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 <[email protected]>
   7  *   Hank Janssen  <[email protected]>
   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 #include <linux/sched/isolation.h>
  25
  26 #include "hyperv_vmbus.h"
  27
  28 static void init_vp_index(struct vmbus_channel *channel);
  29
  30 const struct vmbus_device vmbus_devs[] = {
  31         /* IDE */
  32         { .dev_type = HV_IDE,
  33           HV_IDE_GUID,
  34           .perf_device = true,
  35           .allowed_in_isolated = false,
  36         },
  37
  38         /* SCSI */
  39         { .dev_type = HV_SCSI,
  40           HV_SCSI_GUID,
  41           .perf_device = true,
  42           .allowed_in_isolated = true,
  43         },
  44
  45         /* Fibre Channel */
  46         { .dev_type = HV_FC,
  47           HV_SYNTHFC_GUID,
  48           .perf_device = true,
  49           .allowed_in_isolated = false,
  50         },
  51
  52         /* Synthetic NIC */
  53         { .dev_type = HV_NIC,
  54           HV_NIC_GUID,
  55           .perf_device = true,
  56           .allowed_in_isolated = true,
  57         },
  58
  59         /* Network Direct */
  60         { .dev_type = HV_ND,
  61           HV_ND_GUID,
  62           .perf_device = true,
  63           .allowed_in_isolated = false,
  64         },
  65
  66         /* PCIE */
  67         { .dev_type = HV_PCIE,
  68           HV_PCIE_GUID,
  69           .perf_device = false,
  70           .allowed_in_isolated = true,
  71         },
  72
  73         /* Synthetic Frame Buffer */
  74         { .dev_type = HV_FB,
  75           HV_SYNTHVID_GUID,
  76           .perf_device = false,
  77           .allowed_in_isolated = false,
  78         },
  79
  80         /* Synthetic Keyboard */
  81         { .dev_type = HV_KBD,
  82           HV_KBD_GUID,
  83           .perf_device = false,
  84           .allowed_in_isolated = false,
  85         },
  86
  87         /* Synthetic MOUSE */
  88         { .dev_type = HV_MOUSE,
  89           HV_MOUSE_GUID,
  90           .perf_device = false,
  91           .allowed_in_isolated = false,
  92         },
  93
  94         /* KVP */
  95         { .dev_type = HV_KVP,
  96           HV_KVP_GUID,
  97           .perf_device = false,
  98           .allowed_in_isolated = false,
  99         },
 100
 101         /* Time Synch */
 102         { .dev_type = HV_TS,
 103           HV_TS_GUID,
 104           .perf_device = false,
 105           .allowed_in_isolated = true,
 106         },
 107
 108         /* Heartbeat */
 109         { .dev_type = HV_HB,
 110           HV_HEART_BEAT_GUID,
 111           .perf_device = false,
 112           .allowed_in_isolated = true,
 113         },
 114
 115         /* Shutdown */
 116         { .dev_type = HV_SHUTDOWN,
 117           HV_SHUTDOWN_GUID,
 118           .perf_device = false,
 119           .allowed_in_isolated = true,
 120         },
 121
 122         /* File copy */
 123         /* fcopy always uses 16KB ring buffer size and is working well for last many years */
 124         { .pref_ring_size = 0x4000,
 125           .dev_type = HV_FCOPY,
 126           HV_FCOPY_GUID,
 127           .perf_device = false,
 128           .allowed_in_isolated = false,
 129         },
 130
 131         /* Backup */
 132         { .dev_type = HV_BACKUP,
 133           HV_VSS_GUID,
 134           .perf_device = false,
 135           .allowed_in_isolated = false,
 136         },
 137
 138         /* Dynamic Memory */
 139         { .dev_type = HV_DM,
 140           HV_DM_GUID,
 141           .perf_device = false,
 142           .allowed_in_isolated = false,
 143         },
 144
 145         /*
 146          * Unknown GUID
 147          * 64 KB ring buffer + 4 KB header should be sufficient size for any Hyper-V device apart
 148          * from HV_NIC and HV_SCSI. This case avoid the fallback for unknown devices to allocate
 149          * much bigger (2 MB) of ring size.
 150          */
 151         { .pref_ring_size = 0x11000,
 152           .dev_type = HV_UNKNOWN,
 153           .perf_device = false,
 154           .allowed_in_isolated = false,
 155         },
 156 };
 157 EXPORT_SYMBOL_GPL(vmbus_devs);
 158
 159 static const struct {
 160         guid_t guid;
 161 } vmbus_unsupported_devs[] = {
 162         { HV_AVMA1_GUID },
 163         { HV_AVMA2_GUID },
 164         { HV_RDV_GUID   },
 165         { HV_IMC_GUID   },
 166 };
 167
 168 /*
 169  * The rescinded channel may be blocked waiting for a response from the host;
 170  * take care of that.
 171  */
 172 static void vmbus_rescind_cleanup(struct vmbus_channel *channel)
 173 {
 174         struct vmbus_channel_msginfo *msginfo;
 175         unsigned long flags;
 176
 177
 178         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
 179         channel->rescind = true;
 180         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
 181                                 msglistentry) {
 182
 183                 if (msginfo->waiting_channel == channel) {
 184                         complete(&msginfo->waitevent);
 185                         break;
 186                 }
 187         }
 188         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
 189 }
 190
 191 static bool is_unsupported_vmbus_devs(const guid_t *guid)
 192 {
 193         int i;
 194
 195         for (i = 0; i < ARRAY_SIZE(vmbus_unsupported_devs); i++)
 196                 if (guid_equal(guid, &vmbus_unsupported_devs[i].guid))
 197                         return true;
 198         return false;
 199 }
 200
 201 static u16 hv_get_dev_type(const struct vmbus_channel *channel)
 202 {
 203         const guid_t *guid = &channel->offermsg.offer.if_type;
 204         u16 i;
 205
 206         if (is_hvsock_channel(channel) || is_unsupported_vmbus_devs(guid))
 207                 return HV_UNKNOWN;
 208
 209         for (i = HV_IDE; i < HV_UNKNOWN; i++) {
 210                 if (guid_equal(guid, &vmbus_devs[i].guid))
 211                         return i;
 212         }
 213         pr_info("Unknown GUID: %pUl\n", guid);
 214         return i;
 215 }
 216
 217 /**
 218  * vmbus_prep_negotiate_resp() - Create default response for Negotiate message
 219  * @icmsghdrp: Pointer to msg header structure
 220  * @buf: Raw buffer channel data
 221  * @buflen: Length of the raw buffer channel data.
 222  * @fw_version: The framework versions we can support.
 223  * @fw_vercnt: The size of @fw_version.
 224  * @srv_version: The service versions we can support.
 225  * @srv_vercnt: The size of @srv_version.
 226  * @nego_fw_version: The selected framework version.
 227  * @nego_srv_version: The selected service version.
 228  *
 229  * Note: Versions are given in decreasing order.
 230  *
 231  * Set up and fill in default negotiate response message.
 232  * Mainly used by Hyper-V drivers.
