Loading...
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
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_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_end_pfn;
441 unsigned long ha_end_pfn;
442 unsigned long end_pfn;
443};
444
445struct balloon_state {
446 __u32 num_pages;
447 struct work_struct wrk;
448};
449
450struct hot_add_wrk {
451 union dm_mem_page_range ha_page_range;
452 union dm_mem_page_range ha_region_range;
453 struct work_struct wrk;
454};
455
456static bool hot_add = true;
457static bool do_hot_add;
458/*
459 * Delay reporting memory pressure by
460 * the specified number of seconds.
461 */
462static uint pressure_report_delay = 45;
463
464/*
465 * The last time we posted a pressure report to host.
466 */
467static unsigned long last_post_time;
468
469module_param(hot_add, bool, (S_IRUGO | S_IWUSR));
470MODULE_PARM_DESC(hot_add, "If set attempt memory hot_add");
471
472module_param(pressure_report_delay, uint, (S_IRUGO | S_IWUSR));
473MODULE_PARM_DESC(pressure_report_delay, "Delay in secs in reporting pressure");
474static atomic_t trans_id = ATOMIC_INIT(0);
475
476static int dm_ring_size = (5 * PAGE_SIZE);
477
478/*
479 * Driver specific state.
480 */
481
482enum hv_dm_state {
483 DM_INITIALIZING = 0,
484 DM_INITIALIZED,
485 DM_BALLOON_UP,
486 DM_BALLOON_DOWN,
487 DM_HOT_ADD,
488 DM_INIT_ERROR
489};
490
491
492static __u8 recv_buffer[PAGE_SIZE];
493static __u8 *send_buffer;
494#define PAGES_IN_2M 512
495#define HA_CHUNK (32 * 1024)
496
497struct hv_dynmem_device {
498 struct hv_device *dev;
499 enum hv_dm_state state;
500 struct completion host_event;
501 struct completion config_event;
502
503 /*
504 * Number of pages we have currently ballooned out.
505 */
506 unsigned int num_pages_ballooned;
507 unsigned int num_pages_onlined;
508 unsigned int num_pages_added;
509
510 /*
511 * State to manage the ballooning (up) operation.
512 */
513 struct balloon_state balloon_wrk;
514
515 /*
516 * State to execute the "hot-add" operation.
517 */
518 struct hot_add_wrk ha_wrk;
519
520 /*
521 * This state tracks if the host has specified a hot-add
522 * region.
523 */
524 bool host_specified_ha_region;
525
526 /*
527 * State to synchronize hot-add.
528 */
529 struct completion ol_waitevent;
530 bool ha_waiting;
531 /*
532 * This thread handles hot-add
533 * requests from the host as well as notifying
534 * the host with regards to memory pressure in
535 * the guest.
536 */
537 struct task_struct *thread;
538
539 struct mutex ha_region_mutex;
540
541 /*
542 * A list of hot-add regions.
543 */
544 struct list_head ha_region_list;
545
546 /*
547 * We start with the highest version we can support
548 * and downgrade based on the host; we save here the
549 * next version to try.
550 */
551 __u32 next_version;
552};
553
554static struct hv_dynmem_device dm_device;
555
556static void post_status(struct hv_dynmem_device *dm);
557
558#ifdef CONFIG_MEMORY_HOTPLUG
559static int hv_memory_notifier(struct notifier_block *nb, unsigned long val,
560 void *v)
561{
562 struct memory_notify *mem = (struct memory_notify *)v;
563
564 switch (val) {
565 case MEM_GOING_ONLINE:
566 mutex_lock(&dm_device.ha_region_mutex);
567 break;
568
569 case MEM_ONLINE:
570 dm_device.num_pages_onlined += mem->nr_pages;
571 case MEM_CANCEL_ONLINE:
572 if (val == MEM_ONLINE ||
573 mutex_is_locked(&dm_device.ha_region_mutex))
574 mutex_unlock(&dm_device.ha_region_mutex);
575 if (dm_device.ha_waiting) {
576 dm_device.ha_waiting = false;
577 complete(&dm_device.ol_waitevent);
578 }
579 break;
580
581 case MEM_OFFLINE:
582 mutex_lock(&dm_device.ha_region_mutex);
583 dm_device.num_pages_onlined -= mem->nr_pages;
584 mutex_unlock(&dm_device.ha_region_mutex);
585 break;
586 case MEM_GOING_OFFLINE:
587 case MEM_CANCEL_OFFLINE:
588 break;
589 }
590 return NOTIFY_OK;
591}
592
593static struct notifier_block hv_memory_nb = {
594 .notifier_call = hv_memory_notifier,
595 .priority = 0
596};
597
598
599static void hv_bring_pgs_online(unsigned long start_pfn, unsigned long size)
600{
601 int i;
602
603 for (i = 0; i < size; i++) {
604 struct page *pg;
605 pg = pfn_to_page(start_pfn + i);
606 __online_page_set_limits(pg);
607 __online_page_increment_counters(pg);
608 __online_page_free(pg);
609 }
610}
611
612static void hv_mem_hot_add(unsigned long start, unsigned long size,
613 unsigned long pfn_count,
614 struct hv_hotadd_state *has)
615{
616 int ret = 0;
617 int i, nid;
618 unsigned long start_pfn;
619 unsigned long processed_pfn;
620 unsigned long total_pfn = pfn_count;
621
622 for (i = 0; i < (size/HA_CHUNK); i++) {
623 start_pfn = start + (i * HA_CHUNK);
624 has->ha_end_pfn += HA_CHUNK;
625
626 if (total_pfn > HA_CHUNK) {
627 processed_pfn = HA_CHUNK;
628 total_pfn -= HA_CHUNK;
629 } else {
630 processed_pfn = total_pfn;
631 total_pfn = 0;
632 }
633
634 has->covered_end_pfn += processed_pfn;
635
636 init_completion(&dm_device.ol_waitevent);
637 dm_device.ha_waiting = true;
638
639 mutex_unlock(&dm_device.ha_region_mutex);
640 nid = memory_add_physaddr_to_nid(PFN_PHYS(start_pfn));
641 ret = add_memory(nid, PFN_PHYS((start_pfn)),
642 (HA_CHUNK << PAGE_SHIFT));
643
644 if (ret) {
645 pr_info("hot_add memory failed error is %d\n", ret);
646 if (ret == -EEXIST) {
647 /*
648 * This error indicates that the error
649 * is not a transient failure. This is the
650 * case where the guest's physical address map
651 * precludes hot adding memory. Stop all further
652 * memory hot-add.
653 */
654 do_hot_add = false;
655 }
656 has->ha_end_pfn -= HA_CHUNK;
657 has->covered_end_pfn -= processed_pfn;
658 mutex_lock(&dm_device.ha_region_mutex);
659 break;
660 }
661
662 /*
663 * Wait for the memory block to be onlined.
664 * Since the hot add has succeeded, it is ok to
665 * proceed even if the pages in the hot added region
666 * have not been "onlined" within the allowed time.
667 */
668 wait_for_completion_timeout(&dm_device.ol_waitevent, 5*HZ);
669 mutex_lock(&dm_device.ha_region_mutex);
670 post_status(&dm_device);
671 }
672
673 return;
674}
675
676static void hv_online_page(struct page *pg)
677{
678 struct list_head *cur;
679 struct hv_hotadd_state *has;
680 unsigned long cur_start_pgp;
681 unsigned long cur_end_pgp;
682
683 list_for_each(cur, &dm_device.ha_region_list) {
684 has = list_entry(cur, struct hv_hotadd_state, list);
685 cur_start_pgp = (unsigned long)pfn_to_page(has->start_pfn);
686 cur_end_pgp = (unsigned long)pfn_to_page(has->covered_end_pfn);
687
688 if (((unsigned long)pg >= cur_start_pgp) &&
689 ((unsigned long)pg < cur_end_pgp)) {
690 /*
691 * This frame is currently backed; online the
692 * page.
693 */
694 __online_page_set_limits(pg);
695 __online_page_increment_counters(pg);
696 __online_page_free(pg);
697 }
698 }
699}
700
701static bool pfn_covered(unsigned long start_pfn, unsigned long pfn_cnt)
702{
703 struct list_head *cur;
704 struct hv_hotadd_state *has;
705 unsigned long residual, new_inc;
706
707 if (list_empty(&dm_device.ha_region_list))
708 return false;
709
710 list_for_each(cur, &dm_device.ha_region_list) {
711 has = list_entry(cur, struct hv_hotadd_state, list);
712
713 /*
714 * If the pfn range we are dealing with is not in the current
715 * "hot add block", move on.
716 */
717 if ((start_pfn >= has->end_pfn))
718 continue;
719 /*
720 * If the current hot add-request extends beyond
721 * our current limit; extend it.
722 */
723 if ((start_pfn + pfn_cnt) > has->end_pfn) {
724 residual = (start_pfn + pfn_cnt - has->end_pfn);
725 /*
726 * Extend the region by multiples of HA_CHUNK.
727 */
728 new_inc = (residual / HA_CHUNK) * HA_CHUNK;
729 if (residual % HA_CHUNK)
730 new_inc += HA_CHUNK;
731
732 has->end_pfn += new_inc;
733 }
734
735 /*
736 * If the current start pfn is not where the covered_end
737 * is, update it.
738 */
739
740 if (has->covered_end_pfn != start_pfn)
741 has->covered_end_pfn = start_pfn;
742
743 return true;
744
745 }
746
747 return false;
748}
749
750static unsigned long handle_pg_range(unsigned long pg_start,
751 unsigned long pg_count)
752{
753 unsigned long start_pfn = pg_start;
754 unsigned long pfn_cnt = pg_count;
755 unsigned long size;
756 struct list_head *cur;
757 struct hv_hotadd_state *has;
758 unsigned long pgs_ol = 0;
759 unsigned long old_covered_state;
760
761 if (list_empty(&dm_device.ha_region_list))
762 return 0;
763
764 list_for_each(cur, &dm_device.ha_region_list) {
765 has = list_entry(cur, struct hv_hotadd_state, list);
766
767 /*
768 * If the pfn range we are dealing with is not in the current
769 * "hot add block", move on.
770 */
771 if ((start_pfn >= has->end_pfn))
772 continue;
773
774 old_covered_state = has->covered_end_pfn;
775
776 if (start_pfn < has->ha_end_pfn) {
777 /*
778 * This is the case where we are backing pages
779 * in an already hot added region. Bring
780 * these pages online first.
781 */
782 pgs_ol = has->ha_end_pfn - start_pfn;
783 if (pgs_ol > pfn_cnt)
784 pgs_ol = pfn_cnt;
785
786 /*
787 * Check if the corresponding memory block is already
788 * online by checking its last previously backed page.
789 * In case it is we need to bring rest (which was not
790 * backed previously) online too.
791 */
792 if (start_pfn > has->start_pfn &&
793 !PageReserved(pfn_to_page(start_pfn - 1)))
794 hv_bring_pgs_online(start_pfn, pgs_ol);
795
796 has->covered_end_pfn += pgs_ol;
797 pfn_cnt -= pgs_ol;
798 }
799
800 if ((has->ha_end_pfn < has->end_pfn) && (pfn_cnt > 0)) {
801 /*
802 * We have some residual hot add range
803 * that needs to be hot added; hot add
804 * it now. Hot add a multiple of
805 * of HA_CHUNK that fully covers the pages
806 * we have.
807 */
808 size = (has->end_pfn - has->ha_end_pfn);
809 if (pfn_cnt <= size) {
810 size = ((pfn_cnt / HA_CHUNK) * HA_CHUNK);
811 if (pfn_cnt % HA_CHUNK)
812 size += HA_CHUNK;
813 } else {
814 pfn_cnt = size;
815 }
816 hv_mem_hot_add(has->ha_end_pfn, size, pfn_cnt, has);
817 }
818 /*
819 * If we managed to online any pages that were given to us,
820 * we declare success.
821 */
822 return has->covered_end_pfn - old_covered_state;
823
824 }
825
826 return 0;
827}
828
829static unsigned long process_hot_add(unsigned long pg_start,
830 unsigned long pfn_cnt,
831 unsigned long rg_start,
832 unsigned long rg_size)
833{
834 struct hv_hotadd_state *ha_region = NULL;
835
836 if (pfn_cnt == 0)
837 return 0;
838
839 if (!dm_device.host_specified_ha_region)
840 if (pfn_covered(pg_start, pfn_cnt))
841 goto do_pg_range;
842
843 /*
844 * If the host has specified a hot-add range; deal with it first.
