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         ret = vmbus_device_register(newchannel->device_obj);
 513
 514         if (ret != 0) {
 515                 pr_err("unable to add child device object (relid %d)\n",
 516                         newchannel->offermsg.child_relid);
 517                 kfree(newchannel->device_obj);
 518                 goto err_deq_chan;
 519         }
 520
 521         newchannel->probe_done = true;
 522         return;
 523
 524 err_deq_chan:
 525         mutex_lock(&vmbus_connection.channel_mutex);
 526
 527         /*
 528          * We need to set the flag, otherwise
 529          * vmbus_onoffer_rescind() can be blocked.
 530          */
 531         newchannel->probe_done = true;
 532
 533         if (primary_channel == NULL) {
 534                 list_del(&newchannel->listentry);
 535         } else {
 536                 spin_lock_irqsave(&primary_channel->lock, flags);
 537                 list_del(&newchannel->sc_list);
 538                 spin_unlock_irqrestore(&primary_channel->lock, flags);
 539         }
 540
 541         mutex_unlock(&vmbus_connection.channel_mutex);
 542
 543         if (newchannel->target_cpu != get_cpu()) {
 544                 put_cpu();
 545                 smp_call_function_single(newchannel->target_cpu,
 546                                          percpu_channel_deq,
 547                                          newchannel, true);
 548         } else {
 549                 percpu_channel_deq(newchannel);
 550                 put_cpu();
 551         }
 552
 553         vmbus_release_relid(newchannel->offermsg.child_relid);
 554
 555         free_channel(newchannel);
 556 }
 557
 558 /*
 559  * vmbus_process_offer - Process the offer by creating a channel/device
 560  * associated with this offer
 561  */
 562 static void vmbus_process_offer(struct vmbus_channel *newchannel)
 563 {
 564         struct vmbus_channel *channel;
 565         struct workqueue_struct *wq;
 566         unsigned long flags;
 567         bool fnew = true;
 568
 569         mutex_lock(&vmbus_connection.channel_mutex);
 570
 571         /*
 572          * Now that we have acquired the channel_mutex,
 573          * we can release the potentially racing rescind thread.
 574          */
 575         atomic_dec(&vmbus_connection.offer_in_progress);
 576
 577         list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
 578                 if (!uuid_le_cmp(channel->offermsg.offer.if_type,
 579                                  newchannel->offermsg.offer.if_type) &&
 580                     !uuid_le_cmp(channel->offermsg.offer.if_instance,
 581                                  newchannel->offermsg.offer.if_instance)) {
 582                         fnew = false;
 583                         break;
 584                 }
 585         }
 586
 587         if (fnew)
 588                 list_add_tail(&newchannel->listentry,
 589                               &vmbus_connection.chn_list);
 590         else {
 591                 /*
 592                  * Check to see if this is a valid sub-channel.
 593                  */
 594                 if (newchannel->offermsg.offer.sub_channel_index == 0) {
 595                         mutex_unlock(&vmbus_connection.channel_mutex);
 596                         /*
 597                          * Don't call free_channel(), because newchannel->kobj
 598                          * is not initialized yet.
 599                          */
 600                         kfree(newchannel);
 601                         WARN_ON_ONCE(1);
 602                         return;
 603                 }
 604                 /*
 605                  * Process the sub-channel.
 606                  */
 607                 newchannel->primary_channel = channel;
 608                 spin_lock_irqsave(&channel->lock, flags);
 609                 list_add_tail(&newchannel->sc_list, &channel->sc_list);
 610                 spin_unlock_irqrestore(&channel->lock, flags);
 611         }
 612
 613         mutex_unlock(&vmbus_connection.channel_mutex);
 614
 615         /*
 616          * vmbus_process_offer() mustn't call channel->sc_creation_callback()
 617          * directly for sub-channels, because sc_creation_callback() ->
 618          * vmbus_open() may never get the host's response to the
 619          * OPEN_CHANNEL message (the host may rescind a channel at any time,
 620          * e.g. in the case of hot removing a NIC), and vmbus_onoffer_rescind()
 621          * may not wake up the vmbus_open() as it's blocked due to a non-zero
 622          * vmbus_connection.offer_in_progress, and finally we have a deadlock.
 623          *
 624          * The above is also true for primary channels, if the related device
 625          * drivers use sync probing mode by default.
