drivers/hv/channel_mgmt.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (c) 2009, Microsoft Corporation.
   4  *
   5  * Authors:
   6  *   Haiyang Zhang <haiyangz@microsoft.com>
   7  *   Hank Janssen  <hjanssen@microsoft.com>
   8  */
   9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  10
  11 #include <linux/kernel.h>
  12 #include <linux/interrupt.h>
  13 #include <linux/sched.h>
  14 #include <linux/wait.h>
  15 #include <linux/mm.h>
  16 #include <linux/slab.h>
  17 #include <linux/list.h>
  18 #include <linux/module.h>
  19 #include <linux/completion.h>
  20 #include <linux/delay.h>
  21 #include <linux/cpu.h>
  22 #include <linux/hyperv.h>
  23 #include <asm/mshyperv.h>
  24
  25 #include "hyperv_vmbus.h"
  26
  27 static void init_vp_index(struct vmbus_channel *channel);
  28
  29 const struct vmbus_device vmbus_devs[] = {
  30         /* IDE */
  31         { .dev_type = HV_IDE,
  32           HV_IDE_GUID,
  33           .perf_device = true,
  34           .allowed_in_isolated = false,
  35         },
  36
  37         /* SCSI */
  38         { .dev_type = HV_SCSI,
  39           HV_SCSI_GUID,
  40           .perf_device = true,
  41           .allowed_in_isolated = true,
  42         },
  43
  44         /* Fibre Channel */
  45         { .dev_type = HV_FC,
  46           HV_SYNTHFC_GUID,
  47           .perf_device = true,
  48           .allowed_in_isolated = false,
  49         },
  50
  51         /* Synthetic NIC */
  52         { .dev_type = HV_NIC,
  53           HV_NIC_GUID,
  54           .perf_device = true,
  55           .allowed_in_isolated = true,
  56         },
  57
  58         /* Network Direct */
  59         { .dev_type = HV_ND,
  60           HV_ND_GUID,
  61           .perf_device = true,
  62           .allowed_in_isolated = false,
  63         },
  64
  65         /* PCIE */
  66         { .dev_type = HV_PCIE,
  67           HV_PCIE_GUID,
  68           .perf_device = false,
  69           .allowed_in_isolated = false,
  70         },
  71
  72         /* Synthetic Frame Buffer */
  73         { .dev_type = HV_FB,
  74           HV_SYNTHVID_GUID,
  75           .perf_device = false,
  76           .allowed_in_isolated = false,
  77         },
  78
  79         /* Synthetic Keyboard */
  80         { .dev_type = HV_KBD,
  81           HV_KBD_GUID,
  82           .perf_device = false,
  83           .allowed_in_isolated = false,
  84         },
  85
  86         /* Synthetic MOUSE */
  87         { .dev_type = HV_MOUSE,
  88           HV_MOUSE_GUID,
  89           .perf_device = false,
  90           .allowed_in_isolated = false,
  91         },
  92
  93         /* KVP */
  94         { .dev_type = HV_KVP,
  95           HV_KVP_GUID,
  96           .perf_device = false,
  97           .allowed_in_isolated = false,
  98         },
  99
 100         /* Time Synch */
 101         { .dev_type = HV_TS,
 102           HV_TS_GUID,
 103           .perf_device = false,
 104           .allowed_in_isolated = true,
 105         },
 106
 107         /* Heartbeat */
 108         { .dev_type = HV_HB,
 109           HV_HEART_BEAT_GUID,
 110           .perf_device = false,
 111           .allowed_in_isolated = true,
 112         },
 113
 114         /* Shutdown */
 115         { .dev_type = HV_SHUTDOWN,
 116           HV_SHUTDOWN_GUID,
 117           .perf_device = false,
 118           .allowed_in_isolated = true,
 119         },
 120
 121         /* File copy */
 122         { .dev_type = HV_FCOPY,
 123           HV_FCOPY_GUID,
 124           .perf_device = false,
 125           .allowed_in_isolated = false,
 126         },
 127
 128         /* Backup */
 129         { .dev_type = HV_BACKUP,
 130           HV_VSS_GUID,
 131           .perf_device = false,
 132           .allowed_in_isolated = false,
 133         },
 134
 135         /* Dynamic Memory */
 136         { .dev_type = HV_DM,
 137           HV_DM_GUID,
 138           .perf_device = false,
 139           .allowed_in_isolated = false,
 140         },
 141
 142         /* Unknown GUID */
 143         { .dev_type = HV_UNKNOWN,
 144           .perf_device = false,
 145           .allowed_in_isolated = false,
 146         },
 147 };
 148
 149 static const struct {
 150         guid_t guid;
 151 } vmbus_unsupported_devs[] = {
 152         { HV_AVMA1_GUID },
 153         { HV_AVMA2_GUID },
 154         { HV_RDV_GUID   },
 155 };
 156
 157 /*
 158  * The rescinded channel may be blocked waiting for a response from the host;
 159  * take care of that.
 160  */
 161 static void vmbus_rescind_cleanup(struct vmbus_channel *channel)
 162 {
 163         struct vmbus_channel_msginfo *msginfo;
 164         unsigned long flags;
 165
 166
 167         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
 168         channel->rescind = true;
 169         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
 170                                 msglistentry) {
 171
 172                 if (msginfo->waiting_channel == channel) {
 173                         complete(&msginfo->waitevent);
 174                         break;
 175                 }
 176         }
 177         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
 178 }
 179
 180 static bool is_unsupported_vmbus_devs(const guid_t *guid)
 181 {
 182         int i;
 183
 184         for (i = 0; i < ARRAY_SIZE(vmbus_unsupported_devs); i++)
 185                 if (guid_equal(guid, &vmbus_unsupported_devs[i].guid))
 186                         return true;
 187         return false;
 188 }
 189
 190 static u16 hv_get_dev_type(const struct vmbus_channel *channel)
 191 {
 192         const guid_t *guid = &channel->offermsg.offer.if_type;
 193         u16 i;
 194
 195         if (is_hvsock_channel(channel) || is_unsupported_vmbus_devs(guid))
 196                 return HV_UNKNOWN;
 197
 198         for (i = HV_IDE; i < HV_UNKNOWN; i++) {
 199                 if (guid_equal(guid, &vmbus_devs[i].guid))
 200                         return i;
 201         }
 202         pr_info("Unknown GUID: %pUl\n", guid);
 203         return i;
 204 }
 205
 206 /**
 207  * vmbus_prep_negotiate_resp() - Create default response for Negotiate message
 208  * @icmsghdrp: Pointer to msg header structure
 209  * @buf: Raw buffer channel data
 210  * @buflen: Length of the raw buffer channel data.
