1 // SPDX-License-Identifier: GPL-2.0
3 * VMware Balloon driver.
5 * Copyright (C) 2000-2018, VMware, Inc. All Rights Reserved.
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.
17 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
19 #include <linux/types.h>
21 #include <linux/kernel.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>
39 MODULE_AUTHOR("VMware, Inc.");
40 MODULE_DESCRIPTION("VMware Memory Control (Balloon) Driver");
41 MODULE_ALIAS("dmi:*:svnVMware*:*");
42 MODULE_ALIAS("vmware_vmmemctl");
43 MODULE_LICENSE("GPL");
45 static bool __read_mostly vmwballoon_shrinker_enable;
46 module_param(vmwballoon_shrinker_enable, bool, 0444);
47 MODULE_PARM_DESC(vmwballoon_shrinker_enable,
48 "Enable non-cooperative out-of-memory protection. Disabled by default as it may degrade performance.");
50 /* Delay in seconds after shrink before inflation. */
51 #define VMBALLOON_SHRINK_DELAY (5)
53 /* Maximum number of refused pages we accumulate during inflation cycle */
54 #define VMW_BALLOON_MAX_REFUSED 16
56 /* Magic number for the balloon mount-point */
57 #define BALLOON_VMW_MAGIC 0x0ba11007
60 * Hypervisor communication port definitions.
62 #define VMW_BALLOON_HV_PORT 0x5670
63 #define VMW_BALLOON_HV_MAGIC 0x456c6d6f
64 #define VMW_BALLOON_GUEST_ID 1 /* Linux */
66 enum vmwballoon_capabilities {
68 * Bit 0 is reserved and not associated to any capability.
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)
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)
82 #define VMW_BALLOON_2M_ORDER (PMD_SHIFT - PAGE_SHIFT)
85 * 64-bit targets are only supported in 64-bit
88 #define VMW_BALLOON_CAPABILITIES (VMW_BALLOON_CAPABILITIES_COMMON \
89 | VMW_BALLOON_64_BIT_TARGET)
91 #define VMW_BALLOON_CAPABILITIES VMW_BALLOON_CAPABILITIES_COMMON
94 enum vmballoon_page_size_type {
97 VMW_BALLOON_LAST_SIZE = VMW_BALLOON_2M_PAGE
100 #define VMW_BALLOON_NUM_PAGE_SIZES (VMW_BALLOON_LAST_SIZE + 1)
102 static const char * const vmballoon_page_size_names[] = {
103 [VMW_BALLOON_4K_PAGE] = "4k",
104 [VMW_BALLOON_2M_PAGE] = "2M"
112 enum vmballoon_op_stat_type {
114 VMW_BALLOON_OP_FAIL_STAT
117 #define VMW_BALLOON_OP_STAT_TYPES (VMW_BALLOON_OP_FAIL_STAT + 1)
120 * enum vmballoon_cmd_type - backdoor commands.
122 * Availability of the commands is as followed:
124 * %VMW_BALLOON_CMD_START, %VMW_BALLOON_CMD_GET_TARGET and
125 * %VMW_BALLOON_CMD_GUEST_ID are always available.
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.
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
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
138 * If the host reports VMW_BALLOON_SIGNALLED_WAKEUP_CMD is supported then
139 * VMW_BALLOON_CMD_VMCI_DOORBELL_SET command is supported.
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
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
155 * @VMW_BALLOON_CMD_BATCHED_2M_UNLOCK: Similar to
156 * @VMW_BALLOON_CMD_BATCHED_UNLOCK for 2MB
158 * @VMW_BALLOON_CMD_VMCI_DOORBELL_SET: A command to set doorbell notification
159 * that would be invoked when the balloon
161 * @VMW_BALLOON_CMD_LAST: Value of the last command.
163 enum 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,
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,
178 #define VMW_BALLOON_CMD_NUM (VMW_BALLOON_CMD_LAST + 1)
180 enum vmballoon_error_codes {
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
192 #define VMW_BALLOON_SUCCESS_WITH_CAPABILITIES (0x03000000)
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))
203 static 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"
216 enum 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
225 #define VMW_BALLOON_PAGE_STAT_NUM (VMW_BALLOON_PAGE_STAT_LAST + 1)
227 enum 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
236 #define VMW_BALLOON_STAT_NUM (VMW_BALLOON_STAT_LAST + 1)
238 static DEFINE_STATIC_KEY_TRUE(vmw_balloon_batching);
239 static DEFINE_STATIC_KEY_FALSE(balloon_stat_enabled);
241 struct 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;
252 * struct vmballoon_batch_entry - a batch entry for lock or unlock.
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.
258 struct vmballoon_batch_entry {
260 u64 reserved : PAGE_SHIFT - 5;
266 * @max_page_size: maximum supported page size for ballooning.
268 * Protected by @conf_sem
270 enum vmballoon_page_size_type max_page_size;
273 * @size: balloon actual size in basic page size (frames).
275 * While we currently do not support size which is bigger than 32-bit,
276 * in preparation for future support, use 64-bits.
281 * @target: balloon target size in basic page size (frames).
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.
289 unsigned long target;
292 * @reset_required: reset flag
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.
302 * @capabilities: hypervisor balloon capabilities.
304 * Protected by @conf_sem.
306 unsigned long capabilities;
309 * @batch_page: pointer to communication batch page.
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.
314 struct vmballoon_batch_entry *batch_page;
317 * @batch_max_pages: maximum pages that can be locked/unlocked.
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.
323 * Protected by @conf_sem.
325 unsigned int batch_max_pages;
328 * @page: page to be locked/unlocked by the hypervisor
330 * @page is only used when batching is disabled and a single page is
331 * reclaimed on each iteration.
333 * Protected by @comm_lock.
338 * @shrink_timeout: timeout until the next inflation.
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.
344 unsigned long shrink_timeout;
347 struct vmballoon_stats *stats;
349 #ifdef CONFIG_DEBUG_FS
350 /* debugfs file exporting statistics */
351 struct dentry *dbg_entry;
355 * @b_dev_info: balloon device information descriptor.
357 struct balloon_dev_info b_dev_info;
359 struct delayed_work dwork;
362 * @huge_pages - list of the inflated 2MB pages.
364 * Protected by @b_dev_info.pages_lock .
366 struct list_head huge_pages;
371 * Protected by @conf_sem.
373 struct vmci_handle vmci_doorbell;
376 * @conf_sem: semaphore to protect the configuration and the statistics.
378 struct rw_semaphore conf_sem;
381 * @comm_lock: lock to protect the communication with the host.
383 * Lock ordering: @conf_sem -> @comm_lock .
385 spinlock_t comm_lock;
388 * @shrinker: shrinker interface that is used to avoid over-inflation.
390 struct shrinker shrinker;
393 * @shrinker_registered: whether the shrinker was registered.
395 * The shrinker interface does not handle gracefully the removal of
396 * shrinker that was not registered before. This indication allows to
397 * simplify the unregistration process.
