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