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 <[email protected]>
  19  *   Hank Janssen  <[email protected]>
  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 = true,
  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
 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
 337         INIT_LIST_HEAD(&channel->sc_list);
 338         INIT_LIST_HEAD(&channel->percpu_list);
 339
 340         tasklet_init(&channel->callback_event,
 341                      vmbus_on_event, (unsigned long)channel);
 342
 343         return channel;
 344 }
 345
 346 /*
 347  * free_channel - Release the resources used by the vmbus channel object
 348  */
 349 static void free_channel(struct vmbus_channel *channel)
 350 {
 351         tasklet_kill(&channel->callback_event);
 352
 353         kfree_rcu(channel, rcu);
 354 }
 355
 356 static void percpu_channel_enq(void *arg)
 357 {
 358         struct vmbus_channel *channel = arg;
 359         struct hv_per_cpu_context *hv_cpu
 360                 = this_cpu_ptr(hv_context.cpu_context);
 361
 362         list_add_tail_rcu(&channel->percpu_list, &hv_cpu->chan_list);
 363 }
 364
 365 static void percpu_channel_deq(void *arg)
 366 {
 367         struct vmbus_channel *channel = arg;
 368
 369         list_del_rcu(&channel->percpu_list);
 370 }
 371
 372
 373 static void vmbus_release_relid(u32 relid)
 374 {
 375         struct vmbus_channel_relid_released msg;
 376
 377         memset(&msg, 0, sizeof(struct vmbus_channel_relid_released));
 378         msg.child_relid = relid;
 379         msg.header.msgtype = CHANNELMSG_RELID_RELEASED;
 380         vmbus_post_msg(&msg, sizeof(struct vmbus_channel_relid_released),
 381                        true);
 382 }
 383
 384 void hv_process_channel_removal(struct vmbus_channel *channel, u32 relid)
 385 {
 386         unsigned long flags;
 387         struct vmbus_channel *primary_channel;
 388
 389         BUG_ON(!channel->rescind);
 390         BUG_ON(!mutex_is_locked(&vmbus_connection.channel_mutex));
 391
 392         if (channel->target_cpu != get_cpu()) {
 393                 put_cpu();
 394                 smp_call_function_single(channel->target_cpu,
 395                                          percpu_channel_deq, channel, true);
 396         } else {
 397                 percpu_channel_deq(channel);
 398                 put_cpu();
 399         }
 400
 401         if (channel->primary_channel == NULL) {
 402                 list_del(&channel->listentry);
 403
 404                 primary_channel = channel;
 405         } else {
 406                 primary_channel = channel->primary_channel;
 407                 spin_lock_irqsave(&primary_channel->lock, flags);
 408                 list_del(&channel->sc_list);
 409                 primary_channel->num_sc--;
 410                 spin_unlock_irqrestore(&primary_channel->lock, flags);
 411         }
 412
 413         /*
 414          * We need to free the bit for init_vp_index() to work in the case
 415          * of sub-channel, when we reload drivers like hv_netvsc.
 416          */
 417         if (channel->affinity_policy == HV_LOCALIZED)
 418                 cpumask_clear_cpu(channel->target_cpu,
 419                                   &primary_channel->alloced_cpus_in_node);
 420
 421         vmbus_release_relid(relid);
 422
 423         free_channel(channel);
 424 }
 425
 426 void vmbus_free_channels(void)
 427 {
 428         struct vmbus_channel *channel, *tmp;
 429
 430         list_for_each_entry_safe(channel, tmp, &vmbus_connection.chn_list,
 431                 listentry) {
 432                 /* hv_process_channel_removal() needs this */
 433                 channel->rescind = true;
 434
 435                 vmbus_device_unregister(channel->device_obj);
 436         }
 437 }
 438
 439 /*
 440  * vmbus_process_offer - Process the offer by creating a channel/device
 441  * associated with this offer
 442  */
 443 static void vmbus_process_offer(struct vmbus_channel *newchannel)
 444 {
 445         struct vmbus_channel *channel;
 446         bool fnew = true;
 447         unsigned long flags;
 448         u16 dev_type;
 449         int ret;
 450
 451         /* Make sure this is a new offer */
 452         mutex_lock(&vmbus_connection.channel_mutex);
 453
 454         list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
 455                 if (!uuid_le_cmp(channel->offermsg.offer.if_type,
 456                         newchannel->offermsg.offer.if_type) &&
 457                         !uuid_le_cmp(channel->offermsg.offer.if_instance,
 458                                 newchannel->offermsg.offer.if_instance)) {
 459                         fnew = false;
 460                         break;
 461                 }
 462         }
 463
 464         if (fnew)
 465                 list_add_tail(&newchannel->listentry,
 466                               &vmbus_connection.chn_list);
 467
 468         mutex_unlock(&vmbus_connection.channel_mutex);
 469
 470         if (!fnew) {
 471                 /*
 472                  * Check to see if this is a sub-channel.
