drivers/hv/channel_mgmt.c

   1 /*
   2  * Copyright (c) 2009, Microsoft Corporation.
   3  *
   4  * This program is free software; you can redistribute it and/or modify it
   5  * under the terms and conditions of the GNU General Public License,
   6  * version 2, as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope it will be useful, but WITHOUT
   9  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  10  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  11  * more details.
  12  *
  13  * You should have received a copy of the GNU General Public License along with
  14  * this program; if not, write to the Free Software Foundation, Inc., 59 Temple
  15  * Place - Suite 330, Boston, MA 02111-1307 USA.
  16  *
  17  * Authors:
  18  *   Haiyang Zhang <haiyangz@microsoft.com>
  19  *   Hank Janssen  <hjanssen@microsoft.com>
  20  */
  21 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  22
  23 #include <linux/kernel.h>
  24 #include <linux/interrupt.h>
  25 #include <linux/sched.h>
  26 #include <linux/wait.h>
  27 #include <linux/mm.h>
  28 #include <linux/slab.h>
  29 #include <linux/list.h>
  30 #include <linux/module.h>
  31 #include <linux/completion.h>
  32 #include <linux/delay.h>
  33 #include <linux/hyperv.h>
  34 #include <asm/mshyperv.h>
  35
  36 #include "hyperv_vmbus.h"
  37
  38 static void init_vp_index(struct vmbus_channel *channel, u16 dev_type);
  39
  40 static const struct vmbus_device vmbus_devs[] = {
  41         /* IDE */
  42         { .dev_type = HV_IDE,
  43           HV_IDE_GUID,
  44           .perf_device = true,
  45         },
  46
  47         /* SCSI */
  48         { .dev_type = HV_SCSI,
  49           HV_SCSI_GUID,
  50           .perf_device = true,
  51         },
  52
  53         /* Fibre Channel */
  54         { .dev_type = HV_FC,
  55           HV_SYNTHFC_GUID,
  56           .perf_device = true,
  57         },
  58
  59         /* Synthetic NIC */
  60         { .dev_type = HV_NIC,
  61           HV_NIC_GUID,
  62           .perf_device = true,
  63         },
  64
  65         /* Network Direct */
  66         { .dev_type = HV_ND,
  67           HV_ND_GUID,
  68           .perf_device = true,
  69         },
  70
  71         /* PCIE */
  72         { .dev_type = HV_PCIE,
  73           HV_PCIE_GUID,
  74           .perf_device = false,
  75         },
  76
  77         /* Synthetic Frame Buffer */
  78         { .dev_type = HV_FB,
  79           HV_SYNTHVID_GUID,
  80           .perf_device = false,
  81         },
  82
  83         /* Synthetic Keyboard */
  84         { .dev_type = HV_KBD,
  85           HV_KBD_GUID,
  86           .perf_device = false,
  87         },
  88
  89         /* Synthetic MOUSE */
  90         { .dev_type = HV_MOUSE,
  91           HV_MOUSE_GUID,
  92           .perf_device = false,
  93         },
  94
  95         /* KVP */
  96         { .dev_type = HV_KVP,
  97           HV_KVP_GUID,
  98           .perf_device = false,
  99         },
 100
 101         /* Time Synch */
 102         { .dev_type = HV_TS,
 103           HV_TS_GUID,
 104           .perf_device = false,
 105         },
 106
 107         /* Heartbeat */
 108         { .dev_type = HV_HB,
 109           HV_HEART_BEAT_GUID,
 110           .perf_device = false,
 111         },
 112
 113         /* Shutdown */
 114         { .dev_type = HV_SHUTDOWN,
 115           HV_SHUTDOWN_GUID,
 116           .perf_device = false,
 117         },
 118
 119         /* File copy */
 120         { .dev_type = HV_FCOPY,
 121           HV_FCOPY_GUID,
 122           .perf_device = false,
 123         },
 124
 125         /* Backup */
 126         { .dev_type = HV_BACKUP,
 127           HV_VSS_GUID,
 128           .perf_device = false,
 129         },
 130
 131         /* Dynamic Memory */
 132         { .dev_type = HV_DM,
 133           HV_DM_GUID,
 134           .perf_device = false,
 135         },
 136
 137         /* Unknown GUID */
 138         { .dev_type = HV_UNKNOWN,
 139           .perf_device = false,
 140         },
 141 };
 142
 143 static const struct {
 144         uuid_le guid;
 145 } vmbus_unsupported_devs[] = {
 146         { HV_AVMA1_GUID },
 147         { HV_AVMA2_GUID },
 148         { HV_RDV_GUID   },
 149 };
 150
 151 /*
 152  * The rescinded channel may be blocked waiting for a response from the host;
 153  * take care of that.
 154  */
 155 static void vmbus_rescind_cleanup(struct vmbus_channel *channel)
 156 {
 157         struct vmbus_channel_msginfo *msginfo;
 158         unsigned long flags;
 159
 160
 161         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
 162         channel->rescind = true;
 163         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
 164                                 msglistentry) {
 165
 166                 if (msginfo->waiting_channel == channel) {
 167                         complete(&msginfo->waitevent);
 168                         break;
 169                 }
 170         }
 171         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
 172 }
 173
 174 static bool is_unsupported_vmbus_devs(const uuid_le *guid)
 175 {
 176         int i;
 177
 178         for (i = 0; i < ARRAY_SIZE(vmbus_unsupported_devs); i++)
 179                 if (!uuid_le_cmp(*guid, vmbus_unsupported_devs[i].guid))
 180                         return true;
 181         return false;
 182 }
 183
 184 static u16 hv_get_dev_type(const struct vmbus_channel *channel)
 185 {
 186         const uuid_le *guid = &channel->offermsg.offer.if_type;
 187         u16 i;
 188
 189         if (is_hvsock_channel(channel) || is_unsupported_vmbus_devs(guid))
 190                 return HV_UNKNOWN;
 191
 192         for (i = HV_IDE; i < HV_UNKNOWN; i++) {
 193                 if (!uuid_le_cmp(*guid, vmbus_devs[i].guid))
 194                         return i;
 195         }
 196         pr_info("Unknown GUID: %pUl\n", guid);
 197         return i;
 198 }
 199
 200 /**
 201  * vmbus_prep_negotiate_resp() - Create default response for Hyper-V Negotiate message
 202  * @icmsghdrp: Pointer to msg header structure
 203  * @icmsg_negotiate: Pointer to negotiate message structure
 204  * @buf: Raw buffer channel data
 205  *
 206  * @icmsghdrp is of type &struct icmsg_hdr.
 207  * Set up and fill in default negotiate response message.
 208  *
 209  * The fw_version and fw_vercnt specifies the framework version that
 210  * we can support.
