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 #include <linux/sched/isolation.h>
  25
  26 #include "hyperv_vmbus.h"
  27
  28 static void init_vp_index(struct vmbus_channel *channel);
  29
  30 const struct vmbus_device vmbus_devs[] = {
  31         /* IDE */
  32         { .dev_type = HV_IDE,
  33           HV_IDE_GUID,
  34           .perf_device = true,
  35           .allowed_in_isolated = false,
  36         },
  37
  38         /* SCSI */
  39         { .dev_type = HV_SCSI,
  40           HV_SCSI_GUID,
  41           .perf_device = true,
  42           .allowed_in_isolated = true,
  43         },
  44
  45         /* Fibre Channel */
  46         { .dev_type = HV_FC,
  47           HV_SYNTHFC_GUID,
  48           .perf_device = true,
  49           .allowed_in_isolated = false,
  50         },
  51
  52         /* Synthetic NIC */
  53         { .dev_type = HV_NIC,
  54           HV_NIC_GUID,
  55           .perf_device = true,
  56           .allowed_in_isolated = true,
  57         },
  58
  59         /* Network Direct */
  60         { .dev_type = HV_ND,
  61           HV_ND_GUID,
  62           .perf_device = true,
  63           .allowed_in_isolated = false,
  64         },
  65
  66         /* PCIE */
  67         { .dev_type = HV_PCIE,
  68           HV_PCIE_GUID,
  69           .perf_device = false,
  70           .allowed_in_isolated = true,
  71         },
  72
  73         /* Synthetic Frame Buffer */
  74         { .dev_type = HV_FB,
  75           HV_SYNTHVID_GUID,
  76           .perf_device = false,
  77           .allowed_in_isolated = false,
  78         },
  79
  80         /* Synthetic Keyboard */
  81         { .dev_type = HV_KBD,
  82           HV_KBD_GUID,
  83           .perf_device = false,
  84           .allowed_in_isolated = false,
  85         },
  86
  87         /* Synthetic MOUSE */
  88         { .dev_type = HV_MOUSE,
  89           HV_MOUSE_GUID,
  90           .perf_device = false,
  91           .allowed_in_isolated = false,
  92         },
  93
  94         /* KVP */
  95         { .dev_type = HV_KVP,
  96           HV_KVP_GUID,
  97           .perf_device = false,
  98           .allowed_in_isolated = false,
  99         },
 100
 101         /* Time Synch */
 102         { .dev_type = HV_TS,
 103           HV_TS_GUID,
 104           .perf_device = false,
 105           .allowed_in_isolated = true,
 106         },
 107
 108         /* Heartbeat */
 109         { .dev_type = HV_HB,
 110           HV_HEART_BEAT_GUID,
 111           .perf_device = false,
 112           .allowed_in_isolated = true,
 113         },
 114
 115         /* Shutdown */
 116         { .dev_type = HV_SHUTDOWN,
 117           HV_SHUTDOWN_GUID,
 118           .perf_device = false,
 119           .allowed_in_isolated = true,
 120         },
 121
 122         /* File copy */
 123         { .dev_type = HV_FCOPY,
 124           HV_FCOPY_GUID,
 125           .perf_device = false,
 126           .allowed_in_isolated = false,
 127         },
 128
 129         /* Backup */
 130         { .dev_type = HV_BACKUP,
 131           HV_VSS_GUID,
 132           .perf_device = false,
 133           .allowed_in_isolated = false,
 134         },
 135
 136         /* Dynamic Memory */
 137         { .dev_type = HV_DM,
 138           HV_DM_GUID,
 139           .perf_device = false,
 140           .allowed_in_isolated = false,
 141         },
 142
 143         /* Unknown GUID */
 144         { .dev_type = HV_UNKNOWN,
 145           .perf_device = false,
 146           .allowed_in_isolated = false,
 147         },
 148 };
 149
 150 static const struct {
 151         guid_t guid;
 152 } vmbus_unsupported_devs[] = {
 153         { HV_AVMA1_GUID },
 154         { HV_AVMA2_GUID },
 155         { HV_RDV_GUID   },
 156         { HV_IMC_GUID   },
 157 };
 158
 159 /*
 160  * The rescinded channel may be blocked waiting for a response from the host;
 161  * take care of that.
 162  */
 163 static void vmbus_rescind_cleanup(struct vmbus_channel *channel)
 164 {
 165         struct vmbus_channel_msginfo *msginfo;
 166         unsigned long flags;
 167
 168
 169         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
 170         channel->rescind = true;
 171         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
 172                                 msglistentry) {
 173
 174                 if (msginfo->waiting_channel == channel) {
 175                         complete(&msginfo->waitevent);
 176                         break;
 177                 }
 178         }
 179         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
 180 }
 181
 182 static bool is_unsupported_vmbus_devs(const guid_t *guid)
 183 {
 184         int i;
 185
 186         for (i = 0; i < ARRAY_SIZE(vmbus_unsupported_devs); i++)
 187                 if (guid_equal(guid, &vmbus_unsupported_devs[i].guid))
 188                         return true;
 189         return false;
 190 }
 191
 192 static u16 hv_get_dev_type(const struct vmbus_channel *channel)
 193 {
 194         const guid_t *guid = &channel->offermsg.offer.if_type;
 195         u16 i;
 196
 197         if (is_hvsock_channel(channel) || is_unsupported_vmbus_devs(guid))
 198                 return HV_UNKNOWN;
 199
 200         for (i = HV_IDE; i < HV_UNKNOWN; i++) {
 201                 if (guid_equal(guid, &vmbus_devs[i].guid))
 202                         return i;
 203         }
 204         pr_info("Unknown GUID: %pUl\n", guid);
 205         return i;
 206 }
 207
 208 /**
 209  * vmbus_prep_negotiate_resp() - Create default response for Negotiate message
 210  * @icmsghdrp: Pointer to msg header structure
 211  * @buf: Raw buffer channel data
 212  * @buflen: Length of the raw buffer channel data.
 213  * @fw_version: The framework versions we can support.
 214  * @fw_vercnt: The size of @fw_version.
 215  * @srv_version: The service versions we can support.
 216  * @srv_vercnt: The size of @srv_version.
 217  * @nego_fw_version: The selected framework version.
 218  * @nego_srv_version: The selected service version.
 219  *
 220  * Note: Versions are given in decreasing order.
 221  *
 222  * Set up and fill in default negotiate response message.
 223  * Mainly used by Hyper-V drivers.
 224  */
 225 bool vmbus_prep_negotiate_resp(struct icmsg_hdr *icmsghdrp, u8 *buf,
 226                                 u32 buflen, const int *fw_version, int fw_vercnt,
 227                                 const int *srv_version, int srv_vercnt,
 228                                 int *nego_fw_version, int *nego_srv_version)
 229 {
 230         int icframe_major, icframe_minor;
 231         int icmsg_major, icmsg_minor;
 232         int fw_major, fw_minor;
 233         int srv_major, srv_minor;
 234         int i, j;
 235         bool found_match = false;
 236         struct icmsg_negotiate *negop;
 237
 238         /* Check that there's enough space for icframe_vercnt, icmsg_vercnt */
 239         if (buflen < ICMSG_HDR + offsetof(struct icmsg_negotiate, reserved)) {
 240                 pr_err_ratelimited("Invalid icmsg negotiate\n");
 241                 return false;
 242         }
 243
 244         icmsghdrp->icmsgsize = 0x10;
 245         negop = (struct icmsg_negotiate *)&buf[ICMSG_HDR];
 246
 247         icframe_major = negop->icframe_vercnt;
 248         icframe_minor = 0;
 249
 250         icmsg_major = negop->icmsg_vercnt;
 251         icmsg_minor = 0;
 252
 253         /* Validate negop packet */
 254         if (icframe_major > IC_VERSION_NEGOTIATION_MAX_VER_COUNT ||
 255             icmsg_major > IC_VERSION_NEGOTIATION_MAX_VER_COUNT ||
 256             ICMSG_NEGOTIATE_PKT_SIZE(icframe_major, icmsg_major) > buflen) {
 257                 pr_err_ratelimited("Invalid icmsg negotiate - icframe_major: %u, icmsg_major: %u\n",
 258                                    icframe_major, icmsg_major);
 259                 goto fw_error;
 260         }
 261
 262         /*
 263          * Select the framework version number we will
 264          * support.
