xref: /linux/drivers/misc/vmw_balloon.c (revision 544029862cbb1d7903e19f2e58f48d4884e1201b)
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/kernel.h>
21 #include <linux/mm.h>
22 #include <linux/vmalloc.h>
23 #include <linux/sched.h>
24 #include <linux/module.h>
25 #include <linux/workqueue.h>
26 #include <linux/debugfs.h>
27 #include <linux/seq_file.h>
28 #include <linux/rwsem.h>
29 #include <linux/slab.h>
30 #include <linux/spinlock.h>
31 #include <linux/vmw_vmci_defs.h>
32 #include <linux/vmw_vmci_api.h>
33 #include <asm/hypervisor.h>
34 
35 MODULE_AUTHOR("VMware, Inc.");
36 MODULE_DESCRIPTION("VMware Memory Control (Balloon) Driver");
37 MODULE_VERSION("1.5.0.0-k");
38 MODULE_ALIAS("dmi:*:svnVMware*:*");
39 MODULE_ALIAS("vmware_vmmemctl");
40 MODULE_LICENSE("GPL");
41 
42 /*
43  * Use __GFP_HIGHMEM to allow pages from HIGHMEM zone. We don't allow wait
44  * (__GFP_RECLAIM) for huge page allocations. Use __GFP_NOWARN, to suppress page
45  * allocation failure warnings. Disallow access to emergency low-memory pools.
46  */
47 #define VMW_HUGE_PAGE_ALLOC_FLAGS	(__GFP_HIGHMEM|__GFP_NOWARN|	\
48 					 __GFP_NOMEMALLOC)
49 
50 /*
51  * Use __GFP_HIGHMEM to allow pages from HIGHMEM zone. We allow lightweight
52  * reclamation (__GFP_NORETRY). Use __GFP_NOWARN, to suppress page allocation
53  * failure warnings. Disallow access to emergency low-memory pools.
54  */
55 #define VMW_PAGE_ALLOC_FLAGS		(__GFP_HIGHMEM|__GFP_NOWARN|	\
56 					 __GFP_NOMEMALLOC|__GFP_NORETRY)
57 
58 /* Maximum number of refused pages we accumulate during inflation cycle */
59 #define VMW_BALLOON_MAX_REFUSED		16
60 
61 /*
62  * Hypervisor communication port definitions.
63  */
64 #define VMW_BALLOON_HV_PORT		0x5670
65 #define VMW_BALLOON_HV_MAGIC		0x456c6d6f
66 #define VMW_BALLOON_GUEST_ID		1	/* Linux */
67 
68 enum vmwballoon_capabilities {
69 	/*
70 	 * Bit 0 is reserved and not associated to any capability.
71 	 */
72 	VMW_BALLOON_BASIC_CMDS			= (1 << 1),
73 	VMW_BALLOON_BATCHED_CMDS		= (1 << 2),
74 	VMW_BALLOON_BATCHED_2M_CMDS		= (1 << 3),
75 	VMW_BALLOON_SIGNALLED_WAKEUP_CMD	= (1 << 4),
76 };
77 
78 #define VMW_BALLOON_CAPABILITIES	(VMW_BALLOON_BASIC_CMDS \
79 					| VMW_BALLOON_BATCHED_CMDS \
80 					| VMW_BALLOON_BATCHED_2M_CMDS \
81 					| VMW_BALLOON_SIGNALLED_WAKEUP_CMD)
82 
83 #define VMW_BALLOON_2M_ORDER		(PMD_SHIFT - PAGE_SHIFT)
84 
85 enum vmballoon_page_size_type {
86 	VMW_BALLOON_4K_PAGE,
87 	VMW_BALLOON_2M_PAGE,
88 	VMW_BALLOON_LAST_SIZE = VMW_BALLOON_2M_PAGE
89 };
90 
91 #define VMW_BALLOON_NUM_PAGE_SIZES	(VMW_BALLOON_LAST_SIZE + 1)
92 
93 static const char * const vmballoon_page_size_names[] = {
94 	[VMW_BALLOON_4K_PAGE]			= "4k",
95 	[VMW_BALLOON_2M_PAGE]			= "2M"
96 };
97 
98 enum vmballoon_op {
99 	VMW_BALLOON_INFLATE,
100 	VMW_BALLOON_DEFLATE
101 };
102 
103 enum vmballoon_op_stat_type {
104 	VMW_BALLOON_OP_STAT,
105 	VMW_BALLOON_OP_FAIL_STAT
106 };
107 
108 #define VMW_BALLOON_OP_STAT_TYPES	(VMW_BALLOON_OP_FAIL_STAT + 1)
109 
110 /**
111  * enum vmballoon_cmd_type - backdoor commands.
112  *
113  * Availability of the commands is as followed:
114  *
115  * %VMW_BALLOON_CMD_START, %VMW_BALLOON_CMD_GET_TARGET and
116  * %VMW_BALLOON_CMD_GUEST_ID are always available.
117  *
118  * If the host reports %VMW_BALLOON_BASIC_CMDS are supported then
119  * %VMW_BALLOON_CMD_LOCK and %VMW_BALLOON_CMD_UNLOCK commands are available.
120  *
121  * If the host reports %VMW_BALLOON_BATCHED_CMDS are supported then
122  * %VMW_BALLOON_CMD_BATCHED_LOCK and VMW_BALLOON_CMD_BATCHED_UNLOCK commands
123  * are available.
124  *
125  * If the host reports %VMW_BALLOON_BATCHED_2M_CMDS are supported then
126  * %VMW_BALLOON_CMD_BATCHED_2M_LOCK and %VMW_BALLOON_CMD_BATCHED_2M_UNLOCK
127  * are supported.
128  *
129  * If the host reports  VMW_BALLOON_SIGNALLED_WAKEUP_CMD is supported then
130  * VMW_BALLOON_CMD_VMCI_DOORBELL_SET command is supported.
131  *
132  * @VMW_BALLOON_CMD_START: Communicating supported version with the hypervisor.
133  * @VMW_BALLOON_CMD_GET_TARGET: Gets the balloon target size.
134  * @VMW_BALLOON_CMD_LOCK: Informs the hypervisor about a ballooned page.
135  * @VMW_BALLOON_CMD_UNLOCK: Informs the hypervisor about a page that is about
136  *			    to be deflated from the balloon.
137  * @VMW_BALLOON_CMD_GUEST_ID: Informs the hypervisor about the type of OS that
138  *			      runs in the VM.
139  * @VMW_BALLOON_CMD_BATCHED_LOCK: Inform the hypervisor about a batch of
140  *				  ballooned pages (up to 512).
141  * @VMW_BALLOON_CMD_BATCHED_UNLOCK: Inform the hypervisor about a batch of
142  *				  pages that are about to be deflated from the
143  *				  balloon (up to 512).
144  * @VMW_BALLOON_CMD_BATCHED_2M_LOCK: Similar to @VMW_BALLOON_CMD_BATCHED_LOCK
145  *				     for 2MB pages.
146  * @VMW_BALLOON_CMD_BATCHED_2M_UNLOCK: Similar to
147  *				       @VMW_BALLOON_CMD_BATCHED_UNLOCK for 2MB
148  *				       pages.
149  * @VMW_BALLOON_CMD_VMCI_DOORBELL_SET: A command to set doorbell notification
150  *				       that would be invoked when the balloon
151  *				       size changes.
152  * @VMW_BALLOON_CMD_LAST: Value of the last command.
