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 static struct hv_dynmem_device dm_device;
 570
 571 static void post_status(struct hv_dynmem_device *dm);
 572
 573 #ifdef CONFIG_MEMORY_HOTPLUG
 574 static int hv_memory_notifier(struct notifier_block *nb, unsigned long val,
 575                               void *v)
 576 {
 577         struct memory_notify *mem = (struct memory_notify *)v;
 578         unsigned long flags;
 579
 580         switch (val) {
 581         case MEM_ONLINE:
 582                 spin_lock_irqsave(&dm_device.ha_lock, flags);
 583                 dm_device.num_pages_onlined += mem->nr_pages;
 584                 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 585         case MEM_CANCEL_ONLINE:
 586                 if (dm_device.ha_waiting) {
 587                         dm_device.ha_waiting = false;
 588                         complete(&dm_device.ol_waitevent);
 589                 }
 590                 break;
 591
 592         case MEM_OFFLINE:
 593                 spin_lock_irqsave(&dm_device.ha_lock, flags);
 594                 dm_device.num_pages_onlined -= mem->nr_pages;
 595                 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 596                 break;
 597         case MEM_GOING_ONLINE:
 598         case MEM_GOING_OFFLINE:
 599         case MEM_CANCEL_OFFLINE:
 600                 break;
 601         }
 602         return NOTIFY_OK;
 603 }
 604
 605 static struct notifier_block hv_memory_nb = {
 606         .notifier_call = hv_memory_notifier,
 607         .priority = 0
 608 };
 609
 610 /* Check if the particular page is backed and can be onlined and online it. */
 611 static void hv_page_online_one(struct hv_hotadd_state *has, struct page *pg)
 612 {
 613         unsigned long cur_start_pgp;
 614         unsigned long cur_end_pgp;
 615         struct hv_hotadd_gap *gap;
 616
 617         cur_start_pgp = (unsigned long)pfn_to_page(has->covered_start_pfn);
 618         cur_end_pgp = (unsigned long)pfn_to_page(has->covered_end_pfn);
 619
 620         /* The page is not backed. */
 621         if (((unsigned long)pg < cur_start_pgp) ||
 622             ((unsigned long)pg >= cur_end_pgp))
 623                 return;
 624
 625         /* Check for gaps. */
 626         list_for_each_entry(gap, &has->gap_list, list) {
 627                 cur_start_pgp = (unsigned long)
 628                         pfn_to_page(gap->start_pfn);
 629                 cur_end_pgp = (unsigned long)
 630                         pfn_to_page(gap->end_pfn);
 631                 if (((unsigned long)pg >= cur_start_pgp) &&
 632                     ((unsigned long)pg < cur_end_pgp)) {
 633                         return;
 634                 }
 635         }
 636
 637         /* This frame is currently backed; online the page. */
 638         __online_page_set_limits(pg);
 639         __online_page_increment_counters(pg);
 640         __online_page_free(pg);
 641 }
 642
 643 static void hv_bring_pgs_online(struct hv_hotadd_state *has,
 644                                 unsigned long start_pfn, unsigned long size)
 645 {
 646         int i;
 647
 648         for (i = 0; i < size; i++)
 649                 hv_page_online_one(has, pfn_to_page(start_pfn + i));
 650 }
 651
 652 static void hv_mem_hot_add(unsigned long start, unsigned long size,
 653                                 unsigned long pfn_count,
 654                                 struct hv_hotadd_state *has)
 655 {
 656         int ret = 0;
 657         int i, nid;
 658         unsigned long start_pfn;
 659         unsigned long processed_pfn;
 660         unsigned long total_pfn = pfn_count;
 661         unsigned long flags;
 662
 663         for (i = 0; i < (size/HA_CHUNK); i++) {
 664                 start_pfn = start + (i * HA_CHUNK);
 665
 666                 spin_lock_irqsave(&dm_device.ha_lock, flags);
 667                 has->ha_end_pfn +=  HA_CHUNK;
 668
 669                 if (total_pfn > HA_CHUNK) {
 670                         processed_pfn = HA_CHUNK;
 671                         total_pfn -= HA_CHUNK;
 672                 } else {
 673                         processed_pfn = total_pfn;
 674                         total_pfn = 0;
 675                 }
 676
 677                 has->covered_end_pfn +=  processed_pfn;
 678                 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 679
 680                 init_completion(&dm_device.ol_waitevent);
 681                 dm_device.ha_waiting = !memhp_auto_online;
 682
 683                 nid = memory_add_physaddr_to_nid(PFN_PHYS(start_pfn));
 684                 ret = add_memory(nid, PFN_PHYS((start_pfn)),
 685                                 (HA_CHUNK << PAGE_SHIFT));
 686
 687                 if (ret) {
 688                         pr_info("hot_add memory failed error is %d\n", ret);
 689                         if (ret == -EEXIST) {
 690                                 /*
 691                                  * This error indicates that the error
 692                                  * is not a transient failure. This is the
 693                                  * case where the guest's physical address map
 694                                  * precludes hot adding memory. Stop all further
 695                                  * memory hot-add.
 696                                  */
 697                                 do_hot_add = false;
 698                         }
 699                         spin_lock_irqsave(&dm_device.ha_lock, flags);
 700                         has->ha_end_pfn -= HA_CHUNK;
 701                         has->covered_end_pfn -=  processed_pfn;
 702                         spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 703                         break;
 704                 }
 705
 706                 /*
 707                  * Wait for the memory block to be onlined when memory onlining
 708                  * is done outside of kernel (memhp_auto_online). Since the hot
 709                  * add has succeeded, it is ok to proceed even if the pages in
 710                  * the hot added region have not been "onlined" within the
 711                  * allowed time.