 473                  */
 474                 if (newchannel->offermsg.offer.sub_channel_index != 0) {
 475                         /*
 476                          * Process the sub-channel.
 477                          */
 478                         newchannel->primary_channel = channel;
 479                         spin_lock_irqsave(&channel->lock, flags);
 480                         list_add_tail(&newchannel->sc_list, &channel->sc_list);
 481                         channel->num_sc++;
 482                         spin_unlock_irqrestore(&channel->lock, flags);
 483                 } else {
 484                         atomic_dec(&vmbus_connection.offer_in_progress);
 485                         goto err_free_chan;
 486                 }
 487         }
 488
 489         dev_type = hv_get_dev_type(newchannel);
 490
 491         init_vp_index(newchannel, dev_type);
 492
 493         if (newchannel->target_cpu != get_cpu()) {
 494                 put_cpu();
 495                 smp_call_function_single(newchannel->target_cpu,
 496                                          percpu_channel_enq,
 497                                          newchannel, true);
 498         } else {
 499                 percpu_channel_enq(newchannel);
 500                 put_cpu();
 501         }
 502
 503         /*
 504          * This state is used to indicate a successful open
 505          * so that when we do close the channel normally, we
 506          * can cleanup properly
 507          */
 508         newchannel->state = CHANNEL_OPEN_STATE;
 509
 510         if (!fnew) {
 511                 if (channel->sc_creation_callback != NULL)
 512                         channel->sc_creation_callback(newchannel);
 513                 atomic_dec(&vmbus_connection.offer_in_progress);
 514                 return;
 515         }
 516
 517         /*
 518          * Start the process of binding this offer to the driver
 519          * We need to set the DeviceObject field before calling
 520          * vmbus_child_dev_add()
 521          */
 522         newchannel->device_obj = vmbus_device_create(
 523                 &newchannel->offermsg.offer.if_type,
 524                 &newchannel->offermsg.offer.if_instance,
 525                 newchannel);
 526         if (!newchannel->device_obj)
 527                 goto err_deq_chan;
 528
 529         newchannel->device_obj->device_id = dev_type;
 530         /*
 531          * Add the new device to the bus. This will kick off device-driver
 532          * binding which eventually invokes the device driver's AddDevice()
 533          * method.
 534          */
 535         ret = vmbus_device_register(newchannel->device_obj);
 536
 537         if (ret != 0) {
 538                 pr_err("unable to add child device object (relid %d)\n",
 539                         newchannel->offermsg.child_relid);
 540                 kfree(newchannel->device_obj);
 541                 goto err_deq_chan;
 542         }
 543
 544         atomic_dec(&vmbus_connection.offer_in_progress);
 545         return;
 546
 547 err_deq_chan:
 548         mutex_lock(&vmbus_connection.channel_mutex);
 549         list_del(&newchannel->listentry);
 550         mutex_unlock(&vmbus_connection.channel_mutex);
 551
 552         if (newchannel->target_cpu != get_cpu()) {
 553                 put_cpu();
 554                 smp_call_function_single(newchannel->target_cpu,
 555                                          percpu_channel_deq, newchannel, true);
 556         } else {
 557                 percpu_channel_deq(newchannel);
 558                 put_cpu();
 559         }
 560
 561         vmbus_release_relid(newchannel->offermsg.child_relid);
 562
 563 err_free_chan:
 564         free_channel(newchannel);
 565 }
 566
 567 /*
 568  * We use this state to statically distribute the channel interrupt load.
 569  */
 570 static int next_numa_node_id;
 571
 572 /*
 573  * Starting with Win8, we can statically distribute the incoming
 574  * channel interrupt load by binding a channel to VCPU.
 575  * We do this in a hierarchical fashion:
 576  * First distribute the primary channels across available NUMA nodes
 577  * and then distribute the subchannels amongst the CPUs in the NUMA
 578  * node assigned to the primary channel.
 579  *
 580  * For pre-win8 hosts or non-performance critical channels we assign the
 581  * first CPU in the first NUMA node.