 211  * @fw_version: The framework versions we can support.
 212  * @fw_vercnt: The size of @fw_version.
 213  * @srv_version: The service versions we can support.
 214  * @srv_vercnt: The size of @srv_version.
 215  * @nego_fw_version: The selected framework version.
 216  * @nego_srv_version: The selected service version.
 217  *
 218  * Note: Versions are given in decreasing order.
 219  *
 220  * Set up and fill in default negotiate response message.
 221  * Mainly used by Hyper-V drivers.
 222  */
 223 bool vmbus_prep_negotiate_resp(struct icmsg_hdr *icmsghdrp, u8 *buf,
 224                                 u32 buflen, const int *fw_version, int fw_vercnt,
 225                                 const int *srv_version, int srv_vercnt,
 226                                 int *nego_fw_version, int *nego_srv_version)
 227 {
 228         int icframe_major, icframe_minor;
 229         int icmsg_major, icmsg_minor;
 230         int fw_major, fw_minor;
 231         int srv_major, srv_minor;
 232         int i, j;
 233         bool found_match = false;
 234         struct icmsg_negotiate *negop;
 235
 236         /* Check that there's enough space for icframe_vercnt, icmsg_vercnt */
 237         if (buflen < ICMSG_HDR + offsetof(struct icmsg_negotiate, reserved)) {
 238                 pr_err_ratelimited("Invalid icmsg negotiate\n");
 239                 return false;
 240         }
 241
 242         icmsghdrp->icmsgsize = 0x10;
 243         negop = (struct icmsg_negotiate *)&buf[ICMSG_HDR];
 244
 245         icframe_major = negop->icframe_vercnt;
 246         icframe_minor = 0;
 247
 248         icmsg_major = negop->icmsg_vercnt;
 249         icmsg_minor = 0;
 250
 251         /* Validate negop packet */
 252         if (icframe_major > IC_VERSION_NEGOTIATION_MAX_VER_COUNT ||
 253             icmsg_major > IC_VERSION_NEGOTIATION_MAX_VER_COUNT ||
 254             ICMSG_NEGOTIATE_PKT_SIZE(icframe_major, icmsg_major) > buflen) {
 255                 pr_err_ratelimited("Invalid icmsg negotiate - icframe_major: %u, icmsg_major: %u\n",
 256                                    icframe_major, icmsg_major);
 257                 goto fw_error;
 258         }
 259
 260         /*
 261          * Select the framework version number we will
 262          * support.
 263          */
 264
 265         for (i = 0; i < fw_vercnt; i++) {
 266                 fw_major = (fw_version[i] >> 16);
 267                 fw_minor = (fw_version[i] & 0xFFFF);
 268
 269                 for (j = 0; j < negop->icframe_vercnt; j++) {
 270                         if ((negop->icversion_data[j].major == fw_major) &&
 271                             (negop->icversion_data[j].minor == fw_minor)) {
 272                                 icframe_major = negop->icversion_data[j].major;
 273                                 icframe_minor = negop->icversion_data[j].minor;
 274                                 found_match = true;
 275                                 break;
 276                         }
 277                 }
 278
 279                 if (found_match)
 280                         break;
 281         }
 282
 283         if (!found_match)
 284                 goto fw_error;
 285
 286         found_match = false;
 287
 288         for (i = 0; i < srv_vercnt; i++) {
 289                 srv_major = (srv_version[i] >> 16);
 290                 srv_minor = (srv_version[i] & 0xFFFF);
 291
 292                 for (j = negop->icframe_vercnt;
 293                         (j < negop->icframe_vercnt + negop->icmsg_vercnt);
 294                         j++) {
 295
 296                         if ((negop->icversion_data[j].major == srv_major) &&
 297                                 (negop->icversion_data[j].minor == srv_minor)) {
 298
 299                                 icmsg_major = negop->icversion_data[j].major;
 300                                 icmsg_minor = negop->icversion_data[j].minor;
 301                                 found_match = true;
 302                                 break;
 303                         }
 304                 }
 305
 306                 if (found_match)
 307                         break;
 308         }
 309
 310         /*
 311          * Respond with the framework and service
 312          * version numbers we can support.
 313          */
 314
 315 fw_error:
 316         if (!found_match) {
 317                 negop->icframe_vercnt = 0;
 318                 negop->icmsg_vercnt = 0;
 319         } else {
 320                 negop->icframe_vercnt = 1;
 321                 negop->icmsg_vercnt = 1;
 322         }
 323
 324         if (nego_fw_version)
 325                 *nego_fw_version = (icframe_major << 16) | icframe_minor;
 326
 327         if (nego_srv_version)
 328                 *nego_srv_version = (icmsg_major << 16) | icmsg_minor;
 329
 330         negop->icversion_data[0].major = icframe_major;
 331         negop->icversion_data[0].minor = icframe_minor;
 332         negop->icversion_data[1].major = icmsg_major;
 333         negop->icversion_data[1].minor = icmsg_minor;
 334         return found_match;
 335 }
 336
 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 smp_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 smp_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         smp_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 appplication 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_alloced_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         init_vp_index(newchannel);
 610
 611         /* Remember the channels that should be cleaned up upon suspend. */
 612         if (is_hvsock_channel(newchannel) || is_sub_channel(newchannel))
 613                 atomic_inc(&vmbus_connection.nr_chan_close_on_suspend);
 614
 615         /*
 616          * Now that we have acquired the channel_mutex,
 617          * we can release the potentially racing rescind thread.
