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/cleanup.h>
  12#include <linux/kernel.h>
  13#include <linux/jiffies.h>
  14#include <linux/mman.h>
  15#include <linux/debugfs.h>
  16#include <linux/delay.h>
  17#include <linux/init.h>
  18#include <linux/module.h>
  19#include <linux/slab.h>
  20#include <linux/kthread.h>
  21#include <linux/completion.h>
  22#include <linux/count_zeros.h>
  23#include <linux/memory_hotplug.h>
  24#include <linux/memory.h>
  25#include <linux/notifier.h>
  26#include <linux/percpu_counter.h>
  27#include <linux/page_reporting.h>
  28#include <linux/sizes.h>
  29
  30#include <linux/hyperv.h>
  31#include <hyperv/hvhdk.h>
  32
  33#include <asm/mshyperv.h>
  34
  35#define CREATE_TRACE_POINTS
  36#include "hv_trace_balloon.h"
  37
  38/*
  39 * We begin with definitions supporting the Dynamic Memory protocol
  40 * with the host.
  41 *
  42 * Begin protocol definitions.
  43 */
  44
  45/*
  46 * Protocol versions. The low word is the minor version, the high word the major
  47 * version.
  48 *
  49 * History:
  50 * Initial version 1.0
  51 * Changed to 0.1 on 2009/03/25
  52 * Changes to 0.2 on 2009/05/14
  53 * Changes to 0.3 on 2009/12/03
  54 * Changed to 1.0 on 2011/04/05
  55 */
  56
  57#define DYNMEM_MAKE_VERSION(Major, Minor) ((__u32)(((Major) << 16) | (Minor)))
  58#define DYNMEM_MAJOR_VERSION(Version) ((__u32)(Version) >> 16)
  59#define DYNMEM_MINOR_VERSION(Version) ((__u32)(Version) & 0xff)
  60
  61enum {
  62	DYNMEM_PROTOCOL_VERSION_1 = DYNMEM_MAKE_VERSION(0, 3),
  63	DYNMEM_PROTOCOL_VERSION_2 = DYNMEM_MAKE_VERSION(1, 0),
  64	DYNMEM_PROTOCOL_VERSION_3 = DYNMEM_MAKE_VERSION(2, 0),
  65
  66	DYNMEM_PROTOCOL_VERSION_WIN7 = DYNMEM_PROTOCOL_VERSION_1,
  67	DYNMEM_PROTOCOL_VERSION_WIN8 = DYNMEM_PROTOCOL_VERSION_2,
  68	DYNMEM_PROTOCOL_VERSION_WIN10 = DYNMEM_PROTOCOL_VERSION_3,
  69
  70	DYNMEM_PROTOCOL_VERSION_CURRENT = DYNMEM_PROTOCOL_VERSION_WIN10
  71};
  72
  73/*
  74 * Message Types
  75 */
  76
  77enum dm_message_type {
  78	/*
  79	 * Version 0.3
  80	 */
  81	DM_ERROR			= 0,
  82	DM_VERSION_REQUEST		= 1,
  83	DM_VERSION_RESPONSE		= 2,
  84	DM_CAPABILITIES_REPORT		= 3,
  85	DM_CAPABILITIES_RESPONSE	= 4,
  86	DM_STATUS_REPORT		= 5,
  87	DM_BALLOON_REQUEST		= 6,
  88	DM_BALLOON_RESPONSE		= 7,
  89	DM_UNBALLOON_REQUEST		= 8,
  90	DM_UNBALLOON_RESPONSE		= 9,
  91	DM_MEM_HOT_ADD_REQUEST		= 10,
  92	DM_MEM_HOT_ADD_RESPONSE		= 11,
  93	DM_VERSION_03_MAX		= 11,
  94	/*
  95	 * Version 1.0.
  96	 */
  97	DM_INFO_MESSAGE			= 12,
  98	DM_VERSION_1_MAX		= 12
  99};
 100
 101/*
 102 * Structures defining the dynamic memory management
 103 * protocol.
 104 */
 105
 106union dm_version {
 107	struct {
 108		__u16 minor_version;
 109		__u16 major_version;
 110	};
 111	__u32 version;
 112} __packed;
 113
 114union dm_caps {
 115	struct {
 116		__u64 balloon:1;
 117		__u64 hot_add:1;
 118		/*
 119		 * To support guests that may have alignment
 120		 * limitations on hot-add, the guest can specify
 121		 * its alignment requirements; a value of n
 122		 * represents an alignment of 2^n in mega bytes.
 123		 */
 124		__u64 hot_add_alignment:4;
 125		__u64 reservedz:58;
 126	} cap_bits;
 127	__u64 caps;
 128} __packed;
 129
 130union dm_mem_page_range {
 131	struct  {
 132		/*
 133		 * The PFN number of the first page in the range.
 134		 * 40 bits is the architectural limit of a PFN
 135		 * number for AMD64.
 136		 */
 137		__u64 start_page:40;
 138		/*
 139		 * The number of pages in the range.
 140		 */
 141		__u64 page_cnt:24;
 142	} finfo;
 143	__u64  page_range;
 144} __packed;
 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
 154struct 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
 165struct dm_message {
 166	struct dm_header hdr;
 167	__u8 data[]; /* enclosed message */
 168} __packed;
 169
 170/*
 171 * Specific message types supporting the dynamic memory protocol.
 172 */
 173
 174/*
 175 * Version negotiation message. Sent from the guest to the host.
 176 * The guest is free to try different versions until the host
 177 * accepts the version.
 178 *
 179 * dm_version: The protocol version requested.
 180 * is_last_attempt: If TRUE, this is the last version guest will request.
 181 * reservedz: Reserved field, set to zero.
 182 */
 183
 184struct dm_version_request {
 185	struct dm_header hdr;
 186	union dm_version version;
 187	__u32 is_last_attempt:1;
 188	__u32 reservedz:31;
 189} __packed;
 190
 191/*
 192 * Version response message; Host to Guest and indicates
 193 * if the host has accepted the version sent by the guest.
 194 *
 195 * is_accepted: If TRUE, host has accepted the version and the guest
 196 * should proceed to the next stage of the protocol. FALSE indicates that
 197 * guest should re-try with a different version.
 198 *
 199 * reservedz: Reserved field, set to zero.
 200 */
 201
 202struct dm_version_response {
 203	struct dm_header hdr;
 204	__u64 is_accepted:1;
 205	__u64 reservedz:63;
 206} __packed;
 207
 208/*
 209 * Message reporting capabilities. This is sent from the guest to the
 210 * host.
 211 */
 212
 213struct dm_capabilities {
 214	struct dm_header hdr;
 215	union dm_caps caps;
 216	__u64 min_page_cnt;
 217	__u64 max_page_number;
 218} __packed;
 219
 220/*
 221 * Response to the capabilities message. This is sent from the host to the
 222 * guest. This message notifies if the host has accepted the guest's
 223 * capabilities. If the host has not accepted, the guest must shutdown
 224 * the service.
 225 *
 226 * is_accepted: Indicates if the host has accepted guest's capabilities.
 227 * reservedz: Must be 0.
 228 */
 229
 230struct dm_capabilities_resp_msg {
 231	struct dm_header hdr;
 232	__u64 is_accepted:1;
 233	__u64 reservedz:63;
 234} __packed;
 235
 236/*
 237 * This message is used to report memory pressure from the guest.
 238 * This message is not part of any transaction and there is no
 239 * response to this message.
 240 *
 241 * num_avail: Available memory in pages.
 242 * num_committed: Committed memory in pages.
 243 * page_file_size: The accumulated size of all page files
 244 *		   in the system in pages.
 245 * zero_free: The number of zero and free pages.
 246 * page_file_writes: The writes to the page file in pages.
 247 * io_diff: An indicator of file cache efficiency or page file activity,
 248 *	    calculated as File Cache Page Fault Count - Page Read Count.
 249 *	    This value is in pages.
 250 *
 251 * Some of these metrics are Windows specific and fortunately
 252 * the algorithm on the host side that computes the guest memory
 253 * pressure only uses num_committed value.
 254 */
 255
 256struct dm_status {
 257	struct dm_header hdr;
 258	__u64 num_avail;
 259	__u64 num_committed;
 260	__u64 page_file_size;
 261	__u64 zero_free;
 262	__u32 page_file_writes;
 263	__u32 io_diff;
 264} __packed;
 265
 266/*
 267 * Message to ask the guest to allocate memory - balloon up message.
 268 * This message is sent from the host to the guest. The guest may not be
 269 * able to allocate as much memory as requested.
 270 *
 271 * num_pages: number of pages to allocate.
 272 */
 273
 274struct dm_balloon {
 275	struct dm_header hdr;
 276	__u32 num_pages;
 277	__u32 reservedz;
 278} __packed;
 279
 280/*
 281 * Balloon response message; this message is sent from the guest
 282 * to the host in response to the balloon message.
 283 *
 284 * reservedz: Reserved; must be set to zero.
 285 * more_pages: If FALSE, this is the last message of the transaction.
 286 * if TRUE there will be at least one more message from the guest.
 287 *
 288 * range_count: The number of ranges in the range array.
 289 *
 290 * range_array: An array of page ranges returned to the host.
 291 *
 292 */
 293
 294struct dm_balloon_response {
 295	struct dm_header hdr;
 296	__u32 reservedz;
 297	__u32 more_pages:1;
 298	__u32 range_count:31;
 299	union dm_mem_page_range range_array[];
 300} __packed;
 301
 302/*
 303 * Un-balloon message; this message is sent from the host
 304 * to the guest to give guest more memory.
