drivers/hv/channel_mgmt.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (c) 2009, Microsoft Corporation.
   4  *
   5  * Authors:
   6  *   Haiyang Zhang <haiyangz@microsoft.com>
   7  *   Hank Janssen  <hjanssen@microsoft.com>
   8  */
   9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  10
  11 #include <linux/kernel.h>
  12 #include <linux/interrupt.h>
  13 #include <linux/sched.h>
  14 #include <linux/wait.h>
  15 #include <linux/mm.h>
  16 #include <linux/slab.h>
  17 #include <linux/list.h>
  18 #include <linux/module.h>
  19 #include <linux/completion.h>
  20 #include <linux/delay.h>
  21 #include <linux/cpu.h>
  22 #include <linux/hyperv.h>
  23 #include <asm/mshyperv.h>
  24
  25 #include "hyperv_vmbus.h"
  26
  27 static void init_vp_index(struct vmbus_channel *channel);
  28
  29 const struct vmbus_device vmbus_devs[] = {
  30         /* IDE */
  31         { .dev_type = HV_IDE,
  32           HV_IDE_GUID,
  33           .perf_device = true,
  34         },
  35
  36         /* SCSI */
  37         { .dev_type = HV_SCSI,
  38           HV_SCSI_GUID,
  39           .perf_device = true,
  40         },
  41
  42         /* Fibre Channel */
  43         { .dev_type = HV_FC,
  44           HV_SYNTHFC_GUID,
  45           .perf_device = true,
  46         },
  47
  48         /* Synthetic NIC */
  49         { .dev_type = HV_NIC,
  50           HV_NIC_GUID,
  51           .perf_device = true,
  52         },
  53
  54         /* Network Direct */
  55         { .dev_type = HV_ND,
  56           HV_ND_GUID,
  57           .perf_device = true,
  58         },
  59
  60         /* PCIE */
  61         { .dev_type = HV_PCIE,
  62           HV_PCIE_GUID,
  63           .perf_device = false,
  64         },
  65
  66         /* Synthetic Frame Buffer */
  67         { .dev_type = HV_FB,
  68           HV_SYNTHVID_GUID,
  69           .perf_device = false,
  70         },
  71
  72         /* Synthetic Keyboard */
  73         { .dev_type = HV_KBD,
  74           HV_KBD_GUID,
  75           .perf_device = false,
  76         },
  77
  78         /* Synthetic MOUSE */
  79         { .dev_type = HV_MOUSE,
  80           HV_MOUSE_GUID,
  81           .perf_device = false,
  82         },
  83
  84         /* KVP */
  85         { .dev_type = HV_KVP,
  86           HV_KVP_GUID,
  87           .perf_device = false,
  88         },
  89
  90         /* Time Synch */
  91         { .dev_type = HV_TS,
  92           HV_TS_GUID,
  93           .perf_device = false,
  94         },
  95
  96         /* Heartbeat */
  97         { .dev_type = HV_HB,
  98           HV_HEART_BEAT_GUID,
  99           .perf_device = false,
 100         },
 101
 102         /* Shutdown */
 103         { .dev_type = HV_SHUTDOWN,
 104           HV_SHUTDOWN_GUID,
 105           .perf_device = false,
 106         },
 107
 108         /* File copy */
 109         { .dev_type = HV_FCOPY,
 110           HV_FCOPY_GUID,
 111           .perf_device = false,
 112         },
 113
 114         /* Backup */
 115         { .dev_type = HV_BACKUP,
 116           HV_VSS_GUID,
 117           .perf_device = false,
 118         },
 119
 120         /* Dynamic Memory */
 121         { .dev_type = HV_DM,
 122           HV_DM_GUID,
 123           .perf_device = false,
 124         },
 125
 126         /* Unknown GUID */
 127         { .dev_type = HV_UNKNOWN,
 128           .perf_device = false,
 129         },
 130 };
 131
 132 static const struct {
 133         guid_t guid;
 134 } vmbus_unsupported_devs[] = {
 135         { HV_AVMA1_GUID },
 136         { HV_AVMA2_GUID },
 137         { HV_RDV_GUID   },
 138 };
 139
 140 /*
 141  * The rescinded channel may be blocked waiting for a response from the host;
 142  * take care of that.
 143  */
 144 static void vmbus_rescind_cleanup(struct vmbus_channel *channel)
 145 {
 146         struct vmbus_channel_msginfo *msginfo;
 147         unsigned long flags;
 148
 149
 150         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
 151         channel->rescind = true;
 152         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
 153                                 msglistentry) {
 154
 155                 if (msginfo->waiting_channel == channel) {
 156                         complete(&msginfo->waitevent);
 157                         break;
 158                 }
 159         }
 160         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
 161 }
 162
 163 static bool is_unsupported_vmbus_devs(const guid_t *guid)
 164 {
 165         int i;
 166
 167         for (i = 0; i < ARRAY_SIZE(vmbus_unsupported_devs); i++)
 168                 if (guid_equal(guid, &vmbus_unsupported_devs[i].guid))
 169                         return true;
 170         return false;
 171 }
 172
 173 static u16 hv_get_dev_type(const struct vmbus_channel *channel)
 174 {
 175         const guid_t *guid = &channel->offermsg.offer.if_type;
 176         u16 i;
 177
 178         if (is_hvsock_channel(channel) || is_unsupported_vmbus_devs(guid))
 179                 return HV_UNKNOWN;
 180
 181         for (i = HV_IDE; i < HV_UNKNOWN; i++) {
 182                 if (guid_equal(guid, &vmbus_devs[i].guid))
 183                         return i;
 184         }
 185         pr_info("Unknown GUID: %pUl\n", guid);
 186         return i;
 187 }
 188
 189 /**
 190  * vmbus_prep_negotiate_resp() - Create default response for Negotiate message
 191  * @icmsghdrp: Pointer to msg header structure
 192  * @buf: Raw buffer channel data
 193  * @fw_version: The framework versions we can support.
 194  * @fw_vercnt: The size of @fw_version.
 195  * @srv_version: The service versions we can support.
 196  * @srv_vercnt: The size of @srv_version.
 197  * @nego_fw_version: The selected framework version.
 198  * @nego_srv_version: The selected service version.
 199  *
 200  * Note: Versions are given in decreasing order.
 201  *
 202  * Set up and fill in default negotiate response message.
 203  * Mainly used by Hyper-V drivers.
 204  */
 205 bool vmbus_prep_negotiate_resp(struct icmsg_hdr *icmsghdrp,
 206                                 u8 *buf, const int *fw_version, int fw_vercnt,
 207                                 const int *srv_version, int srv_vercnt,
 208                                 int *nego_fw_version, int *nego_srv_version)
 209 {
 210         int icframe_major, icframe_minor;
 211         int icmsg_major, icmsg_minor;
 212         int fw_major, fw_minor;
 213         int srv_major, srv_minor;
 214         int i, j;
 215         bool found_match = false;
 216         struct icmsg_negotiate *negop;
 217
 218         icmsghdrp->icmsgsize = 0x10;
 219         negop = (struct icmsg_negotiate *)&buf[
 220                 sizeof(struct vmbuspipe_hdr) +
 221                 sizeof(struct icmsg_hdr)];
 222
 223         icframe_major = negop->icframe_vercnt;
 224         icframe_minor = 0;
 225
 226         icmsg_major = negop->icmsg_vercnt;
 227         icmsg_minor = 0;
 228
 229         /*
 230          * Select the framework version number we will
 231          * support.