 582  */
 583 static void init_vp_index(struct vmbus_channel *channel, u16 dev_type)
 584 {
 585         u32 cur_cpu;
 586         bool perf_chn = vmbus_devs[dev_type].perf_device;
 587         struct vmbus_channel *primary = channel->primary_channel;
 588         int next_node;
 589         struct cpumask available_mask;
 590         struct cpumask *alloced_mask;
 591
 592         if ((vmbus_proto_version == VERSION_WS2008) ||
 593             (vmbus_proto_version == VERSION_WIN7) || (!perf_chn)) {
 594                 /*
 595                  * Prior to win8, all channel interrupts are
 596                  * delivered on cpu 0.
 597                  * Also if the channel is not a performance critical
 598                  * channel, bind it to cpu 0.
 599                  */
 600                 channel->numa_node = 0;
 601                 channel->target_cpu = 0;
 602                 channel->target_vp = hv_cpu_number_to_vp_number(0);
 603                 return;
 604         }
 605
 606         /*
 607          * Based on the channel affinity policy, we will assign the NUMA
 608          * nodes.
 609          */
 610
 611         if ((channel->affinity_policy == HV_BALANCED) || (!primary)) {
 612                 while (true) {
 613                         next_node = next_numa_node_id++;
 614                         if (next_node == nr_node_ids) {
 615                                 next_node = next_numa_node_id = 0;
 616                                 continue;
 617                         }
 618                         if (cpumask_empty(cpumask_of_node(next_node)))
 619                                 continue;
 620                         break;
 621                 }
 622                 channel->numa_node = next_node;
 623                 primary = channel;
 624         }
 625         alloced_mask = &hv_context.hv_numa_map[primary->numa_node];
 626
 627         if (cpumask_weight(alloced_mask) ==
 628             cpumask_weight(cpumask_of_node(primary->numa_node))) {
 629                 /*
 630                  * We have cycled through all the CPUs in the node;
 631                  * reset the alloced map.
 632                  */
 633                 cpumask_clear(alloced_mask);
 634         }
 635
 636         cpumask_xor(&available_mask, alloced_mask,
 637                     cpumask_of_node(primary->numa_node));
 638
 639         cur_cpu = -1;
 640
 641         if (primary->affinity_policy == HV_LOCALIZED) {
 642                 /*
 643                  * Normally Hyper-V host doesn't create more subchannels
 644                  * than there are VCPUs on the node but it is possible when not
 645                  * all present VCPUs on the node are initialized by guest.
 646                  * Clear the alloced_cpus_in_node to start over.
 647                  */
 648                 if (cpumask_equal(&primary->alloced_cpus_in_node,
 649                                   cpumask_of_node(primary->numa_node)))
 650                         cpumask_clear(&primary->alloced_cpus_in_node);
 651         }
 652
 653         while (true) {
 654                 cur_cpu = cpumask_next(cur_cpu, &available_mask);
 655                 if (cur_cpu >= nr_cpu_ids) {
 656                         cur_cpu = -1;
 657                         cpumask_copy(&available_mask,
 658                                      cpumask_of_node(primary->numa_node));
 659                         continue;
 660                 }
 661
 662                 if (primary->affinity_policy == HV_LOCALIZED) {
 663                         /*
 664                          * NOTE: in the case of sub-channel, we clear the
 665                          * sub-channel related bit(s) in
 666                          * primary->alloced_cpus_in_node in
 667                          * hv_process_channel_removal(), so when we
 668                          * reload drivers like hv_netvsc in SMP guest, here
 669                          * we're able to re-allocate
 670                          * bit from primary->alloced_cpus_in_node.
 671                          */
 672                         if (!cpumask_test_cpu(cur_cpu,
 673                                               &primary->alloced_cpus_in_node)) {
 674                                 cpumask_set_cpu(cur_cpu,
 675                                                 &primary->alloced_cpus_in_node);
 676                                 cpumask_set_cpu(cur_cpu, alloced_mask);
 677                                 break;
 678                         }
 679                 } else {
 680                         cpumask_set_cpu(cur_cpu, alloced_mask);
 681                         break;
 682                 }
 683         }
 684
 685         channel->target_cpu = cur_cpu;
 686         channel->target_vp = hv_cpu_number_to_vp_number(cur_cpu);
 687 }
 688
 689 static void vmbus_wait_for_unload(void)
 690 {
 691         int cpu;
 692         void *page_addr;
 693         struct hv_message *msg;
 694         struct vmbus_channel_message_header *hdr;
 695         u32 message_type;
 696
 697         /*
 698          * CHANNELMSG_UNLOAD_RESPONSE is always delivered to the CPU which was
 699          * used for initial contact or to CPU0 depending on host version. When
 700          * we're crashing on a different CPU let's hope that IRQ handler on
 701          * the cpu which receives CHANNELMSG_UNLOAD_RESPONSE is still
 702          * functional and vmbus_unload_response() will complete
 703          * vmbus_connection.unload_event. If not, the last thing we can do is
 704          * read message pages for all CPUs directly.
