xref: /openbmc/linux/drivers/misc/vmw_balloon.c (revision be709d48)
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_ALIAS("dmi:*:svnVMware*:*");
38 MODULE_ALIAS("vmware_vmmemctl");
39 MODULE_LICENSE("GPL");
40 
41 /*
42  * Use __GFP_HIGHMEM to allow pages from HIGHMEM zone. We don't allow wait
43  * (__GFP_RECLAIM) for huge page allocations. Use __GFP_NOWARN, to suppress page
44  * allocation failure warnings. Disallow access to emergency low-memory pools.
45  */
46 #define VMW_HUGE_PAGE_ALLOC_FLAGS	(__GFP_HIGHMEM|__GFP_NOWARN|	\
47 					 __GFP_NOMEMALLOC)
48 
49 /*
50  * Use __GFP_HIGHMEM to allow pages from HIGHMEM zone. We allow lightweight
51  * reclamation (__GFP_NORETRY). Use __GFP_NOWARN, to suppress page allocation
52  * failure warnings. Disallow access to emergency low-memory pools.
53  */
54 #define VMW_PAGE_ALLOC_FLAGS		(__GFP_HIGHMEM|__GFP_NOWARN|	\
55 					 __GFP_NOMEMALLOC|__GFP_NORETRY)
56 
57 /* Maximum number of refused pages we accumulate during inflation cycle */
58 #define VMW_BALLOON_MAX_REFUSED		16
59 
60 /*
61  * Hypervisor communication port definitions.
62  */
63 #define VMW_BALLOON_HV_PORT		0x5670
64 #define VMW_BALLOON_HV_MAGIC		0x456c6d6f
65 #define VMW_BALLOON_GUEST_ID		1	/* Linux */
66 
67 enum vmwballoon_capabilities {
68 	/*
69 	 * Bit 0 is reserved and not associated to any capability.
70 	 */
71 	VMW_BALLOON_BASIC_CMDS			= (1 << 1),
72 	VMW_BALLOON_BATCHED_CMDS		= (1 << 2),
73 	VMW_BALLOON_BATCHED_2M_CMDS		= (1 << 3),
74 	VMW_BALLOON_SIGNALLED_WAKEUP_CMD	= (1 << 4),
75 	VMW_BALLOON_64_BIT_TARGET		= (1 << 5)
76 };
77 
78 #define VMW_BALLOON_CAPABILITIES_COMMON	(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 /*
86  * 64-bit targets are only supported in 64-bit
87  */
88 #ifdef CONFIG_64BIT
89 #define VMW_BALLOON_CAPABILITIES	(VMW_BALLOON_CAPABILITIES_COMMON \
90 					| VMW_BALLOON_64_BIT_TARGET)
91 #else
92 #define VMW_BALLOON_CAPABILITIES	VMW_BALLOON_CAPABILITIES_COMMON
93 #endif
94 
95 enum vmballoon_page_size_type {
96 	VMW_BALLOON_4K_PAGE,
97 	VMW_BALLOON_2M_PAGE,
98 	VMW_BALLOON_LAST_SIZE = VMW_BALLOON_2M_PAGE
99 };
100 
101 #define VMW_BALLOON_NUM_PAGE_SIZES	(VMW_BALLOON_LAST_SIZE + 1)
102 
103 static const char * const vmballoon_page_size_names[] = {
104 	[VMW_BALLOON_4K_PAGE]			= "4k",
105 	[VMW_BALLOON_2M_PAGE]			= "2M"
106 };
107 
108 enum vmballoon_op {
109 	VMW_BALLOON_INFLATE,
110 	VMW_BALLOON_DEFLATE
111 };
112 
113 enum vmballoon_op_stat_type {
114 	VMW_BALLOON_OP_STAT,
115 	VMW_BALLOON_OP_FAIL_STAT
116 };
117 
118 #define VMW_BALLOON_OP_STAT_TYPES	(VMW_BALLOON_OP_FAIL_STAT + 1)
119 
120 /**
121  * enum vmballoon_cmd_type - backdoor commands.
122  *
123  * Availability of the commands is as followed:
124  *
125  * %VMW_BALLOON_CMD_START, %VMW_BALLOON_CMD_GET_TARGET and
126  * %VMW_BALLOON_CMD_GUEST_ID are always available.
127  *
128  * If the host reports %VMW_BALLOON_BASIC_CMDS are supported then
129  * %VMW_BALLOON_CMD_LOCK and %VMW_BALLOON_CMD_UNLOCK commands are available.
130  *
131  * If the host reports %VMW_BALLOON_BATCHED_CMDS are supported then
132  * %VMW_BALLOON_CMD_BATCHED_LOCK and VMW_BALLOON_CMD_BATCHED_UNLOCK commands
133  * are available.
134  *
135  * If the host reports %VMW_BALLOON_BATCHED_2M_CMDS are supported then
136  * %VMW_BALLOON_CMD_BATCHED_2M_LOCK and %VMW_BALLOON_CMD_BATCHED_2M_UNLOCK
137  * are supported.
138  *
139  * If the host reports  VMW_BALLOON_SIGNALLED_WAKEUP_CMD is supported then
140  * VMW_BALLOON_CMD_VMCI_DOORBELL_SET command is supported.
141  *
142  * @VMW_BALLOON_CMD_START: Communicating supported version with the hypervisor.
143  * @VMW_BALLOON_CMD_GET_TARGET: Gets the balloon target size.
144  * @VMW_BALLOON_CMD_LOCK: Informs the hypervisor about a ballooned page.
145  * @VMW_BALLOON_CMD_UNLOCK: Informs the hypervisor about a page that is about
146  *			    to be deflated from the balloon.
147  * @VMW_BALLOON_CMD_GUEST_ID: Informs the hypervisor about the type of OS that
148  *			      runs in the VM.
149  * @VMW_BALLOON_CMD_BATCHED_LOCK: Inform the hypervisor about a batch of
150  *				  ballooned pages (up to 512).
151  * @VMW_BALLOON_CMD_BATCHED_UNLOCK: Inform the hypervisor about a batch of
152  *				  pages that are about to be deflated from the
153  *				  balloon (up to 512).
154  * @VMW_BALLOON_CMD_BATCHED_2M_LOCK: Similar to @VMW_BALLOON_CMD_BATCHED_LOCK
155  *				     for 2MB pages.
156  * @VMW_BALLOON_CMD_BATCHED_2M_UNLOCK: Similar to
157  *				       @VMW_BALLOON_CMD_BATCHED_UNLOCK for 2MB
158  *				       pages.
159  * @VMW_BALLOON_CMD_VMCI_DOORBELL_SET: A command to set doorbell notification
160  *				       that would be invoked when the balloon
161  *				       size changes.
162  * @VMW_BALLOON_CMD_LAST: Value of the last command.
