drivers/hv/hv_balloon.c

   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
  65 enum {
  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
  83 enum 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
 113 union dm_version {
 114         struct {
 115                 __u16 minor_version;
 116                 __u16 major_version;
 117         };
 118         __u32 version;
 119 } __packed;
 120
 121
 122 union 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
 138 union 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
 164 struct 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
 175 struct 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
 195 struct 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
 213 struct 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
 224 struct 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
 241 struct 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
 267 struct 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
 286 struct 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
 307 struct 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
 330 struct 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
 344 struct 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
 358 struct 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
 383 struct 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
 393 enum 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
 403 struct 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
 417 struct 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
 440 struct 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
 453 struct hv_hotadd_gap {
 454         struct list_head list;
 455         unsigned long start_pfn;
 456         unsigned long end_pfn;
 457 };
 458
 459 struct balloon_state {
 460         __u32 num_pages;
 461         struct work_struct wrk;
 462 };
 463
 464 struct 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
 470 static bool hot_add = true;
 471 static bool do_hot_add;
 472 /*
 473  * Delay reporting memory pressure by
 474  * the specified number of seconds.
 475  */
 476 static uint pressure_report_delay = 45;
 477
 478 /*
 479  * The last time we posted a pressure report to host.
 480  */
 481 static unsigned long last_post_time;
 482
 483 module_param(hot_add, bool, (S_IRUGO | S_IWUSR));
 484 MODULE_PARM_DESC(hot_add, "If set attempt memory hot_add");
 485
 486 module_param(pressure_report_delay, uint, (S_IRUGO | S_IWUSR));
 487 MODULE_PARM_DESC(pressure_report_delay, "Delay in secs in reporting pressure");
 488 static atomic_t trans_id = ATOMIC_INIT(0);
 489
 490 static int dm_ring_size = (5 * PAGE_SIZE);
 491
 492 /*
 493  * Driver specific state.
 494  */
 495
 496 enum 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
 506 static __u8 recv_buffer[PAGE_SIZE];
 507 static __u8 *send_buffer;
 508 #define PAGES_IN_2M     512
 509 #define HA_CHUNK (32 * 1024)
 510
 511 struct 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
 577 static struct hv_dynmem_device dm_device;
 578
 579 static void post_status(struct hv_dynmem_device *dm);
 580
 581 #ifdef CONFIG_MEMORY_HOTPLUG
 582 static 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
 600 static 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
 636 static 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
 676 static 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. */
 682 static 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         lockdep_assert_held(&dm_device.ha_lock);
 693         dm_device.num_pages_onlined++;
 694 }
 695
 696 static 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
 706 static 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
 774 static 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
 792 static 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
 852 static 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. It is possible to observe struct pages still
 892                          * being uninitialized here so check section instead.
 893                          * In case the section is online we need to bring the
 894                          * rest of pfns (which were not backed previously)
 895                          * online too.
 896                          */
 897                         if (start_pfn > has->start_pfn &&
 898                             online_section_nr(pfn_to_section_nr(start_pfn)))
 899                                 hv_bring_pgs_online(has, start_pfn, pgs_ol);
 900
 901                 }
 902
 903                 if ((has->ha_end_pfn < has->end_pfn) && (pfn_cnt > 0)) {
 904                         /*
 905                          * We have some residual hot add range
 906                          * that needs to be hot added; hot add
 907                          * it now. Hot add a multiple of
 908                          * of HA_CHUNK that fully covers the pages
 909                          * we have.
 910                          */
 911                         size = (has->end_pfn - has->ha_end_pfn);
 912                         if (pfn_cnt <= size) {
 913                                 size = ((pfn_cnt / HA_CHUNK) * HA_CHUNK);
 914                                 if (pfn_cnt % HA_CHUNK)
 915                                         size += HA_CHUNK;
 916                         } else {
 917                                 pfn_cnt = size;
 918                         }
 919                         spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 920                         hv_mem_hot_add(has->ha_end_pfn, size, pfn_cnt, has);
 921                         spin_lock_irqsave(&dm_device.ha_lock, flags);
 922                 }
 923                 /*
 924                  * If we managed to online any pages that were given to us,
 925                  * we declare success.
 926                  */
 927                 res = has->covered_end_pfn - old_covered_state;
 928                 break;
 929         }
 930         spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 931
 932         return res;
 933 }
 934
 935 static unsigned long process_hot_add(unsigned long pg_start,
 936                                         unsigned long pfn_cnt,
 937                                         unsigned long rg_start,
 938                                         unsigned long rg_size)
 939 {
 940         struct hv_hotadd_state *ha_region = NULL;
 941         int covered;
 942         unsigned long flags;
 943
 944         if (pfn_cnt == 0)
 945                 return 0;
 946
 947         if (!dm_device.host_specified_ha_region) {
 948                 covered = pfn_covered(pg_start, pfn_cnt);
 949                 if (covered < 0)
 950                         return 0;
 951
 952                 if (covered)
 953                         goto do_pg_range;
 954         }
 955
 956         /*
 957          * If the host has specified a hot-add range; deal with it first.
