drivers/hv/channel_mgmt.c

   1 /*
   2  * Copyright (c) 2009, Microsoft Corporation.
   3  *
   4  * This program is free software; you can redistribute it and/or modify it
   5  * under the terms and conditions of the GNU General Public License,
   6  * version 2, as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope it will be useful, but WITHOUT
   9  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  10  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  11  * more details.
  12  *
  13  * You should have received a copy of the GNU General Public License along with
  14  * this program; if not, write to the Free Software Foundation, Inc., 59 Temple
  15  * Place - Suite 330, Boston, MA 02111-1307 USA.
  16  *
  17  * Authors:
  18  *   Haiyang Zhang <haiyangz@microsoft.com>
  19  *   Hank Janssen  <hjanssen@microsoft.com>
  20  */
  21 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  22
  23 #include <linux/kernel.h>
  24 #include <linux/interrupt.h>
  25 #include <linux/sched.h>
  26 #include <linux/wait.h>
  27 #include <linux/mm.h>
  28 #include <linux/slab.h>
  29 #include <linux/list.h>
  30 #include <linux/module.h>
  31 #include <linux/completion.h>
  32 #include <linux/delay.h>
  33 #include <linux/hyperv.h>
  34 #include <asm/mshyperv.h>
  35
  36 #include "hyperv_vmbus.h"
  37
  38 static void init_vp_index(struct vmbus_channel *channel, u16 dev_type);
  39
  40 static const struct vmbus_device vmbus_devs[] = {
  41         /* IDE */
  42         { .dev_type = HV_IDE,
  43           HV_IDE_GUID,
  44           .perf_device = true,
  45         },
  46
  47         /* SCSI */
  48         { .dev_type = HV_SCSI,
  49           HV_SCSI_GUID,
  50           .perf_device = true,
  51         },
  52
  53         /* Fibre Channel */
  54         { .dev_type = HV_FC,
  55           HV_SYNTHFC_GUID,
  56           .perf_device = true,
  57         },
  58
  59         /* Synthetic NIC */
  60         { .dev_type = HV_NIC,
  61           HV_NIC_GUID,
  62           .perf_device = true,
  63         },
  64
  65         /* Network Direct */
  66         { .dev_type = HV_ND,
  67           HV_ND_GUID,
  68           .perf_device = true,
  69         },
  70
  71         /* PCIE */
  72         { .dev_type = HV_PCIE,
  73           HV_PCIE_GUID,
  74           .perf_device = 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         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                         atomic_dec(&vmbus_connection.offer_in_progress);
 532                         goto err_free_chan;
 533                 }
 534
 535                 if (channel->sc_creation_callback != NULL)
 536                         channel->sc_creation_callback(newchannel);
 537                 newchannel->probe_done = true;
 538                 return;
 539         }
 540
 541         /*
 542          * Start the process of binding this offer to the driver
 543          * We need to set the DeviceObject field before calling
 544          * vmbus_child_dev_add()
 545          */
 546         newchannel->device_obj = vmbus_device_create(
 547                 &newchannel->offermsg.offer.if_type,
 548                 &newchannel->offermsg.offer.if_instance,
 549                 newchannel);
 550         if (!newchannel->device_obj)
 551                 goto err_deq_chan;
 552
 553         newchannel->device_obj->device_id = dev_type;
 554         /*
 555          * Add the new device to the bus. This will kick off device-driver
 556          * binding which eventually invokes the device driver's AddDevice()
 557          * method.
 558          */
 559         ret = vmbus_device_register(newchannel->device_obj);
 560
 561         if (ret != 0) {
 562                 pr_err("unable to add child device object (relid %d)\n",
 563                         newchannel->offermsg.child_relid);
 564                 kfree(newchannel->device_obj);
 565                 goto err_deq_chan;
 566         }
 567
 568         newchannel->probe_done = true;
 569         return;
 570
 571 err_deq_chan:
 572         mutex_lock(&vmbus_connection.channel_mutex);
 573         list_del(&newchannel->listentry);
 574         mutex_unlock(&vmbus_connection.channel_mutex);
 575
 576         if (newchannel->target_cpu != get_cpu()) {
 577                 put_cpu();
 578                 smp_call_function_single(newchannel->target_cpu,
 579                                          percpu_channel_deq, newchannel, true);
 580         } else {
 581                 percpu_channel_deq(newchannel);
 582                 put_cpu();
 583         }
 584
 585         vmbus_release_relid(newchannel->offermsg.child_relid);
 586
 587 err_free_chan:
 588         free_channel(newchannel);
 589 }
 590
 591 /*
 592  * We use this state to statically distribute the channel interrupt load.
