /*
1157 * Split the refused pages to 4k. This will also empty
1158 * the refused pages list.
1159 */
1160 vmballoon_split_refused_pages(&ctl);
1161 ctl.page_size--;
1162 }
1163
1164 cond_resched();
1165 }
1166
1167 /*
1168 * Release pages that were allocated while attempting to inflate the
1169 * balloon but were refused by the host for one reason or another,
1170 * and update the statistics.
1171 */
1172 if (ctl.n_refused_pages != 0)
1173 vmballoon_release_refused_pages(b, &ctl);
1174
1175 vmballoon_release_page_list(&ctl.prealloc_pages, NULL, ctl.page_size);
1176}
1177
1178/**
1179 * vmballoon_deflate() - Decrease the size of the balloon.
1180 *
1181 * @b: pointer to the balloon
1182 * @n_frames: the number of frames to deflate. If zero, automatically
1183 * calculated according to the target size.
1184 * @coordinated: whether to coordinate with the host
1185 *
1186 * Decrease the size of the balloon allowing guest to use more memory.
1187 *
1188 * Return: The number of deflated frames (i.e., basic page size units)
1189 */
1190static unsigned long vmballoon_deflate(struct vmballoon *b, uint64_t n_frames,
1191 bool coordinated)
1192{
1193 unsigned long deflated_frames = 0;
1194 unsigned long tried_frames = 0;
1195 struct vmballoon_ctl ctl = {
1196 .pages = LIST_HEAD_INIT(ctl.pages),
1197 .refused_pages = LIST_HEAD_INIT(ctl.refused_pages),
1198 .page_size = VMW_BALLOON_4K_PAGE,
1199 .op = VMW_BALLOON_DEFLATE
1200 };
1201
1202 /* free pages to reach target */
1203 while (true) {
1204 unsigned int to_deflate_pages, n_unlocked_frames;
1205 unsigned int page_in_frames;
1206 int64_t to_deflate_frames;
1207 bool deflated_all;
1208
1209 page_in_frames = vmballoon_page_in_frames(ctl.page_size);
1210
1211 VM_BUG_ON(!list_empty(&ctl.pages));
1212 VM_BUG_ON(ctl.n_pages);
1213 VM_BUG_ON(!list_empty(&ctl.refused_pages));
1214 VM_BUG_ON(ctl.n_refused_pages);
1215
1216 /*
1217 * If we were requested a specific number of frames, we try to
1218 * deflate this number of frames. Otherwise, deflation is
1219 * performed according to the target and balloon size.
1220 */
1221 to_deflate_frames = n_frames ? n_frames - tried_frames :
1222 -vmballoon_change(b);
1223
1224 /* break if no work to do */
1225 if (to_deflate_frames <= 0)
1226 break;
1227
1228 /*
1229 * Calculate the number of frames based on current page size,
1230 * but limit the deflated frames to a single chunk
1231 */
1232 to_deflate_pages = min_t(unsigned long, b->batch_max_pages,
1233 DIV_ROUND_UP_ULL(to_deflate_frames,
1234 page_in_frames));
1235
1236 /* First take the pages from the balloon pages. */
1237 vmballoon_dequeue_page_list(b, &ctl.pages, &ctl.n_pages,
1238 ctl.page_size, to_deflate_pages);
1239
1240 /*
1241 * Before pages are moving to the refused list, count their
1242 * frames as frames that we tried to deflate.
1243 */
1244 tried_frames += ctl.n_pages * page_in_frames;
1245
1246 /*
1247 * Unlock the pages by communicating with the hypervisor if the
1248 * communication is coordinated (i.e., not pop). We ignore the
1249 * return code. Instead we check if all the pages we manage to
1250 * unlock all the pages. If we failed, we will move to the next
1251 * page size, and would eventually try again later.
1252 */
1253 if (coordinated)
1254 vmballoon_lock(b, &ctl);
1255
1256 /*
1257 * Check if we deflated enough. We will move to the next page
1258 * size if we did not manage to do so. This calculation takes
1259 * place now, as once the pages are released, the number of
1260 * pages is zeroed.