 265          */
 266
 267         for (i = 0; i < fw_vercnt; i++) {
 268                 fw_major = (fw_version[i] >> 16);
 269                 fw_minor = (fw_version[i] & 0xFFFF);
 270
 271                 for (j = 0; j < negop->icframe_vercnt; j++) {
 272                         if ((negop->icversion_data[j].major == fw_major) &&
 273                             (negop->icversion_data[j].minor == fw_minor)) {
 274                                 icframe_major = negop->icversion_data[j].major;
 275                                 icframe_minor = negop->icversion_data[j].minor;
 276                                 found_match = true;
 277                                 break;
 278                         }
 279                 }
 280
 281                 if (found_match)
 282                         break;
 283         }
 284
 285         if (!found_match)
 286                 goto fw_error;
 287
 288         found_match = false;
 289
 290         for (i = 0; i < srv_vercnt; i++) {
 291                 srv_major = (srv_version[i] >> 16);
 292                 srv_minor = (srv_version[i] & 0xFFFF);
 293
 294                 for (j = negop->icframe_vercnt;
 295                         (j < negop->icframe_vercnt + negop->icmsg_vercnt);
 296                         j++) {
 297
 298                         if ((negop->icversion_data[j].major == srv_major) &&
 299                                 (negop->icversion_data[j].minor == srv_minor)) {
 300
 301                                 icmsg_major = negop->icversion_data[j].major;
 302                                 icmsg_minor = negop->icversion_data[j].minor;
 303                                 found_match = true;
 304                                 break;
 305                         }
 306                 }
 307
 308                 if (found_match)
 309                         break;
 310         }
 311
 312         /*
 313          * Respond with the framework and service
 314          * version numbers we can support.
 315          */
 316
 317 fw_error:
 318         if (!found_match) {
 319                 negop->icframe_vercnt = 0;
 320                 negop->icmsg_vercnt = 0;
 321         } else {
 322                 negop->icframe_vercnt = 1;
 323                 negop->icmsg_vercnt = 1;
 324         }
 325
 326         if (nego_fw_version)
 327                 *nego_fw_version = (icframe_major << 16) | icframe_minor;
 328
 329         if (nego_srv_version)
 330                 *nego_srv_version = (icmsg_major << 16) | icmsg_minor;
 331
 332         negop->icversion_data[0].major = icframe_major;
 333         negop->icversion_data[0].minor = icframe_minor;
 334         negop->icversion_data[1].major = icmsg_major;
 335         negop->icversion_data[1].minor = icmsg_minor;
 336         return found_match;
 337 }
 338 EXPORT_SYMBOL_GPL(vmbus_prep_negotiate_resp);
 339
 340 /*
 341  * alloc_channel - Allocate and initialize a vmbus channel object
 342  */
 343 static struct vmbus_channel *alloc_channel(void)
 344 {
 345         struct vmbus_channel *channel;
 346
 347         channel = kzalloc(sizeof(*channel), GFP_ATOMIC);
 348         if (!channel)
 349                 return NULL;
 350
 351         spin_lock_init(&channel->sched_lock);
 352         init_completion(&channel->rescind_event);
 353
 354         INIT_LIST_HEAD(&channel->sc_list);
 355
 356         tasklet_init(&channel->callback_event,
 357                      vmbus_on_event, (unsigned long)channel);
 358
 359         hv_ringbuffer_pre_init(channel);
 360
 361         return channel;
 362 }
 363
 364 /*
 365  * free_channel - Release the resources used by the vmbus channel object
 366  */
 367 static void free_channel(struct vmbus_channel *channel)
 368 {
 369         tasklet_kill(&channel->callback_event);
 370         vmbus_remove_channel_attr_group(channel);
 371
 372         kobject_put(&channel->kobj);
 373 }
 374
 375 void vmbus_channel_map_relid(struct vmbus_channel *channel)
 376 {
 377         if (WARN_ON(channel->offermsg.child_relid >= MAX_CHANNEL_RELIDS))
 378                 return;
 379         /*
 380          * The mapping of the channel's relid is visible from the CPUs that
 381          * execute vmbus_chan_sched() by the time that vmbus_chan_sched() will
 382          * execute:
 383          *
 384          *  (a) In the "normal (i.e., not resuming from hibernation)" path,
 385          *      the full barrier in virt_store_mb() guarantees that the store
 386          *      is propagated to all CPUs before the add_channel_work work
 387          *      is queued.  In turn, add_channel_work is queued before the
 388          *      channel's ring buffer is allocated/initialized and the
 389          *      OPENCHANNEL message for the channel is sent in vmbus_open().
 390          *      Hyper-V won't start sending the interrupts for the channel
 391          *      before the OPENCHANNEL message is acked.  The memory barrier
 392          *      in vmbus_chan_sched() -> sync_test_and_clear_bit() ensures
 393          *      that vmbus_chan_sched() must find the channel's relid in
 394          *      recv_int_page before retrieving the channel pointer from the
 395          *      array of channels.
 396          *
 397          *  (b) In the "resuming from hibernation" path, the virt_store_mb()
 398          *      guarantees that the store is propagated to all CPUs before
 399          *      the VMBus connection is marked as ready for the resume event
 400          *      (cf. check_ready_for_resume_event()).  The interrupt handler
 401          *      of the VMBus driver and vmbus_chan_sched() can not run before
 402          *      vmbus_bus_resume() has completed execution (cf. resume_noirq).
 403          */
 404         virt_store_mb(
 405                 vmbus_connection.channels[channel->offermsg.child_relid],
 406                 channel);
 407 }
 408
 409 void vmbus_channel_unmap_relid(struct vmbus_channel *channel)
 410 {
 411         if (WARN_ON(channel->offermsg.child_relid >= MAX_CHANNEL_RELIDS))
 412                 return;
 413         WRITE_ONCE(
 414                 vmbus_connection.channels[channel->offermsg.child_relid],
 415                 NULL);
 416 }
 417
 418 static void vmbus_release_relid(u32 relid)
 419 {
 420         struct vmbus_channel_relid_released msg;
 421         int ret;
 422
 423         memset(&msg, 0, sizeof(struct vmbus_channel_relid_released));
 424         msg.child_relid = relid;
 425         msg.header.msgtype = CHANNELMSG_RELID_RELEASED;
 426         ret = vmbus_post_msg(&msg, sizeof(struct vmbus_channel_relid_released),
 427                              true);
 428
 429         trace_vmbus_release_relid(&msg, ret);
 430 }
 431
 432 void hv_process_channel_removal(struct vmbus_channel *channel)
 433 {
 434         lockdep_assert_held(&vmbus_connection.channel_mutex);
 435         BUG_ON(!channel->rescind);
 436
 437         /*
 438          * hv_process_channel_removal() could find INVALID_RELID only for
 439          * hv_sock channels.  See the inline comments in vmbus_onoffer().
 440          */
 441         WARN_ON(channel->offermsg.child_relid == INVALID_RELID &&
 442                 !is_hvsock_channel(channel));
 443
 444         /*
 445          * Upon suspend, an in-use hv_sock channel is removed from the array of
 446          * channels and the relid is invalidated.  After hibernation, when the
 447          * user-space application destroys the channel, it's unnecessary and
 448          * unsafe to remove the channel from the array of channels.  See also
 449          * the inline comments before the call of vmbus_release_relid() below.