 593  */
 594 static int next_numa_node_id;
 595
 596 /*
 597  * Starting with Win8, we can statically distribute the incoming
 598  * channel interrupt load by binding a channel to VCPU.
 599  * We do this in a hierarchical fashion:
 600  * First distribute the primary channels across available NUMA nodes
 601  * and then distribute the subchannels amongst the CPUs in the NUMA
 602  * node assigned to the primary channel.
 603  *
 604  * For pre-win8 hosts or non-performance critical channels we assign the
 605  * first CPU in the first NUMA node.
 606  */
 607 static void init_vp_index(struct vmbus_channel *channel, u16 dev_type)
 608 {
 609         u32 cur_cpu;
 610         bool perf_chn = vmbus_devs[dev_type].perf_device;
 611         struct vmbus_channel *primary = channel->primary_channel;
 612         int next_node;
 613         struct cpumask available_mask;
 614         struct cpumask *alloced_mask;
 615
 616         if ((vmbus_proto_version == VERSION_WS2008) ||
 617             (vmbus_proto_version == VERSION_WIN7) || (!perf_chn)) {
 618                 /*
 619                  * Prior to win8, all channel interrupts are
 620                  * delivered on cpu 0.
 621                  * Also if the channel is not a performance critical
 622                  * channel, bind it to cpu 0.
 623                  */
 624                 channel->numa_node = 0;
 625                 channel->target_cpu = 0;
 626                 channel->target_vp = hv_cpu_number_to_vp_number(0);
 627                 return;
 628         }
 629
 630         /*
 631          * Based on the channel affinity policy, we will assign the NUMA
 632          * nodes.
 633          */
 634
 635         if ((channel->affinity_policy == HV_BALANCED) || (!primary)) {
 636                 while (true) {
 637                         next_node = next_numa_node_id++;
 638                         if (next_node == nr_node_ids) {
 639                                 next_node = next_numa_node_id = 0;
 640                                 continue;
 641                         }
 642                         if (cpumask_empty(cpumask_of_node(next_node)))
 643                                 continue;
 644                         break;
 645                 }
 646                 channel->numa_node = next_node;
 647                 primary = channel;
 648         }
 649         alloced_mask = &hv_context.hv_numa_map[primary->numa_node];
 650
 651         if (cpumask_weight(alloced_mask) ==
 652             cpumask_weight(cpumask_of_node(primary->numa_node))) {
 653                 /*
 654                  * We have cycled through all the CPUs in the node;
 655                  * reset the alloced map.
 656                  */
 657                 cpumask_clear(alloced_mask);
 658         }
 659
 660         cpumask_xor(&available_mask, alloced_mask,
 661                     cpumask_of_node(primary->numa_node));
 662
 663         cur_cpu = -1;
 664
 665         if (primary->affinity_policy == HV_LOCALIZED) {
 666                 /*
 667                  * Normally Hyper-V host doesn't create more subchannels
 668                  * than there are VCPUs on the node but it is possible when not
 669                  * all present VCPUs on the node are initialized by guest.
 670                  * Clear the alloced_cpus_in_node to start over.
 671                  */
 672                 if (cpumask_equal(&primary->alloced_cpus_in_node,
 673                                   cpumask_of_node(primary->numa_node)))
 674                         cpumask_clear(&primary->alloced_cpus_in_node);
 675         }
 676
 677         while (true) {
 678                 cur_cpu = cpumask_next(cur_cpu, &available_mask);
 679                 if (cur_cpu >= nr_cpu_ids) {
 680                         cur_cpu = -1;
 681                         cpumask_copy(&available_mask,
 682                                      cpumask_of_node(primary->numa_node));
 683                         continue;
 684                 }
 685
 686                 if (primary->affinity_policy == HV_LOCALIZED) {
 687                         /*
 688                          * NOTE: in the case of sub-channel, we clear the
 689                          * sub-channel related bit(s) in
 690                          * primary->alloced_cpus_in_node in
 691                          * hv_process_channel_removal(), so when we
 692                          * reload drivers like hv_netvsc in SMP guest, here
 693                          * we're able to re-allocate
 694                          * bit from primary->alloced_cpus_in_node.