1261 */
1262 deflated_all = (ctl.n_pages == to_deflate_pages);
1263
1264 /* Update local and global counters */
1265 n_unlocked_frames = ctl.n_pages * page_in_frames;
1266 atomic64_sub(n_unlocked_frames, &b->size);
1267 deflated_frames += n_unlocked_frames;
1268
1269 vmballoon_stats_page_add(b, VMW_BALLOON_PAGE_STAT_FREE,
1270 ctl.page_size, ctl.n_pages);
1271
1272 /* free the ballooned pages */
1273 vmballoon_release_page_list(&ctl.pages, &ctl.n_pages,
1274 ctl.page_size);
1275
1276 /* Return the refused pages to the ballooned list. */
1277 vmballoon_enqueue_page_list(b, &ctl.refused_pages,
1278 &ctl.n_refused_pages,
1279 ctl.page_size);
1280
1281 /* If we failed to unlock all the pages, move to next size. */
1282 if (!deflated_all) {
1283 if (ctl.page_size == b->max_page_size)
1284 break;
1285 ctl.page_size++;
1286 }
1287
1288 cond_resched();
1289 }
1290
1291 return deflated_frames;
1292}
1293
1294/**
1295 * vmballoon_deinit_batching - disables batching mode.
1296 *
1297 * @b: pointer to &struct vmballoon.
1298 *
1299 * Disables batching, by deallocating the page for communication with the
1300 * hypervisor and disabling the static key to indicate that batching is off.
1301 */
1302static void vmballoon_deinit_batching(struct vmballoon *b)
1303{
1304 free_page((unsigned long)b->batch_page);
1305 b->batch_page = NULL;
1306 static_branch_disable(&vmw_balloon_batching);
1307 b->batch_max_pages = 1;
1308}
1309
1310/**
1311 * vmballoon_init_batching - enable batching mode.
1312 *
1313 * @b: pointer to &struct vmballoon.
1314 *
1315 * Enables batching, by allocating a page for communication with the hypervisor
1316 * and enabling the static_key to use batching.
1317 *
1318 * Return: zero on success or an appropriate error-code.
1319 */
1320static int vmballoon_init_batching(struct vmballoon *b)
1321{
1322 struct page *page;
1323
1324 page = alloc_page(GFP_KERNEL | __GFP_ZERO);
1325 if (!page)
1326 return -ENOMEM;
1327
1328 b->batch_page = page_address(page);
1329 b->batch_max_pages = PAGE_SIZE / sizeof(struct vmballoon_batch_entry);
1330
1331 static_branch_enable(&vmw_balloon_batching);
1332
1333 return 0;
1334}
1335
1336/*
1337 * Receive notification and resize balloon
1338 */
1339static void vmballoon_doorbell(void *client_data)
1340{
1341 struct vmballoon *b = client_data;
1342
1343 vmballoon_stats_gen_inc(b, VMW_BALLOON_STAT_DOORBELL);
1344
1345 mod_delayed_work(system_freezable_wq, &b->dwork, 0);
1346}
1347
1348/*
1349 * Clean up vmci doorbell
1350 */
1351static void vmballoon_vmci_cleanup(struct vmballoon *b)
1352{
1353 vmballoon_cmd(b, VMW_BALLOON_CMD_VMCI_DOORBELL_SET,
1354 VMCI_INVALID_ID, VMCI_INVALID_ID);
1355
1356 if (!vmci_handle_is_invalid(b->vmci_doorbell)) {
1357 vmci_doorbell_destroy(b->vmci_doorbell);
1358 b->vmci_doorbell = VMCI_INVALID_HANDLE;
1359 }
1360}
1361
1362/**
1363 * vmballoon_vmci_init - Initialize vmci doorbell.
1364 *
1365 * @b: pointer to the balloon.
