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                 if (!PageOffline(pg))
 686                         __SetPageOffline(pg);
 687                 return;
 688         }
 689         if (PageOffline(pg))
 690                 __ClearPageOffline(pg);
 691
 692         /* This frame is currently backed; online the page. */
 693         __online_page_set_limits(pg);
 694         __online_page_increment_counters(pg);
 695         __online_page_free(pg);
 696
 697         lockdep_assert_held(&dm_device.ha_lock);
 698         dm_device.num_pages_onlined++;
 699 }
 700
 701 static void hv_bring_pgs_online(struct hv_hotadd_state *has,
 702                                 unsigned long start_pfn, unsigned long size)
 703 {
 704         int i;
 705
 706         pr_debug("Online %lu pages starting at pfn 0x%lx\n", size, start_pfn);
 707         for (i = 0; i < size; i++)
 708                 hv_page_online_one(has, pfn_to_page(start_pfn + i));
 709 }
 710
 711 static void hv_mem_hot_add(unsigned long start, unsigned long size,
 712                                 unsigned long pfn_count,
 713                                 struct hv_hotadd_state *has)
 714 {
 715         int ret = 0;
 716         int i, nid;
 717         unsigned long start_pfn;
 718         unsigned long processed_pfn;
 719         unsigned long total_pfn = pfn_count;
 720         unsigned long flags;
 721
 722         for (i = 0; i < (size/HA_CHUNK); i++) {
 723                 start_pfn = start + (i * HA_CHUNK);
 724
 725                 spin_lock_irqsave(&dm_device.ha_lock, flags);
 726                 has->ha_end_pfn +=  HA_CHUNK;
 727
 728                 if (total_pfn > HA_CHUNK) {
 729                         processed_pfn = HA_CHUNK;
 730                         total_pfn -= HA_CHUNK;
 731                 } else {
 732                         processed_pfn = total_pfn;
 733                         total_pfn = 0;
 734                 }
 735
 736                 has->covered_end_pfn +=  processed_pfn;
 737                 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 738
 739                 init_completion(&dm_device.ol_waitevent);
 740                 dm_device.ha_waiting = !memhp_auto_online;
 741
 742                 nid = memory_add_physaddr_to_nid(PFN_PHYS(start_pfn));
 743                 ret = add_memory(nid, PFN_PHYS((start_pfn)),
 744                                 (HA_CHUNK << PAGE_SHIFT));
 745
 746                 if (ret) {
 747                         pr_err("hot_add memory failed error is %d\n", ret);
 748                         if (ret == -EEXIST) {
 749                                 /*
 750                                  * This error indicates that the error
 751                                  * is not a transient failure. This is the
 752                                  * case where the guest's physical address map
 753                                  * precludes hot adding memory. Stop all further
 754                                  * memory hot-add.
 755                                  */
 756                                 do_hot_add = false;
 757                         }
 758                         spin_lock_irqsave(&dm_device.ha_lock, flags);
 759                         has->ha_end_pfn -= HA_CHUNK;
 760                         has->covered_end_pfn -=  processed_pfn;
 761                         spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 762                         break;
 763                 }
 764
 765                 /*
 766                  * Wait for the memory block to be onlined when memory onlining
 767                  * is done outside of kernel (memhp_auto_online). Since the hot
 768                  * add has succeeded, it is ok to proceed even if the pages in
 769                  * the hot added region have not been "onlined" within the
 770                  * allowed time.
 771                  */
 772                 if (dm_device.ha_waiting)
 773                         wait_for_completion_timeout(&dm_device.ol_waitevent,
 774                                                     5*HZ);
 775                 post_status(&dm_device);
 776         }
 777 }
 778
 779 static void hv_online_page(struct page *pg, unsigned int order)
 780 {
 781         struct hv_hotadd_state *has;
 782         unsigned long flags;
 783         unsigned long pfn = page_to_pfn(pg);
 784
 785         spin_lock_irqsave(&dm_device.ha_lock, flags);
 786         list_for_each_entry(has, &dm_device.ha_region_list, list) {
 787                 /* The page belongs to a different HAS. */
 788                 if ((pfn < has->start_pfn) ||
 789                                 (pfn + (1UL << order) > has->end_pfn))
 790                         continue;
 791
 792                 hv_bring_pgs_online(has, pfn, 1UL << order);
 793                 break;
 794         }
 795         spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 796 }
 797
 798 static int pfn_covered(unsigned long start_pfn, unsigned long pfn_cnt)
 799 {
 800         struct hv_hotadd_state *has;
 801         struct hv_hotadd_gap *gap;
 802         unsigned long residual, new_inc;
 803         int ret = 0;
 804         unsigned long flags;
 805
 806         spin_lock_irqsave(&dm_device.ha_lock, flags);
 807         list_for_each_entry(has, &dm_device.ha_region_list, list) {
 808                 /*
 809                  * If the pfn range we are dealing with is not in the current
 810                  * "hot add block", move on.
 811                  */
 812                 if (start_pfn < has->start_pfn || start_pfn >= has->end_pfn)
 813                         continue;
 814
 815                 /*
 816                  * If the current start pfn is not where the covered_end
 817                  * is, create a gap and update covered_end_pfn.