 618          */
 619         atomic_dec(&vmbus_connection.offer_in_progress);
 620
 621         list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
 622                 if (guid_equal(&channel->offermsg.offer.if_type,
 623                                &newchannel->offermsg.offer.if_type) &&
 624                     guid_equal(&channel->offermsg.offer.if_instance,
 625                                &newchannel->offermsg.offer.if_instance)) {
 626                         fnew = false;
 627                         break;
 628                 }
 629         }
 630
 631         if (fnew) {
 632                 list_add_tail(&newchannel->listentry,
 633                               &vmbus_connection.chn_list);
 634         } else {
 635                 /*
 636                  * Check to see if this is a valid sub-channel.
 637                  */
 638                 if (newchannel->offermsg.offer.sub_channel_index == 0) {
 639                         mutex_unlock(&vmbus_connection.channel_mutex);
 640                         /*
 641                          * Don't call free_channel(), because newchannel->kobj
 642                          * is not initialized yet.
 643                          */
 644                         kfree(newchannel);
 645                         WARN_ON_ONCE(1);
 646                         return;
 647                 }
 648                 /*
 649                  * Process the sub-channel.
 650                  */
 651                 newchannel->primary_channel = channel;
 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  * We use this state to statically distribute the channel interrupt load.
 689  */
 690 static int next_numa_node_id;
 691
 692 /*
 693  * Starting with Win8, we can statically distribute the incoming
 694  * channel interrupt load by binding a channel to VCPU.
 695  *
 696  * For pre-win8 hosts or non-performance critical channels we assign the
 697  * VMBUS_CONNECT_CPU.
 698  *
 699  * Starting with win8, performance critical channels will be distributed
 700  * evenly among all the available NUMA nodes.  Once the node is assigned,
 701  * we will assign the CPU based on a simple round robin scheme.
 702  */
 703 static void init_vp_index(struct vmbus_channel *channel)
 704 {
 705         bool perf_chn = hv_is_perf_channel(channel);
 706         cpumask_var_t available_mask;
 707         struct cpumask *alloced_mask;
 708         u32 target_cpu;
 709         int numa_node;
 710
 711         if ((vmbus_proto_version == VERSION_WS2008) ||
 712             (vmbus_proto_version == VERSION_WIN7) || (!perf_chn) ||
 713             !alloc_cpumask_var(&available_mask, GFP_KERNEL)) {
 714                 /*
 715                  * Prior to win8, all channel interrupts are
 716                  * delivered on VMBUS_CONNECT_CPU.
 717                  * Also if the channel is not a performance critical
 718                  * channel, bind it to VMBUS_CONNECT_CPU.
 719                  * In case alloc_cpumask_var() fails, bind it to
 720                  * VMBUS_CONNECT_CPU.
 721                  */
 722                 channel->target_cpu = VMBUS_CONNECT_CPU;
 723                 if (perf_chn)
 724                         hv_set_alloced_cpu(VMBUS_CONNECT_CPU);
 725                 return;
 726         }
 727
 728         while (true) {
 729                 numa_node = next_numa_node_id++;
 730                 if (numa_node == nr_node_ids) {
 731                         next_numa_node_id = 0;
 732                         continue;
 733                 }
 734                 if (cpumask_empty(cpumask_of_node(numa_node)))
 735                         continue;
 736                 break;
 737         }
 738         alloced_mask = &hv_context.hv_numa_map[numa_node];
 739
 740         if (cpumask_weight(alloced_mask) ==
 741             cpumask_weight(cpumask_of_node(numa_node))) {
 742                 /*
 743                  * We have cycled through all the CPUs in the node;
 744                  * reset the alloced map.
 745                  */
 746                 cpumask_clear(alloced_mask);
 747         }
 748
 749         cpumask_xor(available_mask, alloced_mask, cpumask_of_node(numa_node));
 750
 751         target_cpu = cpumask_first(available_mask);
 752         cpumask_set_cpu(target_cpu, alloced_mask);
 753
 754         channel->target_cpu = target_cpu;
 755
 756         free_cpumask_var(available_mask);
 757 }
 758
 759 static void vmbus_wait_for_unload(void)
 760 {
 761         int cpu;
 762         void *page_addr;
 763         struct hv_message *msg;
 764         struct vmbus_channel_message_header *hdr;
 765         u32 message_type, i;
 766
 767         /*
 768          * CHANNELMSG_UNLOAD_RESPONSE is always delivered to the CPU which was
 769          * used for initial contact or to CPU0 depending on host version. When
 770          * we're crashing on a different CPU let's hope that IRQ handler on
 771          * the cpu which receives CHANNELMSG_UNLOAD_RESPONSE is still
 772          * functional and vmbus_unload_response() will complete
 773          * vmbus_connection.unload_event. If not, the last thing we can do is
 774          * read message pages for all CPUs directly.
 775          *
 776          * Wait no more than 10 seconds so that the panic path can't get
 777          * hung forever in case the response message isn't seen.
