xref: /openbmc/linux/drivers/misc/vmw_balloon.c (revision 209cdbd0)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * VMware Balloon driver.
4  *
5  * Copyright (C) 2000-2018, VMware, Inc. All Rights Reserved.
6  *
7  * This is VMware physical memory management driver for Linux. The driver
8  * acts like a "balloon" that can be inflated to reclaim physical pages by
9  * reserving them in the guest and invalidating them in the monitor,
10  * freeing up the underlying machine pages so they can be allocated to
11  * other guests.  The balloon can also be deflated to allow the guest to
12  * use more physical memory. Higher level policies can control the sizes
13  * of balloons in VMs in order to manage physical memory resources.
14  */
15 
16 //#define DEBUG
17 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
18 
19 #include <linux/types.h>
20 #include <linux/io.h>
21 #include <linux/kernel.h>
22 #include <linux/mm.h>
23 #include <linux/vmalloc.h>
24 #include <linux/sched.h>
25 #include <linux/module.h>
26 #include <linux/workqueue.h>
27 #include <linux/debugfs.h>
28 #include <linux/seq_file.h>
29 #include <linux/rwsem.h>
30 #include <linux/slab.h>
31 #include <linux/spinlock.h>
32 #include <linux/balloon_compaction.h>
33 #include <linux/vmw_vmci_defs.h>
34 #include <linux/vmw_vmci_api.h>
35 #include <asm/hypervisor.h>
36 
37 MODULE_AUTHOR("VMware, Inc.");
38 MODULE_DESCRIPTION("VMware Memory Control (Balloon) Driver");
39 MODULE_ALIAS("dmi:*:svnVMware*:*");
40 MODULE_ALIAS("vmware_vmmemctl");
41 MODULE_LICENSE("GPL");
42 
43 static bool __read_mostly vmwballoon_shrinker_enable;
44 module_param(vmwballoon_shrinker_enable, bool, 0444);
45 MODULE_PARM_DESC(vmwballoon_shrinker_enable,
46 	"Enable non-cooperative out-of-memory protection. Disabled by default as it may degrade performance.");
47 
48 /* Delay in seconds after shrink before inflation. */
49 #define VMBALLOON_SHRINK_DELAY		(5)
50 
51 /* Maximum number of refused pages we accumulate during inflation cycle */
52 #define VMW_BALLOON_MAX_REFUSED		16
53 
54 /* Magic number for the balloon mount-point */
55 #define BALLOON_VMW_MAGIC		0x0ba11007
56 
57 /*
58  * Hypervisor communication port definitions.
59  */
60 #define VMW_BALLOON_HV_PORT		0x5670
61 #define VMW_BALLOON_HV_MAGIC		0x456c6d6f
62 #define VMW_BALLOON_GUEST_ID		1	/* Linux */
63 
64 enum vmwballoon_capabilities {
65 	/*
66 	 * Bit 0 is reserved and not associated to any capability.
67 	 */
68 	VMW_BALLOON_BASIC_CMDS			= (1 << 1),
69 	VMW_BALLOON_BATCHED_CMDS		= (1 << 2),
70 	VMW_BALLOON_BATCHED_2M_CMDS		= (1 << 3),
71 	VMW_BALLOON_SIGNALLED_WAKEUP_CMD	= (1 << 4),
72 	VMW_BALLOON_64_BIT_TARGET		= (1 << 5)
73 };
74 
75 #define VMW_BALLOON_CAPABILITIES_COMMON	(VMW_BALLOON_BASIC_CMDS \
76 					| VMW_BALLOON_BATCHED_CMDS \
77 					| VMW_BALLOON_BATCHED_2M_CMDS \
78 					| VMW_BALLOON_SIGNALLED_WAKEUP_CMD)
79 
80 #define VMW_BALLOON_2M_ORDER		(PMD_SHIFT - PAGE_SHIFT)
81 
82 /*
83  * 64-bit targets are only supported in 64-bit
84  */
85 #ifdef CONFIG_64BIT
86 #define VMW_BALLOON_CAPABILITIES	(VMW_BALLOON_CAPABILITIES_COMMON \
87 					| VMW_BALLOON_64_BIT_TARGET)
88 #else
89 #define VMW_BALLOON_CAPABILITIES	VMW_BALLOON_CAPABILITIES_COMMON
90 #endif
91 
92 enum vmballoon_page_size_type {
93 	VMW_BALLOON_4K_PAGE,
94 	VMW_BALLOON_2M_PAGE,
95 	VMW_BALLOON_LAST_SIZE = VMW_BALLOON_2M_PAGE
96 };
97 
98 #define VMW_BALLOON_NUM_PAGE_SIZES	(VMW_BALLOON_LAST_SIZE + 1)
99 
100 static const char * const vmballoon_page_size_names[] = {
101 	[VMW_BALLOON_4K_PAGE]			= "4k",
102 	[VMW_BALLOON_2M_PAGE]			= "2M"
103 };
104 
105 enum vmballoon_op {
106 	VMW_BALLOON_INFLATE,
107 	VMW_BALLOON_DEFLATE
108 };
109 
110 enum vmballoon_op_stat_type {
111 	VMW_BALLOON_OP_STAT,
112 	VMW_BALLOON_OP_FAIL_STAT
113 };
114 
115 #define VMW_BALLOON_OP_STAT_TYPES	(VMW_BALLOON_OP_FAIL_STAT + 1)
116 
117 /**
118  * enum vmballoon_cmd_type - backdoor commands.
119  *
120  * Availability of the commands is as followed:
121  *
122  * %VMW_BALLOON_CMD_START, %VMW_BALLOON_CMD_GET_TARGET and
123  * %VMW_BALLOON_CMD_GUEST_ID are always available.
124  *
125  * If the host reports %VMW_BALLOON_BASIC_CMDS are supported then
126  * %VMW_BALLOON_CMD_LOCK and %VMW_BALLOON_CMD_UNLOCK commands are available.
127  *
128  * If the host reports %VMW_BALLOON_BATCHED_CMDS are supported then
129  * %VMW_BALLOON_CMD_BATCHED_LOCK and VMW_BALLOON_CMD_BATCHED_UNLOCK commands
130  * are available.
131  *
132  * If the host reports %VMW_BALLOON_BATCHED_2M_CMDS are supported then
133  * %VMW_BALLOON_CMD_BATCHED_2M_LOCK and %VMW_BALLOON_CMD_BATCHED_2M_UNLOCK
134  * are supported.
135  *
136  * If the host reports  VMW_BALLOON_SIGNALLED_WAKEUP_CMD is supported then
137  * VMW_BALLOON_CMD_VMCI_DOORBELL_SET command is supported.
138  *
139  * @VMW_BALLOON_CMD_START: Communicating supported version with the hypervisor.
140  * @VMW_BALLOON_CMD_GET_TARGET: Gets the balloon target size.
141  * @VMW_BALLOON_CMD_LOCK: Informs the hypervisor about a ballooned page.
142  * @VMW_BALLOON_CMD_UNLOCK: Informs the hypervisor about a page that is about
143  *			    to be deflated from the balloon.
144  * @VMW_BALLOON_CMD_GUEST_ID: Informs the hypervisor about the type of OS that
145  *			      runs in the VM.
146  * @VMW_BALLOON_CMD_BATCHED_LOCK: Inform the hypervisor about a batch of
147  *				  ballooned pages (up to 512).
148  * @VMW_BALLOON_CMD_BATCHED_UNLOCK: Inform the hypervisor about a batch of
149  *				  pages that are about to be deflated from the
150  *				  balloon (up to 512).
151  * @VMW_BALLOON_CMD_BATCHED_2M_LOCK: Similar to @VMW_BALLOON_CMD_BATCHED_LOCK
152  *				     for 2MB pages.
153  * @VMW_BALLOON_CMD_BATCHED_2M_UNLOCK: Similar to
154  *				       @VMW_BALLOON_CMD_BATCHED_UNLOCK for 2MB
155  *				       pages.
156  * @VMW_BALLOON_CMD_VMCI_DOORBELL_SET: A command to set doorbell notification
157  *				       that would be invoked when the balloon
158  *				       size changes.
159  * @VMW_BALLOON_CMD_LAST: Value of the last command.