 233  */
 234 bool vmbus_prep_negotiate_resp(struct icmsg_hdr *icmsghdrp, u8 *buf,
 235                                 u32 buflen, const int *fw_version, int fw_vercnt,
 236                                 const int *srv_version, int srv_vercnt,
 237                                 int *nego_fw_version, int *nego_srv_version)
 238 {
 239         int icframe_major, icframe_minor;
 240         int icmsg_major, icmsg_minor;
 241         int fw_major, fw_minor;
 242         int srv_major, srv_minor;
 243         int i, j;
 244         bool found_match = false;
 245         struct icmsg_negotiate *negop;
 246
 247         /* Check that there's enough space for icframe_vercnt, icmsg_vercnt */
 248         if (buflen < ICMSG_HDR + offsetof(struct icmsg_negotiate, reserved)) {
 249                 pr_err_ratelimited("Invalid icmsg negotiate\n");
 250                 return false;
 251         }
 252
 253         icmsghdrp->icmsgsize = 0x10;
 254         negop = (struct icmsg_negotiate *)&buf[ICMSG_HDR];
 255
 256         icframe_major = negop->icframe_vercnt;
 257         icframe_minor = 0;
 258
 259         icmsg_major = negop->icmsg_vercnt;
 260         icmsg_minor = 0;
 261
 262         /* Validate negop packet */
 263         if (icframe_major > IC_VERSION_NEGOTIATION_MAX_VER_COUNT ||
 264             icmsg_major > IC_VERSION_NEGOTIATION_MAX_VER_COUNT ||
 265             ICMSG_NEGOTIATE_PKT_SIZE(icframe_major, icmsg_major) > buflen) {
 266                 pr_err_ratelimited("Invalid icmsg negotiate - icframe_major: %u, icmsg_major: %u\n",
 267                                    icframe_major, icmsg_major);
 268                 goto fw_error;
 269         }
 270
 271         /*
 272          * Select the framework version number we will
 273          * support.
 274          */
 275
 276         for (i = 0; i < fw_vercnt; i++) {
 277                 fw_major = (fw_version[i] >> 16);
 278                 fw_minor = (fw_version[i] & 0xFFFF);
 279
 280                 for (j = 0; j < negop->icframe_vercnt; j++) {
 281                         if ((negop->icversion_data[j].major == fw_major) &&
 282                             (negop->icversion_data[j].minor == fw_minor)) {
 283                                 icframe_major = negop->icversion_data[j].major;
 284                                 icframe_minor = negop->icversion_data[j].minor;
 285                                 found_match = true;
 286                                 break;
 287                         }
 288                 }
 289
 290                 if (found_match)
 291                         break;
 292         }
 293
 294         if (!found_match)
 295                 goto fw_error;
 296
 297         found_match = false;
 298
 299         for (i = 0; i < srv_vercnt; i++) {
 300                 srv_major = (srv_version[i] >> 16);
 301                 srv_minor = (srv_version[i] & 0xFFFF);
 302
 303                 for (j = negop->icframe_vercnt;
 304                         (j < negop->icframe_vercnt + negop->icmsg_vercnt);
 305                         j++) {
 306
 307                         if ((negop->icversion_data[j].major == srv_major) &&
 308                                 (negop->icversion_data[j].minor == srv_minor)) {
 309
 310                                 icmsg_major = negop->icversion_data[j].major;
 311                                 icmsg_minor = negop->icversion_data[j].minor;
 312                                 found_match = true;
 313                                 break;
 314                         }
 315                 }
 316
 317                 if (found_match)
 318                         break;
 319         }
 320
 321         /*
 322          * Respond with the framework and service
 323          * version numbers we can support.
 324          */
 325
 326 fw_error:
 327         if (!found_match) {
 328                 negop->icframe_vercnt = 0;
 329                 negop->icmsg_vercnt = 0;
 330         } else {
 331                 negop->icframe_vercnt = 1;
 332                 negop->icmsg_vercnt = 1;
 333         }
 334
 335         if (nego_fw_version)
 336                 *nego_fw_version = (icframe_major << 16) | icframe_minor;
 337
 338         if (nego_srv_version)
 339                 *nego_srv_version = (icmsg_major << 16) | icmsg_minor;
 340
 341         negop->icversion_data[0].major = icframe_major;
 342         negop->icversion_data[0].minor = icframe_minor;
 343         negop->icversion_data[1].major = icmsg_major;
 344         negop->icversion_data[1].minor = icmsg_minor;
 345         return found_match;
 346 }
 347 EXPORT_SYMBOL_GPL(vmbus_prep_negotiate_resp);
 348
 349 /*
 350  * alloc_channel - Allocate and initialize a vmbus channel object
 351  */
 352 static struct vmbus_channel *alloc_channel(void)
 353 {
 354         struct vmbus_channel *channel;
 355
 356         channel = kzalloc(sizeof(*channel), GFP_ATOMIC);
 357         if (!channel)
 358                 return NULL;
 359
 360         spin_lock_init(&channel->sched_lock);
 361         init_completion(&channel->rescind_event);
 362
 363         INIT_LIST_HEAD(&channel->sc_list);
 364
 365         tasklet_init(&channel->callback_event,
 366                      vmbus_on_event, (unsigned long)channel);
 367
 368         hv_ringbuffer_pre_init(channel);
 369
 370         return channel;
 371 }
 372
 373 /*
 374  * free_channel - Release the resources used by the vmbus channel object
 375  */
 376 static void free_channel(struct vmbus_channel *channel)
 377 {
 378         tasklet_kill(&channel->callback_event);
 379         vmbus_remove_channel_attr_group(channel);
 380
 381         kobject_put(&channel->kobj);
 382 }
 383
 384 void vmbus_channel_map_relid(struct vmbus_channel *channel)
 385 {
 386         if (WARN_ON(channel->offermsg.child_relid >= MAX_CHANNEL_RELIDS))
 387                 return;
 388         /*
 389          * The mapping of the channel's relid is visible from the CPUs that
 390          * execute vmbus_chan_sched() by the time that vmbus_chan_sched() will
 391          * execute:
 392          *
 393          *  (a) In the "normal (i.e., not resuming from hibernation)" path,
 394          *      the full barrier in virt_store_mb() guarantees that the store
 395          *      is propagated to all CPUs before the add_channel_work work
 396          *      is queued.  In turn, add_channel_work is queued before the
 397          *      channel's ring buffer is allocated/initialized and the
 398          *      OPENCHANNEL message for the channel is sent in vmbus_open().
 399          *      Hyper-V won't start sending the interrupts for the channel
 400          *      before the OPENCHANNEL message is acked.  The memory barrier
 401          *      in vmbus_chan_sched() -> sync_test_and_clear_bit() ensures
 402          *      that vmbus_chan_sched() must find the channel's relid in
 403          *      recv_int_page before retrieving the channel pointer from the
 404          *      array of channels.
 405          *
 406          *  (b) In the "resuming from hibernation" path, the virt_store_mb()
 407          *      guarantees that the store is propagated to all CPUs before
 408          *      the VMBus connection is marked as ready for the resume event
 409          *      (cf. check_ready_for_resume_event()).  The interrupt handler
 410          *      of the VMBus driver and vmbus_chan_sched() can not run before
 411          *      vmbus_bus_resume() has completed execution (cf. resume_noirq).
 412          */
 413         virt_store_mb(
 414                 vmbus_connection.channels[channel->offermsg.child_relid],
 415                 channel);
 416 }
 417
 418 void vmbus_channel_unmap_relid(struct vmbus_channel *channel)
 419 {
 420         if (WARN_ON(channel->offermsg.child_relid >= MAX_CHANNEL_RELIDS))
 421                 return;
 422         WRITE_ONCE(
 423                 vmbus_connection.channels[channel->offermsg.child_relid],
 424                 NULL);
 425 }
 426
 427 static void vmbus_release_relid(u32 relid)
 428 {
 429         struct vmbus_channel_relid_released msg;
 430         int ret;
 431
 432         memset(&msg, 0, sizeof(struct vmbus_channel_relid_released));
 433         msg.child_relid = relid;
 434         msg.header.msgtype = CHANNELMSG_RELID_RELEASED;
 435         ret = vmbus_post_msg(&msg, sizeof(struct vmbus_channel_relid_released),
 436                              true);
 437
 438         trace_vmbus_release_relid(&msg, ret);
 439 }
 440
 441 void hv_process_channel_removal(struct vmbus_channel *channel)
 442 {
 443         lockdep_assert_held(&vmbus_connection.channel_mutex);
 444         BUG_ON(!channel->rescind);
 445
 446         /*
 447          * hv_process_channel_removal() could find INVALID_RELID only for
 448          * hv_sock channels.  See the inline comments in vmbus_onoffer().