 958          */
 959
 960         if (rg_size != 0) {
 961                 ha_region = kzalloc(sizeof(struct hv_hotadd_state), GFP_KERNEL);
 962                 if (!ha_region)
 963                         return 0;
 964
 965                 INIT_LIST_HEAD(&ha_region->list);
 966                 INIT_LIST_HEAD(&ha_region->gap_list);
 967
 968                 ha_region->start_pfn = rg_start;
 969                 ha_region->ha_end_pfn = rg_start;
 970                 ha_region->covered_start_pfn = pg_start;
 971                 ha_region->covered_end_pfn = pg_start;
 972                 ha_region->end_pfn = rg_start + rg_size;
 973
 974                 spin_lock_irqsave(&dm_device.ha_lock, flags);
 975                 list_add_tail(&ha_region->list, &dm_device.ha_region_list);
 976                 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 977         }
 978
 979 do_pg_range:
 980         /*
 981          * Process the page range specified; bringing them
 982          * online if possible.
 983          */
 984         return handle_pg_range(pg_start, pfn_cnt);
 985 }
 986
 987 #endif
 988
 989 static void hot_add_req(struct work_struct *dummy)
 990 {
 991         struct dm_hot_add_response resp;
 992 #ifdef CONFIG_MEMORY_HOTPLUG
 993         unsigned long pg_start, pfn_cnt;
 994         unsigned long rg_start, rg_sz;
 995 #endif
 996         struct hv_dynmem_device *dm = &dm_device;
 997
 998         memset(&resp, 0, sizeof(struct dm_hot_add_response));
 999         resp.hdr.type = DM_MEM_HOT_ADD_RESPONSE;
1000         resp.hdr.size = sizeof(struct dm_hot_add_response);
1001
1002 #ifdef CONFIG_MEMORY_HOTPLUG
1003         pg_start = dm->ha_wrk.ha_page_range.finfo.start_page;
1004         pfn_cnt = dm->ha_wrk.ha_page_range.finfo.page_cnt;
1005
1006         rg_start = dm->ha_wrk.ha_region_range.finfo.start_page;
1007         rg_sz = dm->ha_wrk.ha_region_range.finfo.page_cnt;
1008
1009         if ((rg_start == 0) && (!dm->host_specified_ha_region)) {
1010                 unsigned long region_size;
1011                 unsigned long region_start;
1012
1013                 /*
1014                  * The host has not specified the hot-add region.
1015                  * Based on the hot-add page range being specified,
1016                  * compute a hot-add region that can cover the pages
1017                  * that need to be hot-added while ensuring the alignment
1018                  * and size requirements of Linux as it relates to hot-add.
1019                  */
1020                 region_start = pg_start;
1021                 region_size = (pfn_cnt / HA_CHUNK) * HA_CHUNK;
1022                 if (pfn_cnt % HA_CHUNK)
1023                         region_size += HA_CHUNK;
1024
1025                 region_start = (pg_start / HA_CHUNK) * HA_CHUNK;
1026
1027                 rg_start = region_start;
1028                 rg_sz = region_size;
1029         }
1030
1031         if (do_hot_add)
1032                 resp.page_count = process_hot_add(pg_start, pfn_cnt,
1033                                                 rg_start, rg_sz);
1034
1035         dm->num_pages_added += resp.page_count;
1036 #endif
1037         /*
1038          * The result field of the response structure has the
1039          * following semantics:
1040          *
1041          * 1. If all or some pages hot-added: Guest should return success.
1042          *
1043          * 2. If no pages could be hot-added:
1044          *
1045          * If the guest returns success, then the host
1046          * will not attempt any further hot-add operations. This
1047          * signifies a permanent failure.
1048          *
1049          * If the guest returns failure, then this failure will be
1050          * treated as a transient failure and the host may retry the
1051          * hot-add operation after some delay.