 305 *
 306 * more_pages: If FALSE, this is the last message of the transaction.
 307 * if TRUE there will be at least one more message from the guest.
 308 *
 309 * reservedz: Reserved; must be set to zero.
 310 *
 311 * range_count: The number of ranges in the range array.
 312 *
 313 * range_array: An array of page ranges returned to the host.
 314 *
 315 */
 316
 317struct dm_unballoon_request {
 318	struct dm_header hdr;
 319	__u32 more_pages:1;
 320	__u32 reservedz:31;
 321	__u32 range_count;
 322	union dm_mem_page_range range_array[];
 323} __packed;
 324
 325/*
 326 * Un-balloon response message; this message is sent from the guest
 327 * to the host in response to an unballoon request.
 328 *
 329 */
 330
 331struct dm_unballoon_response {
 332	struct dm_header hdr;
 333} __packed;
 334
 335/*
 336 * Hot add request message. Message sent from the host to the guest.
 337 *
 338 * mem_range: Memory range to hot add.
 339 *
 340 */
 341
 342struct dm_hot_add {
 343	struct dm_header hdr;
 344	union dm_mem_page_range range;
 345} __packed;
 346
 347/*
 348 * Hot add response message.
 349 * This message is sent by the guest to report the status of a hot add request.
 350 * If page_count is less than the requested page count, then the host should
 351 * assume all further hot add requests will fail, since this indicates that
 352 * the guest has hit an upper physical memory barrier.
 353 *
 354 * Hot adds may also fail due to low resources; in this case, the guest must
 355 * not complete this message until the hot add can succeed, and the host must
 356 * not send a new hot add request until the response is sent.
 357 * If VSC fails to hot add memory DYNMEM_NUMBER_OF_UNSUCCESSFUL_HOTADD_ATTEMPTS
 358 * times it fails the request.
 359 *
 360 *
 361 * page_count: number of pages that were successfully hot added.
 362 *
 363 * result: result of the operation 1: success, 0: failure.
 364 *
 365 */
 366
 367struct dm_hot_add_response {
 368	struct dm_header hdr;
 369	__u32 page_count;
 370	__u32 result;
 371} __packed;
 372
 373/*
 374 * Types of information sent from host to the guest.
 375 */
 376
 377enum dm_info_type {
 378	INFO_TYPE_MAX_PAGE_CNT = 0,
 379	MAX_INFO_TYPE
 380};
 381
 382/*
 383 * Header for the information message.
 384 */
 385
 386struct dm_info_header {
 387	enum dm_info_type type;
 388	__u32 data_size;
 389} __packed;
 390
 391/*
 392 * This message is sent from the host to the guest to pass
 393 * some relevant information (win8 addition).
 394 *
 395 * reserved: no used.
 396 * info_size: size of the information blob.
 397 * info: information blob.
 398 */
 399
 400struct dm_info_msg {
 401	struct dm_header hdr;
 402	__u32 reserved;
 403	__u32 info_size;
 404	__u8  info[];
 405};
 406
 407/*
 408 * End protocol definitions.
 409 */
 410
 411/*
 412 * State to manage hot adding memory into the guest.
 413 * The range start_pfn : end_pfn specifies the range
 414 * that the host has asked us to hot add. The range
 415 * start_pfn : ha_end_pfn specifies the range that we have
 416 * currently hot added. We hot add in chunks equal to the
 417 * memory block size; it is possible that we may not be able
 418 * to bring online all the pages in the region. The range
 419 * covered_start_pfn:covered_end_pfn defines the pages that can
 420 * be brought online.
 421 */
 422
 423struct hv_hotadd_state {
 424	struct list_head list;
 425	unsigned long start_pfn;
 426	unsigned long covered_start_pfn;
 427	unsigned long covered_end_pfn;
 428	unsigned long ha_end_pfn;
 429	unsigned long end_pfn;
 430	/*
 431	 * A list of gaps.
 432	 */
 433	struct list_head gap_list;
 434};
 435
 436struct hv_hotadd_gap {
 437	struct list_head list;
 438	unsigned long start_pfn;
 439	unsigned long end_pfn;
 440};
 441
 442struct balloon_state {
 443	__u32 num_pages;
 444	struct work_struct wrk;
 445};
 446
 447struct hot_add_wrk {
 448	union dm_mem_page_range ha_page_range;
 449	union dm_mem_page_range ha_region_range;
 450	struct work_struct wrk;
 451};
 452
 453static bool allow_hibernation;
 454static bool hot_add = true;
 455static bool do_hot_add;
 456/*
 457 * Delay reporting memory pressure by
 458 * the specified number of seconds.
 459 */
 460static uint pressure_report_delay = 45;
 461extern unsigned int page_reporting_order;
 462#define HV_MAX_FAILURES	2
 463
 464/*
 465 * The last time we posted a pressure report to host.
 466 */
 467static unsigned long last_post_time;
 468
 469static int hv_hypercall_multi_failure;
 470
 471module_param(hot_add, bool, 0644);
 472MODULE_PARM_DESC(hot_add, "If set attempt memory hot_add");
 473
 474module_param(pressure_report_delay, uint, 0644);
 475MODULE_PARM_DESC(pressure_report_delay, "Delay in secs in reporting pressure");
 476static atomic_t trans_id = ATOMIC_INIT(0);
 477
 478static int dm_ring_size = VMBUS_RING_SIZE(16 * 1024);
 479
 480/*
 481 * Driver specific state.
 482 */
 483
 484enum hv_dm_state {
 485	DM_INITIALIZING = 0,
 486	DM_INITIALIZED,
 487	DM_BALLOON_UP,
 488	DM_BALLOON_DOWN,
 489	DM_HOT_ADD,
 490	DM_INIT_ERROR
 491};
 492
 493static __u8 recv_buffer[HV_HYP_PAGE_SIZE];
 494static __u8 balloon_up_send_buffer[HV_HYP_PAGE_SIZE];
 495
 496static unsigned long ha_pages_in_chunk;
 497#define HA_BYTES_IN_CHUNK (ha_pages_in_chunk << PAGE_SHIFT)
 498
 499#define PAGES_IN_2M (2 * 1024 * 1024 / PAGE_SIZE)
 500
 501struct 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	/*
 535	 * This thread handles hot-add
 536	 * requests from the host as well as notifying
 537	 * the host with regards to memory pressure in
 538	 * the guest.
 539	 */
 540	struct task_struct *thread;
 541
 542	/*
 543	 * Protects ha_region_list, num_pages_onlined counter and individual
 544	 * regions from ha_region_list.
 545	 */
 546	spinlock_t ha_lock;
 547
 548	/*
 549	 * A list of hot-add regions.
 550	 */
 551	struct list_head ha_region_list;
 552
 553	/*
 554	 * We start with the highest version we can support
 555	 * and downgrade based on the host; we save here the
 556	 * next version to try.
 557	 */
 558	__u32 next_version;
 559
 560	/*
 561	 * The negotiated version agreed by host.
 562	 */
 563	__u32 version;
 564
 565	struct page_reporting_dev_info pr_dev_info;
 566
 567	/*
 568	 * Maximum number of pages that can be hot_add-ed
 569	 */
 570	__u64 max_dynamic_page_count;
 571};
 572
 573static struct hv_dynmem_device dm_device;
 574
 575static void post_status(struct hv_dynmem_device *dm);
 576
 577static void enable_page_reporting(void);
 578
 579static void disable_page_reporting(void);
 580
 581#ifdef CONFIG_MEMORY_HOTPLUG
 582static inline bool has_pfn_is_backed(struct hv_hotadd_state *has,
 583				     unsigned long pfn)
 584{
 585	struct hv_hotadd_gap *gap;
 586
 587	/* The page is not backed. */
 588	if (pfn < has->covered_start_pfn || pfn >= has->covered_end_pfn)
 589		return false;
 590
 591	/* Check for gaps. */
 592	list_for_each_entry(gap, &has->gap_list, list) {
 593		if (pfn >= gap->start_pfn && pfn < gap->end_pfn)
 594			return false;
 595	}
 596
 597	return true;
 598}
 599
 600static unsigned long hv_page_offline_check(unsigned long start_pfn,
 601					   unsigned long nr_pages)
 602{
 603	unsigned long pfn = start_pfn, count = 0;
 604	struct hv_hotadd_state *has;
 605	bool found;
 606
 607	while (pfn < start_pfn + nr_pages) {
 608		/*
 609		 * Search for HAS which covers the pfn and when we find one
 610		 * count how many consequitive PFNs are covered.
 611		 */
 612		found = false;
 613		list_for_each_entry(has, &dm_device.ha_region_list, list) {
 614			while ((pfn >= has->start_pfn) &&
 615			       (pfn < has->end_pfn) &&
 616			       (pfn < start_pfn + nr_pages)) {
 617				found = true;
 618				if (has_pfn_is_backed(has, pfn))
 619					count++;
 620				pfn++;
 621			}
 622		}
 623
 624		/*
 625		 * This PFN is not in any HAS (e.g. we're offlining a region
 626		 * which was present at boot), no need to account for it. Go
 627		 * to the next one.