 232          */
 233
 234         for (i = 0; i < fw_vercnt; i++) {
 235                 fw_major = (fw_version[i] >> 16);
 236                 fw_minor = (fw_version[i] & 0xFFFF);
 237
 238                 for (j = 0; j < negop->icframe_vercnt; j++) {
 239                         if ((negop->icversion_data[j].major == fw_major) &&
 240                             (negop->icversion_data[j].minor == fw_minor)) {
 241                                 icframe_major = negop->icversion_data[j].major;
 242                                 icframe_minor = negop->icversion_data[j].minor;
 243                                 found_match = true;
 244                                 break;
 245                         }
 246                 }
 247
 248                 if (found_match)
 249                         break;
 250         }
 251
 252         if (!found_match)
 253                 goto fw_error;
 254
 255         found_match = false;
 256
 257         for (i = 0; i < srv_vercnt; i++) {
 258                 srv_major = (srv_version[i] >> 16);
 259                 srv_minor = (srv_version[i] & 0xFFFF);
 260
 261                 for (j = negop->icframe_vercnt;
 262                         (j < negop->icframe_vercnt + negop->icmsg_vercnt);
 263                         j++) {
 264
 265                         if ((negop->icversion_data[j].major == srv_major) &&
 266                                 (negop->icversion_data[j].minor == srv_minor)) {
 267
 268                                 icmsg_major = negop->icversion_data[j].major;
 269                                 icmsg_minor = negop->icversion_data[j].minor;
 270                                 found_match = true;
 271                                 break;
 272                         }
 273                 }
 274
 275                 if (found_match)
 276                         break;
 277         }
 278
 279         /*
 280          * Respond with the framework and service
 281          * version numbers we can support.
 282          */
 283
 284 fw_error:
 285         if (!found_match) {
 286                 negop->icframe_vercnt = 0;
 287                 negop->icmsg_vercnt = 0;
 288         } else {
 289                 negop->icframe_vercnt = 1;
 290                 negop->icmsg_vercnt = 1;
 291         }
 292
 293         if (nego_fw_version)
 294                 *nego_fw_version = (icframe_major << 16) | icframe_minor;
 295
 296         if (nego_srv_version)
 297                 *nego_srv_version = (icmsg_major << 16) | icmsg_minor;
 298
 299         negop->icversion_data[0].major = icframe_major;
 300         negop->icversion_data[0].minor = icframe_minor;
 301         negop->icversion_data[1].major = icmsg_major;
 302         negop->icversion_data[1].minor = icmsg_minor;
 303         return found_match;
 304 }
 305
 306 EXPORT_SYMBOL_GPL(vmbus_prep_negotiate_resp);
 307
 308 /*
 309  * alloc_channel - Allocate and initialize a vmbus channel object
 310  */
 311 static struct vmbus_channel *alloc_channel(void)
 312 {
 313         struct vmbus_channel *channel;
 314
 315         channel = kzalloc(sizeof(*channel), GFP_ATOMIC);
 316         if (!channel)
 317                 return NULL;
 318
 319         spin_lock_init(&channel->sched_lock);
 320         init_completion(&channel->rescind_event);
 321
 322         INIT_LIST_HEAD(&channel->sc_list);
 323
 324         tasklet_init(&channel->callback_event,
 325                      vmbus_on_event, (unsigned long)channel);
 326
 327         hv_ringbuffer_pre_init(channel);
 328
 329         return channel;
 330 }
 331
 332 /*
 333  * free_channel - Release the resources used by the vmbus channel object
 334  */
 335 static void free_channel(struct vmbus_channel *channel)
 336 {
 337         tasklet_kill(&channel->callback_event);
 338         vmbus_remove_channel_attr_group(channel);
 339
 340         kobject_put(&channel->kobj);
 341 }
 342
 343 void vmbus_channel_map_relid(struct vmbus_channel *channel)
 344 {
 345         if (WARN_ON(channel->offermsg.child_relid >= MAX_CHANNEL_RELIDS))
 346                 return;
 347         /*
 348          * The mapping of the channel's relid is visible from the CPUs that
 349          * execute vmbus_chan_sched() by the time that vmbus_chan_sched() will
 350          * execute:
 351          *
 352          *  (a) In the "normal (i.e., not resuming from hibernation)" path,
 353          *      the full barrier in virt_store_mb() guarantees that the store
 354          *      is propagated to all CPUs before the add_channel_work work
 355          *      is queued.  In turn, add_channel_work is queued before the
 356          *      channel's ring buffer is allocated/initialized and the
 357          *      OPENCHANNEL message for the channel is sent in vmbus_open().
 358          *      Hyper-V won't start sending the interrupts for the channel
 359          *      before the OPENCHANNEL message is acked.  The memory barrier
 360          *      in vmbus_chan_sched() -> sync_test_and_clear_bit() ensures
 361          *      that vmbus_chan_sched() must find the channel's relid in
 362          *      recv_int_page before retrieving the channel pointer from the
 363          *      array of channels.
 364          *
 365          *  (b) In the "resuming from hibernation" path, the virt_store_mb()
 366          *      guarantees that the store is propagated to all CPUs before
 367          *      the VMBus connection is marked as ready for the resume event
 368          *      (cf. check_ready_for_resume_event()).  The interrupt handler
 369          *      of the VMBus driver and vmbus_chan_sched() can not run before
 370          *      vmbus_bus_resume() has completed execution (cf. resume_noirq).
 371          */
 372         virt_store_mb(
 373                 vmbus_connection.channels[channel->offermsg.child_relid],
 374                 channel);
 375 }
 376
 377 void vmbus_channel_unmap_relid(struct vmbus_channel *channel)
 378 {
 379         if (WARN_ON(channel->offermsg.child_relid >= MAX_CHANNEL_RELIDS))
 380                 return;
 381         WRITE_ONCE(
 382                 vmbus_connection.channels[channel->offermsg.child_relid],
 383                 NULL);
 384 }
 385
 386 static void vmbus_release_relid(u32 relid)
 387 {
 388         struct vmbus_channel_relid_released msg;
 389         int ret;
 390
 391         memset(&msg, 0, sizeof(struct vmbus_channel_relid_released));
 392         msg.child_relid = relid;
 393         msg.header.msgtype = CHANNELMSG_RELID_RELEASED;
 394         ret = vmbus_post_msg(&msg, sizeof(struct vmbus_channel_relid_released),
 395                              true);
 396
 397         trace_vmbus_release_relid(&msg, ret);
 398 }
 399
 400 void hv_process_channel_removal(struct vmbus_channel *channel)
 401 {
 402         lockdep_assert_held(&vmbus_connection.channel_mutex);
 403         BUG_ON(!channel->rescind);
 404
 405         /*
 406          * hv_process_channel_removal() could find INVALID_RELID only for
 407          * hv_sock channels.  See the inline comments in vmbus_onoffer().
