drivers/hv/channel_mgmt.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (c) 2009, Microsoft Corporation.
   4  *
   5  * Authors:
   6  *   Haiyang Zhang <haiyangz@microsoft.com>
   7  *   Hank Janssen  <hjanssen@microsoft.com>
   8  */
   9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  10
  11 #include <linux/kernel.h>
  12 #include <linux/interrupt.h>
  13 #include <linux/sched.h>
  14 #include <linux/wait.h>
  15 #include <linux/mm.h>
  16 #include <linux/slab.h>
  17 #include <linux/list.h>
  18 #include <linux/module.h>
  19 #include <linux/completion.h>
  20 #include <linux/delay.h>
  21 #include <linux/cpu.h>
  22 #include <linux/hyperv.h>
  23 #include <asm/mshyperv.h>
  24
  25 #include "hyperv_vmbus.h"
  26
  27 static void init_vp_index(struct vmbus_channel *channel);
  28
  29 const struct vmbus_device vmbus_devs[] = {
  30         /* IDE */
  31         { .dev_type = HV_IDE,
  32           HV_IDE_GUID,
  33           .perf_device = true,
  34           .allowed_in_isolated = false,
  35         },
  36
  37         /* SCSI */
  38         { .dev_type = HV_SCSI,
  39           HV_SCSI_GUID,
  40           .perf_device = true,
  41           .allowed_in_isolated = true,
  42         },
  43
  44         /* Fibre Channel */
  45         { .dev_type = HV_FC,
  46           HV_SYNTHFC_GUID,
  47           .perf_device = true,
  48           .allowed_in_isolated = false,
  49         },
  50
  51         /* Synthetic NIC */
  52         { .dev_type = HV_NIC,
  53           HV_NIC_GUID,
  54           .perf_device = true,
  55           .allowed_in_isolated = true,
  56         },
  57
  58         /* Network Direct */
  59         { .dev_type = HV_ND,
  60           HV_ND_GUID,
  61           .perf_device = true,
  62           .allowed_in_isolated = false,
  63         },
  64
  65         /* PCIE */
  66         { .dev_type = HV_PCIE,
  67           HV_PCIE_GUID,
  68           .perf_device = false,
  69           .allowed_in_isolated = false,
  70         },
  71
  72         /* Synthetic Frame Buffer */
  73         { .dev_type = HV_FB,
  74           HV_SYNTHVID_GUID,
  75           .perf_device = false,
  76           .allowed_in_isolated = false,
  77         },
  78
  79         /* Synthetic Keyboard */
  80         { .dev_type = HV_KBD,
  81           HV_KBD_GUID,
  82           .perf_device = false,
  83           .allowed_in_isolated = false,
  84         },
  85
  86         /* Synthetic MOUSE */
  87         { .dev_type = HV_MOUSE,
  88           HV_MOUSE_GUID,
  89           .perf_device = false,
  90           .allowed_in_isolated = false,
  91         },
  92
  93         /* KVP */
  94         { .dev_type = HV_KVP,
  95           HV_KVP_GUID,
  96           .perf_device = false,
  97           .allowed_in_isolated = false,
  98         },
  99
 100         /* Time Synch */
 101         { .dev_type = HV_TS,
 102           HV_TS_GUID,
 103           .perf_device = false,
 104           .allowed_in_isolated = true,
 105         },
 106
 107         /* Heartbeat */
 108         { .dev_type = HV_HB,
 109           HV_HEART_BEAT_GUID,
 110           .perf_device = false,
 111           .allowed_in_isolated = true,
 112         },
 113
 114         /* Shutdown */
 115         { .dev_type = HV_SHUTDOWN,
 116           HV_SHUTDOWN_GUID,
 117           .perf_device = false,
 118           .allowed_in_isolated = true,
 119         },
 120
 121         /* File copy */
 122         { .dev_type = HV_FCOPY,
 123           HV_FCOPY_GUID,
 124           .perf_device = false,
 125           .allowed_in_isolated = false,
 126         },
 127
 128         /* Backup */
 129         { .dev_type = HV_BACKUP,
 130           HV_VSS_GUID,
 131           .perf_device = false,
 132           .allowed_in_isolated = false,
 133         },
 134
 135         /* Dynamic Memory */
 136         { .dev_type = HV_DM,
 137           HV_DM_GUID,
 138           .perf_device = false,
 139           .allowed_in_isolated = false,
 140         },
 141
 142         /* Unknown GUID */
 143         { .dev_type = HV_UNKNOWN,
 144           .perf_device = false,
 145           .allowed_in_isolated = false,
 146         },
 147 };
 148
 149 static const struct {
 150         guid_t guid;
 151 } vmbus_unsupported_devs[] = {
 152         { HV_AVMA1_GUID },
 153         { HV_AVMA2_GUID },
 154         { HV_RDV_GUID   },
 155 };
 156
 157 /*
 158  * The rescinded channel may be blocked waiting for a response from the host;
 159  * take care of that.
 160  */
 161 static void vmbus_rescind_cleanup(struct vmbus_channel *channel)
 162 {
 163         struct vmbus_channel_msginfo *msginfo;
 164         unsigned long flags;
 165
 166
 167         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
 168         channel->rescind = true;
 169         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
 170                                 msglistentry) {
 171
 172                 if (msginfo->waiting_channel == channel) {
 173                         complete(&msginfo->waitevent);
 174                         break;
 175                 }
 176         }
 177         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
 178 }
 179
 180 static bool is_unsupported_vmbus_devs(const guid_t *guid)
 181 {
 182         int i;
 183
 184         for (i = 0; i < ARRAY_SIZE(vmbus_unsupported_devs); i++)
 185                 if (guid_equal(guid, &vmbus_unsupported_devs[i].guid))
 186                         return true;
 187         return false;
 188 }
 189
 190 static u16 hv_get_dev_type(const struct vmbus_channel *channel)
 191 {
 192         const guid_t *guid = &channel->offermsg.offer.if_type;
 193         u16 i;
 194
 195         if (is_hvsock_channel(channel) || is_unsupported_vmbus_devs(guid))
 196                 return HV_UNKNOWN;
 197
 198         for (i = HV_IDE; i < HV_UNKNOWN; i++) {
 199                 if (guid_equal(guid, &vmbus_devs[i].guid))
 200                         return i;
 201         }
 202         pr_info("Unknown GUID: %pUl\n", guid);
 203         return i;
 204 }
 205
 206 /**
 207  * vmbus_prep_negotiate_resp() - Create default response for Negotiate message
 208  * @icmsghdrp: Pointer to msg header structure
 209  * @buf: Raw buffer channel data
 210  * @buflen: Length of the raw buffer channel data.
 211  * @fw_version: The framework versions we can support.
 212  * @fw_vercnt: The size of @fw_version.
 213  * @srv_version: The service versions we can support.
 214  * @srv_vercnt: The size of @srv_version.
 215  * @nego_fw_version: The selected framework version.
 216  * @nego_srv_version: The selected service version.
 217  *
 218  * Note: Versions are given in decreasing order.
 219  *
 220  * Set up and fill in default negotiate response message.
 221  * Mainly used by Hyper-V drivers.