1052          */
1053         if (resp.page_count > 0)
1054                 resp.result = 1;
1055         else if (!do_hot_add)
1056                 resp.result = 1;
1057         else
1058                 resp.result = 0;
1059
1060         if (!do_hot_add || (resp.page_count == 0))
1061                 pr_err("Memory hot add failed\n");
1062
1063         dm->state = DM_INITIALIZED;
1064         resp.hdr.trans_id = atomic_inc_return(&trans_id);
1065         vmbus_sendpacket(dm->dev->channel, &resp,
1066                         sizeof(struct dm_hot_add_response),
1067                         (unsigned long)NULL,
1068                         VM_PKT_DATA_INBAND, 0);
1069 }
1070
1071 static void process_info(struct hv_dynmem_device *dm, struct dm_info_msg *msg)
1072 {
1073         struct dm_info_header *info_hdr;
1074
1075         info_hdr = (struct dm_info_header *)msg->info;
1076
1077         switch (info_hdr->type) {
1078         case INFO_TYPE_MAX_PAGE_CNT:
1079                 if (info_hdr->data_size == sizeof(__u64)) {
1080                         __u64 *max_page_count = (__u64 *)&info_hdr[1];
1081
1082                         pr_info("Max. dynamic memory size: %llu MB\n",
1083                                 (*max_page_count) >> (20 - PAGE_SHIFT));
1084                 }
1085
1086                 break;
1087         default:
1088                 pr_warn("Received Unknown type: %d\n", info_hdr->type);
1089         }
1090 }
1091
1092 static unsigned long compute_balloon_floor(void)
1093 {
1094         unsigned long min_pages;
1095         unsigned long nr_pages = totalram_pages();
1096 #define MB2PAGES(mb) ((mb) << (20 - PAGE_SHIFT))
1097         /* Simple continuous piecewiese linear function:
1098          *  max MiB -> min MiB  gradient
1099          *       0         0
1100          *      16        16
1101          *      32        24
1102          *     128        72    (1/2)
1103          *     512       168    (1/4)
1104          *    2048       360    (1/8)
1105          *    8192       744    (1/16)
1106          *   32768      1512    (1/32)
1107          */
1108         if (nr_pages < MB2PAGES(128))
1109                 min_pages = MB2PAGES(8) + (nr_pages >> 1);
1110         else if (nr_pages < MB2PAGES(512))
1111                 min_pages = MB2PAGES(40) + (nr_pages >> 2);
1112         else if (nr_pages < MB2PAGES(2048))
1113                 min_pages = MB2PAGES(104) + (nr_pages >> 3);
1114         else if (nr_pages < MB2PAGES(8192))
1115                 min_pages = MB2PAGES(232) + (nr_pages >> 4);
1116         else
1117                 min_pages = MB2PAGES(488) + (nr_pages >> 5);
1118 #undef MB2PAGES
1119         return min_pages;
1120 }
1121
1122 /*
1123  * Post our status as it relates memory pressure to the
1124  * host. Host expects the guests to post this status
1125  * periodically at 1 second intervals.
1126  *
1127  * The metrics specified in this protocol are very Windows
1128  * specific and so we cook up numbers here to convey our memory
1129  * pressure.
1130  */
1131
1132 static void post_status(struct hv_dynmem_device *dm)
1133 {
1134         struct dm_status status;
1135         unsigned long now = jiffies;
1136         unsigned long last_post = last_post_time;
1137
1138         if (pressure_report_delay > 0) {
1139                 --pressure_report_delay;
1140                 return;
1141         }
1142
1143         if (!time_after(now, (last_post_time + HZ)))
1144                 return;
1145
1146         memset(&status, 0, sizeof(struct dm_status));
1147         status.hdr.type = DM_STATUS_REPORT;
1148         status.hdr.size = sizeof(struct dm_status);
1149         status.hdr.trans_id = atomic_inc_return(&trans_id);
1150
1151         /*
1152          * The host expects the guest to report free and committed memory.
1153          * Furthermore, the host expects the pressure information to include
1154          * the ballooned out pages. For a given amount of memory that we are
1155          * managing we need to compute a floor below which we should not
1156          * balloon. Compute this and add it to the pressure report.
1157          * We also need to report all offline pages (num_pages_added -
1158          * num_pages_onlined) as committed to the host, otherwise it can try
1159          * asking us to balloon them out.
1160          */
1161         status.num_avail = si_mem_available();
1162         status.num_committed = vm_memory_committed() +
1163                 dm->num_pages_ballooned +
1164                 (dm->num_pages_added > dm->num_pages_onlined ?
1165                  dm->num_pages_added - dm->num_pages_onlined : 0) +
1166                 compute_balloon_floor();
1167
1168         trace_balloon_status(status.num_avail, status.num_committed,
1169                              vm_memory_committed(), dm->num_pages_ballooned,
1170                              dm->num_pages_added, dm->num_pages_onlined);
1171         /*
1172          * If our transaction ID is no longer current, just don't
1173          * send the status. This can happen if we were interrupted
1174          * after we picked our transaction ID.