 628		 */
 629		if (!found)
 630			pfn++;
 631	}
 632
 633	return count;
 634}
 635
 636static int hv_memory_notifier(struct notifier_block *nb, unsigned long val,
 637			      void *v)
 638{
 639	struct memory_notify *mem = (struct memory_notify *)v;
 640	unsigned long pfn_count;
 641
 642	switch (val) {
 643	case MEM_ONLINE:
 644	case MEM_CANCEL_ONLINE:
 645		complete(&dm_device.ol_waitevent);
 646		break;
 647
 648	case MEM_OFFLINE:
 649		scoped_guard(spinlock_irqsave, &dm_device.ha_lock) {
 650			pfn_count = hv_page_offline_check(mem->start_pfn,
 651							  mem->nr_pages);
 652			if (pfn_count <= dm_device.num_pages_onlined) {
 653				dm_device.num_pages_onlined -= pfn_count;
 654			} else {
 655				/*
 656				 * We're offlining more pages than we
 657				 * managed to online. This is
 658				 * unexpected. In any case don't let
 659				 * num_pages_onlined wrap around zero.
 660				 */
 661				WARN_ON_ONCE(1);
 662				dm_device.num_pages_onlined = 0;
 663			}
 664		}
 665		break;
 666	case MEM_GOING_ONLINE:
 667	case MEM_GOING_OFFLINE:
 668	case MEM_CANCEL_OFFLINE:
 669		break;
 670	}
 671	return NOTIFY_OK;
 672}
 673
 674static struct notifier_block hv_memory_nb = {
 675	.notifier_call = hv_memory_notifier,
 676	.priority = 0
 677};
 678
 679/* Check if the particular page is backed and can be onlined and online it. */
 680static void hv_page_online_one(struct hv_hotadd_state *has, struct page *pg)
 681{
 682	if (!has_pfn_is_backed(has, page_to_pfn(pg))) {
 683		if (!PageOffline(pg))
 684			__SetPageOffline(pg);
 685		return;
 686	} else if (!PageOffline(pg))
 687		return;
 688
 689	/* This frame is currently backed; online the page. */
 690	generic_online_page(pg, 0);
 691
 692	lockdep_assert_held(&dm_device.ha_lock);
 693	dm_device.num_pages_onlined++;
 694}
 695
 696static void hv_bring_pgs_online(struct hv_hotadd_state *has,
 697				unsigned long start_pfn, unsigned long size)
 698{
 699	int i;
 700
 701	pr_debug("Online %lu pages starting at pfn 0x%lx\n", size, start_pfn);
 702	for (i = 0; i < size; i++)
 703		hv_page_online_one(has, pfn_to_page(start_pfn + i));
 704}
 705
 706static void hv_mem_hot_add(unsigned long start, unsigned long size,
 707				unsigned long pfn_count,
 708				struct hv_hotadd_state *has)
 709{
 710	int ret = 0;
 711	int i, nid;
 712	unsigned long start_pfn;
 713	unsigned long processed_pfn;
 714	unsigned long total_pfn = pfn_count;
 715
 716	for (i = 0; i < (size/ha_pages_in_chunk); i++) {
 717		start_pfn = start + (i * ha_pages_in_chunk);
 718
 719		scoped_guard(spinlock_irqsave, &dm_device.ha_lock) {
 720			has->ha_end_pfn += ha_pages_in_chunk;
 721			processed_pfn = umin(total_pfn, ha_pages_in_chunk);
 722			total_pfn -= processed_pfn;
 723			has->covered_end_pfn += processed_pfn;
 724		}
 725
 726		reinit_completion(&dm_device.ol_waitevent);
 727
 728		nid = memory_add_physaddr_to_nid(PFN_PHYS(start_pfn));
 729		ret = add_memory(nid, PFN_PHYS((start_pfn)),
 730				 HA_BYTES_IN_CHUNK, MHP_MERGE_RESOURCE);
 731
 732		if (ret) {
 733			pr_err("hot_add memory failed error is %d\n", ret);
 734			if (ret == -EEXIST) {
 735				/*
 736				 * This error indicates that the error
 737				 * is not a transient failure. This is the
 738				 * case where the guest's physical address map
 739				 * precludes hot adding memory. Stop all further
 740				 * memory hot-add.
 741				 */
 742				do_hot_add = false;
 743			}
 744			scoped_guard(spinlock_irqsave, &dm_device.ha_lock) {
 745				has->ha_end_pfn -= ha_pages_in_chunk;
 746				has->covered_end_pfn -=  processed_pfn;
 747			}
 748			break;
 749		}
 750
 751		/*
 752		 * Wait for memory to get onlined. If the kernel onlined the
 753		 * memory when adding it, this will return directly. Otherwise,
 754		 * it will wait for user space to online the memory. This helps
 755		 * to avoid adding memory faster than it is getting onlined. As
 756		 * adding succeeded, it is ok to proceed even if the memory was
 757		 * not onlined in time.
 758		 */
 759		wait_for_completion_timeout(&dm_device.ol_waitevent, secs_to_jiffies(5));
 760		post_status(&dm_device);
 761	}
 762}
 763
 764static void hv_online_page(struct page *pg, unsigned int order)
 765{
 766	struct hv_hotadd_state *has;
 767	unsigned long pfn = page_to_pfn(pg);
 768
 769	scoped_guard(spinlock_irqsave, &dm_device.ha_lock) {
 770		list_for_each_entry(has, &dm_device.ha_region_list, list) {
 771			/* The page belongs to a different HAS. */
 772			if (pfn < has->start_pfn ||
 773				(pfn + (1UL << order) > has->end_pfn))
 774				continue;
 775
 776			hv_bring_pgs_online(has, pfn, 1UL << order);
 777			return;
 778		}
 779	}
 780	generic_online_page(pg, order);
 781}
 782
 783static int pfn_covered(unsigned long start_pfn, unsigned long pfn_cnt)
 784{
 785	struct hv_hotadd_state *has;
 786	struct hv_hotadd_gap *gap;
 787	unsigned long residual;
 788	int ret = 0;
 789
 790	guard(spinlock_irqsave)(&dm_device.ha_lock);
 791	list_for_each_entry(has, &dm_device.ha_region_list, list) {
 792		/*
 793		 * If the pfn range we are dealing with is not in the current
 794		 * "hot add block", move on.
 795		 */
 796		if (start_pfn < has->start_pfn || start_pfn >= has->end_pfn)
 797			continue;
 798
 799		/*
 800		 * If the current start pfn is not where the covered_end
 801		 * is, create a gap and update covered_end_pfn.
 802		 */
 803		if (has->covered_end_pfn != start_pfn) {
 804			gap = kzalloc(sizeof(struct hv_hotadd_gap), GFP_ATOMIC);
 805			if (!gap) {
 806				ret = -ENOMEM;
 807				break;
 808			}
 809
 810			INIT_LIST_HEAD(&gap->list);
 811			gap->start_pfn = has->covered_end_pfn;
 812			gap->end_pfn = start_pfn;
 813			list_add_tail(&gap->list, &has->gap_list);
 814
 815			has->covered_end_pfn = start_pfn;
 816		}
 817
 818		/*
 819		 * If the current hot add-request extends beyond
 820		 * our current limit; extend it.
 821		 */
 822		if ((start_pfn + pfn_cnt) > has->end_pfn) {
 823			/* Extend the region by multiples of ha_pages_in_chunk */
 824			residual = (start_pfn + pfn_cnt - has->end_pfn);
 825			has->end_pfn += ALIGN(residual, ha_pages_in_chunk);
 826		}
 827
 828		ret = 1;
 829		break;
 830	}
 831
 832	return ret;
 833}
 834
 835static unsigned long handle_pg_range(unsigned long pg_start,
 836				     unsigned long pg_count)
 837{
 838	unsigned long start_pfn = pg_start;
 839	unsigned long pfn_cnt = pg_count;
 840	unsigned long size;
 841	struct hv_hotadd_state *has;
 842	unsigned long pgs_ol = 0;
 843	unsigned long old_covered_state;
 844	unsigned long res = 0, flags;
 845
 846	pr_debug("Hot adding %lu pages starting at pfn 0x%lx.\n", pg_count,
 847		 pg_start);
 848
 849	spin_lock_irqsave(&dm_device.ha_lock, flags);
 850	list_for_each_entry(has, &dm_device.ha_region_list, list) {
 851		/*
 852		 * If the pfn range we are dealing with is not in the current
 853		 * "hot add block", move on.
 854		 */
 855		if (start_pfn < has->start_pfn || start_pfn >= has->end_pfn)
 856			continue;
 857
 858		old_covered_state = has->covered_end_pfn;
 859
 860		if (start_pfn < has->ha_end_pfn) {
 861			/*
 862			 * This is the case where we are backing pages
 863			 * in an already hot added region. Bring
 864			 * these pages online first.
 865			 */
 866			pgs_ol = has->ha_end_pfn - start_pfn;
 867			if (pgs_ol > pfn_cnt)
 868				pgs_ol = pfn_cnt;
 869
 870			has->covered_end_pfn +=  pgs_ol;
 871			pfn_cnt -= pgs_ol;
 872			/*
 873			 * Check if the corresponding memory block is already
 874			 * online. It is possible to observe struct pages still
 875			 * being uninitialized here so check section instead.
 876			 * In case the section is online we need to bring the
 877			 * rest of pfns (which were not backed previously)
 878			 * online too.
 879			 */
 880			if (start_pfn > has->start_pfn &&
 881			    online_section_nr(pfn_to_section_nr(start_pfn)))
 882				hv_bring_pgs_online(has, start_pfn, pgs_ol);
 883		}
 884
 885		if (has->ha_end_pfn < has->end_pfn && pfn_cnt > 0) {
 886			/*
 887			 * We have some residual hot add range
 888			 * that needs to be hot added; hot add
 889			 * it now. Hot add a multiple of
 890			 * ha_pages_in_chunk that fully covers the pages
 891			 * we have.
