xref: /openbmc/linux/drivers/hv/hv_balloon.c (revision 4bce6fce)
1 /*
2  * Copyright (c) 2012, Microsoft Corporation.
3  *
4  * Author:
5  *   K. Y. Srinivasan <kys@microsoft.com>
6  *
7  * This program is free software; you can redistribute it and/or modify it
8  * under the terms of the GNU General Public License version 2 as published
9  * by the Free Software Foundation.
10  *
11  * This program is distributed in the hope that it will be useful, but
12  * WITHOUT ANY WARRANTY; without even the implied warranty of
13  * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
14  * NON INFRINGEMENT.  See the GNU General Public License for more
15  * details.
16  *
17  */
18 
19 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
20 
21 #include <linux/kernel.h>
22 #include <linux/jiffies.h>
23 #include <linux/mman.h>
24 #include <linux/delay.h>
25 #include <linux/init.h>
26 #include <linux/module.h>
27 #include <linux/slab.h>
28 #include <linux/kthread.h>
29 #include <linux/completion.h>
30 #include <linux/memory_hotplug.h>
31 #include <linux/memory.h>
32 #include <linux/notifier.h>
33 #include <linux/percpu_counter.h>
34 
35 #include <linux/hyperv.h>
36 
37 /*
38  * We begin with definitions supporting the Dynamic Memory protocol
39  * with the host.
40  *
41  * Begin protocol definitions.
42  */
43 
44 
45 
46 /*
47  * Protocol versions. The low word is the minor version, the high word the major
48  * version.
49  *
50  * History:
51  * Initial version 1.0
52  * Changed to 0.1 on 2009/03/25
53  * Changes to 0.2 on 2009/05/14
54  * Changes to 0.3 on 2009/12/03
55  * Changed to 1.0 on 2011/04/05
56  */
57 
58 #define DYNMEM_MAKE_VERSION(Major, Minor) ((__u32)(((Major) << 16) | (Minor)))
59 #define DYNMEM_MAJOR_VERSION(Version) ((__u32)(Version) >> 16)
60 #define DYNMEM_MINOR_VERSION(Version) ((__u32)(Version) & 0xff)
61 
62 enum {
63 	DYNMEM_PROTOCOL_VERSION_1 = DYNMEM_MAKE_VERSION(0, 3),
64 	DYNMEM_PROTOCOL_VERSION_2 = DYNMEM_MAKE_VERSION(1, 0),
65 
66 	DYNMEM_PROTOCOL_VERSION_WIN7 = DYNMEM_PROTOCOL_VERSION_1,
67 	DYNMEM_PROTOCOL_VERSION_WIN8 = DYNMEM_PROTOCOL_VERSION_2,
68 
69 	DYNMEM_PROTOCOL_VERSION_CURRENT = DYNMEM_PROTOCOL_VERSION_WIN8
70 };
71 
72 
73 
74 /*
75  * Message Types
76  */
77 
78 enum dm_message_type {
79 	/*
80 	 * Version 0.3
81 	 */
82 	DM_ERROR			= 0,
83 	DM_VERSION_REQUEST		= 1,
84 	DM_VERSION_RESPONSE		= 2,
85 	DM_CAPABILITIES_REPORT		= 3,
86 	DM_CAPABILITIES_RESPONSE	= 4,
87 	DM_STATUS_REPORT		= 5,
88 	DM_BALLOON_REQUEST		= 6,
89 	DM_BALLOON_RESPONSE		= 7,
90 	DM_UNBALLOON_REQUEST		= 8,
91 	DM_UNBALLOON_RESPONSE		= 9,
92 	DM_MEM_HOT_ADD_REQUEST		= 10,
93 	DM_MEM_HOT_ADD_RESPONSE		= 11,
94 	DM_VERSION_03_MAX		= 11,
95 	/*
96 	 * Version 1.0.
97 	 */
98 	DM_INFO_MESSAGE			= 12,
99 	DM_VERSION_1_MAX		= 12
100 };
101 
102 
103 /*
104  * Structures defining the dynamic memory management
105  * protocol.
106  */
107 
108 union dm_version {
109 	struct {
110 		__u16 minor_version;
111 		__u16 major_version;
112 	};
113 	__u32 version;
114 } __packed;
115 
116 
117 union dm_caps {
118 	struct {
119 		__u64 balloon:1;
120 		__u64 hot_add:1;
121 		/*
122 		 * To support guests that may have alignment
123 		 * limitations on hot-add, the guest can specify
124 		 * its alignment requirements; a value of n
125 		 * represents an alignment of 2^n in mega bytes.
126 		 */
127 		__u64 hot_add_alignment:4;
128 		__u64 reservedz:58;
129 	} cap_bits;
130 	__u64 caps;
131 } __packed;
132 
133 union dm_mem_page_range {
134 	struct  {
135 		/*
136 		 * The PFN number of the first page in the range.
137 		 * 40 bits is the architectural limit of a PFN
138 		 * number for AMD64.
139 		 */
140 		__u64 start_page:40;
141 		/*
142 		 * The number of pages in the range.
143 		 */
144 		__u64 page_cnt:24;
145 	} finfo;
146 	__u64  page_range;
147 } __packed;
148 
149 
150 
151 /*
152  * The header for all dynamic memory messages:
153  *
154  * type: Type of the message.
155  * size: Size of the message in bytes; including the header.
156  * trans_id: The guest is responsible for manufacturing this ID.
157  */
158 
159 struct dm_header {
160 	__u16 type;
161 	__u16 size;
162 	__u32 trans_id;
163 } __packed;
164 
165 /*
166  * A generic message format for dynamic memory.
167  * Specific message formats are defined later in the file.
168  */
169 
170 struct dm_message {
171 	struct dm_header hdr;
172 	__u8 data[]; /* enclosed message */
173 } __packed;
174 
175 
176 /*
177  * Specific message types supporting the dynamic memory protocol.
178  */
179 
180 /*
181  * Version negotiation message. Sent from the guest to the host.
182  * The guest is free to try different versions until the host
183  * accepts the version.
184  *
185  * dm_version: The protocol version requested.
186  * is_last_attempt: If TRUE, this is the last version guest will request.
187  * reservedz: Reserved field, set to zero.
188  */
189 
190 struct dm_version_request {
191 	struct dm_header hdr;
192 	union dm_version version;
193 	__u32 is_last_attempt:1;
194 	__u32 reservedz:31;
195 } __packed;
196 
197 /*
198  * Version response message; Host to Guest and indicates
199  * if the host has accepted the version sent by the guest.
200  *
201  * is_accepted: If TRUE, host has accepted the version and the guest
202  * should proceed to the next stage of the protocol. FALSE indicates that
203  * guest should re-try with a different version.
204  *
205  * reservedz: Reserved field, set to zero.
206  */
207 
208 struct dm_version_response {
209 	struct dm_header hdr;
210 	__u64 is_accepted:1;
211 	__u64 reservedz:63;
212 } __packed;
213 
214 /*
215  * Message reporting capabilities. This is sent from the guest to the
216  * host.
