drivers/hv/hv_balloon.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (c) 2012, Microsoft Corporation.
   4  *
   5  * Author:
   6  *   K. Y. Srinivasan <kys@microsoft.com>
   7  */
   8
   9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  10
  11 #include <linux/kernel.h>
  12 #include <linux/jiffies.h>
  13 #include <linux/mman.h>
  14 #include <linux/delay.h>
  15 #include <linux/init.h>
  16 #include <linux/module.h>
  17 #include <linux/slab.h>
  18 #include <linux/kthread.h>
  19 #include <linux/completion.h>
  20 #include <linux/memory_hotplug.h>
  21 #include <linux/memory.h>
  22 #include <linux/notifier.h>
  23 #include <linux/percpu_counter.h>
  24
  25 #include <linux/hyperv.h>
  26
  27 #define CREATE_TRACE_POINTS
  28 #include "hv_trace_balloon.h"
  29
  30 /*
  31  * We begin with definitions supporting the Dynamic Memory protocol
  32  * with the host.
  33  *
  34  * Begin protocol definitions.
  35  */
  36
  37
  38
  39 /*
  40  * Protocol versions. The low word is the minor version, the high word the major
  41  * version.
  42  *
  43  * History:
  44  * Initial version 1.0
  45  * Changed to 0.1 on 2009/03/25
  46  * Changes to 0.2 on 2009/05/14
  47  * Changes to 0.3 on 2009/12/03
  48  * Changed to 1.0 on 2011/04/05
  49  */
  50
  51 #define DYNMEM_MAKE_VERSION(Major, Minor) ((__u32)(((Major) << 16) | (Minor)))
  52 #define DYNMEM_MAJOR_VERSION(Version) ((__u32)(Version) >> 16)
  53 #define DYNMEM_MINOR_VERSION(Version) ((__u32)(Version) & 0xff)
  54
  55 enum {
  56         DYNMEM_PROTOCOL_VERSION_1 = DYNMEM_MAKE_VERSION(0, 3),
  57         DYNMEM_PROTOCOL_VERSION_2 = DYNMEM_MAKE_VERSION(1, 0),
  58         DYNMEM_PROTOCOL_VERSION_3 = DYNMEM_MAKE_VERSION(2, 0),
  59
  60         DYNMEM_PROTOCOL_VERSION_WIN7 = DYNMEM_PROTOCOL_VERSION_1,
  61         DYNMEM_PROTOCOL_VERSION_WIN8 = DYNMEM_PROTOCOL_VERSION_2,
  62         DYNMEM_PROTOCOL_VERSION_WIN10 = DYNMEM_PROTOCOL_VERSION_3,
  63
  64         DYNMEM_PROTOCOL_VERSION_CURRENT = DYNMEM_PROTOCOL_VERSION_WIN10
  65 };
  66
  67
  68
  69 /*
  70  * Message Types
  71  */
  72
  73 enum dm_message_type {
  74         /*
  75          * Version 0.3
  76          */
  77         DM_ERROR                        = 0,
  78         DM_VERSION_REQUEST              = 1,
  79         DM_VERSION_RESPONSE             = 2,
  80         DM_CAPABILITIES_REPORT          = 3,
  81         DM_CAPABILITIES_RESPONSE        = 4,
  82         DM_STATUS_REPORT                = 5,
  83         DM_BALLOON_REQUEST              = 6,
  84         DM_BALLOON_RESPONSE             = 7,
  85         DM_UNBALLOON_REQUEST            = 8,
  86         DM_UNBALLOON_RESPONSE           = 9,
  87         DM_MEM_HOT_ADD_REQUEST          = 10,
  88         DM_MEM_HOT_ADD_RESPONSE         = 11,
  89         DM_VERSION_03_MAX               = 11,
  90         /*
  91          * Version 1.0.
  92          */
  93         DM_INFO_MESSAGE                 = 12,
  94         DM_VERSION_1_MAX                = 12
  95 };
  96
  97
  98 /*
  99  * Structures defining the dynamic memory management
 100  * protocol.
 101  */
 102
 103 union dm_version {
 104         struct {
 105                 __u16 minor_version;
 106                 __u16 major_version;
 107         };
 108         __u32 version;
 109 } __packed;
 110
 111
 112 union dm_caps {
 113         struct {
 114                 __u64 balloon:1;
 115                 __u64 hot_add:1;
 116                 /*
 117                  * To support guests that may have alignment
 118                  * limitations on hot-add, the guest can specify
 119                  * its alignment requirements; a value of n
 120                  * represents an alignment of 2^n in mega bytes.
 121                  */
 122                 __u64 hot_add_alignment:4;
 123                 __u64 reservedz:58;
 124         } cap_bits;
 125         __u64 caps;
 126 } __packed;
 127
 128 union dm_mem_page_range {
 129         struct  {
 130                 /*
 131                  * The PFN number of the first page in the range.
 132                  * 40 bits is the architectural limit of a PFN
 133                  * number for AMD64.
 134                  */
 135                 __u64 start_page:40;
 136                 /*
 137                  * The number of pages in the range.
 138                  */
 139                 __u64 page_cnt:24;
 140         } finfo;
 141         __u64  page_range;
 142 } __packed;
 143
 144
 145
 146 /*
 147  * The header for all dynamic memory messages:
 148  *
 149  * type: Type of the message.
 150  * size: Size of the message in bytes; including the header.
 151  * trans_id: The guest is responsible for manufacturing this ID.
 152  */
 153
 154 struct dm_header {
 155         __u16 type;
 156         __u16 size;
 157         __u32 trans_id;
 158 } __packed;
 159
 160 /*
 161  * A generic message format for dynamic memory.
 162  * Specific message formats are defined later in the file.
 163  */
 164
 165 struct dm_message {
 166         struct dm_header hdr;
 167         __u8 data[]; /* enclosed message */
 168 } __packed;
 169
 170
 171 /*
 172  * Specific message types supporting the dynamic memory protocol.
 173  */
 174
 175 /*
 176  * Version negotiation message. Sent from the guest to the host.
 177  * The guest is free to try different versions until the host
 178  * accepts the version.
 179  *
 180  * dm_version: The protocol version requested.
 181  * is_last_attempt: If TRUE, this is the last version guest will request.
 182  * reservedz: Reserved field, set to zero.
 183  */
 184
 185 struct dm_version_request {
 186         struct dm_header hdr;
 187         union dm_version version;
 188         __u32 is_last_attempt:1;
 189         __u32 reservedz:31;
 190 } __packed;
 191
 192 /*
 193  * Version response message; Host to Guest and indicates
 194  * if the host has accepted the version sent by the guest.
 195  *
 196  * is_accepted: If TRUE, host has accepted the version and the guest
 197  * should proceed to the next stage of the protocol. FALSE indicates that
 198  * guest should re-try with a different version.
 199  *
 200  * reservedz: Reserved field, set to zero.
 201  */
 202
 203 struct dm_version_response {
 204         struct dm_header hdr;
 205         __u64 is_accepted:1;
 206         __u64 reservedz:63;
 207 } __packed;
 208
 209 /*
 210  * Message reporting capabilities. This is sent from the guest to the
 211  * host.
 212  */
 213
 214 struct dm_capabilities {
 215         struct dm_header hdr;
 216         union dm_caps caps;
 217         __u64 min_page_cnt;
 218         __u64 max_page_number;
 219 } __packed;
 220
 221 /*
 222  * Response to the capabilities message. This is sent from the host to the
 223  * guest. This message notifies if the host has accepted the guest's
 224  * capabilities. If the host has not accepted, the guest must shutdown
 225  * the service.
 226  *
 227  * is_accepted: Indicates if the host has accepted guest's capabilities.
 228  * reservedz: Must be 0.
 229  */
 230
 231 struct dm_capabilities_resp_msg {
 232         struct dm_header hdr;
 233         __u64 is_accepted:1;
 234         __u64 reservedz:63;
 235 } __packed;
 236
 237 /*
 238  * This message is used to report memory pressure from the guest.
 239  * This message is not part of any transaction and there is no
 240  * response to this message.