 222  */
 223 bool vmbus_prep_negotiate_resp(struct icmsg_hdr *icmsghdrp, u8 *buf,
 224                                 u32 buflen, const int *fw_version, int fw_vercnt,
 225                                 const int *srv_version, int srv_vercnt,
 226                                 int *nego_fw_version, int *nego_srv_version)
 227 {
 228         int icframe_major, icframe_minor;
 229         int icmsg_major, icmsg_minor;
 230         int fw_major, fw_minor;
 231         int srv_major, srv_minor;
 232         int i, j;
 233         bool found_match = false;
 234         struct icmsg_negotiate *negop;
 235
 236         /* Check that there's enough space for icframe_vercnt, icmsg_vercnt */
 237         if (buflen < ICMSG_HDR + offsetof(struct icmsg_negotiate, reserved)) {
 238                 pr_err_ratelimited("Invalid icmsg negotiate\n");
 239                 return false;
 240         }
 241
 242         icmsghdrp->icmsgsize = 0x10;
 243         negop = (struct icmsg_negotiate *)&buf[ICMSG_HDR];
 244
 245         icframe_major = negop->icframe_vercnt;
 246         icframe_minor = 0;
 247
 248         icmsg_major = negop->icmsg_vercnt;
 249         icmsg_minor = 0;
 250
 251         /* Validate negop packet */
 252         if (icframe_major > IC_VERSION_NEGOTIATION_MAX_VER_COUNT ||
 253             icmsg_major > IC_VERSION_NEGOTIATION_MAX_VER_COUNT ||
 254             ICMSG_NEGOTIATE_PKT_SIZE(icframe_major, icmsg_major) > buflen) {
 255                 pr_err_ratelimited("Invalid icmsg negotiate - icframe_major: %u, icmsg_major: %u\n",
 256                                    icframe_major, icmsg_major);
 257                 goto fw_error;
 258         }
 259
 260         /*
 261          * Select the framework version number we will
 262          * support.
 263          */
 264
 265         for (i = 0; i < fw_vercnt; i++) {
 266                 fw_major = (fw_version[i] >> 16);
 267                 fw_minor = (fw_version[i] & 0xFFFF);
 268
 269                 for (j = 0; j < negop->icframe_vercnt; j++) {
 270                         if ((negop->icversion_data[j].major == fw_major) &&
 271                             (negop->icversion_data[j].minor == fw_minor)) {
 272                                 icframe_major = negop->icversion_data[j].major;
 273                                 icframe_minor = negop->icversion_data[j].minor;
 274                                 found_match = true;
 275                                 break;
 276                         }
 277                 }
 278
 279                 if (found_match)
 280                         break;
 281         }
 282
 283         if (!found_match)
 284                 goto fw_error;
 285
 286         found_match = false;
 287
 288         for (i = 0; i < srv_vercnt; i++) {
 289                 srv_major = (srv_version[i] >> 16);
 290                 srv_minor = (srv_version[i] & 0xFFFF);
 291
 292                 for (j = negop->icframe_vercnt;
 293                         (j < negop->icframe_vercnt + negop->icmsg_vercnt);
 294                         j++) {
 295
 296                         if ((negop->icversion_data[j].major == srv_major) &&
 297                                 (negop->icversion_data[j].minor == srv_minor)) {
 298
 299                                 icmsg_major = negop->icversion_data[j].major;
 300                                 icmsg_minor = negop->icversion_data[j].minor;
 301                                 found_match = true;
 302                                 break;
 303                         }
 304                 }
 305
 306                 if (found_match)
 307                         break;
 308         }
 309
 310         /*
 311          * Respond with the framework and service
 312          * version numbers we can support.
 313          */
 314
 315 fw_error:
 316         if (!found_match) {
 317                 negop->icframe_vercnt = 0;
 318                 negop->icmsg_vercnt = 0;
 319         } else {
 320                 negop->icframe_vercnt = 1;
 321                 negop->icmsg_vercnt = 1;
 322         }
 323
 324         if (nego_fw_version)
 325                 *nego_fw_version = (icframe_major << 16) | icframe_minor;
 326
 327         if (nego_srv_version)
 328                 *nego_srv_version = (icmsg_major << 16) | icmsg_minor;
 329
 330         negop->icversion_data[0].major = icframe_major;
 331         negop->icversion_data[0].minor = icframe_minor;
 332         negop->icversion_data[1].major = icmsg_major;
 333         negop->icversion_data[1].minor = icmsg_minor;
 334         return found_match;
 335 }
 336 EXPORT_SYMBOL_GPL(vmbus_prep_negotiate_resp);
 337
 338 /*
 339  * alloc_channel - Allocate and initialize a vmbus channel object
 340  */
 341 static struct vmbus_channel *alloc_channel(void)
 342 {
 343         struct vmbus_channel *channel;
 344
 345         channel = kzalloc(sizeof(*channel), GFP_ATOMIC);
 346         if (!channel)
 347                 return NULL;
 348
 349         spin_lock_init(&channel->sched_lock);
 350         init_completion(&channel->rescind_event);
 351
 352         INIT_LIST_HEAD(&channel->sc_list);
 353
 354         tasklet_init(&channel->callback_event,
 355                      vmbus_on_event, (unsigned long)channel);
 356
 357         hv_ringbuffer_pre_init(channel);
 358
 359         return channel;
 360 }
 361
 362 /*
 363  * free_channel - Release the resources used by the vmbus channel object
 364  */
 365 static void free_channel(struct vmbus_channel *channel)
 366 {
 367         tasklet_kill(&channel->callback_event);
 368         vmbus_remove_channel_attr_group(channel);
 369
 370         kobject_put(&channel->kobj);
 371 }
 372
 373 void vmbus_channel_map_relid(struct vmbus_channel *channel)
 374 {
 375         if (WARN_ON(channel->offermsg.child_relid >= MAX_CHANNEL_RELIDS))
 376                 return;
 377         /*
 378          * The mapping of the channel's relid is visible from the CPUs that
 379          * execute vmbus_chan_sched() by the time that vmbus_chan_sched() will
 380          * execute:
 381          *
 382          *  (a) In the "normal (i.e., not resuming from hibernation)" path,
 383          *      the full barrier in smp_store_mb() guarantees that the store
 384          *      is propagated to all CPUs before the add_channel_work work
 385          *      is queued.  In turn, add_channel_work is queued before the
 386          *      channel's ring buffer is allocated/initialized and the
 387          *      OPENCHANNEL message for the channel is sent in vmbus_open().
 388          *      Hyper-V won't start sending the interrupts for the channel
 389          *      before the OPENCHANNEL message is acked.  The memory barrier
 390          *      in vmbus_chan_sched() -> sync_test_and_clear_bit() ensures
 391          *      that vmbus_chan_sched() must find the channel's relid in
 392          *      recv_int_page before retrieving the channel pointer from the
 393          *      array of channels.
 394          *
 395          *  (b) In the "resuming from hibernation" path, the smp_store_mb()
 396          *      guarantees that the store is propagated to all CPUs before
 397          *      the VMBus connection is marked as ready for the resume event
 398          *      (cf. check_ready_for_resume_event()).  The interrupt handler
 399          *      of the VMBus driver and vmbus_chan_sched() can not run before
 400          *      vmbus_bus_resume() has completed execution (cf. resume_noirq).
 401          */
 402         smp_store_mb(
 403                 vmbus_connection.channels[channel->offermsg.child_relid],
 404                 channel);
 405 }
 406
 407 void vmbus_channel_unmap_relid(struct vmbus_channel *channel)
 408 {
 409         if (WARN_ON(channel->offermsg.child_relid >= MAX_CHANNEL_RELIDS))
 410                 return;
 411         WRITE_ONCE(
 412                 vmbus_connection.channels[channel->offermsg.child_relid],
 413                 NULL);
 414 }
 415
 416 static void vmbus_release_relid(u32 relid)
 417 {
 418         struct vmbus_channel_relid_released msg;
 419         int ret;
 420
 421         memset(&msg, 0, sizeof(struct vmbus_channel_relid_released));
 422         msg.child_relid = relid;
 423         msg.header.msgtype = CHANNELMSG_RELID_RELEASED;
 424         ret = vmbus_post_msg(&msg, sizeof(struct vmbus_channel_relid_released),
 425                              true);
 426
 427         trace_vmbus_release_relid(&msg, ret);
 428 }
 429
 430 void hv_process_channel_removal(struct vmbus_channel *channel)
 431 {
 432         lockdep_assert_held(&vmbus_connection.channel_mutex);
 433         BUG_ON(!channel->rescind);
 434
 435         /*
 436          * hv_process_channel_removal() could find INVALID_RELID only for
 437          * hv_sock channels.  See the inline comments in vmbus_onoffer().