 705          */
 706         while (1) {
 707                 if (completion_done(&vmbus_connection.unload_event))
 708                         break;
 709
 710                 for_each_online_cpu(cpu) {
 711                         struct hv_per_cpu_context *hv_cpu
 712                                 = per_cpu_ptr(hv_context.cpu_context, cpu);
 713
 714                         page_addr = hv_cpu->synic_message_page;
 715                         msg = (struct hv_message *)page_addr
 716                                 + VMBUS_MESSAGE_SINT;
 717
 718                         message_type = READ_ONCE(msg->header.message_type);
 719                         if (message_type == HVMSG_NONE)
 720                                 continue;
 721
 722                         hdr = (struct vmbus_channel_message_header *)
 723                                 msg->u.payload;
 724
 725                         if (hdr->msgtype == CHANNELMSG_UNLOAD_RESPONSE)
 726                                 complete(&vmbus_connection.unload_event);
 727
 728                         vmbus_signal_eom(msg, message_type);
 729                 }
 730
 731                 mdelay(10);
 732         }
 733
 734         /*
 735          * We're crashing and already got the UNLOAD_RESPONSE, cleanup all
 736          * maybe-pending messages on all CPUs to be able to receive new
 737          * messages after we reconnect.
 738          */
 739         for_each_online_cpu(cpu) {
 740                 struct hv_per_cpu_context *hv_cpu
 741                         = per_cpu_ptr(hv_context.cpu_context, cpu);
 742
 743                 page_addr = hv_cpu->synic_message_page;
 744                 msg = (struct hv_message *)page_addr + VMBUS_MESSAGE_SINT;
 745                 msg->header.message_type = HVMSG_NONE;
 746         }
 747 }
 748
 749 /*
 750  * vmbus_unload_response - Handler for the unload response.
 751  */
 752 static void vmbus_unload_response(struct vmbus_channel_message_header *hdr)
 753 {
 754         /*
 755          * This is a global event; just wakeup the waiting thread.
 756          * Once we successfully unload, we can cleanup the monitor state.
 757          */
 758         complete(&vmbus_connection.unload_event);
 759 }
 760
 761 void vmbus_initiate_unload(bool crash)
 762 {
 763         struct vmbus_channel_message_header hdr;
 764
 765         /* Pre-Win2012R2 hosts don't support reconnect */
 766         if (vmbus_proto_version < VERSION_WIN8_1)
 767                 return;
 768
 769         init_completion(&vmbus_connection.unload_event);
 770         memset(&hdr, 0, sizeof(struct vmbus_channel_message_header));
 771         hdr.msgtype = CHANNELMSG_UNLOAD;
 772         vmbus_post_msg(&hdr, sizeof(struct vmbus_channel_message_header),
 773                        !crash);
 774
 775         /*
 776          * vmbus_initiate_unload() is also called on crash and the crash can be
 777          * happening in an interrupt context, where scheduling is impossible.
 778          */
 779         if (!crash)
 780                 wait_for_completion(&vmbus_connection.unload_event);
 781         else
 782                 vmbus_wait_for_unload();
 783 }
 784
 785 /*
 786  * vmbus_onoffer - Handler for channel offers from vmbus in parent partition.
 787  *
 788  */
 789 static void vmbus_onoffer(struct vmbus_channel_message_header *hdr)
 790 {
 791         struct vmbus_channel_offer_channel *offer;
 792         struct vmbus_channel *newchannel;
 793
 794         offer = (struct vmbus_channel_offer_channel *)hdr;
 795
 796         /* Allocate the channel object and save this offer. */
 797         newchannel = alloc_channel();
 798         if (!newchannel) {
 799                 vmbus_release_relid(offer->child_relid);
 800                 atomic_dec(&vmbus_connection.offer_in_progress);
 801                 pr_err("Unable to allocate channel object\n");
 802                 return;
 803         }
 804
 805         /*
 806          * Setup state for signalling the host.