 778          */
 779         for (i = 0; i < 1000; i++) {
 780                 if (completion_done(&vmbus_connection.unload_event))
 781                         break;
 782
 783                 for_each_online_cpu(cpu) {
 784                         struct hv_per_cpu_context *hv_cpu
 785                                 = per_cpu_ptr(hv_context.cpu_context, cpu);
 786
 787                         page_addr = hv_cpu->synic_message_page;
 788                         msg = (struct hv_message *)page_addr
 789                                 + VMBUS_MESSAGE_SINT;
 790
 791                         message_type = READ_ONCE(msg->header.message_type);
 792                         if (message_type == HVMSG_NONE)
 793                                 continue;
 794
 795                         hdr = (struct vmbus_channel_message_header *)
 796                                 msg->u.payload;
 797
 798                         if (hdr->msgtype == CHANNELMSG_UNLOAD_RESPONSE)
 799                                 complete(&vmbus_connection.unload_event);
 800
 801                         vmbus_signal_eom(msg, message_type);
 802                 }
 803
 804                 mdelay(10);
 805         }
 806
 807         /*
 808          * We're crashing and already got the UNLOAD_RESPONSE, cleanup all
 809          * maybe-pending messages on all CPUs to be able to receive new
 810          * messages after we reconnect.
 811          */
 812         for_each_online_cpu(cpu) {
 813                 struct hv_per_cpu_context *hv_cpu
 814                         = per_cpu_ptr(hv_context.cpu_context, cpu);
 815
 816                 page_addr = hv_cpu->synic_message_page;
 817                 msg = (struct hv_message *)page_addr + VMBUS_MESSAGE_SINT;
 818                 msg->header.message_type = HVMSG_NONE;
 819         }
 820 }
 821
 822 /*
 823  * vmbus_unload_response - Handler for the unload response.
 824  */
 825 static void vmbus_unload_response(struct vmbus_channel_message_header *hdr)
 826 {
 827         /*
 828          * This is a global event; just wakeup the waiting thread.
 829          * Once we successfully unload, we can cleanup the monitor state.
 830          */
 831         complete(&vmbus_connection.unload_event);
 832 }
 833
 834 void vmbus_initiate_unload(bool crash)
 835 {
 836         struct vmbus_channel_message_header hdr;
 837
 838         if (xchg(&vmbus_connection.conn_state, DISCONNECTED) == DISCONNECTED)
 839                 return;
 840
 841         /* Pre-Win2012R2 hosts don't support reconnect */
 842         if (vmbus_proto_version < VERSION_WIN8_1)
 843                 return;
 844
 845         init_completion(&vmbus_connection.unload_event);
 846         memset(&hdr, 0, sizeof(struct vmbus_channel_message_header));
 847         hdr.msgtype = CHANNELMSG_UNLOAD;
 848         vmbus_post_msg(&hdr, sizeof(struct vmbus_channel_message_header),
 849                        !crash);
 850
 851         /*
 852          * vmbus_initiate_unload() is also called on crash and the crash can be
 853          * happening in an interrupt context, where scheduling is impossible.
 854          */
 855         if (!crash)
 856                 wait_for_completion(&vmbus_connection.unload_event);
 857         else
 858                 vmbus_wait_for_unload();
 859 }
 860
 861 static void check_ready_for_resume_event(void)
 862 {
 863         /*
 864          * If all the old primary channels have been fixed up, then it's safe
 865          * to resume.
 866          */
 867         if (atomic_dec_and_test(&vmbus_connection.nr_chan_fixup_on_resume))
 868                 complete(&vmbus_connection.ready_for_resume_event);
 869 }
 870
 871 static void vmbus_setup_channel_state(struct vmbus_channel *channel,
 872                                       struct vmbus_channel_offer_channel *offer)
 873 {
 874         /*
 875          * Setup state for signalling the host.
 876          */
 877         channel->sig_event = VMBUS_EVENT_CONNECTION_ID;
 878
 879         if (vmbus_proto_version != VERSION_WS2008) {
 880                 channel->is_dedicated_interrupt =
 881                                 (offer->is_dedicated_interrupt != 0);
 882                 channel->sig_event = offer->connection_id;
 883         }
 884
 885         memcpy(&channel->offermsg, offer,
 886                sizeof(struct vmbus_channel_offer_channel));
 887         channel->monitor_grp = (u8)offer->monitorid / 32;
 888         channel->monitor_bit = (u8)offer->monitorid % 32;
 889         channel->device_id = hv_get_dev_type(channel);
 890 }
 891
 892 /*
 893  * find_primary_channel_by_offer - Get the channel object given the new offer.
 894  * This is only used in the resume path of hibernation.
 895  */
 896 static struct vmbus_channel *
 897 find_primary_channel_by_offer(const struct vmbus_channel_offer_channel *offer)
 898 {
 899         struct vmbus_channel *channel = NULL, *iter;
 900         const guid_t *inst1, *inst2;
 901
 902         /* Ignore sub-channel offers. */
 903         if (offer->offer.sub_channel_index != 0)
 904                 return NULL;
 905
 906         mutex_lock(&vmbus_connection.channel_mutex);
 907
 908         list_for_each_entry(iter, &vmbus_connection.chn_list, listentry) {
 909                 inst1 = &iter->offermsg.offer.if_instance;
 910                 inst2 = &offer->offer.if_instance;
 911
 912                 if (guid_equal(inst1, inst2)) {
 913                         channel = iter;
 914                         break;
 915                 }
 916         }
 917
 918         mutex_unlock(&vmbus_connection.channel_mutex);
 919
 920         return channel;
 921 }
 922
 923 static bool vmbus_is_valid_device(const guid_t *guid)
 924 {
 925         u16 i;
 926
 927         if (!hv_is_isolation_supported())
 928                 return true;
 929
 930         for (i = 0; i < ARRAY_SIZE(vmbus_devs); i++) {
 931                 if (guid_equal(guid, &vmbus_devs[i].guid))
 932                         return vmbus_devs[i].allowed_in_isolated;
 933         }
 934         return false;
 935 }
 936
 937 /*
 938  * vmbus_onoffer - Handler for channel offers from vmbus in parent partition.