 241  *
 242  * num_avail: Available memory in pages.
 243  * num_committed: Committed memory in pages.
 244  * page_file_size: The accumulated size of all page files
 245  *                 in the system in pages.
 246  * zero_free: The nunber of zero and free pages.
 247  * page_file_writes: The writes to the page file in pages.
 248  * io_diff: An indicator of file cache efficiency or page file activity,
 249  *          calculated as File Cache Page Fault Count - Page Read Count.
 250  *          This value is in pages.
 251  *
 252  * Some of these metrics are Windows specific and fortunately
 253  * the algorithm on the host side that computes the guest memory
 254  * pressure only uses num_committed value.
 255  */
 256
 257 struct dm_status {
 258         struct dm_header hdr;
 259         __u64 num_avail;
 260         __u64 num_committed;
 261         __u64 page_file_size;
 262         __u64 zero_free;
 263         __u32 page_file_writes;
 264         __u32 io_diff;
 265 } __packed;
 266
 267
 268 /*
 269  * Message to ask the guest to allocate memory - balloon up message.
 270  * This message is sent from the host to the guest. The guest may not be
 271  * able to allocate as much memory as requested.
 272  *
 273  * num_pages: number of pages to allocate.
 274  */
 275
 276 struct dm_balloon {
 277         struct dm_header hdr;
 278         __u32 num_pages;
 279         __u32 reservedz;
 280 } __packed;
 281
 282
 283 /*
 284  * Balloon response message; this message is sent from the guest
 285  * to the host in response to the balloon message.
 286  *
 287  * reservedz: Reserved; must be set to zero.
 288  * more_pages: If FALSE, this is the last message of the transaction.
 289  * if TRUE there will atleast one more message from the guest.
 290  *
 291  * range_count: The number of ranges in the range array.
 292  *
 293  * range_array: An array of page ranges returned to the host.
 294  *
 295  */
 296
 297 struct dm_balloon_response {
 298         struct dm_header hdr;
 299         __u32 reservedz;
 300         __u32 more_pages:1;
 301         __u32 range_count:31;
 302         union dm_mem_page_range range_array[];
 303 } __packed;
 304
 305 /*
 306  * Un-balloon message; this message is sent from the host
 307  * to the guest to give guest more memory.
 308  *
 309  * more_pages: If FALSE, this is the last message of the transaction.
 310  * if TRUE there will atleast one more message from the guest.
 311  *
 312  * reservedz: Reserved; must be set to zero.
 313  *
 314  * range_count: The number of ranges in the range array.
 315  *
 316  * range_array: An array of page ranges returned to the host.
 317  *
 318  */
 319
 320 struct dm_unballoon_request {
 321         struct dm_header hdr;
 322         __u32 more_pages:1;
 323         __u32 reservedz:31;
 324         __u32 range_count;
 325         union dm_mem_page_range range_array[];
 326 } __packed;
 327
 328 /*
 329  * Un-balloon response message; this message is sent from the guest
 330  * to the host in response to an unballoon request.
 331  *
 332  */
 333
 334 struct dm_unballoon_response {
 335         struct dm_header hdr;
 336 } __packed;
 337
 338
 339 /*
 340  * Hot add request message. Message sent from the host to the guest.
 341  *
 342  * mem_range: Memory range to hot add.
 343  *
 344  * On Linux we currently don't support this since we cannot hot add
 345  * arbitrary granularity of memory.
 346  */
 347
 348 struct dm_hot_add {
 349         struct dm_header hdr;
 350         union dm_mem_page_range range;
 351 } __packed;
 352
 353 /*
 354  * Hot add response message.
 355  * This message is sent by the guest to report the status of a hot add request.
 356  * If page_count is less than the requested page count, then the host should
 357  * assume all further hot add requests will fail, since this indicates that
 358  * the guest has hit an upper physical memory barrier.
 359  *
 360  * Hot adds may also fail due to low resources; in this case, the guest must
 361  * not complete this message until the hot add can succeed, and the host must
 362  * not send a new hot add request until the response is sent.
 363  * If VSC fails to hot add memory DYNMEM_NUMBER_OF_UNSUCCESSFUL_HOTADD_ATTEMPTS
 364  * times it fails the request.
 365  *
 366  *
 367  * page_count: number of pages that were successfully hot added.
 368  *
 369  * result: result of the operation 1: success, 0: failure.
 370  *
 371  */
 372
 373 struct dm_hot_add_response {
 374         struct dm_header hdr;
 375         __u32 page_count;
 376         __u32 result;
 377 } __packed;
 378
 379 /*
 380  * Types of information sent from host to the guest.
 381  */
 382
 383 enum dm_info_type {
 384         INFO_TYPE_MAX_PAGE_CNT = 0,
 385         MAX_INFO_TYPE
 386 };
 387
 388
 389 /*
 390  * Header for the information message.
 391  */
 392
 393 struct dm_info_header {
 394         enum dm_info_type type;
 395         __u32 data_size;
 396 } __packed;
 397
 398 /*
 399  * This message is sent from the host to the guest to pass
 400  * some relevant information (win8 addition).
 401  *
 402  * reserved: no used.
 403  * info_size: size of the information blob.
 404  * info: information blob.
 405  */
 406
 407 struct dm_info_msg {
 408         struct dm_header hdr;
 409         __u32 reserved;
 410         __u32 info_size;
 411         __u8  info[];
 412 };
 413
 414 /*
 415  * End protocol definitions.
 416  */
 417
 418 /*
 419  * State to manage hot adding memory into the guest.
 420  * The range start_pfn : end_pfn specifies the range
 421  * that the host has asked us to hot add. The range
 422  * start_pfn : ha_end_pfn specifies the range that we have
 423  * currently hot added. We hot add in multiples of 128M
 424  * chunks; it is possible that we may not be able to bring
 425  * online all the pages in the region. The range
 426  * covered_start_pfn:covered_end_pfn defines the pages that can
 427  * be brough online.
 428  */
 429
 430 struct hv_hotadd_state {
 431         struct list_head list;
 432         unsigned long start_pfn;
 433         unsigned long covered_start_pfn;
 434         unsigned long covered_end_pfn;
 435         unsigned long ha_end_pfn;
 436         unsigned long end_pfn;
 437         /*
 438          * A list of gaps.
 439          */
 440         struct list_head gap_list;
 441 };
 442
 443 struct hv_hotadd_gap {
 444         struct list_head list;
 445         unsigned long start_pfn;
 446         unsigned long end_pfn;
 447 };
 448
 449 struct balloon_state {
 450         __u32 num_pages;
 451         struct work_struct wrk;
 452 };
 453
 454 struct hot_add_wrk {
 455         union dm_mem_page_range ha_page_range;
 456         union dm_mem_page_range ha_region_range;
 457         struct work_struct wrk;
 458 };
 459
 460 static bool hot_add = true;
 461 static bool do_hot_add;
 462 /*
 463  * Delay reporting memory pressure by
 464  * the specified number of seconds.
 465  */
 466 static uint pressure_report_delay = 45;
 467
 468 /*
 469  * The last time we posted a pressure report to host.
 470  */
 471 static unsigned long last_post_time;
 472
 473 module_param(hot_add, bool, (S_IRUGO | S_IWUSR));
 474 MODULE_PARM_DESC(hot_add, "If set attempt memory hot_add");
 475
 476 module_param(pressure_report_delay, uint, (S_IRUGO | S_IWUSR));
 477 MODULE_PARM_DESC(pressure_report_delay, "Delay in secs in reporting pressure");
 478 static atomic_t trans_id = ATOMIC_INIT(0);
 479
 480 static int dm_ring_size = (5 * PAGE_SIZE);
 481
 482 /*
 483  * Driver specific state.
 484  */
 485
 486 enum hv_dm_state {
 487         DM_INITIALIZING = 0,
 488         DM_INITIALIZED,
 489         DM_BALLOON_UP,
 490         DM_BALLOON_DOWN,
 491         DM_HOT_ADD,
 492         DM_INIT_ERROR
 493 };
 494
 495
 496 static __u8 recv_buffer[PAGE_SIZE];
 497 static __u8 *send_buffer;
 498 #define PAGES_IN_2M     512
 499 #define HA_CHUNK (32 * 1024)
 500
 501 struct hv_dynmem_device {
 502         struct hv_device *dev;
 503         enum hv_dm_state state;
 504         struct completion host_event;
 505         struct completion config_event;
 506
 507         /*
 508          * Number of pages we have currently ballooned out.