 211  *
 212  * The srv_version and srv_vercnt specifies the service
 213  * versions we can support.
 214  *
 215  * Versions are given in decreasing order.
 216  *
 217  * nego_fw_version and nego_srv_version store the selected protocol versions.
 218  *
 219  * Mainly used by Hyper-V drivers.
 220  */
 221 bool vmbus_prep_negotiate_resp(struct icmsg_hdr *icmsghdrp,
 222                                 u8 *buf, const int *fw_version, int fw_vercnt,
 223                                 const int *srv_version, int srv_vercnt,
 224                                 int *nego_fw_version, int *nego_srv_version)
 225 {
 226         int icframe_major, icframe_minor;
 227         int icmsg_major, icmsg_minor;
 228         int fw_major, fw_minor;
 229         int srv_major, srv_minor;
 230         int i, j;
 231         bool found_match = false;
 232         struct icmsg_negotiate *negop;
 233
 234         icmsghdrp->icmsgsize = 0x10;
 235         negop = (struct icmsg_negotiate *)&buf[
 236                 sizeof(struct vmbuspipe_hdr) +
 237                 sizeof(struct icmsg_hdr)];
 238
 239         icframe_major = negop->icframe_vercnt;
 240         icframe_minor = 0;
 241
 242         icmsg_major = negop->icmsg_vercnt;
 243         icmsg_minor = 0;
 244
 245         /*
 246          * Select the framework version number we will
 247          * support.
 248          */
 249
 250         for (i = 0; i < fw_vercnt; i++) {
 251                 fw_major = (fw_version[i] >> 16);
 252                 fw_minor = (fw_version[i] & 0xFFFF);
 253
 254                 for (j = 0; j < negop->icframe_vercnt; j++) {
 255                         if ((negop->icversion_data[j].major == fw_major) &&
 256                             (negop->icversion_data[j].minor == fw_minor)) {
 257                                 icframe_major = negop->icversion_data[j].major;
 258                                 icframe_minor = negop->icversion_data[j].minor;
 259                                 found_match = true;
 260                                 break;
 261                         }
 262                 }
 263
 264                 if (found_match)
 265                         break;
 266         }
 267
 268         if (!found_match)
 269                 goto fw_error;
 270
 271         found_match = false;
 272
 273         for (i = 0; i < srv_vercnt; i++) {
 274                 srv_major = (srv_version[i] >> 16);
 275                 srv_minor = (srv_version[i] & 0xFFFF);
 276
 277                 for (j = negop->icframe_vercnt;
 278                         (j < negop->icframe_vercnt + negop->icmsg_vercnt);
 279                         j++) {
 280
 281                         if ((negop->icversion_data[j].major == srv_major) &&
 282                                 (negop->icversion_data[j].minor == srv_minor)) {
 283
 284                                 icmsg_major = negop->icversion_data[j].major;
 285                                 icmsg_minor = negop->icversion_data[j].minor;
 286                                 found_match = true;
 287                                 break;
 288                         }
 289                 }
 290
 291                 if (found_match)
 292                         break;
 293         }
 294
 295         /*
 296          * Respond with the framework and service
 297          * version numbers we can support.
 298          */
 299
 300 fw_error:
 301         if (!found_match) {
 302                 negop->icframe_vercnt = 0;
 303                 negop->icmsg_vercnt = 0;
 304         } else {
 305                 negop->icframe_vercnt = 1;
 306                 negop->icmsg_vercnt = 1;
 307         }
 308
 309         if (nego_fw_version)
 310                 *nego_fw_version = (icframe_major << 16) | icframe_minor;
 311
 312         if (nego_srv_version)
 313                 *nego_srv_version = (icmsg_major << 16) | icmsg_minor;
 314
 315         negop->icversion_data[0].major = icframe_major;
 316         negop->icversion_data[0].minor = icframe_minor;
 317         negop->icversion_data[1].major = icmsg_major;
 318         negop->icversion_data[1].minor = icmsg_minor;
 319         return found_match;
 320 }
 321
 322 EXPORT_SYMBOL_GPL(vmbus_prep_negotiate_resp);
 323
 324 /*
 325  * alloc_channel - Allocate and initialize a vmbus channel object
 326  */
 327 static struct vmbus_channel *alloc_channel(void)
 328 {
 329         struct vmbus_channel *channel;
 330
 331         channel = kzalloc(sizeof(*channel), GFP_ATOMIC);
 332         if (!channel)
 333                 return NULL;
 334
 335         spin_lock_init(&channel->lock);
 336         init_completion(&channel->rescind_event);
 337
 338         INIT_LIST_HEAD(&channel->sc_list);
 339         INIT_LIST_HEAD(&channel->percpu_list);
 340
 341         tasklet_init(&channel->callback_event,
 342                      vmbus_on_event, (unsigned long)channel);
 343
 344         return channel;
 345 }
 346
 347 /*
 348  * free_channel - Release the resources used by the vmbus channel object
 349  */
 350 static void free_channel(struct vmbus_channel *channel)
 351 {
 352         tasklet_kill(&channel->callback_event);
 353         vmbus_remove_channel_attr_group(channel);
 354
 355         kobject_put(&channel->kobj);
 356 }
 357
 358 static void percpu_channel_enq(void *arg)
 359 {
 360         struct vmbus_channel *channel = arg;
 361         struct hv_per_cpu_context *hv_cpu
 362                 = this_cpu_ptr(hv_context.cpu_context);
 363
 364         list_add_tail_rcu(&channel->percpu_list, &hv_cpu->chan_list);
 365 }
 366
 367 static void percpu_channel_deq(void *arg)
 368 {
 369         struct vmbus_channel *channel = arg;
 370
 371         list_del_rcu(&channel->percpu_list);
 372 }
 373
 374
 375 static void vmbus_release_relid(u32 relid)
 376 {
 377         struct vmbus_channel_relid_released msg;
 378         int ret;
 379
 380         memset(&msg, 0, sizeof(struct vmbus_channel_relid_released));
 381         msg.child_relid = relid;
 382         msg.header.msgtype = CHANNELMSG_RELID_RELEASED;
 383         ret = vmbus_post_msg(&msg, sizeof(struct vmbus_channel_relid_released),
 384                              true);
 385
 386         trace_vmbus_release_relid(&msg, ret);
 387 }
 388
 389 void hv_process_channel_removal(u32 relid)
 390 {
 391         unsigned long flags;
 392         struct vmbus_channel *primary_channel, *channel;
 393
 394         BUG_ON(!mutex_is_locked(&vmbus_connection.channel_mutex));
 395
 396         /*
 397          * Make sure channel is valid as we may have raced.