 892			 */
 893			size = (has->end_pfn - has->ha_end_pfn);
 894			if (pfn_cnt <= size) {
 895				size = ALIGN(pfn_cnt, ha_pages_in_chunk);
 896			} else {
 897				pfn_cnt = size;
 898			}
 899			spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 900			hv_mem_hot_add(has->ha_end_pfn, size, pfn_cnt, has);
 901			spin_lock_irqsave(&dm_device.ha_lock, flags);
 902		}
 903		/*
 904		 * If we managed to online any pages that were given to us,
 905		 * we declare success.
 906		 */
 907		res = has->covered_end_pfn - old_covered_state;
 908		break;
 909	}
 910	spin_unlock_irqrestore(&dm_device.ha_lock, flags);
 911
 912	return res;
 913}
 914
 915static unsigned long process_hot_add(unsigned long pg_start,
 916					unsigned long pfn_cnt,
 917					unsigned long rg_start,
 918					unsigned long rg_size)
 919{
 920	struct hv_hotadd_state *ha_region = NULL;
 921	int covered;
 922
 923	if (pfn_cnt == 0)
 924		return 0;
 925
 926	if (!dm_device.host_specified_ha_region) {
 927		covered = pfn_covered(pg_start, pfn_cnt);
 928		if (covered < 0)
 929			return 0;
 930
 931		if (covered)
 932			goto do_pg_range;
 933	}
 934
 935	/*
 936	 * If the host has specified a hot-add range; deal with it first.
 937	 */
 938
 939	if (rg_size != 0) {
 940		ha_region = kzalloc(sizeof(struct hv_hotadd_state), GFP_KERNEL);
 941		if (!ha_region)
 942			return 0;
 943
 944		INIT_LIST_HEAD(&ha_region->list);
 945		INIT_LIST_HEAD(&ha_region->gap_list);
 946
 947		ha_region->start_pfn = rg_start;
 948		ha_region->ha_end_pfn = rg_start;
 949		ha_region->covered_start_pfn = pg_start;
 950		ha_region->covered_end_pfn = pg_start;
 951		ha_region->end_pfn = rg_start + rg_size;
 952
 953		scoped_guard(spinlock_irqsave, &dm_device.ha_lock) {
 954			list_add_tail(&ha_region->list, &dm_device.ha_region_list);
 955		}
 956	}
 957
 958do_pg_range:
 959	/*
 960	 * Process the page range specified; bringing them
 961	 * online if possible.
 962	 */
 963	return handle_pg_range(pg_start, pfn_cnt);
 964}
 965
 966#endif
 967
 968static void hot_add_req(struct work_struct *dummy)
 969{
 970	struct dm_hot_add_response resp;
 971#ifdef CONFIG_MEMORY_HOTPLUG
 972	unsigned long pg_start, pfn_cnt;
 973	unsigned long rg_start, rg_sz;
 974#endif
 975	struct hv_dynmem_device *dm = &dm_device;
 976
 977	memset(&resp, 0, sizeof(struct dm_hot_add_response));
 978	resp.hdr.type = DM_MEM_HOT_ADD_RESPONSE;
 979	resp.hdr.size = sizeof(struct dm_hot_add_response);
 980
 981#ifdef CONFIG_MEMORY_HOTPLUG
 982	pg_start = dm->ha_wrk.ha_page_range.finfo.start_page;
 983	pfn_cnt = dm->ha_wrk.ha_page_range.finfo.page_cnt;
 984
 985	rg_start = dm->ha_wrk.ha_region_range.finfo.start_page;
 986	rg_sz = dm->ha_wrk.ha_region_range.finfo.page_cnt;
 987
 988	if (rg_start == 0 && !dm->host_specified_ha_region) {
 989		/*
 990		 * The host has not specified the hot-add region.
 991		 * Based on the hot-add page range being specified,
 992		 * compute a hot-add region that can cover the pages
 993		 * that need to be hot-added while ensuring the alignment
 994		 * and size requirements of Linux as it relates to hot-add.
 995		 */
 996		rg_start = ALIGN_DOWN(pg_start, ha_pages_in_chunk);
 997		rg_sz = ALIGN(pfn_cnt, ha_pages_in_chunk);
 998	}
 999
1000	if (do_hot_add)
1001		resp.page_count = process_hot_add(pg_start, pfn_cnt,
1002						  rg_start, rg_sz);
1003
1004	dm->num_pages_added += resp.page_count;
1005#endif
1006	/*
1007	 * The result field of the response structure has the
1008	 * following semantics:
1009	 *
1010	 * 1. If all or some pages hot-added: Guest should return success.
1011	 *
1012	 * 2. If no pages could be hot-added:
1013	 *
1014	 * If the guest returns success, then the host
1015	 * will not attempt any further hot-add operations. This
1016	 * signifies a permanent failure.
1017	 *
1018	 * If the guest returns failure, then this failure will be
1019	 * treated as a transient failure and the host may retry the
1020	 * hot-add operation after some delay.
1021	 */
1022	if (resp.page_count > 0)
1023		resp.result = 1;
1024	else if (!do_hot_add)
1025		resp.result = 1;
1026	else
1027		resp.result = 0;
1028
1029	if (!do_hot_add || resp.page_count == 0) {
1030		if (!allow_hibernation)
1031			pr_err("Memory hot add failed\n");
1032		else
1033			pr_info("Ignore hot-add request!\n");
1034	}
1035
1036	dm->state = DM_INITIALIZED;
1037	resp.hdr.trans_id = atomic_inc_return(&trans_id);
1038	vmbus_sendpacket(dm->dev->channel, &resp,
1039			sizeof(struct dm_hot_add_response),
1040			(unsigned long)NULL,
1041			VM_PKT_DATA_INBAND, 0);
1042}
1043
1044static void process_info(struct hv_dynmem_device *dm, struct dm_info_msg *msg)
1045{
1046	struct dm_info_header *info_hdr;
1047
1048	info_hdr = (struct dm_info_header *)msg->info;
1049
1050	switch (info_hdr->type) {
1051	case INFO_TYPE_MAX_PAGE_CNT:
1052		if (info_hdr->data_size == sizeof(__u64)) {
1053			__u64 *max_page_count = (__u64 *)&info_hdr[1];
1054
1055			pr_info("Max. dynamic memory size: %llu MB\n",
1056				(*max_page_count) >> (20 - HV_HYP_PAGE_SHIFT));
1057			dm->max_dynamic_page_count = *max_page_count;
1058		}
1059
1060		break;
1061	default:
1062		pr_warn("Received Unknown type: %d\n", info_hdr->type);
1063	}
1064}
1065
1066static unsigned long compute_balloon_floor(void)
1067{
1068	unsigned long min_pages;
1069	unsigned long nr_pages = totalram_pages();
1070#define MB2PAGES(mb) ((mb) << (20 - PAGE_SHIFT))
1071	/* Simple continuous piecewiese linear function:
1072	 *  max MiB -> min MiB  gradient
1073	 *       0         0
1074	 *      16        16
1075	 *      32        24
1076	 *     128        72    (1/2)
1077	 *     512       168    (1/4)
1078	 *    2048       360    (1/8)
1079	 *    8192       744    (1/16)
1080	 *   32768      1512	(1/32)
1081	 */
1082	if (nr_pages < MB2PAGES(128))
1083		min_pages = MB2PAGES(8) + (nr_pages >> 1);
1084	else if (nr_pages < MB2PAGES(512))
1085		min_pages = MB2PAGES(40) + (nr_pages >> 2);
1086	else if (nr_pages < MB2PAGES(2048))
1087		min_pages = MB2PAGES(104) + (nr_pages >> 3);
1088	else if (nr_pages < MB2PAGES(8192))
1089		min_pages = MB2PAGES(232) + (nr_pages >> 4);
1090	else
1091		min_pages = MB2PAGES(488) + (nr_pages >> 5);
1092#undef MB2PAGES
1093	return min_pages;
1094}
1095
1096/*
1097 * Compute total committed memory pages
1098 */
1099
1100static unsigned long get_pages_committed(struct hv_dynmem_device *dm)
1101{
1102	return vm_memory_committed() +
1103		dm->num_pages_ballooned +
1104		(dm->num_pages_added > dm->num_pages_onlined ?
1105		 dm->num_pages_added - dm->num_pages_onlined : 0) +
1106		compute_balloon_floor();
1107}
1108
1109/*
1110 * Post our status as it relates memory pressure to the
1111 * host. Host expects the guests to post this status
1112 * periodically at 1 second intervals.
1113 *
1114 * The metrics specified in this protocol are very Windows
1115 * specific and so we cook up numbers here to convey our memory
1116 * pressure.
1117 */
1118
1119static void post_status(struct hv_dynmem_device *dm)
1120{
1121	struct dm_status status;
1122	unsigned long now = jiffies;
1123	unsigned long last_post = last_post_time;
1124	unsigned long num_pages_avail, num_pages_committed;
1125
1126	if (pressure_report_delay > 0) {
1127		--pressure_report_delay;
1128		return;
1129	}
1130
1131	if (!time_after(now, (last_post_time + HZ)))
1132		return;
1133
1134	memset(&status, 0, sizeof(struct dm_status));
1135	status.hdr.type = DM_STATUS_REPORT;
1136	status.hdr.size = sizeof(struct dm_status);
1137	status.hdr.trans_id = atomic_inc_return(&trans_id);
1138
1139	/*
1140	 * The host expects the guest to report free and committed memory.
1141	 * Furthermore, the host expects the pressure information to include
1142	 * the ballooned out pages. For a given amount of memory that we are
1143	 * managing we need to compute a floor below which we should not
1144	 * balloon. Compute this and add it to the pressure report.
