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 152  * The rescinded channel may be blocked waiting for a response from the host;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