153  */
154 enum vmballoon_cmd_type {
155 	VMW_BALLOON_CMD_START,
156 	VMW_BALLOON_CMD_GET_TARGET,
157 	VMW_BALLOON_CMD_LOCK,
158 	VMW_BALLOON_CMD_UNLOCK,
159 	VMW_BALLOON_CMD_GUEST_ID,
160 	/* No command 5 */
161 	VMW_BALLOON_CMD_BATCHED_LOCK = 6,
162 	VMW_BALLOON_CMD_BATCHED_UNLOCK,
163 	VMW_BALLOON_CMD_BATCHED_2M_LOCK,
164 	VMW_BALLOON_CMD_BATCHED_2M_UNLOCK,
165 	VMW_BALLOON_CMD_VMCI_DOORBELL_SET,
166 	VMW_BALLOON_CMD_LAST = VMW_BALLOON_CMD_VMCI_DOORBELL_SET,
167 };
168 
169 #define VMW_BALLOON_CMD_NUM	(VMW_BALLOON_CMD_LAST + 1)
170 
171 enum vmballoon_error_codes {
172 	VMW_BALLOON_SUCCESS,
173 	VMW_BALLOON_ERROR_CMD_INVALID,
174 	VMW_BALLOON_ERROR_PPN_INVALID,
175 	VMW_BALLOON_ERROR_PPN_LOCKED,
176 	VMW_BALLOON_ERROR_PPN_UNLOCKED,
177 	VMW_BALLOON_ERROR_PPN_PINNED,
178 	VMW_BALLOON_ERROR_PPN_NOTNEEDED,
179 	VMW_BALLOON_ERROR_RESET,
180 	VMW_BALLOON_ERROR_BUSY
181 };
182 
183 #define VMW_BALLOON_SUCCESS_WITH_CAPABILITIES	(0x03000000)
184 
185 #define VMW_BALLOON_CMD_WITH_TARGET_MASK			\
186 	((1UL << VMW_BALLOON_CMD_GET_TARGET)		|	\
187 	 (1UL << VMW_BALLOON_CMD_LOCK)			|	\
188 	 (1UL << VMW_BALLOON_CMD_UNLOCK)		|	\
189 	 (1UL << VMW_BALLOON_CMD_BATCHED_LOCK)		|	\
190 	 (1UL << VMW_BALLOON_CMD_BATCHED_UNLOCK)	|	\
191 	 (1UL << VMW_BALLOON_CMD_BATCHED_2M_LOCK)	|	\
192 	 (1UL << VMW_BALLOON_CMD_BATCHED_2M_UNLOCK))
193 
194 static const char * const vmballoon_cmd_names[] = {
195 	[VMW_BALLOON_CMD_START]			= "start",
196 	[VMW_BALLOON_CMD_GET_TARGET]		= "target",
197 	[VMW_BALLOON_CMD_LOCK]			= "lock",
198 	[VMW_BALLOON_CMD_UNLOCK]		= "unlock",
199 	[VMW_BALLOON_CMD_GUEST_ID]		= "guestType",
200 	[VMW_BALLOON_CMD_BATCHED_LOCK]		= "batchLock",
201 	[VMW_BALLOON_CMD_BATCHED_UNLOCK]	= "batchUnlock",
202 	[VMW_BALLOON_CMD_BATCHED_2M_LOCK]	= "2m-lock",
203 	[VMW_BALLOON_CMD_BATCHED_2M_UNLOCK]	= "2m-unlock",
204 	[VMW_BALLOON_CMD_VMCI_DOORBELL_SET]	= "doorbellSet"
205 };
206 
207 enum vmballoon_stat_page {
208 	VMW_BALLOON_PAGE_STAT_ALLOC,
209 	VMW_BALLOON_PAGE_STAT_ALLOC_FAIL,
210 	VMW_BALLOON_PAGE_STAT_REFUSED_ALLOC,
211 	VMW_BALLOON_PAGE_STAT_REFUSED_FREE,
212 	VMW_BALLOON_PAGE_STAT_FREE,
213 	VMW_BALLOON_PAGE_STAT_LAST = VMW_BALLOON_PAGE_STAT_FREE
214 };
215 
216 #define VMW_BALLOON_PAGE_STAT_NUM	(VMW_BALLOON_PAGE_STAT_LAST + 1)
217 
218 enum vmballoon_stat_general {
219 	VMW_BALLOON_STAT_TIMER,
220 	VMW_BALLOON_STAT_DOORBELL,
221 	VMW_BALLOON_STAT_RESET,
222 	VMW_BALLOON_STAT_LAST = VMW_BALLOON_STAT_RESET
223 };
224 
225 #define VMW_BALLOON_STAT_NUM		(VMW_BALLOON_STAT_LAST + 1)
226 
227 
228 static DEFINE_STATIC_KEY_TRUE(vmw_balloon_batching);
229 static DEFINE_STATIC_KEY_FALSE(balloon_stat_enabled);
230 
231 struct vmballoon_ctl {
232 	struct list_head pages;
233 	struct list_head refused_pages;
234 	unsigned int n_refused_pages;
235 	unsigned int n_pages;
236 	enum vmballoon_page_size_type page_size;
237 	enum vmballoon_op op;
238 };
239 
240 struct vmballoon_page_size {
241 	/* list of reserved physical pages */
242 	struct list_head pages;
243 };
244 
245 /**
246  * struct vmballoon_batch_entry - a batch entry for lock or unlock.
247  *
248  * @status: the status of the operation, which is written by the hypervisor.
249  * @reserved: reserved for future use. Must be set to zero.
250  * @pfn: the physical frame number of the page to be locked or unlocked.
251  */
252 struct vmballoon_batch_entry {
253 	u64 status : 5;
254 	u64 reserved : PAGE_SHIFT - 5;
255 	u64 pfn : 52;
256 } __packed;
257 
258 struct vmballoon {
259 	struct vmballoon_page_size page_sizes[VMW_BALLOON_NUM_PAGE_SIZES];
260 
261 	/**
262 	 * @max_page_size: maximum supported page size for ballooning.
263 	 *
264 	 * Protected by @conf_sem
265 	 */
266 	enum vmballoon_page_size_type max_page_size;
267 
268 	/**
269 	 * @size: balloon actual size in basic page size (frames).
270 	 *
271 	 * While we currently do not support size which is bigger than 32-bit,
272 	 * in preparation for future support, use 64-bits.
273 	 */
274 	atomic64_t size;
275 
276 	/**
277 	 * @target: balloon target size in basic page size (frames).
278 	 *
279 	 * We do not protect the target under the assumption that setting the
280 	 * value is always done through a single write. If this assumption ever
281 	 * breaks, we would have to use X_ONCE for accesses, and suffer the less
282 	 * optimized code. Although we may read stale target value if multiple
283 	 * accesses happen at once, the performance impact should be minor.
284 	 */
285 	unsigned long target;
286 
287 	/**
288 	 * @reset_required: reset flag
289 	 *
290 	 * Setting this flag may introduce races, but the code is expected to
291 	 * handle them gracefully. In the worst case, another operation will
292 	 * fail as reset did not take place. Clearing the flag is done while
293 	 * holding @conf_sem for write.
294 	 */
295 	bool reset_required;
296 
297 	/**
298 	 * @capabilities: hypervisor balloon capabilities.
299 	 *
300 	 * Protected by @conf_sem.
301 	 */
302 	unsigned long capabilities;
303 
304 	/**
305 	 * @batch_page: pointer to communication batch page.
306 	 *
307 	 * When batching is used, batch_page points to a page, which holds up to
308 	 * %VMW_BALLOON_BATCH_MAX_PAGES entries for locking or unlocking.
309 	 */
310 	struct vmballoon_batch_entry *batch_page;
311 
312 	/**
313 	 * @batch_max_pages: maximum pages that can be locked/unlocked.
314 	 *
315 	 * Indicates the number of pages that the hypervisor can lock or unlock
316 	 * at once, according to whether batching is enabled. If batching is
317 	 * disabled, only a single page can be locked/unlock on each operation.
318 	 *
319 	 * Protected by @conf_sem.
320 	 */
321 	unsigned int batch_max_pages;
322 
323 	/**
324 	 * @page: page to be locked/unlocked by the hypervisor
325 	 *
326 	 * @page is only used when batching is disabled and a single page is
327 	 * reclaimed on each iteration.
328 	 *
329 	 * Protected by @comm_lock.
330 	 */
331 	struct page *page;
332 
333 	/* statistics */
334 	struct vmballoon_stats *stats;
335 
336 #ifdef CONFIG_DEBUG_FS
337 	/* debugfs file exporting statistics */
338 	struct dentry *dbg_entry;
339 #endif
340 
341 	struct delayed_work dwork;
342 
343 	/**
344 	 * @vmci_doorbell.