163  */
164 enum vmballoon_cmd_type {
165 	VMW_BALLOON_CMD_START,
166 	VMW_BALLOON_CMD_GET_TARGET,
167 	VMW_BALLOON_CMD_LOCK,
168 	VMW_BALLOON_CMD_UNLOCK,
169 	VMW_BALLOON_CMD_GUEST_ID,
170 	/* No command 5 */
171 	VMW_BALLOON_CMD_BATCHED_LOCK = 6,
172 	VMW_BALLOON_CMD_BATCHED_UNLOCK,
173 	VMW_BALLOON_CMD_BATCHED_2M_LOCK,
174 	VMW_BALLOON_CMD_BATCHED_2M_UNLOCK,
175 	VMW_BALLOON_CMD_VMCI_DOORBELL_SET,
176 	VMW_BALLOON_CMD_LAST = VMW_BALLOON_CMD_VMCI_DOORBELL_SET,
177 };
178 
179 #define VMW_BALLOON_CMD_NUM	(VMW_BALLOON_CMD_LAST + 1)
180 
181 enum vmballoon_error_codes {
182 	VMW_BALLOON_SUCCESS,
183 	VMW_BALLOON_ERROR_CMD_INVALID,
184 	VMW_BALLOON_ERROR_PPN_INVALID,
185 	VMW_BALLOON_ERROR_PPN_LOCKED,
186 	VMW_BALLOON_ERROR_PPN_UNLOCKED,
187 	VMW_BALLOON_ERROR_PPN_PINNED,
188 	VMW_BALLOON_ERROR_PPN_NOTNEEDED,
189 	VMW_BALLOON_ERROR_RESET,
190 	VMW_BALLOON_ERROR_BUSY
191 };
192 
193 #define VMW_BALLOON_SUCCESS_WITH_CAPABILITIES	(0x03000000)
194 
195 #define VMW_BALLOON_CMD_WITH_TARGET_MASK			\
196 	((1UL << VMW_BALLOON_CMD_GET_TARGET)		|	\
197 	 (1UL << VMW_BALLOON_CMD_LOCK)			|	\
198 	 (1UL << VMW_BALLOON_CMD_UNLOCK)		|	\
199 	 (1UL << VMW_BALLOON_CMD_BATCHED_LOCK)		|	\
200 	 (1UL << VMW_BALLOON_CMD_BATCHED_UNLOCK)	|	\
201 	 (1UL << VMW_BALLOON_CMD_BATCHED_2M_LOCK)	|	\
202 	 (1UL << VMW_BALLOON_CMD_BATCHED_2M_UNLOCK))
203 
204 static const char * const vmballoon_cmd_names[] = {
205 	[VMW_BALLOON_CMD_START]			= "start",
206 	[VMW_BALLOON_CMD_GET_TARGET]		= "target",
207 	[VMW_BALLOON_CMD_LOCK]			= "lock",
208 	[VMW_BALLOON_CMD_UNLOCK]		= "unlock",
209 	[VMW_BALLOON_CMD_GUEST_ID]		= "guestType",
210 	[VMW_BALLOON_CMD_BATCHED_LOCK]		= "batchLock",
211 	[VMW_BALLOON_CMD_BATCHED_UNLOCK]	= "batchUnlock",
212 	[VMW_BALLOON_CMD_BATCHED_2M_LOCK]	= "2m-lock",
213 	[VMW_BALLOON_CMD_BATCHED_2M_UNLOCK]	= "2m-unlock",
214 	[VMW_BALLOON_CMD_VMCI_DOORBELL_SET]	= "doorbellSet"
215 };
216 
217 enum vmballoon_stat_page {
218 	VMW_BALLOON_PAGE_STAT_ALLOC,
219 	VMW_BALLOON_PAGE_STAT_ALLOC_FAIL,
220 	VMW_BALLOON_PAGE_STAT_REFUSED_ALLOC,
221 	VMW_BALLOON_PAGE_STAT_REFUSED_FREE,
222 	VMW_BALLOON_PAGE_STAT_FREE,
223 	VMW_BALLOON_PAGE_STAT_LAST = VMW_BALLOON_PAGE_STAT_FREE
224 };
225 
226 #define VMW_BALLOON_PAGE_STAT_NUM	(VMW_BALLOON_PAGE_STAT_LAST + 1)
227 
228 enum vmballoon_stat_general {
229 	VMW_BALLOON_STAT_TIMER,
230 	VMW_BALLOON_STAT_DOORBELL,
231 	VMW_BALLOON_STAT_RESET,
232 	VMW_BALLOON_STAT_LAST = VMW_BALLOON_STAT_RESET
233 };
234 
235 #define VMW_BALLOON_STAT_NUM		(VMW_BALLOON_STAT_LAST + 1)
236 
237 
238 static DEFINE_STATIC_KEY_TRUE(vmw_balloon_batching);
239 static DEFINE_STATIC_KEY_FALSE(balloon_stat_enabled);
240 
241 struct vmballoon_ctl {
242 	struct list_head pages;
243 	struct list_head refused_pages;
244 	unsigned int n_refused_pages;
245 	unsigned int n_pages;
246 	enum vmballoon_page_size_type page_size;
247 	enum vmballoon_op op;
248 };
249 
250 struct vmballoon_page_size {
251 	/* list of reserved physical pages */
252 	struct list_head pages;
253 };
254 
255 /**
256  * struct vmballoon_batch_entry - a batch entry for lock or unlock.
257  *
258  * @status: the status of the operation, which is written by the hypervisor.
259  * @reserved: reserved for future use. Must be set to zero.
260  * @pfn: the physical frame number of the page to be locked or unlocked.
261  */
262 struct vmballoon_batch_entry {
263 	u64 status : 5;
264 	u64 reserved : PAGE_SHIFT - 5;
265 	u64 pfn : 52;
266 } __packed;
267 
268 struct vmballoon {
269 	struct vmballoon_page_size page_sizes[VMW_BALLOON_NUM_PAGE_SIZES];
270 
271 	/**
272 	 * @max_page_size: maximum supported page size for ballooning.
273 	 *
274 	 * Protected by @conf_sem
275 	 */
276 	enum vmballoon_page_size_type max_page_size;
277 
278 	/**
279 	 * @size: balloon actual size in basic page size (frames).
280 	 *
281 	 * While we currently do not support size which is bigger than 32-bit,
282 	 * in preparation for future support, use 64-bits.
283 	 */
284 	atomic64_t size;
285 
286 	/**
287 	 * @target: balloon target size in basic page size (frames).
288 	 *
289 	 * We do not protect the target under the assumption that setting the
290 	 * value is always done through a single write. If this assumption ever
291 	 * breaks, we would have to use X_ONCE for accesses, and suffer the less
292 	 * optimized code. Although we may read stale target value if multiple
293 	 * accesses happen at once, the performance impact should be minor.
294 	 */
295 	unsigned long target;
296 
297 	/**
298 	 * @reset_required: reset flag
299 	 *
300 	 * Setting this flag may introduce races, but the code is expected to
301 	 * handle them gracefully. In the worst case, another operation will
302 	 * fail as reset did not take place. Clearing the flag is done while
303 	 * holding @conf_sem for write.
304 	 */
305 	bool reset_required;
306 
307 	/**
308 	 * @capabilities: hypervisor balloon capabilities.
309 	 *
310 	 * Protected by @conf_sem.
311 	 */
312 	unsigned long capabilities;
313 
314 	/**
315 	 * @batch_page: pointer to communication batch page.
316 	 *
317 	 * When batching is used, batch_page points to a page, which holds up to
318 	 * %VMW_BALLOON_BATCH_MAX_PAGES entries for locking or unlocking.
319 	 */
320 	struct vmballoon_batch_entry *batch_page;
321 
322 	/**
323 	 * @batch_max_pages: maximum pages that can be locked/unlocked.
324 	 *
325 	 * Indicates the number of pages that the hypervisor can lock or unlock
326 	 * at once, according to whether batching is enabled. If batching is
327 	 * disabled, only a single page can be locked/unlock on each operation.
328 	 *
329 	 * Protected by @conf_sem.
330 	 */
331 	unsigned int batch_max_pages;
332 
333 	/**
334 	 * @page: page to be locked/unlocked by the hypervisor
335 	 *
336 	 * @page is only used when batching is disabled and a single page is
337 	 * reclaimed on each iteration.
338 	 *
339 	 * Protected by @comm_lock.
340 	 */
341 	struct page *page;
342 
343 	/* statistics */
344 	struct vmballoon_stats *stats;
345 
346 #ifdef CONFIG_DEBUG_FS
347 	/* debugfs file exporting statistics */
348 	struct dentry *dbg_entry;
349 #endif
350 
351 	struct delayed_work dwork;
352 
353 	/**
354 	 * @vmci_doorbell.
355 	 *
356 	 * Protected by @conf_sem.
357 	 */
358 	struct vmci_handle vmci_doorbell;
359 
360 	/**
361 	 * @conf_sem: semaphore to protect the configuration and the statistics.
362 	 */
363 	struct rw_semaphore conf_sem;
364 
365 	/**
366 	 * @comm_lock: lock to protect the communication with the host.
367 	 *
368 	 * Lock ordering: @conf_sem -> @comm_lock .
