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           .allowed_in_isolated = false,
  35         },
  36
  37         /* SCSI */
  38         { .dev_type = HV_SCSI,
  39           HV_SCSI_GUID,
  40           .perf_device = true,
  41           .allowed_in_isolated = true,
  42         },
  43
  44         /* Fibre Channel */
  45         { .dev_type = HV_FC,
  46           HV_SYNTHFC_GUID,
  47           .perf_device = true,
  48           .allowed_in_isolated = false,
  49         },
  50
  51         /* Synthetic NIC */
  52         { .dev_type = HV_NIC,
  53           HV_NIC_GUID,
  54           .perf_device = true,
  55           .allowed_in_isolated = true,
  56         },
  57
  58         /* Network Direct */
  59         { .dev_type = HV_ND,
  60           HV_ND_GUID,
  61           .perf_device = true,
  62           .allowed_in_isolated = false,
  63         },
  64
  65         /* PCIE */
  66         { .dev_type = HV_PCIE,
  67           HV_PCIE_GUID,
  68           .perf_device = false,
  69           .allowed_in_isolated = false,
  70         },
  71
  72         /* Synthetic Frame Buffer */
  73         { .dev_type = HV_FB,
  74           HV_SYNTHVID_GUID,
  75           .perf_device = false,
  76           .allowed_in_isolated = false,
  77         },
  78
  79         /* Synthetic Keyboard */
  80         { .dev_type = HV_KBD,
  81           HV_KBD_GUID,
  82           .perf_device = false,
  83           .allowed_in_isolated = false,
  84         },
  85
  86         /* Synthetic MOUSE */
  87         { .dev_type = HV_MOUSE,
  88           HV_MOUSE_GUID,
  89           .perf_device = false,
  90           .allowed_in_isolated = false,
  91         },
  92
  93         /* KVP */
  94         { .dev_type = HV_KVP,
  95           HV_KVP_GUID,
  96           .perf_device = false,
  97           .allowed_in_isolated = false,
  98         },
  99
 100         /* Time Synch */
 101         { .dev_type = HV_TS,
 102           HV_TS_GUID,
 103           .perf_device = false,
 104           .allowed_in_isolated = true,
 105         },
 106
 107         /* Heartbeat */
 108         { .dev_type = HV_HB,
 109           HV_HEART_BEAT_GUID,
 110           .perf_device = false,
 111           .allowed_in_isolated = true,
 112         },
 113
 114         /* Shutdown */
 115         { .dev_type = HV_SHUTDOWN,
 116           HV_SHUTDOWN_GUID,
 117           .perf_device = false,
 118           .allowed_in_isolated = true,
 119         },
 120
 121         /* File copy */
 122         { .dev_type = HV_FCOPY,
 123           HV_FCOPY_GUID,
 124           .perf_device = false,
 125           .allowed_in_isolated = false,
 126         },
 127
 128         /* Backup */
 129         { .dev_type = HV_BACKUP,
 130           HV_VSS_GUID,
 131           .perf_device = false,
 132           .allowed_in_isolated = false,
 133         },
 134
 135         /* Dynamic Memory */
 136         { .dev_type = HV_DM,
 137           HV_DM_GUID,
 138           .perf_device = false,
 139           .allowed_in_isolated = false,
 140         },
 141
 142         /* Unknown GUID */
 143         { .dev_type = HV_UNKNOWN,
 144           .perf_device = false,
 145           .allowed_in_isolated = false,
 146         },
 147 };
 148
 149 static const struct {
 150         guid_t guid;
 151 } vmbus_unsupported_devs[] = {
 152         { HV_AVMA1_GUID },
 153         { HV_AVMA2_GUID },
 154         { HV_RDV_GUID   },
 155 };
 156
 157 /*
 158  * The rescinded channel may be blocked waiting for a response from the host;
 159  * take care of that.
 160  */
 161 static void vmbus_rescind_cleanup(struct vmbus_channel *channel)
 162 {
 163         struct vmbus_channel_msginfo *msginfo;
 164         unsigned long flags;
 165
 166
 167         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
 168         channel->rescind = true;
 169         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
 170                                 msglistentry) {
 171
 172                 if (msginfo->waiting_channel == channel) {
 173                         complete(&msginfo->waitevent);
 174                         break;
 175                 }
 176         }
 177         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
 178 }
 179
 180 static bool is_unsupported_vmbus_devs(const guid_t *guid)
 181 {
 182         int i;
 183
 184         for (i = 0; i < ARRAY_SIZE(vmbus_unsupported_devs); i++)
 185                 if (guid_equal(guid, &vmbus_unsupported_devs[i].guid))
 186                         return true;
 187         return false;
 188 }
 189
 190 static u16 hv_get_dev_type(const struct vmbus_channel *channel)
 191 {
 192         const guid_t *guid = &channel->offermsg.offer.if_type;
 193         u16 i;
 194
 195         if (is_hvsock_channel(channel) || is_unsupported_vmbus_devs(guid))
 196                 return HV_UNKNOWN;
 197
 198         for (i = HV_IDE; i < HV_UNKNOWN; i++) {
 199                 if (guid_equal(guid, &vmbus_devs[i].guid))
 200                         return i;
 201         }
 202         pr_info("Unknown GUID: %pUl\n", guid);
 203         return i;
 204 }
 205
 206 /**
 207  * vmbus_prep_negotiate_resp() - Create default response for Negotiate message
 208  * @icmsghdrp: Pointer to msg header structure
 209  * @buf: Raw buffer channel data
 210  * @buflen: Length of the raw buffer channel data.
 211  * @fw_version: The framework versions we can support.
 212  * @fw_vercnt: The size of @fw_version.
 213  * @srv_version: The service versions we can support.
 214  * @srv_vercnt: The size of @srv_version.
 215  * @nego_fw_version: The selected framework version.
 216  * @nego_srv_version: The selected service version.
 217  *
 218  * Note: Versions are given in decreasing order.
 219  *
 220  * Set up and fill in default negotiate response message.
 221  * Mainly used by Hyper-V drivers.
 222  */
 223 bool vmbus_prep_negotiate_resp(struct icmsg_hdr *icmsghdrp, u8 *buf,
 224                                 u32 buflen, const int *fw_version, int fw_vercnt,
 225                                 const int *srv_version, int srv_vercnt,
 226                                 int *nego_fw_version, int *nego_srv_version)
 227 {
 228         int icframe_major, icframe_minor;
 229         int icmsg_major, icmsg_minor;
 230         int fw_major, fw_minor;
 231         int srv_major, srv_minor;
 232         int i, j;
 233         bool found_match = false;
 234         struct icmsg_negotiate *negop;
 235
 236         /* Check that there's enough space for icframe_vercnt, icmsg_vercnt */
 237         if (buflen < ICMSG_HDR + offsetof(struct icmsg_negotiate, reserved)) {
 238                 pr_err_ratelimited("Invalid icmsg negotiate\n");
 239                 return false;
 240         }
 241
 242         icmsghdrp->icmsgsize = 0x10;
 243         negop = (struct icmsg_negotiate *)&buf[ICMSG_HDR];
 244
 245         icframe_major = negop->icframe_vercnt;
 246         icframe_minor = 0;
 247
 248         icmsg_major = negop->icmsg_vercnt;
 249         icmsg_minor = 0;
 250
 251         /* Validate negop packet */
 252         if (icframe_major > IC_VERSION_NEGOTIATION_MAX_VER_COUNT ||
 253             icmsg_major > IC_VERSION_NEGOTIATION_MAX_VER_COUNT ||
 254             ICMSG_NEGOTIATE_PKT_SIZE(icframe_major, icmsg_major) > buflen) {
 255                 pr_err_ratelimited("Invalid icmsg negotiate - icframe_major: %u, icmsg_major: %u\n",
 256                                    icframe_major, icmsg_major);
 257                 goto fw_error;
 258         }
 259
 260         /*
 261          * Select the framework version number we will
 262          * support.
