drivers/hv/hv_balloon.c

   1 /*
   2  * Copyright (c) 2012, Microsoft Corporation.
   3  *
   4  * Author:
   5  *   K. Y. Srinivasan <kys@microsoft.com>
   6  *
   7  * This program is free software; you can redistribute it and/or modify it
   8  * under the terms of the GNU General Public License version 2 as published
   9  * by the Free Software Foundation.
  10  *
  11  * This program is distributed in the hope that it will be useful, but
  12  * WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
  14  * NON INFRINGEMENT.  See the GNU General Public License for more
  15  * details.
  16  *
  17  */
  18
  19 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  20
  21 #include <linux/kernel.h>
  22 #include <linux/jiffies.h>
  23 #include <linux/mman.h>
  24 #include <linux/delay.h>
  25 #include <linux/init.h>
  26 #include <linux/module.h>
  27 #include <linux/slab.h>
  28 #include <linux/kthread.h>
  29 #include <linux/completion.h>
  30 #include <linux/memory_hotplug.h>
  31 #include <linux/memory.h>
  32 #include <linux/notifier.h>
  33 #include <linux/percpu_counter.h>
  34
  35 #include <linux/hyperv.h>
  36
  37 /*
  38  * We begin with definitions supporting the Dynamic Memory protocol
  39  * with the host.
  40  *
  41  * Begin protocol definitions.
  42  */
  43
  44
  45
  46 /*
  47  * Protocol versions. The low word is the minor version, the high word the major
  48  * version.
  49  *
  50  * History:
  51  * Initial version 1.0
  52  * Changed to 0.1 on 2009/03/25
  53  * Changes to 0.2 on 2009/05/14
  54  * Changes to 0.3 on 2009/12/03
  55  * Changed to 1.0 on 2011/04/05
  56  */
  57
  58 #define DYNMEM_MAKE_VERSION(Major, Minor) ((__u32)(((Major) << 16) | (Minor)))
  59 #define DYNMEM_MAJOR_VERSION(Version) ((__u32)(Version) >> 16)
  60 #define DYNMEM_MINOR_VERSION(Version) ((__u32)(Version) & 0xff)
  61
  62 enum {
  63         DYNMEM_PROTOCOL_VERSION_1 = DYNMEM_MAKE_VERSION(0, 3),
  64         DYNMEM_PROTOCOL_VERSION_2 = DYNMEM_MAKE_VERSION(1, 0),
  65         DYNMEM_PROTOCOL_VERSION_3 = DYNMEM_MAKE_VERSION(2, 0),
  66
  67         DYNMEM_PROTOCOL_VERSION_WIN7 = DYNMEM_PROTOCOL_VERSION_1,
  68         DYNMEM_PROTOCOL_VERSION_WIN8 = DYNMEM_PROTOCOL_VERSION_2,
  69         DYNMEM_PROTOCOL_VERSION_WIN10 = DYNMEM_PROTOCOL_VERSION_3,
  70
  71         DYNMEM_PROTOCOL_VERSION_CURRENT = DYNMEM_PROTOCOL_VERSION_WIN10
  72 };
  73
  74
  75
  76 /*
  77  * Message Types
  78  */
  79
  80 enum dm_message_type {
  81         /*
  82          * Version 0.3
  83          */
  84         DM_ERROR                        = 0,
  85         DM_VERSION_REQUEST              = 1,
  86         DM_VERSION_RESPONSE             = 2,
  87         DM_CAPABILITIES_REPORT          = 3,
  88         DM_CAPABILITIES_RESPONSE        = 4,
  89         DM_STATUS_REPORT                = 5,
  90         DM_BALLOON_REQUEST              = 6,
  91         DM_BALLOON_RESPONSE             = 7,
  92         DM_UNBALLOON_REQUEST            = 8,
  93         DM_UNBALLOON_RESPONSE           = 9,
  94         DM_MEM_HOT_ADD_REQUEST          = 10,
  95         DM_MEM_HOT_ADD_RESPONSE         = 11,
  96         DM_VERSION_03_MAX               = 11,
  97         /*
  98          * Version 1.0.
  99          */
 100         DM_INFO_MESSAGE                 = 12,
 101         DM_VERSION_1_MAX                = 12
 102 };
 103
 104
 105 /*
 106  * Structures defining the dynamic memory management
 107  * protocol.
 108  */
 109
 110 union dm_version {
 111         struct {
 112                 __u16 minor_version;
 113                 __u16 major_version;
 114         };
 115         __u32 version;
 116 } __packed;
 117
 118
 119 union dm_caps {
 120         struct {
 121                 __u64 balloon:1;
 122                 __u64 hot_add:1;
 123                 /*
 124                  * To support guests that may have alignment
 125                  * limitations on hot-add, the guest can specify
 126                  * its alignment requirements; a value of n
 127                  * represents an alignment of 2^n in mega bytes.
 128                  */
 129                 __u64 hot_add_alignment:4;
 130                 __u64 reservedz:58;
 131         } cap_bits;
 132         __u64 caps;
 133 } __packed;
 134
 135 union dm_mem_page_range {
 136         struct  {
 137                 /*
 138                  * The PFN number of the first page in the range.
 139                  * 40 bits is the architectural limit of a PFN
 140                  * number for AMD64.
 141                  */
 142                 __u64 start_page:40;
 143                 /*
 144                  * The number of pages in the range.
 145                  */
 146                 __u64 page_cnt:24;
 147         } finfo;
 148         __u64  page_range;
 149 } __packed;
 150
 151
 152
 153 /*
 154  * The header for all dynamic memory messages:
 155  *
 156  * type: Type of the message.
 157  * size: Size of the message in bytes; including the header.
 158  * trans_id: The guest is responsible for manufacturing this ID.
 159  */
 160
 161 struct dm_header {
 162         __u16 type;
 163         __u16 size;
 164         __u32 trans_id;
 165 } __packed;
 166
 167 /*
 168  * A generic message format for dynamic memory.
 169  * Specific message formats are defined later in the file.
 170  */
 171
 172 struct dm_message {
 173         struct dm_header hdr;
 174         __u8 data[]; /* enclosed message */
 175 } __packed;
 176
 177
 178 /*
 179  * Specific message types supporting the dynamic memory protocol.
 180  */
 181
 182 /*
 183  * Version negotiation message. Sent from the guest to the host.
 184  * The guest is free to try different versions until the host
 185  * accepts the version.
 186  *
 187  * dm_version: The protocol version requested.
 188  * is_last_attempt: If TRUE, this is the last version guest will request.
 189  * reservedz: Reserved field, set to zero.
 190  */
 191
 192 struct dm_version_request {
 193         struct dm_header hdr;
 194         union dm_version version;
 195         __u32 is_last_attempt:1;
 196         __u32 reservedz:31;
 197 } __packed;
 198
 199 /*
 200  * Version response message; Host to Guest and indicates
 201  * if the host has accepted the version sent by the guest.
 202  *
 203  * is_accepted: If TRUE, host has accepted the version and the guest
 204  * should proceed to the next stage of the protocol. FALSE indicates that
 205  * guest should re-try with a different version.
 206  *
 207  * reservedz: Reserved field, set to zero.
 208  */
 209
 210 struct dm_version_response {
 211         struct dm_header hdr;
 212         __u64 is_accepted:1;
 213         __u64 reservedz:63;
 214 } __packed;
 215
 216 /*
 217  * Message reporting capabilities. This is sent from the guest to the
 218  * host.
 219  */
 220
 221 struct dm_capabilities {
 222         struct dm_header hdr;
 223         union dm_caps caps;
 224         __u64 min_page_cnt;
 225         __u64 max_page_number;
 226 } __packed;
 227
 228 /*
 229  * Response to the capabilities message. This is sent from the host to the
 230  * guest. This message notifies if the host has accepted the guest's
 231  * capabilities. If the host has not accepted, the guest must shutdown
 232  * the service.