 695                          */
 696                         if (!cpumask_test_cpu(cur_cpu,
 697                                               &primary->alloced_cpus_in_node)) {
 698                                 cpumask_set_cpu(cur_cpu,
 699                                                 &primary->alloced_cpus_in_node);
 700                                 cpumask_set_cpu(cur_cpu, alloced_mask);
 701                                 break;
 702                         }
 703                 } else {
 704                         cpumask_set_cpu(cur_cpu, alloced_mask);
 705                         break;
 706                 }
 707         }
 708
 709         channel->target_cpu = cur_cpu;
 710         channel->target_vp = hv_cpu_number_to_vp_number(cur_cpu);
 711 }
 712
 713 static void vmbus_wait_for_unload(void)
 714 {
 715         int cpu;
 716         void *page_addr;
 717         struct hv_message *msg;
 718         struct vmbus_channel_message_header *hdr;
 719         u32 message_type;
 720
 721         /*
 722          * CHANNELMSG_UNLOAD_RESPONSE is always delivered to the CPU which was
 723          * used for initial contact or to CPU0 depending on host version. When
 724          * we're crashing on a different CPU let's hope that IRQ handler on
 725          * the cpu which receives CHANNELMSG_UNLOAD_RESPONSE is still
 726          * functional and vmbus_unload_response() will complete
 727          * vmbus_connection.unload_event. If not, the last thing we can do is
 728          * read message pages for all CPUs directly.
 729          */
 730         while (1) {
 731                 if (completion_done(&vmbus_connection.unload_event))
 732                         break;
 733
 734                 for_each_online_cpu(cpu) {
 735                         struct hv_per_cpu_context *hv_cpu
 736                                 = per_cpu_ptr(hv_context.cpu_context, cpu);
 737
 738                         page_addr = hv_cpu->synic_message_page;
 739                         msg = (struct hv_message *)page_addr
 740                                 + VMBUS_MESSAGE_SINT;
 741
 742                         message_type = READ_ONCE(msg->header.message_type);
 743                         if (message_type == HVMSG_NONE)
 744                                 continue;
 745
 746                         hdr = (struct vmbus_channel_message_header *)
 747                                 msg->u.payload;
 748
 749                         if (hdr->msgtype == CHANNELMSG_UNLOAD_RESPONSE)
 750                                 complete(&vmbus_connection.unload_event);
 751
 752                         vmbus_signal_eom(msg, message_type);
 753                 }
 754
 755                 mdelay(10);
 756         }
 757
 758         /*
 759          * We're crashing and already got the UNLOAD_RESPONSE, cleanup all
 760          * maybe-pending messages on all CPUs to be able to receive new
 761          * messages after we reconnect.
 762          */
 763         for_each_online_cpu(cpu) {
 764                 struct hv_per_cpu_context *hv_cpu
 765                         = per_cpu_ptr(hv_context.cpu_context, cpu);
 766
 767                 page_addr = hv_cpu->synic_message_page;
 768                 msg = (struct hv_message *)page_addr + VMBUS_MESSAGE_SINT;
 769                 msg->header.message_type = HVMSG_NONE;
 770         }
 771 }
 772
 773 /*
 774  * vmbus_unload_response - Handler for the unload response.
 775  */
 776 static void vmbus_unload_response(struct vmbus_channel_message_header *hdr)
 777 {
 778         /*
 779          * This is a global event; just wakeup the waiting thread.
 780          * Once we successfully unload, we can cleanup the monitor state.
