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
 354         kobject_put(&channel->kobj);
 355 }
 356
 357 static void percpu_channel_enq(void *arg)
 358 {
 359         struct vmbus_channel *channel = arg;
 360         struct hv_per_cpu_context *hv_cpu
 361                 = this_cpu_ptr(hv_context.cpu_context);
 362
 363         list_add_tail_rcu(&channel->percpu_list, &hv_cpu->chan_list);
 364 }
 365
 366 static void percpu_channel_deq(void *arg)
 367 {
 368         struct vmbus_channel *channel = arg;
 369
 370         list_del_rcu(&channel->percpu_list);
 371 }
 372
 373
 374 static void vmbus_release_relid(u32 relid)
 375 {
 376         struct vmbus_channel_relid_released msg;
 377         int ret;
 378
 379         memset(&msg, 0, sizeof(struct vmbus_channel_relid_released));
 380         msg.child_relid = relid;
 381         msg.header.msgtype = CHANNELMSG_RELID_RELEASED;
 382         ret = vmbus_post_msg(&msg, sizeof(struct vmbus_channel_relid_released),
 383                              true);
 384
 385         trace_vmbus_release_relid(&msg, ret);
 386 }
 387
 388 void hv_process_channel_removal(u32 relid)
 389 {
 390         unsigned long flags;
 391         struct vmbus_channel *primary_channel, *channel;
 392
 393         BUG_ON(!mutex_is_locked(&vmbus_connection.channel_mutex));
 394
 395         /*
 396          * Make sure channel is valid as we may have raced.
 397          */
 398         channel = relid2channel(relid);
 399         if (!channel)
 400                 return;
 401
 402         BUG_ON(!channel->rescind);
 403         if (channel->target_cpu != get_cpu()) {
 404                 put_cpu();
 405                 smp_call_function_single(channel->target_cpu,
 406                                          percpu_channel_deq, channel, true);
 407         } else {
 408                 percpu_channel_deq(channel);
 409                 put_cpu();
 410         }
 411
 412         if (channel->primary_channel == NULL) {
 413                 list_del(&channel->listentry);
 414
 415                 primary_channel = channel;
 416         } else {
 417                 primary_channel = channel->primary_channel;
 418                 spin_lock_irqsave(&primary_channel->lock, flags);
 419                 list_del(&channel->sc_list);
 420                 primary_channel->num_sc--;
 421                 spin_unlock_irqrestore(&primary_channel->lock, flags);
 422         }
 423
 424         /*
 425          * We need to free the bit for init_vp_index() to work in the case
 426          * of sub-channel, when we reload drivers like hv_netvsc.
 427          */
 428         if (channel->affinity_policy == HV_LOCALIZED)
 429                 cpumask_clear_cpu(channel->target_cpu,
 430                                   &primary_channel->alloced_cpus_in_node);
 431
 432         vmbus_release_relid(relid);
 433
 434         free_channel(channel);
 435 }
 436
 437 void vmbus_free_channels(void)
 438 {
 439         struct vmbus_channel *channel, *tmp;
 440
 441         list_for_each_entry_safe(channel, tmp, &vmbus_connection.chn_list,
 442                 listentry) {
 443                 /* hv_process_channel_removal() needs this */
 444                 channel->rescind = true;
 445
 446                 vmbus_device_unregister(channel->device_obj);
 447         }
 448 }
 449
 450 /* Note: the function can run concurrently for primary/sub channels. */
 451 static void vmbus_add_channel_work(struct work_struct *work)
 452 {
 453         struct vmbus_channel *newchannel =
 454                 container_of(work, struct vmbus_channel, add_channel_work);
 455         struct vmbus_channel *primary_channel = newchannel->primary_channel;
 456         unsigned long flags;
 457         u16 dev_type;
 458         int ret;
 459
 460         dev_type = hv_get_dev_type(newchannel);
 461
 462         init_vp_index(newchannel, dev_type);
 463
 464         if (newchannel->target_cpu != get_cpu()) {
 465                 put_cpu();
 466                 smp_call_function_single(newchannel->target_cpu,
 467                                          percpu_channel_enq,
 468                                          newchannel, true);
 469         } else {
 470                 percpu_channel_enq(newchannel);
 471                 put_cpu();
 472         }
 473
 474         /*
 475          * This state is used to indicate a successful open
 476          * so that when we do close the channel normally, we
 477          * can cleanup properly.
 478          */
 479         newchannel->state = CHANNEL_OPEN_STATE;
 480
 481         if (primary_channel != NULL) {
 482                 /* newchannel is a sub-channel. */
 483                 struct hv_device *dev = primary_channel->device_obj;
 484
 485                 if (vmbus_add_channel_kobj(dev, newchannel))
 486                         goto err_deq_chan;
 487
 488                 if (primary_channel->sc_creation_callback != NULL)
 489                         primary_channel->sc_creation_callback(newchannel);
 490
 491                 newchannel->probe_done = true;
 492                 return;
 493         }
 494
 495         /*
 496          * Start the process of binding the primary channel to the driver
 497          */
 498         newchannel->device_obj = vmbus_device_create(
 499                 &newchannel->offermsg.offer.if_type,
 500                 &newchannel->offermsg.offer.if_instance,
 501                 newchannel);
 502         if (!newchannel->device_obj)
 503                 goto err_deq_chan;
 504
 505         newchannel->device_obj->device_id = dev_type;
 506         /*
 507          * Add the new device to the bus. This will kick off device-driver
 508          * binding which eventually invokes the device driver's AddDevice()
 509          * method.