 398          */
 399         channel = relid2channel(relid);
 400         if (!channel)
 401                 return;
 402
 403         BUG_ON(!channel->rescind);
 404         if (channel->target_cpu != get_cpu()) {
 405                 put_cpu();
 406                 smp_call_function_single(channel->target_cpu,
 407                                          percpu_channel_deq, channel, true);
 408         } else {
 409                 percpu_channel_deq(channel);
 410                 put_cpu();
 411         }
 412
 413         if (channel->primary_channel == NULL) {
 414                 list_del(&channel->listentry);
 415
 416                 primary_channel = channel;
 417         } else {
 418                 primary_channel = channel->primary_channel;
 419                 spin_lock_irqsave(&primary_channel->lock, flags);
 420                 list_del(&channel->sc_list);
 421                 primary_channel->num_sc--;
 422                 spin_unlock_irqrestore(&primary_channel->lock, flags);
 423         }
 424
 425         /*
 426          * We need to free the bit for init_vp_index() to work in the case
 427          * of sub-channel, when we reload drivers like hv_netvsc.
 428          */
 429         if (channel->affinity_policy == HV_LOCALIZED)
 430                 cpumask_clear_cpu(channel->target_cpu,
 431                                   &primary_channel->alloced_cpus_in_node);
 432
 433         vmbus_release_relid(relid);
 434
 435         free_channel(channel);
 436 }
 437
 438 void vmbus_free_channels(void)
 439 {
 440         struct vmbus_channel *channel, *tmp;
 441
 442         list_for_each_entry_safe(channel, tmp, &vmbus_connection.chn_list,
 443                 listentry) {
 444                 /* hv_process_channel_removal() needs this */
 445                 channel->rescind = true;
 446
 447                 vmbus_device_unregister(channel->device_obj);
 448         }
 449 }
 450
 451 /* Note: the function can run concurrently for primary/sub channels. */
 452 static void vmbus_add_channel_work(struct work_struct *work)
 453 {
 454         struct vmbus_channel *newchannel =
 455                 container_of(work, struct vmbus_channel, add_channel_work);
 456         struct vmbus_channel *primary_channel = newchannel->primary_channel;
 457         unsigned long flags;
 458         u16 dev_type;
 459         int ret;
 460
 461         dev_type = hv_get_dev_type(newchannel);
 462
 463         init_vp_index(newchannel, dev_type);
 464
 465         if (newchannel->target_cpu != get_cpu()) {
 466                 put_cpu();
 467                 smp_call_function_single(newchannel->target_cpu,
 468                                          percpu_channel_enq,
 469                                          newchannel, true);
 470         } else {
 471                 percpu_channel_enq(newchannel);
 472                 put_cpu();
 473         }
 474
 475         /*
 476          * This state is used to indicate a successful open
 477          * so that when we do close the channel normally, we
 478          * can cleanup properly.
 479          */
 480         newchannel->state = CHANNEL_OPEN_STATE;
 481
 482         if (primary_channel != NULL) {
 483                 /* newchannel is a sub-channel. */
 484                 struct hv_device *dev = primary_channel->device_obj;
 485
 486                 if (vmbus_add_channel_kobj(dev, newchannel))
 487                         goto err_deq_chan;
 488
 489                 if (primary_channel->sc_creation_callback != NULL)
 490                         primary_channel->sc_creation_callback(newchannel);
 491
 492                 newchannel->probe_done = true;
 493                 return;
 494         }
 495
 496         /*
 497          * Start the process of binding the primary channel to the driver
 498          */
 499         newchannel->device_obj = vmbus_device_create(
 500                 &newchannel->offermsg.offer.if_type,
 501                 &newchannel->offermsg.offer.if_instance,
 502                 newchannel);
 503         if (!newchannel->device_obj)
 504                 goto err_deq_chan;
 505
 506         newchannel->device_obj->device_id = dev_type;
 507         /*
 508          * Add the new device to the bus. This will kick off device-driver
 509          * binding which eventually invokes the device driver's AddDevice()
 510          * method.
 511          *
 512          * If vmbus_device_register() fails, the 'device_obj' is freed in
 513          * vmbus_device_release() as called by device_unregister() in the
 514          * error path of vmbus_device_register(). In the outside error
 515          * path, there's no need to free it.
 516          */
 517         ret = vmbus_device_register(newchannel->device_obj);
 518
 519         if (ret != 0) {
 520                 pr_err("unable to add child device object (relid %d)\n",
 521                         newchannel->offermsg.child_relid);
 522                 goto err_deq_chan;
 523         }
 524
 525         newchannel->probe_done = true;
 526         return;
 527
 528 err_deq_chan:
 529         mutex_lock(&vmbus_connection.channel_mutex);
 530
 531         /*
 532          * We need to set the flag, otherwise
 533          * vmbus_onoffer_rescind() can be blocked.
 534          */
 535         newchannel->probe_done = true;
 536
 537         if (primary_channel == NULL) {
 538                 list_del(&newchannel->listentry);
 539         } else {
 540                 spin_lock_irqsave(&primary_channel->lock, flags);
 541                 list_del(&newchannel->sc_list);
 542                 spin_unlock_irqrestore(&primary_channel->lock, flags);
 543         }
 544
 545         mutex_unlock(&vmbus_connection.channel_mutex);
 546
 547         if (newchannel->target_cpu != get_cpu()) {
 548                 put_cpu();
 549                 smp_call_function_single(newchannel->target_cpu,
 550                                          percpu_channel_deq,
 551                                          newchannel, true);
 552         } else {
 553                 percpu_channel_deq(newchannel);
 554                 put_cpu();
 555         }
 556
 557         vmbus_release_relid(newchannel->offermsg.child_relid);
 558
 559         free_channel(newchannel);
 560 }
 561
 562 /*
 563  * vmbus_process_offer - Process the offer by creating a channel/device
 564  * associated with this offer
 565  */
 566 static void vmbus_process_offer(struct vmbus_channel *newchannel)
 567 {
 568         struct vmbus_channel *channel;
 569         struct workqueue_struct *wq;
 570         unsigned long flags;
 571         bool fnew = true;
 572
 573         mutex_lock(&vmbus_connection.channel_mutex);
 574
 575         /*
 576          * Now that we have acquired the channel_mutex,
 577          * we can release the potentially racing rescind thread.
 578          */
 579         atomic_dec(&vmbus_connection.offer_in_progress);
 580
 581         list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
 582                 if (!uuid_le_cmp(channel->offermsg.offer.if_type,
 583                                  newchannel->offermsg.offer.if_type) &&
 584                     !uuid_le_cmp(channel->offermsg.offer.if_instance,
 585                                  newchannel->offermsg.offer.if_instance)) {
 586                         fnew = false;
 587                         break;
 588                 }
 589         }
 590
 591         if (fnew)
 592                 list_add_tail(&newchannel->listentry,
 593                               &vmbus_connection.chn_list);
 594         else {
 595                 /*
 596                  * Check to see if this is a valid sub-channel.
