drivers/hv/channel_mgmt.c

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