217  */
218 
219 struct dm_capabilities {
220 	struct dm_header hdr;
221 	union dm_caps caps;
222 	__u64 min_page_cnt;
223 	__u64 max_page_number;
224 } __packed;
225 
226 /*
227  * Response to the capabilities message. This is sent from the host to the
228  * guest. This message notifies if the host has accepted the guest's
229  * capabilities. If the host has not accepted, the guest must shutdown
230  * the service.
231  *
232  * is_accepted: Indicates if the host has accepted guest's capabilities.
233  * reservedz: Must be 0.
234  */
235 
236 struct dm_capabilities_resp_msg {
237 	struct dm_header hdr;
238 	__u64 is_accepted:1;
239 	__u64 reservedz:63;
240 } __packed;
241 
242 /*
243  * This message is used to report memory pressure from the guest.
244  * This message is not part of any transaction and there is no
245  * response to this message.
246  *
247  * num_avail: Available memory in pages.
248  * num_committed: Committed memory in pages.
249  * page_file_size: The accumulated size of all page files
250  *		   in the system in pages.
251  * zero_free: The nunber of zero and free pages.
252  * page_file_writes: The writes to the page file in pages.
253  * io_diff: An indicator of file cache efficiency or page file activity,
254  *	    calculated as File Cache Page Fault Count - Page Read Count.
255  *	    This value is in pages.
256  *
257  * Some of these metrics are Windows specific and fortunately
258  * the algorithm on the host side that computes the guest memory
259  * pressure only uses num_committed value.
260  */
261 
262 struct dm_status {
263 	struct dm_header hdr;
264 	__u64 num_avail;
265 	__u64 num_committed;
266 	__u64 page_file_size;
267 	__u64 zero_free;
268 	__u32 page_file_writes;
269 	__u32 io_diff;
270 } __packed;
271 
272 
273 /*
274  * Message to ask the guest to allocate memory - balloon up message.
275  * This message is sent from the host to the guest. The guest may not be
276  * able to allocate as much memory as requested.
277  *
278  * num_pages: number of pages to allocate.
279  */
280 
281 struct dm_balloon {
282 	struct dm_header hdr;
283 	__u32 num_pages;
284 	__u32 reservedz;
285 } __packed;
286 
287 
288 /*
289  * Balloon response message; this message is sent from the guest
290  * to the host in response to the balloon message.
291  *
292  * reservedz: Reserved; must be set to zero.
293  * more_pages: If FALSE, this is the last message of the transaction.
294  * if TRUE there will atleast one more message from the guest.
295  *
296  * range_count: The number of ranges in the range array.
297  *
298  * range_array: An array of page ranges returned to the host.
299  *
300  */
301 
302 struct dm_balloon_response {
303 	struct dm_header hdr;
304 	__u32 reservedz;
305 	__u32 more_pages:1;
306 	__u32 range_count:31;
307 	union dm_mem_page_range range_array[];
308 } __packed;
309 
310 /*
311  * Un-balloon message; this message is sent from the host
312  * to the guest to give guest more memory.
313  *
314  * more_pages: If FALSE, this is the last message of the transaction.
315  * if TRUE there will atleast one more message from the guest.
316  *
317  * reservedz: Reserved; must be set to zero.
318  *
319  * range_count: The number of ranges in the range array.
320  *
321  * range_array: An array of page ranges returned to the host.
322  *
323  */
324 
325 struct dm_unballoon_request {
326 	struct dm_header hdr;
327 	__u32 more_pages:1;
328 	__u32 reservedz:31;
329 	__u32 range_count;
330 	union dm_mem_page_range range_array[];
331 } __packed;
332 
333 /*
334  * Un-balloon response message; this message is sent from the guest
335  * to the host in response to an unballoon request.
336  *
337  */
338 
339 struct dm_unballoon_response {
340 	struct dm_header hdr;
341 } __packed;
342 
343 
344 /*
345  * Hot add request message. Message sent from the host to the guest.
346  *
347  * mem_range: Memory range to hot add.
348  *
349  * On Linux we currently don't support this since we cannot hot add
350  * arbitrary granularity of memory.
351  */
352 
353 struct dm_hot_add {
354 	struct dm_header hdr;
355 	union dm_mem_page_range range;
356 } __packed;
357 
358 /*
359  * Hot add response message.
360  * This message is sent by the guest to report the status of a hot add request.
361  * If page_count is less than the requested page count, then the host should
362  * assume all further hot add requests will fail, since this indicates that
363  * the guest has hit an upper physical memory barrier.
364  *
365  * Hot adds may also fail due to low resources; in this case, the guest must
366  * not complete this message until the hot add can succeed, and the host must
367  * not send a new hot add request until the response is sent.
368  * If VSC fails to hot add memory DYNMEM_NUMBER_OF_UNSUCCESSFUL_HOTADD_ATTEMPTS
369  * times it fails the request.
370  *
371  *
372  * page_count: number of pages that were successfully hot added.
373  *
374  * result: result of the operation 1: success, 0: failure.
375  *
376  */
377 
378 struct dm_hot_add_response {
379 	struct dm_header hdr;
380 	__u32 page_count;
381 	__u32 result;
382 } __packed;
383 
384 /*
385  * Types of information sent from host to the guest.
386  */
387 
388 enum dm_info_type {
389 	INFO_TYPE_MAX_PAGE_CNT = 0,
390 	MAX_INFO_TYPE
391 };
392 
393 
394 /*
395  * Header for the information message.
396  */
397 
398 struct dm_info_header {
399 	enum dm_info_type type;
400 	__u32 data_size;
401 } __packed;
402 
403 /*
404  * This message is sent from the host to the guest to pass
405  * some relevant information (win8 addition).
406  *
407  * reserved: no used.
408  * info_size: size of the information blob.
409  * info: information blob.
410  */
411 
412 struct dm_info_msg {
413 	struct dm_header hdr;
414 	__u32 reserved;
415 	__u32 info_size;
416 	__u8  info[];
417 };
418 
419 /*
420  * End protocol definitions.
421  */
422 
423 /*
424  * State to manage hot adding memory into the guest.
425  * The range start_pfn : end_pfn specifies the range
426  * that the host has asked us to hot add. The range
427  * start_pfn : ha_end_pfn specifies the range that we have
428  * currently hot added. We hot add in multiples of 128M
429  * chunks; it is possible that we may not be able to bring
430  * online all the pages in the region. The range
431  * covered_end_pfn defines the pages that can
432  * be brough online.
433  */
434 
435 struct hv_hotadd_state {
436 	struct list_head list;
437 	unsigned long start_pfn;
438 	unsigned long covered_end_pfn;
439 	unsigned long ha_end_pfn;
440 	unsigned long end_pfn;
441 };
442 
443 struct balloon_state {
444 	__u32 num_pages;
445 	struct work_struct wrk;
446 };
447 
448 struct hot_add_wrk {
449 	union dm_mem_page_range ha_page_range;
450 	union dm_mem_page_range ha_region_range;
451 	struct work_struct wrk;
452 };
453 
454 static bool hot_add = true;
455 static bool do_hot_add;
456 /*
457  * Delay reporting memory pressure by
458  * the specified number of seconds.
459  */
460 static uint pressure_report_delay = 45;
461 
462 /*
463  * The last time we posted a pressure report to host.
