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/hyperv.h>
  22 #include <asm/mshyperv.h>
  23
  24 #include "hyperv_vmbus.h"
  25
  26 static void init_vp_index(struct vmbus_channel *channel, u16 dev_type);
  27
  28 static const struct vmbus_device vmbus_devs[] = {
  29         /* IDE */
  30         { .dev_type = HV_IDE,
  31           HV_IDE_GUID,
  32           .perf_device = true,
  33         },
  34
  35         /* SCSI */
  36         { .dev_type = HV_SCSI,
  37           HV_SCSI_GUID,
  38           .perf_device = true,
  39         },
  40
  41         /* Fibre Channel */
  42         { .dev_type = HV_FC,
  43           HV_SYNTHFC_GUID,
  44           .perf_device = true,
  45         },
  46
  47         /* Synthetic NIC */
  48         { .dev_type = HV_NIC,
  49           HV_NIC_GUID,
  50           .perf_device = true,
  51         },
  52
  53         /* Network Direct */
  54         { .dev_type = HV_ND,
  55           HV_ND_GUID,
  56           .perf_device = true,
  57         },
  58
  59         /* PCIE */
  60         { .dev_type = HV_PCIE,
  61           HV_PCIE_GUID,
  62           .perf_device = false,
  63         },
  64
  65         /* Synthetic Frame Buffer */
  66         { .dev_type = HV_FB,
  67           HV_SYNTHVID_GUID,
  68           .perf_device = false,
  69         },
  70
  71         /* Synthetic Keyboard */
  72         { .dev_type = HV_KBD,
  73           HV_KBD_GUID,
  74           .perf_device = false,
  75         },
  76
  77         /* Synthetic MOUSE */
  78         { .dev_type = HV_MOUSE,
  79           HV_MOUSE_GUID,
  80           .perf_device = false,
  81         },
  82
  83         /* KVP */
  84         { .dev_type = HV_KVP,
  85           HV_KVP_GUID,
  86           .perf_device = false,
  87         },
  88
  89         /* Time Synch */
  90         { .dev_type = HV_TS,
  91           HV_TS_GUID,
  92           .perf_device = false,
  93         },
  94
  95         /* Heartbeat */
  96         { .dev_type = HV_HB,
  97           HV_HEART_BEAT_GUID,
  98           .perf_device = false,
  99         },
 100
 101         /* Shutdown */
 102         { .dev_type = HV_SHUTDOWN,
 103           HV_SHUTDOWN_GUID,
 104           .perf_device = false,
 105         },
 106
 107         /* File copy */
 108         { .dev_type = HV_FCOPY,
 109           HV_FCOPY_GUID,
 110           .perf_device = false,
 111         },
 112
 113         /* Backup */
 114         { .dev_type = HV_BACKUP,
 115           HV_VSS_GUID,
 116           .perf_device = false,
 117         },
 118
 119         /* Dynamic Memory */
 120         { .dev_type = HV_DM,
 121           HV_DM_GUID,
 122           .perf_device = false,
 123         },
 124
 125         /* Unknown GUID */
 126         { .dev_type = HV_UNKNOWN,
 127           .perf_device = false,
 128         },
 129 };
 130
 131 static const struct {
 132         guid_t guid;
 133 } vmbus_unsupported_devs[] = {
 134         { HV_AVMA1_GUID },
 135         { HV_AVMA2_GUID },
 136         { HV_RDV_GUID   },
 137 };
 138
 139 /*
 140  * The rescinded channel may be blocked waiting for a response from the host;
 141  * take care of that.
 142  */
 143 static void vmbus_rescind_cleanup(struct vmbus_channel *channel)
 144 {
 145         struct vmbus_channel_msginfo *msginfo;
 146         unsigned long flags;
 147
 148
 149         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
 150         channel->rescind = true;
 151         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
 152                                 msglistentry) {
 153
 154                 if (msginfo->waiting_channel == channel) {
 155                         complete(&msginfo->waitevent);
 156                         break;
 157                 }
 158         }
 159         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
 160 }
 161
 162 static bool is_unsupported_vmbus_devs(const guid_t *guid)
 163 {
 164         int i;
 165
 166         for (i = 0; i < ARRAY_SIZE(vmbus_unsupported_devs); i++)
 167                 if (guid_equal(guid, &vmbus_unsupported_devs[i].guid))
 168                         return true;
 169         return false;
 170 }
 171
 172 static u16 hv_get_dev_type(const struct vmbus_channel *channel)
 173 {
 174         const guid_t *guid = &channel->offermsg.offer.if_type;
 175         u16 i;
 176
 177         if (is_hvsock_channel(channel) || is_unsupported_vmbus_devs(guid))
 178                 return HV_UNKNOWN;
 179
 180         for (i = HV_IDE; i < HV_UNKNOWN; i++) {
 181                 if (guid_equal(guid, &vmbus_devs[i].guid))
 182                         return i;
 183         }
 184         pr_info("Unknown GUID: %pUl\n", guid);
 185         return i;
 186 }
 187
 188 /**
 189  * vmbus_prep_negotiate_resp() - Create default response for Negotiate message
 190  * @icmsghdrp: Pointer to msg header structure
 191  * @buf: Raw buffer channel data
 192  * @fw_version: The framework versions we can support.
 193  * @fw_vercnt: The size of @fw_version.
 194  * @srv_version: The service versions we can support.
 195  * @srv_vercnt: The size of @srv_version.
 196  * @nego_fw_version: The selected framework version.
 197  * @nego_srv_version: The selected service version.
 198  *
 199  * Note: Versions are given in decreasing order.
 200  *
 201  * Set up and fill in default negotiate response message.
 202  * Mainly used by Hyper-V drivers.