 807          */
 808         newchannel->sig_event = VMBUS_EVENT_CONNECTION_ID;
 809
 810         if (vmbus_proto_version != VERSION_WS2008) {
 811                 newchannel->is_dedicated_interrupt =
 812                                 (offer->is_dedicated_interrupt != 0);
 813                 newchannel->sig_event = offer->connection_id;
 814         }
 815
 816         memcpy(&newchannel->offermsg, offer,
 817                sizeof(struct vmbus_channel_offer_channel));
 818         newchannel->monitor_grp = (u8)offer->monitorid / 32;
 819         newchannel->monitor_bit = (u8)offer->monitorid % 32;
 820
 821         vmbus_process_offer(newchannel);
 822 }
 823
 824 /*
 825  * vmbus_onoffer_rescind - Rescind offer handler.
 826  *
 827  * We queue a work item to process this offer synchronously
 828  */
 829 static void vmbus_onoffer_rescind(struct vmbus_channel_message_header *hdr)
 830 {
 831         struct vmbus_channel_rescind_offer *rescind;
 832         struct vmbus_channel *channel;
 833         unsigned long flags;
 834         struct device *dev;
 835
 836         rescind = (struct vmbus_channel_rescind_offer *)hdr;
 837
 838         /*
 839          * The offer msg and the corresponding rescind msg
 840          * from the host are guranteed to be ordered -
 841          * offer comes in first and then the rescind.
 842          * Since we process these events in work elements,
 843          * and with preemption, we may end up processing
 844          * the events out of order. Given that we handle these
 845          * work elements on the same CPU, this is possible only
 846          * in the case of preemption. In any case wait here
 847          * until the offer processing has moved beyond the
 848          * point where the channel is discoverable.
 849          */
 850
 851         while (atomic_read(&vmbus_connection.offer_in_progress) != 0) {
 852                 /*
 853                  * We wait here until any channel offer is currently
 854                  * being processed.
 855                  */
 856                 msleep(1);
 857         }
 858
 859         mutex_lock(&vmbus_connection.channel_mutex);
 860         channel = relid2channel(rescind->child_relid);
 861         mutex_unlock(&vmbus_connection.channel_mutex);
 862
 863         if (channel == NULL) {
 864                 /*
 865                  * We failed in processing the offer message;
 866                  * we would have cleaned up the relid in that
 867                  * failure path.
 868                  */
 869                 return;
 870         }
 871
 872         spin_lock_irqsave(&channel->lock, flags);
 873         channel->rescind = true;
 874         spin_unlock_irqrestore(&channel->lock, flags);
 875
 876         vmbus_rescind_cleanup(channel);
 877
 878         if (channel->device_obj) {
 879                 if (channel->chn_rescind_callback) {
 880                         channel->chn_rescind_callback(channel);
 881                         return;
 882                 }
 883                 /*
 884                  * We will have to unregister this device from the
 885                  * driver core.
 886                  */
 887                 dev = get_device(&channel->device_obj->device);
 888                 if (dev) {
 889                         vmbus_device_unregister(channel->device_obj);
 890                         put_device(dev);
 891                 }
 892         }
 893         if (channel->primary_channel != NULL) {
 894                 /*
 895                  * Sub-channel is being rescinded. Following is the channel
 896                  * close sequence when initiated from the driveri (refer to
 897                  * vmbus_close() for details):
 898                  * 1. Close all sub-channels first
 899                  * 2. Then close the primary channel.
 900                  */
 901                 if (channel->state == CHANNEL_OPEN_STATE) {
 902                         /*
 903                          * The channel is currently not open;
 904                          * it is safe for us to cleanup the channel.
 905                          */
 906                         mutex_lock(&vmbus_connection.channel_mutex);
 907                         hv_process_channel_removal(channel,
 908                                                 channel->offermsg.child_relid);
 909                         mutex_unlock(&vmbus_connection.channel_mutex);
 910                 }
 911         }
 912 }
 913
 914 void vmbus_hvsock_device_unregister(struct vmbus_channel *channel)
 915 {
 916         mutex_lock(&vmbus_connection.channel_mutex);
 917
 918         BUG_ON(!is_hvsock_channel(channel));
 919
 920         channel->rescind = true;
 921         vmbus_device_unregister(channel->device_obj);
 922
 923         mutex_unlock(&vmbus_connection.channel_mutex);
 924 }
 925 EXPORT_SYMBOL_GPL(vmbus_hvsock_device_unregister);
 926
 927
 928 /*
 929  * vmbus_onoffers_delivered -
 930  * This is invoked when all offers have been delivered.
 931  *
 932  * Nothing to do here.
 933  */
 934 static void vmbus_onoffers_delivered(
 935                         struct vmbus_channel_message_header *hdr)
 936 {
 937 }
 938
 939 /*
 940  * vmbus_onopen_result - Open result handler.
 941  *
 942  * This is invoked when we received a response to our channel open request.