 263          */
 264
 265         for (i = 0; i < fw_vercnt; i++) {
 266                 fw_major = (fw_version[i] >> 16);
 267                 fw_minor = (fw_version[i] & 0xFFFF);
 268
 269                 for (j = 0; j < negop->icframe_vercnt; j++) {
 270                         if ((negop->icversion_data[j].major == fw_major) &&
 271                             (negop->icversion_data[j].minor == fw_minor)) {
 272                                 icframe_major = negop->icversion_data[j].major;
 273                                 icframe_minor = negop->icversion_data[j].minor;
 274                                 found_match = true;
 275                                 break;
 276                         }
 277                 }
 278
 279                 if (found_match)
 280                         break;
 281         }
 282
 283         if (!found_match)
 284                 goto fw_error;
 285
 286         found_match = false;
 287
 288         for (i = 0; i < srv_vercnt; i++) {
 289                 srv_major = (srv_version[i] >> 16);
 290                 srv_minor = (srv_version[i] & 0xFFFF);
 291
 292                 for (j = negop->icframe_vercnt;
 293                         (j < negop->icframe_vercnt + negop->icmsg_vercnt);
 294                         j++) {
 295
 296                         if ((negop->icversion_data[j].major == srv_major) &&
 297                                 (negop->icversion_data[j].minor == srv_minor)) {
 298
 299                                 icmsg_major = negop->icversion_data[j].major;
 300                                 icmsg_minor = negop->icversion_data[j].minor;
 301                                 found_match = true;
 302                                 break;
 303                         }
 304                 }
 305
 306                 if (found_match)
 307                         break;
 308         }
 309
 310         /*
 311          * Respond with the framework and service
 312          * version numbers we can support.
 313          */
 314
 315 fw_error:
 316         if (!found_match) {
 317                 negop->icframe_vercnt = 0;
 318                 negop->icmsg_vercnt = 0;
 319         } else {
 320                 negop->icframe_vercnt = 1;
 321                 negop->icmsg_vercnt = 1;
 322         }
 323
 324         if (nego_fw_version)
 325                 *nego_fw_version = (icframe_major << 16) | icframe_minor;
 326
 327         if (nego_srv_version)
 328                 *nego_srv_version = (icmsg_major << 16) | icmsg_minor;
 329
 330         negop->icversion_data[0].major = icframe_major;
 331         negop->icversion_data[0].minor = icframe_minor;
 332         negop->icversion_data[1].major = icmsg_major;
 333         negop->icversion_data[1].minor = icmsg_minor;
 334         return found_match;
 335 }
 336 EXPORT_SYMBOL_GPL(vmbus_prep_negotiate_resp);
 337
 338 /*
 339  * alloc_channel - Allocate and initialize a vmbus channel object
 340  */
 341 static struct vmbus_channel *alloc_channel(void)
 342 {
 343         struct vmbus_channel *channel;
 344
 345         channel = kzalloc(sizeof(*channel), GFP_ATOMIC);
 346         if (!channel)
 347                 return NULL;
 348
 349         spin_lock_init(&channel->sched_lock);
 350         init_completion(&channel->rescind_event);
 351
 352         INIT_LIST_HEAD(&channel->sc_list);
 353
 354         tasklet_init(&channel->callback_event,
 355                      vmbus_on_event, (unsigned long)channel);
 356
 357         hv_ringbuffer_pre_init(channel);
 358
 359         return channel;
 360 }
 361
 362 /*
 363  * free_channel - Release the resources used by the vmbus channel object
 364  */
 365 static void free_channel(struct vmbus_channel *channel)
 366 {
 367         tasklet_kill(&channel->callback_event);
 368         vmbus_remove_channel_attr_group(channel);
 369
 370         kobject_put(&channel->kobj);
 371 }
 372
 373 void vmbus_channel_map_relid(struct vmbus_channel *channel)
 374 {
 375         if (WARN_ON(channel->offermsg.child_relid >= MAX_CHANNEL_RELIDS))
 376                 return;
 377         /*
 378          * The mapping of the channel's relid is visible from the CPUs that
 379          * execute vmbus_chan_sched() by the time that vmbus_chan_sched() will
 380          * execute:
 381          *
 382          *  (a) In the "normal (i.e., not resuming from hibernation)" path,
 383          *      the full barrier in virt_store_mb() guarantees that the store
 384          *      is propagated to all CPUs before the add_channel_work work
 385          *      is queued.  In turn, add_channel_work is queued before the
 386          *      channel's ring buffer is allocated/initialized and the
 387          *      OPENCHANNEL message for the channel is sent in vmbus_open().
 388          *      Hyper-V won't start sending the interrupts for the channel
 389          *      before the OPENCHANNEL message is acked.  The memory barrier
 390          *      in vmbus_chan_sched() -> sync_test_and_clear_bit() ensures
 391          *      that vmbus_chan_sched() must find the channel's relid in
 392          *      recv_int_page before retrieving the channel pointer from the
 393          *      array of channels.
 394          *
 395          *  (b) In the "resuming from hibernation" path, the virt_store_mb()
 396          *      guarantees that the store is propagated to all CPUs before
 397          *      the VMBus connection is marked as ready for the resume event
 398          *      (cf. check_ready_for_resume_event()).  The interrupt handler
 399          *      of the VMBus driver and vmbus_chan_sched() can not run before
 400          *      vmbus_bus_resume() has completed execution (cf. resume_noirq).
 401          */
 402         virt_store_mb(
 403                 vmbus_connection.channels[channel->offermsg.child_relid],
 404                 channel);
 405 }
 406
 407 void vmbus_channel_unmap_relid(struct vmbus_channel *channel)
 408 {
 409         if (WARN_ON(channel->offermsg.child_relid >= MAX_CHANNEL_RELIDS))
 410                 return;
 411         WRITE_ONCE(
 412                 vmbus_connection.channels[channel->offermsg.child_relid],
 413                 NULL);
 414 }
 415
 416 static void vmbus_release_relid(u32 relid)
 417 {
 418         struct vmbus_channel_relid_released msg;
 419         int ret;
 420
 421         memset(&msg, 0, sizeof(struct vmbus_channel_relid_released));
 422         msg.child_relid = relid;
 423         msg.header.msgtype = CHANNELMSG_RELID_RELEASED;
 424         ret = vmbus_post_msg(&msg, sizeof(struct vmbus_channel_relid_released),
 425                              true);
 426
 427         trace_vmbus_release_relid(&msg, ret);
 428 }
 429
 430 void hv_process_channel_removal(struct vmbus_channel *channel)
 431 {
 432         lockdep_assert_held(&vmbus_connection.channel_mutex);
 433         BUG_ON(!channel->rescind);
 434
 435         /*
 436          * hv_process_channel_removal() could find INVALID_RELID only for
 437          * hv_sock channels.  See the inline comments in vmbus_onoffer().
 438          */
 439         WARN_ON(channel->offermsg.child_relid == INVALID_RELID &&
 440                 !is_hvsock_channel(channel));
 441
 442         /*
 443          * Upon suspend, an in-use hv_sock channel is removed from the array of
 444          * channels and the relid is invalidated.  After hibernation, when the
 445          * user-space appplication destroys the channel, it's unnecessary and
 446          * unsafe to remove the channel from the array of channels.  See also
 447          * the inline comments before the call of vmbus_release_relid() below.
 448          */
 449         if (channel->offermsg.child_relid != INVALID_RELID)
 450                 vmbus_channel_unmap_relid(channel);
 451
 452         if (channel->primary_channel == NULL)
 453                 list_del(&channel->listentry);
 454         else
 455                 list_del(&channel->sc_list);
 456
 457         /*
 458          * If this is a "perf" channel, updates the hv_numa_map[] masks so that
 459          * init_vp_index() can (re-)use the CPU.