 597                  */
 598                 if (newchannel->offermsg.offer.sub_channel_index == 0) {
 599                         mutex_unlock(&vmbus_connection.channel_mutex);
 600                         /*
 601                          * Don't call free_channel(), because newchannel->kobj
 602                          * is not initialized yet.
 603                          */
 604                         kfree(newchannel);
 605                         WARN_ON_ONCE(1);
 606                         return;
 607                 }
 608                 /*
 609                  * Process the sub-channel.
 610                  */
 611                 newchannel->primary_channel = channel;
 612                 spin_lock_irqsave(&channel->lock, flags);
 613                 list_add_tail(&newchannel->sc_list, &channel->sc_list);
 614                 spin_unlock_irqrestore(&channel->lock, flags);
 615         }
 616
 617         mutex_unlock(&vmbus_connection.channel_mutex);
 618
 619         /*
 620          * vmbus_process_offer() mustn't call channel->sc_creation_callback()
 621          * directly for sub-channels, because sc_creation_callback() ->
 622          * vmbus_open() may never get the host's response to the
 623          * OPEN_CHANNEL message (the host may rescind a channel at any time,
 624          * e.g. in the case of hot removing a NIC), and vmbus_onoffer_rescind()
 625          * may not wake up the vmbus_open() as it's blocked due to a non-zero
 626          * vmbus_connection.offer_in_progress, and finally we have a deadlock.
 627          *
 628          * The above is also true for primary channels, if the related device
 629          * drivers use sync probing mode by default.
 630          *
 631          * And, usually the handling of primary channels and sub-channels can
 632          * depend on each other, so we should offload them to different
 633          * workqueues to avoid possible deadlock, e.g. in sync-probing mode,
 634          * NIC1's netvsc_subchan_work() can race with NIC2's netvsc_probe() ->
 635          * rtnl_lock(), and causes deadlock: the former gets the rtnl_lock
 636          * and waits for all the sub-channels to appear, but the latter
 637          * can't get the rtnl_lock and this blocks the handling of
 638          * sub-channels.
 639          */
 640         INIT_WORK(&newchannel->add_channel_work, vmbus_add_channel_work);
 641         wq = fnew ? vmbus_connection.handle_primary_chan_wq :
 642                     vmbus_connection.handle_sub_chan_wq;
 643         queue_work(wq, &newchannel->add_channel_work);
 644 }
 645
 646 /*
 647  * We use this state to statically distribute the channel interrupt load.
 648  */
 649 static int next_numa_node_id;
 650 /*
 651  * init_vp_index() accesses global variables like next_numa_node_id, and
 652  * it can run concurrently for primary channels and sub-channels: see
 653  * vmbus_process_offer(), so we need the lock to protect the global
 654  * variables.
 655  */
 656 static DEFINE_SPINLOCK(bind_channel_to_cpu_lock);
 657
 658 /*
 659  * Starting with Win8, we can statically distribute the incoming
 660  * channel interrupt load by binding a channel to VCPU.
 661  * We distribute the interrupt loads to one or more NUMA nodes based on
 662  * the channel's affinity_policy.
 663  *
 664  * For pre-win8 hosts or non-performance critical channels we assign the
 665  * first CPU in the first NUMA node.
 666  */
 667 static void init_vp_index(struct vmbus_channel *channel, u16 dev_type)
 668 {
 669         u32 cur_cpu;
 670         bool perf_chn = vmbus_devs[dev_type].perf_device;
 671         struct vmbus_channel *primary = channel->primary_channel;
 672         int next_node;
 673         cpumask_var_t available_mask;
 674         struct cpumask *alloced_mask;
 675
 676         if ((vmbus_proto_version == VERSION_WS2008) ||
 677             (vmbus_proto_version == VERSION_WIN7) || (!perf_chn) ||
 678             !alloc_cpumask_var(&available_mask, GFP_KERNEL)) {
 679                 /*
 680                  * Prior to win8, all channel interrupts are
 681                  * delivered on cpu 0.
 682                  * Also if the channel is not a performance critical
 683                  * channel, bind it to cpu 0.
 684                  * In case alloc_cpumask_var() fails, bind it to cpu 0.
 685                  */
 686                 channel->numa_node = 0;
 687                 channel->target_cpu = 0;
 688                 channel->target_vp = hv_cpu_number_to_vp_number(0);
 689                 return;
 690         }
 691
 692         spin_lock(&bind_channel_to_cpu_lock);
 693
 694         /*
 695          * Based on the channel affinity policy, we will assign the NUMA
 696          * nodes.
 697          */
 698
 699         if ((channel->affinity_policy == HV_BALANCED) || (!primary)) {
 700                 while (true) {
 701                         next_node = next_numa_node_id++;
 702                         if (next_node == nr_node_ids) {
 703                                 next_node = next_numa_node_id = 0;
 704                                 continue;
 705                         }
 706                         if (cpumask_empty(cpumask_of_node(next_node)))
 707                                 continue;
 708                         break;
 709                 }
 710                 channel->numa_node = next_node;
 711                 primary = channel;
 712         }
 713         alloced_mask = &hv_context.hv_numa_map[primary->numa_node];
 714
 715         if (cpumask_weight(alloced_mask) ==
 716             cpumask_weight(cpumask_of_node(primary->numa_node))) {
 717                 /*
 718                  * We have cycled through all the CPUs in the node;
 719                  * reset the alloced map.
 720                  */
 721                 cpumask_clear(alloced_mask);
 722         }
 723
 724         cpumask_xor(available_mask, alloced_mask,
 725                     cpumask_of_node(primary->numa_node));
 726
 727         cur_cpu = -1;
 728
 729         if (primary->affinity_policy == HV_LOCALIZED) {
 730                 /*
 731                  * Normally Hyper-V host doesn't create more subchannels
 732                  * than there are VCPUs on the node but it is possible when not
 733                  * all present VCPUs on the node are initialized by guest.
 734                  * Clear the alloced_cpus_in_node to start over.