1366 *
1367 * Return: zero on success or when wakeup command not supported. Error-code
1368 * otherwise.
1369 *
1370 * Initialize vmci doorbell, to get notified as soon as balloon changes.
1371 */
1372static int vmballoon_vmci_init(struct vmballoon *b)
1373{
1374 unsigned long error;
1375
1376 if ((b->capabilities & VMW_BALLOON_SIGNALLED_WAKEUP_CMD) == 0)
1377 return 0;
1378
1379 error = vmci_doorbell_create(&b->vmci_doorbell, VMCI_FLAG_DELAYED_CB,
1380 VMCI_PRIVILEGE_FLAG_RESTRICTED,
1381 vmballoon_doorbell, b);
1382
1383 if (error != VMCI_SUCCESS)
1384 goto fail;
1385
1386 error = __vmballoon_cmd(b, VMW_BALLOON_CMD_VMCI_DOORBELL_SET,
1387 b->vmci_doorbell.context,
1388 b->vmci_doorbell.resource, NULL);
1389
1390 if (error != VMW_BALLOON_SUCCESS)
1391 goto fail;
1392
1393 return 0;
1394fail:
1395 vmballoon_vmci_cleanup(b);
1396 return -EIO;
1397}
1398
1399/**
1400 * vmballoon_pop - Quickly release all pages allocate for the balloon.
1401 *
1402 * @b: pointer to the balloon.
1403 *
1404 * This function is called when host decides to "reset" balloon for one reason
1405 * or another. Unlike normal "deflate" we do not (shall not) notify host of the
1406 * pages being released.
1407 */
1408static void vmballoon_pop(struct vmballoon *b)
1409{
1410 unsigned long size;
1411
1412 while ((size = atomic64_read(&b->size)))
1413 vmballoon_deflate(b, size, false);
1414}
1415
1416/*
1417 * Perform standard reset sequence by popping the balloon (in case it
1418 * is not empty) and then restarting protocol. This operation normally
1419 * happens when host responds with VMW_BALLOON_ERROR_RESET to a command.
1420 */
1421static void vmballoon_reset(struct vmballoon *b)
1422{
1423 int error;
1424
1425 down_write(&b->conf_sem);
1426
1427 vmballoon_vmci_cleanup(b);
1428
1429 /* free all pages, skipping monitor unlock */
1430 vmballoon_pop(b);
1431
1432 if (vmballoon_send_start(b, VMW_BALLOON_CAPABILITIES))
1433 goto unlock;
1434
1435 if ((b->capabilities & VMW_BALLOON_BATCHED_CMDS) != 0) {
1436 if (vmballoon_init_batching(b)) {
1437 /*
1438 * We failed to initialize batching, inform the monitor
1439 * about it by sending a null capability.
1440 *
1441 * The guest will retry in one second.
1442 */
1443 vmballoon_send_start(b, 0);
1444 goto unlock;
1445 }
1446 } else if ((b->capabilities & VMW_BALLOON_BASIC_CMDS) != 0) {
1447 vmballoon_deinit_batching(b);
1448 }
1449
1450 vmballoon_stats_gen_inc(b, VMW_BALLOON_STAT_RESET);
1451 b->reset_required = false;
1452
1453 error = vmballoon_vmci_init(b);
1454 if (error)
1455 pr_err("failed to initialize vmci doorbell\n");
1456
1457 if (vmballoon_send_guest_id(b))
1458 pr_err("failed to send guest ID to the host\n");
1459
1460unlock:
1461 up_write(&b->conf_sem);
1462}
1463
1464/**
1465 * vmballoon_work - periodic balloon worker for reset, inflation and deflation.
1466 *
1467 * @work: pointer to the &work_struct which is provided by the workqueue.
1468 *
1469 * Resets the protocol if needed, gets the new size and adjusts balloon as
1470 * needed. Repeat in 1 sec.