369 	 */
370 	spinlock_t comm_lock;
371 };
372 
373 static struct vmballoon balloon;
374 
375 struct vmballoon_stats {
376 	/* timer / doorbell operations */
377 	atomic64_t general_stat[VMW_BALLOON_STAT_NUM];
378 
379 	/* allocation statistics for huge and small pages */
380 	atomic64_t
381 	       page_stat[VMW_BALLOON_PAGE_STAT_NUM][VMW_BALLOON_NUM_PAGE_SIZES];
382 
383 	/* Monitor operations: total operations, and failures */
384 	atomic64_t ops[VMW_BALLOON_CMD_NUM][VMW_BALLOON_OP_STAT_TYPES];
385 };
386 
387 static inline bool is_vmballoon_stats_on(void)
388 {
389 	return IS_ENABLED(CONFIG_DEBUG_FS) &&
390 		static_branch_unlikely(&balloon_stat_enabled);
391 }
392 
393 static inline void vmballoon_stats_op_inc(struct vmballoon *b, unsigned int op,
394 					  enum vmballoon_op_stat_type type)
395 {
396 	if (is_vmballoon_stats_on())
397 		atomic64_inc(&b->stats->ops[op][type]);
398 }
399 
400 static inline void vmballoon_stats_gen_inc(struct vmballoon *b,
401 					   enum vmballoon_stat_general stat)
402 {
403 	if (is_vmballoon_stats_on())
404 		atomic64_inc(&b->stats->general_stat[stat]);
405 }
406 
407 static inline void vmballoon_stats_gen_add(struct vmballoon *b,
408 					   enum vmballoon_stat_general stat,
409 					   unsigned int val)
410 {
411 	if (is_vmballoon_stats_on())
412 		atomic64_add(val, &b->stats->general_stat[stat]);
413 }
414 
415 static inline void vmballoon_stats_page_inc(struct vmballoon *b,
416 					    enum vmballoon_stat_page stat,
417 					    enum vmballoon_page_size_type size)
418 {
419 	if (is_vmballoon_stats_on())
420 		atomic64_inc(&b->stats->page_stat[stat][size]);
421 }
422 
423 static inline void vmballoon_stats_page_add(struct vmballoon *b,
424 					    enum vmballoon_stat_page stat,
425 					    enum vmballoon_page_size_type size,
426 					    unsigned int val)
427 {
428 	if (is_vmballoon_stats_on())
429 		atomic64_add(val, &b->stats->page_stat[stat][size]);
430 }
431 
432 static inline unsigned long
433 __vmballoon_cmd(struct vmballoon *b, unsigned long cmd, unsigned long arg1,
434 		unsigned long arg2, unsigned long *result)
435 {
436 	unsigned long status, dummy1, dummy2, dummy3, local_result;
437 
438 	vmballoon_stats_op_inc(b, cmd, VMW_BALLOON_OP_STAT);
439 
440 	asm volatile ("inl %%dx" :
441 		"=a"(status),
442 		"=c"(dummy1),
443 		"=d"(dummy2),
444 		"=b"(local_result),
445 		"=S"(dummy3) :
446 		"0"(VMW_BALLOON_HV_MAGIC),
447 		"1"(cmd),
448 		"2"(VMW_BALLOON_HV_PORT),
449 		"3"(arg1),
450 		"4"(arg2) :
451 		"memory");
452 
453 	/* update the result if needed */
454 	if (result)
455 		*result = (cmd == VMW_BALLOON_CMD_START) ? dummy1 :
456 							   local_result;
457 
458 	/* update target when applicable */
459 	if (status == VMW_BALLOON_SUCCESS &&
460 	    ((1ul << cmd) & VMW_BALLOON_CMD_WITH_TARGET_MASK))
461 		WRITE_ONCE(b->target, local_result);
462 
463 	if (status != VMW_BALLOON_SUCCESS &&
464 	    status != VMW_BALLOON_SUCCESS_WITH_CAPABILITIES) {
465 		vmballoon_stats_op_inc(b, cmd, VMW_BALLOON_OP_FAIL_STAT);
466 		pr_debug("%s: %s [0x%lx,0x%lx) failed, returned %ld\n",
467 			 __func__, vmballoon_cmd_names[cmd], arg1, arg2,
468 			 status);
469 	}
470 
471 	/* mark reset required accordingly */
472 	if (status == VMW_BALLOON_ERROR_RESET)
473 		b->reset_required = true;
474 
475 	return status;
476 }
477 
478 static __always_inline unsigned long
479 vmballoon_cmd(struct vmballoon *b, unsigned long cmd, unsigned long arg1,
480 	      unsigned long arg2)
481 {
482 	unsigned long dummy;
483 
484 	return __vmballoon_cmd(b, cmd, arg1, arg2, &dummy);
485 }
486 
487 /*
488  * Send "start" command to the host, communicating supported version
489  * of the protocol.
490  */
491 static int vmballoon_send_start(struct vmballoon *b, unsigned long req_caps)
492 {
493 	unsigned long status, capabilities;
494 
495 	status = __vmballoon_cmd(b, VMW_BALLOON_CMD_START, req_caps, 0,
496 				 &capabilities);
497 
498 	switch (status) {
499 	case VMW_BALLOON_SUCCESS_WITH_CAPABILITIES:
500 		b->capabilities = capabilities;
501 		break;
502 	case VMW_BALLOON_SUCCESS:
503 		b->capabilities = VMW_BALLOON_BASIC_CMDS;
504 		break;
505 	default:
506 		return -EIO;
507 	}
508 
509 	/*
510 	 * 2MB pages are only supported with batching. If batching is for some
511 	 * reason disabled, do not use 2MB pages, since otherwise the legacy
512 	 * mechanism is used with 2MB pages, causing a failure.
513 	 */
514 	b->max_page_size = VMW_BALLOON_4K_PAGE;
515 	if ((b->capabilities & VMW_BALLOON_BATCHED_2M_CMDS) &&
516 	    (b->capabilities & VMW_BALLOON_BATCHED_CMDS))
517 		b->max_page_size = VMW_BALLOON_2M_PAGE;
518 
519 
520 	return 0;
521 }
522 
523 /**
524  * vmballoon_send_guest_id - communicate guest type to the host.
525  *
526  * @b: pointer to the balloon.
527  *
528  * Communicate guest type to the host so that it can adjust ballooning
529  * algorithm to the one most appropriate for the guest. This command
530  * is normally issued after sending "start" command and is part of
531  * standard reset sequence.
532  *
533  * Return: zero on success or appropriate error code.
534  */
535 static int vmballoon_send_guest_id(struct vmballoon *b)
536 {
537 	unsigned long status;
538 
539 	status = vmballoon_cmd(b, VMW_BALLOON_CMD_GUEST_ID,
540 			       VMW_BALLOON_GUEST_ID, 0);
541 
542 	return status == VMW_BALLOON_SUCCESS ? 0 : -EIO;
543 }
544 
545 /**
546  * vmballoon_page_order() - return the order of the page
547  * @page_size: the size of the page.
548  *
549  * Return: the allocation order.
550  */
551 static inline
552 unsigned int vmballoon_page_order(enum vmballoon_page_size_type page_size)
553 {
554 	return page_size == VMW_BALLOON_2M_PAGE ? VMW_BALLOON_2M_ORDER : 0;
555 }
556 
557 /**
558  * vmballoon_page_in_frames() - returns the number of frames in a page.
559  * @page_size: the size of the page.
560  *
561  * Return: the number of 4k frames.
562  */
563 static inline unsigned int
564 vmballoon_page_in_frames(enum vmballoon_page_size_type page_size)
565 {
566 	return 1 << vmballoon_page_order(page_size);
567 }
568 
569 /**
570  * vmballoon_mark_page_offline() - mark a page as offline
571  * @page: pointer for the page.
572  * @page_size: the size of the page.
573  */
574 static void
575 vmballoon_mark_page_offline(struct page *page,
576 			    enum vmballoon_page_size_type page_size)
577 {
578 	int i;
579 
580 	for (i = 0; i < vmballoon_page_in_frames(page_size); i++)
581 		__SetPageOffline(page + i);
582 }
583 
584 /**
585  * vmballoon_mark_page_online() - mark a page as online
586  * @page: pointer for the page.
