drivers/hv/hv_balloon.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (c) 2012, Microsoft Corporation.
   4  *
   5  * Author:
   6  *   K. Y. Srinivasan <kys@microsoft.com>
   7  */
   8
   9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  10
  11 #include <linux/kernel.h>
  12 #include <linux/jiffies.h>
  13 #include <linux/mman.h>
  14 #include <linux/debugfs.h>
  15 #include <linux/delay.h>
  16 #include <linux/init.h>
  17 #include <linux/module.h>
  18 #include <linux/slab.h>
  19 #include <linux/kthread.h>
  20 #include <linux/completion.h>
  21 #include <linux/count_zeros.h>
  22 #include <linux/memory_hotplug.h>
  23 #include <linux/memory.h>
  24 #include <linux/notifier.h>
  25 #include <linux/percpu_counter.h>
  26 #include <linux/page_reporting.h>
  27
  28 #include <linux/hyperv.h>
  29 #include <asm/hyperv-tlfs.h>
  30
  31 #include <asm/mshyperv.h>
  32
  33 #define CREATE_TRACE_POINTS
  34 #include "hv_trace_balloon.h"
  35
  36 /*
  37  * We begin with definitions supporting the Dynamic Memory protocol
  38  * with the host.
  39  *
  40  * Begin protocol definitions.
  41  */
  42
  43
  44
  45 /*
  46  * Protocol versions. The low word is the minor version, the high word the major
  47  * version.
  48  *
  49  * History:
  50  * Initial version 1.0
  51  * Changed to 0.1 on 2009/03/25
  52  * Changes to 0.2 on 2009/05/14
  53  * Changes to 0.3 on 2009/12/03
  54  * Changed to 1.0 on 2011/04/05
  55  */
  56
  57 #define DYNMEM_MAKE_VERSION(Major, Minor) ((__u32)(((Major) << 16) | (Minor)))
  58 #define DYNMEM_MAJOR_VERSION(Version) ((__u32)(Version) >> 16)
  59 #define DYNMEM_MINOR_VERSION(Version) ((__u32)(Version) & 0xff)
  60
  61 enum {
  62         DYNMEM_PROTOCOL_VERSION_1 = DYNMEM_MAKE_VERSION(0, 3),
  63         DYNMEM_PROTOCOL_VERSION_2 = DYNMEM_MAKE_VERSION(1, 0),
  64         DYNMEM_PROTOCOL_VERSION_3 = DYNMEM_MAKE_VERSION(2, 0),
  65
  66         DYNMEM_PROTOCOL_VERSION_WIN7 = DYNMEM_PROTOCOL_VERSION_1,
  67         DYNMEM_PROTOCOL_VERSION_WIN8 = DYNMEM_PROTOCOL_VERSION_2,
  68         DYNMEM_PROTOCOL_VERSION_WIN10 = DYNMEM_PROTOCOL_VERSION_3,
  69
  70         DYNMEM_PROTOCOL_VERSION_CURRENT = DYNMEM_PROTOCOL_VERSION_WIN10
  71 };
  72
  73
  74
  75 /*
  76  * Message Types
  77  */
  78
  79 enum dm_message_type {
  80         /*
  81          * Version 0.3
  82          */
  83         DM_ERROR                        = 0,
  84         DM_VERSION_REQUEST              = 1,
  85         DM_VERSION_RESPONSE             = 2,
  86         DM_CAPABILITIES_REPORT          = 3,
  87         DM_CAPABILITIES_RESPONSE        = 4,
  88         DM_STATUS_REPORT                = 5,
  89         DM_BALLOON_REQUEST              = 6,
  90         DM_BALLOON_RESPONSE             = 7,
  91         DM_UNBALLOON_REQUEST            = 8,
  92         DM_UNBALLOON_RESPONSE           = 9,
  93         DM_MEM_HOT_ADD_REQUEST          = 10,
  94         DM_MEM_HOT_ADD_RESPONSE         = 11,
  95         DM_VERSION_03_MAX               = 11,
  96         /*
  97          * Version 1.0.
  98          */
  99         DM_INFO_MESSAGE                 = 12,
 100         DM_VERSION_1_MAX                = 12
 101 };
 102
 103
 104 /*
 105  * Structures defining the dynamic memory management
 106  * protocol.
 107  */
 108
 109 union dm_version {
 110         struct {
 111                 __u16 minor_version;
 112                 __u16 major_version;
 113         };
 114         __u32 version;
 115 } __packed;
 116
 117
 118 union dm_caps {
 119         struct {
 120                 __u64 balloon:1;
 121                 __u64 hot_add:1;
 122                 /*
 123                  * To support guests that may have alignment
 124                  * limitations on hot-add, the guest can specify
 125                  * its alignment requirements; a value of n
 126                  * represents an alignment of 2^n in mega bytes.
 127                  */
 128                 __u64 hot_add_alignment:4;
 129                 __u64 reservedz:58;
 130         } cap_bits;
 131         __u64 caps;
 132 } __packed;
 133
 134 union dm_mem_page_range {
 135         struct  {
 136                 /*
 137                  * The PFN number of the first page in the range.
 138                  * 40 bits is the architectural limit of a PFN
 139                  * number for AMD64.
 140                  */
 141                 __u64 start_page:40;
 142                 /*
 143                  * The number of pages in the range.
 144                  */
 145                 __u64 page_cnt:24;
 146         } finfo;
 147         __u64  page_range;
 148 } __packed;
 149
 150
 151
 152 /*
 153  * The header for all dynamic memory messages:
 154  *
 155  * type: Type of the message.
 156  * size: Size of the message in bytes; including the header.
 157  * trans_id: The guest is responsible for manufacturing this ID.
 158  */
 159
 160 struct dm_header {
 161         __u16 type;
 162         __u16 size;
 163         __u32 trans_id;
 164 } __packed;
 165
 166 /*
 167  * A generic message format for dynamic memory.
 168  * Specific message formats are defined later in the file.
 169  */
 170
 171 struct dm_message {
 172         struct dm_header hdr;
 173         __u8 data[]; /* enclosed message */
 174 } __packed;
 175
 176
 177 /*
 178  * Specific message types supporting the dynamic memory protocol.
 179  */
 180
 181 /*
 182  * Version negotiation message. Sent from the guest to the host.
 183  * The guest is free to try different versions until the host
 184  * accepts the version.
 185  *
 186  * dm_version: The protocol version requested.
 187  * is_last_attempt: If TRUE, this is the last version guest will request.
 188  * reservedz: Reserved field, set to zero.
 189  */
 190
 191 struct dm_version_request {
 192         struct dm_header hdr;
 193         union dm_version version;
 194         __u32 is_last_attempt:1;
 195         __u32 reservedz:31;
 196 } __packed;
 197
 198 /*
 199  * Version response message; Host to Guest and indicates
 200  * if the host has accepted the version sent by the guest.
 201  *
 202  * is_accepted: If TRUE, host has accepted the version and the guest
 203  * should proceed to the next stage of the protocol. FALSE indicates that
 204  * guest should re-try with a different version.
 205  *
 206  * reservedz: Reserved field, set to zero.
 207  */
 208
 209 struct dm_version_response {
 210         struct dm_header hdr;
 211         __u64 is_accepted:1;
 212         __u64 reservedz:63;
 213 } __packed;
 214
 215 /*
 216  * Message reporting capabilities. This is sent from the guest to the
 217  * host.
 218  */
 219
 220 struct dm_capabilities {
 221         struct dm_header hdr;
 222         union dm_caps caps;
 223         __u64 min_page_cnt;
 224         __u64 max_page_number;
 225 } __packed;
 226
 227 /*
 228  * Response to the capabilities message. This is sent from the host to the
 229  * guest. This message notifies if the host has accepted the guest's
 230  * capabilities. If the host has not accepted, the guest must shutdown
 231  * the service.
 232  *
 233  * is_accepted: Indicates if the host has accepted guest's capabilities.
 234  * reservedz: Must be 0.
 235  */
 236
 237 struct dm_capabilities_resp_msg {
 238         struct dm_header hdr;
 239         __u64 is_accepted:1;
 240         __u64 reservedz:63;
 241 } __packed;
 242
 243 /*
 244  * This message is used to report memory pressure from the guest.
 245  * This message is not part of any transaction and there is no
 246  * response to this message.
 247  *
 248  * num_avail: Available memory in pages.
 249  * num_committed: Committed memory in pages.
 250  * page_file_size: The accumulated size of all page files
 251  *                 in the system in pages.
 252  * zero_free: The number of zero and free pages.
 253  * page_file_writes: The writes to the page file in pages.
 254  * io_diff: An indicator of file cache efficiency or page file activity,
 255  *          calculated as File Cache Page Fault Count - Page Read Count.
 256  *          This value is in pages.
 257  *
 258  * Some of these metrics are Windows specific and fortunately
 259  * the algorithm on the host side that computes the guest memory
 260  * pressure only uses num_committed value.
 261  */
 262
 263 struct dm_status {
 264         struct dm_header hdr;
 265         __u64 num_avail;
 266         __u64 num_committed;
 267         __u64 page_file_size;
 268         __u64 zero_free;
 269         __u32 page_file_writes;
 270         __u32 io_diff;
 271 } __packed;
 272
 273
 274 /*
 275  * Message to ask the guest to allocate memory - balloon up message.
 276  * This message is sent from the host to the guest. The guest may not be
 277  * able to allocate as much memory as requested.
 278  *
 279  * num_pages: number of pages to allocate.
 280  */
 281
 282 struct dm_balloon {
 283         struct dm_header hdr;
 284         __u32 num_pages;
 285         __u32 reservedz;
 286 } __packed;
 287
 288
 289 /*
 290  * Balloon response message; this message is sent from the guest
 291  * to the host in response to the balloon message.
 292  *
 293  * reservedz: Reserved; must be set to zero.
 294  * more_pages: If FALSE, this is the last message of the transaction.
 295  * if TRUE there will atleast one more message from the guest.
