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         init_vp_index(newchannel);
 609
 610         /* Remember the channels that should be cleaned up upon suspend. */
 611         if (is_hvsock_channel(newchannel) || is_sub_channel(newchannel))
 612                 atomic_inc(&vmbus_connection.nr_chan_close_on_suspend);
 613
 614         /*
 615          * Now that we have acquired the channel_mutex,
 616          * we can release the potentially racing rescind thread.
 617          */
 618         atomic_dec(&vmbus_connection.offer_in_progress);
 619
 620         list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
 621                 if (guid_equal(&channel->offermsg.offer.if_type,
 622                                &newchannel->offermsg.offer.if_type) &&
 623                     guid_equal(&channel->offermsg.offer.if_instance,
 624                                &newchannel->offermsg.offer.if_instance)) {
 625                         fnew = false;
 626                         break;
 627                 }
 628         }
 629
 630         if (fnew) {
 631                 list_add_tail(&newchannel->listentry,
 632                               &vmbus_connection.chn_list);
 633         } else {
 634                 /*
 635                  * Check to see if this is a valid sub-channel.
 636                  */
 637                 if (newchannel->offermsg.offer.sub_channel_index == 0) {
 638                         mutex_unlock(&vmbus_connection.channel_mutex);
 639                         /*
 640                          * Don't call free_channel(), because newchannel->kobj
 641                          * is not initialized yet.
 642                          */
 643                         kfree(newchannel);
 644                         WARN_ON_ONCE(1);
 645                         return;
 646                 }
 647                 /*
 648                  * Process the sub-channel.
 649                  */
 650                 newchannel->primary_channel = channel;
 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  * We use this state to statically distribute the channel interrupt load.
 688  */
 689 static int next_numa_node_id;
 690
 691 /*
 692  * Starting with Win8, we can statically distribute the incoming
 693  * channel interrupt load by binding a channel to VCPU.
 694  *
 695  * For pre-win8 hosts or non-performance critical channels we assign the
 696  * VMBUS_CONNECT_CPU.
 697  *
 698  * Starting with win8, performance critical channels will be distributed
 699  * evenly among all the available NUMA nodes.  Once the node is assigned,
 700  * we will assign the CPU based on a simple round robin scheme.
 701  */
 702 static void init_vp_index(struct vmbus_channel *channel)
 703 {
 704         bool perf_chn = hv_is_perf_channel(channel);
 705         cpumask_var_t available_mask;
 706         struct cpumask *alloced_mask;
 707         u32 target_cpu;
 708         int numa_node;
 709
 710         if ((vmbus_proto_version == VERSION_WS2008) ||
 711             (vmbus_proto_version == VERSION_WIN7) || (!perf_chn) ||
 712             !alloc_cpumask_var(&available_mask, GFP_KERNEL)) {
 713                 /*
 714                  * Prior to win8, all channel interrupts are
 715                  * delivered on VMBUS_CONNECT_CPU.
 716                  * Also if the channel is not a performance critical
 717                  * channel, bind it to VMBUS_CONNECT_CPU.
 718                  * In case alloc_cpumask_var() fails, bind it to
 719                  * VMBUS_CONNECT_CPU.
 720                  */
 721                 channel->target_cpu = VMBUS_CONNECT_CPU;
 722                 if (perf_chn)
 723                         hv_set_alloced_cpu(VMBUS_CONNECT_CPU);
 724                 return;
 725         }
 726
 727         while (true) {
 728                 numa_node = next_numa_node_id++;
 729                 if (numa_node == nr_node_ids) {
 730                         next_numa_node_id = 0;
 731                         continue;
 732                 }
 733                 if (cpumask_empty(cpumask_of_node(numa_node)))
 734                         continue;
 735                 break;
 736         }
 737         alloced_mask = &hv_context.hv_numa_map[numa_node];
 738
 739         if (cpumask_weight(alloced_mask) ==
 740             cpumask_weight(cpumask_of_node(numa_node))) {
 741                 /*
 742                  * We have cycled through all the CPUs in the node;
 743                  * reset the alloced map.
