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