 510          */
 511         ret = vmbus_device_register(newchannel->device_obj);
 512
 513         if (ret != 0) {
 514                 pr_err("unable to add child device object (relid %d)\n",
 515                         newchannel->offermsg.child_relid);
 516                 kfree(newchannel->device_obj);
 517                 goto err_deq_chan;
 518         }
 519
 520         newchannel->probe_done = true;
 521         return;
 522
 523 err_deq_chan:
 524         mutex_lock(&vmbus_connection.channel_mutex);
 525
 526         /*
 527          * We need to set the flag, otherwise
 528          * vmbus_onoffer_rescind() can be blocked.
 529          */
 530         newchannel->probe_done = true;
 531
 532         if (primary_channel == NULL) {
 533                 list_del(&newchannel->listentry);
 534         } else {
 535                 spin_lock_irqsave(&primary_channel->lock, flags);
 536                 list_del(&newchannel->sc_list);
 537                 spin_unlock_irqrestore(&primary_channel->lock, flags);
 538         }
 539
 540         mutex_unlock(&vmbus_connection.channel_mutex);
 541
 542         if (newchannel->target_cpu != get_cpu()) {
 543                 put_cpu();
 544                 smp_call_function_single(newchannel->target_cpu,
 545                                          percpu_channel_deq,
 546                                          newchannel, true);
 547         } else {
 548                 percpu_channel_deq(newchannel);
 549                 put_cpu();
 550         }
 551
 552         vmbus_release_relid(newchannel->offermsg.child_relid);
 553
 554         free_channel(newchannel);
 555 }
 556
 557 /*
 558  * vmbus_process_offer - Process the offer by creating a channel/device
 559  * associated with this offer
 560  */
 561 static void vmbus_process_offer(struct vmbus_channel *newchannel)
 562 {
 563         struct vmbus_channel *channel;
 564         struct workqueue_struct *wq;
 565         unsigned long flags;
 566         bool fnew = true;
 567
 568         mutex_lock(&vmbus_connection.channel_mutex);
 569
 570         /*
 571          * Now that we have acquired the channel_mutex,
 572          * we can release the potentially racing rescind thread.
 573          */
 574         atomic_dec(&vmbus_connection.offer_in_progress);
 575
 576         list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
 577                 if (!uuid_le_cmp(channel->offermsg.offer.if_type,
 578                                  newchannel->offermsg.offer.if_type) &&
 579                     !uuid_le_cmp(channel->offermsg.offer.if_instance,
 580                                  newchannel->offermsg.offer.if_instance)) {
 581                         fnew = false;
 582                         break;
 583                 }
 584         }
 585
 586         if (fnew)
 587                 list_add_tail(&newchannel->listentry,
 588                               &vmbus_connection.chn_list);
 589         else {
 590                 /*
 591                  * Check to see if this is a valid sub-channel.
 592                  */
 593                 if (newchannel->offermsg.offer.sub_channel_index == 0) {
 594                         mutex_unlock(&vmbus_connection.channel_mutex);
 595                         /*
 596                          * Don't call free_channel(), because newchannel->kobj
 597                          * is not initialized yet.
 598                          */
 599                         kfree(newchannel);
 600                         WARN_ON_ONCE(1);
 601                         return;
 602                 }
 603                 /*
 604                  * Process the sub-channel.
 605                  */
 606                 newchannel->primary_channel = channel;
 607                 spin_lock_irqsave(&channel->lock, flags);
 608                 list_add_tail(&newchannel->sc_list, &channel->sc_list);
 609                 spin_unlock_irqrestore(&channel->lock, flags);
 610         }
 611
 612         mutex_unlock(&vmbus_connection.channel_mutex);
 613
 614         /*
 615          * vmbus_process_offer() mustn't call channel->sc_creation_callback()
 616          * directly for sub-channels, because sc_creation_callback() ->
 617          * vmbus_open() may never get the host's response to the
 618          * OPEN_CHANNEL message (the host may rescind a channel at any time,
 619          * e.g. in the case of hot removing a NIC), and vmbus_onoffer_rescind()
 620          * may not wake up the vmbus_open() as it's blocked due to a non-zero
 621          * vmbus_connection.offer_in_progress, and finally we have a deadlock.
 622          *
 623          * The above is also true for primary channels, if the related device
 624          * drivers use sync probing mode by default.
 625          *
 626          * And, usually the handling of primary channels and sub-channels can
 627          * depend on each other, so we should offload them to different
 628          * workqueues to avoid possible deadlock, e.g. in sync-probing mode,
 629          * NIC1's netvsc_subchan_work() can race with NIC2's netvsc_probe() ->
 630          * rtnl_lock(), and causes deadlock: the former gets the rtnl_lock
 631          * and waits for all the sub-channels to appear, but the latter
 632          * can't get the rtnl_lock and this blocks the handling of
 633          * sub-channels.