464  */
465 static unsigned long last_post_time;
466 
467 module_param(hot_add, bool, (S_IRUGO | S_IWUSR));
468 MODULE_PARM_DESC(hot_add, "If set attempt memory hot_add");
469 
470 module_param(pressure_report_delay, uint, (S_IRUGO | S_IWUSR));
471 MODULE_PARM_DESC(pressure_report_delay, "Delay in secs in reporting pressure");
472 static atomic_t trans_id = ATOMIC_INIT(0);
473 
474 static int dm_ring_size = (5 * PAGE_SIZE);
475 
476 /*
477  * Driver specific state.
478  */
479 
480 enum hv_dm_state {
481 	DM_INITIALIZING = 0,
482 	DM_INITIALIZED,
483 	DM_BALLOON_UP,
484 	DM_BALLOON_DOWN,
485 	DM_HOT_ADD,
486 	DM_INIT_ERROR
487 };
488 
489 
490 static __u8 recv_buffer[PAGE_SIZE];
491 static __u8 *send_buffer;
492 #define PAGES_IN_2M	512
493 #define HA_CHUNK (32 * 1024)
494 
495 struct hv_dynmem_device {
496 	struct hv_device *dev;
497 	enum hv_dm_state state;
498 	struct completion host_event;
499 	struct completion config_event;
500 
501 	/*
502 	 * Number of pages we have currently ballooned out.
503 	 */
504 	unsigned int num_pages_ballooned;
505 	unsigned int num_pages_onlined;
506 	unsigned int num_pages_added;
507 
508 	/*
509 	 * State to manage the ballooning (up) operation.
510 	 */
511 	struct balloon_state balloon_wrk;
512 
513 	/*
514 	 * State to execute the "hot-add" operation.
515 	 */
516 	struct hot_add_wrk ha_wrk;
517 
518 	/*
519 	 * This state tracks if the host has specified a hot-add
520 	 * region.
521 	 */
522 	bool host_specified_ha_region;
523 
524 	/*
525 	 * State to synchronize hot-add.
526 	 */
527 	struct completion  ol_waitevent;
528 	bool ha_waiting;
529 	/*
530 	 * This thread handles hot-add
531 	 * requests from the host as well as notifying
532 	 * the host with regards to memory pressure in
533 	 * the guest.
534 	 */
535 	struct task_struct *thread;
536 
537 	struct mutex ha_region_mutex;
538 
539 	/*
540 	 * A list of hot-add regions.
541 	 */
542 	struct list_head ha_region_list;
543 
544 	/*
545 	 * We start with the highest version we can support
546 	 * and downgrade based on the host; we save here the
547 	 * next version to try.
548 	 */
549 	__u32 next_version;
550 };
551 
552 static struct hv_dynmem_device dm_device;
553 
554 static void post_status(struct hv_dynmem_device *dm);
555 
556 #ifdef CONFIG_MEMORY_HOTPLUG
557 static int hv_memory_notifier(struct notifier_block *nb, unsigned long val,
558 			      void *v)
559 {
560 	struct memory_notify *mem = (struct memory_notify *)v;
561 
562 	switch (val) {
563 	case MEM_GOING_ONLINE:
564 		mutex_lock(&dm_device.ha_region_mutex);
565 		break;
566 
567 	case MEM_ONLINE:
568 		dm_device.num_pages_onlined += mem->nr_pages;
569 	case MEM_CANCEL_ONLINE:
570 		mutex_unlock(&dm_device.ha_region_mutex);
571 		if (dm_device.ha_waiting) {
572 			dm_device.ha_waiting = false;
573 			complete(&dm_device.ol_waitevent);
574 		}
575 		break;
576 
577 	case MEM_OFFLINE:
578 		mutex_lock(&dm_device.ha_region_mutex);
579 		dm_device.num_pages_onlined -= mem->nr_pages;
580 		mutex_unlock(&dm_device.ha_region_mutex);
581 		break;
582 	case MEM_GOING_OFFLINE:
583 	case MEM_CANCEL_OFFLINE:
584 		break;
585 	}
586 	return NOTIFY_OK;
587 }
588 
589 static struct notifier_block hv_memory_nb = {
590 	.notifier_call = hv_memory_notifier,
591 	.priority = 0
592 };
593 
594 
595 static void hv_bring_pgs_online(unsigned long start_pfn, unsigned long size)
596 {
597 	int i;
598 
599 	for (i = 0; i < size; i++) {
600 		struct page *pg;
601 		pg = pfn_to_page(start_pfn + i);
602 		__online_page_set_limits(pg);
603 		__online_page_increment_counters(pg);
604 		__online_page_free(pg);
605 	}
606 }
607 
608 static void hv_mem_hot_add(unsigned long start, unsigned long size,
609 				unsigned long pfn_count,
610 				struct hv_hotadd_state *has)
611 {
612 	int ret = 0;
613 	int i, nid;
614 	unsigned long start_pfn;
615 	unsigned long processed_pfn;
616 	unsigned long total_pfn = pfn_count;
617 
618 	for (i = 0; i < (size/HA_CHUNK); i++) {
619 		start_pfn = start + (i * HA_CHUNK);
620 		has->ha_end_pfn +=  HA_CHUNK;
621 
622 		if (total_pfn > HA_CHUNK) {
623 			processed_pfn = HA_CHUNK;
624 			total_pfn -= HA_CHUNK;
625 		} else {
626 			processed_pfn = total_pfn;
627 			total_pfn = 0;
628 		}
629 
630 		has->covered_end_pfn +=  processed_pfn;
631 
632 		init_completion(&dm_device.ol_waitevent);
633 		dm_device.ha_waiting = true;
634 
635 		mutex_unlock(&dm_device.ha_region_mutex);
636 		nid = memory_add_physaddr_to_nid(PFN_PHYS(start_pfn));
637 		ret = add_memory(nid, PFN_PHYS((start_pfn)),
638 				(HA_CHUNK << PAGE_SHIFT));
639 
640 		if (ret) {
641 			pr_info("hot_add memory failed error is %d\n", ret);
642 			if (ret == -EEXIST) {
643 				/*
644 				 * This error indicates that the error
645 				 * is not a transient failure. This is the
646 				 * case where the guest's physical address map
647 				 * precludes hot adding memory. Stop all further
648 				 * memory hot-add.
649 				 */
650 				do_hot_add = false;
651 			}
652 			has->ha_end_pfn -= HA_CHUNK;
653 			has->covered_end_pfn -=  processed_pfn;
654 			mutex_lock(&dm_device.ha_region_mutex);
655 			break;
656 		}
657 
658 		/*
659 		 * Wait for the memory block to be onlined.
660 		 * Since the hot add has succeeded, it is ok to
661 		 * proceed even if the pages in the hot added region
662 		 * have not been "onlined" within the allowed time.
663 		 */
664 		wait_for_completion_timeout(&dm_device.ol_waitevent, 5*HZ);
665 		mutex_lock(&dm_device.ha_region_mutex);
666 		post_status(&dm_device);
667 	}
668 
669 	return;
670 }
671 
672 static void hv_online_page(struct page *pg)
673 {
674 	struct list_head *cur;
675 	struct hv_hotadd_state *has;
676 	unsigned long cur_start_pgp;
677 	unsigned long cur_end_pgp;
678 
679 	list_for_each(cur, &dm_device.ha_region_list) {
680 		has = list_entry(cur, struct hv_hotadd_state, list);
681 		cur_start_pgp = (unsigned long)pfn_to_page(has->start_pfn);
682 		cur_end_pgp = (unsigned long)pfn_to_page(has->covered_end_pfn);
683 
684 		if (((unsigned long)pg >= cur_start_pgp) &&
685 			((unsigned long)pg < cur_end_pgp)) {
686 			/*
687 			 * This frame is currently backed; online the
688 			 * page.