 450          */
 451         if (channel->offermsg.child_relid != INVALID_RELID)
 452                 vmbus_channel_unmap_relid(channel);
 453
 454         if (channel->primary_channel == NULL)
 455                 list_del(&channel->listentry);
 456         else
 457                 list_del(&channel->sc_list);
 458
 459         /*
 460          * If this is a "perf" channel, updates the hv_numa_map[] masks so that
 461          * init_vp_index() can (re-)use the CPU.
 462          */
 463         if (hv_is_perf_channel(channel))
 464                 hv_clear_allocated_cpu(channel->target_cpu);
 465
 466         /*
 467          * Upon suspend, an in-use hv_sock channel is marked as "rescinded" and
 468          * the relid is invalidated; after hibernation, when the user-space app
 469          * destroys the channel, the relid is INVALID_RELID, and in this case
 470          * it's unnecessary and unsafe to release the old relid, since the same
 471          * relid can refer to a completely different channel now.
 472          */
 473         if (channel->offermsg.child_relid != INVALID_RELID)
 474                 vmbus_release_relid(channel->offermsg.child_relid);
 475
 476         free_channel(channel);
 477 }
 478
 479 void vmbus_free_channels(void)
 480 {
 481         struct vmbus_channel *channel, *tmp;
 482
 483         list_for_each_entry_safe(channel, tmp, &vmbus_connection.chn_list,
 484                 listentry) {
 485                 /* hv_process_channel_removal() needs this */
 486                 channel->rescind = true;
 487
 488                 vmbus_device_unregister(channel->device_obj);
 489         }
 490 }
 491
 492 /* Note: the function can run concurrently for primary/sub channels. */
 493 static void vmbus_add_channel_work(struct work_struct *work)
 494 {
 495         struct vmbus_channel *newchannel =
 496                 container_of(work, struct vmbus_channel, add_channel_work);
 497         struct vmbus_channel *primary_channel = newchannel->primary_channel;
 498         int ret;
 499
 500         /*
 501          * This state is used to indicate a successful open
 502          * so that when we do close the channel normally, we
 503          * can cleanup properly.
 504          */
 505         newchannel->state = CHANNEL_OPEN_STATE;
 506
 507         if (primary_channel != NULL) {
 508                 /* newchannel is a sub-channel. */
 509                 struct hv_device *dev = primary_channel->device_obj;
 510
 511                 if (vmbus_add_channel_kobj(dev, newchannel))
 512                         goto err_deq_chan;
 513
 514                 if (primary_channel->sc_creation_callback != NULL)
 515                         primary_channel->sc_creation_callback(newchannel);
 516
 517                 newchannel->probe_done = true;
 518                 return;
 519         }
 520
 521         /*
 522          * Start the process of binding the primary channel to the driver
 523          */
 524         newchannel->device_obj = vmbus_device_create(
 525                 &newchannel->offermsg.offer.if_type,
 526                 &newchannel->offermsg.offer.if_instance,
 527                 newchannel);
 528         if (!newchannel->device_obj)
 529                 goto err_deq_chan;
 530
 531         newchannel->device_obj->device_id = newchannel->device_id;
 532         /*
 533          * Add the new device to the bus. This will kick off device-driver
 534          * binding which eventually invokes the device driver's AddDevice()
 535          * method.
 536          *
 537          * If vmbus_device_register() fails, the 'device_obj' is freed in
 538          * vmbus_device_release() as called by device_unregister() in the
 539          * error path of vmbus_device_register(). In the outside error
 540          * path, there's no need to free it.
 541          */
 542         ret = vmbus_device_register(newchannel->device_obj);
 543
 544         if (ret != 0) {
 545                 pr_err("unable to add child device object (relid %d)\n",
 546                         newchannel->offermsg.child_relid);
 547                 goto err_deq_chan;
 548         }
 549
 550         newchannel->probe_done = true;
 551         return;
 552
 553 err_deq_chan:
 554         mutex_lock(&vmbus_connection.channel_mutex);
 555
 556         /*
 557          * We need to set the flag, otherwise
 558          * vmbus_onoffer_rescind() can be blocked.
 559          */
 560         newchannel->probe_done = true;
 561
 562         if (primary_channel == NULL)
 563                 list_del(&newchannel->listentry);
 564         else
 565                 list_del(&newchannel->sc_list);
 566
 567         /* vmbus_process_offer() has mapped the channel. */
 568         vmbus_channel_unmap_relid(newchannel);
 569
 570         mutex_unlock(&vmbus_connection.channel_mutex);
 571
 572         vmbus_release_relid(newchannel->offermsg.child_relid);
 573
 574         free_channel(newchannel);
 575 }
 576
 577 /*
 578  * vmbus_process_offer - Process the offer by creating a channel/device
 579  * associated with this offer
 580  */
 581 static void vmbus_process_offer(struct vmbus_channel *newchannel)
 582 {
 583         struct vmbus_channel *channel;
 584         struct workqueue_struct *wq;
 585         bool fnew = true;
 586
 587         /*
 588          * Synchronize vmbus_process_offer() and CPU hotplugging:
 589          *
 590          * CPU1                         CPU2
 591          *
 592          * [vmbus_process_offer()]      [Hot removal of the CPU]
 593          *
 594          * CPU_READ_LOCK                CPUS_WRITE_LOCK
 595          * LOAD cpu_online_mask         SEARCH chn_list
 596          * STORE target_cpu             LOAD target_cpu
 597          * INSERT chn_list              STORE cpu_online_mask
 598          * CPUS_READ_UNLOCK             CPUS_WRITE_UNLOCK
 599          *
 600          * Forbids: CPU1's LOAD from *not* seing CPU2's STORE &&
 601          *              CPU2's SEARCH from *not* seeing CPU1's INSERT
 602          *
 603          * Forbids: CPU2's SEARCH from seeing CPU1's INSERT &&
 604          *              CPU2's LOAD from *not* seing CPU1's STORE
 605          */
 606         cpus_read_lock();
 607
 608         /*
 609          * Serializes the modifications of the chn_list list as well as
 610          * the accesses to next_numa_node_id in init_vp_index().
 611          */
 612         mutex_lock(&vmbus_connection.channel_mutex);
 613
 614         list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
 615                 if (guid_equal(&channel->offermsg.offer.if_type,
 616                                &newchannel->offermsg.offer.if_type) &&
 617                     guid_equal(&channel->offermsg.offer.if_instance,
 618                                &newchannel->offermsg.offer.if_instance)) {
 619                         fnew = false;
 620                         newchannel->primary_channel = channel;
 621                         break;
 622                 }
 623         }
 624
 625         init_vp_index(newchannel);
 626
 627         /* Remember the channels that should be cleaned up upon suspend. */
 628         if (is_hvsock_channel(newchannel) || is_sub_channel(newchannel))
 629                 atomic_inc(&vmbus_connection.nr_chan_close_on_suspend);
 630
 631         /*
 632          * Now that we have acquired the channel_mutex,
 633          * we can release the potentially racing rescind thread.
 634          */
 635         atomic_dec(&vmbus_connection.offer_in_progress);
 636
 637         if (fnew) {
 638                 list_add_tail(&newchannel->listentry,
 639                               &vmbus_connection.chn_list);
 640         } else {
 641                 /*
 642                  * Check to see if this is a valid sub-channel.
 643                  */
 644                 if (newchannel->offermsg.offer.sub_channel_index == 0) {
 645                         mutex_unlock(&vmbus_connection.channel_mutex);
 646                         cpus_read_unlock();
 647                         /*
 648                          * Don't call free_channel(), because newchannel->kobj
 649                          * is not initialized yet.