 626          *
 627          * And, usually the handling of primary channels and sub-channels can
 628          * depend on each other, so we should offload them to different
 629          * workqueues to avoid possible deadlock, e.g. in sync-probing mode,
 630          * NIC1's netvsc_subchan_work() can race with NIC2's netvsc_probe() ->
 631          * rtnl_lock(), and causes deadlock: the former gets the rtnl_lock
 632          * and waits for all the sub-channels to appear, but the latter
 633          * can't get the rtnl_lock and this blocks the handling of
 634          * sub-channels.
 635          */
 636         INIT_WORK(&newchannel->add_channel_work, vmbus_add_channel_work);
 637         wq = fnew ? vmbus_connection.handle_primary_chan_wq :
 638                     vmbus_connection.handle_sub_chan_wq;
 639         queue_work(wq, &newchannel->add_channel_work);
 640 }
 641
 642 /*
 643  * We use this state to statically distribute the channel interrupt load.
 644  */
 645 static int next_numa_node_id;
 646 /*
 647  * init_vp_index() accesses global variables like next_numa_node_id, and
 648  * it can run concurrently for primary channels and sub-channels: see
 649  * vmbus_process_offer(), so we need the lock to protect the global
 650  * variables.
 651  */
 652 static DEFINE_SPINLOCK(bind_channel_to_cpu_lock);
 653
 654 /*
 655  * Starting with Win8, we can statically distribute the incoming
 656  * channel interrupt load by binding a channel to VCPU.
 657  * We distribute the interrupt loads to one or more NUMA nodes based on
 658  * the channel's affinity_policy.
 659  *
 660  * For pre-win8 hosts or non-performance critical channels we assign the
 661  * first CPU in the first NUMA node.
 662  */
 663 static void init_vp_index(struct vmbus_channel *channel, u16 dev_type)
 664 {
 665         u32 cur_cpu;
 666         bool perf_chn = vmbus_devs[dev_type].perf_device;
 667         struct vmbus_channel *primary = channel->primary_channel;
 668         int next_node;
 669         cpumask_var_t available_mask;
 670         struct cpumask *alloced_mask;
 671
 672         if ((vmbus_proto_version == VERSION_WS2008) ||
 673             (vmbus_proto_version == VERSION_WIN7) || (!perf_chn) ||
 674             !alloc_cpumask_var(&available_mask, GFP_KERNEL)) {
 675                 /*
 676                  * Prior to win8, all channel interrupts are
 677                  * delivered on cpu 0.
 678                  * Also if the channel is not a performance critical
 679                  * channel, bind it to cpu 0.
 680                  * In case alloc_cpumask_var() fails, bind it to cpu 0.
 681                  */
 682                 channel->numa_node = 0;
 683                 channel->target_cpu = 0;
 684                 channel->target_vp = hv_cpu_number_to_vp_number(0);
 685                 return;
 686         }
 687
 688         spin_lock(&bind_channel_to_cpu_lock);
 689
 690         /*
 691          * Based on the channel affinity policy, we will assign the NUMA
 692          * nodes.
 693          */
 694
 695         if ((channel->affinity_policy == HV_BALANCED) || (!primary)) {
 696                 while (true) {
 697                         next_node = next_numa_node_id++;
 698                         if (next_node == nr_node_ids) {
 699                                 next_node = next_numa_node_id = 0;
 700                                 continue;
 701                         }
 702                         if (cpumask_empty(cpumask_of_node(next_node)))
 703                                 continue;
 704                         break;
 705                 }
 706                 channel->numa_node = next_node;
 707                 primary = channel;
 708         }
 709         alloced_mask = &hv_context.hv_numa_map[primary->numa_node];
 710
 711         if (cpumask_weight(alloced_mask) ==
 712             cpumask_weight(cpumask_of_node(primary->numa_node))) {
 713                 /*
 714                  * We have cycled through all the CPUs in the node;
 715                  * reset the alloced map.
 716                  */
 717                 cpumask_clear(alloced_mask);
 718         }
 719
 720         cpumask_xor(available_mask, alloced_mask,
 721                     cpumask_of_node(primary->numa_node));
 722
 723         cur_cpu = -1;
 724
 725         if (primary->affinity_policy == HV_LOCALIZED) {
 726                 /*
 727                  * Normally Hyper-V host doesn't create more subchannels
 728                  * than there are VCPUs on the node but it is possible when not
 729                  * all present VCPUs on the node are initialized by guest.
 730                  * Clear the alloced_cpus_in_node to start over.