 438          */
 439         WARN_ON(channel->offermsg.child_relid == INVALID_RELID &&
 440                 !is_hvsock_channel(channel));
 441
 442         /*
 443          * Upon suspend, an in-use hv_sock channel is removed from the array of
 444          * channels and the relid is invalidated.  After hibernation, when the
 445          * user-space appplication destroys the channel, it's unnecessary and
 446          * unsafe to remove the channel from the array of channels.  See also
 447          * the inline comments before the call of vmbus_release_relid() below.
 448          */
 449         if (channel->offermsg.child_relid != INVALID_RELID)
 450                 vmbus_channel_unmap_relid(channel);
 451
 452         if (channel->primary_channel == NULL)
 453                 list_del(&channel->listentry);
 454         else
 455                 list_del(&channel->sc_list);
 456
 457         /*
 458          * If this is a "perf" channel, updates the hv_numa_map[] masks so that
 459          * init_vp_index() can (re-)use the CPU.
 460          */
 461         if (hv_is_perf_channel(channel))
 462                 hv_clear_alloced_cpu(channel->target_cpu);
 463
 464         /*
 465          * Upon suspend, an in-use hv_sock channel is marked as "rescinded" and
 466          * the relid is invalidated; after hibernation, when the user-space app
 467          * destroys the channel, the relid is INVALID_RELID, and in this case
 468          * it's unnecessary and unsafe to release the old relid, since the same
 469          * relid can refer to a completely different channel now.
 470          */
 471         if (channel->offermsg.child_relid != INVALID_RELID)
 472                 vmbus_release_relid(channel->offermsg.child_relid);
 473
 474         free_channel(channel);
 475 }
 476
 477 void vmbus_free_channels(void)
 478 {
 479         struct vmbus_channel *channel, *tmp;
 480
 481         list_for_each_entry_safe(channel, tmp, &vmbus_connection.chn_list,
 482                 listentry) {
 483                 /* hv_process_channel_removal() needs this */
 484                 channel->rescind = true;
 485
 486                 vmbus_device_unregister(channel->device_obj);
 487         }
 488 }
 489
 490 /* Note: the function can run concurrently for primary/sub channels. */
 491 static void vmbus_add_channel_work(struct work_struct *work)
 492 {
 493         struct vmbus_channel *newchannel =
 494                 container_of(work, struct vmbus_channel, add_channel_work);
 495         struct vmbus_channel *primary_channel = newchannel->primary_channel;
 496         int ret;
 497
 498         /*
 499          * This state is used to indicate a successful open
 500          * so that when we do close the channel normally, we
 501          * can cleanup properly.
 502          */
 503         newchannel->state = CHANNEL_OPEN_STATE;
 504
 505         if (primary_channel != NULL) {
 506                 /* newchannel is a sub-channel. */
 507                 struct hv_device *dev = primary_channel->device_obj;
 508
 509                 if (vmbus_add_channel_kobj(dev, newchannel))
 510                         goto err_deq_chan;
 511
 512                 if (primary_channel->sc_creation_callback != NULL)
 513                         primary_channel->sc_creation_callback(newchannel);
 514
 515                 newchannel->probe_done = true;
 516                 return;
 517         }
 518
 519         /*
 520          * Start the process of binding the primary channel to the driver
 521          */
 522         newchannel->device_obj = vmbus_device_create(
 523                 &newchannel->offermsg.offer.if_type,
 524                 &newchannel->offermsg.offer.if_instance,
 525                 newchannel);
 526         if (!newchannel->device_obj)
 527                 goto err_deq_chan;
 528
 529         newchannel->device_obj->device_id = newchannel->device_id;
 530         /*
 531          * Add the new device to the bus. This will kick off device-driver
 532          * binding which eventually invokes the device driver's AddDevice()
 533          * method.
 534          */
 535         ret = vmbus_device_register(newchannel->device_obj);
 536
 537         if (ret != 0) {
 538                 pr_err("unable to add child device object (relid %d)\n",
 539                         newchannel->offermsg.child_relid);
 540                 kfree(newchannel->device_obj);
 541                 goto err_deq_chan;
 542         }
 543
 544         newchannel->probe_done = true;
 545         return;
 546
 547 err_deq_chan:
 548         mutex_lock(&vmbus_connection.channel_mutex);
 549
 550         /*
 551          * We need to set the flag, otherwise
 552          * vmbus_onoffer_rescind() can be blocked.
 553          */
 554         newchannel->probe_done = true;
 555
 556         if (primary_channel == NULL)
 557                 list_del(&newchannel->listentry);
 558         else
 559                 list_del(&newchannel->sc_list);
 560
 561         /* vmbus_process_offer() has mapped the channel. */
 562         vmbus_channel_unmap_relid(newchannel);
 563
 564         mutex_unlock(&vmbus_connection.channel_mutex);
 565
 566         vmbus_release_relid(newchannel->offermsg.child_relid);
 567
 568         free_channel(newchannel);
 569 }
 570
 571 /*
 572  * vmbus_process_offer - Process the offer by creating a channel/device
 573  * associated with this offer
 574  */
 575 static void vmbus_process_offer(struct vmbus_channel *newchannel)
 576 {
 577         struct vmbus_channel *channel;
 578         struct workqueue_struct *wq;
 579         bool fnew = true;
 580
 581         /*
 582          * Synchronize vmbus_process_offer() and CPU hotplugging:
 583          *
 584          * CPU1                         CPU2
 585          *
 586          * [vmbus_process_offer()]      [Hot removal of the CPU]
 587          *
 588          * CPU_READ_LOCK                CPUS_WRITE_LOCK
 589          * LOAD cpu_online_mask         SEARCH chn_list
 590          * STORE target_cpu             LOAD target_cpu
 591          * INSERT chn_list              STORE cpu_online_mask
 592          * CPUS_READ_UNLOCK             CPUS_WRITE_UNLOCK
 593          *
 594          * Forbids: CPU1's LOAD from *not* seing CPU2's STORE &&
 595          *              CPU2's SEARCH from *not* seeing CPU1's INSERT
 596          *
 597          * Forbids: CPU2's SEARCH from seeing CPU1's INSERT &&
 598          *              CPU2's LOAD from *not* seing CPU1's STORE
 599          */
 600         cpus_read_lock();
 601
 602         /*
 603          * Serializes the modifications of the chn_list list as well as
 604          * the accesses to next_numa_node_id in init_vp_index().
 605          */
 606         mutex_lock(&vmbus_connection.channel_mutex);
 607
 608         list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
 609                 if (guid_equal(&channel->offermsg.offer.if_type,
 610                                &newchannel->offermsg.offer.if_type) &&
 611                     guid_equal(&channel->offermsg.offer.if_instance,
 612                                &newchannel->offermsg.offer.if_instance)) {
 613                         fnew = false;
 614                         newchannel->primary_channel = channel;
 615                         break;
 616                 }
 617         }
 618
 619         init_vp_index(newchannel);
 620
 621         /* Remember the channels that should be cleaned up upon suspend. */
 622         if (is_hvsock_channel(newchannel) || is_sub_channel(newchannel))
 623                 atomic_inc(&vmbus_connection.nr_chan_close_on_suspend);
 624
 625         /*
 626          * Now that we have acquired the channel_mutex,
 627          * we can release the potentially racing rescind thread.
 628          */
 629         atomic_dec(&vmbus_connection.offer_in_progress);
 630
 631         if (fnew) {
 632                 list_add_tail(&newchannel->listentry,
 633                               &vmbus_connection.chn_list);
 634         } else {
 635                 /*
 636                  * Check to see if this is a valid sub-channel.
 637                  */
 638                 if (newchannel->offermsg.offer.sub_channel_index == 0) {
 639                         mutex_unlock(&vmbus_connection.channel_mutex);
 640                         /*
 641                          * Don't call free_channel(), because newchannel->kobj
 642                          * is not initialized yet.