 650                          */
 651                         kfree(newchannel);
 652                         WARN_ON_ONCE(1);
 653                         return;
 654                 }
 655                 /*
 656                  * Process the sub-channel.
 657                  */
 658                 list_add_tail(&newchannel->sc_list, &channel->sc_list);
 659         }
 660
 661         vmbus_channel_map_relid(newchannel);
 662
 663         mutex_unlock(&vmbus_connection.channel_mutex);
 664         cpus_read_unlock();
 665
 666         /*
 667          * vmbus_process_offer() mustn't call channel->sc_creation_callback()
 668          * directly for sub-channels, because sc_creation_callback() ->
 669          * vmbus_open() may never get the host's response to the
 670          * OPEN_CHANNEL message (the host may rescind a channel at any time,
 671          * e.g. in the case of hot removing a NIC), and vmbus_onoffer_rescind()
 672          * may not wake up the vmbus_open() as it's blocked due to a non-zero
 673          * vmbus_connection.offer_in_progress, and finally we have a deadlock.
 674          *
 675          * The above is also true for primary channels, if the related device
 676          * drivers use sync probing mode by default.
 677          *
 678          * And, usually the handling of primary channels and sub-channels can
 679          * depend on each other, so we should offload them to different
 680          * workqueues to avoid possible deadlock, e.g. in sync-probing mode,
 681          * NIC1's netvsc_subchan_work() can race with NIC2's netvsc_probe() ->
 682          * rtnl_lock(), and causes deadlock: the former gets the rtnl_lock
 683          * and waits for all the sub-channels to appear, but the latter
 684          * can't get the rtnl_lock and this blocks the handling of
 685          * sub-channels.
 686          */
 687         INIT_WORK(&newchannel->add_channel_work, vmbus_add_channel_work);
 688         wq = fnew ? vmbus_connection.handle_primary_chan_wq :
 689                     vmbus_connection.handle_sub_chan_wq;
 690         queue_work(wq, &newchannel->add_channel_work);
 691 }
 692
 693 /*
 694  * Check if CPUs used by other channels of the same device.
 695  * It should only be called by init_vp_index().
 696  */
 697 static bool hv_cpuself_used(u32 cpu, struct vmbus_channel *chn)
 698 {
 699         struct vmbus_channel *primary = chn->primary_channel;
 700         struct vmbus_channel *sc;
 701
 702         lockdep_assert_held(&vmbus_connection.channel_mutex);
 703
 704         if (!primary)
 705                 return false;
 706
 707         if (primary->target_cpu == cpu)
 708                 return true;
 709
 710         list_for_each_entry(sc, &primary->sc_list, sc_list)
 711                 if (sc != chn && sc->target_cpu == cpu)
 712                         return true;
 713
 714         return false;
 715 }
 716
 717 /*
 718  * We use this state to statically distribute the channel interrupt load.
 719  */
 720 static int next_numa_node_id;
 721
 722 /*
 723  * We can statically distribute the incoming channel interrupt load
 724  * by binding a channel to VCPU.
 725  *
 726  * For non-performance critical channels we assign the VMBUS_CONNECT_CPU.
 727  * Performance critical channels will be distributed evenly among all
 728  * the available NUMA nodes.  Once the node is assigned, we will assign
 729  * the CPU based on a simple round robin scheme.
 730  */
 731 static void init_vp_index(struct vmbus_channel *channel)
 732 {
 733         bool perf_chn = hv_is_perf_channel(channel);
 734         u32 i, ncpu = num_online_cpus();
 735         cpumask_var_t available_mask;
 736         struct cpumask *allocated_mask;
 737         const struct cpumask *hk_mask = housekeeping_cpumask(HK_TYPE_MANAGED_IRQ);
 738         u32 target_cpu;
 739         int numa_node;
 740
 741         if (!perf_chn ||
 742             !alloc_cpumask_var(&available_mask, GFP_KERNEL) ||
 743             cpumask_empty(hk_mask)) {
 744                 /*
 745                  * If the channel is not a performance critical
 746                  * channel, bind it to VMBUS_CONNECT_CPU.
 747                  * In case alloc_cpumask_var() fails, bind it to
 748                  * VMBUS_CONNECT_CPU.
 749                  * If all the cpus are isolated, bind it to
 750                  * VMBUS_CONNECT_CPU.
 751                  */
 752                 channel->target_cpu = VMBUS_CONNECT_CPU;
 753                 if (perf_chn)
 754                         hv_set_allocated_cpu(VMBUS_CONNECT_CPU);
 755                 return;
 756         }
 757
 758         for (i = 1; i <= ncpu + 1; i++) {
 759                 while (true) {
 760                         numa_node = next_numa_node_id++;
 761                         if (numa_node == nr_node_ids) {
 762                                 next_numa_node_id = 0;
 763                                 continue;
 764                         }
 765                         if (cpumask_empty(cpumask_of_node(numa_node)))
 766                                 continue;
 767                         break;
 768                 }
 769                 allocated_mask = &hv_context.hv_numa_map[numa_node];
 770
 771 retry:
 772                 cpumask_xor(available_mask, allocated_mask, cpumask_of_node(numa_node));
 773                 cpumask_and(available_mask, available_mask, hk_mask);
 774
 775                 if (cpumask_empty(available_mask)) {
 776                         /*
 777                          * We have cycled through all the CPUs in the node;
 778                          * reset the allocated map.
 779                          */
 780                         cpumask_clear(allocated_mask);
 781                         goto retry;
 782                 }
 783
 784                 target_cpu = cpumask_first(available_mask);
 785                 cpumask_set_cpu(target_cpu, allocated_mask);
 786
 787                 if (channel->offermsg.offer.sub_channel_index >= ncpu ||
 788                     i > ncpu || !hv_cpuself_used(target_cpu, channel))
 789                         break;
 790         }
 791
 792         channel->target_cpu = target_cpu;
 793
 794         free_cpumask_var(available_mask);
 795 }
 796
 797 #define UNLOAD_DELAY_UNIT_MS    10              /* 10 milliseconds */
 798 #define UNLOAD_WAIT_MS          (100*1000)      /* 100 seconds */
 799 #define UNLOAD_WAIT_LOOPS       (UNLOAD_WAIT_MS/UNLOAD_DELAY_UNIT_MS)
 800 #define UNLOAD_MSG_MS           (5*1000)        /* Every 5 seconds */
 801 #define UNLOAD_MSG_LOOPS        (UNLOAD_MSG_MS/UNLOAD_DELAY_UNIT_MS)
 802
 803 static void vmbus_wait_for_unload(void)
 804 {
 805         int cpu;
 806         void *page_addr;
 807         struct hv_message *msg;
 808         struct vmbus_channel_message_header *hdr;
 809         u32 message_type, i;
 810
 811         /*
 812          * CHANNELMSG_UNLOAD_RESPONSE is always delivered to the CPU which was
 813          * used for initial contact or to CPU0 depending on host version. When
 814          * we're crashing on a different CPU let's hope that IRQ handler on
 815          * the cpu which receives CHANNELMSG_UNLOAD_RESPONSE is still
 816          * functional and vmbus_unload_response() will complete
 817          * vmbus_connection.unload_event. If not, the last thing we can do is
 818          * read message pages for all CPUs directly.
 819          *
 820          * Wait up to 100 seconds since an Azure host must writeback any dirty
 821          * data in its disk cache before the VMbus UNLOAD request will
 822          * complete. This flushing has been empirically observed to take up
 823          * to 50 seconds in cases with a lot of dirty data, so allow additional
 824          * leeway and for inaccuracies in mdelay(). But eventually time out so
 825          * that the panic path can't get hung forever in case the response
 826          * message isn't seen.