345 	 *
346 	 * Protected by @conf_sem.
347 	 */
348 	struct vmci_handle vmci_doorbell;
349 
350 	/**
351 	 * @conf_sem: semaphore to protect the configuration and the statistics.
352 	 */
353 	struct rw_semaphore conf_sem;
354 
355 	/**
356 	 * @comm_lock: lock to protect the communication with the host.
357 	 *
358 	 * Lock ordering: @conf_sem -> @comm_lock .
359 	 */
360 	spinlock_t comm_lock;
361 };
362 
363 static struct vmballoon balloon;
364 
365 struct vmballoon_stats {
366 	/* timer / doorbell operations */
367 	atomic64_t general_stat[VMW_BALLOON_STAT_NUM];
368 
369 	/* allocation statistics for huge and small pages */
370 	atomic64_t
371 	       page_stat[VMW_BALLOON_PAGE_STAT_NUM][VMW_BALLOON_NUM_PAGE_SIZES];
372 
373 	/* Monitor operations: total operations, and failures */
374 	atomic64_t ops[VMW_BALLOON_CMD_NUM][VMW_BALLOON_OP_STAT_TYPES];
375 };
376 
377 static inline bool is_vmballoon_stats_on(void)
378 {
379 	return IS_ENABLED(CONFIG_DEBUG_FS) &&
380 		static_branch_unlikely(&balloon_stat_enabled);
381 }
382 
383 static inline void vmballoon_stats_op_inc(struct vmballoon *b, unsigned int op,
384 					  enum vmballoon_op_stat_type type)
385 {
386 	if (is_vmballoon_stats_on())
387 		atomic64_inc(&b->stats->ops[op][type]);
388 }
389 
390 static inline void vmballoon_stats_gen_inc(struct vmballoon *b,
391 					   enum vmballoon_stat_general stat)
392 {
393 	if (is_vmballoon_stats_on())
394 		atomic64_inc(&b->stats->general_stat[stat]);
395 }
396 
397 static inline void vmballoon_stats_gen_add(struct vmballoon *b,
398 					   enum vmballoon_stat_general stat,
399 					   unsigned int val)
400 {
401 	if (is_vmballoon_stats_on())
402 		atomic64_add(val, &b->stats->general_stat[stat]);
403 }
404 
405 static inline void vmballoon_stats_page_inc(struct vmballoon *b,
406 					    enum vmballoon_stat_page stat,
407 					    enum vmballoon_page_size_type size)
408 {
409 	if (is_vmballoon_stats_on())
410 		atomic64_inc(&b->stats->page_stat[stat][size]);
411 }
412 
413 static inline void vmballoon_stats_page_add(struct vmballoon *b,
414 					    enum vmballoon_stat_page stat,
415 					    enum vmballoon_page_size_type size,
416 					    unsigned int val)
417 {
418 	if (is_vmballoon_stats_on())
419 		atomic64_add(val, &b->stats->page_stat[stat][size]);
420 }
421 
422 static inline unsigned long
423 __vmballoon_cmd(struct vmballoon *b, unsigned long cmd, unsigned long arg1,
424 		unsigned long arg2, unsigned long *result)
425 {
426 	unsigned long status, dummy1, dummy2, dummy3, local_result;
427 
428 	vmballoon_stats_op_inc(b, cmd, VMW_BALLOON_OP_STAT);
429 
430 	asm volatile ("inl %%dx" :
431 		"=a"(status),
432 		"=c"(dummy1),
433 		"=d"(dummy2),
434 		"=b"(local_result),
435 		"=S"(dummy3) :
436 		"0"(VMW_BALLOON_HV_MAGIC),
437 		"1"(cmd),
438 		"2"(VMW_BALLOON_HV_PORT),
439 		"3"(arg1),
440 		"4"(arg2) :
441 		"memory");
442 
443 	/* update the result if needed */
444 	if (result)
445 		*result = (cmd == VMW_BALLOON_CMD_START) ? dummy1 :
446 							   local_result;
447 
448 	/* update target when applicable */
449 	if (status == VMW_BALLOON_SUCCESS &&
450 	    ((1ul << cmd) & VMW_BALLOON_CMD_WITH_TARGET_MASK))
451 		WRITE_ONCE(b->target, local_result);
452 
453 	if (status != VMW_BALLOON_SUCCESS &&
454 	    status != VMW_BALLOON_SUCCESS_WITH_CAPABILITIES) {
455 		vmballoon_stats_op_inc(b, cmd, VMW_BALLOON_OP_FAIL_STAT);
456 		pr_debug("%s: %s [0x%lx,0x%lx) failed, returned %ld\n",
457 			 __func__, vmballoon_cmd_names[cmd], arg1, arg2,
458 			 status);
459 	}
460 
461 	/* mark reset required accordingly */
462 	if (status == VMW_BALLOON_ERROR_RESET)
463 		b->reset_required = true;
464 
465 	return status;
466 }
467 
468 static __always_inline unsigned long
469 vmballoon_cmd(struct vmballoon *b, unsigned long cmd, unsigned long arg1,
470 	      unsigned long arg2)
471 {
472 	unsigned long dummy;
473 
474 	return __vmballoon_cmd(b, cmd, arg1, arg2, &dummy);
475 }
476 
477 /*
478  * Send "start" command to the host, communicating supported version
479  * of the protocol.
480  */
481 static int vmballoon_send_start(struct vmballoon *b, unsigned long req_caps)
482 {
483 	unsigned long status, capabilities;
484 
485 	status = __vmballoon_cmd(b, VMW_BALLOON_CMD_START, req_caps, 0,
486 				 &capabilities);
487 
488 	switch (status) {
489 	case VMW_BALLOON_SUCCESS_WITH_CAPABILITIES:
490 		b->capabilities = capabilities;
491 		break;
492 	case VMW_BALLOON_SUCCESS:
493 		b->capabilities = VMW_BALLOON_BASIC_CMDS;
494 		break;
495 	default:
496 		return -EIO;
497 	}
498 
499 	/*
500 	 * 2MB pages are only supported with batching. If batching is for some
501 	 * reason disabled, do not use 2MB pages, since otherwise the legacy
502 	 * mechanism is used with 2MB pages, causing a failure.
503 	 */
504 	b->max_page_size = VMW_BALLOON_4K_PAGE;
505 	if ((b->capabilities & VMW_BALLOON_BATCHED_2M_CMDS) &&
506 	    (b->capabilities & VMW_BALLOON_BATCHED_CMDS))
507 		b->max_page_size = VMW_BALLOON_2M_PAGE;
508 
509 
510 	return 0;
511 }
512 
513 /**
514  * vmballoon_send_guest_id - communicate guest type to the host.
515  *
516  * @b: pointer to the balloon.
517  *
518  * Communicate guest type to the host so that it can adjust ballooning
519  * algorithm to the one most appropriate for the guest. This command
520  * is normally issued after sending "start" command and is part of
521  * standard reset sequence.
522  *
523  * Return: zero on success or appropriate error code.
524  */
525 static int vmballoon_send_guest_id(struct vmballoon *b)
526 {
527 	unsigned long status;
528 
529 	status = vmballoon_cmd(b, VMW_BALLOON_CMD_GUEST_ID,
530 			       VMW_BALLOON_GUEST_ID, 0);
531 
532 	return status == VMW_BALLOON_SUCCESS ? 0 : -EIO;
533 }
534 
535 /**
536  * vmballoon_page_order() - return the order of the page
537  * @page_size: the size of the page.
538  *
539  * Return: the allocation order.
540  */
541 static inline
542 unsigned int vmballoon_page_order(enum vmballoon_page_size_type page_size)
543 {
544 	return page_size == VMW_BALLOON_2M_PAGE ? VMW_BALLOON_2M_ORDER : 0;
545 }
546 
547 /**
548  * vmballoon_page_in_frames() - returns the number of frames in a page.
549  * @page_size: the size of the page.
550  *
551  * Return: the number of 4k frames.
552  */
553 static inline unsigned int
554 vmballoon_page_in_frames(enum vmballoon_page_size_type page_size)
555 {
556 	return 1 << vmballoon_page_order(page_size);
557 }
558 
559 /**
560  * vmballoon_mark_page_offline() - mark a page as offline
561  * @page: pointer for the page.