399 bool shrinker_registered;
402 static struct vmballoon balloon;
404 struct vmballoon_stats {
405 /* timer / doorbell operations */
406 atomic64_t general_stat[VMW_BALLOON_STAT_NUM];
408 /* allocation statistics for huge and small pages */
410 page_stat[VMW_BALLOON_PAGE_STAT_NUM][VMW_BALLOON_NUM_PAGE_SIZES];
412 /* Monitor operations: total operations, and failures */
413 atomic64_t ops[VMW_BALLOON_CMD_NUM][VMW_BALLOON_OP_STAT_TYPES];
416 static inline bool is_vmballoon_stats_on(void)
418 return IS_ENABLED(CONFIG_DEBUG_FS) &&
419 static_branch_unlikely(&balloon_stat_enabled);
422 static inline void vmballoon_stats_op_inc(struct vmballoon *b, unsigned int op,
423 enum vmballoon_op_stat_type type)
425 if (is_vmballoon_stats_on())
426 atomic64_inc(&b->stats->ops[op][type]);
429 static inline void vmballoon_stats_gen_inc(struct vmballoon *b,
430 enum vmballoon_stat_general stat)
432 if (is_vmballoon_stats_on())
433 atomic64_inc(&b->stats->general_stat[stat]);
436 static inline void vmballoon_stats_gen_add(struct vmballoon *b,
437 enum vmballoon_stat_general stat,
440 if (is_vmballoon_stats_on())
441 atomic64_add(val, &b->stats->general_stat[stat]);
444 static inline void vmballoon_stats_page_inc(struct vmballoon *b,
445 enum vmballoon_stat_page stat,
446 enum vmballoon_page_size_type size)
448 if (is_vmballoon_stats_on())
449 atomic64_inc(&b->stats->page_stat[stat][size]);
452 static inline void vmballoon_stats_page_add(struct vmballoon *b,
453 enum vmballoon_stat_page stat,
454 enum vmballoon_page_size_type size,
457 if (is_vmballoon_stats_on())
458 atomic64_add(val, &b->stats->page_stat[stat][size]);
461 static inline unsigned long
462 __vmballoon_cmd(struct vmballoon *b, unsigned long cmd, unsigned long arg1,
463 unsigned long arg2, unsigned long *result)
465 unsigned long status, dummy1, dummy2, dummy3, local_result;
467 vmballoon_stats_op_inc(b, cmd, VMW_BALLOON_OP_STAT);
469 asm volatile ("inl %%dx" :
475 "0"(VMW_BALLOON_HV_MAGIC),
477 "2"(VMW_BALLOON_HV_PORT),
482 /* update the result if needed */
484 *result = (cmd == VMW_BALLOON_CMD_START) ? dummy1 :
487 /* update target when applicable */
488 if (status == VMW_BALLOON_SUCCESS &&
489 ((1ul << cmd) & VMW_BALLOON_CMD_WITH_TARGET_MASK))
490 WRITE_ONCE(b->target, local_result);
492 if (status != VMW_BALLOON_SUCCESS &&
493 status != VMW_BALLOON_SUCCESS_WITH_CAPABILITIES) {
494 vmballoon_stats_op_inc(b, cmd, VMW_BALLOON_OP_FAIL_STAT);
495 pr_debug("%s: %s [0x%lx,0x%lx) failed, returned %ld\n",
496 __func__, vmballoon_cmd_names[cmd], arg1, arg2,
500 /* mark reset required accordingly */
501 if (status == VMW_BALLOON_ERROR_RESET)
502 b->reset_required = true;
507 static __always_inline unsigned long
508 vmballoon_cmd(struct vmballoon *b, unsigned long cmd, unsigned long arg1,
513 return __vmballoon_cmd(b, cmd, arg1, arg2, &dummy);
517 * Send "start" command to the host, communicating supported version
520 static int vmballoon_send_start(struct vmballoon *b, unsigned long req_caps)
522 unsigned long status, capabilities;
524 status = __vmballoon_cmd(b, VMW_BALLOON_CMD_START, req_caps, 0,
528 case VMW_BALLOON_SUCCESS_WITH_CAPABILITIES:
529 b->capabilities = capabilities;
531 case VMW_BALLOON_SUCCESS:
532 b->capabilities = VMW_BALLOON_BASIC_CMDS;
539 * 2MB pages are only supported with batching. If batching is for some
540 * reason disabled, do not use 2MB pages, since otherwise the legacy
541 * mechanism is used with 2MB pages, causing a failure.
543 b->max_page_size = VMW_BALLOON_4K_PAGE;
544 if ((b->capabilities & VMW_BALLOON_BATCHED_2M_CMDS) &&
545 (b->capabilities & VMW_BALLOON_BATCHED_CMDS))
546 b->max_page_size = VMW_BALLOON_2M_PAGE;
553 * vmballoon_send_guest_id - communicate guest type to the host.
555 * @b: pointer to the balloon.
557 * Communicate guest type to the host so that it can adjust ballooning
558 * algorithm to the one most appropriate for the guest. This command
559 * is normally issued after sending "start" command and is part of
560 * standard reset sequence.
562 * Return: zero on success or appropriate error code.
564 static int vmballoon_send_guest_id(struct vmballoon *b)
566 unsigned long status;
568 status = vmballoon_cmd(b, VMW_BALLOON_CMD_GUEST_ID,
569 VMW_BALLOON_GUEST_ID, 0);
571 return status == VMW_BALLOON_SUCCESS ? 0 : -EIO;
575 * vmballoon_page_order() - return the order of the page
576 * @page_size: the size of the page.
578 * Return: the allocation order.
581 unsigned int vmballoon_page_order(enum vmballoon_page_size_type page_size)
583 return page_size == VMW_BALLOON_2M_PAGE ? VMW_BALLOON_2M_ORDER : 0;
587 * vmballoon_page_in_frames() - returns the number of frames in a page.
588 * @page_size: the size of the page.
590 * Return: the number of 4k frames.
592 static inline unsigned int
593 vmballoon_page_in_frames(enum vmballoon_page_size_type page_size)
595 return 1 << vmballoon_page_order(page_size);
599 * vmballoon_mark_page_offline() - mark a page as offline
600 * @page: pointer for the page.
601 * @page_size: the size of the page.
604 vmballoon_mark_page_offline(struct page *page,
605 enum vmballoon_page_size_type page_size)
609 for (i = 0; i < vmballoon_page_in_frames(page_size); i++)
610 __SetPageOffline(page + i);
614 * vmballoon_mark_page_online() - mark a page as online
615 * @page: pointer for the page.
616 * @page_size: the size of the page.
619 vmballoon_mark_page_online(struct page *page,
620 enum vmballoon_page_size_type page_size)
624 for (i = 0; i < vmballoon_page_in_frames(page_size); i++)
625 __ClearPageOffline(page + i);
629 * vmballoon_send_get_target() - Retrieve desired balloon size from the host.
631 * @b: pointer to the balloon.