 233  *
 234  * is_accepted: Indicates if the host has accepted guest's capabilities.
 235  * reservedz: Must be 0.
 236  */
 237
 238 struct dm_capabilities_resp_msg {
 239         struct dm_header hdr;
 240         __u64 is_accepted:1;
 241         __u64 reservedz:63;
 242 } __packed;
 243
 244 /*
 245  * This message is used to report memory pressure from the guest.
 246  * This message is not part of any transaction and there is no
 247  * response to this message.
 248  *
 249  * num_avail: Available memory in pages.
 250  * num_committed: Committed memory in pages.
 251  * page_file_size: The accumulated size of all page files
 252  *                 in the system in pages.
 253  * zero_free: The nunber of zero and free pages.
 254  * page_file_writes: The writes to the page file in pages.
 255  * io_diff: An indicator of file cache efficiency or page file activity,
 256  *          calculated as File Cache Page Fault Count - Page Read Count.
 257  *          This value is in pages.
 258  *
 259  * Some of these metrics are Windows specific and fortunately
 260  * the algorithm on the host side that computes the guest memory
 261  * pressure only uses num_committed value.
 262  */
 263
 264 struct dm_status {
 265         struct dm_header hdr;
 266         __u64 num_avail;
 267         __u64 num_committed;
 268         __u64 page_file_size;
 269         __u64 zero_free;
 270         __u32 page_file_writes;
 271         __u32 io_diff;
 272 } __packed;
 273
 274
 275 /*
 276  * Message to ask the guest to allocate memory - balloon up message.
 277  * This message is sent from the host to the guest. The guest may not be
 278  * able to allocate as much memory as requested.
 279  *
 280  * num_pages: number of pages to allocate.
 281  */
 282
 283 struct dm_balloon {
 284         struct dm_header hdr;
 285         __u32 num_pages;
 286         __u32 reservedz;
 287 } __packed;
 288
 289
 290 /*
 291  * Balloon response message; this message is sent from the guest
 292  * to the host in response to the balloon message.
 293  *
 294  * reservedz: Reserved; must be set to zero.
 295  * more_pages: If FALSE, this is the last message of the transaction.
 296  * if TRUE there will atleast one more message from the guest.
 297  *
 298  * range_count: The number of ranges in the range array.
 299  *
 300  * range_array: An array of page ranges returned to the host.
 301  *
 302  */
 303
 304 struct dm_balloon_response {
 305         struct dm_header hdr;
 306         __u32 reservedz;
 307         __u32 more_pages:1;
 308         __u32 range_count:31;
 309         union dm_mem_page_range range_array[];
 310 } __packed;
 311
 312 /*
 313  * Un-balloon message; this message is sent from the host
 314  * to the guest to give guest more memory.
 315  *
 316  * more_pages: If FALSE, this is the last message of the transaction.
 317  * if TRUE there will atleast one more message from the guest.
 318  *
 319  * reservedz: Reserved; must be set to zero.
 320  *
 321  * range_count: The number of ranges in the range array.
 322  *
 323  * range_array: An array of page ranges returned to the host.
 324  *
 325  */
 326
 327 struct dm_unballoon_request {
 328         struct dm_header hdr;
 329         __u32 more_pages:1;
 330         __u32 reservedz:31;
 331         __u32 range_count;
 332         union dm_mem_page_range range_array[];
 333 } __packed;
 334
 335 /*
 336  * Un-balloon response message; this message is sent from the guest
 337  * to the host in response to an unballoon request.
 338  *
 339  */
 340
 341 struct dm_unballoon_response {
 342         struct dm_header hdr;
 343 } __packed;
 344
 345
 346 /*
 347  * Hot add request message. Message sent from the host to the guest.
 348  *
 349  * mem_range: Memory range to hot add.
 350  *
 351  * On Linux we currently don't support this since we cannot hot add
 352  * arbitrary granularity of memory.
 353  */
 354
 355 struct dm_hot_add {
 356         struct dm_header hdr;
 357         union dm_mem_page_range range;
 358 } __packed;
 359
 360 /*
 361  * Hot add response message.
 362  * This message is sent by the guest to report the status of a hot add request.
 363  * If page_count is less than the requested page count, then the host should
 364  * assume all further hot add requests will fail, since this indicates that
 365  * the guest has hit an upper physical memory barrier.
 366  *
 367  * Hot adds may also fail due to low resources; in this case, the guest must
 368  * not complete this message until the hot add can succeed, and the host must
 369  * not send a new hot add request until the response is sent.
 370  * If VSC fails to hot add memory DYNMEM_NUMBER_OF_UNSUCCESSFUL_HOTADD_ATTEMPTS
 371  * times it fails the request.
 372  *
 373  *
 374  * page_count: number of pages that were successfully hot added.
 375  *
 376  * result: result of the operation 1: success, 0: failure.
 377  *
 378  */
 379
 380 struct dm_hot_add_response {
 381         struct dm_header hdr;
 382         __u32 page_count;
 383         __u32 result;
 384 } __packed;
 385
 386 /*
 387  * Types of information sent from host to the guest.
 388  */
 389
 390 enum dm_info_type {
 391         INFO_TYPE_MAX_PAGE_CNT = 0,
 392         MAX_INFO_TYPE
 393 };
 394
 395
 396 /*
 397  * Header for the information message.
 398  */
 399
 400 struct dm_info_header {
 401         enum dm_info_type type;
 402         __u32 data_size;
 403 } __packed;
 404
 405 /*
 406  * This message is sent from the host to the guest to pass
 407  * some relevant information (win8 addition).
 408  *
 409  * reserved: no used.
 410  * info_size: size of the information blob.
 411  * info: information blob.
 412  */
 413
 414 struct dm_info_msg {
 415         struct dm_header hdr;
 416         __u32 reserved;
 417         __u32 info_size;
 418         __u8  info[];
 419 };
 420
 421 /*
 422  * End protocol definitions.
 423  */
 424
 425 /*
 426  * State to manage hot adding memory into the guest.
 427  * The range start_pfn : end_pfn specifies the range
 428  * that the host has asked us to hot add. The range
 429  * start_pfn : ha_end_pfn specifies the range that we have
 430  * currently hot added. We hot add in multiples of 128M
 431  * chunks; it is possible that we may not be able to bring
 432  * online all the pages in the region. The range
 433  * covered_start_pfn:covered_end_pfn defines the pages that can
 434  * be brough online.
 435  */
 436
 437 struct hv_hotadd_state {
 438         struct list_head list;
 439         unsigned long start_pfn;
 440         unsigned long covered_start_pfn;
 441         unsigned long covered_end_pfn;
 442         unsigned long ha_end_pfn;
 443         unsigned long end_pfn;
 444         /*
 445          * A list of gaps.
 446          */
 447         struct list_head gap_list;
 448 };
 449
 450 struct hv_hotadd_gap {
 451         struct list_head list;
 452         unsigned long start_pfn;
 453         unsigned long end_pfn;
 454 };
 455
 456 struct balloon_state {
 457         __u32 num_pages;
 458         struct work_struct wrk;
 459 };
 460
 461 struct hot_add_wrk {
 462         union dm_mem_page_range ha_page_range;
 463         union dm_mem_page_range ha_region_range;
 464         struct work_struct wrk;
 465 };
 466
 467 static bool hot_add = true;
 468 static bool do_hot_add;
 469 /*
 470  * Delay reporting memory pressure by
 471  * the specified number of seconds.
 472  */
 473 static uint pressure_report_delay = 45;
 474
 475 /*
 476  * The last time we posted a pressure report to host.
 477  */
 478 static unsigned long last_post_time;
 479
 480 module_param(hot_add, bool, (S_IRUGO | S_IWUSR));
 481 MODULE_PARM_DESC(hot_add, "If set attempt memory hot_add");
 482
 483 module_param(pressure_report_delay, uint, (S_IRUGO | S_IWUSR));
 484 MODULE_PARM_DESC(pressure_report_delay, "Delay in secs in reporting pressure");
 485 static atomic_t trans_id = ATOMIC_INIT(0);
 486
 487 static int dm_ring_size = (5 * PAGE_SIZE);
 488
 489 /*
 490  * Driver specific state.