587  * @page_size: the size of the page.
588  */
589 static void
590 vmballoon_mark_page_online(struct page *page,
591 			   enum vmballoon_page_size_type page_size)
592 {
593 	int i;
594 
595 	for (i = 0; i < vmballoon_page_in_frames(page_size); i++)
596 		__ClearPageOffline(page + i);
597 }
598 
599 /**
600  * vmballoon_send_get_target() - Retrieve desired balloon size from the host.
601  *
602  * @b: pointer to the balloon.
603  *
604  * Return: zero on success, EINVAL if limit does not fit in 32-bit, as required
605  * by the host-guest protocol and EIO if an error occurred in communicating with
606  * the host.
607  */
608 static int vmballoon_send_get_target(struct vmballoon *b)
609 {
610 	unsigned long status;
611 	unsigned long limit;
612 
613 	limit = totalram_pages();
614 
615 	/* Ensure limit fits in 32-bits if 64-bit targets are not supported */
616 	if (!(b->capabilities & VMW_BALLOON_64_BIT_TARGET) &&
617 	    limit != (u32)limit)
618 		return -EINVAL;
619 
620 	status = vmballoon_cmd(b, VMW_BALLOON_CMD_GET_TARGET, limit, 0);
621 
622 	return status == VMW_BALLOON_SUCCESS ? 0 : -EIO;
623 }
624 
625 /**
626  * vmballoon_alloc_page_list - allocates a list of pages.
627  *
628  * @b: pointer to the balloon.
629  * @ctl: pointer for the %struct vmballoon_ctl, which defines the operation.
630  * @req_n_pages: the number of requested pages.
631  *
632  * Tries to allocate @req_n_pages. Add them to the list of balloon pages in
633  * @ctl.pages and updates @ctl.n_pages to reflect the number of pages.
634  *
635  * Return: zero on success or error code otherwise.
636  */
637 static int vmballoon_alloc_page_list(struct vmballoon *b,
638 				     struct vmballoon_ctl *ctl,
639 				     unsigned int req_n_pages)
640 {
641 	struct page *page;
642 	unsigned int i;
643 
644 	for (i = 0; i < req_n_pages; i++) {
645 		if (ctl->page_size == VMW_BALLOON_2M_PAGE)
646 			page = alloc_pages(VMW_HUGE_PAGE_ALLOC_FLAGS,
647 					   VMW_BALLOON_2M_ORDER);
648 		else
649 			page = alloc_page(VMW_PAGE_ALLOC_FLAGS);
650 
651 		/* Update statistics */
652 		vmballoon_stats_page_inc(b, VMW_BALLOON_PAGE_STAT_ALLOC,
653 					 ctl->page_size);
654 
655 		if (page) {
656 			vmballoon_mark_page_offline(page, ctl->page_size);
657 			/* Success. Add the page to the list and continue. */
658 			list_add(&page->lru, &ctl->pages);
659 			continue;
660 		}
661 
662 		/* Allocation failed. Update statistics and stop. */
663 		vmballoon_stats_page_inc(b, VMW_BALLOON_PAGE_STAT_ALLOC_FAIL,
664 					 ctl->page_size);
665 		break;
666 	}
667 
668 	ctl->n_pages = i;
669 
670 	return req_n_pages == ctl->n_pages ? 0 : -ENOMEM;
671 }
672 
673 /**
674  * vmballoon_handle_one_result - Handle lock/unlock result for a single page.
675  *
676  * @b: pointer for %struct vmballoon.
677  * @page: pointer for the page whose result should be handled.
678  * @page_size: size of the page.
679  * @status: status of the operation as provided by the hypervisor.
680  */
681 static int vmballoon_handle_one_result(struct vmballoon *b, struct page *page,
682 				       enum vmballoon_page_size_type page_size,
683 				       unsigned long status)
684 {
685 	/* On success do nothing. The page is already on the balloon list. */
686 	if (likely(status == VMW_BALLOON_SUCCESS))
687 		return 0;
688 
689 	pr_debug("%s: failed comm pfn %lx status %lu page_size %s\n", __func__,
690 		 page_to_pfn(page), status,
691 		 vmballoon_page_size_names[page_size]);
692 
693 	/* Error occurred */
694 	vmballoon_stats_page_inc(b, VMW_BALLOON_PAGE_STAT_REFUSED_ALLOC,
695 				 page_size);
696 
697 	return -EIO;
698 }
699 
700 /**
701  * vmballoon_status_page - returns the status of (un)lock operation
702  *
703  * @b: pointer to the balloon.
704  * @idx: index for the page for which the operation is performed.
705  * @p: pointer to where the page struct is returned.
706  *
707  * Following a lock or unlock operation, returns the status of the operation for
708  * an individual page. Provides the page that the operation was performed on on
709  * the @page argument.
710  *
711  * Returns: The status of a lock or unlock operation for an individual page.
712  */
713 static unsigned long vmballoon_status_page(struct vmballoon *b, int idx,
714 					   struct page **p)
715 {
716 	if (static_branch_likely(&vmw_balloon_batching)) {
717 		/* batching mode */
718 		*p = pfn_to_page(b->batch_page[idx].pfn);
719 		return b->batch_page[idx].status;
720 	}
721 
722 	/* non-batching mode */
723 	*p = b->page;
724 
725 	/*
726 	 * If a failure occurs, the indication will be provided in the status
727 	 * of the entire operation, which is considered before the individual
728 	 * page status. So for non-batching mode, the indication is always of
729 	 * success.
730 	 */
731 	return VMW_BALLOON_SUCCESS;
732 }
733 
734 /**
735  * vmballoon_lock_op - notifies the host about inflated/deflated pages.
736  * @b: pointer to the balloon.
737  * @num_pages: number of inflated/deflated pages.
738  * @page_size: size of the page.
739  * @op: the type of operation (lock or unlock).
740  *
741  * Notify the host about page(s) that were ballooned (or removed from the
742  * balloon) so that host can use it without fear that guest will need it (or
743  * stop using them since the VM does). Host may reject some pages, we need to
744  * check the return value and maybe submit a different page. The pages that are
745  * inflated/deflated are pointed by @b->page.
746  *
747  * Return: result as provided by the hypervisor.
748  */
749 static unsigned long vmballoon_lock_op(struct vmballoon *b,
750 				       unsigned int num_pages,
751 				       enum vmballoon_page_size_type page_size,
752 				       enum vmballoon_op op)
753 {
754 	unsigned long cmd, pfn;
755 
756 	lockdep_assert_held(&b->comm_lock);
757 
758 	if (static_branch_likely(&vmw_balloon_batching)) {
759 		if (op == VMW_BALLOON_INFLATE)
760 			cmd = page_size == VMW_BALLOON_2M_PAGE ?
761 				VMW_BALLOON_CMD_BATCHED_2M_LOCK :
762 				VMW_BALLOON_CMD_BATCHED_LOCK;
763 		else
764 			cmd = page_size == VMW_BALLOON_2M_PAGE ?
765 				VMW_BALLOON_CMD_BATCHED_2M_UNLOCK :
766 				VMW_BALLOON_CMD_BATCHED_UNLOCK;
767 
768 		pfn = PHYS_PFN(virt_to_phys(b->batch_page));
769 	} else {
770 		cmd = op == VMW_BALLOON_INFLATE ? VMW_BALLOON_CMD_LOCK :
771 						  VMW_BALLOON_CMD_UNLOCK;
772 		pfn = page_to_pfn(b->page);
773 
774 		/* In non-batching mode, PFNs must fit in 32-bit */
775 		if (unlikely(pfn != (u32)pfn))
776 			return VMW_BALLOON_ERROR_PPN_INVALID;
777 	}
778 
779 	return vmballoon_cmd(b, cmd, pfn, num_pages);
780 }
781 
782 /**
783  * vmballoon_add_page - adds a page towards lock/unlock operation.
784  *
785  * @b: pointer to the balloon.
786  * @idx: index of the page to be ballooned in this batch.