 939  *
 940  */
 941 static void vmbus_onoffer(struct vmbus_channel_message_header *hdr)
 942 {
 943         struct vmbus_channel_offer_channel *offer;
 944         struct vmbus_channel *oldchannel, *newchannel;
 945         size_t offer_sz;
 946
 947         offer = (struct vmbus_channel_offer_channel *)hdr;
 948
 949         trace_vmbus_onoffer(offer);
 950
 951         if (!vmbus_is_valid_device(&offer->offer.if_type)) {
 952                 pr_err_ratelimited("Invalid offer %d from the host supporting isolation\n",
 953                                    offer->child_relid);
 954                 atomic_dec(&vmbus_connection.offer_in_progress);
 955                 return;
 956         }
 957
 958         oldchannel = find_primary_channel_by_offer(offer);
 959
 960         if (oldchannel != NULL) {
 961                 /*
 962                  * We're resuming from hibernation: all the sub-channel and
 963                  * hv_sock channels we had before the hibernation should have
 964                  * been cleaned up, and now we must be seeing a re-offered
 965                  * primary channel that we had before the hibernation.
 966                  */
 967
 968                 /*
 969                  * { Initially: channel relid = INVALID_RELID,
 970                  *              channels[valid_relid] = NULL }
 971                  *
 972                  * CPU1                                 CPU2
 973                  *
 974                  * [vmbus_onoffer()]                    [vmbus_device_release()]
 975                  *
 976                  * LOCK channel_mutex                   LOCK channel_mutex
 977                  * STORE channel relid = valid_relid    LOAD r1 = channel relid
 978                  * MAP_RELID channel                    if (r1 != INVALID_RELID)
 979                  * UNLOCK channel_mutex                   UNMAP_RELID channel
 980                  *                                      UNLOCK channel_mutex
 981                  *
 982                  * Forbids: r1 == valid_relid &&
 983                  *              channels[valid_relid] == channel
 984                  *
 985                  * Note.  r1 can be INVALID_RELID only for an hv_sock channel.
 986                  * None of the hv_sock channels which were present before the
 987                  * suspend are re-offered upon the resume.  See the WARN_ON()
 988                  * in hv_process_channel_removal().
 989                  */
 990                 mutex_lock(&vmbus_connection.channel_mutex);
 991
 992                 atomic_dec(&vmbus_connection.offer_in_progress);
 993
 994                 WARN_ON(oldchannel->offermsg.child_relid != INVALID_RELID);
 995                 /* Fix up the relid. */
 996                 oldchannel->offermsg.child_relid = offer->child_relid;
 997
 998                 offer_sz = sizeof(*offer);
 999                 if (memcmp(offer, &oldchannel->offermsg, offer_sz) != 0) {
1000                         /*
1001                          * This is not an error, since the host can also change
1002                          * the other field(s) of the offer, e.g. on WS RS5
1003                          * (Build 17763), the offer->connection_id of the
1004                          * Mellanox VF vmbus device can change when the host
1005                          * reoffers the device upon resume.
1006                          */
1007                         pr_debug("vmbus offer changed: relid=%d\n",
1008                                  offer->child_relid);
1009
1010                         print_hex_dump_debug("Old vmbus offer: ",
1011                                              DUMP_PREFIX_OFFSET, 16, 4,
1012                                              &oldchannel->offermsg, offer_sz,
1013                                              false);
1014                         print_hex_dump_debug("New vmbus offer: ",
1015                                              DUMP_PREFIX_OFFSET, 16, 4,
1016                                              offer, offer_sz, false);
1017
1018                         /* Fix up the old channel. */
1019                         vmbus_setup_channel_state(oldchannel, offer);
1020                 }
1021
1022                 /* Add the channel back to the array of channels. */
1023                 vmbus_channel_map_relid(oldchannel);
1024                 check_ready_for_resume_event();
1025
1026                 mutex_unlock(&vmbus_connection.channel_mutex);
1027                 return;
1028         }
1029
1030         /* Allocate the channel object and save this offer. */
1031         newchannel = alloc_channel();
1032         if (!newchannel) {
1033                 vmbus_release_relid(offer->child_relid);
1034                 atomic_dec(&vmbus_connection.offer_in_progress);
1035                 pr_err("Unable to allocate channel object\n");
1036                 return;
1037         }
1038
1039         vmbus_setup_channel_state(newchannel, offer);
1040
1041         vmbus_process_offer(newchannel);
1042 }
1043
1044 static void check_ready_for_suspend_event(void)
1045 {
1046         /*
1047          * If all the sub-channels or hv_sock channels have been cleaned up,
1048          * then it's safe to suspend.
1049          */
1050         if (atomic_dec_and_test(&vmbus_connection.nr_chan_close_on_suspend))
1051                 complete(&vmbus_connection.ready_for_suspend_event);
1052 }
1053
1054 /*
1055  * vmbus_onoffer_rescind - Rescind offer handler.
1056  *
1057  * We queue a work item to process this offer synchronously
1058  */
1059 static void vmbus_onoffer_rescind(struct vmbus_channel_message_header *hdr)
1060 {
1061         struct vmbus_channel_rescind_offer *rescind;
1062         struct vmbus_channel *channel;
1063         struct device *dev;
1064         bool clean_up_chan_for_suspend;
1065
1066         rescind = (struct vmbus_channel_rescind_offer *)hdr;
1067
1068         trace_vmbus_onoffer_rescind(rescind);
1069
1070         /*
1071          * The offer msg and the corresponding rescind msg
1072          * from the host are guranteed to be ordered -
1073          * offer comes in first and then the rescind.