 509          */
 510         unsigned int num_pages_ballooned;
 511         unsigned int num_pages_onlined;
 512         unsigned int num_pages_added;
 513
 514         /*
 515          * State to manage the ballooning (up) operation.
 516          */
 517         struct balloon_state balloon_wrk;
 518
 519         /*
 520          * State to execute the "hot-add" operation.
 521          */
 522         struct hot_add_wrk ha_wrk;
 523
 524         /*
 525          * This state tracks if the host has specified a hot-add
 526          * region.
 527          */
 528         bool host_specified_ha_region;
 529
 530         /*
 531          * State to synchronize hot-add.
 532          */
 533         struct completion  ol_waitevent;
 534         bool ha_waiting;
 535         /*
 536          * This thread handles hot-add
 537          * requests from the host as well as notifying
 538          * the host with regards to memory pressure in
 539          * the guest.
 540          */
 541         struct task_struct *thread;
 542
 543         /*
 544          * Protects ha_region_list, num_pages_onlined counter and individual
 545          * regions from ha_region_list.
 546          */
 547         spinlock_t ha_lock;
 548
 549         /*
 550          * A list of hot-add regions.
 551          */
 552         struct list_head ha_region_list;
 553
 554         /*
 555          * We start with the highest version we can support
 556          * and downgrade based on the host; we save here the
 557          * next version to try.
 558          */
 559         __u32 next_version;
 560
 561         /*
 562          * The negotiated version agreed by host.
 563          */
 564         __u32 version;
 565 };
 566
 567 static struct hv_dynmem_device dm_device;
 568
 569 static void post_status(struct hv_dynmem_device *dm);
 570
 571 #ifdef CONFIG_MEMORY_HOTPLUG
 572 static inline bool has_pfn_is_backed(struct hv_hotadd_state *has,
 573                                      unsigned long pfn)
 574 {
 575         struct hv_hotadd_gap *gap;
 576
 577         /* The page is not backed. */
 578         if ((pfn < has->covered_start_pfn) || (pfn >= has->covered_end_pfn))
 579                 return false;
 580
 581         /* Check for gaps. */
 582         list_for_each_entry(gap, &has->gap_list, list) {
 583                 if ((pfn >= gap->start_pfn) && (pfn < gap->end_pfn))
 584                         return false;
 585         }
 586
 587         return true;
 588 }
 589
 590 static unsigned long hv_page_offline_check(unsigned long start_pfn,
 591                                            unsigned long nr_pages)
 592 {
 593         unsigned long pfn = start_pfn, count = 0;
 594         struct hv_hotadd_state *has;
 595         bool found;
 596
 597         while (pfn < start_pfn + nr_pages) {
 598                 /*
 599                  * Search for HAS which covers the pfn and when we find one
 600                  * count how many consequitive PFNs are covered.
 601                  */
 602                 found = false;
 603                 list_for_each_entry(has, &dm_device.ha_region_list, list) {
 604                         while ((pfn >= has->start_pfn) &&
 605                                (pfn < has->end_pfn) &&
 606                                (pfn < start_pfn + nr_pages)) {
 607                                 found = true;
 608                                 if (has_pfn_is_backed(has, pfn))
 609                                         count++;
 610                                 pfn++;
 611                         }
 612                 }
 613
 614                 /*
 615                  * This PFN is not in any HAS (e.g. we're offlining a region
 616                  * which was present at boot), no need to account for it. Go
 617                  * to the next one.
 618                  */
 619                 if (!found)
 620                         pfn++;
 621         }
 622
 623         return count;
 624 }
 625
 626 static int hv_memory_notifier(struct notifier_block *nb, unsigned long val,
 627                               void *v)
 628 {
 629         struct memory_notify *mem = (struct memory_notify *)v;
 630         unsigned long flags, pfn_count;
 631
 632         switch (val) {
 633         case MEM_ONLINE:
 634         case MEM_CANCEL_ONLINE:
 635                 if (dm_device.ha_waiting) {
 636                         dm_device.ha_waiting = false;
 637                         complete(&dm_device.ol_waitevent);
 638                 }
 639                 break;
 640
 641         case MEM_OFFLINE:
 642                 spin_lock_irqsave(&dm_device.ha_lock, flags);
 643                 pfn_count = hv_page_offline_check(mem->start_pfn,
 644                                                   mem->nr_pages);
 645                 if (pfn_count <= dm_device.num_pages_onlined) {
 646                         dm_device.num_pages_onlined -= pfn_count;
 647                 } else {
 648                         /*
 649                          * We're offlining more pages than we managed to online.
 650                          * This is unexpected. In any case don't let
 651                          * num_pages_onlined wrap around zero.
 652                          */
 653                         WARN_ON_ONCE(1);
 654                         dm_device.num_pages_onlined = 0;
 655                 }
 656                 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 657                 break;
 658         case MEM_GOING_ONLINE:
 659         case MEM_GOING_OFFLINE:
 660         case MEM_CANCEL_OFFLINE:
 661                 break;
 662         }
 663         return NOTIFY_OK;
 664 }
 665
 666 static struct notifier_block hv_memory_nb = {
 667         .notifier_call = hv_memory_notifier,
 668         .priority = 0
 669 };
 670
 671 /* Check if the particular page is backed and can be onlined and online it. */
 672 static void hv_page_online_one(struct hv_hotadd_state *has, struct page *pg)
 673 {
 674         if (!has_pfn_is_backed(has, page_to_pfn(pg))) {
 675                 if (!PageOffline(pg))
 676                         __SetPageOffline(pg);
 677                 return;
 678         }
 679         if (PageOffline(pg))
 680                 __ClearPageOffline(pg);
 681
 682         /* This frame is currently backed; online the page. */
 683         __online_page_set_limits(pg);
 684         __online_page_increment_counters(pg);
 685         __online_page_free(pg);
 686
 687         lockdep_assert_held(&dm_device.ha_lock);
 688         dm_device.num_pages_onlined++;
 689 }
 690
 691 static void hv_bring_pgs_online(struct hv_hotadd_state *has,
 692                                 unsigned long start_pfn, unsigned long size)
 693 {
 694         int i;
 695
 696         pr_debug("Online %lu pages starting at pfn 0x%lx\n", size, start_pfn);
 697         for (i = 0; i < size; i++)
 698                 hv_page_online_one(has, pfn_to_page(start_pfn + i));
 699 }
 700
 701 static void hv_mem_hot_add(unsigned long start, unsigned long size,
 702                                 unsigned long pfn_count,
 703                                 struct hv_hotadd_state *has)
 704 {
 705         int ret = 0;
 706         int i, nid;
 707         unsigned long start_pfn;
 708         unsigned long processed_pfn;
 709         unsigned long total_pfn = pfn_count;
 710         unsigned long flags;
 711
 712         for (i = 0; i < (size/HA_CHUNK); i++) {
 713                 start_pfn = start + (i * HA_CHUNK);
 714
 715                 spin_lock_irqsave(&dm_device.ha_lock, flags);
 716                 has->ha_end_pfn +=  HA_CHUNK;
 717
 718                 if (total_pfn > HA_CHUNK) {
 719                         processed_pfn = HA_CHUNK;
 720                         total_pfn -= HA_CHUNK;
 721                 } else {
 722                         processed_pfn = total_pfn;
 723                         total_pfn = 0;
 724                 }
 725
 726                 has->covered_end_pfn +=  processed_pfn;
 727                 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 728
 729                 init_completion(&dm_device.ol_waitevent);
 730                 dm_device.ha_waiting = !memhp_auto_online;
 731
 732                 nid = memory_add_physaddr_to_nid(PFN_PHYS(start_pfn));
 733                 ret = add_memory(nid, PFN_PHYS((start_pfn)),
 734                                 (HA_CHUNK << PAGE_SHIFT));
 735
 736                 if (ret) {
 737                         pr_err("hot_add memory failed error is %d\n", ret);
 738                         if (ret == -EEXIST) {
 739                                 /*
 740                                  * This error indicates that the error
 741                                  * is not a transient failure. This is the
 742                                  * case where the guest's physical address map
 743                                  * precludes hot adding memory. Stop all further
 744                                  * memory hot-add.