689 			 */
690 			__online_page_set_limits(pg);
691 			__online_page_increment_counters(pg);
692 			__online_page_free(pg);
693 		}
694 	}
695 }
696 
697 static bool pfn_covered(unsigned long start_pfn, unsigned long pfn_cnt)
698 {
699 	struct list_head *cur;
700 	struct hv_hotadd_state *has;
701 	unsigned long residual, new_inc;
702 
703 	if (list_empty(&dm_device.ha_region_list))
704 		return false;
705 
706 	list_for_each(cur, &dm_device.ha_region_list) {
707 		has = list_entry(cur, struct hv_hotadd_state, list);
708 
709 		/*
710 		 * If the pfn range we are dealing with is not in the current
711 		 * "hot add block", move on.
712 		 */
713 		if ((start_pfn >= has->end_pfn))
714 			continue;
715 		/*
716 		 * If the current hot add-request extends beyond
717 		 * our current limit; extend it.
718 		 */
719 		if ((start_pfn + pfn_cnt) > has->end_pfn) {
720 			residual = (start_pfn + pfn_cnt - has->end_pfn);
721 			/*
722 			 * Extend the region by multiples of HA_CHUNK.
723 			 */
724 			new_inc = (residual / HA_CHUNK) * HA_CHUNK;
725 			if (residual % HA_CHUNK)
726 				new_inc += HA_CHUNK;
727 
728 			has->end_pfn += new_inc;
729 		}
730 
731 		/*
732 		 * If the current start pfn is not where the covered_end
733 		 * is, update it.
734 		 */
735 
736 		if (has->covered_end_pfn != start_pfn)
737 			has->covered_end_pfn = start_pfn;
738 
739 		return true;
740 
741 	}
742 
743 	return false;
744 }
745 
746 static unsigned long handle_pg_range(unsigned long pg_start,
747 					unsigned long pg_count)
748 {
749 	unsigned long start_pfn = pg_start;
750 	unsigned long pfn_cnt = pg_count;
751 	unsigned long size;
752 	struct list_head *cur;
753 	struct hv_hotadd_state *has;
754 	unsigned long pgs_ol = 0;
755 	unsigned long old_covered_state;
756 
757 	if (list_empty(&dm_device.ha_region_list))
758 		return 0;
759 
760 	list_for_each(cur, &dm_device.ha_region_list) {
761 		has = list_entry(cur, struct hv_hotadd_state, list);
762 
763 		/*
764 		 * If the pfn range we are dealing with is not in the current
765 		 * "hot add block", move on.
766 		 */
767 		if ((start_pfn >= has->end_pfn))
768 			continue;
769 
770 		old_covered_state = has->covered_end_pfn;
771 
772 		if (start_pfn < has->ha_end_pfn) {
773 			/*
774 			 * This is the case where we are backing pages
775 			 * in an already hot added region. Bring
776 			 * these pages online first.
777 			 */
778 			pgs_ol = has->ha_end_pfn - start_pfn;
779 			if (pgs_ol > pfn_cnt)
780 				pgs_ol = pfn_cnt;
781 
782 			/*
783 			 * Check if the corresponding memory block is already
784 			 * online by checking its last previously backed page.
785 			 * In case it is we need to bring rest (which was not
786 			 * backed previously) online too.
787 			 */
788 			if (start_pfn > has->start_pfn &&
789 			    !PageReserved(pfn_to_page(start_pfn - 1)))
790 				hv_bring_pgs_online(start_pfn, pgs_ol);
791 
792 			has->covered_end_pfn +=  pgs_ol;
793 			pfn_cnt -= pgs_ol;
794 		}
795 
796 		if ((has->ha_end_pfn < has->end_pfn) && (pfn_cnt > 0)) {
797 			/*
798 			 * We have some residual hot add range
799 			 * that needs to be hot added; hot add
800 			 * it now. Hot add a multiple of
801 			 * of HA_CHUNK that fully covers the pages
802 			 * we have.
803 			 */
804 			size = (has->end_pfn - has->ha_end_pfn);
805 			if (pfn_cnt <= size) {
806 				size = ((pfn_cnt / HA_CHUNK) * HA_CHUNK);
807 				if (pfn_cnt % HA_CHUNK)
808 					size += HA_CHUNK;
809 			} else {
810 				pfn_cnt = size;
811 			}
812 			hv_mem_hot_add(has->ha_end_pfn, size, pfn_cnt, has);
813 		}
814 		/*
815 		 * If we managed to online any pages that were given to us,
816 		 * we declare success.
817 		 */
818 		return has->covered_end_pfn - old_covered_state;
819 
820 	}
821 
822 	return 0;
823 }
824 
825 static unsigned long process_hot_add(unsigned long pg_start,
826 					unsigned long pfn_cnt,
827 					unsigned long rg_start,
828 					unsigned long rg_size)
829 {
830 	struct hv_hotadd_state *ha_region = NULL;
831 
832 	if (pfn_cnt == 0)
833 		return 0;
834 
835 	if (!dm_device.host_specified_ha_region)
836 		if (pfn_covered(pg_start, pfn_cnt))
837 			goto do_pg_range;
838 
839 	/*
840 	 * If the host has specified a hot-add range; deal with it first.
841 	 */
842 
843 	if (rg_size != 0) {
844 		ha_region = kzalloc(sizeof(struct hv_hotadd_state), GFP_KERNEL);
845 		if (!ha_region)
846 			return 0;
847 
848 		INIT_LIST_HEAD(&ha_region->list);
849 
850 		list_add_tail(&ha_region->list, &dm_device.ha_region_list);
851 		ha_region->start_pfn = rg_start;
852 		ha_region->ha_end_pfn = rg_start;
853 		ha_region->covered_end_pfn = pg_start;
854 		ha_region->end_pfn = rg_start + rg_size;
855 	}
856 
857 do_pg_range:
858 	/*
859 	 * Process the page range specified; bringing them
860 	 * online if possible.
861 	 */
862 	return handle_pg_range(pg_start, pfn_cnt);
863 }
864 
865 #endif
866 
867 static void hot_add_req(struct work_struct *dummy)
868 {
869 	struct dm_hot_add_response resp;
870 #ifdef CONFIG_MEMORY_HOTPLUG
871 	unsigned long pg_start, pfn_cnt;
872 	unsigned long rg_start, rg_sz;
873 #endif
874 	struct hv_dynmem_device *dm = &dm_device;
875 
876 	memset(&resp, 0, sizeof(struct dm_hot_add_response));
877 	resp.hdr.type = DM_MEM_HOT_ADD_RESPONSE;
878 	resp.hdr.size = sizeof(struct dm_hot_add_response);
879 
880 #ifdef CONFIG_MEMORY_HOTPLUG
881 	mutex_lock(&dm_device.ha_region_mutex);
882 	pg_start = dm->ha_wrk.ha_page_range.finfo.start_page;
883 	pfn_cnt = dm->ha_wrk.ha_page_range.finfo.page_cnt;
884 
885 	rg_start = dm->ha_wrk.ha_region_range.finfo.start_page;
886 	rg_sz = dm->ha_wrk.ha_region_range.finfo.page_cnt;
887 
888 	if ((rg_start == 0) && (!dm->host_specified_ha_region)) {
889 		unsigned long region_size;
890 		unsigned long region_start;
891 
892 		/*
893 		 * The host has not specified the hot-add region.