 449          */
 450         WARN_ON(channel->offermsg.child_relid == INVALID_RELID &&
 451                 !is_hvsock_channel(channel));
 452
 453         /*
 454          * Upon suspend, an in-use hv_sock channel is removed from the array of
 455          * channels and the relid is invalidated.  After hibernation, when the
 456          * user-space application destroys the channel, it's unnecessary and
 457          * unsafe to remove the channel from the array of channels.  See also
 458          * the inline comments before the call of vmbus_release_relid() below.
 459          */
 460         if (channel->offermsg.child_relid != INVALID_RELID)
 461                 vmbus_channel_unmap_relid(channel);
 462
 463         if (channel->primary_channel == NULL)
 464                 list_del(&channel->listentry);
 465         else
 466                 list_del(&channel->sc_list);
 467
 468         /*
 469          * If this is a "perf" channel, updates the hv_numa_map[] masks so that
 470          * init_vp_index() can (re-)use the CPU.
 471          */
 472         if (hv_is_perf_channel(channel))
 473                 hv_clear_allocated_cpu(channel->target_cpu);
 474
 475         /*
 476          * Upon suspend, an in-use hv_sock channel is marked as "rescinded" and
 477          * the relid is invalidated; after hibernation, when the user-space app
 478          * destroys the channel, the relid is INVALID_RELID, and in this case
 479          * it's unnecessary and unsafe to release the old relid, since the same
 480          * relid can refer to a completely different channel now.
 481          */
 482         if (channel->offermsg.child_relid != INVALID_RELID)
 483                 vmbus_release_relid(channel->offermsg.child_relid);
 484
 485         free_channel(channel);
 486 }
 487
 488 void vmbus_free_channels(void)
 489 {
 490         struct vmbus_channel *channel, *tmp;
 491
 492         list_for_each_entry_safe(channel, tmp, &vmbus_connection.chn_list,
 493                 listentry) {
 494                 /* hv_process_channel_removal() needs this */
 495                 channel->rescind = true;
 496
 497                 vmbus_device_unregister(channel->device_obj);
 498         }
 499 }
 500
 501 /* Note: the function can run concurrently for primary/sub channels. */
 502 static void vmbus_add_channel_work(struct work_struct *work)
 503 {
 504         struct vmbus_channel *newchannel =
 505                 container_of(work, struct vmbus_channel, add_channel_work);
 506         struct vmbus_channel *primary_channel = newchannel->primary_channel;
 507         int ret;
 508
 509         /*
 510          * This state is used to indicate a successful open
 511          * so that when we do close the channel normally, we
 512          * can cleanup properly.
 513          */
 514         newchannel->state = CHANNEL_OPEN_STATE;
 515
 516         if (primary_channel != NULL) {
 517                 /* newchannel is a sub-channel. */
 518                 struct hv_device *dev = primary_channel->device_obj;
 519
 520                 if (vmbus_add_channel_kobj(dev, newchannel))
 521                         goto err_deq_chan;
 522
 523                 if (primary_channel->sc_creation_callback != NULL)
 524                         primary_channel->sc_creation_callback(newchannel);
 525
 526                 newchannel->probe_done = true;
 527                 return;
 528         }
 529
 530         /*
 531          * Start the process of binding the primary channel to the driver
 532          */
 533         newchannel->device_obj = vmbus_device_create(
 534                 &newchannel->offermsg.offer.if_type,
 535                 &newchannel->offermsg.offer.if_instance,
 536                 newchannel);
 537         if (!newchannel->device_obj)
 538                 goto err_deq_chan;
 539
 540         newchannel->device_obj->device_id = newchannel->device_id;
 541         /*
 542          * Add the new device to the bus. This will kick off device-driver
 543          * binding which eventually invokes the device driver's AddDevice()
 544          * method.
 545          *
 546          * If vmbus_device_register() fails, the 'device_obj' is freed in
 547          * vmbus_device_release() as called by device_unregister() in the
 548          * error path of vmbus_device_register(). In the outside error
 549          * path, there's no need to free it.
 550          */
 551         ret = vmbus_device_register(newchannel->device_obj);
 552
 553         if (ret != 0) {
 554                 pr_err("unable to add child device object (relid %d)\n",
 555                         newchannel->offermsg.child_relid);
 556                 goto err_deq_chan;
 557         }
 558
 559         newchannel->probe_done = true;
 560         return;
 561
 562 err_deq_chan:
 563         mutex_lock(&vmbus_connection.channel_mutex);
 564
 565         /*
 566          * We need to set the flag, otherwise
 567          * vmbus_onoffer_rescind() can be blocked.
 568          */
 569         newchannel->probe_done = true;
 570
 571         if (primary_channel == NULL)
 572                 list_del(&newchannel->listentry);
 573         else
 574                 list_del(&newchannel->sc_list);
 575
 576         /* vmbus_process_offer() has mapped the channel. */
 577         vmbus_channel_unmap_relid(newchannel);
 578
 579         mutex_unlock(&vmbus_connection.channel_mutex);
 580
 581         vmbus_release_relid(newchannel->offermsg.child_relid);
 582
 583         free_channel(newchannel);
 584 }
 585
 586 /*
 587  * vmbus_process_offer - Process the offer by creating a channel/device
 588  * associated with this offer
 589  */
 590 static void vmbus_process_offer(struct vmbus_channel *newchannel)
 591 {
 592         struct vmbus_channel *channel;
 593         struct workqueue_struct *wq;
 594         bool fnew = true;
 595
 596         /*
 597          * Synchronize vmbus_process_offer() and CPU hotplugging:
 598          *
 599          * CPU1                         CPU2
 600          *
 601          * [vmbus_process_offer()]      [Hot removal of the CPU]
 602          *
 603          * CPU_READ_LOCK                CPUS_WRITE_LOCK
 604          * LOAD cpu_online_mask         SEARCH chn_list
 605          * STORE target_cpu             LOAD target_cpu
 606          * INSERT chn_list              STORE cpu_online_mask
 607          * CPUS_READ_UNLOCK             CPUS_WRITE_UNLOCK
 608          *
 609          * Forbids: CPU1's LOAD from *not* seing CPU2's STORE &&
 610          *              CPU2's SEARCH from *not* seeing CPU1's INSERT
 611          *
 612          * Forbids: CPU2's SEARCH from seeing CPU1's INSERT &&
 613          *              CPU2's LOAD from *not* seing CPU1's STORE
 614          */
 615         cpus_read_lock();
 616
 617         /*
 618          * Serializes the modifications of the chn_list list as well as
 619          * the accesses to next_numa_node_id in init_vp_index().
 620          */
 621         mutex_lock(&vmbus_connection.channel_mutex);
 622
 623         list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
 624                 if (guid_equal(&channel->offermsg.offer.if_type,
 625                                &newchannel->offermsg.offer.if_type) &&
 626                     guid_equal(&channel->offermsg.offer.if_instance,
 627                                &newchannel->offermsg.offer.if_instance)) {
 628                         fnew = false;
 629                         newchannel->primary_channel = channel;
 630                         break;
 631                 }
 632         }
 633
 634         init_vp_index(newchannel);
 635
 636         /* Remember the channels that should be cleaned up upon suspend. */
 637         if (is_hvsock_channel(newchannel) || is_sub_channel(newchannel))
 638                 atomic_inc(&vmbus_connection.nr_chan_close_on_suspend);
 639
 640         /*
 641          * Now that we have acquired the channel_mutex,
 642          * we can release the potentially racing rescind thread.
 643          */
 644         atomic_dec(&vmbus_connection.offer_in_progress);
 645
 646         if (fnew) {
 647                 list_add_tail(&newchannel->listentry,
 648                               &vmbus_connection.chn_list);
 649         } else {
 650                 /*
 651                  * Check to see if this is a valid sub-channel.