 643                          */
 644                         kfree(newchannel);
 645                         WARN_ON_ONCE(1);
 646                         return;
 647                 }
 648                 /*
 649                  * Process the sub-channel.
 650                  */
 651                 list_add_tail(&newchannel->sc_list, &channel->sc_list);
 652         }
 653
 654         vmbus_channel_map_relid(newchannel);
 655
 656         mutex_unlock(&vmbus_connection.channel_mutex);
 657         cpus_read_unlock();
 658
 659         /*
 660          * vmbus_process_offer() mustn't call channel->sc_creation_callback()
 661          * directly for sub-channels, because sc_creation_callback() ->
 662          * vmbus_open() may never get the host's response to the
 663          * OPEN_CHANNEL message (the host may rescind a channel at any time,
 664          * e.g. in the case of hot removing a NIC), and vmbus_onoffer_rescind()
 665          * may not wake up the vmbus_open() as it's blocked due to a non-zero
 666          * vmbus_connection.offer_in_progress, and finally we have a deadlock.
 667          *
 668          * The above is also true for primary channels, if the related device
 669          * drivers use sync probing mode by default.
 670          *
 671          * And, usually the handling of primary channels and sub-channels can
 672          * depend on each other, so we should offload them to different
 673          * workqueues to avoid possible deadlock, e.g. in sync-probing mode,
 674          * NIC1's netvsc_subchan_work() can race with NIC2's netvsc_probe() ->
 675          * rtnl_lock(), and causes deadlock: the former gets the rtnl_lock
 676          * and waits for all the sub-channels to appear, but the latter
 677          * can't get the rtnl_lock and this blocks the handling of
 678          * sub-channels.
 679          */
 680         INIT_WORK(&newchannel->add_channel_work, vmbus_add_channel_work);
 681         wq = fnew ? vmbus_connection.handle_primary_chan_wq :
 682                     vmbus_connection.handle_sub_chan_wq;
 683         queue_work(wq, &newchannel->add_channel_work);
 684 }
 685
 686 /*
 687  * Check if CPUs used by other channels of the same device.
 688  * It should only be called by init_vp_index().
 689  */
 690 static bool hv_cpuself_used(u32 cpu, struct vmbus_channel *chn)
 691 {
 692         struct vmbus_channel *primary = chn->primary_channel;
 693         struct vmbus_channel *sc;
 694
 695         lockdep_assert_held(&vmbus_connection.channel_mutex);
 696
 697         if (!primary)
 698                 return false;
 699
 700         if (primary->target_cpu == cpu)
 701                 return true;
 702
 703         list_for_each_entry(sc, &primary->sc_list, sc_list)
 704                 if (sc != chn && sc->target_cpu == cpu)
 705                         return true;
 706
 707         return false;
 708 }
 709
 710 /*
 711  * We use this state to statically distribute the channel interrupt load.
 712  */
 713 static int next_numa_node_id;
 714
 715 /*
 716  * Starting with Win8, we can statically distribute the incoming
 717  * channel interrupt load by binding a channel to VCPU.
 718  *
 719  * For pre-win8 hosts or non-performance critical channels we assign the
 720  * VMBUS_CONNECT_CPU.
 721  *
 722  * Starting with win8, performance critical channels will be distributed
 723  * evenly among all the available NUMA nodes.  Once the node is assigned,
 724  * we will assign the CPU based on a simple round robin scheme.
 725  */
 726 static void init_vp_index(struct vmbus_channel *channel)
 727 {
 728         bool perf_chn = hv_is_perf_channel(channel);
 729         u32 i, ncpu = num_online_cpus();
 730         cpumask_var_t available_mask;
 731         struct cpumask *alloced_mask;
 732         u32 target_cpu;
 733         int numa_node;
 734
 735         if ((vmbus_proto_version == VERSION_WS2008) ||
 736             (vmbus_proto_version == VERSION_WIN7) || (!perf_chn) ||
 737             !alloc_cpumask_var(&available_mask, GFP_KERNEL)) {
 738                 /*
 739                  * Prior to win8, all channel interrupts are
 740                  * delivered on VMBUS_CONNECT_CPU.
 741                  * Also if the channel is not a performance critical
 742                  * channel, bind it to VMBUS_CONNECT_CPU.
 743                  * In case alloc_cpumask_var() fails, bind it to
 744                  * VMBUS_CONNECT_CPU.
 745                  */
 746                 channel->target_cpu = VMBUS_CONNECT_CPU;
 747                 if (perf_chn)
 748                         hv_set_alloced_cpu(VMBUS_CONNECT_CPU);
 749                 return;
 750         }
 751
 752         for (i = 1; i <= ncpu + 1; i++) {
 753                 while (true) {
 754                         numa_node = next_numa_node_id++;
 755                         if (numa_node == nr_node_ids) {
 756                                 next_numa_node_id = 0;
 757                                 continue;
 758                         }
 759                         if (cpumask_empty(cpumask_of_node(numa_node)))
 760                                 continue;
 761                         break;
 762                 }
 763                 alloced_mask = &hv_context.hv_numa_map[numa_node];
 764
 765                 if (cpumask_weight(alloced_mask) ==
 766                     cpumask_weight(cpumask_of_node(numa_node))) {
 767                         /*
 768                          * We have cycled through all the CPUs in the node;
 769                          * reset the alloced map.
 770                          */
 771                         cpumask_clear(alloced_mask);
 772                 }
 773
 774                 cpumask_xor(available_mask, alloced_mask,
 775                             cpumask_of_node(numa_node));
 776
 777                 target_cpu = cpumask_first(available_mask);
 778                 cpumask_set_cpu(target_cpu, alloced_mask);
 779
 780                 if (channel->offermsg.offer.sub_channel_index >= ncpu ||
 781                     i > ncpu || !hv_cpuself_used(target_cpu, channel))
 782                         break;
 783         }
 784
 785         channel->target_cpu = target_cpu;
 786
 787         free_cpumask_var(available_mask);
 788 }
 789
 790 #define UNLOAD_DELAY_UNIT_MS    10              /* 10 milliseconds */
 791 #define UNLOAD_WAIT_MS          (100*1000)      /* 100 seconds */
 792 #define UNLOAD_WAIT_LOOPS       (UNLOAD_WAIT_MS/UNLOAD_DELAY_UNIT_MS)
 793 #define UNLOAD_MSG_MS           (5*1000)        /* Every 5 seconds */
 794 #define UNLOAD_MSG_LOOPS        (UNLOAD_MSG_MS/UNLOAD_DELAY_UNIT_MS)
 795
 796 static void vmbus_wait_for_unload(void)
 797 {
 798         int cpu;
 799         void *page_addr;
 800         struct hv_message *msg;
 801         struct vmbus_channel_message_header *hdr;
 802         u32 message_type, i;
 803
 804         /*
 805          * CHANNELMSG_UNLOAD_RESPONSE is always delivered to the CPU which was
 806          * used for initial contact or to CPU0 depending on host version. When
 807          * we're crashing on a different CPU let's hope that IRQ handler on
 808          * the cpu which receives CHANNELMSG_UNLOAD_RESPONSE is still
 809          * functional and vmbus_unload_response() will complete
 810          * vmbus_connection.unload_event. If not, the last thing we can do is
 811          * read message pages for all CPUs directly.
 812          *
 813          * Wait up to 100 seconds since an Azure host must writeback any dirty
 814          * data in its disk cache before the VMbus UNLOAD request will
 815          * complete. This flushing has been empirically observed to take up
 816          * to 50 seconds in cases with a lot of dirty data, so allow additional
 817          * leeway and for inaccuracies in mdelay(). But eventually time out so
 818          * that the panic path can't get hung forever in case the response
 819          * message isn't seen.
