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