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/hyperv.h>
  22 #include <asm/mshyperv.h>
  23
  24 #include "hyperv_vmbus.h"
  25
  26 static void init_vp_index(struct vmbus_channel *channel, u16 dev_type);
  27
  28 static const struct vmbus_device vmbus_devs[] = {
  29         /* IDE */
  30         { .dev_type = HV_IDE,
  31           HV_IDE_GUID,
  32           .perf_device = true,
  33         },
  34
  35         /* SCSI */
  36         { .dev_type = HV_SCSI,
  37           HV_SCSI_GUID,
  38           .perf_device = true,
  39         },
  40
  41         /* Fibre Channel */
  42         { .dev_type = HV_FC,
  43           HV_SYNTHFC_GUID,
  44           .perf_device = true,
  45         },
  46
  47         /* Synthetic NIC */
  48         { .dev_type = HV_NIC,
  49           HV_NIC_GUID,
  50           .perf_device = true,
  51         },
  52
  53         /* Network Direct */
  54         { .dev_type = HV_ND,
  55           HV_ND_GUID,
  56           .perf_device = true,
  57         },
  58
  59         /* PCIE */
  60         { .dev_type = HV_PCIE,
  61           HV_PCIE_GUID,
  62           .perf_device = false,
  63         },
  64
  65         /* Synthetic Frame Buffer */
  66         { .dev_type = HV_FB,
  67           HV_SYNTHVID_GUID,
  68           .perf_device = false,
  69         },
  70
  71         /* Synthetic Keyboard */
  72         { .dev_type = HV_KBD,
  73           HV_KBD_GUID,
  74           .perf_device = false,
  75         },
  76
  77         /* Synthetic MOUSE */
  78         { .dev_type = HV_MOUSE,
  79           HV_MOUSE_GUID,
  80           .perf_device = false,
  81         },
  82
  83         /* KVP */
  84         { .dev_type = HV_KVP,
  85           HV_KVP_GUID,
  86           .perf_device = false,
  87         },
  88
  89         /* Time Synch */
  90         { .dev_type = HV_TS,
  91           HV_TS_GUID,
  92           .perf_device = false,
  93         },
  94
  95         /* Heartbeat */
  96         { .dev_type = HV_HB,
  97           HV_HEART_BEAT_GUID,
  98           .perf_device = false,
  99         },
 100
 101         /* Shutdown */
 102         { .dev_type = HV_SHUTDOWN,
 103           HV_SHUTDOWN_GUID,
 104           .perf_device = false,
 105         },
 106
 107         /* File copy */
 108         { .dev_type = HV_FCOPY,
 109           HV_FCOPY_GUID,
 110           .perf_device = false,
 111         },
 112
 113         /* Backup */
 114         { .dev_type = HV_BACKUP,
 115           HV_VSS_GUID,
 116           .perf_device = false,
 117         },
 118
 119         /* Dynamic Memory */
 120         { .dev_type = HV_DM,
 121           HV_DM_GUID,
 122           .perf_device = false,
 123         },
 124
 125         /* Unknown GUID */
 126         { .dev_type = HV_UNKNOWN,
 127           .perf_device = false,
 128         },
 129 };
 130
 131 static const struct {
 132         guid_t guid;
 133 } vmbus_unsupported_devs[] = {
 134         { HV_AVMA1_GUID },
 135         { HV_AVMA2_GUID },
 136         { HV_RDV_GUID   },
 137 };
 138
 139 /*
 140  * The rescinded channel may be blocked waiting for a response from the host;
 141  * take care of that.
 142  */
 143 static void vmbus_rescind_cleanup(struct vmbus_channel *channel)
 144 {
 145         struct vmbus_channel_msginfo *msginfo;
 146         unsigned long flags;
 147
 148
 149         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
 150         channel->rescind = true;
 151         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
 152                                 msglistentry) {
 153
 154                 if (msginfo->waiting_channel == channel) {
 155                         complete(&msginfo->waitevent);
 156                         break;
 157                 }
 158         }
 159         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
 160 }
 161
 162 static bool is_unsupported_vmbus_devs(const guid_t *guid)
 163 {
 164         int i;
 165
 166         for (i = 0; i < ARRAY_SIZE(vmbus_unsupported_devs); i++)
 167                 if (guid_equal(guid, &vmbus_unsupported_devs[i].guid))
 168                         return true;
 169         return false;
 170 }
 171
 172 static u16 hv_get_dev_type(const struct vmbus_channel *channel)
 173 {
 174         const guid_t *guid = &channel->offermsg.offer.if_type;
 175         u16 i;
 176
 177         if (is_hvsock_channel(channel) || is_unsupported_vmbus_devs(guid))
 178                 return HV_UNKNOWN;
 179
 180         for (i = HV_IDE; i < HV_UNKNOWN; i++) {
 181                 if (guid_equal(guid, &vmbus_devs[i].guid))
 182                         return i;
 183         }
 184         pr_info("Unknown GUID: %pUl\n", guid);
 185         return i;
 186 }
 187
 188 /**
 189  * vmbus_prep_negotiate_resp() - Create default response for Negotiate message
 190  * @icmsghdrp: Pointer to msg header structure
 191  * @buf: Raw buffer channel data
 192  * @fw_version: The framework versions we can support.
 193  * @fw_vercnt: The size of @fw_version.
 194  * @srv_version: The service versions we can support.
 195  * @srv_vercnt: The size of @srv_version.
 196  * @nego_fw_version: The selected framework version.
 197  * @nego_srv_version: The selected service version.
 198  *
 199  * Note: Versions are given in decreasing order.
 200  *
 201  * Set up and fill in default negotiate response message.
 202  * Mainly used by Hyper-V drivers.
 203  */
 204 bool vmbus_prep_negotiate_resp(struct icmsg_hdr *icmsghdrp,
 205                                 u8 *buf, const int *fw_version, int fw_vercnt,
 206                                 const int *srv_version, int srv_vercnt,
 207                                 int *nego_fw_version, int *nego_srv_version)
 208 {
 209         int icframe_major, icframe_minor;
 210         int icmsg_major, icmsg_minor;
 211         int fw_major, fw_minor;
 212         int srv_major, srv_minor;
 213         int i, j;
 214         bool found_match = false;
 215         struct icmsg_negotiate *negop;
 216
 217         icmsghdrp->icmsgsize = 0x10;
 218         negop = (struct icmsg_negotiate *)&buf[
 219                 sizeof(struct vmbuspipe_hdr) +
 220                 sizeof(struct icmsg_hdr)];
 221
 222         icframe_major = negop->icframe_vercnt;
 223         icframe_minor = 0;
 224
 225         icmsg_major = negop->icmsg_vercnt;
 226         icmsg_minor = 0;
 227
 228         /*
 229          * Select the framework version number we will
 230          * support.