 731                  */
 732                 if (cpumask_equal(&primary->alloced_cpus_in_node,
 733                                   cpumask_of_node(primary->numa_node)))
 734                         cpumask_clear(&primary->alloced_cpus_in_node);
 735         }
 736
 737         while (true) {
 738                 cur_cpu = cpumask_next(cur_cpu, available_mask);
 739                 if (cur_cpu >= nr_cpu_ids) {
 740                         cur_cpu = -1;
 741                         cpumask_copy(available_mask,
 742                                      cpumask_of_node(primary->numa_node));
 743                         continue;
 744                 }
 745
 746                 if (primary->affinity_policy == HV_LOCALIZED) {
 747                         /*
 748                          * NOTE: in the case of sub-channel, we clear the
 749                          * sub-channel related bit(s) in
 750                          * primary->alloced_cpus_in_node in
 751                          * hv_process_channel_removal(), so when we
 752                          * reload drivers like hv_netvsc in SMP guest, here
 753                          * we're able to re-allocate
 754                          * bit from primary->alloced_cpus_in_node.
 755                          */
 756                         if (!cpumask_test_cpu(cur_cpu,
 757                                               &primary->alloced_cpus_in_node)) {
 758                                 cpumask_set_cpu(cur_cpu,
 759                                                 &primary->alloced_cpus_in_node);
 760                                 cpumask_set_cpu(cur_cpu, alloced_mask);
 761                                 break;
 762                         }
 763                 } else {
 764                         cpumask_set_cpu(cur_cpu, alloced_mask);
 765                         break;
 766                 }
 767         }
 768
 769         channel->target_cpu = cur_cpu;
 770         channel->target_vp = hv_cpu_number_to_vp_number(cur_cpu);
 771
 772         spin_unlock(&bind_channel_to_cpu_lock);
 773
 774         free_cpumask_var(available_mask);
 775 }
 776
 777 #define UNLOAD_DELAY_UNIT_MS    10              /* 10 milliseconds */
 778 #define UNLOAD_WAIT_MS          (100*1000)      /* 100 seconds */
 779 #define UNLOAD_WAIT_LOOPS       (UNLOAD_WAIT_MS/UNLOAD_DELAY_UNIT_MS)
 780 #define UNLOAD_MSG_MS           (5*1000)        /* Every 5 seconds */
 781 #define UNLOAD_MSG_LOOPS        (UNLOAD_MSG_MS/UNLOAD_DELAY_UNIT_MS)
 782
 783 static void vmbus_wait_for_unload(void)
 784 {
 785         int cpu;
 786         void *page_addr;
 787         struct hv_message *msg;
 788         struct vmbus_channel_message_header *hdr;
 789         u32 message_type, i;
 790
 791         /*
 792          * CHANNELMSG_UNLOAD_RESPONSE is always delivered to the CPU which was
 793          * used for initial contact or to CPU0 depending on host version. When
 794          * we're crashing on a different CPU let's hope that IRQ handler on
 795          * the cpu which receives CHANNELMSG_UNLOAD_RESPONSE is still
 796          * functional and vmbus_unload_response() will complete
 797          * vmbus_connection.unload_event. If not, the last thing we can do is
 798          * read message pages for all CPUs directly.
 799          *
 800          * Wait up to 100 seconds since an Azure host must writeback any dirty
 801          * data in its disk cache before the VMbus UNLOAD request will
 802          * complete. This flushing has been empirically observed to take up
 803          * to 50 seconds in cases with a lot of dirty data, so allow additional
 804          * leeway and for inaccuracies in mdelay(). But eventually time out so
 805          * that the panic path can't get hung forever in case the response
 806          * message isn't seen.
 807          */
 808         for (i = 1; i <= UNLOAD_WAIT_LOOPS; i++) {
 809                 if (completion_done(&vmbus_connection.unload_event))
 810                         goto completed;
 811
 812                 for_each_online_cpu(cpu) {
 813                         struct hv_per_cpu_context *hv_cpu
 814                                 = per_cpu_ptr(hv_context.cpu_context, cpu);
 815
 816                         page_addr = hv_cpu->synic_message_page;
 817                         msg = (struct hv_message *)page_addr
 818                                 + VMBUS_MESSAGE_SINT;
 819
 820                         message_type = READ_ONCE(msg->header.message_type);
 821                         if (message_type == HVMSG_NONE)
 822                                 continue;
 823
 824                         hdr = (struct vmbus_channel_message_header *)
 825                                 msg->u.payload;
 826
 827                         if (hdr->msgtype == CHANNELMSG_UNLOAD_RESPONSE)
 828                                 complete(&vmbus_connection.unload_event);
 829
 830                         vmbus_signal_eom(msg, message_type);
 831                 }
 832
 833                 /*
 834                  * Give a notice periodically so someone watching the
 835                  * serial output won't think it is completely hung.
