xref: /openbmc/linux/drivers/misc/vmw_balloon.c (revision ccb01374)
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_send_get_target() - Retrieve desired balloon size from the host.
561  *
562  * @b: pointer to the balloon.
563  *
564  * Return: zero on success, EINVAL if limit does not fit in 32-bit, as required
565  * by the host-guest protocol and EIO if an error occurred in communicating with
566  * the host.
567  */
568 static int vmballoon_send_get_target(struct vmballoon *b)
569 {
570 	unsigned long status;
571 	unsigned long limit;
572 
573 	limit = totalram_pages();
574 
575 	/* Ensure limit fits in 32-bits */
576 	if (limit != (u32)limit)
577 		return -EINVAL;
578 
579 	status = vmballoon_cmd(b, VMW_BALLOON_CMD_GET_TARGET, limit, 0);
580 
581 	return status == VMW_BALLOON_SUCCESS ? 0 : -EIO;
582 }
583 
584 /**
585  * vmballoon_alloc_page_list - allocates a list of pages.
586  *
587  * @b: pointer to the balloon.
588  * @ctl: pointer for the %struct vmballoon_ctl, which defines the operation.
589  * @req_n_pages: the number of requested pages.
590  *
591  * Tries to allocate @req_n_pages. Add them to the list of balloon pages in
592  * @ctl.pages and updates @ctl.n_pages to reflect the number of pages.
593  *
594  * Return: zero on success or error code otherwise.
595  */
596 static int vmballoon_alloc_page_list(struct vmballoon *b,
597 				     struct vmballoon_ctl *ctl,
598 				     unsigned int req_n_pages)
599 {
600 	struct page *page;
601 	unsigned int i;
602 
603 	for (i = 0; i < req_n_pages; i++) {
604 		if (ctl->page_size == VMW_BALLOON_2M_PAGE)
605 			page = alloc_pages(VMW_HUGE_PAGE_ALLOC_FLAGS,
606 					   VMW_BALLOON_2M_ORDER);
607 		else
608 			page = alloc_page(VMW_PAGE_ALLOC_FLAGS);
609 
610 		/* Update statistics */
611 		vmballoon_stats_page_inc(b, VMW_BALLOON_PAGE_STAT_ALLOC,
612 					 ctl->page_size);
613 
614 		if (page) {
615 			/* Success. Add the page to the list and continue. */
616 			list_add(&page->lru, &ctl->pages);
617 			continue;
618 		}
619 
620 		/* Allocation failed. Update statistics and stop. */
621 		vmballoon_stats_page_inc(b, VMW_BALLOON_PAGE_STAT_ALLOC_FAIL,
622 					 ctl->page_size);
623 		break;
624 	}
625 
626 	ctl->n_pages = i;
627 
628 	return req_n_pages == ctl->n_pages ? 0 : -ENOMEM;
629 }
630 
631 /**
632  * vmballoon_handle_one_result - Handle lock/unlock result for a single page.
633  *
634  * @b: pointer for %struct vmballoon.
635  * @page: pointer for the page whose result should be handled.
636  * @page_size: size of the page.
637  * @status: status of the operation as provided by the hypervisor.
638  */
639 static int vmballoon_handle_one_result(struct vmballoon *b, struct page *page,
640 				       enum vmballoon_page_size_type page_size,
641 				       unsigned long status)
642 {
643 	/* On success do nothing. The page is already on the balloon list. */
644 	if (likely(status == VMW_BALLOON_SUCCESS))
645 		return 0;
646 
647 	pr_debug("%s: failed comm pfn %lx status %lu page_size %s\n", __func__,
648 		 page_to_pfn(page), status,
649 		 vmballoon_page_size_names[page_size]);
650 
651 	/* Error occurred */
652 	vmballoon_stats_page_inc(b, VMW_BALLOON_PAGE_STAT_REFUSED_ALLOC,
653 				 page_size);
654 
655 	return -EIO;
656 }
657 
658 /**
659  * vmballoon_status_page - returns the status of (un)lock operation
660  *
661  * @b: pointer to the balloon.
662  * @idx: index for the page for which the operation is performed.
663  * @p: pointer to where the page struct is returned.
664  *
665  * Following a lock or unlock operation, returns the status of the operation for
666  * an individual page. Provides the page that the operation was performed on on
667  * the @page argument.
668  *
669  * Returns: The status of a lock or unlock operation for an individual page.
670  */
671 static unsigned long vmballoon_status_page(struct vmballoon *b, int idx,
672 					   struct page **p)
673 {
674 	if (static_branch_likely(&vmw_balloon_batching)) {
675 		/* batching mode */
676 		*p = pfn_to_page(b->batch_page[idx].pfn);
677 		return b->batch_page[idx].status;
678 	}
679 
680 	/* non-batching mode */
681 	*p = b->page;
682 
683 	/*
684 	 * If a failure occurs, the indication will be provided in the status
685 	 * of the entire operation, which is considered before the individual
686 	 * page status. So for non-batching mode, the indication is always of
687 	 * success.