 231          */
 232
 233         for (i = 0; i < fw_vercnt; i++) {
 234                 fw_major = (fw_version[i] >> 16);
 235                 fw_minor = (fw_version[i] & 0xFFFF);
 236
 237                 for (j = 0; j < negop->icframe_vercnt; j++) {
 238                         if ((negop->icversion_data[j].major == fw_major) &&
 239                             (negop->icversion_data[j].minor == fw_minor)) {
 240                                 icframe_major = negop->icversion_data[j].major;
 241                                 icframe_minor = negop->icversion_data[j].minor;
 242                                 found_match = true;
 243                                 break;
 244                         }
 245                 }
 246
 247                 if (found_match)
 248                         break;
 249         }
 250
 251         if (!found_match)
 252                 goto fw_error;
 253
 254         found_match = false;
 255
 256         for (i = 0; i < srv_vercnt; i++) {
 257                 srv_major = (srv_version[i] >> 16);
 258                 srv_minor = (srv_version[i] & 0xFFFF);
 259
 260                 for (j = negop->icframe_vercnt;
 261                         (j < negop->icframe_vercnt + negop->icmsg_vercnt);
 262                         j++) {
 263
 264                         if ((negop->icversion_data[j].major == srv_major) &&
 265                                 (negop->icversion_data[j].minor == srv_minor)) {
 266
 267                                 icmsg_major = negop->icversion_data[j].major;
 268                                 icmsg_minor = negop->icversion_data[j].minor;
 269                                 found_match = true;
 270                                 break;
 271                         }
 272                 }
 273
 274                 if (found_match)
 275                         break;
 276         }
 277
 278         /*
 279          * Respond with the framework and service
 280          * version numbers we can support.
 281          */
 282
 283 fw_error:
 284         if (!found_match) {
 285                 negop->icframe_vercnt = 0;
 286                 negop->icmsg_vercnt = 0;
 287         } else {
 288                 negop->icframe_vercnt = 1;
 289                 negop->icmsg_vercnt = 1;
 290         }
 291
 292         if (nego_fw_version)
 293                 *nego_fw_version = (icframe_major << 16) | icframe_minor;
 294
 295         if (nego_srv_version)
 296                 *nego_srv_version = (icmsg_major << 16) | icmsg_minor;
 297
 298         negop->icversion_data[0].major = icframe_major;
 299         negop->icversion_data[0].minor = icframe_minor;
 300         negop->icversion_data[1].major = icmsg_major;
 301         negop->icversion_data[1].minor = icmsg_minor;
 302         return found_match;
 303 }
 304
 305 EXPORT_SYMBOL_GPL(vmbus_prep_negotiate_resp);
 306
 307 /*
 308  * alloc_channel - Allocate and initialize a vmbus channel object
 309  */
 310 static struct vmbus_channel *alloc_channel(void)
 311 {
 312         struct vmbus_channel *channel;
 313
 314         channel = kzalloc(sizeof(*channel), GFP_ATOMIC);
 315         if (!channel)
 316                 return NULL;
 317
 318         spin_lock_init(&channel->lock);
 319         init_completion(&channel->rescind_event);
 320
 321         INIT_LIST_HEAD(&channel->sc_list);
 322         INIT_LIST_HEAD(&channel->percpu_list);
 323
 324         tasklet_init(&channel->callback_event,
 325                      vmbus_on_event, (unsigned long)channel);
 326
 327         hv_ringbuffer_pre_init(channel);
 328
 329         return channel;
 330 }
 331
 332 /*
 333  * free_channel - Release the resources used by the vmbus channel object
 334  */
 335 static void free_channel(struct vmbus_channel *channel)
 336 {
 337         tasklet_kill(&channel->callback_event);
 338         vmbus_remove_channel_attr_group(channel);
 339
 340         kobject_put(&channel->kobj);
 341 }
 342
 343 static void percpu_channel_enq(void *arg)
 344 {
 345         struct vmbus_channel *channel = arg;
 346         struct hv_per_cpu_context *hv_cpu
 347                 = this_cpu_ptr(hv_context.cpu_context);
 348
 349         list_add_tail_rcu(&channel->percpu_list, &hv_cpu->chan_list);
 350 }
 351
 352 static void percpu_channel_deq(void *arg)
 353 {
 354         struct vmbus_channel *channel = arg;
 355
 356         list_del_rcu(&channel->percpu_list);
 357 }
 358
 359
 360 static void vmbus_release_relid(u32 relid)
 361 {
 362         struct vmbus_channel_relid_released msg;
 363         int ret;
 364
 365         memset(&msg, 0, sizeof(struct vmbus_channel_relid_released));
 366         msg.child_relid = relid;
 367         msg.header.msgtype = CHANNELMSG_RELID_RELEASED;
 368         ret = vmbus_post_msg(&msg, sizeof(struct vmbus_channel_relid_released),
 369                              true);
 370
 371         trace_vmbus_release_relid(&msg, ret);
 372 }
 373
 374 void hv_process_channel_removal(struct vmbus_channel *channel)
 375 {
 376         struct vmbus_channel *primary_channel;
 377         unsigned long flags;
 378
 379         BUG_ON(!mutex_is_locked(&vmbus_connection.channel_mutex));
 380         BUG_ON(!channel->rescind);
 381
 382         if (channel->target_cpu != get_cpu()) {
 383                 put_cpu();
 384                 smp_call_function_single(channel->target_cpu,
 385                                          percpu_channel_deq, channel, true);
 386         } else {
 387                 percpu_channel_deq(channel);
 388                 put_cpu();
 389         }
 390
 391         if (channel->primary_channel == NULL) {
 392                 list_del(&channel->listentry);
 393
 394                 primary_channel = channel;
 395         } else {
 396                 primary_channel = channel->primary_channel;
 397                 spin_lock_irqsave(&primary_channel->lock, flags);
 398                 list_del(&channel->sc_list);
 399                 spin_unlock_irqrestore(&primary_channel->lock, flags);
 400         }
 401
 402         /*
 403          * We need to free the bit for init_vp_index() to work in the case
 404          * of sub-channel, when we reload drivers like hv_netvsc.
 405          */
 406         if (channel->affinity_policy == HV_LOCALIZED)
 407                 cpumask_clear_cpu(channel->target_cpu,
 408                                   &primary_channel->alloced_cpus_in_node);
 409
 410         /*
 411          * Upon suspend, an in-use hv_sock channel is marked as "rescinded" and
 412          * the relid is invalidated; after hibernation, when the user-space app
 413          * destroys the channel, the relid is INVALID_RELID, and in this case
 414          * it's unnecessary and unsafe to release the old relid, since the same
 415          * relid can refer to a completely different channel now.