 836                  */
 837                 if (!(i % UNLOAD_MSG_LOOPS))
 838                         pr_notice("Waiting for VMBus UNLOAD to complete\n");
 839
 840                 mdelay(UNLOAD_DELAY_UNIT_MS);
 841         }
 842         pr_err("Continuing even though VMBus UNLOAD did not complete\n");
 843
 844 completed:
 845         /*
 846          * We're crashing and already got the UNLOAD_RESPONSE, cleanup all
 847          * maybe-pending messages on all CPUs to be able to receive new
 848          * messages after we reconnect.
 849          */
 850         for_each_online_cpu(cpu) {
 851                 struct hv_per_cpu_context *hv_cpu
 852                         = per_cpu_ptr(hv_context.cpu_context, cpu);
 853
 854                 page_addr = hv_cpu->synic_message_page;
 855                 msg = (struct hv_message *)page_addr + VMBUS_MESSAGE_SINT;
 856                 msg->header.message_type = HVMSG_NONE;
 857         }
 858 }
 859
 860 /*
 861  * vmbus_unload_response - Handler for the unload response.
 862  */
 863 static void vmbus_unload_response(struct vmbus_channel_message_header *hdr)
 864 {
 865         /*
 866          * This is a global event; just wakeup the waiting thread.
 867          * Once we successfully unload, we can cleanup the monitor state.
 868          */
 869         complete(&vmbus_connection.unload_event);
 870 }
 871
 872 void vmbus_initiate_unload(bool crash)
 873 {
 874         struct vmbus_channel_message_header hdr;
 875
 876         if (xchg(&vmbus_connection.conn_state, DISCONNECTED) == DISCONNECTED)
 877                 return;
 878
 879         /* Pre-Win2012R2 hosts don't support reconnect */
 880         if (vmbus_proto_version < VERSION_WIN8_1)
 881                 return;
 882
 883         init_completion(&vmbus_connection.unload_event);
 884         memset(&hdr, 0, sizeof(struct vmbus_channel_message_header));
 885         hdr.msgtype = CHANNELMSG_UNLOAD;
 886         vmbus_post_msg(&hdr, sizeof(struct vmbus_channel_message_header),
 887                        !crash);
 888
 889         /*
 890          * vmbus_initiate_unload() is also called on crash and the crash can be
 891          * happening in an interrupt context, where scheduling is impossible.
 892          */
 893         if (!crash)
 894                 wait_for_completion(&vmbus_connection.unload_event);
 895         else
 896                 vmbus_wait_for_unload();
 897 }
 898
 899 /*
 900  * vmbus_onoffer - Handler for channel offers from vmbus in parent partition.
 901  *
 902  */
 903 static void vmbus_onoffer(struct vmbus_channel_message_header *hdr)
 904 {
 905         struct vmbus_channel_offer_channel *offer;
 906         struct vmbus_channel *newchannel;
 907
 908         offer = (struct vmbus_channel_offer_channel *)hdr;
 909
 910         trace_vmbus_onoffer(offer);
 911
 912         /* Allocate the channel object and save this offer. */
 913         newchannel = alloc_channel();
 914         if (!newchannel) {
 915                 vmbus_release_relid(offer->child_relid);
 916                 atomic_dec(&vmbus_connection.offer_in_progress);
 917                 pr_err("Unable to allocate channel object\n");
 918                 return;
 919         }
 920
 921         /*
 922          * Setup state for signalling the host.
 923          */
 924         newchannel->sig_event = VMBUS_EVENT_CONNECTION_ID;
 925
 926         if (vmbus_proto_version != VERSION_WS2008) {
 927                 newchannel->is_dedicated_interrupt =
 928                                 (offer->is_dedicated_interrupt != 0);
 929                 newchannel->sig_event = offer->connection_id;
 930         }
 931
 932         memcpy(&newchannel->offermsg, offer,
 933                sizeof(struct vmbus_channel_offer_channel));
 934         newchannel->monitor_grp = (u8)offer->monitorid / 32;
 935         newchannel->monitor_bit = (u8)offer->monitorid % 32;
 936
 937         vmbus_process_offer(newchannel);
 938 }
 939
 940 /*
 941  * vmbus_onoffer_rescind - Rescind offer handler.