 296  *
 297  * range_count: The number of ranges in the range array.
 298  *
 299  * range_array: An array of page ranges returned to the host.
 300  *
 301  */
 302
 303 struct dm_balloon_response {
 304         struct dm_header hdr;
 305         __u32 reservedz;
 306         __u32 more_pages:1;
 307         __u32 range_count:31;
 308         union dm_mem_page_range range_array[];
 309 } __packed;
 310
 311 /*
 312  * Un-balloon message; this message is sent from the host
 313  * to the guest to give guest more memory.
 314  *
 315  * more_pages: If FALSE, this is the last message of the transaction.
 316  * if TRUE there will atleast one more message from the guest.
 317  *
 318  * reservedz: Reserved; must be set to zero.
 319  *
 320  * range_count: The number of ranges in the range array.
 321  *
 322  * range_array: An array of page ranges returned to the host.
 323  *
 324  */
 325
 326 struct dm_unballoon_request {
 327         struct dm_header hdr;
 328         __u32 more_pages:1;
 329         __u32 reservedz:31;
 330         __u32 range_count;
 331         union dm_mem_page_range range_array[];
 332 } __packed;
 333
 334 /*
 335  * Un-balloon response message; this message is sent from the guest
 336  * to the host in response to an unballoon request.
 337  *
 338  */
 339
 340 struct dm_unballoon_response {
 341         struct dm_header hdr;
 342 } __packed;
 343
 344
 345 /*
 346  * Hot add request message. Message sent from the host to the guest.
 347  *
 348  * mem_range: Memory range to hot add.
 349  *
 350  */
 351
 352 struct dm_hot_add {
 353         struct dm_header hdr;
 354         union dm_mem_page_range range;
 355 } __packed;
 356
 357 /*
 358  * Hot add response message.
 359  * This message is sent by the guest to report the status of a hot add request.
 360  * If page_count is less than the requested page count, then the host should
 361  * assume all further hot add requests will fail, since this indicates that
 362  * the guest has hit an upper physical memory barrier.
 363  *
 364  * Hot adds may also fail due to low resources; in this case, the guest must
 365  * not complete this message until the hot add can succeed, and the host must
 366  * not send a new hot add request until the response is sent.
 367  * If VSC fails to hot add memory DYNMEM_NUMBER_OF_UNSUCCESSFUL_HOTADD_ATTEMPTS
 368  * times it fails the request.
 369  *
 370  *
 371  * page_count: number of pages that were successfully hot added.
 372  *
 373  * result: result of the operation 1: success, 0: failure.
 374  *
 375  */
 376
 377 struct dm_hot_add_response {
 378         struct dm_header hdr;
 379         __u32 page_count;
 380         __u32 result;
 381 } __packed;
 382
 383 /*
 384  * Types of information sent from host to the guest.
 385  */
 386
 387 enum dm_info_type {
 388         INFO_TYPE_MAX_PAGE_CNT = 0,
 389         MAX_INFO_TYPE
 390 };
 391
 392
 393 /*
 394  * Header for the information message.
 395  */
 396
 397 struct dm_info_header {
 398         enum dm_info_type type;
 399         __u32 data_size;
 400 } __packed;
 401
 402 /*
 403  * This message is sent from the host to the guest to pass
 404  * some relevant information (win8 addition).
 405  *
 406  * reserved: no used.
 407  * info_size: size of the information blob.
 408  * info: information blob.
 409  */
 410
 411 struct dm_info_msg {
 412         struct dm_header hdr;
 413         __u32 reserved;
 414         __u32 info_size;
 415         __u8  info[];
 416 };
 417
 418 /*
 419  * End protocol definitions.
 420  */
 421
 422 /*
 423  * State to manage hot adding memory into the guest.
 424  * The range start_pfn : end_pfn specifies the range
 425  * that the host has asked us to hot add. The range
 426  * start_pfn : ha_end_pfn specifies the range that we have
 427  * currently hot added. We hot add in multiples of 128M
 428  * chunks; it is possible that we may not be able to bring
 429  * online all the pages in the region. The range
 430  * covered_start_pfn:covered_end_pfn defines the pages that can
 431  * be brough online.
 432  */
 433
 434 struct hv_hotadd_state {
 435         struct list_head list;
 436         unsigned long start_pfn;
 437         unsigned long covered_start_pfn;
 438         unsigned long covered_end_pfn;
 439         unsigned long ha_end_pfn;
 440         unsigned long end_pfn;
 441         /*
 442          * A list of gaps.
 443          */
 444         struct list_head gap_list;
 445 };
 446
 447 struct hv_hotadd_gap {
 448         struct list_head list;
 449         unsigned long start_pfn;
 450         unsigned long end_pfn;
 451 };
 452
 453 struct balloon_state {
 454         __u32 num_pages;
 455         struct work_struct wrk;
 456 };
 457
 458 struct hot_add_wrk {
 459         union dm_mem_page_range ha_page_range;
 460         union dm_mem_page_range ha_region_range;
 461         struct work_struct wrk;
 462 };
 463
 464 static bool allow_hibernation;
 465 static bool hot_add = true;
 466 static bool do_hot_add;
 467 /*
 468  * Delay reporting memory pressure by
 469  * the specified number of seconds.
 470  */
 471 static uint pressure_report_delay = 45;
 472 extern unsigned int page_reporting_order;
 473 #define HV_MAX_FAILURES 2
 474
 475 /*
 476  * The last time we posted a pressure report to host.
 477  */
 478 static unsigned long last_post_time;
 479
 480 static int hv_hypercall_multi_failure;
 481
 482 module_param(hot_add, bool, (S_IRUGO | S_IWUSR));
 483 MODULE_PARM_DESC(hot_add, "If set attempt memory hot_add");
 484
 485 module_param(pressure_report_delay, uint, (S_IRUGO | S_IWUSR));
 486 MODULE_PARM_DESC(pressure_report_delay, "Delay in secs in reporting pressure");
 487 static atomic_t trans_id = ATOMIC_INIT(0);
 488
 489 static int dm_ring_size = VMBUS_RING_SIZE(16 * 1024);
 490
 491 /*
 492  * Driver specific state.
 493  */
 494
 495 enum hv_dm_state {
 496         DM_INITIALIZING = 0,
 497         DM_INITIALIZED,
 498         DM_BALLOON_UP,
 499         DM_BALLOON_DOWN,
 500         DM_HOT_ADD,
 501         DM_INIT_ERROR
 502 };
 503
 504
 505 static __u8 recv_buffer[HV_HYP_PAGE_SIZE];
 506 static __u8 balloon_up_send_buffer[HV_HYP_PAGE_SIZE];
 507 #define PAGES_IN_2M (2 * 1024 * 1024 / PAGE_SIZE)
 508 #define HA_CHUNK (128 * 1024 * 1024 / PAGE_SIZE)
 509
 510 struct hv_dynmem_device {
 511         struct hv_device *dev;
 512         enum hv_dm_state state;
 513         struct completion host_event;
 514         struct completion config_event;
 515
 516         /*
 517          * Number of pages we have currently ballooned out.
 518          */
 519         unsigned int num_pages_ballooned;
 520         unsigned int num_pages_onlined;
 521         unsigned int num_pages_added;
 522
 523         /*
 524          * State to manage the ballooning (up) operation.
 525          */
 526         struct balloon_state balloon_wrk;
 527
 528         /*
 529          * State to execute the "hot-add" operation.
 530          */
 531         struct hot_add_wrk ha_wrk;
 532
 533         /*
 534          * This state tracks if the host has specified a hot-add
 535          * region.
 536          */
 537         bool host_specified_ha_region;
 538
 539         /*
 540          * State to synchronize hot-add.
 541          */
 542         struct completion  ol_waitevent;
 543         /*
 544          * This thread handles hot-add
 545          * requests from the host as well as notifying
 546          * the host with regards to memory pressure in
 547          * the guest.
 548          */
 549         struct task_struct *thread;
 550
 551         /*
 552          * Protects ha_region_list, num_pages_onlined counter and individual
 553          * regions from ha_region_list.
 554          */
 555         spinlock_t ha_lock;
 556
 557         /*
 558          * A list of hot-add regions.
 559          */
 560         struct list_head ha_region_list;
 561
 562         /*
 563          * We start with the highest version we can support
 564          * and downgrade based on the host; we save here the
 565          * next version to try.
 566          */
 567         __u32 next_version;
 568
 569         /*
 570          * The negotiated version agreed by host.
 571          */
 572         __u32 version;
 573
 574         struct page_reporting_dev_info pr_dev_info;
 575
 576         /*
 577          * Maximum number of pages that can be hot_add-ed
 578          */
 579         __u64 max_dynamic_page_count;
 580 };
 581
 582 static struct hv_dynmem_device dm_device;
 583
 584 static void post_status(struct hv_dynmem_device *dm);
 585
 586 static void enable_page_reporting(void);
 587
 588 static void disable_page_reporting(void);
 589
 590 #ifdef CONFIG_MEMORY_HOTPLUG
 591 static inline bool has_pfn_is_backed(struct hv_hotadd_state *has,
 592                                      unsigned long pfn)
 593 {
 594         struct hv_hotadd_gap *gap;
 595
 596         /* The page is not backed. */
 597         if ((pfn < has->covered_start_pfn) || (pfn >= has->covered_end_pfn))
 598                 return false;
 599
 600         /* Check for gaps. */
 601         list_for_each_entry(gap, &has->gap_list, list) {
 602                 if ((pfn >= gap->start_pfn) && (pfn < gap->end_pfn))
 603                         return false;
 604         }
 605
 606         return true;
 607 }
 608
 609 static unsigned long hv_page_offline_check(unsigned long start_pfn,
 610                                            unsigned long nr_pages)
 611 {
 612         unsigned long pfn = start_pfn, count = 0;
 613         struct hv_hotadd_state *has;
 614         bool found;
 615
 616         while (pfn < start_pfn + nr_pages) {
 617                 /*
 618                  * Search for HAS which covers the pfn and when we find one
 619                  * count how many consequitive PFNs are covered.