 203  */
 204 bool vmbus_prep_negotiate_resp(struct icmsg_hdr *icmsghdrp,
 205                                 u8 *buf, const int *fw_version, int fw_vercnt,
 206                                 const int *srv_version, int srv_vercnt,
 207                                 int *nego_fw_version, int *nego_srv_version)
 208 {
 209         int icframe_major, icframe_minor;
 210         int icmsg_major, icmsg_minor;
 211         int fw_major, fw_minor;
 212         int srv_major, srv_minor;
 213         int i, j;
 214         bool found_match = false;
 215         struct icmsg_negotiate *negop;
 216
 217         icmsghdrp->icmsgsize = 0x10;
 218         negop = (struct icmsg_negotiate *)&buf[
 219                 sizeof(struct vmbuspipe_hdr) +
 220                 sizeof(struct icmsg_hdr)];
 221
 222         icframe_major = negop->icframe_vercnt;
 223         icframe_minor = 0;
 224
 225         icmsg_major = negop->icmsg_vercnt;
 226         icmsg_minor = 0;
 227
 228         /*
 229          * Select the framework version number we will
 230          * support.
 231          */
 232
 233         for (i = 0; i < fw_vercnt; i++) {
 234                 fw_major = (fw_version[i] >> 16);
 235                 fw_minor = (fw_version[i] & 0xFFFF);
 236
 237                 for (j = 0; j < negop->icframe_vercnt; j++) {
 238                         if ((negop->icversion_data[j].major == fw_major) &&
 239                             (negop->icversion_data[j].minor == fw_minor)) {
 240                                 icframe_major = negop->icversion_data[j].major;
 241                                 icframe_minor = negop->icversion_data[j].minor;
 242                                 found_match = true;
 243                                 break;
 244                         }
 245                 }
 246
 247                 if (found_match)
 248                         break;
 249         }
 250
 251         if (!found_match)
 252                 goto fw_error;
 253
 254         found_match = false;
 255
 256         for (i = 0; i < srv_vercnt; i++) {
 257                 srv_major = (srv_version[i] >> 16);
 258                 srv_minor = (srv_version[i] & 0xFFFF);
 259
 260                 for (j = negop->icframe_vercnt;
 261                         (j < negop->icframe_vercnt + negop->icmsg_vercnt);
 262                         j++) {
 263
 264                         if ((negop->icversion_data[j].major == srv_major) &&
 265                                 (negop->icversion_data[j].minor == srv_minor)) {
 266
 267                                 icmsg_major = negop->icversion_data[j].major;
 268                                 icmsg_minor = negop->icversion_data[j].minor;
 269                                 found_match = true;
 270                                 break;
 271                         }
 272                 }
 273
 274                 if (found_match)
 275                         break;
 276         }
 277
 278         /*
 279          * Respond with the framework and service
 280          * version numbers we can support.
 281          */
 282
 283 fw_error:
 284         if (!found_match) {
 285                 negop->icframe_vercnt = 0;
 286                 negop->icmsg_vercnt = 0;
 287         } else {
 288                 negop->icframe_vercnt = 1;
 289                 negop->icmsg_vercnt = 1;
 290         }
 291
 292         if (nego_fw_version)
 293                 *nego_fw_version = (icframe_major << 16) | icframe_minor;
 294
 295         if (nego_srv_version)
 296                 *nego_srv_version = (icmsg_major << 16) | icmsg_minor;
 297
 298         negop->icversion_data[0].major = icframe_major;
 299         negop->icversion_data[0].minor = icframe_minor;
 300         negop->icversion_data[1].major = icmsg_major;
 301         negop->icversion_data[1].minor = icmsg_minor;
 302         return found_match;
 303 }
 304
 305 EXPORT_SYMBOL_GPL(vmbus_prep_negotiate_resp);
 306
 307 /*
 308  * alloc_channel - Allocate and initialize a vmbus channel object
 309  */
 310 static struct vmbus_channel *alloc_channel(void)
 311 {
 312         struct vmbus_channel *channel;
 313
 314         channel = kzalloc(sizeof(*channel), GFP_ATOMIC);
 315         if (!channel)
 316                 return NULL;
 317
 318         spin_lock_init(&channel->lock);
 319         init_completion(&channel->rescind_event);
 320
 321         INIT_LIST_HEAD(&channel->sc_list);
 322         INIT_LIST_HEAD(&channel->percpu_list);
 323
 324         tasklet_init(&channel->callback_event,
 325                      vmbus_on_event, (unsigned long)channel);
 326
 327         hv_ringbuffer_pre_init(channel);
 328
 329         return channel;
 330 }
 331
 332 /*
 333  * free_channel - Release the resources used by the vmbus channel object
 334  */
 335 static void free_channel(struct vmbus_channel *channel)
 336 {
 337         tasklet_kill(&channel->callback_event);
 338         vmbus_remove_channel_attr_group(channel);
 339
 340         kobject_put(&channel->kobj);
 341 }
 342
 343 static void percpu_channel_enq(void *arg)
 344 {
 345         struct vmbus_channel *channel = arg;
 346         struct hv_per_cpu_context *hv_cpu
 347                 = this_cpu_ptr(hv_context.cpu_context);
 348
 349         list_add_tail_rcu(&channel->percpu_list, &hv_cpu->chan_list);
 350 }
 351
 352 static void percpu_channel_deq(void *arg)
 353 {
 354         struct vmbus_channel *channel = arg;
 355
 356         list_del_rcu(&channel->percpu_list);
 357 }
 358
 359
 360 static void vmbus_release_relid(u32 relid)
 361 {
 362         struct vmbus_channel_relid_released msg;
 363         int ret;
 364
 365         memset(&msg, 0, sizeof(struct vmbus_channel_relid_released));
 366         msg.child_relid = relid;
 367         msg.header.msgtype = CHANNELMSG_RELID_RELEASED;
 368         ret = vmbus_post_msg(&msg, sizeof(struct vmbus_channel_relid_released),
 369                              true);
 370
 371         trace_vmbus_release_relid(&msg, ret);
 372 }
 373
 374 void hv_process_channel_removal(struct vmbus_channel *channel)
 375 {
 376         struct vmbus_channel *primary_channel;
 377         unsigned long flags;
 378
 379         BUG_ON(!mutex_is_locked(&vmbus_connection.channel_mutex));
 380         BUG_ON(!channel->rescind);
 381
 382         if (channel->target_cpu != get_cpu()) {
 383                 put_cpu();
 384                 smp_call_function_single(channel->target_cpu,
 385                                          percpu_channel_deq, channel, true);
 386         } else {
 387                 percpu_channel_deq(channel);
 388                 put_cpu();
 389         }
 390
 391         if (channel->primary_channel == NULL) {
 392                 list_del(&channel->listentry);
 393
 394                 primary_channel = channel;
 395         } else {
 396                 primary_channel = channel->primary_channel;
 397                 spin_lock_irqsave(&primary_channel->lock, flags);
 398                 list_del(&channel->sc_list);
 399                 spin_unlock_irqrestore(&primary_channel->lock, flags);
 400         }
 401
 402         /*
 403          * We need to free the bit for init_vp_index() to work in the case
 404          * of sub-channel, when we reload drivers like hv_netvsc.