 827          */
 828         for (i = 1; i <= UNLOAD_WAIT_LOOPS; i++) {
 829                 if (completion_done(&vmbus_connection.unload_event))
 830                         goto completed;
 831
 832                 for_each_online_cpu(cpu) {
 833                         struct hv_per_cpu_context *hv_cpu
 834                                 = per_cpu_ptr(hv_context.cpu_context, cpu);
 835
 836                         page_addr = hv_cpu->synic_message_page;
 837                         msg = (struct hv_message *)page_addr
 838                                 + VMBUS_MESSAGE_SINT;
 839
 840                         message_type = READ_ONCE(msg->header.message_type);
 841                         if (message_type == HVMSG_NONE)
 842                                 continue;
 843
 844                         hdr = (struct vmbus_channel_message_header *)
 845                                 msg->u.payload;
 846
 847                         if (hdr->msgtype == CHANNELMSG_UNLOAD_RESPONSE)
 848                                 complete(&vmbus_connection.unload_event);
 849
 850                         vmbus_signal_eom(msg, message_type);
 851                 }
 852
 853                 /*
 854                  * Give a notice periodically so someone watching the
 855                  * serial output won't think it is completely hung.
 856                  */
 857                 if (!(i % UNLOAD_MSG_LOOPS))
 858                         pr_notice("Waiting for VMBus UNLOAD to complete\n");
 859
 860                 mdelay(UNLOAD_DELAY_UNIT_MS);
 861         }
 862         pr_err("Continuing even though VMBus UNLOAD did not complete\n");
 863
 864 completed:
 865         /*
 866          * We're crashing and already got the UNLOAD_RESPONSE, cleanup all
 867          * maybe-pending messages on all CPUs to be able to receive new
 868          * messages after we reconnect.
 869          */
 870         for_each_online_cpu(cpu) {
 871                 struct hv_per_cpu_context *hv_cpu
 872                         = per_cpu_ptr(hv_context.cpu_context, cpu);
 873
 874                 page_addr = hv_cpu->synic_message_page;
 875                 msg = (struct hv_message *)page_addr + VMBUS_MESSAGE_SINT;
 876                 msg->header.message_type = HVMSG_NONE;
 877         }
 878 }
 879
 880 /*
 881  * vmbus_unload_response - Handler for the unload response.
 882  */
 883 static void vmbus_unload_response(struct vmbus_channel_message_header *hdr)
 884 {
 885         /*
 886          * This is a global event; just wakeup the waiting thread.
 887          * Once we successfully unload, we can cleanup the monitor state.
 888          *
 889          * NB.  A malicious or compromised Hyper-V could send a spurious
 890          * message of type CHANNELMSG_UNLOAD_RESPONSE, and trigger a call
 891          * of the complete() below.  Make sure that unload_event has been
 892          * initialized by the time this complete() is executed.
 893          */
 894         complete(&vmbus_connection.unload_event);
 895 }
 896
 897 void vmbus_initiate_unload(bool crash)
 898 {
 899         struct vmbus_channel_message_header hdr;
 900
 901         if (xchg(&vmbus_connection.conn_state, DISCONNECTED) == DISCONNECTED)
 902                 return;
 903
 904         /* Pre-Win2012R2 hosts don't support reconnect */
 905         if (vmbus_proto_version < VERSION_WIN8_1)
 906                 return;
 907
 908         reinit_completion(&vmbus_connection.unload_event);
 909         memset(&hdr, 0, sizeof(struct vmbus_channel_message_header));
 910         hdr.msgtype = CHANNELMSG_UNLOAD;
 911         vmbus_post_msg(&hdr, sizeof(struct vmbus_channel_message_header),
 912                        !crash);
 913
 914         /*
 915          * vmbus_initiate_unload() is also called on crash and the crash can be
 916          * happening in an interrupt context, where scheduling is impossible.
 917          */
 918         if (!crash)
 919                 wait_for_completion(&vmbus_connection.unload_event);
 920         else
 921                 vmbus_wait_for_unload();
 922 }
 923
 924 static void check_ready_for_resume_event(void)
 925 {
 926         /*
 927          * If all the old primary channels have been fixed up, then it's safe
 928          * to resume.
 929          */
 930         if (atomic_dec_and_test(&vmbus_connection.nr_chan_fixup_on_resume))
 931                 complete(&vmbus_connection.ready_for_resume_event);
 932 }
 933
 934 static void vmbus_setup_channel_state(struct vmbus_channel *channel,
 935                                       struct vmbus_channel_offer_channel *offer)
 936 {
 937         /*
 938          * Setup state for signalling the host.
 939          */
 940         channel->sig_event = VMBUS_EVENT_CONNECTION_ID;
 941
 942         channel->is_dedicated_interrupt =
 943                         (offer->is_dedicated_interrupt != 0);
 944         channel->sig_event = offer->connection_id;
 945
 946         memcpy(&channel->offermsg, offer,
 947                sizeof(struct vmbus_channel_offer_channel));
 948         channel->monitor_grp = (u8)offer->monitorid / 32;
 949         channel->monitor_bit = (u8)offer->monitorid % 32;
 950         channel->device_id = hv_get_dev_type(channel);
 951 }
 952
 953 /*
 954  * find_primary_channel_by_offer - Get the channel object given the new offer.
 955  * This is only used in the resume path of hibernation.
 956  */
 957 static struct vmbus_channel *
 958 find_primary_channel_by_offer(const struct vmbus_channel_offer_channel *offer)
 959 {
 960         struct vmbus_channel *channel = NULL, *iter;
 961         const guid_t *inst1, *inst2;
 962
 963         /* Ignore sub-channel offers. */
 964         if (offer->offer.sub_channel_index != 0)
 965                 return NULL;
 966
 967         mutex_lock(&vmbus_connection.channel_mutex);
 968
 969         list_for_each_entry(iter, &vmbus_connection.chn_list, listentry) {
 970                 inst1 = &iter->offermsg.offer.if_instance;
 971                 inst2 = &offer->offer.if_instance;
 972
 973                 if (guid_equal(inst1, inst2)) {
 974                         channel = iter;
 975                         break;
 976                 }
 977         }
 978
 979         mutex_unlock(&vmbus_connection.channel_mutex);
 980
 981         return channel;
 982 }
 983
 984 static bool vmbus_is_valid_offer(const struct vmbus_channel_offer_channel *offer)
 985 {
 986         const guid_t *guid = &offer->offer.if_type;
 987         u16 i;
 988
 989         if (!hv_is_isolation_supported())
 990                 return true;
 991
 992         if (is_hvsock_offer(offer))
 993                 return true;
 994
 995         for (i = 0; i < ARRAY_SIZE(vmbus_devs); i++) {
 996                 if (guid_equal(guid, &vmbus_devs[i].guid))
 997                         return vmbus_devs[i].allowed_in_isolated;
 998         }
 999         return false;
1000 }
1001
1002 /*
1003  * vmbus_onoffer - Handler for channel offers from vmbus in parent partition.
1004  *
1005  */
1006 static void vmbus_onoffer(struct vmbus_channel_message_header *hdr)
1007 {
1008         struct vmbus_channel_offer_channel *offer;
1009         struct vmbus_channel *oldchannel, *newchannel;
1010         size_t offer_sz;
1011
1012         offer = (struct vmbus_channel_offer_channel *)hdr;
1013
1014         trace_vmbus_onoffer(offer);
1015
1016         if (!vmbus_is_valid_offer(offer)) {
1017                 pr_err_ratelimited("Invalid offer %d from the host supporting isolation\n",
1018                                    offer->child_relid);
1019                 atomic_dec(&vmbus_connection.offer_in_progress);
1020                 return;
1021         }
1022
1023         oldchannel = find_primary_channel_by_offer(offer);
1024
1025         if (oldchannel != NULL) {
1026                 /*
1027                  * We're resuming from hibernation: all the sub-channel and
1028                  * hv_sock channels we had before the hibernation should have
1029                  * been cleaned up, and now we must be seeing a re-offered
1030                  * primary channel that we had before the hibernation.