894 		 * Based on the hot-add page range being specified,
895 		 * compute a hot-add region that can cover the pages
896 		 * that need to be hot-added while ensuring the alignment
897 		 * and size requirements of Linux as it relates to hot-add.
898 		 */
899 		region_start = pg_start;
900 		region_size = (pfn_cnt / HA_CHUNK) * HA_CHUNK;
901 		if (pfn_cnt % HA_CHUNK)
902 			region_size += HA_CHUNK;
903 
904 		region_start = (pg_start / HA_CHUNK) * HA_CHUNK;
905 
906 		rg_start = region_start;
907 		rg_sz = region_size;
908 	}
909 
910 	if (do_hot_add)
911 		resp.page_count = process_hot_add(pg_start, pfn_cnt,
912 						rg_start, rg_sz);
913 
914 	dm->num_pages_added += resp.page_count;
915 	mutex_unlock(&dm_device.ha_region_mutex);
916 #endif
917 	/*
918 	 * The result field of the response structure has the
919 	 * following semantics:
920 	 *
921 	 * 1. If all or some pages hot-added: Guest should return success.
922 	 *
923 	 * 2. If no pages could be hot-added:
924 	 *
925 	 * If the guest returns success, then the host
926 	 * will not attempt any further hot-add operations. This
927 	 * signifies a permanent failure.
928 	 *
929 	 * If the guest returns failure, then this failure will be
930 	 * treated as a transient failure and the host may retry the
931 	 * hot-add operation after some delay.
932 	 */
933 	if (resp.page_count > 0)
934 		resp.result = 1;
935 	else if (!do_hot_add)
936 		resp.result = 1;
937 	else
938 		resp.result = 0;
939 
940 	if (!do_hot_add || (resp.page_count == 0))
941 		pr_info("Memory hot add failed\n");
942 
943 	dm->state = DM_INITIALIZED;
944 	resp.hdr.trans_id = atomic_inc_return(&trans_id);
945 	vmbus_sendpacket(dm->dev->channel, &resp,
946 			sizeof(struct dm_hot_add_response),
947 			(unsigned long)NULL,
948 			VM_PKT_DATA_INBAND, 0);
949 }
950 
951 static void process_info(struct hv_dynmem_device *dm, struct dm_info_msg *msg)
952 {
953 	struct dm_info_header *info_hdr;
954 
955 	info_hdr = (struct dm_info_header *)msg->info;
956 
957 	switch (info_hdr->type) {
958 	case INFO_TYPE_MAX_PAGE_CNT:
959 		pr_info("Received INFO_TYPE_MAX_PAGE_CNT\n");
960 		pr_info("Data Size is %d\n", info_hdr->data_size);
961 		break;
962 	default:
963 		pr_info("Received Unknown type: %d\n", info_hdr->type);
964 	}
965 }
966 
967 static unsigned long compute_balloon_floor(void)
968 {
969 	unsigned long min_pages;
970 #define MB2PAGES(mb) ((mb) << (20 - PAGE_SHIFT))
971 	/* Simple continuous piecewiese linear function:
972 	 *  max MiB -> min MiB  gradient
973 	 *       0         0
974 	 *      16        16
975 	 *      32        24
976 	 *     128        72    (1/2)
977 	 *     512       168    (1/4)
978 	 *    2048       360    (1/8)
979 	 *    8192       744    (1/16)
980 	 *   32768      1512	(1/32)
981 	 */
982 	if (totalram_pages < MB2PAGES(128))
983 		min_pages = MB2PAGES(8) + (totalram_pages >> 1);
984 	else if (totalram_pages < MB2PAGES(512))
985 		min_pages = MB2PAGES(40) + (totalram_pages >> 2);
986 	else if (totalram_pages < MB2PAGES(2048))
987 		min_pages = MB2PAGES(104) + (totalram_pages >> 3);
988 	else if (totalram_pages < MB2PAGES(8192))
989 		min_pages = MB2PAGES(232) + (totalram_pages >> 4);
990 	else
991 		min_pages = MB2PAGES(488) + (totalram_pages >> 5);
992 #undef MB2PAGES
993 	return min_pages;
994 }
995 
996 /*
997  * Post our status as it relates memory pressure to the
998  * host. Host expects the guests to post this status
999  * periodically at 1 second intervals.
1000  *
1001  * The metrics specified in this protocol are very Windows
1002  * specific and so we cook up numbers here to convey our memory
1003  * pressure.
1004  */
1005 
1006 static void post_status(struct hv_dynmem_device *dm)
1007 {
1008 	struct dm_status status;
1009 	struct sysinfo val;
1010 	unsigned long now = jiffies;
1011 	unsigned long last_post = last_post_time;
1012 
1013 	if (pressure_report_delay > 0) {
1014 		--pressure_report_delay;
1015 		return;
1016 	}
1017 
1018 	if (!time_after(now, (last_post_time + HZ)))
1019 		return;
1020 
1021 	si_meminfo(&val);
1022 	memset(&status, 0, sizeof(struct dm_status));
1023 	status.hdr.type = DM_STATUS_REPORT;
1024 	status.hdr.size = sizeof(struct dm_status);
1025 	status.hdr.trans_id = atomic_inc_return(&trans_id);
1026 
1027 	/*
1028 	 * The host expects the guest to report free and committed memory.
1029 	 * Furthermore, the host expects the pressure information to include
1030 	 * the ballooned out pages. For a given amount of memory that we are
1031 	 * managing we need to compute a floor below which we should not
1032 	 * balloon. Compute this and add it to the pressure report.
1033 	 * We also need to report all offline pages (num_pages_added -
1034 	 * num_pages_onlined) as committed to the host, otherwise it can try
1035 	 * asking us to balloon them out.
1036 	 */
1037 	status.num_avail = val.freeram;
1038 	status.num_committed = vm_memory_committed() +
1039 		dm->num_pages_ballooned +
1040 		(dm->num_pages_added > dm->num_pages_onlined ?
1041 		 dm->num_pages_added - dm->num_pages_onlined : 0) +
1042 		compute_balloon_floor();
1043 
1044 	/*
1045 	 * If our transaction ID is no longer current, just don't
1046 	 * send the status. This can happen if we were interrupted
1047 	 * after we picked our transaction ID.
1048 	 */
1049 	if (status.hdr.trans_id != atomic_read(&trans_id))
1050 		return;
1051 
1052 	/*
1053 	 * If the last post time that we sampled has changed,
1054 	 * we have raced, don't post the status.