562  * @page_size: the size of the page.
563  */
564 static void
565 vmballoon_mark_page_offline(struct page *page,
566 			    enum vmballoon_page_size_type page_size)
567 {
568 	int i;
569 
570 	for (i = 0; i < vmballoon_page_in_frames(page_size); i++)
571 		__SetPageOffline(page + i);
572 }
573 
574 /**
575  * vmballoon_mark_page_online() - mark a page as online
576  * @page: pointer for the page.
577  * @page_size: the size of the page.
578  */
579 static void
580 vmballoon_mark_page_online(struct page *page,
581 			   enum vmballoon_page_size_type page_size)
582 {
583 	int i;
584 
585 	for (i = 0; i < vmballoon_page_in_frames(page_size); i++)
586 		__ClearPageOffline(page + i);
587 }
588 
589 /**
590  * vmballoon_send_get_target() - Retrieve desired balloon size from the host.
591  *
592  * @b: pointer to the balloon.
593  *
594  * Return: zero on success, EINVAL if limit does not fit in 32-bit, as required
595  * by the host-guest protocol and EIO if an error occurred in communicating with
596  * the host.
597  */
598 static int vmballoon_send_get_target(struct vmballoon *b)
599 {
600 	unsigned long status;
601 	unsigned long limit;
602 
603 	limit = totalram_pages();
604 
605 	/* Ensure limit fits in 32-bits */
606 	if (limit != (u32)limit)
607 		return -EINVAL;
608 
609 	status = vmballoon_cmd(b, VMW_BALLOON_CMD_GET_TARGET, limit, 0);
610 
611 	return status == VMW_BALLOON_SUCCESS ? 0 : -EIO;
612 }
613 
614 /**
615  * vmballoon_alloc_page_list - allocates a list of pages.
616  *
617  * @b: pointer to the balloon.
618  * @ctl: pointer for the %struct vmballoon_ctl, which defines the operation.
619  * @req_n_pages: the number of requested pages.
620  *
621  * Tries to allocate @req_n_pages. Add them to the list of balloon pages in
622  * @ctl.pages and updates @ctl.n_pages to reflect the number of pages.
623  *
624  * Return: zero on success or error code otherwise.
625  */
626 static int vmballoon_alloc_page_list(struct vmballoon *b,
627 				     struct vmballoon_ctl *ctl,
628 				     unsigned int req_n_pages)
629 {
630 	struct page *page;
631 	unsigned int i;
632 
633 	for (i = 0; i < req_n_pages; i++) {
634 		if (ctl->page_size == VMW_BALLOON_2M_PAGE)
635 			page = alloc_pages(VMW_HUGE_PAGE_ALLOC_FLAGS,
636 					   VMW_BALLOON_2M_ORDER);
637 		else
638 			page = alloc_page(VMW_PAGE_ALLOC_FLAGS);
639 
640 		/* Update statistics */
641 		vmballoon_stats_page_inc(b, VMW_BALLOON_PAGE_STAT_ALLOC,
642 					 ctl->page_size);
643 
644 		if (page) {
645 			vmballoon_mark_page_offline(page, ctl->page_size);
646 			/* Success. Add the page to the list and continue. */
647 			list_add(&page->lru, &ctl->pages);
648 			continue;
649 		}
650 
651 		/* Allocation failed. Update statistics and stop. */
652 		vmballoon_stats_page_inc(b, VMW_BALLOON_PAGE_STAT_ALLOC_FAIL,
653 					 ctl->page_size);
654 		break;
655 	}
656 
657 	ctl->n_pages = i;
658 
659 	return req_n_pages == ctl->n_pages ? 0 : -ENOMEM;
660 }
661 
662 /**
663  * vmballoon_handle_one_result - Handle lock/unlock result for a single page.
664  *
665  * @b: pointer for %struct vmballoon.
666  * @page: pointer for the page whose result should be handled.
667  * @page_size: size of the page.
668  * @status: status of the operation as provided by the hypervisor.
669  */
670 static int vmballoon_handle_one_result(struct vmballoon *b, struct page *page,
671 				       enum vmballoon_page_size_type page_size,
672 				       unsigned long status)
673 {
674 	/* On success do nothing. The page is already on the balloon list. */
675 	if (likely(status == VMW_BALLOON_SUCCESS))
676 		return 0;
677 
678 	pr_debug("%s: failed comm pfn %lx status %lu page_size %s\n", __func__,
679 		 page_to_pfn(page), status,
680 		 vmballoon_page_size_names[page_size]);
681 
682 	/* Error occurred */
683 	vmballoon_stats_page_inc(b, VMW_BALLOON_PAGE_STAT_REFUSED_ALLOC,
684 				 page_size);
685 
686 	return -EIO;
687 }
688 
689 /**
690  * vmballoon_status_page - returns the status of (un)lock operation
691  *
692  * @b: pointer to the balloon.
693  * @idx: index for the page for which the operation is performed.
694  * @p: pointer to where the page struct is returned.
695  *
696  * Following a lock or unlock operation, returns the status of the operation for
697  * an individual page. Provides the page that the operation was performed on on
698  * the @page argument.
699  *
700  * Returns: The status of a lock or unlock operation for an individual page.
701  */
702 static unsigned long vmballoon_status_page(struct vmballoon *b, int idx,
703 					   struct page **p)
704 {
705 	if (static_branch_likely(&vmw_balloon_batching)) {
706 		/* batching mode */
707 		*p = pfn_to_page(b->batch_page[idx].pfn);
708 		return b->batch_page[idx].status;
709 	}
710 
711 	/* non-batching mode */
712 	*p = b->page;
713 
714 	/*
715 	 * If a failure occurs, the indication will be provided in the status
716 	 * of the entire operation, which is considered before the individual
717 	 * page status. So for non-batching mode, the indication is always of
718 	 * success.
719 	 */
720 	return VMW_BALLOON_SUCCESS;
721 }
722 
723 /**
724  * vmballoon_lock_op - notifies the host about inflated/deflated pages.
725  * @b: pointer to the balloon.
726  * @num_pages: number of inflated/deflated pages.
727  * @page_size: size of the page.
728  * @op: the type of operation (lock or unlock).
729  *
730  * Notify the host about page(s) that were ballooned (or removed from the
731  * balloon) so that host can use it without fear that guest will need it (or
732  * stop using them since the VM does). Host may reject some pages, we need to
733  * check the return value and maybe submit a different page. The pages that are
734  * inflated/deflated are pointed by @b->page.
735  *
736  * Return: result as provided by the hypervisor.
737  */
738 static unsigned long vmballoon_lock_op(struct vmballoon *b,
739 				       unsigned int num_pages,
740 				       enum vmballoon_page_size_type page_size,
741 				       enum vmballoon_op op)
742 {
743 	unsigned long cmd, pfn;
744 
745 	lockdep_assert_held(&b->comm_lock);
746 
747 	if (static_branch_likely(&vmw_balloon_batching)) {
748 		if (op == VMW_BALLOON_INFLATE)
749 			cmd = page_size == VMW_BALLOON_2M_PAGE ?
750 				VMW_BALLOON_CMD_BATCHED_2M_LOCK :
751 				VMW_BALLOON_CMD_BATCHED_LOCK;
752 		else
753 			cmd = page_size == VMW_BALLOON_2M_PAGE ?
754 				VMW_BALLOON_CMD_BATCHED_2M_UNLOCK :
755 				VMW_BALLOON_CMD_BATCHED_UNLOCK;
756 
757 		pfn = PHYS_PFN(virt_to_phys(b->batch_page));
758 	} else {
759 		cmd = op == VMW_BALLOON_INFLATE ? VMW_BALLOON_CMD_LOCK :
760 						  VMW_BALLOON_CMD_UNLOCK;
761 		pfn = page_to_pfn(b->page);
762 
763 		/* In non-batching mode, PFNs must fit in 32-bit */
764 		if (unlikely(pfn != (u32)pfn))
765 			return VMW_BALLOON_ERROR_PPN_INVALID;
766 	}
767 
768 	return vmballoon_cmd(b, cmd, pfn, num_pages);
769 }
770 
771 /**
772  * vmballoon_add_page - adds a page towards lock/unlock operation.
