blob: 354da7f207b740db3aace8c54a5600e08b00b95c [file] [log] [blame]
Kyle Swenson8d8f6542021-03-15 11:02:55 -06001/*
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
62enum {
63 DYNMEM_PROTOCOL_VERSION_1 = DYNMEM_MAKE_VERSION(0, 3),
64 DYNMEM_PROTOCOL_VERSION_2 = DYNMEM_MAKE_VERSION(1, 0),
65 DYNMEM_PROTOCOL_VERSION_3 = DYNMEM_MAKE_VERSION(2, 0),
66
67 DYNMEM_PROTOCOL_VERSION_WIN7 = DYNMEM_PROTOCOL_VERSION_1,
68 DYNMEM_PROTOCOL_VERSION_WIN8 = DYNMEM_PROTOCOL_VERSION_2,
69 DYNMEM_PROTOCOL_VERSION_WIN10 = DYNMEM_PROTOCOL_VERSION_3,
70
71 DYNMEM_PROTOCOL_VERSION_CURRENT = DYNMEM_PROTOCOL_VERSION_WIN10
72};
73
74
75
76/*
77 * Message Types
78 */
79
80enum dm_message_type {
81 /*
82 * Version 0.3
83 */
84 DM_ERROR = 0,
85 DM_VERSION_REQUEST = 1,
86 DM_VERSION_RESPONSE = 2,
87 DM_CAPABILITIES_REPORT = 3,
88 DM_CAPABILITIES_RESPONSE = 4,
89 DM_STATUS_REPORT = 5,
90 DM_BALLOON_REQUEST = 6,
91 DM_BALLOON_RESPONSE = 7,
92 DM_UNBALLOON_REQUEST = 8,
93 DM_UNBALLOON_RESPONSE = 9,
94 DM_MEM_HOT_ADD_REQUEST = 10,
95 DM_MEM_HOT_ADD_RESPONSE = 11,
96 DM_VERSION_03_MAX = 11,
97 /*
98 * Version 1.0.
99 */
100 DM_INFO_MESSAGE = 12,
101 DM_VERSION_1_MAX = 12
102};
103
104
105/*
106 * Structures defining the dynamic memory management
107 * protocol.
108 */
109
110union dm_version {
111 struct {
112 __u16 minor_version;
113 __u16 major_version;
114 };
115 __u32 version;
116} __packed;
117
118
119union dm_caps {
120 struct {
121 __u64 balloon:1;
122 __u64 hot_add:1;
123 /*
124 * To support guests that may have alignment
125 * limitations on hot-add, the guest can specify
126 * its alignment requirements; a value of n
127 * represents an alignment of 2^n in mega bytes.
128 */
129 __u64 hot_add_alignment:4;
130 __u64 reservedz:58;
131 } cap_bits;
132 __u64 caps;
133} __packed;
134
135union dm_mem_page_range {
136 struct {
137 /*
138 * The PFN number of the first page in the range.
139 * 40 bits is the architectural limit of a PFN
140 * number for AMD64.
141 */
142 __u64 start_page:40;
143 /*
144 * The number of pages in the range.
145 */
146 __u64 page_cnt:24;
147 } finfo;
148 __u64 page_range;
149} __packed;
150
151
152
153/*
154 * The header for all dynamic memory messages:
155 *
156 * type: Type of the message.
157 * size: Size of the message in bytes; including the header.
158 * trans_id: The guest is responsible for manufacturing this ID.
159 */
160
161struct dm_header {
162 __u16 type;
163 __u16 size;
164 __u32 trans_id;
165} __packed;
166
167/*
168 * A generic message format for dynamic memory.
169 * Specific message formats are defined later in the file.
170 */
171
172struct dm_message {
173 struct dm_header hdr;
174 __u8 data[]; /* enclosed message */
175} __packed;
176
177
178/*
179 * Specific message types supporting the dynamic memory protocol.
180 */
181
182/*
183 * Version negotiation message. Sent from the guest to the host.
184 * The guest is free to try different versions until the host
185 * accepts the version.
186 *
187 * dm_version: The protocol version requested.
188 * is_last_attempt: If TRUE, this is the last version guest will request.
189 * reservedz: Reserved field, set to zero.
190 */
191
192struct dm_version_request {
193 struct dm_header hdr;
194 union dm_version version;
195 __u32 is_last_attempt:1;
196 __u32 reservedz:31;
197} __packed;
198
199/*
200 * Version response message; Host to Guest and indicates
201 * if the host has accepted the version sent by the guest.
202 *
203 * is_accepted: If TRUE, host has accepted the version and the guest
204 * should proceed to the next stage of the protocol. FALSE indicates that
205 * guest should re-try with a different version.
206 *
207 * reservedz: Reserved field, set to zero.
208 */
209
210struct dm_version_response {
211 struct dm_header hdr;
212 __u64 is_accepted:1;
213 __u64 reservedz:63;
214} __packed;
215
216/*
217 * Message reporting capabilities. This is sent from the guest to the
218 * host.
219 */
220
221struct dm_capabilities {
222 struct dm_header hdr;
223 union dm_caps caps;
224 __u64 min_page_cnt;
225 __u64 max_page_number;
226} __packed;
227
228/*
229 * Response to the capabilities message. This is sent from the host to the
230 * guest. This message notifies if the host has accepted the guest's
231 * capabilities. If the host has not accepted, the guest must shutdown
232 * the service.
233 *
234 * is_accepted: Indicates if the host has accepted guest's capabilities.
235 * reservedz: Must be 0.
236 */
237
238struct dm_capabilities_resp_msg {
239 struct dm_header hdr;
240 __u64 is_accepted:1;
241 __u64 reservedz:63;
242} __packed;
243
244/*
245 * This message is used to report memory pressure from the guest.
246 * This message is not part of any transaction and there is no
247 * response to this message.
248 *
249 * num_avail: Available memory in pages.
250 * num_committed: Committed memory in pages.
251 * page_file_size: The accumulated size of all page files
252 * in the system in pages.
253 * zero_free: The nunber of zero and free pages.
254 * page_file_writes: The writes to the page file in pages.
255 * io_diff: An indicator of file cache efficiency or page file activity,
256 * calculated as File Cache Page Fault Count - Page Read Count.
257 * This value is in pages.
258 *
259 * Some of these metrics are Windows specific and fortunately
260 * the algorithm on the host side that computes the guest memory
261 * pressure only uses num_committed value.
262 */
263
264struct dm_status {
265 struct dm_header hdr;
266 __u64 num_avail;
267 __u64 num_committed;
268 __u64 page_file_size;
269 __u64 zero_free;
270 __u32 page_file_writes;
271 __u32 io_diff;
272} __packed;
273
274
275/*
276 * Message to ask the guest to allocate memory - balloon up message.
277 * This message is sent from the host to the guest. The guest may not be
278 * able to allocate as much memory as requested.
279 *
280 * num_pages: number of pages to allocate.