 744                  */
 745                 cpumask_clear(alloced_mask);
 746         }
 747
 748         cpumask_xor(available_mask, alloced_mask, cpumask_of_node(numa_node));
 749
 750         target_cpu = cpumask_first(available_mask);
 751         cpumask_set_cpu(target_cpu, alloced_mask);
 752
 753         channel->target_cpu = target_cpu;
 754
 755         free_cpumask_var(available_mask);
 756 }
 757
 758 #define UNLOAD_DELAY_UNIT_MS    10              /* 10 milliseconds */
 759 #define UNLOAD_WAIT_MS          (100*1000)      /* 100 seconds */
 760 #define UNLOAD_WAIT_LOOPS       (UNLOAD_WAIT_MS/UNLOAD_DELAY_UNIT_MS)
 761 #define UNLOAD_MSG_MS           (5*1000)        /* Every 5 seconds */
 762 #define UNLOAD_MSG_LOOPS        (UNLOAD_MSG_MS/UNLOAD_DELAY_UNIT_MS)
 763
 764 static void vmbus_wait_for_unload(void)
 765 {
 766         int cpu;
 767         void *page_addr;
 768         struct hv_message *msg;
 769         struct vmbus_channel_message_header *hdr;
 770         u32 message_type, i;
 771
 772         /*
 773          * CHANNELMSG_UNLOAD_RESPONSE is always delivered to the CPU which was
 774          * used for initial contact or to CPU0 depending on host version. When
 775          * we're crashing on a different CPU let's hope that IRQ handler on
 776          * the cpu which receives CHANNELMSG_UNLOAD_RESPONSE is still
 777          * functional and vmbus_unload_response() will complete
 778          * vmbus_connection.unload_event. If not, the last thing we can do is
 779          * read message pages for all CPUs directly.
 780          *
 781          * Wait up to 100 seconds since an Azure host must writeback any dirty
 782          * data in its disk cache before the VMbus UNLOAD request will
 783          * complete. This flushing has been empirically observed to take up
 784          * to 50 seconds in cases with a lot of dirty data, so allow additional
 785          * leeway and for inaccuracies in mdelay(). But eventually time out so
 786          * that the panic path can't get hung forever in case the response
 787          * message isn't seen.
 788          */
 789         for (i = 1; i <= UNLOAD_WAIT_LOOPS; i++) {
 790                 if (completion_done(&vmbus_connection.unload_event))
 791                         goto completed;
 792
 793                 for_each_online_cpu(cpu) {
 794                         struct hv_per_cpu_context *hv_cpu
 795                                 = per_cpu_ptr(hv_context.cpu_context, cpu);
 796
 797                         page_addr = hv_cpu->synic_message_page;
 798                         msg = (struct hv_message *)page_addr
 799                                 + VMBUS_MESSAGE_SINT;
 800
 801                         message_type = READ_ONCE(msg->header.message_type);
 802                         if (message_type == HVMSG_NONE)
 803                                 continue;
 804
 805                         hdr = (struct vmbus_channel_message_header *)
 806                                 msg->u.payload;
 807
 808                         if (hdr->msgtype == CHANNELMSG_UNLOAD_RESPONSE)
 809                                 complete(&vmbus_connection.unload_event);
 810
 811                         vmbus_signal_eom(msg, message_type);
 812                 }
 813
 814                 /*
 815                  * Give a notice periodically so someone watching the
 816                  * serial output won't think it is completely hung.
 817                  */
 818                 if (!(i % UNLOAD_MSG_LOOPS))
 819                         pr_notice("Waiting for VMBus UNLOAD to complete\n");
 820
 821                 mdelay(UNLOAD_DELAY_UNIT_MS);
 822         }
 823         pr_err("Continuing even though VMBus UNLOAD did not complete\n");
 824
 825 completed:
 826         /*
 827          * We're crashing and already got the UNLOAD_RESPONSE, cleanup all
 828          * maybe-pending messages on all CPUs to be able to receive new
 829          * messages after we reconnect.
 830          */
 831         for_each_online_cpu(cpu) {
 832                 struct hv_per_cpu_context *hv_cpu
 833                         = per_cpu_ptr(hv_context.cpu_context, cpu);
 834
 835                 page_addr = hv_cpu->synic_message_page;
 836                 msg = (struct hv_message *)page_addr + VMBUS_MESSAGE_SINT;
 837                 msg->header.message_type = HVMSG_NONE;
 838         }
 839 }
 840
 841 /*
 842  * vmbus_unload_response - Handler for the unload response.
 843  */
 844 static void vmbus_unload_response(struct vmbus_channel_message_header *hdr)
 845 {
 846         /*
 847          * This is a global event; just wakeup the waiting thread.
 848          * Once we successfully unload, we can cleanup the monitor state.
 849          *
 850          * NB.  A malicious or compromised Hyper-V could send a spurious
 851          * message of type CHANNELMSG_UNLOAD_RESPONSE, and trigger a call
 852          * of the complete() below.  Make sure that unload_event has been
 853          * initialized by the time this complete() is executed.
 854          */
 855         complete(&vmbus_connection.unload_event);
 856 }
 857
 858 void vmbus_initiate_unload(bool crash)
 859 {
 860         struct vmbus_channel_message_header hdr;
 861
 862         if (xchg(&vmbus_connection.conn_state, DISCONNECTED) == DISCONNECTED)
 863                 return;
 864
 865         /* Pre-Win2012R2 hosts don't support reconnect */
 866         if (vmbus_proto_version < VERSION_WIN8_1)
 867                 return;
 868
 869         reinit_completion(&vmbus_connection.unload_event);
 870         memset(&hdr, 0, sizeof(struct vmbus_channel_message_header));
 871         hdr.msgtype = CHANNELMSG_UNLOAD;
 872         vmbus_post_msg(&hdr, sizeof(struct vmbus_channel_message_header),
 873                        !crash);
 874
 875         /*
 876          * vmbus_initiate_unload() is also called on crash and the crash can be
 877          * happening in an interrupt context, where scheduling is impossible.