 634          */
 635         INIT_WORK(&newchannel->add_channel_work, vmbus_add_channel_work);
 636         wq = fnew ? vmbus_connection.handle_primary_chan_wq :
 637                     vmbus_connection.handle_sub_chan_wq;
 638         queue_work(wq, &newchannel->add_channel_work);
 639 }
 640
 641 /*
 642  * We use this state to statically distribute the channel interrupt load.
 643  */
 644 static int next_numa_node_id;
 645 /*
 646  * init_vp_index() accesses global variables like next_numa_node_id, and
 647  * it can run concurrently for primary channels and sub-channels: see
 648  * vmbus_process_offer(), so we need the lock to protect the global
 649  * variables.
 650  */
 651 static DEFINE_SPINLOCK(bind_channel_to_cpu_lock);
 652
 653 /*
 654  * Starting with Win8, we can statically distribute the incoming
 655  * channel interrupt load by binding a channel to VCPU.
 656  * We distribute the interrupt loads to one or more NUMA nodes based on
 657  * the channel's affinity_policy.
 658  *
 659  * For pre-win8 hosts or non-performance critical channels we assign the
 660  * first CPU in the first NUMA node.
 661  */
 662 static void init_vp_index(struct vmbus_channel *channel, u16 dev_type)
 663 {
 664         u32 cur_cpu;
 665         bool perf_chn = vmbus_devs[dev_type].perf_device;
 666         struct vmbus_channel *primary = channel->primary_channel;
 667         int next_node;
 668         cpumask_var_t available_mask;
 669         struct cpumask *alloced_mask;
 670
 671         if ((vmbus_proto_version == VERSION_WS2008) ||
 672             (vmbus_proto_version == VERSION_WIN7) || (!perf_chn) ||
 673             !alloc_cpumask_var(&available_mask, GFP_KERNEL)) {
 674                 /*
 675                  * Prior to win8, all channel interrupts are
 676                  * delivered on cpu 0.
 677                  * Also if the channel is not a performance critical
 678                  * channel, bind it to cpu 0.
 679                  * In case alloc_cpumask_var() fails, bind it to cpu 0.
 680                  */
 681                 channel->numa_node = 0;
 682                 channel->target_cpu = 0;
 683                 channel->target_vp = hv_cpu_number_to_vp_number(0);
 684                 return;
 685         }
 686
 687         spin_lock(&bind_channel_to_cpu_lock);
 688
 689         /*
 690          * Based on the channel affinity policy, we will assign the NUMA
 691          * nodes.
 692          */
 693
 694         if ((channel->affinity_policy == HV_BALANCED) || (!primary)) {
 695                 while (true) {
 696                         next_node = next_numa_node_id++;
 697                         if (next_node == nr_node_ids) {
 698                                 next_node = next_numa_node_id = 0;
 699                                 continue;
 700                         }
 701                         if (cpumask_empty(cpumask_of_node(next_node)))
 702                                 continue;
 703                         break;
 704                 }
 705                 channel->numa_node = next_node;
 706                 primary = channel;
 707         }
 708         alloced_mask = &hv_context.hv_numa_map[primary->numa_node];
 709
 710         if (cpumask_weight(alloced_mask) ==
 711             cpumask_weight(cpumask_of_node(primary->numa_node))) {
 712                 /*
 713                  * We have cycled through all the CPUs in the node;
 714                  * reset the alloced map.
 715                  */
 716                 cpumask_clear(alloced_mask);
 717         }
 718
 719         cpumask_xor(available_mask, alloced_mask,
 720                     cpumask_of_node(primary->numa_node));
 721
 722         cur_cpu = -1;
 723
 724         if (primary->affinity_policy == HV_LOCALIZED) {
 725                 /*
 726                  * Normally Hyper-V host doesn't create more subchannels
 727                  * than there are VCPUs on the node but it is possible when not
 728                  * all present VCPUs on the node are initialized by guest.
 729                  * Clear the alloced_cpus_in_node to start over.
 730                  */
 731                 if (cpumask_equal(&primary->alloced_cpus_in_node,
 732                                   cpumask_of_node(primary->numa_node)))
 733                         cpumask_clear(&primary->alloced_cpus_in_node);
 734         }
 735
 736         while (true) {
 737                 cur_cpu = cpumask_next(cur_cpu, available_mask);
 738                 if (cur_cpu >= nr_cpu_ids) {
 739                         cur_cpu = -1;
 740                         cpumask_copy(available_mask,
 741                                      cpumask_of_node(primary->numa_node));
 742                         continue;
 743                 }
 744
 745                 if (primary->affinity_policy == HV_LOCALIZED) {
 746                         /*
 747                          * NOTE: in the case of sub-channel, we clear the
 748                          * sub-channel related bit(s) in
 749                          * primary->alloced_cpus_in_node in
 750                          * hv_process_channel_removal(), so when we
 751                          * reload drivers like hv_netvsc in SMP guest, here
 752                          * we're able to re-allocate
 753                          * bit from primary->alloced_cpus_in_node.