281 */
282
283struct dm_balloon {
284 struct dm_header hdr;
285 __u32 num_pages;
286 __u32 reservedz;
287} __packed;
288
289
290/*
291 * Balloon response message; this message is sent from the guest
292 * to the host in response to the balloon message.
293 *
294 * reservedz: Reserved; must be set to zero.
295 * more_pages: If FALSE, this is the last message of the transaction.
296 * if TRUE there will atleast one more message from the guest.
297 *
298 * range_count: The number of ranges in the range array.
299 *
300 * range_array: An array of page ranges returned to the host.
301 *
302 */
303
304struct dm_balloon_response {
305 struct dm_header hdr;
306 __u32 reservedz;
307 __u32 more_pages:1;
308 __u32 range_count:31;
309 union dm_mem_page_range range_array[];
310} __packed;
311
312/*
313 * Un-balloon message; this message is sent from the host
314 * to the guest to give guest more memory.
315 *
316 * more_pages: If FALSE, this is the last message of the transaction.
317 * if TRUE there will atleast one more message from the guest.
318 *
319 * reservedz: Reserved; must be set to zero.
320 *
321 * range_count: The number of ranges in the range array.
322 *
323 * range_array: An array of page ranges returned to the host.
324 *
325 */
326
327struct dm_unballoon_request {
328 struct dm_header hdr;
329 __u32 more_pages:1;
330 __u32 reservedz:31;
331 __u32 range_count;
332 union dm_mem_page_range range_array[];
333} __packed;
334
335/*
336 * Un-balloon response message; this message is sent from the guest
337 * to the host in response to an unballoon request.
338 *
339 */
340
341struct dm_unballoon_response {
342 struct dm_header hdr;
343} __packed;
344
345
346/*
347 * Hot add request message. Message sent from the host to the guest.
348 *
349 * mem_range: Memory range to hot add.
350 *
351 * On Linux we currently don't support this since we cannot hot add
352 * arbitrary granularity of memory.
353 */
354
355struct dm_hot_add {
356 struct dm_header hdr;
357 union dm_mem_page_range range;
358} __packed;
359
360/*
361 * Hot add response message.
362 * This message is sent by the guest to report the status of a hot add request.
363 * If page_count is less than the requested page count, then the host should
364 * assume all further hot add requests will fail, since this indicates that
365 * the guest has hit an upper physical memory barrier.
366 *
367 * Hot adds may also fail due to low resources; in this case, the guest must
368 * not complete this message until the hot add can succeed, and the host must
369 * not send a new hot add request until the response is sent.
370 * If VSC fails to hot add memory DYNMEM_NUMBER_OF_UNSUCCESSFUL_HOTADD_ATTEMPTS
371 * times it fails the request.
372 *
373 *
374 * page_count: number of pages that were successfully hot added.
375 *
376 * result: result of the operation 1: success, 0: failure.
377 *
378 */
379
380struct dm_hot_add_response {
381 struct dm_header hdr;
382 __u32 page_count;
383 __u32 result;
384} __packed;
385
386/*
387 * Types of information sent from host to the guest.
388 */
389
390enum dm_info_type {
391 INFO_TYPE_MAX_PAGE_CNT = 0,
392 MAX_INFO_TYPE
393};
394
395
396/*
397 * Header for the information message.
398 */
399
400struct dm_info_header {
401 enum dm_info_type type;
402 __u32 data_size;
403} __packed;
404
405/*
406 * This message is sent from the host to the guest to pass
407 * some relevant information (win8 addition).
408 *
409 * reserved: no used.
410 * info_size: size of the information blob.
411 * info: information blob.
412 */
413
414struct dm_info_msg {
415 struct dm_header hdr;
416 __u32 reserved;
417 __u32 info_size;
418 __u8 info[];
419};
420
421/*
422 * End protocol definitions.
423 */
424
425/*
426 * State to manage hot adding memory into the guest.
427 * The range start_pfn : end_pfn specifies the range
428 * that the host has asked us to hot add. The range
429 * start_pfn : ha_end_pfn specifies the range that we have
430 * currently hot added. We hot add in multiples of 128M
431 * chunks; it is possible that we may not be able to bring
432 * online all the pages in the region. The range
433 * covered_start_pfn:covered_end_pfn defines the pages that can
434 * be brough online.
435 */
436
437struct hv_hotadd_state {
438 struct list_head list;
439 unsigned long start_pfn;
440 unsigned long covered_start_pfn;
441 unsigned long covered_end_pfn;
442 unsigned long ha_end_pfn;
443 unsigned long end_pfn;
444 /*
445 * A list of gaps.
446 */
447 struct list_head gap_list;
448};
449
450struct hv_hotadd_gap {
451 struct list_head list;
452 unsigned long start_pfn;
453 unsigned long end_pfn;
454};
455
456struct balloon_state {
457 __u32 num_pages;
458 struct work_struct wrk;
459};
460
461struct hot_add_wrk {
462 union dm_mem_page_range ha_page_range;
463 union dm_mem_page_range ha_region_range;
464 struct work_struct wrk;
465};
466
467static bool hot_add = true;
468static bool do_hot_add;
469/*
470 * Delay reporting memory pressure by
471 * the specified number of seconds.
472 */
473static uint pressure_report_delay = 45;
474
475/*
476 * The last time we posted a pressure report to host.
477 */
478static unsigned long last_post_time;
479
480module_param(hot_add, bool, (S_IRUGO | S_IWUSR));
481MODULE_PARM_DESC(hot_add, "If set attempt memory hot_add");
482
483module_param(pressure_report_delay, uint, (S_IRUGO | S_IWUSR));
484MODULE_PARM_DESC(pressure_report_delay, "Delay in secs in reporting pressure");
485static atomic_t trans_id = ATOMIC_INIT(0);
486
487static int dm_ring_size = (5 * PAGE_SIZE);
488
489/*
490 * Driver specific state.
491 */
492
493enum hv_dm_state {
494 DM_INITIALIZING = 0,
495 DM_INITIALIZED,
496 DM_BALLOON_UP,
497 DM_BALLOON_DOWN,
498 DM_HOT_ADD,
499 DM_INIT_ERROR
500};
501
502
503static __u8 recv_buffer[PAGE_SIZE];
504static __u8 *send_buffer;
505#define PAGES_IN_2M 512
506#define HA_CHUNK (32 * 1024)
507
508struct hv_dynmem_device {
509 struct hv_device *dev;
510 enum hv_dm_state state;
511 struct completion host_event;
512 struct completion config_event;
513
514 /*
515 * Number of pages we have currently ballooned out.
516 */
517 unsigned int num_pages_ballooned;
518 unsigned int num_pages_onlined;
519 unsigned int num_pages_added;
520
521 /*
522 * State to manage the ballooning (up) operation.
523 */
524 struct balloon_state balloon_wrk;
525
526 /*
527 * State to execute the "hot-add" operation.