 491  */
 492
 493 enum hv_dm_state {
 494         DM_INITIALIZING = 0,
 495         DM_INITIALIZED,
 496         DM_BALLOON_UP,
 497         DM_BALLOON_DOWN,
 498         DM_HOT_ADD,
 499         DM_INIT_ERROR
 500 };
 501
 502
 503 static __u8 recv_buffer[PAGE_SIZE];
 504 static __u8 *send_buffer;
 505 #define PAGES_IN_2M     512
 506 #define HA_CHUNK (32 * 1024)
 507
 508 struct hv_dynmem_device {
 509         struct hv_device *dev;
 510         enum hv_dm_state state;
 511         struct completion host_event;
 512         struct completion config_event;
 513
 514         /*
 515          * Number of pages we have currently ballooned out.
 516          */
 517         unsigned int num_pages_ballooned;
 518         unsigned int num_pages_onlined;
 519         unsigned int num_pages_added;
 520
 521         /*
 522          * State to manage the ballooning (up) operation.
 523          */
 524         struct balloon_state balloon_wrk;
 525
 526         /*
 527          * State to execute the "hot-add" operation.
 528          */
 529         struct hot_add_wrk ha_wrk;
 530
 531         /*
 532          * This state tracks if the host has specified a hot-add
 533          * region.
 534          */
 535         bool host_specified_ha_region;
 536
 537         /*
 538          * State to synchronize hot-add.
 539          */
 540         struct completion  ol_waitevent;
 541         bool ha_waiting;
 542         /*
 543          * This thread handles hot-add
 544          * requests from the host as well as notifying
 545          * the host with regards to memory pressure in
 546          * the guest.
 547          */
 548         struct task_struct *thread;
 549
 550         /*
 551          * Protects ha_region_list, num_pages_onlined counter and individual
 552          * regions from ha_region_list.
 553          */
 554         spinlock_t ha_lock;
 555
 556         /*
 557          * A list of hot-add regions.
 558          */
 559         struct list_head ha_region_list;
 560
 561         /*
 562          * We start with the highest version we can support
 563          * and downgrade based on the host; we save here the
 564          * next version to try.
 565          */
 566         __u32 next_version;
 567
 568         /*
 569          * The negotiated version agreed by host.
 570          */
 571         __u32 version;
 572 };
 573
 574 static struct hv_dynmem_device dm_device;
 575
 576 static void post_status(struct hv_dynmem_device *dm);
 577
 578 #ifdef CONFIG_MEMORY_HOTPLUG
 579 static int hv_memory_notifier(struct notifier_block *nb, unsigned long val,
 580                               void *v)
 581 {
 582         struct memory_notify *mem = (struct memory_notify *)v;
 583         unsigned long flags;
 584
 585         switch (val) {
 586         case MEM_ONLINE:
 587         case MEM_CANCEL_ONLINE:
 588                 if (dm_device.ha_waiting) {
 589                         dm_device.ha_waiting = false;
 590                         complete(&dm_device.ol_waitevent);
 591                 }
 592                 break;
 593
 594         case MEM_OFFLINE:
 595                 spin_lock_irqsave(&dm_device.ha_lock, flags);
 596                 dm_device.num_pages_onlined -= mem->nr_pages;
 597                 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 598                 break;
 599         case MEM_GOING_ONLINE:
 600         case MEM_GOING_OFFLINE:
 601         case MEM_CANCEL_OFFLINE:
 602                 break;
 603         }
 604         return NOTIFY_OK;
 605 }
 606
 607 static struct notifier_block hv_memory_nb = {
 608         .notifier_call = hv_memory_notifier,
 609         .priority = 0
 610 };
 611
 612 /* Check if the particular page is backed and can be onlined and online it. */
 613 static void hv_page_online_one(struct hv_hotadd_state *has, struct page *pg)
 614 {
 615         unsigned long cur_start_pgp;
 616         unsigned long cur_end_pgp;
 617         struct hv_hotadd_gap *gap;
 618
 619         cur_start_pgp = (unsigned long)pfn_to_page(has->covered_start_pfn);
 620         cur_end_pgp = (unsigned long)pfn_to_page(has->covered_end_pfn);
 621
 622         /* The page is not backed. */
 623         if (((unsigned long)pg < cur_start_pgp) ||
 624             ((unsigned long)pg >= cur_end_pgp))
 625                 return;
 626
 627         /* Check for gaps. */
 628         list_for_each_entry(gap, &has->gap_list, list) {
 629                 cur_start_pgp = (unsigned long)
 630                         pfn_to_page(gap->start_pfn);
 631                 cur_end_pgp = (unsigned long)
 632                         pfn_to_page(gap->end_pfn);
 633                 if (((unsigned long)pg >= cur_start_pgp) &&
 634                     ((unsigned long)pg < cur_end_pgp)) {
 635                         return;
 636                 }
 637         }
 638
 639         /* This frame is currently backed; online the page. */
 640         __online_page_set_limits(pg);
 641         __online_page_increment_counters(pg);
 642         __online_page_free(pg);
 643
 644         WARN_ON_ONCE(!spin_is_locked(&dm_device.ha_lock));
 645         dm_device.num_pages_onlined++;
 646 }
 647
 648 static void hv_bring_pgs_online(struct hv_hotadd_state *has,
 649                                 unsigned long start_pfn, unsigned long size)
 650 {
 651         int i;
 652
 653         pr_debug("Online %lu pages starting at pfn 0x%lx\n", size, start_pfn);
 654         for (i = 0; i < size; i++)
 655                 hv_page_online_one(has, pfn_to_page(start_pfn + i));
 656 }
 657
 658 static void hv_mem_hot_add(unsigned long start, unsigned long size,
 659                                 unsigned long pfn_count,
 660                                 struct hv_hotadd_state *has)
 661 {
 662         int ret = 0;
 663         int i, nid;
 664         unsigned long start_pfn;
 665         unsigned long processed_pfn;
 666         unsigned long total_pfn = pfn_count;
 667         unsigned long flags;
 668
 669         for (i = 0; i < (size/HA_CHUNK); i++) {
 670                 start_pfn = start + (i * HA_CHUNK);
 671
 672                 spin_lock_irqsave(&dm_device.ha_lock, flags);
 673                 has->ha_end_pfn +=  HA_CHUNK;
 674
 675                 if (total_pfn > HA_CHUNK) {
 676                         processed_pfn = HA_CHUNK;
 677                         total_pfn -= HA_CHUNK;
 678                 } else {
 679                         processed_pfn = total_pfn;
 680                         total_pfn = 0;
 681                 }
 682
 683                 has->covered_end_pfn +=  processed_pfn;
 684                 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 685
 686                 init_completion(&dm_device.ol_waitevent);
 687                 dm_device.ha_waiting = !memhp_auto_online;
 688
 689                 nid = memory_add_physaddr_to_nid(PFN_PHYS(start_pfn));
 690                 ret = add_memory(nid, PFN_PHYS((start_pfn)),
 691                                 (HA_CHUNK << PAGE_SHIFT));
 692
 693                 if (ret) {
 694                         pr_warn("hot_add memory failed error is %d\n", ret);
 695                         if (ret == -EEXIST) {
 696                                 /*
 697                                  * This error indicates that the error
 698                                  * is not a transient failure. This is the
 699                                  * case where the guest's physical address map
 700                                  * precludes hot adding memory. Stop all further
 701                                  * memory hot-add.