1031                  */
1032
1033                 /*
1034                  * { Initially: channel relid = INVALID_RELID,
1035                  *              channels[valid_relid] = NULL }
1036                  *
1037                  * CPU1                                 CPU2
1038                  *
1039                  * [vmbus_onoffer()]                    [vmbus_device_release()]
1040                  *
1041                  * LOCK channel_mutex                   LOCK channel_mutex
1042                  * STORE channel relid = valid_relid    LOAD r1 = channel relid
1043                  * MAP_RELID channel                    if (r1 != INVALID_RELID)
1044                  * UNLOCK channel_mutex                   UNMAP_RELID channel
1045                  *                                      UNLOCK channel_mutex
1046                  *
1047                  * Forbids: r1 == valid_relid &&
1048                  *              channels[valid_relid] == channel
1049                  *
1050                  * Note.  r1 can be INVALID_RELID only for an hv_sock channel.
1051                  * None of the hv_sock channels which were present before the
1052                  * suspend are re-offered upon the resume.  See the WARN_ON()
1053                  * in hv_process_channel_removal().
1054                  */
1055                 mutex_lock(&vmbus_connection.channel_mutex);
1056
1057                 atomic_dec(&vmbus_connection.offer_in_progress);
1058
1059                 WARN_ON(oldchannel->offermsg.child_relid != INVALID_RELID);
1060                 /* Fix up the relid. */
1061                 oldchannel->offermsg.child_relid = offer->child_relid;
1062
1063                 offer_sz = sizeof(*offer);
1064                 if (memcmp(offer, &oldchannel->offermsg, offer_sz) != 0) {
1065                         /*
1066                          * This is not an error, since the host can also change
1067                          * the other field(s) of the offer, e.g. on WS RS5
1068                          * (Build 17763), the offer->connection_id of the
1069                          * Mellanox VF vmbus device can change when the host
1070                          * reoffers the device upon resume.
1071                          */
1072                         pr_debug("vmbus offer changed: relid=%d\n",
1073                                  offer->child_relid);
1074
1075                         print_hex_dump_debug("Old vmbus offer: ",
1076                                              DUMP_PREFIX_OFFSET, 16, 4,
1077                                              &oldchannel->offermsg, offer_sz,
1078                                              false);
1079                         print_hex_dump_debug("New vmbus offer: ",
1080                                              DUMP_PREFIX_OFFSET, 16, 4,
1081                                              offer, offer_sz, false);
1082
1083                         /* Fix up the old channel. */
1084                         vmbus_setup_channel_state(oldchannel, offer);
1085                 }
1086
1087                 /* Add the channel back to the array of channels. */
1088                 vmbus_channel_map_relid(oldchannel);
1089                 check_ready_for_resume_event();
1090
1091                 mutex_unlock(&vmbus_connection.channel_mutex);
1092                 return;
1093         }
1094
1095         /* Allocate the channel object and save this offer. */
1096         newchannel = alloc_channel();
1097         if (!newchannel) {
1098                 vmbus_release_relid(offer->child_relid);
1099                 atomic_dec(&vmbus_connection.offer_in_progress);
1100                 pr_err("Unable to allocate channel object\n");
1101                 return;
1102         }
1103
1104         vmbus_setup_channel_state(newchannel, offer);
1105
1106         vmbus_process_offer(newchannel);
1107 }
1108
1109 static void check_ready_for_suspend_event(void)
1110 {
1111         /*
1112          * If all the sub-channels or hv_sock channels have been cleaned up,
1113          * then it's safe to suspend.
1114          */
1115         if (atomic_dec_and_test(&vmbus_connection.nr_chan_close_on_suspend))
1116                 complete(&vmbus_connection.ready_for_suspend_event);
1117 }
1118
1119 /*
1120  * vmbus_onoffer_rescind - Rescind offer handler.
1121  *
1122  * We queue a work item to process this offer synchronously
1123  */
1124 static void vmbus_onoffer_rescind(struct vmbus_channel_message_header *hdr)
1125 {
1126         struct vmbus_channel_rescind_offer *rescind;
1127         struct vmbus_channel *channel;
1128         struct device *dev;
1129         bool clean_up_chan_for_suspend;
1130
1131         rescind = (struct vmbus_channel_rescind_offer *)hdr;
1132
1133         trace_vmbus_onoffer_rescind(rescind);
1134
1135         /*
1136          * The offer msg and the corresponding rescind msg
1137          * from the host are guranteed to be ordered -
1138          * offer comes in first and then the rescind.
1139          * Since we process these events in work elements,
1140          * and with preemption, we may end up processing
1141          * the events out of order.  We rely on the synchronization
1142          * provided by offer_in_progress and by channel_mutex for
1143          * ordering these events:
1144          *
1145          * { Initially: offer_in_progress = 1 }
1146          *
1147          * CPU1                         CPU2
1148          *
1149          * [vmbus_onoffer()]            [vmbus_onoffer_rescind()]
1150          *
1151          * LOCK channel_mutex           WAIT_ON offer_in_progress == 0
1152          * DECREMENT offer_in_progress  LOCK channel_mutex
1153          * STORE channels[]             LOAD channels[]
1154          * UNLOCK channel_mutex         UNLOCK channel_mutex
1155          *
1156          * Forbids: CPU2's LOAD from *not* seeing CPU1's STORE
1157          */
1158
1159         while (atomic_read(&vmbus_connection.offer_in_progress) != 0) {
1160                 /*
1161                  * We wait here until any channel offer is currently
1162                  * being processed.
1163                  */
1164                 msleep(1);
1165         }
1166
1167         mutex_lock(&vmbus_connection.channel_mutex);
1168         channel = relid2channel(rescind->child_relid);
1169         if (channel != NULL) {
1170                 /*
1171                  * Guarantee that no other instance of vmbus_onoffer_rescind()
1172                  * has got a reference to the channel object.  Synchronize on
1173                  * &vmbus_connection.channel_mutex.
1174                  */
1175                 if (channel->rescind_ref) {
1176                         mutex_unlock(&vmbus_connection.channel_mutex);
1177                         return;
1178                 }
1179                 channel->rescind_ref = true;
1180         }
1181         mutex_unlock(&vmbus_connection.channel_mutex);
1182
1183         if (channel == NULL) {
1184                 /*
1185                  * We failed in processing the offer message;
1186                  * we would have cleaned up the relid in that
1187                  * failure path.
1188                  */
1189                 return;
1190         }
1191
1192         clean_up_chan_for_suspend = is_hvsock_channel(channel) ||
1193                                     is_sub_channel(channel);
1194         /*
1195          * Before setting channel->rescind in vmbus_rescind_cleanup(), we
1196          * should make sure the channel callback is not running any more.
1197          */
1198         vmbus_reset_channel_cb(channel);
1199
1200         /*
1201          * Now wait for offer handling to complete.
1202          */
1203         vmbus_rescind_cleanup(channel);
1204         while (READ_ONCE(channel->probe_done) == false) {
1205                 /*
1206                  * We wait here until any channel offer is currently
1207                  * being processed.
1208                  */
1209                 msleep(1);
1210         }
1211
1212         /*
1213          * At this point, the rescind handling can proceed safely.
1214          */
1215
1216         if (channel->device_obj) {
1217                 if (channel->chn_rescind_callback) {
1218                         channel->chn_rescind_callback(channel);
1219
1220                         if (clean_up_chan_for_suspend)
1221                                 check_ready_for_suspend_event();
1222
1223                         return;
1224                 }
1225                 /*
1226                  * We will have to unregister this device from the
1227                  * driver core.