633 * Return: zero on success, EINVAL if limit does not fit in 32-bit, as required
634 * by the host-guest protocol and EIO if an error occurred in communicating with
637 static int vmballoon_send_get_target(struct vmballoon *b)
639 unsigned long status;
642 limit = totalram_pages();
644 /* Ensure limit fits in 32-bits if 64-bit targets are not supported */
645 if (!(b->capabilities & VMW_BALLOON_64_BIT_TARGET) &&
649 status = vmballoon_cmd(b, VMW_BALLOON_CMD_GET_TARGET, limit, 0);
651 return status == VMW_BALLOON_SUCCESS ? 0 : -EIO;
655 * vmballoon_alloc_page_list - allocates a list of pages.
657 * @b: pointer to the balloon.
658 * @ctl: pointer for the %struct vmballoon_ctl, which defines the operation.
659 * @req_n_pages: the number of requested pages.
661 * Tries to allocate @req_n_pages. Add them to the list of balloon pages in
662 * @ctl.pages and updates @ctl.n_pages to reflect the number of pages.
664 * Return: zero on success or error code otherwise.
666 static int vmballoon_alloc_page_list(struct vmballoon *b,
667 struct vmballoon_ctl *ctl,
668 unsigned int req_n_pages)
673 for (i = 0; i < req_n_pages; i++) {
675 * First check if we happen to have pages that were allocated
676 * before. This happens when 2MB page rejected during inflation
677 * by the hypervisor, and then split into 4KB pages.
679 if (!list_empty(&ctl->prealloc_pages)) {
680 page = list_first_entry(&ctl->prealloc_pages,
682 list_del(&page->lru);
684 if (ctl->page_size == VMW_BALLOON_2M_PAGE)
685 page = alloc_pages(__GFP_HIGHMEM|__GFP_NOWARN|
686 __GFP_NOMEMALLOC, VMW_BALLOON_2M_ORDER);
688 page = balloon_page_alloc();
690 vmballoon_stats_page_inc(b, VMW_BALLOON_PAGE_STAT_ALLOC,
695 /* Success. Add the page to the list and continue. */
696 list_add(&page->lru, &ctl->pages);
700 /* Allocation failed. Update statistics and stop. */
701 vmballoon_stats_page_inc(b, VMW_BALLOON_PAGE_STAT_ALLOC_FAIL,
708 return req_n_pages == ctl->n_pages ? 0 : -ENOMEM;
712 * vmballoon_handle_one_result - Handle lock/unlock result for a single page.
714 * @b: pointer for %struct vmballoon.
715 * @page: pointer for the page whose result should be handled.
716 * @page_size: size of the page.
717 * @status: status of the operation as provided by the hypervisor.
719 static int vmballoon_handle_one_result(struct vmballoon *b, struct page *page,
720 enum vmballoon_page_size_type page_size,
721 unsigned long status)
723 /* On success do nothing. The page is already on the balloon list. */
724 if (likely(status == VMW_BALLOON_SUCCESS))
727 pr_debug("%s: failed comm pfn %lx status %lu page_size %s\n", __func__,
728 page_to_pfn(page), status,
729 vmballoon_page_size_names[page_size]);
732 vmballoon_stats_page_inc(b, VMW_BALLOON_PAGE_STAT_REFUSED_ALLOC,
739 * vmballoon_status_page - returns the status of (un)lock operation
741 * @b: pointer to the balloon.
742 * @idx: index for the page for which the operation is performed.
743 * @p: pointer to where the page struct is returned.
745 * Following a lock or unlock operation, returns the status of the operation for
746 * an individual page. Provides the page that the operation was performed on on
747 * the @page argument.
749 * Returns: The status of a lock or unlock operation for an individual page.
751 static unsigned long vmballoon_status_page(struct vmballoon *b, int idx,
754 if (static_branch_likely(&vmw_balloon_batching)) {
756 *p = pfn_to_page(b->batch_page[idx].pfn);
757 return b->batch_page[idx].status;
760 /* non-batching mode */
764 * If a failure occurs, the indication will be provided in the status
765 * of the entire operation, which is considered before the individual
766 * page status. So for non-batching mode, the indication is always of
769 return VMW_BALLOON_SUCCESS;
773 * vmballoon_lock_op - notifies the host about inflated/deflated pages.
774 * @b: pointer to the balloon.
775 * @num_pages: number of inflated/deflated pages.
776 * @page_size: size of the page.
777 * @op: the type of operation (lock or unlock).
779 * Notify the host about page(s) that were ballooned (or removed from the
780 * balloon) so that host can use it without fear that guest will need it (or
781 * stop using them since the VM does). Host may reject some pages, we need to
782 * check the return value and maybe submit a different page. The pages that are
783 * inflated/deflated are pointed by @b->page.
785 * Return: result as provided by the hypervisor.
787 static unsigned long vmballoon_lock_op(struct vmballoon *b,
788 unsigned int num_pages,
789 enum vmballoon_page_size_type page_size,
790 enum vmballoon_op op)
792 unsigned long cmd, pfn;
794 lockdep_assert_held(&b->comm_lock);
796 if (static_branch_likely(&vmw_balloon_batching)) {
797 if (op == VMW_BALLOON_INFLATE)
798 cmd = page_size == VMW_BALLOON_2M_PAGE ?
799 VMW_BALLOON_CMD_BATCHED_2M_LOCK :
800 VMW_BALLOON_CMD_BATCHED_LOCK;
802 cmd = page_size == VMW_BALLOON_2M_PAGE ?
803 VMW_BALLOON_CMD_BATCHED_2M_UNLOCK :
804 VMW_BALLOON_CMD_BATCHED_UNLOCK;
806 pfn = PHYS_PFN(virt_to_phys(b->batch_page));
808 cmd = op == VMW_BALLOON_INFLATE ? VMW_BALLOON_CMD_LOCK :
809 VMW_BALLOON_CMD_UNLOCK;
810 pfn = page_to_pfn(b->page);
812 /* In non-batching mode, PFNs must fit in 32-bit */
813 if (unlikely(pfn != (u32)pfn))
814 return VMW_BALLOON_ERROR_PPN_INVALID;
817 return vmballoon_cmd(b, cmd, pfn, num_pages);
821 * vmballoon_add_page - adds a page towards lock/unlock operation.
823 * @b: pointer to the balloon.
824 * @idx: index of the page to be ballooned in this batch.
825 * @p: pointer to the page that is about to be ballooned.
827 * Adds the page to be ballooned. Must be called while holding @comm_lock.
829 static void vmballoon_add_page(struct vmballoon *b, unsigned int idx,
832 lockdep_assert_held(&b->comm_lock);
834 if (static_branch_likely(&vmw_balloon_batching))
835 b->batch_page[idx] = (struct vmballoon_batch_entry)
836 { .pfn = page_to_pfn(p) };
842 * vmballoon_lock - lock or unlock a batch of pages.
844 * @b: pointer to the balloon.
845 * @ctl: pointer for the %struct vmballoon_ctl, which defines the operation.
847 * Notifies the host of about ballooned pages (after inflation or deflation,
848 * according to @ctl). If the host rejects the page put it on the
849 * @ctl refuse list. These refused page are then released when moving to the
850 * next size of pages.
852 * Note that we neither free any @page here nor put them back on the ballooned
853 * pages list. Instead we queue it for later processing. We do that for several
854 * reasons. First, we do not want to free the page under the lock. Second, it
855 * allows us to unify the handling of lock and unlock. In the inflate case, the
856 * caller will check if there are too many refused pages and release them.