 878          */
 879         if (!crash)
 880                 wait_for_completion(&vmbus_connection.unload_event);
 881         else
 882                 vmbus_wait_for_unload();
 883 }
 884
 885 static void check_ready_for_resume_event(void)
 886 {
 887         /*
 888          * If all the old primary channels have been fixed up, then it's safe
 889          * to resume.
 890          */
 891         if (atomic_dec_and_test(&vmbus_connection.nr_chan_fixup_on_resume))
 892                 complete(&vmbus_connection.ready_for_resume_event);
 893 }
 894
 895 static void vmbus_setup_channel_state(struct vmbus_channel *channel,
 896                                       struct vmbus_channel_offer_channel *offer)
 897 {
 898         /*
 899          * Setup state for signalling the host.
 900          */
 901         channel->sig_event = VMBUS_EVENT_CONNECTION_ID;
 902
 903         if (vmbus_proto_version != VERSION_WS2008) {
 904                 channel->is_dedicated_interrupt =
 905                                 (offer->is_dedicated_interrupt != 0);
 906                 channel->sig_event = offer->connection_id;
 907         }
 908
 909         memcpy(&channel->offermsg, offer,
 910                sizeof(struct vmbus_channel_offer_channel));
 911         channel->monitor_grp = (u8)offer->monitorid / 32;
 912         channel->monitor_bit = (u8)offer->monitorid % 32;
 913         channel->device_id = hv_get_dev_type(channel);
 914 }
 915
 916 /*
 917  * find_primary_channel_by_offer - Get the channel object given the new offer.
 918  * This is only used in the resume path of hibernation.
 919  */
 920 static struct vmbus_channel *
 921 find_primary_channel_by_offer(const struct vmbus_channel_offer_channel *offer)
 922 {
 923         struct vmbus_channel *channel = NULL, *iter;
 924         const guid_t *inst1, *inst2;
 925
 926         /* Ignore sub-channel offers. */
 927         if (offer->offer.sub_channel_index != 0)
 928                 return NULL;
 929
 930         mutex_lock(&vmbus_connection.channel_mutex);
 931
 932         list_for_each_entry(iter, &vmbus_connection.chn_list, listentry) {
 933                 inst1 = &iter->offermsg.offer.if_instance;
 934                 inst2 = &offer->offer.if_instance;
 935
 936                 if (guid_equal(inst1, inst2)) {
 937                         channel = iter;
 938                         break;
 939                 }
 940         }
 941
 942         mutex_unlock(&vmbus_connection.channel_mutex);
 943
 944         return channel;
 945 }
 946
 947 static bool vmbus_is_valid_device(const guid_t *guid)
 948 {
 949         u16 i;
 950
 951         if (!hv_is_isolation_supported())
 952                 return true;
 953
 954         for (i = 0; i < ARRAY_SIZE(vmbus_devs); i++) {
 955                 if (guid_equal(guid, &vmbus_devs[i].guid))
 956                         return vmbus_devs[i].allowed_in_isolated;
 957         }
 958         return false;
 959 }
 960
 961 /*
 962  * vmbus_onoffer - Handler for channel offers from vmbus in parent partition.
 963  *
 964  */
 965 static void vmbus_onoffer(struct vmbus_channel_message_header *hdr)
 966 {
 967         struct vmbus_channel_offer_channel *offer;
 968         struct vmbus_channel *oldchannel, *newchannel;
 969         size_t offer_sz;
 970
 971         offer = (struct vmbus_channel_offer_channel *)hdr;
 972
 973         trace_vmbus_onoffer(offer);
 974
 975         if (!vmbus_is_valid_device(&offer->offer.if_type)) {
 976                 pr_err_ratelimited("Invalid offer %d from the host supporting isolation\n",
 977                                    offer->child_relid);
 978                 atomic_dec(&vmbus_connection.offer_in_progress);
 979                 return;
 980         }
 981
 982         oldchannel = find_primary_channel_by_offer(offer);
 983
 984         if (oldchannel != NULL) {
 985                 /*
 986                  * We're resuming from hibernation: all the sub-channel and
 987                  * hv_sock channels we had before the hibernation should have
 988                  * been cleaned up, and now we must be seeing a re-offered
 989                  * primary channel that we had before the hibernation.