 620                  */
 621                 found = false;
 622                 list_for_each_entry(has, &dm_device.ha_region_list, list) {
 623                         while ((pfn >= has->start_pfn) &&
 624                                (pfn < has->end_pfn) &&
 625                                (pfn < start_pfn + nr_pages)) {
 626                                 found = true;
 627                                 if (has_pfn_is_backed(has, pfn))
 628                                         count++;
 629                                 pfn++;
 630                         }
 631                 }
 632
 633                 /*
 634                  * This PFN is not in any HAS (e.g. we're offlining a region
 635                  * which was present at boot), no need to account for it. Go
 636                  * to the next one.
 637                  */
 638                 if (!found)
 639                         pfn++;
 640         }
 641
 642         return count;
 643 }
 644
 645 static int hv_memory_notifier(struct notifier_block *nb, unsigned long val,
 646                               void *v)
 647 {
 648         struct memory_notify *mem = (struct memory_notify *)v;
 649         unsigned long flags, pfn_count;
 650
 651         switch (val) {
 652         case MEM_ONLINE:
 653         case MEM_CANCEL_ONLINE:
 654                 complete(&dm_device.ol_waitevent);
 655                 break;
 656
 657         case MEM_OFFLINE:
 658                 spin_lock_irqsave(&dm_device.ha_lock, flags);
 659                 pfn_count = hv_page_offline_check(mem->start_pfn,
 660                                                   mem->nr_pages);
 661                 if (pfn_count <= dm_device.num_pages_onlined) {
 662                         dm_device.num_pages_onlined -= pfn_count;
 663                 } else {
 664                         /*
 665                          * We're offlining more pages than we managed to online.
 666                          * This is unexpected. In any case don't let
 667                          * num_pages_onlined wrap around zero.
 668                          */
 669                         WARN_ON_ONCE(1);
 670                         dm_device.num_pages_onlined = 0;
 671                 }
 672                 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 673                 break;
 674         case MEM_GOING_ONLINE:
 675         case MEM_GOING_OFFLINE:
 676         case MEM_CANCEL_OFFLINE:
 677                 break;
 678         }
 679         return NOTIFY_OK;
 680 }
 681
 682 static struct notifier_block hv_memory_nb = {
 683         .notifier_call = hv_memory_notifier,
 684         .priority = 0
 685 };
 686
 687 /* Check if the particular page is backed and can be onlined and online it. */
 688 static void hv_page_online_one(struct hv_hotadd_state *has, struct page *pg)
 689 {
 690         if (!has_pfn_is_backed(has, page_to_pfn(pg))) {
 691                 if (!PageOffline(pg))
 692                         __SetPageOffline(pg);
 693                 return;
 694         }
 695         if (PageOffline(pg))
 696                 __ClearPageOffline(pg);
 697
 698         /* This frame is currently backed; online the page. */
 699         generic_online_page(pg, 0);
 700
 701         lockdep_assert_held(&dm_device.ha_lock);
 702         dm_device.num_pages_onlined++;
 703 }
 704
 705 static void hv_bring_pgs_online(struct hv_hotadd_state *has,
 706                                 unsigned long start_pfn, unsigned long size)
 707 {
 708         int i;
 709
 710         pr_debug("Online %lu pages starting at pfn 0x%lx\n", size, start_pfn);
 711         for (i = 0; i < size; i++)
 712                 hv_page_online_one(has, pfn_to_page(start_pfn + i));
 713 }
 714
 715 static void hv_mem_hot_add(unsigned long start, unsigned long size,
 716                                 unsigned long pfn_count,
 717                                 struct hv_hotadd_state *has)
 718 {
 719         int ret = 0;
 720         int i, nid;
 721         unsigned long start_pfn;
 722         unsigned long processed_pfn;
 723         unsigned long total_pfn = pfn_count;
 724         unsigned long flags;
 725
 726         for (i = 0; i < (size/HA_CHUNK); i++) {
 727                 start_pfn = start + (i * HA_CHUNK);
 728
 729                 spin_lock_irqsave(&dm_device.ha_lock, flags);
 730                 has->ha_end_pfn +=  HA_CHUNK;
 731
 732                 if (total_pfn > HA_CHUNK) {
 733                         processed_pfn = HA_CHUNK;
 734                         total_pfn -= HA_CHUNK;
 735                 } else {
 736                         processed_pfn = total_pfn;
 737                         total_pfn = 0;
 738                 }
 739
 740                 has->covered_end_pfn +=  processed_pfn;
 741                 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 742
 743                 reinit_completion(&dm_device.ol_waitevent);
 744
 745                 nid = memory_add_physaddr_to_nid(PFN_PHYS(start_pfn));
 746                 ret = add_memory(nid, PFN_PHYS((start_pfn)),
 747                                 (HA_CHUNK << PAGE_SHIFT), MHP_MERGE_RESOURCE);
 748
 749                 if (ret) {
 750                         pr_err("hot_add memory failed error is %d\n", ret);
 751                         if (ret == -EEXIST) {
 752                                 /*
 753                                  * This error indicates that the error
 754                                  * is not a transient failure. This is the
 755                                  * case where the guest's physical address map
 756                                  * precludes hot adding memory. Stop all further
 757                                  * memory hot-add.
 758                                  */
 759                                 do_hot_add = false;
 760                         }
 761                         spin_lock_irqsave(&dm_device.ha_lock, flags);
 762                         has->ha_end_pfn -= HA_CHUNK;
 763                         has->covered_end_pfn -=  processed_pfn;
 764                         spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 765                         break;
 766                 }
 767
 768                 /*
 769                  * Wait for memory to get onlined. If the kernel onlined the
 770                  * memory when adding it, this will return directly. Otherwise,
 771                  * it will wait for user space to online the memory. This helps
 772                  * to avoid adding memory faster than it is getting onlined. As
 773                  * adding succeeded, it is ok to proceed even if the memory was
 774                  * not onlined in time.
 775                  */
 776                 wait_for_completion_timeout(&dm_device.ol_waitevent, 5 * HZ);
 777                 post_status(&dm_device);
 778         }
 779 }
 780
 781 static void hv_online_page(struct page *pg, unsigned int order)
 782 {
 783         struct hv_hotadd_state *has;
 784         unsigned long flags;
 785         unsigned long pfn = page_to_pfn(pg);
 786
 787         spin_lock_irqsave(&dm_device.ha_lock, flags);
 788         list_for_each_entry(has, &dm_device.ha_region_list, list) {
 789                 /* The page belongs to a different HAS. */
 790                 if ((pfn < has->start_pfn) ||
 791                                 (pfn + (1UL << order) > has->end_pfn))
 792                         continue;
 793
 794                 hv_bring_pgs_online(has, pfn, 1UL << order);
 795                 break;
 796         }
 797         spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 798 }
 799
 800 static int pfn_covered(unsigned long start_pfn, unsigned long pfn_cnt)
 801 {
 802         struct hv_hotadd_state *has;
 803         struct hv_hotadd_gap *gap;
 804         unsigned long residual, new_inc;
 805         int ret = 0;
 806         unsigned long flags;
 807
 808         spin_lock_irqsave(&dm_device.ha_lock, flags);
 809         list_for_each_entry(has, &dm_device.ha_region_list, list) {
 810                 /*
 811                  * If the pfn range we are dealing with is not in the current
 812                  * "hot add block", move on.
 813                  */
 814                 if (start_pfn < has->start_pfn || start_pfn >= has->end_pfn)
 815                         continue;
 816
 817                 /*
 818                  * If the current start pfn is not where the covered_end
 819                  * is, create a gap and update covered_end_pfn.
 820                  */
 821                 if (has->covered_end_pfn != start_pfn) {
 822                         gap = kzalloc(sizeof(struct hv_hotadd_gap), GFP_ATOMIC);
 823                         if (!gap) {
 824                                 ret = -ENOMEM;
 825                                 break;
 826                         }
 827
 828                         INIT_LIST_HEAD(&gap->list);
 829                         gap->start_pfn = has->covered_end_pfn;
 830                         gap->end_pfn = start_pfn;
 831                         list_add_tail(&gap->list, &has->gap_list);
 832
 833                         has->covered_end_pfn = start_pfn;
 834                 }
 835
 836                 /*
 837                  * If the current hot add-request extends beyond
 838                  * our current limit; extend it.
 839                  */
 840                 if ((start_pfn + pfn_cnt) > has->end_pfn) {
 841                         residual = (start_pfn + pfn_cnt - has->end_pfn);
 842                         /*
 843                          * Extend the region by multiples of HA_CHUNK.
 844                          */
 845                         new_inc = (residual / HA_CHUNK) * HA_CHUNK;
 846                         if (residual % HA_CHUNK)
 847                                 new_inc += HA_CHUNK;
 848
 849                         has->end_pfn += new_inc;
 850                 }
 851
 852                 ret = 1;
 853                 break;
 854         }
 855         spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 856
 857         return ret;
 858 }
 859
 860 static unsigned long handle_pg_range(unsigned long pg_start,
 861                                         unsigned long pg_count)
 862 {
 863         unsigned long start_pfn = pg_start;
 864         unsigned long pfn_cnt = pg_count;
 865         unsigned long size;
 866         struct hv_hotadd_state *has;
 867         unsigned long pgs_ol = 0;
 868         unsigned long old_covered_state;
 869         unsigned long res = 0, flags;
 870
 871         pr_debug("Hot adding %lu pages starting at pfn 0x%lx.\n", pg_count,
 872                 pg_start);
 873
 874         spin_lock_irqsave(&dm_device.ha_lock, flags);
 875         list_for_each_entry(has, &dm_device.ha_region_list, list) {
 876                 /*
 877                  * If the pfn range we are dealing with is not in the current
 878                  * "hot add block", move on.