 408          */
 409         WARN_ON(channel->offermsg.child_relid == INVALID_RELID &&
 410                 !is_hvsock_channel(channel));
 411
 412         /*
 413          * Upon suspend, an in-use hv_sock channel is removed from the array of
 414          * channels and the relid is invalidated.  After hibernation, when the
 415          * user-space appplication destroys the channel, it's unnecessary and
 416          * unsafe to remove the channel from the array of channels.  See also
 417          * the inline comments before the call of vmbus_release_relid() below.
 418          */
 419         if (channel->offermsg.child_relid != INVALID_RELID)
 420                 vmbus_channel_unmap_relid(channel);
 421
 422         if (channel->primary_channel == NULL)
 423                 list_del(&channel->listentry);
 424         else
 425                 list_del(&channel->sc_list);
 426
 427         /*
 428          * If this is a "perf" channel, updates the hv_numa_map[] masks so that
 429          * init_vp_index() can (re-)use the CPU.
 430          */
 431         if (hv_is_perf_channel(channel))
 432                 hv_clear_alloced_cpu(channel->target_cpu);
 433
 434         /*
 435          * Upon suspend, an in-use hv_sock channel is marked as "rescinded" and
 436          * the relid is invalidated; after hibernation, when the user-space app
 437          * destroys the channel, the relid is INVALID_RELID, and in this case
 438          * it's unnecessary and unsafe to release the old relid, since the same
 439          * relid can refer to a completely different channel now.
 440          */
 441         if (channel->offermsg.child_relid != INVALID_RELID)
 442                 vmbus_release_relid(channel->offermsg.child_relid);
 443
 444         free_channel(channel);
 445 }
 446
 447 void vmbus_free_channels(void)
 448 {
 449         struct vmbus_channel *channel, *tmp;
 450
 451         list_for_each_entry_safe(channel, tmp, &vmbus_connection.chn_list,
 452                 listentry) {
 453                 /* hv_process_channel_removal() needs this */
 454                 channel->rescind = true;
 455
 456                 vmbus_device_unregister(channel->device_obj);
 457         }
 458 }
 459
 460 /* Note: the function can run concurrently for primary/sub channels. */
 461 static void vmbus_add_channel_work(struct work_struct *work)
 462 {
 463         struct vmbus_channel *newchannel =
 464                 container_of(work, struct vmbus_channel, add_channel_work);
 465         struct vmbus_channel *primary_channel = newchannel->primary_channel;
 466         int ret;
 467
 468         /*
 469          * This state is used to indicate a successful open
 470          * so that when we do close the channel normally, we
 471          * can cleanup properly.
 472          */
 473         newchannel->state = CHANNEL_OPEN_STATE;
 474
 475         if (primary_channel != NULL) {
 476                 /* newchannel is a sub-channel. */
 477                 struct hv_device *dev = primary_channel->device_obj;
 478
 479                 if (vmbus_add_channel_kobj(dev, newchannel))
 480                         goto err_deq_chan;
 481
 482                 if (primary_channel->sc_creation_callback != NULL)
 483                         primary_channel->sc_creation_callback(newchannel);
 484
 485                 newchannel->probe_done = true;
 486                 return;
 487         }
 488
 489         /*
 490          * Start the process of binding the primary channel to the driver
 491          */
 492         newchannel->device_obj = vmbus_device_create(
 493                 &newchannel->offermsg.offer.if_type,
 494                 &newchannel->offermsg.offer.if_instance,
 495                 newchannel);
 496         if (!newchannel->device_obj)
 497                 goto err_deq_chan;
 498
 499         newchannel->device_obj->device_id = newchannel->device_id;
 500         /*
 501          * Add the new device to the bus. This will kick off device-driver
 502          * binding which eventually invokes the device driver's AddDevice()
 503          * method.
 504          */
 505         ret = vmbus_device_register(newchannel->device_obj);
 506
 507         if (ret != 0) {
 508                 pr_err("unable to add child device object (relid %d)\n",
 509                         newchannel->offermsg.child_relid);
 510                 kfree(newchannel->device_obj);
 511                 goto err_deq_chan;
 512         }
 513
 514         newchannel->probe_done = true;
 515         return;
 516
 517 err_deq_chan:
 518         mutex_lock(&vmbus_connection.channel_mutex);
 519
 520         /*
 521          * We need to set the flag, otherwise
 522          * vmbus_onoffer_rescind() can be blocked.
 523          */
 524         newchannel->probe_done = true;
 525
 526         if (primary_channel == NULL)
 527                 list_del(&newchannel->listentry);
 528         else
 529                 list_del(&newchannel->sc_list);
 530
 531         /* vmbus_process_offer() has mapped the channel. */
 532         vmbus_channel_unmap_relid(newchannel);
 533
 534         mutex_unlock(&vmbus_connection.channel_mutex);
 535
 536         vmbus_release_relid(newchannel->offermsg.child_relid);
 537
 538         free_channel(newchannel);
 539 }
 540
 541 /*
 542  * vmbus_process_offer - Process the offer by creating a channel/device
 543  * associated with this offer
 544  */
 545 static void vmbus_process_offer(struct vmbus_channel *newchannel)
 546 {
 547         struct vmbus_channel *channel;
 548         struct workqueue_struct *wq;
 549         bool fnew = true;
 550
 551         /*
 552          * Synchronize vmbus_process_offer() and CPU hotplugging:
 553          *
 554          * CPU1                         CPU2
 555          *
 556          * [vmbus_process_offer()]      [Hot removal of the CPU]
 557          *
 558          * CPU_READ_LOCK                CPUS_WRITE_LOCK
 559          * LOAD cpu_online_mask         SEARCH chn_list
 560          * STORE target_cpu             LOAD target_cpu
 561          * INSERT chn_list              STORE cpu_online_mask
 562          * CPUS_READ_UNLOCK             CPUS_WRITE_UNLOCK
 563          *
 564          * Forbids: CPU1's LOAD from *not* seing CPU2's STORE &&
 565          *              CPU2's SEARCH from *not* seeing CPU1's INSERT
 566          *
 567          * Forbids: CPU2's SEARCH from seeing CPU1's INSERT &&
 568          *              CPU2's LOAD from *not* seing CPU1's STORE
 569          */
 570         cpus_read_lock();
 571
 572         /*
 573          * Serializes the modifications of the chn_list list as well as
 574          * the accesses to next_numa_node_id in init_vp_index().
 575          */
 576         mutex_lock(&vmbus_connection.channel_mutex);
 577
 578         init_vp_index(newchannel);
 579
 580         /* Remember the channels that should be cleaned up upon suspend. */
 581         if (is_hvsock_channel(newchannel) || is_sub_channel(newchannel))
 582                 atomic_inc(&vmbus_connection.nr_chan_close_on_suspend);
 583
 584         /*
 585          * Now that we have acquired the channel_mutex,
 586          * we can release the potentially racing rescind thread.
 587          */
 588         atomic_dec(&vmbus_connection.offer_in_progress);
 589
 590         list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
 591                 if (guid_equal(&channel->offermsg.offer.if_type,
 592                                &newchannel->offermsg.offer.if_type) &&
 593                     guid_equal(&channel->offermsg.offer.if_instance,
 594                                &newchannel->offermsg.offer.if_instance)) {
 595                         fnew = false;
 596                         break;
 597                 }
 598         }
 599
 600         if (fnew) {
 601                 list_add_tail(&newchannel->listentry,
 602                               &vmbus_connection.chn_list);
 603         } else {
 604                 /*
 605                  * Check to see if this is a valid sub-channel.