688 	 */
689 	return VMW_BALLOON_SUCCESS;
690 }
691 
692 /**
693  * vmballoon_lock_op - notifies the host about inflated/deflated pages.
694  * @b: pointer to the balloon.
695  * @num_pages: number of inflated/deflated pages.
696  * @page_size: size of the page.
697  * @op: the type of operation (lock or unlock).
698  *
699  * Notify the host about page(s) that were ballooned (or removed from the
700  * balloon) so that host can use it without fear that guest will need it (or
701  * stop using them since the VM does). Host may reject some pages, we need to
702  * check the return value and maybe submit a different page. The pages that are
703  * inflated/deflated are pointed by @b->page.
704  *
705  * Return: result as provided by the hypervisor.
706  */
707 static unsigned long vmballoon_lock_op(struct vmballoon *b,
708 				       unsigned int num_pages,
709 				       enum vmballoon_page_size_type page_size,
710 				       enum vmballoon_op op)
711 {
712 	unsigned long cmd, pfn;
713 
714 	lockdep_assert_held(&b->comm_lock);
715 
716 	if (static_branch_likely(&vmw_balloon_batching)) {
717 		if (op == VMW_BALLOON_INFLATE)
718 			cmd = page_size == VMW_BALLOON_2M_PAGE ?
719 				VMW_BALLOON_CMD_BATCHED_2M_LOCK :
720 				VMW_BALLOON_CMD_BATCHED_LOCK;
721 		else
722 			cmd = page_size == VMW_BALLOON_2M_PAGE ?
723 				VMW_BALLOON_CMD_BATCHED_2M_UNLOCK :
724 				VMW_BALLOON_CMD_BATCHED_UNLOCK;
725 
726 		pfn = PHYS_PFN(virt_to_phys(b->batch_page));
727 	} else {
728 		cmd = op == VMW_BALLOON_INFLATE ? VMW_BALLOON_CMD_LOCK :
729 						  VMW_BALLOON_CMD_UNLOCK;
730 		pfn = page_to_pfn(b->page);
731 
732 		/* In non-batching mode, PFNs must fit in 32-bit */
733 		if (unlikely(pfn != (u32)pfn))
734 			return VMW_BALLOON_ERROR_PPN_INVALID;
735 	}
736 
737 	return vmballoon_cmd(b, cmd, pfn, num_pages);
738 }
739 
740 /**
741  * vmballoon_add_page - adds a page towards lock/unlock operation.
742  *
743  * @b: pointer to the balloon.
744  * @idx: index of the page to be ballooned in this batch.
745  * @p: pointer to the page that is about to be ballooned.
746  *
747  * Adds the page to be ballooned. Must be called while holding @comm_lock.
748  */
749 static void vmballoon_add_page(struct vmballoon *b, unsigned int idx,
750 			       struct page *p)
751 {
752 	lockdep_assert_held(&b->comm_lock);
753 
754 	if (static_branch_likely(&vmw_balloon_batching))
755 		b->batch_page[idx] = (struct vmballoon_batch_entry)
756 					{ .pfn = page_to_pfn(p) };
757 	else
758 		b->page = p;
759 }
760 
761 /**
762  * vmballoon_lock - lock or unlock a batch of pages.
763  *
764  * @b: pointer to the balloon.
765  * @ctl: pointer for the %struct vmballoon_ctl, which defines the operation.
766  *
767  * Notifies the host of about ballooned pages (after inflation or deflation,
768  * according to @ctl). If the host rejects the page put it on the
769  * @ctl refuse list. These refused page are then released when moving to the
770  * next size of pages.
771  *
772  * Note that we neither free any @page here nor put them back on the ballooned
773  * pages list. Instead we queue it for later processing. We do that for several
774  * reasons. First, we do not want to free the page under the lock. Second, it
775  * allows us to unify the handling of lock and unlock. In the inflate case, the
776  * caller will check if there are too many refused pages and release them.
777  * Although it is not identical to the past behavior, it should not affect
778  * performance.
779  */
780 static int vmballoon_lock(struct vmballoon *b, struct vmballoon_ctl *ctl)
781 {
782 	unsigned long batch_status;
783 	struct page *page;
784 	unsigned int i, num_pages;
785 
786 	num_pages = ctl->n_pages;
787 	if (num_pages == 0)
788 		return 0;
789 
790 	/* communication with the host is done under the communication lock */
791 	spin_lock(&b->comm_lock);
792 
793 	i = 0;
794 	list_for_each_entry(page, &ctl->pages, lru)
795 		vmballoon_add_page(b, i++, page);
796 
797 	batch_status = vmballoon_lock_op(b, ctl->n_pages, ctl->page_size,
798 					 ctl->op);
799 
800 	/*
801 	 * Iterate over the pages in the provided list. Since we are changing
802 	 * @ctl->n_pages we are saving the original value in @num_pages and
803 	 * use this value to bound the loop.