528 */
529 struct hot_add_wrk ha_wrk;
530
531 /*
532 * This state tracks if the host has specified a hot-add
533 * region.
534 */
535 bool host_specified_ha_region;
536
537 /*
538 * State to synchronize hot-add.
539 */
540 struct completion ol_waitevent;
541 bool ha_waiting;
542 /*
543 * This thread handles hot-add
544 * requests from the host as well as notifying
545 * the host with regards to memory pressure in
546 * the guest.
547 */
548 struct task_struct *thread;
549
550 struct mutex ha_region_mutex;
551
552 /*
553 * A list of hot-add regions.
554 */
555 struct list_head ha_region_list;
556
557 /*
558 * We start with the highest version we can support
559 * and downgrade based on the host; we save here the
560 * next version to try.
561 */
562 __u32 next_version;
563};
564
565static struct hv_dynmem_device dm_device;
566
567static void post_status(struct hv_dynmem_device *dm);
568
569#ifdef CONFIG_MEMORY_HOTPLUG
570static int hv_memory_notifier(struct notifier_block *nb, unsigned long val,
571 void *v)
572{
573 struct memory_notify *mem = (struct memory_notify *)v;
574
575 switch (val) {
576 case MEM_GOING_ONLINE:
577 mutex_lock(&dm_device.ha_region_mutex);
578 break;
579
580 case MEM_ONLINE:
581 dm_device.num_pages_onlined += mem->nr_pages;
582 case MEM_CANCEL_ONLINE:
583 if (val == MEM_ONLINE ||
584 mutex_is_locked(&dm_device.ha_region_mutex))
585 mutex_unlock(&dm_device.ha_region_mutex);
586 if (dm_device.ha_waiting) {
587 dm_device.ha_waiting = false;
588 complete(&dm_device.ol_waitevent);
589 }
590 break;
591
592 case MEM_OFFLINE:
593 mutex_lock(&dm_device.ha_region_mutex);
594 dm_device.num_pages_onlined -= mem->nr_pages;
595 mutex_unlock(&dm_device.ha_region_mutex);
596 break;
597 case MEM_GOING_OFFLINE:
598 case MEM_CANCEL_OFFLINE:
599 break;
600 }
601 return NOTIFY_OK;
602}
603
604static struct notifier_block hv_memory_nb = {
605 .notifier_call = hv_memory_notifier,
606 .priority = 0
607};
608
609/* Check if the particular page is backed and can be onlined and online it. */
610static void hv_page_online_one(struct hv_hotadd_state *has, struct page *pg)
611{
612 unsigned long cur_start_pgp;
613 unsigned long cur_end_pgp;
614 struct hv_hotadd_gap *gap;
615
616 cur_start_pgp = (unsigned long)pfn_to_page(has->covered_start_pfn);
617 cur_end_pgp = (unsigned long)pfn_to_page(has->covered_end_pfn);
618
619 /* The page is not backed. */
620 if (((unsigned long)pg < cur_start_pgp) ||
621 ((unsigned long)pg >= cur_end_pgp))
622 return;
623
624 /* Check for gaps. */
625 list_for_each_entry(gap, &has->gap_list, list) {
626 cur_start_pgp = (unsigned long)
627 pfn_to_page(gap->start_pfn);
628 cur_end_pgp = (unsigned long)
629 pfn_to_page(gap->end_pfn);
630 if (((unsigned long)pg >= cur_start_pgp) &&
631 ((unsigned long)pg < cur_end_pgp)) {
632 return;
633 }
634 }
635
636 /* This frame is currently backed; online the page. */
637 __online_page_set_limits(pg);
638 __online_page_increment_counters(pg);
639 __online_page_free(pg);
640}
641
642static void hv_bring_pgs_online(struct hv_hotadd_state *has,
643 unsigned long start_pfn, unsigned long size)
644{
645 int i;
646
647 for (i = 0; i < size; i++)
648 hv_page_online_one(has, pfn_to_page(start_pfn + i));
649}
650
651static void hv_mem_hot_add(unsigned long start, unsigned long size,
652 unsigned long pfn_count,
653 struct hv_hotadd_state *has)
654{
655 int ret = 0;
656 int i, nid;
657 unsigned long start_pfn;
658 unsigned long processed_pfn;
659 unsigned long total_pfn = pfn_count;
660
661 for (i = 0; i < (size/HA_CHUNK); i++) {
662 start_pfn = start + (i * HA_CHUNK);
663 has->ha_end_pfn += HA_CHUNK;
664
665 if (total_pfn > HA_CHUNK) {
666 processed_pfn = HA_CHUNK;
667 total_pfn -= HA_CHUNK;
668 } else {
669 processed_pfn = total_pfn;
670 total_pfn = 0;
671 }
672
673 has->covered_end_pfn += processed_pfn;
674
675 init_completion(&dm_device.ol_waitevent);
676 dm_device.ha_waiting = true;
677
678 mutex_unlock(&dm_device.ha_region_mutex);
679 nid = memory_add_physaddr_to_nid(PFN_PHYS(start_pfn));
680 ret = add_memory(nid, PFN_PHYS((start_pfn)),
681 (HA_CHUNK << PAGE_SHIFT));
682
683 if (ret) {
684 pr_info("hot_add memory failed error is %d\n", ret);
685 if (ret == -EEXIST) {
686 /*
687 * This error indicates that the error
688 * is not a transient failure. This is the
689 * case where the guest's physical address map
690 * precludes hot adding memory. Stop all further
691 * memory hot-add.
692 */
693 do_hot_add = false;
694 }
695 has->ha_end_pfn -= HA_CHUNK;
696 has->covered_end_pfn -= processed_pfn;
697 mutex_lock(&dm_device.ha_region_mutex);
698 break;
699 }
700
701 /*
702 * Wait for the memory block to be onlined.
703 * Since the hot add has succeeded, it is ok to
704 * proceed even if the pages in the hot added region
705 * have not been "onlined" within the allowed time.
706 */
707 wait_for_completion_timeout(&dm_device.ol_waitevent, 5*HZ);
708 mutex_lock(&dm_device.ha_region_mutex);
709 post_status(&dm_device);
710 }
711
712 return;
713}
714
715static void hv_online_page(struct page *pg)
716{
717 struct list_head *cur;
718 struct hv_hotadd_state *has;
719 unsigned long cur_start_pgp;
720 unsigned long cur_end_pgp;
721
722 list_for_each(cur, &dm_device.ha_region_list) {
723 has = list_entry(cur, struct hv_hotadd_state, list);
724 cur_start_pgp = (unsigned long)
725 pfn_to_page(has->start_pfn);
726 cur_end_pgp = (unsigned long)pfn_to_page(has->end_pfn);
727
728 /* The page belongs to a different HAS. */
729 if (((unsigned long)pg < cur_start_pgp) ||
730 ((unsigned long)pg >= cur_end_pgp))
731 continue;
732
733 hv_page_online_one(has, pg);
734 break;
735 }
736}
737
738static int pfn_covered(unsigned long start_pfn, unsigned long pfn_cnt)
739{
740 struct list_head *cur;
741 struct hv_hotadd_state *has;
742 struct hv_hotadd_gap *gap;
743 unsigned long residual, new_inc;
744
745 if (list_empty(&dm_device.ha_region_list))
746 return false;
747
748 list_for_each(cur, &dm_device.ha_region_list) {
749 has = list_entry(cur, struct hv_hotadd_state, list);
750
751 /*
752 * If the pfn range we are dealing with is not in the current
753 * "hot add block", move on.