 460          */
 461         if (hv_is_perf_channel(channel))
 462                 hv_clear_alloced_cpu(channel->target_cpu);
 463
 464         /*
 465          * Upon suspend, an in-use hv_sock channel is marked as "rescinded" and
 466          * the relid is invalidated; after hibernation, when the user-space app
 467          * destroys the channel, the relid is INVALID_RELID, and in this case
 468          * it's unnecessary and unsafe to release the old relid, since the same
 469          * relid can refer to a completely different channel now.
 470          */
 471         if (channel->offermsg.child_relid != INVALID_RELID)
 472                 vmbus_release_relid(channel->offermsg.child_relid);
 473
 474         free_channel(channel);
 475 }
 476
 477 void vmbus_free_channels(void)
 478 {
 479         struct vmbus_channel *channel, *tmp;
 480
 481         list_for_each_entry_safe(channel, tmp, &vmbus_connection.chn_list,
 482                 listentry) {
 483                 /* hv_process_channel_removal() needs this */
 484                 channel->rescind = true;
 485
 486                 vmbus_device_unregister(channel->device_obj);
 487         }
 488 }
 489
 490 /* Note: the function can run concurrently for primary/sub channels. */
 491 static void vmbus_add_channel_work(struct work_struct *work)
 492 {
 493         struct vmbus_channel *newchannel =
 494                 container_of(work, struct vmbus_channel, add_channel_work);
 495         struct vmbus_channel *primary_channel = newchannel->primary_channel;
 496         int ret;
 497
 498         /*
 499          * This state is used to indicate a successful open
 500          * so that when we do close the channel normally, we
 501          * can cleanup properly.
 502          */
 503         newchannel->state = CHANNEL_OPEN_STATE;
 504
 505         if (primary_channel != NULL) {
 506                 /* newchannel is a sub-channel. */
 507                 struct hv_device *dev = primary_channel->device_obj;
 508
 509                 if (vmbus_add_channel_kobj(dev, newchannel))
 510                         goto err_deq_chan;
 511
 512                 if (primary_channel->sc_creation_callback != NULL)
 513                         primary_channel->sc_creation_callback(newchannel);
 514
 515                 newchannel->probe_done = true;
 516                 return;
 517         }
 518
 519         /*
 520          * Start the process of binding the primary channel to the driver
 521          */
 522         newchannel->device_obj = vmbus_device_create(
 523                 &newchannel->offermsg.offer.if_type,
 524                 &newchannel->offermsg.offer.if_instance,
 525                 newchannel);
 526         if (!newchannel->device_obj)
 527                 goto err_deq_chan;
 528
 529         newchannel->device_obj->device_id = newchannel->device_id;
 530         /*
 531          * Add the new device to the bus. This will kick off device-driver
 532          * binding which eventually invokes the device driver's AddDevice()
 533          * method.
 534          */
 535         ret = vmbus_device_register(newchannel->device_obj);
 536
 537         if (ret != 0) {
 538                 pr_err("unable to add child device object (relid %d)\n",
 539                         newchannel->offermsg.child_relid);
 540                 kfree(newchannel->device_obj);
 541                 goto err_deq_chan;
 542         }
 543
 544         newchannel->probe_done = true;
 545         return;
 546
 547 err_deq_chan:
 548         mutex_lock(&vmbus_connection.channel_mutex);
 549
 550         /*
 551          * We need to set the flag, otherwise
 552          * vmbus_onoffer_rescind() can be blocked.
 553          */
 554         newchannel->probe_done = true;
 555
 556         if (primary_channel == NULL)
 557                 list_del(&newchannel->listentry);
 558         else
 559                 list_del(&newchannel->sc_list);
 560
 561         /* vmbus_process_offer() has mapped the channel. */
 562         vmbus_channel_unmap_relid(newchannel);
 563
 564         mutex_unlock(&vmbus_connection.channel_mutex);
 565
 566         vmbus_release_relid(newchannel->offermsg.child_relid);
 567
 568         free_channel(newchannel);
 569 }
 570
 571 /*
 572  * vmbus_process_offer - Process the offer by creating a channel/device
 573  * associated with this offer
 574  */
 575 static void vmbus_process_offer(struct vmbus_channel *newchannel)
 576 {
 577         struct vmbus_channel *channel;
 578         struct workqueue_struct *wq;
 579         bool fnew = true;
 580
 581         /*
 582          * Synchronize vmbus_process_offer() and CPU hotplugging:
 583          *
 584          * CPU1                         CPU2
 585          *
 586          * [vmbus_process_offer()]      [Hot removal of the CPU]
 587          *
 588          * CPU_READ_LOCK                CPUS_WRITE_LOCK
 589          * LOAD cpu_online_mask         SEARCH chn_list
 590          * STORE target_cpu             LOAD target_cpu
 591          * INSERT chn_list              STORE cpu_online_mask
 592          * CPUS_READ_UNLOCK             CPUS_WRITE_UNLOCK
 593          *
 594          * Forbids: CPU1's LOAD from *not* seing CPU2's STORE &&
 595          *              CPU2's SEARCH from *not* seeing CPU1's INSERT
 596          *
 597          * Forbids: CPU2's SEARCH from seeing CPU1's INSERT &&
 598          *              CPU2's LOAD from *not* seing CPU1's STORE
 599          */
 600         cpus_read_lock();
 601
 602         /*
 603          * Serializes the modifications of the chn_list list as well as
 604          * the accesses to next_numa_node_id in init_vp_index().
 605          */
 606         mutex_lock(&vmbus_connection.channel_mutex);
 607
 608         list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
 609                 if (guid_equal(&channel->offermsg.offer.if_type,
 610                                &newchannel->offermsg.offer.if_type) &&
 611                     guid_equal(&channel->offermsg.offer.if_instance,
 612                                &newchannel->offermsg.offer.if_instance)) {
 613                         fnew = false;
 614                         newchannel->primary_channel = channel;
 615                         break;
 616                 }
 617         }
 618
 619         init_vp_index(newchannel);
 620
 621         /* Remember the channels that should be cleaned up upon suspend. */
 622         if (is_hvsock_channel(newchannel) || is_sub_channel(newchannel))
 623                 atomic_inc(&vmbus_connection.nr_chan_close_on_suspend);
 624
 625         /*
 626          * Now that we have acquired the channel_mutex,
 627          * we can release the potentially racing rescind thread.
 628          */
 629         atomic_dec(&vmbus_connection.offer_in_progress);
 630
 631         if (fnew) {
 632                 list_add_tail(&newchannel->listentry,
 633                               &vmbus_connection.chn_list);
 634         } else {
 635                 /*
 636                  * Check to see if this is a valid sub-channel.
 637                  */
 638                 if (newchannel->offermsg.offer.sub_channel_index == 0) {
 639                         mutex_unlock(&vmbus_connection.channel_mutex);
 640                         cpus_read_unlock();
 641                         /*
 642                          * Don't call free_channel(), because newchannel->kobj
 643                          * is not initialized yet.
 644                          */
 645                         kfree(newchannel);
 646                         WARN_ON_ONCE(1);
 647                         return;
 648                 }
 649                 /*
 650                  * Process the sub-channel.
 651                  */
 652                 list_add_tail(&newchannel->sc_list, &channel->sc_list);
 653         }
 654
 655         vmbus_channel_map_relid(newchannel);
 656
 657         mutex_unlock(&vmbus_connection.channel_mutex);
 658         cpus_read_unlock();
 659
 660         /*
 661          * vmbus_process_offer() mustn't call channel->sc_creation_callback()
 662          * directly for sub-channels, because sc_creation_callback() ->
 663          * vmbus_open() may never get the host's response to the
 664          * OPEN_CHANNEL message (the host may rescind a channel at any time,
 665          * e.g. in the case of hot removing a NIC), and vmbus_onoffer_rescind()
 666          * may not wake up the vmbus_open() as it's blocked due to a non-zero
 667          * vmbus_connection.offer_in_progress, and finally we have a deadlock.