160  */
161 enum vmballoon_cmd_type {
162 	VMW_BALLOON_CMD_START,
163 	VMW_BALLOON_CMD_GET_TARGET,
164 	VMW_BALLOON_CMD_LOCK,
165 	VMW_BALLOON_CMD_UNLOCK,
166 	VMW_BALLOON_CMD_GUEST_ID,
167 	/* No command 5 */
168 	VMW_BALLOON_CMD_BATCHED_LOCK = 6,
169 	VMW_BALLOON_CMD_BATCHED_UNLOCK,
170 	VMW_BALLOON_CMD_BATCHED_2M_LOCK,
171 	VMW_BALLOON_CMD_BATCHED_2M_UNLOCK,
172 	VMW_BALLOON_CMD_VMCI_DOORBELL_SET,
173 	VMW_BALLOON_CMD_LAST = VMW_BALLOON_CMD_VMCI_DOORBELL_SET,
174 };
175 
176 #define VMW_BALLOON_CMD_NUM	(VMW_BALLOON_CMD_LAST + 1)
177 
178 enum vmballoon_error_codes {
179 	VMW_BALLOON_SUCCESS,
180 	VMW_BALLOON_ERROR_CMD_INVALID,
181 	VMW_BALLOON_ERROR_PPN_INVALID,
182 	VMW_BALLOON_ERROR_PPN_LOCKED,
183 	VMW_BALLOON_ERROR_PPN_UNLOCKED,
184 	VMW_BALLOON_ERROR_PPN_PINNED,
185 	VMW_BALLOON_ERROR_PPN_NOTNEEDED,
186 	VMW_BALLOON_ERROR_RESET,
187 	VMW_BALLOON_ERROR_BUSY
188 };
189 
190 #define VMW_BALLOON_SUCCESS_WITH_CAPABILITIES	(0x03000000)
191 
192 #define VMW_BALLOON_CMD_WITH_TARGET_MASK			\
193 	((1UL << VMW_BALLOON_CMD_GET_TARGET)		|	\
194 	 (1UL << VMW_BALLOON_CMD_LOCK)			|	\
195 	 (1UL << VMW_BALLOON_CMD_UNLOCK)		|	\
196 	 (1UL << VMW_BALLOON_CMD_BATCHED_LOCK)		|	\
197 	 (1UL << VMW_BALLOON_CMD_BATCHED_UNLOCK)	|	\
198 	 (1UL << VMW_BALLOON_CMD_BATCHED_2M_LOCK)	|	\
199 	 (1UL << VMW_BALLOON_CMD_BATCHED_2M_UNLOCK))
200 
201 static const char * const vmballoon_cmd_names[] = {
202 	[VMW_BALLOON_CMD_START]			= "start",
203 	[VMW_BALLOON_CMD_GET_TARGET]		= "target",
204 	[VMW_BALLOON_CMD_LOCK]			= "lock",
205 	[VMW_BALLOON_CMD_UNLOCK]		= "unlock",
206 	[VMW_BALLOON_CMD_GUEST_ID]		= "guestType",
207 	[VMW_BALLOON_CMD_BATCHED_LOCK]		= "batchLock",
208 	[VMW_BALLOON_CMD_BATCHED_UNLOCK]	= "batchUnlock",
209 	[VMW_BALLOON_CMD_BATCHED_2M_LOCK]	= "2m-lock",
210 	[VMW_BALLOON_CMD_BATCHED_2M_UNLOCK]	= "2m-unlock",
211 	[VMW_BALLOON_CMD_VMCI_DOORBELL_SET]	= "doorbellSet"
212 };
213 
214 enum vmballoon_stat_page {
215 	VMW_BALLOON_PAGE_STAT_ALLOC,
216 	VMW_BALLOON_PAGE_STAT_ALLOC_FAIL,
217 	VMW_BALLOON_PAGE_STAT_REFUSED_ALLOC,
218 	VMW_BALLOON_PAGE_STAT_REFUSED_FREE,
219 	VMW_BALLOON_PAGE_STAT_FREE,
220 	VMW_BALLOON_PAGE_STAT_LAST = VMW_BALLOON_PAGE_STAT_FREE
221 };
222 
223 #define VMW_BALLOON_PAGE_STAT_NUM	(VMW_BALLOON_PAGE_STAT_LAST + 1)
224 
225 enum vmballoon_stat_general {
226 	VMW_BALLOON_STAT_TIMER,
227 	VMW_BALLOON_STAT_DOORBELL,
228 	VMW_BALLOON_STAT_RESET,
229 	VMW_BALLOON_STAT_SHRINK,
230 	VMW_BALLOON_STAT_SHRINK_FREE,
231 	VMW_BALLOON_STAT_LAST = VMW_BALLOON_STAT_SHRINK_FREE
232 };
233 
234 #define VMW_BALLOON_STAT_NUM		(VMW_BALLOON_STAT_LAST + 1)
235 
236 static DEFINE_STATIC_KEY_TRUE(vmw_balloon_batching);
237 static DEFINE_STATIC_KEY_FALSE(balloon_stat_enabled);
238 
239 struct vmballoon_ctl {
240 	struct list_head pages;
241 	struct list_head refused_pages;
242 	struct list_head prealloc_pages;
243 	unsigned int n_refused_pages;
244 	unsigned int n_pages;
245 	enum vmballoon_page_size_type page_size;
246 	enum vmballoon_op op;
247 };
248 
249 /**
250  * struct vmballoon_batch_entry - a batch entry for lock or unlock.
251  *
252  * @status: the status of the operation, which is written by the hypervisor.
253  * @reserved: reserved for future use. Must be set to zero.
254  * @pfn: the physical frame number of the page to be locked or unlocked.
255  */
256 struct vmballoon_batch_entry {
257 	u64 status : 5;
258 	u64 reserved : PAGE_SHIFT - 5;
259 	u64 pfn : 52;
260 } __packed;
261 
262 struct vmballoon {
263 	/**
264 	 * @max_page_size: maximum supported page size for ballooning.
265 	 *
266 	 * Protected by @conf_sem
267 	 */
268 	enum vmballoon_page_size_type max_page_size;
269 
270 	/**
271 	 * @size: balloon actual size in basic page size (frames).
272 	 *
273 	 * While we currently do not support size which is bigger than 32-bit,
274 	 * in preparation for future support, use 64-bits.
275 	 */
276 	atomic64_t size;
277 
278 	/**
279 	 * @target: balloon target size in basic page size (frames).
280 	 *
281 	 * We do not protect the target under the assumption that setting the
282 	 * value is always done through a single write. If this assumption ever
283 	 * breaks, we would have to use X_ONCE for accesses, and suffer the less
284 	 * optimized code. Although we may read stale target value if multiple
285 	 * accesses happen at once, the performance impact should be minor.
286 	 */
287 	unsigned long target;
288 
289 	/**
290 	 * @reset_required: reset flag
291 	 *
292 	 * Setting this flag may introduce races, but the code is expected to
293 	 * handle them gracefully. In the worst case, another operation will
294 	 * fail as reset did not take place. Clearing the flag is done while
295 	 * holding @conf_sem for write.
296 	 */
297 	bool reset_required;
298 
299 	/**
300 	 * @capabilities: hypervisor balloon capabilities.
301 	 *
302 	 * Protected by @conf_sem.
303 	 */
304 	unsigned long capabilities;
305 
306 	/**
307 	 * @batch_page: pointer to communication batch page.
308 	 *
309 	 * When batching is used, batch_page points to a page, which holds up to
310 	 * %VMW_BALLOON_BATCH_MAX_PAGES entries for locking or unlocking.
311 	 */
312 	struct vmballoon_batch_entry *batch_page;
313 
314 	/**
315 	 * @batch_max_pages: maximum pages that can be locked/unlocked.
316 	 *
317 	 * Indicates the number of pages that the hypervisor can lock or unlock
318 	 * at once, according to whether batching is enabled. If batching is
319 	 * disabled, only a single page can be locked/unlock on each operation.
320 	 *
321 	 * Protected by @conf_sem.
322 	 */
323 	unsigned int batch_max_pages;
324 
325 	/**
326 	 * @page: page to be locked/unlocked by the hypervisor
327 	 *
328 	 * @page is only used when batching is disabled and a single page is
329 	 * reclaimed on each iteration.
330 	 *
331 	 * Protected by @comm_lock.
332 	 */
333 	struct page *page;
334 
335 	/**
336 	 * @shrink_timeout: timeout until the next inflation.
337 	 *
338 	 * After an shrink event, indicates the time in jiffies after which
339 	 * inflation is allowed again. Can be written concurrently with reads,
340 	 * so must use READ_ONCE/WRITE_ONCE when accessing.
341 	 */
342 	unsigned long shrink_timeout;
343 
344 	/* statistics */
345 	struct vmballoon_stats *stats;
346 
347 	/**
348 	 * @b_dev_info: balloon device information descriptor.
349 	 */
350 	struct balloon_dev_info b_dev_info;
351 
352 	struct delayed_work dwork;
353 
354 	/**
355 	 * @huge_pages - list of the inflated 2MB pages.
356 	 *
357 	 * Protected by @b_dev_info.pages_lock .
358 	 */
359 	struct list_head huge_pages;
360 
361 	/**
362 	 * @vmci_doorbell.
363 	 *
364 	 * Protected by @conf_sem.
365 	 */
366 	struct vmci_handle vmci_doorbell;
367 
368 	/**
369 	 * @conf_sem: semaphore to protect the configuration and the statistics.
370 	 */
371 	struct rw_semaphore conf_sem;
372 
373 	/**
374 	 * @comm_lock: lock to protect the communication with the host.
375 	 *
376 	 * Lock ordering: @conf_sem -> @comm_lock .
377 	 */
378 	spinlock_t comm_lock;
379 
380 	/**
381 	 * @shrinker: shrinker interface that is used to avoid over-inflation.
382 	 */
383 	struct shrinker shrinker;
384 
385 	/**
386 	 * @shrinker_registered: whether the shrinker was registered.
387 	 *
388 	 * The shrinker interface does not handle gracefully the removal of
389 	 * shrinker that was not registered before. This indication allows to
390 	 * simplify the unregistration process.