 606                  */
 607                 if (newchannel->offermsg.offer.sub_channel_index == 0) {
 608                         mutex_unlock(&vmbus_connection.channel_mutex);
 609                         cpus_read_unlock();
 610                         /*
 611                          * Don't call free_channel(), because newchannel->kobj
 612                          * is not initialized yet.
 613                          */
 614                         kfree(newchannel);
 615                         WARN_ON_ONCE(1);
 616                         return;
 617                 }
 618                 /*
 619                  * Process the sub-channel.
 620                  */
 621                 newchannel->primary_channel = channel;
 622                 list_add_tail(&newchannel->sc_list, &channel->sc_list);
 623         }
 624
 625         vmbus_channel_map_relid(newchannel);
 626
 627         mutex_unlock(&vmbus_connection.channel_mutex);
 628         cpus_read_unlock();
 629
 630         /*
 631          * vmbus_process_offer() mustn't call channel->sc_creation_callback()
 632          * directly for sub-channels, because sc_creation_callback() ->
 633          * vmbus_open() may never get the host's response to the
 634          * OPEN_CHANNEL message (the host may rescind a channel at any time,
 635          * e.g. in the case of hot removing a NIC), and vmbus_onoffer_rescind()
 636          * may not wake up the vmbus_open() as it's blocked due to a non-zero
 637          * vmbus_connection.offer_in_progress, and finally we have a deadlock.
 638          *
 639          * The above is also true for primary channels, if the related device
 640          * drivers use sync probing mode by default.
 641          *
 642          * And, usually the handling of primary channels and sub-channels can
 643          * depend on each other, so we should offload them to different
 644          * workqueues to avoid possible deadlock, e.g. in sync-probing mode,
 645          * NIC1's netvsc_subchan_work() can race with NIC2's netvsc_probe() ->
 646          * rtnl_lock(), and causes deadlock: the former gets the rtnl_lock
 647          * and waits for all the sub-channels to appear, but the latter
 648          * can't get the rtnl_lock and this blocks the handling of
 649          * sub-channels.
 650          */
 651         INIT_WORK(&newchannel->add_channel_work, vmbus_add_channel_work);
 652         wq = fnew ? vmbus_connection.handle_primary_chan_wq :
 653                     vmbus_connection.handle_sub_chan_wq;
 654         queue_work(wq, &newchannel->add_channel_work);
 655 }
 656
 657 /*
 658  * We use this state to statically distribute the channel interrupt load.
 659  */
 660 static int next_numa_node_id;
 661
 662 /*
 663  * Starting with Win8, we can statically distribute the incoming
 664  * channel interrupt load by binding a channel to VCPU.
 665  *
 666  * For pre-win8 hosts or non-performance critical channels we assign the
 667  * VMBUS_CONNECT_CPU.
 668  *
 669  * Starting with win8, performance critical channels will be distributed
 670  * evenly among all the available NUMA nodes.  Once the node is assigned,
 671  * we will assign the CPU based on a simple round robin scheme.
 672  */
 673 static void init_vp_index(struct vmbus_channel *channel)
 674 {
 675         bool perf_chn = hv_is_perf_channel(channel);
 676         cpumask_var_t available_mask;
 677         struct cpumask *alloced_mask;
 678         u32 target_cpu;
 679         int numa_node;
 680
 681         if ((vmbus_proto_version == VERSION_WS2008) ||
 682             (vmbus_proto_version == VERSION_WIN7) || (!perf_chn) ||
 683             !alloc_cpumask_var(&available_mask, GFP_KERNEL)) {
 684                 /*
 685                  * Prior to win8, all channel interrupts are
 686                  * delivered on VMBUS_CONNECT_CPU.
 687                  * Also if the channel is not a performance critical
 688                  * channel, bind it to VMBUS_CONNECT_CPU.
 689                  * In case alloc_cpumask_var() fails, bind it to
 690                  * VMBUS_CONNECT_CPU.
 691                  */
 692                 channel->target_cpu = VMBUS_CONNECT_CPU;
 693                 if (perf_chn)
 694                         hv_set_alloced_cpu(VMBUS_CONNECT_CPU);
 695                 return;
 696         }
 697
 698         while (true) {
 699                 numa_node = next_numa_node_id++;
 700                 if (numa_node == nr_node_ids) {
 701                         next_numa_node_id = 0;
 702                         continue;
 703                 }
 704                 if (cpumask_empty(cpumask_of_node(numa_node)))
 705                         continue;
 706                 break;
 707         }
 708         alloced_mask = &hv_context.hv_numa_map[numa_node];
 709
 710         if (cpumask_weight(alloced_mask) ==
 711             cpumask_weight(cpumask_of_node(numa_node))) {
 712                 /*
 713                  * We have cycled through all the CPUs in the node;
 714                  * reset the alloced map.
 715                  */
 716                 cpumask_clear(alloced_mask);
 717         }
 718
 719         cpumask_xor(available_mask, alloced_mask, cpumask_of_node(numa_node));
 720
 721         target_cpu = cpumask_first(available_mask);
 722         cpumask_set_cpu(target_cpu, alloced_mask);
 723
 724         channel->target_cpu = target_cpu;
 725
 726         free_cpumask_var(available_mask);
 727 }
 728
 729 #define UNLOAD_DELAY_UNIT_MS    10              /* 10 milliseconds */
 730 #define UNLOAD_WAIT_MS          (100*1000)      /* 100 seconds */
 731 #define UNLOAD_WAIT_LOOPS       (UNLOAD_WAIT_MS/UNLOAD_DELAY_UNIT_MS)
 732 #define UNLOAD_MSG_MS           (5*1000)        /* Every 5 seconds */
 733 #define UNLOAD_MSG_LOOPS        (UNLOAD_MSG_MS/UNLOAD_DELAY_UNIT_MS)
 734
 735 static void vmbus_wait_for_unload(void)
 736 {
 737         int cpu;
 738         void *page_addr;
 739         struct hv_message *msg;
 740         struct vmbus_channel_message_header *hdr;
 741         u32 message_type, i;
 742
 743         /*
 744          * CHANNELMSG_UNLOAD_RESPONSE is always delivered to the CPU which was
 745          * used for initial contact or to CPU0 depending on host version. When
 746          * we're crashing on a different CPU let's hope that IRQ handler on
 747          * the cpu which receives CHANNELMSG_UNLOAD_RESPONSE is still
 748          * functional and vmbus_unload_response() will complete
 749          * vmbus_connection.unload_event. If not, the last thing we can do is
 750          * read message pages for all CPUs directly.
 751          *
 752          * Wait up to 100 seconds since an Azure host must writeback any dirty
 753          * data in its disk cache before the VMbus UNLOAD request will
 754          * complete. This flushing has been empirically observed to take up
 755          * to 50 seconds in cases with a lot of dirty data, so allow additional
 756          * leeway and for inaccuracies in mdelay(). But eventually time out so
 757          * that the panic path can't get hung forever in case the response
 758          * message isn't seen.