 405          */
 406         if (channel->affinity_policy == HV_LOCALIZED)
 407                 cpumask_clear_cpu(channel->target_cpu,
 408                                   &primary_channel->alloced_cpus_in_node);
 409
 410         vmbus_release_relid(channel->offermsg.child_relid);
 411
 412         free_channel(channel);
 413 }
 414
 415 void vmbus_free_channels(void)
 416 {
 417         struct vmbus_channel *channel, *tmp;
 418
 419         list_for_each_entry_safe(channel, tmp, &vmbus_connection.chn_list,
 420                 listentry) {
 421                 /* hv_process_channel_removal() needs this */
 422                 channel->rescind = true;
 423
 424                 vmbus_device_unregister(channel->device_obj);
 425         }
 426 }
 427
 428 /* Note: the function can run concurrently for primary/sub channels. */
 429 static void vmbus_add_channel_work(struct work_struct *work)
 430 {
 431         struct vmbus_channel *newchannel =
 432                 container_of(work, struct vmbus_channel, add_channel_work);
 433         struct vmbus_channel *primary_channel = newchannel->primary_channel;
 434         unsigned long flags;
 435         u16 dev_type;
 436         int ret;
 437
 438         dev_type = hv_get_dev_type(newchannel);
 439
 440         init_vp_index(newchannel, dev_type);
 441
 442         if (newchannel->target_cpu != get_cpu()) {
 443                 put_cpu();
 444                 smp_call_function_single(newchannel->target_cpu,
 445                                          percpu_channel_enq,
 446                                          newchannel, true);
 447         } else {
 448                 percpu_channel_enq(newchannel);
 449                 put_cpu();
 450         }
 451
 452         /*
 453          * This state is used to indicate a successful open
 454          * so that when we do close the channel normally, we
 455          * can cleanup properly.
 456          */
 457         newchannel->state = CHANNEL_OPEN_STATE;
 458
 459         if (primary_channel != NULL) {
 460                 /* newchannel is a sub-channel. */
 461                 struct hv_device *dev = primary_channel->device_obj;
 462
 463                 if (vmbus_add_channel_kobj(dev, newchannel))
 464                         goto err_deq_chan;
 465
 466                 if (primary_channel->sc_creation_callback != NULL)
 467                         primary_channel->sc_creation_callback(newchannel);
 468
 469                 newchannel->probe_done = true;
 470                 return;
 471         }
 472
 473         /*
 474          * Start the process of binding the primary channel to the driver
 475          */
 476         newchannel->device_obj = vmbus_device_create(
 477                 &newchannel->offermsg.offer.if_type,
 478                 &newchannel->offermsg.offer.if_instance,
 479                 newchannel);
 480         if (!newchannel->device_obj)
 481                 goto err_deq_chan;
 482
 483         newchannel->device_obj->device_id = dev_type;
 484         /*
 485          * Add the new device to the bus. This will kick off device-driver
 486          * binding which eventually invokes the device driver's AddDevice()
 487          * method.
 488          */
 489         ret = vmbus_device_register(newchannel->device_obj);
 490
 491         if (ret != 0) {
 492                 pr_err("unable to add child device object (relid %d)\n",
 493                         newchannel->offermsg.child_relid);
 494                 kfree(newchannel->device_obj);
 495                 goto err_deq_chan;
 496         }
 497
 498         newchannel->probe_done = true;
 499         return;
 500
 501 err_deq_chan:
 502         mutex_lock(&vmbus_connection.channel_mutex);
 503
 504         /*
 505          * We need to set the flag, otherwise
 506          * vmbus_onoffer_rescind() can be blocked.
 507          */
 508         newchannel->probe_done = true;
 509
 510         if (primary_channel == NULL) {
 511                 list_del(&newchannel->listentry);
 512         } else {
 513                 spin_lock_irqsave(&primary_channel->lock, flags);
 514                 list_del(&newchannel->sc_list);
 515                 spin_unlock_irqrestore(&primary_channel->lock, flags);
 516         }
 517
 518         mutex_unlock(&vmbus_connection.channel_mutex);
 519
 520         if (newchannel->target_cpu != get_cpu()) {
 521                 put_cpu();
 522                 smp_call_function_single(newchannel->target_cpu,
 523                                          percpu_channel_deq,
 524                                          newchannel, true);
 525         } else {
 526                 percpu_channel_deq(newchannel);
 527                 put_cpu();
 528         }
 529
 530         vmbus_release_relid(newchannel->offermsg.child_relid);
 531
 532         free_channel(newchannel);
 533 }
 534
 535 /*
 536  * vmbus_process_offer - Process the offer by creating a channel/device
 537  * associated with this offer
 538  */
 539 static void vmbus_process_offer(struct vmbus_channel *newchannel)
 540 {
 541         struct vmbus_channel *channel;
 542         struct workqueue_struct *wq;
 543         unsigned long flags;
 544         bool fnew = true;
 545
 546         mutex_lock(&vmbus_connection.channel_mutex);
 547
 548         /*
 549          * Now that we have acquired the channel_mutex,
 550          * we can release the potentially racing rescind thread.
 551          */
 552         atomic_dec(&vmbus_connection.offer_in_progress);
 553
 554         list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
 555                 if (guid_equal(&channel->offermsg.offer.if_type,
 556                                &newchannel->offermsg.offer.if_type) &&
 557                     guid_equal(&channel->offermsg.offer.if_instance,
 558                                &newchannel->offermsg.offer.if_instance)) {
 559                         fnew = false;
 560                         break;
 561                 }
 562         }
 563
 564         if (fnew)
 565                 list_add_tail(&newchannel->listentry,
 566                               &vmbus_connection.chn_list);
 567         else {
 568                 /*
 569                  * Check to see if this is a valid sub-channel.