391 	 */
392 	bool shrinker_registered;
393 };
394 
395 static struct vmballoon balloon;
396 
397 struct vmballoon_stats {
398 	/* timer / doorbell operations */
399 	atomic64_t general_stat[VMW_BALLOON_STAT_NUM];
400 
401 	/* allocation statistics for huge and small pages */
402 	atomic64_t
403 	       page_stat[VMW_BALLOON_PAGE_STAT_NUM][VMW_BALLOON_NUM_PAGE_SIZES];
404 
405 	/* Monitor operations: total operations, and failures */
406 	atomic64_t ops[VMW_BALLOON_CMD_NUM][VMW_BALLOON_OP_STAT_TYPES];
407 };
408 
is_vmballoon_stats_on(void)409 static inline bool is_vmballoon_stats_on(void)
410 {
411 	return IS_ENABLED(CONFIG_DEBUG_FS) &&
412 		static_branch_unlikely(&balloon_stat_enabled);
413 }
414 
vmballoon_stats_op_inc(struct vmballoon * b,unsigned int op,enum vmballoon_op_stat_type type)415 static inline void vmballoon_stats_op_inc(struct vmballoon *b, unsigned int op,
416 					  enum vmballoon_op_stat_type type)
417 {
418 	if (is_vmballoon_stats_on())
419 		atomic64_inc(&b->stats->ops[op][type]);
420 }
421 
vmballoon_stats_gen_inc(struct vmballoon * b,enum vmballoon_stat_general stat)422 static inline void vmballoon_stats_gen_inc(struct vmballoon *b,
423 					   enum vmballoon_stat_general stat)
424 {
425 	if (is_vmballoon_stats_on())
426 		atomic64_inc(&b->stats->general_stat[stat]);
427 }
428 
vmballoon_stats_gen_add(struct vmballoon * b,enum vmballoon_stat_general stat,unsigned int val)429 static inline void vmballoon_stats_gen_add(struct vmballoon *b,
430 					   enum vmballoon_stat_general stat,
431 					   unsigned int val)
432 {
433 	if (is_vmballoon_stats_on())
434 		atomic64_add(val, &b->stats->general_stat[stat]);
435 }
436 
vmballoon_stats_page_inc(struct vmballoon * b,enum vmballoon_stat_page stat,enum vmballoon_page_size_type size)437 static inline void vmballoon_stats_page_inc(struct vmballoon *b,
438 					    enum vmballoon_stat_page stat,
439 					    enum vmballoon_page_size_type size)
440 {
441 	if (is_vmballoon_stats_on())
442 		atomic64_inc(&b->stats->page_stat[stat][size]);
443 }
444 
vmballoon_stats_page_add(struct vmballoon * b,enum vmballoon_stat_page stat,enum vmballoon_page_size_type size,unsigned int val)445 static inline void vmballoon_stats_page_add(struct vmballoon *b,
446 					    enum vmballoon_stat_page stat,
447 					    enum vmballoon_page_size_type size,
448 					    unsigned int val)
449 {
450 	if (is_vmballoon_stats_on())
451 		atomic64_add(val, &b->stats->page_stat[stat][size]);
452 }
453 
454 static inline unsigned long
__vmballoon_cmd(struct vmballoon * b,unsigned long cmd,unsigned long arg1,unsigned long arg2,unsigned long * result)455 __vmballoon_cmd(struct vmballoon *b, unsigned long cmd, unsigned long arg1,
456 		unsigned long arg2, unsigned long *result)
457 {
458 	unsigned long status, dummy1, dummy2, dummy3, local_result;
459 
460 	vmballoon_stats_op_inc(b, cmd, VMW_BALLOON_OP_STAT);
461 
462 	asm volatile ("inl %%dx" :
463 		"=a"(status),
464 		"=c"(dummy1),
465 		"=d"(dummy2),
466 		"=b"(local_result),
467 		"=S"(dummy3) :
468 		"0"(VMW_BALLOON_HV_MAGIC),
469 		"1"(cmd),
470 		"2"(VMW_BALLOON_HV_PORT),
471 		"3"(arg1),
472 		"4"(arg2) :
473 		"memory");
474 
475 	/* update the result if needed */
476 	if (result)
477 		*result = (cmd == VMW_BALLOON_CMD_START) ? dummy1 :
478 							   local_result;
479 
480 	/* update target when applicable */
481 	if (status == VMW_BALLOON_SUCCESS &&
482 	    ((1ul << cmd) & VMW_BALLOON_CMD_WITH_TARGET_MASK))
483 		WRITE_ONCE(b->target, local_result);
484 
485 	if (status != VMW_BALLOON_SUCCESS &&
486 	    status != VMW_BALLOON_SUCCESS_WITH_CAPABILITIES) {
487 		vmballoon_stats_op_inc(b, cmd, VMW_BALLOON_OP_FAIL_STAT);
488 		pr_debug("%s: %s [0x%lx,0x%lx) failed, returned %ld\n",
489 			 __func__, vmballoon_cmd_names[cmd], arg1, arg2,
490 			 status);
491 	}
492 
493 	/* mark reset required accordingly */
494 	if (status == VMW_BALLOON_ERROR_RESET)
495 		b->reset_required = true;
496 
497 	return status;
498 }
499 
500 static __always_inline unsigned long
vmballoon_cmd(struct vmballoon * b,unsigned long cmd,unsigned long arg1,unsigned long arg2)501 vmballoon_cmd(struct vmballoon *b, unsigned long cmd, unsigned long arg1,
502 	      unsigned long arg2)
503 {
504 	unsigned long dummy;
505 
506 	return __vmballoon_cmd(b, cmd, arg1, arg2, &dummy);
507 }
508 
509 /*
510  * Send "start" command to the host, communicating supported version
511  * of the protocol.
512  */
vmballoon_send_start(struct vmballoon * b,unsigned long req_caps)513 static int vmballoon_send_start(struct vmballoon *b, unsigned long req_caps)
514 {
515 	unsigned long status, capabilities;
516 
517 	status = __vmballoon_cmd(b, VMW_BALLOON_CMD_START, req_caps, 0,
518 				 &capabilities);
519 
520 	switch (status) {
521 	case VMW_BALLOON_SUCCESS_WITH_CAPABILITIES:
522 		b->capabilities = capabilities;
523 		break;
524 	case VMW_BALLOON_SUCCESS:
525 		b->capabilities = VMW_BALLOON_BASIC_CMDS;
526 		break;
527 	default:
528 		return -EIO;
529 	}
530 
531 	/*
532 	 * 2MB pages are only supported with batching. If batching is for some
533 	 * reason disabled, do not use 2MB pages, since otherwise the legacy
534 	 * mechanism is used with 2MB pages, causing a failure.
535 	 */
536 	b->max_page_size = VMW_BALLOON_4K_PAGE;
537 	if ((b->capabilities & VMW_BALLOON_BATCHED_2M_CMDS) &&
538 	    (b->capabilities & VMW_BALLOON_BATCHED_CMDS))
539 		b->max_page_size = VMW_BALLOON_2M_PAGE;
540 
541 
542 	return 0;
543 }
544 
545 /**
546  * vmballoon_send_guest_id - communicate guest type to the host.
547  *
548  * @b: pointer to the balloon.
549  *
550  * Communicate guest type to the host so that it can adjust ballooning
551  * algorithm to the one most appropriate for the guest. This command
552  * is normally issued after sending "start" command and is part of
553  * standard reset sequence.
554  *
555  * Return: zero on success or appropriate error code.
556  */
vmballoon_send_guest_id(struct vmballoon * b)557 static int vmballoon_send_guest_id(struct vmballoon *b)
558 {
559 	unsigned long status;
560 
561 	status = vmballoon_cmd(b, VMW_BALLOON_CMD_GUEST_ID,
562 			       VMW_BALLOON_GUEST_ID, 0);
563 
564 	return status == VMW_BALLOON_SUCCESS ? 0 : -EIO;
565 }
566 
567 /**
568  * vmballoon_page_order() - return the order of the page
569  * @page_size: the size of the page.
570  *
571  * Return: the allocation order.
572  */
573 static inline
vmballoon_page_order(enum vmballoon_page_size_type page_size)574 unsigned int vmballoon_page_order(enum vmballoon_page_size_type page_size)
575 {
576 	return page_size == VMW_BALLOON_2M_PAGE ? VMW_BALLOON_2M_ORDER : 0;
577 }
578 
579 /**
580  * vmballoon_page_in_frames() - returns the number of frames in a page.
581  * @page_size: the size of the page.
582  *
583  * Return: the number of 4k frames.
584  */
585 static inline unsigned int
vmballoon_page_in_frames(enum vmballoon_page_size_type page_size)586 vmballoon_page_in_frames(enum vmballoon_page_size_type page_size)
587 {
588 	return 1 << vmballoon_page_order(page_size);
589 }
590 
591 /**
592  * vmballoon_mark_page_offline() - mark a page as offline
593  * @page: pointer for the page.
594  * @page_size: the size of the page.
595  */
596 static void
vmballoon_mark_page_offline(struct page * page,enum vmballoon_page_size_type page_size)597 vmballoon_mark_page_offline(struct page *page,
598 			    enum vmballoon_page_size_type page_size)
599 {
600 	int i;
601 
602 	for (i = 0; i < vmballoon_page_in_frames(page_size); i++)
603 		__SetPageOffline(page + i);
604 }
605 
606 /**
607  * vmballoon_mark_page_online() - mark a page as online
608  * @page: pointer for the page.
609  * @page_size: the size of the page.
610  */
611 static void
vmballoon_mark_page_online(struct page * page,enum vmballoon_page_size_type page_size)612 vmballoon_mark_page_online(struct page *page,
613 			   enum vmballoon_page_size_type page_size)
614 {
615 	int i;
616 
617 	for (i = 0; i < vmballoon_page_in_frames(page_size); i++)
618 		__ClearPageOffline(page + i);
619 }
620 
621 /**
622  * vmballoon_send_get_target() - Retrieve desired balloon size from the host.
623  *
624  * @b: pointer to the balloon.
625  *
626  * Return: zero on success, EINVAL if limit does not fit in 32-bit, as required
627  * by the host-guest protocol and EIO if an error occurred in communicating with
628  * the host.
629  */
vmballoon_send_get_target(struct vmballoon * b)630 static int vmballoon_send_get_target(struct vmballoon *b)
631 {
632 	unsigned long status;
633 	unsigned long limit;
634 
635 	limit = totalram_pages();
636 
637 	/* Ensure limit fits in 32-bits if 64-bit targets are not supported */
638 	if (!(b->capabilities & VMW_BALLOON_64_BIT_TARGET) &&
639 	    limit != (u32)limit)
640 		return -EINVAL;
641 
642 	status = vmballoon_cmd(b, VMW_BALLOON_CMD_GET_TARGET, limit, 0);
643 
644 	return status == VMW_BALLOON_SUCCESS ? 0 : -EIO;
645 }
646 
647 /**
648  * vmballoon_alloc_page_list - allocates a list of pages.
649  *
650  * @b: pointer to the balloon.
651  * @ctl: pointer for the %struct vmballoon_ctl, which defines the operation.
652  * @req_n_pages: the number of requested pages.
653  *
654  * Tries to allocate @req_n_pages. Add them to the list of balloon pages in
655  * @ctl.pages and updates @ctl.n_pages to reflect the number of pages.
656  *
657  * Return: zero on success or error code otherwise.
658  */
vmballoon_alloc_page_list(struct vmballoon * b,struct vmballoon_ctl * ctl,unsigned int req_n_pages)659 static int vmballoon_alloc_page_list(struct vmballoon *b,
660 				     struct vmballoon_ctl *ctl,
661 				     unsigned int req_n_pages)
662 {
663 	struct page *page;
664 	unsigned int i;
665 
666 	for (i = 0; i < req_n_pages; i++) {
667 		/*
668 		 * First check if we happen to have pages that were allocated
669 		 * before. This happens when 2MB page rejected during inflation
670 		 * by the hypervisor, and then split into 4KB pages.
671 		 */
672 		if (!list_empty(&ctl->prealloc_pages)) {
673 			page = list_first_entry(&ctl->prealloc_pages,
674 						struct page, lru);
675 			list_del(&page->lru);
676 		} else {
677 			if (ctl->page_size == VMW_BALLOON_2M_PAGE)
678 				page = alloc_pages(__GFP_HIGHMEM|__GFP_NOWARN|
679 					__GFP_NOMEMALLOC, VMW_BALLOON_2M_ORDER);
680 			else
681 				page = balloon_page_alloc();
682 
683 			vmballoon_stats_page_inc(b, VMW_BALLOON_PAGE_STAT_ALLOC,
684 						 ctl->page_size);
685 		}
686 
687 		if (page) {
688 			/* Success. Add the page to the list and continue. */
689 			list_add(&page->lru, &ctl->pages);
690 			continue;
691 		}
692 
693 		/* Allocation failed. Update statistics and stop. */
694 		vmballoon_stats_page_inc(b, VMW_BALLOON_PAGE_STAT_ALLOC_FAIL,
695 					 ctl->page_size);
696 		break;
697 	}
698 
699 	ctl->n_pages = i;
700 
701 	return req_n_pages == ctl->n_pages ? 0 : -ENOMEM;
702 }
703 
704 /**
705  * vmballoon_handle_one_result - Handle lock/unlock result for a single page.