 943  * Find the matching request, copy the response and signal the requesting
 944  * thread.
 945  */
 946 static void vmbus_onopen_result(struct vmbus_channel_message_header *hdr)
 947 {
 948         struct vmbus_channel_open_result *result;
 949         struct vmbus_channel_msginfo *msginfo;
 950         struct vmbus_channel_message_header *requestheader;
 951         struct vmbus_channel_open_channel *openmsg;
 952         unsigned long flags;
 953
 954         result = (struct vmbus_channel_open_result *)hdr;
 955
 956         /*
 957          * Find the open msg, copy the result and signal/unblock the wait event
 958          */
 959         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
 960
 961         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
 962                                 msglistentry) {
 963                 requestheader =
 964                         (struct vmbus_channel_message_header *)msginfo->msg;
 965
 966                 if (requestheader->msgtype == CHANNELMSG_OPENCHANNEL) {
 967                         openmsg =
 968                         (struct vmbus_channel_open_channel *)msginfo->msg;
 969                         if (openmsg->child_relid == result->child_relid &&
 970                             openmsg->openid == result->openid) {
 971                                 memcpy(&msginfo->response.open_result,
 972                                        result,
 973                                        sizeof(
 974                                         struct vmbus_channel_open_result));
 975                                 complete(&msginfo->waitevent);
 976                                 break;
 977                         }
 978                 }
 979         }
 980         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
 981 }
 982
 983 /*
 984  * vmbus_ongpadl_created - GPADL created handler.
 985  *
 986  * This is invoked when we received a response to our gpadl create request.
 987  * Find the matching request, copy the response and signal the requesting
 988  * thread.
 989  */
 990 static void vmbus_ongpadl_created(struct vmbus_channel_message_header *hdr)
 991 {
 992         struct vmbus_channel_gpadl_created *gpadlcreated;
 993         struct vmbus_channel_msginfo *msginfo;
 994         struct vmbus_channel_message_header *requestheader;
 995         struct vmbus_channel_gpadl_header *gpadlheader;
 996         unsigned long flags;
 997
 998         gpadlcreated = (struct vmbus_channel_gpadl_created *)hdr;
 999
1000         /*
1001          * Find the establish msg, copy the result and signal/unblock the wait
1002          * event
1003          */
1004         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1005
1006         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1007                                 msglistentry) {
1008                 requestheader =
1009                         (struct vmbus_channel_message_header *)msginfo->msg;
1010
1011                 if (requestheader->msgtype == CHANNELMSG_GPADL_HEADER) {
1012                         gpadlheader =
1013                         (struct vmbus_channel_gpadl_header *)requestheader;
1014
1015                         if ((gpadlcreated->child_relid ==
1016                              gpadlheader->child_relid) &&
1017                             (gpadlcreated->gpadl == gpadlheader->gpadl)) {
1018                                 memcpy(&msginfo->response.gpadl_created,
1019                                        gpadlcreated,
1020                                        sizeof(
1021                                         struct vmbus_channel_gpadl_created));
1022                                 complete(&msginfo->waitevent);
1023                                 break;
1024                         }
1025                 }
1026         }
1027         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1028 }
1029
1030 /*
1031  * vmbus_ongpadl_torndown - GPADL torndown handler.
1032  *
1033  * This is invoked when we received a response to our gpadl teardown request.
1034  * Find the matching request, copy the response and signal the requesting
1035  * thread.
1036  */
1037 static void vmbus_ongpadl_torndown(
1038                         struct vmbus_channel_message_header *hdr)
1039 {
1040         struct vmbus_channel_gpadl_torndown *gpadl_torndown;
1041         struct vmbus_channel_msginfo *msginfo;
1042         struct vmbus_channel_message_header *requestheader;
1043         struct vmbus_channel_gpadl_teardown *gpadl_teardown;
1044         unsigned long flags;
1045
1046         gpadl_torndown = (struct vmbus_channel_gpadl_torndown *)hdr;
1047
1048         /*
1049          * Find the open msg, copy the result and signal/unblock the wait event
1050          */
1051         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1052
1053         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1054                                 msglistentry) {
1055                 requestheader =
1056                         (struct vmbus_channel_message_header *)msginfo->msg;
1057
1058                 if (requestheader->msgtype == CHANNELMSG_GPADL_TEARDOWN) {
1059                         gpadl_teardown =
1060                         (struct vmbus_channel_gpadl_teardown *)requestheader;
1061
1062                         if (gpadl_torndown->gpadl == gpadl_teardown->gpadl) {
1063                                 memcpy(&msginfo->response.gpadl_torndown,
1064                                        gpadl_torndown,
1065                                        sizeof(
1066                                         struct vmbus_channel_gpadl_torndown));
1067                                 complete(&msginfo->waitevent);
1068                                 break;
1069                         }
1070                 }
1071         }
1072         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1073 }
1074
1075 /*
1076  * vmbus_onversion_response - Version response handler
1077  *
1078  * This is invoked when we received a response to our initiate contact request.