1175          */
1176         if (status.hdr.trans_id != atomic_read(&trans_id))
1177                 return;
1178
1179         /*
1180          * If the last post time that we sampled has changed,
1181          * we have raced, don't post the status.
1182          */
1183         if (last_post != last_post_time)
1184                 return;
1185
1186         last_post_time = jiffies;
1187         vmbus_sendpacket(dm->dev->channel, &status,
1188                                 sizeof(struct dm_status),
1189                                 (unsigned long)NULL,
1190                                 VM_PKT_DATA_INBAND, 0);
1191
1192 }
1193
1194 static void free_balloon_pages(struct hv_dynmem_device *dm,
1195                          union dm_mem_page_range *range_array)
1196 {
1197         int num_pages = range_array->finfo.page_cnt;
1198         __u64 start_frame = range_array->finfo.start_page;
1199         struct page *pg;
1200         int i;
1201
1202         for (i = 0; i < num_pages; i++) {
1203                 pg = pfn_to_page(i + start_frame);
1204                 __free_page(pg);
1205                 dm->num_pages_ballooned--;
1206         }
1207 }
1208
1209
1210
1211 static unsigned int alloc_balloon_pages(struct hv_dynmem_device *dm,
1212                                         unsigned int num_pages,
1213                                         struct dm_balloon_response *bl_resp,
1214                                         int alloc_unit)
1215 {
1216         unsigned int i = 0;
1217         struct page *pg;
1218
1219         if (num_pages < alloc_unit)
1220                 return 0;
1221
1222         for (i = 0; (i * alloc_unit) < num_pages; i++) {
1223                 if (bl_resp->hdr.size + sizeof(union dm_mem_page_range) >
1224                         PAGE_SIZE)
1225                         return i * alloc_unit;
1226
1227                 /*
1228                  * We execute this code in a thread context. Furthermore,
1229                  * we don't want the kernel to try too hard.
1230                  */
1231                 pg = alloc_pages(GFP_HIGHUSER | __GFP_NORETRY |
1232                                 __GFP_NOMEMALLOC | __GFP_NOWARN,
1233                                 get_order(alloc_unit << PAGE_SHIFT));
1234
1235                 if (!pg)
1236                         return i * alloc_unit;
1237
1238                 dm->num_pages_ballooned += alloc_unit;
1239
1240                 /*
1241                  * If we allocatted 2M pages; split them so we
1242                  * can free them in any order we get.
1243                  */
1244
1245                 if (alloc_unit != 1)
1246                         split_page(pg, get_order(alloc_unit << PAGE_SHIFT));
1247
1248                 bl_resp->range_count++;
1249                 bl_resp->range_array[i].finfo.start_page =
1250                         page_to_pfn(pg);
1251                 bl_resp->range_array[i].finfo.page_cnt = alloc_unit;
1252                 bl_resp->hdr.size += sizeof(union dm_mem_page_range);
1253
1254         }
1255
1256         return num_pages;
1257 }
1258
1259 static void balloon_up(struct work_struct *dummy)
1260 {
1261         unsigned int num_pages = dm_device.balloon_wrk.num_pages;
1262         unsigned int num_ballooned = 0;
1263         struct dm_balloon_response *bl_resp;
1264         int alloc_unit;
1265         int ret;
1266         bool done = false;
1267         int i;
1268         long avail_pages;
1269         unsigned long floor;
1270
1271         /* The host balloons pages in 2M granularity. */
1272         WARN_ON_ONCE(num_pages % PAGES_IN_2M != 0);
1273
1274         /*
1275          * We will attempt 2M allocations. However, if we fail to
1276          * allocate 2M chunks, we will go back to 4k allocations.
1277          */
1278         alloc_unit = 512;
1279
1280         avail_pages = si_mem_available();
1281         floor = compute_balloon_floor();
1282
1283         /* Refuse to balloon below the floor, keep the 2M granularity. */
1284         if (avail_pages < num_pages || avail_pages - num_pages < floor) {
1285                 pr_warn("Balloon request will be partially fulfilled. %s\n",
1286                         avail_pages < num_pages ? "Not enough memory." :
1287                         "Balloon floor reached.");
1288
1289                 num_pages = avail_pages > floor ? (avail_pages - floor) : 0;
1290                 num_pages -= num_pages % PAGES_IN_2M;
1291         }
1292
1293         while (!done) {
1294                 bl_resp = (struct dm_balloon_response *)send_buffer;
1295                 memset(send_buffer, 0, PAGE_SIZE);
1296                 bl_resp->hdr.type = DM_BALLOON_RESPONSE;
1297                 bl_resp->hdr.size = sizeof(struct dm_balloon_response);
1298                 bl_resp->more_pages = 1;
1299
1300                 num_pages -= num_ballooned;
1301                 num_ballooned = alloc_balloon_pages(&dm_device, num_pages,
1302                                                     bl_resp, alloc_unit);
1303
1304                 if (alloc_unit != 1 && num_ballooned == 0) {
1305                         alloc_unit = 1;
1306                         continue;
1307                 }
1308
1309                 if (num_ballooned == 0 || num_ballooned == num_pages) {
1310                         pr_debug("Ballooned %u out of %u requested pages.\n",
1311                                 num_pages, dm_device.balloon_wrk.num_pages);
1312
1313                         bl_resp->more_pages = 0;
1314                         done = true;
1315                         dm_device.state = DM_INITIALIZED;
1316                 }
1317
1318                 /*
1319                  * We are pushing a lot of data through the channel;
1320                  * deal with transient failures caused because of the
1321                  * lack of space in the ring buffer.