 668          *
 669          * The above is also true for primary channels, if the related device
 670          * drivers use sync probing mode by default.
 671          *
 672          * And, usually the handling of primary channels and sub-channels can
 673          * depend on each other, so we should offload them to different
 674          * workqueues to avoid possible deadlock, e.g. in sync-probing mode,
 675          * NIC1's netvsc_subchan_work() can race with NIC2's netvsc_probe() ->
 676          * rtnl_lock(), and causes deadlock: the former gets the rtnl_lock
 677          * and waits for all the sub-channels to appear, but the latter
 678          * can't get the rtnl_lock and this blocks the handling of
 679          * sub-channels.
 680          */
 681         INIT_WORK(&newchannel->add_channel_work, vmbus_add_channel_work);
 682         wq = fnew ? vmbus_connection.handle_primary_chan_wq :
 683                     vmbus_connection.handle_sub_chan_wq;
 684         queue_work(wq, &newchannel->add_channel_work);
 685 }
 686
 687 /*
 688  * Check if CPUs used by other channels of the same device.
 689  * It should only be called by init_vp_index().
 690  */
 691 static bool hv_cpuself_used(u32 cpu, struct vmbus_channel *chn)
 692 {
 693         struct vmbus_channel *primary = chn->primary_channel;
 694         struct vmbus_channel *sc;
 695
 696         lockdep_assert_held(&vmbus_connection.channel_mutex);
 697
 698         if (!primary)
 699                 return false;
 700
 701         if (primary->target_cpu == cpu)
 702                 return true;
 703
 704         list_for_each_entry(sc, &primary->sc_list, sc_list)
 705                 if (sc != chn && sc->target_cpu == cpu)
 706                         return true;
 707
 708         return false;
 709 }
 710
 711 /*
 712  * We use this state to statically distribute the channel interrupt load.
 713  */
 714 static int next_numa_node_id;
 715
 716 /*
 717  * Starting with Win8, we can statically distribute the incoming
 718  * channel interrupt load by binding a channel to VCPU.
 719  *
 720  * For pre-win8 hosts or non-performance critical channels we assign the
 721  * VMBUS_CONNECT_CPU.
 722  *
 723  * Starting with win8, performance critical channels will be distributed
 724  * evenly among all the available NUMA nodes.  Once the node is assigned,
 725  * we will assign the CPU based on a simple round robin scheme.
 726  */
 727 static void init_vp_index(struct vmbus_channel *channel)
 728 {
 729         bool perf_chn = hv_is_perf_channel(channel);
 730         u32 i, ncpu = num_online_cpus();
 731         cpumask_var_t available_mask;
 732         struct cpumask *alloced_mask;
 733         u32 target_cpu;
 734         int numa_node;
 735
 736         if ((vmbus_proto_version == VERSION_WS2008) ||
 737             (vmbus_proto_version == VERSION_WIN7) || (!perf_chn) ||
 738             !alloc_cpumask_var(&available_mask, GFP_KERNEL)) {
 739                 /*
 740                  * Prior to win8, all channel interrupts are
 741                  * delivered on VMBUS_CONNECT_CPU.
 742                  * Also if the channel is not a performance critical
 743                  * channel, bind it to VMBUS_CONNECT_CPU.
 744                  * In case alloc_cpumask_var() fails, bind it to
 745                  * VMBUS_CONNECT_CPU.
 746                  */
 747                 channel->target_cpu = VMBUS_CONNECT_CPU;
 748                 if (perf_chn)
 749                         hv_set_alloced_cpu(VMBUS_CONNECT_CPU);
 750                 return;
 751         }
 752
 753         for (i = 1; i <= ncpu + 1; i++) {
 754                 while (true) {
 755                         numa_node = next_numa_node_id++;
 756                         if (numa_node == nr_node_ids) {
 757                                 next_numa_node_id = 0;
 758                                 continue;
 759                         }
 760                         if (cpumask_empty(cpumask_of_node(numa_node)))
 761                                 continue;
 762                         break;
 763                 }
 764                 alloced_mask = &hv_context.hv_numa_map[numa_node];
 765
 766                 if (cpumask_weight(alloced_mask) ==
 767                     cpumask_weight(cpumask_of_node(numa_node))) {
 768                         /*
 769                          * We have cycled through all the CPUs in the node;
 770                          * reset the alloced map.
 771                          */
 772                         cpumask_clear(alloced_mask);
 773                 }
 774
 775                 cpumask_xor(available_mask, alloced_mask,
 776                             cpumask_of_node(numa_node));
 777
 778                 target_cpu = cpumask_first(available_mask);
 779                 cpumask_set_cpu(target_cpu, alloced_mask);
 780
 781                 if (channel->offermsg.offer.sub_channel_index >= ncpu ||
 782                     i > ncpu || !hv_cpuself_used(target_cpu, channel))
 783                         break;
 784         }
 785
 786         channel->target_cpu = target_cpu;
 787
 788         free_cpumask_var(available_mask);
 789 }
 790
 791 #define UNLOAD_DELAY_UNIT_MS    10              /* 10 milliseconds */
 792 #define UNLOAD_WAIT_MS          (100*1000)      /* 100 seconds */
 793 #define UNLOAD_WAIT_LOOPS       (UNLOAD_WAIT_MS/UNLOAD_DELAY_UNIT_MS)
 794 #define UNLOAD_MSG_MS           (5*1000)        /* Every 5 seconds */
 795 #define UNLOAD_MSG_LOOPS        (UNLOAD_MSG_MS/UNLOAD_DELAY_UNIT_MS)
 796
 797 static void vmbus_wait_for_unload(void)
 798 {
 799         int cpu;
 800         void *page_addr;
 801         struct hv_message *msg;
 802         struct vmbus_channel_message_header *hdr;
 803         u32 message_type, i;
 804
 805         /*
 806          * CHANNELMSG_UNLOAD_RESPONSE is always delivered to the CPU which was
 807          * used for initial contact or to CPU0 depending on host version. When
 808          * we're crashing on a different CPU let's hope that IRQ handler on
 809          * the cpu which receives CHANNELMSG_UNLOAD_RESPONSE is still
 810          * functional and vmbus_unload_response() will complete
 811          * vmbus_connection.unload_event. If not, the last thing we can do is
 812          * read message pages for all CPUs directly.
 813          *
 814          * Wait up to 100 seconds since an Azure host must writeback any dirty
 815          * data in its disk cache before the VMbus UNLOAD request will
 816          * complete. This flushing has been empirically observed to take up
 817          * to 50 seconds in cases with a lot of dirty data, so allow additional
 818          * leeway and for inaccuracies in mdelay(). But eventually time out so
 819          * that the panic path can't get hung forever in case the response
 820          * message isn't seen.
 821          */
 822         for (i = 1; i <= UNLOAD_WAIT_LOOPS; i++) {
 823                 if (completion_done(&vmbus_connection.unload_event))
 824                         goto completed;
 825
 826                 for_each_online_cpu(cpu) {
 827                         struct hv_per_cpu_context *hv_cpu
 828                                 = per_cpu_ptr(hv_context.cpu_context, cpu);
 829
 830                         page_addr = hv_cpu->synic_message_page;
 831                         msg = (struct hv_message *)page_addr
 832                                 + VMBUS_MESSAGE_SINT;
 833
 834                         message_type = READ_ONCE(msg->header.message_type);
 835                         if (message_type == HVMSG_NONE)
 836                                 continue;
 837
 838                         hdr = (struct vmbus_channel_message_header *)
 839                                 msg->u.payload;
 840
 841                         if (hdr->msgtype == CHANNELMSG_UNLOAD_RESPONSE)
 842                                 complete(&vmbus_connection.unload_event);
 843
 844                         vmbus_signal_eom(msg, message_type);
 845                 }
 846
 847                 /*
 848                  * Give a notice periodically so someone watching the
 849                  * serial output won't think it is completely hung.