706  *
707  * @b: pointer for %struct vmballoon.
708  * @page: pointer for the page whose result should be handled.
709  * @page_size: size of the page.
710  * @status: status of the operation as provided by the hypervisor.
711  */
vmballoon_handle_one_result(struct vmballoon * b,struct page * page,enum vmballoon_page_size_type page_size,unsigned long status)712 static int vmballoon_handle_one_result(struct vmballoon *b, struct page *page,
713 				       enum vmballoon_page_size_type page_size,
714 				       unsigned long status)
715 {
716 	/* On success do nothing. The page is already on the balloon list. */
717 	if (likely(status == VMW_BALLOON_SUCCESS))
718 		return 0;
719 
720 	pr_debug("%s: failed comm pfn %lx status %lu page_size %s\n", __func__,
721 		 page_to_pfn(page), status,
722 		 vmballoon_page_size_names[page_size]);
723 
724 	/* Error occurred */
725 	vmballoon_stats_page_inc(b, VMW_BALLOON_PAGE_STAT_REFUSED_ALLOC,
726 				 page_size);
727 
728 	return -EIO;
729 }
730 
731 /**
732  * vmballoon_status_page - returns the status of (un)lock operation
733  *
734  * @b: pointer to the balloon.
735  * @idx: index for the page for which the operation is performed.
736  * @p: pointer to where the page struct is returned.
737  *
738  * Following a lock or unlock operation, returns the status of the operation for
739  * an individual page. Provides the page that the operation was performed on on
740  * the @page argument.
741  *
742  * Returns: The status of a lock or unlock operation for an individual page.
743  */
vmballoon_status_page(struct vmballoon * b,int idx,struct page ** p)744 static unsigned long vmballoon_status_page(struct vmballoon *b, int idx,
745 					   struct page **p)
746 {
747 	if (static_branch_likely(&vmw_balloon_batching)) {
748 		/* batching mode */
749 		*p = pfn_to_page(b->batch_page[idx].pfn);
750 		return b->batch_page[idx].status;
751 	}
752 
753 	/* non-batching mode */
754 	*p = b->page;
755 
756 	/*
757 	 * If a failure occurs, the indication will be provided in the status
758 	 * of the entire operation, which is considered before the individual
759 	 * page status. So for non-batching mode, the indication is always of
760 	 * success.
761 	 */
762 	return VMW_BALLOON_SUCCESS;
763 }
764 
765 /**
766  * vmballoon_lock_op - notifies the host about inflated/deflated pages.
767  * @b: pointer to the balloon.
768  * @num_pages: number of inflated/deflated pages.
769  * @page_size: size of the page.
770  * @op: the type of operation (lock or unlock).
771  *
772  * Notify the host about page(s) that were ballooned (or removed from the
773  * balloon) so that host can use it without fear that guest will need it (or
774  * stop using them since the VM does). Host may reject some pages, we need to
775  * check the return value and maybe submit a different page. The pages that are
776  * inflated/deflated are pointed by @b->page.
777  *
778  * Return: result as provided by the hypervisor.
779  */
vmballoon_lock_op(struct vmballoon * b,unsigned int num_pages,enum vmballoon_page_size_type page_size,enum vmballoon_op op)780 static unsigned long vmballoon_lock_op(struct vmballoon *b,
781 				       unsigned int num_pages,
782 				       enum vmballoon_page_size_type page_size,
783 				       enum vmballoon_op op)
784 {
785 	unsigned long cmd, pfn;
786 
787 	lockdep_assert_held(&b->comm_lock);
788 
789 	if (static_branch_likely(&vmw_balloon_batching)) {
790 		if (op == VMW_BALLOON_INFLATE)
791 			cmd = page_size == VMW_BALLOON_2M_PAGE ?
792 				VMW_BALLOON_CMD_BATCHED_2M_LOCK :
793 				VMW_BALLOON_CMD_BATCHED_LOCK;
794 		else
795 			cmd = page_size == VMW_BALLOON_2M_PAGE ?
796 				VMW_BALLOON_CMD_BATCHED_2M_UNLOCK :
797 				VMW_BALLOON_CMD_BATCHED_UNLOCK;
798 
799 		pfn = PHYS_PFN(virt_to_phys(b->batch_page));
800 	} else {
801 		cmd = op == VMW_BALLOON_INFLATE ? VMW_BALLOON_CMD_LOCK :
802 						  VMW_BALLOON_CMD_UNLOCK;
803 		pfn = page_to_pfn(b->page);
804 
805 		/* In non-batching mode, PFNs must fit in 32-bit */
806 		if (unlikely(pfn != (u32)pfn))
807 			return VMW_BALLOON_ERROR_PPN_INVALID;
808 	}
809 
810 	return vmballoon_cmd(b, cmd, pfn, num_pages);
811 }
812 
813 /**
814  * vmballoon_add_page - adds a page towards lock/unlock operation.
815  *
816  * @b: pointer to the balloon.
817  * @idx: index of the page to be ballooned in this batch.
818  * @p: pointer to the page that is about to be ballooned.
819  *
820  * Adds the page to be ballooned. Must be called while holding @comm_lock.
821  */
vmballoon_add_page(struct vmballoon * b,unsigned int idx,struct page * p)822 static void vmballoon_add_page(struct vmballoon *b, unsigned int idx,
823 			       struct page *p)
824 {
825 	lockdep_assert_held(&b->comm_lock);
826 
827 	if (static_branch_likely(&vmw_balloon_batching))
828 		b->batch_page[idx] = (struct vmballoon_batch_entry)
829 					{ .pfn = page_to_pfn(p) };
830 	else
831 		b->page = p;
832 }
833 
834 /**
835  * vmballoon_lock - lock or unlock a batch of pages.
836  *
837  * @b: pointer to the balloon.
838  * @ctl: pointer for the %struct vmballoon_ctl, which defines the operation.
839  *
840  * Notifies the host of about ballooned pages (after inflation or deflation,
841  * according to @ctl). If the host rejects the page put it on the
842  * @ctl refuse list. These refused page are then released when moving to the
843  * next size of pages.
844  *
845  * Note that we neither free any @page here nor put them back on the ballooned
846  * pages list. Instead we queue it for later processing. We do that for several
847  * reasons. First, we do not want to free the page under the lock. Second, it
848  * allows us to unify the handling of lock and unlock. In the inflate case, the
849  * caller will check if there are too many refused pages and release them.
850  * Although it is not identical to the past behavior, it should not affect
851  * performance.
852  */
vmballoon_lock(struct vmballoon * b,struct vmballoon_ctl * ctl)853 static int vmballoon_lock(struct vmballoon *b, struct vmballoon_ctl *ctl)
854 {
855 	unsigned long batch_status;
856 	struct page *page;
857 	unsigned int i, num_pages;
858 
859 	num_pages = ctl->n_pages;
860 	if (num_pages == 0)
861 		return 0;
862 
863 	/* communication with the host is done under the communication lock */
864 	spin_lock(&b->comm_lock);
865 
866 	i = 0;
867 	list_for_each_entry(page, &ctl->pages, lru)
868 		vmballoon_add_page(b, i++, page);
869 
870 	batch_status = vmballoon_lock_op(b, ctl->n_pages, ctl->page_size,
871 					 ctl->op);
872 
873 	/*
874 	 * Iterate over the pages in the provided list. Since we are changing
875 	 * @ctl->n_pages we are saving the original value in @num_pages and
876 	 * use this value to bound the loop.
877 	 */
878 	for (i = 0; i < num_pages; i++) {
879 		unsigned long status;
880 
881 		status = vmballoon_status_page(b, i, &page);
882 
883 		/*
884 		 * Failure of the whole batch overrides a single operation
885 		 * results.
886 		 */
887 		if (batch_status != VMW_BALLOON_SUCCESS)
888 			status = batch_status;
889 
890 		/* Continue if no error happened */
891 		if (!vmballoon_handle_one_result(b, page, ctl->page_size,
892 						 status))
893 			continue;
894 
895 		/*
896 		 * Error happened. Move the pages to the refused list and update
897 		 * the pages number.
898 		 */
899 		list_move(&page->lru, &ctl->refused_pages);
900 		ctl->n_pages--;
901 		ctl->n_refused_pages++;
902 	}
903 
904 	spin_unlock(&b->comm_lock);
905 
906 	return batch_status == VMW_BALLOON_SUCCESS ? 0 : -EIO;
907 }
908 
909 /**
910  * vmballoon_release_page_list() - Releases a page list
911  *
912  * @page_list: list of pages to release.
913  * @n_pages: pointer to the number of pages.
914  * @page_size: whether the pages in the list are 2MB (or else 4KB).
915  *
916  * Releases the list of pages and zeros the number of pages.
917  */
vmballoon_release_page_list(struct list_head * page_list,int * n_pages,enum vmballoon_page_size_type page_size)918 static void vmballoon_release_page_list(struct list_head *page_list,
919 				       int *n_pages,
920 				       enum vmballoon_page_size_type page_size)
921 {
922 	struct page *page, *tmp;
923 
924 	list_for_each_entry_safe(page, tmp, page_list, lru) {
925 		list_del(&page->lru);
926 		__free_pages(page, vmballoon_page_order(page_size));
927 	}
928 
929 	if (n_pages)
930 		*n_pages = 0;
931 }
932 
933 
934 /*
935  * Release pages that were allocated while attempting to inflate the
936  * balloon but were refused by the host for one reason or another.
937  */
vmballoon_release_refused_pages(struct vmballoon * b,struct vmballoon_ctl * ctl)938 static void vmballoon_release_refused_pages(struct vmballoon *b,
939 					    struct vmballoon_ctl *ctl)
940 {
941 	vmballoon_stats_page_inc(b, VMW_BALLOON_PAGE_STAT_REFUSED_FREE,
942 				 ctl->page_size);
943 
944 	vmballoon_release_page_list(&ctl->refused_pages, &ctl->n_refused_pages,
945 				    ctl->page_size);
946 }
947 
948 /**
949  * vmballoon_change - retrieve the required balloon change
950  *
951  * @b: pointer for the balloon.
