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 159  * The rescinded channel may be blocked waiting for a response from the host;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