 781          */
 782         complete(&vmbus_connection.unload_event);
 783 }
 784
 785 void vmbus_initiate_unload(bool crash)
 786 {
 787         struct vmbus_channel_message_header hdr;
 788
 789         /* Pre-Win2012R2 hosts don't support reconnect */
 790         if (vmbus_proto_version < VERSION_WIN8_1)
 791                 return;
 792
 793         init_completion(&vmbus_connection.unload_event);
 794         memset(&hdr, 0, sizeof(struct vmbus_channel_message_header));
 795         hdr.msgtype = CHANNELMSG_UNLOAD;
 796         vmbus_post_msg(&hdr, sizeof(struct vmbus_channel_message_header),
 797                        !crash);
 798
 799         /*
 800          * vmbus_initiate_unload() is also called on crash and the crash can be
 801          * happening in an interrupt context, where scheduling is impossible.
 802          */
 803         if (!crash)
 804                 wait_for_completion(&vmbus_connection.unload_event);
 805         else
 806                 vmbus_wait_for_unload();
 807 }
 808
 809 /*
 810  * vmbus_onoffer - Handler for channel offers from vmbus in parent partition.
 811  *
 812  */
 813 static void vmbus_onoffer(struct vmbus_channel_message_header *hdr)
 814 {
 815         struct vmbus_channel_offer_channel *offer;
 816         struct vmbus_channel *newchannel;
 817
 818         offer = (struct vmbus_channel_offer_channel *)hdr;
 819
 820         trace_vmbus_onoffer(offer);
 821
 822         /* Allocate the channel object and save this offer. */
 823         newchannel = alloc_channel();
 824         if (!newchannel) {
 825                 vmbus_release_relid(offer->child_relid);
 826                 atomic_dec(&vmbus_connection.offer_in_progress);
 827                 pr_err("Unable to allocate channel object\n");
 828                 return;
 829         }
 830
 831         /*
 832          * Setup state for signalling the host.
 833          */
 834         newchannel->sig_event = VMBUS_EVENT_CONNECTION_ID;
 835
 836         if (vmbus_proto_version != VERSION_WS2008) {
 837                 newchannel->is_dedicated_interrupt =
 838                                 (offer->is_dedicated_interrupt != 0);
 839                 newchannel->sig_event = offer->connection_id;
 840         }
 841
 842         memcpy(&newchannel->offermsg, offer,
 843                sizeof(struct vmbus_channel_offer_channel));
 844         newchannel->monitor_grp = (u8)offer->monitorid / 32;
 845         newchannel->monitor_bit = (u8)offer->monitorid % 32;
 846
 847         vmbus_process_offer(newchannel);
 848 }
 849
 850 /*
 851  * vmbus_onoffer_rescind - Rescind offer handler.
 852  *
 853  * We queue a work item to process this offer synchronously
 854  */
 855 static void vmbus_onoffer_rescind(struct vmbus_channel_message_header *hdr)
 856 {
 857         struct vmbus_channel_rescind_offer *rescind;
 858         struct vmbus_channel *channel;
 859         struct device *dev;
 860
 861         rescind = (struct vmbus_channel_rescind_offer *)hdr;
 862
 863         trace_vmbus_onoffer_rescind(rescind);
 864
 865         /*
 866          * The offer msg and the corresponding rescind msg
 867          * from the host are guranteed to be ordered -
 868          * offer comes in first and then the rescind.
 869          * Since we process these events in work elements,
 870          * and with preemption, we may end up processing
 871          * the events out of order. Given that we handle these
 872          * work elements on the same CPU, this is possible only
 873          * in the case of preemption. In any case wait here
 874          * until the offer processing has moved beyond the
 875          * point where the channel is discoverable.
 876          */
 877
 878         while (atomic_read(&vmbus_connection.offer_in_progress) != 0) {
 879                 /*
 880                  * We wait here until any channel offer is currently
 881                  * being processed.
 882                  */
 883                 msleep(1);
 884         }
 885
 886         mutex_lock(&vmbus_connection.channel_mutex);
 887         channel = relid2channel(rescind->child_relid);
 888         mutex_unlock(&vmbus_connection.channel_mutex);
 889
 890         if (channel == NULL) {
 891                 /*
 892                  * We failed in processing the offer message;
 893                  * we would have cleaned up the relid in that
 894                  * failure path.
 895                  */
 896                 return;
 897         }
 898
 899         /*
 900          * Now wait for offer handling to complete.