 652                  */
 653                 if (newchannel->offermsg.offer.sub_channel_index == 0) {
 654                         mutex_unlock(&vmbus_connection.channel_mutex);
 655                         cpus_read_unlock();
 656                         /*
 657                          * Don't call free_channel(), because newchannel->kobj
 658                          * is not initialized yet.
 659                          */
 660                         kfree(newchannel);
 661                         WARN_ON_ONCE(1);
 662                         return;
 663                 }
 664                 /*
 665                  * Process the sub-channel.
 666                  */
 667                 list_add_tail(&newchannel->sc_list, &channel->sc_list);
 668         }
 669
 670         vmbus_channel_map_relid(newchannel);
 671
 672         mutex_unlock(&vmbus_connection.channel_mutex);
 673         cpus_read_unlock();
 674
 675         /*
 676          * vmbus_process_offer() mustn't call channel->sc_creation_callback()
 677          * directly for sub-channels, because sc_creation_callback() ->
 678          * vmbus_open() may never get the host's response to the
 679          * OPEN_CHANNEL message (the host may rescind a channel at any time,
 680          * e.g. in the case of hot removing a NIC), and vmbus_onoffer_rescind()
 681          * may not wake up the vmbus_open() as it's blocked due to a non-zero
 682          * vmbus_connection.offer_in_progress, and finally we have a deadlock.
 683          *
 684          * The above is also true for primary channels, if the related device
 685          * drivers use sync probing mode by default.
 686          *
 687          * And, usually the handling of primary channels and sub-channels can
 688          * depend on each other, so we should offload them to different
 689          * workqueues to avoid possible deadlock, e.g. in sync-probing mode,
 690          * NIC1's netvsc_subchan_work() can race with NIC2's netvsc_probe() ->
 691          * rtnl_lock(), and causes deadlock: the former gets the rtnl_lock
 692          * and waits for all the sub-channels to appear, but the latter
 693          * can't get the rtnl_lock and this blocks the handling of
 694          * sub-channels.
 695          */
 696         INIT_WORK(&newchannel->add_channel_work, vmbus_add_channel_work);
 697         wq = fnew ? vmbus_connection.handle_primary_chan_wq :
 698                     vmbus_connection.handle_sub_chan_wq;
 699         queue_work(wq, &newchannel->add_channel_work);
 700 }
 701
 702 /*
 703  * Check if CPUs used by other channels of the same device.
 704  * It should only be called by init_vp_index().
 705  */
 706 static bool hv_cpuself_used(u32 cpu, struct vmbus_channel *chn)
 707 {
 708         struct vmbus_channel *primary = chn->primary_channel;
 709         struct vmbus_channel *sc;
 710
 711         lockdep_assert_held(&vmbus_connection.channel_mutex);
 712
 713         if (!primary)
 714                 return false;
 715
 716         if (primary->target_cpu == cpu)
 717                 return true;
 718
 719         list_for_each_entry(sc, &primary->sc_list, sc_list)
 720                 if (sc != chn && sc->target_cpu == cpu)
 721                         return true;
 722
 723         return false;
 724 }
 725
 726 /*
 727  * We use this state to statically distribute the channel interrupt load.
 728  */
 729 static int next_numa_node_id;
 730
 731 /*
 732  * We can statically distribute the incoming channel interrupt load
 733  * by binding a channel to VCPU.
 734  *
 735  * For non-performance critical channels we assign the VMBUS_CONNECT_CPU.
 736  * Performance critical channels will be distributed evenly among all
 737  * the available NUMA nodes.  Once the node is assigned, we will assign
 738  * the CPU based on a simple round robin scheme.
 739  */
 740 static void init_vp_index(struct vmbus_channel *channel)
 741 {
 742         bool perf_chn = hv_is_perf_channel(channel);
 743         u32 i, ncpu = num_online_cpus();
 744         cpumask_var_t available_mask;
 745         struct cpumask *allocated_mask;
 746         const struct cpumask *hk_mask = housekeeping_cpumask(HK_TYPE_MANAGED_IRQ);
 747         u32 target_cpu;
 748         int numa_node;
 749
 750         if (!perf_chn ||
 751             !alloc_cpumask_var(&available_mask, GFP_KERNEL) ||
 752             cpumask_empty(hk_mask)) {
 753                 /*
 754                  * If the channel is not a performance critical
 755                  * channel, bind it to VMBUS_CONNECT_CPU.
 756                  * In case alloc_cpumask_var() fails, bind it to
 757                  * VMBUS_CONNECT_CPU.
 758                  * If all the cpus are isolated, bind it to
 759                  * VMBUS_CONNECT_CPU.
 760                  */
 761                 channel->target_cpu = VMBUS_CONNECT_CPU;
 762                 if (perf_chn)
 763                         hv_set_allocated_cpu(VMBUS_CONNECT_CPU);
 764                 return;
 765         }
 766
 767         for (i = 1; i <= ncpu + 1; i++) {
 768                 while (true) {
 769                         numa_node = next_numa_node_id++;
 770                         if (numa_node == nr_node_ids) {
 771                                 next_numa_node_id = 0;
 772                                 continue;
 773                         }
 774                         if (cpumask_empty(cpumask_of_node(numa_node)))
 775                                 continue;
 776                         break;
 777                 }
 778                 allocated_mask = &hv_context.hv_numa_map[numa_node];
 779
 780 retry:
 781                 cpumask_xor(available_mask, allocated_mask, cpumask_of_node(numa_node));
 782                 cpumask_and(available_mask, available_mask, hk_mask);
 783
 784                 if (cpumask_empty(available_mask)) {
 785                         /*
 786                          * We have cycled through all the CPUs in the node;
 787                          * reset the allocated map.
 788                          */
 789                         cpumask_clear(allocated_mask);
 790                         goto retry;
 791                 }
 792
 793                 target_cpu = cpumask_first(available_mask);
 794                 cpumask_set_cpu(target_cpu, allocated_mask);
 795
 796                 if (channel->offermsg.offer.sub_channel_index >= ncpu ||
 797                     i > ncpu || !hv_cpuself_used(target_cpu, channel))
 798                         break;
 799         }
 800
 801         channel->target_cpu = target_cpu;
 802
 803         free_cpumask_var(available_mask);
 804 }
 805
 806 #define UNLOAD_DELAY_UNIT_MS    10              /* 10 milliseconds */
 807 #define UNLOAD_WAIT_MS          (100*1000)      /* 100 seconds */
 808 #define UNLOAD_WAIT_LOOPS       (UNLOAD_WAIT_MS/UNLOAD_DELAY_UNIT_MS)
 809 #define UNLOAD_MSG_MS           (5*1000)        /* Every 5 seconds */
 810 #define UNLOAD_MSG_LOOPS        (UNLOAD_MSG_MS/UNLOAD_DELAY_UNIT_MS)
 811
 812 static void vmbus_wait_for_unload(void)
 813 {
 814         int cpu;
 815         void *page_addr;
 816         struct hv_message *msg;
 817         struct vmbus_channel_message_header *hdr;
 818         u32 message_type, i;
 819
 820         /*
 821          * CHANNELMSG_UNLOAD_RESPONSE is always delivered to the CPU which was
 822          * used for initial contact or to CPU0 depending on host version. When
 823          * we're crashing on a different CPU let's hope that IRQ handler on
 824          * the cpu which receives CHANNELMSG_UNLOAD_RESPONSE is still
 825          * functional and vmbus_unload_response() will complete
 826          * vmbus_connection.unload_event. If not, the last thing we can do is
 827          * read message pages for all CPUs directly.
 828          *
 829          * Wait up to 100 seconds since an Azure host must writeback any dirty
 830          * data in its disk cache before the VMbus UNLOAD request will
 831          * complete. This flushing has been empirically observed to take up
 832          * to 50 seconds in cases with a lot of dirty data, so allow additional
 833          * leeway and for inaccuracies in mdelay(). But eventually time out so
 834          * that the panic path can't get hung forever in case the response
 835          * message isn't seen.