 754                          */
 755                         if (!cpumask_test_cpu(cur_cpu,
 756                                               &primary->alloced_cpus_in_node)) {
 757                                 cpumask_set_cpu(cur_cpu,
 758                                                 &primary->alloced_cpus_in_node);
 759                                 cpumask_set_cpu(cur_cpu, alloced_mask);
 760                                 break;
 761                         }
 762                 } else {
 763                         cpumask_set_cpu(cur_cpu, alloced_mask);
 764                         break;
 765                 }
 766         }
 767
 768         channel->target_cpu = cur_cpu;
 769         channel->target_vp = hv_cpu_number_to_vp_number(cur_cpu);
 770
 771         spin_unlock(&bind_channel_to_cpu_lock);
 772
 773         free_cpumask_var(available_mask);
 774 }
 775
 776 #define UNLOAD_DELAY_UNIT_MS    10              /* 10 milliseconds */
 777 #define UNLOAD_WAIT_MS          (100*1000)      /* 100 seconds */
 778 #define UNLOAD_WAIT_LOOPS       (UNLOAD_WAIT_MS/UNLOAD_DELAY_UNIT_MS)
 779 #define UNLOAD_MSG_MS           (5*1000)        /* Every 5 seconds */
 780 #define UNLOAD_MSG_LOOPS        (UNLOAD_MSG_MS/UNLOAD_DELAY_UNIT_MS)
 781
 782 static void vmbus_wait_for_unload(void)
 783 {
 784         int cpu;
 785         void *page_addr;
 786         struct hv_message *msg;
 787         struct vmbus_channel_message_header *hdr;
 788         u32 message_type, i;
 789
 790         /*
 791          * CHANNELMSG_UNLOAD_RESPONSE is always delivered to the CPU which was
 792          * used for initial contact or to CPU0 depending on host version. When
 793          * we're crashing on a different CPU let's hope that IRQ handler on
 794          * the cpu which receives CHANNELMSG_UNLOAD_RESPONSE is still
 795          * functional and vmbus_unload_response() will complete
 796          * vmbus_connection.unload_event. If not, the last thing we can do is
 797          * read message pages for all CPUs directly.
 798          *
 799          * Wait up to 100 seconds since an Azure host must writeback any dirty
 800          * data in its disk cache before the VMbus UNLOAD request will
 801          * complete. This flushing has been empirically observed to take up
 802          * to 50 seconds in cases with a lot of dirty data, so allow additional
 803          * leeway and for inaccuracies in mdelay(). But eventually time out so
 804          * that the panic path can't get hung forever in case the response
 805          * message isn't seen.
 806          */
 807         for (i = 1; i <= UNLOAD_WAIT_LOOPS; i++) {
 808                 if (completion_done(&vmbus_connection.unload_event))
 809                         goto completed;
 810
 811                 for_each_online_cpu(cpu) {
 812                         struct hv_per_cpu_context *hv_cpu
 813                                 = per_cpu_ptr(hv_context.cpu_context, cpu);
 814
 815                         page_addr = hv_cpu->synic_message_page;
 816                         msg = (struct hv_message *)page_addr
 817                                 + VMBUS_MESSAGE_SINT;
 818
 819                         message_type = READ_ONCE(msg->header.message_type);
 820                         if (message_type == HVMSG_NONE)
 821                                 continue;
 822
 823                         hdr = (struct vmbus_channel_message_header *)
 824                                 msg->u.payload;
 825
 826                         if (hdr->msgtype == CHANNELMSG_UNLOAD_RESPONSE)
 827                                 complete(&vmbus_connection.unload_event);
 828
 829                         vmbus_signal_eom(msg, message_type);
 830                 }
 831
 832                 /*
 833                  * Give a notice periodically so someone watching the
 834                  * serial output won't think it is completely hung.
 835                  */
 836                 if (!(i % UNLOAD_MSG_LOOPS))
 837                         pr_notice("Waiting for VMBus UNLOAD to complete\n");
 838
 839                 mdelay(UNLOAD_DELAY_UNIT_MS);
 840         }
 841         pr_err("Continuing even though VMBus UNLOAD did not complete\n");
 842
 843 completed:
 844         /*
 845          * We're crashing and already got the UNLOAD_RESPONSE, cleanup all
 846          * maybe-pending messages on all CPUs to be able to receive new
 847          * messages after we reconnect.
 848          */
 849         for_each_online_cpu(cpu) {
 850                 struct hv_per_cpu_context *hv_cpu
 851                         = per_cpu_ptr(hv_context.cpu_context, cpu);
 852
 853                 page_addr = hv_cpu->synic_message_page;
 854                 msg = (struct hv_message *)page_addr + VMBUS_MESSAGE_SINT;
 855                 msg->header.message_type = HVMSG_NONE;
 856         }
 857 }
 858
 859 /*
 860  * vmbus_unload_response - Handler for the unload response.
 861  */
 862 static void vmbus_unload_response(struct vmbus_channel_message_header *hdr)
 863 {
 864         /*
 865          * This is a global event; just wakeup the waiting thread.
 866          * Once we successfully unload, we can cleanup the monitor state.