 990                  */
 991
 992                 /*
 993                  * { Initially: channel relid = INVALID_RELID,
 994                  *              channels[valid_relid] = NULL }
 995                  *
 996                  * CPU1                                 CPU2
 997                  *
 998                  * [vmbus_onoffer()]                    [vmbus_device_release()]
 999                  *
1000                  * LOCK channel_mutex                   LOCK channel_mutex
1001                  * STORE channel relid = valid_relid    LOAD r1 = channel relid
1002                  * MAP_RELID channel                    if (r1 != INVALID_RELID)
1003                  * UNLOCK channel_mutex                   UNMAP_RELID channel
1004                  *                                      UNLOCK channel_mutex
1005                  *
1006                  * Forbids: r1 == valid_relid &&
1007                  *              channels[valid_relid] == channel
1008                  *
1009                  * Note.  r1 can be INVALID_RELID only for an hv_sock channel.
1010                  * None of the hv_sock channels which were present before the
1011                  * suspend are re-offered upon the resume.  See the WARN_ON()
1012                  * in hv_process_channel_removal().
1013                  */
1014                 mutex_lock(&vmbus_connection.channel_mutex);
1015
1016                 atomic_dec(&vmbus_connection.offer_in_progress);
1017
1018                 WARN_ON(oldchannel->offermsg.child_relid != INVALID_RELID);
1019                 /* Fix up the relid. */
1020                 oldchannel->offermsg.child_relid = offer->child_relid;
1021
1022                 offer_sz = sizeof(*offer);
1023                 if (memcmp(offer, &oldchannel->offermsg, offer_sz) != 0) {
1024                         /*
1025                          * This is not an error, since the host can also change
1026                          * the other field(s) of the offer, e.g. on WS RS5
1027                          * (Build 17763), the offer->connection_id of the
1028                          * Mellanox VF vmbus device can change when the host
1029                          * reoffers the device upon resume.
1030                          */
1031                         pr_debug("vmbus offer changed: relid=%d\n",
1032                                  offer->child_relid);
1033
1034                         print_hex_dump_debug("Old vmbus offer: ",
1035                                              DUMP_PREFIX_OFFSET, 16, 4,
1036                                              &oldchannel->offermsg, offer_sz,
1037                                              false);
1038                         print_hex_dump_debug("New vmbus offer: ",
1039                                              DUMP_PREFIX_OFFSET, 16, 4,
1040                                              offer, offer_sz, false);
1041
1042                         /* Fix up the old channel. */
1043                         vmbus_setup_channel_state(oldchannel, offer);
1044                 }
1045
1046                 /* Add the channel back to the array of channels. */
1047                 vmbus_channel_map_relid(oldchannel);
1048                 check_ready_for_resume_event();
1049
1050                 mutex_unlock(&vmbus_connection.channel_mutex);
1051                 return;
1052         }
1053
1054         /* Allocate the channel object and save this offer. */
1055         newchannel = alloc_channel();
1056         if (!newchannel) {
1057                 vmbus_release_relid(offer->child_relid);
1058                 atomic_dec(&vmbus_connection.offer_in_progress);
1059                 pr_err("Unable to allocate channel object\n");
1060                 return;
1061         }
1062
1063         vmbus_setup_channel_state(newchannel, offer);
1064
1065         vmbus_process_offer(newchannel);
1066 }
1067
1068 static void check_ready_for_suspend_event(void)
1069 {
1070         /*
1071          * If all the sub-channels or hv_sock channels have been cleaned up,
1072          * then it's safe to suspend.
1073          */
1074         if (atomic_dec_and_test(&vmbus_connection.nr_chan_close_on_suspend))
1075                 complete(&vmbus_connection.ready_for_suspend_event);
1076 }
1077
1078 /*
1079  * vmbus_onoffer_rescind - Rescind offer handler.
1080  *
1081  * We queue a work item to process this offer synchronously
1082  */
1083 static void vmbus_onoffer_rescind(struct vmbus_channel_message_header *hdr)
1084 {
1085         struct vmbus_channel_rescind_offer *rescind;
1086         struct vmbus_channel *channel;
1087         struct device *dev;
1088         bool clean_up_chan_for_suspend;
1089
1090         rescind = (struct vmbus_channel_rescind_offer *)hdr;
1091
1092         trace_vmbus_onoffer_rescind(rescind);
1093
1094         /*
1095          * The offer msg and the corresponding rescind msg
1096          * from the host are guranteed to be ordered -
1097          * offer comes in first and then the rescind.