 712                  */
 713                 if (dm_device.ha_waiting)
 714                         wait_for_completion_timeout(&dm_device.ol_waitevent,
 715                                                     5*HZ);
 716                 post_status(&dm_device);
 717         }
 718
 719         return;
 720 }
 721
 722 static void hv_online_page(struct page *pg)
 723 {
 724         struct hv_hotadd_state *has;
 725         unsigned long cur_start_pgp;
 726         unsigned long cur_end_pgp;
 727         unsigned long flags;
 728
 729         spin_lock_irqsave(&dm_device.ha_lock, flags);
 730         list_for_each_entry(has, &dm_device.ha_region_list, list) {
 731                 cur_start_pgp = (unsigned long)
 732                         pfn_to_page(has->start_pfn);
 733                 cur_end_pgp = (unsigned long)pfn_to_page(has->end_pfn);
 734
 735                 /* The page belongs to a different HAS. */
 736                 if (((unsigned long)pg < cur_start_pgp) ||
 737                     ((unsigned long)pg >= cur_end_pgp))
 738                         continue;
 739
 740                 hv_page_online_one(has, pg);
 741                 break;
 742         }
 743         spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 744 }
 745
 746 static int pfn_covered(unsigned long start_pfn, unsigned long pfn_cnt)
 747 {
 748         struct hv_hotadd_state *has;
 749         struct hv_hotadd_gap *gap;
 750         unsigned long residual, new_inc;
 751         int ret = 0;
 752         unsigned long flags;
 753
 754         spin_lock_irqsave(&dm_device.ha_lock, flags);
 755         list_for_each_entry(has, &dm_device.ha_region_list, list) {
 756                 /*
 757                  * If the pfn range we are dealing with is not in the current
 758                  * "hot add block", move on.
 759                  */
 760                 if (start_pfn < has->start_pfn || start_pfn >= has->end_pfn)
 761                         continue;
 762
 763                 /*
 764                  * If the current start pfn is not where the covered_end
 765                  * is, create a gap and update covered_end_pfn.
 766                  */
 767                 if (has->covered_end_pfn != start_pfn) {
 768                         gap = kzalloc(sizeof(struct hv_hotadd_gap), GFP_ATOMIC);
 769                         if (!gap) {
 770                                 ret = -ENOMEM;
 771                                 break;
 772                         }
 773
 774                         INIT_LIST_HEAD(&gap->list);
 775                         gap->start_pfn = has->covered_end_pfn;
 776                         gap->end_pfn = start_pfn;
 777                         list_add_tail(&gap->list, &has->gap_list);
 778
 779                         has->covered_end_pfn = start_pfn;
 780                 }
 781
 782                 /*
 783                  * If the current hot add-request extends beyond
 784                  * our current limit; extend it.
 785                  */
 786                 if ((start_pfn + pfn_cnt) > has->end_pfn) {
 787                         residual = (start_pfn + pfn_cnt - has->end_pfn);
 788                         /*
 789                          * Extend the region by multiples of HA_CHUNK.
 790                          */
 791                         new_inc = (residual / HA_CHUNK) * HA_CHUNK;
 792                         if (residual % HA_CHUNK)
 793                                 new_inc += HA_CHUNK;
 794
 795                         has->end_pfn += new_inc;
 796                 }
 797
 798                 ret = 1;
 799                 break;
 800         }
 801         spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 802
 803         return ret;
 804 }
 805
 806 static unsigned long handle_pg_range(unsigned long pg_start,
 807                                         unsigned long pg_count)
 808 {
 809         unsigned long start_pfn = pg_start;
 810         unsigned long pfn_cnt = pg_count;
 811         unsigned long size;
 812         struct hv_hotadd_state *has;
 813         unsigned long pgs_ol = 0;
 814         unsigned long old_covered_state;
 815         unsigned long res = 0, flags;
 816
 817         spin_lock_irqsave(&dm_device.ha_lock, flags);
 818         list_for_each_entry(has, &dm_device.ha_region_list, list) {
 819                 /*
 820                  * If the pfn range we are dealing with is not in the current
 821                  * "hot add block", move on.
 822                  */
 823                 if (start_pfn < has->start_pfn || start_pfn >= has->end_pfn)
 824                         continue;
 825
 826                 old_covered_state = has->covered_end_pfn;
 827
 828                 if (start_pfn < has->ha_end_pfn) {
 829                         /*
 830                          * This is the case where we are backing pages
 831                          * in an already hot added region. Bring
 832                          * these pages online first.
 833                          */
 834                         pgs_ol = has->ha_end_pfn - start_pfn;
 835                         if (pgs_ol > pfn_cnt)
 836                                 pgs_ol = pfn_cnt;
 837
 838                         has->covered_end_pfn +=  pgs_ol;
 839                         pfn_cnt -= pgs_ol;
 840                         /*
 841                          * Check if the corresponding memory block is already
 842                          * online by checking its last previously backed page.
 843                          * In case it is we need to bring rest (which was not
 844                          * backed previously) online too.