952  *
953  * Return: the required change for the balloon size. A positive number
954  * indicates inflation, a negative number indicates a deflation.
955  */
vmballoon_change(struct vmballoon * b)956 static int64_t vmballoon_change(struct vmballoon *b)
957 {
958 	int64_t size, target;
959 
960 	size = atomic64_read(&b->size);
961 	target = READ_ONCE(b->target);
962 
963 	/*
964 	 * We must cast first because of int sizes
965 	 * Otherwise we might get huge positives instead of negatives
966 	 */
967 
968 	if (b->reset_required)
969 		return 0;
970 
971 	/* consider a 2MB slack on deflate, unless the balloon is emptied */
972 	if (target < size && target != 0 &&
973 	    size - target < vmballoon_page_in_frames(VMW_BALLOON_2M_PAGE))
974 		return 0;
975 
976 	/* If an out-of-memory recently occurred, inflation is disallowed. */
977 	if (target > size && time_before(jiffies, READ_ONCE(b->shrink_timeout)))
978 		return 0;
979 
980 	return target - size;
981 }
982 
983 /**
984  * vmballoon_enqueue_page_list() - Enqueues list of pages after inflation.
985  *
986  * @b: pointer to balloon.
987  * @pages: list of pages to enqueue.
988  * @n_pages: pointer to number of pages in list. The value is zeroed.
989  * @page_size: whether the pages are 2MB or 4KB pages.
990  *
991  * Enqueues the provides list of pages in the ballooned page list, clears the
992  * list and zeroes the number of pages that was provided.
993  */
vmballoon_enqueue_page_list(struct vmballoon * b,struct list_head * pages,unsigned int * n_pages,enum vmballoon_page_size_type page_size)994 static void vmballoon_enqueue_page_list(struct vmballoon *b,
995 					struct list_head *pages,
996 					unsigned int *n_pages,
997 					enum vmballoon_page_size_type page_size)
998 {
999 	unsigned long flags;
1000 	struct page *page;
1001 
1002 	if (page_size == VMW_BALLOON_4K_PAGE) {
1003 		balloon_page_list_enqueue(&b->b_dev_info, pages);
1004 	} else {
1005 		/*
1006 		 * Keep the huge pages in a local list which is not available
1007 		 * for the balloon compaction mechanism.
1008 		 */
1009 		spin_lock_irqsave(&b->b_dev_info.pages_lock, flags);
1010 
1011 		list_for_each_entry(page, pages, lru) {
1012 			vmballoon_mark_page_offline(page, VMW_BALLOON_2M_PAGE);
1013 		}
1014 
1015 		list_splice_init(pages, &b->huge_pages);
1016 		__count_vm_events(BALLOON_INFLATE, *n_pages *
1017 				  vmballoon_page_in_frames(VMW_BALLOON_2M_PAGE));
1018 		spin_unlock_irqrestore(&b->b_dev_info.pages_lock, flags);
1019 	}
1020 
1021 	*n_pages = 0;
1022 }
1023 
1024 /**
1025  * vmballoon_dequeue_page_list() - Dequeues page lists for deflation.
1026  *
1027  * @b: pointer to balloon.
1028  * @pages: list of pages to enqueue.
1029  * @n_pages: pointer to number of pages in list. The value is zeroed.
1030  * @page_size: whether the pages are 2MB or 4KB pages.
1031  * @n_req_pages: the number of requested pages.
1032  *
1033  * Dequeues the number of requested pages from the balloon for deflation. The
1034  * number of dequeued pages may be lower, if not enough pages in the requested
1035  * size are available.
1036  */
vmballoon_dequeue_page_list(struct vmballoon * b,struct list_head * pages,unsigned int * n_pages,enum vmballoon_page_size_type page_size,unsigned int n_req_pages)1037 static void vmballoon_dequeue_page_list(struct vmballoon *b,
1038 					struct list_head *pages,
1039 					unsigned int *n_pages,
1040 					enum vmballoon_page_size_type page_size,
1041 					unsigned int n_req_pages)
1042 {
1043 	struct page *page, *tmp;
1044 	unsigned int i = 0;
1045 	unsigned long flags;
1046 
1047 	/* In the case of 4k pages, use the compaction infrastructure */
1048 	if (page_size == VMW_BALLOON_4K_PAGE) {
1049 		*n_pages = balloon_page_list_dequeue(&b->b_dev_info, pages,
1050 						     n_req_pages);
1051 		return;
1052 	}
1053 
1054 	/* 2MB pages */
1055 	spin_lock_irqsave(&b->b_dev_info.pages_lock, flags);
1056 	list_for_each_entry_safe(page, tmp, &b->huge_pages, lru) {
1057 		vmballoon_mark_page_online(page, VMW_BALLOON_2M_PAGE);
1058 
1059 		list_move(&page->lru, pages);
1060 		if (++i == n_req_pages)
1061 			break;
1062 	}
1063 
1064 	__count_vm_events(BALLOON_DEFLATE,
1065 			  i * vmballoon_page_in_frames(VMW_BALLOON_2M_PAGE));
1066 	spin_unlock_irqrestore(&b->b_dev_info.pages_lock, flags);
1067 	*n_pages = i;
1068 }
1069 
1070 /**
1071  * vmballoon_split_refused_pages() - Split the 2MB refused pages to 4k.
1072  *
1073  * If inflation of 2MB pages was denied by the hypervisor, it is likely to be
1074  * due to one or few 4KB pages. These 2MB pages may keep being allocated and
1075  * then being refused. To prevent this case, this function splits the refused
1076  * pages into 4KB pages and adds them into @prealloc_pages list.
1077  *
1078  * @ctl: pointer for the %struct vmballoon_ctl, which defines the operation.
1079  */
vmballoon_split_refused_pages(struct vmballoon_ctl * ctl)1080 static void vmballoon_split_refused_pages(struct vmballoon_ctl *ctl)
1081 {
1082 	struct page *page, *tmp;
1083 	unsigned int i, order;
1084 
1085 	order = vmballoon_page_order(ctl->page_size);
1086 
1087 	list_for_each_entry_safe(page, tmp, &ctl->refused_pages, lru) {
1088 		list_del(&page->lru);
1089 		split_page(page, order);
1090 		for (i = 0; i < (1 << order); i++)
1091 			list_add(&page[i].lru, &ctl->prealloc_pages);
1092 	}
1093 	ctl->n_refused_pages = 0;
1094 }
1095 
1096 /**
1097  * vmballoon_inflate() - Inflate the balloon towards its target size.
1098  *
1099  * @b: pointer to the balloon.
1100  */
vmballoon_inflate(struct vmballoon * b)1101 static void vmballoon_inflate(struct vmballoon *b)
1102 {
1103 	int64_t to_inflate_frames;
1104 	struct vmballoon_ctl ctl = {
1105 		.pages = LIST_HEAD_INIT(ctl.pages),
1106 		.refused_pages = LIST_HEAD_INIT(ctl.refused_pages),
1107 		.prealloc_pages = LIST_HEAD_INIT(ctl.prealloc_pages),
1108 		.page_size = b->max_page_size,
1109 		.op = VMW_BALLOON_INFLATE
1110 	};
1111 
1112 	while ((to_inflate_frames = vmballoon_change(b)) > 0) {
1113 		unsigned int to_inflate_pages, page_in_frames;
1114 		int alloc_error, lock_error = 0;
1115 
1116 		VM_BUG_ON(!list_empty(&ctl.pages));
1117 		VM_BUG_ON(ctl.n_pages != 0);
1118 
1119 		page_in_frames = vmballoon_page_in_frames(ctl.page_size);
1120 
1121 		to_inflate_pages = min_t(unsigned long, b->batch_max_pages,
1122 					 DIV_ROUND_UP_ULL(to_inflate_frames,
1123 							  page_in_frames));
1124 
1125 		/* Start by allocating */
1126 		alloc_error = vmballoon_alloc_page_list(b, &ctl,
1127 							to_inflate_pages);
1128 
1129 		/* Actually lock the pages by telling the hypervisor */
1130 		lock_error = vmballoon_lock(b, &ctl);
1131 
1132 		/*
1133 		 * If an error indicates that something serious went wrong,
1134 		 * stop the inflation.
1135 		 */
1136 		if (lock_error)
1137 			break;
1138 
1139 		/* Update the balloon size */
1140 		atomic64_add(ctl.n_pages * page_in_frames, &b->size);
1141 
1142 		vmballoon_enqueue_page_list(b, &ctl.pages, &ctl.n_pages,
1143 					    ctl.page_size);
1144 
1145 		/*
1146 		 * If allocation failed or the number of refused pages exceeds
1147 		 * the maximum allowed, move to the next page size.
1148 		 */
1149 		if (alloc_error ||
1150 		    ctl.n_refused_pages >= VMW_BALLOON_MAX_REFUSED) {
1151 			if (ctl.page_size == VMW_BALLOON_4K_PAGE)
1152 				break;
1153 
1154 			/*
1155 			 * Split the refused pages to 4k. This will also empty
1156 			 * the refused pages list.
1157 			 */
1158 			vmballoon_split_refused_pages(&ctl);
1159 			ctl.page_size--;
1160 		}
1161 
1162 		cond_resched();
1163 	}
1164 
1165 	/*
1166 	 * Release pages that were allocated while attempting to inflate the
1167 	 * balloon but were refused by the host for one reason or another,
1168 	 * and update the statistics.
1169 	 */
1170 	if (ctl.n_refused_pages != 0)
1171 		vmballoon_release_refused_pages(b, &ctl);
1172 
1173 	vmballoon_release_page_list(&ctl.prealloc_pages, NULL, ctl.page_size);
1174 }
1175 
1176 /**
1177  * vmballoon_deflate() - Decrease the size of the balloon.
1178  *
1179  * @b: pointer to the balloon
1180  * @n_frames: the number of frames to deflate. If zero, automatically
1181  * calculated according to the target size.
1182  * @coordinated: whether to coordinate with the host
1183  *
1184  * Decrease the size of the balloon allowing guest to use more memory.