 702                                  */
 703                                 do_hot_add = false;
 704                         }
 705                         spin_lock_irqsave(&dm_device.ha_lock, flags);
 706                         has->ha_end_pfn -= HA_CHUNK;
 707                         has->covered_end_pfn -=  processed_pfn;
 708                         spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 709                         break;
 710                 }
 711
 712                 /*
 713                  * Wait for the memory block to be onlined when memory onlining
 714                  * is done outside of kernel (memhp_auto_online). Since the hot
 715                  * add has succeeded, it is ok to proceed even if the pages in
 716                  * the hot added region have not been "onlined" within the
 717                  * allowed time.
 718                  */
 719                 if (dm_device.ha_waiting)
 720                         wait_for_completion_timeout(&dm_device.ol_waitevent,
 721                                                     5*HZ);
 722                 post_status(&dm_device);
 723         }
 724 }
 725
 726 static void hv_online_page(struct page *pg)
 727 {
 728         struct hv_hotadd_state *has;
 729         unsigned long cur_start_pgp;
 730         unsigned long cur_end_pgp;
 731         unsigned long flags;
 732
 733         spin_lock_irqsave(&dm_device.ha_lock, flags);
 734         list_for_each_entry(has, &dm_device.ha_region_list, list) {
 735                 cur_start_pgp = (unsigned long)
 736                         pfn_to_page(has->start_pfn);
 737                 cur_end_pgp = (unsigned long)pfn_to_page(has->end_pfn);
 738
 739                 /* The page belongs to a different HAS. */
 740                 if (((unsigned long)pg < cur_start_pgp) ||
 741                     ((unsigned long)pg >= cur_end_pgp))
 742                         continue;
 743
 744                 hv_page_online_one(has, pg);
 745                 break;
 746         }
 747         spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 748 }
 749
 750 static int pfn_covered(unsigned long start_pfn, unsigned long pfn_cnt)
 751 {
 752         struct hv_hotadd_state *has;
 753         struct hv_hotadd_gap *gap;
 754         unsigned long residual, new_inc;
 755         int ret = 0;
 756         unsigned long flags;
 757
 758         spin_lock_irqsave(&dm_device.ha_lock, flags);
 759         list_for_each_entry(has, &dm_device.ha_region_list, list) {
 760                 /*
 761                  * If the pfn range we are dealing with is not in the current
 762                  * "hot add block", move on.
 763                  */
 764                 if (start_pfn < has->start_pfn || start_pfn >= has->end_pfn)
 765                         continue;
 766
 767                 /*
 768                  * If the current start pfn is not where the covered_end
 769                  * is, create a gap and update covered_end_pfn.
 770                  */
 771                 if (has->covered_end_pfn != start_pfn) {
 772                         gap = kzalloc(sizeof(struct hv_hotadd_gap), GFP_ATOMIC);
 773                         if (!gap) {
 774                                 ret = -ENOMEM;
 775                                 break;
 776                         }
 777
 778                         INIT_LIST_HEAD(&gap->list);
 779                         gap->start_pfn = has->covered_end_pfn;
 780                         gap->end_pfn = start_pfn;
 781                         list_add_tail(&gap->list, &has->gap_list);
 782
 783                         has->covered_end_pfn = start_pfn;
 784                 }
 785
 786                 /*
 787                  * If the current hot add-request extends beyond
 788                  * our current limit; extend it.
 789                  */
 790                 if ((start_pfn + pfn_cnt) > has->end_pfn) {
 791                         residual = (start_pfn + pfn_cnt - has->end_pfn);
 792                         /*
 793                          * Extend the region by multiples of HA_CHUNK.
 794                          */
 795                         new_inc = (residual / HA_CHUNK) * HA_CHUNK;
 796                         if (residual % HA_CHUNK)
 797                                 new_inc += HA_CHUNK;
 798
 799                         has->end_pfn += new_inc;
 800                 }
 801
 802                 ret = 1;
 803                 break;
 804         }
 805         spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 806
 807         return ret;
 808 }
 809
 810 static unsigned long handle_pg_range(unsigned long pg_start,
 811                                         unsigned long pg_count)
 812 {
 813         unsigned long start_pfn = pg_start;
 814         unsigned long pfn_cnt = pg_count;
 815         unsigned long size;
 816         struct hv_hotadd_state *has;
 817         unsigned long pgs_ol = 0;
 818         unsigned long old_covered_state;
 819         unsigned long res = 0, flags;
 820
 821         pr_debug("Hot adding %lu pages starting at pfn 0x%lx.\n", pg_count,
 822                 pg_start);
 823
 824         spin_lock_irqsave(&dm_device.ha_lock, flags);
 825         list_for_each_entry(has, &dm_device.ha_region_list, list) {
 826                 /*
 827                  * If the pfn range we are dealing with is not in the current
 828                  * "hot add block", move on.
 829                  */
 830                 if (start_pfn < has->start_pfn || start_pfn >= has->end_pfn)
 831                         continue;
 832
 833                 old_covered_state = has->covered_end_pfn;
 834
 835                 if (start_pfn < has->ha_end_pfn) {
 836                         /*
 837                          * This is the case where we are backing pages
 838                          * in an already hot added region. Bring
 839                          * these pages online first.
 840                          */
 841                         pgs_ol = has->ha_end_pfn - start_pfn;
 842                         if (pgs_ol > pfn_cnt)
 843                                 pgs_ol = pfn_cnt;
 844
 845                         has->covered_end_pfn +=  pgs_ol;
 846                         pfn_cnt -= pgs_ol;
 847                         /*
 848                          * Check if the corresponding memory block is already
 849                          * online by checking its last previously backed page.
 850                          * In case it is we need to bring rest (which was not
 851                          * backed previously) online too.
 852                          */
 853                         if (start_pfn > has->start_pfn &&
 854                             !PageReserved(pfn_to_page(start_pfn - 1)))
 855                                 hv_bring_pgs_online(has, start_pfn, pgs_ol);
 856
 857                 }
 858
 859                 if ((has->ha_end_pfn < has->end_pfn) && (pfn_cnt > 0)) {
 860                         /*
 861                          * We have some residual hot add range
 862                          * that needs to be hot added; hot add
 863                          * it now. Hot add a multiple of
 864                          * of HA_CHUNK that fully covers the pages
 865                          * we have.
 866                          */
 867                         size = (has->end_pfn - has->ha_end_pfn);
 868                         if (pfn_cnt <= size) {
 869                                 size = ((pfn_cnt / HA_CHUNK) * HA_CHUNK);
 870                                 if (pfn_cnt % HA_CHUNK)
 871                                         size += HA_CHUNK;
 872                         } else {
 873                                 pfn_cnt = size;
 874                         }
 875                         spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 876                         hv_mem_hot_add(has->ha_end_pfn, size, pfn_cnt, has);
 877                         spin_lock_irqsave(&dm_device.ha_lock, flags);
 878                 }
 879                 /*
 880                  * If we managed to online any pages that were given to us,
 881                  * we declare success.
 882                  */
 883                 res = has->covered_end_pfn - old_covered_state;
 884                 break;
 885         }
 886         spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 887
 888         return res;
 889 }
 890
 891 static unsigned long process_hot_add(unsigned long pg_start,
 892                                         unsigned long pfn_cnt,
 893                                         unsigned long rg_start,
 894                                         unsigned long rg_size)
 895 {
 896         struct hv_hotadd_state *ha_region = NULL;
 897         int covered;
 898         unsigned long flags;
 899
 900         if (pfn_cnt == 0)
 901                 return 0;
 902
 903         if (!dm_device.host_specified_ha_region) {
 904                 covered = pfn_covered(pg_start, pfn_cnt);
 905                 if (covered < 0)
 906                         return 0;
 907
 908                 if (covered)
 909                         goto do_pg_range;
 910         }
 911
 912         /*
 913          * If the host has specified a hot-add range; deal with it first.