804 	 */
805 	for (i = 0; i < num_pages; i++) {
806 		unsigned long status;
807 
808 		status = vmballoon_status_page(b, i, &page);
809 
810 		/*
811 		 * Failure of the whole batch overrides a single operation
812 		 * results.
813 		 */
814 		if (batch_status != VMW_BALLOON_SUCCESS)
815 			status = batch_status;
816 
817 		/* Continue if no error happened */
818 		if (!vmballoon_handle_one_result(b, page, ctl->page_size,
819 						 status))
820 			continue;
821 
822 		/*
823 		 * Error happened. Move the pages to the refused list and update
824 		 * the pages number.
825 		 */
826 		list_move(&page->lru, &ctl->refused_pages);
827 		ctl->n_pages--;
828 		ctl->n_refused_pages++;
829 	}
830 
831 	spin_unlock(&b->comm_lock);
832 
833 	return batch_status == VMW_BALLOON_SUCCESS ? 0 : -EIO;
834 }
835 
836 /**
837  * vmballoon_release_page_list() - Releases a page list
838  *
839  * @page_list: list of pages to release.
840  * @n_pages: pointer to the number of pages.
841  * @page_size: whether the pages in the list are 2MB (or else 4KB).
842  *
843  * Releases the list of pages and zeros the number of pages.
844  */
845 static void vmballoon_release_page_list(struct list_head *page_list,
846 				       int *n_pages,
847 				       enum vmballoon_page_size_type page_size)
848 {
849 	struct page *page, *tmp;
850 
851 	list_for_each_entry_safe(page, tmp, page_list, lru) {
852 		list_del(&page->lru);
853 		__free_pages(page, vmballoon_page_order(page_size));
854 	}
855 
856 	*n_pages = 0;
857 }
858 
859 
860 /*
861  * Release pages that were allocated while attempting to inflate the
862  * balloon but were refused by the host for one reason or another.
863  */
864 static void vmballoon_release_refused_pages(struct vmballoon *b,
865 					    struct vmballoon_ctl *ctl)
866 {
867 	vmballoon_stats_page_inc(b, VMW_BALLOON_PAGE_STAT_REFUSED_FREE,
868 				 ctl->page_size);
869 
870 	vmballoon_release_page_list(&ctl->refused_pages, &ctl->n_refused_pages,
871 				    ctl->page_size);
872 }
873 
874 /**
875  * vmballoon_change - retrieve the required balloon change
876  *
877  * @b: pointer for the balloon.
878  *
879  * Return: the required change for the balloon size. A positive number
880  * indicates inflation, a negative number indicates a deflation.
881  */
882 static int64_t vmballoon_change(struct vmballoon *b)
883 {
884 	int64_t size, target;
885 
886 	size = atomic64_read(&b->size);
887 	target = READ_ONCE(b->target);
888 
889 	/*
890 	 * We must cast first because of int sizes
891 	 * Otherwise we might get huge positives instead of negatives
892 	 */
893 
894 	if (b->reset_required)
895 		return 0;
896 
897 	/* consider a 2MB slack on deflate, unless the balloon is emptied */
898 	if (target < size && target != 0 &&
899 	    size - target < vmballoon_page_in_frames(VMW_BALLOON_2M_PAGE))
900 		return 0;
901 
902 	return target - size;
903 }
904 
905 /**
906  * vmballoon_enqueue_page_list() - Enqueues list of pages after inflation.
907  *
908  * @b: pointer to balloon.
909  * @pages: list of pages to enqueue.
910  * @n_pages: pointer to number of pages in list. The value is zeroed.
911  * @page_size: whether the pages are 2MB or 4KB pages.
912  *
913  * Enqueues the provides list of pages in the ballooned page list, clears the
914  * list and zeroes the number of pages that was provided.
915  */
916 static void vmballoon_enqueue_page_list(struct vmballoon *b,
917 					struct list_head *pages,
918 					unsigned int *n_pages,
919 					enum vmballoon_page_size_type page_size)
920 {
921 	struct vmballoon_page_size *page_size_info = &b->page_sizes[page_size];
922 
923 	list_splice_init(pages, &page_size_info->pages);
924 	*n_pages = 0;
925 }
926 
927 /**
928  * vmballoon_dequeue_page_list() - Dequeues page lists for deflation.
929  *
930  * @b: pointer to balloon.
931  * @pages: list of pages to enqueue.
932  * @n_pages: pointer to number of pages in list. The value is zeroed.
933  * @page_size: whether the pages are 2MB or 4KB pages.
934  * @n_req_pages: the number of requested pages.
935  *
936  * Dequeues the number of requested pages from the balloon for deflation. The
937  * number of dequeued pages may be lower, if not enough pages in the requested
938  * size are available.