1185  *
1186  * Return: The number of deflated frames (i.e., basic page size units)
1187  */
vmballoon_deflate(struct vmballoon * b,uint64_t n_frames,bool coordinated)1188 static unsigned long vmballoon_deflate(struct vmballoon *b, uint64_t n_frames,
1189 				       bool coordinated)
1190 {
1191 	unsigned long deflated_frames = 0;
1192 	unsigned long tried_frames = 0;
1193 	struct vmballoon_ctl ctl = {
1194 		.pages = LIST_HEAD_INIT(ctl.pages),
1195 		.refused_pages = LIST_HEAD_INIT(ctl.refused_pages),
1196 		.page_size = VMW_BALLOON_4K_PAGE,
1197 		.op = VMW_BALLOON_DEFLATE
1198 	};
1199 
1200 	/* free pages to reach target */
1201 	while (true) {
1202 		unsigned int to_deflate_pages, n_unlocked_frames;
1203 		unsigned int page_in_frames;
1204 		int64_t to_deflate_frames;
1205 		bool deflated_all;
1206 
1207 		page_in_frames = vmballoon_page_in_frames(ctl.page_size);
1208 
1209 		VM_BUG_ON(!list_empty(&ctl.pages));
1210 		VM_BUG_ON(ctl.n_pages);
1211 		VM_BUG_ON(!list_empty(&ctl.refused_pages));
1212 		VM_BUG_ON(ctl.n_refused_pages);
1213 
1214 		/*
1215 		 * If we were requested a specific number of frames, we try to
1216 		 * deflate this number of frames. Otherwise, deflation is
1217 		 * performed according to the target and balloon size.
1218 		 */
1219 		to_deflate_frames = n_frames ? n_frames - tried_frames :
1220 					       -vmballoon_change(b);
1221 
1222 		/* break if no work to do */
1223 		if (to_deflate_frames <= 0)
1224 			break;
1225 
1226 		/*
1227 		 * Calculate the number of frames based on current page size,
1228 		 * but limit the deflated frames to a single chunk
1229 		 */
1230 		to_deflate_pages = min_t(unsigned long, b->batch_max_pages,
1231 					 DIV_ROUND_UP_ULL(to_deflate_frames,
1232 							  page_in_frames));
1233 
1234 		/* First take the pages from the balloon pages. */
1235 		vmballoon_dequeue_page_list(b, &ctl.pages, &ctl.n_pages,
1236 					    ctl.page_size, to_deflate_pages);
1237 
1238 		/*
1239 		 * Before pages are moving to the refused list, count their
1240 		 * frames as frames that we tried to deflate.
1241 		 */
1242 		tried_frames += ctl.n_pages * page_in_frames;
1243 
1244 		/*
1245 		 * Unlock the pages by communicating with the hypervisor if the
1246 		 * communication is coordinated (i.e., not pop). We ignore the
1247 		 * return code. Instead we check if all the pages we manage to
1248 		 * unlock all the pages. If we failed, we will move to the next
1249 		 * page size, and would eventually try again later.
1250 		 */
1251 		if (coordinated)
1252 			vmballoon_lock(b, &ctl);
1253 
1254 		/*
1255 		 * Check if we deflated enough. We will move to the next page
1256 		 * size if we did not manage to do so. This calculation takes
1257 		 * place now, as once the pages are released, the number of
1258 		 * pages is zeroed.
1259 		 */
1260 		deflated_all = (ctl.n_pages == to_deflate_pages);
1261 
1262 		/* Update local and global counters */
1263 		n_unlocked_frames = ctl.n_pages * page_in_frames;
1264 		atomic64_sub(n_unlocked_frames, &b->size);
1265 		deflated_frames += n_unlocked_frames;
1266 
1267 		vmballoon_stats_page_add(b, VMW_BALLOON_PAGE_STAT_FREE,
1268 					 ctl.page_size, ctl.n_pages);
1269 
1270 		/* free the ballooned pages */
1271 		vmballoon_release_page_list(&ctl.pages, &ctl.n_pages,
1272 					    ctl.page_size);
1273 
1274 		/* Return the refused pages to the ballooned list. */
1275 		vmballoon_enqueue_page_list(b, &ctl.refused_pages,
1276 					    &ctl.n_refused_pages,
1277 					    ctl.page_size);
1278 
1279 		/* If we failed to unlock all the pages, move to next size. */
1280 		if (!deflated_all) {
1281 			if (ctl.page_size == b->max_page_size)
1282 				break;
1283 			ctl.page_size++;
1284 		}
1285 
1286 		cond_resched();
1287 	}
1288 
1289 	return deflated_frames;
1290 }
1291 
1292 /**
1293  * vmballoon_deinit_batching - disables batching mode.
1294  *
1295  * @b: pointer to &struct vmballoon.
1296  *
1297  * Disables batching, by deallocating the page for communication with the
1298  * hypervisor and disabling the static key to indicate that batching is off.
1299  */
vmballoon_deinit_batching(struct vmballoon * b)1300 static void vmballoon_deinit_batching(struct vmballoon *b)
1301 {
1302 	free_page((unsigned long)b->batch_page);
1303 	b->batch_page = NULL;
1304 	static_branch_disable(&vmw_balloon_batching);
1305 	b->batch_max_pages = 1;
1306 }
1307 
1308 /**
1309  * vmballoon_init_batching - enable batching mode.
1310  *
1311  * @b: pointer to &struct vmballoon.
1312  *
1313  * Enables batching, by allocating a page for communication with the hypervisor
1314  * and enabling the static_key to use batching.
1315  *
1316  * Return: zero on success or an appropriate error-code.
1317  */
vmballoon_init_batching(struct vmballoon * b)1318 static int vmballoon_init_batching(struct vmballoon *b)
1319 {
1320 	struct page *page;
1321 
1322 	page = alloc_page(GFP_KERNEL | __GFP_ZERO);
1323 	if (!page)
1324 		return -ENOMEM;
1325 
1326 	b->batch_page = page_address(page);
1327 	b->batch_max_pages = PAGE_SIZE / sizeof(struct vmballoon_batch_entry);
1328 
1329 	static_branch_enable(&vmw_balloon_batching);
1330 
1331 	return 0;
1332 }
1333 
1334 /*
1335  * Receive notification and resize balloon
1336  */
vmballoon_doorbell(void * client_data)1337 static void vmballoon_doorbell(void *client_data)
1338 {
1339 	struct vmballoon *b = client_data;
1340 
1341 	vmballoon_stats_gen_inc(b, VMW_BALLOON_STAT_DOORBELL);
1342 
1343 	mod_delayed_work(system_freezable_wq, &b->dwork, 0);
1344 }
1345 
1346 /*
1347  * Clean up vmci doorbell
1348  */
vmballoon_vmci_cleanup(struct vmballoon * b)1349 static void vmballoon_vmci_cleanup(struct vmballoon *b)
1350 {
1351 	vmballoon_cmd(b, VMW_BALLOON_CMD_VMCI_DOORBELL_SET,
1352 		      VMCI_INVALID_ID, VMCI_INVALID_ID);
1353 
1354 	if (!vmci_handle_is_invalid(b->vmci_doorbell)) {
1355 		vmci_doorbell_destroy(b->vmci_doorbell);
1356 		b->vmci_doorbell = VMCI_INVALID_HANDLE;
1357 	}
1358 }
1359 
1360 /**
1361  * vmballoon_vmci_init - Initialize vmci doorbell.
1362  *
1363  * @b: pointer to the balloon.
1364  *
1365  * Return: zero on success or when wakeup command not supported. Error-code
1366  * otherwise.
1367  *
1368  * Initialize vmci doorbell, to get notified as soon as balloon changes.
1369  */
vmballoon_vmci_init(struct vmballoon * b)1370 static int vmballoon_vmci_init(struct vmballoon *b)
1371 {
1372 	unsigned long error;
1373 
1374 	if ((b->capabilities & VMW_BALLOON_SIGNALLED_WAKEUP_CMD) == 0)
1375 		return 0;
1376 
1377 	error = vmci_doorbell_create(&b->vmci_doorbell, VMCI_FLAG_DELAYED_CB,
1378 				     VMCI_PRIVILEGE_FLAG_RESTRICTED,
1379 				     vmballoon_doorbell, b);
1380 
1381 	if (error != VMCI_SUCCESS)
1382 		goto fail;
1383 
1384 	error =	__vmballoon_cmd(b, VMW_BALLOON_CMD_VMCI_DOORBELL_SET,
1385 				b->vmci_doorbell.context,
1386 				b->vmci_doorbell.resource, NULL);
1387 
1388 	if (error != VMW_BALLOON_SUCCESS)
1389 		goto fail;
1390 
1391 	return 0;
1392 fail:
1393 	vmballoon_vmci_cleanup(b);
1394 	return -EIO;
1395 }
1396 
1397 /**
1398  * vmballoon_pop - Quickly release all pages allocate for the balloon.
1399  *
1400  * @b: pointer to the balloon.
1401  *
1402  * This function is called when host decides to "reset" balloon for one reason
1403  * or another. Unlike normal "deflate" we do not (shall not) notify host of the
1404  * pages being released.
1405  */
vmballoon_pop(struct vmballoon * b)1406 static void vmballoon_pop(struct vmballoon *b)
1407 {
1408 	unsigned long size;
1409 
1410 	while ((size = atomic64_read(&b->size)))
1411 		vmballoon_deflate(b, size, false);
1412 }
1413 
1414 /*
1415  * Perform standard reset sequence by popping the balloon (in case it
1416  * is not  empty) and then restarting protocol. This operation normally
1417  * happens when host responds with VMW_BALLOON_ERROR_RESET to a command.
1418  */
vmballoon_reset(struct vmballoon * b)1419 static void vmballoon_reset(struct vmballoon *b)
1420 {
1421 	int error;
1422 
1423 	down_write(&b->conf_sem);
1424 
1425 	vmballoon_vmci_cleanup(b);
1426 
1427 	/* free all pages, skipping monitor unlock */
1428 	vmballoon_pop(b);
1429 
1430 	if (vmballoon_send_start(b, VMW_BALLOON_CAPABILITIES))
1431 		goto unlock;
1432 
1433 	if ((b->capabilities & VMW_BALLOON_BATCHED_CMDS) != 0) {
1434 		if (vmballoon_init_batching(b)) {
1435 			/*
1436 			 * We failed to initialize batching, inform the monitor
1437 			 * about it by sending a null capability.
1438 			 *
1439 			 * The guest will retry in one second.