1471 */
1472static void vmballoon_work(struct work_struct *work)
1473{
1474 struct delayed_work *dwork = to_delayed_work(work);
1475 struct vmballoon *b = container_of(dwork, struct vmballoon, dwork);
1476 int64_t change = 0;
1477
1478 if (b->reset_required)
1479 vmballoon_reset(b);
1480
1481 down_read(&b->conf_sem);
1482
1483 /*
1484 * Update the stats while holding the semaphore to ensure that
1485 * @stats_enabled is consistent with whether the stats are actually
1486 * enabled
1487 */
1488 vmballoon_stats_gen_inc(b, VMW_BALLOON_STAT_TIMER);
1489
1490 if (!vmballoon_send_get_target(b))
1491 change = vmballoon_change(b);
1492
1493 if (change != 0) {
1494 pr_debug("%s - size: %llu, target %lu\n", __func__,
1495 atomic64_read(&b->size), READ_ONCE(b->target));
1496
1497 if (change > 0)
1498 vmballoon_inflate(b);
1499 else /* (change < 0) */
1500 vmballoon_deflate(b, 0, true);
1501 }
1502
1503 up_read(&b->conf_sem);
1504
1505 /*
1506 * We are using a freezable workqueue so that balloon operations are
1507 * stopped while the system transitions to/from sleep/hibernation.
1508 */
1509 queue_delayed_work(system_freezable_wq,
1510 dwork, round_jiffies_relative(HZ));
1511
1512}
1513
1514/**
1515 * vmballoon_shrinker_scan() - deflate the balloon due to memory pressure.
1516 * @shrinker: pointer to the balloon shrinker.
1517 * @sc: page reclaim information.
1518 *
1519 * Returns: number of pages that were freed during deflation.
1520 */
1521static unsigned long vmballoon_shrinker_scan(struct shrinker *shrinker,
1522 struct shrink_control *sc)
1523{
1524 struct vmballoon *b = &balloon;
1525 unsigned long deflated_frames;
1526
1527 pr_debug("%s - size: %llu", __func__, atomic64_read(&b->size));
1528
1529 vmballoon_stats_gen_inc(b, VMW_BALLOON_STAT_SHRINK);
1530
1531 /*
1532 * If the lock is also contended for read, we cannot easily reclaim and
1533 * we bail out.
1534 */
1535 if (!down_read_trylock(&b->conf_sem))
1536 return 0;
1537
1538 deflated_frames = vmballoon_deflate(b, sc->nr_to_scan, true);
1539
1540 vmballoon_stats_gen_add(b, VMW_BALLOON_STAT_SHRINK_FREE,
1541 deflated_frames);
1542
1543 /*
1544 * Delay future inflation for some time to mitigate the situations in
1545 * which balloon continuously grows and shrinks. Use WRITE_ONCE() since
1546 * the access is asynchronous.
1547 */
1548 WRITE_ONCE(b->shrink_timeout, jiffies + HZ * VMBALLOON_SHRINK_DELAY);
1549
1550 up_read(&b->conf_sem);
1551
1552 return deflated_frames;
1553}
1554
1555/**
1556 * vmballoon_shrinker_count() - return the number of ballooned pages.
1557 * @shrinker: pointer to the balloon shrinker.
1558 * @sc: page reclaim information.
1559 *
1560 * Returns: number of 4k pages that are allocated for the balloon and can
1561 * therefore be reclaimed under pressure.