1322                  */
1323
1324                 do {
1325                         bl_resp->hdr.trans_id = atomic_inc_return(&trans_id);
1326                         ret = vmbus_sendpacket(dm_device.dev->channel,
1327                                                 bl_resp,
1328                                                 bl_resp->hdr.size,
1329                                                 (unsigned long)NULL,
1330                                                 VM_PKT_DATA_INBAND, 0);
1331
1332                         if (ret == -EAGAIN)
1333                                 msleep(20);
1334                         post_status(&dm_device);
1335                 } while (ret == -EAGAIN);
1336
1337                 if (ret) {
1338                         /*
1339                          * Free up the memory we allocatted.
1340                          */
1341                         pr_err("Balloon response failed\n");
1342
1343                         for (i = 0; i < bl_resp->range_count; i++)
1344                                 free_balloon_pages(&dm_device,
1345                                                  &bl_resp->range_array[i]);
1346
1347                         done = true;
1348                 }
1349         }
1350
1351 }
1352
1353 static void balloon_down(struct hv_dynmem_device *dm,
1354                         struct dm_unballoon_request *req)
1355 {
1356         union dm_mem_page_range *range_array = req->range_array;
1357         int range_count = req->range_count;
1358         struct dm_unballoon_response resp;
1359         int i;
1360         unsigned int prev_pages_ballooned = dm->num_pages_ballooned;
1361
1362         for (i = 0; i < range_count; i++) {
1363                 free_balloon_pages(dm, &range_array[i]);
1364                 complete(&dm_device.config_event);
1365         }
1366
1367         pr_debug("Freed %u ballooned pages.\n",
1368                 prev_pages_ballooned - dm->num_pages_ballooned);
1369
1370         if (req->more_pages == 1)
1371                 return;
1372
1373         memset(&resp, 0, sizeof(struct dm_unballoon_response));
1374         resp.hdr.type = DM_UNBALLOON_RESPONSE;
1375         resp.hdr.trans_id = atomic_inc_return(&trans_id);
1376         resp.hdr.size = sizeof(struct dm_unballoon_response);
1377
1378         vmbus_sendpacket(dm_device.dev->channel, &resp,
1379                                 sizeof(struct dm_unballoon_response),
1380                                 (unsigned long)NULL,
1381                                 VM_PKT_DATA_INBAND, 0);
1382
1383         dm->state = DM_INITIALIZED;
1384 }
1385
1386 static void balloon_onchannelcallback(void *context);
1387
1388 static int dm_thread_func(void *dm_dev)
1389 {
1390         struct hv_dynmem_device *dm = dm_dev;
1391
1392         while (!kthread_should_stop()) {
1393                 wait_for_completion_interruptible_timeout(
1394                                                 &dm_device.config_event, 1*HZ);
1395                 /*
1396                  * The host expects us to post information on the memory
1397                  * pressure every second.
1398                  */
1399                 reinit_completion(&dm_device.config_event);
1400                 post_status(dm);
1401         }
1402
1403         return 0;
1404 }
1405
1406
1407 static void version_resp(struct hv_dynmem_device *dm,
1408                         struct dm_version_response *vresp)
1409 {
1410         struct dm_version_request version_req;
1411         int ret;
1412
1413         if (vresp->is_accepted) {
1414                 /*
1415                  * We are done; wakeup the
1416                  * context waiting for version
1417                  * negotiation.
1418                  */
1419                 complete(&dm->host_event);
1420                 return;
1421         }
1422         /*
1423          * If there are more versions to try, continue
1424          * with negotiations; if not
1425          * shutdown the service since we are not able
1426          * to negotiate a suitable version number
1427          * with the host.