 759          */
 760         for (i = 1; i <= UNLOAD_WAIT_LOOPS; i++) {
 761                 if (completion_done(&vmbus_connection.unload_event))
 762                         goto completed;
 763
 764                 for_each_online_cpu(cpu) {
 765                         struct hv_per_cpu_context *hv_cpu
 766                                 = per_cpu_ptr(hv_context.cpu_context, cpu);
 767
 768                         page_addr = hv_cpu->synic_message_page;
 769                         msg = (struct hv_message *)page_addr
 770                                 + VMBUS_MESSAGE_SINT;
 771
 772                         message_type = READ_ONCE(msg->header.message_type);
 773                         if (message_type == HVMSG_NONE)
 774                                 continue;
 775
 776                         hdr = (struct vmbus_channel_message_header *)
 777                                 msg->u.payload;
 778
 779                         if (hdr->msgtype == CHANNELMSG_UNLOAD_RESPONSE)
 780                                 complete(&vmbus_connection.unload_event);
 781
 782                         vmbus_signal_eom(msg, message_type);
 783                 }
 784
 785                 /*
 786                  * Give a notice periodically so someone watching the
 787                  * serial output won't think it is completely hung.
 788                  */
 789                 if (!(i % UNLOAD_MSG_LOOPS))
 790                         pr_notice("Waiting for VMBus UNLOAD to complete\n");
 791
 792                 mdelay(UNLOAD_DELAY_UNIT_MS);
 793         }
 794         pr_err("Continuing even though VMBus UNLOAD did not complete\n");
 795
 796 completed:
 797         /*
 798          * We're crashing and already got the UNLOAD_RESPONSE, cleanup all
 799          * maybe-pending messages on all CPUs to be able to receive new
 800          * messages after we reconnect.
 801          */
 802         for_each_online_cpu(cpu) {
 803                 struct hv_per_cpu_context *hv_cpu
 804                         = per_cpu_ptr(hv_context.cpu_context, cpu);
 805
 806                 page_addr = hv_cpu->synic_message_page;
 807                 msg = (struct hv_message *)page_addr + VMBUS_MESSAGE_SINT;
 808                 msg->header.message_type = HVMSG_NONE;
 809         }
 810 }
 811
 812 /*
 813  * vmbus_unload_response - Handler for the unload response.
 814  */
 815 static void vmbus_unload_response(struct vmbus_channel_message_header *hdr)
 816 {
 817         /*
 818          * This is a global event; just wakeup the waiting thread.
 819          * Once we successfully unload, we can cleanup the monitor state.
 820          */
 821         complete(&vmbus_connection.unload_event);
 822 }
 823
 824 void vmbus_initiate_unload(bool crash)
 825 {
 826         struct vmbus_channel_message_header hdr;
 827
 828         if (xchg(&vmbus_connection.conn_state, DISCONNECTED) == DISCONNECTED)
 829                 return;
 830
 831         /* Pre-Win2012R2 hosts don't support reconnect */
 832         if (vmbus_proto_version < VERSION_WIN8_1)
 833                 return;
 834
 835         init_completion(&vmbus_connection.unload_event);
 836         memset(&hdr, 0, sizeof(struct vmbus_channel_message_header));
 837         hdr.msgtype = CHANNELMSG_UNLOAD;
 838         vmbus_post_msg(&hdr, sizeof(struct vmbus_channel_message_header),
 839                        !crash);
 840
 841         /*
 842          * vmbus_initiate_unload() is also called on crash and the crash can be
 843          * happening in an interrupt context, where scheduling is impossible.
 844          */
 845         if (!crash)
 846                 wait_for_completion(&vmbus_connection.unload_event);
 847         else
 848                 vmbus_wait_for_unload();
 849 }
 850
 851 static void check_ready_for_resume_event(void)
 852 {
 853         /*
 854          * If all the old primary channels have been fixed up, then it's safe
 855          * to resume.
 856          */
 857         if (atomic_dec_and_test(&vmbus_connection.nr_chan_fixup_on_resume))
 858                 complete(&vmbus_connection.ready_for_resume_event);
 859 }
 860
 861 static void vmbus_setup_channel_state(struct vmbus_channel *channel,
 862                                       struct vmbus_channel_offer_channel *offer)
 863 {
 864         /*
 865          * Setup state for signalling the host.
 866          */
 867         channel->sig_event = VMBUS_EVENT_CONNECTION_ID;
 868
 869         if (vmbus_proto_version != VERSION_WS2008) {
 870                 channel->is_dedicated_interrupt =
 871                                 (offer->is_dedicated_interrupt != 0);
 872                 channel->sig_event = offer->connection_id;
 873         }
 874
 875         memcpy(&channel->offermsg, offer,
 876                sizeof(struct vmbus_channel_offer_channel));
 877         channel->monitor_grp = (u8)offer->monitorid / 32;
 878         channel->monitor_bit = (u8)offer->monitorid % 32;
 879         channel->device_id = hv_get_dev_type(channel);
 880 }
 881
 882 /*
 883  * find_primary_channel_by_offer - Get the channel object given the new offer.
 884  * This is only used in the resume path of hibernation.
 885  */
 886 static struct vmbus_channel *
 887 find_primary_channel_by_offer(const struct vmbus_channel_offer_channel *offer)
 888 {
 889         struct vmbus_channel *channel = NULL, *iter;
 890         const guid_t *inst1, *inst2;
 891
 892         /* Ignore sub-channel offers. */
 893         if (offer->offer.sub_channel_index != 0)
 894                 return NULL;
 895
 896         mutex_lock(&vmbus_connection.channel_mutex);
 897
 898         list_for_each_entry(iter, &vmbus_connection.chn_list, listentry) {
 899                 inst1 = &iter->offermsg.offer.if_instance;
 900                 inst2 = &offer->offer.if_instance;
 901
 902                 if (guid_equal(inst1, inst2)) {
 903                         channel = iter;
 904                         break;
 905                 }
 906         }
 907
 908         mutex_unlock(&vmbus_connection.channel_mutex);
 909
 910         return channel;
 911 }
 912
 913 /*
 914  * vmbus_onoffer - Handler for channel offers from vmbus in parent partition.
 915  *
 916  */
 917 static void vmbus_onoffer(struct vmbus_channel_message_header *hdr)
 918 {
 919         struct vmbus_channel_offer_channel *offer;
 920         struct vmbus_channel *oldchannel, *newchannel;
 921         size_t offer_sz;
 922
 923         offer = (struct vmbus_channel_offer_channel *)hdr;
 924
 925         trace_vmbus_onoffer(offer);
 926
 927         oldchannel = find_primary_channel_by_offer(offer);
 928
 929         if (oldchannel != NULL) {
 930                 /*
 931                  * We're resuming from hibernation: all the sub-channel and
 932                  * hv_sock channels we had before the hibernation should have
 933                  * been cleaned up, and now we must be seeing a re-offered
 934                  * primary channel that we had before the hibernation.