1562 */
1563static unsigned long vmballoon_shrinker_count(struct shrinker *shrinker,
1564 struct shrink_control *sc)
1565{
1566 struct vmballoon *b = &balloon;
1567
1568 return atomic64_read(&b->size);
1569}
1570
1571static void vmballoon_unregister_shrinker(struct vmballoon *b)
1572{
1573 if (b->shrinker_registered)
1574 unregister_shrinker(&b->shrinker);
1575 b->shrinker_registered = false;
1576}
1577
1578static int vmballoon_register_shrinker(struct vmballoon *b)
1579{
1580 int r;
1581
1582 /* Do nothing if the shrinker is not enabled */
1583 if (!vmwballoon_shrinker_enable)
1584 return 0;
1585
1586 b->shrinker.scan_objects = vmballoon_shrinker_scan;
1587 b->shrinker.count_objects = vmballoon_shrinker_count;
1588 b->shrinker.seeks = DEFAULT_SEEKS;
1589
1590 r = register_shrinker(&b->shrinker);
1591
1592 if (r == 0)
1593 b->shrinker_registered = true;
1594
1595 return r;
1596}
1597
1598/*
1599 * DEBUGFS Interface
1600 */
1601#ifdef CONFIG_DEBUG_FS
1602
1603static const char * const vmballoon_stat_page_names[] = {
1604 [VMW_BALLOON_PAGE_STAT_ALLOC] = "alloc",
1605 [VMW_BALLOON_PAGE_STAT_ALLOC_FAIL] = "allocFail",
1606 [VMW_BALLOON_PAGE_STAT_REFUSED_ALLOC] = "errAlloc",
1607 [VMW_BALLOON_PAGE_STAT_REFUSED_FREE] = "errFree",
1608 [VMW_BALLOON_PAGE_STAT_FREE] = "free"
1609};
1610
1611static const char * const vmballoon_stat_names[] = {
1612 [VMW_BALLOON_STAT_TIMER] = "timer",
1613 [VMW_BALLOON_STAT_DOORBELL] = "doorbell",
1614 [VMW_BALLOON_STAT_RESET] = "reset",
1615 [VMW_BALLOON_STAT_SHRINK] = "shrink",
1616 [VMW_BALLOON_STAT_SHRINK_FREE] = "shrinkFree"
1617};
1618
1619static int vmballoon_enable_stats(struct vmballoon *b)
1620{
1621 int r = 0;
1622
1623 down_write(&b->conf_sem);
1624
1625 /* did we somehow race with another reader which enabled stats? */
1626 if (b->stats)
1627 goto out;
1628
1629 b->stats = kzalloc(sizeof(*b->stats), GFP_KERNEL);
1630
1631 if (!b->stats) {
1632 /* allocation failed */
1633 r = -ENOMEM;
1634 goto out;
1635 }
1636 static_key_enable(&balloon_stat_enabled.key);
1637out:
1638 up_write(&b->conf_sem);
1639 return r;
1640}
1641
1642/**
1643 * vmballoon_debug_show - shows statistics of balloon operations.
1644 * @f: pointer to the &struct seq_file.
1645 * @offset: ignored.
1646 *
1647 * Provides the statistics that can be accessed in vmmemctl in the debugfs.
1648 * To avoid the overhead - mainly that of memory - of collecting the statistics,
1649 * we only collect statistics after the first time the counters are read.
1650 *
1651 * Return: zero on success or an error code.
1652 */
1653static int vmballoon_debug_show(struct seq_file *f, void *offset)
1654{
1655 struct vmballoon *b = f->private;
1656 int i, j;
1657
1658 /* enables stats if they are disabled */
1659 if (!b->stats) {
1660 int r = vmballoon_enable_stats(b);
1661
1662 if (r)
1663 return r;
1664 }
1665
1666 /* format capabilities info */
1667 seq_printf(f, "%-22s: %#16x\n", "balloon capabilities",
1668 VMW_BALLOON_CAPABILITIES);
1669 seq_printf(f, "%-22s: %#16lx\n", "used capabilities", b->capabilities);
1670 seq_printf(f, "%-22s: %16s\n", "is resetting",
1671 b->reset_required ? "y" : "n");
1672
1673 /* format size info */
1674 seq_printf(f, "%-22s: %16lu\n", "target", READ_ONCE(b->target));
1675 seq_printf(f, "%-22s: %16llu\n", "current", atomic64_read(&b->size));
1676