857 * Although it is not identical to the past behavior, it should not affect
860 static int vmballoon_lock(struct vmballoon *b, struct vmballoon_ctl *ctl)
862 unsigned long batch_status;
864 unsigned int i, num_pages;
866 num_pages = ctl->n_pages;
870 /* communication with the host is done under the communication lock */
871 spin_lock(&b->comm_lock);
874 list_for_each_entry(page, &ctl->pages, lru)
875 vmballoon_add_page(b, i++, page);
877 batch_status = vmballoon_lock_op(b, ctl->n_pages, ctl->page_size,
881 * Iterate over the pages in the provided list. Since we are changing
882 * @ctl->n_pages we are saving the original value in @num_pages and
883 * use this value to bound the loop.
885 for (i = 0; i < num_pages; i++) {
886 unsigned long status;
888 status = vmballoon_status_page(b, i, &page);
891 * Failure of the whole batch overrides a single operation
894 if (batch_status != VMW_BALLOON_SUCCESS)
895 status = batch_status;
897 /* Continue if no error happened */
898 if (!vmballoon_handle_one_result(b, page, ctl->page_size,
903 * Error happened. Move the pages to the refused list and update
906 list_move(&page->lru, &ctl->refused_pages);
908 ctl->n_refused_pages++;
911 spin_unlock(&b->comm_lock);
913 return batch_status == VMW_BALLOON_SUCCESS ? 0 : -EIO;
917 * vmballoon_release_page_list() - Releases a page list
919 * @page_list: list of pages to release.
920 * @n_pages: pointer to the number of pages.
921 * @page_size: whether the pages in the list are 2MB (or else 4KB).
923 * Releases the list of pages and zeros the number of pages.
925 static void vmballoon_release_page_list(struct list_head *page_list,
927 enum vmballoon_page_size_type page_size)
929 struct page *page, *tmp;
931 list_for_each_entry_safe(page, tmp, page_list, lru) {
932 list_del(&page->lru);
933 __free_pages(page, vmballoon_page_order(page_size));
942 * Release pages that were allocated while attempting to inflate the
943 * balloon but were refused by the host for one reason or another.
945 static void vmballoon_release_refused_pages(struct vmballoon *b,
946 struct vmballoon_ctl *ctl)
948 vmballoon_stats_page_inc(b, VMW_BALLOON_PAGE_STAT_REFUSED_FREE,
951 vmballoon_release_page_list(&ctl->refused_pages, &ctl->n_refused_pages,
956 * vmballoon_change - retrieve the required balloon change
958 * @b: pointer for the balloon.
960 * Return: the required change for the balloon size. A positive number
961 * indicates inflation, a negative number indicates a deflation.
963 static int64_t vmballoon_change(struct vmballoon *b)
965 int64_t size, target;
967 size = atomic64_read(&b->size);
968 target = READ_ONCE(b->target);
971 * We must cast first because of int sizes
972 * Otherwise we might get huge positives instead of negatives
975 if (b->reset_required)
978 /* consider a 2MB slack on deflate, unless the balloon is emptied */
979 if (target < size && target != 0 &&
980 size - target < vmballoon_page_in_frames(VMW_BALLOON_2M_PAGE))
983 /* If an out-of-memory recently occurred, inflation is disallowed. */
984 if (target > size && time_before(jiffies, READ_ONCE(b->shrink_timeout)))
987 return target - size;
991 * vmballoon_enqueue_page_list() - Enqueues list of pages after inflation.
993 * @b: pointer to balloon.
994 * @pages: list of pages to enqueue.
995 * @n_pages: pointer to number of pages in list. The value is zeroed.
996 * @page_size: whether the pages are 2MB or 4KB pages.
998 * Enqueues the provides list of pages in the ballooned page list, clears the
999 * list and zeroes the number of pages that was provided.
1001 static void vmballoon_enqueue_page_list(struct vmballoon *b,
1002 struct list_head *pages,
1003 unsigned int *n_pages,
1004 enum vmballoon_page_size_type page_size)
1006 unsigned long flags;
1009 if (page_size == VMW_BALLOON_4K_PAGE) {
1010 balloon_page_list_enqueue(&b->b_dev_info, pages);
1013 * Keep the huge pages in a local list which is not available
1014 * for the balloon compaction mechanism.
1016 spin_lock_irqsave(&b->b_dev_info.pages_lock, flags);
1018 list_for_each_entry(page, pages, lru) {
1019 vmballoon_mark_page_offline(page, VMW_BALLOON_2M_PAGE);
1022 list_splice_init(pages, &b->huge_pages);
1023 __count_vm_events(BALLOON_INFLATE, *n_pages *
1024 vmballoon_page_in_frames(VMW_BALLOON_2M_PAGE));
1025 spin_unlock_irqrestore(&b->b_dev_info.pages_lock, flags);
1032 * vmballoon_dequeue_page_list() - Dequeues page lists for deflation.
1034 * @b: pointer to balloon.
1035 * @pages: list of pages to enqueue.
1036 * @n_pages: pointer to number of pages in list. The value is zeroed.
1037 * @page_size: whether the pages are 2MB or 4KB pages.
1038 * @n_req_pages: the number of requested pages.
1040 * Dequeues the number of requested pages from the balloon for deflation. The
1041 * number of dequeued pages may be lower, if not enough pages in the requested
1042 * size are available.
1044 static void vmballoon_dequeue_page_list(struct vmballoon *b,
1045 struct list_head *pages,
1046 unsigned int *n_pages,
1047 enum vmballoon_page_size_type page_size,
1048 unsigned int n_req_pages)
1050 struct page *page, *tmp;
1052 unsigned long flags;
1054 /* In the case of 4k pages, use the compaction infrastructure */
1055 if (page_size == VMW_BALLOON_4K_PAGE) {
1056 *n_pages = balloon_page_list_dequeue(&b->b_dev_info, pages,
1062 spin_lock_irqsave(&b->b_dev_info.pages_lock, flags);
1063 list_for_each_entry_safe(page, tmp, &b->huge_pages, lru) {
1064 vmballoon_mark_page_online(page, VMW_BALLOON_2M_PAGE);
1066 list_move(&page->lru, pages);
1067 if (++i == n_req_pages)
1071 __count_vm_events(BALLOON_DEFLATE,
1072 i * vmballoon_page_in_frames(VMW_BALLOON_2M_PAGE));
1073 spin_unlock_irqrestore(&b->b_dev_info.pages_lock, flags);
1078 * vmballoon_split_refused_pages() - Split the 2MB refused pages to 4k.
1080 * If inflation of 2MB pages was denied by the hypervisor, it is likely to be
1081 * due to one or few 4KB pages. These 2MB pages may keep being allocated and
1082 * then being refused. To prevent this case, this function splits the refused
1083 * pages into 4KB pages and adds them into @prealloc_pages list.
1085 * @ctl: pointer for the %struct vmballoon_ctl, which defines the operation.