754 */
755 if (start_pfn < has->start_pfn || start_pfn >= has->end_pfn)
756 continue;
757
758 /*
759 * If the current start pfn is not where the covered_end
760 * is, create a gap and update covered_end_pfn.
761 */
762 if (has->covered_end_pfn != start_pfn) {
763 gap = kzalloc(sizeof(struct hv_hotadd_gap), GFP_ATOMIC);
764 if (!gap)
765 return -ENOMEM;
766
767 INIT_LIST_HEAD(&gap->list);
768 gap->start_pfn = has->covered_end_pfn;
769 gap->end_pfn = start_pfn;
770 list_add_tail(&gap->list, &has->gap_list);
771
772 has->covered_end_pfn = start_pfn;
773 }
774
775 /*
776 * If the current hot add-request extends beyond
777 * our current limit; extend it.
778 */
779 if ((start_pfn + pfn_cnt) > has->end_pfn) {
780 residual = (start_pfn + pfn_cnt - has->end_pfn);
781 /*
782 * Extend the region by multiples of HA_CHUNK.
783 */
784 new_inc = (residual / HA_CHUNK) * HA_CHUNK;
785 if (residual % HA_CHUNK)
786 new_inc += HA_CHUNK;
787
788 has->end_pfn += new_inc;
789 }
790
791 return 1;
792 }
793
794 return 0;
795}
796
797static unsigned long handle_pg_range(unsigned long pg_start,
798 unsigned long pg_count)
799{
800 unsigned long start_pfn = pg_start;
801 unsigned long pfn_cnt = pg_count;
802 unsigned long size;
803 struct list_head *cur;
804 struct hv_hotadd_state *has;
805 unsigned long pgs_ol = 0;
806 unsigned long old_covered_state;
807
808 if (list_empty(&dm_device.ha_region_list))
809 return 0;
810
811 list_for_each(cur, &dm_device.ha_region_list) {
812 has = list_entry(cur, struct hv_hotadd_state, list);
813
814 /*
815 * If the pfn range we are dealing with is not in the current
816 * "hot add block", move on.
817 */
818 if (start_pfn < has->start_pfn || start_pfn >= has->end_pfn)
819 continue;
820
821 old_covered_state = has->covered_end_pfn;
822
823 if (start_pfn < has->ha_end_pfn) {
824 /*
825 * This is the case where we are backing pages
826 * in an already hot added region. Bring
827 * these pages online first.
828 */
829 pgs_ol = has->ha_end_pfn - start_pfn;
830 if (pgs_ol > pfn_cnt)
831 pgs_ol = pfn_cnt;
832
833 has->covered_end_pfn += pgs_ol;
834 pfn_cnt -= pgs_ol;
835 /*
836 * Check if the corresponding memory block is already
837 * online by checking its last previously backed page.
838 * In case it is we need to bring rest (which was not
839 * backed previously) online too.
840 */
841 if (start_pfn > has->start_pfn &&
842 !PageReserved(pfn_to_page(start_pfn - 1)))
843 hv_bring_pgs_online(has, start_pfn, pgs_ol);
844
845 }
846
847 if ((has->ha_end_pfn < has->end_pfn) && (pfn_cnt > 0)) {
848 /*
849 * We have some residual hot add range
850 * that needs to be hot added; hot add
851 * it now. Hot add a multiple of
852 * of HA_CHUNK that fully covers the pages
853 * we have.
854 */
855 size = (has->end_pfn - has->ha_end_pfn);
856 if (pfn_cnt <= size) {
857 size = ((pfn_cnt / HA_CHUNK) * HA_CHUNK);
858 if (pfn_cnt % HA_CHUNK)
859 size += HA_CHUNK;
860 } else {
861 pfn_cnt = size;
862 }
863 hv_mem_hot_add(has->ha_end_pfn, size, pfn_cnt, has);
864 }
865 /*
866 * If we managed to online any pages that were given to us,
867 * we declare success.
868 */
869 return has->covered_end_pfn - old_covered_state;
870
871 }
872
873 return 0;
874}
875
876static unsigned long process_hot_add(unsigned long pg_start,
877 unsigned long pfn_cnt,
878 unsigned long rg_start,
879 unsigned long rg_size)
880{
881 struct hv_hotadd_state *ha_region = NULL;
882 int covered;
883
884 if (pfn_cnt == 0)
885 return 0;
886
887 if (!dm_device.host_specified_ha_region) {
888 covered = pfn_covered(pg_start, pfn_cnt);
889 if (covered < 0)
890 return 0;
891
892 if (covered)
893 goto do_pg_range;
894 }
895
896 /*
897 * If the host has specified a hot-add range; deal with it first.
898 */
899
900 if (rg_size != 0) {
901 ha_region = kzalloc(sizeof(struct hv_hotadd_state), GFP_KERNEL);
902 if (!ha_region)
903 return 0;
904
905 INIT_LIST_HEAD(&ha_region->list);
906 INIT_LIST_HEAD(&ha_region->gap_list);
907
908 list_add_tail(&ha_region->list, &dm_device.ha_region_list);
909 ha_region->start_pfn = rg_start;
910 ha_region->ha_end_pfn = rg_start;
911 ha_region->covered_start_pfn = pg_start;
912 ha_region->covered_end_pfn = pg_start;
913 ha_region->end_pfn = rg_start + rg_size;
914 }
915
916do_pg_range:
917 /*
918 * Process the page range specified; bringing them
919 * online if possible.
920 */
921 return handle_pg_range(pg_start, pfn_cnt);
922}
923
924#endif
925
926static void hot_add_req(struct work_struct *dummy)
927{
928 struct dm_hot_add_response resp;
929#ifdef CONFIG_MEMORY_HOTPLUG
930 unsigned long pg_start, pfn_cnt;
931 unsigned long rg_start, rg_sz;
932#endif
933 struct hv_dynmem_device *dm = &dm_device;
934
935 memset(&resp, 0, sizeof(struct dm_hot_add_response));
936 resp.hdr.type = DM_MEM_HOT_ADD_RESPONSE;
937 resp.hdr.size = sizeof(struct dm_hot_add_response);
938
939#ifdef CONFIG_MEMORY_HOTPLUG
940 mutex_lock(&dm_device.ha_region_mutex);
941 pg_start = dm->ha_wrk.ha_page_range.finfo.start_page;
942 pfn_cnt = dm->ha_wrk.ha_page_range.finfo.page_cnt;
943
944 rg_start = dm->ha_wrk.ha_region_range.finfo.start_page;
945 rg_sz = dm->ha_wrk.ha_region_range.finfo.page_cnt;
946
947 if ((rg_start == 0) && (!dm->host_specified_ha_region)) {
948 unsigned long region_size;
949 unsigned long region_start;
950
951 /*
952 * The host has not specified the hot-add region.