 850                  */
 851                 if (!(i % UNLOAD_MSG_LOOPS))
 852                         pr_notice("Waiting for VMBus UNLOAD to complete\n");
 853
 854                 mdelay(UNLOAD_DELAY_UNIT_MS);
 855         }
 856         pr_err("Continuing even though VMBus UNLOAD did not complete\n");
 857
 858 completed:
 859         /*
 860          * We're crashing and already got the UNLOAD_RESPONSE, cleanup all
 861          * maybe-pending messages on all CPUs to be able to receive new
 862          * messages after we reconnect.
 863          */
 864         for_each_online_cpu(cpu) {
 865                 struct hv_per_cpu_context *hv_cpu
 866                         = per_cpu_ptr(hv_context.cpu_context, cpu);
 867
 868                 page_addr = hv_cpu->synic_message_page;
 869                 msg = (struct hv_message *)page_addr + VMBUS_MESSAGE_SINT;
 870                 msg->header.message_type = HVMSG_NONE;
 871         }
 872 }
 873
 874 /*
 875  * vmbus_unload_response - Handler for the unload response.
 876  */
 877 static void vmbus_unload_response(struct vmbus_channel_message_header *hdr)
 878 {
 879         /*
 880          * This is a global event; just wakeup the waiting thread.
 881          * Once we successfully unload, we can cleanup the monitor state.
 882          *
 883          * NB.  A malicious or compromised Hyper-V could send a spurious
 884          * message of type CHANNELMSG_UNLOAD_RESPONSE, and trigger a call
 885          * of the complete() below.  Make sure that unload_event has been
 886          * initialized by the time this complete() is executed.
 887          */
 888         complete(&vmbus_connection.unload_event);
 889 }
 890
 891 void vmbus_initiate_unload(bool crash)
 892 {
 893         struct vmbus_channel_message_header hdr;
 894
 895         if (xchg(&vmbus_connection.conn_state, DISCONNECTED) == DISCONNECTED)
 896                 return;
 897
 898         /* Pre-Win2012R2 hosts don't support reconnect */
 899         if (vmbus_proto_version < VERSION_WIN8_1)
 900                 return;
 901
 902         reinit_completion(&vmbus_connection.unload_event);
 903         memset(&hdr, 0, sizeof(struct vmbus_channel_message_header));
 904         hdr.msgtype = CHANNELMSG_UNLOAD;
 905         vmbus_post_msg(&hdr, sizeof(struct vmbus_channel_message_header),
 906                        !crash);
 907
 908         /*
 909          * vmbus_initiate_unload() is also called on crash and the crash can be
 910          * happening in an interrupt context, where scheduling is impossible.
 911          */
 912         if (!crash)
 913                 wait_for_completion(&vmbus_connection.unload_event);
 914         else
 915                 vmbus_wait_for_unload();
 916 }
 917
 918 static void check_ready_for_resume_event(void)
 919 {
 920         /*
 921          * If all the old primary channels have been fixed up, then it's safe
 922          * to resume.
 923          */
 924         if (atomic_dec_and_test(&vmbus_connection.nr_chan_fixup_on_resume))
 925                 complete(&vmbus_connection.ready_for_resume_event);
 926 }
 927
 928 static void vmbus_setup_channel_state(struct vmbus_channel *channel,
 929                                       struct vmbus_channel_offer_channel *offer)
 930 {
 931         /*
 932          * Setup state for signalling the host.
 933          */
 934         channel->sig_event = VMBUS_EVENT_CONNECTION_ID;
 935
 936         if (vmbus_proto_version != VERSION_WS2008) {
 937                 channel->is_dedicated_interrupt =
 938                                 (offer->is_dedicated_interrupt != 0);
 939                 channel->sig_event = offer->connection_id;
 940         }
 941
 942         memcpy(&channel->offermsg, offer,
 943                sizeof(struct vmbus_channel_offer_channel));
 944         channel->monitor_grp = (u8)offer->monitorid / 32;
 945         channel->monitor_bit = (u8)offer->monitorid % 32;
 946         channel->device_id = hv_get_dev_type(channel);
 947 }
 948
 949 /*
 950  * find_primary_channel_by_offer - Get the channel object given the new offer.
 951  * This is only used in the resume path of hibernation.
 952  */
 953 static struct vmbus_channel *
 954 find_primary_channel_by_offer(const struct vmbus_channel_offer_channel *offer)
 955 {
 956         struct vmbus_channel *channel = NULL, *iter;
 957         const guid_t *inst1, *inst2;
 958
 959         /* Ignore sub-channel offers. */
 960         if (offer->offer.sub_channel_index != 0)
 961                 return NULL;
 962
 963         mutex_lock(&vmbus_connection.channel_mutex);
 964
 965         list_for_each_entry(iter, &vmbus_connection.chn_list, listentry) {
 966                 inst1 = &iter->offermsg.offer.if_instance;
 967                 inst2 = &offer->offer.if_instance;
 968
 969                 if (guid_equal(inst1, inst2)) {
 970                         channel = iter;
 971                         break;
 972                 }
 973         }
 974
 975         mutex_unlock(&vmbus_connection.channel_mutex);
 976
 977         return channel;
 978 }
 979
 980 static bool vmbus_is_valid_device(const guid_t *guid)
 981 {
 982         u16 i;
 983
 984         if (!hv_is_isolation_supported())
 985                 return true;
 986
 987         for (i = 0; i < ARRAY_SIZE(vmbus_devs); i++) {
 988                 if (guid_equal(guid, &vmbus_devs[i].guid))
 989                         return vmbus_devs[i].allowed_in_isolated;
 990         }
 991         return false;
 992 }
 993
 994 /*
 995  * vmbus_onoffer - Handler for channel offers from vmbus in parent partition.
 996  *
 997  */
 998 static void vmbus_onoffer(struct vmbus_channel_message_header *hdr)
 999 {
1000         struct vmbus_channel_offer_channel *offer;
1001         struct vmbus_channel *oldchannel, *newchannel;
1002         size_t offer_sz;
1003
1004         offer = (struct vmbus_channel_offer_channel *)hdr;
1005
1006         trace_vmbus_onoffer(offer);
1007
1008         if (!vmbus_is_valid_device(&offer->offer.if_type)) {
1009                 pr_err_ratelimited("Invalid offer %d from the host supporting isolation\n",
1010                                    offer->child_relid);
1011                 atomic_dec(&vmbus_connection.offer_in_progress);
1012                 return;
1013         }
1014
1015         oldchannel = find_primary_channel_by_offer(offer);
1016
1017         if (oldchannel != NULL) {
1018                 /*
1019                  * We're resuming from hibernation: all the sub-channel and
1020                  * hv_sock channels we had before the hibernation should have
1021                  * been cleaned up, and now we must be seeing a re-offered
1022                  * primary channel that we had before the hibernation.
1023                  */
1024
1025                 /*
1026                  * { Initially: channel relid = INVALID_RELID,
1027                  *              channels[valid_relid] = NULL }
1028                  *
1029                  * CPU1                                 CPU2
1030                  *
1031                  * [vmbus_onoffer()]                    [vmbus_device_release()]
1032                  *
1033                  * LOCK channel_mutex                   LOCK channel_mutex
1034                  * STORE channel relid = valid_relid    LOAD r1 = channel relid
1035                  * MAP_RELID channel                    if (r1 != INVALID_RELID)
1036                  * UNLOCK channel_mutex                   UNMAP_RELID channel
1037                  *                                      UNLOCK channel_mutex
1038                  *
1039                  * Forbids: r1 == valid_relid &&
1040                  *              channels[valid_relid] == channel
1041                  *
1042                  * Note.  r1 can be INVALID_RELID only for an hv_sock channel.