 416          */
 417         if (channel->offermsg.child_relid != INVALID_RELID)
 418                 vmbus_release_relid(channel->offermsg.child_relid);
 419
 420         free_channel(channel);
 421 }
 422
 423 void vmbus_free_channels(void)
 424 {
 425         struct vmbus_channel *channel, *tmp;
 426
 427         list_for_each_entry_safe(channel, tmp, &vmbus_connection.chn_list,
 428                 listentry) {
 429                 /* hv_process_channel_removal() needs this */
 430                 channel->rescind = true;
 431
 432                 vmbus_device_unregister(channel->device_obj);
 433         }
 434 }
 435
 436 /* Note: the function can run concurrently for primary/sub channels. */
 437 static void vmbus_add_channel_work(struct work_struct *work)
 438 {
 439         struct vmbus_channel *newchannel =
 440                 container_of(work, struct vmbus_channel, add_channel_work);
 441         struct vmbus_channel *primary_channel = newchannel->primary_channel;
 442         unsigned long flags;
 443         u16 dev_type;
 444         int ret;
 445
 446         dev_type = hv_get_dev_type(newchannel);
 447
 448         init_vp_index(newchannel, dev_type);
 449
 450         if (newchannel->target_cpu != get_cpu()) {
 451                 put_cpu();
 452                 smp_call_function_single(newchannel->target_cpu,
 453                                          percpu_channel_enq,
 454                                          newchannel, true);
 455         } else {
 456                 percpu_channel_enq(newchannel);
 457                 put_cpu();
 458         }
 459
 460         /*
 461          * This state is used to indicate a successful open
 462          * so that when we do close the channel normally, we
 463          * can cleanup properly.
 464          */
 465         newchannel->state = CHANNEL_OPEN_STATE;
 466
 467         if (primary_channel != NULL) {
 468                 /* newchannel is a sub-channel. */
 469                 struct hv_device *dev = primary_channel->device_obj;
 470
 471                 if (vmbus_add_channel_kobj(dev, newchannel))
 472                         goto err_deq_chan;
 473
 474                 if (primary_channel->sc_creation_callback != NULL)
 475                         primary_channel->sc_creation_callback(newchannel);
 476
 477                 newchannel->probe_done = true;
 478                 return;
 479         }
 480
 481         /*
 482          * Start the process of binding the primary channel to the driver
 483          */
 484         newchannel->device_obj = vmbus_device_create(
 485                 &newchannel->offermsg.offer.if_type,
 486                 &newchannel->offermsg.offer.if_instance,
 487                 newchannel);
 488         if (!newchannel->device_obj)
 489                 goto err_deq_chan;
 490
 491         newchannel->device_obj->device_id = dev_type;
 492         /*
 493          * Add the new device to the bus. This will kick off device-driver
 494          * binding which eventually invokes the device driver's AddDevice()
 495          * method.
 496          *
 497          * If vmbus_device_register() fails, the 'device_obj' is freed in
 498          * vmbus_device_release() as called by device_unregister() in the
 499          * error path of vmbus_device_register(). In the outside error
 500          * path, there's no need to free it.
 501          */
 502         ret = vmbus_device_register(newchannel->device_obj);
 503
 504         if (ret != 0) {
 505                 pr_err("unable to add child device object (relid %d)\n",
 506                         newchannel->offermsg.child_relid);
 507                 goto err_deq_chan;
 508         }
 509
 510         newchannel->probe_done = true;
 511         return;
 512
 513 err_deq_chan:
 514         mutex_lock(&vmbus_connection.channel_mutex);
 515
 516         /*
 517          * We need to set the flag, otherwise
 518          * vmbus_onoffer_rescind() can be blocked.
 519          */
 520         newchannel->probe_done = true;
 521
 522         if (primary_channel == NULL) {
 523                 list_del(&newchannel->listentry);
 524         } else {
 525                 spin_lock_irqsave(&primary_channel->lock, flags);
 526                 list_del(&newchannel->sc_list);
 527                 spin_unlock_irqrestore(&primary_channel->lock, flags);
 528         }
 529
 530         mutex_unlock(&vmbus_connection.channel_mutex);
 531
 532         if (newchannel->target_cpu != get_cpu()) {
 533                 put_cpu();
 534                 smp_call_function_single(newchannel->target_cpu,
 535                                          percpu_channel_deq,
 536                                          newchannel, true);
 537         } else {
 538                 percpu_channel_deq(newchannel);
 539                 put_cpu();
 540         }
 541
 542         vmbus_release_relid(newchannel->offermsg.child_relid);
 543
 544         free_channel(newchannel);
 545 }
 546
 547 /*
 548  * vmbus_process_offer - Process the offer by creating a channel/device
 549  * associated with this offer
 550  */
 551 static void vmbus_process_offer(struct vmbus_channel *newchannel)
 552 {
 553         struct vmbus_channel *channel;
 554         struct workqueue_struct *wq;
 555         unsigned long flags;
 556         bool fnew = true;
 557
 558         mutex_lock(&vmbus_connection.channel_mutex);
 559
 560         /* Remember the channels that should be cleaned up upon suspend. */
 561         if (is_hvsock_channel(newchannel) || is_sub_channel(newchannel))
 562                 atomic_inc(&vmbus_connection.nr_chan_close_on_suspend);
 563
 564         /*
 565          * Now that we have acquired the channel_mutex,
 566          * we can release the potentially racing rescind thread.
 567          */
 568         atomic_dec(&vmbus_connection.offer_in_progress);
 569
 570         list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
 571                 if (guid_equal(&channel->offermsg.offer.if_type,
 572                                &newchannel->offermsg.offer.if_type) &&
 573                     guid_equal(&channel->offermsg.offer.if_instance,
 574                                &newchannel->offermsg.offer.if_instance)) {
 575                         fnew = false;
 576                         break;
 577                 }
 578         }
 579
 580         if (fnew)
 581                 list_add_tail(&newchannel->listentry,
 582                               &vmbus_connection.chn_list);
 583         else {
 584                 /*
 585                  * Check to see if this is a valid sub-channel.
 586                  */
 587                 if (newchannel->offermsg.offer.sub_channel_index == 0) {
 588                         mutex_unlock(&vmbus_connection.channel_mutex);
 589                         /*
 590                          * Don't call free_channel(), because newchannel->kobj
 591                          * is not initialized yet.
 592                          */
 593                         kfree(newchannel);
 594                         WARN_ON_ONCE(1);
 595                         return;
 596                 }
 597                 /*
 598                  * Process the sub-channel.
 599                  */
 600                 newchannel->primary_channel = channel;
 601                 spin_lock_irqsave(&channel->lock, flags);
 602                 list_add_tail(&newchannel->sc_list, &channel->sc_list);
 603                 spin_unlock_irqrestore(&channel->lock, flags);
 604         }
 605
 606         mutex_unlock(&vmbus_connection.channel_mutex);
 607
 608         /*
 609          * vmbus_process_offer() mustn't call channel->sc_creation_callback()
 610          * directly for sub-channels, because sc_creation_callback() ->
 611          * vmbus_open() may never get the host's response to the
 612          * OPEN_CHANNEL message (the host may rescind a channel at any time,
 613          * e.g. in the case of hot removing a NIC), and vmbus_onoffer_rescind()
 614          * may not wake up the vmbus_open() as it's blocked due to a non-zero
 615          * vmbus_connection.offer_in_progress, and finally we have a deadlock.