 735                  */
 736                 if (cpumask_equal(&primary->alloced_cpus_in_node,
 737                                   cpumask_of_node(primary->numa_node)))
 738                         cpumask_clear(&primary->alloced_cpus_in_node);
 739         }
 740
 741         while (true) {
 742                 cur_cpu = cpumask_next(cur_cpu, available_mask);
 743                 if (cur_cpu >= nr_cpu_ids) {
 744                         cur_cpu = -1;
 745                         cpumask_copy(available_mask,
 746                                      cpumask_of_node(primary->numa_node));
 747                         continue;
 748                 }
 749
 750                 if (primary->affinity_policy == HV_LOCALIZED) {
 751                         /*
 752                          * NOTE: in the case of sub-channel, we clear the
 753                          * sub-channel related bit(s) in
 754                          * primary->alloced_cpus_in_node in
 755                          * hv_process_channel_removal(), so when we
 756                          * reload drivers like hv_netvsc in SMP guest, here
 757                          * we're able to re-allocate
 758                          * bit from primary->alloced_cpus_in_node.
 759                          */
 760                         if (!cpumask_test_cpu(cur_cpu,
 761                                               &primary->alloced_cpus_in_node)) {
 762                                 cpumask_set_cpu(cur_cpu,
 763                                                 &primary->alloced_cpus_in_node);
 764                                 cpumask_set_cpu(cur_cpu, alloced_mask);
 765                                 break;
 766                         }
 767                 } else {
 768                         cpumask_set_cpu(cur_cpu, alloced_mask);
 769                         break;
 770                 }
 771         }
 772
 773         channel->target_cpu = cur_cpu;
 774         channel->target_vp = hv_cpu_number_to_vp_number(cur_cpu);
 775
 776         spin_unlock(&bind_channel_to_cpu_lock);
 777
 778         free_cpumask_var(available_mask);
 779 }
 780
 781 #define UNLOAD_DELAY_UNIT_MS    10              /* 10 milliseconds */
 782 #define UNLOAD_WAIT_MS          (100*1000)      /* 100 seconds */
 783 #define UNLOAD_WAIT_LOOPS       (UNLOAD_WAIT_MS/UNLOAD_DELAY_UNIT_MS)
 784 #define UNLOAD_MSG_MS           (5*1000)        /* Every 5 seconds */
 785 #define UNLOAD_MSG_LOOPS        (UNLOAD_MSG_MS/UNLOAD_DELAY_UNIT_MS)
 786
 787 static void vmbus_wait_for_unload(void)
 788 {
 789         int cpu;
 790         void *page_addr;
 791         struct hv_message *msg;
 792         struct vmbus_channel_message_header *hdr;
 793         u32 message_type, i;
 794
 795         /*
 796          * CHANNELMSG_UNLOAD_RESPONSE is always delivered to the CPU which was
 797          * used for initial contact or to CPU0 depending on host version. When
 798          * we're crashing on a different CPU let's hope that IRQ handler on
 799          * the cpu which receives CHANNELMSG_UNLOAD_RESPONSE is still
 800          * functional and vmbus_unload_response() will complete
 801          * vmbus_connection.unload_event. If not, the last thing we can do is
 802          * read message pages for all CPUs directly.
 803          *
 804          * Wait up to 100 seconds since an Azure host must writeback any dirty
 805          * data in its disk cache before the VMbus UNLOAD request will
 806          * complete. This flushing has been empirically observed to take up
 807          * to 50 seconds in cases with a lot of dirty data, so allow additional
 808          * leeway and for inaccuracies in mdelay(). But eventually time out so
 809          * that the panic path can't get hung forever in case the response
 810          * message isn't seen.
 811          */
 812         for (i = 1; i <= UNLOAD_WAIT_LOOPS; i++) {
 813                 if (completion_done(&vmbus_connection.unload_event))
 814                         goto completed;
 815
 816                 for_each_online_cpu(cpu) {
 817                         struct hv_per_cpu_context *hv_cpu
 818                                 = per_cpu_ptr(hv_context.cpu_context, cpu);
 819
 820                         page_addr = hv_cpu->synic_message_page;
 821                         msg = (struct hv_message *)page_addr
 822                                 + VMBUS_MESSAGE_SINT;
 823
 824                         message_type = READ_ONCE(msg->header.message_type);
 825                         if (message_type == HVMSG_NONE)
 826                                 continue;
 827
 828                         hdr = (struct vmbus_channel_message_header *)
 829                                 msg->u.payload;
 830
 831                         if (hdr->msgtype == CHANNELMSG_UNLOAD_RESPONSE)
 832                                 complete(&vmbus_connection.unload_event);
 833
 834                         vmbus_signal_eom(msg, message_type);
 835                 }
 836
 837                 /*
 838                  * Give a notice periodically so someone watching the
 839                  * serial output won't think it is completely hung.
 840                  */
 841                 if (!(i % UNLOAD_MSG_LOOPS))
 842                         pr_notice("Waiting for VMBus UNLOAD to complete\n");
 843
 844                 mdelay(UNLOAD_DELAY_UNIT_MS);
 845         }
 846         pr_err("Continuing even though VMBus UNLOAD did not complete\n");
 847
 848 completed:
 849         /*
 850          * We're crashing and already got the UNLOAD_RESPONSE, cleanup all
 851          * maybe-pending messages on all CPUs to be able to receive new
 852          * messages after we reconnect.
 853          */
 854         for_each_online_cpu(cpu) {
 855                 struct hv_per_cpu_context *hv_cpu
 856                         = per_cpu_ptr(hv_context.cpu_context, cpu);
 857
 858                 page_addr = hv_cpu->synic_message_page;
 859                 msg = (struct hv_message *)page_addr + VMBUS_MESSAGE_SINT;
 860                 msg->header.message_type = HVMSG_NONE;
 861         }
 862 }
 863
 864 /*
 865  * vmbus_unload_response - Handler for the unload response.
 866  */
 867 static void vmbus_unload_response(struct vmbus_channel_message_header *hdr)
 868 {
 869         /*
 870          * This is a global event; just wakeup the waiting thread.
 871          * Once we successfully unload, we can cleanup the monitor state.
 872          */
 873         complete(&vmbus_connection.unload_event);
 874 }
 875
 876 void vmbus_initiate_unload(bool crash)
 877 {
 878         struct vmbus_channel_message_header hdr;
 879
 880         if (xchg(&vmbus_connection.conn_state, DISCONNECTED) == DISCONNECTED)
 881                 return;
 882
 883         /* Pre-Win2012R2 hosts don't support reconnect */
 884         if (vmbus_proto_version < VERSION_WIN8_1)
 885                 return;
 886
 887         init_completion(&vmbus_connection.unload_event);
 888         memset(&hdr, 0, sizeof(struct vmbus_channel_message_header));
 889         hdr.msgtype = CHANNELMSG_UNLOAD;
 890         vmbus_post_msg(&hdr, sizeof(struct vmbus_channel_message_header),
 891                        !crash);
 892
 893         /*
 894          * vmbus_initiate_unload() is also called on crash and the crash can be
 895          * happening in an interrupt context, where scheduling is impossible.