1043                  * None of the hv_sock channels which were present before the
1044                  * suspend are re-offered upon the resume.  See the WARN_ON()
1045                  * in hv_process_channel_removal().
1046                  */
1047                 mutex_lock(&vmbus_connection.channel_mutex);
1048
1049                 atomic_dec(&vmbus_connection.offer_in_progress);
1050
1051                 WARN_ON(oldchannel->offermsg.child_relid != INVALID_RELID);
1052                 /* Fix up the relid. */
1053                 oldchannel->offermsg.child_relid = offer->child_relid;
1054
1055                 offer_sz = sizeof(*offer);
1056                 if (memcmp(offer, &oldchannel->offermsg, offer_sz) != 0) {
1057                         /*
1058                          * This is not an error, since the host can also change
1059                          * the other field(s) of the offer, e.g. on WS RS5
1060                          * (Build 17763), the offer->connection_id of the
1061                          * Mellanox VF vmbus device can change when the host
1062                          * reoffers the device upon resume.
1063                          */
1064                         pr_debug("vmbus offer changed: relid=%d\n",
1065                                  offer->child_relid);
1066
1067                         print_hex_dump_debug("Old vmbus offer: ",
1068                                              DUMP_PREFIX_OFFSET, 16, 4,
1069                                              &oldchannel->offermsg, offer_sz,
1070                                              false);
1071                         print_hex_dump_debug("New vmbus offer: ",
1072                                              DUMP_PREFIX_OFFSET, 16, 4,
1073                                              offer, offer_sz, false);
1074
1075                         /* Fix up the old channel. */
1076                         vmbus_setup_channel_state(oldchannel, offer);
1077                 }
1078
1079                 /* Add the channel back to the array of channels. */
1080                 vmbus_channel_map_relid(oldchannel);
1081                 check_ready_for_resume_event();
1082
1083                 mutex_unlock(&vmbus_connection.channel_mutex);
1084                 return;
1085         }
1086
1087         /* Allocate the channel object and save this offer. */
1088         newchannel = alloc_channel();
1089         if (!newchannel) {
1090                 vmbus_release_relid(offer->child_relid);
1091                 atomic_dec(&vmbus_connection.offer_in_progress);
1092                 pr_err("Unable to allocate channel object\n");
1093                 return;
1094         }
1095
1096         vmbus_setup_channel_state(newchannel, offer);
1097
1098         vmbus_process_offer(newchannel);
1099 }
1100
1101 static void check_ready_for_suspend_event(void)
1102 {
1103         /*
1104          * If all the sub-channels or hv_sock channels have been cleaned up,
1105          * then it's safe to suspend.
1106          */
1107         if (atomic_dec_and_test(&vmbus_connection.nr_chan_close_on_suspend))
1108                 complete(&vmbus_connection.ready_for_suspend_event);
1109 }
1110
1111 /*
1112  * vmbus_onoffer_rescind - Rescind offer handler.
1113  *
1114  * We queue a work item to process this offer synchronously
1115  */
1116 static void vmbus_onoffer_rescind(struct vmbus_channel_message_header *hdr)
1117 {
1118         struct vmbus_channel_rescind_offer *rescind;
1119         struct vmbus_channel *channel;
1120         struct device *dev;
1121         bool clean_up_chan_for_suspend;
1122
1123         rescind = (struct vmbus_channel_rescind_offer *)hdr;
1124
1125         trace_vmbus_onoffer_rescind(rescind);
1126
1127         /*
1128          * The offer msg and the corresponding rescind msg
1129          * from the host are guranteed to be ordered -
1130          * offer comes in first and then the rescind.
1131          * Since we process these events in work elements,
1132          * and with preemption, we may end up processing
1133          * the events out of order.  We rely on the synchronization
1134          * provided by offer_in_progress and by channel_mutex for
1135          * ordering these events:
1136          *
1137          * { Initially: offer_in_progress = 1 }
1138          *
1139          * CPU1                         CPU2
1140          *
1141          * [vmbus_onoffer()]            [vmbus_onoffer_rescind()]
1142          *
1143          * LOCK channel_mutex           WAIT_ON offer_in_progress == 0
1144          * DECREMENT offer_in_progress  LOCK channel_mutex
1145          * STORE channels[]             LOAD channels[]
1146          * UNLOCK channel_mutex         UNLOCK channel_mutex
1147          *
1148          * Forbids: CPU2's LOAD from *not* seeing CPU1's STORE
1149          */
1150
1151         while (atomic_read(&vmbus_connection.offer_in_progress) != 0) {
1152                 /*
1153                  * We wait here until any channel offer is currently
1154                  * being processed.
1155                  */
1156                 msleep(1);
1157         }
1158
1159         mutex_lock(&vmbus_connection.channel_mutex);
1160         channel = relid2channel(rescind->child_relid);
1161         if (channel != NULL) {
1162                 /*
1163                  * Guarantee that no other instance of vmbus_onoffer_rescind()
1164                  * has got a reference to the channel object.  Synchronize on
1165                  * &vmbus_connection.channel_mutex.
1166                  */
1167                 if (channel->rescind_ref) {
1168                         mutex_unlock(&vmbus_connection.channel_mutex);
1169                         return;
1170                 }
1171                 channel->rescind_ref = true;
1172         }
1173         mutex_unlock(&vmbus_connection.channel_mutex);
1174
1175         if (channel == NULL) {
1176                 /*
1177                  * We failed in processing the offer message;
1178                  * we would have cleaned up the relid in that
1179                  * failure path.
1180                  */
1181                 return;
1182         }
1183
1184         clean_up_chan_for_suspend = is_hvsock_channel(channel) ||
1185                                     is_sub_channel(channel);
1186         /*
1187          * Before setting channel->rescind in vmbus_rescind_cleanup(), we
1188          * should make sure the channel callback is not running any more.
1189          */
1190         vmbus_reset_channel_cb(channel);
1191
1192         /*
1193          * Now wait for offer handling to complete.
1194          */
1195         vmbus_rescind_cleanup(channel);
1196         while (READ_ONCE(channel->probe_done) == false) {
1197                 /*
1198                  * We wait here until any channel offer is currently
1199                  * being processed.
1200                  */
1201                 msleep(1);
1202         }
1203
1204         /*
1205          * At this point, the rescind handling can proceed safely.
1206          */
1207
1208         if (channel->device_obj) {
1209                 if (channel->chn_rescind_callback) {
1210                         channel->chn_rescind_callback(channel);
1211
1212                         if (clean_up_chan_for_suspend)
1213                                 check_ready_for_suspend_event();
1214
1215                         return;
1216                 }
1217                 /*
1218                  * We will have to unregister this device from the
1219                  * driver core.
1220                  */
1221                 dev = get_device(&channel->device_obj->device);
1222                 if (dev) {
1223                         vmbus_device_unregister(channel->device_obj);
1224                         put_device(dev);
1225                 }
1226         } else if (channel->primary_channel != NULL) {
1227                 /*
1228                  * Sub-channel is being rescinded. Following is the channel
1229                  * close sequence when initiated from the driveri (refer to
1230                  * vmbus_close() for details):
1231                  * 1. Close all sub-channels first
1232                  * 2. Then close the primary channel.
1233                  */
1234                 mutex_lock(&vmbus_connection.channel_mutex);
1235                 if (channel->state == CHANNEL_OPEN_STATE) {
1236                         /*
1237                          * The channel is currently not open;
1238                          * it is safe for us to cleanup the channel.