1428          */
1429         if (dm->next_version == 0)
1430                 goto version_error;
1431
1432         memset(&version_req, 0, sizeof(struct dm_version_request));
1433         version_req.hdr.type = DM_VERSION_REQUEST;
1434         version_req.hdr.size = sizeof(struct dm_version_request);
1435         version_req.hdr.trans_id = atomic_inc_return(&trans_id);
1436         version_req.version.version = dm->next_version;
1437         dm->version = version_req.version.version;
1438
1439         /*
1440          * Set the next version to try in case current version fails.
1441          * Win7 protocol ought to be the last one to try.
1442          */
1443         switch (version_req.version.version) {
1444         case DYNMEM_PROTOCOL_VERSION_WIN8:
1445                 dm->next_version = DYNMEM_PROTOCOL_VERSION_WIN7;
1446                 version_req.is_last_attempt = 0;
1447                 break;
1448         default:
1449                 dm->next_version = 0;
1450                 version_req.is_last_attempt = 1;
1451         }
1452
1453         ret = vmbus_sendpacket(dm->dev->channel, &version_req,
1454                                 sizeof(struct dm_version_request),
1455                                 (unsigned long)NULL,
1456                                 VM_PKT_DATA_INBAND, 0);
1457
1458         if (ret)
1459                 goto version_error;
1460
1461         return;
1462
1463 version_error:
1464         dm->state = DM_INIT_ERROR;
1465         complete(&dm->host_event);
1466 }
1467
1468 static void cap_resp(struct hv_dynmem_device *dm,
1469                         struct dm_capabilities_resp_msg *cap_resp)
1470 {
1471         if (!cap_resp->is_accepted) {
1472                 pr_err("Capabilities not accepted by host\n");
1473                 dm->state = DM_INIT_ERROR;
1474         }
1475         complete(&dm->host_event);
1476 }
1477
1478 static void balloon_onchannelcallback(void *context)
1479 {
1480         struct hv_device *dev = context;
1481         u32 recvlen;
1482         u64 requestid;
1483         struct dm_message *dm_msg;
1484         struct dm_header *dm_hdr;
1485         struct hv_dynmem_device *dm = hv_get_drvdata(dev);
1486         struct dm_balloon *bal_msg;
1487         struct dm_hot_add *ha_msg;
1488         union dm_mem_page_range *ha_pg_range;
1489         union dm_mem_page_range *ha_region;
1490
1491         memset(recv_buffer, 0, sizeof(recv_buffer));
1492         vmbus_recvpacket(dev->channel, recv_buffer,
1493                          PAGE_SIZE, &recvlen, &requestid);
1494
1495         if (recvlen > 0) {
1496                 dm_msg = (struct dm_message *)recv_buffer;
1497                 dm_hdr = &dm_msg->hdr;
1498
1499                 switch (dm_hdr->type) {
1500                 case DM_VERSION_RESPONSE:
1501                         version_resp(dm,
1502                                  (struct dm_version_response *)dm_msg);
1503                         break;
1504
1505                 case DM_CAPABILITIES_RESPONSE:
1506                         cap_resp(dm,
1507                                  (struct dm_capabilities_resp_msg *)dm_msg);
1508                         break;
1509
1510                 case DM_BALLOON_REQUEST:
1511                         if (dm->state == DM_BALLOON_UP)
1512                                 pr_warn("Currently ballooning\n");
1513                         bal_msg = (struct dm_balloon *)recv_buffer;
1514                         dm->state = DM_BALLOON_UP;
1515                         dm_device.balloon_wrk.num_pages = bal_msg->num_pages;
1516                         schedule_work(&dm_device.balloon_wrk.wrk);
1517                         break;
1518
1519                 case DM_UNBALLOON_REQUEST:
1520                         dm->state = DM_BALLOON_DOWN;
1521                         balloon_down(dm,
1522                                  (struct dm_unballoon_request *)recv_buffer);
1523                         break;
1524
1525                 case DM_MEM_HOT_ADD_REQUEST:
1526                         if (dm->state == DM_HOT_ADD)
1527                                 pr_warn("Currently hot-adding\n");
1528                         dm->state = DM_HOT_ADD;
1529                         ha_msg = (struct dm_hot_add *)recv_buffer;
1530                         if (ha_msg->hdr.size == sizeof(struct dm_hot_add)) {
1531                                 /*
1532                                  * This is a normal hot-add request specifying
1533                                  * hot-add memory.
1534                                  */
1535                                 dm->host_specified_ha_region = false;
1536                                 ha_pg_range = &ha_msg->range;
1537                                 dm->ha_wrk.ha_page_range = *ha_pg_range;
1538                                 dm->ha_wrk.ha_region_range.page_range = 0;
1539                         } else {
1540                                 /*
1541                                  * Host is specifying that we first hot-add
1542                                  * a region and then partially populate this
1543                                  * region.