787  * @p: pointer to the page that is about to be ballooned.
788  *
789  * Adds the page to be ballooned. Must be called while holding @comm_lock.
790  */
791 static void vmballoon_add_page(struct vmballoon *b, unsigned int idx,
792 			       struct page *p)
793 {
794 	lockdep_assert_held(&b->comm_lock);
795 
796 	if (static_branch_likely(&vmw_balloon_batching))
797 		b->batch_page[idx] = (struct vmballoon_batch_entry)
798 					{ .pfn = page_to_pfn(p) };
799 	else
800 		b->page = p;
801 }
802 
803 /**
804  * vmballoon_lock - lock or unlock a batch of pages.
805  *
806  * @b: pointer to the balloon.
807  * @ctl: pointer for the %struct vmballoon_ctl, which defines the operation.
808  *
809  * Notifies the host of about ballooned pages (after inflation or deflation,
810  * according to @ctl). If the host rejects the page put it on the
811  * @ctl refuse list. These refused page are then released when moving to the
812  * next size of pages.
813  *
814  * Note that we neither free any @page here nor put them back on the ballooned
815  * pages list. Instead we queue it for later processing. We do that for several
816  * reasons. First, we do not want to free the page under the lock. Second, it
817  * allows us to unify the handling of lock and unlock. In the inflate case, the
818  * caller will check if there are too many refused pages and release them.
819  * Although it is not identical to the past behavior, it should not affect
820  * performance.
821  */
822 static int vmballoon_lock(struct vmballoon *b, struct vmballoon_ctl *ctl)
823 {
824 	unsigned long batch_status;
825 	struct page *page;
826 	unsigned int i, num_pages;
827 
828 	num_pages = ctl->n_pages;
829 	if (num_pages == 0)
830 		return 0;
831 
832 	/* communication with the host is done under the communication lock */
833 	spin_lock(&b->comm_lock);
834 
835 	i = 0;
836 	list_for_each_entry(page, &ctl->pages, lru)
837 		vmballoon_add_page(b, i++, page);
838 
839 	batch_status = vmballoon_lock_op(b, ctl->n_pages, ctl->page_size,
840 					 ctl->op);
841 
842 	/*
843 	 * Iterate over the pages in the provided list. Since we are changing
844 	 * @ctl->n_pages we are saving the original value in @num_pages and
845 	 * use this value to bound the loop.
846 	 */
847 	for (i = 0; i < num_pages; i++) {
848 		unsigned long status;
849 
850 		status = vmballoon_status_page(b, i, &page);
851 
852 		/*
853 		 * Failure of the whole batch overrides a single operation
854 		 * results.
855 		 */
856 		if (batch_status != VMW_BALLOON_SUCCESS)
857 			status = batch_status;
858 
859 		/* Continue if no error happened */
860 		if (!vmballoon_handle_one_result(b, page, ctl->page_size,
861 						 status))
862 			continue;
863 
864 		/*
865 		 * Error happened. Move the pages to the refused list and update
866 		 * the pages number.
867 		 */
868 		list_move(&page->lru, &ctl->refused_pages);
869 		ctl->n_pages--;
870 		ctl->n_refused_pages++;
871 	}
872 
873 	spin_unlock(&b->comm_lock);
874 
875 	return batch_status == VMW_BALLOON_SUCCESS ? 0 : -EIO;
876 }
877 
878 /**
879  * vmballoon_release_page_list() - Releases a page list
880  *
881  * @page_list: list of pages to release.
882  * @n_pages: pointer to the number of pages.
883  * @page_size: whether the pages in the list are 2MB (or else 4KB).
884  *
885  * Releases the list of pages and zeros the number of pages.
886  */
887 static void vmballoon_release_page_list(struct list_head *page_list,
888 				       int *n_pages,
889 				       enum vmballoon_page_size_type page_size)
890 {
891 	struct page *page, *tmp;
892 
893 	list_for_each_entry_safe(page, tmp, page_list, lru) {
894 		list_del(&page->lru);
895 		vmballoon_mark_page_online(page, page_size);
896 		__free_pages(page, vmballoon_page_order(page_size));
897 	}
898 
899 	*n_pages = 0;
900 }
901 
902 
903 /*
904  * Release pages that were allocated while attempting to inflate the
905  * balloon but were refused by the host for one reason or another.
906  */
907 static void vmballoon_release_refused_pages(struct vmballoon *b,
908 					    struct vmballoon_ctl *ctl)
909 {
910 	vmballoon_stats_page_inc(b, VMW_BALLOON_PAGE_STAT_REFUSED_FREE,
911 				 ctl->page_size);
912 
913 	vmballoon_release_page_list(&ctl->refused_pages, &ctl->n_refused_pages,
914 				    ctl->page_size);
915 }
916 
917 /**
918  * vmballoon_change - retrieve the required balloon change
919  *
920  * @b: pointer for the balloon.
921  *
922  * Return: the required change for the balloon size. A positive number
923  * indicates inflation, a negative number indicates a deflation.
924  */
925 static int64_t vmballoon_change(struct vmballoon *b)
926 {
927 	int64_t size, target;
928 
929 	size = atomic64_read(&b->size);
930 	target = READ_ONCE(b->target);
931 
932 	/*
933 	 * We must cast first because of int sizes
934 	 * Otherwise we might get huge positives instead of negatives
935 	 */
936 
937 	if (b->reset_required)
938 		return 0;
939 
940 	/* consider a 2MB slack on deflate, unless the balloon is emptied */
941 	if (target < size && target != 0 &&
942 	    size - target < vmballoon_page_in_frames(VMW_BALLOON_2M_PAGE))
943 		return 0;
944 
945 	return target - size;
946 }
947 
948 /**
949  * vmballoon_enqueue_page_list() - Enqueues list of pages after inflation.
950  *
951  * @b: pointer to balloon.
952  * @pages: list of pages to enqueue.
953  * @n_pages: pointer to number of pages in list. The value is zeroed.
954  * @page_size: whether the pages are 2MB or 4KB pages.
955  *
956  * Enqueues the provides list of pages in the ballooned page list, clears the
957  * list and zeroes the number of pages that was provided.
958  */
959 static void vmballoon_enqueue_page_list(struct vmballoon *b,
960 					struct list_head *pages,
961 					unsigned int *n_pages,
962 					enum vmballoon_page_size_type page_size)
963 {
964 	struct vmballoon_page_size *page_size_info = &b->page_sizes[page_size];
965 
966 	list_splice_init(pages, &page_size_info->pages);
967 	*n_pages = 0;
968 }
969 
970 /**
971  * vmballoon_dequeue_page_list() - Dequeues page lists for deflation.
972  *
973  * @b: pointer to balloon.
974  * @pages: list of pages to enqueue.
975  * @n_pages: pointer to number of pages in list. The value is zeroed.
976  * @page_size: whether the pages are 2MB or 4KB pages.
977  * @n_req_pages: the number of requested pages.
978  *
979  * Dequeues the number of requested pages from the balloon for deflation. The
980  * number of dequeued pages may be lower, if not enough pages in the requested
981  * size are available.
982  */
983 static void vmballoon_dequeue_page_list(struct vmballoon *b,
984 					struct list_head *pages,
985 					unsigned int *n_pages,
986 					enum vmballoon_page_size_type page_size,
987 					unsigned int n_req_pages)
988 {
989 	struct vmballoon_page_size *page_size_info = &b->page_sizes[page_size];
990 	struct page *page, *tmp;
991 	unsigned int i = 0;
992 
993 	list_for_each_entry_safe(page, tmp, &page_size_info->pages, lru) {
994 		list_move(&page->lru, pages);
995 		if (++i == n_req_pages)
996 			break;
997 	}
998 	*n_pages = i;
999 }
1000 
1001 /**
1002  * vmballoon_inflate() - Inflate the balloon towards its target size.
1003  *
1004  * @b: pointer to the balloon.