 879                  */
 880                 if (start_pfn < has->start_pfn || start_pfn >= has->end_pfn)
 881                         continue;
 882
 883                 old_covered_state = has->covered_end_pfn;
 884
 885                 if (start_pfn < has->ha_end_pfn) {
 886                         /*
 887                          * This is the case where we are backing pages
 888                          * in an already hot added region. Bring
 889                          * these pages online first.
 890                          */
 891                         pgs_ol = has->ha_end_pfn - start_pfn;
 892                         if (pgs_ol > pfn_cnt)
 893                                 pgs_ol = pfn_cnt;
 894
 895                         has->covered_end_pfn +=  pgs_ol;
 896                         pfn_cnt -= pgs_ol;
 897                         /*
 898                          * Check if the corresponding memory block is already
 899                          * online. It is possible to observe struct pages still
 900                          * being uninitialized here so check section instead.
 901                          * In case the section is online we need to bring the
 902                          * rest of pfns (which were not backed previously)
 903                          * online too.
 904                          */
 905                         if (start_pfn > has->start_pfn &&
 906                             online_section_nr(pfn_to_section_nr(start_pfn)))
 907                                 hv_bring_pgs_online(has, start_pfn, pgs_ol);
 908
 909                 }
 910
 911                 if ((has->ha_end_pfn < has->end_pfn) && (pfn_cnt > 0)) {
 912                         /*
 913                          * We have some residual hot add range
 914                          * that needs to be hot added; hot add
 915                          * it now. Hot add a multiple of
 916                          * HA_CHUNK that fully covers the pages
 917                          * we have.
 918                          */
 919                         size = (has->end_pfn - has->ha_end_pfn);
 920                         if (pfn_cnt <= size) {
 921                                 size = ((pfn_cnt / HA_CHUNK) * HA_CHUNK);
 922                                 if (pfn_cnt % HA_CHUNK)
 923                                         size += HA_CHUNK;
 924                         } else {
 925                                 pfn_cnt = size;
 926                         }
 927                         spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 928                         hv_mem_hot_add(has->ha_end_pfn, size, pfn_cnt, has);
 929                         spin_lock_irqsave(&dm_device.ha_lock, flags);
 930                 }
 931                 /*
 932                  * If we managed to online any pages that were given to us,
 933                  * we declare success.
 934                  */
 935                 res = has->covered_end_pfn - old_covered_state;
 936                 break;
 937         }
 938         spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 939
 940         return res;
 941 }
 942
 943 static unsigned long process_hot_add(unsigned long pg_start,
 944                                         unsigned long pfn_cnt,
 945                                         unsigned long rg_start,
 946                                         unsigned long rg_size)
 947 {
 948         struct hv_hotadd_state *ha_region = NULL;
 949         int covered;
 950         unsigned long flags;
 951
 952         if (pfn_cnt == 0)
 953                 return 0;
 954
 955         if (!dm_device.host_specified_ha_region) {
 956                 covered = pfn_covered(pg_start, pfn_cnt);
 957                 if (covered < 0)
 958                         return 0;
 959
 960                 if (covered)
 961                         goto do_pg_range;
 962         }
 963
 964         /*
 965          * If the host has specified a hot-add range; deal with it first.
 966          */
 967
 968         if (rg_size != 0) {
 969                 ha_region = kzalloc(sizeof(struct hv_hotadd_state), GFP_KERNEL);
 970                 if (!ha_region)
 971                         return 0;
 972
 973                 INIT_LIST_HEAD(&ha_region->list);
 974                 INIT_LIST_HEAD(&ha_region->gap_list);
 975
 976                 ha_region->start_pfn = rg_start;
 977                 ha_region->ha_end_pfn = rg_start;
 978                 ha_region->covered_start_pfn = pg_start;
 979                 ha_region->covered_end_pfn = pg_start;
 980                 ha_region->end_pfn = rg_start + rg_size;
 981
 982                 spin_lock_irqsave(&dm_device.ha_lock, flags);
 983                 list_add_tail(&ha_region->list, &dm_device.ha_region_list);
 984                 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 985         }
 986
 987 do_pg_range:
 988         /*
 989          * Process the page range specified; bringing them
 990          * online if possible.
 991          */
 992         return handle_pg_range(pg_start, pfn_cnt);
 993 }
 994
 995 #endif
 996
 997 static void hot_add_req(struct work_struct *dummy)
 998 {
 999         struct dm_hot_add_response resp;
1000 #ifdef CONFIG_MEMORY_HOTPLUG
1001         unsigned long pg_start, pfn_cnt;
1002         unsigned long rg_start, rg_sz;
1003 #endif
1004         struct hv_dynmem_device *dm = &dm_device;
1005
1006         memset(&resp, 0, sizeof(struct dm_hot_add_response));
1007         resp.hdr.type = DM_MEM_HOT_ADD_RESPONSE;
1008         resp.hdr.size = sizeof(struct dm_hot_add_response);
1009
1010 #ifdef CONFIG_MEMORY_HOTPLUG
1011         pg_start = dm->ha_wrk.ha_page_range.finfo.start_page;
1012         pfn_cnt = dm->ha_wrk.ha_page_range.finfo.page_cnt;
1013
1014         rg_start = dm->ha_wrk.ha_region_range.finfo.start_page;
1015         rg_sz = dm->ha_wrk.ha_region_range.finfo.page_cnt;
1016
1017         if ((rg_start == 0) && (!dm->host_specified_ha_region)) {
1018                 unsigned long region_size;
1019                 unsigned long region_start;
1020
1021                 /*
1022                  * The host has not specified the hot-add region.
1023                  * Based on the hot-add page range being specified,
1024                  * compute a hot-add region that can cover the pages
1025                  * that need to be hot-added while ensuring the alignment
1026                  * and size requirements of Linux as it relates to hot-add.
1027                  */
1028                 region_size = (pfn_cnt / HA_CHUNK) * HA_CHUNK;
1029                 if (pfn_cnt % HA_CHUNK)
1030                         region_size += HA_CHUNK;
1031
1032                 region_start = (pg_start / HA_CHUNK) * HA_CHUNK;
1033
1034                 rg_start = region_start;
1035                 rg_sz = region_size;
1036         }
1037
1038         if (do_hot_add)
1039                 resp.page_count = process_hot_add(pg_start, pfn_cnt,
1040                                                 rg_start, rg_sz);
1041
1042         dm->num_pages_added += resp.page_count;
1043 #endif
1044         /*
1045          * The result field of the response structure has the
1046          * following semantics:
1047          *
1048          * 1. If all or some pages hot-added: Guest should return success.
1049          *
1050          * 2. If no pages could be hot-added:
1051          *
1052          * If the guest returns success, then the host
1053          * will not attempt any further hot-add operations. This
1054          * signifies a permanent failure.
1055          *
1056          * If the guest returns failure, then this failure will be
1057          * treated as a transient failure and the host may retry the
1058          * hot-add operation after some delay.
1059          */
1060         if (resp.page_count > 0)
1061                 resp.result = 1;
1062         else if (!do_hot_add)
1063                 resp.result = 1;
1064         else
1065                 resp.result = 0;
1066
1067         if (!do_hot_add || resp.page_count == 0) {
1068                 if (!allow_hibernation)
1069                         pr_err("Memory hot add failed\n");
1070                 else
1071                         pr_info("Ignore hot-add request!\n");
1072         }
1073
1074         dm->state = DM_INITIALIZED;
1075         resp.hdr.trans_id = atomic_inc_return(&trans_id);
1076         vmbus_sendpacket(dm->dev->channel, &resp,
1077                         sizeof(struct dm_hot_add_response),
1078                         (unsigned long)NULL,
1079                         VM_PKT_DATA_INBAND, 0);
1080 }
1081
1082 static void process_info(struct hv_dynmem_device *dm, struct dm_info_msg *msg)
1083 {
1084         struct dm_info_header *info_hdr;
1085
1086         info_hdr = (struct dm_info_header *)msg->info;
1087
1088         switch (info_hdr->type) {
1089         case INFO_TYPE_MAX_PAGE_CNT:
1090                 if (info_hdr->data_size == sizeof(__u64)) {
1091                         __u64 *max_page_count = (__u64 *)&info_hdr[1];
1092
1093                         pr_info("Max. dynamic memory size: %llu MB\n",
1094                                 (*max_page_count) >> (20 - HV_HYP_PAGE_SHIFT));
1095                         dm->max_dynamic_page_count = *max_page_count;
1096                 }
1097
1098                 break;
1099         default:
1100                 pr_warn("Received Unknown type: %d\n", info_hdr->type);
1101         }
1102 }
1103
1104 static unsigned long compute_balloon_floor(void)
1105 {
1106         unsigned long min_pages;
1107         unsigned long nr_pages = totalram_pages();
1108 #define MB2PAGES(mb) ((mb) << (20 - PAGE_SHIFT))
1109         /* Simple continuous piecewiese linear function:
1110          *  max MiB -> min MiB  gradient
1111          *       0         0
1112          *      16        16
1113          *      32        24
1114          *     128        72    (1/2)
1115          *     512       168    (1/4)
1116          *    2048       360    (1/8)
1117          *    8192       744    (1/16)
1118          *   32768      1512    (1/32)
1119          */
1120         if (nr_pages < MB2PAGES(128))
1121                 min_pages = MB2PAGES(8) + (nr_pages >> 1);
1122         else if (nr_pages < MB2PAGES(512))
1123                 min_pages = MB2PAGES(40) + (nr_pages >> 2);
1124         else if (nr_pages < MB2PAGES(2048))
1125                 min_pages = MB2PAGES(104) + (nr_pages >> 3);
1126         else if (nr_pages < MB2PAGES(8192))
1127                 min_pages = MB2PAGES(232) + (nr_pages >> 4);
1128         else
1129                 min_pages = MB2PAGES(488) + (nr_pages >> 5);
1130 #undef MB2PAGES
1131         return min_pages;
1132 }
1133
1134 /*
1135  * Compute total committed memory pages
1136  */
1137
1138 static unsigned long get_pages_committed(struct hv_dynmem_device *dm)
1139 {
1140         return vm_memory_committed() +
1141                 dm->num_pages_ballooned +
1142                 (dm->num_pages_added > dm->num_pages_onlined ?