1079  * Find the matching request, copy the response and signal the requesting
1080  * thread.
1081  */
1082 static void vmbus_onversion_response(
1083                 struct vmbus_channel_message_header *hdr)
1084 {
1085         struct vmbus_channel_msginfo *msginfo;
1086         struct vmbus_channel_message_header *requestheader;
1087         struct vmbus_channel_version_response *version_response;
1088         unsigned long flags;
1089
1090         version_response = (struct vmbus_channel_version_response *)hdr;
1091         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1092
1093         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1094                                 msglistentry) {
1095                 requestheader =
1096                         (struct vmbus_channel_message_header *)msginfo->msg;
1097
1098                 if (requestheader->msgtype ==
1099                     CHANNELMSG_INITIATE_CONTACT) {
1100                         memcpy(&msginfo->response.version_response,
1101                               version_response,
1102                               sizeof(struct vmbus_channel_version_response));
1103                         complete(&msginfo->waitevent);
1104                 }
1105         }
1106         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1107 }
1108
1109 /* Channel message dispatch table */
1110 const struct vmbus_channel_message_table_entry
1111 channel_message_table[CHANNELMSG_COUNT] = {
1112         { CHANNELMSG_INVALID,                   0, NULL },
1113         { CHANNELMSG_OFFERCHANNEL,              0, vmbus_onoffer },
1114         { CHANNELMSG_RESCIND_CHANNELOFFER,      0, vmbus_onoffer_rescind },
1115         { CHANNELMSG_REQUESTOFFERS,             0, NULL },
1116         { CHANNELMSG_ALLOFFERS_DELIVERED,       1, vmbus_onoffers_delivered },
1117         { CHANNELMSG_OPENCHANNEL,               0, NULL },
1118         { CHANNELMSG_OPENCHANNEL_RESULT,        1, vmbus_onopen_result },
1119         { CHANNELMSG_CLOSECHANNEL,              0, NULL },
1120         { CHANNELMSG_GPADL_HEADER,              0, NULL },
1121         { CHANNELMSG_GPADL_BODY,                0, NULL },
1122         { CHANNELMSG_GPADL_CREATED,             1, vmbus_ongpadl_created },
1123         { CHANNELMSG_GPADL_TEARDOWN,            0, NULL },
1124         { CHANNELMSG_GPADL_TORNDOWN,            1, vmbus_ongpadl_torndown },
1125         { CHANNELMSG_RELID_RELEASED,            0, NULL },
1126         { CHANNELMSG_INITIATE_CONTACT,          0, NULL },
1127         { CHANNELMSG_VERSION_RESPONSE,          1, vmbus_onversion_response },
1128         { CHANNELMSG_UNLOAD,                    0, NULL },
1129         { CHANNELMSG_UNLOAD_RESPONSE,           1, vmbus_unload_response },
1130         { CHANNELMSG_18,                        0, NULL },
1131         { CHANNELMSG_19,                        0, NULL },
1132         { CHANNELMSG_20,                        0, NULL },
1133         { CHANNELMSG_TL_CONNECT_REQUEST,        0, NULL },
1134 };
1135
1136 /*
1137  * vmbus_onmessage - Handler for channel protocol messages.
1138  *
1139  * This is invoked in the vmbus worker thread context.
1140  */
1141 void vmbus_onmessage(void *context)
1142 {
1143         struct hv_message *msg = context;
1144         struct vmbus_channel_message_header *hdr;
1145         int size;
1146
1147         hdr = (struct vmbus_channel_message_header *)msg->u.payload;
1148         size = msg->header.payload_size;
1149
1150         if (hdr->msgtype >= CHANNELMSG_COUNT) {
1151                 pr_err("Received invalid channel message type %d size %d\n",
1152                            hdr->msgtype, size);
1153                 print_hex_dump_bytes("", DUMP_PREFIX_NONE,
1154                                      (unsigned char *)msg->u.payload, size);
1155                 return;
1156         }
1157
1158         if (channel_message_table[hdr->msgtype].message_handler)
1159                 channel_message_table[hdr->msgtype].message_handler(hdr);
1160         else
1161                 pr_err("Unhandled channel message type %d\n", hdr->msgtype);
1162 }
1163
1164 /*
1165  * vmbus_request_offers - Send a request to get all our pending offers.