1098          * Since we process these events in work elements,
1099          * and with preemption, we may end up processing
1100          * the events out of order.  We rely on the synchronization
1101          * provided by offer_in_progress and by channel_mutex for
1102          * ordering these events:
1103          *
1104          * { Initially: offer_in_progress = 1 }
1105          *
1106          * CPU1                         CPU2
1107          *
1108          * [vmbus_onoffer()]            [vmbus_onoffer_rescind()]
1109          *
1110          * LOCK channel_mutex           WAIT_ON offer_in_progress == 0
1111          * DECREMENT offer_in_progress  LOCK channel_mutex
1112          * STORE channels[]             LOAD channels[]
1113          * UNLOCK channel_mutex         UNLOCK channel_mutex
1114          *
1115          * Forbids: CPU2's LOAD from *not* seeing CPU1's STORE
1116          */
1117
1118         while (atomic_read(&vmbus_connection.offer_in_progress) != 0) {
1119                 /*
1120                  * We wait here until any channel offer is currently
1121                  * being processed.
1122                  */
1123                 msleep(1);
1124         }
1125
1126         mutex_lock(&vmbus_connection.channel_mutex);
1127         channel = relid2channel(rescind->child_relid);
1128         if (channel != NULL) {
1129                 /*
1130                  * Guarantee that no other instance of vmbus_onoffer_rescind()
1131                  * has got a reference to the channel object.  Synchronize on
1132                  * &vmbus_connection.channel_mutex.
1133                  */
1134                 if (channel->rescind_ref) {
1135                         mutex_unlock(&vmbus_connection.channel_mutex);
1136                         return;
1137                 }
1138                 channel->rescind_ref = true;
1139         }
1140         mutex_unlock(&vmbus_connection.channel_mutex);
1141
1142         if (channel == NULL) {
1143                 /*
1144                  * We failed in processing the offer message;
1145                  * we would have cleaned up the relid in that
1146                  * failure path.
1147                  */
1148                 return;
1149         }
1150
1151         clean_up_chan_for_suspend = is_hvsock_channel(channel) ||
1152                                     is_sub_channel(channel);
1153         /*
1154          * Before setting channel->rescind in vmbus_rescind_cleanup(), we
1155          * should make sure the channel callback is not running any more.
1156          */
1157         vmbus_reset_channel_cb(channel);
1158
1159         /*
1160          * Now wait for offer handling to complete.
1161          */
1162         vmbus_rescind_cleanup(channel);
1163         while (READ_ONCE(channel->probe_done) == false) {
1164                 /*
1165                  * We wait here until any channel offer is currently
1166                  * being processed.
1167                  */
1168                 msleep(1);
1169         }
1170
1171         /*
1172          * At this point, the rescind handling can proceed safely.
1173          */
1174
1175         if (channel->device_obj) {
1176                 if (channel->chn_rescind_callback) {
1177                         channel->chn_rescind_callback(channel);
1178
1179                         if (clean_up_chan_for_suspend)
1180                                 check_ready_for_suspend_event();
1181
1182                         return;
1183                 }
1184                 /*
1185                  * We will have to unregister this device from the
1186                  * driver core.
1187                  */
1188                 dev = get_device(&channel->device_obj->device);
1189                 if (dev) {
1190                         vmbus_device_unregister(channel->device_obj);
1191                         put_device(dev);
1192                 }
1193         } else if (channel->primary_channel != NULL) {
1194                 /*
1195                  * Sub-channel is being rescinded. Following is the channel
1196                  * close sequence when initiated from the driveri (refer to
1197                  * vmbus_close() for details):
1198                  * 1. Close all sub-channels first
1199                  * 2. Then close the primary channel.
1200                  */
1201                 mutex_lock(&vmbus_connection.channel_mutex);
1202                 if (channel->state == CHANNEL_OPEN_STATE) {
1203                         /*
1204                          * The channel is currently not open;
1205                          * it is safe for us to cleanup the channel.
1206                          */
1207                         hv_process_channel_removal(channel);
1208                 } else {
1209                         complete(&channel->rescind_event);
1210                 }
1211                 mutex_unlock(&vmbus_connection.channel_mutex);
1212         }
1213
1214         /* The "channel" may have been freed. Do not access it any longer. */
1215
1216         if (clean_up_chan_for_suspend)
1217                 check_ready_for_suspend_event();
1218 }
1219
1220 void vmbus_hvsock_device_unregister(struct vmbus_channel *channel)
1221 {
1222         BUG_ON(!is_hvsock_channel(channel));
1223
1224         /* We always get a rescind msg when a connection is closed. */
1225         while (!READ_ONCE(channel->probe_done) || !READ_ONCE(channel->rescind))
1226                 msleep(1);
1227
1228         vmbus_device_unregister(channel->device_obj);
1229 }
1230 EXPORT_SYMBOL_GPL(vmbus_hvsock_device_unregister);
1231
1232
1233 /*
1234  * vmbus_onoffers_delivered -
1235  * This is invoked when all offers have been delivered.
1236  *
1237  * Nothing to do here.
1238  */
1239 static void vmbus_onoffers_delivered(
1240                         struct vmbus_channel_message_header *hdr)
1241 {
1242 }
1243
1244 /*
1245  * vmbus_onopen_result - Open result handler.