 818                  */
 819                 if (has->covered_end_pfn != start_pfn) {
 820                         gap = kzalloc(sizeof(struct hv_hotadd_gap), GFP_ATOMIC);
 821                         if (!gap) {
 822                                 ret = -ENOMEM;
 823                                 break;
 824                         }
 825
 826                         INIT_LIST_HEAD(&gap->list);
 827                         gap->start_pfn = has->covered_end_pfn;
 828                         gap->end_pfn = start_pfn;
 829                         list_add_tail(&gap->list, &has->gap_list);
 830
 831                         has->covered_end_pfn = start_pfn;
 832                 }
 833
 834                 /*
 835                  * If the current hot add-request extends beyond
 836                  * our current limit; extend it.
 837                  */
 838                 if ((start_pfn + pfn_cnt) > has->end_pfn) {
 839                         residual = (start_pfn + pfn_cnt - has->end_pfn);
 840                         /*
 841                          * Extend the region by multiples of HA_CHUNK.
 842                          */
 843                         new_inc = (residual / HA_CHUNK) * HA_CHUNK;
 844                         if (residual % HA_CHUNK)
 845                                 new_inc += HA_CHUNK;
 846
 847                         has->end_pfn += new_inc;
 848                 }
 849
 850                 ret = 1;
 851                 break;
 852         }
 853         spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 854
 855         return ret;
 856 }
 857
 858 static unsigned long handle_pg_range(unsigned long pg_start,
 859                                         unsigned long pg_count)
 860 {
 861         unsigned long start_pfn = pg_start;
 862         unsigned long pfn_cnt = pg_count;
 863         unsigned long size;
 864         struct hv_hotadd_state *has;
 865         unsigned long pgs_ol = 0;
 866         unsigned long old_covered_state;
 867         unsigned long res = 0, flags;
 868
 869         pr_debug("Hot adding %lu pages starting at pfn 0x%lx.\n", pg_count,
 870                 pg_start);
 871
 872         spin_lock_irqsave(&dm_device.ha_lock, flags);
 873         list_for_each_entry(has, &dm_device.ha_region_list, list) {
 874                 /*
 875                  * If the pfn range we are dealing with is not in the current
 876                  * "hot add block", move on.
 877                  */
 878                 if (start_pfn < has->start_pfn || start_pfn >= has->end_pfn)
 879                         continue;
 880
 881                 old_covered_state = has->covered_end_pfn;
 882
 883                 if (start_pfn < has->ha_end_pfn) {
 884                         /*
 885                          * This is the case where we are backing pages
 886                          * in an already hot added region. Bring
 887                          * these pages online first.
 888                          */
 889                         pgs_ol = has->ha_end_pfn - start_pfn;
 890                         if (pgs_ol > pfn_cnt)
 891                                 pgs_ol = pfn_cnt;
 892
 893                         has->covered_end_pfn +=  pgs_ol;
 894                         pfn_cnt -= pgs_ol;
 895                         /*
 896                          * Check if the corresponding memory block is already
 897                          * online. It is possible to observe struct pages still
 898                          * being uninitialized here so check section instead.
 899                          * In case the section is online we need to bring the
 900                          * rest of pfns (which were not backed previously)
 901                          * online too.
 902                          */
 903                         if (start_pfn > has->start_pfn &&
 904                             online_section_nr(pfn_to_section_nr(start_pfn)))
 905                                 hv_bring_pgs_online(has, start_pfn, pgs_ol);
 906
 907                 }
 908
 909                 if ((has->ha_end_pfn < has->end_pfn) && (pfn_cnt > 0)) {
 910                         /*
 911                          * We have some residual hot add range
 912                          * that needs to be hot added; hot add
 913                          * it now. Hot add a multiple of
 914                          * of HA_CHUNK that fully covers the pages
 915                          * we have.
 916                          */
 917                         size = (has->end_pfn - has->ha_end_pfn);
 918                         if (pfn_cnt <= size) {
 919                                 size = ((pfn_cnt / HA_CHUNK) * HA_CHUNK);
 920                                 if (pfn_cnt % HA_CHUNK)
 921                                         size += HA_CHUNK;
 922                         } else {
 923                                 pfn_cnt = size;
 924                         }
 925                         spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 926                         hv_mem_hot_add(has->ha_end_pfn, size, pfn_cnt, has);
 927                         spin_lock_irqsave(&dm_device.ha_lock, flags);
 928                 }
 929                 /*
 930                  * If we managed to online any pages that were given to us,
 931                  * we declare success.
 932                  */
 933                 res = has->covered_end_pfn - old_covered_state;
 934                 break;
 935         }
 936         spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 937
 938         return res;
 939 }
 940
 941 static unsigned long process_hot_add(unsigned long pg_start,
 942                                         unsigned long pfn_cnt,
 943                                         unsigned long rg_start,
 944                                         unsigned long rg_size)
 945 {
 946         struct hv_hotadd_state *ha_region = NULL;
 947         int covered;
 948         unsigned long flags;
 949
 950         if (pfn_cnt == 0)
 951                 return 0;
 952
 953         if (!dm_device.host_specified_ha_region) {
 954                 covered = pfn_covered(pg_start, pfn_cnt);
 955                 if (covered < 0)
 956                         return 0;
 957
 958                 if (covered)
 959                         goto do_pg_range;
 960         }
 961
 962         /*
 963          * If the host has specified a hot-add range; deal with it first.