 942  *
 943  * We queue a work item to process this offer synchronously
 944  */
 945 static void vmbus_onoffer_rescind(struct vmbus_channel_message_header *hdr)
 946 {
 947         struct vmbus_channel_rescind_offer *rescind;
 948         struct vmbus_channel *channel;
 949         struct device *dev;
 950
 951         rescind = (struct vmbus_channel_rescind_offer *)hdr;
 952
 953         trace_vmbus_onoffer_rescind(rescind);
 954
 955         /*
 956          * The offer msg and the corresponding rescind msg
 957          * from the host are guranteed to be ordered -
 958          * offer comes in first and then the rescind.
 959          * Since we process these events in work elements,
 960          * and with preemption, we may end up processing
 961          * the events out of order. Given that we handle these
 962          * work elements on the same CPU, this is possible only
 963          * in the case of preemption. In any case wait here
 964          * until the offer processing has moved beyond the
 965          * point where the channel is discoverable.
 966          */
 967
 968         while (atomic_read(&vmbus_connection.offer_in_progress) != 0) {
 969                 /*
 970                  * We wait here until any channel offer is currently
 971                  * being processed.
 972                  */
 973                 msleep(1);
 974         }
 975
 976         mutex_lock(&vmbus_connection.channel_mutex);
 977         channel = relid2channel(rescind->child_relid);
 978         mutex_unlock(&vmbus_connection.channel_mutex);
 979
 980         if (channel == NULL) {
 981                 /*
 982                  * We failed in processing the offer message;
 983                  * we would have cleaned up the relid in that
 984                  * failure path.
 985                  */
 986                 return;
 987         }
 988
 989         /*
 990          * Before setting channel->rescind in vmbus_rescind_cleanup(), we
 991          * should make sure the channel callback is not running any more.
 992          */
 993         vmbus_reset_channel_cb(channel);
 994
 995         /*
 996          * Now wait for offer handling to complete.
 997          */
 998         vmbus_rescind_cleanup(channel);
 999         while (READ_ONCE(channel->probe_done) == false) {
1000                 /*
1001                  * We wait here until any channel offer is currently
1002                  * being processed.
1003                  */
1004                 msleep(1);
1005         }
1006
1007         /*
1008          * At this point, the rescind handling can proceed safely.
1009          */
1010
1011         if (channel->device_obj) {
1012                 if (channel->chn_rescind_callback) {
1013                         channel->chn_rescind_callback(channel);
1014                         return;
1015                 }
1016                 /*
1017                  * We will have to unregister this device from the
1018                  * driver core.
1019                  */
1020                 dev = get_device(&channel->device_obj->device);
1021                 if (dev) {
1022                         vmbus_device_unregister(channel->device_obj);
1023                         put_device(dev);
1024                 }
1025         } else if (channel->primary_channel != NULL) {
1026                 /*
1027                  * Sub-channel is being rescinded. Following is the channel
1028                  * close sequence when initiated from the driveri (refer to
1029                  * vmbus_close() for details):
1030                  * 1. Close all sub-channels first
1031                  * 2. Then close the primary channel.
1032                  */
1033                 mutex_lock(&vmbus_connection.channel_mutex);
1034                 if (channel->state == CHANNEL_OPEN_STATE) {
1035                         /*
1036                          * The channel is currently not open;
1037                          * it is safe for us to cleanup the channel.
1038                          */
1039                         hv_process_channel_removal(rescind->child_relid);
1040                 } else {
1041                         complete(&channel->rescind_event);
1042                 }
1043                 mutex_unlock(&vmbus_connection.channel_mutex);
1044         }
1045 }
1046
1047 void vmbus_hvsock_device_unregister(struct vmbus_channel *channel)
1048 {
1049         BUG_ON(!is_hvsock_channel(channel));
1050
1051         /* We always get a rescind msg when a connection is closed. */
1052         while (!READ_ONCE(channel->probe_done) || !READ_ONCE(channel->rescind))
1053                 msleep(1);
1054
1055         vmbus_device_unregister(channel->device_obj);
1056 }
1057 EXPORT_SYMBOL_GPL(vmbus_hvsock_device_unregister);
1058
1059
1060 /*
1061  * vmbus_onoffers_delivered -
1062  * This is invoked when all offers have been delivered.
1063  *
1064  * Nothing to do here.
1065  */
1066 static void vmbus_onoffers_delivered(
1067                         struct vmbus_channel_message_header *hdr)
1068 {
1069 }
1070
1071 /*
1072  * vmbus_onopen_result - Open result handler.
1073  *
1074  * This is invoked when we received a response to our channel open request.
1075  * Find the matching request, copy the response and signal the requesting
1076  * thread.