845 */
846
847 if (rg_size != 0) {
848 ha_region = kzalloc(sizeof(struct hv_hotadd_state), GFP_KERNEL);
849 if (!ha_region)
850 return 0;
851
852 INIT_LIST_HEAD(&ha_region->list);
853
854 list_add_tail(&ha_region->list, &dm_device.ha_region_list);
855 ha_region->start_pfn = rg_start;
856 ha_region->ha_end_pfn = rg_start;
857 ha_region->covered_end_pfn = pg_start;
858 ha_region->end_pfn = rg_start + rg_size;
859 }
860
861do_pg_range:
862 /*
863 * Process the page range specified; bringing them
864 * online if possible.
865 */
866 return handle_pg_range(pg_start, pfn_cnt);
867}
868
869#endif
870
871static void hot_add_req(struct work_struct *dummy)
872{
873 struct dm_hot_add_response resp;
874#ifdef CONFIG_MEMORY_HOTPLUG
875 unsigned long pg_start, pfn_cnt;
876 unsigned long rg_start, rg_sz;
877#endif
878 struct hv_dynmem_device *dm = &dm_device;
879
880 memset(&resp, 0, sizeof(struct dm_hot_add_response));
881 resp.hdr.type = DM_MEM_HOT_ADD_RESPONSE;
882 resp.hdr.size = sizeof(struct dm_hot_add_response);
883
884#ifdef CONFIG_MEMORY_HOTPLUG
885 mutex_lock(&dm_device.ha_region_mutex);
886 pg_start = dm->ha_wrk.ha_page_range.finfo.start_page;
887 pfn_cnt = dm->ha_wrk.ha_page_range.finfo.page_cnt;
888
889 rg_start = dm->ha_wrk.ha_region_range.finfo.start_page;
890 rg_sz = dm->ha_wrk.ha_region_range.finfo.page_cnt;
891
892 if ((rg_start == 0) && (!dm->host_specified_ha_region)) {
893 unsigned long region_size;
894 unsigned long region_start;
895
896 /*
897 * The host has not specified the hot-add region.
898 * Based on the hot-add page range being specified,
899 * compute a hot-add region that can cover the pages
900 * that need to be hot-added while ensuring the alignment
901 * and size requirements of Linux as it relates to hot-add.
902 */
903 region_start = pg_start;
904 region_size = (pfn_cnt / HA_CHUNK) * HA_CHUNK;
905 if (pfn_cnt % HA_CHUNK)
906 region_size += HA_CHUNK;
907
908 region_start = (pg_start / HA_CHUNK) * HA_CHUNK;
909
910 rg_start = region_start;
911 rg_sz = region_size;
912 }
913
914 if (do_hot_add)
915 resp.page_count = process_hot_add(pg_start, pfn_cnt,
916 rg_start, rg_sz);
917
918 dm->num_pages_added += resp.page_count;
919 mutex_unlock(&dm_device.ha_region_mutex);
920#endif
921 /*
922 * The result field of the response structure has the
923 * following semantics:
924 *
925 * 1. If all or some pages hot-added: Guest should return success.
926 *
927 * 2. If no pages could be hot-added:
928 *
929 * If the guest returns success, then the host
930 * will not attempt any further hot-add operations. This
931 * signifies a permanent failure.
932 *
933 * If the guest returns failure, then this failure will be
934 * treated as a transient failure and the host may retry the
935 * hot-add operation after some delay.
936 */
937 if (resp.page_count > 0)
938 resp.result = 1;
939 else if (!do_hot_add)
940 resp.result = 1;
941 else
942 resp.result = 0;
943
944 if (!do_hot_add || (resp.page_count == 0))
945 pr_info("Memory hot add failed\n");
946
947 dm->state = DM_INITIALIZED;
948 resp.hdr.trans_id = atomic_inc_return(&trans_id);
949 vmbus_sendpacket(dm->dev->channel, &resp,
950 sizeof(struct dm_hot_add_response),
951 (unsigned long)NULL,
952 VM_PKT_DATA_INBAND, 0);
953}
954
955static void process_info(struct hv_dynmem_device *dm, struct dm_info_msg *msg)
956{
957 struct dm_info_header *info_hdr;
958
959 info_hdr = (struct dm_info_header *)msg->info;
960
961 switch (info_hdr->type) {
962 case INFO_TYPE_MAX_PAGE_CNT:
963 pr_info("Received INFO_TYPE_MAX_PAGE_CNT\n");
964 pr_info("Data Size is %d\n", info_hdr->data_size);
965 break;
966 default:
967 pr_info("Received Unknown type: %d\n", info_hdr->type);
968 }
969}
970
971static unsigned long compute_balloon_floor(void)
972{
973 unsigned long min_pages;
974#define MB2PAGES(mb) ((mb) << (20 - PAGE_SHIFT))
975 /* Simple continuous piecewiese linear function:
976 * max MiB -> min MiB gradient
977 * 0 0
978 * 16 16
979 * 32 24
980 * 128 72 (1/2)
981 * 512 168 (1/4)
982 * 2048 360 (1/8)
983 * 8192 744 (1/16)
984 * 32768 1512 (1/32)
985 */
986 if (totalram_pages < MB2PAGES(128))
987 min_pages = MB2PAGES(8) + (totalram_pages >> 1);
988 else if (totalram_pages < MB2PAGES(512))
989 min_pages = MB2PAGES(40) + (totalram_pages >> 2);
990 else if (totalram_pages < MB2PAGES(2048))
991 min_pages = MB2PAGES(104) + (totalram_pages >> 3);
992 else if (totalram_pages < MB2PAGES(8192))
993 min_pages = MB2PAGES(232) + (totalram_pages >> 4);
994 else
995 min_pages = MB2PAGES(488) + (totalram_pages >> 5);
996#undef MB2PAGES
997 return min_pages;
998}
999
1000/*
1001 * Post our status as it relates memory pressure to the
1002 * host. Host expects the guests to post this status
1003 * periodically at 1 second intervals.
1004 *
1005 * The metrics specified in this protocol are very Windows
1006 * specific and so we cook up numbers here to convey our memory
1007 * pressure.
1008 */
1009
1010static void post_status(struct hv_dynmem_device *dm)
1011{
1012 struct dm_status status;
1013 struct sysinfo val;
1014 unsigned long now = jiffies;
1015 unsigned long last_post = last_post_time;
1016
1017 if (pressure_report_delay > 0) {
1018 --pressure_report_delay;
1019 return;
1020 }
1021
1022 if (!time_after(now, (last_post_time + HZ)))
1023 return;
1024
1025 si_meminfo(&val);
1026 memset(&status, 0, sizeof(struct dm_status));
1027 status.hdr.type = DM_STATUS_REPORT;
1028 status.hdr.size = sizeof(struct dm_status);
1029 status.hdr.trans_id = atomic_inc_return(&trans_id);
1030
1031 /*
1032 * The host expects the guest to report free and committed memory.
1033 * Furthermore, the host expects the pressure information to include
1034 * the ballooned out pages. For a given amount of memory that we are
1035 * managing we need to compute a floor below which we should not
1036 * balloon. Compute this and add it to the pressure report.
1037 * We also need to report all offline pages (num_pages_added -
1038 * num_pages_onlined) as committed to the host, otherwise it can try
1039 * asking us to balloon them out.
1040 */
1041 status.num_avail = val.freeram;
1042 status.num_committed = vm_memory_committed() +
1043 dm->num_pages_ballooned +
1044 (dm->num_pages_added > dm->num_pages_onlined ?
1045 dm->num_pages_added - dm->num_pages_onlined : 0) +
1046 compute_balloon_floor();
1047
1048 /*
1049 * If our transaction ID is no longer current, just don't
1050 * send the status. This can happen if we were interrupted
1051 * after we picked our transaction ID.
1052 */
1053 if (status.hdr.trans_id != atomic_read(&trans_id))
1054 return;
1055
1056 /*
1057 * If the last post time that we sampled has changed,
1058 * we have raced, don't post the status.
1059 */
1060 if (last_post != last_post_time)
1061 return;
1062
1063 last_post_time = jiffies;
1064 vmbus_sendpacket(dm->dev->channel, &status,
1065 sizeof(struct dm_status),
1066 (unsigned long)NULL,
1067 VM_PKT_DATA_INBAND, 0);
1068
1069}
1070
1071static void free_balloon_pages(struct hv_dynmem_device *dm,
1072 union dm_mem_page_range *range_array)
1073{
1074 int num_pages = range_array->finfo.page_cnt;
1075 __u64 start_frame = range_array->finfo.start_page;
1076 struct page *pg;
1077 int i;
1078
1079 for (i = 0; i < num_pages; i++) {
1080 pg = pfn_to_page(i + start_frame);
1081 __free_page(pg);
1082 dm->num_pages_ballooned--;
1083 }
1084}
1085
1086
1087
1088static unsigned int alloc_balloon_pages(struct hv_dynmem_device *dm,
1089 unsigned int num_pages,
1090 struct dm_balloon_response *bl_resp,
1091 int alloc_unit)
1092{
1093 unsigned int i = 0;
1094 struct page *pg;
1095
1096 if (num_pages < alloc_unit)
1097 return 0;
1098
1099 for (i = 0; (i * alloc_unit) < num_pages; i++) {
1100 if (bl_resp->hdr.size + sizeof(union dm_mem_page_range) >
1101 PAGE_SIZE)
1102 return i * alloc_unit;
1103
1104 /*
1105 * We execute this code in a thread context. Furthermore,
1106 * we don't want the kernel to try too hard.
1107 */
1108 pg = alloc_pages(GFP_HIGHUSER | __GFP_NORETRY |
1109 __GFP_NOMEMALLOC | __GFP_NOWARN,
1110 get_order(alloc_unit << PAGE_SHIFT));
1111
1112 if (!pg)
1113 return i * alloc_unit;
1114
1115 dm->num_pages_ballooned += alloc_unit;
1116
1117 /*
1118 * If we allocatted 2M pages; split them so we
1119 * can free them in any order we get.
1120 */
1121
1122 if (alloc_unit != 1)
1123 split_page(pg, get_order(alloc_unit << PAGE_SHIFT));
1124
1125 bl_resp->range_count++;
1126 bl_resp->range_array[i].finfo.start_page =
1127 page_to_pfn(pg);
1128 bl_resp->range_array[i].finfo.page_cnt = alloc_unit;
1129 bl_resp->hdr.size += sizeof(union dm_mem_page_range);
1130
1131 }
1132
1133 return num_pages;
1134}
1135
1136
1137
1138static void balloon_up(struct work_struct *dummy)
1139{
1140 unsigned int num_pages = dm_device.balloon_wrk.num_pages;
1141 unsigned int num_ballooned = 0;
1142 struct dm_balloon_response *bl_resp;
1143 int alloc_unit;
1144 int ret;
1145 bool done = false;
1146 int i;
1147 struct sysinfo val;
1148 unsigned long floor;
1149
1150 /* The host balloons pages in 2M granularity. */
1151 WARN_ON_ONCE(num_pages % PAGES_IN_2M != 0);
1152
1153 /*
1154 * We will attempt 2M allocations. However, if we fail to
1155 * allocate 2M chunks, we will go back to 4k allocations.
1156 */
1157 alloc_unit = 512;
1158
1159 si_meminfo(&val);
1160 floor = compute_balloon_floor();
1161
1162 /* Refuse to balloon below the floor, keep the 2M granularity. */
1163 if (val.freeram < num_pages || val.freeram - num_pages < floor) {
1164 num_pages = val.freeram > floor ? (val.freeram - floor) : 0;
1165 num_pages -= num_pages % PAGES_IN_2M;
1166 }
1167
1168 while (!done) {
1169 bl_resp = (struct dm_balloon_response *)send_buffer;
1170 memset(send_buffer, 0, PAGE_SIZE);
1171 bl_resp->hdr.type = DM_BALLOON_RESPONSE;
1172 bl_resp->hdr.size = sizeof(struct dm_balloon_response);
1173 bl_resp->more_pages = 1;
1174
1175
1176 num_pages -= num_ballooned;
1177 num_ballooned = alloc_balloon_pages(&dm_device, num_pages,
1178 bl_resp, alloc_unit);
1179
1180 if (alloc_unit != 1 && num_ballooned == 0) {
1181 alloc_unit = 1;
1182 continue;
1183 }
1184
1185 if (num_ballooned == 0 || num_ballooned == num_pages) {
1186 bl_resp->more_pages = 0;
1187 done = true;
1188 dm_device.state = DM_INITIALIZED;
1189 }
1190
1191 /*
1192 * We are pushing a lot of data through the channel;
1193 * deal with transient failures caused because of the
1194 * lack of space in the ring buffer.