 845                          */
 846                         if (start_pfn > has->start_pfn &&
 847                             !PageReserved(pfn_to_page(start_pfn - 1)))
 848                                 hv_bring_pgs_online(has, start_pfn, pgs_ol);
 849
 850                 }
 851
 852                 if ((has->ha_end_pfn < has->end_pfn) && (pfn_cnt > 0)) {
 853                         /*
 854                          * We have some residual hot add range
 855                          * that needs to be hot added; hot add
 856                          * it now. Hot add a multiple of
 857                          * of HA_CHUNK that fully covers the pages
 858                          * we have.
 859                          */
 860                         size = (has->end_pfn - has->ha_end_pfn);
 861                         if (pfn_cnt <= size) {
 862                                 size = ((pfn_cnt / HA_CHUNK) * HA_CHUNK);
 863                                 if (pfn_cnt % HA_CHUNK)
 864                                         size += HA_CHUNK;
 865                         } else {
 866                                 pfn_cnt = size;
 867                         }
 868                         spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 869                         hv_mem_hot_add(has->ha_end_pfn, size, pfn_cnt, has);
 870                         spin_lock_irqsave(&dm_device.ha_lock, flags);
 871                 }
 872                 /*
 873                  * If we managed to online any pages that were given to us,
 874                  * we declare success.
 875                  */
 876                 res = has->covered_end_pfn - old_covered_state;
 877                 break;
 878         }
 879         spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 880
 881         return res;
 882 }
 883
 884 static unsigned long process_hot_add(unsigned long pg_start,
 885                                         unsigned long pfn_cnt,
 886                                         unsigned long rg_start,
 887                                         unsigned long rg_size)
 888 {
 889         struct hv_hotadd_state *ha_region = NULL;
 890         int covered;
 891         unsigned long flags;
 892
 893         if (pfn_cnt == 0)
 894                 return 0;
 895
 896         if (!dm_device.host_specified_ha_region) {
 897                 covered = pfn_covered(pg_start, pfn_cnt);
 898                 if (covered < 0)
 899                         return 0;
 900
 901                 if (covered)
 902                         goto do_pg_range;
 903         }
 904
 905         /*
 906          * If the host has specified a hot-add range; deal with it first.
 907          */
 908
 909         if (rg_size != 0) {
 910                 ha_region = kzalloc(sizeof(struct hv_hotadd_state), GFP_KERNEL);
 911                 if (!ha_region)
 912                         return 0;
 913
 914                 INIT_LIST_HEAD(&ha_region->list);
 915                 INIT_LIST_HEAD(&ha_region->gap_list);
 916
 917                 ha_region->start_pfn = rg_start;
 918                 ha_region->ha_end_pfn = rg_start;
 919                 ha_region->covered_start_pfn = pg_start;
 920                 ha_region->covered_end_pfn = pg_start;
 921                 ha_region->end_pfn = rg_start + rg_size;
 922
 923                 spin_lock_irqsave(&dm_device.ha_lock, flags);
 924                 list_add_tail(&ha_region->list, &dm_device.ha_region_list);
 925                 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 926         }
 927
 928 do_pg_range:
 929         /*
 930          * Process the page range specified; bringing them
 931          * online if possible.
 932          */
 933         return handle_pg_range(pg_start, pfn_cnt);
 934 }
 935
 936 #endif
 937
 938 static void hot_add_req(struct work_struct *dummy)
 939 {
 940         struct dm_hot_add_response resp;
 941 #ifdef CONFIG_MEMORY_HOTPLUG
 942         unsigned long pg_start, pfn_cnt;
 943         unsigned long rg_start, rg_sz;
 944 #endif
 945         struct hv_dynmem_device *dm = &dm_device;
 946
 947         memset(&resp, 0, sizeof(struct dm_hot_add_response));
 948         resp.hdr.type = DM_MEM_HOT_ADD_RESPONSE;
 949         resp.hdr.size = sizeof(struct dm_hot_add_response);
 950
 951 #ifdef CONFIG_MEMORY_HOTPLUG
 952         pg_start = dm->ha_wrk.ha_page_range.finfo.start_page;
 953         pfn_cnt = dm->ha_wrk.ha_page_range.finfo.page_cnt;
 954
 955         rg_start = dm->ha_wrk.ha_region_range.finfo.start_page;
 956         rg_sz = dm->ha_wrk.ha_region_range.finfo.page_cnt;
 957
 958         if ((rg_start == 0) && (!dm->host_specified_ha_region)) {
 959                 unsigned long region_size;
 960                 unsigned long region_start;
 961
 962                 /*
 963                  * The host has not specified the hot-add region.
 964                  * Based on the hot-add page range being specified,
 965                  * compute a hot-add region that can cover the pages
 966                  * that need to be hot-added while ensuring the alignment
 967                  * and size requirements of Linux as it relates to hot-add.
 968                  */
 969                 region_start = pg_start;
 970                 region_size = (pfn_cnt / HA_CHUNK) * HA_CHUNK;
 971                 if (pfn_cnt % HA_CHUNK)
 972                         region_size += HA_CHUNK;
 973
 974                 region_start = (pg_start / HA_CHUNK) * HA_CHUNK;
 975
 976                 rg_start = region_start;
 977                 rg_sz = region_size;
 978         }
 979
 980         if (do_hot_add)
 981                 resp.page_count = process_hot_add(pg_start, pfn_cnt,
 982                                                 rg_start, rg_sz);
 983
 984         dm->num_pages_added += resp.page_count;
 985 #endif
 986         /*
 987          * The result field of the response structure has the
 988          * following semantics:
 989          *
 990          * 1. If all or some pages hot-added: Guest should return success.