 570                  */
 571                 if (newchannel->offermsg.offer.sub_channel_index == 0) {
 572                         mutex_unlock(&vmbus_connection.channel_mutex);
 573                         /*
 574                          * Don't call free_channel(), because newchannel->kobj
 575                          * is not initialized yet.
 576                          */
 577                         kfree(newchannel);
 578                         WARN_ON_ONCE(1);
 579                         return;
 580                 }
 581                 /*
 582                  * Process the sub-channel.
 583                  */
 584                 newchannel->primary_channel = channel;
 585                 spin_lock_irqsave(&channel->lock, flags);
 586                 list_add_tail(&newchannel->sc_list, &channel->sc_list);
 587                 spin_unlock_irqrestore(&channel->lock, flags);
 588         }
 589
 590         mutex_unlock(&vmbus_connection.channel_mutex);
 591
 592         /*
 593          * vmbus_process_offer() mustn't call channel->sc_creation_callback()
 594          * directly for sub-channels, because sc_creation_callback() ->
 595          * vmbus_open() may never get the host's response to the
 596          * OPEN_CHANNEL message (the host may rescind a channel at any time,
 597          * e.g. in the case of hot removing a NIC), and vmbus_onoffer_rescind()
 598          * may not wake up the vmbus_open() as it's blocked due to a non-zero
 599          * vmbus_connection.offer_in_progress, and finally we have a deadlock.
 600          *
 601          * The above is also true for primary channels, if the related device
 602          * drivers use sync probing mode by default.
 603          *
 604          * And, usually the handling of primary channels and sub-channels can
 605          * depend on each other, so we should offload them to different
 606          * workqueues to avoid possible deadlock, e.g. in sync-probing mode,
 607          * NIC1's netvsc_subchan_work() can race with NIC2's netvsc_probe() ->
 608          * rtnl_lock(), and causes deadlock: the former gets the rtnl_lock
 609          * and waits for all the sub-channels to appear, but the latter
 610          * can't get the rtnl_lock and this blocks the handling of
 611          * sub-channels.
 612          */
 613         INIT_WORK(&newchannel->add_channel_work, vmbus_add_channel_work);
 614         wq = fnew ? vmbus_connection.handle_primary_chan_wq :
 615                     vmbus_connection.handle_sub_chan_wq;
 616         queue_work(wq, &newchannel->add_channel_work);
 617 }
 618
 619 /*
 620  * We use this state to statically distribute the channel interrupt load.
 621  */
 622 static int next_numa_node_id;
 623 /*
 624  * init_vp_index() accesses global variables like next_numa_node_id, and
 625  * it can run concurrently for primary channels and sub-channels: see
 626  * vmbus_process_offer(), so we need the lock to protect the global
 627  * variables.
 628  */
 629 static DEFINE_SPINLOCK(bind_channel_to_cpu_lock);
 630
 631 /*
 632  * Starting with Win8, we can statically distribute the incoming
 633  * channel interrupt load by binding a channel to VCPU.
 634  * We distribute the interrupt loads to one or more NUMA nodes based on
 635  * the channel's affinity_policy.
 636  *
 637  * For pre-win8 hosts or non-performance critical channels we assign the
 638  * first CPU in the first NUMA node.
 639  */
 640 static void init_vp_index(struct vmbus_channel *channel, u16 dev_type)
 641 {
 642         u32 cur_cpu;
 643         bool perf_chn = vmbus_devs[dev_type].perf_device;
 644         struct vmbus_channel *primary = channel->primary_channel;
 645         int next_node;
 646         cpumask_var_t available_mask;
 647         struct cpumask *alloced_mask;
 648
 649         if ((vmbus_proto_version == VERSION_WS2008) ||
 650             (vmbus_proto_version == VERSION_WIN7) || (!perf_chn) ||
 651             !alloc_cpumask_var(&available_mask, GFP_KERNEL)) {
 652                 /*
 653                  * Prior to win8, all channel interrupts are
 654                  * delivered on cpu 0.
 655                  * Also if the channel is not a performance critical
 656                  * channel, bind it to cpu 0.
 657                  * In case alloc_cpumask_var() fails, bind it to cpu 0.
 658                  */
 659                 channel->numa_node = 0;
 660                 channel->target_cpu = 0;
 661                 channel->target_vp = hv_cpu_number_to_vp_number(0);
 662                 return;
 663         }
 664
 665         spin_lock(&bind_channel_to_cpu_lock);
 666
 667         /*
 668          * Based on the channel affinity policy, we will assign the NUMA
 669          * nodes.
 670          */
 671
 672         if ((channel->affinity_policy == HV_BALANCED) || (!primary)) {
 673                 while (true) {
 674                         next_node = next_numa_node_id++;
 675                         if (next_node == nr_node_ids) {
 676                                 next_node = next_numa_node_id = 0;
 677                                 continue;
 678                         }
 679                         if (cpumask_empty(cpumask_of_node(next_node)))
 680                                 continue;
 681                         break;
 682                 }
 683                 channel->numa_node = next_node;
 684                 primary = channel;
 685         }
 686         alloced_mask = &hv_context.hv_numa_map[primary->numa_node];
 687
 688         if (cpumask_weight(alloced_mask) ==
 689             cpumask_weight(cpumask_of_node(primary->numa_node))) {
 690                 /*
 691                  * We have cycled through all the CPUs in the node;
 692                  * reset the alloced map.
 693                  */
 694                 cpumask_clear(alloced_mask);
 695         }
 696
 697         cpumask_xor(available_mask, alloced_mask,
 698                     cpumask_of_node(primary->numa_node));
 699
 700         cur_cpu = -1;
 701
 702         if (primary->affinity_policy == HV_LOCALIZED) {
 703                 /*
 704                  * Normally Hyper-V host doesn't create more subchannels
 705                  * than there are VCPUs on the node but it is possible when not
 706                  * all present VCPUs on the node are initialized by guest.