 867          */
 868         complete(&vmbus_connection.unload_event);
 869 }
 870
 871 void vmbus_initiate_unload(bool crash)
 872 {
 873         struct vmbus_channel_message_header hdr;
 874
 875         if (xchg(&vmbus_connection.conn_state, DISCONNECTED) == DISCONNECTED)
 876                 return;
 877
 878         /* Pre-Win2012R2 hosts don't support reconnect */
 879         if (vmbus_proto_version < VERSION_WIN8_1)
 880                 return;
 881
 882         init_completion(&vmbus_connection.unload_event);
 883         memset(&hdr, 0, sizeof(struct vmbus_channel_message_header));
 884         hdr.msgtype = CHANNELMSG_UNLOAD;
 885         vmbus_post_msg(&hdr, sizeof(struct vmbus_channel_message_header),
 886                        !crash);
 887
 888         /*
 889          * vmbus_initiate_unload() is also called on crash and the crash can be
 890          * happening in an interrupt context, where scheduling is impossible.
 891          */
 892         if (!crash)
 893                 wait_for_completion(&vmbus_connection.unload_event);
 894         else
 895                 vmbus_wait_for_unload();
 896 }
 897
 898 /*
 899  * vmbus_onoffer - Handler for channel offers from vmbus in parent partition.
 900  *
 901  */
 902 static void vmbus_onoffer(struct vmbus_channel_message_header *hdr)
 903 {
 904         struct vmbus_channel_offer_channel *offer;
 905         struct vmbus_channel *newchannel;
 906
 907         offer = (struct vmbus_channel_offer_channel *)hdr;
 908
 909         trace_vmbus_onoffer(offer);
 910
 911         /* Allocate the channel object and save this offer. */
 912         newchannel = alloc_channel();
 913         if (!newchannel) {
 914                 vmbus_release_relid(offer->child_relid);
 915                 atomic_dec(&vmbus_connection.offer_in_progress);
 916                 pr_err("Unable to allocate channel object\n");
 917                 return;
 918         }
 919
 920         /*
 921          * Setup state for signalling the host.
 922          */
 923         newchannel->sig_event = VMBUS_EVENT_CONNECTION_ID;
 924
 925         if (vmbus_proto_version != VERSION_WS2008) {
 926                 newchannel->is_dedicated_interrupt =
 927                                 (offer->is_dedicated_interrupt != 0);
 928                 newchannel->sig_event = offer->connection_id;
 929         }
 930
 931         memcpy(&newchannel->offermsg, offer,
 932                sizeof(struct vmbus_channel_offer_channel));
 933         newchannel->monitor_grp = (u8)offer->monitorid / 32;
 934         newchannel->monitor_bit = (u8)offer->monitorid % 32;
 935
 936         vmbus_process_offer(newchannel);
 937 }
 938
 939 /*
 940  * vmbus_onoffer_rescind - Rescind offer handler.
 941  *
 942  * We queue a work item to process this offer synchronously
 943  */
 944 static void vmbus_onoffer_rescind(struct vmbus_channel_message_header *hdr)
 945 {
 946         struct vmbus_channel_rescind_offer *rescind;
 947         struct vmbus_channel *channel;
 948         struct device *dev;
 949
 950         rescind = (struct vmbus_channel_rescind_offer *)hdr;
 951
 952         trace_vmbus_onoffer_rescind(rescind);
 953
 954         /*
 955          * The offer msg and the corresponding rescind msg
 956          * from the host are guranteed to be ordered -
 957          * offer comes in first and then the rescind.
 958          * Since we process these events in work elements,
 959          * and with preemption, we may end up processing
 960          * the events out of order. Given that we handle these
 961          * work elements on the same CPU, this is possible only
 962          * in the case of preemption. In any case wait here
 963          * until the offer processing has moved beyond the
 964          * point where the channel is discoverable.
 965          */
 966
 967         while (atomic_read(&vmbus_connection.offer_in_progress) != 0) {
 968                 /*
 969                  * We wait here until any channel offer is currently
 970                  * being processed.
 971                  */
 972                 msleep(1);
 973         }
 974
 975         mutex_lock(&vmbus_connection.channel_mutex);
 976         channel = relid2channel(rescind->child_relid);
 977         mutex_unlock(&vmbus_connection.channel_mutex);
 978
 979         if (channel == NULL) {
 980                 /*
 981                  * We failed in processing the offer message;
 982                  * we would have cleaned up the relid in that
 983                  * failure path.
 984                  */
 985                 return;
 986         }
 987
 988         /*
 989          * Before setting channel->rescind in vmbus_rescind_cleanup(), we
 990          * should make sure the channel callback is not running any more.
 991          */
 992         vmbus_reset_channel_cb(channel);
 993
 994         /*
 995          * Now wait for offer handling to complete.
 996          */
 997         vmbus_rescind_cleanup(channel);
 998         while (READ_ONCE(channel->probe_done) == false) {
 999                 /*
1000                  * We wait here until any channel offer is currently
1001                  * being processed.
1002                  */
1003                 msleep(1);
1004         }
1005
1006         /*
1007          * At this point, the rescind handling can proceed safely.
1008          */
1009
1010         if (channel->device_obj) {
1011                 if (channel->chn_rescind_callback) {
1012                         channel->chn_rescind_callback(channel);
1013                         return;
1014                 }
1015                 /*
1016                  * We will have to unregister this device from the
1017                  * driver core.