1055 	 */
1056 	if (last_post != last_post_time)
1057 		return;
1058 
1059 	last_post_time = jiffies;
1060 	vmbus_sendpacket(dm->dev->channel, &status,
1061 				sizeof(struct dm_status),
1062 				(unsigned long)NULL,
1063 				VM_PKT_DATA_INBAND, 0);
1064 
1065 }
1066 
1067 static void free_balloon_pages(struct hv_dynmem_device *dm,
1068 			 union dm_mem_page_range *range_array)
1069 {
1070 	int num_pages = range_array->finfo.page_cnt;
1071 	__u64 start_frame = range_array->finfo.start_page;
1072 	struct page *pg;
1073 	int i;
1074 
1075 	for (i = 0; i < num_pages; i++) {
1076 		pg = pfn_to_page(i + start_frame);
1077 		__free_page(pg);
1078 		dm->num_pages_ballooned--;
1079 	}
1080 }
1081 
1082 
1083 
1084 static unsigned int alloc_balloon_pages(struct hv_dynmem_device *dm,
1085 					unsigned int num_pages,
1086 					struct dm_balloon_response *bl_resp,
1087 					int alloc_unit)
1088 {
1089 	unsigned int i = 0;
1090 	struct page *pg;
1091 
1092 	if (num_pages < alloc_unit)
1093 		return 0;
1094 
1095 	for (i = 0; (i * alloc_unit) < num_pages; i++) {
1096 		if (bl_resp->hdr.size + sizeof(union dm_mem_page_range) >
1097 			PAGE_SIZE)
1098 			return i * alloc_unit;
1099 
1100 		/*
1101 		 * We execute this code in a thread context. Furthermore,
1102 		 * we don't want the kernel to try too hard.
1103 		 */
1104 		pg = alloc_pages(GFP_HIGHUSER | __GFP_NORETRY |
1105 				__GFP_NOMEMALLOC | __GFP_NOWARN,
1106 				get_order(alloc_unit << PAGE_SHIFT));
1107 
1108 		if (!pg)
1109 			return i * alloc_unit;
1110 
1111 		dm->num_pages_ballooned += alloc_unit;
1112 
1113 		/*
1114 		 * If we allocatted 2M pages; split them so we
1115 		 * can free them in any order we get.
1116 		 */
1117 
1118 		if (alloc_unit != 1)
1119 			split_page(pg, get_order(alloc_unit << PAGE_SHIFT));
1120 
1121 		bl_resp->range_count++;
1122 		bl_resp->range_array[i].finfo.start_page =
1123 			page_to_pfn(pg);
1124 		bl_resp->range_array[i].finfo.page_cnt = alloc_unit;
1125 		bl_resp->hdr.size += sizeof(union dm_mem_page_range);
1126 
1127 	}
1128 
1129 	return num_pages;
1130 }
1131 
1132 
1133 
1134 static void balloon_up(struct work_struct *dummy)
1135 {
1136 	unsigned int num_pages = dm_device.balloon_wrk.num_pages;
1137 	unsigned int num_ballooned = 0;
1138 	struct dm_balloon_response *bl_resp;
1139 	int alloc_unit;
1140 	int ret;
1141 	bool done = false;
1142 	int i;
1143 	struct sysinfo val;
1144 	unsigned long floor;
1145 
1146 	/* The host balloons pages in 2M granularity. */
1147 	WARN_ON_ONCE(num_pages % PAGES_IN_2M != 0);
1148 
1149 	/*
1150 	 * We will attempt 2M allocations. However, if we fail to
1151 	 * allocate 2M chunks, we will go back to 4k allocations.
1152 	 */
1153 	alloc_unit = 512;
1154 
1155 	si_meminfo(&val);
1156 	floor = compute_balloon_floor();
1157 
1158 	/* Refuse to balloon below the floor, keep the 2M granularity. */
1159 	if (val.freeram < num_pages || val.freeram - num_pages < floor) {
1160 		num_pages = val.freeram > floor ? (val.freeram - floor) : 0;
1161 		num_pages -= num_pages % PAGES_IN_2M;
1162 	}
1163 
1164 	while (!done) {
1165 		bl_resp = (struct dm_balloon_response *)send_buffer;
1166 		memset(send_buffer, 0, PAGE_SIZE);
1167 		bl_resp->hdr.type = DM_BALLOON_RESPONSE;
1168 		bl_resp->hdr.size = sizeof(struct dm_balloon_response);
1169 		bl_resp->more_pages = 1;
1170 
1171 
1172 		num_pages -= num_ballooned;
1173 		num_ballooned = alloc_balloon_pages(&dm_device, num_pages,
1174 						    bl_resp, alloc_unit);
1175 
1176 		if (alloc_unit != 1 && num_ballooned == 0) {
1177 			alloc_unit = 1;
1178 			continue;
1179 		}
1180 
1181 		if (num_ballooned == 0 || num_ballooned == num_pages) {
1182 			bl_resp->more_pages = 0;
1183 			done = true;
1184 			dm_device.state = DM_INITIALIZED;
1185 		}
1186 
1187 		/*
1188 		 * We are pushing a lot of data through the channel;
1189 		 * deal with transient failures caused because of the
1190 		 * lack of space in the ring buffer.
1191 		 */
1192 
1193 		do {
1194 			bl_resp->hdr.trans_id = atomic_inc_return(&trans_id);
1195 			ret = vmbus_sendpacket(dm_device.dev->channel,
1196 						bl_resp,
1197 						bl_resp->hdr.size,
1198 						(unsigned long)NULL,
1199 						VM_PKT_DATA_INBAND, 0);
1200 
1201 			if (ret == -EAGAIN)
1202 				msleep(20);
1203 			post_status(&dm_device);
1204 		} while (ret == -EAGAIN);
1205 
1206 		if (ret) {
1207 			/*
1208 			 * Free up the memory we allocatted.
1209 			 */
1210 			pr_info("Balloon response failed\n");
1211 
1212 			for (i = 0; i < bl_resp->range_count; i++)
1213 				free_balloon_pages(&dm_device,
1214 						 &bl_resp->range_array[i]);
1215 
1216 			done = true;
1217 		}
1218 	}
1219 
1220 }
1221 
1222 static void balloon_down(struct hv_dynmem_device *dm,
1223 			struct dm_unballoon_request *req)
1224 {
1225 	union dm_mem_page_range *range_array = req->range_array;
1226 	int range_count = req->range_count;
1227 	struct dm_unballoon_response resp;
1228 	int i;
1229 
1230 	for (i = 0; i < range_count; i++) {
1231 		free_balloon_pages(dm, &range_array[i]);
1232 		complete(&dm_device.config_event);
1233 	}
1234 
1235 	if (req->more_pages == 1)
1236 		return;
1237 
1238 	memset(&resp, 0, sizeof(struct dm_unballoon_response));
1239 	resp.hdr.type = DM_UNBALLOON_RESPONSE;
1240 	resp.hdr.trans_id = atomic_inc_return(&trans_id);
1241 	resp.hdr.size = sizeof(struct dm_unballoon_response);
1242 
1243 	vmbus_sendpacket(dm_device.dev->channel, &resp,
1244 				sizeof(struct dm_unballoon_response),
1245 				(unsigned long)NULL,
1246 				VM_PKT_DATA_INBAND, 0);
1247 
1248 	dm->state = DM_INITIALIZED;
1249 }
1250 
1251 static void balloon_onchannelcallback(void *context);
1252 
1253 static int dm_thread_func(void *dm_dev)
1254 {
1255 	struct hv_dynmem_device *dm = dm_dev;
1256 
1257 	while (!kthread_should_stop()) {
1258 		wait_for_completion_interruptible_timeout(
1259 						&dm_device.config_event, 1*HZ);
1260 		/*
1261 		 * The host expects us to post information on the memory
1262 		 * pressure every second.