 836          */
 837         for (i = 1; i <= UNLOAD_WAIT_LOOPS; i++) {
 838                 if (completion_done(&vmbus_connection.unload_event))
 839                         goto completed;
 840
 841                 for_each_present_cpu(cpu) {
 842                         struct hv_per_cpu_context *hv_cpu
 843                                 = per_cpu_ptr(hv_context.cpu_context, cpu);
 844
 845                         /*
 846                          * In a CoCo VM the synic_message_page is not allocated
 847                          * in hv_synic_alloc(). Instead it is set/cleared in
 848                          * hv_synic_enable_regs() and hv_synic_disable_regs()
 849                          * such that it is set only when the CPU is online. If
 850                          * not all present CPUs are online, the message page
 851                          * might be NULL, so skip such CPUs.
 852                          */
 853                         page_addr = hv_cpu->synic_message_page;
 854                         if (!page_addr)
 855                                 continue;
 856
 857                         msg = (struct hv_message *)page_addr
 858                                 + VMBUS_MESSAGE_SINT;
 859
 860                         message_type = READ_ONCE(msg->header.message_type);
 861                         if (message_type == HVMSG_NONE)
 862                                 continue;
 863
 864                         hdr = (struct vmbus_channel_message_header *)
 865                                 msg->u.payload;
 866
 867                         if (hdr->msgtype == CHANNELMSG_UNLOAD_RESPONSE)
 868                                 complete(&vmbus_connection.unload_event);
 869
 870                         vmbus_signal_eom(msg, message_type);
 871                 }
 872
 873                 /*
 874                  * Give a notice periodically so someone watching the
 875                  * serial output won't think it is completely hung.
 876                  */
 877                 if (!(i % UNLOAD_MSG_LOOPS))
 878                         pr_notice("Waiting for VMBus UNLOAD to complete\n");
 879
 880                 mdelay(UNLOAD_DELAY_UNIT_MS);
 881         }
 882         pr_err("Continuing even though VMBus UNLOAD did not complete\n");
 883
 884 completed:
 885         /*
 886          * We're crashing and already got the UNLOAD_RESPONSE, cleanup all
 887          * maybe-pending messages on all CPUs to be able to receive new
 888          * messages after we reconnect.
 889          */
 890         for_each_present_cpu(cpu) {
 891                 struct hv_per_cpu_context *hv_cpu
 892                         = per_cpu_ptr(hv_context.cpu_context, cpu);
 893
 894                 page_addr = hv_cpu->synic_message_page;
 895                 if (!page_addr)
 896                         continue;
 897
 898                 msg = (struct hv_message *)page_addr + VMBUS_MESSAGE_SINT;
 899                 msg->header.message_type = HVMSG_NONE;
 900         }
 901 }
 902
 903 /*
 904  * vmbus_unload_response - Handler for the unload response.
 905  */
 906 static void vmbus_unload_response(struct vmbus_channel_message_header *hdr)
 907 {
 908         /*
 909          * This is a global event; just wakeup the waiting thread.
 910          * Once we successfully unload, we can cleanup the monitor state.
 911          *
 912          * NB.  A malicious or compromised Hyper-V could send a spurious
 913          * message of type CHANNELMSG_UNLOAD_RESPONSE, and trigger a call
 914          * of the complete() below.  Make sure that unload_event has been
 915          * initialized by the time this complete() is executed.
 916          */
 917         complete(&vmbus_connection.unload_event);
 918 }
 919
 920 void vmbus_initiate_unload(bool crash)
 921 {
 922         struct vmbus_channel_message_header hdr;
 923
 924         if (xchg(&vmbus_connection.conn_state, DISCONNECTED) == DISCONNECTED)
 925                 return;
 926
 927         /* Pre-Win2012R2 hosts don't support reconnect */
 928         if (vmbus_proto_version < VERSION_WIN8_1)
 929                 return;
 930
 931         reinit_completion(&vmbus_connection.unload_event);
 932         memset(&hdr, 0, sizeof(struct vmbus_channel_message_header));
 933         hdr.msgtype = CHANNELMSG_UNLOAD;
 934         vmbus_post_msg(&hdr, sizeof(struct vmbus_channel_message_header),
 935                        !crash);
 936
 937         /*
 938          * vmbus_initiate_unload() is also called on crash and the crash can be
 939          * happening in an interrupt context, where scheduling is impossible.
 940          */
 941         if (!crash)
 942                 wait_for_completion(&vmbus_connection.unload_event);
 943         else
 944                 vmbus_wait_for_unload();
 945 }
 946
 947 static void check_ready_for_resume_event(void)
 948 {
 949         /*
 950          * If all the old primary channels have been fixed up, then it's safe
 951          * to resume.
 952          */
 953         if (atomic_dec_and_test(&vmbus_connection.nr_chan_fixup_on_resume))
 954                 complete(&vmbus_connection.ready_for_resume_event);
 955 }
 956
 957 static void vmbus_setup_channel_state(struct vmbus_channel *channel,
 958                                       struct vmbus_channel_offer_channel *offer)
 959 {
 960         /*
 961          * Setup state for signalling the host.
 962          */
 963         channel->sig_event = VMBUS_EVENT_CONNECTION_ID;
 964
 965         channel->is_dedicated_interrupt =
 966                         (offer->is_dedicated_interrupt != 0);
 967         channel->sig_event = offer->connection_id;
 968
 969         memcpy(&channel->offermsg, offer,
 970                sizeof(struct vmbus_channel_offer_channel));
 971         channel->monitor_grp = (u8)offer->monitorid / 32;
 972         channel->monitor_bit = (u8)offer->monitorid % 32;
 973         channel->device_id = hv_get_dev_type(channel);
 974 }
 975
 976 /*
 977  * find_primary_channel_by_offer - Get the channel object given the new offer.
 978  * This is only used in the resume path of hibernation.
 979  */
 980 static struct vmbus_channel *
 981 find_primary_channel_by_offer(const struct vmbus_channel_offer_channel *offer)
 982 {
 983         struct vmbus_channel *channel = NULL, *iter;
 984         const guid_t *inst1, *inst2;
 985
 986         /* Ignore sub-channel offers. */
 987         if (offer->offer.sub_channel_index != 0)
 988                 return NULL;
 989
 990         mutex_lock(&vmbus_connection.channel_mutex);
 991
 992         list_for_each_entry(iter, &vmbus_connection.chn_list, listentry) {
 993                 inst1 = &iter->offermsg.offer.if_instance;
 994                 inst2 = &offer->offer.if_instance;
 995
 996                 if (guid_equal(inst1, inst2)) {
 997                         channel = iter;
 998                         break;
 999                 }
1000         }
1001
1002         mutex_unlock(&vmbus_connection.channel_mutex);
1003
1004         return channel;
1005 }
1006
1007 static bool vmbus_is_valid_offer(const struct vmbus_channel_offer_channel *offer)
1008 {
1009         const guid_t *guid = &offer->offer.if_type;
1010         u16 i;
1011
1012         if (!hv_is_isolation_supported())
1013                 return true;
1014
1015         if (is_hvsock_offer(offer))
1016                 return true;
1017
1018         for (i = 0; i < ARRAY_SIZE(vmbus_devs); i++) {
1019                 if (guid_equal(guid, &vmbus_devs[i].guid))
1020                         return vmbus_devs[i].allowed_in_isolated;
1021         }
1022         return false;
1023 }
1024
1025 /*
1026  * vmbus_onoffer - Handler for channel offers from vmbus in parent partition.