1195 */
1196
1197 do {
1198 bl_resp->hdr.trans_id = atomic_inc_return(&trans_id);
1199 ret = vmbus_sendpacket(dm_device.dev->channel,
1200 bl_resp,
1201 bl_resp->hdr.size,
1202 (unsigned long)NULL,
1203 VM_PKT_DATA_INBAND, 0);
1204
1205 if (ret == -EAGAIN)
1206 msleep(20);
1207 post_status(&dm_device);
1208 } while (ret == -EAGAIN);
1209
1210 if (ret) {
1211 /*
1212 * Free up the memory we allocatted.
1213 */
1214 pr_info("Balloon response failed\n");
1215
1216 for (i = 0; i < bl_resp->range_count; i++)
1217 free_balloon_pages(&dm_device,
1218 &bl_resp->range_array[i]);
1219
1220 done = true;
1221 }
1222 }
1223
1224}
1225
1226static void balloon_down(struct hv_dynmem_device *dm,
1227 struct dm_unballoon_request *req)
1228{
1229 union dm_mem_page_range *range_array = req->range_array;
1230 int range_count = req->range_count;
1231 struct dm_unballoon_response resp;
1232 int i;
1233
1234 for (i = 0; i < range_count; i++) {
1235 free_balloon_pages(dm, &range_array[i]);
1236 complete(&dm_device.config_event);
1237 }
1238
1239 if (req->more_pages == 1)
1240 return;
1241
1242 memset(&resp, 0, sizeof(struct dm_unballoon_response));
1243 resp.hdr.type = DM_UNBALLOON_RESPONSE;
1244 resp.hdr.trans_id = atomic_inc_return(&trans_id);
1245 resp.hdr.size = sizeof(struct dm_unballoon_response);
1246
1247 vmbus_sendpacket(dm_device.dev->channel, &resp,
1248 sizeof(struct dm_unballoon_response),
1249 (unsigned long)NULL,
1250 VM_PKT_DATA_INBAND, 0);
1251
1252 dm->state = DM_INITIALIZED;
1253}
1254
1255static void balloon_onchannelcallback(void *context);
1256
1257static int dm_thread_func(void *dm_dev)
1258{
1259 struct hv_dynmem_device *dm = dm_dev;
1260
1261 while (!kthread_should_stop()) {
1262 wait_for_completion_interruptible_timeout(
1263 &dm_device.config_event, 1*HZ);
1264 /*
1265 * The host expects us to post information on the memory
1266 * pressure every second.
1267 */
1268 reinit_completion(&dm_device.config_event);
1269 post_status(dm);
1270 }
1271
1272 return 0;
1273}
1274
1275
1276static void version_resp(struct hv_dynmem_device *dm,
1277 struct dm_version_response *vresp)
1278{
1279 struct dm_version_request version_req;
1280 int ret;
1281
1282 if (vresp->is_accepted) {
1283 /*
1284 * We are done; wakeup the
1285 * context waiting for version
1286 * negotiation.
1287 */
1288 complete(&dm->host_event);
1289 return;
1290 }
1291 /*
1292 * If there are more versions to try, continue
1293 * with negotiations; if not
1294 * shutdown the service since we are not able
1295 * to negotiate a suitable version number
1296 * with the host.
1297 */
1298 if (dm->next_version == 0)
1299 goto version_error;
1300
1301 memset(&version_req, 0, sizeof(struct dm_version_request));
1302 version_req.hdr.type = DM_VERSION_REQUEST;
1303 version_req.hdr.size = sizeof(struct dm_version_request);
1304 version_req.hdr.trans_id = atomic_inc_return(&trans_id);
1305 version_req.version.version = dm->next_version;
1306
1307 /*
1308 * Set the next version to try in case current version fails.
1309 * Win7 protocol ought to be the last one to try.
1310 */
1311 switch (version_req.version.version) {
1312 case DYNMEM_PROTOCOL_VERSION_WIN8:
1313 dm->next_version = DYNMEM_PROTOCOL_VERSION_WIN7;
1314 version_req.is_last_attempt = 0;
1315 break;
1316 default:
1317 dm->next_version = 0;
1318 version_req.is_last_attempt = 1;
1319 }
1320
1321 ret = vmbus_sendpacket(dm->dev->channel, &version_req,
1322 sizeof(struct dm_version_request),
1323 (unsigned long)NULL,
1324 VM_PKT_DATA_INBAND, 0);
1325
1326 if (ret)
1327 goto version_error;
1328
1329 return;
1330
1331version_error:
1332 dm->state = DM_INIT_ERROR;
1333 complete(&dm->host_event);
1334}
1335
1336static void cap_resp(struct hv_dynmem_device *dm,
1337 struct dm_capabilities_resp_msg *cap_resp)
1338{
1339 if (!cap_resp->is_accepted) {
1340 pr_info("Capabilities not accepted by host\n");
1341 dm->state = DM_INIT_ERROR;
1342 }
1343 complete(&dm->host_event);
1344}
1345
1346static void balloon_onchannelcallback(void *context)
1347{
1348 struct hv_device *dev = context;
1349 u32 recvlen;
1350 u64 requestid;
1351 struct dm_message *dm_msg;
1352 struct dm_header *dm_hdr;
1353 struct hv_dynmem_device *dm = hv_get_drvdata(dev);
1354 struct dm_balloon *bal_msg;
1355 struct dm_hot_add *ha_msg;
1356 union dm_mem_page_range *ha_pg_range;
1357 union dm_mem_page_range *ha_region;
1358
1359 memset(recv_buffer, 0, sizeof(recv_buffer));
1360 vmbus_recvpacket(dev->channel, recv_buffer,
1361 PAGE_SIZE, &recvlen, &requestid);
1362
1363 if (recvlen > 0) {
1364 dm_msg = (struct dm_message *)recv_buffer;
1365 dm_hdr = &dm_msg->hdr;
1366
1367 switch (dm_hdr->type) {
1368 case DM_VERSION_RESPONSE:
1369 version_resp(dm,
1370 (struct dm_version_response *)dm_msg);
1371 break;
1372
1373 case DM_CAPABILITIES_RESPONSE:
1374 cap_resp(dm,
1375 (struct dm_capabilities_resp_msg *)dm_msg);
1376 break;
1377
1378 case DM_BALLOON_REQUEST:
1379 if (dm->state == DM_BALLOON_UP)
1380 pr_warn("Currently ballooning\n");
1381 bal_msg = (struct dm_balloon *)recv_buffer;
1382 dm->state = DM_BALLOON_UP;
1383 dm_device.balloon_wrk.num_pages = bal_msg->num_pages;
1384 schedule_work(&dm_device.balloon_wrk.wrk);
1385 break;
1386
1387 case DM_UNBALLOON_REQUEST:
1388 dm->state = DM_BALLOON_DOWN;
1389 balloon_down(dm,
1390 (struct dm_unballoon_request *)recv_buffer);
1391 break;
1392
1393 case DM_MEM_HOT_ADD_REQUEST:
1394 if (dm->state == DM_HOT_ADD)
1395 pr_warn("Currently hot-adding\n");
1396 dm->state = DM_HOT_ADD;
1397 ha_msg = (struct dm_hot_add *)recv_buffer;
1398 if (ha_msg->hdr.size == sizeof(struct dm_hot_add)) {
1399 /*
1400 * This is a normal hot-add request specifying
1401 * hot-add memory.
1402 */
1403 ha_pg_range = &ha_msg->range;
1404 dm->ha_wrk.ha_page_range = *ha_pg_range;
1405 dm->ha_wrk.ha_region_range.page_range = 0;
1406 } else {
1407 /*
1408 * Host is specifying that we first hot-add
1409 * a region and then partially populate this
1410 * region.
1411 */
1412 dm->host_specified_ha_region = true;
1413 ha_pg_range = &ha_msg->range;
1414 ha_region = &ha_pg_range[1];
1415 dm->ha_wrk.ha_page_range = *ha_pg_range;
1416 dm->ha_wrk.ha_region_range = *ha_region;
1417 }
1418 schedule_work(&dm_device.ha_wrk.wrk);
1419 break;
1420
1421 case DM_INFO_MESSAGE:
1422 process_info(dm, (struct dm_info_msg *)dm_msg);
1423 break;
1424
1425 default:
1426 pr_err("Unhandled message: type: %d\n", dm_hdr->type);
1427
1428 }
1429 }
1430
1431}
1432
1433static int balloon_probe(struct hv_device *dev,
1434 const struct hv_vmbus_device_id *dev_id)
1435{
1436 int ret;
1437 unsigned long t;
1438 struct dm_version_request version_req;
1439 struct dm_capabilities cap_msg;
1440
1441 do_hot_add = hot_add;
1442
1443 /*
1444 * First allocate a send buffer.
1445 */
1446
1447 send_buffer = kmalloc(PAGE_SIZE, GFP_KERNEL);
1448 if (!send_buffer)
1449 return -ENOMEM;
1450
1451 ret = vmbus_open(dev->channel, dm_ring_size, dm_ring_size, NULL, 0,
1452 balloon_onchannelcallback, dev);
1453
1454 if (ret)
1455 goto probe_error0;
1456
1457 dm_device.dev = dev;
1458 dm_device.state = DM_INITIALIZING;
1459 dm_device.next_version = DYNMEM_PROTOCOL_VERSION_WIN8;
1460 init_completion(&dm_device.host_event);
1461 init_completion(&dm_device.config_event);
1462 INIT_LIST_HEAD(&dm_device.ha_region_list);
1463 mutex_init(&dm_device.ha_region_mutex);
1464 INIT_WORK(&dm_device.balloon_wrk.wrk, balloon_up);
1465 INIT_WORK(&dm_device.ha_wrk.wrk, hot_add_req);
1466 dm_device.host_specified_ha_region = false;
1467
1468 dm_device.thread =
1469 kthread_run(dm_thread_func, &dm_device, "hv_balloon");
1470 if (IS_ERR(dm_device.thread)) {
1471 ret = PTR_ERR(dm_device.thread);
1472 goto probe_error1;
1473 }
1474
1475#ifdef CONFIG_MEMORY_HOTPLUG
1476 set_online_page_callback(&hv_online_page);
1477 register_memory_notifier(&hv_memory_nb);
1478#endif
1479
1480 hv_set_drvdata(dev, &dm_device);
1481 /*
1482 * Initiate the hand shake with the host and negotiate
1483 * a version that the host can support. We start with the
1484 * highest version number and go down if the host cannot
1485 * support it.
1486 */
1487 memset(&version_req, 0, sizeof(struct dm_version_request));
1488 version_req.hdr.type = DM_VERSION_REQUEST;
1489 version_req.hdr.size = sizeof(struct dm_version_request);
1490 version_req.hdr.trans_id = atomic_inc_return(&trans_id);
1491 version_req.version.version = DYNMEM_PROTOCOL_VERSION_WIN10;
1492 version_req.is_last_attempt = 0;
1493
1494 ret = vmbus_sendpacket(dev->channel, &version_req,
1495 sizeof(struct dm_version_request),
1496 (unsigned long)NULL,
1497 VM_PKT_DATA_INBAND, 0);
1498 if (ret)
1499 goto probe_error2;
1500
1501 t = wait_for_completion_timeout(&dm_device.host_event, 5*HZ);
1502 if (t == 0) {
1503 ret = -ETIMEDOUT;
1504 goto probe_error2;
1505 }
1506
1507 /*
1508 * If we could not negotiate a compatible version with the host
1509 * fail the probe function.
1510 */
1511 if (dm_device.state == DM_INIT_ERROR) {
1512 ret = -ETIMEDOUT;
1513 goto probe_error2;
1514 }
1515 /*
1516 * Now submit our capabilities to the host.