 707                  * Clear the alloced_cpus_in_node to start over.
 708                  */
 709                 if (cpumask_equal(&primary->alloced_cpus_in_node,
 710                                   cpumask_of_node(primary->numa_node)))
 711                         cpumask_clear(&primary->alloced_cpus_in_node);
 712         }
 713
 714         while (true) {
 715                 cur_cpu = cpumask_next(cur_cpu, available_mask);
 716                 if (cur_cpu >= nr_cpu_ids) {
 717                         cur_cpu = -1;
 718                         cpumask_copy(available_mask,
 719                                      cpumask_of_node(primary->numa_node));
 720                         continue;
 721                 }
 722
 723                 if (primary->affinity_policy == HV_LOCALIZED) {
 724                         /*
 725                          * NOTE: in the case of sub-channel, we clear the
 726                          * sub-channel related bit(s) in
 727                          * primary->alloced_cpus_in_node in
 728                          * hv_process_channel_removal(), so when we
 729                          * reload drivers like hv_netvsc in SMP guest, here
 730                          * we're able to re-allocate
 731                          * bit from primary->alloced_cpus_in_node.
 732                          */
 733                         if (!cpumask_test_cpu(cur_cpu,
 734                                               &primary->alloced_cpus_in_node)) {
 735                                 cpumask_set_cpu(cur_cpu,
 736                                                 &primary->alloced_cpus_in_node);
 737                                 cpumask_set_cpu(cur_cpu, alloced_mask);
 738                                 break;
 739                         }
 740                 } else {
 741                         cpumask_set_cpu(cur_cpu, alloced_mask);
 742                         break;
 743                 }
 744         }
 745
 746         channel->target_cpu = cur_cpu;
 747         channel->target_vp = hv_cpu_number_to_vp_number(cur_cpu);
 748
 749         spin_unlock(&bind_channel_to_cpu_lock);
 750
 751         free_cpumask_var(available_mask);
 752 }
 753
 754 static void vmbus_wait_for_unload(void)
 755 {
 756         int cpu;
 757         void *page_addr;
 758         struct hv_message *msg;
 759         struct vmbus_channel_message_header *hdr;
 760         u32 message_type;
 761
 762         /*
 763          * CHANNELMSG_UNLOAD_RESPONSE is always delivered to the CPU which was
 764          * used for initial contact or to CPU0 depending on host version. When
 765          * we're crashing on a different CPU let's hope that IRQ handler on
 766          * the cpu which receives CHANNELMSG_UNLOAD_RESPONSE is still
 767          * functional and vmbus_unload_response() will complete
 768          * vmbus_connection.unload_event. If not, the last thing we can do is
 769          * read message pages for all CPUs directly.
 770          */
 771         while (1) {
 772                 if (completion_done(&vmbus_connection.unload_event))
 773                         break;
 774
 775                 for_each_online_cpu(cpu) {
 776                         struct hv_per_cpu_context *hv_cpu
 777                                 = per_cpu_ptr(hv_context.cpu_context, cpu);
 778
 779                         page_addr = hv_cpu->synic_message_page;
 780                         msg = (struct hv_message *)page_addr
 781                                 + VMBUS_MESSAGE_SINT;
 782
 783                         message_type = READ_ONCE(msg->header.message_type);
 784                         if (message_type == HVMSG_NONE)
 785                                 continue;
 786
 787                         hdr = (struct vmbus_channel_message_header *)
 788                                 msg->u.payload;
 789
 790                         if (hdr->msgtype == CHANNELMSG_UNLOAD_RESPONSE)
 791                                 complete(&vmbus_connection.unload_event);
 792
 793                         vmbus_signal_eom(msg, message_type);
 794                 }
 795
 796                 mdelay(10);
 797         }
 798
 799         /*
 800          * We're crashing and already got the UNLOAD_RESPONSE, cleanup all
 801          * maybe-pending messages on all CPUs to be able to receive new
 802          * messages after we reconnect.
 803          */
 804         for_each_online_cpu(cpu) {
 805                 struct hv_per_cpu_context *hv_cpu
 806                         = per_cpu_ptr(hv_context.cpu_context, cpu);
 807
 808                 page_addr = hv_cpu->synic_message_page;
 809                 msg = (struct hv_message *)page_addr + VMBUS_MESSAGE_SINT;
 810                 msg->header.message_type = HVMSG_NONE;
 811         }
 812 }
 813
 814 /*
 815  * vmbus_unload_response - Handler for the unload response.
 816  */
 817 static void vmbus_unload_response(struct vmbus_channel_message_header *hdr)
 818 {
 819         /*
 820          * This is a global event; just wakeup the waiting thread.
 821          * Once we successfully unload, we can cleanup the monitor state.
 822          */
 823         complete(&vmbus_connection.unload_event);
 824 }
 825
 826 void vmbus_initiate_unload(bool crash)
 827 {
 828         struct vmbus_channel_message_header hdr;
 829
 830         /* Pre-Win2012R2 hosts don't support reconnect */
 831         if (vmbus_proto_version < VERSION_WIN8_1)
 832                 return;
 833
 834         init_completion(&vmbus_connection.unload_event);
 835         memset(&hdr, 0, sizeof(struct vmbus_channel_message_header));
 836         hdr.msgtype = CHANNELMSG_UNLOAD;
 837         vmbus_post_msg(&hdr, sizeof(struct vmbus_channel_message_header),
 838                        !crash);
 839
 840         /*
 841          * vmbus_initiate_unload() is also called on crash and the crash can be
 842          * happening in an interrupt context, where scheduling is impossible.
 843          */
 844         if (!crash)
 845                 wait_for_completion(&vmbus_connection.unload_event);
 846         else
 847                 vmbus_wait_for_unload();
 848 }
 849
 850 /*
 851  * vmbus_onoffer - Handler for channel offers from vmbus in parent partition.