1544                                  */
1545                                 dm->host_specified_ha_region = true;
1546                                 ha_pg_range = &ha_msg->range;
1547                                 ha_region = &ha_pg_range[1];
1548                                 dm->ha_wrk.ha_page_range = *ha_pg_range;
1549                                 dm->ha_wrk.ha_region_range = *ha_region;
1550                         }
1551                         schedule_work(&dm_device.ha_wrk.wrk);
1552                         break;
1553
1554                 case DM_INFO_MESSAGE:
1555                         process_info(dm, (struct dm_info_msg *)dm_msg);
1556                         break;
1557
1558                 default:
1559                         pr_warn("Unhandled message: type: %d\n", dm_hdr->type);
1560
1561                 }
1562         }
1563
1564 }
1565
1566 static int balloon_probe(struct hv_device *dev,
1567                         const struct hv_vmbus_device_id *dev_id)
1568 {
1569         int ret;
1570         unsigned long t;
1571         struct dm_version_request version_req;
1572         struct dm_capabilities cap_msg;
1573
1574 #ifdef CONFIG_MEMORY_HOTPLUG
1575         do_hot_add = hot_add;
1576 #else
1577         do_hot_add = false;
1578 #endif
1579
1580         /*
1581          * First allocate a send buffer.
1582          */
1583
1584         send_buffer = kmalloc(PAGE_SIZE, GFP_KERNEL);
1585         if (!send_buffer)
1586                 return -ENOMEM;
1587
1588         ret = vmbus_open(dev->channel, dm_ring_size, dm_ring_size, NULL, 0,
1589                         balloon_onchannelcallback, dev);
1590
1591         if (ret)
1592                 goto probe_error0;
1593
1594         dm_device.dev = dev;
1595         dm_device.state = DM_INITIALIZING;
1596         dm_device.next_version = DYNMEM_PROTOCOL_VERSION_WIN8;
1597         init_completion(&dm_device.host_event);
1598         init_completion(&dm_device.config_event);
1599         INIT_LIST_HEAD(&dm_device.ha_region_list);
1600         spin_lock_init(&dm_device.ha_lock);
1601         INIT_WORK(&dm_device.balloon_wrk.wrk, balloon_up);
1602         INIT_WORK(&dm_device.ha_wrk.wrk, hot_add_req);
1603         dm_device.host_specified_ha_region = false;
1604
1605         dm_device.thread =
1606                  kthread_run(dm_thread_func, &dm_device, "hv_balloon");
1607         if (IS_ERR(dm_device.thread)) {
1608                 ret = PTR_ERR(dm_device.thread);
1609                 goto probe_error1;
1610         }
1611
1612 #ifdef CONFIG_MEMORY_HOTPLUG
1613         set_online_page_callback(&hv_online_page);
1614         register_memory_notifier(&hv_memory_nb);
1615 #endif
1616
1617         hv_set_drvdata(dev, &dm_device);
1618         /*
1619          * Initiate the hand shake with the host and negotiate
1620          * a version that the host can support. We start with the
1621          * highest version number and go down if the host cannot
1622          * support it.
1623          */
1624         memset(&version_req, 0, sizeof(struct dm_version_request));
1625         version_req.hdr.type = DM_VERSION_REQUEST;
1626         version_req.hdr.size = sizeof(struct dm_version_request);
1627         version_req.hdr.trans_id = atomic_inc_return(&trans_id);
1628         version_req.version.version = DYNMEM_PROTOCOL_VERSION_WIN10;
1629         version_req.is_last_attempt = 0;
1630         dm_device.version = version_req.version.version;
1631
1632         ret = vmbus_sendpacket(dev->channel, &version_req,
1633                                 sizeof(struct dm_version_request),
1634                                 (unsigned long)NULL,
1635                                 VM_PKT_DATA_INBAND, 0);
1636         if (ret)
1637                 goto probe_error2;
1638
1639         t = wait_for_completion_timeout(&dm_device.host_event, 5*HZ);
1640         if (t == 0) {
1641                 ret = -ETIMEDOUT;
1642                 goto probe_error2;
1643         }
1644
1645         /*
1646          * If we could not negotiate a compatible version with the host
1647          * fail the probe function.
1648          */
1649         if (dm_device.state == DM_INIT_ERROR) {
1650                 ret = -ETIMEDOUT;
1651                 goto probe_error2;
1652         }
1653
1654         pr_info("Using Dynamic Memory protocol version %u.%u\n",
1655                 DYNMEM_MAJOR_VERSION(dm_device.version),
1656                 DYNMEM_MINOR_VERSION(dm_device.version));
1657
1658         /*
1659          * Now submit our capabilities to the host.