 745                                  */
 746                                 do_hot_add = false;
 747                         }
 748                         spin_lock_irqsave(&dm_device.ha_lock, flags);
 749                         has->ha_end_pfn -= HA_CHUNK;
 750                         has->covered_end_pfn -=  processed_pfn;
 751                         spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 752                         break;
 753                 }
 754
 755                 /*
 756                  * Wait for the memory block to be onlined when memory onlining
 757                  * is done outside of kernel (memhp_auto_online). Since the hot
 758                  * add has succeeded, it is ok to proceed even if the pages in
 759                  * the hot added region have not been "onlined" within the
 760                  * allowed time.
 761                  */
 762                 if (dm_device.ha_waiting)
 763                         wait_for_completion_timeout(&dm_device.ol_waitevent,
 764                                                     5*HZ);
 765                 post_status(&dm_device);
 766         }
 767 }
 768
 769 static void hv_online_page(struct page *pg, unsigned int order)
 770 {
 771         struct hv_hotadd_state *has;
 772         unsigned long flags;
 773         unsigned long pfn = page_to_pfn(pg);
 774
 775         spin_lock_irqsave(&dm_device.ha_lock, flags);
 776         list_for_each_entry(has, &dm_device.ha_region_list, list) {
 777                 /* The page belongs to a different HAS. */
 778                 if ((pfn < has->start_pfn) ||
 779                                 (pfn + (1UL << order) > has->end_pfn))
 780                         continue;
 781
 782                 hv_bring_pgs_online(has, pfn, 1UL << order);
 783                 break;
 784         }
 785         spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 786 }
 787
 788 static int pfn_covered(unsigned long start_pfn, unsigned long pfn_cnt)
 789 {
 790         struct hv_hotadd_state *has;
 791         struct hv_hotadd_gap *gap;
 792         unsigned long residual, new_inc;
 793         int ret = 0;
 794         unsigned long flags;
 795
 796         spin_lock_irqsave(&dm_device.ha_lock, flags);
 797         list_for_each_entry(has, &dm_device.ha_region_list, list) {
 798                 /*
 799                  * If the pfn range we are dealing with is not in the current
 800                  * "hot add block", move on.
 801                  */
 802                 if (start_pfn < has->start_pfn || start_pfn >= has->end_pfn)
 803                         continue;
 804
 805                 /*
 806                  * If the current start pfn is not where the covered_end
 807                  * is, create a gap and update covered_end_pfn.
 808                  */
 809                 if (has->covered_end_pfn != start_pfn) {
 810                         gap = kzalloc(sizeof(struct hv_hotadd_gap), GFP_ATOMIC);
 811                         if (!gap) {
 812                                 ret = -ENOMEM;
 813                                 break;
 814                         }
 815
 816                         INIT_LIST_HEAD(&gap->list);
 817                         gap->start_pfn = has->covered_end_pfn;
 818                         gap->end_pfn = start_pfn;
 819                         list_add_tail(&gap->list, &has->gap_list);
 820
 821                         has->covered_end_pfn = start_pfn;
 822                 }
 823
 824                 /*
 825                  * If the current hot add-request extends beyond
 826                  * our current limit; extend it.
 827                  */
 828                 if ((start_pfn + pfn_cnt) > has->end_pfn) {
 829                         residual = (start_pfn + pfn_cnt - has->end_pfn);
 830                         /*
 831                          * Extend the region by multiples of HA_CHUNK.
 832                          */
 833                         new_inc = (residual / HA_CHUNK) * HA_CHUNK;
 834                         if (residual % HA_CHUNK)
 835                                 new_inc += HA_CHUNK;
 836
 837                         has->end_pfn += new_inc;
 838                 }
 839
 840                 ret = 1;
 841                 break;
 842         }
 843         spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 844
 845         return ret;
 846 }
 847
 848 static unsigned long handle_pg_range(unsigned long pg_start,
 849                                         unsigned long pg_count)
 850 {
 851         unsigned long start_pfn = pg_start;
 852         unsigned long pfn_cnt = pg_count;
 853         unsigned long size;
 854         struct hv_hotadd_state *has;
 855         unsigned long pgs_ol = 0;
 856         unsigned long old_covered_state;
 857         unsigned long res = 0, flags;
 858
 859         pr_debug("Hot adding %lu pages starting at pfn 0x%lx.\n", pg_count,
 860                 pg_start);
 861
 862         spin_lock_irqsave(&dm_device.ha_lock, flags);
 863         list_for_each_entry(has, &dm_device.ha_region_list, list) {
 864                 /*
 865                  * If the pfn range we are dealing with is not in the current
 866                  * "hot add block", move on.
 867                  */
 868                 if (start_pfn < has->start_pfn || start_pfn >= has->end_pfn)
 869                         continue;
 870
 871                 old_covered_state = has->covered_end_pfn;
 872
 873                 if (start_pfn < has->ha_end_pfn) {
 874                         /*
 875                          * This is the case where we are backing pages
 876                          * in an already hot added region. Bring
 877                          * these pages online first.
 878                          */
 879                         pgs_ol = has->ha_end_pfn - start_pfn;
 880                         if (pgs_ol > pfn_cnt)
 881                                 pgs_ol = pfn_cnt;
 882
 883                         has->covered_end_pfn +=  pgs_ol;
 884                         pfn_cnt -= pgs_ol;
 885                         /*
 886                          * Check if the corresponding memory block is already
 887                          * online. It is possible to observe struct pages still
 888                          * being uninitialized here so check section instead.
 889                          * In case the section is online we need to bring the
 890                          * rest of pfns (which were not backed previously)
 891                          * online too.
 892                          */
 893                         if (start_pfn > has->start_pfn &&
 894                             online_section_nr(pfn_to_section_nr(start_pfn)))
 895                                 hv_bring_pgs_online(has, start_pfn, pgs_ol);
 896
 897                 }
 898
 899                 if ((has->ha_end_pfn < has->end_pfn) && (pfn_cnt > 0)) {
 900                         /*
 901                          * We have some residual hot add range
 902                          * that needs to be hot added; hot add
 903                          * it now. Hot add a multiple of
 904                          * of HA_CHUNK that fully covers the pages
 905                          * we have.
 906                          */
 907                         size = (has->end_pfn - has->ha_end_pfn);
 908                         if (pfn_cnt <= size) {
 909                                 size = ((pfn_cnt / HA_CHUNK) * HA_CHUNK);
 910                                 if (pfn_cnt % HA_CHUNK)
 911                                         size += HA_CHUNK;
 912                         } else {
 913                                 pfn_cnt = size;
 914                         }
 915                         spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 916                         hv_mem_hot_add(has->ha_end_pfn, size, pfn_cnt, has);
 917                         spin_lock_irqsave(&dm_device.ha_lock, flags);
 918                 }
 919                 /*
 920                  * If we managed to online any pages that were given to us,
 921                  * we declare success.
 922                  */
 923                 res = has->covered_end_pfn - old_covered_state;
 924                 break;
 925         }
 926         spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 927
 928         return res;
 929 }
 930
 931 static unsigned long process_hot_add(unsigned long pg_start,
 932                                         unsigned long pfn_cnt,
 933                                         unsigned long rg_start,
 934                                         unsigned long rg_size)
 935 {
 936         struct hv_hotadd_state *ha_region = NULL;
 937         int covered;
 938         unsigned long flags;
 939
 940         if (pfn_cnt == 0)
 941                 return 0;
 942
 943         if (!dm_device.host_specified_ha_region) {
 944                 covered = pfn_covered(pg_start, pfn_cnt);
 945                 if (covered < 0)
 946                         return 0;
 947
 948                 if (covered)
 949                         goto do_pg_range;
 950         }
 951
 952         /*
 953          * If the host has specified a hot-add range; deal with it first.