1440 			 */
1441 			vmballoon_send_start(b, 0);
1442 			goto unlock;
1443 		}
1444 	} else if ((b->capabilities & VMW_BALLOON_BASIC_CMDS) != 0) {
1445 		vmballoon_deinit_batching(b);
1446 	}
1447 
1448 	vmballoon_stats_gen_inc(b, VMW_BALLOON_STAT_RESET);
1449 	b->reset_required = false;
1450 
1451 	error = vmballoon_vmci_init(b);
1452 	if (error)
1453 		pr_err_once("failed to initialize vmci doorbell\n");
1454 
1455 	if (vmballoon_send_guest_id(b))
1456 		pr_err_once("failed to send guest ID to the host\n");
1457 
1458 unlock:
1459 	up_write(&b->conf_sem);
1460 }
1461 
1462 /**
1463  * vmballoon_work - periodic balloon worker for reset, inflation and deflation.
1464  *
1465  * @work: pointer to the &work_struct which is provided by the workqueue.
1466  *
1467  * Resets the protocol if needed, gets the new size and adjusts balloon as
1468  * needed. Repeat in 1 sec.
1469  */
vmballoon_work(struct work_struct * work)1470 static void vmballoon_work(struct work_struct *work)
1471 {
1472 	struct delayed_work *dwork = to_delayed_work(work);
1473 	struct vmballoon *b = container_of(dwork, struct vmballoon, dwork);
1474 	int64_t change = 0;
1475 
1476 	if (b->reset_required)
1477 		vmballoon_reset(b);
1478 
1479 	down_read(&b->conf_sem);
1480 
1481 	/*
1482 	 * Update the stats while holding the semaphore to ensure that
1483 	 * @stats_enabled is consistent with whether the stats are actually
1484 	 * enabled
1485 	 */
1486 	vmballoon_stats_gen_inc(b, VMW_BALLOON_STAT_TIMER);
1487 
1488 	if (!vmballoon_send_get_target(b))
1489 		change = vmballoon_change(b);
1490 
1491 	if (change != 0) {
1492 		pr_debug("%s - size: %llu, target %lu\n", __func__,
1493 			 atomic64_read(&b->size), READ_ONCE(b->target));
1494 
1495 		if (change > 0)
1496 			vmballoon_inflate(b);
1497 		else  /* (change < 0) */
1498 			vmballoon_deflate(b, 0, true);
1499 	}
1500 
1501 	up_read(&b->conf_sem);
1502 
1503 	/*
1504 	 * We are using a freezable workqueue so that balloon operations are
1505 	 * stopped while the system transitions to/from sleep/hibernation.
1506 	 */
1507 	queue_delayed_work(system_freezable_wq,
1508 			   dwork, round_jiffies_relative(HZ));
1509 
1510 }
1511 
1512 /**
1513  * vmballoon_shrinker_scan() - deflate the balloon due to memory pressure.
1514  * @shrinker: pointer to the balloon shrinker.
1515  * @sc: page reclaim information.
1516  *
1517  * Returns: number of pages that were freed during deflation.
1518  */
vmballoon_shrinker_scan(struct shrinker * shrinker,struct shrink_control * sc)1519 static unsigned long vmballoon_shrinker_scan(struct shrinker *shrinker,
1520 					     struct shrink_control *sc)
1521 {
1522 	struct vmballoon *b = &balloon;
1523 	unsigned long deflated_frames;
1524 
1525 	pr_debug("%s - size: %llu", __func__, atomic64_read(&b->size));
1526 
1527 	vmballoon_stats_gen_inc(b, VMW_BALLOON_STAT_SHRINK);
1528 
1529 	/*
1530 	 * If the lock is also contended for read, we cannot easily reclaim and
1531 	 * we bail out.
1532 	 */
1533 	if (!down_read_trylock(&b->conf_sem))
1534 		return 0;
1535 
1536 	deflated_frames = vmballoon_deflate(b, sc->nr_to_scan, true);
1537 
1538 	vmballoon_stats_gen_add(b, VMW_BALLOON_STAT_SHRINK_FREE,
1539 				deflated_frames);
1540 
1541 	/*
1542 	 * Delay future inflation for some time to mitigate the situations in
1543 	 * which balloon continuously grows and shrinks. Use WRITE_ONCE() since
1544 	 * the access is asynchronous.
1545 	 */
1546 	WRITE_ONCE(b->shrink_timeout, jiffies + HZ * VMBALLOON_SHRINK_DELAY);
1547 
1548 	up_read(&b->conf_sem);
1549 
1550 	return deflated_frames;
1551 }
1552 
1553 /**
1554  * vmballoon_shrinker_count() - return the number of ballooned pages.
1555  * @shrinker: pointer to the balloon shrinker.
1556  * @sc: page reclaim information.
1557  *
1558  * Returns: number of 4k pages that are allocated for the balloon and can
1559  *	    therefore be reclaimed under pressure.
1560  */
vmballoon_shrinker_count(struct shrinker * shrinker,struct shrink_control * sc)1561 static unsigned long vmballoon_shrinker_count(struct shrinker *shrinker,
1562 					      struct shrink_control *sc)
1563 {
1564 	struct vmballoon *b = &balloon;
1565 
1566 	return atomic64_read(&b->size);
1567 }
1568 
vmballoon_unregister_shrinker(struct vmballoon * b)1569 static void vmballoon_unregister_shrinker(struct vmballoon *b)
1570 {
1571 	if (b->shrinker_registered)
1572 		unregister_shrinker(&b->shrinker);
1573 	b->shrinker_registered = false;
1574 }
1575 
vmballoon_register_shrinker(struct vmballoon * b)1576 static int vmballoon_register_shrinker(struct vmballoon *b)
1577 {
1578 	int r;
1579 
1580 	/* Do nothing if the shrinker is not enabled */
1581 	if (!vmwballoon_shrinker_enable)
1582 		return 0;
1583 
1584 	b->shrinker.scan_objects = vmballoon_shrinker_scan;
1585 	b->shrinker.count_objects = vmballoon_shrinker_count;
1586 	b->shrinker.seeks = DEFAULT_SEEKS;
1587 
1588 	r = register_shrinker(&b->shrinker, "vmw-balloon");
1589 
1590 	if (r == 0)
1591 		b->shrinker_registered = true;
1592 
1593 	return r;
1594 }
1595 
1596 /*
1597  * DEBUGFS Interface
1598  */
1599 #ifdef CONFIG_DEBUG_FS
1600 
1601 static const char * const vmballoon_stat_page_names[] = {
1602 	[VMW_BALLOON_PAGE_STAT_ALLOC]		= "alloc",
1603 	[VMW_BALLOON_PAGE_STAT_ALLOC_FAIL]	= "allocFail",
1604 	[VMW_BALLOON_PAGE_STAT_REFUSED_ALLOC]	= "errAlloc",
1605 	[VMW_BALLOON_PAGE_STAT_REFUSED_FREE]	= "errFree",
1606 	[VMW_BALLOON_PAGE_STAT_FREE]		= "free"
1607 };
1608 
1609 static const char * const vmballoon_stat_names[] = {
1610 	[VMW_BALLOON_STAT_TIMER]		= "timer",
1611 	[VMW_BALLOON_STAT_DOORBELL]		= "doorbell",
1612 	[VMW_BALLOON_STAT_RESET]		= "reset",
1613 	[VMW_BALLOON_STAT_SHRINK]		= "shrink",
1614 	[VMW_BALLOON_STAT_SHRINK_FREE]		= "shrinkFree"
1615 };
1616 
vmballoon_enable_stats(struct vmballoon * b)1617 static int vmballoon_enable_stats(struct vmballoon *b)
1618 {
1619 	int r = 0;
1620 
1621 	down_write(&b->conf_sem);
1622 
1623 	/* did we somehow race with another reader which enabled stats? */
1624 	if (b->stats)
1625 		goto out;
1626 
1627 	b->stats = kzalloc(sizeof(*b->stats), GFP_KERNEL);
1628 
1629 	if (!b->stats) {
1630 		/* allocation failed */
1631 		r = -ENOMEM;
1632 		goto out;
1633 	}
1634 	static_key_enable(&balloon_stat_enabled.key);
1635 out:
1636 	up_write(&b->conf_sem);
1637 	return r;
1638 }
1639 
1640 /**
1641  * vmballoon_debug_show - shows statistics of balloon operations.
1642  * @f: pointer to the &struct seq_file.
1643  * @offset: ignored.
1644  *
1645  * Provides the statistics that can be accessed in vmmemctl in the debugfs.
1646  * To avoid the overhead - mainly that of memory - of collecting the statistics,
1647  * we only collect statistics after the first time the counters are read.
1648  *
1649  * Return: zero on success or an error code.