953 * Based on the hot-add page range being specified,
954 * compute a hot-add region that can cover the pages
955 * that need to be hot-added while ensuring the alignment
956 * and size requirements of Linux as it relates to hot-add.
957 */
958 region_start = pg_start;
959 region_size = (pfn_cnt / HA_CHUNK) * HA_CHUNK;
960 if (pfn_cnt % HA_CHUNK)
961 region_size += HA_CHUNK;
962
963 region_start = (pg_start / HA_CHUNK) * HA_CHUNK;
964
965 rg_start = region_start;
966 rg_sz = region_size;
967 }
968
969 if (do_hot_add)
970 resp.page_count = process_hot_add(pg_start, pfn_cnt,
971 rg_start, rg_sz);
972
973 dm->num_pages_added += resp.page_count;
974 mutex_unlock(&dm_device.ha_region_mutex);
975#endif
976 /*
977 * The result field of the response structure has the
978 * following semantics:
979 *
980 * 1. If all or some pages hot-added: Guest should return success.
981 *
982 * 2. If no pages could be hot-added:
983 *
984 * If the guest returns success, then the host
985 * will not attempt any further hot-add operations. This
986 * signifies a permanent failure.
987 *
988 * If the guest returns failure, then this failure will be
989 * treated as a transient failure and the host may retry the
990 * hot-add operation after some delay.
991 */
992 if (resp.page_count > 0)
993 resp.result = 1;
994 else if (!do_hot_add)
995 resp.result = 1;
996 else
997 resp.result = 0;
998
999 if (!do_hot_add || (resp.page_count == 0))
1000 pr_info("Memory hot add failed\n");
1001
1002 dm->state = DM_INITIALIZED;
1003 resp.hdr.trans_id = atomic_inc_return(&trans_id);
1004 vmbus_sendpacket(dm->dev->channel, &resp,
1005 sizeof(struct dm_hot_add_response),
1006 (unsigned long)NULL,
1007 VM_PKT_DATA_INBAND, 0);
1008}
1009
1010static void process_info(struct hv_dynmem_device *dm, struct dm_info_msg *msg)
1011{
1012 struct dm_info_header *info_hdr;
1013
1014 info_hdr = (struct dm_info_header *)msg->info;
1015
1016 switch (info_hdr->type) {
1017 case INFO_TYPE_MAX_PAGE_CNT:
1018 pr_info("Received INFO_TYPE_MAX_PAGE_CNT\n");
1019 pr_info("Data Size is %d\n", info_hdr->data_size);
1020 break;
1021 default:
1022 pr_info("Received Unknown type: %d\n", info_hdr->type);
1023 }
1024}
1025
1026static unsigned long compute_balloon_floor(void)
1027{
1028 unsigned long min_pages;
1029#define MB2PAGES(mb) ((mb) << (20 - PAGE_SHIFT))
1030 /* Simple continuous piecewiese linear function:
1031 * max MiB -> min MiB gradient
1032 * 0 0
1033 * 16 16
1034 * 32 24
1035 * 128 72 (1/2)
1036 * 512 168 (1/4)
1037 * 2048 360 (1/8)
1038 * 8192 744 (1/16)
1039 * 32768 1512 (1/32)
1040 */
1041 if (totalram_pages < MB2PAGES(128))
1042 min_pages = MB2PAGES(8) + (totalram_pages >> 1);
1043 else if (totalram_pages < MB2PAGES(512))
1044 min_pages = MB2PAGES(40) + (totalram_pages >> 2);
1045 else if (totalram_pages < MB2PAGES(2048))
1046 min_pages = MB2PAGES(104) + (totalram_pages >> 3);
1047 else if (totalram_pages < MB2PAGES(8192))
1048 min_pages = MB2PAGES(232) + (totalram_pages >> 4);
1049 else
1050 min_pages = MB2PAGES(488) + (totalram_pages >> 5);
1051#undef MB2PAGES
1052 return min_pages;
1053}
1054
1055/*
1056 * Post our status as it relates memory pressure to the
1057 * host. Host expects the guests to post this status
1058 * periodically at 1 second intervals.
1059 *
1060 * The metrics specified in this protocol are very Windows
1061 * specific and so we cook up numbers here to convey our memory
1062 * pressure.
1063 */
1064
1065static void post_status(struct hv_dynmem_device *dm)
1066{
1067 struct dm_status status;
1068 struct sysinfo val;
1069 unsigned long now = jiffies;
1070 unsigned long last_post = last_post_time;
1071
1072 if (pressure_report_delay > 0) {
1073 --pressure_report_delay;
1074 return;
1075 }
1076
1077 if (!time_after(now, (last_post_time + HZ)))
1078 return;
1079
1080 si_meminfo(&val);
1081 memset(&status, 0, sizeof(struct dm_status));
1082 status.hdr.type = DM_STATUS_REPORT;
1083 status.hdr.size = sizeof(struct dm_status);
1084 status.hdr.trans_id = atomic_inc_return(&trans_id);
1085
1086 /*
1087 * The host expects the guest to report free and committed memory.
1088 * Furthermore, the host expects the pressure information to include
1089 * the ballooned out pages. For a given amount of memory that we are
1090 * managing we need to compute a floor below which we should not
1091 * balloon. Compute this and add it to the pressure report.
1092 * We also need to report all offline pages (num_pages_added -
1093 * num_pages_onlined) as committed to the host, otherwise it can try
1094 * asking us to balloon them out.
1095 */
1096 status.num_avail = val.freeram;
1097 status.num_committed = vm_memory_committed() +
1098 dm->num_pages_ballooned +
1099 (dm->num_pages_added > dm->num_pages_onlined ?
1100 dm->num_pages_added - dm->num_pages_onlined : 0) +
1101 compute_balloon_floor();
1102
1103 /*
1104 * If our transaction ID is no longer current, just don't
1105 * send the status. This can happen if we were interrupted
1106 * after we picked our transaction ID.
1107 */
1108 if (status.hdr.trans_id != atomic_read(&trans_id))
1109 return;
1110
1111 /*
1112 * If the last post time that we sampled has changed,
1113 * we have raced, don't post the status.