1143                  dm->num_pages_added - dm->num_pages_onlined : 0) +
1144                 compute_balloon_floor();
1145 }
1146
1147 /*
1148  * Post our status as it relates memory pressure to the
1149  * host. Host expects the guests to post this status
1150  * periodically at 1 second intervals.
1151  *
1152  * The metrics specified in this protocol are very Windows
1153  * specific and so we cook up numbers here to convey our memory
1154  * pressure.
1155  */
1156
1157 static void post_status(struct hv_dynmem_device *dm)
1158 {
1159         struct dm_status status;
1160         unsigned long now = jiffies;
1161         unsigned long last_post = last_post_time;
1162         unsigned long num_pages_avail, num_pages_committed;
1163
1164         if (pressure_report_delay > 0) {
1165                 --pressure_report_delay;
1166                 return;
1167         }
1168
1169         if (!time_after(now, (last_post_time + HZ)))
1170                 return;
1171
1172         memset(&status, 0, sizeof(struct dm_status));
1173         status.hdr.type = DM_STATUS_REPORT;
1174         status.hdr.size = sizeof(struct dm_status);
1175         status.hdr.trans_id = atomic_inc_return(&trans_id);
1176
1177         /*
1178          * The host expects the guest to report free and committed memory.
1179          * Furthermore, the host expects the pressure information to include
1180          * the ballooned out pages. For a given amount of memory that we are
1181          * managing we need to compute a floor below which we should not
1182          * balloon. Compute this and add it to the pressure report.
1183          * We also need to report all offline pages (num_pages_added -
1184          * num_pages_onlined) as committed to the host, otherwise it can try
1185          * asking us to balloon them out.
1186          */
1187         num_pages_avail = si_mem_available();
1188         num_pages_committed = get_pages_committed(dm);
1189
1190         trace_balloon_status(num_pages_avail, num_pages_committed,
1191                              vm_memory_committed(), dm->num_pages_ballooned,
1192                              dm->num_pages_added, dm->num_pages_onlined);
1193
1194         /* Convert numbers of pages into numbers of HV_HYP_PAGEs. */
1195         status.num_avail = num_pages_avail * NR_HV_HYP_PAGES_IN_PAGE;
1196         status.num_committed = num_pages_committed * NR_HV_HYP_PAGES_IN_PAGE;
1197
1198         /*
1199          * If our transaction ID is no longer current, just don't
1200          * send the status. This can happen if we were interrupted
1201          * after we picked our transaction ID.
1202          */
1203         if (status.hdr.trans_id != atomic_read(&trans_id))
1204                 return;
1205
1206         /*
1207          * If the last post time that we sampled has changed,
1208          * we have raced, don't post the status.
1209          */
1210         if (last_post != last_post_time)
1211                 return;
1212
1213         last_post_time = jiffies;
1214         vmbus_sendpacket(dm->dev->channel, &status,
1215                                 sizeof(struct dm_status),
1216                                 (unsigned long)NULL,
1217                                 VM_PKT_DATA_INBAND, 0);
1218
1219 }
1220
1221 static void free_balloon_pages(struct hv_dynmem_device *dm,
1222                          union dm_mem_page_range *range_array)
1223 {
1224         int num_pages = range_array->finfo.page_cnt;
1225         __u64 start_frame = range_array->finfo.start_page;
1226         struct page *pg;
1227         int i;
1228
1229         for (i = 0; i < num_pages; i++) {
1230                 pg = pfn_to_page(i + start_frame);
1231                 __ClearPageOffline(pg);
1232                 __free_page(pg);
1233                 dm->num_pages_ballooned--;
1234                 adjust_managed_page_count(pg, 1);
1235         }
1236 }
1237
1238
1239
1240 static unsigned int alloc_balloon_pages(struct hv_dynmem_device *dm,
1241                                         unsigned int num_pages,
1242                                         struct dm_balloon_response *bl_resp,
1243                                         int alloc_unit)
1244 {
1245         unsigned int i, j;
1246         struct page *pg;
1247
1248         for (i = 0; i < num_pages / alloc_unit; i++) {
1249                 if (bl_resp->hdr.size + sizeof(union dm_mem_page_range) >
1250                         HV_HYP_PAGE_SIZE)
1251                         return i * alloc_unit;
1252
1253                 /*
1254                  * We execute this code in a thread context. Furthermore,
1255                  * we don't want the kernel to try too hard.
1256                  */
1257                 pg = alloc_pages(GFP_HIGHUSER | __GFP_NORETRY |
1258                                 __GFP_NOMEMALLOC | __GFP_NOWARN,
1259                                 get_order(alloc_unit << PAGE_SHIFT));
1260
1261                 if (!pg)
1262                         return i * alloc_unit;
1263
1264                 dm->num_pages_ballooned += alloc_unit;
1265
1266                 /*
1267                  * If we allocatted 2M pages; split them so we
1268                  * can free them in any order we get.
1269                  */
1270
1271                 if (alloc_unit != 1)
1272                         split_page(pg, get_order(alloc_unit << PAGE_SHIFT));
1273
1274                 /* mark all pages offline */
1275                 for (j = 0; j < alloc_unit; j++) {
1276                         __SetPageOffline(pg + j);
1277                         adjust_managed_page_count(pg + j, -1);
1278                 }
1279
1280                 bl_resp->range_count++;
1281                 bl_resp->range_array[i].finfo.start_page =
1282                         page_to_pfn(pg);
1283                 bl_resp->range_array[i].finfo.page_cnt = alloc_unit;
1284                 bl_resp->hdr.size += sizeof(union dm_mem_page_range);
1285
1286         }
1287
1288         return i * alloc_unit;
1289 }
1290
1291 static void balloon_up(struct work_struct *dummy)
1292 {
1293         unsigned int num_pages = dm_device.balloon_wrk.num_pages;
1294         unsigned int num_ballooned = 0;
1295         struct dm_balloon_response *bl_resp;
1296         int alloc_unit;
1297         int ret;
1298         bool done = false;
1299         int i;
1300         long avail_pages;
1301         unsigned long floor;
1302
1303         /*
1304          * We will attempt 2M allocations. However, if we fail to
1305          * allocate 2M chunks, we will go back to PAGE_SIZE allocations.
1306          */
1307         alloc_unit = PAGES_IN_2M;
1308
1309         avail_pages = si_mem_available();
1310         floor = compute_balloon_floor();
1311
1312         /* Refuse to balloon below the floor. */
1313         if (avail_pages < num_pages || avail_pages - num_pages < floor) {
1314                 pr_info("Balloon request will be partially fulfilled. %s\n",
1315                         avail_pages < num_pages ? "Not enough memory." :
1316                         "Balloon floor reached.");
1317
1318                 num_pages = avail_pages > floor ? (avail_pages - floor) : 0;
1319         }
1320
1321         while (!done) {
1322                 memset(balloon_up_send_buffer, 0, HV_HYP_PAGE_SIZE);
1323                 bl_resp = (struct dm_balloon_response *)balloon_up_send_buffer;
1324                 bl_resp->hdr.type = DM_BALLOON_RESPONSE;
1325                 bl_resp->hdr.size = sizeof(struct dm_balloon_response);
1326                 bl_resp->more_pages = 1;
1327
1328                 num_pages -= num_ballooned;
1329                 num_ballooned = alloc_balloon_pages(&dm_device, num_pages,
1330                                                     bl_resp, alloc_unit);
1331
1332                 if (alloc_unit != 1 && num_ballooned == 0) {
1333                         alloc_unit = 1;
1334                         continue;
1335                 }
1336
1337                 if (num_ballooned == 0 || num_ballooned == num_pages) {
1338                         pr_debug("Ballooned %u out of %u requested pages.\n",
1339                                 num_pages, dm_device.balloon_wrk.num_pages);
1340
1341                         bl_resp->more_pages = 0;
1342                         done = true;
1343                         dm_device.state = DM_INITIALIZED;
1344                 }
1345
1346                 /*
1347                  * We are pushing a lot of data through the channel;
1348                  * deal with transient failures caused because of the
1349                  * lack of space in the ring buffer.
1350                  */
1351
1352                 do {
1353                         bl_resp->hdr.trans_id = atomic_inc_return(&trans_id);
1354                         ret = vmbus_sendpacket(dm_device.dev->channel,
1355                                                 bl_resp,
1356                                                 bl_resp->hdr.size,
1357                                                 (unsigned long)NULL,
1358                                                 VM_PKT_DATA_INBAND, 0);
1359
1360                         if (ret == -EAGAIN)
1361                                 msleep(20);
1362                         post_status(&dm_device);
1363                 } while (ret == -EAGAIN);
1364
1365                 if (ret) {
1366                         /*
1367                          * Free up the memory we allocatted.
1368                          */
1369                         pr_err("Balloon response failed\n");
1370
1371                         for (i = 0; i < bl_resp->range_count; i++)
1372                                 free_balloon_pages(&dm_device,
1373                                                  &bl_resp->range_array[i]);
1374
1375                         done = true;
1376                 }
1377         }
1378
1379 }
1380
1381 static void balloon_down(struct hv_dynmem_device *dm,
1382                         struct dm_unballoon_request *req)
1383 {
1384         union dm_mem_page_range *range_array = req->range_array;
1385         int range_count = req->range_count;
1386         struct dm_unballoon_response resp;
1387         int i;
1388         unsigned int prev_pages_ballooned = dm->num_pages_ballooned;
1389
1390         for (i = 0; i < range_count; i++) {
1391                 free_balloon_pages(dm, &range_array[i]);
1392                 complete(&dm_device.config_event);
1393         }
1394
1395         pr_debug("Freed %u ballooned pages.\n",
1396                 prev_pages_ballooned - dm->num_pages_ballooned);
1397
1398         if (req->more_pages == 1)
1399                 return;
1400
1401         memset(&resp, 0, sizeof(struct dm_unballoon_response));
1402         resp.hdr.type = DM_UNBALLOON_RESPONSE;
1403         resp.hdr.trans_id = atomic_inc_return(&trans_id);
1404         resp.hdr.size = sizeof(struct dm_unballoon_response);
1405
1406         vmbus_sendpacket(dm_device.dev->channel, &resp,
1407                                 sizeof(struct dm_unballoon_response),
1408                                 (unsigned long)NULL,
1409                                 VM_PKT_DATA_INBAND, 0);
1410
1411         dm->state = DM_INITIALIZED;
1412 }
1413
1414 static void balloon_onchannelcallback(void *context);
1415
1416 static int dm_thread_func(void *dm_dev)
1417 {
1418         struct hv_dynmem_device *dm = dm_dev;
1419
1420         while (!kthread_should_stop()) {
1421                 wait_for_completion_interruptible_timeout(
1422                                                 &dm_device.config_event, 1*HZ);
1423                 /*
1424                  * The host expects us to post information on the memory
1425                  * pressure every second.