1027  *
1028  */
1029 static void vmbus_onoffer(struct vmbus_channel_message_header *hdr)
1030 {
1031         struct vmbus_channel_offer_channel *offer;
1032         struct vmbus_channel *oldchannel, *newchannel;
1033         size_t offer_sz;
1034
1035         offer = (struct vmbus_channel_offer_channel *)hdr;
1036
1037         trace_vmbus_onoffer(offer);
1038
1039         if (!vmbus_is_valid_offer(offer)) {
1040                 pr_err_ratelimited("Invalid offer %d from the host supporting isolation\n",
1041                                    offer->child_relid);
1042                 atomic_dec(&vmbus_connection.offer_in_progress);
1043                 return;
1044         }
1045
1046         oldchannel = find_primary_channel_by_offer(offer);
1047
1048         if (oldchannel != NULL) {
1049                 /*
1050                  * We're resuming from hibernation: all the sub-channel and
1051                  * hv_sock channels we had before the hibernation should have
1052                  * been cleaned up, and now we must be seeing a re-offered
1053                  * primary channel that we had before the hibernation.
1054                  */
1055
1056                 /*
1057                  * { Initially: channel relid = INVALID_RELID,
1058                  *              channels[valid_relid] = NULL }
1059                  *
1060                  * CPU1                                 CPU2
1061                  *
1062                  * [vmbus_onoffer()]                    [vmbus_device_release()]
1063                  *
1064                  * LOCK channel_mutex                   LOCK channel_mutex
1065                  * STORE channel relid = valid_relid    LOAD r1 = channel relid
1066                  * MAP_RELID channel                    if (r1 != INVALID_RELID)
1067                  * UNLOCK channel_mutex                   UNMAP_RELID channel
1068                  *                                      UNLOCK channel_mutex
1069                  *
1070                  * Forbids: r1 == valid_relid &&
1071                  *              channels[valid_relid] == channel
1072                  *
1073                  * Note.  r1 can be INVALID_RELID only for an hv_sock channel.
1074                  * None of the hv_sock channels which were present before the
1075                  * suspend are re-offered upon the resume.  See the WARN_ON()
1076                  * in hv_process_channel_removal().
1077                  */
1078                 mutex_lock(&vmbus_connection.channel_mutex);
1079
1080                 atomic_dec(&vmbus_connection.offer_in_progress);
1081
1082                 WARN_ON(oldchannel->offermsg.child_relid != INVALID_RELID);
1083                 /* Fix up the relid. */
1084                 oldchannel->offermsg.child_relid = offer->child_relid;
1085
1086                 offer_sz = sizeof(*offer);
1087                 if (memcmp(offer, &oldchannel->offermsg, offer_sz) != 0) {
1088                         /*
1089                          * This is not an error, since the host can also change
1090                          * the other field(s) of the offer, e.g. on WS RS5
1091                          * (Build 17763), the offer->connection_id of the
1092                          * Mellanox VF vmbus device can change when the host
1093                          * reoffers the device upon resume.
1094                          */
1095                         pr_debug("vmbus offer changed: relid=%d\n",
1096                                  offer->child_relid);
1097
1098                         print_hex_dump_debug("Old vmbus offer: ",
1099                                              DUMP_PREFIX_OFFSET, 16, 4,
1100                                              &oldchannel->offermsg, offer_sz,
1101                                              false);
1102                         print_hex_dump_debug("New vmbus offer: ",
1103                                              DUMP_PREFIX_OFFSET, 16, 4,
1104                                              offer, offer_sz, false);
1105
1106                         /* Fix up the old channel. */
1107                         vmbus_setup_channel_state(oldchannel, offer);
1108                 }
1109
1110                 /* Add the channel back to the array of channels. */
1111                 vmbus_channel_map_relid(oldchannel);
1112                 check_ready_for_resume_event();
1113
1114                 mutex_unlock(&vmbus_connection.channel_mutex);
1115                 return;
1116         }
1117
1118         /* Allocate the channel object and save this offer. */
1119         newchannel = alloc_channel();
1120         if (!newchannel) {
1121                 vmbus_release_relid(offer->child_relid);
1122                 atomic_dec(&vmbus_connection.offer_in_progress);
1123                 pr_err("Unable to allocate channel object\n");
1124                 return;
1125         }
1126
1127         vmbus_setup_channel_state(newchannel, offer);
1128
1129         vmbus_process_offer(newchannel);
1130 }
1131
1132 static void check_ready_for_suspend_event(void)
1133 {
1134         /*
1135          * If all the sub-channels or hv_sock channels have been cleaned up,
1136          * then it's safe to suspend.
1137          */
1138         if (atomic_dec_and_test(&vmbus_connection.nr_chan_close_on_suspend))
1139                 complete(&vmbus_connection.ready_for_suspend_event);
1140 }
1141
1142 /*
1143  * vmbus_onoffer_rescind - Rescind offer handler.
1144  *
1145  * We queue a work item to process this offer synchronously
1146  */
1147 static void vmbus_onoffer_rescind(struct vmbus_channel_message_header *hdr)
1148 {
1149         struct vmbus_channel_rescind_offer *rescind;
1150         struct vmbus_channel *channel;
1151         struct device *dev;
1152         bool clean_up_chan_for_suspend;
1153
1154         rescind = (struct vmbus_channel_rescind_offer *)hdr;
1155
1156         trace_vmbus_onoffer_rescind(rescind);
1157
1158         /*
1159          * The offer msg and the corresponding rescind msg
1160          * from the host are guranteed to be ordered -
1161          * offer comes in first and then the rescind.
1162          * Since we process these events in work elements,
1163          * and with preemption, we may end up processing
1164          * the events out of order.  We rely on the synchronization
1165          * provided by offer_in_progress and by channel_mutex for
1166          * ordering these events:
1167          *
1168          * { Initially: offer_in_progress = 1 }
1169          *
1170          * CPU1                         CPU2
1171          *
1172          * [vmbus_onoffer()]            [vmbus_onoffer_rescind()]
1173          *
1174          * LOCK channel_mutex           WAIT_ON offer_in_progress == 0
1175          * DECREMENT offer_in_progress  LOCK channel_mutex
1176          * STORE channels[]             LOAD channels[]
1177          * UNLOCK channel_mutex         UNLOCK channel_mutex
1178          *
1179          * Forbids: CPU2's LOAD from *not* seeing CPU1's STORE
1180          */
1181
1182         while (atomic_read(&vmbus_connection.offer_in_progress) != 0) {
1183                 /*
1184                  * We wait here until any channel offer is currently
1185                  * being processed.
1186                  */
1187                 msleep(1);
1188         }
1189
1190         mutex_lock(&vmbus_connection.channel_mutex);
1191         channel = relid2channel(rescind->child_relid);
1192         if (channel != NULL) {
1193                 /*
1194                  * Guarantee that no other instance of vmbus_onoffer_rescind()
1195                  * has got a reference to the channel object.  Synchronize on
1196                  * &vmbus_connection.channel_mutex.
1197                  */
1198                 if (channel->rescind_ref) {
1199                         mutex_unlock(&vmbus_connection.channel_mutex);
1200                         return;
1201                 }
1202                 channel->rescind_ref = true;
1203         }
1204         mutex_unlock(&vmbus_connection.channel_mutex);
1205
1206         if (channel == NULL) {
1207                 /*
1208                  * We failed in processing the offer message;
1209                  * we would have cleaned up the relid in that
1210                  * failure path.
1211                  */
1212                 return;
1213         }
1214
1215         clean_up_chan_for_suspend = is_hvsock_channel(channel) ||
1216                                     is_sub_channel(channel);
1217         /*
1218          * Before setting channel->rescind in vmbus_rescind_cleanup(), we
1219          * should make sure the channel callback is not running any more.
1220          */
1221         vmbus_reset_channel_cb(channel);
1222
1223         /*
1224          * Now wait for offer handling to complete.
1225          */
1226         vmbus_rescind_cleanup(channel);
1227         while (READ_ONCE(channel->probe_done) == false) {
1228                 /*
1229                  * We wait here until any channel offer is currently
1230                  * being processed.
1231                  */
1232                 msleep(1);
1233         }
1234
1235         /*
1236          * At this point, the rescind handling can proceed safely.