1018                  */
1019                 dev = get_device(&channel->device_obj->device);
1020                 if (dev) {
1021                         vmbus_device_unregister(channel->device_obj);
1022                         put_device(dev);
1023                 }
1024         } else if (channel->primary_channel != NULL) {
1025                 /*
1026                  * Sub-channel is being rescinded. Following is the channel
1027                  * close sequence when initiated from the driveri (refer to
1028                  * vmbus_close() for details):
1029                  * 1. Close all sub-channels first
1030                  * 2. Then close the primary channel.
1031                  */
1032                 mutex_lock(&vmbus_connection.channel_mutex);
1033                 if (channel->state == CHANNEL_OPEN_STATE) {
1034                         /*
1035                          * The channel is currently not open;
1036                          * it is safe for us to cleanup the channel.
1037                          */
1038                         hv_process_channel_removal(rescind->child_relid);
1039                 } else {
1040                         complete(&channel->rescind_event);
1041                 }
1042                 mutex_unlock(&vmbus_connection.channel_mutex);
1043         }
1044 }
1045
1046 void vmbus_hvsock_device_unregister(struct vmbus_channel *channel)
1047 {
1048         BUG_ON(!is_hvsock_channel(channel));
1049
1050         /* We always get a rescind msg when a connection is closed. */
1051         while (!READ_ONCE(channel->probe_done) || !READ_ONCE(channel->rescind))
1052                 msleep(1);
1053
1054         vmbus_device_unregister(channel->device_obj);
1055 }
1056 EXPORT_SYMBOL_GPL(vmbus_hvsock_device_unregister);
1057
1058
1059 /*
1060  * vmbus_onoffers_delivered -
1061  * This is invoked when all offers have been delivered.
1062  *
1063  * Nothing to do here.
1064  */
1065 static void vmbus_onoffers_delivered(
1066                         struct vmbus_channel_message_header *hdr)
1067 {
1068 }
1069
1070 /*
1071  * vmbus_onopen_result - Open result handler.
1072  *
1073  * This is invoked when we received a response to our channel open request.
1074  * Find the matching request, copy the response and signal the requesting
1075  * thread.
1076  */
1077 static void vmbus_onopen_result(struct vmbus_channel_message_header *hdr)
1078 {
1079         struct vmbus_channel_open_result *result;
1080         struct vmbus_channel_msginfo *msginfo;
1081         struct vmbus_channel_message_header *requestheader;
1082         struct vmbus_channel_open_channel *openmsg;
1083         unsigned long flags;
1084
1085         result = (struct vmbus_channel_open_result *)hdr;
1086
1087         trace_vmbus_onopen_result(result);
1088
1089         /*
1090          * Find the open msg, copy the result and signal/unblock the wait event
1091          */
1092         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1093
1094         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1095                                 msglistentry) {
1096                 requestheader =
1097                         (struct vmbus_channel_message_header *)msginfo->msg;
1098
1099                 if (requestheader->msgtype == CHANNELMSG_OPENCHANNEL) {
1100                         openmsg =
1101                         (struct vmbus_channel_open_channel *)msginfo->msg;
1102                         if (openmsg->child_relid == result->child_relid &&
1103                             openmsg->openid == result->openid) {
1104                                 memcpy(&msginfo->response.open_result,
1105                                        result,
1106                                        sizeof(
1107                                         struct vmbus_channel_open_result));
1108                                 complete(&msginfo->waitevent);
1109                                 break;
1110                         }
1111                 }
1112         }
1113         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1114 }
1115
1116 /*
1117  * vmbus_ongpadl_created - GPADL created handler.
1118  *
1119  * This is invoked when we received a response to our gpadl create request.
1120  * Find the matching request, copy the response and signal the requesting
1121  * thread.
1122  */
1123 static void vmbus_ongpadl_created(struct vmbus_channel_message_header *hdr)
1124 {
1125         struct vmbus_channel_gpadl_created *gpadlcreated;
1126         struct vmbus_channel_msginfo *msginfo;
1127         struct vmbus_channel_message_header *requestheader;
1128         struct vmbus_channel_gpadl_header *gpadlheader;
1129         unsigned long flags;
1130
1131         gpadlcreated = (struct vmbus_channel_gpadl_created *)hdr;
1132
1133         trace_vmbus_ongpadl_created(gpadlcreated);
1134
1135         /*
1136          * Find the establish msg, copy the result and signal/unblock the wait
1137          * event
1138          */
1139         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1140
1141         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1142                                 msglistentry) {
1143                 requestheader =
1144                         (struct vmbus_channel_message_header *)msginfo->msg;
1145
1146                 if (requestheader->msgtype == CHANNELMSG_GPADL_HEADER) {
1147                         gpadlheader =
1148                         (struct vmbus_channel_gpadl_header *)requestheader;
1149
1150                         if ((gpadlcreated->child_relid ==
1151                              gpadlheader->child_relid) &&
1152                             (gpadlcreated->gpadl == gpadlheader->gpadl)) {
1153                                 memcpy(&msginfo->response.gpadl_created,
1154                                        gpadlcreated,
1155                                        sizeof(
1156                                         struct vmbus_channel_gpadl_created));
1157                                 complete(&msginfo->waitevent);
1158                                 break;
1159                         }
1160                 }
1161         }
1162         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1163 }
1164
1165 /*
1166  * vmbus_ongpadl_torndown - GPADL torndown handler.