773  *
774  * @b: pointer to the balloon.
775  * @idx: index of the page to be ballooned in this batch.
776  * @p: pointer to the page that is about to be ballooned.
777  *
778  * Adds the page to be ballooned. Must be called while holding @comm_lock.
779  */
780 static void vmballoon_add_page(struct vmballoon *b, unsigned int idx,
781 			       struct page *p)
782 {
783 	lockdep_assert_held(&b->comm_lock);
784 
785 	if (static_branch_likely(&vmw_balloon_batching))
786 		b->batch_page[idx] = (struct vmballoon_batch_entry)
787 					{ .pfn = page_to_pfn(p) };
788 	else
789 		b->page = p;
790 }
791 
792 /**
793  * vmballoon_lock - lock or unlock a batch of pages.
794  *
795  * @b: pointer to the balloon.
796  * @ctl: pointer for the %struct vmballoon_ctl, which defines the operation.
797  *
798  * Notifies the host of about ballooned pages (after inflation or deflation,
799  * according to @ctl). If the host rejects the page put it on the
800  * @ctl refuse list. These refused page are then released when moving to the
801  * next size of pages.
802  *
803  * Note that we neither free any @page here nor put them back on the ballooned
804  * pages list. Instead we queue it for later processing. We do that for several
805  * reasons. First, we do not want to free the page under the lock. Second, it
806  * allows us to unify the handling of lock and unlock. In the inflate case, the
807  * caller will check if there are too many refused pages and release them.
808  * Although it is not identical to the past behavior, it should not affect
809  * performance.
810  */
811 static int vmballoon_lock(struct vmballoon *b, struct vmballoon_ctl *ctl)
812 {
813 	unsigned long batch_status;
814 	struct page *page;
815 	unsigned int i, num_pages;
816 
817 	num_pages = ctl->n_pages;
818 	if (num_pages == 0)
819 		return 0;
820 
821 	/* communication with the host is done under the communication lock */
822 	spin_lock(&b->comm_lock);
823 
824 	i = 0;
825 	list_for_each_entry(page, &ctl->pages, lru)
826 		vmballoon_add_page(b, i++, page);
827 
828 	batch_status = vmballoon_lock_op(b, ctl->n_pages, ctl->page_size,
829 					 ctl->op);
830 
831 	/*
832 	 * Iterate over the pages in the provided list. Since we are changing
833 	 * @ctl->n_pages we are saving the original value in @num_pages and
834 	 * use this value to bound the loop.
835 	 */
836 	for (i = 0; i < num_pages; i++) {
837 		unsigned long status;
838 
839 		status = vmballoon_status_page(b, i, &page);
840 
841 		/*
842 		 * Failure of the whole batch overrides a single operation
843 		 * results.
844 		 */
845 		if (batch_status != VMW_BALLOON_SUCCESS)
846 			status = batch_status;
847 
848 		/* Continue if no error happened */
849 		if (!vmballoon_handle_one_result(b, page, ctl->page_size,
850 						 status))
851 			continue;
852 
853 		/*
854 		 * Error happened. Move the pages to the refused list and update
855 		 * the pages number.
856 		 */
857 		list_move(&page->lru, &ctl->refused_pages);
858 		ctl->n_pages--;
859 		ctl->n_refused_pages++;
860 	}
861 
862 	spin_unlock(&b->comm_lock);
863 
864 	return batch_status == VMW_BALLOON_SUCCESS ? 0 : -EIO;
865 }
866 
867 /**
868  * vmballoon_release_page_list() - Releases a page list
869  *
870  * @page_list: list of pages to release.
871  * @n_pages: pointer to the number of pages.
872  * @page_size: whether the pages in the list are 2MB (or else 4KB).
873  *
874  * Releases the list of pages and zeros the number of pages.
875  */
876 static void vmballoon_release_page_list(struct list_head *page_list,
877 				       int *n_pages,
878 				       enum vmballoon_page_size_type page_size)
879 {
880 	struct page *page, *tmp;
881 
882 	list_for_each_entry_safe(page, tmp, page_list, lru) {
883 		list_del(&page->lru);
884 		vmballoon_mark_page_online(page, page_size);
885 		__free_pages(page, vmballoon_page_order(page_size));
886 	}
887 
888 	*n_pages = 0;
889 }
890 
891 
892 /*
893  * Release pages that were allocated while attempting to inflate the
894  * balloon but were refused by the host for one reason or another.
895  */
896 static void vmballoon_release_refused_pages(struct vmballoon *b,
897 					    struct vmballoon_ctl *ctl)
898 {
899 	vmballoon_stats_page_inc(b, VMW_BALLOON_PAGE_STAT_REFUSED_FREE,
900 				 ctl->page_size);
901 
902 	vmballoon_release_page_list(&ctl->refused_pages, &ctl->n_refused_pages,
903 				    ctl->page_size);
904 }
905 
906 /**
907  * vmballoon_change - retrieve the required balloon change
908  *
909  * @b: pointer for the balloon.
910  *
911  * Return: the required change for the balloon size. A positive number
912  * indicates inflation, a negative number indicates a deflation.
913  */
914 static int64_t vmballoon_change(struct vmballoon *b)
915 {
916 	int64_t size, target;
917 
918 	size = atomic64_read(&b->size);
919 	target = READ_ONCE(b->target);
920 
921 	/*
922 	 * We must cast first because of int sizes
923 	 * Otherwise we might get huge positives instead of negatives
924 	 */
925 
926 	if (b->reset_required)
927 		return 0;
928 
929 	/* consider a 2MB slack on deflate, unless the balloon is emptied */
930 	if (target < size && target != 0 &&
931 	    size - target < vmballoon_page_in_frames(VMW_BALLOON_2M_PAGE))
932 		return 0;
933 
934 	return target - size;
935 }
936 
937 /**
938  * vmballoon_enqueue_page_list() - Enqueues list of pages after inflation.
939  *
940  * @b: pointer to balloon.
941  * @pages: list of pages to enqueue.
942  * @n_pages: pointer to number of pages in list. The value is zeroed.
943  * @page_size: whether the pages are 2MB or 4KB pages.
944  *
945  * Enqueues the provides list of pages in the ballooned page list, clears the
946  * list and zeroes the number of pages that was provided.
947  */
948 static void vmballoon_enqueue_page_list(struct vmballoon *b,
949 					struct list_head *pages,
950 					unsigned int *n_pages,
951 					enum vmballoon_page_size_type page_size)
952 {
953 	struct vmballoon_page_size *page_size_info = &b->page_sizes[page_size];
954 
955 	list_splice_init(pages, &page_size_info->pages);
956 	*n_pages = 0;
957 }
958 
959 /**
960  * vmballoon_dequeue_page_list() - Dequeues page lists for deflation.
961  *
962  * @b: pointer to balloon.
963  * @pages: list of pages to enqueue.
964  * @n_pages: pointer to number of pages in list. The value is zeroed.
965  * @page_size: whether the pages are 2MB or 4KB pages.
966  * @n_req_pages: the number of requested pages.
967  *
968  * Dequeues the number of requested pages from the balloon for deflation. The
969  * number of dequeued pages may be lower, if not enough pages in the requested
970  * size are available.
971  */
972 static void vmballoon_dequeue_page_list(struct vmballoon *b,
973 					struct list_head *pages,
974 					unsigned int *n_pages,
975 					enum vmballoon_page_size_type page_size,
976 					unsigned int n_req_pages)
977 {
978 	struct vmballoon_page_size *page_size_info = &b->page_sizes[page_size];
979 	struct page *page, *tmp;
980 	unsigned int i = 0;
981 
982 	list_for_each_entry_safe(page, tmp, &page_size_info->pages, lru) {
983 		list_move(&page->lru, pages);
984 		if (++i == n_req_pages)
985 			break;
986 	}
987 	*n_pages = i;
988 }
989 
990 /**
991  * vmballoon_inflate() - Inflate the balloon towards its target size.