 914          */
 915
 916         if (rg_size != 0) {
 917                 ha_region = kzalloc(sizeof(struct hv_hotadd_state), GFP_KERNEL);
 918                 if (!ha_region)
 919                         return 0;
 920
 921                 INIT_LIST_HEAD(&ha_region->list);
 922                 INIT_LIST_HEAD(&ha_region->gap_list);
 923
 924                 ha_region->start_pfn = rg_start;
 925                 ha_region->ha_end_pfn = rg_start;
 926                 ha_region->covered_start_pfn = pg_start;
 927                 ha_region->covered_end_pfn = pg_start;
 928                 ha_region->end_pfn = rg_start + rg_size;
 929
 930                 spin_lock_irqsave(&dm_device.ha_lock, flags);
 931                 list_add_tail(&ha_region->list, &dm_device.ha_region_list);
 932                 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 933         }
 934
 935 do_pg_range:
 936         /*
 937          * Process the page range specified; bringing them
 938          * online if possible.
 939          */
 940         return handle_pg_range(pg_start, pfn_cnt);
 941 }
 942
 943 #endif
 944
 945 static void hot_add_req(struct work_struct *dummy)
 946 {
 947         struct dm_hot_add_response resp;
 948 #ifdef CONFIG_MEMORY_HOTPLUG
 949         unsigned long pg_start, pfn_cnt;
 950         unsigned long rg_start, rg_sz;
 951 #endif
 952         struct hv_dynmem_device *dm = &dm_device;
 953
 954         memset(&resp, 0, sizeof(struct dm_hot_add_response));
 955         resp.hdr.type = DM_MEM_HOT_ADD_RESPONSE;
 956         resp.hdr.size = sizeof(struct dm_hot_add_response);
 957
 958 #ifdef CONFIG_MEMORY_HOTPLUG
 959         pg_start = dm->ha_wrk.ha_page_range.finfo.start_page;
 960         pfn_cnt = dm->ha_wrk.ha_page_range.finfo.page_cnt;
 961
 962         rg_start = dm->ha_wrk.ha_region_range.finfo.start_page;
 963         rg_sz = dm->ha_wrk.ha_region_range.finfo.page_cnt;
 964
 965         if ((rg_start == 0) && (!dm->host_specified_ha_region)) {
 966                 unsigned long region_size;
 967                 unsigned long region_start;
 968
 969                 /*
 970                  * The host has not specified the hot-add region.
 971                  * Based on the hot-add page range being specified,
 972                  * compute a hot-add region that can cover the pages
 973                  * that need to be hot-added while ensuring the alignment
 974                  * and size requirements of Linux as it relates to hot-add.
 975                  */
 976                 region_start = pg_start;
 977                 region_size = (pfn_cnt / HA_CHUNK) * HA_CHUNK;
 978                 if (pfn_cnt % HA_CHUNK)
 979                         region_size += HA_CHUNK;
 980
 981                 region_start = (pg_start / HA_CHUNK) * HA_CHUNK;
 982
 983                 rg_start = region_start;
 984                 rg_sz = region_size;
 985         }
 986
 987         if (do_hot_add)
 988                 resp.page_count = process_hot_add(pg_start, pfn_cnt,
 989                                                 rg_start, rg_sz);
 990
 991         dm->num_pages_added += resp.page_count;
 992 #endif
 993         /*
 994          * The result field of the response structure has the
 995          * following semantics:
 996          *
 997          * 1. If all or some pages hot-added: Guest should return success.
 998          *
 999          * 2. If no pages could be hot-added:
1000          *
1001          * If the guest returns success, then the host
1002          * will not attempt any further hot-add operations. This
1003          * signifies a permanent failure.
1004          *
1005          * If the guest returns failure, then this failure will be
1006          * treated as a transient failure and the host may retry the
1007          * hot-add operation after some delay.
1008          */
1009         if (resp.page_count > 0)
1010                 resp.result = 1;
1011         else if (!do_hot_add)
1012                 resp.result = 1;
1013         else
1014                 resp.result = 0;
1015
1016         if (!do_hot_add || (resp.page_count == 0))
1017                 pr_info("Memory hot add failed\n");
1018
1019         dm->state = DM_INITIALIZED;
1020         resp.hdr.trans_id = atomic_inc_return(&trans_id);
1021         vmbus_sendpacket(dm->dev->channel, &resp,
1022                         sizeof(struct dm_hot_add_response),
1023                         (unsigned long)NULL,
1024                         VM_PKT_DATA_INBAND, 0);
1025 }
1026
1027 static void process_info(struct hv_dynmem_device *dm, struct dm_info_msg *msg)
1028 {
1029         struct dm_info_header *info_hdr;
1030
1031         info_hdr = (struct dm_info_header *)msg->info;
1032
1033         switch (info_hdr->type) {
1034         case INFO_TYPE_MAX_PAGE_CNT:
1035                 if (info_hdr->data_size == sizeof(__u64)) {
1036                         __u64 *max_page_count = (__u64 *)&info_hdr[1];
1037
1038                         pr_info("Max. dynamic memory size: %llu MB\n",
1039                                 (*max_page_count) >> (20 - PAGE_SHIFT));
1040                 }
1041
1042                 break;
1043         default:
1044                 pr_info("Received Unknown type: %d\n", info_hdr->type);
1045         }
1046 }
1047
1048 static unsigned long compute_balloon_floor(void)
1049 {
1050         unsigned long min_pages;
1051 #define MB2PAGES(mb) ((mb) << (20 - PAGE_SHIFT))
1052         /* Simple continuous piecewiese linear function:
1053          *  max MiB -> min MiB  gradient
1054          *       0         0
1055          *      16        16
1056          *      32        24
1057          *     128        72    (1/2)
1058          *     512       168    (1/4)
1059          *    2048       360    (1/8)
1060          *    8192       744    (1/16)
1061          *   32768      1512    (1/32)
1062          */
1063         if (totalram_pages < MB2PAGES(128))
1064                 min_pages = MB2PAGES(8) + (totalram_pages >> 1);
1065         else if (totalram_pages < MB2PAGES(512))
1066                 min_pages = MB2PAGES(40) + (totalram_pages >> 2);
1067         else if (totalram_pages < MB2PAGES(2048))
1068                 min_pages = MB2PAGES(104) + (totalram_pages >> 3);
1069         else if (totalram_pages < MB2PAGES(8192))
1070                 min_pages = MB2PAGES(232) + (totalram_pages >> 4);
1071         else
1072                 min_pages = MB2PAGES(488) + (totalram_pages >> 5);
1073 #undef MB2PAGES
1074         return min_pages;
1075 }
1076
1077 /*
1078  * Post our status as it relates memory pressure to the
1079  * host. Host expects the guests to post this status
1080  * periodically at 1 second intervals.
1081  *
1082  * The metrics specified in this protocol are very Windows
1083  * specific and so we cook up numbers here to convey our memory
1084  * pressure.
1085  */
1086
1087 static void post_status(struct hv_dynmem_device *dm)
1088 {
1089         struct dm_status status;
1090         unsigned long now = jiffies;
1091         unsigned long last_post = last_post_time;
1092
1093         if (pressure_report_delay > 0) {
1094                 --pressure_report_delay;
1095                 return;
1096         }
1097
1098         if (!time_after(now, (last_post_time + HZ)))
1099                 return;
1100
1101         memset(&status, 0, sizeof(struct dm_status));
1102         status.hdr.type = DM_STATUS_REPORT;
1103         status.hdr.size = sizeof(struct dm_status);
1104         status.hdr.trans_id = atomic_inc_return(&trans_id);
1105
1106         /*
1107          * The host expects the guest to report free and committed memory.
1108          * Furthermore, the host expects the pressure information to include
1109          * the ballooned out pages. For a given amount of memory that we are
1110          * managing we need to compute a floor below which we should not
1111          * balloon. Compute this and add it to the pressure report.
1112          * We also need to report all offline pages (num_pages_added -
1113          * num_pages_onlined) as committed to the host, otherwise it can try
1114          * asking us to balloon them out.