 896          */
 897         if (!crash)
 898                 wait_for_completion(&vmbus_connection.unload_event);
 899         else
 900                 vmbus_wait_for_unload();
 901 }
 902
 903 /*
 904  * vmbus_onoffer - Handler for channel offers from vmbus in parent partition.
 905  *
 906  */
 907 static void vmbus_onoffer(struct vmbus_channel_message_header *hdr)
 908 {
 909         struct vmbus_channel_offer_channel *offer;
 910         struct vmbus_channel *newchannel;
 911
 912         offer = (struct vmbus_channel_offer_channel *)hdr;
 913
 914         trace_vmbus_onoffer(offer);
 915
 916         /* Allocate the channel object and save this offer. */
 917         newchannel = alloc_channel();
 918         if (!newchannel) {
 919                 vmbus_release_relid(offer->child_relid);
 920                 atomic_dec(&vmbus_connection.offer_in_progress);
 921                 pr_err("Unable to allocate channel object\n");
 922                 return;
 923         }
 924
 925         /*
 926          * Setup state for signalling the host.
 927          */
 928         newchannel->sig_event = VMBUS_EVENT_CONNECTION_ID;
 929
 930         if (vmbus_proto_version != VERSION_WS2008) {
 931                 newchannel->is_dedicated_interrupt =
 932                                 (offer->is_dedicated_interrupt != 0);
 933                 newchannel->sig_event = offer->connection_id;
 934         }
 935
 936         memcpy(&newchannel->offermsg, offer,
 937                sizeof(struct vmbus_channel_offer_channel));
 938         newchannel->monitor_grp = (u8)offer->monitorid / 32;
 939         newchannel->monitor_bit = (u8)offer->monitorid % 32;
 940
 941         vmbus_process_offer(newchannel);
 942 }
 943
 944 /*
 945  * vmbus_onoffer_rescind - Rescind offer handler.
 946  *
 947  * We queue a work item to process this offer synchronously
 948  */
 949 static void vmbus_onoffer_rescind(struct vmbus_channel_message_header *hdr)
 950 {
 951         struct vmbus_channel_rescind_offer *rescind;
 952         struct vmbus_channel *channel;
 953         struct device *dev;
 954
 955         rescind = (struct vmbus_channel_rescind_offer *)hdr;
 956
 957         trace_vmbus_onoffer_rescind(rescind);
 958
 959         /*
 960          * The offer msg and the corresponding rescind msg
 961          * from the host are guranteed to be ordered -
 962          * offer comes in first and then the rescind.
 963          * Since we process these events in work elements,
 964          * and with preemption, we may end up processing
 965          * the events out of order. Given that we handle these
 966          * work elements on the same CPU, this is possible only
 967          * in the case of preemption. In any case wait here
 968          * until the offer processing has moved beyond the
 969          * point where the channel is discoverable.
 970          */
 971
 972         while (atomic_read(&vmbus_connection.offer_in_progress) != 0) {
 973                 /*
 974                  * We wait here until any channel offer is currently
 975                  * being processed.
 976                  */
 977                 msleep(1);
 978         }
 979
 980         mutex_lock(&vmbus_connection.channel_mutex);
 981         channel = relid2channel(rescind->child_relid);
 982         mutex_unlock(&vmbus_connection.channel_mutex);
 983
 984         if (channel == NULL) {
 985                 /*
 986                  * We failed in processing the offer message;
 987                  * we would have cleaned up the relid in that
 988                  * failure path.
 989                  */
 990                 return;
 991         }
 992
 993         /*
 994          * Before setting channel->rescind in vmbus_rescind_cleanup(), we
 995          * should make sure the channel callback is not running any more.
 996          */
 997         vmbus_reset_channel_cb(channel);
 998
 999         /*
1000          * Now wait for offer handling to complete.
1001          */
1002         vmbus_rescind_cleanup(channel);
1003         while (READ_ONCE(channel->probe_done) == false) {
1004                 /*
1005                  * We wait here until any channel offer is currently
1006                  * being processed.
1007                  */
1008                 msleep(1);
1009         }
1010
1011         /*
1012          * At this point, the rescind handling can proceed safely.
1013          */
1014
1015         if (channel->device_obj) {
1016                 if (channel->chn_rescind_callback) {
1017                         channel->chn_rescind_callback(channel);
1018                         return;
1019                 }
1020                 /*
1021                  * We will have to unregister this device from the
1022                  * driver core.
1023                  */
1024                 dev = get_device(&channel->device_obj->device);
1025                 if (dev) {
1026                         vmbus_device_unregister(channel->device_obj);
1027                         put_device(dev);
1028                 }
1029         } else if (channel->primary_channel != NULL) {
1030                 /*
1031                  * Sub-channel is being rescinded. Following is the channel
1032                  * close sequence when initiated from the driveri (refer to
1033                  * vmbus_close() for details):
1034                  * 1. Close all sub-channels first
1035                  * 2. Then close the primary channel.
1036                  */
1037                 mutex_lock(&vmbus_connection.channel_mutex);
1038                 if (channel->state == CHANNEL_OPEN_STATE) {
1039                         /*
1040                          * The channel is currently not open;
1041                          * it is safe for us to cleanup the channel.
1042                          */
1043                         hv_process_channel_removal(rescind->child_relid);
1044                 } else {
1045                         complete(&channel->rescind_event);
1046                 }
1047                 mutex_unlock(&vmbus_connection.channel_mutex);
1048         }
1049 }
1050
1051 void vmbus_hvsock_device_unregister(struct vmbus_channel *channel)
1052 {
1053         BUG_ON(!is_hvsock_channel(channel));
1054
1055         /* We always get a rescind msg when a connection is closed. */
1056         while (!READ_ONCE(channel->probe_done) || !READ_ONCE(channel->rescind))
1057                 msleep(1);
1058
1059         vmbus_device_unregister(channel->device_obj);
1060 }
1061 EXPORT_SYMBOL_GPL(vmbus_hvsock_device_unregister);
1062
1063
1064 /*
1065  * vmbus_onoffers_delivered -
1066  * This is invoked when all offers have been delivered.
1067  *
1068  * Nothing to do here.
1069  */
1070 static void vmbus_onoffers_delivered(
1071                         struct vmbus_channel_message_header *hdr)
1072 {
1073 }
1074
1075 /*
1076  * vmbus_onopen_result - Open result handler.
1077  *
1078  * This is invoked when we received a response to our channel open request.
1079  * Find the matching request, copy the response and signal the requesting
1080  * thread.