 820          */
 821         for (i = 1; i <= UNLOAD_WAIT_LOOPS; i++) {
 822                 if (completion_done(&vmbus_connection.unload_event))
 823                         goto completed;
 824
 825                 for_each_online_cpu(cpu) {
 826                         struct hv_per_cpu_context *hv_cpu
 827                                 = per_cpu_ptr(hv_context.cpu_context, cpu);
 828
 829                         page_addr = hv_cpu->synic_message_page;
 830                         msg = (struct hv_message *)page_addr
 831                                 + VMBUS_MESSAGE_SINT;
 832
 833                         message_type = READ_ONCE(msg->header.message_type);
 834                         if (message_type == HVMSG_NONE)
 835                                 continue;
 836
 837                         hdr = (struct vmbus_channel_message_header *)
 838                                 msg->u.payload;
 839
 840                         if (hdr->msgtype == CHANNELMSG_UNLOAD_RESPONSE)
 841                                 complete(&vmbus_connection.unload_event);
 842
 843                         vmbus_signal_eom(msg, message_type);
 844                 }
 845
 846                 /*
 847                  * Give a notice periodically so someone watching the
 848                  * serial output won't think it is completely hung.
 849                  */
 850                 if (!(i % UNLOAD_MSG_LOOPS))
 851                         pr_notice("Waiting for VMBus UNLOAD to complete\n");
 852
 853                 mdelay(UNLOAD_DELAY_UNIT_MS);
 854         }
 855         pr_err("Continuing even though VMBus UNLOAD did not complete\n");
 856
 857 completed:
 858         /*
 859          * We're crashing and already got the UNLOAD_RESPONSE, cleanup all
 860          * maybe-pending messages on all CPUs to be able to receive new
 861          * messages after we reconnect.
 862          */
 863         for_each_online_cpu(cpu) {
 864                 struct hv_per_cpu_context *hv_cpu
 865                         = per_cpu_ptr(hv_context.cpu_context, cpu);
 866
 867                 page_addr = hv_cpu->synic_message_page;
 868                 msg = (struct hv_message *)page_addr + VMBUS_MESSAGE_SINT;
 869                 msg->header.message_type = HVMSG_NONE;
 870         }
 871 }
 872
 873 /*
 874  * vmbus_unload_response - Handler for the unload response.
 875  */
 876 static void vmbus_unload_response(struct vmbus_channel_message_header *hdr)
 877 {
 878         /*
 879          * This is a global event; just wakeup the waiting thread.
 880          * Once we successfully unload, we can cleanup the monitor state.
 881          *
 882          * NB.  A malicious or compromised Hyper-V could send a spurious
 883          * message of type CHANNELMSG_UNLOAD_RESPONSE, and trigger a call
 884          * of the complete() below.  Make sure that unload_event has been
 885          * initialized by the time this complete() is executed.
 886          */
 887         complete(&vmbus_connection.unload_event);
 888 }
 889
 890 void vmbus_initiate_unload(bool crash)
 891 {
 892         struct vmbus_channel_message_header hdr;
 893
 894         if (xchg(&vmbus_connection.conn_state, DISCONNECTED) == DISCONNECTED)
 895                 return;
 896
 897         /* Pre-Win2012R2 hosts don't support reconnect */
 898         if (vmbus_proto_version < VERSION_WIN8_1)
 899                 return;
 900
 901         reinit_completion(&vmbus_connection.unload_event);
 902         memset(&hdr, 0, sizeof(struct vmbus_channel_message_header));
 903         hdr.msgtype = CHANNELMSG_UNLOAD;
 904         vmbus_post_msg(&hdr, sizeof(struct vmbus_channel_message_header),
 905                        !crash);
 906
 907         /*
 908          * vmbus_initiate_unload() is also called on crash and the crash can be
 909          * happening in an interrupt context, where scheduling is impossible.
 910          */
 911         if (!crash)
 912                 wait_for_completion(&vmbus_connection.unload_event);
 913         else
 914                 vmbus_wait_for_unload();
 915 }
 916
 917 static void check_ready_for_resume_event(void)
 918 {
 919         /*
 920          * If all the old primary channels have been fixed up, then it's safe
 921          * to resume.
 922          */
 923         if (atomic_dec_and_test(&vmbus_connection.nr_chan_fixup_on_resume))
 924                 complete(&vmbus_connection.ready_for_resume_event);
 925 }
 926
 927 static void vmbus_setup_channel_state(struct vmbus_channel *channel,
 928                                       struct vmbus_channel_offer_channel *offer)
 929 {
 930         /*
 931          * Setup state for signalling the host.
 932          */
 933         channel->sig_event = VMBUS_EVENT_CONNECTION_ID;
 934
 935         if (vmbus_proto_version != VERSION_WS2008) {
 936                 channel->is_dedicated_interrupt =
 937                                 (offer->is_dedicated_interrupt != 0);
 938                 channel->sig_event = offer->connection_id;
 939         }
 940
 941         memcpy(&channel->offermsg, offer,
 942                sizeof(struct vmbus_channel_offer_channel));
 943         channel->monitor_grp = (u8)offer->monitorid / 32;
 944         channel->monitor_bit = (u8)offer->monitorid % 32;
 945         channel->device_id = hv_get_dev_type(channel);
 946 }
 947
 948 /*
 949  * find_primary_channel_by_offer - Get the channel object given the new offer.
 950  * This is only used in the resume path of hibernation.
 951  */
 952 static struct vmbus_channel *
 953 find_primary_channel_by_offer(const struct vmbus_channel_offer_channel *offer)
 954 {
 955         struct vmbus_channel *channel = NULL, *iter;
 956         const guid_t *inst1, *inst2;
 957
 958         /* Ignore sub-channel offers. */
 959         if (offer->offer.sub_channel_index != 0)
 960                 return NULL;
 961
 962         mutex_lock(&vmbus_connection.channel_mutex);
 963
 964         list_for_each_entry(iter, &vmbus_connection.chn_list, listentry) {
 965                 inst1 = &iter->offermsg.offer.if_instance;
 966                 inst2 = &offer->offer.if_instance;
 967
 968                 if (guid_equal(inst1, inst2)) {
 969                         channel = iter;
 970                         break;
 971                 }
 972         }
 973
 974         mutex_unlock(&vmbus_connection.channel_mutex);
 975
 976         return channel;
 977 }
 978
 979 static bool vmbus_is_valid_device(const guid_t *guid)
 980 {
 981         u16 i;
 982
 983         if (!hv_is_isolation_supported())
 984                 return true;
 985
 986         for (i = 0; i < ARRAY_SIZE(vmbus_devs); i++) {
 987                 if (guid_equal(guid, &vmbus_devs[i].guid))
 988                         return vmbus_devs[i].allowed_in_isolated;
 989         }
 990         return false;
 991 }
 992
 993 /*
 994  * vmbus_onoffer - Handler for channel offers from vmbus in parent partition.
 995  *
 996  */
 997 static void vmbus_onoffer(struct vmbus_channel_message_header *hdr)
 998 {
 999         struct vmbus_channel_offer_channel *offer;
1000         struct vmbus_channel *oldchannel, *newchannel;
1001         size_t offer_sz;
1002
1003         offer = (struct vmbus_channel_offer_channel *)hdr;
1004
1005         trace_vmbus_onoffer(offer);
1006
1007         if (!vmbus_is_valid_device(&offer->offer.if_type)) {
1008                 pr_err_ratelimited("Invalid offer %d from the host supporting isolation\n",
1009                                    offer->child_relid);
1010                 atomic_dec(&vmbus_connection.offer_in_progress);
1011                 return;
1012         }
1013
1014         oldchannel = find_primary_channel_by_offer(offer);
1015
1016         if (oldchannel != NULL) {
1017                 /*
1018                  * We're resuming from hibernation: all the sub-channel and
1019                  * hv_sock channels we had before the hibernation should have
1020                  * been cleaned up, and now we must be seeing a re-offered
1021                  * primary channel that we had before the hibernation.