 964          */
 965
 966         if (rg_size != 0) {
 967                 ha_region = kzalloc(sizeof(struct hv_hotadd_state), GFP_KERNEL);
 968                 if (!ha_region)
 969                         return 0;
 970
 971                 INIT_LIST_HEAD(&ha_region->list);
 972                 INIT_LIST_HEAD(&ha_region->gap_list);
 973
 974                 ha_region->start_pfn = rg_start;
 975                 ha_region->ha_end_pfn = rg_start;
 976                 ha_region->covered_start_pfn = pg_start;
 977                 ha_region->covered_end_pfn = pg_start;
 978                 ha_region->end_pfn = rg_start + rg_size;
 979
 980                 spin_lock_irqsave(&dm_device.ha_lock, flags);
 981                 list_add_tail(&ha_region->list, &dm_device.ha_region_list);
 982                 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 983         }
 984
 985 do_pg_range:
 986         /*
 987          * Process the page range specified; bringing them
 988          * online if possible.
 989          */
 990         return handle_pg_range(pg_start, pfn_cnt);
 991 }
 992
 993 #endif
 994
 995 static void hot_add_req(struct work_struct *dummy)
 996 {
 997         struct dm_hot_add_response resp;
 998 #ifdef CONFIG_MEMORY_HOTPLUG
 999         unsigned long pg_start, pfn_cnt;
1000         unsigned long rg_start, rg_sz;
1001 #endif
1002         struct hv_dynmem_device *dm = &dm_device;
1003
1004         memset(&resp, 0, sizeof(struct dm_hot_add_response));
1005         resp.hdr.type = DM_MEM_HOT_ADD_RESPONSE;
1006         resp.hdr.size = sizeof(struct dm_hot_add_response);
1007
1008 #ifdef CONFIG_MEMORY_HOTPLUG
1009         pg_start = dm->ha_wrk.ha_page_range.finfo.start_page;
1010         pfn_cnt = dm->ha_wrk.ha_page_range.finfo.page_cnt;
1011
1012         rg_start = dm->ha_wrk.ha_region_range.finfo.start_page;
1013         rg_sz = dm->ha_wrk.ha_region_range.finfo.page_cnt;
1014
1015         if ((rg_start == 0) && (!dm->host_specified_ha_region)) {
1016                 unsigned long region_size;
1017                 unsigned long region_start;
1018
1019                 /*
1020                  * The host has not specified the hot-add region.
1021                  * Based on the hot-add page range being specified,
1022                  * compute a hot-add region that can cover the pages
1023                  * that need to be hot-added while ensuring the alignment
1024                  * and size requirements of Linux as it relates to hot-add.
1025                  */
1026                 region_start = pg_start;
1027                 region_size = (pfn_cnt / HA_CHUNK) * HA_CHUNK;
1028                 if (pfn_cnt % HA_CHUNK)
1029                         region_size += HA_CHUNK;
1030
1031                 region_start = (pg_start / HA_CHUNK) * HA_CHUNK;
1032
1033                 rg_start = region_start;
1034                 rg_sz = region_size;
1035         }
1036
1037         if (do_hot_add)
1038                 resp.page_count = process_hot_add(pg_start, pfn_cnt,
1039                                                 rg_start, rg_sz);
1040
1041         dm->num_pages_added += resp.page_count;
1042 #endif
1043         /*
1044          * The result field of the response structure has the
1045          * following semantics:
1046          *
1047          * 1. If all or some pages hot-added: Guest should return success.
1048          *
1049          * 2. If no pages could be hot-added:
1050          *
1051          * If the guest returns success, then the host
1052          * will not attempt any further hot-add operations. This
1053          * signifies a permanent failure.
1054          *
1055          * If the guest returns failure, then this failure will be
1056          * treated as a transient failure and the host may retry the
1057          * hot-add operation after some delay.
1058          */
1059         if (resp.page_count > 0)
1060                 resp.result = 1;
1061         else if (!do_hot_add)
1062                 resp.result = 1;
1063         else
1064                 resp.result = 0;
1065
1066         if (!do_hot_add || (resp.page_count == 0))
1067                 pr_err("Memory hot add failed\n");
1068
1069         dm->state = DM_INITIALIZED;
1070         resp.hdr.trans_id = atomic_inc_return(&trans_id);
1071         vmbus_sendpacket(dm->dev->channel, &resp,
1072                         sizeof(struct dm_hot_add_response),
1073                         (unsigned long)NULL,
1074                         VM_PKT_DATA_INBAND, 0);
1075 }
1076
1077 static void process_info(struct hv_dynmem_device *dm, struct dm_info_msg *msg)
1078 {
1079         struct dm_info_header *info_hdr;
1080
1081         info_hdr = (struct dm_info_header *)msg->info;
1082
1083         switch (info_hdr->type) {
1084         case INFO_TYPE_MAX_PAGE_CNT:
1085                 if (info_hdr->data_size == sizeof(__u64)) {
1086                         __u64 *max_page_count = (__u64 *)&info_hdr[1];
1087
1088                         pr_info("Max. dynamic memory size: %llu MB\n",
1089                                 (*max_page_count) >> (20 - PAGE_SHIFT));
1090                 }
1091
1092                 break;
1093         default:
1094                 pr_warn("Received Unknown type: %d\n", info_hdr->type);
1095         }
1096 }
1097
1098 static unsigned long compute_balloon_floor(void)
1099 {
1100         unsigned long min_pages;
1101         unsigned long nr_pages = totalram_pages();
1102 #define MB2PAGES(mb) ((mb) << (20 - PAGE_SHIFT))
1103         /* Simple continuous piecewiese linear function:
1104          *  max MiB -> min MiB  gradient
1105          *       0         0
1106          *      16        16
1107          *      32        24
1108          *     128        72    (1/2)
1109          *     512       168    (1/4)
1110          *    2048       360    (1/8)
1111          *    8192       744    (1/16)
1112          *   32768      1512    (1/32)
1113          */
1114         if (nr_pages < MB2PAGES(128))
1115                 min_pages = MB2PAGES(8) + (nr_pages >> 1);
1116         else if (nr_pages < MB2PAGES(512))
1117                 min_pages = MB2PAGES(40) + (nr_pages >> 2);
1118         else if (nr_pages < MB2PAGES(2048))
1119                 min_pages = MB2PAGES(104) + (nr_pages >> 3);
1120         else if (nr_pages < MB2PAGES(8192))
1121                 min_pages = MB2PAGES(232) + (nr_pages >> 4);
1122         else
1123                 min_pages = MB2PAGES(488) + (nr_pages >> 5);
1124 #undef MB2PAGES
1125         return min_pages;
1126 }
1127
1128 /*
1129  * Post our status as it relates memory pressure to the
1130  * host. Host expects the guests to post this status
1131  * periodically at 1 second intervals.