1517 */
1518 memset(&cap_msg, 0, sizeof(struct dm_capabilities));
1519 cap_msg.hdr.type = DM_CAPABILITIES_REPORT;
1520 cap_msg.hdr.size = sizeof(struct dm_capabilities);
1521 cap_msg.hdr.trans_id = atomic_inc_return(&trans_id);
1522
1523 cap_msg.caps.cap_bits.balloon = 1;
1524 cap_msg.caps.cap_bits.hot_add = 1;
1525
1526 /*
1527 * Specify our alignment requirements as it relates
1528 * memory hot-add. Specify 128MB alignment.
1529 */
1530 cap_msg.caps.cap_bits.hot_add_alignment = 7;
1531
1532 /*
1533 * Currently the host does not use these
1534 * values and we set them to what is done in the
1535 * Windows driver.
1536 */
1537 cap_msg.min_page_cnt = 0;
1538 cap_msg.max_page_number = -1;
1539
1540 ret = vmbus_sendpacket(dev->channel, &cap_msg,
1541 sizeof(struct dm_capabilities),
1542 (unsigned long)NULL,
1543 VM_PKT_DATA_INBAND, 0);
1544 if (ret)
1545 goto probe_error2;
1546
1547 t = wait_for_completion_timeout(&dm_device.host_event, 5*HZ);
1548 if (t == 0) {
1549 ret = -ETIMEDOUT;
1550 goto probe_error2;
1551 }
1552
1553 /*
1554 * If the host does not like our capabilities,
1555 * fail the probe function.
1556 */
1557 if (dm_device.state == DM_INIT_ERROR) {
1558 ret = -ETIMEDOUT;
1559 goto probe_error2;
1560 }
1561
1562 dm_device.state = DM_INITIALIZED;
1563
1564 return 0;
1565
1566probe_error2:
1567#ifdef CONFIG_MEMORY_HOTPLUG
1568 restore_online_page_callback(&hv_online_page);
1569#endif
1570 kthread_stop(dm_device.thread);
1571
1572probe_error1:
1573 vmbus_close(dev->channel);
1574probe_error0:
1575 kfree(send_buffer);
1576 return ret;
1577}
1578
1579static int balloon_remove(struct hv_device *dev)
1580{
1581 struct hv_dynmem_device *dm = hv_get_drvdata(dev);
1582 struct list_head *cur, *tmp;
1583 struct hv_hotadd_state *has;
1584
1585 if (dm->num_pages_ballooned != 0)
1586 pr_warn("Ballooned pages: %d\n", dm->num_pages_ballooned);
1587
1588 cancel_work_sync(&dm->balloon_wrk.wrk);
1589 cancel_work_sync(&dm->ha_wrk.wrk);
1590
1591 vmbus_close(dev->channel);
1592 kthread_stop(dm->thread);
1593 kfree(send_buffer);
1594#ifdef CONFIG_MEMORY_HOTPLUG
1595 restore_online_page_callback(&hv_online_page);
1596 unregister_memory_notifier(&hv_memory_nb);
1597#endif
1598 list_for_each_safe(cur, tmp, &dm->ha_region_list) {
1599 has = list_entry(cur, struct hv_hotadd_state, list);
1600 list_del(&has->list);
1601 kfree(has);
1602 }
1603
1604 return 0;
1605}
1606
1607static const struct hv_vmbus_device_id id_table[] = {
1608 /* Dynamic Memory Class ID */
1609 /* 525074DC-8985-46e2-8057-A307DC18A502 */
1610 { HV_DM_GUID, },
1611 { },
1612};
1613
1614MODULE_DEVICE_TABLE(vmbus, id_table);
1615
1616static struct hv_driver balloon_drv = {
1617 .name = "hv_balloon",
1618 .id_table = id_table,
1619 .probe = balloon_probe,
1620 .remove = balloon_remove,
1621};
1622
1623static int __init init_balloon_drv(void)
1624{
1625
1626 return vmbus_driver_register(&balloon_drv);
1627}
1628
1629module_init(init_balloon_drv);
1630
1631MODULE_DESCRIPTION("Hyper-V Balloon");
1632MODULE_LICENSE("GPL");
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#define CREATE_TRACE_POINTS
38#include "hv_trace_balloon.h"
39
40/*
41 * We begin with definitions supporting the Dynamic Memory protocol
42 * with the host.
43 *
44 * Begin protocol definitions.
45 */
46
47
48
49/*
50 * Protocol versions. The low word is the minor version, the high word the major
51 * version.
52 *
53 * History:
54 * Initial version 1.0
55 * Changed to 0.1 on 2009/03/25
56 * Changes to 0.2 on 2009/05/14
57 * Changes to 0.3 on 2009/12/03
58 * Changed to 1.0 on 2011/04/05
59 */
60
61#define DYNMEM_MAKE_VERSION(Major, Minor) ((__u32)(((Major) << 16) | (Minor)))
62#define DYNMEM_MAJOR_VERSION(Version) ((__u32)(Version) >> 16)
63#define DYNMEM_MINOR_VERSION(Version) ((__u32)(Version) & 0xff)
64
65enum {
66 DYNMEM_PROTOCOL_VERSION_1 = DYNMEM_MAKE_VERSION(0, 3),
67 DYNMEM_PROTOCOL_VERSION_2 = DYNMEM_MAKE_VERSION(1, 0),
68 DYNMEM_PROTOCOL_VERSION_3 = DYNMEM_MAKE_VERSION(2, 0),
69
70 DYNMEM_PROTOCOL_VERSION_WIN7 = DYNMEM_PROTOCOL_VERSION_1,
71 DYNMEM_PROTOCOL_VERSION_WIN8 = DYNMEM_PROTOCOL_VERSION_2,
72 DYNMEM_PROTOCOL_VERSION_WIN10 = DYNMEM_PROTOCOL_VERSION_3,
73
74 DYNMEM_PROTOCOL_VERSION_CURRENT = DYNMEM_PROTOCOL_VERSION_WIN10
75};
76
77
78
79/*
80 * Message Types
81 */
82
83enum dm_message_type {
84 /*
85 * Version 0.3
86 */
87 DM_ERROR = 0,
88 DM_VERSION_REQUEST = 1,
89 DM_VERSION_RESPONSE = 2,
90 DM_CAPABILITIES_REPORT = 3,
91 DM_CAPABILITIES_RESPONSE = 4,
92 DM_STATUS_REPORT = 5,
93 DM_BALLOON_REQUEST = 6,
94 DM_BALLOON_RESPONSE = 7,
95 DM_UNBALLOON_REQUEST = 8,
96 DM_UNBALLOON_RESPONSE = 9,
97 DM_MEM_HOT_ADD_REQUEST = 10,
98 DM_MEM_HOT_ADD_RESPONSE = 11,
99 DM_VERSION_03_MAX = 11,
100 /*
101 * Version 1.0.
102 */
103 DM_INFO_MESSAGE = 12,
104 DM_VERSION_1_MAX = 12
105};
106
107
108/*
109 * Structures defining the dynamic memory management
110 * protocol.
111 */
112
113union dm_version {
114 struct {
115 __u16 minor_version;
116 __u16 major_version;
117 };
118 __u32 version;
119} __packed;
120
121
122union dm_caps {
123 struct {
124 __u64 balloon:1;
125 __u64 hot_add:1;
126 /*
127 * To support guests that may have alignment
128 * limitations on hot-add, the guest can specify
129 * its alignment requirements; a value of n
130 * represents an alignment of 2^n in mega bytes.
131 */
132 __u64 hot_add_alignment:4;
133 __u64 reservedz:58;
134 } cap_bits;
135 __u64 caps;
136} __packed;
137
138union dm_mem_page_range {
139 struct {
140 /*
141 * The PFN number of the first page in the range.
142 * 40 bits is the architectural limit of a PFN
143 * number for AMD64.
144 */
145 __u64 start_page:40;
146 /*
147 * The number of pages in the range.
148 */
149 __u64 page_cnt:24;
150 } finfo;
151 __u64 page_range;
152} __packed;
153
154
155
156/*
157 * The header for all dynamic memory messages:
158 *
159 * type: Type of the message.
160 * size: Size of the message in bytes; including the header.
161 * trans_id: The guest is responsible for manufacturing this ID.
162 */
163
164struct dm_header {
165 __u16 type;
166 __u16 size;
167 __u32 trans_id;
168} __packed;
169
170/*
171 * A generic message format for dynamic memory.
172 * Specific message formats are defined later in the file.
173 */
174
175struct dm_message {
176 struct dm_header hdr;
177 __u8 data[]; /* enclosed message */
178} __packed;
179
180
181/*
182 * Specific message types supporting the dynamic memory protocol.
183 */
184
185/*
186 * Version negotiation message. Sent from the guest to the host.
187 * The guest is free to try different versions until the host
188 * accepts the version.
189 *
190 * dm_version: The protocol version requested.
191 * is_last_attempt: If TRUE, this is the last version guest will request.
192 * reservedz: Reserved field, set to zero.
193 */
194
195struct dm_version_request {
196 struct dm_header hdr;
197 union dm_version version;
198 __u32 is_last_attempt:1;
199 __u32 reservedz:31;
200} __packed;
201
202/*
203 * Version response message; Host to Guest and indicates
204 * if the host has accepted the version sent by the guest.
205 *
206 * is_accepted: If TRUE, host has accepted the version and the guest
207 * should proceed to the next stage of the protocol. FALSE indicates that
208 * guest should re-try with a different version.
209 *
210 * reservedz: Reserved field, set to zero.
211 */
212
213struct dm_version_response {
214 struct dm_header hdr;
215 __u64 is_accepted:1;
216 __u64 reservedz:63;
217} __packed;
218
219/*
220 * Message reporting capabilities. This is sent from the guest to the
221 * host.
222 */
223
224struct dm_capabilities {
225 struct dm_header hdr;
226 union dm_caps caps;
227 __u64 min_page_cnt;
228 __u64 max_page_number;
229} __packed;
230
231/*
232 * Response to the capabilities message. This is sent from the host to the
233 * guest. This message notifies if the host has accepted the guest's
234 * capabilities. If the host has not accepted, the guest must shutdown
235 * the service.
236 *
237 * is_accepted: Indicates if the host has accepted guest's capabilities.
238 * reservedz: Must be 0.
239 */
240
241struct dm_capabilities_resp_msg {
242 struct dm_header hdr;
243 __u64 is_accepted:1;
244 __u64 reservedz:63;
245} __packed;
246
247/*
248 * This message is used to report memory pressure from the guest.
249 * This message is not part of any transaction and there is no
250 * response to this message.
251 *
252 * num_avail: Available memory in pages.
253 * num_committed: Committed memory in pages.
254 * page_file_size: The accumulated size of all page files
255 * in the system in pages.
256 * zero_free: The nunber of zero and free pages.
257 * page_file_writes: The writes to the page file in pages.
258 * io_diff: An indicator of file cache efficiency or page file activity,
259 * calculated as File Cache Page Fault Count - Page Read Count.
260 * This value is in pages.
261 *
262 * Some of these metrics are Windows specific and fortunately
263 * the algorithm on the host side that computes the guest memory
264 * pressure only uses num_committed value.
265 */
266
267struct dm_status {
268 struct dm_header hdr;
269 __u64 num_avail;
270 __u64 num_committed;
271 __u64 page_file_size;
272 __u64 zero_free;
273 __u32 page_file_writes;
274 __u32 io_diff;
275} __packed;
276
277
278/*
279 * Message to ask the guest to allocate memory - balloon up message.
280 * This message is sent from the host to the guest. The guest may not be
281 * able to allocate as much memory as requested.
282 *
283 * num_pages: number of pages to allocate.
284 */
285
286struct dm_balloon {
287 struct dm_header hdr;
288 __u32 num_pages;
289 __u32 reservedz;
290} __packed;
291
292
293/*
294 * Balloon response message; this message is sent from the guest
295 * to the host in response to the balloon message.
296 *
297 * reservedz: Reserved; must be set to zero.
298 * more_pages: If FALSE, this is the last message of the transaction.
299 * if TRUE there will atleast one more message from the guest.
300 *
301 * range_count: The number of ranges in the range array.
302 *
303 * range_array: An array of page ranges returned to the host.
304 *
305 */
306
307struct dm_balloon_response {
308 struct dm_header hdr;
309 __u32 reservedz;
310 __u32 more_pages:1;
311 __u32 range_count:31;
312 union dm_mem_page_range range_array[];
313} __packed;
314
315/*
316 * Un-balloon message; this message is sent from the host
317 * to the guest to give guest more memory.
318 *
319 * more_pages: If FALSE, this is the last message of the transaction.
320 * if TRUE there will atleast one more message from the guest.