1074          * Since we process these events in work elements,
1075          * and with preemption, we may end up processing
1076          * the events out of order.  We rely on the synchronization
1077          * provided by offer_in_progress and by channel_mutex for
1078          * ordering these events:
1079          *
1080          * { Initially: offer_in_progress = 1 }
1081          *
1082          * CPU1                         CPU2
1083          *
1084          * [vmbus_onoffer()]            [vmbus_onoffer_rescind()]
1085          *
1086          * LOCK channel_mutex           WAIT_ON offer_in_progress == 0
1087          * DECREMENT offer_in_progress  LOCK channel_mutex
1088          * STORE channels[]             LOAD channels[]
1089          * UNLOCK channel_mutex         UNLOCK channel_mutex
1090          *
1091          * Forbids: CPU2's LOAD from *not* seeing CPU1's STORE
1092          */
1093
1094         while (atomic_read(&vmbus_connection.offer_in_progress) != 0) {
1095                 /*
1096                  * We wait here until any channel offer is currently
1097                  * being processed.
1098                  */
1099                 msleep(1);
1100         }
1101
1102         mutex_lock(&vmbus_connection.channel_mutex);
1103         channel = relid2channel(rescind->child_relid);
1104         if (channel != NULL) {
1105                 /*
1106                  * Guarantee that no other instance of vmbus_onoffer_rescind()
1107                  * has got a reference to the channel object.  Synchronize on
1108                  * &vmbus_connection.channel_mutex.
1109                  */
1110                 if (channel->rescind_ref) {
1111                         mutex_unlock(&vmbus_connection.channel_mutex);
1112                         return;
1113                 }
1114                 channel->rescind_ref = true;
1115         }
1116         mutex_unlock(&vmbus_connection.channel_mutex);
1117
1118         if (channel == NULL) {
1119                 /*
1120                  * We failed in processing the offer message;
1121                  * we would have cleaned up the relid in that
1122                  * failure path.
1123                  */
1124                 return;
1125         }
1126
1127         clean_up_chan_for_suspend = is_hvsock_channel(channel) ||
1128                                     is_sub_channel(channel);
1129         /*
1130          * Before setting channel->rescind in vmbus_rescind_cleanup(), we
1131          * should make sure the channel callback is not running any more.
1132          */
1133         vmbus_reset_channel_cb(channel);
1134
1135         /*
1136          * Now wait for offer handling to complete.
1137          */
1138         vmbus_rescind_cleanup(channel);
1139         while (READ_ONCE(channel->probe_done) == false) {
1140                 /*
1141                  * We wait here until any channel offer is currently
1142                  * being processed.
1143                  */
1144                 msleep(1);
1145         }
1146
1147         /*
1148          * At this point, the rescind handling can proceed safely.
1149          */
1150
1151         if (channel->device_obj) {
1152                 if (channel->chn_rescind_callback) {
1153                         channel->chn_rescind_callback(channel);
1154
1155                         if (clean_up_chan_for_suspend)
1156                                 check_ready_for_suspend_event();
1157
1158                         return;
1159                 }
1160                 /*
1161                  * We will have to unregister this device from the
1162                  * driver core.
1163                  */
1164                 dev = get_device(&channel->device_obj->device);
1165                 if (dev) {
1166                         vmbus_device_unregister(channel->device_obj);
1167                         put_device(dev);
1168                 }
1169         } else if (channel->primary_channel != NULL) {
1170                 /*
1171                  * Sub-channel is being rescinded. Following is the channel
1172                  * close sequence when initiated from the driveri (refer to
1173                  * vmbus_close() for details):
1174                  * 1. Close all sub-channels first
1175                  * 2. Then close the primary channel.
1176                  */
1177                 mutex_lock(&vmbus_connection.channel_mutex);
1178                 if (channel->state == CHANNEL_OPEN_STATE) {
1179                         /*
1180                          * The channel is currently not open;
1181                          * it is safe for us to cleanup the channel.
1182                          */
1183                         hv_process_channel_removal(channel);
1184                 } else {
1185                         complete(&channel->rescind_event);
1186                 }
1187                 mutex_unlock(&vmbus_connection.channel_mutex);
1188         }
1189
1190         /* The "channel" may have been freed. Do not access it any longer. */
1191
1192         if (clean_up_chan_for_suspend)
1193                 check_ready_for_suspend_event();
1194 }
1195
1196 void vmbus_hvsock_device_unregister(struct vmbus_channel *channel)
1197 {
1198         BUG_ON(!is_hvsock_channel(channel));
1199
1200         /* We always get a rescind msg when a connection is closed. */
1201         while (!READ_ONCE(channel->probe_done) || !READ_ONCE(channel->rescind))
1202                 msleep(1);
1203
1204         vmbus_device_unregister(channel->device_obj);
1205 }
1206 EXPORT_SYMBOL_GPL(vmbus_hvsock_device_unregister);
1207
1208
1209 /*
1210  * vmbus_onoffers_delivered -
1211  * This is invoked when all offers have been delivered.
1212  *
1213  * Nothing to do here.
1214  */
1215 static void vmbus_onoffers_delivered(
1216                         struct vmbus_channel_message_header *hdr)
1217 {
1218 }
1219
1220 /*
1221  * vmbus_onopen_result - Open result handler.
1222  *
1223  * This is invoked when we received a response to our channel open request.
1224  * Find the matching request, copy the response and signal the requesting
1225  * thread.
1226  */
1227 static void vmbus_onopen_result(struct vmbus_channel_message_header *hdr)
1228 {
1229         struct vmbus_channel_open_result *result;
1230         struct vmbus_channel_msginfo *msginfo;
1231         struct vmbus_channel_message_header *requestheader;
1232         struct vmbus_channel_open_channel *openmsg;
1233         unsigned long flags;
1234
1235         result = (struct vmbus_channel_open_result *)hdr;
1236
1237         trace_vmbus_onopen_result(result);
1238
1239         /*
1240          * Find the open msg, copy the result and signal/unblock the wait event
1241          */
1242         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1243
1244         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1245                                 msglistentry) {
1246                 requestheader =
1247                         (struct vmbus_channel_message_header *)msginfo->msg;
1248
1249                 if (requestheader->msgtype == CHANNELMSG_OPENCHANNEL) {
1250                         openmsg =
1251                         (struct vmbus_channel_open_channel *)msginfo->msg;
1252                         if (openmsg->child_relid == result->child_relid &&
1253                             openmsg->openid == result->openid) {
1254                                 memcpy(&msginfo->response.open_result,
1255                                        result,
1256                                        sizeof(
1257                                         struct vmbus_channel_open_result));
1258                                 complete(&msginfo->waitevent);
1259                                 break;
1260                         }
1261                 }
1262         }
1263         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1264 }
1265
1266 /*
1267  * vmbus_ongpadl_created - GPADL created handler.