1145	 * We also need to report all offline pages (num_pages_added -
1146	 * num_pages_onlined) as committed to the host, otherwise it can try
1147	 * asking us to balloon them out.
1148	 */
1149	num_pages_avail = si_mem_available();
1150	num_pages_committed = get_pages_committed(dm);
1151
1152	trace_balloon_status(num_pages_avail, num_pages_committed,
1153			     vm_memory_committed(), dm->num_pages_ballooned,
1154			     dm->num_pages_added, dm->num_pages_onlined);
1155
1156	/* Convert numbers of pages into numbers of HV_HYP_PAGEs. */
1157	status.num_avail = num_pages_avail * NR_HV_HYP_PAGES_IN_PAGE;
1158	status.num_committed = num_pages_committed * NR_HV_HYP_PAGES_IN_PAGE;
1159
1160	/*
1161	 * If our transaction ID is no longer current, just don't
1162	 * send the status. This can happen if we were interrupted
1163	 * after we picked our transaction ID.
1164	 */
1165	if (status.hdr.trans_id != atomic_read(&trans_id))
1166		return;
1167
1168	/*
1169	 * If the last post time that we sampled has changed,
1170	 * we have raced, don't post the status.
1171	 */
1172	if (last_post != last_post_time)
1173		return;
1174
1175	last_post_time = jiffies;
1176	vmbus_sendpacket(dm->dev->channel, &status,
1177				sizeof(struct dm_status),
1178				(unsigned long)NULL,
1179				VM_PKT_DATA_INBAND, 0);
1180}
1181
1182static void free_balloon_pages(struct hv_dynmem_device *dm,
1183			       union dm_mem_page_range *range_array)
1184{
1185	int num_pages = range_array->finfo.page_cnt;
1186	__u64 start_frame = range_array->finfo.start_page;
1187	struct page *pg;
1188	int i;
1189
1190	for (i = 0; i < num_pages; i++) {
1191		pg = pfn_to_page(i + start_frame);
1192		__ClearPageOffline(pg);
1193		__free_page(pg);
1194		dm->num_pages_ballooned--;
1195		mod_node_page_state(page_pgdat(pg), NR_BALLOON_PAGES, -1);
1196		adjust_managed_page_count(pg, 1);
1197	}
1198}
1199
1200static unsigned int alloc_balloon_pages(struct hv_dynmem_device *dm,
1201					unsigned int num_pages,
1202					struct dm_balloon_response *bl_resp,
1203					int alloc_unit)
1204{
1205	unsigned int i, j;
1206	struct page *pg;
1207
1208	for (i = 0; i < num_pages / alloc_unit; i++) {
1209		if (bl_resp->hdr.size + sizeof(union dm_mem_page_range) >
1210			HV_HYP_PAGE_SIZE)
1211			return i * alloc_unit;
1212
1213		/*
1214		 * We execute this code in a thread context. Furthermore,
1215		 * we don't want the kernel to try too hard.
1216		 */
1217		pg = alloc_pages(GFP_HIGHUSER | __GFP_NORETRY |
1218				__GFP_NOMEMALLOC | __GFP_NOWARN,
1219				get_order(alloc_unit << PAGE_SHIFT));
1220
1221		if (!pg)
1222			return i * alloc_unit;
1223
1224		dm->num_pages_ballooned += alloc_unit;
1225		mod_node_page_state(page_pgdat(pg), NR_BALLOON_PAGES, alloc_unit);
1226
1227		/*
1228		 * If we allocatted 2M pages; split them so we
1229		 * can free them in any order we get.
1230		 */
1231
1232		if (alloc_unit != 1)
1233			split_page(pg, get_order(alloc_unit << PAGE_SHIFT));
1234
1235		/* mark all pages offline */
1236		for (j = 0; j < alloc_unit; j++) {
1237			__SetPageOffline(pg + j);
1238			adjust_managed_page_count(pg + j, -1);
1239		}
1240
1241		bl_resp->range_count++;
1242		bl_resp->range_array[i].finfo.start_page =
1243			page_to_pfn(pg);
1244		bl_resp->range_array[i].finfo.page_cnt = alloc_unit;
1245		bl_resp->hdr.size += sizeof(union dm_mem_page_range);
1246	}
1247
1248	return i * alloc_unit;
1249}
1250
1251static void balloon_up(struct work_struct *dummy)
1252{
1253	unsigned int num_pages = dm_device.balloon_wrk.num_pages;
1254	unsigned int num_ballooned = 0;
1255	struct dm_balloon_response *bl_resp;
1256	int alloc_unit;
1257	int ret;
1258	bool done = false;
1259	int i;
1260	long avail_pages;
1261	unsigned long floor;
1262
1263	/*
1264	 * We will attempt 2M allocations. However, if we fail to
1265	 * allocate 2M chunks, we will go back to PAGE_SIZE allocations.
1266	 */
1267	alloc_unit = PAGES_IN_2M;
1268
1269	avail_pages = si_mem_available();
1270	floor = compute_balloon_floor();
1271
1272	/* Refuse to balloon below the floor. */
1273	if (avail_pages < num_pages || avail_pages - num_pages < floor) {
1274		pr_info("Balloon request will be partially fulfilled. %s\n",
1275			avail_pages < num_pages ? "Not enough memory." :
1276			"Balloon floor reached.");
1277
1278		num_pages = avail_pages > floor ? (avail_pages - floor) : 0;
1279	}
1280
1281	while (!done) {
1282		memset(balloon_up_send_buffer, 0, HV_HYP_PAGE_SIZE);
1283		bl_resp = (struct dm_balloon_response *)balloon_up_send_buffer;
1284		bl_resp->hdr.type = DM_BALLOON_RESPONSE;
1285		bl_resp->hdr.size = sizeof(struct dm_balloon_response);
1286		bl_resp->more_pages = 1;
1287
1288		num_pages -= num_ballooned;
1289		num_ballooned = alloc_balloon_pages(&dm_device, num_pages,
1290						    bl_resp, alloc_unit);
1291
1292		if (alloc_unit != 1 && num_ballooned == 0) {
1293			alloc_unit = 1;
1294			continue;
1295		}
1296
1297		if (num_ballooned == 0 || num_ballooned == num_pages) {
1298			pr_debug("Ballooned %u out of %u requested pages.\n",
1299				 num_pages, dm_device.balloon_wrk.num_pages);
1300
1301			bl_resp->more_pages = 0;
1302			done = true;
1303			dm_device.state = DM_INITIALIZED;
1304		}
1305
1306		/*
1307		 * We are pushing a lot of data through the channel;
1308		 * deal with transient failures caused because of the
1309		 * lack of space in the ring buffer.
1310		 */
1311
1312		do {
1313			bl_resp->hdr.trans_id = atomic_inc_return(&trans_id);
1314			ret = vmbus_sendpacket(dm_device.dev->channel,
1315						bl_resp,
1316						bl_resp->hdr.size,
1317						(unsigned long)NULL,
1318						VM_PKT_DATA_INBAND, 0);
1319
1320			if (ret == -EAGAIN)
1321				msleep(20);
1322			post_status(&dm_device);
1323		} while (ret == -EAGAIN);
1324
1325		if (ret) {
1326			/*
1327			 * Free up the memory we allocatted.
1328			 */
1329			pr_err("Balloon response failed\n");
1330
1331			for (i = 0; i < bl_resp->range_count; i++)
1332				free_balloon_pages(&dm_device,
1333						   &bl_resp->range_array[i]);
1334
1335			done = true;
1336		}
1337	}
1338}
1339
1340static void balloon_down(struct hv_dynmem_device *dm,
1341			 struct dm_unballoon_request *req)
1342{
1343	union dm_mem_page_range *range_array = req->range_array;
1344	int range_count = req->range_count;
1345	struct dm_unballoon_response resp;
1346	int i;
1347	unsigned int prev_pages_ballooned = dm->num_pages_ballooned;
1348
1349	for (i = 0; i < range_count; i++) {
1350		free_balloon_pages(dm, &range_array[i]);
1351		complete(&dm_device.config_event);
1352	}
1353
1354	pr_debug("Freed %u ballooned pages.\n",
1355		 prev_pages_ballooned - dm->num_pages_ballooned);
1356
1357	if (req->more_pages == 1)
1358		return;
1359
1360	memset(&resp, 0, sizeof(struct dm_unballoon_response));
1361	resp.hdr.type = DM_UNBALLOON_RESPONSE;
1362	resp.hdr.trans_id = atomic_inc_return(&trans_id);
1363	resp.hdr.size = sizeof(struct dm_unballoon_response);
1364
1365	vmbus_sendpacket(dm_device.dev->channel, &resp,
1366				sizeof(struct dm_unballoon_response),
1367				(unsigned long)NULL,
1368				VM_PKT_DATA_INBAND, 0);
1369
1370	dm->state = DM_INITIALIZED;
1371}
1372
1373static void balloon_onchannelcallback(void *context);
1374
1375static int dm_thread_func(void *dm_dev)
1376{
1377	struct hv_dynmem_device *dm = dm_dev;
1378
1379	while (!kthread_should_stop()) {
1380		wait_for_completion_interruptible_timeout(&dm_device.config_event,
1381								secs_to_jiffies(1));
1382		/*
1383		 * The host expects us to post information on the memory
1384		 * pressure every second.