 616          *
 617          * The above is also true for primary channels, if the related device
 618          * drivers use sync probing mode by default.
 619          *
 620          * And, usually the handling of primary channels and sub-channels can
 621          * depend on each other, so we should offload them to different
 622          * workqueues to avoid possible deadlock, e.g. in sync-probing mode,
 623          * NIC1's netvsc_subchan_work() can race with NIC2's netvsc_probe() ->
 624          * rtnl_lock(), and causes deadlock: the former gets the rtnl_lock
 625          * and waits for all the sub-channels to appear, but the latter
 626          * can't get the rtnl_lock and this blocks the handling of
 627          * sub-channels.
 628          */
 629         INIT_WORK(&newchannel->add_channel_work, vmbus_add_channel_work);
 630         wq = fnew ? vmbus_connection.handle_primary_chan_wq :
 631                     vmbus_connection.handle_sub_chan_wq;
 632         queue_work(wq, &newchannel->add_channel_work);
 633 }
 634
 635 /*
 636  * We use this state to statically distribute the channel interrupt load.
 637  */
 638 static int next_numa_node_id;
 639 /*
 640  * init_vp_index() accesses global variables like next_numa_node_id, and
 641  * it can run concurrently for primary channels and sub-channels: see
 642  * vmbus_process_offer(), so we need the lock to protect the global
 643  * variables.
 644  */
 645 static DEFINE_SPINLOCK(bind_channel_to_cpu_lock);
 646
 647 /*
 648  * Starting with Win8, we can statically distribute the incoming
 649  * channel interrupt load by binding a channel to VCPU.
 650  * We distribute the interrupt loads to one or more NUMA nodes based on
 651  * the channel's affinity_policy.
 652  *
 653  * For pre-win8 hosts or non-performance critical channels we assign the
 654  * first CPU in the first NUMA node.
 655  */
 656 static void init_vp_index(struct vmbus_channel *channel, u16 dev_type)
 657 {
 658         u32 cur_cpu;
 659         bool perf_chn = vmbus_devs[dev_type].perf_device;
 660         struct vmbus_channel *primary = channel->primary_channel;
 661         int next_node;
 662         cpumask_var_t available_mask;
 663         struct cpumask *alloced_mask;
 664
 665         if ((vmbus_proto_version == VERSION_WS2008) ||
 666             (vmbus_proto_version == VERSION_WIN7) || (!perf_chn) ||
 667             !alloc_cpumask_var(&available_mask, GFP_KERNEL)) {
 668                 /*
 669                  * Prior to win8, all channel interrupts are
 670                  * delivered on cpu 0.
 671                  * Also if the channel is not a performance critical
 672                  * channel, bind it to cpu 0.
 673                  * In case alloc_cpumask_var() fails, bind it to cpu 0.
 674                  */
 675                 channel->numa_node = 0;
 676                 channel->target_cpu = 0;
 677                 channel->target_vp = hv_cpu_number_to_vp_number(0);
 678                 return;
 679         }
 680
 681         spin_lock(&bind_channel_to_cpu_lock);
 682
 683         /*
 684          * Based on the channel affinity policy, we will assign the NUMA
 685          * nodes.
 686          */
 687
 688         if ((channel->affinity_policy == HV_BALANCED) || (!primary)) {
 689                 while (true) {
 690                         next_node = next_numa_node_id++;
 691                         if (next_node == nr_node_ids) {
 692                                 next_node = next_numa_node_id = 0;
 693                                 continue;
 694                         }
 695                         if (cpumask_empty(cpumask_of_node(next_node)))
 696                                 continue;
 697                         break;
 698                 }
 699                 channel->numa_node = next_node;
 700                 primary = channel;
 701         }
 702         alloced_mask = &hv_context.hv_numa_map[primary->numa_node];
 703
 704         if (cpumask_weight(alloced_mask) ==
 705             cpumask_weight(cpumask_of_node(primary->numa_node))) {
 706                 /*
 707                  * We have cycled through all the CPUs in the node;
 708                  * reset the alloced map.
 709                  */
 710                 cpumask_clear(alloced_mask);
 711         }
 712
 713         cpumask_xor(available_mask, alloced_mask,
 714                     cpumask_of_node(primary->numa_node));
 715
 716         cur_cpu = -1;
 717
 718         if (primary->affinity_policy == HV_LOCALIZED) {
 719                 /*
 720                  * Normally Hyper-V host doesn't create more subchannels
 721                  * than there are VCPUs on the node but it is possible when not
 722                  * all present VCPUs on the node are initialized by guest.
 723                  * Clear the alloced_cpus_in_node to start over.
 724                  */
 725                 if (cpumask_equal(&primary->alloced_cpus_in_node,
 726                                   cpumask_of_node(primary->numa_node)))
 727                         cpumask_clear(&primary->alloced_cpus_in_node);
 728         }
 729
 730         while (true) {
 731                 cur_cpu = cpumask_next(cur_cpu, available_mask);
 732                 if (cur_cpu >= nr_cpu_ids) {
 733                         cur_cpu = -1;
 734                         cpumask_copy(available_mask,
 735                                      cpumask_of_node(primary->numa_node));
 736                         continue;
 737                 }
 738
 739                 if (primary->affinity_policy == HV_LOCALIZED) {
 740                         /*
 741                          * NOTE: in the case of sub-channel, we clear the
 742                          * sub-channel related bit(s) in
 743                          * primary->alloced_cpus_in_node in
 744                          * hv_process_channel_removal(), so when we
 745                          * reload drivers like hv_netvsc in SMP guest, here
 746                          * we're able to re-allocate
 747                          * bit from primary->alloced_cpus_in_node.
 748                          */
 749                         if (!cpumask_test_cpu(cur_cpu,
 750                                               &primary->alloced_cpus_in_node)) {
 751                                 cpumask_set_cpu(cur_cpu,
 752                                                 &primary->alloced_cpus_in_node);
 753                                 cpumask_set_cpu(cur_cpu, alloced_mask);
 754                                 break;
 755                         }
 756                 } else {
 757                         cpumask_set_cpu(cur_cpu, alloced_mask);
 758                         break;
 759                 }
 760         }
 761
 762         channel->target_cpu = cur_cpu;
 763         channel->target_vp = hv_cpu_number_to_vp_number(cur_cpu);
 764
 765         spin_unlock(&bind_channel_to_cpu_lock);
 766
 767         free_cpumask_var(available_mask);
 768 }
 769
 770 #define UNLOAD_DELAY_UNIT_MS    10              /* 10 milliseconds */
 771 #define UNLOAD_WAIT_MS          (100*1000)      /* 100 seconds */
 772 #define UNLOAD_WAIT_LOOPS       (UNLOAD_WAIT_MS/UNLOAD_DELAY_UNIT_MS)
 773 #define UNLOAD_MSG_MS           (5*1000)        /* Every 5 seconds */
 774 #define UNLOAD_MSG_LOOPS        (UNLOAD_MSG_MS/UNLOAD_DELAY_UNIT_MS)
 775
 776 static void vmbus_wait_for_unload(void)
 777 {
 778         int cpu;
 779         void *page_addr;
 780         struct hv_message *msg;
 781         struct vmbus_channel_message_header *hdr;
 782         u32 message_type, i;
 783
 784         /*
 785          * CHANNELMSG_UNLOAD_RESPONSE is always delivered to the CPU which was
 786          * used for initial contact or to CPU0 depending on host version. When
 787          * we're crashing on a different CPU let's hope that IRQ handler on
 788          * the cpu which receives CHANNELMSG_UNLOAD_RESPONSE is still
 789          * functional and vmbus_unload_response() will complete
 790          * vmbus_connection.unload_event. If not, the last thing we can do is
 791          * read message pages for all CPUs directly.