1246  *
1247  * This is invoked when we received a response to our channel open request.
1248  * Find the matching request, copy the response and signal the requesting
1249  * thread.
1250  */
1251 static void vmbus_onopen_result(struct vmbus_channel_message_header *hdr)
1252 {
1253         struct vmbus_channel_open_result *result;
1254         struct vmbus_channel_msginfo *msginfo;
1255         struct vmbus_channel_message_header *requestheader;
1256         struct vmbus_channel_open_channel *openmsg;
1257         unsigned long flags;
1258
1259         result = (struct vmbus_channel_open_result *)hdr;
1260
1261         trace_vmbus_onopen_result(result);
1262
1263         /*
1264          * Find the open msg, copy the result and signal/unblock the wait event
1265          */
1266         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1267
1268         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1269                                 msglistentry) {
1270                 requestheader =
1271                         (struct vmbus_channel_message_header *)msginfo->msg;
1272
1273                 if (requestheader->msgtype == CHANNELMSG_OPENCHANNEL) {
1274                         openmsg =
1275                         (struct vmbus_channel_open_channel *)msginfo->msg;
1276                         if (openmsg->child_relid == result->child_relid &&
1277                             openmsg->openid == result->openid) {
1278                                 memcpy(&msginfo->response.open_result,
1279                                        result,
1280                                        sizeof(
1281                                         struct vmbus_channel_open_result));
1282                                 complete(&msginfo->waitevent);
1283                                 break;
1284                         }
1285                 }
1286         }
1287         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1288 }
1289
1290 /*
1291  * vmbus_ongpadl_created - GPADL created handler.
1292  *
1293  * This is invoked when we received a response to our gpadl create request.
1294  * Find the matching request, copy the response and signal the requesting
1295  * thread.
1296  */
1297 static void vmbus_ongpadl_created(struct vmbus_channel_message_header *hdr)
1298 {
1299         struct vmbus_channel_gpadl_created *gpadlcreated;
1300         struct vmbus_channel_msginfo *msginfo;
1301         struct vmbus_channel_message_header *requestheader;
1302         struct vmbus_channel_gpadl_header *gpadlheader;
1303         unsigned long flags;
1304
1305         gpadlcreated = (struct vmbus_channel_gpadl_created *)hdr;
1306
1307         trace_vmbus_ongpadl_created(gpadlcreated);
1308
1309         /*
1310          * Find the establish msg, copy the result and signal/unblock the wait
1311          * event
1312          */
1313         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1314
1315         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1316                                 msglistentry) {
1317                 requestheader =
1318                         (struct vmbus_channel_message_header *)msginfo->msg;
1319
1320                 if (requestheader->msgtype == CHANNELMSG_GPADL_HEADER) {
1321                         gpadlheader =
1322                         (struct vmbus_channel_gpadl_header *)requestheader;
1323
1324                         if ((gpadlcreated->child_relid ==
1325                              gpadlheader->child_relid) &&
1326                             (gpadlcreated->gpadl == gpadlheader->gpadl)) {
1327                                 memcpy(&msginfo->response.gpadl_created,
1328                                        gpadlcreated,
1329                                        sizeof(
1330                                         struct vmbus_channel_gpadl_created));
1331                                 complete(&msginfo->waitevent);
1332                                 break;
1333                         }
1334                 }
1335         }
1336         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1337 }
1338
1339 /*
1340  * vmbus_onmodifychannel_response - Modify Channel response handler.
1341  *
1342  * This is invoked when we received a response to our channel modify request.
1343  * Find the matching request, copy the response and signal the requesting thread.
1344  */
1345 static void vmbus_onmodifychannel_response(struct vmbus_channel_message_header *hdr)
1346 {
1347         struct vmbus_channel_modifychannel_response *response;
1348         struct vmbus_channel_msginfo *msginfo;
1349         unsigned long flags;
1350
1351         response = (struct vmbus_channel_modifychannel_response *)hdr;
1352
1353         trace_vmbus_onmodifychannel_response(response);
1354
1355         /*
1356          * Find the modify msg, copy the response and signal/unblock the wait event.
1357          */
1358         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1359
1360         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list, msglistentry) {
1361                 struct vmbus_channel_message_header *responseheader =
1362                                 (struct vmbus_channel_message_header *)msginfo->msg;
1363
1364                 if (responseheader->msgtype == CHANNELMSG_MODIFYCHANNEL) {
1365                         struct vmbus_channel_modifychannel *modifymsg;
1366
1367                         modifymsg = (struct vmbus_channel_modifychannel *)msginfo->msg;
1368                         if (modifymsg->child_relid == response->child_relid) {
1369                                 memcpy(&msginfo->response.modify_response, response,
1370                                        sizeof(*response));
1371                                 complete(&msginfo->waitevent);
1372                                 break;
1373                         }
1374                 }
1375         }
1376         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1377 }
1378
1379 /*
1380  * vmbus_ongpadl_torndown - GPADL torndown handler.