1228                  */
1229                 dev = get_device(&channel->device_obj->device);
1230                 if (dev) {
1231                         vmbus_device_unregister(channel->device_obj);
1232                         put_device(dev);
1233                 }
1234         } else if (channel->primary_channel != NULL) {
1235                 /*
1236                  * Sub-channel is being rescinded. Following is the channel
1237                  * close sequence when initiated from the driveri (refer to
1238                  * vmbus_close() for details):
1239                  * 1. Close all sub-channels first
1240                  * 2. Then close the primary channel.
1241                  */
1242                 mutex_lock(&vmbus_connection.channel_mutex);
1243                 if (channel->state == CHANNEL_OPEN_STATE) {
1244                         /*
1245                          * The channel is currently not open;
1246                          * it is safe for us to cleanup the channel.
1247                          */
1248                         hv_process_channel_removal(channel);
1249                 } else {
1250                         complete(&channel->rescind_event);
1251                 }
1252                 mutex_unlock(&vmbus_connection.channel_mutex);
1253         }
1254
1255         /* The "channel" may have been freed. Do not access it any longer. */
1256
1257         if (clean_up_chan_for_suspend)
1258                 check_ready_for_suspend_event();
1259 }
1260
1261 void vmbus_hvsock_device_unregister(struct vmbus_channel *channel)
1262 {
1263         BUG_ON(!is_hvsock_channel(channel));
1264
1265         /* We always get a rescind msg when a connection is closed. */
1266         while (!READ_ONCE(channel->probe_done) || !READ_ONCE(channel->rescind))
1267                 msleep(1);
1268
1269         vmbus_device_unregister(channel->device_obj);
1270 }
1271 EXPORT_SYMBOL_GPL(vmbus_hvsock_device_unregister);
1272
1273
1274 /*
1275  * vmbus_onoffers_delivered -
1276  * This is invoked when all offers have been delivered.
1277  *
1278  * Nothing to do here.
1279  */
1280 static void vmbus_onoffers_delivered(
1281                         struct vmbus_channel_message_header *hdr)
1282 {
1283 }
1284
1285 /*
1286  * vmbus_onopen_result - Open result handler.
1287  *
1288  * This is invoked when we received a response to our channel open request.
1289  * Find the matching request, copy the response and signal the requesting
1290  * thread.
1291  */
1292 static void vmbus_onopen_result(struct vmbus_channel_message_header *hdr)
1293 {
1294         struct vmbus_channel_open_result *result;
1295         struct vmbus_channel_msginfo *msginfo;
1296         struct vmbus_channel_message_header *requestheader;
1297         struct vmbus_channel_open_channel *openmsg;
1298         unsigned long flags;
1299
1300         result = (struct vmbus_channel_open_result *)hdr;
1301
1302         trace_vmbus_onopen_result(result);
1303
1304         /*
1305          * Find the open msg, copy the result and signal/unblock the wait event
1306          */
1307         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1308
1309         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1310                                 msglistentry) {
1311                 requestheader =
1312                         (struct vmbus_channel_message_header *)msginfo->msg;
1313
1314                 if (requestheader->msgtype == CHANNELMSG_OPENCHANNEL) {
1315                         openmsg =
1316                         (struct vmbus_channel_open_channel *)msginfo->msg;
1317                         if (openmsg->child_relid == result->child_relid &&
1318                             openmsg->openid == result->openid) {
1319                                 memcpy(&msginfo->response.open_result,
1320                                        result,
1321                                        sizeof(
1322                                         struct vmbus_channel_open_result));
1323                                 complete(&msginfo->waitevent);
1324                                 break;
1325                         }
1326                 }
1327         }
1328         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1329 }
1330
1331 /*
1332  * vmbus_ongpadl_created - GPADL created handler.
1333  *
1334  * This is invoked when we received a response to our gpadl create request.
1335  * Find the matching request, copy the response and signal the requesting
1336  * thread.
1337  */
1338 static void vmbus_ongpadl_created(struct vmbus_channel_message_header *hdr)
1339 {
1340         struct vmbus_channel_gpadl_created *gpadlcreated;
1341         struct vmbus_channel_msginfo *msginfo;
1342         struct vmbus_channel_message_header *requestheader;
1343         struct vmbus_channel_gpadl_header *gpadlheader;
1344         unsigned long flags;
1345
1346         gpadlcreated = (struct vmbus_channel_gpadl_created *)hdr;
1347
1348         trace_vmbus_ongpadl_created(gpadlcreated);
1349
1350         /*
1351          * Find the establish msg, copy the result and signal/unblock the wait
1352          * event
1353          */
1354         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1355
1356         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1357                                 msglistentry) {
1358                 requestheader =
1359                         (struct vmbus_channel_message_header *)msginfo->msg;
1360
1361                 if (requestheader->msgtype == CHANNELMSG_GPADL_HEADER) {
1362                         gpadlheader =
1363                         (struct vmbus_channel_gpadl_header *)requestheader;
1364
1365                         if ((gpadlcreated->child_relid ==
1366                              gpadlheader->child_relid) &&
1367                             (gpadlcreated->gpadl == gpadlheader->gpadl)) {
1368                                 memcpy(&msginfo->response.gpadl_created,
1369                                        gpadlcreated,
1370                                        sizeof(
1371                                         struct vmbus_channel_gpadl_created));
1372                                 complete(&msginfo->waitevent);
1373                                 break;
1374                         }
1375                 }
1376         }
1377         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1378 }
1379
1380 /*
1381  * vmbus_onmodifychannel_response - Modify Channel response handler.
1382  *
1383  * This is invoked when we received a response to our channel modify request.
1384  * Find the matching request, copy the response and signal the requesting thread.
1385  */
1386 static void vmbus_onmodifychannel_response(struct vmbus_channel_message_header *hdr)
1387 {
1388         struct vmbus_channel_modifychannel_response *response;
1389         struct vmbus_channel_msginfo *msginfo;
1390         unsigned long flags;
1391
1392         response = (struct vmbus_channel_modifychannel_response *)hdr;
1393
1394         trace_vmbus_onmodifychannel_response(response);
1395
1396         /*
1397          * Find the modify msg, copy the response and signal/unblock the wait event.
1398          */
1399         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1400
1401         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list, msglistentry) {
1402                 struct vmbus_channel_message_header *responseheader =
1403                                 (struct vmbus_channel_message_header *)msginfo->msg;
1404
1405                 if (responseheader->msgtype == CHANNELMSG_MODIFYCHANNEL) {
1406                         struct vmbus_channel_modifychannel *modifymsg;
1407
1408                         modifymsg = (struct vmbus_channel_modifychannel *)msginfo->msg;
1409                         if (modifymsg->child_relid == response->child_relid) {
1410                                 memcpy(&msginfo->response.modify_response, response,
1411                                        sizeof(*response));
1412                                 complete(&msginfo->waitevent);
1413                                 break;
1414                         }
1415                 }
1416         }
1417         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1418 }
1419
1420 /*
1421  * vmbus_ongpadl_torndown - GPADL torndown handler.
1422  *
1423  * This is invoked when we received a response to our gpadl teardown request.
1424  * Find the matching request, copy the response and signal the requesting
1425  * thread.