1660          */
1661         memset(&cap_msg, 0, sizeof(struct dm_capabilities));
1662         cap_msg.hdr.type = DM_CAPABILITIES_REPORT;
1663         cap_msg.hdr.size = sizeof(struct dm_capabilities);
1664         cap_msg.hdr.trans_id = atomic_inc_return(&trans_id);
1665
1666         cap_msg.caps.cap_bits.balloon = 1;
1667         cap_msg.caps.cap_bits.hot_add = 1;
1668
1669         /*
1670          * Specify our alignment requirements as it relates
1671          * memory hot-add. Specify 128MB alignment.
1672          */
1673         cap_msg.caps.cap_bits.hot_add_alignment = 7;
1674
1675         /*
1676          * Currently the host does not use these
1677          * values and we set them to what is done in the
1678          * Windows driver.
1679          */
1680         cap_msg.min_page_cnt = 0;
1681         cap_msg.max_page_number = -1;
1682
1683         ret = vmbus_sendpacket(dev->channel, &cap_msg,
1684                                 sizeof(struct dm_capabilities),
1685                                 (unsigned long)NULL,
1686                                 VM_PKT_DATA_INBAND, 0);
1687         if (ret)
1688                 goto probe_error2;
1689
1690         t = wait_for_completion_timeout(&dm_device.host_event, 5*HZ);
1691         if (t == 0) {
1692                 ret = -ETIMEDOUT;
1693                 goto probe_error2;
1694         }
1695
1696         /*
1697          * If the host does not like our capabilities,
1698          * fail the probe function.
1699          */
1700         if (dm_device.state == DM_INIT_ERROR) {
1701                 ret = -ETIMEDOUT;
1702                 goto probe_error2;
1703         }
1704
1705         dm_device.state = DM_INITIALIZED;
1706         last_post_time = jiffies;
1707
1708         return 0;
1709
1710 probe_error2:
1711 #ifdef CONFIG_MEMORY_HOTPLUG
1712         restore_online_page_callback(&hv_online_page);
1713 #endif
1714         kthread_stop(dm_device.thread);
1715
1716 probe_error1:
1717         vmbus_close(dev->channel);
1718 probe_error0:
1719         kfree(send_buffer);
1720         return ret;
1721 }
1722
1723 static int balloon_remove(struct hv_device *dev)
1724 {
1725         struct hv_dynmem_device *dm = hv_get_drvdata(dev);
1726         struct hv_hotadd_state *has, *tmp;
1727         struct hv_hotadd_gap *gap, *tmp_gap;
1728         unsigned long flags;
1729
1730         if (dm->num_pages_ballooned != 0)
1731                 pr_warn("Ballooned pages: %d\n", dm->num_pages_ballooned);
1732
1733         cancel_work_sync(&dm->balloon_wrk.wrk);
1734         cancel_work_sync(&dm->ha_wrk.wrk);
1735
1736         vmbus_close(dev->channel);
1737         kthread_stop(dm->thread);
1738         kfree(send_buffer);
1739 #ifdef CONFIG_MEMORY_HOTPLUG
1740         restore_online_page_callback(&hv_online_page);
1741         unregister_memory_notifier(&hv_memory_nb);
1742 #endif
1743         spin_lock_irqsave(&dm_device.ha_lock, flags);
1744         list_for_each_entry_safe(has, tmp, &dm->ha_region_list, list) {
1745                 list_for_each_entry_safe(gap, tmp_gap, &has->gap_list, list) {
1746                         list_del(&gap->list);
1747                         kfree(gap);
1748                 }
1749                 list_del(&has->list);
1750                 kfree(has);
1751         }
1752         spin_unlock_irqrestore(&dm_device.ha_lock, flags);
1753
1754         return 0;
1755 }
1756
1757 static const struct hv_vmbus_device_id id_table[] = {
1758         /* Dynamic Memory Class ID */
1759         /* 525074DC-8985-46e2-8057-A307DC18A502 */
1760         { HV_DM_GUID, },
1761         { },
1762 };
1763
1764 MODULE_DEVICE_TABLE(vmbus, id_table);
1765
1766 static  struct hv_driver balloon_drv = {
1767         .name = "hv_balloon",
1768         .id_table = id_table,
1769         .probe =  balloon_probe,
1770         .remove =  balloon_remove,
1771         .driver = {
1772                 .probe_type = PROBE_PREFER_ASYNCHRONOUS,
1773         },
1774 };
1775
1776 static int __init init_balloon_drv(void)
1777 {
1778
1779         return vmbus_driver_register(&balloon_drv);
1780 }
1781
1782 module_init(init_balloon_drv);
1783
1784 MODULE_DESCRIPTION("Hyper-V Balloon");
1785 MODULE_LICENSE("GPL");