1115          */
1116         status.num_avail = si_mem_available();
1117         status.num_committed = vm_memory_committed() +
1118                 dm->num_pages_ballooned +
1119                 (dm->num_pages_added > dm->num_pages_onlined ?
1120                  dm->num_pages_added - dm->num_pages_onlined : 0) +
1121                 compute_balloon_floor();
1122
1123         /*
1124          * If our transaction ID is no longer current, just don't
1125          * send the status. This can happen if we were interrupted
1126          * after we picked our transaction ID.
1127          */
1128         if (status.hdr.trans_id != atomic_read(&trans_id))
1129                 return;
1130
1131         /*
1132          * If the last post time that we sampled has changed,
1133          * we have raced, don't post the status.
1134          */
1135         if (last_post != last_post_time)
1136                 return;
1137
1138         last_post_time = jiffies;
1139         vmbus_sendpacket(dm->dev->channel, &status,
1140                                 sizeof(struct dm_status),
1141                                 (unsigned long)NULL,
1142                                 VM_PKT_DATA_INBAND, 0);
1143
1144 }
1145
1146 static void free_balloon_pages(struct hv_dynmem_device *dm,
1147                          union dm_mem_page_range *range_array)
1148 {
1149         int num_pages = range_array->finfo.page_cnt;
1150         __u64 start_frame = range_array->finfo.start_page;
1151         struct page *pg;
1152         int i;
1153
1154         for (i = 0; i < num_pages; i++) {
1155                 pg = pfn_to_page(i + start_frame);
1156                 __free_page(pg);
1157                 dm->num_pages_ballooned--;
1158         }
1159 }
1160
1161
1162
1163 static unsigned int alloc_balloon_pages(struct hv_dynmem_device *dm,
1164                                         unsigned int num_pages,
1165                                         struct dm_balloon_response *bl_resp,
1166                                         int alloc_unit)
1167 {
1168         unsigned int i = 0;
1169         struct page *pg;
1170
1171         if (num_pages < alloc_unit)
1172                 return 0;
1173
1174         for (i = 0; (i * alloc_unit) < num_pages; i++) {
1175                 if (bl_resp->hdr.size + sizeof(union dm_mem_page_range) >
1176                         PAGE_SIZE)
1177                         return i * alloc_unit;
1178
1179                 /*
1180                  * We execute this code in a thread context. Furthermore,
1181                  * we don't want the kernel to try too hard.
1182                  */
1183                 pg = alloc_pages(GFP_HIGHUSER | __GFP_NORETRY |
1184                                 __GFP_NOMEMALLOC | __GFP_NOWARN,
1185                                 get_order(alloc_unit << PAGE_SHIFT));
1186
1187                 if (!pg)
1188                         return i * alloc_unit;
1189
1190                 dm->num_pages_ballooned += alloc_unit;
1191
1192                 /*
1193                  * If we allocatted 2M pages; split them so we
1194                  * can free them in any order we get.
1195                  */
1196
1197                 if (alloc_unit != 1)
1198                         split_page(pg, get_order(alloc_unit << PAGE_SHIFT));
1199
1200                 bl_resp->range_count++;
1201                 bl_resp->range_array[i].finfo.start_page =
1202                         page_to_pfn(pg);
1203                 bl_resp->range_array[i].finfo.page_cnt = alloc_unit;
1204                 bl_resp->hdr.size += sizeof(union dm_mem_page_range);
1205
1206         }
1207
1208         return num_pages;
1209 }
1210
1211 static void balloon_up(struct work_struct *dummy)
1212 {
1213         unsigned int num_pages = dm_device.balloon_wrk.num_pages;
1214         unsigned int num_ballooned = 0;
1215         struct dm_balloon_response *bl_resp;
1216         int alloc_unit;
1217         int ret;
1218         bool done = false;
1219         int i;
1220         long avail_pages;
1221         unsigned long floor;
1222
1223         /* The host balloons pages in 2M granularity. */
1224         WARN_ON_ONCE(num_pages % PAGES_IN_2M != 0);
1225
1226         /*
1227          * We will attempt 2M allocations. However, if we fail to
1228          * allocate 2M chunks, we will go back to 4k allocations.
1229          */
1230         alloc_unit = 512;
1231
1232         avail_pages = si_mem_available();
1233         floor = compute_balloon_floor();
1234
1235         /* Refuse to balloon below the floor, keep the 2M granularity. */
1236         if (avail_pages < num_pages || avail_pages - num_pages < floor) {
1237                 pr_warn("Balloon request will be partially fulfilled. %s\n",
1238                         avail_pages < num_pages ? "Not enough memory." :
1239                         "Balloon floor reached.");
1240
1241                 num_pages = avail_pages > floor ? (avail_pages - floor) : 0;
1242                 num_pages -= num_pages % PAGES_IN_2M;
1243         }
1244
1245         while (!done) {
1246                 bl_resp = (struct dm_balloon_response *)send_buffer;
1247                 memset(send_buffer, 0, PAGE_SIZE);
1248                 bl_resp->hdr.type = DM_BALLOON_RESPONSE;
1249                 bl_resp->hdr.size = sizeof(struct dm_balloon_response);
1250                 bl_resp->more_pages = 1;
1251
1252                 num_pages -= num_ballooned;
1253                 num_ballooned = alloc_balloon_pages(&dm_device, num_pages,
1254                                                     bl_resp, alloc_unit);
1255
1256                 if (alloc_unit != 1 && num_ballooned == 0) {
1257                         alloc_unit = 1;
1258                         continue;
1259                 }
1260
1261                 if (num_ballooned == 0 || num_ballooned == num_pages) {
1262                         pr_debug("Ballooned %u out of %u requested pages.\n",
1263                                 num_pages, dm_device.balloon_wrk.num_pages);
1264
1265                         bl_resp->more_pages = 0;
1266                         done = true;
1267                         dm_device.state = DM_INITIALIZED;
1268                 }
1269
1270                 /*
1271                  * We are pushing a lot of data through the channel;
1272                  * deal with transient failures caused because of the
1273                  * lack of space in the ring buffer.
1274                  */
1275
1276                 do {
1277                         bl_resp->hdr.trans_id = atomic_inc_return(&trans_id);
1278                         ret = vmbus_sendpacket(dm_device.dev->channel,
1279                                                 bl_resp,
1280                                                 bl_resp->hdr.size,
1281                                                 (unsigned long)NULL,
1282                                                 VM_PKT_DATA_INBAND, 0);
1283
1284                         if (ret == -EAGAIN)
1285                                 msleep(20);
1286                         post_status(&dm_device);
1287                 } while (ret == -EAGAIN);
1288
1289                 if (ret) {
1290                         /*
1291                          * Free up the memory we allocatted.
1292                          */
1293                         pr_info("Balloon response failed\n");
1294
1295                         for (i = 0; i < bl_resp->range_count; i++)
1296                                 free_balloon_pages(&dm_device,
1297                                                  &bl_resp->range_array[i]);
1298
1299                         done = true;
1300                 }
1301         }
1302
1303 }
1304
1305 static void balloon_down(struct hv_dynmem_device *dm,
1306                         struct dm_unballoon_request *req)
1307 {
1308         union dm_mem_page_range *range_array = req->range_array;
1309         int range_count = req->range_count;
1310         struct dm_unballoon_response resp;
1311         int i;
1312         unsigned int prev_pages_ballooned = dm->num_pages_ballooned;
1313
1314         for (i = 0; i < range_count; i++) {
1315                 free_balloon_pages(dm, &range_array[i]);
1316                 complete(&dm_device.config_event);
1317         }
1318
1319         pr_debug("Freed %u ballooned pages.\n",
1320                 prev_pages_ballooned - dm->num_pages_ballooned);
1321
1322         if (req->more_pages == 1)
1323                 return;
1324
1325         memset(&resp, 0, sizeof(struct dm_unballoon_response));
1326         resp.hdr.type = DM_UNBALLOON_RESPONSE;
1327         resp.hdr.trans_id = atomic_inc_return(&trans_id);
1328         resp.hdr.size = sizeof(struct dm_unballoon_response);
1329
1330         vmbus_sendpacket(dm_device.dev->channel, &resp,
1331                                 sizeof(struct dm_unballoon_response),
1332                                 (unsigned long)NULL,
1333                                 VM_PKT_DATA_INBAND, 0);
1334
1335         dm->state = DM_INITIALIZED;
1336 }
1337
1338 static void balloon_onchannelcallback(void *context);
1339
1340 static int dm_thread_func(void *dm_dev)
1341 {
1342         struct hv_dynmem_device *dm = dm_dev;
1343
1344         while (!kthread_should_stop()) {
1345                 wait_for_completion_interruptible_timeout(
1346                                                 &dm_device.config_event, 1*HZ);
1347                 /*
1348                  * The host expects us to post information on the memory
1349                  * pressure every second.