939  */
940 static void vmballoon_dequeue_page_list(struct vmballoon *b,
941 					struct list_head *pages,
942 					unsigned int *n_pages,
943 					enum vmballoon_page_size_type page_size,
944 					unsigned int n_req_pages)
945 {
946 	struct vmballoon_page_size *page_size_info = &b->page_sizes[page_size];
947 	struct page *page, *tmp;
948 	unsigned int i = 0;
949 
950 	list_for_each_entry_safe(page, tmp, &page_size_info->pages, lru) {
951 		list_move(&page->lru, pages);
952 		if (++i == n_req_pages)
953 			break;
954 	}
955 	*n_pages = i;
956 }
957 
958 /**
959  * vmballoon_inflate() - Inflate the balloon towards its target size.
960  *
961  * @b: pointer to the balloon.
962  */
963 static void vmballoon_inflate(struct vmballoon *b)
964 {
965 	int64_t to_inflate_frames;
966 	struct vmballoon_ctl ctl = {
967 		.pages = LIST_HEAD_INIT(ctl.pages),
968 		.refused_pages = LIST_HEAD_INIT(ctl.refused_pages),
969 		.page_size = b->max_page_size,
970 		.op = VMW_BALLOON_INFLATE
971 	};
972 
973 	while ((to_inflate_frames = vmballoon_change(b)) > 0) {
974 		unsigned int to_inflate_pages, page_in_frames;
975 		int alloc_error, lock_error = 0;
976 
977 		VM_BUG_ON(!list_empty(&ctl.pages));
978 		VM_BUG_ON(ctl.n_pages != 0);
979 
980 		page_in_frames = vmballoon_page_in_frames(ctl.page_size);
981 
982 		to_inflate_pages = min_t(unsigned long, b->batch_max_pages,
983 					 DIV_ROUND_UP_ULL(to_inflate_frames,
984 							  page_in_frames));
985 
986 		/* Start by allocating */
987 		alloc_error = vmballoon_alloc_page_list(b, &ctl,
988 							to_inflate_pages);
989 
990 		/* Actually lock the pages by telling the hypervisor */
991 		lock_error = vmballoon_lock(b, &ctl);
992 
993 		/*
994 		 * If an error indicates that something serious went wrong,
995 		 * stop the inflation.
996 		 */
997 		if (lock_error)
998 			break;
999 
1000 		/* Update the balloon size */
1001 		atomic64_add(ctl.n_pages * page_in_frames, &b->size);
1002 
1003 		vmballoon_enqueue_page_list(b, &ctl.pages, &ctl.n_pages,
1004 					    ctl.page_size);
1005 
1006 		/*
1007 		 * If allocation failed or the number of refused pages exceeds
1008 		 * the maximum allowed, move to the next page size.
1009 		 */
1010 		if (alloc_error ||
1011 		    ctl.n_refused_pages >= VMW_BALLOON_MAX_REFUSED) {
1012 			if (ctl.page_size == VMW_BALLOON_4K_PAGE)
1013 				break;
1014 
1015 			/*
1016 			 * Ignore errors from locking as we now switch to 4k
1017 			 * pages and we might get different errors.
1018 			 */
1019 			vmballoon_release_refused_pages(b, &ctl);
1020 			ctl.page_size--;
1021 		}
1022 
1023 		cond_resched();
1024 	}
1025 
1026 	/*
1027 	 * Release pages that were allocated while attempting to inflate the
1028 	 * balloon but were refused by the host for one reason or another,
1029 	 * and update the statistics.
1030 	 */
1031 	if (ctl.n_refused_pages != 0)
1032 		vmballoon_release_refused_pages(b, &ctl);
1033 }
1034 
1035 /**
1036  * vmballoon_deflate() - Decrease the size of the balloon.
1037  *
1038  * @b: pointer to the balloon
1039  * @n_frames: the number of frames to deflate. If zero, automatically
1040  * calculated according to the target size.
1041  * @coordinated: whether to coordinate with the host
1042  *
1043  * Decrease the size of the balloon allowing guest to use more memory.
1044  *
1045  * Return: The number of deflated frames (i.e., basic page size units)
1046  */
1047 static unsigned long vmballoon_deflate(struct vmballoon *b, uint64_t n_frames,
1048 				       bool coordinated)
1049 {
1050 	unsigned long deflated_frames = 0;
1051 	unsigned long tried_frames = 0;
1052 	struct vmballoon_ctl ctl = {
1053 		.pages = LIST_HEAD_INIT(ctl.pages),
1054 		.refused_pages = LIST_HEAD_INIT(ctl.refused_pages),
1055 		.page_size = VMW_BALLOON_4K_PAGE,
1056 		.op = VMW_BALLOON_DEFLATE
1057 	};
1058 
1059 	/* free pages to reach target */
1060 	while (true) {
1061 		unsigned int to_deflate_pages, n_unlocked_frames;
1062 		unsigned int page_in_frames;
1063 		int64_t to_deflate_frames;
1064 		bool deflated_all;
1065 
1066 		page_in_frames = vmballoon_page_in_frames(ctl.page_size);
1067 
1068 		VM_BUG_ON(!list_empty(&ctl.pages));
1069 		VM_BUG_ON(ctl.n_pages);
1070 		VM_BUG_ON(!list_empty(&ctl.refused_pages));
1071 		VM_BUG_ON(ctl.n_refused_pages);
1072 
1073 		/*
1074 		 * If we were requested a specific number of frames, we try to
1075 		 * deflate this number of frames. Otherwise, deflation is
1076 		 * performed according to the target and balloon size.