1077  */
1078 static void vmbus_onopen_result(struct vmbus_channel_message_header *hdr)
1079 {
1080         struct vmbus_channel_open_result *result;
1081         struct vmbus_channel_msginfo *msginfo;
1082         struct vmbus_channel_message_header *requestheader;
1083         struct vmbus_channel_open_channel *openmsg;
1084         unsigned long flags;
1085
1086         result = (struct vmbus_channel_open_result *)hdr;
1087
1088         trace_vmbus_onopen_result(result);
1089
1090         /*
1091          * Find the open msg, copy the result and signal/unblock the wait event
1092          */
1093         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1094
1095         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1096                                 msglistentry) {
1097                 requestheader =
1098                         (struct vmbus_channel_message_header *)msginfo->msg;
1099
1100                 if (requestheader->msgtype == CHANNELMSG_OPENCHANNEL) {
1101                         openmsg =
1102                         (struct vmbus_channel_open_channel *)msginfo->msg;
1103                         if (openmsg->child_relid == result->child_relid &&
1104                             openmsg->openid == result->openid) {
1105                                 memcpy(&msginfo->response.open_result,
1106                                        result,
1107                                        sizeof(
1108                                         struct vmbus_channel_open_result));
1109                                 complete(&msginfo->waitevent);
1110                                 break;
1111                         }
1112                 }
1113         }
1114         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1115 }
1116
1117 /*
1118  * vmbus_ongpadl_created - GPADL created handler.
1119  *
1120  * This is invoked when we received a response to our gpadl create request.
1121  * Find the matching request, copy the response and signal the requesting
1122  * thread.
1123  */
1124 static void vmbus_ongpadl_created(struct vmbus_channel_message_header *hdr)
1125 {
1126         struct vmbus_channel_gpadl_created *gpadlcreated;
1127         struct vmbus_channel_msginfo *msginfo;
1128         struct vmbus_channel_message_header *requestheader;
1129         struct vmbus_channel_gpadl_header *gpadlheader;
1130         unsigned long flags;
1131
1132         gpadlcreated = (struct vmbus_channel_gpadl_created *)hdr;
1133
1134         trace_vmbus_ongpadl_created(gpadlcreated);
1135
1136         /*
1137          * Find the establish msg, copy the result and signal/unblock the wait
1138          * event
1139          */
1140         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1141
1142         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1143                                 msglistentry) {
1144                 requestheader =
1145                         (struct vmbus_channel_message_header *)msginfo->msg;
1146
1147                 if (requestheader->msgtype == CHANNELMSG_GPADL_HEADER) {
1148                         gpadlheader =
1149                         (struct vmbus_channel_gpadl_header *)requestheader;
1150
1151                         if ((gpadlcreated->child_relid ==
1152                              gpadlheader->child_relid) &&
1153                             (gpadlcreated->gpadl == gpadlheader->gpadl)) {
1154                                 memcpy(&msginfo->response.gpadl_created,
1155                                        gpadlcreated,
1156                                        sizeof(
1157                                         struct vmbus_channel_gpadl_created));
1158                                 complete(&msginfo->waitevent);
1159                                 break;
1160                         }
1161                 }
1162         }
1163         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1164 }
1165
1166 /*
1167  * vmbus_ongpadl_torndown - GPADL torndown handler.
1168  *
1169  * This is invoked when we received a response to our gpadl teardown request.
1170  * Find the matching request, copy the response and signal the requesting
1171  * thread.
1172  */
1173 static void vmbus_ongpadl_torndown(
1174                         struct vmbus_channel_message_header *hdr)
1175 {
1176         struct vmbus_channel_gpadl_torndown *gpadl_torndown;
1177         struct vmbus_channel_msginfo *msginfo;
1178         struct vmbus_channel_message_header *requestheader;
1179         struct vmbus_channel_gpadl_teardown *gpadl_teardown;
1180         unsigned long flags;
1181
1182         gpadl_torndown = (struct vmbus_channel_gpadl_torndown *)hdr;
1183
1184         trace_vmbus_ongpadl_torndown(gpadl_torndown);
1185
1186         /*
1187          * Find the open msg, copy the result and signal/unblock the wait event
1188          */
1189         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1190
1191         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1192                                 msglistentry) {
1193                 requestheader =
1194                         (struct vmbus_channel_message_header *)msginfo->msg;
1195
1196                 if (requestheader->msgtype == CHANNELMSG_GPADL_TEARDOWN) {
1197                         gpadl_teardown =
1198                         (struct vmbus_channel_gpadl_teardown *)requestheader;
1199
1200                         if (gpadl_torndown->gpadl == gpadl_teardown->gpadl) {
1201                                 memcpy(&msginfo->response.gpadl_torndown,
1202                                        gpadl_torndown,
1203                                        sizeof(
1204                                         struct vmbus_channel_gpadl_torndown));
1205                                 complete(&msginfo->waitevent);
1206                                 break;
1207                         }
1208                 }
1209         }
1210         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1211 }
1212
1213 /*
1214  * vmbus_onversion_response - Version response handler
1215  *
1216  * This is invoked when we received a response to our initiate contact request.