1005  */
1006 static void vmballoon_inflate(struct vmballoon *b)
1007 {
1008 	int64_t to_inflate_frames;
1009 	struct vmballoon_ctl ctl = {
1010 		.pages = LIST_HEAD_INIT(ctl.pages),
1011 		.refused_pages = LIST_HEAD_INIT(ctl.refused_pages),
1012 		.page_size = b->max_page_size,
1013 		.op = VMW_BALLOON_INFLATE
1014 	};
1015 
1016 	while ((to_inflate_frames = vmballoon_change(b)) > 0) {
1017 		unsigned int to_inflate_pages, page_in_frames;
1018 		int alloc_error, lock_error = 0;
1019 
1020 		VM_BUG_ON(!list_empty(&ctl.pages));
1021 		VM_BUG_ON(ctl.n_pages != 0);
1022 
1023 		page_in_frames = vmballoon_page_in_frames(ctl.page_size);
1024 
1025 		to_inflate_pages = min_t(unsigned long, b->batch_max_pages,
1026 					 DIV_ROUND_UP_ULL(to_inflate_frames,
1027 							  page_in_frames));
1028 
1029 		/* Start by allocating */
1030 		alloc_error = vmballoon_alloc_page_list(b, &ctl,
1031 							to_inflate_pages);
1032 
1033 		/* Actually lock the pages by telling the hypervisor */
1034 		lock_error = vmballoon_lock(b, &ctl);
1035 
1036 		/*
1037 		 * If an error indicates that something serious went wrong,
1038 		 * stop the inflation.
1039 		 */
1040 		if (lock_error)
1041 			break;
1042 
1043 		/* Update the balloon size */
1044 		atomic64_add(ctl.n_pages * page_in_frames, &b->size);
1045 
1046 		vmballoon_enqueue_page_list(b, &ctl.pages, &ctl.n_pages,
1047 					    ctl.page_size);
1048 
1049 		/*
1050 		 * If allocation failed or the number of refused pages exceeds
1051 		 * the maximum allowed, move to the next page size.
1052 		 */
1053 		if (alloc_error ||
1054 		    ctl.n_refused_pages >= VMW_BALLOON_MAX_REFUSED) {
1055 			if (ctl.page_size == VMW_BALLOON_4K_PAGE)
1056 				break;
1057 
1058 			/*
1059 			 * Ignore errors from locking as we now switch to 4k
1060 			 * pages and we might get different errors.
1061 			 */
1062 			vmballoon_release_refused_pages(b, &ctl);
1063 			ctl.page_size--;
1064 		}
1065 
1066 		cond_resched();
1067 	}
1068 
1069 	/*
1070 	 * Release pages that were allocated while attempting to inflate the
1071 	 * balloon but were refused by the host for one reason or another,
1072 	 * and update the statistics.
1073 	 */
1074 	if (ctl.n_refused_pages != 0)
1075 		vmballoon_release_refused_pages(b, &ctl);
1076 }
1077 
1078 /**
1079  * vmballoon_deflate() - Decrease the size of the balloon.
1080  *
1081  * @b: pointer to the balloon
1082  * @n_frames: the number of frames to deflate. If zero, automatically
1083  * calculated according to the target size.
1084  * @coordinated: whether to coordinate with the host
1085  *
1086  * Decrease the size of the balloon allowing guest to use more memory.
1087  *
1088  * Return: The number of deflated frames (i.e., basic page size units)
1089  */
1090 static unsigned long vmballoon_deflate(struct vmballoon *b, uint64_t n_frames,
1091 				       bool coordinated)
1092 {
1093 	unsigned long deflated_frames = 0;
1094 	unsigned long tried_frames = 0;
1095 	struct vmballoon_ctl ctl = {
1096 		.pages = LIST_HEAD_INIT(ctl.pages),
1097 		.refused_pages = LIST_HEAD_INIT(ctl.refused_pages),
1098 		.page_size = VMW_BALLOON_4K_PAGE,
1099 		.op = VMW_BALLOON_DEFLATE
1100 	};
1101 
1102 	/* free pages to reach target */
1103 	while (true) {
1104 		unsigned int to_deflate_pages, n_unlocked_frames;
1105 		unsigned int page_in_frames;
1106 		int64_t to_deflate_frames;
1107 		bool deflated_all;
1108 
1109 		page_in_frames = vmballoon_page_in_frames(ctl.page_size);
1110 
1111 		VM_BUG_ON(!list_empty(&ctl.pages));
1112 		VM_BUG_ON(ctl.n_pages);
1113 		VM_BUG_ON(!list_empty(&ctl.refused_pages));
1114 		VM_BUG_ON(ctl.n_refused_pages);
1115 
1116 		/*
1117 		 * If we were requested a specific number of frames, we try to
1118 		 * deflate this number of frames. Otherwise, deflation is
1119 		 * performed according to the target and balloon size.
1120 		 */
1121 		to_deflate_frames = n_frames ? n_frames - tried_frames :
1122 					       -vmballoon_change(b);
1123 
1124 		/* break if no work to do */
1125 		if (to_deflate_frames <= 0)
1126 			break;
1127 
1128 		/*
1129 		 * Calculate the number of frames based on current page size,
1130 		 * but limit the deflated frames to a single chunk
1131 		 */
1132 		to_deflate_pages = min_t(unsigned long, b->batch_max_pages,
1133 					 DIV_ROUND_UP_ULL(to_deflate_frames,
1134 							  page_in_frames));
1135 
1136 		/* First take the pages from the balloon pages. */
1137 		vmballoon_dequeue_page_list(b, &ctl.pages, &ctl.n_pages,
1138 					    ctl.page_size, to_deflate_pages);
1139 
1140 		/*
1141 		 * Before pages are moving to the refused list, count their
1142 		 * frames as frames that we tried to deflate.
1143 		 */
1144 		tried_frames += ctl.n_pages * page_in_frames;
1145 
1146 		/*
1147 		 * Unlock the pages by communicating with the hypervisor if the
1148 		 * communication is coordinated (i.e., not pop). We ignore the
1149 		 * return code. Instead we check if all the pages we manage to
1150 		 * unlock all the pages. If we failed, we will move to the next
1151 		 * page size, and would eventually try again later.
1152 		 */
1153 		if (coordinated)
1154 			vmballoon_lock(b, &ctl);
1155 
1156 		/*
1157 		 * Check if we deflated enough. We will move to the next page
1158 		 * size if we did not manage to do so. This calculation takes
1159 		 * place now, as once the pages are released, the number of
1160 		 * pages is zeroed.
1161 		 */
1162 		deflated_all = (ctl.n_pages == to_deflate_pages);
1163 
1164 		/* Update local and global counters */
1165 		n_unlocked_frames = ctl.n_pages * page_in_frames;
1166 		atomic64_sub(n_unlocked_frames, &b->size);
1167 		deflated_frames += n_unlocked_frames;
1168 
1169 		vmballoon_stats_page_add(b, VMW_BALLOON_PAGE_STAT_FREE,
1170 					 ctl.page_size, ctl.n_pages);
1171 
1172 		/* free the ballooned pages */
1173 		vmballoon_release_page_list(&ctl.pages, &ctl.n_pages,
1174 					    ctl.page_size);
1175 
1176 		/* Return the refused pages to the ballooned list. */
1177 		vmballoon_enqueue_page_list(b, &ctl.refused_pages,
1178 					    &ctl.n_refused_pages,
1179 					    ctl.page_size);
1180 
1181 		/* If we failed to unlock all the pages, move to next size. */
1182 		if (!deflated_all) {
1183 			if (ctl.page_size == b->max_page_size)
1184 				break;
1185 			ctl.page_size++;
1186 		}
1187 
1188 		cond_resched();
1189 	}
1190 
1191 	return deflated_frames;
1192 }
1193 
1194 /**
1195  * vmballoon_deinit_batching - disables batching mode.
1196  *
1197  * @b: pointer to &struct vmballoon.
1198  *
1199  * Disables batching, by deallocating the page for communication with the
1200  * hypervisor and disabling the static key to indicate that batching is off.