1022                  */
1023
1024                 /*
1025                  * { Initially: channel relid = INVALID_RELID,
1026                  *              channels[valid_relid] = NULL }
1027                  *
1028                  * CPU1                                 CPU2
1029                  *
1030                  * [vmbus_onoffer()]                    [vmbus_device_release()]
1031                  *
1032                  * LOCK channel_mutex                   LOCK channel_mutex
1033                  * STORE channel relid = valid_relid    LOAD r1 = channel relid
1034                  * MAP_RELID channel                    if (r1 != INVALID_RELID)
1035                  * UNLOCK channel_mutex                   UNMAP_RELID channel
1036                  *                                      UNLOCK channel_mutex
1037                  *
1038                  * Forbids: r1 == valid_relid &&
1039                  *              channels[valid_relid] == channel
1040                  *
1041                  * Note.  r1 can be INVALID_RELID only for an hv_sock channel.
1042                  * None of the hv_sock channels which were present before the
1043                  * suspend are re-offered upon the resume.  See the WARN_ON()
1044                  * in hv_process_channel_removal().
1045                  */
1046                 mutex_lock(&vmbus_connection.channel_mutex);
1047
1048                 atomic_dec(&vmbus_connection.offer_in_progress);
1049
1050                 WARN_ON(oldchannel->offermsg.child_relid != INVALID_RELID);
1051                 /* Fix up the relid. */
1052                 oldchannel->offermsg.child_relid = offer->child_relid;
1053
1054                 offer_sz = sizeof(*offer);
1055                 if (memcmp(offer, &oldchannel->offermsg, offer_sz) != 0) {
1056                         /*
1057                          * This is not an error, since the host can also change
1058                          * the other field(s) of the offer, e.g. on WS RS5
1059                          * (Build 17763), the offer->connection_id of the
1060                          * Mellanox VF vmbus device can change when the host
1061                          * reoffers the device upon resume.
1062                          */
1063                         pr_debug("vmbus offer changed: relid=%d\n",
1064                                  offer->child_relid);
1065
1066                         print_hex_dump_debug("Old vmbus offer: ",
1067                                              DUMP_PREFIX_OFFSET, 16, 4,
1068                                              &oldchannel->offermsg, offer_sz,
1069                                              false);
1070                         print_hex_dump_debug("New vmbus offer: ",
1071                                              DUMP_PREFIX_OFFSET, 16, 4,
1072                                              offer, offer_sz, false);
1073
1074                         /* Fix up the old channel. */
1075                         vmbus_setup_channel_state(oldchannel, offer);
1076                 }
1077
1078                 /* Add the channel back to the array of channels. */
1079                 vmbus_channel_map_relid(oldchannel);
1080                 check_ready_for_resume_event();
1081
1082                 mutex_unlock(&vmbus_connection.channel_mutex);
1083                 return;
1084         }
1085
1086         /* Allocate the channel object and save this offer. */
1087         newchannel = alloc_channel();
1088         if (!newchannel) {
1089                 vmbus_release_relid(offer->child_relid);
1090                 atomic_dec(&vmbus_connection.offer_in_progress);
1091                 pr_err("Unable to allocate channel object\n");
1092                 return;
1093         }
1094
1095         vmbus_setup_channel_state(newchannel, offer);
1096
1097         vmbus_process_offer(newchannel);
1098 }
1099
1100 static void check_ready_for_suspend_event(void)
1101 {
1102         /*
1103          * If all the sub-channels or hv_sock channels have been cleaned up,
1104          * then it's safe to suspend.
1105          */
1106         if (atomic_dec_and_test(&vmbus_connection.nr_chan_close_on_suspend))
1107                 complete(&vmbus_connection.ready_for_suspend_event);
1108 }
1109
1110 /*
1111  * vmbus_onoffer_rescind - Rescind offer handler.
1112  *
1113  * We queue a work item to process this offer synchronously
1114  */
1115 static void vmbus_onoffer_rescind(struct vmbus_channel_message_header *hdr)
1116 {
1117         struct vmbus_channel_rescind_offer *rescind;
1118         struct vmbus_channel *channel;
1119         struct device *dev;
1120         bool clean_up_chan_for_suspend;
1121
1122         rescind = (struct vmbus_channel_rescind_offer *)hdr;
1123
1124         trace_vmbus_onoffer_rescind(rescind);
1125
1126         /*
1127          * The offer msg and the corresponding rescind msg
1128          * from the host are guranteed to be ordered -
1129          * offer comes in first and then the rescind.
1130          * Since we process these events in work elements,
1131          * and with preemption, we may end up processing
1132          * the events out of order.  We rely on the synchronization
1133          * provided by offer_in_progress and by channel_mutex for
1134          * ordering these events:
1135          *
1136          * { Initially: offer_in_progress = 1 }
1137          *
1138          * CPU1                         CPU2
1139          *
1140          * [vmbus_onoffer()]            [vmbus_onoffer_rescind()]
1141          *
1142          * LOCK channel_mutex           WAIT_ON offer_in_progress == 0
1143          * DECREMENT offer_in_progress  LOCK channel_mutex
1144          * STORE channels[]             LOAD channels[]
1145          * UNLOCK channel_mutex         UNLOCK channel_mutex
1146          *
1147          * Forbids: CPU2's LOAD from *not* seeing CPU1's STORE
1148          */
1149
1150         while (atomic_read(&vmbus_connection.offer_in_progress) != 0) {
1151                 /*
1152                  * We wait here until any channel offer is currently
1153                  * being processed.
1154                  */
1155                 msleep(1);
1156         }
1157
1158         mutex_lock(&vmbus_connection.channel_mutex);
1159         channel = relid2channel(rescind->child_relid);
1160         if (channel != NULL) {
1161                 /*
1162                  * Guarantee that no other instance of vmbus_onoffer_rescind()
1163                  * has got a reference to the channel object.  Synchronize on
1164                  * &vmbus_connection.channel_mutex.
1165                  */
1166                 if (channel->rescind_ref) {
1167                         mutex_unlock(&vmbus_connection.channel_mutex);
1168                         return;
1169                 }
1170                 channel->rescind_ref = true;
1171         }
1172         mutex_unlock(&vmbus_connection.channel_mutex);
1173
1174         if (channel == NULL) {
1175                 /*
1176                  * We failed in processing the offer message;
1177                  * we would have cleaned up the relid in that
1178                  * failure path.
1179                  */
1180                 return;
1181         }
1182
1183         clean_up_chan_for_suspend = is_hvsock_channel(channel) ||
1184                                     is_sub_channel(channel);
1185         /*
1186          * Before setting channel->rescind in vmbus_rescind_cleanup(), we
1187          * should make sure the channel callback is not running any more.
1188          */
1189         vmbus_reset_channel_cb(channel);
1190
1191         /*
1192          * Now wait for offer handling to complete.
1193          */
1194         vmbus_rescind_cleanup(channel);
1195         while (READ_ONCE(channel->probe_done) == false) {
1196                 /*
1197                  * We wait here until any channel offer is currently
1198                  * being processed.
1199                  */
1200                 msleep(1);
1201         }
1202
1203         /*
1204          * At this point, the rescind handling can proceed safely.
1205          */
1206
1207         if (channel->device_obj) {
1208                 if (channel->chn_rescind_callback) {
1209                         channel->chn_rescind_callback(channel);
1210
1211                         if (clean_up_chan_for_suspend)
1212                                 check_ready_for_suspend_event();
1213
1214                         return;
1215                 }
1216                 /*
1217                  * We will have to unregister this device from the
1218                  * driver core.
1219                  */
1220                 dev = get_device(&channel->device_obj->device);
1221                 if (dev) {
1222                         vmbus_device_unregister(channel->device_obj);
1223                         put_device(dev);
1224                 }
1225         } else if (channel->primary_channel != NULL) {
1226                 /*
1227                  * Sub-channel is being rescinded. Following is the channel
1228                  * close sequence when initiated from the driveri (refer to
1229                  * vmbus_close() for details):
1230                  * 1. Close all sub-channels first
1231                  * 2. Then close the primary channel.