 852  *
 853  */
 854 static void vmbus_onoffer(struct vmbus_channel_message_header *hdr)
 855 {
 856         struct vmbus_channel_offer_channel *offer;
 857         struct vmbus_channel *newchannel;
 858
 859         offer = (struct vmbus_channel_offer_channel *)hdr;
 860
 861         trace_vmbus_onoffer(offer);
 862
 863         /* Allocate the channel object and save this offer. */
 864         newchannel = alloc_channel();
 865         if (!newchannel) {
 866                 vmbus_release_relid(offer->child_relid);
 867                 atomic_dec(&vmbus_connection.offer_in_progress);
 868                 pr_err("Unable to allocate channel object\n");
 869                 return;
 870         }
 871
 872         /*
 873          * Setup state for signalling the host.
 874          */
 875         newchannel->sig_event = VMBUS_EVENT_CONNECTION_ID;
 876
 877         if (vmbus_proto_version != VERSION_WS2008) {
 878                 newchannel->is_dedicated_interrupt =
 879                                 (offer->is_dedicated_interrupt != 0);
 880                 newchannel->sig_event = offer->connection_id;
 881         }
 882
 883         memcpy(&newchannel->offermsg, offer,
 884                sizeof(struct vmbus_channel_offer_channel));
 885         newchannel->monitor_grp = (u8)offer->monitorid / 32;
 886         newchannel->monitor_bit = (u8)offer->monitorid % 32;
 887
 888         vmbus_process_offer(newchannel);
 889 }
 890
 891 /*
 892  * vmbus_onoffer_rescind - Rescind offer handler.
 893  *
 894  * We queue a work item to process this offer synchronously
 895  */
 896 static void vmbus_onoffer_rescind(struct vmbus_channel_message_header *hdr)
 897 {
 898         struct vmbus_channel_rescind_offer *rescind;
 899         struct vmbus_channel *channel;
 900         struct device *dev;
 901
 902         rescind = (struct vmbus_channel_rescind_offer *)hdr;
 903
 904         trace_vmbus_onoffer_rescind(rescind);
 905
 906         /*
 907          * The offer msg and the corresponding rescind msg
 908          * from the host are guranteed to be ordered -
 909          * offer comes in first and then the rescind.
 910          * Since we process these events in work elements,
 911          * and with preemption, we may end up processing
 912          * the events out of order. Given that we handle these
 913          * work elements on the same CPU, this is possible only
 914          * in the case of preemption. In any case wait here
 915          * until the offer processing has moved beyond the
 916          * point where the channel is discoverable.
 917          */
 918
 919         while (atomic_read(&vmbus_connection.offer_in_progress) != 0) {
 920                 /*
 921                  * We wait here until any channel offer is currently
 922                  * being processed.
 923                  */
 924                 msleep(1);
 925         }
 926
 927         mutex_lock(&vmbus_connection.channel_mutex);
 928         channel = relid2channel(rescind->child_relid);
 929         mutex_unlock(&vmbus_connection.channel_mutex);
 930
 931         if (channel == NULL) {
 932                 /*
 933                  * We failed in processing the offer message;
 934                  * we would have cleaned up the relid in that
 935                  * failure path.
 936                  */
 937                 return;
 938         }
 939
 940         /*
 941          * Before setting channel->rescind in vmbus_rescind_cleanup(), we
 942          * should make sure the channel callback is not running any more.
 943          */
 944         vmbus_reset_channel_cb(channel);
 945
 946         /*
 947          * Now wait for offer handling to complete.
 948          */
 949         vmbus_rescind_cleanup(channel);
 950         while (READ_ONCE(channel->probe_done) == false) {
 951                 /*
 952                  * We wait here until any channel offer is currently
 953                  * being processed.
 954                  */
 955                 msleep(1);
 956         }
 957
 958         /*
 959          * At this point, the rescind handling can proceed safely.
 960          */
 961
 962         if (channel->device_obj) {
 963                 if (channel->chn_rescind_callback) {
 964                         channel->chn_rescind_callback(channel);
 965                         return;
 966                 }
 967                 /*
 968                  * We will have to unregister this device from the
 969                  * driver core.
 970                  */
 971                 dev = get_device(&channel->device_obj->device);
 972                 if (dev) {
 973                         vmbus_device_unregister(channel->device_obj);
 974                         put_device(dev);
 975                 }
 976         }
 977         if (channel->primary_channel != NULL) {
 978                 /*
 979                  * Sub-channel is being rescinded. Following is the channel
 980                  * close sequence when initiated from the driveri (refer to
 981                  * vmbus_close() for details):
 982                  * 1. Close all sub-channels first
 983                  * 2. Then close the primary channel.
 984                  */
 985                 mutex_lock(&vmbus_connection.channel_mutex);
 986                 if (channel->state == CHANNEL_OPEN_STATE) {
 987                         /*
 988                          * The channel is currently not open;
 989                          * it is safe for us to cleanup the channel.
 990                          */
 991                         hv_process_channel_removal(channel);
 992                 } else {
 993                         complete(&channel->rescind_event);
 994                 }
 995                 mutex_unlock(&vmbus_connection.channel_mutex);
 996         }
 997 }
 998
 999 void vmbus_hvsock_device_unregister(struct vmbus_channel *channel)
1000 {
1001         BUG_ON(!is_hvsock_channel(channel));
1002
1003         /* We always get a rescind msg when a connection is closed. */
1004         while (!READ_ONCE(channel->probe_done) || !READ_ONCE(channel->rescind))
1005                 msleep(1);
1006
1007         vmbus_device_unregister(channel->device_obj);
1008 }
1009 EXPORT_SYMBOL_GPL(vmbus_hvsock_device_unregister);
1010
1011
1012 /*
1013  * vmbus_onoffers_delivered -
1014  * This is invoked when all offers have been delivered.
1015  *
1016  * Nothing to do here.
1017  */
1018 static void vmbus_onoffers_delivered(
1019                         struct vmbus_channel_message_header *hdr)
1020 {
1021 }
1022
1023 /*
1024  * vmbus_onopen_result - Open result handler.
1025  *
1026  * This is invoked when we received a response to our channel open request.
1027  * Find the matching request, copy the response and signal the requesting
1028  * thread.