 954          */
 955
 956         if (rg_size != 0) {
 957                 ha_region = kzalloc(sizeof(struct hv_hotadd_state), GFP_KERNEL);
 958                 if (!ha_region)
 959                         return 0;
 960
 961                 INIT_LIST_HEAD(&ha_region->list);
 962                 INIT_LIST_HEAD(&ha_region->gap_list);
 963
 964                 ha_region->start_pfn = rg_start;
 965                 ha_region->ha_end_pfn = rg_start;
 966                 ha_region->covered_start_pfn = pg_start;
 967                 ha_region->covered_end_pfn = pg_start;
 968                 ha_region->end_pfn = rg_start + rg_size;
 969
 970                 spin_lock_irqsave(&dm_device.ha_lock, flags);
 971                 list_add_tail(&ha_region->list, &dm_device.ha_region_list);
 972                 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 973         }
 974
 975 do_pg_range:
 976         /*
 977          * Process the page range specified; bringing them
 978          * online if possible.
 979          */
 980         return handle_pg_range(pg_start, pfn_cnt);
 981 }
 982
 983 #endif
 984
 985 static void hot_add_req(struct work_struct *dummy)
 986 {
 987         struct dm_hot_add_response resp;
 988 #ifdef CONFIG_MEMORY_HOTPLUG
 989         unsigned long pg_start, pfn_cnt;
 990         unsigned long rg_start, rg_sz;
 991 #endif
 992         struct hv_dynmem_device *dm = &dm_device;
 993
 994         memset(&resp, 0, sizeof(struct dm_hot_add_response));
 995         resp.hdr.type = DM_MEM_HOT_ADD_RESPONSE;
 996         resp.hdr.size = sizeof(struct dm_hot_add_response);
 997
 998 #ifdef CONFIG_MEMORY_HOTPLUG
 999         pg_start = dm->ha_wrk.ha_page_range.finfo.start_page;
1000         pfn_cnt = dm->ha_wrk.ha_page_range.finfo.page_cnt;
1001
1002         rg_start = dm->ha_wrk.ha_region_range.finfo.start_page;
1003         rg_sz = dm->ha_wrk.ha_region_range.finfo.page_cnt;
1004
1005         if ((rg_start == 0) && (!dm->host_specified_ha_region)) {
1006                 unsigned long region_size;
1007                 unsigned long region_start;
1008
1009                 /*
1010                  * The host has not specified the hot-add region.
1011                  * Based on the hot-add page range being specified,
1012                  * compute a hot-add region that can cover the pages
1013                  * that need to be hot-added while ensuring the alignment
1014                  * and size requirements of Linux as it relates to hot-add.
1015                  */
1016                 region_start = pg_start;
1017                 region_size = (pfn_cnt / HA_CHUNK) * HA_CHUNK;
1018                 if (pfn_cnt % HA_CHUNK)
1019                         region_size += HA_CHUNK;
1020
1021                 region_start = (pg_start / HA_CHUNK) * HA_CHUNK;
1022
1023                 rg_start = region_start;
1024                 rg_sz = region_size;
1025         }
1026
1027         if (do_hot_add)
1028                 resp.page_count = process_hot_add(pg_start, pfn_cnt,
1029                                                 rg_start, rg_sz);
1030
1031         dm->num_pages_added += resp.page_count;
1032 #endif
1033         /*
1034          * The result field of the response structure has the
1035          * following semantics:
1036          *
1037          * 1. If all or some pages hot-added: Guest should return success.
1038          *
1039          * 2. If no pages could be hot-added:
1040          *
1041          * If the guest returns success, then the host
1042          * will not attempt any further hot-add operations. This
1043          * signifies a permanent failure.
1044          *
1045          * If the guest returns failure, then this failure will be
1046          * treated as a transient failure and the host may retry the
1047          * hot-add operation after some delay.
1048          */
1049         if (resp.page_count > 0)
1050                 resp.result = 1;
1051         else if (!do_hot_add)
1052                 resp.result = 1;
1053         else
1054                 resp.result = 0;
1055
1056         if (!do_hot_add || (resp.page_count == 0))
1057                 pr_err("Memory hot add failed\n");
1058
1059         dm->state = DM_INITIALIZED;
1060         resp.hdr.trans_id = atomic_inc_return(&trans_id);
1061         vmbus_sendpacket(dm->dev->channel, &resp,
1062                         sizeof(struct dm_hot_add_response),
1063                         (unsigned long)NULL,
1064                         VM_PKT_DATA_INBAND, 0);
1065 }
1066
1067 static void process_info(struct hv_dynmem_device *dm, struct dm_info_msg *msg)
1068 {
1069         struct dm_info_header *info_hdr;
1070
1071         info_hdr = (struct dm_info_header *)msg->info;
1072
1073         switch (info_hdr->type) {
1074         case INFO_TYPE_MAX_PAGE_CNT:
1075                 if (info_hdr->data_size == sizeof(__u64)) {
1076                         __u64 *max_page_count = (__u64 *)&info_hdr[1];
1077
1078                         pr_info("Max. dynamic memory size: %llu MB\n",
1079                                 (*max_page_count) >> (20 - PAGE_SHIFT));
1080                 }
1081
1082                 break;
1083         default:
1084                 pr_warn("Received Unknown type: %d\n", info_hdr->type);
1085         }
1086 }
1087
1088 static unsigned long compute_balloon_floor(void)
1089 {
1090         unsigned long min_pages;
1091         unsigned long nr_pages = totalram_pages();
1092 #define MB2PAGES(mb) ((mb) << (20 - PAGE_SHIFT))
1093         /* Simple continuous piecewiese linear function:
1094          *  max MiB -> min MiB  gradient
1095          *       0         0
1096          *      16        16
1097          *      32        24
1098          *     128        72    (1/2)
1099          *     512       168    (1/4)
1100          *    2048       360    (1/8)
1101          *    8192       744    (1/16)
1102          *   32768      1512    (1/32)
1103          */
1104         if (nr_pages < MB2PAGES(128))
1105                 min_pages = MB2PAGES(8) + (nr_pages >> 1);
1106         else if (nr_pages < MB2PAGES(512))
1107                 min_pages = MB2PAGES(40) + (nr_pages >> 2);
1108         else if (nr_pages < MB2PAGES(2048))
1109                 min_pages = MB2PAGES(104) + (nr_pages >> 3);
1110         else if (nr_pages < MB2PAGES(8192))
1111                 min_pages = MB2PAGES(232) + (nr_pages >> 4);
1112         else
1113                 min_pages = MB2PAGES(488) + (nr_pages >> 5);
1114 #undef MB2PAGES
1115         return min_pages;
1116 }
1117
1118 /*
1119  * Post our status as it relates memory pressure to the
1120  * host. Host expects the guests to post this status
1121  * periodically at 1 second intervals.
1122  *
1123  * The metrics specified in this protocol are very Windows
1124  * specific and so we cook up numbers here to convey our memory
1125  * pressure.
1126  */
1127
1128 static void post_status(struct hv_dynmem_device *dm)
1129 {
1130         struct dm_status status;
1131         unsigned long now = jiffies;
1132         unsigned long last_post = last_post_time;
1133
1134         if (pressure_report_delay > 0) {
1135                 --pressure_report_delay;
1136                 return;
1137         }
1138
1139         if (!time_after(now, (last_post_time + HZ)))
1140                 return;
1141
1142         memset(&status, 0, sizeof(struct dm_status));
1143         status.hdr.type = DM_STATUS_REPORT;
1144         status.hdr.size = sizeof(struct dm_status);
1145         status.hdr.trans_id = atomic_inc_return(&trans_id);
1146
1147         /*
1148          * The host expects the guest to report free and committed memory.
1149          * Furthermore, the host expects the pressure information to include
1150          * the ballooned out pages. For a given amount of memory that we are
1151          * managing we need to compute a floor below which we should not
1152          * balloon. Compute this and add it to the pressure report.