1385		 */
1386		reinit_completion(&dm_device.config_event);
1387		post_status(dm);
1388		/*
1389		 * disable free page reporting if multiple hypercall
1390		 * failure flag set. It is not done in the page_reporting
1391		 * callback context as that causes a deadlock between
1392		 * page_reporting_process() and page_reporting_unregister()
1393		 */
1394		if (hv_hypercall_multi_failure >= HV_MAX_FAILURES) {
1395			pr_err("Multiple failures in cold memory discard hypercall, disabling page reporting\n");
1396			disable_page_reporting();
1397			/* Reset the flag after disabling reporting */
1398			hv_hypercall_multi_failure = 0;
1399		}
1400	}
1401
1402	return 0;
1403}
1404
1405static void version_resp(struct hv_dynmem_device *dm,
1406			 struct dm_version_response *vresp)
1407{
1408	struct dm_version_request version_req;
1409	int ret;
1410
1411	if (vresp->is_accepted) {
1412		/*
1413		 * We are done; wakeup the
1414		 * context waiting for version
1415		 * negotiation.
1416		 */
1417		complete(&dm->host_event);
1418		return;
1419	}
1420	/*
1421	 * If there are more versions to try, continue
1422	 * with negotiations; if not
1423	 * shutdown the service since we are not able
1424	 * to negotiate a suitable version number
1425	 * with the host.
1426	 */
1427	if (dm->next_version == 0)
1428		goto version_error;
1429
1430	memset(&version_req, 0, sizeof(struct dm_version_request));
1431	version_req.hdr.type = DM_VERSION_REQUEST;
1432	version_req.hdr.size = sizeof(struct dm_version_request);
1433	version_req.hdr.trans_id = atomic_inc_return(&trans_id);
1434	version_req.version.version = dm->next_version;
1435	dm->version = version_req.version.version;
1436
1437	/*
1438	 * Set the next version to try in case current version fails.
1439	 * Win7 protocol ought to be the last one to try.
1440	 */
1441	switch (version_req.version.version) {
1442	case DYNMEM_PROTOCOL_VERSION_WIN8:
1443		dm->next_version = DYNMEM_PROTOCOL_VERSION_WIN7;
1444		version_req.is_last_attempt = 0;
1445		break;
1446	default:
1447		dm->next_version = 0;
1448		version_req.is_last_attempt = 1;
1449	}
1450
1451	ret = vmbus_sendpacket(dm->dev->channel, &version_req,
1452				sizeof(struct dm_version_request),
1453				(unsigned long)NULL,
1454				VM_PKT_DATA_INBAND, 0);
1455
1456	if (ret)
1457		goto version_error;
1458
1459	return;
1460
1461version_error:
1462	dm->state = DM_INIT_ERROR;
1463	complete(&dm->host_event);
1464}
1465
1466static void cap_resp(struct hv_dynmem_device *dm,
1467		     struct dm_capabilities_resp_msg *cap_resp)
1468{
1469	if (!cap_resp->is_accepted) {
1470		pr_err("Capabilities not accepted by host\n");
1471		dm->state = DM_INIT_ERROR;
1472	}
1473	complete(&dm->host_event);
1474}
1475
1476static void balloon_onchannelcallback(void *context)
1477{
1478	struct hv_device *dev = context;
1479	u32 recvlen;
1480	u64 requestid;
1481	struct dm_message *dm_msg;
1482	struct dm_header *dm_hdr;
1483	struct hv_dynmem_device *dm = hv_get_drvdata(dev);
1484	struct dm_balloon *bal_msg;
1485	struct dm_hot_add *ha_msg;
1486	union dm_mem_page_range *ha_pg_range;
1487	union dm_mem_page_range *ha_region;
1488
1489	memset(recv_buffer, 0, sizeof(recv_buffer));
1490	vmbus_recvpacket(dev->channel, recv_buffer,
1491			 HV_HYP_PAGE_SIZE, &recvlen, &requestid);
1492
1493	if (recvlen > 0) {
1494		dm_msg = (struct dm_message *)recv_buffer;
1495		dm_hdr = &dm_msg->hdr;
1496
1497		switch (dm_hdr->type) {
1498		case DM_VERSION_RESPONSE:
1499			version_resp(dm,
1500				     (struct dm_version_response *)dm_msg);
1501			break;
1502
1503		case DM_CAPABILITIES_RESPONSE:
1504			cap_resp(dm,
1505				 (struct dm_capabilities_resp_msg *)dm_msg);
1506			break;
1507
1508		case DM_BALLOON_REQUEST:
1509			if (allow_hibernation) {
1510				pr_info("Ignore balloon-up request!\n");
1511				break;
1512			}
1513
1514			if (dm->state == DM_BALLOON_UP)
1515				pr_warn("Currently ballooning\n");
1516			bal_msg = (struct dm_balloon *)recv_buffer;
1517			dm->state = DM_BALLOON_UP;
1518			dm_device.balloon_wrk.num_pages = bal_msg->num_pages;
1519			schedule_work(&dm_device.balloon_wrk.wrk);
1520			break;
1521
1522		case DM_UNBALLOON_REQUEST:
1523			if (allow_hibernation) {
1524				pr_info("Ignore balloon-down request!\n");
1525				break;
1526			}
1527
1528			dm->state = DM_BALLOON_DOWN;
1529			balloon_down(dm,
1530				     (struct dm_unballoon_request *)recv_buffer);
1531			break;
1532
1533		case DM_MEM_HOT_ADD_REQUEST:
1534			if (dm->state == DM_HOT_ADD)
1535				pr_warn("Currently hot-adding\n");
1536			dm->state = DM_HOT_ADD;
1537			ha_msg = (struct dm_hot_add *)recv_buffer;
1538			if (ha_msg->hdr.size == sizeof(struct dm_hot_add)) {
1539				/*
1540				 * This is a normal hot-add request specifying
1541				 * hot-add memory.
1542				 */
1543				dm->host_specified_ha_region = false;
1544				ha_pg_range = &ha_msg->range;
1545				dm->ha_wrk.ha_page_range = *ha_pg_range;
1546				dm->ha_wrk.ha_region_range.page_range = 0;
1547			} else {
1548				/*
1549				 * Host is specifying that we first hot-add
1550				 * a region and then partially populate this
1551				 * region.
1552				 */
1553				dm->host_specified_ha_region = true;
1554				ha_pg_range = &ha_msg->range;
1555				ha_region = &ha_pg_range[1];
1556				dm->ha_wrk.ha_page_range = *ha_pg_range;
1557				dm->ha_wrk.ha_region_range = *ha_region;
1558			}
1559			schedule_work(&dm_device.ha_wrk.wrk);
1560			break;
1561
1562		case DM_INFO_MESSAGE:
1563			process_info(dm, (struct dm_info_msg *)dm_msg);
1564			break;
1565
1566		default:
1567			pr_warn_ratelimited("Unhandled message: type: %d\n", dm_hdr->type);
1568		}
1569	}
1570}
1571
1572#define HV_LARGE_REPORTING_ORDER	9
1573#define HV_LARGE_REPORTING_LEN (HV_HYP_PAGE_SIZE << \
1574		HV_LARGE_REPORTING_ORDER)
1575static int hv_free_page_report(struct page_reporting_dev_info *pr_dev_info,
1576			       struct scatterlist *sgl, unsigned int nents)
1577{
1578	unsigned long flags;
1579	struct hv_memory_hint *hint;
1580	int i, order;
1581	u64 status;
1582	struct scatterlist *sg;
1583
1584	WARN_ON_ONCE(nents > HV_MEMORY_HINT_MAX_GPA_PAGE_RANGES);
1585	WARN_ON_ONCE(sgl->length < (HV_HYP_PAGE_SIZE << page_reporting_order));
1586	local_irq_save(flags);
1587	hint = *this_cpu_ptr(hyperv_pcpu_input_arg);
1588	if (!hint) {
1589		local_irq_restore(flags);
1590		return -ENOSPC;
1591	}
1592
1593	hint->heat_type = HV_EXTMEM_HEAT_HINT_COLD_DISCARD;
1594	hint->reserved = 0;
1595	for_each_sg(sgl, sg, nents, i) {
1596		union hv_gpa_page_range *range;
1597
1598		range = &hint->ranges[i];
1599		range->address_space = 0;
1600		order = get_order(sg->length);
1601		/*
1602		 * Hyper-V expects the additional_pages field in the units
1603		 * of one of these 3 sizes, 4Kbytes, 2Mbytes or 1Gbytes.
1604		 * This is dictated by the values of the fields page.largesize
1605		 * and page_size.
1606		 * This code however, only uses 4Kbytes and 2Mbytes units
1607		 * and not 1Gbytes unit.