 792          *
 793          * Wait up to 100 seconds since an Azure host must writeback any dirty
 794          * data in its disk cache before the VMbus UNLOAD request will
 795          * complete. This flushing has been empirically observed to take up
 796          * to 50 seconds in cases with a lot of dirty data, so allow additional
 797          * leeway and for inaccuracies in mdelay(). But eventually time out so
 798          * that the panic path can't get hung forever in case the response
 799          * message isn't seen.
 800          */
 801         for (i = 1; i <= UNLOAD_WAIT_LOOPS; i++) {
 802                 if (completion_done(&vmbus_connection.unload_event))
 803                         goto completed;
 804
 805                 for_each_online_cpu(cpu) {
 806                         struct hv_per_cpu_context *hv_cpu
 807                                 = per_cpu_ptr(hv_context.cpu_context, cpu);
 808
 809                         page_addr = hv_cpu->synic_message_page;
 810                         msg = (struct hv_message *)page_addr
 811                                 + VMBUS_MESSAGE_SINT;
 812
 813                         message_type = READ_ONCE(msg->header.message_type);
 814                         if (message_type == HVMSG_NONE)
 815                                 continue;
 816
 817                         hdr = (struct vmbus_channel_message_header *)
 818                                 msg->u.payload;
 819
 820                         if (hdr->msgtype == CHANNELMSG_UNLOAD_RESPONSE)
 821                                 complete(&vmbus_connection.unload_event);
 822
 823                         vmbus_signal_eom(msg, message_type);
 824                 }
 825
 826                 /*
 827                  * Give a notice periodically so someone watching the
 828                  * serial output won't think it is completely hung.
 829                  */
 830                 if (!(i % UNLOAD_MSG_LOOPS))
 831                         pr_notice("Waiting for VMBus UNLOAD to complete\n");
 832
 833                 mdelay(UNLOAD_DELAY_UNIT_MS);
 834         }
 835         pr_err("Continuing even though VMBus UNLOAD did not complete\n");
 836
 837 completed:
 838         /*
 839          * We're crashing and already got the UNLOAD_RESPONSE, cleanup all
 840          * maybe-pending messages on all CPUs to be able to receive new
 841          * messages after we reconnect.
 842          */
 843         for_each_online_cpu(cpu) {
 844                 struct hv_per_cpu_context *hv_cpu
 845                         = per_cpu_ptr(hv_context.cpu_context, cpu);
 846
 847                 page_addr = hv_cpu->synic_message_page;
 848                 msg = (struct hv_message *)page_addr + VMBUS_MESSAGE_SINT;
 849                 msg->header.message_type = HVMSG_NONE;
 850         }
 851 }
 852
 853 /*
 854  * vmbus_unload_response - Handler for the unload response.
 855  */
 856 static void vmbus_unload_response(struct vmbus_channel_message_header *hdr)
 857 {
 858         /*
 859          * This is a global event; just wakeup the waiting thread.
 860          * Once we successfully unload, we can cleanup the monitor state.
 861          */
 862         complete(&vmbus_connection.unload_event);
 863 }
 864
 865 void vmbus_initiate_unload(bool crash)
 866 {
 867         struct vmbus_channel_message_header hdr;
 868
 869         if (xchg(&vmbus_connection.conn_state, DISCONNECTED) == DISCONNECTED)
 870                 return;
 871
 872         /* Pre-Win2012R2 hosts don't support reconnect */
 873         if (vmbus_proto_version < VERSION_WIN8_1)
 874                 return;
 875
 876         init_completion(&vmbus_connection.unload_event);
 877         memset(&hdr, 0, sizeof(struct vmbus_channel_message_header));
 878         hdr.msgtype = CHANNELMSG_UNLOAD;
 879         vmbus_post_msg(&hdr, sizeof(struct vmbus_channel_message_header),
 880                        !crash);
 881
 882         /*
 883          * vmbus_initiate_unload() is also called on crash and the crash can be
 884          * happening in an interrupt context, where scheduling is impossible.
 885          */
 886         if (!crash)
 887                 wait_for_completion(&vmbus_connection.unload_event);
 888         else
 889                 vmbus_wait_for_unload();
 890 }
 891
 892 static void check_ready_for_resume_event(void)
 893 {
 894         /*
 895          * If all the old primary channels have been fixed up, then it's safe
 896          * to resume.
 897          */
 898         if (atomic_dec_and_test(&vmbus_connection.nr_chan_fixup_on_resume))
 899                 complete(&vmbus_connection.ready_for_resume_event);
 900 }
 901
 902 static void vmbus_setup_channel_state(struct vmbus_channel *channel,
 903                                       struct vmbus_channel_offer_channel *offer)
 904 {
 905         /*
 906          * Setup state for signalling the host.
 907          */
 908         channel->sig_event = VMBUS_EVENT_CONNECTION_ID;
 909
 910         if (vmbus_proto_version != VERSION_WS2008) {
 911                 channel->is_dedicated_interrupt =
 912                                 (offer->is_dedicated_interrupt != 0);
 913                 channel->sig_event = offer->connection_id;
 914         }
 915
 916         memcpy(&channel->offermsg, offer,
 917                sizeof(struct vmbus_channel_offer_channel));
 918         channel->monitor_grp = (u8)offer->monitorid / 32;
 919         channel->monitor_bit = (u8)offer->monitorid % 32;
 920 }
 921
 922 /*
 923  * find_primary_channel_by_offer - Get the channel object given the new offer.
 924  * This is only used in the resume path of hibernation.