 991          *
 992          * 2. If no pages could be hot-added:
 993          *
 994          * If the guest returns success, then the host
 995          * will not attempt any further hot-add operations. This
 996          * signifies a permanent failure.
 997          *
 998          * If the guest returns failure, then this failure will be
 999          * treated as a transient failure and the host may retry the
1000          * hot-add operation after some delay.
1001          */
1002         if (resp.page_count > 0)
1003                 resp.result = 1;
1004         else if (!do_hot_add)
1005                 resp.result = 1;
1006         else
1007                 resp.result = 0;
1008
1009         if (!do_hot_add || (resp.page_count == 0))
1010                 pr_info("Memory hot add failed\n");
1011
1012         dm->state = DM_INITIALIZED;
1013         resp.hdr.trans_id = atomic_inc_return(&trans_id);
1014         vmbus_sendpacket(dm->dev->channel, &resp,
1015                         sizeof(struct dm_hot_add_response),
1016                         (unsigned long)NULL,
1017                         VM_PKT_DATA_INBAND, 0);
1018 }
1019
1020 static void process_info(struct hv_dynmem_device *dm, struct dm_info_msg *msg)
1021 {
1022         struct dm_info_header *info_hdr;
1023
1024         info_hdr = (struct dm_info_header *)msg->info;
1025
1026         switch (info_hdr->type) {
1027         case INFO_TYPE_MAX_PAGE_CNT:
1028                 pr_info("Received INFO_TYPE_MAX_PAGE_CNT\n");
1029                 pr_info("Data Size is %d\n", info_hdr->data_size);
1030                 break;
1031         default:
1032                 pr_info("Received Unknown type: %d\n", info_hdr->type);
1033         }
1034 }
1035
1036 static unsigned long compute_balloon_floor(void)
1037 {
1038         unsigned long min_pages;
1039 #define MB2PAGES(mb) ((mb) << (20 - PAGE_SHIFT))
1040         /* Simple continuous piecewiese linear function:
1041          *  max MiB -> min MiB  gradient
1042          *       0         0
1043          *      16        16
1044          *      32        24
1045          *     128        72    (1/2)
1046          *     512       168    (1/4)
1047          *    2048       360    (1/8)
1048          *    8192       744    (1/16)
1049          *   32768      1512    (1/32)
1050          */
1051         if (totalram_pages < MB2PAGES(128))
1052                 min_pages = MB2PAGES(8) + (totalram_pages >> 1);
1053         else if (totalram_pages < MB2PAGES(512))
1054                 min_pages = MB2PAGES(40) + (totalram_pages >> 2);
1055         else if (totalram_pages < MB2PAGES(2048))
1056                 min_pages = MB2PAGES(104) + (totalram_pages >> 3);
1057         else if (totalram_pages < MB2PAGES(8192))
1058                 min_pages = MB2PAGES(232) + (totalram_pages >> 4);
1059         else
1060                 min_pages = MB2PAGES(488) + (totalram_pages >> 5);
1061 #undef MB2PAGES
1062         return min_pages;
1063 }
1064
1065 /*
1066  * Post our status as it relates memory pressure to the
1067  * host. Host expects the guests to post this status
1068  * periodically at 1 second intervals.
1069  *
1070  * The metrics specified in this protocol are very Windows
1071  * specific and so we cook up numbers here to convey our memory
1072  * pressure.
1073  */
1074
1075 static void post_status(struct hv_dynmem_device *dm)
1076 {
1077         struct dm_status status;
1078         struct sysinfo val;
1079         unsigned long now = jiffies;
1080         unsigned long last_post = last_post_time;
1081
1082         if (pressure_report_delay > 0) {
1083                 --pressure_report_delay;
1084                 return;
1085         }
1086
1087         if (!time_after(now, (last_post_time + HZ)))
1088                 return;
1089
1090         si_meminfo(&val);
1091         memset(&status, 0, sizeof(struct dm_status));
1092         status.hdr.type = DM_STATUS_REPORT;
1093         status.hdr.size = sizeof(struct dm_status);
1094         status.hdr.trans_id = atomic_inc_return(&trans_id);
1095
1096         /*
1097          * The host expects the guest to report free and committed memory.
1098          * Furthermore, the host expects the pressure information to include
1099          * the ballooned out pages. For a given amount of memory that we are
1100          * managing we need to compute a floor below which we should not
1101          * balloon. Compute this and add it to the pressure report.
1102          * We also need to report all offline pages (num_pages_added -
1103          * num_pages_onlined) as committed to the host, otherwise it can try
1104          * asking us to balloon them out.
1105          */
1106         status.num_avail = val.freeram;
1107         status.num_committed = vm_memory_committed() +
1108                 dm->num_pages_ballooned +
1109                 (dm->num_pages_added > dm->num_pages_onlined ?
1110                  dm->num_pages_added - dm->num_pages_onlined : 0) +
1111                 compute_balloon_floor();
1112
1113         /*
1114          * If our transaction ID is no longer current, just don't
1115          * send the status. This can happen if we were interrupted
1116          * after we picked our transaction ID.
1117          */
1118         if (status.hdr.trans_id != atomic_read(&trans_id))
1119                 return;
1120
1121         /*
1122          * If the last post time that we sampled has changed,
1123          * we have raced, don't post the status.