1029  */
1030 static void vmbus_onopen_result(struct vmbus_channel_message_header *hdr)
1031 {
1032         struct vmbus_channel_open_result *result;
1033         struct vmbus_channel_msginfo *msginfo;
1034         struct vmbus_channel_message_header *requestheader;
1035         struct vmbus_channel_open_channel *openmsg;
1036         unsigned long flags;
1037
1038         result = (struct vmbus_channel_open_result *)hdr;
1039
1040         trace_vmbus_onopen_result(result);
1041
1042         /*
1043          * Find the open msg, copy the result and signal/unblock the wait event
1044          */
1045         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1046
1047         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1048                                 msglistentry) {
1049                 requestheader =
1050                         (struct vmbus_channel_message_header *)msginfo->msg;
1051
1052                 if (requestheader->msgtype == CHANNELMSG_OPENCHANNEL) {
1053                         openmsg =
1054                         (struct vmbus_channel_open_channel *)msginfo->msg;
1055                         if (openmsg->child_relid == result->child_relid &&
1056                             openmsg->openid == result->openid) {
1057                                 memcpy(&msginfo->response.open_result,
1058                                        result,
1059                                        sizeof(
1060                                         struct vmbus_channel_open_result));
1061                                 complete(&msginfo->waitevent);
1062                                 break;
1063                         }
1064                 }
1065         }
1066         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1067 }
1068
1069 /*
1070  * vmbus_ongpadl_created - GPADL created handler.
1071  *
1072  * This is invoked when we received a response to our gpadl create request.
1073  * Find the matching request, copy the response and signal the requesting
1074  * thread.
1075  */
1076 static void vmbus_ongpadl_created(struct vmbus_channel_message_header *hdr)
1077 {
1078         struct vmbus_channel_gpadl_created *gpadlcreated;
1079         struct vmbus_channel_msginfo *msginfo;
1080         struct vmbus_channel_message_header *requestheader;
1081         struct vmbus_channel_gpadl_header *gpadlheader;
1082         unsigned long flags;
1083
1084         gpadlcreated = (struct vmbus_channel_gpadl_created *)hdr;
1085
1086         trace_vmbus_ongpadl_created(gpadlcreated);
1087
1088         /*
1089          * Find the establish msg, copy the result and signal/unblock the wait
1090          * event
1091          */
1092         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1093
1094         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1095                                 msglistentry) {
1096                 requestheader =
1097                         (struct vmbus_channel_message_header *)msginfo->msg;
1098
1099                 if (requestheader->msgtype == CHANNELMSG_GPADL_HEADER) {
1100                         gpadlheader =
1101                         (struct vmbus_channel_gpadl_header *)requestheader;
1102
1103                         if ((gpadlcreated->child_relid ==
1104                              gpadlheader->child_relid) &&
1105                             (gpadlcreated->gpadl == gpadlheader->gpadl)) {
1106                                 memcpy(&msginfo->response.gpadl_created,
1107                                        gpadlcreated,
1108                                        sizeof(
1109                                         struct vmbus_channel_gpadl_created));
1110                                 complete(&msginfo->waitevent);
1111                                 break;
1112                         }
1113                 }
1114         }
1115         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1116 }
1117
1118 /*
1119  * vmbus_ongpadl_torndown - GPADL torndown handler.
1120  *
1121  * This is invoked when we received a response to our gpadl teardown request.
1122  * Find the matching request, copy the response and signal the requesting
1123  * thread.
1124  */
1125 static void vmbus_ongpadl_torndown(
1126                         struct vmbus_channel_message_header *hdr)
1127 {
1128         struct vmbus_channel_gpadl_torndown *gpadl_torndown;
1129         struct vmbus_channel_msginfo *msginfo;
1130         struct vmbus_channel_message_header *requestheader;
1131         struct vmbus_channel_gpadl_teardown *gpadl_teardown;
1132         unsigned long flags;
1133
1134         gpadl_torndown = (struct vmbus_channel_gpadl_torndown *)hdr;
1135
1136         trace_vmbus_ongpadl_torndown(gpadl_torndown);
1137
1138         /*
1139          * Find the open msg, copy the result and signal/unblock the wait event
1140          */
1141         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1142
1143         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1144                                 msglistentry) {
1145                 requestheader =
1146                         (struct vmbus_channel_message_header *)msginfo->msg;
1147
1148                 if (requestheader->msgtype == CHANNELMSG_GPADL_TEARDOWN) {
1149                         gpadl_teardown =
1150                         (struct vmbus_channel_gpadl_teardown *)requestheader;
1151
1152                         if (gpadl_torndown->gpadl == gpadl_teardown->gpadl) {
1153                                 memcpy(&msginfo->response.gpadl_torndown,
1154                                        gpadl_torndown,
1155                                        sizeof(
1156                                         struct vmbus_channel_gpadl_torndown));
1157                                 complete(&msginfo->waitevent);
1158                                 break;
1159                         }
1160                 }
1161         }
1162         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1163 }
1164
1165 /*
1166  * vmbus_onversion_response - Version response handler
1167  *
1168  * This is invoked when we received a response to our initiate contact request.
1169  * Find the matching request, copy the response and signal the requesting
1170  * thread.