1153          * We also need to report all offline pages (num_pages_added -
1154          * num_pages_onlined) as committed to the host, otherwise it can try
1155          * asking us to balloon them out.
1156          */
1157         status.num_avail = si_mem_available();
1158         status.num_committed = vm_memory_committed() +
1159                 dm->num_pages_ballooned +
1160                 (dm->num_pages_added > dm->num_pages_onlined ?
1161                  dm->num_pages_added - dm->num_pages_onlined : 0) +
1162                 compute_balloon_floor();
1163
1164         trace_balloon_status(status.num_avail, status.num_committed,
1165                              vm_memory_committed(), dm->num_pages_ballooned,
1166                              dm->num_pages_added, dm->num_pages_onlined);
1167         /*
1168          * If our transaction ID is no longer current, just don't
1169          * send the status. This can happen if we were interrupted
1170          * after we picked our transaction ID.
1171          */
1172         if (status.hdr.trans_id != atomic_read(&trans_id))
1173                 return;
1174
1175         /*
1176          * If the last post time that we sampled has changed,
1177          * we have raced, don't post the status.
1178          */
1179         if (last_post != last_post_time)
1180                 return;
1181
1182         last_post_time = jiffies;
1183         vmbus_sendpacket(dm->dev->channel, &status,
1184                                 sizeof(struct dm_status),
1185                                 (unsigned long)NULL,
1186                                 VM_PKT_DATA_INBAND, 0);
1187
1188 }
1189
1190 static void free_balloon_pages(struct hv_dynmem_device *dm,
1191                          union dm_mem_page_range *range_array)
1192 {
1193         int num_pages = range_array->finfo.page_cnt;
1194         __u64 start_frame = range_array->finfo.start_page;
1195         struct page *pg;
1196         int i;
1197
1198         for (i = 0; i < num_pages; i++) {
1199                 pg = pfn_to_page(i + start_frame);
1200                 __ClearPageOffline(pg);
1201                 __free_page(pg);
1202                 dm->num_pages_ballooned--;
1203         }
1204 }
1205
1206
1207
1208 static unsigned int alloc_balloon_pages(struct hv_dynmem_device *dm,
1209                                         unsigned int num_pages,
1210                                         struct dm_balloon_response *bl_resp,
1211                                         int alloc_unit)
1212 {
1213         unsigned int i, j;
1214         struct page *pg;
1215
1216         if (num_pages < alloc_unit)
1217                 return 0;
1218
1219         for (i = 0; (i * alloc_unit) < num_pages; i++) {
1220                 if (bl_resp->hdr.size + sizeof(union dm_mem_page_range) >
1221                         PAGE_SIZE)
1222                         return i * alloc_unit;
1223
1224                 /*
1225                  * We execute this code in a thread context. Furthermore,
1226                  * we don't want the kernel to try too hard.
1227                  */
1228                 pg = alloc_pages(GFP_HIGHUSER | __GFP_NORETRY |
1229                                 __GFP_NOMEMALLOC | __GFP_NOWARN,
1230                                 get_order(alloc_unit << PAGE_SHIFT));
1231
1232                 if (!pg)
1233                         return i * alloc_unit;
1234
1235                 dm->num_pages_ballooned += alloc_unit;
1236
1237                 /*
1238                  * If we allocatted 2M pages; split them so we
1239                  * can free them in any order we get.
1240                  */
1241
1242                 if (alloc_unit != 1)
1243                         split_page(pg, get_order(alloc_unit << PAGE_SHIFT));
1244
1245                 /* mark all pages offline */
1246                 for (j = 0; j < (1 << get_order(alloc_unit << PAGE_SHIFT)); j++)
1247                         __SetPageOffline(pg + j);
1248
1249                 bl_resp->range_count++;
1250                 bl_resp->range_array[i].finfo.start_page =
1251                         page_to_pfn(pg);
1252                 bl_resp->range_array[i].finfo.page_cnt = alloc_unit;
1253                 bl_resp->hdr.size += sizeof(union dm_mem_page_range);
1254
1255         }
1256
1257         return num_pages;
1258 }
1259
1260 static void balloon_up(struct work_struct *dummy)
1261 {
1262         unsigned int num_pages = dm_device.balloon_wrk.num_pages;
1263         unsigned int num_ballooned = 0;
1264         struct dm_balloon_response *bl_resp;
1265         int alloc_unit;
1266         int ret;
1267         bool done = false;
1268         int i;
1269         long avail_pages;
1270         unsigned long floor;
1271
1272         /* The host balloons pages in 2M granularity. */
1273         WARN_ON_ONCE(num_pages % PAGES_IN_2M != 0);
1274
1275         /*
1276          * We will attempt 2M allocations. However, if we fail to
1277          * allocate 2M chunks, we will go back to 4k allocations.
1278          */
1279         alloc_unit = 512;
1280
1281         avail_pages = si_mem_available();
1282         floor = compute_balloon_floor();
1283
1284         /* Refuse to balloon below the floor, keep the 2M granularity. */
1285         if (avail_pages < num_pages || avail_pages - num_pages < floor) {
1286                 pr_warn("Balloon request will be partially fulfilled. %s\n",
1287                         avail_pages < num_pages ? "Not enough memory." :
1288                         "Balloon floor reached.");
1289
1290                 num_pages = avail_pages > floor ? (avail_pages - floor) : 0;
1291                 num_pages -= num_pages % PAGES_IN_2M;
1292         }
1293
1294         while (!done) {
1295                 bl_resp = (struct dm_balloon_response *)send_buffer;
1296                 memset(send_buffer, 0, PAGE_SIZE);
1297                 bl_resp->hdr.type = DM_BALLOON_RESPONSE;
1298                 bl_resp->hdr.size = sizeof(struct dm_balloon_response);
1299                 bl_resp->more_pages = 1;
1300
1301                 num_pages -= num_ballooned;
1302                 num_ballooned = alloc_balloon_pages(&dm_device, num_pages,
1303                                                     bl_resp, alloc_unit);
1304
1305                 if (alloc_unit != 1 && num_ballooned == 0) {
1306                         alloc_unit = 1;
1307                         continue;
1308                 }
1309
1310                 if (num_ballooned == 0 || num_ballooned == num_pages) {
1311                         pr_debug("Ballooned %u out of %u requested pages.\n",
1312                                 num_pages, dm_device.balloon_wrk.num_pages);
1313
1314                         bl_resp->more_pages = 0;
1315                         done = true;
1316                         dm_device.state = DM_INITIALIZED;
1317                 }
1318
1319                 /*
1320                  * We are pushing a lot of data through the channel;
1321                  * deal with transient failures caused because of the
1322                  * lack of space in the ring buffer.
1323                  */
1324
1325                 do {
1326                         bl_resp->hdr.trans_id = atomic_inc_return(&trans_id);
1327                         ret = vmbus_sendpacket(dm_device.dev->channel,
1328                                                 bl_resp,
1329                                                 bl_resp->hdr.size,
1330                                                 (unsigned long)NULL,
1331                                                 VM_PKT_DATA_INBAND, 0);
1332
1333                         if (ret == -EAGAIN)
1334                                 msleep(20);
1335                         post_status(&dm_device);
1336                 } while (ret == -EAGAIN);
1337
1338                 if (ret) {
1339                         /*
1340                          * Free up the memory we allocatted.