1124          */
1125         if (last_post != last_post_time)
1126                 return;
1127
1128         last_post_time = jiffies;
1129         vmbus_sendpacket(dm->dev->channel, &status,
1130                                 sizeof(struct dm_status),
1131                                 (unsigned long)NULL,
1132                                 VM_PKT_DATA_INBAND, 0);
1133
1134 }
1135
1136 static void free_balloon_pages(struct hv_dynmem_device *dm,
1137                          union dm_mem_page_range *range_array)
1138 {
1139         int num_pages = range_array->finfo.page_cnt;
1140         __u64 start_frame = range_array->finfo.start_page;
1141         struct page *pg;
1142         int i;
1143
1144         for (i = 0; i < num_pages; i++) {
1145                 pg = pfn_to_page(i + start_frame);
1146                 __free_page(pg);
1147                 dm->num_pages_ballooned--;
1148         }
1149 }
1150
1151
1152
1153 static unsigned int alloc_balloon_pages(struct hv_dynmem_device *dm,
1154                                         unsigned int num_pages,
1155                                         struct dm_balloon_response *bl_resp,
1156                                         int alloc_unit)
1157 {
1158         unsigned int i = 0;
1159         struct page *pg;
1160
1161         if (num_pages < alloc_unit)
1162                 return 0;
1163
1164         for (i = 0; (i * alloc_unit) < num_pages; i++) {
1165                 if (bl_resp->hdr.size + sizeof(union dm_mem_page_range) >
1166                         PAGE_SIZE)
1167                         return i * alloc_unit;
1168
1169                 /*
1170                  * We execute this code in a thread context. Furthermore,
1171                  * we don't want the kernel to try too hard.
1172                  */
1173                 pg = alloc_pages(GFP_HIGHUSER | __GFP_NORETRY |
1174                                 __GFP_NOMEMALLOC | __GFP_NOWARN,
1175                                 get_order(alloc_unit << PAGE_SHIFT));
1176
1177                 if (!pg)
1178                         return i * alloc_unit;
1179
1180                 dm->num_pages_ballooned += alloc_unit;
1181
1182                 /*
1183                  * If we allocatted 2M pages; split them so we
1184                  * can free them in any order we get.
1185                  */
1186
1187                 if (alloc_unit != 1)
1188                         split_page(pg, get_order(alloc_unit << PAGE_SHIFT));
1189
1190                 bl_resp->range_count++;
1191                 bl_resp->range_array[i].finfo.start_page =
1192                         page_to_pfn(pg);
1193                 bl_resp->range_array[i].finfo.page_cnt = alloc_unit;
1194                 bl_resp->hdr.size += sizeof(union dm_mem_page_range);
1195
1196         }
1197
1198         return num_pages;
1199 }
1200
1201
1202
1203 static void balloon_up(struct work_struct *dummy)
1204 {
1205         unsigned int num_pages = dm_device.balloon_wrk.num_pages;
1206         unsigned int num_ballooned = 0;
1207         struct dm_balloon_response *bl_resp;
1208         int alloc_unit;
1209         int ret;
1210         bool done = false;
1211         int i;
1212         struct sysinfo val;
1213         unsigned long floor;
1214
1215         /* The host balloons pages in 2M granularity. */
1216         WARN_ON_ONCE(num_pages % PAGES_IN_2M != 0);
1217
1218         /*
1219          * We will attempt 2M allocations. However, if we fail to
1220          * allocate 2M chunks, we will go back to 4k allocations.
1221          */
1222         alloc_unit = 512;
1223
1224         si_meminfo(&val);
1225         floor = compute_balloon_floor();
1226
1227         /* Refuse to balloon below the floor, keep the 2M granularity. */
1228         if (val.freeram < num_pages || val.freeram - num_pages < floor) {
1229                 num_pages = val.freeram > floor ? (val.freeram - floor) : 0;
1230                 num_pages -= num_pages % PAGES_IN_2M;
1231         }
1232
1233         while (!done) {
1234                 bl_resp = (struct dm_balloon_response *)send_buffer;
1235                 memset(send_buffer, 0, PAGE_SIZE);
1236                 bl_resp->hdr.type = DM_BALLOON_RESPONSE;
1237                 bl_resp->hdr.size = sizeof(struct dm_balloon_response);
1238                 bl_resp->more_pages = 1;
1239
1240
1241                 num_pages -= num_ballooned;
1242                 num_ballooned = alloc_balloon_pages(&dm_device, num_pages,
1243                                                     bl_resp, alloc_unit);
1244
1245                 if (alloc_unit != 1 && num_ballooned == 0) {
1246                         alloc_unit = 1;
1247                         continue;
1248                 }
1249
1250                 if (num_ballooned == 0 || num_ballooned == num_pages) {
1251                         bl_resp->more_pages = 0;
1252                         done = true;
1253                         dm_device.state = DM_INITIALIZED;
1254                 }
1255
1256                 /*
1257                  * We are pushing a lot of data through the channel;
1258                  * deal with transient failures caused because of the
1259                  * lack of space in the ring buffer.
1260                  */
1261
1262                 do {
1263                         bl_resp->hdr.trans_id = atomic_inc_return(&trans_id);
1264                         ret = vmbus_sendpacket(dm_device.dev->channel,
1265                                                 bl_resp,
1266                                                 bl_resp->hdr.size,
1267                                                 (unsigned long)NULL,
1268                                                 VM_PKT_DATA_INBAND, 0);
1269
1270                         if (ret == -EAGAIN)
1271                                 msleep(20);
1272                         post_status(&dm_device);
1273                 } while (ret == -EAGAIN);
1274
1275                 if (ret) {
1276                         /*
1277                          * Free up the memory we allocatted.