 935                  */
 936
 937                 /*
 938                  * { Initially: channel relid = INVALID_RELID,
 939                  *              channels[valid_relid] = NULL }
 940                  *
 941                  * CPU1                                 CPU2
 942                  *
 943                  * [vmbus_onoffer()]                    [vmbus_device_release()]
 944                  *
 945                  * LOCK channel_mutex                   LOCK channel_mutex
 946                  * STORE channel relid = valid_relid    LOAD r1 = channel relid
 947                  * MAP_RELID channel                    if (r1 != INVALID_RELID)
 948                  * UNLOCK channel_mutex                   UNMAP_RELID channel
 949                  *                                      UNLOCK channel_mutex
 950                  *
 951                  * Forbids: r1 == valid_relid &&
 952                  *              channels[valid_relid] == channel
 953                  *
 954                  * Note.  r1 can be INVALID_RELID only for an hv_sock channel.
 955                  * None of the hv_sock channels which were present before the
 956                  * suspend are re-offered upon the resume.  See the WARN_ON()
 957                  * in hv_process_channel_removal().
 958                  */
 959                 mutex_lock(&vmbus_connection.channel_mutex);
 960
 961                 atomic_dec(&vmbus_connection.offer_in_progress);
 962
 963                 WARN_ON(oldchannel->offermsg.child_relid != INVALID_RELID);
 964                 /* Fix up the relid. */
 965                 oldchannel->offermsg.child_relid = offer->child_relid;
 966
 967                 offer_sz = sizeof(*offer);
 968                 if (memcmp(offer, &oldchannel->offermsg, offer_sz) != 0) {
 969                         /*
 970                          * This is not an error, since the host can also change
 971                          * the other field(s) of the offer, e.g. on WS RS5
 972                          * (Build 17763), the offer->connection_id of the
 973                          * Mellanox VF vmbus device can change when the host
 974                          * reoffers the device upon resume.
 975                          */
 976                         pr_debug("vmbus offer changed: relid=%d\n",
 977                                  offer->child_relid);
 978
 979                         print_hex_dump_debug("Old vmbus offer: ",
 980                                              DUMP_PREFIX_OFFSET, 16, 4,
 981                                              &oldchannel->offermsg, offer_sz,
 982                                              false);
 983                         print_hex_dump_debug("New vmbus offer: ",
 984                                              DUMP_PREFIX_OFFSET, 16, 4,
 985                                              offer, offer_sz, false);
 986
 987                         /* Fix up the old channel. */
 988                         vmbus_setup_channel_state(oldchannel, offer);
 989                 }
 990
 991                 /* Add the channel back to the array of channels. */
 992                 vmbus_channel_map_relid(oldchannel);
 993                 check_ready_for_resume_event();
 994
 995                 mutex_unlock(&vmbus_connection.channel_mutex);
 996                 return;
 997         }
 998
 999         /* Allocate the channel object and save this offer. */
1000         newchannel = alloc_channel();
1001         if (!newchannel) {
1002                 vmbus_release_relid(offer->child_relid);
1003                 atomic_dec(&vmbus_connection.offer_in_progress);
1004                 pr_err("Unable to allocate channel object\n");
1005                 return;
1006         }
1007
1008         vmbus_setup_channel_state(newchannel, offer);
1009
1010         vmbus_process_offer(newchannel);
1011 }
1012
1013 static void check_ready_for_suspend_event(void)
1014 {
1015         /*
1016          * If all the sub-channels or hv_sock channels have been cleaned up,
1017          * then it's safe to suspend.
1018          */
1019         if (atomic_dec_and_test(&vmbus_connection.nr_chan_close_on_suspend))
1020                 complete(&vmbus_connection.ready_for_suspend_event);
1021 }
1022
1023 /*
1024  * vmbus_onoffer_rescind - Rescind offer handler.
1025  *
1026  * We queue a work item to process this offer synchronously
1027  */
1028 static void vmbus_onoffer_rescind(struct vmbus_channel_message_header *hdr)
1029 {
1030         struct vmbus_channel_rescind_offer *rescind;
1031         struct vmbus_channel *channel;
1032         struct device *dev;
1033         bool clean_up_chan_for_suspend;
1034
1035         rescind = (struct vmbus_channel_rescind_offer *)hdr;
1036
1037         trace_vmbus_onoffer_rescind(rescind);
1038
1039         /*
1040          * The offer msg and the corresponding rescind msg
1041          * from the host are guranteed to be ordered -
1042          * offer comes in first and then the rescind.
1043          * Since we process these events in work elements,
1044          * and with preemption, we may end up processing
1045          * the events out of order.  We rely on the synchronization
1046          * provided by offer_in_progress and by channel_mutex for
1047          * ordering these events:
1048          *
1049          * { Initially: offer_in_progress = 1 }
1050          *
1051          * CPU1                         CPU2
1052          *
1053          * [vmbus_onoffer()]            [vmbus_onoffer_rescind()]
1054          *
1055          * LOCK channel_mutex           WAIT_ON offer_in_progress == 0
1056          * DECREMENT offer_in_progress  LOCK channel_mutex
1057          * STORE channels[]             LOAD channels[]
1058          * UNLOCK channel_mutex         UNLOCK channel_mutex
1059          *
1060          * Forbids: CPU2's LOAD from *not* seeing CPU1's STORE
1061          */
1062
1063         while (atomic_read(&vmbus_connection.offer_in_progress) != 0) {
1064                 /*
1065                  * We wait here until any channel offer is currently
1066                  * being processed.
1067                  */
1068                 msleep(1);
1069         }
1070
1071         mutex_lock(&vmbus_connection.channel_mutex);
1072         channel = relid2channel(rescind->child_relid);
1073         mutex_unlock(&vmbus_connection.channel_mutex);
1074
1075         if (channel == NULL) {
1076                 /*
1077                  * We failed in processing the offer message;
1078                  * we would have cleaned up the relid in that
1079                  * failure path.
1080                  */
1081                 return;
1082         }
1083
1084         clean_up_chan_for_suspend = is_hvsock_channel(channel) ||
1085                                     is_sub_channel(channel);
1086         /*
1087          * Before setting channel->rescind in vmbus_rescind_cleanup(), we
1088          * should make sure the channel callback is not running any more.
1089          */
1090         vmbus_reset_channel_cb(channel);
1091
1092         /*
1093          * Now wait for offer handling to complete.
1094          */
1095         vmbus_rescind_cleanup(channel);
1096         while (READ_ONCE(channel->probe_done) == false) {
1097                 /*
1098                  * We wait here until any channel offer is currently
1099                  * being processed.
1100                  */
1101                 msleep(1);
1102         }
1103
1104         /*
1105          * At this point, the rescind handling can proceed safely.
1106          */
1107
1108         if (channel->device_obj) {
1109                 if (channel->chn_rescind_callback) {
1110                         channel->chn_rescind_callback(channel);
1111
1112                         if (clean_up_chan_for_suspend)
1113                                 check_ready_for_suspend_event();
1114
1115                         return;
1116                 }
1117                 /*
1118                  * We will have to unregister this device from the
1119                  * driver core.
1120                  */
1121                 dev = get_device(&channel->device_obj->device);
1122                 if (dev) {
1123                         vmbus_device_unregister(channel->device_obj);
1124                         put_device(dev);
1125                 }
1126         } else if (channel->primary_channel != NULL) {
1127                 /*
1128                  * Sub-channel is being rescinded. Following is the channel
1129                  * close sequence when initiated from the driveri (refer to
1130                  * vmbus_close() for details):
1131                  * 1. Close all sub-channels first
1132                  * 2. Then close the primary channel.