321 *
322 * reservedz: Reserved; must be set to zero.
323 *
324 * range_count: The number of ranges in the range array.
325 *
326 * range_array: An array of page ranges returned to the host.
327 *
328 */
329
330struct dm_unballoon_request {
331 struct dm_header hdr;
332 __u32 more_pages:1;
333 __u32 reservedz:31;
334 __u32 range_count;
335 union dm_mem_page_range range_array[];
336} __packed;
337
338/*
339 * Un-balloon response message; this message is sent from the guest
340 * to the host in response to an unballoon request.
341 *
342 */
343
344struct dm_unballoon_response {
345 struct dm_header hdr;
346} __packed;
347
348
349/*
350 * Hot add request message. Message sent from the host to the guest.
351 *
352 * mem_range: Memory range to hot add.
353 *
354 * On Linux we currently don't support this since we cannot hot add
355 * arbitrary granularity of memory.
356 */
357
358struct dm_hot_add {
359 struct dm_header hdr;
360 union dm_mem_page_range range;
361} __packed;
362
363/*
364 * Hot add response message.
365 * This message is sent by the guest to report the status of a hot add request.
366 * If page_count is less than the requested page count, then the host should
367 * assume all further hot add requests will fail, since this indicates that
368 * the guest has hit an upper physical memory barrier.
369 *
370 * Hot adds may also fail due to low resources; in this case, the guest must
371 * not complete this message until the hot add can succeed, and the host must
372 * not send a new hot add request until the response is sent.
373 * If VSC fails to hot add memory DYNMEM_NUMBER_OF_UNSUCCESSFUL_HOTADD_ATTEMPTS
374 * times it fails the request.
375 *
376 *
377 * page_count: number of pages that were successfully hot added.
378 *
379 * result: result of the operation 1: success, 0: failure.
380 *
381 */
382
383struct dm_hot_add_response {
384 struct dm_header hdr;
385 __u32 page_count;
386 __u32 result;
387} __packed;
388
389/*
390 * Types of information sent from host to the guest.
391 */
392
393enum dm_info_type {
394 INFO_TYPE_MAX_PAGE_CNT = 0,
395 MAX_INFO_TYPE
396};
397
398
399/*
400 * Header for the information message.
401 */
402
403struct dm_info_header {
404 enum dm_info_type type;
405 __u32 data_size;
406} __packed;
407
408/*
409 * This message is sent from the host to the guest to pass
410 * some relevant information (win8 addition).
411 *
412 * reserved: no used.
413 * info_size: size of the information blob.
414 * info: information blob.
415 */
416
417struct dm_info_msg {
418 struct dm_header hdr;
419 __u32 reserved;
420 __u32 info_size;
421 __u8 info[];
422};
423
424/*
425 * End protocol definitions.
426 */
427
428/*
429 * State to manage hot adding memory into the guest.
430 * The range start_pfn : end_pfn specifies the range
431 * that the host has asked us to hot add. The range
432 * start_pfn : ha_end_pfn specifies the range that we have
433 * currently hot added. We hot add in multiples of 128M
434 * chunks; it is possible that we may not be able to bring
435 * online all the pages in the region. The range
436 * covered_start_pfn:covered_end_pfn defines the pages that can
437 * be brough online.
438 */
439
440struct hv_hotadd_state {
441 struct list_head list;
442 unsigned long start_pfn;
443 unsigned long covered_start_pfn;
444 unsigned long covered_end_pfn;
445 unsigned long ha_end_pfn;
446 unsigned long end_pfn;
447 /*
448 * A list of gaps.
449 */
450 struct list_head gap_list;
451};
452
453struct hv_hotadd_gap {
454 struct list_head list;
455 unsigned long start_pfn;
456 unsigned long end_pfn;
457};
458
459struct balloon_state {
460 __u32 num_pages;
461 struct work_struct wrk;
462};
463
464struct hot_add_wrk {
465 union dm_mem_page_range ha_page_range;
466 union dm_mem_page_range ha_region_range;
467 struct work_struct wrk;
468};
469
470static bool hot_add = true;
471static bool do_hot_add;
472/*
473 * Delay reporting memory pressure by
474 * the specified number of seconds.
475 */
476static uint pressure_report_delay = 45;
477
478/*
479 * The last time we posted a pressure report to host.
480 */
481static unsigned long last_post_time;
482
483module_param(hot_add, bool, (S_IRUGO | S_IWUSR));
484MODULE_PARM_DESC(hot_add, "If set attempt memory hot_add");
485
486module_param(pressure_report_delay, uint, (S_IRUGO | S_IWUSR));
487MODULE_PARM_DESC(pressure_report_delay, "Delay in secs in reporting pressure");
488static atomic_t trans_id = ATOMIC_INIT(0);
489
490static int dm_ring_size = (5 * PAGE_SIZE);
491
492/*
493 * Driver specific state.
494 */
495
496enum hv_dm_state {
497 DM_INITIALIZING = 0,
498 DM_INITIALIZED,
499 DM_BALLOON_UP,
500 DM_BALLOON_DOWN,
501 DM_HOT_ADD,
502 DM_INIT_ERROR
503};
504
505
506static __u8 recv_buffer[PAGE_SIZE];
507static __u8 *send_buffer;
508#define PAGES_IN_2M 512
509#define HA_CHUNK (32 * 1024)
510
511struct hv_dynmem_device {
512 struct hv_device *dev;
513 enum hv_dm_state state;
514 struct completion host_event;
515 struct completion config_event;
516
517 /*
518 * Number of pages we have currently ballooned out.
519 */
520 unsigned int num_pages_ballooned;
521 unsigned int num_pages_onlined;
522 unsigned int num_pages_added;
523
524 /*
525 * State to manage the ballooning (up) operation.
526 */
527 struct balloon_state balloon_wrk;
528
529 /*
530 * State to execute the "hot-add" operation.
531 */
532 struct hot_add_wrk ha_wrk;
533
534 /*
535 * This state tracks if the host has specified a hot-add
536 * region.
537 */
538 bool host_specified_ha_region;
539
540 /*
541 * State to synchronize hot-add.
542 */
543 struct completion ol_waitevent;
544 bool ha_waiting;
545 /*
546 * This thread handles hot-add
547 * requests from the host as well as notifying
548 * the host with regards to memory pressure in
549 * the guest.
550 */
551 struct task_struct *thread;
552
553 /*
554 * Protects ha_region_list, num_pages_onlined counter and individual
555 * regions from ha_region_list.
556 */
557 spinlock_t ha_lock;
558
559 /*
560 * A list of hot-add regions.
561 */
562 struct list_head ha_region_list;
563
564 /*
565 * We start with the highest version we can support
566 * and downgrade based on the host; we save here the
567 * next version to try.
568 */
569 __u32 next_version;
570
571 /*
572 * The negotiated version agreed by host.
573 */
574 __u32 version;
575};
576
577static struct hv_dynmem_device dm_device;
578
579static void post_status(struct hv_dynmem_device *dm);
580
581#ifdef CONFIG_MEMORY_HOTPLUG
582static inline bool has_pfn_is_backed(struct hv_hotadd_state *has,
583 unsigned long pfn)
584{
585 struct hv_hotadd_gap *gap;
586
587 /* The page is not backed. */
588 if ((pfn < has->covered_start_pfn) || (pfn >= has->covered_end_pfn))
589 return false;
590
591 /* Check for gaps. */
592 list_for_each_entry(gap, &has->gap_list, list) {
593 if ((pfn >= gap->start_pfn) && (pfn < gap->end_pfn))
594 return false;
595 }
596
597 return true;
598}
599
600static unsigned long hv_page_offline_check(unsigned long start_pfn,
601 unsigned long nr_pages)
602{
603 unsigned long pfn = start_pfn, count = 0;
604 struct hv_hotadd_state *has;
605 bool found;
606
607 while (pfn < start_pfn + nr_pages) {
608 /*
609 * Search for HAS which covers the pfn and when we find one
610 * count how many consequitive PFNs are covered.
611 */
612 found = false;
613 list_for_each_entry(has, &dm_device.ha_region_list, list) {
614 while ((pfn >= has->start_pfn) &&
615 (pfn < has->end_pfn) &&
616 (pfn < start_pfn + nr_pages)) {
617 found = true;
618 if (has_pfn_is_backed(has, pfn))
619 count++;
620 pfn++;
621 }
622 }
623
624 /*
625 * This PFN is not in any HAS (e.g. we're offlining a region
626 * which was present at boot), no need to account for it. Go
627 * to the next one.
628 */
629 if (!found)
630 pfn++;
631 }
632
633 return count;
634}
635
636static int hv_memory_notifier(struct notifier_block *nb, unsigned long val,
637 void *v)
638{
639 struct memory_notify *mem = (struct memory_notify *)v;
640 unsigned long flags, pfn_count;
641
642 switch (val) {
643 case MEM_ONLINE:
644 case MEM_CANCEL_ONLINE:
645 if (dm_device.ha_waiting) {
646 dm_device.ha_waiting = false;
647 complete(&dm_device.ol_waitevent);
648 }
649 break;
650
651 case MEM_OFFLINE:
652 spin_lock_irqsave(&dm_device.ha_lock, flags);
653 pfn_count = hv_page_offline_check(mem->start_pfn,
654 mem->nr_pages);
655 if (pfn_count <= dm_device.num_pages_onlined) {
656 dm_device.num_pages_onlined -= pfn_count;
657 } else {
658 /*
659 * We're offlining more pages than we managed to online.
660 * This is unexpected. In any case don't let
661 * num_pages_onlined wrap around zero.
662 */
663 WARN_ON_ONCE(1);
664 dm_device.num_pages_onlined = 0;
665 }
666 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
667 break;
668 case MEM_GOING_ONLINE:
669 case MEM_GOING_OFFLINE:
670 case MEM_CANCEL_OFFLINE:
671 break;
672 }
673 return NOTIFY_OK;
674}
675
676static struct notifier_block hv_memory_nb = {
677 .notifier_call = hv_memory_notifier,
678 .priority = 0
679};
680
681/* Check if the particular page is backed and can be onlined and online it. */
682static void hv_page_online_one(struct hv_hotadd_state *has, struct page *pg)
683{
684 if (!has_pfn_is_backed(has, page_to_pfn(pg)))
685 return;
686
687 /* This frame is currently backed; online the page. */
688 __online_page_set_limits(pg);
689 __online_page_increment_counters(pg);
690 __online_page_free(pg);
691
692 WARN_ON_ONCE(!spin_is_locked(&dm_device.ha_lock));
693 dm_device.num_pages_onlined++;
694}
695
696static void hv_bring_pgs_online(struct hv_hotadd_state *has,
697 unsigned long start_pfn, unsigned long size)
698{
699 int i;
700
701 pr_debug("Online %lu pages starting at pfn 0x%lx\n", size, start_pfn);
702 for (i = 0; i < size; i++)
703 hv_page_online_one(has, pfn_to_page(start_pfn + i));
704}
705
706static void hv_mem_hot_add(unsigned long start, unsigned long size,
707 unsigned long pfn_count,
708 struct hv_hotadd_state *has)
709{
710 int ret = 0;
711 int i, nid;
712 unsigned long start_pfn;
713 unsigned long processed_pfn;
714 unsigned long total_pfn = pfn_count;
715 unsigned long flags;
716
717 for (i = 0; i < (size/HA_CHUNK); i++) {
718 start_pfn = start + (i * HA_CHUNK);
719
720 spin_lock_irqsave(&dm_device.ha_lock, flags);
721 has->ha_end_pfn += HA_CHUNK;
722
723 if (total_pfn > HA_CHUNK) {
724 processed_pfn = HA_CHUNK;
725 total_pfn -= HA_CHUNK;
726 } else {
727 processed_pfn = total_pfn;
728 total_pfn = 0;
729 }
730
731 has->covered_end_pfn += processed_pfn;
732 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
733
734 init_completion(&dm_device.ol_waitevent);
735 dm_device.ha_waiting = !memhp_auto_online;
736
737 nid = memory_add_physaddr_to_nid(PFN_PHYS(start_pfn));
738 ret = add_memory(nid, PFN_PHYS((start_pfn)),
739 (HA_CHUNK << PAGE_SHIFT));
740
741 if (ret) {
742 pr_err("hot_add memory failed error is %d\n", ret);
743 if (ret == -EEXIST) {
744 /*
745 * This error indicates that the error
746 * is not a transient failure. This is the
747 * case where the guest's physical address map
748 * precludes hot adding memory. Stop all further
749 * memory hot-add.