1426                  */
1427                 reinit_completion(&dm_device.config_event);
1428                 post_status(dm);
1429                 /*
1430                  * disable free page reporting if multiple hypercall
1431                  * failure flag set. It is not done in the page_reporting
1432                  * callback context as that causes a deadlock between
1433                  * page_reporting_process() and page_reporting_unregister()
1434                  */
1435                 if (hv_hypercall_multi_failure >= HV_MAX_FAILURES) {
1436                         pr_err("Multiple failures in cold memory discard hypercall, disabling page reporting\n");
1437                         disable_page_reporting();
1438                         /* Reset the flag after disabling reporting */
1439                         hv_hypercall_multi_failure = 0;
1440                 }
1441         }
1442
1443         return 0;
1444 }
1445
1446
1447 static void version_resp(struct hv_dynmem_device *dm,
1448                         struct dm_version_response *vresp)
1449 {
1450         struct dm_version_request version_req;
1451         int ret;
1452
1453         if (vresp->is_accepted) {
1454                 /*
1455                  * We are done; wakeup the
1456                  * context waiting for version
1457                  * negotiation.
1458                  */
1459                 complete(&dm->host_event);
1460                 return;
1461         }
1462         /*
1463          * If there are more versions to try, continue
1464          * with negotiations; if not
1465          * shutdown the service since we are not able
1466          * to negotiate a suitable version number
1467          * with the host.
1468          */
1469         if (dm->next_version == 0)
1470                 goto version_error;
1471
1472         memset(&version_req, 0, sizeof(struct dm_version_request));
1473         version_req.hdr.type = DM_VERSION_REQUEST;
1474         version_req.hdr.size = sizeof(struct dm_version_request);
1475         version_req.hdr.trans_id = atomic_inc_return(&trans_id);
1476         version_req.version.version = dm->next_version;
1477         dm->version = version_req.version.version;
1478
1479         /*
1480          * Set the next version to try in case current version fails.
1481          * Win7 protocol ought to be the last one to try.
1482          */
1483         switch (version_req.version.version) {
1484         case DYNMEM_PROTOCOL_VERSION_WIN8:
1485                 dm->next_version = DYNMEM_PROTOCOL_VERSION_WIN7;
1486                 version_req.is_last_attempt = 0;
1487                 break;
1488         default:
1489                 dm->next_version = 0;
1490                 version_req.is_last_attempt = 1;
1491         }
1492
1493         ret = vmbus_sendpacket(dm->dev->channel, &version_req,
1494                                 sizeof(struct dm_version_request),
1495                                 (unsigned long)NULL,
1496                                 VM_PKT_DATA_INBAND, 0);
1497
1498         if (ret)
1499                 goto version_error;
1500
1501         return;
1502
1503 version_error:
1504         dm->state = DM_INIT_ERROR;
1505         complete(&dm->host_event);
1506 }
1507
1508 static void cap_resp(struct hv_dynmem_device *dm,
1509                         struct dm_capabilities_resp_msg *cap_resp)
1510 {
1511         if (!cap_resp->is_accepted) {
1512                 pr_err("Capabilities not accepted by host\n");
1513                 dm->state = DM_INIT_ERROR;
1514         }
1515         complete(&dm->host_event);
1516 }
1517
1518 static void balloon_onchannelcallback(void *context)
1519 {
1520         struct hv_device *dev = context;
1521         u32 recvlen;
1522         u64 requestid;
1523         struct dm_message *dm_msg;
1524         struct dm_header *dm_hdr;
1525         struct hv_dynmem_device *dm = hv_get_drvdata(dev);
1526         struct dm_balloon *bal_msg;
1527         struct dm_hot_add *ha_msg;
1528         union dm_mem_page_range *ha_pg_range;
1529         union dm_mem_page_range *ha_region;
1530
1531         memset(recv_buffer, 0, sizeof(recv_buffer));
1532         vmbus_recvpacket(dev->channel, recv_buffer,
1533                          HV_HYP_PAGE_SIZE, &recvlen, &requestid);
1534
1535         if (recvlen > 0) {
1536                 dm_msg = (struct dm_message *)recv_buffer;
1537                 dm_hdr = &dm_msg->hdr;
1538
1539                 switch (dm_hdr->type) {
1540                 case DM_VERSION_RESPONSE:
1541                         version_resp(dm,
1542                                  (struct dm_version_response *)dm_msg);
1543                         break;
1544
1545                 case DM_CAPABILITIES_RESPONSE:
1546                         cap_resp(dm,
1547                                  (struct dm_capabilities_resp_msg *)dm_msg);
1548                         break;
1549
1550                 case DM_BALLOON_REQUEST:
1551                         if (allow_hibernation) {
1552                                 pr_info("Ignore balloon-up request!\n");
1553                                 break;
1554                         }
1555
1556                         if (dm->state == DM_BALLOON_UP)
1557                                 pr_warn("Currently ballooning\n");
1558                         bal_msg = (struct dm_balloon *)recv_buffer;
1559                         dm->state = DM_BALLOON_UP;
1560                         dm_device.balloon_wrk.num_pages = bal_msg->num_pages;
1561                         schedule_work(&dm_device.balloon_wrk.wrk);
1562                         break;
1563
1564                 case DM_UNBALLOON_REQUEST:
1565                         if (allow_hibernation) {
1566                                 pr_info("Ignore balloon-down request!\n");
1567                                 break;
1568                         }
1569
1570                         dm->state = DM_BALLOON_DOWN;
1571                         balloon_down(dm,
1572                                  (struct dm_unballoon_request *)recv_buffer);
1573                         break;
1574
1575                 case DM_MEM_HOT_ADD_REQUEST:
1576                         if (dm->state == DM_HOT_ADD)
1577                                 pr_warn("Currently hot-adding\n");
1578                         dm->state = DM_HOT_ADD;
1579                         ha_msg = (struct dm_hot_add *)recv_buffer;
1580                         if (ha_msg->hdr.size == sizeof(struct dm_hot_add)) {
1581                                 /*
1582                                  * This is a normal hot-add request specifying
1583                                  * hot-add memory.
1584                                  */
1585                                 dm->host_specified_ha_region = false;
1586                                 ha_pg_range = &ha_msg->range;
1587                                 dm->ha_wrk.ha_page_range = *ha_pg_range;
1588                                 dm->ha_wrk.ha_region_range.page_range = 0;
1589                         } else {
1590                                 /*
1591                                  * Host is specifying that we first hot-add
1592                                  * a region and then partially populate this
1593                                  * region.
1594                                  */
1595                                 dm->host_specified_ha_region = true;
1596                                 ha_pg_range = &ha_msg->range;
1597                                 ha_region = &ha_pg_range[1];
1598                                 dm->ha_wrk.ha_page_range = *ha_pg_range;
1599                                 dm->ha_wrk.ha_region_range = *ha_region;
1600                         }
1601                         schedule_work(&dm_device.ha_wrk.wrk);
1602                         break;
1603
1604                 case DM_INFO_MESSAGE:
1605                         process_info(dm, (struct dm_info_msg *)dm_msg);
1606                         break;
1607
1608                 default:
1609                         pr_warn_ratelimited("Unhandled message: type: %d\n", dm_hdr->type);
1610
1611                 }
1612         }
1613
1614 }
1615
1616 #define HV_LARGE_REPORTING_ORDER        9
1617 #define HV_LARGE_REPORTING_LEN (HV_HYP_PAGE_SIZE << \
1618                 HV_LARGE_REPORTING_ORDER)
1619 static int hv_free_page_report(struct page_reporting_dev_info *pr_dev_info,
1620                     struct scatterlist *sgl, unsigned int nents)
1621 {
1622         unsigned long flags;
1623         struct hv_memory_hint *hint;
1624         int i, order;
1625         u64 status;
1626         struct scatterlist *sg;
1627
1628         WARN_ON_ONCE(nents > HV_MEMORY_HINT_MAX_GPA_PAGE_RANGES);
1629         WARN_ON_ONCE(sgl->length < (HV_HYP_PAGE_SIZE << page_reporting_order));
1630         local_irq_save(flags);
1631         hint = *this_cpu_ptr(hyperv_pcpu_input_arg);
1632         if (!hint) {
1633                 local_irq_restore(flags);
1634                 return -ENOSPC;
1635         }
1636
1637         hint->type = HV_EXT_MEMORY_HEAT_HINT_TYPE_COLD_DISCARD;
1638         hint->reserved = 0;
1639         for_each_sg(sgl, sg, nents, i) {
1640                 union hv_gpa_page_range *range;
1641
1642                 range = &hint->ranges[i];
1643                 range->address_space = 0;
1644                 order = get_order(sg->length);
1645                 /*
1646                  * Hyper-V expects the additional_pages field in the units
1647                  * of one of these 3 sizes, 4Kbytes, 2Mbytes or 1Gbytes.
1648                  * This is dictated by the values of the fields page.largesize
1649                  * and page_size.