1114 */
1115 if (last_post != last_post_time)
1116 return;
1117
1118 last_post_time = jiffies;
1119 vmbus_sendpacket(dm->dev->channel, &status,
1120 sizeof(struct dm_status),
1121 (unsigned long)NULL,
1122 VM_PKT_DATA_INBAND, 0);
1123
1124}
1125
1126static void free_balloon_pages(struct hv_dynmem_device *dm,
1127 union dm_mem_page_range *range_array)
1128{
1129 int num_pages = range_array->finfo.page_cnt;
1130 __u64 start_frame = range_array->finfo.start_page;
1131 struct page *pg;
1132 int i;
1133
1134 for (i = 0; i < num_pages; i++) {
1135 pg = pfn_to_page(i + start_frame);
1136 __free_page(pg);
1137 dm->num_pages_ballooned--;
1138 }
1139}
1140
1141
1142
1143static unsigned int alloc_balloon_pages(struct hv_dynmem_device *dm,
1144 unsigned int num_pages,
1145 struct dm_balloon_response *bl_resp,
1146 int alloc_unit)
1147{
1148 unsigned int i = 0;
1149 struct page *pg;
1150
1151 if (num_pages < alloc_unit)
1152 return 0;
1153
1154 for (i = 0; (i * alloc_unit) < num_pages; i++) {
1155 if (bl_resp->hdr.size + sizeof(union dm_mem_page_range) >
1156 PAGE_SIZE)
1157 return i * alloc_unit;
1158
1159 /*
1160 * We execute this code in a thread context. Furthermore,
1161 * we don't want the kernel to try too hard.
1162 */
1163 pg = alloc_pages(GFP_HIGHUSER | __GFP_NORETRY |
1164 __GFP_NOMEMALLOC | __GFP_NOWARN,
1165 get_order(alloc_unit << PAGE_SHIFT));
1166
1167 if (!pg)
1168 return i * alloc_unit;
1169
1170 dm->num_pages_ballooned += alloc_unit;
1171
1172 /*
1173 * If we allocatted 2M pages; split them so we
1174 * can free them in any order we get.
1175 */
1176
1177 if (alloc_unit != 1)
1178 split_page(pg, get_order(alloc_unit << PAGE_SHIFT));
1179
1180 bl_resp->range_count++;
1181 bl_resp->range_array[i].finfo.start_page =
1182 page_to_pfn(pg);
1183 bl_resp->range_array[i].finfo.page_cnt = alloc_unit;
1184 bl_resp->hdr.size += sizeof(union dm_mem_page_range);
1185
1186 }
1187
1188 return num_pages;
1189}
1190
1191
1192
1193static void balloon_up(struct work_struct *dummy)
1194{
1195 unsigned int num_pages = dm_device.balloon_wrk.num_pages;
1196 unsigned int num_ballooned = 0;
1197 struct dm_balloon_response *bl_resp;
1198 int alloc_unit;
1199 int ret;
1200 bool done = false;
1201 int i;
1202 struct sysinfo val;
1203 unsigned long floor;
1204
1205 /* The host balloons pages in 2M granularity. */
1206 WARN_ON_ONCE(num_pages % PAGES_IN_2M != 0);
1207
1208 /*
1209 * We will attempt 2M allocations. However, if we fail to
1210 * allocate 2M chunks, we will go back to 4k allocations.
1211 */
1212 alloc_unit = 512;
1213
1214 si_meminfo(&val);
1215 floor = compute_balloon_floor();
1216
1217 /* Refuse to balloon below the floor, keep the 2M granularity. */
1218 if (val.freeram < num_pages || val.freeram - num_pages < floor) {
1219 num_pages = val.freeram > floor ? (val.freeram - floor) : 0;
1220 num_pages -= num_pages % PAGES_IN_2M;
1221 }
1222
1223 while (!done) {
1224 bl_resp = (struct dm_balloon_response *)send_buffer;
1225 memset(send_buffer, 0, PAGE_SIZE);
1226 bl_resp->hdr.type = DM_BALLOON_RESPONSE;
1227 bl_resp->hdr.size = sizeof(struct dm_balloon_response);
1228 bl_resp->more_pages = 1;
1229
1230
1231 num_pages -= num_ballooned;
1232 num_ballooned = alloc_balloon_pages(&dm_device, num_pages,
1233 bl_resp, alloc_unit);
1234
1235 if (alloc_unit != 1 && num_ballooned == 0) {
1236 alloc_unit = 1;
1237 continue;
1238 }
1239
1240 if (num_ballooned == 0 || num_ballooned == num_pages) {
1241 bl_resp->more_pages = 0;
1242 done = true;
1243 dm_device.state = DM_INITIALIZED;
1244 }
1245
1246 /*
1247 * We are pushing a lot of data through the channel;
1248 * deal with transient failures caused because of the
1249 * lack of space in the ring buffer.
1250 */
1251
1252 do {
1253 bl_resp->hdr.trans_id = atomic_inc_return(&trans_id);
1254 ret = vmbus_sendpacket(dm_device.dev->channel,
1255 bl_resp,
1256 bl_resp->hdr.size,
1257 (unsigned long)NULL,
1258 VM_PKT_DATA_INBAND, 0);
1259
1260 if (ret == -EAGAIN)
1261 msleep(20);
1262 post_status(&dm_device);
1263 } while (ret == -EAGAIN);
1264
1265 if (ret) {
1266 /*
1267 * Free up the memory we allocatted.
1268 */
1269 pr_info("Balloon response failed\n");
1270
1271 for (i = 0; i < bl_resp->range_count; i++)
1272 free_balloon_pages(&dm_device,
1273 &bl_resp->range_array[i]);
1274
1275 done = true;
1276 }
1277 }
1278
1279}
1280
1281static void balloon_down(struct hv_dynmem_device *dm,
1282 struct dm_unballoon_request *req)
1283{
1284 union dm_mem_page_range *range_array = req->range_array;
1285 int range_count = req->range_count;
1286 struct dm_unballoon_response resp;
1287 int i;
1288
1289 for (i = 0; i < range_count; i++) {
1290 free_balloon_pages(dm, &range_array[i]);
1291 complete(&dm_device.config_event);
1292 }
1293
1294 if (req->more_pages == 1)
1295 return;
1296
1297 memset(&resp, 0, sizeof(struct dm_unballoon_response));
1298 resp.hdr.type = DM_UNBALLOON_RESPONSE;
1299 resp.hdr.trans_id = atomic_inc_return(&trans_id);
1300 resp.hdr.size = sizeof(struct dm_unballoon_response);
1301
1302 vmbus_sendpacket(dm_device.dev->channel, &resp,
1303 sizeof(struct dm_unballoon_response),
1304 (unsigned long)NULL,
1305 VM_PKT_DATA_INBAND, 0);
1306
1307 dm->state = DM_INITIALIZED;
1308}
1309
1310static void balloon_onchannelcallback(void *context);
1311
1312static int dm_thread_func(void *dm_dev)
1313{
1314 struct hv_dynmem_device *dm = dm_dev;
1315
1316 while (!kthread_should_stop()) {
1317 wait_for_completion_interruptible_timeout(
1318 &dm_device.config_event, 1*HZ);
1319 /*
1320 * The host expects us to post information on the memory
1321 * pressure every second.