1677 for (i = 0; i < VMW_BALLOON_CMD_NUM; i++) {
1678 if (vmballoon_cmd_names[i] == NULL)
1679 continue;
1680
1681 seq_printf(f, "%-22s: %16llu (%llu failed)\n",
1682 vmballoon_cmd_names[i],
1683 atomic64_read(&b->stats->ops[i][VMW_BALLOON_OP_STAT]),
1684 atomic64_read(&b->stats->ops[i][VMW_BALLOON_OP_FAIL_STAT]));
1685 }
1686
1687 for (i = 0; i < VMW_BALLOON_STAT_NUM; i++)
1688 seq_printf(f, "%-22s: %16llu\n",
1689 vmballoon_stat_names[i],
1690 atomic64_read(&b->stats->general_stat[i]));
1691
1692 for (i = 0; i < VMW_BALLOON_PAGE_STAT_NUM; i++) {
1693 for (j = 0; j < VMW_BALLOON_NUM_PAGE_SIZES; j++)
1694 seq_printf(f, "%-18s(%s): %16llu\n",
1695 vmballoon_stat_page_names[i],
1696 vmballoon_page_size_names[j],
1697 atomic64_read(&b->stats->page_stat[i][j]));
1698 }
1699
1700 return 0;
1701}
1702
1703DEFINE_SHOW_ATTRIBUTE(vmballoon_debug);
1704
1705static void __init vmballoon_debugfs_init(struct vmballoon *b)
1706{
1707 debugfs_create_file("vmmemctl", S_IRUGO, NULL, b,
1708 &vmballoon_debug_fops);
1709}
1710
1711static void __exit vmballoon_debugfs_exit(struct vmballoon *b)
1712{
1713 static_key_disable(&balloon_stat_enabled.key);
1714 debugfs_remove(debugfs_lookup("vmmemctl", NULL));
1715 kfree(b->stats);
1716 b->stats = NULL;
1717}
1718
1719#else
1720
1721static inline void vmballoon_debugfs_init(struct vmballoon *b)
1722{
1723}
1724
1725static inline void vmballoon_debugfs_exit(struct vmballoon *b)
1726{
1727}
1728
1729#endif /* CONFIG_DEBUG_FS */
1730
1731
1732#ifdef CONFIG_BALLOON_COMPACTION
1733
1734static int vmballoon_init_fs_context(struct fs_context *fc)
1735{
1736 return init_pseudo(fc, BALLOON_VMW_MAGIC) ? 0 : -ENOMEM;
1737}
1738
1739static struct file_system_type vmballoon_fs = {
1740 .name = "balloon-vmware",
1741 .init_fs_context = vmballoon_init_fs_context,
1742 .kill_sb = kill_anon_super,
1743};
1744
1745static struct vfsmount *vmballoon_mnt;
1746
1747/**
1748 * vmballoon_migratepage() - migrates a balloon page.
1749 * @b_dev_info: balloon device information descriptor.
1750 * @newpage: the page to which @page should be migrated.
1751 * @page: a ballooned page that should be migrated.
1752 * @mode: migration mode, ignored.
1753 *
1754 * This function is really open-coded, but that is according to the interface
1755 * that balloon_compaction provides.
1756 *
1757 * Return: zero on success, -EAGAIN when migration cannot be performed
1758 * momentarily, and -EBUSY if migration failed and should be retried
1759 * with that specific page.
1760 */
1761static int vmballoon_migratepage(struct balloon_dev_info *b_dev_info,
1762 struct page *newpage, struct page *page,
1763 enum migrate_mode mode)
1764{
1765 unsigned long status, flags;
1766 struct vmballoon *b;
1767 int ret;
1768
1769 b = container_of(b_dev_info, struct vmballoon, b_dev_info);
1770
1771 /*
1772 * If the semaphore is taken, there is ongoing configuration change
1773 * (i.e., balloon reset), so try again.
1774 */
1775 if (!down_read_trylock(&b->conf_sem))
1776 return -EAGAIN;
1777
1778 spin_lock(&b->comm_lock);
1779 /*
1780 * We must start by deflating and not inflating, as otherwise the
1781 * hypervisor may tell us that it has enough memory and the new page is
1782 * not needed. Since the old page is isolated, we cannot use the list
1783 * interface to unlock it, as the LRU field is used for isolation.