1650                  * This code however, only uses 4Kbytes and 2Mbytes units
1651                  * and not 1Gbytes unit.
1652                  */
1653
1654                 /* page reporting for pages 2MB or higher */
1655                 if (order >= HV_LARGE_REPORTING_ORDER ) {
1656                         range->page.largepage = 1;
1657                         range->page_size = HV_GPA_PAGE_RANGE_PAGE_SIZE_2MB;
1658                         range->base_large_pfn = page_to_hvpfn(
1659                                         sg_page(sg)) >> HV_LARGE_REPORTING_ORDER;
1660                         range->page.additional_pages =
1661                                 (sg->length / HV_LARGE_REPORTING_LEN) - 1;
1662                 } else {
1663                         /* Page reporting for pages below 2MB */
1664                         range->page.basepfn = page_to_hvpfn(sg_page(sg));
1665                         range->page.largepage = false;
1666                         range->page.additional_pages =
1667                                 (sg->length / HV_HYP_PAGE_SIZE) - 1;
1668                 }
1669
1670         }
1671
1672         status = hv_do_rep_hypercall(HV_EXT_CALL_MEMORY_HEAT_HINT, nents, 0,
1673                                      hint, NULL);
1674         local_irq_restore(flags);
1675         if (!hv_result_success(status)) {
1676
1677                 pr_err("Cold memory discard hypercall failed with status %llx\n",
1678                                 status);
1679                 if (hv_hypercall_multi_failure > 0)
1680                         hv_hypercall_multi_failure++;
1681
1682                 if (hv_result(status) == HV_STATUS_INVALID_PARAMETER) {
1683                         pr_err("Underlying Hyper-V does not support order less than 9. Hypercall failed\n");
1684                         pr_err("Defaulting to page_reporting_order %d\n",
1685                                         pageblock_order);
1686                         page_reporting_order = pageblock_order;
1687                         hv_hypercall_multi_failure++;
1688                         return -EINVAL;
1689                 }
1690
1691                 return -EINVAL;
1692         }
1693
1694         return 0;
1695 }
1696
1697 static void enable_page_reporting(void)
1698 {
1699         int ret;
1700
1701         if (!hv_query_ext_cap(HV_EXT_CAPABILITY_MEMORY_COLD_DISCARD_HINT)) {
1702                 pr_debug("Cold memory discard hint not supported by Hyper-V\n");
1703                 return;
1704         }
1705
1706         BUILD_BUG_ON(PAGE_REPORTING_CAPACITY > HV_MEMORY_HINT_MAX_GPA_PAGE_RANGES);
1707         dm_device.pr_dev_info.report = hv_free_page_report;
1708         /*
1709          * We let the page_reporting_order parameter decide the order
1710          * in the page_reporting code
1711          */
1712         dm_device.pr_dev_info.order = 0;
1713         ret = page_reporting_register(&dm_device.pr_dev_info);
1714         if (ret < 0) {
1715                 dm_device.pr_dev_info.report = NULL;
1716                 pr_err("Failed to enable cold memory discard: %d\n", ret);
1717         } else {
1718                 pr_info("Cold memory discard hint enabled with order %d\n",
1719                                 page_reporting_order);
1720         }
1721 }
1722
1723 static void disable_page_reporting(void)
1724 {
1725         if (dm_device.pr_dev_info.report) {
1726                 page_reporting_unregister(&dm_device.pr_dev_info);
1727                 dm_device.pr_dev_info.report = NULL;
1728         }
1729 }
1730
1731 static int ballooning_enabled(void)
1732 {
1733         /*
1734          * Disable ballooning if the page size is not 4k (HV_HYP_PAGE_SIZE),
1735          * since currently it's unclear to us whether an unballoon request can
1736          * make sure all page ranges are guest page size aligned.
1737          */
1738         if (PAGE_SIZE != HV_HYP_PAGE_SIZE) {
1739                 pr_info("Ballooning disabled because page size is not 4096 bytes\n");
1740                 return 0;
1741         }
1742
1743         return 1;
1744 }
1745
1746 static int hot_add_enabled(void)
1747 {
1748         /*
1749          * Disable hot add on ARM64, because we currently rely on
1750          * memory_add_physaddr_to_nid() to get a node id of a hot add range,
1751          * however ARM64's memory_add_physaddr_to_nid() always return 0 and
1752          * DM_MEM_HOT_ADD_REQUEST doesn't have the NUMA node information for
1753          * add_memory().
1754          */
1755         if (IS_ENABLED(CONFIG_ARM64)) {
1756                 pr_info("Memory hot add disabled on ARM64\n");
1757                 return 0;
1758         }
1759
1760         return 1;
1761 }
1762
1763 static int balloon_connect_vsp(struct hv_device *dev)
1764 {
1765         struct dm_version_request version_req;
1766         struct dm_capabilities cap_msg;
1767         unsigned long t;
1768         int ret;
1769
1770         /*
1771          * max_pkt_size should be large enough for one vmbus packet header plus
1772          * our receive buffer size. Hyper-V sends messages up to
1773          * HV_HYP_PAGE_SIZE bytes long on balloon channel.
1774          */
1775         dev->channel->max_pkt_size = HV_HYP_PAGE_SIZE * 2;
1776
1777         ret = vmbus_open(dev->channel, dm_ring_size, dm_ring_size, NULL, 0,
1778                          balloon_onchannelcallback, dev);
1779         if (ret)
1780                 return ret;
1781
1782         /*
1783          * Initiate the hand shake with the host and negotiate
1784          * a version that the host can support. We start with the
1785          * highest version number and go down if the host cannot
1786          * support it.
1787          */
1788         memset(&version_req, 0, sizeof(struct dm_version_request));
1789         version_req.hdr.type = DM_VERSION_REQUEST;
1790         version_req.hdr.size = sizeof(struct dm_version_request);
1791         version_req.hdr.trans_id = atomic_inc_return(&trans_id);
1792         version_req.version.version = DYNMEM_PROTOCOL_VERSION_WIN10;
1793         version_req.is_last_attempt = 0;
1794         dm_device.version = version_req.version.version;
1795
1796         ret = vmbus_sendpacket(dev->channel, &version_req,
1797                                sizeof(struct dm_version_request),
1798                                (unsigned long)NULL, VM_PKT_DATA_INBAND, 0);
1799         if (ret)
1800                 goto out;
1801
1802         t = wait_for_completion_timeout(&dm_device.host_event, 5*HZ);
1803         if (t == 0) {
1804                 ret = -ETIMEDOUT;
1805                 goto out;
1806         }
1807
1808         /*
1809          * If we could not negotiate a compatible version with the host
1810          * fail the probe function.
1811          */
1812         if (dm_device.state == DM_INIT_ERROR) {
1813                 ret = -EPROTO;
1814                 goto out;
1815         }
1816
1817         pr_info("Using Dynamic Memory protocol version %u.%u\n",
1818                 DYNMEM_MAJOR_VERSION(dm_device.version),
1819                 DYNMEM_MINOR_VERSION(dm_device.version));
1820
1821         /*
1822          * Now submit our capabilities to the host.
1823          */
1824         memset(&cap_msg, 0, sizeof(struct dm_capabilities));
1825         cap_msg.hdr.type = DM_CAPABILITIES_REPORT;
1826         cap_msg.hdr.size = sizeof(struct dm_capabilities);
1827         cap_msg.hdr.trans_id = atomic_inc_return(&trans_id);
1828
1829         /*
1830          * When hibernation (i.e. virtual ACPI S4 state) is enabled, the host
1831          * currently still requires the bits to be set, so we have to add code
1832          * to fail the host's hot-add and balloon up/down requests, if any.
1833          */
1834         cap_msg.caps.cap_bits.balloon = ballooning_enabled();
1835         cap_msg.caps.cap_bits.hot_add = hot_add_enabled();
1836
1837         /*
1838          * Specify our alignment requirements as it relates
1839          * memory hot-add. Specify 128MB alignment.
1840          */
1841         cap_msg.caps.cap_bits.hot_add_alignment = 7;
1842
1843         /*
1844          * Currently the host does not use these
1845          * values and we set them to what is done in the
1846          * Windows driver.
1847          */
1848         cap_msg.min_page_cnt = 0;
1849         cap_msg.max_page_number = -1;
1850
1851         ret = vmbus_sendpacket(dev->channel, &cap_msg,
1852                                sizeof(struct dm_capabilities),
1853                                (unsigned long)NULL, VM_PKT_DATA_INBAND, 0);
1854         if (ret)
1855                 goto out;
1856
1857         t = wait_for_completion_timeout(&dm_device.host_event, 5*HZ);
1858         if (t == 0) {
1859                 ret = -ETIMEDOUT;
1860                 goto out;
1861         }
1862
1863         /*
1864          * If the host does not like our capabilities,
1865          * fail the probe function.
1866          */
1867         if (dm_device.state == DM_INIT_ERROR) {
1868                 ret = -EPROTO;
1869                 goto out;
1870         }
1871
1872         return 0;
1873 out:
1874         vmbus_close(dev->channel);
1875         return ret;
1876 }
1877
1878 /*
1879  * DEBUGFS Interface
1880  */
1881 #ifdef CONFIG_DEBUG_FS
1882
1883 /**
1884  * hv_balloon_debug_show - shows statistics of balloon operations.
1885  * @f: pointer to the &struct seq_file.
1886  * @offset: ignored.
1887  *
1888  * Provides the statistics that can be accessed in hv-balloon in the debugfs.
1889  *
1890  * Return: zero on success or an error code.