1350                  */
1351                 reinit_completion(&dm_device.config_event);
1352                 post_status(dm);
1353         }
1354
1355         return 0;
1356 }
1357
1358
1359 static void version_resp(struct hv_dynmem_device *dm,
1360                         struct dm_version_response *vresp)
1361 {
1362         struct dm_version_request version_req;
1363         int ret;
1364
1365         if (vresp->is_accepted) {
1366                 /*
1367                  * We are done; wakeup the
1368                  * context waiting for version
1369                  * negotiation.
1370                  */
1371                 complete(&dm->host_event);
1372                 return;
1373         }
1374         /*
1375          * If there are more versions to try, continue
1376          * with negotiations; if not
1377          * shutdown the service since we are not able
1378          * to negotiate a suitable version number
1379          * with the host.
1380          */
1381         if (dm->next_version == 0)
1382                 goto version_error;
1383
1384         memset(&version_req, 0, sizeof(struct dm_version_request));
1385         version_req.hdr.type = DM_VERSION_REQUEST;
1386         version_req.hdr.size = sizeof(struct dm_version_request);
1387         version_req.hdr.trans_id = atomic_inc_return(&trans_id);
1388         version_req.version.version = dm->next_version;
1389         dm->version = version_req.version.version;
1390
1391         /*
1392          * Set the next version to try in case current version fails.
1393          * Win7 protocol ought to be the last one to try.
1394          */
1395         switch (version_req.version.version) {
1396         case DYNMEM_PROTOCOL_VERSION_WIN8:
1397                 dm->next_version = DYNMEM_PROTOCOL_VERSION_WIN7;
1398                 version_req.is_last_attempt = 0;
1399                 break;
1400         default:
1401                 dm->next_version = 0;
1402                 version_req.is_last_attempt = 1;
1403         }
1404
1405         ret = vmbus_sendpacket(dm->dev->channel, &version_req,
1406                                 sizeof(struct dm_version_request),
1407                                 (unsigned long)NULL,
1408                                 VM_PKT_DATA_INBAND, 0);
1409
1410         if (ret)
1411                 goto version_error;
1412
1413         return;
1414
1415 version_error:
1416         dm->state = DM_INIT_ERROR;
1417         complete(&dm->host_event);
1418 }
1419
1420 static void cap_resp(struct hv_dynmem_device *dm,
1421                         struct dm_capabilities_resp_msg *cap_resp)
1422 {
1423         if (!cap_resp->is_accepted) {
1424                 pr_info("Capabilities not accepted by host\n");
1425                 dm->state = DM_INIT_ERROR;
1426         }
1427         complete(&dm->host_event);
1428 }
1429
1430 static void balloon_onchannelcallback(void *context)
1431 {
1432         struct hv_device *dev = context;
1433         u32 recvlen;
1434         u64 requestid;
1435         struct dm_message *dm_msg;
1436         struct dm_header *dm_hdr;
1437         struct hv_dynmem_device *dm = hv_get_drvdata(dev);
1438         struct dm_balloon *bal_msg;
1439         struct dm_hot_add *ha_msg;
1440         union dm_mem_page_range *ha_pg_range;
1441         union dm_mem_page_range *ha_region;
1442
1443         memset(recv_buffer, 0, sizeof(recv_buffer));
1444         vmbus_recvpacket(dev->channel, recv_buffer,
1445                          PAGE_SIZE, &recvlen, &requestid);
1446
1447         if (recvlen > 0) {
1448                 dm_msg = (struct dm_message *)recv_buffer;
1449                 dm_hdr = &dm_msg->hdr;
1450
1451                 switch (dm_hdr->type) {
1452                 case DM_VERSION_RESPONSE:
1453                         version_resp(dm,
1454                                  (struct dm_version_response *)dm_msg);
1455                         break;
1456
1457                 case DM_CAPABILITIES_RESPONSE:
1458                         cap_resp(dm,
1459                                  (struct dm_capabilities_resp_msg *)dm_msg);
1460                         break;
1461
1462                 case DM_BALLOON_REQUEST:
1463                         if (dm->state == DM_BALLOON_UP)
1464                                 pr_warn("Currently ballooning\n");
1465                         bal_msg = (struct dm_balloon *)recv_buffer;
1466                         dm->state = DM_BALLOON_UP;
1467                         dm_device.balloon_wrk.num_pages = bal_msg->num_pages;
1468                         schedule_work(&dm_device.balloon_wrk.wrk);
1469                         break;
1470
1471                 case DM_UNBALLOON_REQUEST:
1472                         dm->state = DM_BALLOON_DOWN;
1473                         balloon_down(dm,
1474                                  (struct dm_unballoon_request *)recv_buffer);
1475                         break;
1476
1477                 case DM_MEM_HOT_ADD_REQUEST:
1478                         if (dm->state == DM_HOT_ADD)
1479                                 pr_warn("Currently hot-adding\n");
1480                         dm->state = DM_HOT_ADD;
1481                         ha_msg = (struct dm_hot_add *)recv_buffer;
1482                         if (ha_msg->hdr.size == sizeof(struct dm_hot_add)) {
1483                                 /*
1484                                  * This is a normal hot-add request specifying
1485                                  * hot-add memory.
1486                                  */
1487                                 dm->host_specified_ha_region = false;
1488                                 ha_pg_range = &ha_msg->range;
1489                                 dm->ha_wrk.ha_page_range = *ha_pg_range;
1490                                 dm->ha_wrk.ha_region_range.page_range = 0;
1491                         } else {
1492                                 /*
1493                                  * Host is specifying that we first hot-add
1494                                  * a region and then partially populate this
1495                                  * region.
1496                                  */
1497                                 dm->host_specified_ha_region = true;
1498                                 ha_pg_range = &ha_msg->range;
1499                                 ha_region = &ha_pg_range[1];
1500                                 dm->ha_wrk.ha_page_range = *ha_pg_range;
1501                                 dm->ha_wrk.ha_region_range = *ha_region;
1502                         }
1503                         schedule_work(&dm_device.ha_wrk.wrk);
1504                         break;
1505
1506                 case DM_INFO_MESSAGE:
1507                         process_info(dm, (struct dm_info_msg *)dm_msg);
1508                         break;
1509
1510                 default:
1511                         pr_err("Unhandled message: type: %d\n", dm_hdr->type);
1512
1513                 }
1514         }
1515
1516 }
1517
1518 static int balloon_probe(struct hv_device *dev,
1519                         const struct hv_vmbus_device_id *dev_id)
1520 {
1521         int ret;
1522         unsigned long t;
1523         struct dm_version_request version_req;
1524         struct dm_capabilities cap_msg;
1525
1526 #ifdef CONFIG_MEMORY_HOTPLUG
1527         do_hot_add = hot_add;
1528 #else
1529         do_hot_add = false;
1530 #endif
1531
1532         /*
1533          * First allocate a send buffer.