1239                          */
1240                         hv_process_channel_removal(channel);
1241                 } else {
1242                         complete(&channel->rescind_event);
1243                 }
1244                 mutex_unlock(&vmbus_connection.channel_mutex);
1245         }
1246
1247         /* The "channel" may have been freed. Do not access it any longer. */
1248
1249         if (clean_up_chan_for_suspend)
1250                 check_ready_for_suspend_event();
1251 }
1252
1253 void vmbus_hvsock_device_unregister(struct vmbus_channel *channel)
1254 {
1255         BUG_ON(!is_hvsock_channel(channel));
1256
1257         /* We always get a rescind msg when a connection is closed. */
1258         while (!READ_ONCE(channel->probe_done) || !READ_ONCE(channel->rescind))
1259                 msleep(1);
1260
1261         vmbus_device_unregister(channel->device_obj);
1262 }
1263 EXPORT_SYMBOL_GPL(vmbus_hvsock_device_unregister);
1264
1265
1266 /*
1267  * vmbus_onoffers_delivered -
1268  * This is invoked when all offers have been delivered.
1269  *
1270  * Nothing to do here.
1271  */
1272 static void vmbus_onoffers_delivered(
1273                         struct vmbus_channel_message_header *hdr)
1274 {
1275 }
1276
1277 /*
1278  * vmbus_onopen_result - Open result handler.
1279  *
1280  * This is invoked when we received a response to our channel open request.
1281  * Find the matching request, copy the response and signal the requesting
1282  * thread.
1283  */
1284 static void vmbus_onopen_result(struct vmbus_channel_message_header *hdr)
1285 {
1286         struct vmbus_channel_open_result *result;
1287         struct vmbus_channel_msginfo *msginfo;
1288         struct vmbus_channel_message_header *requestheader;
1289         struct vmbus_channel_open_channel *openmsg;
1290         unsigned long flags;
1291
1292         result = (struct vmbus_channel_open_result *)hdr;
1293
1294         trace_vmbus_onopen_result(result);
1295
1296         /*
1297          * Find the open msg, copy the result and signal/unblock the wait event
1298          */
1299         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1300
1301         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1302                                 msglistentry) {
1303                 requestheader =
1304                         (struct vmbus_channel_message_header *)msginfo->msg;
1305
1306                 if (requestheader->msgtype == CHANNELMSG_OPENCHANNEL) {
1307                         openmsg =
1308                         (struct vmbus_channel_open_channel *)msginfo->msg;
1309                         if (openmsg->child_relid == result->child_relid &&
1310                             openmsg->openid == result->openid) {
1311                                 memcpy(&msginfo->response.open_result,
1312                                        result,
1313                                        sizeof(
1314                                         struct vmbus_channel_open_result));
1315                                 complete(&msginfo->waitevent);
1316                                 break;
1317                         }
1318                 }
1319         }
1320         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1321 }
1322
1323 /*
1324  * vmbus_ongpadl_created - GPADL created handler.
1325  *
1326  * This is invoked when we received a response to our gpadl create request.
1327  * Find the matching request, copy the response and signal the requesting
1328  * thread.
1329  */
1330 static void vmbus_ongpadl_created(struct vmbus_channel_message_header *hdr)
1331 {
1332         struct vmbus_channel_gpadl_created *gpadlcreated;
1333         struct vmbus_channel_msginfo *msginfo;
1334         struct vmbus_channel_message_header *requestheader;
1335         struct vmbus_channel_gpadl_header *gpadlheader;
1336         unsigned long flags;
1337
1338         gpadlcreated = (struct vmbus_channel_gpadl_created *)hdr;
1339
1340         trace_vmbus_ongpadl_created(gpadlcreated);
1341
1342         /*
1343          * Find the establish msg, copy the result and signal/unblock the wait
1344          * event
1345          */
1346         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1347
1348         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1349                                 msglistentry) {
1350                 requestheader =
1351                         (struct vmbus_channel_message_header *)msginfo->msg;
1352
1353                 if (requestheader->msgtype == CHANNELMSG_GPADL_HEADER) {
1354                         gpadlheader =
1355                         (struct vmbus_channel_gpadl_header *)requestheader;
1356
1357                         if ((gpadlcreated->child_relid ==
1358                              gpadlheader->child_relid) &&
1359                             (gpadlcreated->gpadl == gpadlheader->gpadl)) {
1360                                 memcpy(&msginfo->response.gpadl_created,
1361                                        gpadlcreated,
1362                                        sizeof(
1363                                         struct vmbus_channel_gpadl_created));
1364                                 complete(&msginfo->waitevent);
1365                                 break;
1366                         }
1367                 }
1368         }
1369         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1370 }
1371
1372 /*
1373  * vmbus_onmodifychannel_response - Modify Channel response handler.
1374  *
1375  * This is invoked when we received a response to our channel modify request.
1376  * Find the matching request, copy the response and signal the requesting thread.
1377  */
1378 static void vmbus_onmodifychannel_response(struct vmbus_channel_message_header *hdr)
1379 {
1380         struct vmbus_channel_modifychannel_response *response;
1381         struct vmbus_channel_msginfo *msginfo;
1382         unsigned long flags;
1383
1384         response = (struct vmbus_channel_modifychannel_response *)hdr;
1385
1386         trace_vmbus_onmodifychannel_response(response);
1387
1388         /*
1389          * Find the modify msg, copy the response and signal/unblock the wait event.
1390          */
1391         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1392
1393         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list, msglistentry) {
1394                 struct vmbus_channel_message_header *responseheader =
1395                                 (struct vmbus_channel_message_header *)msginfo->msg;
1396
1397                 if (responseheader->msgtype == CHANNELMSG_MODIFYCHANNEL) {
1398                         struct vmbus_channel_modifychannel *modifymsg;
1399
1400                         modifymsg = (struct vmbus_channel_modifychannel *)msginfo->msg;
1401                         if (modifymsg->child_relid == response->child_relid) {
1402                                 memcpy(&msginfo->response.modify_response, response,
1403                                        sizeof(*response));
1404                                 complete(&msginfo->waitevent);
1405                                 break;
1406                         }
1407                 }
1408         }
1409         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1410 }
1411
1412 /*
1413  * vmbus_ongpadl_torndown - GPADL torndown handler.
1414  *
1415  * This is invoked when we received a response to our gpadl teardown request.
1416  * Find the matching request, copy the response and signal the requesting
1417  * thread.
1418  */
1419 static void vmbus_ongpadl_torndown(
1420                         struct vmbus_channel_message_header *hdr)
1421 {
1422         struct vmbus_channel_gpadl_torndown *gpadl_torndown;
1423         struct vmbus_channel_msginfo *msginfo;
1424         struct vmbus_channel_message_header *requestheader;
1425         struct vmbus_channel_gpadl_teardown *gpadl_teardown;
1426         unsigned long flags;
1427
1428         gpadl_torndown = (struct vmbus_channel_gpadl_torndown *)hdr;
1429
1430         trace_vmbus_ongpadl_torndown(gpadl_torndown);
1431
1432         /*
1433          * Find the open msg, copy the result and signal/unblock the wait event
1434          */
1435         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1436
1437         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1438                                 msglistentry) {
1439                 requestheader =
1440                         (struct vmbus_channel_message_header *)msginfo->msg;
1441
1442                 if (requestheader->msgtype == CHANNELMSG_GPADL_TEARDOWN) {
1443                         gpadl_teardown =
1444                         (struct vmbus_channel_gpadl_teardown *)requestheader;
1445
1446                         if (gpadl_torndown->gpadl == gpadl_teardown->gpadl) {
1447                                 memcpy(&msginfo->response.gpadl_torndown,
1448                                        gpadl_torndown,
1449                                        sizeof(
1450                                         struct vmbus_channel_gpadl_torndown));
1451                                 complete(&msginfo->waitevent);
1452                                 break;
1453                         }
1454                 }
1455         }
1456         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1457 }
1458
1459 /*
1460  * vmbus_onversion_response - Version response handler
1461  *
1462  * This is invoked when we received a response to our initiate contact request.