1077 		 */
1078 		to_deflate_frames = n_frames ? n_frames - tried_frames :
1079 					       -vmballoon_change(b);
1080 
1081 		/* break if no work to do */
1082 		if (to_deflate_frames <= 0)
1083 			break;
1084 
1085 		/*
1086 		 * Calculate the number of frames based on current page size,
1087 		 * but limit the deflated frames to a single chunk
1088 		 */
1089 		to_deflate_pages = min_t(unsigned long, b->batch_max_pages,
1090 					 DIV_ROUND_UP_ULL(to_deflate_frames,
1091 							  page_in_frames));
1092 
1093 		/* First take the pages from the balloon pages. */
1094 		vmballoon_dequeue_page_list(b, &ctl.pages, &ctl.n_pages,
1095 					    ctl.page_size, to_deflate_pages);
1096 
1097 		/*
1098 		 * Before pages are moving to the refused list, count their
1099 		 * frames as frames that we tried to deflate.
1100 		 */
1101 		tried_frames += ctl.n_pages * page_in_frames;
1102 
1103 		/*
1104 		 * Unlock the pages by communicating with the hypervisor if the
1105 		 * communication is coordinated (i.e., not pop). We ignore the
1106 		 * return code. Instead we check if all the pages we manage to
1107 		 * unlock all the pages. If we failed, we will move to the next
1108 		 * page size, and would eventually try again later.
1109 		 */
1110 		if (coordinated)
1111 			vmballoon_lock(b, &ctl);
1112 
1113 		/*
1114 		 * Check if we deflated enough. We will move to the next page
1115 		 * size if we did not manage to do so. This calculation takes
1116 		 * place now, as once the pages are released, the number of
1117 		 * pages is zeroed.
1118 		 */
1119 		deflated_all = (ctl.n_pages == to_deflate_pages);
1120 
1121 		/* Update local and global counters */
1122 		n_unlocked_frames = ctl.n_pages * page_in_frames;
1123 		atomic64_sub(n_unlocked_frames, &b->size);
1124 		deflated_frames += n_unlocked_frames;
1125 
1126 		vmballoon_stats_page_add(b, VMW_BALLOON_PAGE_STAT_FREE,
1127 					 ctl.page_size, ctl.n_pages);
1128 
1129 		/* free the ballooned pages */
1130 		vmballoon_release_page_list(&ctl.pages, &ctl.n_pages,
1131 					    ctl.page_size);
1132 
1133 		/* Return the refused pages to the ballooned list. */
1134 		vmballoon_enqueue_page_list(b, &ctl.refused_pages,
1135 					    &ctl.n_refused_pages,
1136 					    ctl.page_size);
1137 
1138 		/* If we failed to unlock all the pages, move to next size. */
1139 		if (!deflated_all) {
1140 			if (ctl.page_size == b->max_page_size)
1141 				break;
1142 			ctl.page_size++;
1143 		}
1144 
1145 		cond_resched();
1146 	}
1147 
1148 	return deflated_frames;
1149 }
1150 
1151 /**
1152  * vmballoon_deinit_batching - disables batching mode.
1153  *
1154  * @b: pointer to &struct vmballoon.
1155  *
1156  * Disables batching, by deallocating the page for communication with the
1157  * hypervisor and disabling the static key to indicate that batching is off.
1158  */
1159 static void vmballoon_deinit_batching(struct vmballoon *b)
1160 {
1161 	free_page((unsigned long)b->batch_page);
1162 	b->batch_page = NULL;
1163 	static_branch_disable(&vmw_balloon_batching);
1164 	b->batch_max_pages = 1;
1165 }
1166 
1167 /**
1168  * vmballoon_init_batching - enable batching mode.
1169  *
1170  * @b: pointer to &struct vmballoon.
1171  *
1172  * Enables batching, by allocating a page for communication with the hypervisor
1173  * and enabling the static_key to use batching.
1174  *
1175  * Return: zero on success or an appropriate error-code.