1381  *
1382  * This is invoked when we received a response to our gpadl teardown request.
1383  * Find the matching request, copy the response and signal the requesting
1384  * thread.
1385  */
1386 static void vmbus_ongpadl_torndown(
1387                         struct vmbus_channel_message_header *hdr)
1388 {
1389         struct vmbus_channel_gpadl_torndown *gpadl_torndown;
1390         struct vmbus_channel_msginfo *msginfo;
1391         struct vmbus_channel_message_header *requestheader;
1392         struct vmbus_channel_gpadl_teardown *gpadl_teardown;
1393         unsigned long flags;
1394
1395         gpadl_torndown = (struct vmbus_channel_gpadl_torndown *)hdr;
1396
1397         trace_vmbus_ongpadl_torndown(gpadl_torndown);
1398
1399         /*
1400          * Find the open msg, copy the result and signal/unblock the wait event
1401          */
1402         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1403
1404         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1405                                 msglistentry) {
1406                 requestheader =
1407                         (struct vmbus_channel_message_header *)msginfo->msg;
1408
1409                 if (requestheader->msgtype == CHANNELMSG_GPADL_TEARDOWN) {
1410                         gpadl_teardown =
1411                         (struct vmbus_channel_gpadl_teardown *)requestheader;
1412
1413                         if (gpadl_torndown->gpadl == gpadl_teardown->gpadl) {
1414                                 memcpy(&msginfo->response.gpadl_torndown,
1415                                        gpadl_torndown,
1416                                        sizeof(
1417                                         struct vmbus_channel_gpadl_torndown));
1418                                 complete(&msginfo->waitevent);
1419                                 break;
1420                         }
1421                 }
1422         }
1423         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1424 }
1425
1426 /*
1427  * vmbus_onversion_response - Version response handler
1428  *
1429  * This is invoked when we received a response to our initiate contact request.
1430  * Find the matching request, copy the response and signal the requesting
1431  * thread.
1432  */
1433 static void vmbus_onversion_response(
1434                 struct vmbus_channel_message_header *hdr)
1435 {
1436         struct vmbus_channel_msginfo *msginfo;
1437         struct vmbus_channel_message_header *requestheader;
1438         struct vmbus_channel_version_response *version_response;
1439         unsigned long flags;
1440
1441         version_response = (struct vmbus_channel_version_response *)hdr;
1442
1443         trace_vmbus_onversion_response(version_response);
1444
1445         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1446
1447         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1448                                 msglistentry) {
1449                 requestheader =
1450                         (struct vmbus_channel_message_header *)msginfo->msg;
1451
1452                 if (requestheader->msgtype ==
1453                     CHANNELMSG_INITIATE_CONTACT) {
1454                         memcpy(&msginfo->response.version_response,
1455                               version_response,
1456                               sizeof(struct vmbus_channel_version_response));
1457                         complete(&msginfo->waitevent);
1458                 }
1459         }
1460         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1461 }
1462
1463 /* Channel message dispatch table */
1464 const struct vmbus_channel_message_table_entry
1465 channel_message_table[CHANNELMSG_COUNT] = {
1466         { CHANNELMSG_INVALID,                   0, NULL, 0},
1467         { CHANNELMSG_OFFERCHANNEL,              0, vmbus_onoffer,
1468                 sizeof(struct vmbus_channel_offer_channel)},
1469         { CHANNELMSG_RESCIND_CHANNELOFFER,      0, vmbus_onoffer_rescind,
1470                 sizeof(struct vmbus_channel_rescind_offer) },
1471         { CHANNELMSG_REQUESTOFFERS,             0, NULL, 0},
1472         { CHANNELMSG_ALLOFFERS_DELIVERED,       1, vmbus_onoffers_delivered, 0},
1473         { CHANNELMSG_OPENCHANNEL,               0, NULL, 0},
1474         { CHANNELMSG_OPENCHANNEL_RESULT,        1, vmbus_onopen_result,
1475                 sizeof(struct vmbus_channel_open_result)},
1476         { CHANNELMSG_CLOSECHANNEL,              0, NULL, 0},
1477         { CHANNELMSG_GPADL_HEADER,              0, NULL, 0},
1478         { CHANNELMSG_GPADL_BODY,                0, NULL, 0},
1479         { CHANNELMSG_GPADL_CREATED,             1, vmbus_ongpadl_created,
1480                 sizeof(struct vmbus_channel_gpadl_created)},
1481         { CHANNELMSG_GPADL_TEARDOWN,            0, NULL, 0},
1482         { CHANNELMSG_GPADL_TORNDOWN,            1, vmbus_ongpadl_torndown,
1483                 sizeof(struct vmbus_channel_gpadl_torndown) },
1484         { CHANNELMSG_RELID_RELEASED,            0, NULL, 0},
1485         { CHANNELMSG_INITIATE_CONTACT,          0, NULL, 0},
1486         { CHANNELMSG_VERSION_RESPONSE,          1, vmbus_onversion_response,
1487                 sizeof(struct vmbus_channel_version_response)},
1488         { CHANNELMSG_UNLOAD,                    0, NULL, 0},
1489         { CHANNELMSG_UNLOAD_RESPONSE,           1, vmbus_unload_response, 0},
1490         { CHANNELMSG_18,                        0, NULL, 0},
1491         { CHANNELMSG_19,                        0, NULL, 0},
1492         { CHANNELMSG_20,                        0, NULL, 0},
1493         { CHANNELMSG_TL_CONNECT_REQUEST,        0, NULL, 0},
1494         { CHANNELMSG_MODIFYCHANNEL,             0, NULL, 0},
1495         { CHANNELMSG_TL_CONNECT_RESULT,         0, NULL, 0},
1496         { CHANNELMSG_MODIFYCHANNEL_RESPONSE,    1, vmbus_onmodifychannel_response,
1497                 sizeof(struct vmbus_channel_modifychannel_response)},
1498 };
1499
1500 /*
1501  * vmbus_onmessage - Handler for channel protocol messages.