992  *
993  * @b: pointer to the balloon.
994  */
995 static void vmballoon_inflate(struct vmballoon *b)
996 {
997 	int64_t to_inflate_frames;
998 	struct vmballoon_ctl ctl = {
999 		.pages = LIST_HEAD_INIT(ctl.pages),
1000 		.refused_pages = LIST_HEAD_INIT(ctl.refused_pages),
1001 		.page_size = b->max_page_size,
1002 		.op = VMW_BALLOON_INFLATE
1003 	};
1004 
1005 	while ((to_inflate_frames = vmballoon_change(b)) > 0) {
1006 		unsigned int to_inflate_pages, page_in_frames;
1007 		int alloc_error, lock_error = 0;
1008 
1009 		VM_BUG_ON(!list_empty(&ctl.pages));
1010 		VM_BUG_ON(ctl.n_pages != 0);
1011 
1012 		page_in_frames = vmballoon_page_in_frames(ctl.page_size);
1013 
1014 		to_inflate_pages = min_t(unsigned long, b->batch_max_pages,
1015 					 DIV_ROUND_UP_ULL(to_inflate_frames,
1016 							  page_in_frames));
1017 
1018 		/* Start by allocating */
1019 		alloc_error = vmballoon_alloc_page_list(b, &ctl,
1020 							to_inflate_pages);
1021 
1022 		/* Actually lock the pages by telling the hypervisor */
1023 		lock_error = vmballoon_lock(b, &ctl);
1024 
1025 		/*
1026 		 * If an error indicates that something serious went wrong,
1027 		 * stop the inflation.
1028 		 */
1029 		if (lock_error)
1030 			break;
1031 
1032 		/* Update the balloon size */
1033 		atomic64_add(ctl.n_pages * page_in_frames, &b->size);
1034 
1035 		vmballoon_enqueue_page_list(b, &ctl.pages, &ctl.n_pages,
1036 					    ctl.page_size);
1037 
1038 		/*
1039 		 * If allocation failed or the number of refused pages exceeds
1040 		 * the maximum allowed, move to the next page size.
1041 		 */
1042 		if (alloc_error ||
1043 		    ctl.n_refused_pages >= VMW_BALLOON_MAX_REFUSED) {
1044 			if (ctl.page_size == VMW_BALLOON_4K_PAGE)
1045 				break;
1046 
1047 			/*
1048 			 * Ignore errors from locking as we now switch to 4k
1049 			 * pages and we might get different errors.
1050 			 */
1051 			vmballoon_release_refused_pages(b, &ctl);
1052 			ctl.page_size--;
1053 		}
1054 
1055 		cond_resched();
1056 	}
1057 
1058 	/*
1059 	 * Release pages that were allocated while attempting to inflate the
1060 	 * balloon but were refused by the host for one reason or another,
1061 	 * and update the statistics.
1062 	 */
1063 	if (ctl.n_refused_pages != 0)
1064 		vmballoon_release_refused_pages(b, &ctl);
1065 }
1066 
1067 /**
1068  * vmballoon_deflate() - Decrease the size of the balloon.
1069  *
1070  * @b: pointer to the balloon
1071  * @n_frames: the number of frames to deflate. If zero, automatically
1072  * calculated according to the target size.
1073  * @coordinated: whether to coordinate with the host
1074  *
1075  * Decrease the size of the balloon allowing guest to use more memory.
1076  *
1077  * Return: The number of deflated frames (i.e., basic page size units)
1078  */
1079 static unsigned long vmballoon_deflate(struct vmballoon *b, uint64_t n_frames,
1080 				       bool coordinated)
1081 {
1082 	unsigned long deflated_frames = 0;
1083 	unsigned long tried_frames = 0;
1084 	struct vmballoon_ctl ctl = {
1085 		.pages = LIST_HEAD_INIT(ctl.pages),
1086 		.refused_pages = LIST_HEAD_INIT(ctl.refused_pages),
1087 		.page_size = VMW_BALLOON_4K_PAGE,
1088 		.op = VMW_BALLOON_DEFLATE
1089 	};
1090 
1091 	/* free pages to reach target */
1092 	while (true) {
1093 		unsigned int to_deflate_pages, n_unlocked_frames;
1094 		unsigned int page_in_frames;
1095 		int64_t to_deflate_frames;
1096 		bool deflated_all;
1097 
1098 		page_in_frames = vmballoon_page_in_frames(ctl.page_size);
1099 
1100 		VM_BUG_ON(!list_empty(&ctl.pages));
1101 		VM_BUG_ON(ctl.n_pages);
1102 		VM_BUG_ON(!list_empty(&ctl.refused_pages));
1103 		VM_BUG_ON(ctl.n_refused_pages);
1104 
1105 		/*
1106 		 * If we were requested a specific number of frames, we try to
1107 		 * deflate this number of frames. Otherwise, deflation is
1108 		 * performed according to the target and balloon size.
1109 		 */
1110 		to_deflate_frames = n_frames ? n_frames - tried_frames :
1111 					       -vmballoon_change(b);
1112 
1113 		/* break if no work to do */
1114 		if (to_deflate_frames <= 0)
1115 			break;
1116 
1117 		/*
1118 		 * Calculate the number of frames based on current page size,
1119 		 * but limit the deflated frames to a single chunk
1120 		 */
1121 		to_deflate_pages = min_t(unsigned long, b->batch_max_pages,
1122 					 DIV_ROUND_UP_ULL(to_deflate_frames,
1123 							  page_in_frames));
1124 
1125 		/* First take the pages from the balloon pages. */
1126 		vmballoon_dequeue_page_list(b, &ctl.pages, &ctl.n_pages,
1127 					    ctl.page_size, to_deflate_pages);
1128 
1129 		/*
1130 		 * Before pages are moving to the refused list, count their
1131 		 * frames as frames that we tried to deflate.
1132 		 */
1133 		tried_frames += ctl.n_pages * page_in_frames;
1134 
1135 		/*
1136 		 * Unlock the pages by communicating with the hypervisor if the
1137 		 * communication is coordinated (i.e., not pop). We ignore the
1138 		 * return code. Instead we check if all the pages we manage to
1139 		 * unlock all the pages. If we failed, we will move to the next
1140 		 * page size, and would eventually try again later.
1141 		 */
1142 		if (coordinated)
1143 			vmballoon_lock(b, &ctl);
1144 
1145 		/*
1146 		 * Check if we deflated enough. We will move to the next page
1147 		 * size if we did not manage to do so. This calculation takes
1148 		 * place now, as once the pages are released, the number of
1149 		 * pages is zeroed.
1150 		 */
1151 		deflated_all = (ctl.n_pages == to_deflate_pages);
1152 
1153 		/* Update local and global counters */
1154 		n_unlocked_frames = ctl.n_pages * page_in_frames;
1155 		atomic64_sub(n_unlocked_frames, &b->size);
1156 		deflated_frames += n_unlocked_frames;
1157 
1158 		vmballoon_stats_page_add(b, VMW_BALLOON_PAGE_STAT_FREE,
1159 					 ctl.page_size, ctl.n_pages);
1160 
1161 		/* free the ballooned pages */
1162 		vmballoon_release_page_list(&ctl.pages, &ctl.n_pages,
1163 					    ctl.page_size);
1164 
1165 		/* Return the refused pages to the ballooned list. */
1166 		vmballoon_enqueue_page_list(b, &ctl.refused_pages,
1167 					    &ctl.n_refused_pages,
1168 					    ctl.page_size);
1169 
1170 		/* If we failed to unlock all the pages, move to next size. */
1171 		if (!deflated_all) {
1172 			if (ctl.page_size == b->max_page_size)
1173 				break;
1174 			ctl.page_size++;
1175 		}
1176 
1177 		cond_resched();
1178 	}
1179 
1180 	return deflated_frames;
1181 }
1182 
1183 /**
1184  * vmballoon_deinit_batching - disables batching mode.
1185  *
1186  * @b: pointer to &struct vmballoon.
1187  *
1188  * Disables batching, by deallocating the page for communication with the
1189  * hypervisor and disabling the static key to indicate that batching is off.