1341                          */
1342                         pr_err("Balloon response failed\n");
1343
1344                         for (i = 0; i < bl_resp->range_count; i++)
1345                                 free_balloon_pages(&dm_device,
1346                                                  &bl_resp->range_array[i]);
1347
1348                         done = true;
1349                 }
1350         }
1351
1352 }
1353
1354 static void balloon_down(struct hv_dynmem_device *dm,
1355                         struct dm_unballoon_request *req)
1356 {
1357         union dm_mem_page_range *range_array = req->range_array;
1358         int range_count = req->range_count;
1359         struct dm_unballoon_response resp;
1360         int i;
1361         unsigned int prev_pages_ballooned = dm->num_pages_ballooned;
1362
1363         for (i = 0; i < range_count; i++) {
1364                 free_balloon_pages(dm, &range_array[i]);
1365                 complete(&dm_device.config_event);
1366         }
1367
1368         pr_debug("Freed %u ballooned pages.\n",
1369                 prev_pages_ballooned - dm->num_pages_ballooned);
1370
1371         if (req->more_pages == 1)
1372                 return;
1373
1374         memset(&resp, 0, sizeof(struct dm_unballoon_response));
1375         resp.hdr.type = DM_UNBALLOON_RESPONSE;
1376         resp.hdr.trans_id = atomic_inc_return(&trans_id);
1377         resp.hdr.size = sizeof(struct dm_unballoon_response);
1378
1379         vmbus_sendpacket(dm_device.dev->channel, &resp,
1380                                 sizeof(struct dm_unballoon_response),
1381                                 (unsigned long)NULL,
1382                                 VM_PKT_DATA_INBAND, 0);
1383
1384         dm->state = DM_INITIALIZED;
1385 }
1386
1387 static void balloon_onchannelcallback(void *context);
1388
1389 static int dm_thread_func(void *dm_dev)
1390 {
1391         struct hv_dynmem_device *dm = dm_dev;
1392
1393         while (!kthread_should_stop()) {
1394                 wait_for_completion_interruptible_timeout(
1395                                                 &dm_device.config_event, 1*HZ);
1396                 /*
1397                  * The host expects us to post information on the memory
1398                  * pressure every second.
1399                  */
1400                 reinit_completion(&dm_device.config_event);
1401                 post_status(dm);
1402         }
1403
1404         return 0;
1405 }
1406
1407
1408 static void version_resp(struct hv_dynmem_device *dm,
1409                         struct dm_version_response *vresp)
1410 {
1411         struct dm_version_request version_req;
1412         int ret;
1413
1414         if (vresp->is_accepted) {
1415                 /*
1416                  * We are done; wakeup the
1417                  * context waiting for version
1418                  * negotiation.
1419                  */
1420                 complete(&dm->host_event);
1421                 return;
1422         }
1423         /*
1424          * If there are more versions to try, continue
1425          * with negotiations; if not
1426          * shutdown the service since we are not able
1427          * to negotiate a suitable version number
1428          * with the host.
1429          */
1430         if (dm->next_version == 0)
1431                 goto version_error;
1432
1433         memset(&version_req, 0, sizeof(struct dm_version_request));
1434         version_req.hdr.type = DM_VERSION_REQUEST;
1435         version_req.hdr.size = sizeof(struct dm_version_request);
1436         version_req.hdr.trans_id = atomic_inc_return(&trans_id);
1437         version_req.version.version = dm->next_version;
1438         dm->version = version_req.version.version;
1439
1440         /*
1441          * Set the next version to try in case current version fails.
1442          * Win7 protocol ought to be the last one to try.
1443          */
1444         switch (version_req.version.version) {
1445         case DYNMEM_PROTOCOL_VERSION_WIN8:
1446                 dm->next_version = DYNMEM_PROTOCOL_VERSION_WIN7;
1447                 version_req.is_last_attempt = 0;
1448                 break;
1449         default:
1450                 dm->next_version = 0;
1451                 version_req.is_last_attempt = 1;
1452         }
1453
1454         ret = vmbus_sendpacket(dm->dev->channel, &version_req,
1455                                 sizeof(struct dm_version_request),
1456                                 (unsigned long)NULL,
1457                                 VM_PKT_DATA_INBAND, 0);
1458
1459         if (ret)
1460                 goto version_error;
1461
1462         return;
1463
1464 version_error:
1465         dm->state = DM_INIT_ERROR;
1466         complete(&dm->host_event);
1467 }
1468
1469 static void cap_resp(struct hv_dynmem_device *dm,
1470                         struct dm_capabilities_resp_msg *cap_resp)
1471 {
1472         if (!cap_resp->is_accepted) {
1473                 pr_err("Capabilities not accepted by host\n");
1474                 dm->state = DM_INIT_ERROR;
1475         }
1476         complete(&dm->host_event);
1477 }
1478
1479 static void balloon_onchannelcallback(void *context)
1480 {
1481         struct hv_device *dev = context;
1482         u32 recvlen;
1483         u64 requestid;
1484         struct dm_message *dm_msg;
1485         struct dm_header *dm_hdr;
1486         struct hv_dynmem_device *dm = hv_get_drvdata(dev);
1487         struct dm_balloon *bal_msg;
1488         struct dm_hot_add *ha_msg;
1489         union dm_mem_page_range *ha_pg_range;
1490         union dm_mem_page_range *ha_region;
1491
1492         memset(recv_buffer, 0, sizeof(recv_buffer));
1493         vmbus_recvpacket(dev->channel, recv_buffer,
1494                          PAGE_SIZE, &recvlen, &requestid);
1495
1496         if (recvlen > 0) {
1497                 dm_msg = (struct dm_message *)recv_buffer;
1498                 dm_hdr = &dm_msg->hdr;
1499
1500                 switch (dm_hdr->type) {
1501                 case DM_VERSION_RESPONSE:
1502                         version_resp(dm,
1503                                  (struct dm_version_response *)dm_msg);
1504                         break;
1505
1506                 case DM_CAPABILITIES_RESPONSE:
1507                         cap_resp(dm,
1508                                  (struct dm_capabilities_resp_msg *)dm_msg);
1509                         break;
1510
1511                 case DM_BALLOON_REQUEST:
1512                         if (dm->state == DM_BALLOON_UP)
1513                                 pr_warn("Currently ballooning\n");
1514                         bal_msg = (struct dm_balloon *)recv_buffer;
1515                         dm->state = DM_BALLOON_UP;
1516                         dm_device.balloon_wrk.num_pages = bal_msg->num_pages;
1517                         schedule_work(&dm_device.balloon_wrk.wrk);
1518                         break;
1519
1520                 case DM_UNBALLOON_REQUEST:
1521                         dm->state = DM_BALLOON_DOWN;
1522                         balloon_down(dm,
1523                                  (struct dm_unballoon_request *)recv_buffer);
1524                         break;
1525
1526                 case DM_MEM_HOT_ADD_REQUEST:
1527                         if (dm->state == DM_HOT_ADD)
1528                                 pr_warn("Currently hot-adding\n");
1529                         dm->state = DM_HOT_ADD;
1530                         ha_msg = (struct dm_hot_add *)recv_buffer;
1531                         if (ha_msg->hdr.size == sizeof(struct dm_hot_add)) {
1532                                 /*
1533                                  * This is a normal hot-add request specifying
1534                                  * hot-add memory.
1535                                  */
1536                                 dm->host_specified_ha_region = false;
1537                                 ha_pg_range = &ha_msg->range;
1538                                 dm->ha_wrk.ha_page_range = *ha_pg_range;
1539                                 dm->ha_wrk.ha_region_range.page_range = 0;
1540                         } else {
1541                                 /*
1542                                  * Host is specifying that we first hot-add
1543                                  * a region and then partially populate this
1544                                  * region.
1545                                  */
1546                                 dm->host_specified_ha_region = true;
1547                                 ha_pg_range = &ha_msg->range;
1548                                 ha_region = &ha_pg_range[1];
1549                                 dm->ha_wrk.ha_page_range = *ha_pg_range;
1550                                 dm->ha_wrk.ha_region_range = *ha_region;
1551                         }
1552                         schedule_work(&dm_device.ha_wrk.wrk);
1553                         break;
1554
1555                 case DM_INFO_MESSAGE:
1556                         process_info(dm, (struct dm_info_msg *)dm_msg);
1557                         break;
1558
1559                 default:
1560                         pr_warn("Unhandled message: type: %d\n", dm_hdr->type);
1561
1562                 }
1563         }
1564
1565 }
1566
1567 static int balloon_probe(struct hv_device *dev,
1568                         const struct hv_vmbus_device_id *dev_id)
1569 {
1570         int ret;
1571         unsigned long t;
1572         struct dm_version_request version_req;
1573         struct dm_capabilities cap_msg;
1574
1575 #ifdef CONFIG_MEMORY_HOTPLUG
1576         do_hot_add = hot_add;
1577 #else
1578         do_hot_add = false;
1579 #endif
1580
1581         /*
1582          * First allocate a send buffer.
