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