 901          */
 902         vmbus_rescind_cleanup(channel);
 903         while (READ_ONCE(channel->probe_done) == false) {
 904                 /*
 905                  * We wait here until any channel offer is currently
 906                  * being processed.
 907                  */
 908                 msleep(1);
 909         }
 910
 911         /*
 912          * At this point, the rescind handling can proceed safely.
 913          */
 914
 915         if (channel->device_obj) {
 916                 if (channel->chn_rescind_callback) {
 917                         channel->chn_rescind_callback(channel);
 918                         return;
 919                 }
 920                 /*
 921                  * We will have to unregister this device from the
 922                  * driver core.
 923                  */
 924                 dev = get_device(&channel->device_obj->device);
 925                 if (dev) {
 926                         vmbus_device_unregister(channel->device_obj);
 927                         put_device(dev);
 928                 }
 929         }
 930         if (channel->primary_channel != NULL) {
 931                 /*
 932                  * Sub-channel is being rescinded. Following is the channel
 933                  * close sequence when initiated from the driveri (refer to
 934                  * vmbus_close() for details):
 935                  * 1. Close all sub-channels first
 936                  * 2. Then close the primary channel.
 937                  */
 938                 mutex_lock(&vmbus_connection.channel_mutex);
 939                 if (channel->state == CHANNEL_OPEN_STATE) {
 940                         /*
 941                          * The channel is currently not open;
 942                          * it is safe for us to cleanup the channel.
 943                          */
 944                         hv_process_channel_removal(rescind->child_relid);
 945                 } else {
 946                         complete(&channel->rescind_event);
 947                 }
 948                 mutex_unlock(&vmbus_connection.channel_mutex);
 949         }
 950 }
 951
 952 void vmbus_hvsock_device_unregister(struct vmbus_channel *channel)
 953 {
 954         BUG_ON(!is_hvsock_channel(channel));
 955
 956         /* We always get a rescind msg when a connection is closed. */
 957         while (!READ_ONCE(channel->probe_done) || !READ_ONCE(channel->rescind))
 958                 msleep(1);
 959
 960         vmbus_device_unregister(channel->device_obj);
 961 }
 962 EXPORT_SYMBOL_GPL(vmbus_hvsock_device_unregister);
 963
 964
 965 /*
 966  * vmbus_onoffers_delivered -
 967  * This is invoked when all offers have been delivered.
 968  *
 969  * Nothing to do here.
 970  */
 971 static void vmbus_onoffers_delivered(
 972                         struct vmbus_channel_message_header *hdr)
 973 {
 974 }
 975
 976 /*
 977  * vmbus_onopen_result - Open result handler.
 978  *
 979  * This is invoked when we received a response to our channel open request.
 980  * Find the matching request, copy the response and signal the requesting
 981  * thread.
 982  */
 983 static void vmbus_onopen_result(struct vmbus_channel_message_header *hdr)
 984 {
 985         struct vmbus_channel_open_result *result;
 986         struct vmbus_channel_msginfo *msginfo;
 987         struct vmbus_channel_message_header *requestheader;
 988         struct vmbus_channel_open_channel *openmsg;
 989         unsigned long flags;
 990
 991         result = (struct vmbus_channel_open_result *)hdr;
 992
 993         trace_vmbus_onopen_result(result);
 994
 995         /*
 996          * Find the open msg, copy the result and signal/unblock the wait event
 997          */
 998         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
 999
1000         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1001                                 msglistentry) {
1002                 requestheader =
1003                         (struct vmbus_channel_message_header *)msginfo->msg;
1004
1005                 if (requestheader->msgtype == CHANNELMSG_OPENCHANNEL) {
1006                         openmsg =
1007                         (struct vmbus_channel_open_channel *)msginfo->msg;
1008                         if (openmsg->child_relid == result->child_relid &&
1009                             openmsg->openid == result->openid) {
1010                                 memcpy(&msginfo->response.open_result,
1011                                        result,
1012                                        sizeof(
1013                                         struct vmbus_channel_open_result));
1014                                 complete(&msginfo->waitevent);
1015                                 break;
1016                         }
1017                 }
1018         }
1019         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1020 }
1021
1022 /*
1023  * vmbus_ongpadl_created - GPADL created handler.