1087 static void vmballoon_split_refused_pages(struct vmballoon_ctl *ctl)
1089 struct page *page, *tmp;
1090 unsigned int i, order;
1092 order = vmballoon_page_order(ctl->page_size);
1094 list_for_each_entry_safe(page, tmp, &ctl->refused_pages, lru) {
1095 list_del(&page->lru);
1096 split_page(page, order);
1097 for (i = 0; i < (1 << order); i++)
1098 list_add(&page[i].lru, &ctl->prealloc_pages);
1100 ctl->n_refused_pages = 0;
1104 * vmballoon_inflate() - Inflate the balloon towards its target size.
1106 * @b: pointer to the balloon.
1108 static void vmballoon_inflate(struct vmballoon *b)
1110 int64_t to_inflate_frames;
1111 struct vmballoon_ctl ctl = {
1112 .pages = LIST_HEAD_INIT(ctl.pages),
1113 .refused_pages = LIST_HEAD_INIT(ctl.refused_pages),
1114 .prealloc_pages = LIST_HEAD_INIT(ctl.prealloc_pages),
1115 .page_size = b->max_page_size,
1116 .op = VMW_BALLOON_INFLATE
1119 while ((to_inflate_frames = vmballoon_change(b)) > 0) {
1120 unsigned int to_inflate_pages, page_in_frames;
1121 int alloc_error, lock_error = 0;
1123 VM_BUG_ON(!list_empty(&ctl.pages));
1124 VM_BUG_ON(ctl.n_pages != 0);
1126 page_in_frames = vmballoon_page_in_frames(ctl.page_size);
1128 to_inflate_pages = min_t(unsigned long, b->batch_max_pages,
1129 DIV_ROUND_UP_ULL(to_inflate_frames,
1132 /* Start by allocating */
1133 alloc_error = vmballoon_alloc_page_list(b, &ctl,
1136 /* Actually lock the pages by telling the hypervisor */
1137 lock_error = vmballoon_lock(b, &ctl);
1140 * If an error indicates that something serious went wrong,
1141 * stop the inflation.
1146 /* Update the balloon size */
1147 atomic64_add(ctl.n_pages * page_in_frames, &b->size);
1149 vmballoon_enqueue_page_list(b, &ctl.pages, &ctl.n_pages,
1153 * If allocation failed or the number of refused pages exceeds
1154 * the maximum allowed, move to the next page size.
1157 ctl.n_refused_pages >= VMW_BALLOON_MAX_REFUSED) {
1158 if (ctl.page_size == VMW_BALLOON_4K_PAGE)
1162 * Split the refused pages to 4k. This will also empty
1163 * the refused pages list.
1165 vmballoon_split_refused_pages(&ctl);
1173 * Release pages that were allocated while attempting to inflate the
1174 * balloon but were refused by the host for one reason or another,
1175 * and update the statistics.
1177 if (ctl.n_refused_pages != 0)
1178 vmballoon_release_refused_pages(b, &ctl);
1180 vmballoon_release_page_list(&ctl.prealloc_pages, NULL, ctl.page_size);
1184 * vmballoon_deflate() - Decrease the size of the balloon.
1186 * @b: pointer to the balloon
1187 * @n_frames: the number of frames to deflate. If zero, automatically
1188 * calculated according to the target size.
1189 * @coordinated: whether to coordinate with the host
1191 * Decrease the size of the balloon allowing guest to use more memory.
1193 * Return: The number of deflated frames (i.e., basic page size units)
1195 static unsigned long vmballoon_deflate(struct vmballoon *b, uint64_t n_frames,
1198 unsigned long deflated_frames = 0;
1199 unsigned long tried_frames = 0;
1200 struct vmballoon_ctl ctl = {
1201 .pages = LIST_HEAD_INIT(ctl.pages),
1202 .refused_pages = LIST_HEAD_INIT(ctl.refused_pages),
1203 .page_size = VMW_BALLOON_4K_PAGE,
1204 .op = VMW_BALLOON_DEFLATE
1207 /* free pages to reach target */
1209 unsigned int to_deflate_pages, n_unlocked_frames;
1210 unsigned int page_in_frames;
1211 int64_t to_deflate_frames;
1214 page_in_frames = vmballoon_page_in_frames(ctl.page_size);
1216 VM_BUG_ON(!list_empty(&ctl.pages));
1217 VM_BUG_ON(ctl.n_pages);
1218 VM_BUG_ON(!list_empty(&ctl.refused_pages));
1219 VM_BUG_ON(ctl.n_refused_pages);
1222 * If we were requested a specific number of frames, we try to
1223 * deflate this number of frames. Otherwise, deflation is
1224 * performed according to the target and balloon size.
1226 to_deflate_frames = n_frames ? n_frames - tried_frames :
1227 -vmballoon_change(b);
1229 /* break if no work to do */
1230 if (to_deflate_frames <= 0)
1234 * Calculate the number of frames based on current page size,
1235 * but limit the deflated frames to a single chunk
1237 to_deflate_pages = min_t(unsigned long, b->batch_max_pages,
1238 DIV_ROUND_UP_ULL(to_deflate_frames,
1241 /* First take the pages from the balloon pages. */
1242 vmballoon_dequeue_page_list(b, &ctl.pages, &ctl.n_pages,
1243 ctl.page_size, to_deflate_pages);
1246 * Before pages are moving to the refused list, count their
1247 * frames as frames that we tried to deflate.
1249 tried_frames += ctl.n_pages * page_in_frames;
1252 * Unlock the pages by communicating with the hypervisor if the
1253 * communication is coordinated (i.e., not pop). We ignore the
1254 * return code. Instead we check if all the pages we manage to
1255 * unlock all the pages. If we failed, we will move to the next
1256 * page size, and would eventually try again later.
1259 vmballoon_lock(b, &ctl);
1262 * Check if we deflated enough. We will move to the next page
1263 * size if we did not manage to do so. This calculation takes
1264 * place now, as once the pages are released, the number of
1267 deflated_all = (ctl.n_pages == to_deflate_pages);
1269 /* Update local and global counters */
1270 n_unlocked_frames = ctl.n_pages * page_in_frames;
1271 atomic64_sub(n_unlocked_frames, &b->size);
1272 deflated_frames += n_unlocked_frames;
1274 vmballoon_stats_page_add(b, VMW_BALLOON_PAGE_STAT_FREE,
1275 ctl.page_size, ctl.n_pages);
1277 /* free the ballooned pages */
1278 vmballoon_release_page_list(&ctl.pages, &ctl.n_pages,
1281 /* Return the refused pages to the ballooned list. */
1282 vmballoon_enqueue_page_list(b, &ctl.refused_pages,
1283 &ctl.n_refused_pages,
1286 /* If we failed to unlock all the pages, move to next size. */
1287 if (!deflated_all) {
1288 if (ctl.page_size == b->max_page_size)
1296 return deflated_frames;
1300 * vmballoon_deinit_batching - disables batching mode.
1302 * @b: pointer to &struct vmballoon.
1304 * Disables batching, by deallocating the page for communication with the
1305 * hypervisor and disabling the static key to indicate that batching is off.