1322 */
1323 reinit_completion(&dm_device.config_event);
1324 post_status(dm);
1325 }
1326
1327 return 0;
1328}
1329
1330
1331static void version_resp(struct hv_dynmem_device *dm,
1332 struct dm_version_response *vresp)
1333{
1334 struct dm_version_request version_req;
1335 int ret;
1336
1337 if (vresp->is_accepted) {
1338 /*
1339 * We are done; wakeup the
1340 * context waiting for version
1341 * negotiation.
1342 */
1343 complete(&dm->host_event);
1344 return;
1345 }
1346 /*
1347 * If there are more versions to try, continue
1348 * with negotiations; if not
1349 * shutdown the service since we are not able
1350 * to negotiate a suitable version number
1351 * with the host.
1352 */
1353 if (dm->next_version == 0)
1354 goto version_error;
1355
1356 memset(&version_req, 0, sizeof(struct dm_version_request));
1357 version_req.hdr.type = DM_VERSION_REQUEST;
1358 version_req.hdr.size = sizeof(struct dm_version_request);
1359 version_req.hdr.trans_id = atomic_inc_return(&trans_id);
1360 version_req.version.version = dm->next_version;
1361
1362 /*
1363 * Set the next version to try in case current version fails.
1364 * Win7 protocol ought to be the last one to try.
1365 */
1366 switch (version_req.version.version) {
1367 case DYNMEM_PROTOCOL_VERSION_WIN8:
1368 dm->next_version = DYNMEM_PROTOCOL_VERSION_WIN7;
1369 version_req.is_last_attempt = 0;
1370 break;
1371 default:
1372 dm->next_version = 0;
1373 version_req.is_last_attempt = 1;
1374 }
1375
1376 ret = vmbus_sendpacket(dm->dev->channel, &version_req,
1377 sizeof(struct dm_version_request),
1378 (unsigned long)NULL,
1379 VM_PKT_DATA_INBAND, 0);
1380
1381 if (ret)
1382 goto version_error;
1383
1384 return;
1385
1386version_error:
1387 dm->state = DM_INIT_ERROR;
1388 complete(&dm->host_event);
1389}
1390
1391static void cap_resp(struct hv_dynmem_device *dm,
1392 struct dm_capabilities_resp_msg *cap_resp)
1393{
1394 if (!cap_resp->is_accepted) {
1395 pr_info("Capabilities not accepted by host\n");
1396 dm->state = DM_INIT_ERROR;
1397 }
1398 complete(&dm->host_event);
1399}
1400
1401static void balloon_onchannelcallback(void *context)
1402{
1403 struct hv_device *dev = context;
1404 u32 recvlen;
1405 u64 requestid;
1406 struct dm_message *dm_msg;
1407 struct dm_header *dm_hdr;
1408 struct hv_dynmem_device *dm = hv_get_drvdata(dev);
1409 struct dm_balloon *bal_msg;
1410 struct dm_hot_add *ha_msg;
1411 union dm_mem_page_range *ha_pg_range;
1412 union dm_mem_page_range *ha_region;
1413
1414 memset(recv_buffer, 0, sizeof(recv_buffer));
1415 vmbus_recvpacket(dev->channel, recv_buffer,
1416 PAGE_SIZE, &recvlen, &requestid);
1417
1418 if (recvlen > 0) {
1419 dm_msg = (struct dm_message *)recv_buffer;
1420 dm_hdr = &dm_msg->hdr;
1421
1422 switch (dm_hdr->type) {
1423 case DM_VERSION_RESPONSE:
1424 version_resp(dm,
1425 (struct dm_version_response *)dm_msg);
1426 break;
1427
1428 case DM_CAPABILITIES_RESPONSE:
1429 cap_resp(dm,
1430 (struct dm_capabilities_resp_msg *)dm_msg);
1431 break;
1432
1433 case DM_BALLOON_REQUEST:
1434 if (dm->state == DM_BALLOON_UP)
1435 pr_warn("Currently ballooning\n");
1436 bal_msg = (struct dm_balloon *)recv_buffer;
1437 dm->state = DM_BALLOON_UP;
1438 dm_device.balloon_wrk.num_pages = bal_msg->num_pages;
1439 schedule_work(&dm_device.balloon_wrk.wrk);
1440 break;
1441
1442 case DM_UNBALLOON_REQUEST:
1443 dm->state = DM_BALLOON_DOWN;
1444 balloon_down(dm,
1445 (struct dm_unballoon_request *)recv_buffer);
1446 break;
1447
1448 case DM_MEM_HOT_ADD_REQUEST:
1449 if (dm->state == DM_HOT_ADD)
1450 pr_warn("Currently hot-adding\n");
1451 dm->state = DM_HOT_ADD;
1452 ha_msg = (struct dm_hot_add *)recv_buffer;
1453 if (ha_msg->hdr.size == sizeof(struct dm_hot_add)) {
1454 /*
1455 * This is a normal hot-add request specifying
1456 * hot-add memory.
1457 */
1458 ha_pg_range = &ha_msg->range;
1459 dm->ha_wrk.ha_page_range = *ha_pg_range;
1460 dm->ha_wrk.ha_region_range.page_range = 0;
1461 } else {
1462 /*
1463 * Host is specifying that we first hot-add
1464 * a region and then partially populate this
1465 * region.
1466 */
1467 dm->host_specified_ha_region = true;
1468 ha_pg_range = &ha_msg->range;
1469 ha_region = &ha_pg_range[1];
1470 dm->ha_wrk.ha_page_range = *ha_pg_range;
1471 dm->ha_wrk.ha_region_range = *ha_region;
1472 }
1473 schedule_work(&dm_device.ha_wrk.wrk);
1474 break;
1475
1476 case DM_INFO_MESSAGE:
1477 process_info(dm, (struct dm_info_msg *)dm_msg);
1478 break;
1479
1480 default:
1481 pr_err("Unhandled message: type: %d\n", dm_hdr->type);
1482
1483 }
1484 }
1485
1486}
1487
1488static int balloon_probe(struct hv_device *dev,
1489 const struct hv_vmbus_device_id *dev_id)
1490{
1491 int ret;
1492 unsigned long t;
1493 struct dm_version_request version_req;
1494 struct dm_capabilities cap_msg;
1495
1496 do_hot_add = hot_add;
1497
1498 /*
1499 * First allocate a send buffer.