1171  */
1172 static void vmbus_onversion_response(
1173                 struct vmbus_channel_message_header *hdr)
1174 {
1175         struct vmbus_channel_msginfo *msginfo;
1176         struct vmbus_channel_message_header *requestheader;
1177         struct vmbus_channel_version_response *version_response;
1178         unsigned long flags;
1179
1180         version_response = (struct vmbus_channel_version_response *)hdr;
1181
1182         trace_vmbus_onversion_response(version_response);
1183
1184         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1185
1186         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1187                                 msglistentry) {
1188                 requestheader =
1189                         (struct vmbus_channel_message_header *)msginfo->msg;
1190
1191                 if (requestheader->msgtype ==
1192                     CHANNELMSG_INITIATE_CONTACT) {
1193                         memcpy(&msginfo->response.version_response,
1194                               version_response,
1195                               sizeof(struct vmbus_channel_version_response));
1196                         complete(&msginfo->waitevent);
1197                 }
1198         }
1199         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1200 }
1201
1202 /* Channel message dispatch table */
1203 const struct vmbus_channel_message_table_entry
1204 channel_message_table[CHANNELMSG_COUNT] = {
1205         { CHANNELMSG_INVALID,                   0, NULL },
1206         { CHANNELMSG_OFFERCHANNEL,              0, vmbus_onoffer },
1207         { CHANNELMSG_RESCIND_CHANNELOFFER,      0, vmbus_onoffer_rescind },
1208         { CHANNELMSG_REQUESTOFFERS,             0, NULL },
1209         { CHANNELMSG_ALLOFFERS_DELIVERED,       1, vmbus_onoffers_delivered },
1210         { CHANNELMSG_OPENCHANNEL,               0, NULL },
1211         { CHANNELMSG_OPENCHANNEL_RESULT,        1, vmbus_onopen_result },
1212         { CHANNELMSG_CLOSECHANNEL,              0, NULL },
1213         { CHANNELMSG_GPADL_HEADER,              0, NULL },
1214         { CHANNELMSG_GPADL_BODY,                0, NULL },
1215         { CHANNELMSG_GPADL_CREATED,             1, vmbus_ongpadl_created },
1216         { CHANNELMSG_GPADL_TEARDOWN,            0, NULL },
1217         { CHANNELMSG_GPADL_TORNDOWN,            1, vmbus_ongpadl_torndown },
1218         { CHANNELMSG_RELID_RELEASED,            0, NULL },
1219         { CHANNELMSG_INITIATE_CONTACT,          0, NULL },
1220         { CHANNELMSG_VERSION_RESPONSE,          1, vmbus_onversion_response },
1221         { CHANNELMSG_UNLOAD,                    0, NULL },
1222         { CHANNELMSG_UNLOAD_RESPONSE,           1, vmbus_unload_response },
1223         { CHANNELMSG_18,                        0, NULL },
1224         { CHANNELMSG_19,                        0, NULL },
1225         { CHANNELMSG_20,                        0, NULL },
1226         { CHANNELMSG_TL_CONNECT_REQUEST,        0, NULL },
1227 };
1228
1229 /*
1230  * vmbus_onmessage - Handler for channel protocol messages.
1231  *
1232  * This is invoked in the vmbus worker thread context.
1233  */
1234 void vmbus_onmessage(void *context)
1235 {
1236         struct hv_message *msg = context;
1237         struct vmbus_channel_message_header *hdr;
1238         int size;
1239
1240         hdr = (struct vmbus_channel_message_header *)msg->u.payload;
1241         size = msg->header.payload_size;
1242
1243         trace_vmbus_on_message(hdr);
1244
1245         if (hdr->msgtype >= CHANNELMSG_COUNT) {
1246                 pr_err("Received invalid channel message type %d size %d\n",
1247                            hdr->msgtype, size);
1248                 print_hex_dump_bytes("", DUMP_PREFIX_NONE,
1249                                      (unsigned char *)msg->u.payload, size);
1250                 return;
1251         }
1252
1253         if (channel_message_table[hdr->msgtype].message_handler)
1254                 channel_message_table[hdr->msgtype].message_handler(hdr);
1255         else
1256                 pr_err("Unhandled channel message type %d\n", hdr->msgtype);
1257 }
1258
1259 /*
1260  * vmbus_request_offers - Send a request to get all our pending offers.
1261  */
1262 int vmbus_request_offers(void)
1263 {
1264         struct vmbus_channel_message_header *msg;
1265         struct vmbus_channel_msginfo *msginfo;
1266         int ret;
1267
1268         msginfo = kmalloc(sizeof(*msginfo) +
1269                           sizeof(struct vmbus_channel_message_header),
1270                           GFP_KERNEL);
1271         if (!msginfo)
1272                 return -ENOMEM;
1273
1274         msg = (struct vmbus_channel_message_header *)msginfo->msg;
1275
1276         msg->msgtype = CHANNELMSG_REQUESTOFFERS;
1277
1278         ret = vmbus_post_msg(msg, sizeof(struct vmbus_channel_message_header),
1279                              true);
1280
1281         trace_vmbus_request_offers(ret);
1282
1283         if (ret != 0) {
1284                 pr_err("Unable to request offers - %d\n", ret);
1285
1286                 goto cleanup;
1287         }
1288
1289 cleanup:
1290         kfree(msginfo);
1291
1292         return ret;
1293 }
1294
1295 static void invoke_sc_cb(struct vmbus_channel *primary_channel)
1296 {
1297         struct list_head *cur, *tmp;
1298         struct vmbus_channel *cur_channel;
1299
1300         if (primary_channel->sc_creation_callback == NULL)
1301                 return;
1302
1303         list_for_each_safe(cur, tmp, &primary_channel->sc_list) {
1304                 cur_channel = list_entry(cur, struct vmbus_channel, sc_list);
1305
1306                 primary_channel->sc_creation_callback(cur_channel);
1307         }
1308 }
1309
1310 void vmbus_set_sc_create_callback(struct vmbus_channel *primary_channel,
1311                                 void (*sc_cr_cb)(struct vmbus_channel *new_sc))
1312 {
1313         primary_channel->sc_creation_callback = sc_cr_cb;
1314 }
1315 EXPORT_SYMBOL_GPL(vmbus_set_sc_create_callback);
1316
1317 bool vmbus_are_subchannels_present(struct vmbus_channel *primary)
1318 {
1319         bool ret;
1320
1321         ret = !list_empty(&primary->sc_list);
1322
1323         if (ret) {
1324                 /*
1325                  * Invoke the callback on sub-channel creation.
1326                  * This will present a uniform interface to the
1327                  * clients.
1328                  */
1329                 invoke_sc_cb(primary);
1330         }
1331
1332         return ret;
1333 }
1334 EXPORT_SYMBOL_GPL(vmbus_are_subchannels_present);
1335
1336 void vmbus_set_chn_rescind_callback(struct vmbus_channel *channel,
1337                 void (*chn_rescind_cb)(struct vmbus_channel *))
1338 {
1339         channel->chn_rescind_callback = chn_rescind_cb;
1340 }
1341 EXPORT_SYMBOL_GPL(vmbus_set_chn_rescind_callback);