1133                  */
1134                 mutex_lock(&vmbus_connection.channel_mutex);
1135                 if (channel->state == CHANNEL_OPEN_STATE) {
1136                         /*
1137                          * The channel is currently not open;
1138                          * it is safe for us to cleanup the channel.
1139                          */
1140                         hv_process_channel_removal(channel);
1141                 } else {
1142                         complete(&channel->rescind_event);
1143                 }
1144                 mutex_unlock(&vmbus_connection.channel_mutex);
1145         }
1146
1147         /* The "channel" may have been freed. Do not access it any longer. */
1148
1149         if (clean_up_chan_for_suspend)
1150                 check_ready_for_suspend_event();
1151 }
1152
1153 void vmbus_hvsock_device_unregister(struct vmbus_channel *channel)
1154 {
1155         BUG_ON(!is_hvsock_channel(channel));
1156
1157         /* We always get a rescind msg when a connection is closed. */
1158         while (!READ_ONCE(channel->probe_done) || !READ_ONCE(channel->rescind))
1159                 msleep(1);
1160
1161         vmbus_device_unregister(channel->device_obj);
1162 }
1163 EXPORT_SYMBOL_GPL(vmbus_hvsock_device_unregister);
1164
1165
1166 /*
1167  * vmbus_onoffers_delivered -
1168  * This is invoked when all offers have been delivered.
1169  *
1170  * Nothing to do here.
1171  */
1172 static void vmbus_onoffers_delivered(
1173                         struct vmbus_channel_message_header *hdr)
1174 {
1175 }
1176
1177 /*
1178  * vmbus_onopen_result - Open result handler.
1179  *
1180  * This is invoked when we received a response to our channel open request.
1181  * Find the matching request, copy the response and signal the requesting
1182  * thread.
1183  */
1184 static void vmbus_onopen_result(struct vmbus_channel_message_header *hdr)
1185 {
1186         struct vmbus_channel_open_result *result;
1187         struct vmbus_channel_msginfo *msginfo;
1188         struct vmbus_channel_message_header *requestheader;
1189         struct vmbus_channel_open_channel *openmsg;
1190         unsigned long flags;
1191
1192         result = (struct vmbus_channel_open_result *)hdr;
1193
1194         trace_vmbus_onopen_result(result);
1195
1196         /*
1197          * Find the open msg, copy the result and signal/unblock the wait event
1198          */
1199         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1200
1201         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1202                                 msglistentry) {
1203                 requestheader =
1204                         (struct vmbus_channel_message_header *)msginfo->msg;
1205
1206                 if (requestheader->msgtype == CHANNELMSG_OPENCHANNEL) {
1207                         openmsg =
1208                         (struct vmbus_channel_open_channel *)msginfo->msg;
1209                         if (openmsg->child_relid == result->child_relid &&
1210                             openmsg->openid == result->openid) {
1211                                 memcpy(&msginfo->response.open_result,
1212                                        result,
1213                                        sizeof(
1214                                         struct vmbus_channel_open_result));
1215                                 complete(&msginfo->waitevent);
1216                                 break;
1217                         }
1218                 }
1219         }
1220         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1221 }
1222
1223 /*
1224  * vmbus_ongpadl_created - GPADL created handler.
1225  *
1226  * This is invoked when we received a response to our gpadl create request.
1227  * Find the matching request, copy the response and signal the requesting
1228  * thread.
1229  */
1230 static void vmbus_ongpadl_created(struct vmbus_channel_message_header *hdr)
1231 {
1232         struct vmbus_channel_gpadl_created *gpadlcreated;
1233         struct vmbus_channel_msginfo *msginfo;
1234         struct vmbus_channel_message_header *requestheader;
1235         struct vmbus_channel_gpadl_header *gpadlheader;
1236         unsigned long flags;
1237
1238         gpadlcreated = (struct vmbus_channel_gpadl_created *)hdr;
1239
1240         trace_vmbus_ongpadl_created(gpadlcreated);
1241
1242         /*
1243          * Find the establish msg, copy the result and signal/unblock the wait
1244          * event
1245          */
1246         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1247
1248         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1249                                 msglistentry) {
1250                 requestheader =
1251                         (struct vmbus_channel_message_header *)msginfo->msg;
1252
1253                 if (requestheader->msgtype == CHANNELMSG_GPADL_HEADER) {
1254                         gpadlheader =
1255                         (struct vmbus_channel_gpadl_header *)requestheader;
1256
1257                         if ((gpadlcreated->child_relid ==
1258                              gpadlheader->child_relid) &&
1259                             (gpadlcreated->gpadl == gpadlheader->gpadl)) {
1260                                 memcpy(&msginfo->response.gpadl_created,
1261                                        gpadlcreated,
1262                                        sizeof(
1263                                         struct vmbus_channel_gpadl_created));
1264                                 complete(&msginfo->waitevent);
1265                                 break;
1266                         }
1267                 }
1268         }
1269         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1270 }
1271
1272 /*
1273  * vmbus_ongpadl_torndown - GPADL torndown handler.
1274  *
1275  * This is invoked when we received a response to our gpadl teardown request.
1276  * Find the matching request, copy the response and signal the requesting
1277  * thread.
1278  */
1279 static void vmbus_ongpadl_torndown(
1280                         struct vmbus_channel_message_header *hdr)
1281 {
1282         struct vmbus_channel_gpadl_torndown *gpadl_torndown;
1283         struct vmbus_channel_msginfo *msginfo;
1284         struct vmbus_channel_message_header *requestheader;
1285         struct vmbus_channel_gpadl_teardown *gpadl_teardown;
1286         unsigned long flags;
1287
1288         gpadl_torndown = (struct vmbus_channel_gpadl_torndown *)hdr;
1289
1290         trace_vmbus_ongpadl_torndown(gpadl_torndown);
1291
1292         /*
1293          * Find the open msg, copy the result and signal/unblock the wait event
1294          */
1295         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1296
1297         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1298                                 msglistentry) {
1299                 requestheader =
1300                         (struct vmbus_channel_message_header *)msginfo->msg;
1301
1302                 if (requestheader->msgtype == CHANNELMSG_GPADL_TEARDOWN) {
1303                         gpadl_teardown =
1304                         (struct vmbus_channel_gpadl_teardown *)requestheader;
1305
1306                         if (gpadl_torndown->gpadl == gpadl_teardown->gpadl) {
1307                                 memcpy(&msginfo->response.gpadl_torndown,
1308                                        gpadl_torndown,
1309                                        sizeof(
1310                                         struct vmbus_channel_gpadl_torndown));
1311                                 complete(&msginfo->waitevent);
1312                                 break;
1313                         }
1314                 }
1315         }
1316         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1317 }
1318
1319 /*
1320  * vmbus_onversion_response - Version response handler
1321  *
1322  * This is invoked when we received a response to our initiate contact request.
1323  * Find the matching request, copy the response and signal the requesting
1324  * thread.