1307 static void vmballoon_deinit_batching(struct vmballoon *b)
1309 free_page((unsigned long)b->batch_page);
1310 b->batch_page = NULL;
1311 static_branch_disable(&vmw_balloon_batching);
1312 b->batch_max_pages = 1;
1316 * vmballoon_init_batching - enable batching mode.
1318 * @b: pointer to &struct vmballoon.
1320 * Enables batching, by allocating a page for communication with the hypervisor
1321 * and enabling the static_key to use batching.
1323 * Return: zero on success or an appropriate error-code.
1325 static int vmballoon_init_batching(struct vmballoon *b)
1329 page = alloc_page(GFP_KERNEL | __GFP_ZERO);
1333 b->batch_page = page_address(page);
1334 b->batch_max_pages = PAGE_SIZE / sizeof(struct vmballoon_batch_entry);
1336 static_branch_enable(&vmw_balloon_batching);
1342 * Receive notification and resize balloon
1344 static void vmballoon_doorbell(void *client_data)
1346 struct vmballoon *b = client_data;
1348 vmballoon_stats_gen_inc(b, VMW_BALLOON_STAT_DOORBELL);
1350 mod_delayed_work(system_freezable_wq, &b->dwork, 0);
1354 * Clean up vmci doorbell
1356 static void vmballoon_vmci_cleanup(struct vmballoon *b)
1358 vmballoon_cmd(b, VMW_BALLOON_CMD_VMCI_DOORBELL_SET,
1359 VMCI_INVALID_ID, VMCI_INVALID_ID);
1361 if (!vmci_handle_is_invalid(b->vmci_doorbell)) {
1362 vmci_doorbell_destroy(b->vmci_doorbell);
1363 b->vmci_doorbell = VMCI_INVALID_HANDLE;
1368 * vmballoon_vmci_init - Initialize vmci doorbell.
1370 * @b: pointer to the balloon.
1372 * Return: zero on success or when wakeup command not supported. Error-code
1375 * Initialize vmci doorbell, to get notified as soon as balloon changes.
1377 static int vmballoon_vmci_init(struct vmballoon *b)
1379 unsigned long error;
1381 if ((b->capabilities & VMW_BALLOON_SIGNALLED_WAKEUP_CMD) == 0)
1384 error = vmci_doorbell_create(&b->vmci_doorbell, VMCI_FLAG_DELAYED_CB,
1385 VMCI_PRIVILEGE_FLAG_RESTRICTED,
1386 vmballoon_doorbell, b);
1388 if (error != VMCI_SUCCESS)
1391 error = __vmballoon_cmd(b, VMW_BALLOON_CMD_VMCI_DOORBELL_SET,
1392 b->vmci_doorbell.context,
1393 b->vmci_doorbell.resource, NULL);
1395 if (error != VMW_BALLOON_SUCCESS)
1400 vmballoon_vmci_cleanup(b);
1405 * vmballoon_pop - Quickly release all pages allocate for the balloon.
1407 * @b: pointer to the balloon.
1409 * This function is called when host decides to "reset" balloon for one reason
1410 * or another. Unlike normal "deflate" we do not (shall not) notify host of the
1411 * pages being released.
1413 static void vmballoon_pop(struct vmballoon *b)
1417 while ((size = atomic64_read(&b->size)))
1418 vmballoon_deflate(b, size, false);
1422 * Perform standard reset sequence by popping the balloon (in case it
1423 * is not empty) and then restarting protocol. This operation normally
1424 * happens when host responds with VMW_BALLOON_ERROR_RESET to a command.
1426 static void vmballoon_reset(struct vmballoon *b)
1430 down_write(&b->conf_sem);
1432 vmballoon_vmci_cleanup(b);
1434 /* free all pages, skipping monitor unlock */
1437 if (vmballoon_send_start(b, VMW_BALLOON_CAPABILITIES))
1440 if ((b->capabilities & VMW_BALLOON_BATCHED_CMDS) != 0) {
1441 if (vmballoon_init_batching(b)) {
1443 * We failed to initialize batching, inform the monitor
1444 * about it by sending a null capability.
1446 * The guest will retry in one second.
1448 vmballoon_send_start(b, 0);
1451 } else if ((b->capabilities & VMW_BALLOON_BASIC_CMDS) != 0) {
1452 vmballoon_deinit_batching(b);
1455 vmballoon_stats_gen_inc(b, VMW_BALLOON_STAT_RESET);
1456 b->reset_required = false;
1458 error = vmballoon_vmci_init(b);
1460 pr_err("failed to initialize vmci doorbell\n");
1462 if (vmballoon_send_guest_id(b))
1463 pr_err("failed to send guest ID to the host\n");
1466 up_write(&b->conf_sem);
1470 * vmballoon_work - periodic balloon worker for reset, inflation and deflation.
1472 * @work: pointer to the &work_struct which is provided by the workqueue.
1474 * Resets the protocol if needed, gets the new size and adjusts balloon as
1475 * needed. Repeat in 1 sec.
1477 static void vmballoon_work(struct work_struct *work)
1479 struct delayed_work *dwork = to_delayed_work(work);
1480 struct vmballoon *b = container_of(dwork, struct vmballoon, dwork);
1483 if (b->reset_required)
1486 down_read(&b->conf_sem);
1489 * Update the stats while holding the semaphore to ensure that
1490 * @stats_enabled is consistent with whether the stats are actually
1493 vmballoon_stats_gen_inc(b, VMW_BALLOON_STAT_TIMER);
1495 if (!vmballoon_send_get_target(b))
1496 change = vmballoon_change(b);
1499 pr_debug("%s - size: %llu, target %lu\n", __func__,
1500 atomic64_read(&b->size), READ_ONCE(b->target));
1503 vmballoon_inflate(b);
1504 else /* (change < 0) */
1505 vmballoon_deflate(b, 0, true);
1508 up_read(&b->conf_sem);
1511 * We are using a freezable workqueue so that balloon operations are
1512 * stopped while the system transitions to/from sleep/hibernation.
1514 queue_delayed_work(system_freezable_wq,
1515 dwork, round_jiffies_relative(HZ));
1520 * vmballoon_shrinker_scan() - deflate the balloon due to memory pressure.
1521 * @shrinker: pointer to the balloon shrinker.
1522 * @sc: page reclaim information.
1524 * Returns: number of pages that were freed during deflation.
1526 static unsigned long vmballoon_shrinker_scan(struct shrinker *shrinker,
1527 struct shrink_control *sc)
1529 struct vmballoon *b = &balloon;
1530 unsigned long deflated_frames;
1532 pr_debug("%s - size: %llu", __func__, atomic64_read(&b->size));
1534 vmballoon_stats_gen_inc(b, VMW_BALLOON_STAT_SHRINK);
1537 * If the lock is also contended for read, we cannot easily reclaim and
1540 if (!down_read_trylock(&b->conf_sem))
1543 deflated_frames = vmballoon_deflate(b, sc->nr_to_scan, true);
1545 vmballoon_stats_gen_add(b, VMW_BALLOON_STAT_SHRINK_FREE,
1549 * Delay future inflation for some time to mitigate the situations in
1550 * which balloon continuously grows and shrinks. Use WRITE_ONCE() since
1551 * the access is asynchronous.