1650  */
vmballoon_debug_show(struct seq_file * f,void * offset)1651 static int vmballoon_debug_show(struct seq_file *f, void *offset)
1652 {
1653 	struct vmballoon *b = f->private;
1654 	int i, j;
1655 
1656 	/* enables stats if they are disabled */
1657 	if (!b->stats) {
1658 		int r = vmballoon_enable_stats(b);
1659 
1660 		if (r)
1661 			return r;
1662 	}
1663 
1664 	/* format capabilities info */
1665 	seq_printf(f, "%-22s: %#16x\n", "balloon capabilities",
1666 		   VMW_BALLOON_CAPABILITIES);
1667 	seq_printf(f, "%-22s: %#16lx\n", "used capabilities", b->capabilities);
1668 	seq_printf(f, "%-22s: %16s\n", "is resetting",
1669 		   b->reset_required ? "y" : "n");
1670 
1671 	/* format size info */
1672 	seq_printf(f, "%-22s: %16lu\n", "target", READ_ONCE(b->target));
1673 	seq_printf(f, "%-22s: %16llu\n", "current", atomic64_read(&b->size));
1674 
1675 	for (i = 0; i < VMW_BALLOON_CMD_NUM; i++) {
1676 		if (vmballoon_cmd_names[i] == NULL)
1677 			continue;
1678 
1679 		seq_printf(f, "%-22s: %16llu (%llu failed)\n",
1680 			   vmballoon_cmd_names[i],
1681 			   atomic64_read(&b->stats->ops[i][VMW_BALLOON_OP_STAT]),
1682 			   atomic64_read(&b->stats->ops[i][VMW_BALLOON_OP_FAIL_STAT]));
1683 	}
1684 
1685 	for (i = 0; i < VMW_BALLOON_STAT_NUM; i++)
1686 		seq_printf(f, "%-22s: %16llu\n",
1687 			   vmballoon_stat_names[i],
1688 			   atomic64_read(&b->stats->general_stat[i]));
1689 
1690 	for (i = 0; i < VMW_BALLOON_PAGE_STAT_NUM; i++) {
1691 		for (j = 0; j < VMW_BALLOON_NUM_PAGE_SIZES; j++)
1692 			seq_printf(f, "%-18s(%s): %16llu\n",
1693 				   vmballoon_stat_page_names[i],
1694 				   vmballoon_page_size_names[j],
1695 				   atomic64_read(&b->stats->page_stat[i][j]));
1696 	}
1697 
1698 	return 0;
1699 }
1700 
1701 DEFINE_SHOW_ATTRIBUTE(vmballoon_debug);
1702 
vmballoon_debugfs_init(struct vmballoon * b)1703 static void __init vmballoon_debugfs_init(struct vmballoon *b)
1704 {
1705 	debugfs_create_file("vmmemctl", S_IRUGO, NULL, b,
1706 			    &vmballoon_debug_fops);
1707 }
1708 
vmballoon_debugfs_exit(struct vmballoon * b)1709 static void __exit vmballoon_debugfs_exit(struct vmballoon *b)
1710 {
1711 	static_key_disable(&balloon_stat_enabled.key);
1712 	debugfs_lookup_and_remove("vmmemctl", NULL);
1713 	kfree(b->stats);
1714 	b->stats = NULL;
1715 }
1716 
1717 #else
1718 
vmballoon_debugfs_init(struct vmballoon * b)1719 static inline void vmballoon_debugfs_init(struct vmballoon *b)
1720 {
1721 }
1722 
vmballoon_debugfs_exit(struct vmballoon * b)1723 static inline void vmballoon_debugfs_exit(struct vmballoon *b)
1724 {
1725 }
1726 
1727 #endif	/* CONFIG_DEBUG_FS */
1728 
1729 
1730 #ifdef CONFIG_BALLOON_COMPACTION
1731 /**
1732  * vmballoon_migratepage() - migrates a balloon page.
1733  * @b_dev_info: balloon device information descriptor.
1734  * @newpage: the page to which @page should be migrated.
1735  * @page: a ballooned page that should be migrated.
1736  * @mode: migration mode, ignored.
1737  *
1738  * This function is really open-coded, but that is according to the interface
1739  * that balloon_compaction provides.
1740  *
1741  * Return: zero on success, -EAGAIN when migration cannot be performed
1742  *	   momentarily, and -EBUSY if migration failed and should be retried
1743  *	   with that specific page.
1744  */
vmballoon_migratepage(struct balloon_dev_info * b_dev_info,struct page * newpage,struct page * page,enum migrate_mode mode)1745 static int vmballoon_migratepage(struct balloon_dev_info *b_dev_info,
1746 				 struct page *newpage, struct page *page,
1747 				 enum migrate_mode mode)
1748 {
1749 	unsigned long status, flags;
1750 	struct vmballoon *b;
1751 	int ret;
1752 
1753 	b = container_of(b_dev_info, struct vmballoon, b_dev_info);
1754 
1755 	/*
1756 	 * If the semaphore is taken, there is ongoing configuration change
1757 	 * (i.e., balloon reset), so try again.
1758 	 */
1759 	if (!down_read_trylock(&b->conf_sem))
1760 		return -EAGAIN;
1761 
1762 	spin_lock(&b->comm_lock);
1763 	/*
1764 	 * We must start by deflating and not inflating, as otherwise the
1765 	 * hypervisor may tell us that it has enough memory and the new page is
1766 	 * not needed. Since the old page is isolated, we cannot use the list
1767 	 * interface to unlock it, as the LRU field is used for isolation.
1768 	 * Instead, we use the native interface directly.
1769 	 */
1770 	vmballoon_add_page(b, 0, page);
1771 	status = vmballoon_lock_op(b, 1, VMW_BALLOON_4K_PAGE,
1772 				   VMW_BALLOON_DEFLATE);
1773 
1774 	if (status == VMW_BALLOON_SUCCESS)
1775 		status = vmballoon_status_page(b, 0, &page);
1776 
1777 	/*
1778 	 * If a failure happened, let the migration mechanism know that it
1779 	 * should not retry.
1780 	 */
1781 	if (status != VMW_BALLOON_SUCCESS) {
1782 		spin_unlock(&b->comm_lock);
1783 		ret = -EBUSY;
1784 		goto out_unlock;
1785 	}
1786 
1787 	/*
1788 	 * The page is isolated, so it is safe to delete it without holding
1789 	 * @pages_lock . We keep holding @comm_lock since we will need it in a
1790 	 * second.
1791 	 */
1792 	balloon_page_delete(page);
1793 
1794 	put_page(page);
1795 
1796 	/* Inflate */
1797 	vmballoon_add_page(b, 0, newpage);
1798 	status = vmballoon_lock_op(b, 1, VMW_BALLOON_4K_PAGE,
1799 				   VMW_BALLOON_INFLATE);
1800 
1801 	if (status == VMW_BALLOON_SUCCESS)
1802 		status = vmballoon_status_page(b, 0, &newpage);
1803 
1804 	spin_unlock(&b->comm_lock);
1805 
1806 	if (status != VMW_BALLOON_SUCCESS) {
1807 		/*
1808 		 * A failure happened. While we can deflate the page we just
1809 		 * inflated, this deflation can also encounter an error. Instead
1810 		 * we will decrease the size of the balloon to reflect the
1811 		 * change and report failure.
1812 		 */
1813 		atomic64_dec(&b->size);
1814 		ret = -EBUSY;
1815 	} else {
1816 		/*
1817 		 * Success. Take a reference for the page, and we will add it to
1818 		 * the list after acquiring the lock.
1819 		 */
1820 		get_page(newpage);
1821 		ret = MIGRATEPAGE_SUCCESS;
1822 	}
1823 
1824 	/* Update the balloon list under the @pages_lock */
1825 	spin_lock_irqsave(&b->b_dev_info.pages_lock, flags);
1826 
1827 	/*
1828 	 * On inflation success, we already took a reference for the @newpage.
1829 	 * If we succeed just insert it to the list and update the statistics
1830 	 * under the lock.
1831 	 */
1832 	if (ret == MIGRATEPAGE_SUCCESS) {
1833 		balloon_page_insert(&b->b_dev_info, newpage);
1834 		__count_vm_event(BALLOON_MIGRATE);
1835 	}
1836 
1837 	/*
1838 	 * We deflated successfully, so regardless to the inflation success, we
1839 	 * need to reduce the number of isolated_pages.
1840 	 */
1841 	b->b_dev_info.isolated_pages--;
1842 	spin_unlock_irqrestore(&b->b_dev_info.pages_lock, flags);
1843 
1844 out_unlock:
1845 	up_read(&b->conf_sem);
1846 	return ret;
1847 }
1848 
1849 /**
1850  * vmballoon_compaction_init() - initialized compaction for the balloon.
1851  *
1852  * @b: pointer to the balloon.
1853  *
1854  * If during the initialization a failure occurred, this function does not
1855  * perform cleanup. The caller must call vmballoon_compaction_deinit() in this
1856  * case.
1857  *
1858  * Return: zero on success or error code on failure.
1859  */
vmballoon_compaction_init(struct vmballoon * b)1860 static __init void vmballoon_compaction_init(struct vmballoon *b)
1861 {
1862 	b->b_dev_info.migratepage = vmballoon_migratepage;
1863 }
1864 
1865 #else /* CONFIG_BALLOON_COMPACTION */
vmballoon_compaction_init(struct vmballoon * b)1866 static inline void vmballoon_compaction_init(struct vmballoon *b)
1867 {
1868 }
1869 #endif /* CONFIG_BALLOON_COMPACTION */
1870 
vmballoon_init(void)1871 static int __init vmballoon_init(void)
1872 {
1873 	int error;
1874 
1875 	/*
1876 	 * Check if we are running on VMware's hypervisor and bail out
1877 	 * if we are not.
1878 	 */
1879 	if (x86_hyper_type != X86_HYPER_VMWARE)
1880 		return -ENODEV;
1881 
1882 	INIT_DELAYED_WORK(&balloon.dwork, vmballoon_work);
1883 
1884 	error = vmballoon_register_shrinker(&balloon);
1885 	if (error)
1886 		goto fail;
1887 
1888 	/*
1889 	 * Initialization of compaction must be done after the call to
1890 	 * balloon_devinfo_init() .
1891 	 */
1892 	balloon_devinfo_init(&balloon.b_dev_info);
1893 	vmballoon_compaction_init(&balloon);
1894 
1895 	INIT_LIST_HEAD(&balloon.huge_pages);
1896 	spin_lock_init(&balloon.comm_lock);
1897 	init_rwsem(&balloon.conf_sem);
1898 	balloon.vmci_doorbell = VMCI_INVALID_HANDLE;
1899 	balloon.batch_page = NULL;
1900 	balloon.page = NULL;
1901 	balloon.reset_required = true;
1902 
1903 	queue_delayed_work(system_freezable_wq, &balloon.dwork, 0);
1904 
1905 	vmballoon_debugfs_init(&balloon);
1906 
1907 	return 0;
1908 fail:
1909 	vmballoon_unregister_shrinker(&balloon);
1910 	return error;
1911 }
1912 
1913 /*
1914  * Using late_initcall() instead of module_init() allows the balloon to use the
1915  * VMCI doorbell even when the balloon is built into the kernel. Otherwise the
1916  * VMCI is probed only after the balloon is initialized. If the balloon is used
1917  * as a module, late_initcall() is equivalent to module_init().
1918  */
1919 late_initcall(vmballoon_init);
1920 
vmballoon_exit(void)1921 static void __exit vmballoon_exit(void)
1922 {
1923 	vmballoon_unregister_shrinker(&balloon);
1924 	vmballoon_vmci_cleanup(&balloon);
1925 	cancel_delayed_work_sync(&balloon.dwork);
1926 
1927 	vmballoon_debugfs_exit(&balloon);
1928 
1929 	/*
1930 	 * Deallocate all reserved memory, and reset connection with monitor.
1931 	 * Reset connection before deallocating memory to avoid potential for
1932 	 * additional spurious resets from guest touching deallocated pages.
1933 	 */
1934 	vmballoon_send_start(&balloon, 0);
1935 	vmballoon_pop(&balloon);
1936 }
1937 module_exit(vmballoon_exit);
1938