 925  */
 926 static struct vmbus_channel *
 927 find_primary_channel_by_offer(const struct vmbus_channel_offer_channel *offer)
 928 {
 929         struct vmbus_channel *channel = NULL, *iter;
 930         const guid_t *inst1, *inst2;
 931
 932         /* Ignore sub-channel offers. */
 933         if (offer->offer.sub_channel_index != 0)
 934                 return NULL;
 935
 936         mutex_lock(&vmbus_connection.channel_mutex);
 937
 938         list_for_each_entry(iter, &vmbus_connection.chn_list, listentry) {
 939                 inst1 = &iter->offermsg.offer.if_instance;
 940                 inst2 = &offer->offer.if_instance;
 941
 942                 if (guid_equal(inst1, inst2)) {
 943                         channel = iter;
 944                         break;
 945                 }
 946         }
 947
 948         mutex_unlock(&vmbus_connection.channel_mutex);
 949
 950         return channel;
 951 }
 952
 953 /*
 954  * vmbus_onoffer - Handler for channel offers from vmbus in parent partition.
 955  *
 956  */
 957 static void vmbus_onoffer(struct vmbus_channel_message_header *hdr)
 958 {
 959         struct vmbus_channel_offer_channel *offer;
 960         struct vmbus_channel *oldchannel, *newchannel;
 961         size_t offer_sz;
 962
 963         offer = (struct vmbus_channel_offer_channel *)hdr;
 964
 965         trace_vmbus_onoffer(offer);
 966
 967         oldchannel = find_primary_channel_by_offer(offer);
 968
 969         if (oldchannel != NULL) {
 970                 atomic_dec(&vmbus_connection.offer_in_progress);
 971
 972                 /*
 973                  * We're resuming from hibernation: all the sub-channel and
 974                  * hv_sock channels we had before the hibernation should have
 975                  * been cleaned up, and now we must be seeing a re-offered
 976                  * primary channel that we had before the hibernation.
 977                  */
 978
 979                 WARN_ON(oldchannel->offermsg.child_relid != INVALID_RELID);
 980                 /* Fix up the relid. */
 981                 oldchannel->offermsg.child_relid = offer->child_relid;
 982
 983                 offer_sz = sizeof(*offer);
 984                 if (memcmp(offer, &oldchannel->offermsg, offer_sz) == 0) {
 985                         check_ready_for_resume_event();
 986                         return;
 987                 }
 988
 989                 /*
 990                  * This is not an error, since the host can also change the
 991                  * other field(s) of the offer, e.g. on WS RS5 (Build 17763),
 992                  * the offer->connection_id of the Mellanox VF vmbus device
 993                  * can change when the host reoffers the device upon resume.
 994                  */
 995                 pr_debug("vmbus offer changed: relid=%d\n",
 996                          offer->child_relid);
 997
 998                 print_hex_dump_debug("Old vmbus offer: ", DUMP_PREFIX_OFFSET,
 999                                      16, 4, &oldchannel->offermsg, offer_sz,
1000                                      false);
1001                 print_hex_dump_debug("New vmbus offer: ", DUMP_PREFIX_OFFSET,
1002                                      16, 4, offer, offer_sz, false);
1003
1004                 /* Fix up the old channel. */
1005                 vmbus_setup_channel_state(oldchannel, offer);
1006
1007                 check_ready_for_resume_event();
1008
1009                 return;
1010         }
1011
1012         /* Allocate the channel object and save this offer. */
1013         newchannel = alloc_channel();
1014         if (!newchannel) {
1015                 vmbus_release_relid(offer->child_relid);
1016                 atomic_dec(&vmbus_connection.offer_in_progress);
1017                 pr_err("Unable to allocate channel object\n");
1018                 return;
1019         }
1020
1021         vmbus_setup_channel_state(newchannel, offer);
1022
1023         vmbus_process_offer(newchannel);
1024 }
1025
1026 static void check_ready_for_suspend_event(void)
1027 {
1028         /*
1029          * If all the sub-channels or hv_sock channels have been cleaned up,
1030          * then it's safe to suspend.
1031          */
1032         if (atomic_dec_and_test(&vmbus_connection.nr_chan_close_on_suspend))
1033                 complete(&vmbus_connection.ready_for_suspend_event);
1034 }
1035
1036 /*
1037  * vmbus_onoffer_rescind - Rescind offer handler.
1038  *
1039  * We queue a work item to process this offer synchronously
1040  */
1041 static void vmbus_onoffer_rescind(struct vmbus_channel_message_header *hdr)
1042 {
1043         struct vmbus_channel_rescind_offer *rescind;
1044         struct vmbus_channel *channel;
1045         struct device *dev;
1046         bool clean_up_chan_for_suspend;
1047
1048         rescind = (struct vmbus_channel_rescind_offer *)hdr;
1049
1050         trace_vmbus_onoffer_rescind(rescind);
1051
1052         /*
1053          * The offer msg and the corresponding rescind msg
1054          * from the host are guranteed to be ordered -
1055          * offer comes in first and then the rescind.
1056          * Since we process these events in work elements,
1057          * and with preemption, we may end up processing
1058          * the events out of order. Given that we handle these
1059          * work elements on the same CPU, this is possible only
1060          * in the case of preemption. In any case wait here
1061          * until the offer processing has moved beyond the
1062          * point where the channel is discoverable.
1063          */
1064
1065         while (atomic_read(&vmbus_connection.offer_in_progress) != 0) {
1066                 /*
1067                  * We wait here until any channel offer is currently
1068                  * being processed.
1069                  */
1070                 msleep(1);
1071         }
1072
1073         mutex_lock(&vmbus_connection.channel_mutex);
1074         channel = relid2channel(rescind->child_relid);
1075         mutex_unlock(&vmbus_connection.channel_mutex);
1076
1077         if (channel == NULL) {
1078                 /*
1079                  * We failed in processing the offer message;
1080                  * we would have cleaned up the relid in that
1081                  * failure path.
1082                  */
1083                 return;
1084         }
1085
1086         clean_up_chan_for_suspend = is_hvsock_channel(channel) ||
1087                                     is_sub_channel(channel);
1088         /*
1089          * Before setting channel->rescind in vmbus_rescind_cleanup(), we
1090          * should make sure the channel callback is not running any more.
1091          */
1092         vmbus_reset_channel_cb(channel);
1093
1094         /*
1095          * Now wait for offer handling to complete.
1096          */
1097         vmbus_rescind_cleanup(channel);
1098         while (READ_ONCE(channel->probe_done) == false) {
1099                 /*
1100                  * We wait here until any channel offer is currently
1101                  * being processed.
1102                  */
1103                 msleep(1);
1104         }
1105
1106         /*
1107          * At this point, the rescind handling can proceed safely.
1108          */
1109
1110         if (channel->device_obj) {
1111                 if (channel->chn_rescind_callback) {
1112                         channel->chn_rescind_callback(channel);
1113
1114                         if (clean_up_chan_for_suspend)
1115                                 check_ready_for_suspend_event();
1116
1117                         return;
1118                 }
1119                 /*
1120                  * We will have to unregister this device from the
1121                  * driver core.
1122                  */
1123                 dev = get_device(&channel->device_obj->device);
1124                 if (dev) {
1125                         vmbus_device_unregister(channel->device_obj);
1126                         put_device(dev);
1127                 }
1128         } else if (channel->primary_channel != NULL) {
1129                 /*
1130                  * Sub-channel is being rescinded. Following is the channel
1131                  * close sequence when initiated from the driveri (refer to
1132                  * vmbus_close() for details):
1133                  * 1. Close all sub-channels first
1134                  * 2. Then close the primary channel.