1278                          */
1279                         pr_info("Balloon response failed\n");
1280
1281                         for (i = 0; i < bl_resp->range_count; i++)
1282                                 free_balloon_pages(&dm_device,
1283                                                  &bl_resp->range_array[i]);
1284
1285                         done = true;
1286                 }
1287         }
1288
1289 }
1290
1291 static void balloon_down(struct hv_dynmem_device *dm,
1292                         struct dm_unballoon_request *req)
1293 {
1294         union dm_mem_page_range *range_array = req->range_array;
1295         int range_count = req->range_count;
1296         struct dm_unballoon_response resp;
1297         int i;
1298
1299         for (i = 0; i < range_count; i++) {
1300                 free_balloon_pages(dm, &range_array[i]);
1301                 complete(&dm_device.config_event);
1302         }
1303
1304         if (req->more_pages == 1)
1305                 return;
1306
1307         memset(&resp, 0, sizeof(struct dm_unballoon_response));
1308         resp.hdr.type = DM_UNBALLOON_RESPONSE;
1309         resp.hdr.trans_id = atomic_inc_return(&trans_id);
1310         resp.hdr.size = sizeof(struct dm_unballoon_response);
1311
1312         vmbus_sendpacket(dm_device.dev->channel, &resp,
1313                                 sizeof(struct dm_unballoon_response),
1314                                 (unsigned long)NULL,
1315                                 VM_PKT_DATA_INBAND, 0);
1316
1317         dm->state = DM_INITIALIZED;
1318 }
1319
1320 static void balloon_onchannelcallback(void *context);
1321
1322 static int dm_thread_func(void *dm_dev)
1323 {
1324         struct hv_dynmem_device *dm = dm_dev;
1325
1326         while (!kthread_should_stop()) {
1327                 wait_for_completion_interruptible_timeout(
1328                                                 &dm_device.config_event, 1*HZ);
1329                 /*
1330                  * The host expects us to post information on the memory
1331                  * pressure every second.
1332                  */
1333                 reinit_completion(&dm_device.config_event);
1334                 post_status(dm);
1335         }
1336
1337         return 0;
1338 }
1339
1340
1341 static void version_resp(struct hv_dynmem_device *dm,
1342                         struct dm_version_response *vresp)
1343 {
1344         struct dm_version_request version_req;
1345         int ret;
1346
1347         if (vresp->is_accepted) {
1348                 /*
1349                  * We are done; wakeup the
1350                  * context waiting for version
1351                  * negotiation.
1352                  */
1353                 complete(&dm->host_event);
1354                 return;
1355         }
1356         /*
1357          * If there are more versions to try, continue
1358          * with negotiations; if not
1359          * shutdown the service since we are not able
1360          * to negotiate a suitable version number
1361          * with the host.
1362          */
1363         if (dm->next_version == 0)
1364                 goto version_error;
1365
1366         memset(&version_req, 0, sizeof(struct dm_version_request));
1367         version_req.hdr.type = DM_VERSION_REQUEST;
1368         version_req.hdr.size = sizeof(struct dm_version_request);
1369         version_req.hdr.trans_id = atomic_inc_return(&trans_id);
1370         version_req.version.version = dm->next_version;
1371
1372         /*
1373          * Set the next version to try in case current version fails.
1374          * Win7 protocol ought to be the last one to try.
1375          */
1376         switch (version_req.version.version) {
1377         case DYNMEM_PROTOCOL_VERSION_WIN8:
1378                 dm->next_version = DYNMEM_PROTOCOL_VERSION_WIN7;
1379                 version_req.is_last_attempt = 0;
1380                 break;
1381         default:
1382                 dm->next_version = 0;
1383                 version_req.is_last_attempt = 1;
1384         }
1385
1386         ret = vmbus_sendpacket(dm->dev->channel, &version_req,
1387                                 sizeof(struct dm_version_request),
1388                                 (unsigned long)NULL,
1389                                 VM_PKT_DATA_INBAND, 0);
1390
1391         if (ret)
1392                 goto version_error;
1393
1394         return;
1395
1396 version_error:
1397         dm->state = DM_INIT_ERROR;
1398         complete(&dm->host_event);
1399 }
1400
1401 static void cap_resp(struct hv_dynmem_device *dm,
1402                         struct dm_capabilities_resp_msg *cap_resp)
1403 {
1404         if (!cap_resp->is_accepted) {
1405                 pr_info("Capabilities not accepted by host\n");
1406                 dm->state = DM_INIT_ERROR;
1407         }
1408         complete(&dm->host_event);
1409 }
1410
1411 static void balloon_onchannelcallback(void *context)
1412 {
1413         struct hv_device *dev = context;
1414         u32 recvlen;
1415         u64 requestid;
1416         struct dm_message *dm_msg;
1417         struct dm_header *dm_hdr;
1418         struct hv_dynmem_device *dm = hv_get_drvdata(dev);
1419         struct dm_balloon *bal_msg;
1420         struct dm_hot_add *ha_msg;
1421         union dm_mem_page_range *ha_pg_range;
1422         union dm_mem_page_range *ha_region;
1423
1424         memset(recv_buffer, 0, sizeof(recv_buffer));
1425         vmbus_recvpacket(dev->channel, recv_buffer,