1024  *
1025  * This is invoked when we received a response to our gpadl create request.
1026  * Find the matching request, copy the response and signal the requesting
1027  * thread.
1028  */
1029 static void vmbus_ongpadl_created(struct vmbus_channel_message_header *hdr)
1030 {
1031         struct vmbus_channel_gpadl_created *gpadlcreated;
1032         struct vmbus_channel_msginfo *msginfo;
1033         struct vmbus_channel_message_header *requestheader;
1034         struct vmbus_channel_gpadl_header *gpadlheader;
1035         unsigned long flags;
1036
1037         gpadlcreated = (struct vmbus_channel_gpadl_created *)hdr;
1038
1039         trace_vmbus_ongpadl_created(gpadlcreated);
1040
1041         /*
1042          * Find the establish msg, copy the result and signal/unblock the wait
1043          * event
1044          */
1045         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1046
1047         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1048                                 msglistentry) {
1049                 requestheader =
1050                         (struct vmbus_channel_message_header *)msginfo->msg;
1051
1052                 if (requestheader->msgtype == CHANNELMSG_GPADL_HEADER) {
1053                         gpadlheader =
1054                         (struct vmbus_channel_gpadl_header *)requestheader;
1055
1056                         if ((gpadlcreated->child_relid ==
1057                              gpadlheader->child_relid) &&
1058                             (gpadlcreated->gpadl == gpadlheader->gpadl)) {
1059                                 memcpy(&msginfo->response.gpadl_created,
1060                                        gpadlcreated,
1061                                        sizeof(
1062                                         struct vmbus_channel_gpadl_created));
1063                                 complete(&msginfo->waitevent);
1064                                 break;
1065                         }
1066                 }
1067         }
1068         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1069 }
1070
1071 /*
1072  * vmbus_ongpadl_torndown - GPADL torndown handler.
1073  *
1074  * This is invoked when we received a response to our gpadl teardown request.
1075  * Find the matching request, copy the response and signal the requesting
1076  * thread.
1077  */
1078 static void vmbus_ongpadl_torndown(
1079                         struct vmbus_channel_message_header *hdr)
1080 {
1081         struct vmbus_channel_gpadl_torndown *gpadl_torndown;
1082         struct vmbus_channel_msginfo *msginfo;
1083         struct vmbus_channel_message_header *requestheader;
1084         struct vmbus_channel_gpadl_teardown *gpadl_teardown;
1085         unsigned long flags;
1086
1087         gpadl_torndown = (struct vmbus_channel_gpadl_torndown *)hdr;
1088
1089         trace_vmbus_ongpadl_torndown(gpadl_torndown);
1090
1091         /*
1092          * Find the open msg, copy the result and signal/unblock the wait event
1093          */
1094         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1095
1096         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1097                                 msglistentry) {
1098                 requestheader =
1099                         (struct vmbus_channel_message_header *)msginfo->msg;
1100
1101                 if (requestheader->msgtype == CHANNELMSG_GPADL_TEARDOWN) {
1102                         gpadl_teardown =
1103                         (struct vmbus_channel_gpadl_teardown *)requestheader;
1104
1105                         if (gpadl_torndown->gpadl == gpadl_teardown->gpadl) {
1106                                 memcpy(&msginfo->response.gpadl_torndown,
1107                                        gpadl_torndown,
1108                                        sizeof(
1109                                         struct vmbus_channel_gpadl_torndown));
1110                                 complete(&msginfo->waitevent);
1111                                 break;
1112                         }
1113                 }
1114         }
1115         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1116 }
1117
1118 /*
1119  * vmbus_onversion_response - Version response handler
1120  *
1121  * This is invoked when we received a response to our initiate contact request.
1122  * Find the matching request, copy the response and signal the requesting
1123  * thread.