1176  */
1177 static int vmballoon_init_batching(struct vmballoon *b)
1178 {
1179 	struct page *page;
1180 
1181 	page = alloc_page(GFP_KERNEL | __GFP_ZERO);
1182 	if (!page)
1183 		return -ENOMEM;
1184 
1185 	b->batch_page = page_address(page);
1186 	b->batch_max_pages = PAGE_SIZE / sizeof(struct vmballoon_batch_entry);
1187 
1188 	static_branch_enable(&vmw_balloon_batching);
1189 
1190 	return 0;
1191 }
1192 
1193 /*
1194  * Receive notification and resize balloon
1195  */
1196 static void vmballoon_doorbell(void *client_data)
1197 {
1198 	struct vmballoon *b = client_data;
1199 
1200 	vmballoon_stats_gen_inc(b, VMW_BALLOON_STAT_DOORBELL);
1201 
1202 	mod_delayed_work(system_freezable_wq, &b->dwork, 0);
1203 }
1204 
1205 /*
1206  * Clean up vmci doorbell
1207  */
1208 static void vmballoon_vmci_cleanup(struct vmballoon *b)
1209 {
1210 	vmballoon_cmd(b, VMW_BALLOON_CMD_VMCI_DOORBELL_SET,
1211 		      VMCI_INVALID_ID, VMCI_INVALID_ID);
1212 
1213 	if (!vmci_handle_is_invalid(b->vmci_doorbell)) {
1214 		vmci_doorbell_destroy(b->vmci_doorbell);
1215 		b->vmci_doorbell = VMCI_INVALID_HANDLE;
1216 	}
1217 }
1218 
1219 /**
1220  * vmballoon_vmci_init - Initialize vmci doorbell.
1221  *
1222  * @b: pointer to the balloon.
1223  *
1224  * Return: zero on success or when wakeup command not supported. Error-code
1225  * otherwise.
1226  *
1227  * Initialize vmci doorbell, to get notified as soon as balloon changes.
1228  */
1229 static int vmballoon_vmci_init(struct vmballoon *b)
1230 {
1231 	unsigned long error;
1232 
1233 	if ((b->capabilities & VMW_BALLOON_SIGNALLED_WAKEUP_CMD) == 0)
1234 		return 0;
1235 
1236 	error = vmci_doorbell_create(&b->vmci_doorbell, VMCI_FLAG_DELAYED_CB,
1237 				     VMCI_PRIVILEGE_FLAG_RESTRICTED,
1238 				     vmballoon_doorbell, b);
1239 
1240 	if (error != VMCI_SUCCESS)
1241 		goto fail;
1242 
1243 	error =	__vmballoon_cmd(b, VMW_BALLOON_CMD_VMCI_DOORBELL_SET,
1244 				b->vmci_doorbell.context,
1245 				b->vmci_doorbell.resource, NULL);
1246 
1247 	if (error != VMW_BALLOON_SUCCESS)
1248 		goto fail;
1249 
1250 	return 0;
1251 fail:
1252 	vmballoon_vmci_cleanup(b);
1253 	return -EIO;
1254 }
1255 
1256 /**
1257  * vmballoon_pop - Quickly release all pages allocate for the balloon.
1258  *
1259  * @b: pointer to the balloon.
1260  *
1261  * This function is called when host decides to "reset" balloon for one reason
1262  * or another. Unlike normal "deflate" we do not (shall not) notify host of the
1263  * pages being released.
1264  */
1265 static void vmballoon_pop(struct vmballoon *b)
1266 {
1267 	unsigned long size;
1268 
1269 	while ((size = atomic64_read(&b->size)))
1270 		vmballoon_deflate(b, size, false);
1271 }
1272 
1273 /*
1274  * Perform standard reset sequence by popping the balloon (in case it
1275  * is not  empty) and then restarting protocol. This operation normally
1276  * happens when host responds with VMW_BALLOON_ERROR_RESET to a command.
1277  */
1278 static void vmballoon_reset(struct vmballoon *b)
1279 {
1280 	int error;
1281 
1282 	down_write(&b->conf_sem);
1283 
1284 	vmballoon_vmci_cleanup(b);
1285 
1286 	/* free all pages, skipping monitor unlock */
1287 	vmballoon_pop(b);
1288 
1289 	if (vmballoon_send_start(b, VMW_BALLOON_CAPABILITIES))
1290 		return;
1291 
1292 	if ((b->capabilities & VMW_BALLOON_BATCHED_CMDS) != 0) {
1293 		if (vmballoon_init_batching(b)) {
1294 			/*
1295 			 * We failed to initialize batching, inform the monitor
1296 			 * about it by sending a null capability.
1297 			 *
1298 			 * The guest will retry in one second.
1299 			 */
1300 			vmballoon_send_start(b, 0);
1301 			return;
1302 		}
1303 	} else if ((b->capabilities & VMW_BALLOON_BASIC_CMDS) != 0) {
1304 		vmballoon_deinit_batching(b);
1305 	}
1306 
1307 	vmballoon_stats_gen_inc(b, VMW_BALLOON_STAT_RESET);
1308 	b->reset_required = false;
1309 
1310 	error = vmballoon_vmci_init(b);
1311 	if (error)
1312 		pr_err("failed to initialize vmci doorbell\n");
1313 
1314 	if (vmballoon_send_guest_id(b))
1315 		pr_err("failed to send guest ID to the host\n");
1316 
1317 	up_write(&b->conf_sem);
1318 }
1319 
1320 /**
1321  * vmballoon_work - periodic balloon worker for reset, inflation and deflation.
1322  *
1323  * @work: pointer to the &work_struct which is provided by the workqueue.
1324  *
1325  * Resets the protocol if needed, gets the new size and adjusts balloon as
1326  * needed. Repeat in 1 sec.