1268  *
1269  * This is invoked when we received a response to our gpadl create request.
1270  * Find the matching request, copy the response and signal the requesting
1271  * thread.
1272  */
1273 static void vmbus_ongpadl_created(struct vmbus_channel_message_header *hdr)
1274 {
1275         struct vmbus_channel_gpadl_created *gpadlcreated;
1276         struct vmbus_channel_msginfo *msginfo;
1277         struct vmbus_channel_message_header *requestheader;
1278         struct vmbus_channel_gpadl_header *gpadlheader;
1279         unsigned long flags;
1280
1281         gpadlcreated = (struct vmbus_channel_gpadl_created *)hdr;
1282
1283         trace_vmbus_ongpadl_created(gpadlcreated);
1284
1285         /*
1286          * Find the establish msg, copy the result and signal/unblock the wait
1287          * event
1288          */
1289         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1290
1291         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1292                                 msglistentry) {
1293                 requestheader =
1294                         (struct vmbus_channel_message_header *)msginfo->msg;
1295
1296                 if (requestheader->msgtype == CHANNELMSG_GPADL_HEADER) {
1297                         gpadlheader =
1298                         (struct vmbus_channel_gpadl_header *)requestheader;
1299
1300                         if ((gpadlcreated->child_relid ==
1301                              gpadlheader->child_relid) &&
1302                             (gpadlcreated->gpadl == gpadlheader->gpadl)) {
1303                                 memcpy(&msginfo->response.gpadl_created,
1304                                        gpadlcreated,
1305                                        sizeof(
1306                                         struct vmbus_channel_gpadl_created));
1307                                 complete(&msginfo->waitevent);
1308                                 break;
1309                         }
1310                 }
1311         }
1312         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1313 }
1314
1315 /*
1316  * vmbus_ongpadl_torndown - GPADL torndown handler.
1317  *
1318  * This is invoked when we received a response to our gpadl teardown request.
1319  * Find the matching request, copy the response and signal the requesting
1320  * thread.
1321  */
1322 static void vmbus_ongpadl_torndown(
1323                         struct vmbus_channel_message_header *hdr)
1324 {
1325         struct vmbus_channel_gpadl_torndown *gpadl_torndown;
1326         struct vmbus_channel_msginfo *msginfo;
1327         struct vmbus_channel_message_header *requestheader;
1328         struct vmbus_channel_gpadl_teardown *gpadl_teardown;
1329         unsigned long flags;
1330
1331         gpadl_torndown = (struct vmbus_channel_gpadl_torndown *)hdr;
1332
1333         trace_vmbus_ongpadl_torndown(gpadl_torndown);
1334
1335         /*
1336          * Find the open msg, copy the result and signal/unblock the wait event
1337          */
1338         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1339
1340         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1341                                 msglistentry) {
1342                 requestheader =
1343                         (struct vmbus_channel_message_header *)msginfo->msg;
1344
1345                 if (requestheader->msgtype == CHANNELMSG_GPADL_TEARDOWN) {
1346                         gpadl_teardown =
1347                         (struct vmbus_channel_gpadl_teardown *)requestheader;
1348
1349                         if (gpadl_torndown->gpadl == gpadl_teardown->gpadl) {
1350                                 memcpy(&msginfo->response.gpadl_torndown,
1351                                        gpadl_torndown,
1352                                        sizeof(
1353                                         struct vmbus_channel_gpadl_torndown));
1354                                 complete(&msginfo->waitevent);
1355                                 break;
1356                         }
1357                 }
1358         }
1359         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1360 }
1361
1362 /*
1363  * vmbus_onversion_response - Version response handler
1364  *
1365  * This is invoked when we received a response to our initiate contact request.
1366  * Find the matching request, copy the response and signal the requesting
1367  * thread.