1500 */
1501
1502 send_buffer = kmalloc(PAGE_SIZE, GFP_KERNEL);
1503 if (!send_buffer)
1504 return -ENOMEM;
1505
1506 ret = vmbus_open(dev->channel, dm_ring_size, dm_ring_size, NULL, 0,
1507 balloon_onchannelcallback, dev);
1508
1509 if (ret)
1510 goto probe_error0;
1511
1512 dm_device.dev = dev;
1513 dm_device.state = DM_INITIALIZING;
1514 dm_device.next_version = DYNMEM_PROTOCOL_VERSION_WIN8;
1515 init_completion(&dm_device.host_event);
1516 init_completion(&dm_device.config_event);
1517 INIT_LIST_HEAD(&dm_device.ha_region_list);
1518 mutex_init(&dm_device.ha_region_mutex);
1519 INIT_WORK(&dm_device.balloon_wrk.wrk, balloon_up);
1520 INIT_WORK(&dm_device.ha_wrk.wrk, hot_add_req);
1521 dm_device.host_specified_ha_region = false;
1522
1523 dm_device.thread =
1524 kthread_run(dm_thread_func, &dm_device, "hv_balloon");
1525 if (IS_ERR(dm_device.thread)) {
1526 ret = PTR_ERR(dm_device.thread);
1527 goto probe_error1;
1528 }
1529
1530#ifdef CONFIG_MEMORY_HOTPLUG
1531 set_online_page_callback(&hv_online_page);
1532 register_memory_notifier(&hv_memory_nb);
1533#endif
1534
1535 hv_set_drvdata(dev, &dm_device);
1536 /*
1537 * Initiate the hand shake with the host and negotiate
1538 * a version that the host can support. We start with the
1539 * highest version number and go down if the host cannot
1540 * support it.
1541 */
1542 memset(&version_req, 0, sizeof(struct dm_version_request));
1543 version_req.hdr.type = DM_VERSION_REQUEST;
1544 version_req.hdr.size = sizeof(struct dm_version_request);
1545 version_req.hdr.trans_id = atomic_inc_return(&trans_id);
1546 version_req.version.version = DYNMEM_PROTOCOL_VERSION_WIN10;
1547 version_req.is_last_attempt = 0;
1548
1549 ret = vmbus_sendpacket(dev->channel, &version_req,
1550 sizeof(struct dm_version_request),
1551 (unsigned long)NULL,
1552 VM_PKT_DATA_INBAND, 0);
1553 if (ret)
1554 goto probe_error2;
1555
1556 t = wait_for_completion_timeout(&dm_device.host_event, 5*HZ);
1557 if (t == 0) {
1558 ret = -ETIMEDOUT;
1559 goto probe_error2;
1560 }
1561
1562 /*
1563 * If we could not negotiate a compatible version with the host
1564 * fail the probe function.
1565 */
1566 if (dm_device.state == DM_INIT_ERROR) {
1567 ret = -ETIMEDOUT;
1568 goto probe_error2;
1569 }
1570 /*
1571 * Now submit our capabilities to the host.
1572 */
1573 memset(&cap_msg, 0, sizeof(struct dm_capabilities));
1574 cap_msg.hdr.type = DM_CAPABILITIES_REPORT;
1575 cap_msg.hdr.size = sizeof(struct dm_capabilities);
1576 cap_msg.hdr.trans_id = atomic_inc_return(&trans_id);
1577
1578 cap_msg.caps.cap_bits.balloon = 1;
1579 cap_msg.caps.cap_bits.hot_add = 1;
1580
1581 /*
1582 * Specify our alignment requirements as it relates
1583 * memory hot-add. Specify 128MB alignment.
1584 */
1585 cap_msg.caps.cap_bits.hot_add_alignment = 7;
1586
1587 /*
1588 * Currently the host does not use these
1589 * values and we set them to what is done in the
1590 * Windows driver.
1591 */
1592 cap_msg.min_page_cnt = 0;
1593 cap_msg.max_page_number = -1;
1594
1595 ret = vmbus_sendpacket(dev->channel, &cap_msg,
1596 sizeof(struct dm_capabilities),
1597 (unsigned long)NULL,
1598 VM_PKT_DATA_INBAND, 0);
1599 if (ret)
1600 goto probe_error2;
1601
1602 t = wait_for_completion_timeout(&dm_device.host_event, 5*HZ);
1603 if (t == 0) {
1604 ret = -ETIMEDOUT;
1605 goto probe_error2;
1606 }
1607
1608 /*
1609 * If the host does not like our capabilities,
1610 * fail the probe function.
1611 */
1612 if (dm_device.state == DM_INIT_ERROR) {
1613 ret = -ETIMEDOUT;
1614 goto probe_error2;
1615 }
1616
1617 dm_device.state = DM_INITIALIZED;
1618
1619 return 0;
1620
1621probe_error2:
1622#ifdef CONFIG_MEMORY_HOTPLUG
1623 restore_online_page_callback(&hv_online_page);
1624#endif
1625 kthread_stop(dm_device.thread);
1626
1627probe_error1:
1628 vmbus_close(dev->channel);
1629probe_error0:
1630 kfree(send_buffer);
1631 return ret;
1632}
1633
1634static int balloon_remove(struct hv_device *dev)
1635{
1636 struct hv_dynmem_device *dm = hv_get_drvdata(dev);
1637 struct list_head *cur, *tmp;
1638 struct hv_hotadd_state *has;
1639 struct hv_hotadd_gap *gap, *tmp_gap;
1640
1641 if (dm->num_pages_ballooned != 0)
1642 pr_warn("Ballooned pages: %d\n", dm->num_pages_ballooned);
1643
1644 cancel_work_sync(&dm->balloon_wrk.wrk);
1645 cancel_work_sync(&dm->ha_wrk.wrk);
1646
1647 vmbus_close(dev->channel);
1648 kthread_stop(dm->thread);
1649 kfree(send_buffer);
1650#ifdef CONFIG_MEMORY_HOTPLUG
1651 restore_online_page_callback(&hv_online_page);
1652 unregister_memory_notifier(&hv_memory_nb);
1653#endif
1654 list_for_each_safe(cur, tmp, &dm->ha_region_list) {
1655 has = list_entry(cur, struct hv_hotadd_state, list);
1656 list_for_each_entry_safe(gap, tmp_gap, &has->gap_list, list) {
1657 list_del(&gap->list);
1658 kfree(gap);
1659 }
1660 list_del(&has->list);
1661 kfree(has);
1662 }
1663
1664 return 0;
1665}
1666
1667static const struct hv_vmbus_device_id id_table[] = {
1668 /* Dynamic Memory Class ID */
1669 /* 525074DC-8985-46e2-8057-A307DC18A502 */
1670 { HV_DM_GUID, },
1671 { },
1672};
1673
1674MODULE_DEVICE_TABLE(vmbus, id_table);
1675
1676static struct hv_driver balloon_drv = {
1677 .name = "hv_balloon",
1678 .id_table = id_table,
1679 .probe = balloon_probe,
1680 .remove = balloon_remove,
1681};
1682
1683static int __init init_balloon_drv(void)
1684{
1685
1686 return vmbus_driver_register(&balloon_drv);
1687}
1688
1689module_init(init_balloon_drv);
1690
1691MODULE_DESCRIPTION("Hyper-V Balloon");
1692MODULE_LICENSE("GPL");