1132  *
1133  * The metrics specified in this protocol are very Windows
1134  * specific and so we cook up numbers here to convey our memory
1135  * pressure.
1136  */
1137
1138 static void post_status(struct hv_dynmem_device *dm)
1139 {
1140         struct dm_status status;
1141         unsigned long now = jiffies;
1142         unsigned long last_post = last_post_time;
1143
1144         if (pressure_report_delay > 0) {
1145                 --pressure_report_delay;
1146                 return;
1147         }
1148
1149         if (!time_after(now, (last_post_time + HZ)))
1150                 return;
1151
1152         memset(&status, 0, sizeof(struct dm_status));
1153         status.hdr.type = DM_STATUS_REPORT;
1154         status.hdr.size = sizeof(struct dm_status);
1155         status.hdr.trans_id = atomic_inc_return(&trans_id);
1156
1157         /*
1158          * The host expects the guest to report free and committed memory.
1159          * Furthermore, the host expects the pressure information to include
1160          * the ballooned out pages. For a given amount of memory that we are
1161          * managing we need to compute a floor below which we should not
1162          * balloon. Compute this and add it to the pressure report.
1163          * We also need to report all offline pages (num_pages_added -
1164          * num_pages_onlined) as committed to the host, otherwise it can try
1165          * asking us to balloon them out.
1166          */
1167         status.num_avail = si_mem_available();
1168         status.num_committed = vm_memory_committed() +
1169                 dm->num_pages_ballooned +
1170                 (dm->num_pages_added > dm->num_pages_onlined ?
1171                  dm->num_pages_added - dm->num_pages_onlined : 0) +
1172                 compute_balloon_floor();
1173
1174         trace_balloon_status(status.num_avail, status.num_committed,
1175                              vm_memory_committed(), dm->num_pages_ballooned,
1176                              dm->num_pages_added, dm->num_pages_onlined);
1177         /*
1178          * If our transaction ID is no longer current, just don't
1179          * send the status. This can happen if we were interrupted
1180          * after we picked our transaction ID.
1181          */
1182         if (status.hdr.trans_id != atomic_read(&trans_id))
1183                 return;
1184
1185         /*
1186          * If the last post time that we sampled has changed,
1187          * we have raced, don't post the status.
1188          */
1189         if (last_post != last_post_time)
1190                 return;
1191
1192         last_post_time = jiffies;
1193         vmbus_sendpacket(dm->dev->channel, &status,
1194                                 sizeof(struct dm_status),
1195                                 (unsigned long)NULL,
1196                                 VM_PKT_DATA_INBAND, 0);
1197
1198 }
1199
1200 static void free_balloon_pages(struct hv_dynmem_device *dm,
1201                          union dm_mem_page_range *range_array)
1202 {
1203         int num_pages = range_array->finfo.page_cnt;
1204         __u64 start_frame = range_array->finfo.start_page;
1205         struct page *pg;
1206         int i;
1207
1208         for (i = 0; i < num_pages; i++) {
1209                 pg = pfn_to_page(i + start_frame);
1210                 __ClearPageOffline(pg);
1211                 __free_page(pg);
1212                 dm->num_pages_ballooned--;
1213         }
1214 }
1215
1216
1217
1218 static unsigned int alloc_balloon_pages(struct hv_dynmem_device *dm,
1219                                         unsigned int num_pages,
1220                                         struct dm_balloon_response *bl_resp,
1221                                         int alloc_unit)
1222 {
1223         unsigned int i, j;
1224         struct page *pg;
1225
1226         if (num_pages < alloc_unit)
1227                 return 0;
1228
1229         for (i = 0; (i * alloc_unit) < num_pages; i++) {
1230                 if (bl_resp->hdr.size + sizeof(union dm_mem_page_range) >
1231                         PAGE_SIZE)
1232                         return i * alloc_unit;
1233
1234                 /*
1235                  * We execute this code in a thread context. Furthermore,
1236                  * we don't want the kernel to try too hard.