1081  */
1082 static void vmbus_onopen_result(struct vmbus_channel_message_header *hdr)
1083 {
1084         struct vmbus_channel_open_result *result;
1085         struct vmbus_channel_msginfo *msginfo;
1086         struct vmbus_channel_message_header *requestheader;
1087         struct vmbus_channel_open_channel *openmsg;
1088         unsigned long flags;
1089
1090         result = (struct vmbus_channel_open_result *)hdr;
1091
1092         trace_vmbus_onopen_result(result);
1093
1094         /*
1095          * Find the open msg, copy the result and signal/unblock the wait event
1096          */
1097         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1098
1099         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1100                                 msglistentry) {
1101                 requestheader =
1102                         (struct vmbus_channel_message_header *)msginfo->msg;
1103
1104                 if (requestheader->msgtype == CHANNELMSG_OPENCHANNEL) {
1105                         openmsg =
1106                         (struct vmbus_channel_open_channel *)msginfo->msg;
1107                         if (openmsg->child_relid == result->child_relid &&
1108                             openmsg->openid == result->openid) {
1109                                 memcpy(&msginfo->response.open_result,
1110                                        result,
1111                                        sizeof(
1112                                         struct vmbus_channel_open_result));
1113                                 complete(&msginfo->waitevent);
1114                                 break;
1115                         }
1116                 }
1117         }
1118         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1119 }
1120
1121 /*
1122  * vmbus_ongpadl_created - GPADL created handler.
1123  *
1124  * This is invoked when we received a response to our gpadl create request.
1125  * Find the matching request, copy the response and signal the requesting
1126  * thread.
1127  */
1128 static void vmbus_ongpadl_created(struct vmbus_channel_message_header *hdr)
1129 {
1130         struct vmbus_channel_gpadl_created *gpadlcreated;
1131         struct vmbus_channel_msginfo *msginfo;
1132         struct vmbus_channel_message_header *requestheader;
1133         struct vmbus_channel_gpadl_header *gpadlheader;
1134         unsigned long flags;
1135
1136         gpadlcreated = (struct vmbus_channel_gpadl_created *)hdr;
1137
1138         trace_vmbus_ongpadl_created(gpadlcreated);
1139
1140         /*
1141          * Find the establish msg, copy the result and signal/unblock the wait
1142          * event
1143          */
1144         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1145
1146         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1147                                 msglistentry) {
1148                 requestheader =
1149                         (struct vmbus_channel_message_header *)msginfo->msg;
1150
1151                 if (requestheader->msgtype == CHANNELMSG_GPADL_HEADER) {
1152                         gpadlheader =
1153                         (struct vmbus_channel_gpadl_header *)requestheader;
1154
1155                         if ((gpadlcreated->child_relid ==
1156                              gpadlheader->child_relid) &&
1157                             (gpadlcreated->gpadl == gpadlheader->gpadl)) {
1158                                 memcpy(&msginfo->response.gpadl_created,
1159                                        gpadlcreated,
1160                                        sizeof(
1161                                         struct vmbus_channel_gpadl_created));
1162                                 complete(&msginfo->waitevent);
1163                                 break;
1164                         }
1165                 }
1166         }
1167         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1168 }
1169
1170 /*
1171  * vmbus_ongpadl_torndown - GPADL torndown handler.
1172  *
1173  * This is invoked when we received a response to our gpadl teardown request.
1174  * Find the matching request, copy the response and signal the requesting
1175  * thread.
1176  */
1177 static void vmbus_ongpadl_torndown(
1178                         struct vmbus_channel_message_header *hdr)
1179 {
1180         struct vmbus_channel_gpadl_torndown *gpadl_torndown;
1181         struct vmbus_channel_msginfo *msginfo;
1182         struct vmbus_channel_message_header *requestheader;
1183         struct vmbus_channel_gpadl_teardown *gpadl_teardown;
1184         unsigned long flags;
1185
1186         gpadl_torndown = (struct vmbus_channel_gpadl_torndown *)hdr;
1187
1188         trace_vmbus_ongpadl_torndown(gpadl_torndown);
1189
1190         /*
1191          * Find the open msg, copy the result and signal/unblock the wait event
1192          */
1193         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1194
1195         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1196                                 msglistentry) {
1197                 requestheader =
1198                         (struct vmbus_channel_message_header *)msginfo->msg;
1199
1200                 if (requestheader->msgtype == CHANNELMSG_GPADL_TEARDOWN) {
1201                         gpadl_teardown =
1202                         (struct vmbus_channel_gpadl_teardown *)requestheader;
1203
1204                         if (gpadl_torndown->gpadl == gpadl_teardown->gpadl) {
1205                                 memcpy(&msginfo->response.gpadl_torndown,
1206                                        gpadl_torndown,
1207                                        sizeof(
1208                                         struct vmbus_channel_gpadl_torndown));
1209                                 complete(&msginfo->waitevent);
1210                                 break;
1211                         }
1212                 }
1213         }
1214         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1215 }
1216
1217 /*
1218  * vmbus_onversion_response - Version response handler
1219  *
1220  * This is invoked when we received a response to our initiate contact request.
1221  * Find the matching request, copy the response and signal the requesting
1222  * thread.
1223  */
1224 static void vmbus_onversion_response(
1225                 struct vmbus_channel_message_header *hdr)
1226 {
1227         struct vmbus_channel_msginfo *msginfo;
1228         struct vmbus_channel_message_header *requestheader;
1229         struct vmbus_channel_version_response *version_response;
1230         unsigned long flags;
1231
1232         version_response = (struct vmbus_channel_version_response *)hdr;
1233
1234         trace_vmbus_onversion_response(version_response);
1235
1236         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1237
1238         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1239                                 msglistentry) {
1240                 requestheader =
1241                         (struct vmbus_channel_message_header *)msginfo->msg;
1242
1243                 if (requestheader->msgtype ==
1244                     CHANNELMSG_INITIATE_CONTACT) {
1245                         memcpy(&msginfo->response.version_response,
1246                               version_response,
1247                               sizeof(struct vmbus_channel_version_response));
1248                         complete(&msginfo->waitevent);
1249                 }
1250         }
1251         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1252 }
1253
1254 /* Channel message dispatch table */
1255 const struct vmbus_channel_message_table_entry
1256 channel_message_table[CHANNELMSG_COUNT] = {
1257         { CHANNELMSG_INVALID,                   0, NULL },
1258         { CHANNELMSG_OFFERCHANNEL,              0, vmbus_onoffer },
1259         { CHANNELMSG_RESCIND_CHANNELOFFER,      0, vmbus_onoffer_rescind },
1260         { CHANNELMSG_REQUESTOFFERS,             0, NULL },
1261         { CHANNELMSG_ALLOFFERS_DELIVERED,       1, vmbus_onoffers_delivered },
1262         { CHANNELMSG_OPENCHANNEL,               0, NULL },
1263         { CHANNELMSG_OPENCHANNEL_RESULT,        1, vmbus_onopen_result },
1264         { CHANNELMSG_CLOSECHANNEL,              0, NULL },
1265         { CHANNELMSG_GPADL_HEADER,              0, NULL },
1266         { CHANNELMSG_GPADL_BODY,                0, NULL },
1267         { CHANNELMSG_GPADL_CREATED,             1, vmbus_ongpadl_created },
1268         { CHANNELMSG_GPADL_TEARDOWN,            0, NULL },
1269         { CHANNELMSG_GPADL_TORNDOWN,            1, vmbus_ongpadl_torndown },
1270         { CHANNELMSG_RELID_RELEASED,            0, NULL },
1271         { CHANNELMSG_INITIATE_CONTACT,          0, NULL },
1272         { CHANNELMSG_VERSION_RESPONSE,          1, vmbus_onversion_response },
1273         { CHANNELMSG_UNLOAD,                    0, NULL },
1274         { CHANNELMSG_UNLOAD_RESPONSE,           1, vmbus_unload_response },
1275         { CHANNELMSG_18,                        0, NULL },
1276         { CHANNELMSG_19,                        0, NULL },
1277         { CHANNELMSG_20,                        0, NULL },
1278         { CHANNELMSG_TL_CONNECT_REQUEST,        0, NULL },
1279         { CHANNELMSG_22,                        0, NULL },
1280         { CHANNELMSG_TL_CONNECT_RESULT,         0, NULL },
1281 };
1282
1283 /*
1284  * vmbus_onmessage - Handler for channel protocol messages.