1135                  */
1136                 mutex_lock(&vmbus_connection.channel_mutex);
1137                 if (channel->state == CHANNEL_OPEN_STATE) {
1138                         /*
1139                          * The channel is currently not open;
1140                          * it is safe for us to cleanup the channel.
1141                          */
1142                         hv_process_channel_removal(channel);
1143                 } else {
1144                         complete(&channel->rescind_event);
1145                 }
1146                 mutex_unlock(&vmbus_connection.channel_mutex);
1147         }
1148
1149         /* The "channel" may have been freed. Do not access it any longer. */
1150
1151         if (clean_up_chan_for_suspend)
1152                 check_ready_for_suspend_event();
1153 }
1154
1155 void vmbus_hvsock_device_unregister(struct vmbus_channel *channel)
1156 {
1157         BUG_ON(!is_hvsock_channel(channel));
1158
1159         /* We always get a rescind msg when a connection is closed. */
1160         while (!READ_ONCE(channel->probe_done) || !READ_ONCE(channel->rescind))
1161                 msleep(1);
1162
1163         vmbus_device_unregister(channel->device_obj);
1164 }
1165 EXPORT_SYMBOL_GPL(vmbus_hvsock_device_unregister);
1166
1167
1168 /*
1169  * vmbus_onoffers_delivered -
1170  * This is invoked when all offers have been delivered.
1171  *
1172  * Nothing to do here.
1173  */
1174 static void vmbus_onoffers_delivered(
1175                         struct vmbus_channel_message_header *hdr)
1176 {
1177 }
1178
1179 /*
1180  * vmbus_onopen_result - Open result handler.
1181  *
1182  * This is invoked when we received a response to our channel open request.
1183  * Find the matching request, copy the response and signal the requesting
1184  * thread.
1185  */
1186 static void vmbus_onopen_result(struct vmbus_channel_message_header *hdr)
1187 {
1188         struct vmbus_channel_open_result *result;
1189         struct vmbus_channel_msginfo *msginfo;
1190         struct vmbus_channel_message_header *requestheader;
1191         struct vmbus_channel_open_channel *openmsg;
1192         unsigned long flags;
1193
1194         result = (struct vmbus_channel_open_result *)hdr;
1195
1196         trace_vmbus_onopen_result(result);
1197
1198         /*
1199          * Find the open msg, copy the result and signal/unblock the wait event
1200          */
1201         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1202
1203         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1204                                 msglistentry) {
1205                 requestheader =
1206                         (struct vmbus_channel_message_header *)msginfo->msg;
1207
1208                 if (requestheader->msgtype == CHANNELMSG_OPENCHANNEL) {
1209                         openmsg =
1210                         (struct vmbus_channel_open_channel *)msginfo->msg;
1211                         if (openmsg->child_relid == result->child_relid &&
1212                             openmsg->openid == result->openid) {
1213                                 memcpy(&msginfo->response.open_result,
1214                                        result,
1215                                        sizeof(
1216                                         struct vmbus_channel_open_result));
1217                                 complete(&msginfo->waitevent);
1218                                 break;
1219                         }
1220                 }
1221         }
1222         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1223 }
1224
1225 /*
1226  * vmbus_ongpadl_created - GPADL created handler.
1227  *
1228  * This is invoked when we received a response to our gpadl create request.
1229  * Find the matching request, copy the response and signal the requesting
1230  * thread.
1231  */
1232 static void vmbus_ongpadl_created(struct vmbus_channel_message_header *hdr)
1233 {
1234         struct vmbus_channel_gpadl_created *gpadlcreated;
1235         struct vmbus_channel_msginfo *msginfo;
1236         struct vmbus_channel_message_header *requestheader;
1237         struct vmbus_channel_gpadl_header *gpadlheader;
1238         unsigned long flags;
1239
1240         gpadlcreated = (struct vmbus_channel_gpadl_created *)hdr;
1241
1242         trace_vmbus_ongpadl_created(gpadlcreated);
1243
1244         /*
1245          * Find the establish msg, copy the result and signal/unblock the wait
1246          * event
1247          */
1248         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1249
1250         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1251                                 msglistentry) {
1252                 requestheader =
1253                         (struct vmbus_channel_message_header *)msginfo->msg;
1254
1255                 if (requestheader->msgtype == CHANNELMSG_GPADL_HEADER) {
1256                         gpadlheader =
1257                         (struct vmbus_channel_gpadl_header *)requestheader;
1258
1259                         if ((gpadlcreated->child_relid ==
1260                              gpadlheader->child_relid) &&
1261                             (gpadlcreated->gpadl == gpadlheader->gpadl)) {
1262                                 memcpy(&msginfo->response.gpadl_created,
1263                                        gpadlcreated,
1264                                        sizeof(
1265                                         struct vmbus_channel_gpadl_created));
1266                                 complete(&msginfo->waitevent);
1267                                 break;
1268                         }
1269                 }
1270         }
1271         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1272 }
1273
1274 /*
1275  * vmbus_ongpadl_torndown - GPADL torndown handler.
1276  *
1277  * This is invoked when we received a response to our gpadl teardown request.
1278  * Find the matching request, copy the response and signal the requesting
1279  * thread.
1280  */
1281 static void vmbus_ongpadl_torndown(
1282                         struct vmbus_channel_message_header *hdr)
1283 {
1284         struct vmbus_channel_gpadl_torndown *gpadl_torndown;
1285         struct vmbus_channel_msginfo *msginfo;
1286         struct vmbus_channel_message_header *requestheader;
1287         struct vmbus_channel_gpadl_teardown *gpadl_teardown;
1288         unsigned long flags;
1289
1290         gpadl_torndown = (struct vmbus_channel_gpadl_torndown *)hdr;
1291
1292         trace_vmbus_ongpadl_torndown(gpadl_torndown);
1293
1294         /*
1295          * Find the open msg, copy the result and signal/unblock the wait event
1296          */
1297         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1298
1299         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1300                                 msglistentry) {
1301                 requestheader =
1302                         (struct vmbus_channel_message_header *)msginfo->msg;
1303
1304                 if (requestheader->msgtype == CHANNELMSG_GPADL_TEARDOWN) {
1305                         gpadl_teardown =
1306                         (struct vmbus_channel_gpadl_teardown *)requestheader;
1307
1308                         if (gpadl_torndown->gpadl == gpadl_teardown->gpadl) {
1309                                 memcpy(&msginfo->response.gpadl_torndown,
1310                                        gpadl_torndown,
1311                                        sizeof(
1312                                         struct vmbus_channel_gpadl_torndown));
1313                                 complete(&msginfo->waitevent);
1314                                 break;
1315                         }
1316                 }
1317         }
1318         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1319 }
1320
1321 /*
1322  * vmbus_onversion_response - Version response handler
1323  *
1324  * This is invoked when we received a response to our initiate contact request.