1784 * Instead, we use the native interface directly.
1785 */
1786 vmballoon_add_page(b, 0, page);
1787 status = vmballoon_lock_op(b, 1, VMW_BALLOON_4K_PAGE,
1788 VMW_BALLOON_DEFLATE);
1789
1790 if (status == VMW_BALLOON_SUCCESS)
1791 status = vmballoon_status_page(b, 0, &page);
1792
1793 /*
1794 * If a failure happened, let the migration mechanism know that it
1795 * should not retry.
1796 */
1797 if (status != VMW_BALLOON_SUCCESS) {
1798 spin_unlock(&b->comm_lock);
1799 ret = -EBUSY;
1800 goto out_unlock;
1801 }
1802
1803 /*
1804 * The page is isolated, so it is safe to delete it without holding
1805 * @pages_lock . We keep holding @comm_lock since we will need it in a
1806 * second.
1807 */
1808 balloon_page_delete(page);
1809
1810 put_page(page);
1811
1812 /* Inflate */
1813 vmballoon_add_page(b, 0, newpage);
1814 status = vmballoon_lock_op(b, 1, VMW_BALLOON_4K_PAGE,
1815 VMW_BALLOON_INFLATE);
1816
1817 if (status == VMW_BALLOON_SUCCESS)
1818 status = vmballoon_status_page(b, 0, &newpage);
1819
1820 spin_unlock(&b->comm_lock);
1821
1822 if (status != VMW_BALLOON_SUCCESS) {
1823 /*
1824 * A failure happened. While we can deflate the page we just
1825 * inflated, this deflation can also encounter an error. Instead
1826 * we will decrease the size of the balloon to reflect the
1827 * change and report failure.
1828 */
1829 atomic64_dec(&b->size);
1830 ret = -EBUSY;
1831 } else {
1832 /*
1833 * Success. Take a reference for the page, and we will add it to
1834 * the list after acquiring the lock.
1835 */
1836 get_page(newpage);
1837 ret = MIGRATEPAGE_SUCCESS;
1838 }
1839
1840 /* Update the balloon list under the @pages_lock */
1841 spin_lock_irqsave(&b->b_dev_info.pages_lock, flags);
1842
1843 /*
1844 * On inflation success, we already took a reference for the @newpage.
1845 * If we succeed just insert it to the list and update the statistics
1846 * under the lock.
1847 */
1848 if (ret == MIGRATEPAGE_SUCCESS) {
1849 balloon_page_insert(&b->b_dev_info, newpage);
1850 __count_vm_event(BALLOON_MIGRATE);
1851 }
1852
1853 /*
1854 * We deflated successfully, so regardless to the inflation success, we
1855 * need to reduce the number of isolated_pages.
1856 */
1857 b->b_dev_info.isolated_pages--;
1858 spin_unlock_irqrestore(&b->b_dev_info.pages_lock, flags);
1859
1860out_unlock:
1861 up_read(&b->conf_sem);
1862 return ret;
1863}
1864
1865/**
1866 * vmballoon_compaction_deinit() - removes compaction related data.
1867 *
1868 * @b: pointer to the balloon.
1869 */
1870static void vmballoon_compaction_deinit(struct vmballoon *b)
1871{
1872 if (!IS_ERR(b->b_dev_info.inode))
1873 iput(b->b_dev_info.inode);
1874
1875 b->b_dev_info.inode = NULL;
1876 kern_unmount(vmballoon_mnt);
1877 vmballoon_mnt = NULL;
1878}
1879
1880/**
1881 * vmballoon_compaction_init() - initialized compaction for the balloon.
1882 *
1883 * @b: pointer to the balloon.
1884 *
1885 * If during the initialization a failure occurred, this function does not
1886 * perform cleanup. The caller must call vmballoon_compaction_deinit() in this
1887 * case.