1463  * Find the matching request, copy the response and signal the requesting
1464  * thread.
1465  */
1466 static void vmbus_onversion_response(
1467                 struct vmbus_channel_message_header *hdr)
1468 {
1469         struct vmbus_channel_msginfo *msginfo;
1470         struct vmbus_channel_message_header *requestheader;
1471         struct vmbus_channel_version_response *version_response;
1472         unsigned long flags;
1473
1474         version_response = (struct vmbus_channel_version_response *)hdr;
1475
1476         trace_vmbus_onversion_response(version_response);
1477
1478         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1479
1480         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1481                                 msglistentry) {
1482                 requestheader =
1483                         (struct vmbus_channel_message_header *)msginfo->msg;
1484
1485                 if (requestheader->msgtype ==
1486                     CHANNELMSG_INITIATE_CONTACT) {
1487                         memcpy(&msginfo->response.version_response,
1488                               version_response,
1489                               sizeof(struct vmbus_channel_version_response));
1490                         complete(&msginfo->waitevent);
1491                 }
1492         }
1493         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1494 }
1495
1496 /* Channel message dispatch table */
1497 const struct vmbus_channel_message_table_entry
1498 channel_message_table[CHANNELMSG_COUNT] = {
1499         { CHANNELMSG_INVALID,                   0, NULL, 0},
1500         { CHANNELMSG_OFFERCHANNEL,              0, vmbus_onoffer,
1501                 sizeof(struct vmbus_channel_offer_channel)},
1502         { CHANNELMSG_RESCIND_CHANNELOFFER,      0, vmbus_onoffer_rescind,
1503                 sizeof(struct vmbus_channel_rescind_offer) },
1504         { CHANNELMSG_REQUESTOFFERS,             0, NULL, 0},
1505         { CHANNELMSG_ALLOFFERS_DELIVERED,       1, vmbus_onoffers_delivered, 0},
1506         { CHANNELMSG_OPENCHANNEL,               0, NULL, 0},
1507         { CHANNELMSG_OPENCHANNEL_RESULT,        1, vmbus_onopen_result,
1508                 sizeof(struct vmbus_channel_open_result)},
1509         { CHANNELMSG_CLOSECHANNEL,              0, NULL, 0},
1510         { CHANNELMSG_GPADL_HEADER,              0, NULL, 0},
1511         { CHANNELMSG_GPADL_BODY,                0, NULL, 0},
1512         { CHANNELMSG_GPADL_CREATED,             1, vmbus_ongpadl_created,
1513                 sizeof(struct vmbus_channel_gpadl_created)},
1514         { CHANNELMSG_GPADL_TEARDOWN,            0, NULL, 0},
1515         { CHANNELMSG_GPADL_TORNDOWN,            1, vmbus_ongpadl_torndown,
1516                 sizeof(struct vmbus_channel_gpadl_torndown) },
1517         { CHANNELMSG_RELID_RELEASED,            0, NULL, 0},
1518         { CHANNELMSG_INITIATE_CONTACT,          0, NULL, 0},
1519         { CHANNELMSG_VERSION_RESPONSE,          1, vmbus_onversion_response,
1520                 sizeof(struct vmbus_channel_version_response)},
1521         { CHANNELMSG_UNLOAD,                    0, NULL, 0},
1522         { CHANNELMSG_UNLOAD_RESPONSE,           1, vmbus_unload_response, 0},
1523         { CHANNELMSG_18,                        0, NULL, 0},
1524         { CHANNELMSG_19,                        0, NULL, 0},
1525         { CHANNELMSG_20,                        0, NULL, 0},
1526         { CHANNELMSG_TL_CONNECT_REQUEST,        0, NULL, 0},
1527         { CHANNELMSG_MODIFYCHANNEL,             0, NULL, 0},
1528         { CHANNELMSG_TL_CONNECT_RESULT,         0, NULL, 0},
1529         { CHANNELMSG_MODIFYCHANNEL_RESPONSE,    1, vmbus_onmodifychannel_response,
1530                 sizeof(struct vmbus_channel_modifychannel_response)},
1531 };
1532
1533 /*
1534  * vmbus_onmessage - Handler for channel protocol messages.
1535  *
1536  * This is invoked in the vmbus worker thread context.
1537  */
1538 void vmbus_onmessage(struct vmbus_channel_message_header *hdr)
1539 {
1540         trace_vmbus_on_message(hdr);
1541
1542         /*
1543          * vmbus_on_msg_dpc() makes sure the hdr->msgtype here can not go
1544          * out of bound and the message_handler pointer can not be NULL.
1545          */
1546         channel_message_table[hdr->msgtype].message_handler(hdr);
1547 }
1548
1549 /*
1550  * vmbus_request_offers - Send a request to get all our pending offers.
1551  */
1552 int vmbus_request_offers(void)
1553 {
1554         struct vmbus_channel_message_header *msg;
1555         struct vmbus_channel_msginfo *msginfo;
1556         int ret;
1557
1558         msginfo = kmalloc(sizeof(*msginfo) +
1559                           sizeof(struct vmbus_channel_message_header),
1560                           GFP_KERNEL);
1561         if (!msginfo)
1562                 return -ENOMEM;
1563
1564         msg = (struct vmbus_channel_message_header *)msginfo->msg;
1565
1566         msg->msgtype = CHANNELMSG_REQUESTOFFERS;
1567
1568         ret = vmbus_post_msg(msg, sizeof(struct vmbus_channel_message_header),
1569                              true);
1570
1571         trace_vmbus_request_offers(ret);
1572
1573         if (ret != 0) {
1574                 pr_err("Unable to request offers - %d\n", ret);
1575
1576                 goto cleanup;
1577         }
1578
1579 cleanup:
1580         kfree(msginfo);
1581
1582         return ret;
1583 }
1584
1585 static void invoke_sc_cb(struct vmbus_channel *primary_channel)
1586 {
1587         struct list_head *cur, *tmp;
1588         struct vmbus_channel *cur_channel;
1589
1590         if (primary_channel->sc_creation_callback == NULL)
1591                 return;
1592
1593         list_for_each_safe(cur, tmp, &primary_channel->sc_list) {
1594                 cur_channel = list_entry(cur, struct vmbus_channel, sc_list);
1595
1596                 primary_channel->sc_creation_callback(cur_channel);
1597         }
1598 }
1599
1600 void vmbus_set_sc_create_callback(struct vmbus_channel *primary_channel,
1601                                 void (*sc_cr_cb)(struct vmbus_channel *new_sc))
1602 {
1603         primary_channel->sc_creation_callback = sc_cr_cb;
1604 }
1605 EXPORT_SYMBOL_GPL(vmbus_set_sc_create_callback);
1606
1607 bool vmbus_are_subchannels_present(struct vmbus_channel *primary)
1608 {
1609         bool ret;
1610
1611         ret = !list_empty(&primary->sc_list);
1612
1613         if (ret) {
1614                 /*
1615                  * Invoke the callback on sub-channel creation.
1616                  * This will present a uniform interface to the
1617                  * clients.
1618                  */
1619                 invoke_sc_cb(primary);
1620         }
1621
1622         return ret;
1623 }
1624 EXPORT_SYMBOL_GPL(vmbus_are_subchannels_present);
1625
1626 void vmbus_set_chn_rescind_callback(struct vmbus_channel *channel,
1627                 void (*chn_rescind_cb)(struct vmbus_channel *))
1628 {
1629         channel->chn_rescind_callback = chn_rescind_cb;
1630 }
1631 EXPORT_SYMBOL_GPL(vmbus_set_chn_rescind_callback);