1426                          PAGE_SIZE, &recvlen, &requestid);
1427
1428         if (recvlen > 0) {
1429                 dm_msg = (struct dm_message *)recv_buffer;
1430                 dm_hdr = &dm_msg->hdr;
1431
1432                 switch (dm_hdr->type) {
1433                 case DM_VERSION_RESPONSE:
1434                         version_resp(dm,
1435                                  (struct dm_version_response *)dm_msg);
1436                         break;
1437
1438                 case DM_CAPABILITIES_RESPONSE:
1439                         cap_resp(dm,
1440                                  (struct dm_capabilities_resp_msg *)dm_msg);
1441                         break;
1442
1443                 case DM_BALLOON_REQUEST:
1444                         if (dm->state == DM_BALLOON_UP)
1445                                 pr_warn("Currently ballooning\n");
1446                         bal_msg = (struct dm_balloon *)recv_buffer;
1447                         dm->state = DM_BALLOON_UP;
1448                         dm_device.balloon_wrk.num_pages = bal_msg->num_pages;
1449                         schedule_work(&dm_device.balloon_wrk.wrk);
1450                         break;
1451
1452                 case DM_UNBALLOON_REQUEST:
1453                         dm->state = DM_BALLOON_DOWN;
1454                         balloon_down(dm,
1455                                  (struct dm_unballoon_request *)recv_buffer);
1456                         break;
1457
1458                 case DM_MEM_HOT_ADD_REQUEST:
1459                         if (dm->state == DM_HOT_ADD)
1460                                 pr_warn("Currently hot-adding\n");
1461                         dm->state = DM_HOT_ADD;
1462                         ha_msg = (struct dm_hot_add *)recv_buffer;
1463                         if (ha_msg->hdr.size == sizeof(struct dm_hot_add)) {
1464                                 /*
1465                                  * This is a normal hot-add request specifying
1466                                  * hot-add memory.
1467                                  */
1468                                 dm->host_specified_ha_region = false;
1469                                 ha_pg_range = &ha_msg->range;
1470                                 dm->ha_wrk.ha_page_range = *ha_pg_range;
1471                                 dm->ha_wrk.ha_region_range.page_range = 0;
1472                         } else {
1473                                 /*
1474                                  * Host is specifying that we first hot-add
1475                                  * a region and then partially populate this
1476                                  * region.
1477                                  */
1478                                 dm->host_specified_ha_region = true;
1479                                 ha_pg_range = &ha_msg->range;
1480                                 ha_region = &ha_pg_range[1];
1481                                 dm->ha_wrk.ha_page_range = *ha_pg_range;
1482                                 dm->ha_wrk.ha_region_range = *ha_region;
1483                         }
1484                         schedule_work(&dm_device.ha_wrk.wrk);
1485                         break;
1486
1487                 case DM_INFO_MESSAGE:
1488                         process_info(dm, (struct dm_info_msg *)dm_msg);
1489                         break;
1490
1491                 default:
1492                         pr_err("Unhandled message: type: %d\n", dm_hdr->type);
1493
1494                 }
1495         }
1496
1497 }
1498
1499 static int balloon_probe(struct hv_device *dev,
1500                         const struct hv_vmbus_device_id *dev_id)
1501 {
1502         int ret;
1503         unsigned long t;
1504         struct dm_version_request version_req;
1505         struct dm_capabilities cap_msg;
1506
1507         do_hot_add = hot_add;
1508
1509         /*
1510          * First allocate a send buffer.
1511          */
1512
1513         send_buffer = kmalloc(PAGE_SIZE, GFP_KERNEL);
1514         if (!send_buffer)
1515                 return -ENOMEM;
1516
1517         ret = vmbus_open(dev->channel, dm_ring_size, dm_ring_size, NULL, 0,
1518                         balloon_onchannelcallback, dev);
1519
1520         if (ret)
1521                 goto probe_error0;
1522
1523         dm_device.dev = dev;
1524         dm_device.state = DM_INITIALIZING;
1525         dm_device.next_version = DYNMEM_PROTOCOL_VERSION_WIN8;
1526         init_completion(&dm_device.host_event);
1527         init_completion(&dm_device.config_event);
1528         INIT_LIST_HEAD(&dm_device.ha_region_list);
1529         spin_lock_init(&dm_device.ha_lock);
1530         INIT_WORK(&dm_device.balloon_wrk.wrk, balloon_up);
1531         INIT_WORK(&dm_device.ha_wrk.wrk, hot_add_req);
1532         dm_device.host_specified_ha_region = false;
1533
1534         dm_device.thread =
1535                  kthread_run(dm_thread_func, &dm_device, "hv_balloon");
1536         if (IS_ERR(dm_device.thread)) {
1537                 ret = PTR_ERR(dm_device.thread);
1538                 goto probe_error1;
1539         }
1540
1541 #ifdef CONFIG_MEMORY_HOTPLUG
1542         set_online_page_callback(&hv_online_page);
1543         register_memory_notifier(&hv_memory_nb);
1544 #endif
1545
1546         hv_set_drvdata(dev, &dm_device);
1547         /*
1548          * Initiate the hand shake with the host and negotiate
1549          * a version that the host can support. We start with the
1550          * highest version number and go down if the host cannot
1551          * support it.