1201  */
1202 static void vmballoon_deinit_batching(struct vmballoon *b)
1203 {
1204 	free_page((unsigned long)b->batch_page);
1205 	b->batch_page = NULL;
1206 	static_branch_disable(&vmw_balloon_batching);
1207 	b->batch_max_pages = 1;
1208 }
1209 
1210 /**
1211  * vmballoon_init_batching - enable batching mode.
1212  *
1213  * @b: pointer to &struct vmballoon.
1214  *
1215  * Enables batching, by allocating a page for communication with the hypervisor
1216  * and enabling the static_key to use batching.
1217  *
1218  * Return: zero on success or an appropriate error-code.
1219  */
1220 static int vmballoon_init_batching(struct vmballoon *b)
1221 {
1222 	struct page *page;
1223 
1224 	page = alloc_page(GFP_KERNEL | __GFP_ZERO);
1225 	if (!page)
1226 		return -ENOMEM;
1227 
1228 	b->batch_page = page_address(page);
1229 	b->batch_max_pages = PAGE_SIZE / sizeof(struct vmballoon_batch_entry);
1230 
1231 	static_branch_enable(&vmw_balloon_batching);
1232 
1233 	return 0;
1234 }
1235 
1236 /*
1237  * Receive notification and resize balloon
1238  */
1239 static void vmballoon_doorbell(void *client_data)
1240 {
1241 	struct vmballoon *b = client_data;
1242 
1243 	vmballoon_stats_gen_inc(b, VMW_BALLOON_STAT_DOORBELL);
1244 
1245 	mod_delayed_work(system_freezable_wq, &b->dwork, 0);
1246 }
1247 
1248 /*
1249  * Clean up vmci doorbell
1250  */
1251 static void vmballoon_vmci_cleanup(struct vmballoon *b)
1252 {
1253 	vmballoon_cmd(b, VMW_BALLOON_CMD_VMCI_DOORBELL_SET,
1254 		      VMCI_INVALID_ID, VMCI_INVALID_ID);
1255 
1256 	if (!vmci_handle_is_invalid(b->vmci_doorbell)) {
1257 		vmci_doorbell_destroy(b->vmci_doorbell);
1258 		b->vmci_doorbell = VMCI_INVALID_HANDLE;
1259 	}
1260 }
1261 
1262 /**
1263  * vmballoon_vmci_init - Initialize vmci doorbell.
1264  *
1265  * @b: pointer to the balloon.
1266  *
1267  * Return: zero on success or when wakeup command not supported. Error-code
1268  * otherwise.
1269  *
1270  * Initialize vmci doorbell, to get notified as soon as balloon changes.
1271  */
1272 static int vmballoon_vmci_init(struct vmballoon *b)
1273 {
1274 	unsigned long error;
1275 
1276 	if ((b->capabilities & VMW_BALLOON_SIGNALLED_WAKEUP_CMD) == 0)
1277 		return 0;
1278 
1279 	error = vmci_doorbell_create(&b->vmci_doorbell, VMCI_FLAG_DELAYED_CB,
1280 				     VMCI_PRIVILEGE_FLAG_RESTRICTED,
1281 				     vmballoon_doorbell, b);
1282 
1283 	if (error != VMCI_SUCCESS)
1284 		goto fail;
1285 
1286 	error =	__vmballoon_cmd(b, VMW_BALLOON_CMD_VMCI_DOORBELL_SET,
1287 				b->vmci_doorbell.context,
1288 				b->vmci_doorbell.resource, NULL);
1289 
1290 	if (error != VMW_BALLOON_SUCCESS)
1291 		goto fail;
1292 
1293 	return 0;
1294 fail:
1295 	vmballoon_vmci_cleanup(b);
1296 	return -EIO;
1297 }
1298 
1299 /**
1300  * vmballoon_pop - Quickly release all pages allocate for the balloon.
1301  *
1302  * @b: pointer to the balloon.
1303  *
1304  * This function is called when host decides to "reset" balloon for one reason
1305  * or another. Unlike normal "deflate" we do not (shall not) notify host of the
1306  * pages being released.
1307  */
1308 static void vmballoon_pop(struct vmballoon *b)
1309 {
1310 	unsigned long size;
1311 
1312 	while ((size = atomic64_read(&b->size)))
1313 		vmballoon_deflate(b, size, false);
1314 }
1315 
1316 /*
1317  * Perform standard reset sequence by popping the balloon (in case it
1318  * is not  empty) and then restarting protocol. This operation normally
1319  * happens when host responds with VMW_BALLOON_ERROR_RESET to a command.
1320  */
1321 static void vmballoon_reset(struct vmballoon *b)
1322 {
1323 	int error;
1324 
1325 	down_write(&b->conf_sem);
1326 
1327 	vmballoon_vmci_cleanup(b);
1328 
1329 	/* free all pages, skipping monitor unlock */
1330 	vmballoon_pop(b);
1331 
1332 	if (vmballoon_send_start(b, VMW_BALLOON_CAPABILITIES))
1333 		goto unlock;
1334 
1335 	if ((b->capabilities & VMW_BALLOON_BATCHED_CMDS) != 0) {
1336 		if (vmballoon_init_batching(b)) {
1337 			/*
1338 			 * We failed to initialize batching, inform the monitor
1339 			 * about it by sending a null capability.
1340 			 *
1341 			 * The guest will retry in one second.
1342 			 */
1343 			vmballoon_send_start(b, 0);
1344 			goto unlock;
1345 		}
1346 	} else if ((b->capabilities & VMW_BALLOON_BASIC_CMDS) != 0) {
1347 		vmballoon_deinit_batching(b);
1348 	}
1349 
1350 	vmballoon_stats_gen_inc(b, VMW_BALLOON_STAT_RESET);
1351 	b->reset_required = false;
1352 
1353 	error = vmballoon_vmci_init(b);
1354 	if (error)
1355 		pr_err("failed to initialize vmci doorbell\n");
1356 
1357 	if (vmballoon_send_guest_id(b))
1358 		pr_err("failed to send guest ID to the host\n");
1359 
1360 unlock:
1361 	up_write(&b->conf_sem);
1362 }
1363 
1364 /**
1365  * vmballoon_work - periodic balloon worker for reset, inflation and deflation.
1366  *
1367  * @work: pointer to the &work_struct which is provided by the workqueue.
1368  *
1369  * Resets the protocol if needed, gets the new size and adjusts balloon as
1370  * needed. Repeat in 1 sec.
1371  */
1372 static void vmballoon_work(struct work_struct *work)
1373 {
1374 	struct delayed_work *dwork = to_delayed_work(work);
1375 	struct vmballoon *b = container_of(dwork, struct vmballoon, dwork);
1376 	int64_t change = 0;
1377 
1378 	if (b->reset_required)
1379 		vmballoon_reset(b);
1380 
1381 	down_read(&b->conf_sem);
1382 
1383 	/*
1384 	 * Update the stats while holding the semaphore to ensure that
1385 	 * @stats_enabled is consistent with whether the stats are actually
1386 	 * enabled
1387 	 */
1388 	vmballoon_stats_gen_inc(b, VMW_BALLOON_STAT_TIMER);
1389 
1390 	if (!vmballoon_send_get_target(b))
1391 		change = vmballoon_change(b);
1392 
1393 	if (change != 0) {
1394 		pr_debug("%s - size: %llu, target %lu\n", __func__,
1395 			 atomic64_read(&b->size), READ_ONCE(b->target));
1396 
1397 		if (change > 0)
1398 			vmballoon_inflate(b);
1399 		else  /* (change < 0) */
1400 			vmballoon_deflate(b, 0, true);
1401 	}
1402 
1403 	up_read(&b->conf_sem);
1404 
1405 	/*
1406 	 * We are using a freezable workqueue so that balloon operations are
1407 	 * stopped while the system transitions to/from sleep/hibernation.