1263 		 */
1264 		reinit_completion(&dm_device.config_event);
1265 		post_status(dm);
1266 	}
1267 
1268 	return 0;
1269 }
1270 
1271 
1272 static void version_resp(struct hv_dynmem_device *dm,
1273 			struct dm_version_response *vresp)
1274 {
1275 	struct dm_version_request version_req;
1276 	int ret;
1277 
1278 	if (vresp->is_accepted) {
1279 		/*
1280 		 * We are done; wakeup the
1281 		 * context waiting for version
1282 		 * negotiation.
1283 		 */
1284 		complete(&dm->host_event);
1285 		return;
1286 	}
1287 	/*
1288 	 * If there are more versions to try, continue
1289 	 * with negotiations; if not
1290 	 * shutdown the service since we are not able
1291 	 * to negotiate a suitable version number
1292 	 * with the host.
1293 	 */
1294 	if (dm->next_version == 0)
1295 		goto version_error;
1296 
1297 	dm->next_version = 0;
1298 	memset(&version_req, 0, sizeof(struct dm_version_request));
1299 	version_req.hdr.type = DM_VERSION_REQUEST;
1300 	version_req.hdr.size = sizeof(struct dm_version_request);
1301 	version_req.hdr.trans_id = atomic_inc_return(&trans_id);
1302 	version_req.version.version = DYNMEM_PROTOCOL_VERSION_WIN7;
1303 	version_req.is_last_attempt = 1;
1304 
1305 	ret = vmbus_sendpacket(dm->dev->channel, &version_req,
1306 				sizeof(struct dm_version_request),
1307 				(unsigned long)NULL,
1308 				VM_PKT_DATA_INBAND, 0);
1309 
1310 	if (ret)
1311 		goto version_error;
1312 
1313 	return;
1314 
1315 version_error:
1316 	dm->state = DM_INIT_ERROR;
1317 	complete(&dm->host_event);
1318 }
1319 
1320 static void cap_resp(struct hv_dynmem_device *dm,
1321 			struct dm_capabilities_resp_msg *cap_resp)
1322 {
1323 	if (!cap_resp->is_accepted) {
1324 		pr_info("Capabilities not accepted by host\n");
1325 		dm->state = DM_INIT_ERROR;
1326 	}
1327 	complete(&dm->host_event);
1328 }
1329 
1330 static void balloon_onchannelcallback(void *context)
1331 {
1332 	struct hv_device *dev = context;
1333 	u32 recvlen;
1334 	u64 requestid;
1335 	struct dm_message *dm_msg;
1336 	struct dm_header *dm_hdr;
1337 	struct hv_dynmem_device *dm = hv_get_drvdata(dev);
1338 	struct dm_balloon *bal_msg;
1339 	struct dm_hot_add *ha_msg;
1340 	union dm_mem_page_range *ha_pg_range;
1341 	union dm_mem_page_range *ha_region;
1342 
1343 	memset(recv_buffer, 0, sizeof(recv_buffer));
1344 	vmbus_recvpacket(dev->channel, recv_buffer,
1345 			 PAGE_SIZE, &recvlen, &requestid);
1346 
1347 	if (recvlen > 0) {
1348 		dm_msg = (struct dm_message *)recv_buffer;
1349 		dm_hdr = &dm_msg->hdr;
1350 
1351 		switch (dm_hdr->type) {
1352 		case DM_VERSION_RESPONSE:
1353 			version_resp(dm,
1354 				 (struct dm_version_response *)dm_msg);
1355 			break;
1356 
1357 		case DM_CAPABILITIES_RESPONSE:
1358 			cap_resp(dm,
1359 				 (struct dm_capabilities_resp_msg *)dm_msg);
1360 			break;
1361 
1362 		case DM_BALLOON_REQUEST:
1363 			if (dm->state == DM_BALLOON_UP)
1364 				pr_warn("Currently ballooning\n");
1365 			bal_msg = (struct dm_balloon *)recv_buffer;
1366 			dm->state = DM_BALLOON_UP;
1367 			dm_device.balloon_wrk.num_pages = bal_msg->num_pages;
1368 			schedule_work(&dm_device.balloon_wrk.wrk);
1369 			break;
1370 
1371 		case DM_UNBALLOON_REQUEST:
1372 			dm->state = DM_BALLOON_DOWN;
1373 			balloon_down(dm,
1374 				 (struct dm_unballoon_request *)recv_buffer);
1375 			break;
1376 
1377 		case DM_MEM_HOT_ADD_REQUEST:
1378 			if (dm->state == DM_HOT_ADD)
1379 				pr_warn("Currently hot-adding\n");
1380 			dm->state = DM_HOT_ADD;
1381 			ha_msg = (struct dm_hot_add *)recv_buffer;
1382 			if (ha_msg->hdr.size == sizeof(struct dm_hot_add)) {
1383 				/*
1384 				 * This is a normal hot-add request specifying
1385 				 * hot-add memory.
1386 				 */
1387 				ha_pg_range = &ha_msg->range;
1388 				dm->ha_wrk.ha_page_range = *ha_pg_range;
1389 				dm->ha_wrk.ha_region_range.page_range = 0;
1390 			} else {
1391 				/*
1392 				 * Host is specifying that we first hot-add
1393 				 * a region and then partially populate this
1394 				 * region.
1395 				 */
1396 				dm->host_specified_ha_region = true;
1397 				ha_pg_range = &ha_msg->range;
1398 				ha_region = &ha_pg_range[1];
1399 				dm->ha_wrk.ha_page_range = *ha_pg_range;
1400 				dm->ha_wrk.ha_region_range = *ha_region;
1401 			}
1402 			schedule_work(&dm_device.ha_wrk.wrk);
1403 			break;
1404 
1405 		case DM_INFO_MESSAGE:
1406 			process_info(dm, (struct dm_info_msg *)dm_msg);
1407 			break;
1408 
1409 		default:
1410 			pr_err("Unhandled message: type: %d\n", dm_hdr->type);
1411 
1412 		}
1413 	}
1414 
1415 }
1416 
1417 static int balloon_probe(struct hv_device *dev,
1418 			const struct hv_vmbus_device_id *dev_id)
1419 {
1420 	int ret;
1421 	unsigned long t;
1422 	struct dm_version_request version_req;
1423 	struct dm_capabilities cap_msg;
1424 
1425 	do_hot_add = hot_add;
1426 
1427 	/*
1428 	 * First allocate a send buffer.