1552          */
1553         memset(&version_req, 0, sizeof(struct dm_version_request));
1554         version_req.hdr.type = DM_VERSION_REQUEST;
1555         version_req.hdr.size = sizeof(struct dm_version_request);
1556         version_req.hdr.trans_id = atomic_inc_return(&trans_id);
1557         version_req.version.version = DYNMEM_PROTOCOL_VERSION_WIN10;
1558         version_req.is_last_attempt = 0;
1559
1560         ret = vmbus_sendpacket(dev->channel, &version_req,
1561                                 sizeof(struct dm_version_request),
1562                                 (unsigned long)NULL,
1563                                 VM_PKT_DATA_INBAND, 0);
1564         if (ret)
1565                 goto probe_error2;
1566
1567         t = wait_for_completion_timeout(&dm_device.host_event, 5*HZ);
1568         if (t == 0) {
1569                 ret = -ETIMEDOUT;
1570                 goto probe_error2;
1571         }
1572
1573         /*
1574          * If we could not negotiate a compatible version with the host
1575          * fail the probe function.
1576          */
1577         if (dm_device.state == DM_INIT_ERROR) {
1578                 ret = -ETIMEDOUT;
1579                 goto probe_error2;
1580         }
1581         /*
1582          * Now submit our capabilities to the host.
1583          */
1584         memset(&cap_msg, 0, sizeof(struct dm_capabilities));
1585         cap_msg.hdr.type = DM_CAPABILITIES_REPORT;
1586         cap_msg.hdr.size = sizeof(struct dm_capabilities);
1587         cap_msg.hdr.trans_id = atomic_inc_return(&trans_id);
1588
1589         cap_msg.caps.cap_bits.balloon = 1;
1590         cap_msg.caps.cap_bits.hot_add = 1;
1591
1592         /*
1593          * Specify our alignment requirements as it relates
1594          * memory hot-add. Specify 128MB alignment.
1595          */
1596         cap_msg.caps.cap_bits.hot_add_alignment = 7;
1597
1598         /*
1599          * Currently the host does not use these
1600          * values and we set them to what is done in the
1601          * Windows driver.
1602          */
1603         cap_msg.min_page_cnt = 0;
1604         cap_msg.max_page_number = -1;
1605
1606         ret = vmbus_sendpacket(dev->channel, &cap_msg,
1607                                 sizeof(struct dm_capabilities),
1608                                 (unsigned long)NULL,
1609                                 VM_PKT_DATA_INBAND, 0);
1610         if (ret)
1611                 goto probe_error2;
1612
1613         t = wait_for_completion_timeout(&dm_device.host_event, 5*HZ);
1614         if (t == 0) {
1615                 ret = -ETIMEDOUT;
1616                 goto probe_error2;
1617         }
1618
1619         /*
1620          * If the host does not like our capabilities,
1621          * fail the probe function.
1622          */
1623         if (dm_device.state == DM_INIT_ERROR) {
1624                 ret = -ETIMEDOUT;
1625                 goto probe_error2;
1626         }
1627
1628         dm_device.state = DM_INITIALIZED;
1629
1630         return 0;
1631
1632 probe_error2:
1633 #ifdef CONFIG_MEMORY_HOTPLUG
1634         restore_online_page_callback(&hv_online_page);
1635 #endif
1636         kthread_stop(dm_device.thread);
1637
1638 probe_error1:
1639         vmbus_close(dev->channel);
1640 probe_error0:
1641         kfree(send_buffer);
1642         return ret;
1643 }
1644
1645 static int balloon_remove(struct hv_device *dev)
1646 {
1647         struct hv_dynmem_device *dm = hv_get_drvdata(dev);
1648         struct hv_hotadd_state *has, *tmp;
1649         struct hv_hotadd_gap *gap, *tmp_gap;
1650         unsigned long flags;
1651
1652         if (dm->num_pages_ballooned != 0)
1653                 pr_warn("Ballooned pages: %d\n", dm->num_pages_ballooned);
1654
1655         cancel_work_sync(&dm->balloon_wrk.wrk);
1656         cancel_work_sync(&dm->ha_wrk.wrk);
1657
1658         vmbus_close(dev->channel);
1659         kthread_stop(dm->thread);
1660         kfree(send_buffer);
1661 #ifdef CONFIG_MEMORY_HOTPLUG
1662         restore_online_page_callback(&hv_online_page);
1663         unregister_memory_notifier(&hv_memory_nb);
1664 #endif
1665         spin_lock_irqsave(&dm_device.ha_lock, flags);
1666         list_for_each_entry_safe(has, tmp, &dm->ha_region_list, list) {
1667                 list_for_each_entry_safe(gap, tmp_gap, &has->gap_list, list) {
1668                         list_del(&gap->list);
1669                         kfree(gap);
1670                 }
1671                 list_del(&has->list);
1672                 kfree(has);
1673         }
1674         spin_unlock_irqrestore(&dm_device.ha_lock, flags);
1675
1676         return 0;
1677 }
1678
1679 static const struct hv_vmbus_device_id id_table[] = {
1680         /* Dynamic Memory Class ID */
1681         /* 525074DC-8985-46e2-8057-A307DC18A502 */
1682         { HV_DM_GUID, },
1683         { },
1684 };
1685
1686 MODULE_DEVICE_TABLE(vmbus, id_table);
1687
1688 static  struct hv_driver balloon_drv = {
1689         .name = "hv_balloon",
1690         .id_table = id_table,
1691         .probe =  balloon_probe,
1692         .remove =  balloon_remove,
1693 };
1694
1695 static int __init init_balloon_drv(void)
1696 {
1697
1698         return vmbus_driver_register(&balloon_drv);
1699 }
1700
1701 module_init(init_balloon_drv);
1702
1703 MODULE_DESCRIPTION("Hyper-V Balloon");
1704 MODULE_LICENSE("GPL");