1325  */
1326 static void vmbus_onversion_response(
1327                 struct vmbus_channel_message_header *hdr)
1328 {
1329         struct vmbus_channel_msginfo *msginfo;
1330         struct vmbus_channel_message_header *requestheader;
1331         struct vmbus_channel_version_response *version_response;
1332         unsigned long flags;
1333
1334         version_response = (struct vmbus_channel_version_response *)hdr;
1335
1336         trace_vmbus_onversion_response(version_response);
1337
1338         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1339
1340         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1341                                 msglistentry) {
1342                 requestheader =
1343                         (struct vmbus_channel_message_header *)msginfo->msg;
1344
1345                 if (requestheader->msgtype ==
1346                     CHANNELMSG_INITIATE_CONTACT) {
1347                         memcpy(&msginfo->response.version_response,
1348                               version_response,
1349                               sizeof(struct vmbus_channel_version_response));
1350                         complete(&msginfo->waitevent);
1351                 }
1352         }
1353         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1354 }
1355
1356 /* Channel message dispatch table */
1357 const struct vmbus_channel_message_table_entry
1358 channel_message_table[CHANNELMSG_COUNT] = {
1359         { CHANNELMSG_INVALID,                   0, NULL, 0},
1360         { CHANNELMSG_OFFERCHANNEL,              0, vmbus_onoffer,
1361                 sizeof(struct vmbus_channel_offer_channel)},
1362         { CHANNELMSG_RESCIND_CHANNELOFFER,      0, vmbus_onoffer_rescind,
1363                 sizeof(struct vmbus_channel_rescind_offer) },
1364         { CHANNELMSG_REQUESTOFFERS,             0, NULL, 0},
1365         { CHANNELMSG_ALLOFFERS_DELIVERED,       1, vmbus_onoffers_delivered, 0},
1366         { CHANNELMSG_OPENCHANNEL,               0, NULL, 0},
1367         { CHANNELMSG_OPENCHANNEL_RESULT,        1, vmbus_onopen_result,
1368                 sizeof(struct vmbus_channel_open_result)},
1369         { CHANNELMSG_CLOSECHANNEL,              0, NULL, 0},
1370         { CHANNELMSG_GPADL_HEADER,              0, NULL, 0},
1371         { CHANNELMSG_GPADL_BODY,                0, NULL, 0},
1372         { CHANNELMSG_GPADL_CREATED,             1, vmbus_ongpadl_created,
1373                 sizeof(struct vmbus_channel_gpadl_created)},
1374         { CHANNELMSG_GPADL_TEARDOWN,            0, NULL, 0},
1375         { CHANNELMSG_GPADL_TORNDOWN,            1, vmbus_ongpadl_torndown,
1376                 sizeof(struct vmbus_channel_gpadl_torndown) },
1377         { CHANNELMSG_RELID_RELEASED,            0, NULL, 0},
1378         { CHANNELMSG_INITIATE_CONTACT,          0, NULL, 0},
1379         { CHANNELMSG_VERSION_RESPONSE,          1, vmbus_onversion_response,
1380                 sizeof(struct vmbus_channel_version_response)},
1381         { CHANNELMSG_UNLOAD,                    0, NULL, 0},
1382         { CHANNELMSG_UNLOAD_RESPONSE,           1, vmbus_unload_response, 0},
1383         { CHANNELMSG_18,                        0, NULL, 0},
1384         { CHANNELMSG_19,                        0, NULL, 0},
1385         { CHANNELMSG_20,                        0, NULL, 0},
1386         { CHANNELMSG_TL_CONNECT_REQUEST,        0, NULL, 0},
1387         { CHANNELMSG_MODIFYCHANNEL,             0, NULL, 0},
1388         { CHANNELMSG_TL_CONNECT_RESULT,         0, NULL, 0},
1389 };
1390
1391 /*
1392  * vmbus_onmessage - Handler for channel protocol messages.
1393  *
1394  * This is invoked in the vmbus worker thread context.
1395  */
1396 void vmbus_onmessage(struct vmbus_channel_message_header *hdr)
1397 {
1398         trace_vmbus_on_message(hdr);
1399
1400         /*
1401          * vmbus_on_msg_dpc() makes sure the hdr->msgtype here can not go
1402          * out of bound and the message_handler pointer can not be NULL.
1403          */
1404         channel_message_table[hdr->msgtype].message_handler(hdr);
1405 }
1406
1407 /*
1408  * vmbus_request_offers - Send a request to get all our pending offers.
1409  */
1410 int vmbus_request_offers(void)
1411 {
1412         struct vmbus_channel_message_header *msg;
1413         struct vmbus_channel_msginfo *msginfo;
1414         int ret;
1415
1416         msginfo = kmalloc(sizeof(*msginfo) +
1417                           sizeof(struct vmbus_channel_message_header),
1418                           GFP_KERNEL);
1419         if (!msginfo)
1420                 return -ENOMEM;
1421
1422         msg = (struct vmbus_channel_message_header *)msginfo->msg;
1423
1424         msg->msgtype = CHANNELMSG_REQUESTOFFERS;
1425
1426         ret = vmbus_post_msg(msg, sizeof(struct vmbus_channel_message_header),
1427                              true);
1428
1429         trace_vmbus_request_offers(ret);
1430
1431         if (ret != 0) {
1432                 pr_err("Unable to request offers - %d\n", ret);
1433
1434                 goto cleanup;
1435         }
1436
1437 cleanup:
1438         kfree(msginfo);
1439
1440         return ret;
1441 }
1442
1443 static void invoke_sc_cb(struct vmbus_channel *primary_channel)
1444 {
1445         struct list_head *cur, *tmp;
1446         struct vmbus_channel *cur_channel;
1447
1448         if (primary_channel->sc_creation_callback == NULL)
1449                 return;
1450
1451         list_for_each_safe(cur, tmp, &primary_channel->sc_list) {
1452                 cur_channel = list_entry(cur, struct vmbus_channel, sc_list);
1453
1454                 primary_channel->sc_creation_callback(cur_channel);
1455         }
1456 }
1457
1458 void vmbus_set_sc_create_callback(struct vmbus_channel *primary_channel,
1459                                 void (*sc_cr_cb)(struct vmbus_channel *new_sc))
1460 {
1461         primary_channel->sc_creation_callback = sc_cr_cb;
1462 }
1463 EXPORT_SYMBOL_GPL(vmbus_set_sc_create_callback);
1464
1465 bool vmbus_are_subchannels_present(struct vmbus_channel *primary)
1466 {
1467         bool ret;
1468
1469         ret = !list_empty(&primary->sc_list);
1470
1471         if (ret) {
1472                 /*
1473                  * Invoke the callback on sub-channel creation.
1474                  * This will present a uniform interface to the
1475                  * clients.
1476                  */
1477                 invoke_sc_cb(primary);
1478         }
1479
1480         return ret;
1481 }
1482 EXPORT_SYMBOL_GPL(vmbus_are_subchannels_present);
1483
1484 void vmbus_set_chn_rescind_callback(struct vmbus_channel *channel,
1485                 void (*chn_rescind_cb)(struct vmbus_channel *))
1486 {
1487         channel->chn_rescind_callback = chn_rescind_cb;
1488 }
1489 EXPORT_SYMBOL_GPL(vmbus_set_chn_rescind_callback);