750 */
751 do_hot_add = false;
752 }
753 spin_lock_irqsave(&dm_device.ha_lock, flags);
754 has->ha_end_pfn -= HA_CHUNK;
755 has->covered_end_pfn -= processed_pfn;
756 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
757 break;
758 }
759
760 /*
761 * Wait for the memory block to be onlined when memory onlining
762 * is done outside of kernel (memhp_auto_online). Since the hot
763 * add has succeeded, it is ok to proceed even if the pages in
764 * the hot added region have not been "onlined" within the
765 * allowed time.
766 */
767 if (dm_device.ha_waiting)
768 wait_for_completion_timeout(&dm_device.ol_waitevent,
769 5*HZ);
770 post_status(&dm_device);
771 }
772}
773
774static void hv_online_page(struct page *pg)
775{
776 struct hv_hotadd_state *has;
777 unsigned long flags;
778 unsigned long pfn = page_to_pfn(pg);
779
780 spin_lock_irqsave(&dm_device.ha_lock, flags);
781 list_for_each_entry(has, &dm_device.ha_region_list, list) {
782 /* The page belongs to a different HAS. */
783 if ((pfn < has->start_pfn) || (pfn >= has->end_pfn))
784 continue;
785
786 hv_page_online_one(has, pg);
787 break;
788 }
789 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
790}
791
792static int pfn_covered(unsigned long start_pfn, unsigned long pfn_cnt)
793{
794 struct hv_hotadd_state *has;
795 struct hv_hotadd_gap *gap;
796 unsigned long residual, new_inc;
797 int ret = 0;
798 unsigned long flags;
799
800 spin_lock_irqsave(&dm_device.ha_lock, flags);
801 list_for_each_entry(has, &dm_device.ha_region_list, list) {
802 /*
803 * If the pfn range we are dealing with is not in the current
804 * "hot add block", move on.
805 */
806 if (start_pfn < has->start_pfn || start_pfn >= has->end_pfn)
807 continue;
808
809 /*
810 * If the current start pfn is not where the covered_end
811 * is, create a gap and update covered_end_pfn.
812 */
813 if (has->covered_end_pfn != start_pfn) {
814 gap = kzalloc(sizeof(struct hv_hotadd_gap), GFP_ATOMIC);
815 if (!gap) {
816 ret = -ENOMEM;
817 break;
818 }
819
820 INIT_LIST_HEAD(&gap->list);
821 gap->start_pfn = has->covered_end_pfn;
822 gap->end_pfn = start_pfn;
823 list_add_tail(&gap->list, &has->gap_list);
824
825 has->covered_end_pfn = start_pfn;
826 }
827
828 /*
829 * If the current hot add-request extends beyond
830 * our current limit; extend it.
831 */
832 if ((start_pfn + pfn_cnt) > has->end_pfn) {
833 residual = (start_pfn + pfn_cnt - has->end_pfn);
834 /*
835 * Extend the region by multiples of HA_CHUNK.
836 */
837 new_inc = (residual / HA_CHUNK) * HA_CHUNK;
838 if (residual % HA_CHUNK)
839 new_inc += HA_CHUNK;
840
841 has->end_pfn += new_inc;
842 }
843
844 ret = 1;
845 break;
846 }
847 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
848
849 return ret;
850}
851
852static unsigned long handle_pg_range(unsigned long pg_start,
853 unsigned long pg_count)
854{
855 unsigned long start_pfn = pg_start;
856 unsigned long pfn_cnt = pg_count;
857 unsigned long size;
858 struct hv_hotadd_state *has;
859 unsigned long pgs_ol = 0;
860 unsigned long old_covered_state;
861 unsigned long res = 0, flags;
862
863 pr_debug("Hot adding %lu pages starting at pfn 0x%lx.\n", pg_count,
864 pg_start);
865
866 spin_lock_irqsave(&dm_device.ha_lock, flags);
867 list_for_each_entry(has, &dm_device.ha_region_list, list) {
868 /*
869 * If the pfn range we are dealing with is not in the current
870 * "hot add block", move on.
871 */
872 if (start_pfn < has->start_pfn || start_pfn >= has->end_pfn)
873 continue;
874
875 old_covered_state = has->covered_end_pfn;
876
877 if (start_pfn < has->ha_end_pfn) {
878 /*
879 * This is the case where we are backing pages
880 * in an already hot added region. Bring
881 * these pages online first.
882 */
883 pgs_ol = has->ha_end_pfn - start_pfn;
884 if (pgs_ol > pfn_cnt)
885 pgs_ol = pfn_cnt;
886
887 has->covered_end_pfn += pgs_ol;
888 pfn_cnt -= pgs_ol;
889 /*
890 * Check if the corresponding memory block is already
891 * online by checking its last previously backed page.
892 * In case it is we need to bring rest (which was not
893 * backed previously) online too.
894 */
895 if (start_pfn > has->start_pfn &&
896 !PageReserved(pfn_to_page(start_pfn - 1)))
897 hv_bring_pgs_online(has, start_pfn, pgs_ol);
898
899 }
900
901 if ((has->ha_end_pfn < has->end_pfn) && (pfn_cnt > 0)) {
902 /*
903 * We have some residual hot add range
904 * that needs to be hot added; hot add
905 * it now. Hot add a multiple of
906 * of HA_CHUNK that fully covers the pages
907 * we have.
908 */
909 size = (has->end_pfn - has->ha_end_pfn);
910 if (pfn_cnt <= size) {
911 size = ((pfn_cnt / HA_CHUNK) * HA_CHUNK);
912 if (pfn_cnt % HA_CHUNK)
913 size += HA_CHUNK;
914 } else {
915 pfn_cnt = size;
916 }
917 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
918 hv_mem_hot_add(has->ha_end_pfn, size, pfn_cnt, has);
919 spin_lock_irqsave(&dm_device.ha_lock, flags);
920 }
921 /*
922 * If we managed to online any pages that were given to us,
923 * we declare success.
924 */
925 res = has->covered_end_pfn - old_covered_state;
926 break;
927 }
928 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
929
930 return res;
931}
932
933static unsigned long process_hot_add(unsigned long pg_start,
934 unsigned long pfn_cnt,
935 unsigned long rg_start,
936 unsigned long rg_size)
937{
938 struct hv_hotadd_state *ha_region = NULL;
939 int covered;
940 unsigned long flags;
941
942 if (pfn_cnt == 0)
943 return 0;
944
945 if (!dm_device.host_specified_ha_region) {
946 covered = pfn_covered(pg_start, pfn_cnt);
947 if (covered < 0)
948 return 0;
949
950 if (covered)
951 goto do_pg_range;
952 }
953
954 /*
955 * If the host has specified a hot-add range; deal with it first.
956 */
957
958 if (rg_size != 0) {
959 ha_region = kzalloc(sizeof(struct hv_hotadd_state), GFP_KERNEL);
960 if (!ha_region)
961 return 0;
962
963 INIT_LIST_HEAD(&ha_region->list);
964 INIT_LIST_HEAD(&ha_region->gap_list);
965
966 ha_region->start_pfn = rg_start;
967 ha_region->ha_end_pfn = rg_start;
968 ha_region->covered_start_pfn = pg_start;
969 ha_region->covered_end_pfn = pg_start;
970 ha_region->end_pfn = rg_start + rg_size;
971
972 spin_lock_irqsave(&dm_device.ha_lock, flags);
973 list_add_tail(&ha_region->list, &dm_device.ha_region_list);
974 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
975 }
976
977do_pg_range:
978 /*
979 * Process the page range specified; bringing them
980 * online if possible.
981 */
982 return handle_pg_range(pg_start, pfn_cnt);
983}
984
985#endif
986
987static void hot_add_req(struct work_struct *dummy)
988{
989 struct dm_hot_add_response resp;
990#ifdef CONFIG_MEMORY_HOTPLUG
991 unsigned long pg_start, pfn_cnt;
992 unsigned long rg_start, rg_sz;
993#endif
994 struct hv_dynmem_device *dm = &dm_device;
995
996 memset(&resp, 0, sizeof(struct dm_hot_add_response));
997 resp.hdr.type = DM_MEM_HOT_ADD_RESPONSE;
998 resp.hdr.size = sizeof(struct dm_hot_add_response);
999
1000#ifdef CONFIG_MEMORY_HOTPLUG
1001 pg_start = dm->ha_wrk.ha_page_range.finfo.start_page;
1002 pfn_cnt = dm->ha_wrk.ha_page_range.finfo.page_cnt;
1003
1004 rg_start = dm->ha_wrk.ha_region_range.finfo.start_page;
1005 rg_sz = dm->ha_wrk.ha_region_range.finfo.page_cnt;
1006
1007 if ((rg_start == 0) && (!dm->host_specified_ha_region)) {
1008 unsigned long region_size;
1009 unsigned long region_start;
1010
1011 /*
1012 * The host has not specified the hot-add region.
1013 * Based on the hot-add page range being specified,
1014 * compute a hot-add region that can cover the pages
1015 * that need to be hot-added while ensuring the alignment
1016 * and size requirements of Linux as it relates to hot-add.
1017 */
1018 region_start = pg_start;
1019 region_size = (pfn_cnt / HA_CHUNK) * HA_CHUNK;
1020 if (pfn_cnt % HA_CHUNK)
1021 region_size += HA_CHUNK;
1022
1023 region_start = (pg_start / HA_CHUNK) * HA_CHUNK;
1024
1025 rg_start = region_start;
1026 rg_sz = region_size;
1027 }
1028
1029 if (do_hot_add)
1030 resp.page_count = process_hot_add(pg_start, pfn_cnt,
1031 rg_start, rg_sz);
1032
1033 dm->num_pages_added += resp.page_count;
1034#endif
1035 /*
1036 * The result field of the response structure has the
1037 * following semantics:
1038 *
1039 * 1. If all or some pages hot-added: Guest should return success.
1040 *
1041 * 2. If no pages could be hot-added:
1042 *
1043 * If the guest returns success, then the host
1044 * will not attempt any further hot-add operations. This
1045 * signifies a permanent failure.
1046 *
1047 * If the guest returns failure, then this failure will be
1048 * treated as a transient failure and the host may retry the
1049 * hot-add operation after some delay.
1050 */
1051 if (resp.page_count > 0)
1052 resp.result = 1;
1053 else if (!do_hot_add)
1054 resp.result = 1;
1055 else
1056 resp.result = 0;
1057
1058 if (!do_hot_add || (resp.page_count == 0))
1059 pr_err("Memory hot add failed\n");
1060
1061 dm->state = DM_INITIALIZED;
1062 resp.hdr.trans_id = atomic_inc_return(&trans_id);
1063 vmbus_sendpacket(dm->dev->channel, &resp,
1064 sizeof(struct dm_hot_add_response),
1065 (unsigned long)NULL,
1066 VM_PKT_DATA_INBAND, 0);
1067}
1068
1069static void process_info(struct hv_dynmem_device *dm, struct dm_info_msg *msg)
1070{
1071 struct dm_info_header *info_hdr;
1072
1073 info_hdr = (struct dm_info_header *)msg->info;
1074
1075 switch (info_hdr->type) {
1076 case INFO_TYPE_MAX_PAGE_CNT:
1077 if (info_hdr->data_size == sizeof(__u64)) {
1078 __u64 *max_page_count = (__u64 *)&info_hdr[1];
1079
1080 pr_info("Max. dynamic memory size: %llu MB\n",
1081 (*max_page_count) >> (20 - PAGE_SHIFT));
1082 }
1083
1084 break;
1085 default:
1086 pr_warn("Received Unknown type: %d\n", info_hdr->type);
1087 }
1088}
1089
1090static unsigned long compute_balloon_floor(void)
1091{
1092 unsigned long min_pages;
1093#define MB2PAGES(mb) ((mb) << (20 - PAGE_SHIFT))
1094 /* Simple continuous piecewiese linear function:
1095 * max MiB -> min MiB gradient
1096 * 0 0
1097 * 16 16
1098 * 32 24
1099 * 128 72 (1/2)
1100 * 512 168 (1/4)
1101 * 2048 360 (1/8)
1102 * 8192 744 (1/16)
1103 * 32768 1512 (1/32)
1104 */
1105 if (totalram_pages < MB2PAGES(128))
1106 min_pages = MB2PAGES(8) + (totalram_pages >> 1);
1107 else if (totalram_pages < MB2PAGES(512))
1108 min_pages = MB2PAGES(40) + (totalram_pages >> 2);
1109 else if (totalram_pages < MB2PAGES(2048))
1110 min_pages = MB2PAGES(104) + (totalram_pages >> 3);
1111 else if (totalram_pages < MB2PAGES(8192))
1112 min_pages = MB2PAGES(232) + (totalram_pages >> 4);
1113 else
1114 min_pages = MB2PAGES(488) + (totalram_pages >> 5);
1115#undef MB2PAGES
1116 return min_pages;
1117}
1118
1119/*
1120 * Post our status as it relates memory pressure to the
1121 * host. Host expects the guests to post this status
1122 * periodically at 1 second intervals.