1237          */
1238
1239         if (channel->device_obj) {
1240                 if (channel->chn_rescind_callback) {
1241                         channel->chn_rescind_callback(channel);
1242
1243                         if (clean_up_chan_for_suspend)
1244                                 check_ready_for_suspend_event();
1245
1246                         return;
1247                 }
1248                 /*
1249                  * We will have to unregister this device from the
1250                  * driver core.
1251                  */
1252                 dev = get_device(&channel->device_obj->device);
1253                 if (dev) {
1254                         vmbus_device_unregister(channel->device_obj);
1255                         put_device(dev);
1256                 }
1257         } else if (channel->primary_channel != NULL) {
1258                 /*
1259                  * Sub-channel is being rescinded. Following is the channel
1260                  * close sequence when initiated from the driveri (refer to
1261                  * vmbus_close() for details):
1262                  * 1. Close all sub-channels first
1263                  * 2. Then close the primary channel.
1264                  */
1265                 mutex_lock(&vmbus_connection.channel_mutex);
1266                 if (channel->state == CHANNEL_OPEN_STATE) {
1267                         /*
1268                          * The channel is currently not open;
1269                          * it is safe for us to cleanup the channel.
1270                          */
1271                         hv_process_channel_removal(channel);
1272                 } else {
1273                         complete(&channel->rescind_event);
1274                 }
1275                 mutex_unlock(&vmbus_connection.channel_mutex);
1276         }
1277
1278         /* The "channel" may have been freed. Do not access it any longer. */
1279
1280         if (clean_up_chan_for_suspend)
1281                 check_ready_for_suspend_event();
1282 }
1283
1284 void vmbus_hvsock_device_unregister(struct vmbus_channel *channel)
1285 {
1286         BUG_ON(!is_hvsock_channel(channel));
1287
1288         /* We always get a rescind msg when a connection is closed. */
1289         while (!READ_ONCE(channel->probe_done) || !READ_ONCE(channel->rescind))
1290                 msleep(1);
1291
1292         vmbus_device_unregister(channel->device_obj);
1293 }
1294 EXPORT_SYMBOL_GPL(vmbus_hvsock_device_unregister);
1295
1296
1297 /*
1298  * vmbus_onoffers_delivered -
1299  * This is invoked when all offers have been delivered.
1300  *
1301  * Nothing to do here.
1302  */
1303 static void vmbus_onoffers_delivered(
1304                         struct vmbus_channel_message_header *hdr)
1305 {
1306 }
1307
1308 /*
1309  * vmbus_onopen_result - Open result handler.
1310  *
1311  * This is invoked when we received a response to our channel open request.
1312  * Find the matching request, copy the response and signal the requesting
1313  * thread.
1314  */
1315 static void vmbus_onopen_result(struct vmbus_channel_message_header *hdr)
1316 {
1317         struct vmbus_channel_open_result *result;
1318         struct vmbus_channel_msginfo *msginfo;
1319         struct vmbus_channel_message_header *requestheader;
1320         struct vmbus_channel_open_channel *openmsg;
1321         unsigned long flags;
1322
1323         result = (struct vmbus_channel_open_result *)hdr;
1324
1325         trace_vmbus_onopen_result(result);
1326
1327         /*
1328          * Find the open msg, copy the result and signal/unblock the wait event
1329          */
1330         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1331
1332         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1333                                 msglistentry) {
1334                 requestheader =
1335                         (struct vmbus_channel_message_header *)msginfo->msg;
1336
1337                 if (requestheader->msgtype == CHANNELMSG_OPENCHANNEL) {
1338                         openmsg =
1339                         (struct vmbus_channel_open_channel *)msginfo->msg;
1340                         if (openmsg->child_relid == result->child_relid &&
1341                             openmsg->openid == result->openid) {
1342                                 memcpy(&msginfo->response.open_result,
1343                                        result,
1344                                        sizeof(
1345                                         struct vmbus_channel_open_result));
1346                                 complete(&msginfo->waitevent);
1347                                 break;
1348                         }
1349                 }
1350         }
1351         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1352 }
1353
1354 /*
1355  * vmbus_ongpadl_created - GPADL created handler.
1356  *
1357  * This is invoked when we received a response to our gpadl create request.
1358  * Find the matching request, copy the response and signal the requesting
1359  * thread.
1360  */
1361 static void vmbus_ongpadl_created(struct vmbus_channel_message_header *hdr)
1362 {
1363         struct vmbus_channel_gpadl_created *gpadlcreated;
1364         struct vmbus_channel_msginfo *msginfo;
1365         struct vmbus_channel_message_header *requestheader;
1366         struct vmbus_channel_gpadl_header *gpadlheader;
1367         unsigned long flags;
1368
1369         gpadlcreated = (struct vmbus_channel_gpadl_created *)hdr;
1370
1371         trace_vmbus_ongpadl_created(gpadlcreated);
1372
1373         /*
1374          * Find the establish msg, copy the result and signal/unblock the wait
1375          * event
1376          */
1377         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1378
1379         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1380                                 msglistentry) {
1381                 requestheader =
1382                         (struct vmbus_channel_message_header *)msginfo->msg;
1383
1384                 if (requestheader->msgtype == CHANNELMSG_GPADL_HEADER) {
1385                         gpadlheader =
1386                         (struct vmbus_channel_gpadl_header *)requestheader;
1387
1388                         if ((gpadlcreated->child_relid ==
1389                              gpadlheader->child_relid) &&
1390                             (gpadlcreated->gpadl == gpadlheader->gpadl)) {
1391                                 memcpy(&msginfo->response.gpadl_created,
1392                                        gpadlcreated,
1393                                        sizeof(
1394                                         struct vmbus_channel_gpadl_created));
1395                                 complete(&msginfo->waitevent);
1396                                 break;
1397                         }
1398                 }
1399         }
1400         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1401 }
1402
1403 /*
1404  * vmbus_onmodifychannel_response - Modify Channel response handler.
1405  *
1406  * This is invoked when we received a response to our channel modify request.
1407  * Find the matching request, copy the response and signal the requesting thread.
1408  */
1409 static void vmbus_onmodifychannel_response(struct vmbus_channel_message_header *hdr)
1410 {
1411         struct vmbus_channel_modifychannel_response *response;
1412         struct vmbus_channel_msginfo *msginfo;
1413         unsigned long flags;
1414
1415         response = (struct vmbus_channel_modifychannel_response *)hdr;
1416
1417         trace_vmbus_onmodifychannel_response(response);
1418
1419         /*
1420          * Find the modify msg, copy the response and signal/unblock the wait event.
1421          */
1422         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1423
1424         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list, msglistentry) {
1425                 struct vmbus_channel_message_header *responseheader =
1426                                 (struct vmbus_channel_message_header *)msginfo->msg;
1427
1428                 if (responseheader->msgtype == CHANNELMSG_MODIFYCHANNEL) {
1429                         struct vmbus_channel_modifychannel *modifymsg;
1430
1431                         modifymsg = (struct vmbus_channel_modifychannel *)msginfo->msg;
1432                         if (modifymsg->child_relid == response->child_relid) {
1433                                 memcpy(&msginfo->response.modify_response, response,
1434                                        sizeof(*response));
1435                                 complete(&msginfo->waitevent);
1436                                 break;
1437                         }
1438                 }
1439         }
1440         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1441 }
1442
1443 /*
1444  * vmbus_ongpadl_torndown - GPADL torndown handler.
1445  *
1446  * This is invoked when we received a response to our gpadl teardown request.
1447  * Find the matching request, copy the response and signal the requesting
1448  * thread.