1237                  */
1238                 pg = alloc_pages(GFP_HIGHUSER | __GFP_NORETRY |
1239                                 __GFP_NOMEMALLOC | __GFP_NOWARN,
1240                                 get_order(alloc_unit << PAGE_SHIFT));
1241
1242                 if (!pg)
1243                         return i * alloc_unit;
1244
1245                 dm->num_pages_ballooned += alloc_unit;
1246
1247                 /*
1248                  * If we allocatted 2M pages; split them so we
1249                  * can free them in any order we get.
1250                  */
1251
1252                 if (alloc_unit != 1)
1253                         split_page(pg, get_order(alloc_unit << PAGE_SHIFT));
1254
1255                 /* mark all pages offline */
1256                 for (j = 0; j < (1 << get_order(alloc_unit << PAGE_SHIFT)); j++)
1257                         __SetPageOffline(pg + j);
1258
1259                 bl_resp->range_count++;
1260                 bl_resp->range_array[i].finfo.start_page =
1261                         page_to_pfn(pg);
1262                 bl_resp->range_array[i].finfo.page_cnt = alloc_unit;
1263                 bl_resp->hdr.size += sizeof(union dm_mem_page_range);
1264
1265         }
1266
1267         return num_pages;
1268 }
1269
1270 static void balloon_up(struct work_struct *dummy)
1271 {
1272         unsigned int num_pages = dm_device.balloon_wrk.num_pages;
1273         unsigned int num_ballooned = 0;
1274         struct dm_balloon_response *bl_resp;
1275         int alloc_unit;
1276         int ret;
1277         bool done = false;
1278         int i;
1279         long avail_pages;
1280         unsigned long floor;
1281
1282         /* The host balloons pages in 2M granularity. */
1283         WARN_ON_ONCE(num_pages % PAGES_IN_2M != 0);
1284
1285         /*
1286          * We will attempt 2M allocations. However, if we fail to
1287          * allocate 2M chunks, we will go back to 4k allocations.
1288          */
1289         alloc_unit = 512;
1290
1291         avail_pages = si_mem_available();
1292         floor = compute_balloon_floor();
1293
1294         /* Refuse to balloon below the floor, keep the 2M granularity. */
1295         if (avail_pages < num_pages || avail_pages - num_pages < floor) {
1296                 pr_warn("Balloon request will be partially fulfilled. %s\n",
1297                         avail_pages < num_pages ? "Not enough memory." :
1298                         "Balloon floor reached.");
1299
1300                 num_pages = avail_pages > floor ? (avail_pages - floor) : 0;
1301                 num_pages -= num_pages % PAGES_IN_2M;
1302         }
1303
1304         while (!done) {
1305                 bl_resp = (struct dm_balloon_response *)send_buffer;
1306                 memset(send_buffer, 0, PAGE_SIZE);
1307                 bl_resp->hdr.type = DM_BALLOON_RESPONSE;
1308                 bl_resp->hdr.size = sizeof(struct dm_balloon_response);
1309                 bl_resp->more_pages = 1;
1310
1311                 num_pages -= num_ballooned;
1312                 num_ballooned = alloc_balloon_pages(&dm_device, num_pages,
1313                                                     bl_resp, alloc_unit);
1314
1315                 if (alloc_unit != 1 && num_ballooned == 0) {
1316                         alloc_unit = 1;
1317                         continue;
1318                 }
1319
1320                 if (num_ballooned == 0 || num_ballooned == num_pages) {
1321                         pr_debug("Ballooned %u out of %u requested pages.\n",
1322                                 num_pages, dm_device.balloon_wrk.num_pages);
1323
1324                         bl_resp->more_pages = 0;
1325                         done = true;
1326                         dm_device.state = DM_INITIALIZED;
1327                 }
1328
1329                 /*
1330                  * We are pushing a lot of data through the channel;
1331                  * deal with transient failures caused because of the
1332                  * lack of space in the ring buffer.
1333                  */
1334
1335                 do {
1336                         bl_resp->hdr.trans_id = atomic_inc_return(&trans_id);
1337                         ret = vmbus_sendpacket(dm_device.dev->channel,
1338                                                 bl_resp,
1339                                                 bl_resp->hdr.size,
1340                                                 (unsigned long)NULL,
1341                                                 VM_PKT_DATA_INBAND, 0);
1342
1343                         if (ret == -EAGAIN)
1344                                 msleep(20);
1345                         post_status(&dm_device);
1346                 } while (ret == -EAGAIN);
1347
1348                 if (ret) {
1349                         /*
1350                          * Free up the memory we allocatted.