1891  */
1892 static int hv_balloon_debug_show(struct seq_file *f, void *offset)
1893 {
1894         struct hv_dynmem_device *dm = f->private;
1895         char *sname;
1896
1897         seq_printf(f, "%-22s: %u.%u\n", "host_version",
1898                                 DYNMEM_MAJOR_VERSION(dm->version),
1899                                 DYNMEM_MINOR_VERSION(dm->version));
1900
1901         seq_printf(f, "%-22s:", "capabilities");
1902         if (ballooning_enabled())
1903                 seq_puts(f, " enabled");
1904
1905         if (hot_add_enabled())
1906                 seq_puts(f, " hot_add");
1907
1908         seq_puts(f, "\n");
1909
1910         seq_printf(f, "%-22s: %u", "state", dm->state);
1911         switch (dm->state) {
1912         case DM_INITIALIZING:
1913                         sname = "Initializing";
1914                         break;
1915         case DM_INITIALIZED:
1916                         sname = "Initialized";
1917                         break;
1918         case DM_BALLOON_UP:
1919                         sname = "Balloon Up";
1920                         break;
1921         case DM_BALLOON_DOWN:
1922                         sname = "Balloon Down";
1923                         break;
1924         case DM_HOT_ADD:
1925                         sname = "Hot Add";
1926                         break;
1927         case DM_INIT_ERROR:
1928                         sname = "Error";
1929                         break;
1930         default:
1931                         sname = "Unknown";
1932         }
1933         seq_printf(f, " (%s)\n", sname);
1934
1935         /* HV Page Size */
1936         seq_printf(f, "%-22s: %ld\n", "page_size", HV_HYP_PAGE_SIZE);
1937
1938         /* Pages added with hot_add */
1939         seq_printf(f, "%-22s: %u\n", "pages_added", dm->num_pages_added);
1940
1941         /* pages that are "onlined"/used from pages_added */
1942         seq_printf(f, "%-22s: %u\n", "pages_onlined", dm->num_pages_onlined);
1943
1944         /* pages we have given back to host */
1945         seq_printf(f, "%-22s: %u\n", "pages_ballooned", dm->num_pages_ballooned);
1946
1947         seq_printf(f, "%-22s: %lu\n", "total_pages_committed",
1948                                 get_pages_committed(dm));
1949
1950         seq_printf(f, "%-22s: %llu\n", "max_dynamic_page_count",
1951                                 dm->max_dynamic_page_count);
1952
1953         return 0;
1954 }
1955
1956 DEFINE_SHOW_ATTRIBUTE(hv_balloon_debug);
1957
1958 static void  hv_balloon_debugfs_init(struct hv_dynmem_device *b)
1959 {
1960         debugfs_create_file("hv-balloon", 0444, NULL, b,
1961                         &hv_balloon_debug_fops);
1962 }
1963
1964 static void  hv_balloon_debugfs_exit(struct hv_dynmem_device *b)
1965 {
1966         debugfs_lookup_and_remove("hv-balloon", NULL);
1967 }
1968
1969 #else
1970
1971 static inline void hv_balloon_debugfs_init(struct hv_dynmem_device  *b)
1972 {
1973 }
1974
1975 static inline void hv_balloon_debugfs_exit(struct hv_dynmem_device *b)
1976 {
1977 }
1978
1979 #endif  /* CONFIG_DEBUG_FS */
1980
1981 static int balloon_probe(struct hv_device *dev,
1982                          const struct hv_vmbus_device_id *dev_id)
1983 {
1984         int ret;
1985
1986         allow_hibernation = hv_is_hibernation_supported();
1987         if (allow_hibernation)
1988                 hot_add = false;
1989
1990 #ifdef CONFIG_MEMORY_HOTPLUG
1991         do_hot_add = hot_add;
1992 #else
1993         do_hot_add = false;
1994 #endif
1995         dm_device.dev = dev;
1996         dm_device.state = DM_INITIALIZING;
1997         dm_device.next_version = DYNMEM_PROTOCOL_VERSION_WIN8;
1998         init_completion(&dm_device.host_event);
1999         init_completion(&dm_device.config_event);
2000         INIT_LIST_HEAD(&dm_device.ha_region_list);
2001         spin_lock_init(&dm_device.ha_lock);
2002         INIT_WORK(&dm_device.balloon_wrk.wrk, balloon_up);
2003         INIT_WORK(&dm_device.ha_wrk.wrk, hot_add_req);
2004         dm_device.host_specified_ha_region = false;
2005
2006 #ifdef CONFIG_MEMORY_HOTPLUG
2007         set_online_page_callback(&hv_online_page);
2008         init_completion(&dm_device.ol_waitevent);
2009         register_memory_notifier(&hv_memory_nb);
2010 #endif
2011
2012         hv_set_drvdata(dev, &dm_device);
2013
2014         ret = balloon_connect_vsp(dev);
2015         if (ret != 0)
2016                 goto connect_error;
2017
2018         enable_page_reporting();
2019         dm_device.state = DM_INITIALIZED;
2020
2021         dm_device.thread =
2022                  kthread_run(dm_thread_func, &dm_device, "hv_balloon");
2023         if (IS_ERR(dm_device.thread)) {
2024                 ret = PTR_ERR(dm_device.thread);
2025                 goto probe_error;
2026         }
2027
2028         hv_balloon_debugfs_init(&dm_device);
2029
2030         return 0;
2031
2032 probe_error:
2033         dm_device.state = DM_INIT_ERROR;
2034         dm_device.thread  = NULL;
2035         disable_page_reporting();
2036         vmbus_close(dev->channel);
2037 connect_error:
2038 #ifdef CONFIG_MEMORY_HOTPLUG
2039         unregister_memory_notifier(&hv_memory_nb);
2040         restore_online_page_callback(&hv_online_page);
2041 #endif
2042         return ret;
2043 }
2044
2045 static void balloon_remove(struct hv_device *dev)
2046 {
2047         struct hv_dynmem_device *dm = hv_get_drvdata(dev);
2048         struct hv_hotadd_state *has, *tmp;
2049         struct hv_hotadd_gap *gap, *tmp_gap;
2050         unsigned long flags;
2051
2052         if (dm->num_pages_ballooned != 0)
2053                 pr_warn("Ballooned pages: %d\n", dm->num_pages_ballooned);
2054
2055         hv_balloon_debugfs_exit(dm);
2056
2057         cancel_work_sync(&dm->balloon_wrk.wrk);
2058         cancel_work_sync(&dm->ha_wrk.wrk);
2059
2060         kthread_stop(dm->thread);
2061
2062         /*
2063          * This is to handle the case when balloon_resume()
2064          * call has failed and some cleanup has been done as
2065          * a part of the error handling.
2066          */
2067         if (dm_device.state != DM_INIT_ERROR) {
2068                 disable_page_reporting();
2069                 vmbus_close(dev->channel);
2070 #ifdef CONFIG_MEMORY_HOTPLUG
2071                 unregister_memory_notifier(&hv_memory_nb);
2072                 restore_online_page_callback(&hv_online_page);
2073 #endif
2074         }
2075
2076         spin_lock_irqsave(&dm_device.ha_lock, flags);
2077         list_for_each_entry_safe(has, tmp, &dm->ha_region_list, list) {
2078                 list_for_each_entry_safe(gap, tmp_gap, &has->gap_list, list) {
2079                         list_del(&gap->list);
2080                         kfree(gap);
2081                 }
2082                 list_del(&has->list);
2083                 kfree(has);
2084         }
2085         spin_unlock_irqrestore(&dm_device.ha_lock, flags);
2086 }
2087
2088 static int balloon_suspend(struct hv_device *hv_dev)
2089 {
2090         struct hv_dynmem_device *dm = hv_get_drvdata(hv_dev);
2091
2092         tasklet_disable(&hv_dev->channel->callback_event);
2093
2094         cancel_work_sync(&dm->balloon_wrk.wrk);
2095         cancel_work_sync(&dm->ha_wrk.wrk);
2096
2097         if (dm->thread) {
2098                 kthread_stop(dm->thread);
2099                 dm->thread = NULL;
2100                 vmbus_close(hv_dev->channel);
2101         }
2102
2103         tasklet_enable(&hv_dev->channel->callback_event);
2104
2105         return 0;
2106
2107 }
2108
2109 static int balloon_resume(struct hv_device *dev)
2110 {
2111         int ret;
2112
2113         dm_device.state = DM_INITIALIZING;
2114
2115         ret = balloon_connect_vsp(dev);
2116
2117         if (ret != 0)
2118                 goto out;
2119
2120         dm_device.thread =
2121                  kthread_run(dm_thread_func, &dm_device, "hv_balloon");
2122         if (IS_ERR(dm_device.thread)) {
2123                 ret = PTR_ERR(dm_device.thread);
2124                 dm_device.thread = NULL;
2125                 goto close_channel;
2126         }
2127
2128         dm_device.state = DM_INITIALIZED;
2129         return 0;
2130 close_channel:
2131         vmbus_close(dev->channel);
2132 out:
2133         dm_device.state = DM_INIT_ERROR;
2134         disable_page_reporting();
2135 #ifdef CONFIG_MEMORY_HOTPLUG
2136         unregister_memory_notifier(&hv_memory_nb);
2137         restore_online_page_callback(&hv_online_page);
2138 #endif
2139         return ret;
2140 }
2141
2142 static const struct hv_vmbus_device_id id_table[] = {
2143         /* Dynamic Memory Class ID */
2144         /* 525074DC-8985-46e2-8057-A307DC18A502 */
2145         { HV_DM_GUID, },
2146         { },
2147 };
2148
2149 MODULE_DEVICE_TABLE(vmbus, id_table);
2150
2151 static  struct hv_driver balloon_drv = {
2152         .name = "hv_balloon",
2153         .id_table = id_table,
2154         .probe =  balloon_probe,
2155         .remove =  balloon_remove,
2156         .suspend = balloon_suspend,
2157         .resume = balloon_resume,
2158         .driver = {
2159                 .probe_type = PROBE_PREFER_ASYNCHRONOUS,
2160         },
2161 };
2162
2163 static int __init init_balloon_drv(void)
2164 {
2165
2166         return vmbus_driver_register(&balloon_drv);
2167 }
2168
2169 module_init(init_balloon_drv);
2170
2171 MODULE_DESCRIPTION("Hyper-V Balloon");
2172 MODULE_LICENSE("GPL");