1553 WRITE_ONCE(b->shrink_timeout, jiffies + HZ * VMBALLOON_SHRINK_DELAY);
1555 up_read(&b->conf_sem);
1557 return deflated_frames;
1561 * vmballoon_shrinker_count() - return the number of ballooned pages.
1562 * @shrinker: pointer to the balloon shrinker.
1563 * @sc: page reclaim information.
1565 * Returns: number of 4k pages that are allocated for the balloon and can
1566 * therefore be reclaimed under pressure.
1568 static unsigned long vmballoon_shrinker_count(struct shrinker *shrinker,
1569 struct shrink_control *sc)
1571 struct vmballoon *b = &balloon;
1573 return atomic64_read(&b->size);
1576 static void vmballoon_unregister_shrinker(struct vmballoon *b)
1578 if (b->shrinker_registered)
1579 unregister_shrinker(&b->shrinker);
1580 b->shrinker_registered = false;
1583 static int vmballoon_register_shrinker(struct vmballoon *b)
1587 /* Do nothing if the shrinker is not enabled */
1588 if (!vmwballoon_shrinker_enable)
1591 b->shrinker.scan_objects = vmballoon_shrinker_scan;
1592 b->shrinker.count_objects = vmballoon_shrinker_count;
1593 b->shrinker.seeks = DEFAULT_SEEKS;
1595 r = register_shrinker(&b->shrinker);
1598 b->shrinker_registered = true;
1606 #ifdef CONFIG_DEBUG_FS
1608 static const char * const vmballoon_stat_page_names[] = {
1609 [VMW_BALLOON_PAGE_STAT_ALLOC] = "alloc",
1610 [VMW_BALLOON_PAGE_STAT_ALLOC_FAIL] = "allocFail",
1611 [VMW_BALLOON_PAGE_STAT_REFUSED_ALLOC] = "errAlloc",
1612 [VMW_BALLOON_PAGE_STAT_REFUSED_FREE] = "errFree",
1613 [VMW_BALLOON_PAGE_STAT_FREE] = "free"
1616 static const char * const vmballoon_stat_names[] = {
1617 [VMW_BALLOON_STAT_TIMER] = "timer",
1618 [VMW_BALLOON_STAT_DOORBELL] = "doorbell",
1619 [VMW_BALLOON_STAT_RESET] = "reset",
1620 [VMW_BALLOON_STAT_SHRINK] = "shrink",
1621 [VMW_BALLOON_STAT_SHRINK_FREE] = "shrinkFree"
1624 static int vmballoon_enable_stats(struct vmballoon *b)
1628 down_write(&b->conf_sem);
1630 /* did we somehow race with another reader which enabled stats? */
1634 b->stats = kzalloc(sizeof(*b->stats), GFP_KERNEL);
1637 /* allocation failed */
1641 static_key_enable(&balloon_stat_enabled.key);
1643 up_write(&b->conf_sem);
1648 * vmballoon_debug_show - shows statistics of balloon operations.
1649 * @f: pointer to the &struct seq_file.
1652 * Provides the statistics that can be accessed in vmmemctl in the debugfs.
1653 * To avoid the overhead - mainly that of memory - of collecting the statistics,
1654 * we only collect statistics after the first time the counters are read.
1656 * Return: zero on success or an error code.
1658 static int vmballoon_debug_show(struct seq_file *f, void *offset)
1660 struct vmballoon *b = f->private;
1663 /* enables stats if they are disabled */
1665 int r = vmballoon_enable_stats(b);
1671 /* format capabilities info */
1672 seq_printf(f, "%-22s: %#16x\n", "balloon capabilities",
1673 VMW_BALLOON_CAPABILITIES);
1674 seq_printf(f, "%-22s: %#16lx\n", "used capabilities", b->capabilities);
1675 seq_printf(f, "%-22s: %16s\n", "is resetting",
1676 b->reset_required ? "y" : "n");
1678 /* format size info */
1679 seq_printf(f, "%-22s: %16lu\n", "target", READ_ONCE(b->target));
1680 seq_printf(f, "%-22s: %16llu\n", "current", atomic64_read(&b->size));
1682 for (i = 0; i < VMW_BALLOON_CMD_NUM; i++) {
1683 if (vmballoon_cmd_names[i] == NULL)
1686 seq_printf(f, "%-22s: %16llu (%llu failed)\n",
1687 vmballoon_cmd_names[i],
1688 atomic64_read(&b->stats->ops[i][VMW_BALLOON_OP_STAT]),
1689 atomic64_read(&b->stats->ops[i][VMW_BALLOON_OP_FAIL_STAT]));
1692 for (i = 0; i < VMW_BALLOON_STAT_NUM; i++)
1693 seq_printf(f, "%-22s: %16llu\n",
1694 vmballoon_stat_names[i],
1695 atomic64_read(&b->stats->general_stat[i]));
1697 for (i = 0; i < VMW_BALLOON_PAGE_STAT_NUM; i++) {
1698 for (j = 0; j < VMW_BALLOON_NUM_PAGE_SIZES; j++)
1699 seq_printf(f, "%-18s(%s): %16llu\n",
1700 vmballoon_stat_page_names[i],
1701 vmballoon_page_size_names[j],
1702 atomic64_read(&b->stats->page_stat[i][j]));
1708 DEFINE_SHOW_ATTRIBUTE(vmballoon_debug);
1710 static void __init vmballoon_debugfs_init(struct vmballoon *b)
1712 b->dbg_entry = debugfs_create_file("vmmemctl", S_IRUGO, NULL, b,
1713 &vmballoon_debug_fops);
1716 static void __exit vmballoon_debugfs_exit(struct vmballoon *b)
1718 static_key_disable(&balloon_stat_enabled.key);
1719 debugfs_remove(b->dbg_entry);
1726 static inline void vmballoon_debugfs_init(struct vmballoon *b)
1730 static inline void vmballoon_debugfs_exit(struct vmballoon *b)
1734 #endif /* CONFIG_DEBUG_FS */
1737 #ifdef CONFIG_BALLOON_COMPACTION
1739 static int vmballoon_init_fs_context(struct fs_context *fc)
1741 return init_pseudo(fc, BALLOON_VMW_MAGIC) ? 0 : -ENOMEM;
1744 static struct file_system_type vmballoon_fs = {
1745 .name = "balloon-vmware",
1746 .init_fs_context = vmballoon_init_fs_context,
1747 .kill_sb = kill_anon_super,
1750 static struct vfsmount *vmballoon_mnt;
1753 * vmballoon_migratepage() - migrates a balloon page.
1754 * @b_dev_info: balloon device information descriptor.
1755 * @newpage: the page to which @page should be migrated.
1756 * @page: a ballooned page that should be migrated.
1757 * @mode: migration mode, ignored.
1759 * This function is really open-coded, but that is according to the interface
1760 * that balloon_compaction provides.
1762 * Return: zero on success, -EAGAIN when migration cannot be performed
1763 * momentarily, and -EBUSY if migration failed and should be retried
1764 * with that specific page.