1124  */
1125 static void vmbus_onversion_response(
1126                 struct vmbus_channel_message_header *hdr)
1127 {
1128         struct vmbus_channel_msginfo *msginfo;
1129         struct vmbus_channel_message_header *requestheader;
1130         struct vmbus_channel_version_response *version_response;
1131         unsigned long flags;
1132
1133         version_response = (struct vmbus_channel_version_response *)hdr;
1134
1135         trace_vmbus_onversion_response(version_response);
1136
1137         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1138
1139         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1140                                 msglistentry) {
1141                 requestheader =
1142                         (struct vmbus_channel_message_header *)msginfo->msg;
1143
1144                 if (requestheader->msgtype ==
1145                     CHANNELMSG_INITIATE_CONTACT) {
1146                         memcpy(&msginfo->response.version_response,
1147                               version_response,
1148                               sizeof(struct vmbus_channel_version_response));
1149                         complete(&msginfo->waitevent);
1150                 }
1151         }
1152         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1153 }
1154
1155 /* Channel message dispatch table */
1156 const struct vmbus_channel_message_table_entry
1157 channel_message_table[CHANNELMSG_COUNT] = {
1158         { CHANNELMSG_INVALID,                   0, NULL },
1159         { CHANNELMSG_OFFERCHANNEL,              0, vmbus_onoffer },
1160         { CHANNELMSG_RESCIND_CHANNELOFFER,      0, vmbus_onoffer_rescind },
1161         { CHANNELMSG_REQUESTOFFERS,             0, NULL },
1162         { CHANNELMSG_ALLOFFERS_DELIVERED,       1, vmbus_onoffers_delivered },
1163         { CHANNELMSG_OPENCHANNEL,               0, NULL },
1164         { CHANNELMSG_OPENCHANNEL_RESULT,        1, vmbus_onopen_result },
1165         { CHANNELMSG_CLOSECHANNEL,              0, NULL },
1166         { CHANNELMSG_GPADL_HEADER,              0, NULL },
1167         { CHANNELMSG_GPADL_BODY,                0, NULL },
1168         { CHANNELMSG_GPADL_CREATED,             1, vmbus_ongpadl_created },
1169         { CHANNELMSG_GPADL_TEARDOWN,            0, NULL },
1170         { CHANNELMSG_GPADL_TORNDOWN,            1, vmbus_ongpadl_torndown },
1171         { CHANNELMSG_RELID_RELEASED,            0, NULL },
1172         { CHANNELMSG_INITIATE_CONTACT,          0, NULL },
1173         { CHANNELMSG_VERSION_RESPONSE,          1, vmbus_onversion_response },
1174         { CHANNELMSG_UNLOAD,                    0, NULL },
1175         { CHANNELMSG_UNLOAD_RESPONSE,           1, vmbus_unload_response },
1176         { CHANNELMSG_18,                        0, NULL },
1177         { CHANNELMSG_19,                        0, NULL },
1178         { CHANNELMSG_20,                        0, NULL },
1179         { CHANNELMSG_TL_CONNECT_REQUEST,        0, NULL },
1180 };
1181
1182 /*
1183  * vmbus_onmessage - Handler for channel protocol messages.
1184  *
1185  * This is invoked in the vmbus worker thread context.
1186  */
1187 void vmbus_onmessage(void *context)
1188 {
1189         struct hv_message *msg = context;
1190         struct vmbus_channel_message_header *hdr;
1191         int size;
1192
1193         hdr = (struct vmbus_channel_message_header *)msg->u.payload;
1194         size = msg->header.payload_size;
1195
1196         trace_vmbus_on_message(hdr);
1197
1198         if (hdr->msgtype >= CHANNELMSG_COUNT) {
1199                 pr_err("Received invalid channel message type %d size %d\n",
1200                            hdr->msgtype, size);
1201                 print_hex_dump_bytes("", DUMP_PREFIX_NONE,
1202                                      (unsigned char *)msg->u.payload, size);
1203                 return;
1204         }
1205
1206         if (channel_message_table[hdr->msgtype].message_handler)
1207                 channel_message_table[hdr->msgtype].message_handler(hdr);
1208         else
1209                 pr_err("Unhandled channel message type %d\n", hdr->msgtype);
1210 }
1211
1212 /*
1213  * vmbus_request_offers - Send a request to get all our pending offers.