1888 *
1889 * Return: zero on success or error code on failure.
1890 */
1891static __init int vmballoon_compaction_init(struct vmballoon *b)
1892{
1893 vmballoon_mnt = kern_mount(&vmballoon_fs);
1894 if (IS_ERR(vmballoon_mnt))
1895 return PTR_ERR(vmballoon_mnt);
1896
1897 b->b_dev_info.migratepage = vmballoon_migratepage;
1898 b->b_dev_info.inode = alloc_anon_inode(vmballoon_mnt->mnt_sb);
1899
1900 if (IS_ERR(b->b_dev_info.inode))
1901 return PTR_ERR(b->b_dev_info.inode);
1902
1903 b->b_dev_info.inode->i_mapping->a_ops = &balloon_aops;
1904 return 0;
1905}
1906
1907#else /* CONFIG_BALLOON_COMPACTION */
1908
1909static void vmballoon_compaction_deinit(struct vmballoon *b)
1910{
1911}
1912
1913static int vmballoon_compaction_init(struct vmballoon *b)
1914{
1915 return 0;
1916}
1917
1918#endif /* CONFIG_BALLOON_COMPACTION */
1919
1920static int __init vmballoon_init(void)
1921{
1922 int error;
1923
1924 /*
1925 * Check if we are running on VMware's hypervisor and bail out
1926 * if we are not.
1927 */
1928 if (x86_hyper_type != X86_HYPER_VMWARE)
1929 return -ENODEV;
1930
1931 INIT_DELAYED_WORK(&balloon.dwork, vmballoon_work);
1932
1933 error = vmballoon_register_shrinker(&balloon);
1934 if (error)
1935 goto fail;
1936
1937 /*
1938 * Initialization of compaction must be done after the call to
1939 * balloon_devinfo_init() .
1940 */
1941 balloon_devinfo_init(&balloon.b_dev_info);
1942 error = vmballoon_compaction_init(&balloon);
1943 if (error)
1944 goto fail;
1945
1946 INIT_LIST_HEAD(&balloon.huge_pages);
1947 spin_lock_init(&balloon.comm_lock);
1948 init_rwsem(&balloon.conf_sem);
1949 balloon.vmci_doorbell = VMCI_INVALID_HANDLE;
1950 balloon.batch_page = NULL;
1951 balloon.page = NULL;
1952 balloon.reset_required = true;
1953
1954 queue_delayed_work(system_freezable_wq, &balloon.dwork, 0);
1955
1956 vmballoon_debugfs_init(&balloon);
1957
1958 return 0;
1959fail:
1960 vmballoon_unregister_shrinker(&balloon);
1961 vmballoon_compaction_deinit(&balloon);
1962 return error;
1963}
1964
1965/*
1966 * Using late_initcall() instead of module_init() allows the balloon to use the
1967 * VMCI doorbell even when the balloon is built into the kernel. Otherwise the
1968 * VMCI is probed only after the balloon is initialized. If the balloon is used
1969 * as a module, late_initcall() is equivalent to module_init().
1970 */
1971late_initcall(vmballoon_init);
1972
1973static void __exit vmballoon_exit(void)
1974{
1975 vmballoon_unregister_shrinker(&balloon);
1976 vmballoon_vmci_cleanup(&balloon);
1977 cancel_delayed_work_sync(&balloon.dwork);
1978
1979 vmballoon_debugfs_exit(&balloon);
1980
1981 /*
1982 * Deallocate all reserved memory, and reset connection with monitor.
1983 * Reset connection before deallocating memory to avoid potential for
1984 * additional spurious resets from guest touching deallocated pages.
1985 */
1986 vmballoon_send_start(&balloon, 0);
1987 vmballoon_pop(&balloon);
1988
1989 /* Only once we popped the balloon, compaction can be deinit */
1990 vmballoon_compaction_deinit(&balloon);
1991}
1992module_exit(vmballoon_exit);