1232                  */
1233                 mutex_lock(&vmbus_connection.channel_mutex);
1234                 if (channel->state == CHANNEL_OPEN_STATE) {
1235                         /*
1236                          * The channel is currently not open;
1237                          * it is safe for us to cleanup the channel.
1238                          */
1239                         hv_process_channel_removal(channel);
1240                 } else {
1241                         complete(&channel->rescind_event);
1242                 }
1243                 mutex_unlock(&vmbus_connection.channel_mutex);
1244         }
1245
1246         /* The "channel" may have been freed. Do not access it any longer. */
1247
1248         if (clean_up_chan_for_suspend)
1249                 check_ready_for_suspend_event();
1250 }
1251
1252 void vmbus_hvsock_device_unregister(struct vmbus_channel *channel)
1253 {
1254         BUG_ON(!is_hvsock_channel(channel));
1255
1256         /* We always get a rescind msg when a connection is closed. */
1257         while (!READ_ONCE(channel->probe_done) || !READ_ONCE(channel->rescind))
1258                 msleep(1);
1259
1260         vmbus_device_unregister(channel->device_obj);
1261 }
1262 EXPORT_SYMBOL_GPL(vmbus_hvsock_device_unregister);
1263
1264
1265 /*
1266  * vmbus_onoffers_delivered -
1267  * This is invoked when all offers have been delivered.
1268  *
1269  * Nothing to do here.
1270  */
1271 static void vmbus_onoffers_delivered(
1272                         struct vmbus_channel_message_header *hdr)
1273 {
1274 }
1275
1276 /*
1277  * vmbus_onopen_result - Open result handler.
1278  *
1279  * This is invoked when we received a response to our channel open request.
1280  * Find the matching request, copy the response and signal the requesting
1281  * thread.
1282  */
1283 static void vmbus_onopen_result(struct vmbus_channel_message_header *hdr)
1284 {
1285         struct vmbus_channel_open_result *result;
1286         struct vmbus_channel_msginfo *msginfo;
1287         struct vmbus_channel_message_header *requestheader;
1288         struct vmbus_channel_open_channel *openmsg;
1289         unsigned long flags;
1290
1291         result = (struct vmbus_channel_open_result *)hdr;
1292
1293         trace_vmbus_onopen_result(result);
1294
1295         /*
1296          * Find the open msg, copy the result and signal/unblock the wait event
1297          */
1298         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1299
1300         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1301                                 msglistentry) {
1302                 requestheader =
1303                         (struct vmbus_channel_message_header *)msginfo->msg;
1304
1305                 if (requestheader->msgtype == CHANNELMSG_OPENCHANNEL) {
1306                         openmsg =
1307                         (struct vmbus_channel_open_channel *)msginfo->msg;
1308                         if (openmsg->child_relid == result->child_relid &&
1309                             openmsg->openid == result->openid) {
1310                                 memcpy(&msginfo->response.open_result,
1311                                        result,
1312                                        sizeof(
1313                                         struct vmbus_channel_open_result));
1314                                 complete(&msginfo->waitevent);
1315                                 break;
1316                         }
1317                 }
1318         }
1319         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1320 }
1321
1322 /*
1323  * vmbus_ongpadl_created - GPADL created handler.
1324  *
1325  * This is invoked when we received a response to our gpadl create request.
1326  * Find the matching request, copy the response and signal the requesting
1327  * thread.
1328  */
1329 static void vmbus_ongpadl_created(struct vmbus_channel_message_header *hdr)
1330 {
1331         struct vmbus_channel_gpadl_created *gpadlcreated;
1332         struct vmbus_channel_msginfo *msginfo;
1333         struct vmbus_channel_message_header *requestheader;
1334         struct vmbus_channel_gpadl_header *gpadlheader;
1335         unsigned long flags;
1336
1337         gpadlcreated = (struct vmbus_channel_gpadl_created *)hdr;
1338
1339         trace_vmbus_ongpadl_created(gpadlcreated);
1340
1341         /*
1342          * Find the establish msg, copy the result and signal/unblock the wait
1343          * event
1344          */
1345         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1346
1347         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1348                                 msglistentry) {
1349                 requestheader =
1350                         (struct vmbus_channel_message_header *)msginfo->msg;
1351
1352                 if (requestheader->msgtype == CHANNELMSG_GPADL_HEADER) {
1353                         gpadlheader =
1354                         (struct vmbus_channel_gpadl_header *)requestheader;
1355
1356                         if ((gpadlcreated->child_relid ==
1357                              gpadlheader->child_relid) &&
1358                             (gpadlcreated->gpadl == gpadlheader->gpadl)) {
1359                                 memcpy(&msginfo->response.gpadl_created,
1360                                        gpadlcreated,
1361                                        sizeof(
1362                                         struct vmbus_channel_gpadl_created));
1363                                 complete(&msginfo->waitevent);
1364                                 break;
1365                         }
1366                 }
1367         }
1368         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1369 }
1370
1371 /*
1372  * vmbus_onmodifychannel_response - Modify Channel response handler.
1373  *
1374  * This is invoked when we received a response to our channel modify request.
1375  * Find the matching request, copy the response and signal the requesting thread.
1376  */
1377 static void vmbus_onmodifychannel_response(struct vmbus_channel_message_header *hdr)
1378 {
1379         struct vmbus_channel_modifychannel_response *response;
1380         struct vmbus_channel_msginfo *msginfo;
1381         unsigned long flags;
1382
1383         response = (struct vmbus_channel_modifychannel_response *)hdr;
1384
1385         trace_vmbus_onmodifychannel_response(response);
1386
1387         /*
1388          * Find the modify msg, copy the response and signal/unblock the wait event.
1389          */
1390         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1391
1392         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list, msglistentry) {
1393                 struct vmbus_channel_message_header *responseheader =
1394                                 (struct vmbus_channel_message_header *)msginfo->msg;
1395
1396                 if (responseheader->msgtype == CHANNELMSG_MODIFYCHANNEL) {
1397                         struct vmbus_channel_modifychannel *modifymsg;
1398
1399                         modifymsg = (struct vmbus_channel_modifychannel *)msginfo->msg;
1400                         if (modifymsg->child_relid == response->child_relid) {
1401                                 memcpy(&msginfo->response.modify_response, response,
1402                                        sizeof(*response));
1403                                 complete(&msginfo->waitevent);
1404                                 break;
1405                         }
1406                 }
1407         }
1408         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1409 }
1410
1411 /*
1412  * vmbus_ongpadl_torndown - GPADL torndown handler.
1413  *
1414  * This is invoked when we received a response to our gpadl teardown request.
1415  * Find the matching request, copy the response and signal the requesting
1416  * thread.
1417  */
1418 static void vmbus_ongpadl_torndown(
1419                         struct vmbus_channel_message_header *hdr)
1420 {
1421         struct vmbus_channel_gpadl_torndown *gpadl_torndown;
1422         struct vmbus_channel_msginfo *msginfo;
1423         struct vmbus_channel_message_header *requestheader;
1424         struct vmbus_channel_gpadl_teardown *gpadl_teardown;
1425         unsigned long flags;
1426
1427         gpadl_torndown = (struct vmbus_channel_gpadl_torndown *)hdr;
1428
1429         trace_vmbus_ongpadl_torndown(gpadl_torndown);
1430
1431         /*
1432          * Find the open msg, copy the result and signal/unblock the wait event
1433          */
1434         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1435
1436         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1437                                 msglistentry) {
1438                 requestheader =
1439                         (struct vmbus_channel_message_header *)msginfo->msg;
1440
1441                 if (requestheader->msgtype == CHANNELMSG_GPADL_TEARDOWN) {
1442                         gpadl_teardown =
1443                         (struct vmbus_channel_gpadl_teardown *)requestheader;
1444
1445                         if (gpadl_torndown->gpadl == gpadl_teardown->gpadl) {
1446                                 memcpy(&msginfo->response.gpadl_torndown,
1447                                        gpadl_torndown,
1448                                        sizeof(
1449                                         struct vmbus_channel_gpadl_torndown));
1450                                 complete(&msginfo->waitevent);
1451                                 break;
1452                         }
1453                 }
1454         }
1455         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1456 }
1457
1458 /*
1459  * vmbus_onversion_response - Version response handler
1460  *
1461  * This is invoked when we received a response to our initiate contact request.