1327  */
1328 static void vmballoon_work(struct work_struct *work)
1329 {
1330 	struct delayed_work *dwork = to_delayed_work(work);
1331 	struct vmballoon *b = container_of(dwork, struct vmballoon, dwork);
1332 	int64_t change = 0;
1333 
1334 	if (b->reset_required)
1335 		vmballoon_reset(b);
1336 
1337 	down_read(&b->conf_sem);
1338 
1339 	/*
1340 	 * Update the stats while holding the semaphore to ensure that
1341 	 * @stats_enabled is consistent with whether the stats are actually
1342 	 * enabled
1343 	 */
1344 	vmballoon_stats_gen_inc(b, VMW_BALLOON_STAT_TIMER);
1345 
1346 	if (!vmballoon_send_get_target(b))
1347 		change = vmballoon_change(b);
1348 
1349 	if (change != 0) {
1350 		pr_debug("%s - size: %llu, target %lu\n", __func__,
1351 			 atomic64_read(&b->size), READ_ONCE(b->target));
1352 
1353 		if (change > 0)
1354 			vmballoon_inflate(b);
1355 		else  /* (change < 0) */
1356 			vmballoon_deflate(b, 0, true);
1357 	}
1358 
1359 	up_read(&b->conf_sem);
1360 
1361 	/*
1362 	 * We are using a freezable workqueue so that balloon operations are
1363 	 * stopped while the system transitions to/from sleep/hibernation.
1364 	 */
1365 	queue_delayed_work(system_freezable_wq,
1366 			   dwork, round_jiffies_relative(HZ));
1367 
1368 }
1369 
1370 /*
1371  * DEBUGFS Interface
1372  */
1373 #ifdef CONFIG_DEBUG_FS
1374 
1375 static const char * const vmballoon_stat_page_names[] = {
1376 	[VMW_BALLOON_PAGE_STAT_ALLOC]		= "alloc",
1377 	[VMW_BALLOON_PAGE_STAT_ALLOC_FAIL]	= "allocFail",
1378 	[VMW_BALLOON_PAGE_STAT_REFUSED_ALLOC]	= "errAlloc",
1379 	[VMW_BALLOON_PAGE_STAT_REFUSED_FREE]	= "errFree",
1380 	[VMW_BALLOON_PAGE_STAT_FREE]		= "free"
1381 };
1382 
1383 static const char * const vmballoon_stat_names[] = {
1384 	[VMW_BALLOON_STAT_TIMER]		= "timer",
1385 	[VMW_BALLOON_STAT_DOORBELL]		= "doorbell",
1386 	[VMW_BALLOON_STAT_RESET]		= "reset",
1387 };
1388 
1389 static int vmballoon_enable_stats(struct vmballoon *b)
1390 {
1391 	int r = 0;
1392 
1393 	down_write(&b->conf_sem);
1394 
1395 	/* did we somehow race with another reader which enabled stats? */
1396 	if (b->stats)
1397 		goto out;
1398 
1399 	b->stats = kzalloc(sizeof(*b->stats), GFP_KERNEL);
1400 
1401 	if (!b->stats) {
1402 		/* allocation failed */
1403 		r = -ENOMEM;
1404 		goto out;
1405 	}
1406 	static_key_enable(&balloon_stat_enabled.key);
1407 out:
1408 	up_write(&b->conf_sem);
1409 	return r;
1410 }
1411 
1412 /**
1413  * vmballoon_debug_show - shows statistics of balloon operations.
1414  * @f: pointer to the &struct seq_file.
1415  * @offset: ignored.
1416  *
1417  * Provides the statistics that can be accessed in vmmemctl in the debugfs.
1418  * To avoid the overhead - mainly that of memory - of collecting the statistics,
1419  * we only collect statistics after the first time the counters are read.
1420  *
1421  * Return: zero on success or an error code.