1351                          */
1352                         pr_err("Balloon response failed\n");
1353
1354                         for (i = 0; i < bl_resp->range_count; i++)
1355                                 free_balloon_pages(&dm_device,
1356                                                  &bl_resp->range_array[i]);
1357
1358                         done = true;
1359                 }
1360         }
1361
1362 }
1363
1364 static void balloon_down(struct hv_dynmem_device *dm,
1365                         struct dm_unballoon_request *req)
1366 {
1367         union dm_mem_page_range *range_array = req->range_array;
1368         int range_count = req->range_count;
1369         struct dm_unballoon_response resp;
1370         int i;
1371         unsigned int prev_pages_ballooned = dm->num_pages_ballooned;
1372
1373         for (i = 0; i < range_count; i++) {
1374                 free_balloon_pages(dm, &range_array[i]);
1375                 complete(&dm_device.config_event);
1376         }
1377
1378         pr_debug("Freed %u ballooned pages.\n",
1379                 prev_pages_ballooned - dm->num_pages_ballooned);
1380
1381         if (req->more_pages == 1)
1382                 return;
1383
1384         memset(&resp, 0, sizeof(struct dm_unballoon_response));
1385         resp.hdr.type = DM_UNBALLOON_RESPONSE;
1386         resp.hdr.trans_id = atomic_inc_return(&trans_id);
1387         resp.hdr.size = sizeof(struct dm_unballoon_response);
1388
1389         vmbus_sendpacket(dm_device.dev->channel, &resp,
1390                                 sizeof(struct dm_unballoon_response),
1391                                 (unsigned long)NULL,
1392                                 VM_PKT_DATA_INBAND, 0);
1393
1394         dm->state = DM_INITIALIZED;
1395 }
1396
1397 static void balloon_onchannelcallback(void *context);
1398
1399 static int dm_thread_func(void *dm_dev)
1400 {
1401         struct hv_dynmem_device *dm = dm_dev;
1402
1403         while (!kthread_should_stop()) {
1404                 wait_for_completion_interruptible_timeout(
1405                                                 &dm_device.config_event, 1*HZ);
1406                 /*
1407                  * The host expects us to post information on the memory
1408                  * pressure every second.
1409                  */
1410                 reinit_completion(&dm_device.config_event);
1411                 post_status(dm);
1412         }
1413
1414         return 0;
1415 }
1416
1417
1418 static void version_resp(struct hv_dynmem_device *dm,
1419                         struct dm_version_response *vresp)
1420 {
1421         struct dm_version_request version_req;
1422         int ret;
1423
1424         if (vresp->is_accepted) {
1425                 /*
1426                  * We are done; wakeup the
1427                  * context waiting for version
1428                  * negotiation.
1429                  */
1430                 complete(&dm->host_event);
1431                 return;
1432         }
1433         /*
1434          * If there are more versions to try, continue
1435          * with negotiations; if not
1436          * shutdown the service since we are not able
1437          * to negotiate a suitable version number
1438          * with the host.
1439          */
1440         if (dm->next_version == 0)
1441                 goto version_error;
1442
1443         memset(&version_req, 0, sizeof(struct dm_version_request));
1444         version_req.hdr.type = DM_VERSION_REQUEST;
1445         version_req.hdr.size = sizeof(struct dm_version_request);
1446         version_req.hdr.trans_id = atomic_inc_return(&trans_id);
1447         version_req.version.version = dm->next_version;
1448         dm->version = version_req.version.version;
1449
1450         /*
1451          * Set the next version to try in case current version fails.
1452          * Win7 protocol ought to be the last one to try.
1453          */
1454         switch (version_req.version.version) {
1455         case DYNMEM_PROTOCOL_VERSION_WIN8:
1456                 dm->next_version = DYNMEM_PROTOCOL_VERSION_WIN7;
1457                 version_req.is_last_attempt = 0;
1458                 break;
1459         default:
1460                 dm->next_version = 0;
1461                 version_req.is_last_attempt = 1;
1462         }
1463
1464         ret = vmbus_sendpacket(dm->dev->channel, &version_req,
1465                                 sizeof(struct dm_version_request),
1466                                 (unsigned long)NULL,
1467                                 VM_PKT_DATA_INBAND, 0);
1468
1469         if (ret)
1470                 goto version_error;
1471
1472         return;
1473
1474 version_error:
1475         dm->state = DM_INIT_ERROR;
1476         complete(&dm->host_event);
1477 }
1478
1479 static void cap_resp(struct hv_dynmem_device *dm,
1480                         struct dm_capabilities_resp_msg *cap_resp)
1481 {
1482         if (!cap_resp->is_accepted) {
1483                 pr_err("Capabilities not accepted by host\n");
1484                 dm->state = DM_INIT_ERROR;
1485         }
1486         complete(&dm->host_event);
1487 }
1488
1489 static void balloon_onchannelcallback(void *context)
1490 {
1491         struct hv_device *dev = context;
1492         u32 recvlen;
1493         u64 requestid;
1494         struct dm_message *dm_msg;
1495         struct dm_header *dm_hdr;
1496         struct hv_dynmem_device *dm = hv_get_drvdata(dev);
1497         struct dm_balloon *bal_msg;
1498         struct dm_hot_add *ha_msg;
1499         union dm_mem_page_range *ha_pg_range;
1500         union dm_mem_page_range *ha_region;
1501
1502         memset(recv_buffer, 0, sizeof(recv_buffer));
1503         vmbus_recvpacket(dev->channel, recv_buffer,
1504                          PAGE_SIZE, &recvlen, &requestid);
1505
1506         if (recvlen > 0) {
1507                 dm_msg = (struct dm_message *)recv_buffer;
1508                 dm_hdr = &dm_msg->hdr;
1509
1510                 switch (dm_hdr->type) {
1511                 case DM_VERSION_RESPONSE:
1512                         version_resp(dm,
1513                                  (struct dm_version_response *)dm_msg);
1514                         break;
1515
1516                 case DM_CAPABILITIES_RESPONSE:
1517                         cap_resp(dm,
1518                                  (struct dm_capabilities_resp_msg *)dm_msg);
1519                         break;
1520
1521                 case DM_BALLOON_REQUEST:
1522                         if (dm->state == DM_BALLOON_UP)
1523                                 pr_warn("Currently ballooning\n");
1524                         bal_msg = (struct dm_balloon *)recv_buffer;
1525                         dm->state = DM_BALLOON_UP;
1526                         dm_device.balloon_wrk.num_pages = bal_msg->num_pages;
1527                         schedule_work(&dm_device.balloon_wrk.wrk);
1528                         break;
1529
1530                 case DM_UNBALLOON_REQUEST:
1531                         dm->state = DM_BALLOON_DOWN;
1532                         balloon_down(dm,
1533                                  (struct dm_unballoon_request *)recv_buffer);
1534                         break;
1535
1536                 case DM_MEM_HOT_ADD_REQUEST:
1537                         if (dm->state == DM_HOT_ADD)
1538                                 pr_warn("Currently hot-adding\n");
1539                         dm->state = DM_HOT_ADD;
1540                         ha_msg = (struct dm_hot_add *)recv_buffer;
1541                         if (ha_msg->hdr.size == sizeof(struct dm_hot_add)) {
1542                                 /*
1543                                  * This is a normal hot-add request specifying
1544                                  * hot-add memory.