1167  *
1168  * This is invoked when we received a response to our gpadl teardown request.
1169  * Find the matching request, copy the response and signal the requesting
1170  * thread.
1171  */
1172 static void vmbus_ongpadl_torndown(
1173                         struct vmbus_channel_message_header *hdr)
1174 {
1175         struct vmbus_channel_gpadl_torndown *gpadl_torndown;
1176         struct vmbus_channel_msginfo *msginfo;
1177         struct vmbus_channel_message_header *requestheader;
1178         struct vmbus_channel_gpadl_teardown *gpadl_teardown;
1179         unsigned long flags;
1180
1181         gpadl_torndown = (struct vmbus_channel_gpadl_torndown *)hdr;
1182
1183         trace_vmbus_ongpadl_torndown(gpadl_torndown);
1184
1185         /*
1186          * Find the open msg, copy the result and signal/unblock the wait event
1187          */
1188         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1189
1190         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1191                                 msglistentry) {
1192                 requestheader =
1193                         (struct vmbus_channel_message_header *)msginfo->msg;
1194
1195                 if (requestheader->msgtype == CHANNELMSG_GPADL_TEARDOWN) {
1196                         gpadl_teardown =
1197                         (struct vmbus_channel_gpadl_teardown *)requestheader;
1198
1199                         if (gpadl_torndown->gpadl == gpadl_teardown->gpadl) {
1200                                 memcpy(&msginfo->response.gpadl_torndown,
1201                                        gpadl_torndown,
1202                                        sizeof(
1203                                         struct vmbus_channel_gpadl_torndown));
1204                                 complete(&msginfo->waitevent);
1205                                 break;
1206                         }
1207                 }
1208         }
1209         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1210 }
1211
1212 /*
1213  * vmbus_onversion_response - Version response handler
1214  *
1215  * This is invoked when we received a response to our initiate contact request.
1216  * Find the matching request, copy the response and signal the requesting
1217  * thread.
1218  */
1219 static void vmbus_onversion_response(
1220                 struct vmbus_channel_message_header *hdr)
1221 {
1222         struct vmbus_channel_msginfo *msginfo;
1223         struct vmbus_channel_message_header *requestheader;
1224         struct vmbus_channel_version_response *version_response;
1225         unsigned long flags;
1226
1227         version_response = (struct vmbus_channel_version_response *)hdr;
1228
1229         trace_vmbus_onversion_response(version_response);
1230
1231         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1232
1233         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1234                                 msglistentry) {
1235                 requestheader =
1236                         (struct vmbus_channel_message_header *)msginfo->msg;
1237
1238                 if (requestheader->msgtype ==
1239                     CHANNELMSG_INITIATE_CONTACT) {
1240                         memcpy(&msginfo->response.version_response,
1241                               version_response,
1242                               sizeof(struct vmbus_channel_version_response));
1243                         complete(&msginfo->waitevent);
1244                 }
1245         }
1246         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1247 }
1248
1249 /* Channel message dispatch table */
1250 const struct vmbus_channel_message_table_entry
1251 channel_message_table[CHANNELMSG_COUNT] = {
1252         { CHANNELMSG_INVALID,                   0, NULL },
1253         { CHANNELMSG_OFFERCHANNEL,              0, vmbus_onoffer },
1254         { CHANNELMSG_RESCIND_CHANNELOFFER,      0, vmbus_onoffer_rescind },
1255         { CHANNELMSG_REQUESTOFFERS,             0, NULL },
1256         { CHANNELMSG_ALLOFFERS_DELIVERED,       1, vmbus_onoffers_delivered },
1257         { CHANNELMSG_OPENCHANNEL,               0, NULL },
1258         { CHANNELMSG_OPENCHANNEL_RESULT,        1, vmbus_onopen_result },
1259         { CHANNELMSG_CLOSECHANNEL,              0, NULL },
1260         { CHANNELMSG_GPADL_HEADER,              0, NULL },
1261         { CHANNELMSG_GPADL_BODY,                0, NULL },
1262         { CHANNELMSG_GPADL_CREATED,             1, vmbus_ongpadl_created },
1263         { CHANNELMSG_GPADL_TEARDOWN,            0, NULL },
1264         { CHANNELMSG_GPADL_TORNDOWN,            1, vmbus_ongpadl_torndown },
1265         { CHANNELMSG_RELID_RELEASED,            0, NULL },
1266         { CHANNELMSG_INITIATE_CONTACT,          0, NULL },
1267         { CHANNELMSG_VERSION_RESPONSE,          1, vmbus_onversion_response },
1268         { CHANNELMSG_UNLOAD,                    0, NULL },
1269         { CHANNELMSG_UNLOAD_RESPONSE,           1, vmbus_unload_response },
1270         { CHANNELMSG_18,                        0, NULL },
1271         { CHANNELMSG_19,                        0, NULL },
1272         { CHANNELMSG_20,                        0, NULL },
1273         { CHANNELMSG_TL_CONNECT_REQUEST,        0, NULL },
1274         { CHANNELMSG_22,                        0, NULL },
1275         { CHANNELMSG_TL_CONNECT_RESULT,         0, NULL },
1276 };
1277
1278 /*
1279  * vmbus_onmessage - Handler for channel protocol messages.