1368  */
1369 static void vmbus_onversion_response(
1370                 struct vmbus_channel_message_header *hdr)
1371 {
1372         struct vmbus_channel_msginfo *msginfo;
1373         struct vmbus_channel_message_header *requestheader;
1374         struct vmbus_channel_version_response *version_response;
1375         unsigned long flags;
1376
1377         version_response = (struct vmbus_channel_version_response *)hdr;
1378
1379         trace_vmbus_onversion_response(version_response);
1380
1381         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1382
1383         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1384                                 msglistentry) {
1385                 requestheader =
1386                         (struct vmbus_channel_message_header *)msginfo->msg;
1387
1388                 if (requestheader->msgtype ==
1389                     CHANNELMSG_INITIATE_CONTACT) {
1390                         memcpy(&msginfo->response.version_response,
1391                               version_response,
1392                               sizeof(struct vmbus_channel_version_response));
1393                         complete(&msginfo->waitevent);
1394                 }
1395         }
1396         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1397 }
1398
1399 /* Channel message dispatch table */
1400 const struct vmbus_channel_message_table_entry
1401 channel_message_table[CHANNELMSG_COUNT] = {
1402         { CHANNELMSG_INVALID,                   0, NULL, 0},
1403         { CHANNELMSG_OFFERCHANNEL,              0, vmbus_onoffer,
1404                 sizeof(struct vmbus_channel_offer_channel)},
1405         { CHANNELMSG_RESCIND_CHANNELOFFER,      0, vmbus_onoffer_rescind,
1406                 sizeof(struct vmbus_channel_rescind_offer) },
1407         { CHANNELMSG_REQUESTOFFERS,             0, NULL, 0},
1408         { CHANNELMSG_ALLOFFERS_DELIVERED,       1, vmbus_onoffers_delivered, 0},
1409         { CHANNELMSG_OPENCHANNEL,               0, NULL, 0},
1410         { CHANNELMSG_OPENCHANNEL_RESULT,        1, vmbus_onopen_result,
1411                 sizeof(struct vmbus_channel_open_result)},
1412         { CHANNELMSG_CLOSECHANNEL,              0, NULL, 0},
1413         { CHANNELMSG_GPADL_HEADER,              0, NULL, 0},
1414         { CHANNELMSG_GPADL_BODY,                0, NULL, 0},
1415         { CHANNELMSG_GPADL_CREATED,             1, vmbus_ongpadl_created,
1416                 sizeof(struct vmbus_channel_gpadl_created)},
1417         { CHANNELMSG_GPADL_TEARDOWN,            0, NULL, 0},
1418         { CHANNELMSG_GPADL_TORNDOWN,            1, vmbus_ongpadl_torndown,
1419                 sizeof(struct vmbus_channel_gpadl_torndown) },
1420         { CHANNELMSG_RELID_RELEASED,            0, NULL, 0},
1421         { CHANNELMSG_INITIATE_CONTACT,          0, NULL, 0},
1422         { CHANNELMSG_VERSION_RESPONSE,          1, vmbus_onversion_response,
1423                 sizeof(struct vmbus_channel_version_response)},
1424         { CHANNELMSG_UNLOAD,                    0, NULL, 0},
1425         { CHANNELMSG_UNLOAD_RESPONSE,           1, vmbus_unload_response, 0},
1426         { CHANNELMSG_18,                        0, NULL, 0},
1427         { CHANNELMSG_19,                        0, NULL, 0},
1428         { CHANNELMSG_20,                        0, NULL, 0},
1429         { CHANNELMSG_TL_CONNECT_REQUEST,        0, NULL, 0},
1430         { CHANNELMSG_MODIFYCHANNEL,             0, NULL, 0},
1431         { CHANNELMSG_TL_CONNECT_RESULT,         0, NULL, 0},
1432 };
1433
1434 /*
1435  * vmbus_onmessage - Handler for channel protocol messages.
1436  *
1437  * This is invoked in the vmbus worker thread context.
1438  */
1439 void vmbus_onmessage(struct vmbus_channel_message_header *hdr)
1440 {
1441         trace_vmbus_on_message(hdr);
1442
1443         /*
1444          * vmbus_on_msg_dpc() makes sure the hdr->msgtype here can not go
1445          * out of bound and the message_handler pointer can not be NULL.
1446          */
1447         channel_message_table[hdr->msgtype].message_handler(hdr);
1448 }
1449
1450 /*
1451  * vmbus_request_offers - Send a request to get all our pending offers.
1452  */
1453 int vmbus_request_offers(void)
1454 {
1455         struct vmbus_channel_message_header *msg;
1456         struct vmbus_channel_msginfo *msginfo;
1457         int ret;
1458
1459         msginfo = kmalloc(sizeof(*msginfo) +
1460                           sizeof(struct vmbus_channel_message_header),
1461                           GFP_KERNEL);
1462         if (!msginfo)
1463                 return -ENOMEM;
1464
1465         msg = (struct vmbus_channel_message_header *)msginfo->msg;
1466
1467         msg->msgtype = CHANNELMSG_REQUESTOFFERS;
1468
1469         ret = vmbus_post_msg(msg, sizeof(struct vmbus_channel_message_header),
1470                              true);
1471
1472         trace_vmbus_request_offers(ret);
1473
1474         if (ret != 0) {
1475                 pr_err("Unable to request offers - %d\n", ret);
1476
1477                 goto cleanup;
1478         }
1479
1480 cleanup:
1481         kfree(msginfo);
1482
1483         return ret;
1484 }
1485
1486 static void invoke_sc_cb(struct vmbus_channel *primary_channel)
1487 {
1488         struct list_head *cur, *tmp;
1489         struct vmbus_channel *cur_channel;
1490
1491         if (primary_channel->sc_creation_callback == NULL)
1492                 return;
1493
1494         list_for_each_safe(cur, tmp, &primary_channel->sc_list) {
1495                 cur_channel = list_entry(cur, struct vmbus_channel, sc_list);
1496
1497                 primary_channel->sc_creation_callback(cur_channel);
1498         }
1499 }
1500
1501 void vmbus_set_sc_create_callback(struct vmbus_channel *primary_channel,
1502                                 void (*sc_cr_cb)(struct vmbus_channel *new_sc))
1503 {
1504         primary_channel->sc_creation_callback = sc_cr_cb;
1505 }
1506 EXPORT_SYMBOL_GPL(vmbus_set_sc_create_callback);
1507
1508 bool vmbus_are_subchannels_present(struct vmbus_channel *primary)
1509 {
1510         bool ret;
1511
1512         ret = !list_empty(&primary->sc_list);
1513
1514         if (ret) {
1515                 /*
1516                  * Invoke the callback on sub-channel creation.
1517                  * This will present a uniform interface to the
1518                  * clients.
1519                  */
1520                 invoke_sc_cb(primary);
1521         }
1522
1523         return ret;
1524 }
1525 EXPORT_SYMBOL_GPL(vmbus_are_subchannels_present);
1526
1527 void vmbus_set_chn_rescind_callback(struct vmbus_channel *channel,
1528                 void (*chn_rescind_cb)(struct vmbus_channel *))
1529 {
1530         channel->chn_rescind_callback = chn_rescind_cb;
1531 }
1532 EXPORT_SYMBOL_GPL(vmbus_set_chn_rescind_callback);