1608		 */
1609
1610		/* page reporting for pages 2MB or higher */
1611		if (order >= HV_LARGE_REPORTING_ORDER) {
1612			range->page.largepage = 1;
1613			range->page_size = HV_GPA_PAGE_RANGE_PAGE_SIZE_2MB;
1614			range->base_large_pfn = page_to_hvpfn(
1615					sg_page(sg)) >> HV_LARGE_REPORTING_ORDER;
1616			range->page.additional_pages =
1617				(sg->length / HV_LARGE_REPORTING_LEN) - 1;
1618		} else {
1619			/* Page reporting for pages below 2MB */
1620			range->page.basepfn = page_to_hvpfn(sg_page(sg));
1621			range->page.largepage = false;
1622			range->page.additional_pages =
1623				(sg->length / HV_HYP_PAGE_SIZE) - 1;
1624		}
1625	}
1626
1627	status = hv_do_rep_hypercall(HV_EXT_CALL_MEMORY_HEAT_HINT, nents, 0,
1628				     hint, NULL);
1629	local_irq_restore(flags);
1630	if (!hv_result_success(status)) {
1631		pr_err("Cold memory discard hypercall failed with status %llx\n",
1632		       status);
1633		if (hv_hypercall_multi_failure > 0)
1634			hv_hypercall_multi_failure++;
1635
1636		if (hv_result(status) == HV_STATUS_INVALID_PARAMETER) {
1637			pr_err("Underlying Hyper-V does not support order less than 9. Hypercall failed\n");
1638			pr_err("Defaulting to page_reporting_order %d\n",
1639			       pageblock_order);
1640			page_reporting_order = pageblock_order;
1641			hv_hypercall_multi_failure++;
1642			return -EINVAL;
1643		}
1644
1645		return -EINVAL;
1646	}
1647
1648	return 0;
1649}
1650
1651static void enable_page_reporting(void)
1652{
1653	int ret;
1654
1655	if (!hv_query_ext_cap(HV_EXT_CAPABILITY_MEMORY_COLD_DISCARD_HINT)) {
1656		pr_debug("Cold memory discard hint not supported by Hyper-V\n");
1657		return;
1658	}
1659
1660	BUILD_BUG_ON(PAGE_REPORTING_CAPACITY > HV_MEMORY_HINT_MAX_GPA_PAGE_RANGES);
1661	dm_device.pr_dev_info.report = hv_free_page_report;
1662	/*
1663	 * We let the page_reporting_order parameter decide the order
1664	 * in the page_reporting code
1665	 */
1666	dm_device.pr_dev_info.order = 0;
1667	ret = page_reporting_register(&dm_device.pr_dev_info);
1668	if (ret < 0) {
1669		dm_device.pr_dev_info.report = NULL;
1670		pr_err("Failed to enable cold memory discard: %d\n", ret);
1671	} else {
1672		pr_info("Cold memory discard hint enabled with order %d\n",
1673			page_reporting_order);
1674	}
1675}
1676
1677static void disable_page_reporting(void)
1678{
1679	if (dm_device.pr_dev_info.report) {
1680		page_reporting_unregister(&dm_device.pr_dev_info);
1681		dm_device.pr_dev_info.report = NULL;
1682	}
1683}
1684
1685static int ballooning_enabled(void)
1686{
1687	/*
1688	 * Disable ballooning if the page size is not 4k (HV_HYP_PAGE_SIZE),
1689	 * since currently it's unclear to us whether an unballoon request can
1690	 * make sure all page ranges are guest page size aligned.
1691	 */
1692	if (PAGE_SIZE != HV_HYP_PAGE_SIZE) {
1693		pr_info("Ballooning disabled because page size is not 4096 bytes\n");
1694		return 0;
1695	}
1696
1697	return 1;
1698}
1699
1700static int hot_add_enabled(void)
1701{
1702	/*
1703	 * Disable hot add on ARM64, because we currently rely on
1704	 * memory_add_physaddr_to_nid() to get a node id of a hot add range,
1705	 * however ARM64's memory_add_physaddr_to_nid() always return 0 and
1706	 * DM_MEM_HOT_ADD_REQUEST doesn't have the NUMA node information for
1707	 * add_memory().
1708	 */
1709	if (IS_ENABLED(CONFIG_ARM64)) {
1710		pr_info("Memory hot add disabled on ARM64\n");
1711		return 0;
1712	}
1713
1714	return 1;
1715}
1716
1717static int balloon_connect_vsp(struct hv_device *dev)
1718{
1719	struct dm_version_request version_req;
1720	struct dm_capabilities cap_msg;
1721	unsigned long t;
1722	int ret;
1723
1724	/*
1725	 * max_pkt_size should be large enough for one vmbus packet header plus
1726	 * our receive buffer size. Hyper-V sends messages up to
1727	 * HV_HYP_PAGE_SIZE bytes long on balloon channel.
1728	 */
1729	dev->channel->max_pkt_size = HV_HYP_PAGE_SIZE * 2;
1730
1731	ret = vmbus_open(dev->channel, dm_ring_size, dm_ring_size, NULL, 0,
1732			 balloon_onchannelcallback, dev);
1733	if (ret)
1734		return ret;
1735
1736	/*
1737	 * Initiate the hand shake with the host and negotiate
1738	 * a version that the host can support. We start with the
1739	 * highest version number and go down if the host cannot
1740	 * support it.
1741	 */
1742	memset(&version_req, 0, sizeof(struct dm_version_request));
1743	version_req.hdr.type = DM_VERSION_REQUEST;
1744	version_req.hdr.size = sizeof(struct dm_version_request);
1745	version_req.hdr.trans_id = atomic_inc_return(&trans_id);
1746	version_req.version.version = DYNMEM_PROTOCOL_VERSION_WIN10;
1747	version_req.is_last_attempt = 0;
1748	dm_device.version = version_req.version.version;
1749
1750	ret = vmbus_sendpacket(dev->channel, &version_req,
1751			       sizeof(struct dm_version_request),
1752			       (unsigned long)NULL, VM_PKT_DATA_INBAND, 0);
1753	if (ret)
1754		goto out;
1755
1756	t = wait_for_completion_timeout(&dm_device.host_event, secs_to_jiffies(5));
1757	if (t == 0) {
1758		ret = -ETIMEDOUT;
1759		goto out;
1760	}
1761
1762	/*
1763	 * If we could not negotiate a compatible version with the host
1764	 * fail the probe function.
1765	 */
1766	if (dm_device.state == DM_INIT_ERROR) {
1767		ret = -EPROTO;
1768		goto out;
1769	}
1770
1771	pr_info("Using Dynamic Memory protocol version %u.%u\n",
1772		DYNMEM_MAJOR_VERSION(dm_device.version),
1773		DYNMEM_MINOR_VERSION(dm_device.version));
1774
1775	/*
1776	 * Now submit our capabilities to the host.
1777	 */
1778	memset(&cap_msg, 0, sizeof(struct dm_capabilities));
1779	cap_msg.hdr.type = DM_CAPABILITIES_REPORT;
1780	cap_msg.hdr.size = sizeof(struct dm_capabilities);
1781	cap_msg.hdr.trans_id = atomic_inc_return(&trans_id);
1782
1783	/*
1784	 * When hibernation (i.e. virtual ACPI S4 state) is enabled, the host
1785	 * currently still requires the bits to be set, so we have to add code
1786	 * to fail the host's hot-add and balloon up/down requests, if any.
1787	 */
1788	cap_msg.caps.cap_bits.balloon = ballooning_enabled();
1789	cap_msg.caps.cap_bits.hot_add = hot_add_enabled();
1790
1791	/*
1792	 * Specify our alignment requirements for memory hot-add. The value is
1793	 * the log base 2 of the number of megabytes in a chunk. For example,
1794	 * with 256 MiB chunks, the value is 8. The number of MiB in a chunk
1795	 * must be a power of 2.
1796	 */
1797	cap_msg.caps.cap_bits.hot_add_alignment =
1798					ilog2(HA_BYTES_IN_CHUNK / SZ_1M);
1799
1800	/*
1801	 * Currently the host does not use these
1802	 * values and we set them to what is done in the
1803	 * Windows driver.
1804	 */
1805	cap_msg.min_page_cnt = 0;
1806	cap_msg.max_page_number = -1;
1807
1808	ret = vmbus_sendpacket(dev->channel, &cap_msg,
1809			       sizeof(struct dm_capabilities),
1810			       (unsigned long)NULL, VM_PKT_DATA_INBAND, 0);
1811	if (ret)
1812		goto out;
1813
1814	t = wait_for_completion_timeout(&dm_device.host_event, secs_to_jiffies(5));
1815	if (t == 0) {
1816		ret = -ETIMEDOUT;
1817		goto out;
1818	}
1819
1820	/*
1821	 * If the host does not like our capabilities,
1822	 * fail the probe function.
1823	 */
1824	if (dm_device.state == DM_INIT_ERROR) {
1825		ret = -EPROTO;
1826		goto out;
1827	}
1828
1829	return 0;
1830out:
1831	vmbus_close(dev->channel);
1832	return ret;
1833}
1834
1835/*
1836 * DEBUGFS Interface
1837 */
1838#ifdef CONFIG_DEBUG_FS
1839
1840/**
1841 * hv_balloon_debug_show - shows statistics of balloon operations.
1842 * @f: pointer to the &struct seq_file.
1843 * @offset: ignored.
1844 *
1845 * Provides the statistics that can be accessed in hv-balloon in the debugfs.
1846 *
1847 * Return: zero on success or an error code.