1422  */
1423 static int vmballoon_debug_show(struct seq_file *f, void *offset)
1424 {
1425 	struct vmballoon *b = f->private;
1426 	int i, j;
1427 
1428 	/* enables stats if they are disabled */
1429 	if (!b->stats) {
1430 		int r = vmballoon_enable_stats(b);
1431 
1432 		if (r)
1433 			return r;
1434 	}
1435 
1436 	/* format capabilities info */
1437 	seq_printf(f, "%-22s: %#16x\n", "balloon capabilities",
1438 		   VMW_BALLOON_CAPABILITIES);
1439 	seq_printf(f, "%-22s: %#16lx\n", "used capabilities", b->capabilities);
1440 	seq_printf(f, "%-22s: %16s\n", "is resetting",
1441 		   b->reset_required ? "y" : "n");
1442 
1443 	/* format size info */
1444 	seq_printf(f, "%-22s: %16lu\n", "target", READ_ONCE(b->target));
1445 	seq_printf(f, "%-22s: %16llu\n", "current", atomic64_read(&b->size));
1446 
1447 	for (i = 0; i < VMW_BALLOON_CMD_NUM; i++) {
1448 		if (vmballoon_cmd_names[i] == NULL)
1449 			continue;
1450 
1451 		seq_printf(f, "%-22s: %16llu (%llu failed)\n",
1452 			   vmballoon_cmd_names[i],
1453 			   atomic64_read(&b->stats->ops[i][VMW_BALLOON_OP_STAT]),
1454 			   atomic64_read(&b->stats->ops[i][VMW_BALLOON_OP_FAIL_STAT]));
1455 	}
1456 
1457 	for (i = 0; i < VMW_BALLOON_STAT_NUM; i++)
1458 		seq_printf(f, "%-22s: %16llu\n",
1459 			   vmballoon_stat_names[i],
1460 			   atomic64_read(&b->stats->general_stat[i]));
1461 
1462 	for (i = 0; i < VMW_BALLOON_PAGE_STAT_NUM; i++) {
1463 		for (j = 0; j < VMW_BALLOON_NUM_PAGE_SIZES; j++)
1464 			seq_printf(f, "%-18s(%s): %16llu\n",
1465 				   vmballoon_stat_page_names[i],
1466 				   vmballoon_page_size_names[j],
1467 				   atomic64_read(&b->stats->page_stat[i][j]));
1468 	}
1469 
1470 	return 0;
1471 }
1472 
1473 DEFINE_SHOW_ATTRIBUTE(vmballoon_debug);
1474 
1475 static int __init vmballoon_debugfs_init(struct vmballoon *b)
1476 {
1477 	int error;
1478 
1479 	b->dbg_entry = debugfs_create_file("vmmemctl", S_IRUGO, NULL, b,
1480 					   &vmballoon_debug_fops);
1481 	if (IS_ERR(b->dbg_entry)) {
1482 		error = PTR_ERR(b->dbg_entry);
1483 		pr_err("failed to create debugfs entry, error: %d\n", error);
1484 		return error;
1485 	}
1486 
1487 	return 0;
1488 }
1489 
1490 static void __exit vmballoon_debugfs_exit(struct vmballoon *b)
1491 {
1492 	static_key_disable(&balloon_stat_enabled.key);
1493 	debugfs_remove(b->dbg_entry);
1494 	kfree(b->stats);
1495 	b->stats = NULL;
1496 }
1497 
1498 #else
1499 
1500 static inline int vmballoon_debugfs_init(struct vmballoon *b)
1501 {
1502 	return 0;
1503 }
1504 
1505 static inline void vmballoon_debugfs_exit(struct vmballoon *b)
1506 {
1507 }
1508 
1509 #endif	/* CONFIG_DEBUG_FS */
1510 
1511 static int __init vmballoon_init(void)
1512 {
1513 	enum vmballoon_page_size_type page_size;
1514 	int error;
1515 
1516 	/*
1517 	 * Check if we are running on VMware's hypervisor and bail out
1518 	 * if we are not.
1519 	 */
1520 	if (x86_hyper_type != X86_HYPER_VMWARE)
1521 		return -ENODEV;
1522 
1523 	for (page_size = VMW_BALLOON_4K_PAGE;
1524 	     page_size <= VMW_BALLOON_LAST_SIZE; page_size++)
1525 		INIT_LIST_HEAD(&balloon.page_sizes[page_size].pages);
1526 
1527 
1528 	INIT_DELAYED_WORK(&balloon.dwork, vmballoon_work);
1529 
1530 	error = vmballoon_debugfs_init(&balloon);
1531 	if (error)
1532 		return error;
1533 
1534 	spin_lock_init(&balloon.comm_lock);
1535 	init_rwsem(&balloon.conf_sem);
1536 	balloon.vmci_doorbell = VMCI_INVALID_HANDLE;
1537 	balloon.batch_page = NULL;
1538 	balloon.page = NULL;
1539 	balloon.reset_required = true;
1540 
1541 	queue_delayed_work(system_freezable_wq, &balloon.dwork, 0);
1542 
1543 	return 0;
1544 }
1545 
1546 /*
1547  * Using late_initcall() instead of module_init() allows the balloon to use the
1548  * VMCI doorbell even when the balloon is built into the kernel. Otherwise the
1549  * VMCI is probed only after the balloon is initialized. If the balloon is used
1550  * as a module, late_initcall() is equivalent to module_init().
1551  */
1552 late_initcall(vmballoon_init);
1553 
1554 static void __exit vmballoon_exit(void)
1555 {
1556 	vmballoon_vmci_cleanup(&balloon);
1557 	cancel_delayed_work_sync(&balloon.dwork);
1558 
1559 	vmballoon_debugfs_exit(&balloon);
1560 
1561 	/*
1562 	 * Deallocate all reserved memory, and reset connection with monitor.
1563 	 * Reset connection before deallocating memory to avoid potential for
1564 	 * additional spurious resets from guest touching deallocated pages.
1565 	 */
1566 	vmballoon_send_start(&balloon, 0);
1567 	vmballoon_pop(&balloon);
1568 }
1569 module_exit(vmballoon_exit);
1570