1545                                  */
1546                                 dm->host_specified_ha_region = false;
1547                                 ha_pg_range = &ha_msg->range;
1548                                 dm->ha_wrk.ha_page_range = *ha_pg_range;
1549                                 dm->ha_wrk.ha_region_range.page_range = 0;
1550                         } else {
1551                                 /*
1552                                  * Host is specifying that we first hot-add
1553                                  * a region and then partially populate this
1554                                  * region.
1555                                  */
1556                                 dm->host_specified_ha_region = true;
1557                                 ha_pg_range = &ha_msg->range;
1558                                 ha_region = &ha_pg_range[1];
1559                                 dm->ha_wrk.ha_page_range = *ha_pg_range;
1560                                 dm->ha_wrk.ha_region_range = *ha_region;
1561                         }
1562                         schedule_work(&dm_device.ha_wrk.wrk);
1563                         break;
1564
1565                 case DM_INFO_MESSAGE:
1566                         process_info(dm, (struct dm_info_msg *)dm_msg);
1567                         break;
1568
1569                 default:
1570                         pr_warn("Unhandled message: type: %d\n", dm_hdr->type);
1571
1572                 }
1573         }
1574
1575 }
1576
1577 static int balloon_probe(struct hv_device *dev,
1578                         const struct hv_vmbus_device_id *dev_id)
1579 {
1580         int ret;
1581         unsigned long t;
1582         struct dm_version_request version_req;
1583         struct dm_capabilities cap_msg;
1584
1585 #ifdef CONFIG_MEMORY_HOTPLUG
1586         do_hot_add = hot_add;
1587 #else
1588         do_hot_add = false;
1589 #endif
1590
1591         /*
1592          * First allocate a send buffer.
1593          */
1594
1595         send_buffer = kmalloc(PAGE_SIZE, GFP_KERNEL);
1596         if (!send_buffer)
1597                 return -ENOMEM;
1598
1599         ret = vmbus_open(dev->channel, dm_ring_size, dm_ring_size, NULL, 0,
1600                         balloon_onchannelcallback, dev);
1601
1602         if (ret)
1603                 goto probe_error0;
1604
1605         dm_device.dev = dev;
1606         dm_device.state = DM_INITIALIZING;
1607         dm_device.next_version = DYNMEM_PROTOCOL_VERSION_WIN8;
1608         init_completion(&dm_device.host_event);
1609         init_completion(&dm_device.config_event);
1610         INIT_LIST_HEAD(&dm_device.ha_region_list);
1611         spin_lock_init(&dm_device.ha_lock);
1612         INIT_WORK(&dm_device.balloon_wrk.wrk, balloon_up);
1613         INIT_WORK(&dm_device.ha_wrk.wrk, hot_add_req);
1614         dm_device.host_specified_ha_region = false;
1615
1616         dm_device.thread =
1617                  kthread_run(dm_thread_func, &dm_device, "hv_balloon");
1618         if (IS_ERR(dm_device.thread)) {
1619                 ret = PTR_ERR(dm_device.thread);
1620                 goto probe_error1;
1621         }
1622
1623 #ifdef CONFIG_MEMORY_HOTPLUG
1624         set_online_page_callback(&hv_online_page);
1625         register_memory_notifier(&hv_memory_nb);
1626 #endif
1627
1628         hv_set_drvdata(dev, &dm_device);
1629         /*
1630          * Initiate the hand shake with the host and negotiate
1631          * a version that the host can support. We start with the
1632          * highest version number and go down if the host cannot
1633          * support it.
1634          */
1635         memset(&version_req, 0, sizeof(struct dm_version_request));
1636         version_req.hdr.type = DM_VERSION_REQUEST;
1637         version_req.hdr.size = sizeof(struct dm_version_request);
1638         version_req.hdr.trans_id = atomic_inc_return(&trans_id);
1639         version_req.version.version = DYNMEM_PROTOCOL_VERSION_WIN10;
1640         version_req.is_last_attempt = 0;
1641         dm_device.version = version_req.version.version;
1642
1643         ret = vmbus_sendpacket(dev->channel, &version_req,
1644                                 sizeof(struct dm_version_request),
1645                                 (unsigned long)NULL,
1646                                 VM_PKT_DATA_INBAND, 0);
1647         if (ret)
1648                 goto probe_error2;
1649
1650         t = wait_for_completion_timeout(&dm_device.host_event, 5*HZ);
1651         if (t == 0) {
1652                 ret = -ETIMEDOUT;
1653                 goto probe_error2;
1654         }
1655
1656         /*
1657          * If we could not negotiate a compatible version with the host
1658          * fail the probe function.