1462  * Find the matching request, copy the response and signal the requesting
1463  * thread.
1464  */
1465 static void vmbus_onversion_response(
1466                 struct vmbus_channel_message_header *hdr)
1467 {
1468         struct vmbus_channel_msginfo *msginfo;
1469         struct vmbus_channel_message_header *requestheader;
1470         struct vmbus_channel_version_response *version_response;
1471         unsigned long flags;
1472
1473         version_response = (struct vmbus_channel_version_response *)hdr;
1474
1475         trace_vmbus_onversion_response(version_response);
1476
1477         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1478
1479         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1480                                 msglistentry) {
1481                 requestheader =
1482                         (struct vmbus_channel_message_header *)msginfo->msg;
1483
1484                 if (requestheader->msgtype ==
1485                     CHANNELMSG_INITIATE_CONTACT) {
1486                         memcpy(&msginfo->response.version_response,
1487                               version_response,
1488                               sizeof(struct vmbus_channel_version_response));
1489                         complete(&msginfo->waitevent);
1490                 }
1491         }
1492         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1493 }
1494
1495 /* Channel message dispatch table */
1496 const struct vmbus_channel_message_table_entry
1497 channel_message_table[CHANNELMSG_COUNT] = {
1498         { CHANNELMSG_INVALID,                   0, NULL, 0},
1499         { CHANNELMSG_OFFERCHANNEL,              0, vmbus_onoffer,
1500                 sizeof(struct vmbus_channel_offer_channel)},
1501         { CHANNELMSG_RESCIND_CHANNELOFFER,      0, vmbus_onoffer_rescind,
1502                 sizeof(struct vmbus_channel_rescind_offer) },
1503         { CHANNELMSG_REQUESTOFFERS,             0, NULL, 0},
1504         { CHANNELMSG_ALLOFFERS_DELIVERED,       1, vmbus_onoffers_delivered, 0},
1505         { CHANNELMSG_OPENCHANNEL,               0, NULL, 0},
1506         { CHANNELMSG_OPENCHANNEL_RESULT,        1, vmbus_onopen_result,
1507                 sizeof(struct vmbus_channel_open_result)},
1508         { CHANNELMSG_CLOSECHANNEL,              0, NULL, 0},
1509         { CHANNELMSG_GPADL_HEADER,              0, NULL, 0},
1510         { CHANNELMSG_GPADL_BODY,                0, NULL, 0},
1511         { CHANNELMSG_GPADL_CREATED,             1, vmbus_ongpadl_created,
1512                 sizeof(struct vmbus_channel_gpadl_created)},
1513         { CHANNELMSG_GPADL_TEARDOWN,            0, NULL, 0},
1514         { CHANNELMSG_GPADL_TORNDOWN,            1, vmbus_ongpadl_torndown,
1515                 sizeof(struct vmbus_channel_gpadl_torndown) },
1516         { CHANNELMSG_RELID_RELEASED,            0, NULL, 0},
1517         { CHANNELMSG_INITIATE_CONTACT,          0, NULL, 0},
1518         { CHANNELMSG_VERSION_RESPONSE,          1, vmbus_onversion_response,
1519                 sizeof(struct vmbus_channel_version_response)},
1520         { CHANNELMSG_UNLOAD,                    0, NULL, 0},
1521         { CHANNELMSG_UNLOAD_RESPONSE,           1, vmbus_unload_response, 0},
1522         { CHANNELMSG_18,                        0, NULL, 0},
1523         { CHANNELMSG_19,                        0, NULL, 0},
1524         { CHANNELMSG_20,                        0, NULL, 0},
1525         { CHANNELMSG_TL_CONNECT_REQUEST,        0, NULL, 0},
1526         { CHANNELMSG_MODIFYCHANNEL,             0, NULL, 0},
1527         { CHANNELMSG_TL_CONNECT_RESULT,         0, NULL, 0},
1528         { CHANNELMSG_MODIFYCHANNEL_RESPONSE,    1, vmbus_onmodifychannel_response,
1529                 sizeof(struct vmbus_channel_modifychannel_response)},
1530 };
1531
1532 /*
1533  * vmbus_onmessage - Handler for channel protocol messages.
1534  *
1535  * This is invoked in the vmbus worker thread context.
1536  */
1537 void vmbus_onmessage(struct vmbus_channel_message_header *hdr)
1538 {
1539         trace_vmbus_on_message(hdr);
1540
1541         /*
1542          * vmbus_on_msg_dpc() makes sure the hdr->msgtype here can not go
1543          * out of bound and the message_handler pointer can not be NULL.
1544          */
1545         channel_message_table[hdr->msgtype].message_handler(hdr);
1546 }
1547
1548 /*
1549  * vmbus_request_offers - Send a request to get all our pending offers.
1550  */
1551 int vmbus_request_offers(void)
1552 {
1553         struct vmbus_channel_message_header *msg;
1554         struct vmbus_channel_msginfo *msginfo;
1555         int ret;
1556
1557         msginfo = kmalloc(sizeof(*msginfo) +
1558                           sizeof(struct vmbus_channel_message_header),
1559                           GFP_KERNEL);
1560         if (!msginfo)
1561                 return -ENOMEM;
1562
1563         msg = (struct vmbus_channel_message_header *)msginfo->msg;
1564
1565         msg->msgtype = CHANNELMSG_REQUESTOFFERS;
1566
1567         ret = vmbus_post_msg(msg, sizeof(struct vmbus_channel_message_header),
1568                              true);
1569
1570         trace_vmbus_request_offers(ret);
1571
1572         if (ret != 0) {
1573                 pr_err("Unable to request offers - %d\n", ret);
1574
1575                 goto cleanup;
1576         }
1577
1578 cleanup:
1579         kfree(msginfo);
1580
1581         return ret;
1582 }
1583
1584 static void invoke_sc_cb(struct vmbus_channel *primary_channel)
1585 {
1586         struct list_head *cur, *tmp;
1587         struct vmbus_channel *cur_channel;
1588
1589         if (primary_channel->sc_creation_callback == NULL)
1590                 return;
1591
1592         list_for_each_safe(cur, tmp, &primary_channel->sc_list) {
1593                 cur_channel = list_entry(cur, struct vmbus_channel, sc_list);
1594
1595                 primary_channel->sc_creation_callback(cur_channel);
1596         }
1597 }
1598
1599 void vmbus_set_sc_create_callback(struct vmbus_channel *primary_channel,
1600                                 void (*sc_cr_cb)(struct vmbus_channel *new_sc))
1601 {
1602         primary_channel->sc_creation_callback = sc_cr_cb;
1603 }
1604 EXPORT_SYMBOL_GPL(vmbus_set_sc_create_callback);
1605
1606 bool vmbus_are_subchannels_present(struct vmbus_channel *primary)
1607 {
1608         bool ret;
1609
1610         ret = !list_empty(&primary->sc_list);
1611
1612         if (ret) {
1613                 /*
1614                  * Invoke the callback on sub-channel creation.
1615                  * This will present a uniform interface to the
1616                  * clients.
1617                  */
1618                 invoke_sc_cb(primary);
1619         }
1620
1621         return ret;
1622 }
1623 EXPORT_SYMBOL_GPL(vmbus_are_subchannels_present);
1624
1625 void vmbus_set_chn_rescind_callback(struct vmbus_channel *channel,
1626                 void (*chn_rescind_cb)(struct vmbus_channel *))
1627 {
1628         channel->chn_rescind_callback = chn_rescind_cb;
1629 }
1630 EXPORT_SYMBOL_GPL(vmbus_set_chn_rescind_callback);