1214  */
1215 int vmbus_request_offers(void)
1216 {
1217         struct vmbus_channel_message_header *msg;
1218         struct vmbus_channel_msginfo *msginfo;
1219         int ret;
1220
1221         msginfo = kmalloc(sizeof(*msginfo) +
1222                           sizeof(struct vmbus_channel_message_header),
1223                           GFP_KERNEL);
1224         if (!msginfo)
1225                 return -ENOMEM;
1226
1227         msg = (struct vmbus_channel_message_header *)msginfo->msg;
1228
1229         msg->msgtype = CHANNELMSG_REQUESTOFFERS;
1230
1231         ret = vmbus_post_msg(msg, sizeof(struct vmbus_channel_message_header),
1232                              true);
1233
1234         trace_vmbus_request_offers(ret);
1235
1236         if (ret != 0) {
1237                 pr_err("Unable to request offers - %d\n", ret);
1238
1239                 goto cleanup;
1240         }
1241
1242 cleanup:
1243         kfree(msginfo);
1244
1245         return ret;
1246 }
1247
1248 /*
1249  * Retrieve the (sub) channel on which to send an outgoing request.
1250  * When a primary channel has multiple sub-channels, we try to
1251  * distribute the load equally amongst all available channels.
1252  */
1253 struct vmbus_channel *vmbus_get_outgoing_channel(struct vmbus_channel *primary)
1254 {
1255         struct list_head *cur, *tmp;
1256         int cur_cpu;
1257         struct vmbus_channel *cur_channel;
1258         struct vmbus_channel *outgoing_channel = primary;
1259         int next_channel;
1260         int i = 1;
1261
1262         if (list_empty(&primary->sc_list))
1263                 return outgoing_channel;
1264
1265         next_channel = primary->next_oc++;
1266
1267         if (next_channel > (primary->num_sc)) {
1268                 primary->next_oc = 0;
1269                 return outgoing_channel;
1270         }
1271
1272         cur_cpu = hv_cpu_number_to_vp_number(smp_processor_id());
1273         list_for_each_safe(cur, tmp, &primary->sc_list) {
1274                 cur_channel = list_entry(cur, struct vmbus_channel, sc_list);
1275                 if (cur_channel->state != CHANNEL_OPENED_STATE)
1276                         continue;
1277
1278                 if (cur_channel->target_vp == cur_cpu)
1279                         return cur_channel;
1280
1281                 if (i == next_channel)
1282                         return cur_channel;
1283
1284                 i++;
1285         }
1286
1287         return outgoing_channel;
1288 }
1289 EXPORT_SYMBOL_GPL(vmbus_get_outgoing_channel);
1290
1291 static void invoke_sc_cb(struct vmbus_channel *primary_channel)
1292 {
1293         struct list_head *cur, *tmp;
1294         struct vmbus_channel *cur_channel;
1295
1296         if (primary_channel->sc_creation_callback == NULL)
1297                 return;
1298
1299         list_for_each_safe(cur, tmp, &primary_channel->sc_list) {
1300                 cur_channel = list_entry(cur, struct vmbus_channel, sc_list);
1301
1302                 primary_channel->sc_creation_callback(cur_channel);
1303         }
1304 }
1305
1306 void vmbus_set_sc_create_callback(struct vmbus_channel *primary_channel,
1307                                 void (*sc_cr_cb)(struct vmbus_channel *new_sc))
1308 {
1309         primary_channel->sc_creation_callback = sc_cr_cb;
1310 }
1311 EXPORT_SYMBOL_GPL(vmbus_set_sc_create_callback);
1312
1313 bool vmbus_are_subchannels_present(struct vmbus_channel *primary)
1314 {
1315         bool ret;
1316
1317         ret = !list_empty(&primary->sc_list);
1318
1319         if (ret) {
1320                 /*
1321                  * Invoke the callback on sub-channel creation.
1322                  * This will present a uniform interface to the
1323                  * clients.
1324                  */
1325                 invoke_sc_cb(primary);
1326         }
1327
1328         return ret;
1329 }
1330 EXPORT_SYMBOL_GPL(vmbus_are_subchannels_present);
1331
1332 void vmbus_set_chn_rescind_callback(struct vmbus_channel *channel,
1333                 void (*chn_rescind_cb)(struct vmbus_channel *))
1334 {
1335         channel->chn_rescind_callback = chn_rescind_cb;
1336 }
1337 EXPORT_SYMBOL_GPL(vmbus_set_chn_rescind_callback);