1449  */
1450 static void vmbus_ongpadl_torndown(
1451                         struct vmbus_channel_message_header *hdr)
1452 {
1453         struct vmbus_channel_gpadl_torndown *gpadl_torndown;
1454         struct vmbus_channel_msginfo *msginfo;
1455         struct vmbus_channel_message_header *requestheader;
1456         struct vmbus_channel_gpadl_teardown *gpadl_teardown;
1457         unsigned long flags;
1458
1459         gpadl_torndown = (struct vmbus_channel_gpadl_torndown *)hdr;
1460
1461         trace_vmbus_ongpadl_torndown(gpadl_torndown);
1462
1463         /*
1464          * Find the open msg, copy the result and signal/unblock the wait event
1465          */
1466         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1467
1468         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1469                                 msglistentry) {
1470                 requestheader =
1471                         (struct vmbus_channel_message_header *)msginfo->msg;
1472
1473                 if (requestheader->msgtype == CHANNELMSG_GPADL_TEARDOWN) {
1474                         gpadl_teardown =
1475                         (struct vmbus_channel_gpadl_teardown *)requestheader;
1476
1477                         if (gpadl_torndown->gpadl == gpadl_teardown->gpadl) {
1478                                 memcpy(&msginfo->response.gpadl_torndown,
1479                                        gpadl_torndown,
1480                                        sizeof(
1481                                         struct vmbus_channel_gpadl_torndown));
1482                                 complete(&msginfo->waitevent);
1483                                 break;
1484                         }
1485                 }
1486         }
1487         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1488 }
1489
1490 /*
1491  * vmbus_onversion_response - Version response handler
1492  *
1493  * This is invoked when we received a response to our initiate contact request.
1494  * Find the matching request, copy the response and signal the requesting
1495  * thread.
1496  */
1497 static void vmbus_onversion_response(
1498                 struct vmbus_channel_message_header *hdr)
1499 {
1500         struct vmbus_channel_msginfo *msginfo;
1501         struct vmbus_channel_message_header *requestheader;
1502         struct vmbus_channel_version_response *version_response;
1503         unsigned long flags;
1504
1505         version_response = (struct vmbus_channel_version_response *)hdr;
1506
1507         trace_vmbus_onversion_response(version_response);
1508
1509         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1510
1511         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1512                                 msglistentry) {
1513                 requestheader =
1514                         (struct vmbus_channel_message_header *)msginfo->msg;
1515
1516                 if (requestheader->msgtype ==
1517                     CHANNELMSG_INITIATE_CONTACT) {
1518                         memcpy(&msginfo->response.version_response,
1519                               version_response,
1520                               sizeof(struct vmbus_channel_version_response));
1521                         complete(&msginfo->waitevent);
1522                 }
1523         }
1524         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1525 }
1526
1527 /* Channel message dispatch table */
1528 const struct vmbus_channel_message_table_entry
1529 channel_message_table[CHANNELMSG_COUNT] = {
1530         { CHANNELMSG_INVALID,                   0, NULL, 0},
1531         { CHANNELMSG_OFFERCHANNEL,              0, vmbus_onoffer,
1532                 sizeof(struct vmbus_channel_offer_channel)},
1533         { CHANNELMSG_RESCIND_CHANNELOFFER,      0, vmbus_onoffer_rescind,
1534                 sizeof(struct vmbus_channel_rescind_offer) },
1535         { CHANNELMSG_REQUESTOFFERS,             0, NULL, 0},
1536         { CHANNELMSG_ALLOFFERS_DELIVERED,       1, vmbus_onoffers_delivered, 0},
1537         { CHANNELMSG_OPENCHANNEL,               0, NULL, 0},
1538         { CHANNELMSG_OPENCHANNEL_RESULT,        1, vmbus_onopen_result,
1539                 sizeof(struct vmbus_channel_open_result)},
1540         { CHANNELMSG_CLOSECHANNEL,              0, NULL, 0},
1541         { CHANNELMSG_GPADL_HEADER,              0, NULL, 0},
1542         { CHANNELMSG_GPADL_BODY,                0, NULL, 0},
1543         { CHANNELMSG_GPADL_CREATED,             1, vmbus_ongpadl_created,
1544                 sizeof(struct vmbus_channel_gpadl_created)},
1545         { CHANNELMSG_GPADL_TEARDOWN,            0, NULL, 0},
1546         { CHANNELMSG_GPADL_TORNDOWN,            1, vmbus_ongpadl_torndown,
1547                 sizeof(struct vmbus_channel_gpadl_torndown) },
1548         { CHANNELMSG_RELID_RELEASED,            0, NULL, 0},
1549         { CHANNELMSG_INITIATE_CONTACT,          0, NULL, 0},
1550         { CHANNELMSG_VERSION_RESPONSE,          1, vmbus_onversion_response,
1551                 sizeof(struct vmbus_channel_version_response)},
1552         { CHANNELMSG_UNLOAD,                    0, NULL, 0},
1553         { CHANNELMSG_UNLOAD_RESPONSE,           1, vmbus_unload_response, 0},
1554         { CHANNELMSG_18,                        0, NULL, 0},
1555         { CHANNELMSG_19,                        0, NULL, 0},
1556         { CHANNELMSG_20,                        0, NULL, 0},
1557         { CHANNELMSG_TL_CONNECT_REQUEST,        0, NULL, 0},
1558         { CHANNELMSG_MODIFYCHANNEL,             0, NULL, 0},
1559         { CHANNELMSG_TL_CONNECT_RESULT,         0, NULL, 0},
1560         { CHANNELMSG_MODIFYCHANNEL_RESPONSE,    1, vmbus_onmodifychannel_response,
1561                 sizeof(struct vmbus_channel_modifychannel_response)},
1562 };
1563
1564 /*
1565  * vmbus_onmessage - Handler for channel protocol messages.
1566  *
1567  * This is invoked in the vmbus worker thread context.
1568  */
1569 void vmbus_onmessage(struct vmbus_channel_message_header *hdr)
1570 {
1571         trace_vmbus_on_message(hdr);
1572
1573         /*
1574          * vmbus_on_msg_dpc() makes sure the hdr->msgtype here can not go
1575          * out of bound and the message_handler pointer can not be NULL.
1576          */
1577         channel_message_table[hdr->msgtype].message_handler(hdr);
1578 }
1579
1580 /*
1581  * vmbus_request_offers - Send a request to get all our pending offers.
1582  */
1583 int vmbus_request_offers(void)
1584 {
1585         struct vmbus_channel_message_header *msg;
1586         struct vmbus_channel_msginfo *msginfo;
1587         int ret;
1588
1589         msginfo = kzalloc(sizeof(*msginfo) +
1590                           sizeof(struct vmbus_channel_message_header),
1591                           GFP_KERNEL);
1592         if (!msginfo)
1593                 return -ENOMEM;
1594
1595         msg = (struct vmbus_channel_message_header *)msginfo->msg;
1596
1597         msg->msgtype = CHANNELMSG_REQUESTOFFERS;
1598
1599         ret = vmbus_post_msg(msg, sizeof(struct vmbus_channel_message_header),
1600                              true);
1601
1602         trace_vmbus_request_offers(ret);
1603
1604         if (ret != 0) {
1605                 pr_err("Unable to request offers - %d\n", ret);
1606
1607                 goto cleanup;
1608         }
1609
1610 cleanup:
1611         kfree(msginfo);
1612
1613         return ret;
1614 }
1615
1616 void vmbus_set_sc_create_callback(struct vmbus_channel *primary_channel,
1617                                 void (*sc_cr_cb)(struct vmbus_channel *new_sc))
1618 {
1619         primary_channel->sc_creation_callback = sc_cr_cb;
1620 }
1621 EXPORT_SYMBOL_GPL(vmbus_set_sc_create_callback);
1622
1623 void vmbus_set_chn_rescind_callback(struct vmbus_channel *channel,
1624                 void (*chn_rescind_cb)(struct vmbus_channel *))
1625 {
1626         channel->chn_rescind_callback = chn_rescind_cb;
1627 }
1628 EXPORT_SYMBOL_GPL(vmbus_set_chn_rescind_callback);