1190  */
1191 static void vmballoon_deinit_batching(struct vmballoon *b)
1192 {
1193 	free_page((unsigned long)b->batch_page);
1194 	b->batch_page = NULL;
1195 	static_branch_disable(&vmw_balloon_batching);
1196 	b->batch_max_pages = 1;
1197 }
1198 
1199 /**
1200  * vmballoon_init_batching - enable batching mode.
1201  *
1202  * @b: pointer to &struct vmballoon.
1203  *
1204  * Enables batching, by allocating a page for communication with the hypervisor
1205  * and enabling the static_key to use batching.
1206  *
1207  * Return: zero on success or an appropriate error-code.
1208  */
1209 static int vmballoon_init_batching(struct vmballoon *b)
1210 {
1211 	struct page *page;
1212 
1213 	page = alloc_page(GFP_KERNEL | __GFP_ZERO);
1214 	if (!page)
1215 		return -ENOMEM;
1216 
1217 	b->batch_page = page_address(page);
1218 	b->batch_max_pages = PAGE_SIZE / sizeof(struct vmballoon_batch_entry);
1219 
1220 	static_branch_enable(&vmw_balloon_batching);
1221 
1222 	return 0;
1223 }
1224 
1225 /*
1226  * Receive notification and resize balloon
1227  */
1228 static void vmballoon_doorbell(void *client_data)
1229 {
1230 	struct vmballoon *b = client_data;
1231 
1232 	vmballoon_stats_gen_inc(b, VMW_BALLOON_STAT_DOORBELL);
1233 
1234 	mod_delayed_work(system_freezable_wq, &b->dwork, 0);
1235 }
1236 
1237 /*
1238  * Clean up vmci doorbell
1239  */
1240 static void vmballoon_vmci_cleanup(struct vmballoon *b)
1241 {
1242 	vmballoon_cmd(b, VMW_BALLOON_CMD_VMCI_DOORBELL_SET,
1243 		      VMCI_INVALID_ID, VMCI_INVALID_ID);
1244 
1245 	if (!vmci_handle_is_invalid(b->vmci_doorbell)) {
1246 		vmci_doorbell_destroy(b->vmci_doorbell);
1247 		b->vmci_doorbell = VMCI_INVALID_HANDLE;
1248 	}
1249 }
1250 
1251 /**
1252  * vmballoon_vmci_init - Initialize vmci doorbell.
1253  *
1254  * @b: pointer to the balloon.
1255  *
1256  * Return: zero on success or when wakeup command not supported. Error-code
1257  * otherwise.
1258  *
1259  * Initialize vmci doorbell, to get notified as soon as balloon changes.
1260  */
1261 static int vmballoon_vmci_init(struct vmballoon *b)
1262 {
1263 	unsigned long error;
1264 
1265 	if ((b->capabilities & VMW_BALLOON_SIGNALLED_WAKEUP_CMD) == 0)
1266 		return 0;
1267 
1268 	error = vmci_doorbell_create(&b->vmci_doorbell, VMCI_FLAG_DELAYED_CB,
1269 				     VMCI_PRIVILEGE_FLAG_RESTRICTED,
1270 				     vmballoon_doorbell, b);
1271 
1272 	if (error != VMCI_SUCCESS)
1273 		goto fail;
1274 
1275 	error =	__vmballoon_cmd(b, VMW_BALLOON_CMD_VMCI_DOORBELL_SET,
1276 				b->vmci_doorbell.context,
1277 				b->vmci_doorbell.resource, NULL);
1278 
1279 	if (error != VMW_BALLOON_SUCCESS)
1280 		goto fail;
1281 
1282 	return 0;
1283 fail:
1284 	vmballoon_vmci_cleanup(b);
1285 	return -EIO;
1286 }
1287 
1288 /**
1289  * vmballoon_pop - Quickly release all pages allocate for the balloon.
1290  *
1291  * @b: pointer to the balloon.
1292  *
1293  * This function is called when host decides to "reset" balloon for one reason
1294  * or another. Unlike normal "deflate" we do not (shall not) notify host of the
1295  * pages being released.
1296  */
1297 static void vmballoon_pop(struct vmballoon *b)
1298 {
1299 	unsigned long size;
1300 
1301 	while ((size = atomic64_read(&b->size)))
1302 		vmballoon_deflate(b, size, false);
1303 }
1304 
1305 /*
1306  * Perform standard reset sequence by popping the balloon (in case it
1307  * is not  empty) and then restarting protocol. This operation normally
1308  * happens when host responds with VMW_BALLOON_ERROR_RESET to a command.
1309  */
1310 static void vmballoon_reset(struct vmballoon *b)
1311 {
1312 	int error;
1313 
1314 	down_write(&b->conf_sem);
1315 
1316 	vmballoon_vmci_cleanup(b);
1317 
1318 	/* free all pages, skipping monitor unlock */
1319 	vmballoon_pop(b);
1320 
1321 	if (vmballoon_send_start(b, VMW_BALLOON_CAPABILITIES))
1322 		return;
1323 
1324 	if ((b->capabilities & VMW_BALLOON_BATCHED_CMDS) != 0) {
1325 		if (vmballoon_init_batching(b)) {
1326 			/*
1327 			 * We failed to initialize batching, inform the monitor
1328 			 * about it by sending a null capability.
1329 			 *
1330 			 * The guest will retry in one second.
1331 			 */
1332 			vmballoon_send_start(b, 0);
1333 			return;
1334 		}
1335 	} else if ((b->capabilities & VMW_BALLOON_BASIC_CMDS) != 0) {
1336 		vmballoon_deinit_batching(b);
1337 	}
1338 
1339 	vmballoon_stats_gen_inc(b, VMW_BALLOON_STAT_RESET);
1340 	b->reset_required = false;
1341 
1342 	error = vmballoon_vmci_init(b);
1343 	if (error)
1344 		pr_err("failed to initialize vmci doorbell\n");
1345 
1346 	if (vmballoon_send_guest_id(b))
1347 		pr_err("failed to send guest ID to the host\n");
1348 
1349 	up_write(&b->conf_sem);
1350 }
1351 
1352 /**
1353  * vmballoon_work - periodic balloon worker for reset, inflation and deflation.
1354  *
1355  * @work: pointer to the &work_struct which is provided by the workqueue.
1356  *
1357  * Resets the protocol if needed, gets the new size and adjusts balloon as
1358  * needed. Repeat in 1 sec.
1359  */
1360 static void vmballoon_work(struct work_struct *work)
1361 {
1362 	struct delayed_work *dwork = to_delayed_work(work);
1363 	struct vmballoon *b = container_of(dwork, struct vmballoon, dwork);
1364 	int64_t change = 0;
1365 
1366 	if (b->reset_required)
1367 		vmballoon_reset(b);
1368 
1369 	down_read(&b->conf_sem);
1370 
1371 	/*
1372 	 * Update the stats while holding the semaphore to ensure that
1373 	 * @stats_enabled is consistent with whether the stats are actually
1374 	 * enabled
1375 	 */
1376 	vmballoon_stats_gen_inc(b, VMW_BALLOON_STAT_TIMER);
1377 
1378 	if (!vmballoon_send_get_target(b))
1379 		change = vmballoon_change(b);
1380 
1381 	if (change != 0) {
1382 		pr_debug("%s - size: %llu, target %lu\n", __func__,
1383 			 atomic64_read(&b->size), READ_ONCE(b->target));
1384 
1385 		if (change > 0)
1386 			vmballoon_inflate(b);
1387 		else  /* (change < 0) */
1388 			vmballoon_deflate(b, 0, true);
1389 	}
1390 
1391 	up_read(&b->conf_sem);
1392 
1393 	/*
1394 	 * We are using a freezable workqueue so that balloon operations are
1395 	 * stopped while the system transitions to/from sleep/hibernation.