1534          */
1535
1536         send_buffer = kmalloc(PAGE_SIZE, GFP_KERNEL);
1537         if (!send_buffer)
1538                 return -ENOMEM;
1539
1540         ret = vmbus_open(dev->channel, dm_ring_size, dm_ring_size, NULL, 0,
1541                         balloon_onchannelcallback, dev);
1542
1543         if (ret)
1544                 goto probe_error0;
1545
1546         dm_device.dev = dev;
1547         dm_device.state = DM_INITIALIZING;
1548         dm_device.next_version = DYNMEM_PROTOCOL_VERSION_WIN8;
1549         init_completion(&dm_device.host_event);
1550         init_completion(&dm_device.config_event);
1551         INIT_LIST_HEAD(&dm_device.ha_region_list);
1552         spin_lock_init(&dm_device.ha_lock);
1553         INIT_WORK(&dm_device.balloon_wrk.wrk, balloon_up);
1554         INIT_WORK(&dm_device.ha_wrk.wrk, hot_add_req);
1555         dm_device.host_specified_ha_region = false;
1556
1557         dm_device.thread =
1558                  kthread_run(dm_thread_func, &dm_device, "hv_balloon");
1559         if (IS_ERR(dm_device.thread)) {
1560                 ret = PTR_ERR(dm_device.thread);
1561                 goto probe_error1;
1562         }
1563
1564 #ifdef CONFIG_MEMORY_HOTPLUG
1565         set_online_page_callback(&hv_online_page);
1566         register_memory_notifier(&hv_memory_nb);
1567 #endif
1568
1569         hv_set_drvdata(dev, &dm_device);
1570         /*
1571          * Initiate the hand shake with the host and negotiate
1572          * a version that the host can support. We start with the
1573          * highest version number and go down if the host cannot
1574          * support it.
1575          */
1576         memset(&version_req, 0, sizeof(struct dm_version_request));
1577         version_req.hdr.type = DM_VERSION_REQUEST;
1578         version_req.hdr.size = sizeof(struct dm_version_request);
1579         version_req.hdr.trans_id = atomic_inc_return(&trans_id);
1580         version_req.version.version = DYNMEM_PROTOCOL_VERSION_WIN10;
1581         version_req.is_last_attempt = 0;
1582         dm_device.version = version_req.version.version;
1583
1584         ret = vmbus_sendpacket(dev->channel, &version_req,
1585                                 sizeof(struct dm_version_request),
1586                                 (unsigned long)NULL,
1587                                 VM_PKT_DATA_INBAND, 0);
1588         if (ret)
1589                 goto probe_error2;
1590
1591         t = wait_for_completion_timeout(&dm_device.host_event, 5*HZ);
1592         if (t == 0) {
1593                 ret = -ETIMEDOUT;
1594                 goto probe_error2;
1595         }
1596
1597         /*
1598          * If we could not negotiate a compatible version with the host
1599          * fail the probe function.
1600          */
1601         if (dm_device.state == DM_INIT_ERROR) {
1602                 ret = -ETIMEDOUT;
1603                 goto probe_error2;
1604         }
1605
1606         pr_info("Using Dynamic Memory protocol version %u.%u\n",
1607                 DYNMEM_MAJOR_VERSION(dm_device.version),
1608                 DYNMEM_MINOR_VERSION(dm_device.version));
1609
1610         /*
1611          * Now submit our capabilities to the host.
1612          */
1613         memset(&cap_msg, 0, sizeof(struct dm_capabilities));
1614         cap_msg.hdr.type = DM_CAPABILITIES_REPORT;
1615         cap_msg.hdr.size = sizeof(struct dm_capabilities);
1616         cap_msg.hdr.trans_id = atomic_inc_return(&trans_id);
1617
1618         cap_msg.caps.cap_bits.balloon = 1;
1619         cap_msg.caps.cap_bits.hot_add = 1;
1620
1621         /*
1622          * Specify our alignment requirements as it relates
1623          * memory hot-add. Specify 128MB alignment.
1624          */
1625         cap_msg.caps.cap_bits.hot_add_alignment = 7;
1626
1627         /*
1628          * Currently the host does not use these
1629          * values and we set them to what is done in the
1630          * Windows driver.
1631          */
1632         cap_msg.min_page_cnt = 0;
1633         cap_msg.max_page_number = -1;
1634
1635         ret = vmbus_sendpacket(dev->channel, &cap_msg,
1636                                 sizeof(struct dm_capabilities),
1637                                 (unsigned long)NULL,
1638                                 VM_PKT_DATA_INBAND, 0);
1639         if (ret)
1640                 goto probe_error2;
1641
1642         t = wait_for_completion_timeout(&dm_device.host_event, 5*HZ);
1643         if (t == 0) {
1644                 ret = -ETIMEDOUT;
1645                 goto probe_error2;
1646         }
1647
1648         /*
1649          * If the host does not like our capabilities,
1650          * fail the probe function.
1651          */
1652         if (dm_device.state == DM_INIT_ERROR) {
1653                 ret = -ETIMEDOUT;
1654                 goto probe_error2;
1655         }
1656
1657         dm_device.state = DM_INITIALIZED;
1658         last_post_time = jiffies;
1659
1660         return 0;
1661
1662 probe_error2:
1663 #ifdef CONFIG_MEMORY_HOTPLUG
1664         restore_online_page_callback(&hv_online_page);
1665 #endif
1666         kthread_stop(dm_device.thread);
1667
1668 probe_error1:
1669         vmbus_close(dev->channel);
1670 probe_error0:
1671         kfree(send_buffer);
1672         return ret;
1673 }
1674
1675 static int balloon_remove(struct hv_device *dev)
1676 {
1677         struct hv_dynmem_device *dm = hv_get_drvdata(dev);
1678         struct hv_hotadd_state *has, *tmp;
1679         struct hv_hotadd_gap *gap, *tmp_gap;
1680         unsigned long flags;
1681
1682         if (dm->num_pages_ballooned != 0)
1683                 pr_warn("Ballooned pages: %d\n", dm->num_pages_ballooned);
1684
1685         cancel_work_sync(&dm->balloon_wrk.wrk);
1686         cancel_work_sync(&dm->ha_wrk.wrk);
1687
1688         vmbus_close(dev->channel);
1689         kthread_stop(dm->thread);
1690         kfree(send_buffer);
1691 #ifdef CONFIG_MEMORY_HOTPLUG
1692         restore_online_page_callback(&hv_online_page);
1693         unregister_memory_notifier(&hv_memory_nb);
1694 #endif
1695         spin_lock_irqsave(&dm_device.ha_lock, flags);
1696         list_for_each_entry_safe(has, tmp, &dm->ha_region_list, list) {
1697                 list_for_each_entry_safe(gap, tmp_gap, &has->gap_list, list) {
1698                         list_del(&gap->list);
1699                         kfree(gap);
1700                 }
1701                 list_del(&has->list);
1702                 kfree(has);
1703         }
1704         spin_unlock_irqrestore(&dm_device.ha_lock, flags);
1705
1706         return 0;
1707 }
1708
1709 static const struct hv_vmbus_device_id id_table[] = {
1710         /* Dynamic Memory Class ID */
1711         /* 525074DC-8985-46e2-8057-A307DC18A502 */
1712         { HV_DM_GUID, },
1713         { },
1714 };
1715
1716 MODULE_DEVICE_TABLE(vmbus, id_table);
1717
1718 static  struct hv_driver balloon_drv = {
1719         .name = "hv_balloon",
1720         .id_table = id_table,
1721         .probe =  balloon_probe,
1722         .remove =  balloon_remove,
1723 };
1724
1725 static int __init init_balloon_drv(void)
1726 {
1727
1728         return vmbus_driver_register(&balloon_drv);
1729 }
1730
1731 module_init(init_balloon_drv);
1732
1733 MODULE_DESCRIPTION("Hyper-V Balloon");
1734 MODULE_LICENSE("GPL");