1285  *
1286  * This is invoked in the vmbus worker thread context.
1287  */
1288 void vmbus_onmessage(void *context)
1289 {
1290         struct hv_message *msg = context;
1291         struct vmbus_channel_message_header *hdr;
1292
1293         hdr = (struct vmbus_channel_message_header *)msg->u.payload;
1294
1295         trace_vmbus_on_message(hdr);
1296
1297         /*
1298          * vmbus_on_msg_dpc() makes sure the hdr->msgtype here can not go
1299          * out of bound and the message_handler pointer can not be NULL.
1300          */
1301         channel_message_table[hdr->msgtype].message_handler(hdr);
1302 }
1303
1304 /*
1305  * vmbus_request_offers - Send a request to get all our pending offers.
1306  */
1307 int vmbus_request_offers(void)
1308 {
1309         struct vmbus_channel_message_header *msg;
1310         struct vmbus_channel_msginfo *msginfo;
1311         int ret;
1312
1313         msginfo = kmalloc(sizeof(*msginfo) +
1314                           sizeof(struct vmbus_channel_message_header),
1315                           GFP_KERNEL);
1316         if (!msginfo)
1317                 return -ENOMEM;
1318
1319         msg = (struct vmbus_channel_message_header *)msginfo->msg;
1320
1321         msg->msgtype = CHANNELMSG_REQUESTOFFERS;
1322
1323         ret = vmbus_post_msg(msg, sizeof(struct vmbus_channel_message_header),
1324                              true);
1325
1326         trace_vmbus_request_offers(ret);
1327
1328         if (ret != 0) {
1329                 pr_err("Unable to request offers - %d\n", ret);
1330
1331                 goto cleanup;
1332         }
1333
1334 cleanup:
1335         kfree(msginfo);
1336
1337         return ret;
1338 }
1339
1340 /*
1341  * Retrieve the (sub) channel on which to send an outgoing request.
1342  * When a primary channel has multiple sub-channels, we try to
1343  * distribute the load equally amongst all available channels.
1344  */
1345 struct vmbus_channel *vmbus_get_outgoing_channel(struct vmbus_channel *primary)
1346 {
1347         struct list_head *cur, *tmp;
1348         int cur_cpu;
1349         struct vmbus_channel *cur_channel;
1350         struct vmbus_channel *outgoing_channel = primary;
1351         int next_channel;
1352         int i = 1;
1353
1354         if (list_empty(&primary->sc_list))
1355                 return outgoing_channel;
1356
1357         next_channel = primary->next_oc++;
1358
1359         if (next_channel > (primary->num_sc)) {
1360                 primary->next_oc = 0;
1361                 return outgoing_channel;
1362         }
1363
1364         cur_cpu = hv_cpu_number_to_vp_number(smp_processor_id());
1365         list_for_each_safe(cur, tmp, &primary->sc_list) {
1366                 cur_channel = list_entry(cur, struct vmbus_channel, sc_list);
1367                 if (cur_channel->state != CHANNEL_OPENED_STATE)
1368                         continue;
1369
1370                 if (cur_channel->target_vp == cur_cpu)
1371                         return cur_channel;
1372
1373                 if (i == next_channel)
1374                         return cur_channel;
1375
1376                 i++;
1377         }
1378
1379         return outgoing_channel;
1380 }
1381 EXPORT_SYMBOL_GPL(vmbus_get_outgoing_channel);
1382
1383 static void invoke_sc_cb(struct vmbus_channel *primary_channel)
1384 {
1385         struct list_head *cur, *tmp;
1386         struct vmbus_channel *cur_channel;
1387
1388         if (primary_channel->sc_creation_callback == NULL)
1389                 return;
1390
1391         list_for_each_safe(cur, tmp, &primary_channel->sc_list) {
1392                 cur_channel = list_entry(cur, struct vmbus_channel, sc_list);
1393
1394                 primary_channel->sc_creation_callback(cur_channel);
1395         }
1396 }
1397
1398 void vmbus_set_sc_create_callback(struct vmbus_channel *primary_channel,
1399                                 void (*sc_cr_cb)(struct vmbus_channel *new_sc))
1400 {
1401         primary_channel->sc_creation_callback = sc_cr_cb;
1402 }
1403 EXPORT_SYMBOL_GPL(vmbus_set_sc_create_callback);
1404
1405 bool vmbus_are_subchannels_present(struct vmbus_channel *primary)
1406 {
1407         bool ret;
1408
1409         ret = !list_empty(&primary->sc_list);
1410
1411         if (ret) {
1412                 /*
1413                  * Invoke the callback on sub-channel creation.
1414                  * This will present a uniform interface to the
1415                  * clients.
1416                  */
1417                 invoke_sc_cb(primary);
1418         }
1419
1420         return ret;
1421 }
1422 EXPORT_SYMBOL_GPL(vmbus_are_subchannels_present);
1423
1424 void vmbus_set_chn_rescind_callback(struct vmbus_channel *channel,
1425                 void (*chn_rescind_cb)(struct vmbus_channel *))
1426 {
1427         channel->chn_rescind_callback = chn_rescind_cb;
1428 }
1429 EXPORT_SYMBOL_GPL(vmbus_set_chn_rescind_callback);