1583          */
1584
1585         send_buffer = kmalloc(PAGE_SIZE, GFP_KERNEL);
1586         if (!send_buffer)
1587                 return -ENOMEM;
1588
1589         ret = vmbus_open(dev->channel, dm_ring_size, dm_ring_size, NULL, 0,
1590                         balloon_onchannelcallback, dev);
1591
1592         if (ret)
1593                 goto probe_error0;
1594
1595         dm_device.dev = dev;
1596         dm_device.state = DM_INITIALIZING;
1597         dm_device.next_version = DYNMEM_PROTOCOL_VERSION_WIN8;
1598         init_completion(&dm_device.host_event);
1599         init_completion(&dm_device.config_event);
1600         INIT_LIST_HEAD(&dm_device.ha_region_list);
1601         spin_lock_init(&dm_device.ha_lock);
1602         INIT_WORK(&dm_device.balloon_wrk.wrk, balloon_up);
1603         INIT_WORK(&dm_device.ha_wrk.wrk, hot_add_req);
1604         dm_device.host_specified_ha_region = false;
1605
1606         dm_device.thread =
1607                  kthread_run(dm_thread_func, &dm_device, "hv_balloon");
1608         if (IS_ERR(dm_device.thread)) {
1609                 ret = PTR_ERR(dm_device.thread);
1610                 goto probe_error1;
1611         }
1612
1613 #ifdef CONFIG_MEMORY_HOTPLUG
1614         set_online_page_callback(&hv_online_page);
1615         register_memory_notifier(&hv_memory_nb);
1616 #endif
1617
1618         hv_set_drvdata(dev, &dm_device);
1619         /*
1620          * Initiate the hand shake with the host and negotiate
1621          * a version that the host can support. We start with the
1622          * highest version number and go down if the host cannot
1623          * support it.
1624          */
1625         memset(&version_req, 0, sizeof(struct dm_version_request));
1626         version_req.hdr.type = DM_VERSION_REQUEST;
1627         version_req.hdr.size = sizeof(struct dm_version_request);
1628         version_req.hdr.trans_id = atomic_inc_return(&trans_id);
1629         version_req.version.version = DYNMEM_PROTOCOL_VERSION_WIN10;
1630         version_req.is_last_attempt = 0;
1631         dm_device.version = version_req.version.version;
1632
1633         ret = vmbus_sendpacket(dev->channel, &version_req,
1634                                 sizeof(struct dm_version_request),
1635                                 (unsigned long)NULL,
1636                                 VM_PKT_DATA_INBAND, 0);
1637         if (ret)
1638                 goto probe_error2;
1639
1640         t = wait_for_completion_timeout(&dm_device.host_event, 5*HZ);
1641         if (t == 0) {
1642                 ret = -ETIMEDOUT;
1643                 goto probe_error2;
1644         }
1645
1646         /*
1647          * If we could not negotiate a compatible version with the host
1648          * fail the probe function.
1649          */
1650         if (dm_device.state == DM_INIT_ERROR) {
1651                 ret = -ETIMEDOUT;
1652                 goto probe_error2;
1653         }
1654
1655         pr_info("Using Dynamic Memory protocol version %u.%u\n",
1656                 DYNMEM_MAJOR_VERSION(dm_device.version),
1657                 DYNMEM_MINOR_VERSION(dm_device.version));
1658
1659         /*
1660          * Now submit our capabilities to the host.
1661          */
1662         memset(&cap_msg, 0, sizeof(struct dm_capabilities));
1663         cap_msg.hdr.type = DM_CAPABILITIES_REPORT;
1664         cap_msg.hdr.size = sizeof(struct dm_capabilities);
1665         cap_msg.hdr.trans_id = atomic_inc_return(&trans_id);
1666
1667         cap_msg.caps.cap_bits.balloon = 1;
1668         cap_msg.caps.cap_bits.hot_add = 1;
1669
1670         /*
1671          * Specify our alignment requirements as it relates
1672          * memory hot-add. Specify 128MB alignment.
1673          */
1674         cap_msg.caps.cap_bits.hot_add_alignment = 7;
1675
1676         /*
1677          * Currently the host does not use these
1678          * values and we set them to what is done in the
1679          * Windows driver.
1680          */
1681         cap_msg.min_page_cnt = 0;
1682         cap_msg.max_page_number = -1;
1683
1684         ret = vmbus_sendpacket(dev->channel, &cap_msg,
1685                                 sizeof(struct dm_capabilities),
1686                                 (unsigned long)NULL,
1687                                 VM_PKT_DATA_INBAND, 0);
1688         if (ret)
1689                 goto probe_error2;
1690
1691         t = wait_for_completion_timeout(&dm_device.host_event, 5*HZ);
1692         if (t == 0) {
1693                 ret = -ETIMEDOUT;
1694                 goto probe_error2;
1695         }
1696
1697         /*
1698          * If the host does not like our capabilities,
1699          * fail the probe function.
1700          */
1701         if (dm_device.state == DM_INIT_ERROR) {
1702                 ret = -ETIMEDOUT;
1703                 goto probe_error2;
1704         }
1705
1706         dm_device.state = DM_INITIALIZED;
1707         last_post_time = jiffies;
1708
1709         return 0;
1710
1711 probe_error2:
1712 #ifdef CONFIG_MEMORY_HOTPLUG
1713         restore_online_page_callback(&hv_online_page);
1714 #endif
1715         kthread_stop(dm_device.thread);
1716
1717 probe_error1:
1718         vmbus_close(dev->channel);
1719 probe_error0:
1720         kfree(send_buffer);
1721         return ret;
1722 }
1723
1724 static int balloon_remove(struct hv_device *dev)
1725 {
1726         struct hv_dynmem_device *dm = hv_get_drvdata(dev);
1727         struct hv_hotadd_state *has, *tmp;
1728         struct hv_hotadd_gap *gap, *tmp_gap;
1729         unsigned long flags;
1730
1731         if (dm->num_pages_ballooned != 0)
1732                 pr_warn("Ballooned pages: %d\n", dm->num_pages_ballooned);
1733
1734         cancel_work_sync(&dm->balloon_wrk.wrk);
1735         cancel_work_sync(&dm->ha_wrk.wrk);
1736
1737         vmbus_close(dev->channel);
1738         kthread_stop(dm->thread);
1739         kfree(send_buffer);
1740 #ifdef CONFIG_MEMORY_HOTPLUG
1741         restore_online_page_callback(&hv_online_page);
1742         unregister_memory_notifier(&hv_memory_nb);
1743 #endif
1744         spin_lock_irqsave(&dm_device.ha_lock, flags);
1745         list_for_each_entry_safe(has, tmp, &dm->ha_region_list, list) {
1746                 list_for_each_entry_safe(gap, tmp_gap, &has->gap_list, list) {
1747                         list_del(&gap->list);
1748                         kfree(gap);
1749                 }
1750                 list_del(&has->list);
1751                 kfree(has);
1752         }
1753         spin_unlock_irqrestore(&dm_device.ha_lock, flags);
1754
1755         return 0;
1756 }
1757
1758 static const struct hv_vmbus_device_id id_table[] = {
1759         /* Dynamic Memory Class ID */
1760         /* 525074DC-8985-46e2-8057-A307DC18A502 */
1761         { HV_DM_GUID, },
1762         { },
1763 };
1764
1765 MODULE_DEVICE_TABLE(vmbus, id_table);
1766
1767 static  struct hv_driver balloon_drv = {
1768         .name = "hv_balloon",
1769         .id_table = id_table,
1770         .probe =  balloon_probe,
1771         .remove =  balloon_remove,
1772         .driver = {
1773                 .probe_type = PROBE_PREFER_ASYNCHRONOUS,
1774         },
1775 };
1776
1777 static int __init init_balloon_drv(void)
1778 {
1779
1780         return vmbus_driver_register(&balloon_drv);
1781 }
1782
1783 module_init(init_balloon_drv);
1784
1785 MODULE_DESCRIPTION("Hyper-V Balloon");
1786 MODULE_LICENSE("GPL");