1325  * Find the matching request, copy the response and signal the requesting
1326  * thread.
1327  */
1328 static void vmbus_onversion_response(
1329                 struct vmbus_channel_message_header *hdr)
1330 {
1331         struct vmbus_channel_msginfo *msginfo;
1332         struct vmbus_channel_message_header *requestheader;
1333         struct vmbus_channel_version_response *version_response;
1334         unsigned long flags;
1335
1336         version_response = (struct vmbus_channel_version_response *)hdr;
1337
1338         trace_vmbus_onversion_response(version_response);
1339
1340         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1341
1342         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1343                                 msglistentry) {
1344                 requestheader =
1345                         (struct vmbus_channel_message_header *)msginfo->msg;
1346
1347                 if (requestheader->msgtype ==
1348                     CHANNELMSG_INITIATE_CONTACT) {
1349                         memcpy(&msginfo->response.version_response,
1350                               version_response,
1351                               sizeof(struct vmbus_channel_version_response));
1352                         complete(&msginfo->waitevent);
1353                 }
1354         }
1355         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1356 }
1357
1358 /* Channel message dispatch table */
1359 const struct vmbus_channel_message_table_entry
1360 channel_message_table[CHANNELMSG_COUNT] = {
1361         { CHANNELMSG_INVALID,                   0, NULL },
1362         { CHANNELMSG_OFFERCHANNEL,              0, vmbus_onoffer },
1363         { CHANNELMSG_RESCIND_CHANNELOFFER,      0, vmbus_onoffer_rescind },
1364         { CHANNELMSG_REQUESTOFFERS,             0, NULL },
1365         { CHANNELMSG_ALLOFFERS_DELIVERED,       1, vmbus_onoffers_delivered },
1366         { CHANNELMSG_OPENCHANNEL,               0, NULL },
1367         { CHANNELMSG_OPENCHANNEL_RESULT,        1, vmbus_onopen_result },
1368         { CHANNELMSG_CLOSECHANNEL,              0, NULL },
1369         { CHANNELMSG_GPADL_HEADER,              0, NULL },
1370         { CHANNELMSG_GPADL_BODY,                0, NULL },
1371         { CHANNELMSG_GPADL_CREATED,             1, vmbus_ongpadl_created },
1372         { CHANNELMSG_GPADL_TEARDOWN,            0, NULL },
1373         { CHANNELMSG_GPADL_TORNDOWN,            1, vmbus_ongpadl_torndown },
1374         { CHANNELMSG_RELID_RELEASED,            0, NULL },
1375         { CHANNELMSG_INITIATE_CONTACT,          0, NULL },
1376         { CHANNELMSG_VERSION_RESPONSE,          1, vmbus_onversion_response },
1377         { CHANNELMSG_UNLOAD,                    0, NULL },
1378         { CHANNELMSG_UNLOAD_RESPONSE,           1, vmbus_unload_response },
1379         { CHANNELMSG_18,                        0, NULL },
1380         { CHANNELMSG_19,                        0, NULL },
1381         { CHANNELMSG_20,                        0, NULL },
1382         { CHANNELMSG_TL_CONNECT_REQUEST,        0, NULL },
1383         { CHANNELMSG_22,                        0, NULL },
1384         { CHANNELMSG_TL_CONNECT_RESULT,         0, NULL },
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(void *context)
1393 {
1394         struct hv_message *msg = context;
1395         struct vmbus_channel_message_header *hdr;
1396
1397         hdr = (struct vmbus_channel_message_header *)msg->u.payload;
1398
1399         trace_vmbus_on_message(hdr);
1400
1401         /*
1402          * vmbus_on_msg_dpc() makes sure the hdr->msgtype here can not go
1403          * out of bound and the message_handler pointer can not be NULL.
1404          */
1405         channel_message_table[hdr->msgtype].message_handler(hdr);
1406 }
1407
1408 /*
1409  * vmbus_request_offers - Send a request to get all our pending offers.
1410  */
1411 int vmbus_request_offers(void)
1412 {
1413         struct vmbus_channel_message_header *msg;
1414         struct vmbus_channel_msginfo *msginfo;
1415         int ret;
1416
1417         msginfo = kmalloc(sizeof(*msginfo) +
1418                           sizeof(struct vmbus_channel_message_header),
1419                           GFP_KERNEL);
1420         if (!msginfo)
1421                 return -ENOMEM;
1422
1423         msg = (struct vmbus_channel_message_header *)msginfo->msg;
1424
1425         msg->msgtype = CHANNELMSG_REQUESTOFFERS;
1426
1427         ret = vmbus_post_msg(msg, sizeof(struct vmbus_channel_message_header),
1428                              true);
1429
1430         trace_vmbus_request_offers(ret);
1431
1432         if (ret != 0) {
1433                 pr_err("Unable to request offers - %d\n", ret);
1434
1435                 goto cleanup;
1436         }
1437
1438 cleanup:
1439         kfree(msginfo);
1440
1441         return ret;
1442 }
1443
1444 static void invoke_sc_cb(struct vmbus_channel *primary_channel)
1445 {
1446         struct list_head *cur, *tmp;
1447         struct vmbus_channel *cur_channel;
1448
1449         if (primary_channel->sc_creation_callback == NULL)
1450                 return;
1451
1452         list_for_each_safe(cur, tmp, &primary_channel->sc_list) {
1453                 cur_channel = list_entry(cur, struct vmbus_channel, sc_list);
1454
1455                 primary_channel->sc_creation_callback(cur_channel);
1456         }
1457 }
1458
1459 void vmbus_set_sc_create_callback(struct vmbus_channel *primary_channel,
1460                                 void (*sc_cr_cb)(struct vmbus_channel *new_sc))
1461 {
1462         primary_channel->sc_creation_callback = sc_cr_cb;
1463 }
1464 EXPORT_SYMBOL_GPL(vmbus_set_sc_create_callback);
1465
1466 bool vmbus_are_subchannels_present(struct vmbus_channel *primary)
1467 {
1468         bool ret;
1469
1470         ret = !list_empty(&primary->sc_list);
1471
1472         if (ret) {
1473                 /*
1474                  * Invoke the callback on sub-channel creation.
1475                  * This will present a uniform interface to the
1476                  * clients.
1477                  */
1478                 invoke_sc_cb(primary);
1479         }
1480
1481         return ret;
1482 }
1483 EXPORT_SYMBOL_GPL(vmbus_are_subchannels_present);
1484
1485 void vmbus_set_chn_rescind_callback(struct vmbus_channel *channel,
1486                 void (*chn_rescind_cb)(struct vmbus_channel *))
1487 {
1488         channel->chn_rescind_callback = chn_rescind_cb;
1489 }
1490 EXPORT_SYMBOL_GPL(vmbus_set_chn_rescind_callback);