1123 *
1124 * The metrics specified in this protocol are very Windows
1125 * specific and so we cook up numbers here to convey our memory
1126 * pressure.
1127 */
1128
1129static void post_status(struct hv_dynmem_device *dm)
1130{
1131 struct dm_status status;
1132 unsigned long now = jiffies;
1133 unsigned long last_post = last_post_time;
1134
1135 if (pressure_report_delay > 0) {
1136 --pressure_report_delay;
1137 return;
1138 }
1139
1140 if (!time_after(now, (last_post_time + HZ)))
1141 return;
1142
1143 memset(&status, 0, sizeof(struct dm_status));
1144 status.hdr.type = DM_STATUS_REPORT;
1145 status.hdr.size = sizeof(struct dm_status);
1146 status.hdr.trans_id = atomic_inc_return(&trans_id);
1147
1148 /*
1149 * The host expects the guest to report free and committed memory.
1150 * Furthermore, the host expects the pressure information to include
1151 * the ballooned out pages. For a given amount of memory that we are
1152 * managing we need to compute a floor below which we should not
1153 * balloon. Compute this and add it to the pressure report.
1154 * We also need to report all offline pages (num_pages_added -
1155 * num_pages_onlined) as committed to the host, otherwise it can try
1156 * asking us to balloon them out.
1157 */
1158 status.num_avail = si_mem_available();
1159 status.num_committed = vm_memory_committed() +
1160 dm->num_pages_ballooned +
1161 (dm->num_pages_added > dm->num_pages_onlined ?
1162 dm->num_pages_added - dm->num_pages_onlined : 0) +
1163 compute_balloon_floor();
1164
1165 trace_balloon_status(status.num_avail, status.num_committed,
1166 vm_memory_committed(), dm->num_pages_ballooned,
1167 dm->num_pages_added, dm->num_pages_onlined);
1168 /*
1169 * If our transaction ID is no longer current, just don't
1170 * send the status. This can happen if we were interrupted
1171 * after we picked our transaction ID.
1172 */
1173 if (status.hdr.trans_id != atomic_read(&trans_id))
1174 return;
1175
1176 /*
1177 * If the last post time that we sampled has changed,
1178 * we have raced, don't post the status.
1179 */
1180 if (last_post != last_post_time)
1181 return;
1182
1183 last_post_time = jiffies;
1184 vmbus_sendpacket(dm->dev->channel, &status,
1185 sizeof(struct dm_status),
1186 (unsigned long)NULL,
1187 VM_PKT_DATA_INBAND, 0);
1188
1189}
1190
1191static void free_balloon_pages(struct hv_dynmem_device *dm,
1192 union dm_mem_page_range *range_array)
1193{
1194 int num_pages = range_array->finfo.page_cnt;
1195 __u64 start_frame = range_array->finfo.start_page;
1196 struct page *pg;
1197 int i;
1198
1199 for (i = 0; i < num_pages; i++) {
1200 pg = pfn_to_page(i + start_frame);
1201 __free_page(pg);
1202 dm->num_pages_ballooned--;
1203 }
1204}
1205
1206
1207
1208static unsigned int alloc_balloon_pages(struct hv_dynmem_device *dm,
1209 unsigned int num_pages,
1210 struct dm_balloon_response *bl_resp,
1211 int alloc_unit)
1212{
1213 unsigned int i = 0;
1214 struct page *pg;
1215
1216 if (num_pages < alloc_unit)
1217 return 0;
1218
1219 for (i = 0; (i * alloc_unit) < num_pages; i++) {
1220 if (bl_resp->hdr.size + sizeof(union dm_mem_page_range) >
1221 PAGE_SIZE)
1222 return i * alloc_unit;
1223
1224 /*
1225 * We execute this code in a thread context. Furthermore,
1226 * we don't want the kernel to try too hard.
1227 */
1228 pg = alloc_pages(GFP_HIGHUSER | __GFP_NORETRY |
1229 __GFP_NOMEMALLOC | __GFP_NOWARN,
1230 get_order(alloc_unit << PAGE_SHIFT));
1231
1232 if (!pg)
1233 return i * alloc_unit;
1234
1235 dm->num_pages_ballooned += alloc_unit;
1236
1237 /*
1238 * If we allocatted 2M pages; split them so we
1239 * can free them in any order we get.
1240 */
1241
1242 if (alloc_unit != 1)
1243 split_page(pg, get_order(alloc_unit << PAGE_SHIFT));
1244
1245 bl_resp->range_count++;
1246 bl_resp->range_array[i].finfo.start_page =
1247 page_to_pfn(pg);
1248 bl_resp->range_array[i].finfo.page_cnt = alloc_unit;
1249 bl_resp->hdr.size += sizeof(union dm_mem_page_range);
1250
1251 }
1252
1253 return num_pages;
1254}
1255
1256static void balloon_up(struct work_struct *dummy)
1257{
1258 unsigned int num_pages = dm_device.balloon_wrk.num_pages;
1259 unsigned int num_ballooned = 0;
1260 struct dm_balloon_response *bl_resp;
1261 int alloc_unit;
1262 int ret;
1263 bool done = false;
1264 int i;
1265 long avail_pages;
1266 unsigned long floor;
1267
1268 /* The host balloons pages in 2M granularity. */
1269 WARN_ON_ONCE(num_pages % PAGES_IN_2M != 0);
1270
1271 /*
1272 * We will attempt 2M allocations. However, if we fail to
1273 * allocate 2M chunks, we will go back to 4k allocations.
1274 */
1275 alloc_unit = 512;
1276
1277 avail_pages = si_mem_available();
1278 floor = compute_balloon_floor();
1279
1280 /* Refuse to balloon below the floor, keep the 2M granularity. */
1281 if (avail_pages < num_pages || avail_pages - num_pages < floor) {
1282 pr_warn("Balloon request will be partially fulfilled. %s\n",
1283 avail_pages < num_pages ? "Not enough memory." :
1284 "Balloon floor reached.");
1285
1286 num_pages = avail_pages > floor ? (avail_pages - floor) : 0;
1287 num_pages -= num_pages % PAGES_IN_2M;
1288 }
1289
1290 while (!done) {
1291 bl_resp = (struct dm_balloon_response *)send_buffer;
1292 memset(send_buffer, 0, PAGE_SIZE);
1293 bl_resp->hdr.type = DM_BALLOON_RESPONSE;
1294 bl_resp->hdr.size = sizeof(struct dm_balloon_response);
1295 bl_resp->more_pages = 1;
1296
1297 num_pages -= num_ballooned;
1298 num_ballooned = alloc_balloon_pages(&dm_device, num_pages,
1299 bl_resp, alloc_unit);
1300
1301 if (alloc_unit != 1 && num_ballooned == 0) {
1302 alloc_unit = 1;
1303 continue;
1304 }
1305
1306 if (num_ballooned == 0 || num_ballooned == num_pages) {
1307 pr_debug("Ballooned %u out of %u requested pages.\n",
1308 num_pages, dm_device.balloon_wrk.num_pages);
1309
1310 bl_resp->more_pages = 0;
1311 done = true;
1312 dm_device.state = DM_INITIALIZED;
1313 }
1314
1315 /*
1316 * We are pushing a lot of data through the channel;
1317 * deal with transient failures caused because of the
1318 * lack of space in the ring buffer.
1319 */
1320
1321 do {
1322 bl_resp->hdr.trans_id = atomic_inc_return(&trans_id);
1323 ret = vmbus_sendpacket(dm_device.dev->channel,
1324 bl_resp,
1325 bl_resp->hdr.size,
1326 (unsigned long)NULL,
1327 VM_PKT_DATA_INBAND, 0);
1328
1329 if (ret == -EAGAIN)
1330 msleep(20);
1331 post_status(&dm_device);
1332 } while (ret == -EAGAIN);
1333
1334 if (ret) {
1335 /*
1336 * Free up the memory we allocatted.
1337 */
1338 pr_err("Balloon response failed\n");
1339
1340 for (i = 0; i < bl_resp->range_count; i++)
1341 free_balloon_pages(&dm_device,
1342 &bl_resp->range_array[i]);
1343
1344 done = true;
1345 }
1346 }
1347
1348}
1349
1350static void balloon_down(struct hv_dynmem_device *dm,
1351 struct dm_unballoon_request *req)
1352{
1353 union dm_mem_page_range *range_array = req->range_array;
1354 int range_count = req->range_count;
1355 struct dm_unballoon_response resp;
1356 int i;
1357 unsigned int prev_pages_ballooned = dm->num_pages_ballooned;
1358
1359 for (i = 0; i < range_count; i++) {
1360 free_balloon_pages(dm, &range_array[i]);
1361 complete(&dm_device.config_event);
1362 }
1363
1364 pr_debug("Freed %u ballooned pages.\n",
1365 prev_pages_ballooned - dm->num_pages_ballooned);
1366
1367 if (req->more_pages == 1)
1368 return;
1369
1370 memset(&resp, 0, sizeof(struct dm_unballoon_response));
1371 resp.hdr.type = DM_UNBALLOON_RESPONSE;
1372 resp.hdr.trans_id = atomic_inc_return(&trans_id);
1373 resp.hdr.size = sizeof(struct dm_unballoon_response);
1374
1375 vmbus_sendpacket(dm_device.dev->channel, &resp,
1376 sizeof(struct dm_unballoon_response),
1377 (unsigned long)NULL,
1378 VM_PKT_DATA_INBAND, 0);
1379
1380 dm->state = DM_INITIALIZED;
1381}
1382
1383static void balloon_onchannelcallback(void *context);
1384
1385static int dm_thread_func(void *dm_dev)
1386{
1387 struct hv_dynmem_device *dm = dm_dev;
1388
1389 while (!kthread_should_stop()) {
1390 wait_for_completion_interruptible_timeout(
1391 &dm_device.config_event, 1*HZ);
1392 /*
1393 * The host expects us to post information on the memory
1394 * pressure every second.
1395 */
1396 reinit_completion(&dm_device.config_event);
1397 post_status(dm);
1398 }
1399
1400 return 0;
1401}
1402
1403
1404static void version_resp(struct hv_dynmem_device *dm,
1405 struct dm_version_response *vresp)
1406{
1407 struct dm_version_request version_req;
1408 int ret;
1409
1410 if (vresp->is_accepted) {
1411 /*
1412 * We are done; wakeup the
1413 * context waiting for version
1414 * negotiation.
1415 */
1416 complete(&dm->host_event);
1417 return;
1418 }
1419 /*
1420 * If there are more versions to try, continue
1421 * with negotiations; if not
1422 * shutdown the service since we are not able
1423 * to negotiate a suitable version number
1424 * with the host.
1425 */
1426 if (dm->next_version == 0)
1427 goto version_error;
1428
1429 memset(&version_req, 0, sizeof(struct dm_version_request));
1430 version_req.hdr.type = DM_VERSION_REQUEST;
1431 version_req.hdr.size = sizeof(struct dm_version_request);
1432 version_req.hdr.trans_id = atomic_inc_return(&trans_id);
1433 version_req.version.version = dm->next_version;
1434 dm->version = version_req.version.version;
1435
1436 /*
1437 * Set the next version to try in case current version fails.
1438 * Win7 protocol ought to be the last one to try.