1166  */
1167 int vmbus_request_offers(void)
1168 {
1169         struct vmbus_channel_message_header *msg;
1170         struct vmbus_channel_msginfo *msginfo;
1171         int ret;
1172
1173         msginfo = kmalloc(sizeof(*msginfo) +
1174                           sizeof(struct vmbus_channel_message_header),
1175                           GFP_KERNEL);
1176         if (!msginfo)
1177                 return -ENOMEM;
1178
1179         msg = (struct vmbus_channel_message_header *)msginfo->msg;
1180
1181         msg->msgtype = CHANNELMSG_REQUESTOFFERS;
1182
1183
1184         ret = vmbus_post_msg(msg, sizeof(struct vmbus_channel_message_header),
1185                              true);
1186         if (ret != 0) {
1187                 pr_err("Unable to request offers - %d\n", ret);
1188
1189                 goto cleanup;
1190         }
1191
1192 cleanup:
1193         kfree(msginfo);
1194
1195         return ret;
1196 }
1197
1198 /*
1199  * Retrieve the (sub) channel on which to send an outgoing request.
1200  * When a primary channel has multiple sub-channels, we try to
1201  * distribute the load equally amongst all available channels.
1202  */
1203 struct vmbus_channel *vmbus_get_outgoing_channel(struct vmbus_channel *primary)
1204 {
1205         struct list_head *cur, *tmp;
1206         int cur_cpu;
1207         struct vmbus_channel *cur_channel;
1208         struct vmbus_channel *outgoing_channel = primary;
1209         int next_channel;
1210         int i = 1;
1211
1212         if (list_empty(&primary->sc_list))
1213                 return outgoing_channel;
1214
1215         next_channel = primary->next_oc++;
1216
1217         if (next_channel > (primary->num_sc)) {
1218                 primary->next_oc = 0;
1219                 return outgoing_channel;
1220         }
1221
1222         cur_cpu = hv_cpu_number_to_vp_number(smp_processor_id());
1223         list_for_each_safe(cur, tmp, &primary->sc_list) {
1224                 cur_channel = list_entry(cur, struct vmbus_channel, sc_list);
1225                 if (cur_channel->state != CHANNEL_OPENED_STATE)
1226                         continue;
1227
1228                 if (cur_channel->target_vp == cur_cpu)
1229                         return cur_channel;
1230
1231                 if (i == next_channel)
1232                         return cur_channel;
1233
1234                 i++;
1235         }
1236
1237         return outgoing_channel;
1238 }
1239 EXPORT_SYMBOL_GPL(vmbus_get_outgoing_channel);
1240
1241 static void invoke_sc_cb(struct vmbus_channel *primary_channel)
1242 {
1243         struct list_head *cur, *tmp;
1244         struct vmbus_channel *cur_channel;
1245
1246         if (primary_channel->sc_creation_callback == NULL)
1247                 return;
1248
1249         list_for_each_safe(cur, tmp, &primary_channel->sc_list) {
1250                 cur_channel = list_entry(cur, struct vmbus_channel, sc_list);
1251
1252                 primary_channel->sc_creation_callback(cur_channel);
1253         }
1254 }
1255
1256 void vmbus_set_sc_create_callback(struct vmbus_channel *primary_channel,
1257                                 void (*sc_cr_cb)(struct vmbus_channel *new_sc))
1258 {
1259         primary_channel->sc_creation_callback = sc_cr_cb;
1260 }
1261 EXPORT_SYMBOL_GPL(vmbus_set_sc_create_callback);
1262
1263 bool vmbus_are_subchannels_present(struct vmbus_channel *primary)
1264 {
1265         bool ret;
1266
1267         ret = !list_empty(&primary->sc_list);
1268
1269         if (ret) {
1270                 /*
1271                  * Invoke the callback on sub-channel creation.
1272                  * This will present a uniform interface to the
1273                  * clients.
1274                  */
1275                 invoke_sc_cb(primary);
1276         }
1277
1278         return ret;
1279 }
1280 EXPORT_SYMBOL_GPL(vmbus_are_subchannels_present);
1281
1282 void vmbus_set_chn_rescind_callback(struct vmbus_channel *channel,
1283                 void (*chn_rescind_cb)(struct vmbus_channel *))
1284 {
1285         channel->chn_rescind_callback = chn_rescind_cb;
1286 }
1287 EXPORT_SYMBOL_GPL(vmbus_set_chn_rescind_callback);