1848 */
1849static int hv_balloon_debug_show(struct seq_file *f, void *offset)
1850{
1851	struct hv_dynmem_device *dm = f->private;
1852	char *sname;
1853
1854	seq_printf(f, "%-22s: %u.%u\n", "host_version",
1855			DYNMEM_MAJOR_VERSION(dm->version),
1856			DYNMEM_MINOR_VERSION(dm->version));
1857
1858	seq_printf(f, "%-22s:", "capabilities");
1859	if (ballooning_enabled())
1860		seq_puts(f, " enabled");
1861
1862	if (hot_add_enabled())
1863		seq_puts(f, " hot_add");
1864
1865	seq_puts(f, "\n");
1866
1867	seq_printf(f, "%-22s: %u", "state", dm->state);
1868	switch (dm->state) {
1869	case DM_INITIALIZING:
1870			sname = "Initializing";
1871			break;
1872	case DM_INITIALIZED:
1873			sname = "Initialized";
1874			break;
1875	case DM_BALLOON_UP:
1876			sname = "Balloon Up";
1877			break;
1878	case DM_BALLOON_DOWN:
1879			sname = "Balloon Down";
1880			break;
1881	case DM_HOT_ADD:
1882			sname = "Hot Add";
1883			break;
1884	case DM_INIT_ERROR:
1885			sname = "Error";
1886			break;
1887	default:
1888			sname = "Unknown";
1889	}
1890	seq_printf(f, " (%s)\n", sname);
1891
1892	/* HV Page Size */
1893	seq_printf(f, "%-22s: %ld\n", "page_size", HV_HYP_PAGE_SIZE);
1894
1895	/* Pages added with hot_add */
1896	seq_printf(f, "%-22s: %u\n", "pages_added", dm->num_pages_added);
1897
1898	/* pages that are "onlined"/used from pages_added */
1899	seq_printf(f, "%-22s: %u\n", "pages_onlined", dm->num_pages_onlined);
1900
1901	/* pages we have given back to host */
1902	seq_printf(f, "%-22s: %u\n", "pages_ballooned", dm->num_pages_ballooned);
1903
1904	seq_printf(f, "%-22s: %lu\n", "total_pages_committed",
1905		   get_pages_committed(dm));
1906
1907	seq_printf(f, "%-22s: %llu\n", "max_dynamic_page_count",
1908		   dm->max_dynamic_page_count);
1909
1910	return 0;
1911}
1912
1913DEFINE_SHOW_ATTRIBUTE(hv_balloon_debug);
1914
1915static void  hv_balloon_debugfs_init(struct hv_dynmem_device *b)
1916{
1917	debugfs_create_file("hv-balloon", 0444, NULL, b,
1918			    &hv_balloon_debug_fops);
1919}
1920
1921static void  hv_balloon_debugfs_exit(struct hv_dynmem_device *b)
1922{
1923	debugfs_lookup_and_remove("hv-balloon", NULL);
1924}
1925
1926#else
1927
1928static inline void hv_balloon_debugfs_init(struct hv_dynmem_device  *b)
1929{
1930}
1931
1932static inline void hv_balloon_debugfs_exit(struct hv_dynmem_device *b)
1933{
1934}
1935
1936#endif	/* CONFIG_DEBUG_FS */
1937
1938static int balloon_probe(struct hv_device *dev,
1939			 const struct hv_vmbus_device_id *dev_id)
1940{
1941	int ret;
1942
1943	allow_hibernation = hv_is_hibernation_supported();
1944	if (allow_hibernation)
1945		hot_add = false;
1946
1947#ifdef CONFIG_MEMORY_HOTPLUG
1948	/*
1949	 * Hot-add must operate in chunks that are of size equal to the
1950	 * memory block size because that's what the core add_memory()
1951	 * interface requires. The Hyper-V interface requires that the memory
1952	 * block size be a power of 2, which is guaranteed by the check in
1953	 * memory_dev_init().
1954	 */
1955	ha_pages_in_chunk = memory_block_size_bytes() / PAGE_SIZE;
1956	do_hot_add = hot_add;
1957#else
1958	/*
1959	 * Without MEMORY_HOTPLUG, the guest returns a failure status for all
1960	 * hot add requests from Hyper-V, and the chunk size is used only to
1961	 * specify alignment to Hyper-V as required by the host/guest protocol.
1962	 * Somewhat arbitrarily, use 128 MiB.
1963	 */
1964	ha_pages_in_chunk = SZ_128M / PAGE_SIZE;
1965	do_hot_add = false;
1966#endif
1967	dm_device.dev = dev;
1968	dm_device.state = DM_INITIALIZING;
1969	dm_device.next_version = DYNMEM_PROTOCOL_VERSION_WIN8;
1970	init_completion(&dm_device.host_event);
1971	init_completion(&dm_device.config_event);
1972	INIT_LIST_HEAD(&dm_device.ha_region_list);
1973	spin_lock_init(&dm_device.ha_lock);
1974	INIT_WORK(&dm_device.balloon_wrk.wrk, balloon_up);
1975	INIT_WORK(&dm_device.ha_wrk.wrk, hot_add_req);
1976	dm_device.host_specified_ha_region = false;
1977
1978#ifdef CONFIG_MEMORY_HOTPLUG
1979	set_online_page_callback(&hv_online_page);
1980	init_completion(&dm_device.ol_waitevent);
1981	register_memory_notifier(&hv_memory_nb);
1982#endif
1983
1984	hv_set_drvdata(dev, &dm_device);
1985
1986	ret = balloon_connect_vsp(dev);
1987	if (ret != 0)
1988		goto connect_error;
1989
1990	enable_page_reporting();
1991	dm_device.state = DM_INITIALIZED;
1992
1993	dm_device.thread =
1994		 kthread_run(dm_thread_func, &dm_device, "hv_balloon");
1995	if (IS_ERR(dm_device.thread)) {
1996		ret = PTR_ERR(dm_device.thread);
1997		goto probe_error;
1998	}
1999
2000	hv_balloon_debugfs_init(&dm_device);
2001
2002	return 0;
2003
2004probe_error:
2005	dm_device.state = DM_INIT_ERROR;
2006	dm_device.thread  = NULL;
2007	disable_page_reporting();
2008	vmbus_close(dev->channel);
2009connect_error:
2010#ifdef CONFIG_MEMORY_HOTPLUG
2011	unregister_memory_notifier(&hv_memory_nb);
2012	restore_online_page_callback(&hv_online_page);
2013#endif
2014	return ret;
2015}
2016
2017static void balloon_remove(struct hv_device *dev)
2018{
2019	struct hv_dynmem_device *dm = hv_get_drvdata(dev);
2020	struct hv_hotadd_state *has, *tmp;
2021	struct hv_hotadd_gap *gap, *tmp_gap;
2022
2023	if (dm->num_pages_ballooned != 0)
2024		pr_warn("Ballooned pages: %d\n", dm->num_pages_ballooned);
2025
2026	hv_balloon_debugfs_exit(dm);
2027
2028	cancel_work_sync(&dm->balloon_wrk.wrk);
2029	cancel_work_sync(&dm->ha_wrk.wrk);
2030
2031	kthread_stop(dm->thread);
2032
2033	/*
2034	 * This is to handle the case when balloon_resume()
2035	 * call has failed and some cleanup has been done as
2036	 * a part of the error handling.
2037	 */
2038	if (dm_device.state != DM_INIT_ERROR) {
2039		disable_page_reporting();
2040		vmbus_close(dev->channel);
2041#ifdef CONFIG_MEMORY_HOTPLUG
2042		unregister_memory_notifier(&hv_memory_nb);
2043		restore_online_page_callback(&hv_online_page);
2044#endif
2045	}
2046
2047	guard(spinlock_irqsave)(&dm_device.ha_lock);
2048	list_for_each_entry_safe(has, tmp, &dm->ha_region_list, list) {
2049		list_for_each_entry_safe(gap, tmp_gap, &has->gap_list, list) {
2050			list_del(&gap->list);
2051			kfree(gap);
2052		}
2053		list_del(&has->list);
2054		kfree(has);
2055	}
2056}
2057
2058static int balloon_suspend(struct hv_device *hv_dev)
2059{
2060	struct hv_dynmem_device *dm = hv_get_drvdata(hv_dev);
2061
2062	tasklet_disable(&hv_dev->channel->callback_event);
2063
2064	cancel_work_sync(&dm->balloon_wrk.wrk);
2065	cancel_work_sync(&dm->ha_wrk.wrk);
2066
2067	if (dm->thread) {
2068		kthread_stop(dm->thread);
2069		dm->thread = NULL;
2070		vmbus_close(hv_dev->channel);
2071	}
2072
2073	tasklet_enable(&hv_dev->channel->callback_event);
2074
2075	return 0;
2076}
2077
2078static int balloon_resume(struct hv_device *dev)
2079{
2080	int ret;
2081
2082	dm_device.state = DM_INITIALIZING;
2083
2084	ret = balloon_connect_vsp(dev);
2085
2086	if (ret != 0)
2087		goto out;
2088
2089	dm_device.thread =
2090		 kthread_run(dm_thread_func, &dm_device, "hv_balloon");
2091	if (IS_ERR(dm_device.thread)) {
2092		ret = PTR_ERR(dm_device.thread);
2093		dm_device.thread = NULL;
2094		goto close_channel;
2095	}
2096
2097	dm_device.state = DM_INITIALIZED;
2098	return 0;
2099close_channel:
2100	vmbus_close(dev->channel);
2101out:
2102	dm_device.state = DM_INIT_ERROR;
2103	disable_page_reporting();
2104#ifdef CONFIG_MEMORY_HOTPLUG
2105	unregister_memory_notifier(&hv_memory_nb);
2106	restore_online_page_callback(&hv_online_page);
2107#endif
2108	return ret;
2109}
2110
2111static const struct hv_vmbus_device_id id_table[] = {
2112	/* Dynamic Memory Class ID */
2113	/* 525074DC-8985-46e2-8057-A307DC18A502 */
2114	{ HV_DM_GUID, },
2115	{ },
2116};
2117
2118MODULE_DEVICE_TABLE(vmbus, id_table);
2119
2120static  struct hv_driver balloon_drv = {
2121	.name = "hv_balloon",
2122	.id_table = id_table,
2123	.probe =  balloon_probe,
2124	.remove =  balloon_remove,
2125	.suspend = balloon_suspend,
2126	.resume = balloon_resume,
2127	.driver = {
2128		.probe_type = PROBE_PREFER_ASYNCHRONOUS,
2129	},
2130};
2131
2132static int __init init_balloon_drv(void)
2133{
2134	return vmbus_driver_register(&balloon_drv);
2135}
2136
2137module_init(init_balloon_drv);
2138
2139MODULE_DESCRIPTION("Hyper-V Balloon");
2140MODULE_LICENSE("GPL");