1439 */
1440 switch (version_req.version.version) {
1441 case DYNMEM_PROTOCOL_VERSION_WIN8:
1442 dm->next_version = DYNMEM_PROTOCOL_VERSION_WIN7;
1443 version_req.is_last_attempt = 0;
1444 break;
1445 default:
1446 dm->next_version = 0;
1447 version_req.is_last_attempt = 1;
1448 }
1449
1450 ret = vmbus_sendpacket(dm->dev->channel, &version_req,
1451 sizeof(struct dm_version_request),
1452 (unsigned long)NULL,
1453 VM_PKT_DATA_INBAND, 0);
1454
1455 if (ret)
1456 goto version_error;
1457
1458 return;
1459
1460version_error:
1461 dm->state = DM_INIT_ERROR;
1462 complete(&dm->host_event);
1463}
1464
1465static void cap_resp(struct hv_dynmem_device *dm,
1466 struct dm_capabilities_resp_msg *cap_resp)
1467{
1468 if (!cap_resp->is_accepted) {
1469 pr_err("Capabilities not accepted by host\n");
1470 dm->state = DM_INIT_ERROR;
1471 }
1472 complete(&dm->host_event);
1473}
1474
1475static void balloon_onchannelcallback(void *context)
1476{
1477 struct hv_device *dev = context;
1478 u32 recvlen;
1479 u64 requestid;
1480 struct dm_message *dm_msg;
1481 struct dm_header *dm_hdr;
1482 struct hv_dynmem_device *dm = hv_get_drvdata(dev);
1483 struct dm_balloon *bal_msg;
1484 struct dm_hot_add *ha_msg;
1485 union dm_mem_page_range *ha_pg_range;
1486 union dm_mem_page_range *ha_region;
1487
1488 memset(recv_buffer, 0, sizeof(recv_buffer));
1489 vmbus_recvpacket(dev->channel, recv_buffer,
1490 PAGE_SIZE, &recvlen, &requestid);
1491
1492 if (recvlen > 0) {
1493 dm_msg = (struct dm_message *)recv_buffer;
1494 dm_hdr = &dm_msg->hdr;
1495
1496 switch (dm_hdr->type) {
1497 case DM_VERSION_RESPONSE:
1498 version_resp(dm,
1499 (struct dm_version_response *)dm_msg);
1500 break;
1501
1502 case DM_CAPABILITIES_RESPONSE:
1503 cap_resp(dm,
1504 (struct dm_capabilities_resp_msg *)dm_msg);
1505 break;
1506
1507 case DM_BALLOON_REQUEST:
1508 if (dm->state == DM_BALLOON_UP)
1509 pr_warn("Currently ballooning\n");
1510 bal_msg = (struct dm_balloon *)recv_buffer;
1511 dm->state = DM_BALLOON_UP;
1512 dm_device.balloon_wrk.num_pages = bal_msg->num_pages;
1513 schedule_work(&dm_device.balloon_wrk.wrk);
1514 break;
1515
1516 case DM_UNBALLOON_REQUEST:
1517 dm->state = DM_BALLOON_DOWN;
1518 balloon_down(dm,
1519 (struct dm_unballoon_request *)recv_buffer);
1520 break;
1521
1522 case DM_MEM_HOT_ADD_REQUEST:
1523 if (dm->state == DM_HOT_ADD)
1524 pr_warn("Currently hot-adding\n");
1525 dm->state = DM_HOT_ADD;
1526 ha_msg = (struct dm_hot_add *)recv_buffer;
1527 if (ha_msg->hdr.size == sizeof(struct dm_hot_add)) {
1528 /*
1529 * This is a normal hot-add request specifying
1530 * hot-add memory.
1531 */
1532 dm->host_specified_ha_region = false;
1533 ha_pg_range = &ha_msg->range;
1534 dm->ha_wrk.ha_page_range = *ha_pg_range;
1535 dm->ha_wrk.ha_region_range.page_range = 0;
1536 } else {
1537 /*
1538 * Host is specifying that we first hot-add
1539 * a region and then partially populate this
1540 * region.
1541 */
1542 dm->host_specified_ha_region = true;
1543 ha_pg_range = &ha_msg->range;
1544 ha_region = &ha_pg_range[1];
1545 dm->ha_wrk.ha_page_range = *ha_pg_range;
1546 dm->ha_wrk.ha_region_range = *ha_region;
1547 }
1548 schedule_work(&dm_device.ha_wrk.wrk);
1549 break;
1550
1551 case DM_INFO_MESSAGE:
1552 process_info(dm, (struct dm_info_msg *)dm_msg);
1553 break;
1554
1555 default:
1556 pr_warn("Unhandled message: type: %d\n", dm_hdr->type);
1557
1558 }
1559 }
1560
1561}
1562
1563static int balloon_probe(struct hv_device *dev,
1564 const struct hv_vmbus_device_id *dev_id)
1565{
1566 int ret;
1567 unsigned long t;
1568 struct dm_version_request version_req;
1569 struct dm_capabilities cap_msg;
1570
1571#ifdef CONFIG_MEMORY_HOTPLUG
1572 do_hot_add = hot_add;
1573#else
1574 do_hot_add = false;
1575#endif
1576
1577 /*
1578 * First allocate a send buffer.
1579 */
1580
1581 send_buffer = kmalloc(PAGE_SIZE, GFP_KERNEL);
1582 if (!send_buffer)
1583 return -ENOMEM;
1584
1585 ret = vmbus_open(dev->channel, dm_ring_size, dm_ring_size, NULL, 0,
1586 balloon_onchannelcallback, dev);
1587
1588 if (ret)
1589 goto probe_error0;
1590
1591 dm_device.dev = dev;
1592 dm_device.state = DM_INITIALIZING;
1593 dm_device.next_version = DYNMEM_PROTOCOL_VERSION_WIN8;
1594 init_completion(&dm_device.host_event);
1595 init_completion(&dm_device.config_event);
1596 INIT_LIST_HEAD(&dm_device.ha_region_list);
1597 spin_lock_init(&dm_device.ha_lock);
1598 INIT_WORK(&dm_device.balloon_wrk.wrk, balloon_up);
1599 INIT_WORK(&dm_device.ha_wrk.wrk, hot_add_req);
1600 dm_device.host_specified_ha_region = false;
1601
1602 dm_device.thread =
1603 kthread_run(dm_thread_func, &dm_device, "hv_balloon");
1604 if (IS_ERR(dm_device.thread)) {
1605 ret = PTR_ERR(dm_device.thread);
1606 goto probe_error1;
1607 }
1608
1609#ifdef CONFIG_MEMORY_HOTPLUG
1610 set_online_page_callback(&hv_online_page);
1611 register_memory_notifier(&hv_memory_nb);
1612#endif
1613
1614 hv_set_drvdata(dev, &dm_device);
1615 /*
1616 * Initiate the hand shake with the host and negotiate
1617 * a version that the host can support. We start with the
1618 * highest version number and go down if the host cannot
1619 * support it.
1620 */
1621 memset(&version_req, 0, sizeof(struct dm_version_request));
1622 version_req.hdr.type = DM_VERSION_REQUEST;
1623 version_req.hdr.size = sizeof(struct dm_version_request);
1624 version_req.hdr.trans_id = atomic_inc_return(&trans_id);
1625 version_req.version.version = DYNMEM_PROTOCOL_VERSION_WIN10;
1626 version_req.is_last_attempt = 0;
1627 dm_device.version = version_req.version.version;
1628
1629 ret = vmbus_sendpacket(dev->channel, &version_req,
1630 sizeof(struct dm_version_request),
1631 (unsigned long)NULL,
1632 VM_PKT_DATA_INBAND, 0);
1633 if (ret)
1634 goto probe_error2;
1635
1636 t = wait_for_completion_timeout(&dm_device.host_event, 5*HZ);
1637 if (t == 0) {
1638 ret = -ETIMEDOUT;
1639 goto probe_error2;
1640 }
1641
1642 /*
1643 * If we could not negotiate a compatible version with the host
1644 * fail the probe function.
1645 */
1646 if (dm_device.state == DM_INIT_ERROR) {
1647 ret = -ETIMEDOUT;
1648 goto probe_error2;
1649 }
1650
1651 pr_info("Using Dynamic Memory protocol version %u.%u\n",
1652 DYNMEM_MAJOR_VERSION(dm_device.version),
1653 DYNMEM_MINOR_VERSION(dm_device.version));
1654
1655 /*
1656 * Now submit our capabilities to the host.
1657 */
1658 memset(&cap_msg, 0, sizeof(struct dm_capabilities));
1659 cap_msg.hdr.type = DM_CAPABILITIES_REPORT;
1660 cap_msg.hdr.size = sizeof(struct dm_capabilities);
1661 cap_msg.hdr.trans_id = atomic_inc_return(&trans_id);
1662
1663 cap_msg.caps.cap_bits.balloon = 1;
1664 cap_msg.caps.cap_bits.hot_add = 1;
1665
1666 /*
1667 * Specify our alignment requirements as it relates
1668 * memory hot-add. Specify 128MB alignment.
1669 */
1670 cap_msg.caps.cap_bits.hot_add_alignment = 7;
1671
1672 /*
1673 * Currently the host does not use these
1674 * values and we set them to what is done in the
1675 * Windows driver.
1676 */
1677 cap_msg.min_page_cnt = 0;
1678 cap_msg.max_page_number = -1;
1679
1680 ret = vmbus_sendpacket(dev->channel, &cap_msg,
1681 sizeof(struct dm_capabilities),
1682 (unsigned long)NULL,
1683 VM_PKT_DATA_INBAND, 0);
1684 if (ret)
1685 goto probe_error2;
1686
1687 t = wait_for_completion_timeout(&dm_device.host_event, 5*HZ);
1688 if (t == 0) {
1689 ret = -ETIMEDOUT;
1690 goto probe_error2;
1691 }
1692
1693 /*
1694 * If the host does not like our capabilities,
1695 * fail the probe function.
1696 */
1697 if (dm_device.state == DM_INIT_ERROR) {
1698 ret = -ETIMEDOUT;
1699 goto probe_error2;
1700 }
1701
1702 dm_device.state = DM_INITIALIZED;
1703 last_post_time = jiffies;
1704
1705 return 0;
1706
1707probe_error2:
1708#ifdef CONFIG_MEMORY_HOTPLUG
1709 restore_online_page_callback(&hv_online_page);
1710#endif
1711 kthread_stop(dm_device.thread);
1712
1713probe_error1:
1714 vmbus_close(dev->channel);
1715probe_error0:
1716 kfree(send_buffer);
1717 return ret;
1718}
1719
1720static int balloon_remove(struct hv_device *dev)
1721{
1722 struct hv_dynmem_device *dm = hv_get_drvdata(dev);
1723 struct hv_hotadd_state *has, *tmp;
1724 struct hv_hotadd_gap *gap, *tmp_gap;
1725 unsigned long flags;
1726
1727 if (dm->num_pages_ballooned != 0)
1728 pr_warn("Ballooned pages: %d\n", dm->num_pages_ballooned);
1729
1730 cancel_work_sync(&dm->balloon_wrk.wrk);
1731 cancel_work_sync(&dm->ha_wrk.wrk);
1732
1733 vmbus_close(dev->channel);
1734 kthread_stop(dm->thread);
1735 kfree(send_buffer);
1736#ifdef CONFIG_MEMORY_HOTPLUG
1737 restore_online_page_callback(&hv_online_page);
1738 unregister_memory_notifier(&hv_memory_nb);
1739#endif
1740 spin_lock_irqsave(&dm_device.ha_lock, flags);
1741 list_for_each_entry_safe(has, tmp, &dm->ha_region_list, list) {
1742 list_for_each_entry_safe(gap, tmp_gap, &has->gap_list, list) {
1743 list_del(&gap->list);
1744 kfree(gap);
1745 }
1746 list_del(&has->list);
1747 kfree(has);
1748 }
1749 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
1750
1751 return 0;
1752}
1753
1754static const struct hv_vmbus_device_id id_table[] = {
1755 /* Dynamic Memory Class ID */
1756 /* 525074DC-8985-46e2-8057-A307DC18A502 */
1757 { HV_DM_GUID, },
1758 { },
1759};
1760
1761MODULE_DEVICE_TABLE(vmbus, id_table);
1762
1763static struct hv_driver balloon_drv = {
1764 .name = "hv_balloon",
1765 .id_table = id_table,
1766 .probe = balloon_probe,
1767 .remove = balloon_remove,
1768};
1769
1770static int __init init_balloon_drv(void)
1771{
1772
1773 return vmbus_driver_register(&balloon_drv);
1774}
1775
1776module_init(init_balloon_drv);
1777
1778MODULE_DESCRIPTION("Hyper-V Balloon");
1779MODULE_LICENSE("GPL");