1426  */
1427 static void vmbus_ongpadl_torndown(
1428                         struct vmbus_channel_message_header *hdr)
1429 {
1430         struct vmbus_channel_gpadl_torndown *gpadl_torndown;
1431         struct vmbus_channel_msginfo *msginfo;
1432         struct vmbus_channel_message_header *requestheader;
1433         struct vmbus_channel_gpadl_teardown *gpadl_teardown;
1434         unsigned long flags;
1435
1436         gpadl_torndown = (struct vmbus_channel_gpadl_torndown *)hdr;
1437
1438         trace_vmbus_ongpadl_torndown(gpadl_torndown);
1439
1440         /*
1441          * Find the open msg, copy the result and signal/unblock the wait event
1442          */
1443         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1444
1445         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1446                                 msglistentry) {
1447                 requestheader =
1448                         (struct vmbus_channel_message_header *)msginfo->msg;
1449
1450                 if (requestheader->msgtype == CHANNELMSG_GPADL_TEARDOWN) {
1451                         gpadl_teardown =
1452                         (struct vmbus_channel_gpadl_teardown *)requestheader;
1453
1454                         if (gpadl_torndown->gpadl == gpadl_teardown->gpadl) {
1455                                 memcpy(&msginfo->response.gpadl_torndown,
1456                                        gpadl_torndown,
1457                                        sizeof(
1458                                         struct vmbus_channel_gpadl_torndown));
1459                                 complete(&msginfo->waitevent);
1460                                 break;
1461                         }
1462                 }
1463         }
1464         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1465 }
1466
1467 /*
1468  * vmbus_onversion_response - Version response handler
1469  *
1470  * This is invoked when we received a response to our initiate contact request.
1471  * Find the matching request, copy the response and signal the requesting
1472  * thread.
1473  */
1474 static void vmbus_onversion_response(
1475                 struct vmbus_channel_message_header *hdr)
1476 {
1477         struct vmbus_channel_msginfo *msginfo;
1478         struct vmbus_channel_message_header *requestheader;
1479         struct vmbus_channel_version_response *version_response;
1480         unsigned long flags;
1481
1482         version_response = (struct vmbus_channel_version_response *)hdr;
1483
1484         trace_vmbus_onversion_response(version_response);
1485
1486         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1487
1488         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1489                                 msglistentry) {
1490                 requestheader =
1491                         (struct vmbus_channel_message_header *)msginfo->msg;
1492
1493                 if (requestheader->msgtype ==
1494                     CHANNELMSG_INITIATE_CONTACT) {
1495                         memcpy(&msginfo->response.version_response,
1496                               version_response,
1497                               sizeof(struct vmbus_channel_version_response));
1498                         complete(&msginfo->waitevent);
1499                 }
1500         }
1501         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1502 }
1503
1504 /* Channel message dispatch table */
1505 const struct vmbus_channel_message_table_entry
1506 channel_message_table[CHANNELMSG_COUNT] = {
1507         { CHANNELMSG_INVALID,                   0, NULL, 0},
1508         { CHANNELMSG_OFFERCHANNEL,              0, vmbus_onoffer,
1509                 sizeof(struct vmbus_channel_offer_channel)},
1510         { CHANNELMSG_RESCIND_CHANNELOFFER,      0, vmbus_onoffer_rescind,
1511                 sizeof(struct vmbus_channel_rescind_offer) },
1512         { CHANNELMSG_REQUESTOFFERS,             0, NULL, 0},
1513         { CHANNELMSG_ALLOFFERS_DELIVERED,       1, vmbus_onoffers_delivered, 0},
1514         { CHANNELMSG_OPENCHANNEL,               0, NULL, 0},
1515         { CHANNELMSG_OPENCHANNEL_RESULT,        1, vmbus_onopen_result,
1516                 sizeof(struct vmbus_channel_open_result)},
1517         { CHANNELMSG_CLOSECHANNEL,              0, NULL, 0},
1518         { CHANNELMSG_GPADL_HEADER,              0, NULL, 0},
1519         { CHANNELMSG_GPADL_BODY,                0, NULL, 0},
1520         { CHANNELMSG_GPADL_CREATED,             1, vmbus_ongpadl_created,
1521                 sizeof(struct vmbus_channel_gpadl_created)},
1522         { CHANNELMSG_GPADL_TEARDOWN,            0, NULL, 0},
1523         { CHANNELMSG_GPADL_TORNDOWN,            1, vmbus_ongpadl_torndown,
1524                 sizeof(struct vmbus_channel_gpadl_torndown) },
1525         { CHANNELMSG_RELID_RELEASED,            0, NULL, 0},
1526         { CHANNELMSG_INITIATE_CONTACT,          0, NULL, 0},
1527         { CHANNELMSG_VERSION_RESPONSE,          1, vmbus_onversion_response,
1528                 sizeof(struct vmbus_channel_version_response)},
1529         { CHANNELMSG_UNLOAD,                    0, NULL, 0},
1530         { CHANNELMSG_UNLOAD_RESPONSE,           1, vmbus_unload_response, 0},
1531         { CHANNELMSG_18,                        0, NULL, 0},
1532         { CHANNELMSG_19,                        0, NULL, 0},
1533         { CHANNELMSG_20,                        0, NULL, 0},
1534         { CHANNELMSG_TL_CONNECT_REQUEST,        0, NULL, 0},
1535         { CHANNELMSG_MODIFYCHANNEL,             0, NULL, 0},
1536         { CHANNELMSG_TL_CONNECT_RESULT,         0, NULL, 0},
1537         { CHANNELMSG_MODIFYCHANNEL_RESPONSE,    1, vmbus_onmodifychannel_response,
1538                 sizeof(struct vmbus_channel_modifychannel_response)},
1539 };
1540
1541 /*
1542  * vmbus_onmessage - Handler for channel protocol messages.
1543  *
1544  * This is invoked in the vmbus worker thread context.
1545  */
1546 void vmbus_onmessage(struct vmbus_channel_message_header *hdr)
1547 {
1548         trace_vmbus_on_message(hdr);
1549
1550         /*
1551          * vmbus_on_msg_dpc() makes sure the hdr->msgtype here can not go
1552          * out of bound and the message_handler pointer can not be NULL.
1553          */
1554         channel_message_table[hdr->msgtype].message_handler(hdr);
1555 }
1556
1557 /*
1558  * vmbus_request_offers - Send a request to get all our pending offers.
1559  */
1560 int vmbus_request_offers(void)
1561 {
1562         struct vmbus_channel_message_header *msg;
1563         struct vmbus_channel_msginfo *msginfo;
1564         int ret;
1565
1566         msginfo = kzalloc(sizeof(*msginfo) +
1567                           sizeof(struct vmbus_channel_message_header),
1568                           GFP_KERNEL);
1569         if (!msginfo)
1570                 return -ENOMEM;
1571
1572         msg = (struct vmbus_channel_message_header *)msginfo->msg;
1573
1574         msg->msgtype = CHANNELMSG_REQUESTOFFERS;
1575
1576         ret = vmbus_post_msg(msg, sizeof(struct vmbus_channel_message_header),
1577                              true);
1578
1579         trace_vmbus_request_offers(ret);
1580
1581         if (ret != 0) {
1582                 pr_err("Unable to request offers - %d\n", ret);
1583
1584                 goto cleanup;
1585         }
1586
1587 cleanup:
1588         kfree(msginfo);
1589
1590         return ret;
1591 }
1592
1593 void vmbus_set_sc_create_callback(struct vmbus_channel *primary_channel,
1594                                 void (*sc_cr_cb)(struct vmbus_channel *new_sc))
1595 {
1596         primary_channel->sc_creation_callback = sc_cr_cb;
1597 }
1598 EXPORT_SYMBOL_GPL(vmbus_set_sc_create_callback);
1599
1600 void vmbus_set_chn_rescind_callback(struct vmbus_channel *channel,
1601                 void (*chn_rescind_cb)(struct vmbus_channel *))
1602 {
1603         channel->chn_rescind_callback = chn_rescind_cb;
1604 }
1605 EXPORT_SYMBOL_GPL(vmbus_set_chn_rescind_callback);