1217  * Find the matching request, copy the response and signal the requesting
1218  * thread.
1219  */
1220 static void vmbus_onversion_response(
1221                 struct vmbus_channel_message_header *hdr)
1222 {
1223         struct vmbus_channel_msginfo *msginfo;
1224         struct vmbus_channel_message_header *requestheader;
1225         struct vmbus_channel_version_response *version_response;
1226         unsigned long flags;
1227
1228         version_response = (struct vmbus_channel_version_response *)hdr;
1229
1230         trace_vmbus_onversion_response(version_response);
1231
1232         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1233
1234         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1235                                 msglistentry) {
1236                 requestheader =
1237                         (struct vmbus_channel_message_header *)msginfo->msg;
1238
1239                 if (requestheader->msgtype ==
1240                     CHANNELMSG_INITIATE_CONTACT) {
1241                         memcpy(&msginfo->response.version_response,
1242                               version_response,
1243                               sizeof(struct vmbus_channel_version_response));
1244                         complete(&msginfo->waitevent);
1245                 }
1246         }
1247         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1248 }
1249
1250 /* Channel message dispatch table */
1251 const struct vmbus_channel_message_table_entry
1252 channel_message_table[CHANNELMSG_COUNT] = {
1253         { CHANNELMSG_INVALID,                   0, NULL },
1254         { CHANNELMSG_OFFERCHANNEL,              0, vmbus_onoffer },
1255         { CHANNELMSG_RESCIND_CHANNELOFFER,      0, vmbus_onoffer_rescind },
1256         { CHANNELMSG_REQUESTOFFERS,             0, NULL },
1257         { CHANNELMSG_ALLOFFERS_DELIVERED,       1, vmbus_onoffers_delivered },
1258         { CHANNELMSG_OPENCHANNEL,               0, NULL },
1259         { CHANNELMSG_OPENCHANNEL_RESULT,        1, vmbus_onopen_result },
1260         { CHANNELMSG_CLOSECHANNEL,              0, NULL },
1261         { CHANNELMSG_GPADL_HEADER,              0, NULL },
1262         { CHANNELMSG_GPADL_BODY,                0, NULL },
1263         { CHANNELMSG_GPADL_CREATED,             1, vmbus_ongpadl_created },
1264         { CHANNELMSG_GPADL_TEARDOWN,            0, NULL },
1265         { CHANNELMSG_GPADL_TORNDOWN,            1, vmbus_ongpadl_torndown },
1266         { CHANNELMSG_RELID_RELEASED,            0, NULL },
1267         { CHANNELMSG_INITIATE_CONTACT,          0, NULL },
1268         { CHANNELMSG_VERSION_RESPONSE,          1, vmbus_onversion_response },
1269         { CHANNELMSG_UNLOAD,                    0, NULL },
1270         { CHANNELMSG_UNLOAD_RESPONSE,           1, vmbus_unload_response },
1271         { CHANNELMSG_18,                        0, NULL },
1272         { CHANNELMSG_19,                        0, NULL },
1273         { CHANNELMSG_20,                        0, NULL },
1274         { CHANNELMSG_TL_CONNECT_REQUEST,        0, NULL },
1275         { CHANNELMSG_22,                        0, NULL },
1276         { CHANNELMSG_TL_CONNECT_RESULT,         0, NULL },
1277 };
1278
1279 /*
1280  * vmbus_onmessage - Handler for channel protocol messages.
1281  *
1282  * This is invoked in the vmbus worker thread context.
1283  */
1284 void vmbus_onmessage(void *context)
1285 {
1286         struct hv_message *msg = context;
1287         struct vmbus_channel_message_header *hdr;
1288
1289         hdr = (struct vmbus_channel_message_header *)msg->u.payload;
1290
1291         trace_vmbus_on_message(hdr);
1292
1293         /*
1294          * vmbus_on_msg_dpc() makes sure the hdr->msgtype here can not go
1295          * out of bound and the message_handler pointer can not be NULL.
1296          */
1297         channel_message_table[hdr->msgtype].message_handler(hdr);
1298 }
1299
1300 /*
1301  * vmbus_request_offers - Send a request to get all our pending offers.