1408 	 */
1409 	queue_delayed_work(system_freezable_wq,
1410 			   dwork, round_jiffies_relative(HZ));
1411 
1412 }
1413 
1414 /*
1415  * DEBUGFS Interface
1416  */
1417 #ifdef CONFIG_DEBUG_FS
1418 
1419 static const char * const vmballoon_stat_page_names[] = {
1420 	[VMW_BALLOON_PAGE_STAT_ALLOC]		= "alloc",
1421 	[VMW_BALLOON_PAGE_STAT_ALLOC_FAIL]	= "allocFail",
1422 	[VMW_BALLOON_PAGE_STAT_REFUSED_ALLOC]	= "errAlloc",
1423 	[VMW_BALLOON_PAGE_STAT_REFUSED_FREE]	= "errFree",
1424 	[VMW_BALLOON_PAGE_STAT_FREE]		= "free"
1425 };
1426 
1427 static const char * const vmballoon_stat_names[] = {
1428 	[VMW_BALLOON_STAT_TIMER]		= "timer",
1429 	[VMW_BALLOON_STAT_DOORBELL]		= "doorbell",
1430 	[VMW_BALLOON_STAT_RESET]		= "reset",
1431 };
1432 
1433 static int vmballoon_enable_stats(struct vmballoon *b)
1434 {
1435 	int r = 0;
1436 
1437 	down_write(&b->conf_sem);
1438 
1439 	/* did we somehow race with another reader which enabled stats? */
1440 	if (b->stats)
1441 		goto out;
1442 
1443 	b->stats = kzalloc(sizeof(*b->stats), GFP_KERNEL);
1444 
1445 	if (!b->stats) {
1446 		/* allocation failed */
1447 		r = -ENOMEM;
1448 		goto out;
1449 	}
1450 	static_key_enable(&balloon_stat_enabled.key);
1451 out:
1452 	up_write(&b->conf_sem);
1453 	return r;
1454 }
1455 
1456 /**
1457  * vmballoon_debug_show - shows statistics of balloon operations.
1458  * @f: pointer to the &struct seq_file.
1459  * @offset: ignored.
1460  *
1461  * Provides the statistics that can be accessed in vmmemctl in the debugfs.
1462  * To avoid the overhead - mainly that of memory - of collecting the statistics,
1463  * we only collect statistics after the first time the counters are read.
1464  *
1465  * Return: zero on success or an error code.
1466  */
1467 static int vmballoon_debug_show(struct seq_file *f, void *offset)
1468 {
1469 	struct vmballoon *b = f->private;
1470 	int i, j;
1471 
1472 	/* enables stats if they are disabled */
1473 	if (!b->stats) {
1474 		int r = vmballoon_enable_stats(b);
1475 
1476 		if (r)
1477 			return r;
1478 	}
1479 
1480 	/* format capabilities info */
1481 	seq_printf(f, "%-22s: %#16x\n", "balloon capabilities",
1482 		   VMW_BALLOON_CAPABILITIES);
1483 	seq_printf(f, "%-22s: %#16lx\n", "used capabilities", b->capabilities);
1484 	seq_printf(f, "%-22s: %16s\n", "is resetting",
1485 		   b->reset_required ? "y" : "n");
1486 
1487 	/* format size info */
1488 	seq_printf(f, "%-22s: %16lu\n", "target", READ_ONCE(b->target));
1489 	seq_printf(f, "%-22s: %16llu\n", "current", atomic64_read(&b->size));
1490 
1491 	for (i = 0; i < VMW_BALLOON_CMD_NUM; i++) {
1492 		if (vmballoon_cmd_names[i] == NULL)
1493 			continue;
1494 
1495 		seq_printf(f, "%-22s: %16llu (%llu failed)\n",
1496 			   vmballoon_cmd_names[i],
1497 			   atomic64_read(&b->stats->ops[i][VMW_BALLOON_OP_STAT]),
1498 			   atomic64_read(&b->stats->ops[i][VMW_BALLOON_OP_FAIL_STAT]));
1499 	}
1500 
1501 	for (i = 0; i < VMW_BALLOON_STAT_NUM; i++)
1502 		seq_printf(f, "%-22s: %16llu\n",
1503 			   vmballoon_stat_names[i],
1504 			   atomic64_read(&b->stats->general_stat[i]));
1505 
1506 	for (i = 0; i < VMW_BALLOON_PAGE_STAT_NUM; i++) {
1507 		for (j = 0; j < VMW_BALLOON_NUM_PAGE_SIZES; j++)
1508 			seq_printf(f, "%-18s(%s): %16llu\n",
1509 				   vmballoon_stat_page_names[i],
1510 				   vmballoon_page_size_names[j],
1511 				   atomic64_read(&b->stats->page_stat[i][j]));
1512 	}
1513 
1514 	return 0;
1515 }
1516 
1517 DEFINE_SHOW_ATTRIBUTE(vmballoon_debug);
1518 
1519 static int __init vmballoon_debugfs_init(struct vmballoon *b)
1520 {
1521 	int error;
1522 
1523 	b->dbg_entry = debugfs_create_file("vmmemctl", S_IRUGO, NULL, b,
1524 					   &vmballoon_debug_fops);
1525 	if (IS_ERR(b->dbg_entry)) {
1526 		error = PTR_ERR(b->dbg_entry);
1527 		pr_err("failed to create debugfs entry, error: %d\n", error);
1528 		return error;
1529 	}
1530 
1531 	return 0;
1532 }
1533 
1534 static void __exit vmballoon_debugfs_exit(struct vmballoon *b)
1535 {
1536 	static_key_disable(&balloon_stat_enabled.key);
1537 	debugfs_remove(b->dbg_entry);
1538 	kfree(b->stats);
1539 	b->stats = NULL;
1540 }
1541 
1542 #else
1543 
1544 static inline int vmballoon_debugfs_init(struct vmballoon *b)
1545 {
1546 	return 0;
1547 }
1548 
1549 static inline void vmballoon_debugfs_exit(struct vmballoon *b)
1550 {
1551 }
1552 
1553 #endif	/* CONFIG_DEBUG_FS */
1554 
1555 static int __init vmballoon_init(void)
1556 {
1557 	enum vmballoon_page_size_type page_size;
1558 	int error;
1559 
1560 	/*
1561 	 * Check if we are running on VMware's hypervisor and bail out
1562 	 * if we are not.
1563 	 */
1564 	if (x86_hyper_type != X86_HYPER_VMWARE)
1565 		return -ENODEV;
1566 
1567 	for (page_size = VMW_BALLOON_4K_PAGE;
1568 	     page_size <= VMW_BALLOON_LAST_SIZE; page_size++)
1569 		INIT_LIST_HEAD(&balloon.page_sizes[page_size].pages);
1570 
1571 
1572 	INIT_DELAYED_WORK(&balloon.dwork, vmballoon_work);
1573 
1574 	error = vmballoon_debugfs_init(&balloon);
1575 	if (error)
1576 		return error;
1577 
1578 	spin_lock_init(&balloon.comm_lock);
1579 	init_rwsem(&balloon.conf_sem);
1580 	balloon.vmci_doorbell = VMCI_INVALID_HANDLE;
1581 	balloon.batch_page = NULL;
1582 	balloon.page = NULL;
1583 	balloon.reset_required = true;
1584 
1585 	queue_delayed_work(system_freezable_wq, &balloon.dwork, 0);
1586 
1587 	return 0;
1588 }
1589 
1590 /*
1591  * Using late_initcall() instead of module_init() allows the balloon to use the
1592  * VMCI doorbell even when the balloon is built into the kernel. Otherwise the
1593  * VMCI is probed only after the balloon is initialized. If the balloon is used
1594  * as a module, late_initcall() is equivalent to module_init().
1595  */
1596 late_initcall(vmballoon_init);
1597 
1598 static void __exit vmballoon_exit(void)
1599 {
1600 	vmballoon_vmci_cleanup(&balloon);
1601 	cancel_delayed_work_sync(&balloon.dwork);
1602 
1603 	vmballoon_debugfs_exit(&balloon);
1604 
1605 	/*
1606 	 * Deallocate all reserved memory, and reset connection with monitor.
1607 	 * Reset connection before deallocating memory to avoid potential for
1608 	 * additional spurious resets from guest touching deallocated pages.
1609 	 */
1610 	vmballoon_send_start(&balloon, 0);
1611 	vmballoon_pop(&balloon);
1612 }
1613 module_exit(vmballoon_exit);
1614