1396 	 */
1397 	queue_delayed_work(system_freezable_wq,
1398 			   dwork, round_jiffies_relative(HZ));
1399 
1400 }
1401 
1402 /*
1403  * DEBUGFS Interface
1404  */
1405 #ifdef CONFIG_DEBUG_FS
1406 
1407 static const char * const vmballoon_stat_page_names[] = {
1408 	[VMW_BALLOON_PAGE_STAT_ALLOC]		= "alloc",
1409 	[VMW_BALLOON_PAGE_STAT_ALLOC_FAIL]	= "allocFail",
1410 	[VMW_BALLOON_PAGE_STAT_REFUSED_ALLOC]	= "errAlloc",
1411 	[VMW_BALLOON_PAGE_STAT_REFUSED_FREE]	= "errFree",
1412 	[VMW_BALLOON_PAGE_STAT_FREE]		= "free"
1413 };
1414 
1415 static const char * const vmballoon_stat_names[] = {
1416 	[VMW_BALLOON_STAT_TIMER]		= "timer",
1417 	[VMW_BALLOON_STAT_DOORBELL]		= "doorbell",
1418 	[VMW_BALLOON_STAT_RESET]		= "reset",
1419 };
1420 
1421 static int vmballoon_enable_stats(struct vmballoon *b)
1422 {
1423 	int r = 0;
1424 
1425 	down_write(&b->conf_sem);
1426 
1427 	/* did we somehow race with another reader which enabled stats? */
1428 	if (b->stats)
1429 		goto out;
1430 
1431 	b->stats = kzalloc(sizeof(*b->stats), GFP_KERNEL);
1432 
1433 	if (!b->stats) {
1434 		/* allocation failed */
1435 		r = -ENOMEM;
1436 		goto out;
1437 	}
1438 	static_key_enable(&balloon_stat_enabled.key);
1439 out:
1440 	up_write(&b->conf_sem);
1441 	return r;
1442 }
1443 
1444 /**
1445  * vmballoon_debug_show - shows statistics of balloon operations.
1446  * @f: pointer to the &struct seq_file.
1447  * @offset: ignored.
1448  *
1449  * Provides the statistics that can be accessed in vmmemctl in the debugfs.
1450  * To avoid the overhead - mainly that of memory - of collecting the statistics,
1451  * we only collect statistics after the first time the counters are read.
1452  *
1453  * Return: zero on success or an error code.
1454  */
1455 static int vmballoon_debug_show(struct seq_file *f, void *offset)
1456 {
1457 	struct vmballoon *b = f->private;
1458 	int i, j;
1459 
1460 	/* enables stats if they are disabled */
1461 	if (!b->stats) {
1462 		int r = vmballoon_enable_stats(b);
1463 
1464 		if (r)
1465 			return r;
1466 	}
1467 
1468 	/* format capabilities info */
1469 	seq_printf(f, "%-22s: %#16x\n", "balloon capabilities",
1470 		   VMW_BALLOON_CAPABILITIES);
1471 	seq_printf(f, "%-22s: %#16lx\n", "used capabilities", b->capabilities);
1472 	seq_printf(f, "%-22s: %16s\n", "is resetting",
1473 		   b->reset_required ? "y" : "n");
1474 
1475 	/* format size info */
1476 	seq_printf(f, "%-22s: %16lu\n", "target", READ_ONCE(b->target));
1477 	seq_printf(f, "%-22s: %16llu\n", "current", atomic64_read(&b->size));
1478 
1479 	for (i = 0; i < VMW_BALLOON_CMD_NUM; i++) {
1480 		if (vmballoon_cmd_names[i] == NULL)
1481 			continue;
1482 
1483 		seq_printf(f, "%-22s: %16llu (%llu failed)\n",
1484 			   vmballoon_cmd_names[i],
1485 			   atomic64_read(&b->stats->ops[i][VMW_BALLOON_OP_STAT]),
1486 			   atomic64_read(&b->stats->ops[i][VMW_BALLOON_OP_FAIL_STAT]));
1487 	}
1488 
1489 	for (i = 0; i < VMW_BALLOON_STAT_NUM; i++)
1490 		seq_printf(f, "%-22s: %16llu\n",
1491 			   vmballoon_stat_names[i],
1492 			   atomic64_read(&b->stats->general_stat[i]));
1493 
1494 	for (i = 0; i < VMW_BALLOON_PAGE_STAT_NUM; i++) {
1495 		for (j = 0; j < VMW_BALLOON_NUM_PAGE_SIZES; j++)
1496 			seq_printf(f, "%-18s(%s): %16llu\n",
1497 				   vmballoon_stat_page_names[i],
1498 				   vmballoon_page_size_names[j],
1499 				   atomic64_read(&b->stats->page_stat[i][j]));
1500 	}
1501 
1502 	return 0;
1503 }
1504 
1505 DEFINE_SHOW_ATTRIBUTE(vmballoon_debug);
1506 
1507 static int __init vmballoon_debugfs_init(struct vmballoon *b)
1508 {
1509 	int error;
1510 
1511 	b->dbg_entry = debugfs_create_file("vmmemctl", S_IRUGO, NULL, b,
1512 					   &vmballoon_debug_fops);
1513 	if (IS_ERR(b->dbg_entry)) {
1514 		error = PTR_ERR(b->dbg_entry);
1515 		pr_err("failed to create debugfs entry, error: %d\n", error);
1516 		return error;
1517 	}
1518 
1519 	return 0;
1520 }
1521 
1522 static void __exit vmballoon_debugfs_exit(struct vmballoon *b)
1523 {
1524 	static_key_disable(&balloon_stat_enabled.key);
1525 	debugfs_remove(b->dbg_entry);
1526 	kfree(b->stats);
1527 	b->stats = NULL;
1528 }
1529 
1530 #else
1531 
1532 static inline int vmballoon_debugfs_init(struct vmballoon *b)
1533 {
1534 	return 0;
1535 }
1536 
1537 static inline void vmballoon_debugfs_exit(struct vmballoon *b)
1538 {
1539 }
1540 
1541 #endif	/* CONFIG_DEBUG_FS */
1542 
1543 static int __init vmballoon_init(void)
1544 {
1545 	enum vmballoon_page_size_type page_size;
1546 	int error;
1547 
1548 	/*
1549 	 * Check if we are running on VMware's hypervisor and bail out
1550 	 * if we are not.
1551 	 */
1552 	if (x86_hyper_type != X86_HYPER_VMWARE)
1553 		return -ENODEV;
1554 
1555 	for (page_size = VMW_BALLOON_4K_PAGE;
1556 	     page_size <= VMW_BALLOON_LAST_SIZE; page_size++)
1557 		INIT_LIST_HEAD(&balloon.page_sizes[page_size].pages);
1558 
1559 
1560 	INIT_DELAYED_WORK(&balloon.dwork, vmballoon_work);
1561 
1562 	error = vmballoon_debugfs_init(&balloon);
1563 	if (error)
1564 		return error;
1565 
1566 	spin_lock_init(&balloon.comm_lock);
1567 	init_rwsem(&balloon.conf_sem);
1568 	balloon.vmci_doorbell = VMCI_INVALID_HANDLE;
1569 	balloon.batch_page = NULL;
1570 	balloon.page = NULL;
1571 	balloon.reset_required = true;
1572 
1573 	queue_delayed_work(system_freezable_wq, &balloon.dwork, 0);
1574 
1575 	return 0;
1576 }
1577 
1578 /*
1579  * Using late_initcall() instead of module_init() allows the balloon to use the
1580  * VMCI doorbell even when the balloon is built into the kernel. Otherwise the
1581  * VMCI is probed only after the balloon is initialized. If the balloon is used
1582  * as a module, late_initcall() is equivalent to module_init().
1583  */
1584 late_initcall(vmballoon_init);
1585 
1586 static void __exit vmballoon_exit(void)
1587 {
1588 	vmballoon_vmci_cleanup(&balloon);
1589 	cancel_delayed_work_sync(&balloon.dwork);
1590 
1591 	vmballoon_debugfs_exit(&balloon);
1592 
1593 	/*
1594 	 * Deallocate all reserved memory, and reset connection with monitor.
1595 	 * Reset connection before deallocating memory to avoid potential for
1596 	 * additional spurious resets from guest touching deallocated pages.
1597 	 */
1598 	vmballoon_send_start(&balloon, 0);
1599 	vmballoon_pop(&balloon);
1600 }
1601 module_exit(vmballoon_exit);
1602