1280  *
1281  * This is invoked in the vmbus worker thread context.
1282  */
1283 void vmbus_onmessage(void *context)
1284 {
1285         struct hv_message *msg = context;
1286         struct vmbus_channel_message_header *hdr;
1287
1288         hdr = (struct vmbus_channel_message_header *)msg->u.payload;
1289
1290         trace_vmbus_on_message(hdr);
1291
1292         /*
1293          * vmbus_on_msg_dpc() makes sure the hdr->msgtype here can not go
1294          * out of bound and the message_handler pointer can not be NULL.
1295          */
1296         channel_message_table[hdr->msgtype].message_handler(hdr);
1297 }
1298
1299 /*
1300  * vmbus_request_offers - Send a request to get all our pending offers.
1301  */
1302 int vmbus_request_offers(void)
1303 {
1304         struct vmbus_channel_message_header *msg;
1305         struct vmbus_channel_msginfo *msginfo;
1306         int ret;
1307
1308         msginfo = kmalloc(sizeof(*msginfo) +
1309                           sizeof(struct vmbus_channel_message_header),
1310                           GFP_KERNEL);
1311         if (!msginfo)
1312                 return -ENOMEM;
1313
1314         msg = (struct vmbus_channel_message_header *)msginfo->msg;
1315
1316         msg->msgtype = CHANNELMSG_REQUESTOFFERS;
1317
1318         ret = vmbus_post_msg(msg, sizeof(struct vmbus_channel_message_header),
1319                              true);
1320
1321         trace_vmbus_request_offers(ret);
1322
1323         if (ret != 0) {
1324                 pr_err("Unable to request offers - %d\n", ret);
1325
1326                 goto cleanup;
1327         }
1328
1329 cleanup:
1330         kfree(msginfo);
1331
1332         return ret;
1333 }
1334
1335 /*
1336  * Retrieve the (sub) channel on which to send an outgoing request.
1337  * When a primary channel has multiple sub-channels, we try to
1338  * distribute the load equally amongst all available channels.
1339  */
1340 struct vmbus_channel *vmbus_get_outgoing_channel(struct vmbus_channel *primary)
1341 {
1342         struct list_head *cur, *tmp;
1343         int cur_cpu;
1344         struct vmbus_channel *cur_channel;
1345         struct vmbus_channel *outgoing_channel = primary;
1346         int next_channel;
1347         int i = 1;
1348
1349         if (list_empty(&primary->sc_list))
1350                 return outgoing_channel;
1351
1352         next_channel = primary->next_oc++;
1353
1354         if (next_channel > (primary->num_sc)) {
1355                 primary->next_oc = 0;
1356                 return outgoing_channel;
1357         }
1358
1359         cur_cpu = hv_cpu_number_to_vp_number(smp_processor_id());
1360         list_for_each_safe(cur, tmp, &primary->sc_list) {
1361                 cur_channel = list_entry(cur, struct vmbus_channel, sc_list);
1362                 if (cur_channel->state != CHANNEL_OPENED_STATE)
1363                         continue;
1364
1365                 if (cur_channel->target_vp == cur_cpu)
1366                         return cur_channel;
1367
1368                 if (i == next_channel)
1369                         return cur_channel;
1370
1371                 i++;
1372         }
1373
1374         return outgoing_channel;
1375 }
1376 EXPORT_SYMBOL_GPL(vmbus_get_outgoing_channel);
1377
1378 static void invoke_sc_cb(struct vmbus_channel *primary_channel)
1379 {
1380         struct list_head *cur, *tmp;
1381         struct vmbus_channel *cur_channel;
1382
1383         if (primary_channel->sc_creation_callback == NULL)
1384                 return;
1385
1386         list_for_each_safe(cur, tmp, &primary_channel->sc_list) {
1387                 cur_channel = list_entry(cur, struct vmbus_channel, sc_list);
1388
1389                 primary_channel->sc_creation_callback(cur_channel);
1390         }
1391 }
1392
1393 void vmbus_set_sc_create_callback(struct vmbus_channel *primary_channel,
1394                                 void (*sc_cr_cb)(struct vmbus_channel *new_sc))
1395 {
1396         primary_channel->sc_creation_callback = sc_cr_cb;
1397 }
1398 EXPORT_SYMBOL_GPL(vmbus_set_sc_create_callback);
1399
1400 bool vmbus_are_subchannels_present(struct vmbus_channel *primary)
1401 {
1402         bool ret;
1403
1404         ret = !list_empty(&primary->sc_list);
1405
1406         if (ret) {
1407                 /*
1408                  * Invoke the callback on sub-channel creation.
1409                  * This will present a uniform interface to the
1410                  * clients.
1411                  */
1412                 invoke_sc_cb(primary);
1413         }
1414
1415         return ret;
1416 }
1417 EXPORT_SYMBOL_GPL(vmbus_are_subchannels_present);
1418
1419 void vmbus_set_chn_rescind_callback(struct vmbus_channel *channel,
1420                 void (*chn_rescind_cb)(struct vmbus_channel *))
1421 {
1422         channel->chn_rescind_callback = chn_rescind_cb;
1423 }
1424 EXPORT_SYMBOL_GPL(vmbus_set_chn_rescind_callback);