1659          */
1660         if (dm_device.state == DM_INIT_ERROR) {
1661                 ret = -ETIMEDOUT;
1662                 goto probe_error2;
1663         }
1664
1665         pr_info("Using Dynamic Memory protocol version %u.%u\n",
1666                 DYNMEM_MAJOR_VERSION(dm_device.version),
1667                 DYNMEM_MINOR_VERSION(dm_device.version));
1668
1669         /*
1670          * Now submit our capabilities to the host.
1671          */
1672         memset(&cap_msg, 0, sizeof(struct dm_capabilities));
1673         cap_msg.hdr.type = DM_CAPABILITIES_REPORT;
1674         cap_msg.hdr.size = sizeof(struct dm_capabilities);
1675         cap_msg.hdr.trans_id = atomic_inc_return(&trans_id);
1676
1677         cap_msg.caps.cap_bits.balloon = 1;
1678         cap_msg.caps.cap_bits.hot_add = 1;
1679
1680         /*
1681          * Specify our alignment requirements as it relates
1682          * memory hot-add. Specify 128MB alignment.
1683          */
1684         cap_msg.caps.cap_bits.hot_add_alignment = 7;
1685
1686         /*
1687          * Currently the host does not use these
1688          * values and we set them to what is done in the
1689          * Windows driver.
1690          */
1691         cap_msg.min_page_cnt = 0;
1692         cap_msg.max_page_number = -1;
1693
1694         ret = vmbus_sendpacket(dev->channel, &cap_msg,
1695                                 sizeof(struct dm_capabilities),
1696                                 (unsigned long)NULL,
1697                                 VM_PKT_DATA_INBAND, 0);
1698         if (ret)
1699                 goto probe_error2;
1700
1701         t = wait_for_completion_timeout(&dm_device.host_event, 5*HZ);
1702         if (t == 0) {
1703                 ret = -ETIMEDOUT;
1704                 goto probe_error2;
1705         }
1706
1707         /*
1708          * If the host does not like our capabilities,
1709          * fail the probe function.
1710          */
1711         if (dm_device.state == DM_INIT_ERROR) {
1712                 ret = -ETIMEDOUT;
1713                 goto probe_error2;
1714         }
1715
1716         dm_device.state = DM_INITIALIZED;
1717         last_post_time = jiffies;
1718
1719         return 0;
1720
1721 probe_error2:
1722 #ifdef CONFIG_MEMORY_HOTPLUG
1723         restore_online_page_callback(&hv_online_page);
1724 #endif
1725         kthread_stop(dm_device.thread);
1726
1727 probe_error1:
1728         vmbus_close(dev->channel);
1729 probe_error0:
1730         kfree(send_buffer);
1731         return ret;
1732 }
1733
1734 static int balloon_remove(struct hv_device *dev)
1735 {
1736         struct hv_dynmem_device *dm = hv_get_drvdata(dev);
1737         struct hv_hotadd_state *has, *tmp;
1738         struct hv_hotadd_gap *gap, *tmp_gap;
1739         unsigned long flags;
1740
1741         if (dm->num_pages_ballooned != 0)
1742                 pr_warn("Ballooned pages: %d\n", dm->num_pages_ballooned);
1743
1744         cancel_work_sync(&dm->balloon_wrk.wrk);
1745         cancel_work_sync(&dm->ha_wrk.wrk);
1746
1747         vmbus_close(dev->channel);
1748         kthread_stop(dm->thread);
1749         kfree(send_buffer);
1750 #ifdef CONFIG_MEMORY_HOTPLUG
1751         restore_online_page_callback(&hv_online_page);
1752         unregister_memory_notifier(&hv_memory_nb);
1753 #endif
1754         spin_lock_irqsave(&dm_device.ha_lock, flags);
1755         list_for_each_entry_safe(has, tmp, &dm->ha_region_list, list) {
1756                 list_for_each_entry_safe(gap, tmp_gap, &has->gap_list, list) {
1757                         list_del(&gap->list);
1758                         kfree(gap);
1759                 }
1760                 list_del(&has->list);
1761                 kfree(has);
1762         }
1763         spin_unlock_irqrestore(&dm_device.ha_lock, flags);
1764
1765         return 0;
1766 }
1767
1768 static const struct hv_vmbus_device_id id_table[] = {
1769         /* Dynamic Memory Class ID */
1770         /* 525074DC-8985-46e2-8057-A307DC18A502 */
1771         { HV_DM_GUID, },
1772         { },
1773 };
1774
1775 MODULE_DEVICE_TABLE(vmbus, id_table);
1776
1777 static  struct hv_driver balloon_drv = {
1778         .name = "hv_balloon",
1779         .id_table = id_table,
1780         .probe =  balloon_probe,
1781         .remove =  balloon_remove,
1782         .driver = {
1783                 .probe_type = PROBE_PREFER_ASYNCHRONOUS,
1784         },
1785 };
1786
1787 static int __init init_balloon_drv(void)
1788 {
1789
1790         return vmbus_driver_register(&balloon_drv);
1791 }
1792
1793 module_init(init_balloon_drv);
1794
1795 MODULE_DESCRIPTION("Hyper-V Balloon");
1796 MODULE_LICENSE("GPL");