1502  *
1503  * This is invoked in the vmbus worker thread context.
1504  */
1505 void vmbus_onmessage(struct vmbus_channel_message_header *hdr)
1506 {
1507         trace_vmbus_on_message(hdr);
1508
1509         /*
1510          * vmbus_on_msg_dpc() makes sure the hdr->msgtype here can not go
1511          * out of bound and the message_handler pointer can not be NULL.
1512          */
1513         channel_message_table[hdr->msgtype].message_handler(hdr);
1514 }
1515
1516 /*
1517  * vmbus_request_offers - Send a request to get all our pending offers.
1518  */
1519 int vmbus_request_offers(void)
1520 {
1521         struct vmbus_channel_message_header *msg;
1522         struct vmbus_channel_msginfo *msginfo;
1523         int ret;
1524
1525         msginfo = kmalloc(sizeof(*msginfo) +
1526                           sizeof(struct vmbus_channel_message_header),
1527                           GFP_KERNEL);
1528         if (!msginfo)
1529                 return -ENOMEM;
1530
1531         msg = (struct vmbus_channel_message_header *)msginfo->msg;
1532
1533         msg->msgtype = CHANNELMSG_REQUESTOFFERS;
1534
1535         ret = vmbus_post_msg(msg, sizeof(struct vmbus_channel_message_header),
1536                              true);
1537
1538         trace_vmbus_request_offers(ret);
1539
1540         if (ret != 0) {
1541                 pr_err("Unable to request offers - %d\n", ret);
1542
1543                 goto cleanup;
1544         }
1545
1546 cleanup:
1547         kfree(msginfo);
1548
1549         return ret;
1550 }
1551
1552 static void invoke_sc_cb(struct vmbus_channel *primary_channel)
1553 {
1554         struct list_head *cur, *tmp;
1555         struct vmbus_channel *cur_channel;
1556
1557         if (primary_channel->sc_creation_callback == NULL)
1558                 return;
1559
1560         list_for_each_safe(cur, tmp, &primary_channel->sc_list) {
1561                 cur_channel = list_entry(cur, struct vmbus_channel, sc_list);
1562
1563                 primary_channel->sc_creation_callback(cur_channel);
1564         }
1565 }
1566
1567 void vmbus_set_sc_create_callback(struct vmbus_channel *primary_channel,
1568                                 void (*sc_cr_cb)(struct vmbus_channel *new_sc))
1569 {
1570         primary_channel->sc_creation_callback = sc_cr_cb;
1571 }
1572 EXPORT_SYMBOL_GPL(vmbus_set_sc_create_callback);
1573
1574 bool vmbus_are_subchannels_present(struct vmbus_channel *primary)
1575 {
1576         bool ret;
1577
1578         ret = !list_empty(&primary->sc_list);
1579
1580         if (ret) {
1581                 /*
1582                  * Invoke the callback on sub-channel creation.
1583                  * This will present a uniform interface to the
1584                  * clients.
1585                  */
1586                 invoke_sc_cb(primary);
1587         }
1588
1589         return ret;
1590 }
1591 EXPORT_SYMBOL_GPL(vmbus_are_subchannels_present);
1592
1593 void vmbus_set_chn_rescind_callback(struct vmbus_channel *channel,
1594                 void (*chn_rescind_cb)(struct vmbus_channel *))
1595 {
1596         channel->chn_rescind_callback = chn_rescind_cb;
1597 }
1598 EXPORT_SYMBOL_GPL(vmbus_set_chn_rescind_callback);