1766 static int vmballoon_migratepage(struct balloon_dev_info *b_dev_info,
1767 struct page *newpage, struct page *page,
1768 enum migrate_mode mode)
1770 unsigned long status, flags;
1771 struct vmballoon *b;
1774 b = container_of(b_dev_info, struct vmballoon, b_dev_info);
1777 * If the semaphore is taken, there is ongoing configuration change
1778 * (i.e., balloon reset), so try again.
1780 if (!down_read_trylock(&b->conf_sem))
1783 spin_lock(&b->comm_lock);
1785 * We must start by deflating and not inflating, as otherwise the
1786 * hypervisor may tell us that it has enough memory and the new page is
1787 * not needed. Since the old page is isolated, we cannot use the list
1788 * interface to unlock it, as the LRU field is used for isolation.
1789 * Instead, we use the native interface directly.
1791 vmballoon_add_page(b, 0, page);
1792 status = vmballoon_lock_op(b, 1, VMW_BALLOON_4K_PAGE,
1793 VMW_BALLOON_DEFLATE);
1795 if (status == VMW_BALLOON_SUCCESS)
1796 status = vmballoon_status_page(b, 0, &page);
1799 * If a failure happened, let the migration mechanism know that it
1802 if (status != VMW_BALLOON_SUCCESS) {
1803 spin_unlock(&b->comm_lock);
1809 * The page is isolated, so it is safe to delete it without holding
1810 * @pages_lock . We keep holding @comm_lock since we will need it in a
1813 balloon_page_delete(page);
1818 vmballoon_add_page(b, 0, newpage);
1819 status = vmballoon_lock_op(b, 1, VMW_BALLOON_4K_PAGE,
1820 VMW_BALLOON_INFLATE);
1822 if (status == VMW_BALLOON_SUCCESS)
1823 status = vmballoon_status_page(b, 0, &newpage);
1825 spin_unlock(&b->comm_lock);
1827 if (status != VMW_BALLOON_SUCCESS) {
1829 * A failure happened. While we can deflate the page we just
1830 * inflated, this deflation can also encounter an error. Instead
1831 * we will decrease the size of the balloon to reflect the
1832 * change and report failure.
1834 atomic64_dec(&b->size);
1838 * Success. Take a reference for the page, and we will add it to
1839 * the list after acquiring the lock.
1842 ret = MIGRATEPAGE_SUCCESS;
1845 /* Update the balloon list under the @pages_lock */
1846 spin_lock_irqsave(&b->b_dev_info.pages_lock, flags);
1849 * On inflation success, we already took a reference for the @newpage.
1850 * If we succeed just insert it to the list and update the statistics
1853 if (ret == MIGRATEPAGE_SUCCESS) {
1854 balloon_page_insert(&b->b_dev_info, newpage);
1855 __count_vm_event(BALLOON_MIGRATE);
1859 * We deflated successfully, so regardless to the inflation success, we
1860 * need to reduce the number of isolated_pages.
1862 b->b_dev_info.isolated_pages--;
1863 spin_unlock_irqrestore(&b->b_dev_info.pages_lock, flags);
1866 up_read(&b->conf_sem);
1871 * vmballoon_compaction_deinit() - removes compaction related data.
1873 * @b: pointer to the balloon.
1875 static void vmballoon_compaction_deinit(struct vmballoon *b)
1877 if (!IS_ERR(b->b_dev_info.inode))
1878 iput(b->b_dev_info.inode);
1880 b->b_dev_info.inode = NULL;
1881 kern_unmount(vmballoon_mnt);
1882 vmballoon_mnt = NULL;
1886 * vmballoon_compaction_init() - initialized compaction for the balloon.
1888 * @b: pointer to the balloon.
1890 * If during the initialization a failure occurred, this function does not
1891 * perform cleanup. The caller must call vmballoon_compaction_deinit() in this
1894 * Return: zero on success or error code on failure.
1896 static __init int vmballoon_compaction_init(struct vmballoon *b)
1898 vmballoon_mnt = kern_mount(&vmballoon_fs);
1899 if (IS_ERR(vmballoon_mnt))
1900 return PTR_ERR(vmballoon_mnt);
1902 b->b_dev_info.migratepage = vmballoon_migratepage;
1903 b->b_dev_info.inode = alloc_anon_inode(vmballoon_mnt->mnt_sb);
1905 if (IS_ERR(b->b_dev_info.inode))
1906 return PTR_ERR(b->b_dev_info.inode);
1908 b->b_dev_info.inode->i_mapping->a_ops = &balloon_aops;
1912 #else /* CONFIG_BALLOON_COMPACTION */
1914 static void vmballoon_compaction_deinit(struct vmballoon *b)
1918 static int vmballoon_compaction_init(struct vmballoon *b)
1923 #endif /* CONFIG_BALLOON_COMPACTION */
1925 static int __init vmballoon_init(void)
1930 * Check if we are running on VMware's hypervisor and bail out
1933 if (x86_hyper_type != X86_HYPER_VMWARE)
1936 INIT_DELAYED_WORK(&balloon.dwork, vmballoon_work);
1938 error = vmballoon_register_shrinker(&balloon);
1943 * Initialization of compaction must be done after the call to
1944 * balloon_devinfo_init() .
1946 balloon_devinfo_init(&balloon.b_dev_info);
1947 error = vmballoon_compaction_init(&balloon);
1951 INIT_LIST_HEAD(&balloon.huge_pages);
1952 spin_lock_init(&balloon.comm_lock);
1953 init_rwsem(&balloon.conf_sem);
1954 balloon.vmci_doorbell = VMCI_INVALID_HANDLE;
1955 balloon.batch_page = NULL;
1956 balloon.page = NULL;
1957 balloon.reset_required = true;
1959 queue_delayed_work(system_freezable_wq, &balloon.dwork, 0);
1961 vmballoon_debugfs_init(&balloon);
1965 vmballoon_unregister_shrinker(&balloon);
1966 vmballoon_compaction_deinit(&balloon);
1971 * Using late_initcall() instead of module_init() allows the balloon to use the
1972 * VMCI doorbell even when the balloon is built into the kernel. Otherwise the
1973 * VMCI is probed only after the balloon is initialized. If the balloon is used
1974 * as a module, late_initcall() is equivalent to module_init().
1976 late_initcall(vmballoon_init);
1978 static void __exit vmballoon_exit(void)
1980 vmballoon_unregister_shrinker(&balloon);
1981 vmballoon_vmci_cleanup(&balloon);
1982 cancel_delayed_work_sync(&balloon.dwork);
1984 vmballoon_debugfs_exit(&balloon);
1987 * Deallocate all reserved memory, and reset connection with monitor.
1988 * Reset connection before deallocating memory to avoid potential for
1989 * additional spurious resets from guest touching deallocated pages.
1991 vmballoon_send_start(&balloon, 0);
1992 vmballoon_pop(&balloon);
1994 /* Only once we popped the balloon, compaction can be deinit */
1995 vmballoon_compaction_deinit(&balloon);
1997 module_exit(vmballoon_exit);