1302  */
1303 int vmbus_request_offers(void)
1304 {
1305         struct vmbus_channel_message_header *msg;
1306         struct vmbus_channel_msginfo *msginfo;
1307         int ret;
1308
1309         msginfo = kmalloc(sizeof(*msginfo) +
1310                           sizeof(struct vmbus_channel_message_header),
1311                           GFP_KERNEL);
1312         if (!msginfo)
1313                 return -ENOMEM;
1314
1315         msg = (struct vmbus_channel_message_header *)msginfo->msg;
1316
1317         msg->msgtype = CHANNELMSG_REQUESTOFFERS;
1318
1319         ret = vmbus_post_msg(msg, sizeof(struct vmbus_channel_message_header),
1320                              true);
1321
1322         trace_vmbus_request_offers(ret);
1323
1324         if (ret != 0) {
1325                 pr_err("Unable to request offers - %d\n", ret);
1326
1327                 goto cleanup;
1328         }
1329
1330 cleanup:
1331         kfree(msginfo);
1332
1333         return ret;
1334 }
1335
1336 /*
1337  * Retrieve the (sub) channel on which to send an outgoing request.
1338  * When a primary channel has multiple sub-channels, we try to
1339  * distribute the load equally amongst all available channels.
1340  */
1341 struct vmbus_channel *vmbus_get_outgoing_channel(struct vmbus_channel *primary)
1342 {
1343         struct list_head *cur, *tmp;
1344         int cur_cpu;
1345         struct vmbus_channel *cur_channel;
1346         struct vmbus_channel *outgoing_channel = primary;
1347         int next_channel;
1348         int i = 1;
1349
1350         if (list_empty(&primary->sc_list))
1351                 return outgoing_channel;
1352
1353         next_channel = primary->next_oc++;
1354
1355         if (next_channel > (primary->num_sc)) {
1356                 primary->next_oc = 0;
1357                 return outgoing_channel;
1358         }
1359
1360         cur_cpu = hv_cpu_number_to_vp_number(smp_processor_id());
1361         list_for_each_safe(cur, tmp, &primary->sc_list) {
1362                 cur_channel = list_entry(cur, struct vmbus_channel, sc_list);
1363                 if (cur_channel->state != CHANNEL_OPENED_STATE)
1364                         continue;
1365
1366                 if (cur_channel->target_vp == cur_cpu)
1367                         return cur_channel;
1368
1369                 if (i == next_channel)
1370                         return cur_channel;
1371
1372                 i++;
1373         }
1374
1375         return outgoing_channel;
1376 }
1377 EXPORT_SYMBOL_GPL(vmbus_get_outgoing_channel);
1378
1379 static void invoke_sc_cb(struct vmbus_channel *primary_channel)
1380 {
1381         struct list_head *cur, *tmp;
1382         struct vmbus_channel *cur_channel;
1383
1384         if (primary_channel->sc_creation_callback == NULL)
1385                 return;
1386
1387         list_for_each_safe(cur, tmp, &primary_channel->sc_list) {
1388                 cur_channel = list_entry(cur, struct vmbus_channel, sc_list);
1389
1390                 primary_channel->sc_creation_callback(cur_channel);
1391         }
1392 }
1393
1394 void vmbus_set_sc_create_callback(struct vmbus_channel *primary_channel,
1395                                 void (*sc_cr_cb)(struct vmbus_channel *new_sc))
1396 {
1397         primary_channel->sc_creation_callback = sc_cr_cb;
1398 }
1399 EXPORT_SYMBOL_GPL(vmbus_set_sc_create_callback);
1400
1401 bool vmbus_are_subchannels_present(struct vmbus_channel *primary)
1402 {
1403         bool ret;
1404
1405         ret = !list_empty(&primary->sc_list);
1406
1407         if (ret) {
1408                 /*
1409                  * Invoke the callback on sub-channel creation.
1410                  * This will present a uniform interface to the
1411                  * clients.
1412                  */
1413                 invoke_sc_cb(primary);
1414         }
1415
1416         return ret;
1417 }
1418 EXPORT_SYMBOL_GPL(vmbus_are_subchannels_present);
1419
1420 void vmbus_set_chn_rescind_callback(struct vmbus_channel *channel,
1421                 void (*chn_rescind_cb)(struct vmbus_channel *))
1422 {
1423         channel->chn_rescind_callback = chn_rescind_cb;
1424 }
1425 EXPORT_SYMBOL_GPL(vmbus_set_chn_rescind_callback);