1429 	 */
1430 
1431 	send_buffer = kmalloc(PAGE_SIZE, GFP_KERNEL);
1432 	if (!send_buffer)
1433 		return -ENOMEM;
1434 
1435 	ret = vmbus_open(dev->channel, dm_ring_size, dm_ring_size, NULL, 0,
1436 			balloon_onchannelcallback, dev);
1437 
1438 	if (ret)
1439 		goto probe_error0;
1440 
1441 	dm_device.dev = dev;
1442 	dm_device.state = DM_INITIALIZING;
1443 	dm_device.next_version = DYNMEM_PROTOCOL_VERSION_WIN7;
1444 	init_completion(&dm_device.host_event);
1445 	init_completion(&dm_device.config_event);
1446 	INIT_LIST_HEAD(&dm_device.ha_region_list);
1447 	mutex_init(&dm_device.ha_region_mutex);
1448 	INIT_WORK(&dm_device.balloon_wrk.wrk, balloon_up);
1449 	INIT_WORK(&dm_device.ha_wrk.wrk, hot_add_req);
1450 	dm_device.host_specified_ha_region = false;
1451 
1452 	dm_device.thread =
1453 		 kthread_run(dm_thread_func, &dm_device, "hv_balloon");
1454 	if (IS_ERR(dm_device.thread)) {
1455 		ret = PTR_ERR(dm_device.thread);
1456 		goto probe_error1;
1457 	}
1458 
1459 #ifdef CONFIG_MEMORY_HOTPLUG
1460 	set_online_page_callback(&hv_online_page);
1461 	register_memory_notifier(&hv_memory_nb);
1462 #endif
1463 
1464 	hv_set_drvdata(dev, &dm_device);
1465 	/*
1466 	 * Initiate the hand shake with the host and negotiate
1467 	 * a version that the host can support. We start with the
1468 	 * highest version number and go down if the host cannot
1469 	 * support it.
1470 	 */
1471 	memset(&version_req, 0, sizeof(struct dm_version_request));
1472 	version_req.hdr.type = DM_VERSION_REQUEST;
1473 	version_req.hdr.size = sizeof(struct dm_version_request);
1474 	version_req.hdr.trans_id = atomic_inc_return(&trans_id);
1475 	version_req.version.version = DYNMEM_PROTOCOL_VERSION_WIN8;
1476 	version_req.is_last_attempt = 0;
1477 
1478 	ret = vmbus_sendpacket(dev->channel, &version_req,
1479 				sizeof(struct dm_version_request),
1480 				(unsigned long)NULL,
1481 				VM_PKT_DATA_INBAND, 0);
1482 	if (ret)
1483 		goto probe_error2;
1484 
1485 	t = wait_for_completion_timeout(&dm_device.host_event, 5*HZ);
1486 	if (t == 0) {
1487 		ret = -ETIMEDOUT;
1488 		goto probe_error2;
1489 	}
1490 
1491 	/*
1492 	 * If we could not negotiate a compatible version with the host
1493 	 * fail the probe function.
1494 	 */
1495 	if (dm_device.state == DM_INIT_ERROR) {
1496 		ret = -ETIMEDOUT;
1497 		goto probe_error2;
1498 	}
1499 	/*
1500 	 * Now submit our capabilities to the host.
1501 	 */
1502 	memset(&cap_msg, 0, sizeof(struct dm_capabilities));
1503 	cap_msg.hdr.type = DM_CAPABILITIES_REPORT;
1504 	cap_msg.hdr.size = sizeof(struct dm_capabilities);
1505 	cap_msg.hdr.trans_id = atomic_inc_return(&trans_id);
1506 
1507 	cap_msg.caps.cap_bits.balloon = 1;
1508 	cap_msg.caps.cap_bits.hot_add = 1;
1509 
1510 	/*
1511 	 * Specify our alignment requirements as it relates
1512 	 * memory hot-add. Specify 128MB alignment.
1513 	 */
1514 	cap_msg.caps.cap_bits.hot_add_alignment = 7;
1515 
1516 	/*
1517 	 * Currently the host does not use these
1518 	 * values and we set them to what is done in the
1519 	 * Windows driver.
1520 	 */
1521 	cap_msg.min_page_cnt = 0;
1522 	cap_msg.max_page_number = -1;
1523 
1524 	ret = vmbus_sendpacket(dev->channel, &cap_msg,
1525 				sizeof(struct dm_capabilities),
1526 				(unsigned long)NULL,
1527 				VM_PKT_DATA_INBAND, 0);
1528 	if (ret)
1529 		goto probe_error2;
1530 
1531 	t = wait_for_completion_timeout(&dm_device.host_event, 5*HZ);
1532 	if (t == 0) {
1533 		ret = -ETIMEDOUT;
1534 		goto probe_error2;
1535 	}
1536 
1537 	/*
1538 	 * If the host does not like our capabilities,
1539 	 * fail the probe function.
1540 	 */
1541 	if (dm_device.state == DM_INIT_ERROR) {
1542 		ret = -ETIMEDOUT;
1543 		goto probe_error2;
1544 	}
1545 
1546 	dm_device.state = DM_INITIALIZED;
1547 
1548 	return 0;
1549 
1550 probe_error2:
1551 #ifdef CONFIG_MEMORY_HOTPLUG
1552 	restore_online_page_callback(&hv_online_page);
1553 #endif
1554 	kthread_stop(dm_device.thread);
1555 
1556 probe_error1:
1557 	vmbus_close(dev->channel);
1558 probe_error0:
1559 	kfree(send_buffer);
1560 	return ret;
1561 }
1562 
1563 static int balloon_remove(struct hv_device *dev)
1564 {
1565 	struct hv_dynmem_device *dm = hv_get_drvdata(dev);
1566 	struct list_head *cur, *tmp;
1567 	struct hv_hotadd_state *has;
1568 
1569 	if (dm->num_pages_ballooned != 0)
1570 		pr_warn("Ballooned pages: %d\n", dm->num_pages_ballooned);
1571 
1572 	cancel_work_sync(&dm->balloon_wrk.wrk);
1573 	cancel_work_sync(&dm->ha_wrk.wrk);
1574 
1575 	vmbus_close(dev->channel);
1576 	kthread_stop(dm->thread);
1577 	kfree(send_buffer);
1578 #ifdef CONFIG_MEMORY_HOTPLUG
1579 	restore_online_page_callback(&hv_online_page);
1580 	unregister_memory_notifier(&hv_memory_nb);
1581 #endif
1582 	list_for_each_safe(cur, tmp, &dm->ha_region_list) {
1583 		has = list_entry(cur, struct hv_hotadd_state, list);
1584 		list_del(&has->list);
1585 		kfree(has);
1586 	}
1587 
1588 	return 0;
1589 }
1590 
1591 static const struct hv_vmbus_device_id id_table[] = {
1592 	/* Dynamic Memory Class ID */
1593 	/* 525074DC-8985-46e2-8057-A307DC18A502 */
1594 	{ HV_DM_GUID, },
1595 	{ },
1596 };
1597 
1598 MODULE_DEVICE_TABLE(vmbus, id_table);
1599 
1600 static  struct hv_driver balloon_drv = {
1601 	.name = "hv_balloon",
1602 	.id_table = id_table,
1603 	.probe =  balloon_probe,
1604 	.remove =  balloon_remove,
1605 };
1606 
1607 static int __init init_balloon_drv(void)
1608 {
1609 
1610 	return vmbus_driver_register(&balloon_drv);
1611 }
1612 
1613 module_init(init_balloon_drv);
1614 
1615 MODULE_DESCRIPTION("Hyper-V Balloon");
1616 MODULE_LICENSE("GPL");
1617