drivers/hv/channel_mgmt.c

   1 /*
   2  * Copyright (c) 2009, Microsoft Corporation.
   3  *
   4  * This program is free software; you can redistribute it and/or modify it
   5  * under the terms and conditions of the GNU General Public License,
   6  * version 2, as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope it will be useful, but WITHOUT
   9  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  10  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  11  * more details.
  12  *
  13  * You should have received a copy of the GNU General Public License along with
  14  * this program; if not, write to the Free Software Foundation, Inc., 59 Temple
  15  * Place - Suite 330, Boston, MA 02111-1307 USA.
  16  *
  17  * Authors:
  18  *   Haiyang Zhang <haiyangz@microsoft.com>
  19  *   Hank Janssen  <hjanssen@microsoft.com>
  20  */
  21 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  22
  23 #include <linux/kernel.h>
  24 #include <linux/interrupt.h>
  25 #include <linux/sched.h>
  26 #include <linux/wait.h>
  27 #include <linux/mm.h>
  28 #include <linux/slab.h>
  29 #include <linux/list.h>
  30 #include <linux/module.h>
  31 #include <linux/completion.h>
  32 #include <linux/delay.h>
  33 #include <linux/hyperv.h>
  34 #include <asm/mshyperv.h>
  35
  36 #include "hyperv_vmbus.h"
  37
  38 static void init_vp_index(struct vmbus_channel *channel, u16 dev_type);
  39
  40 static const struct vmbus_device vmbus_devs[] = {
  41         /* IDE */
  42         { .dev_type = HV_IDE,
  43           HV_IDE_GUID,
  44           .perf_device = true,
  45         },
  46
  47         /* SCSI */
  48         { .dev_type = HV_SCSI,
  49           HV_SCSI_GUID,
  50           .perf_device = true,
  51         },
  52
  53         /* Fibre Channel */
  54         { .dev_type = HV_FC,
  55           HV_SYNTHFC_GUID,
  56           .perf_device = true,
  57         },
  58
  59         /* Synthetic NIC */
  60         { .dev_type = HV_NIC,
  61           HV_NIC_GUID,
  62           .perf_device = true,
  63         },
  64
  65         /* Network Direct */
  66         { .dev_type = HV_ND,
  67           HV_ND_GUID,
  68           .perf_device = true,
  69         },
  70
  71         /* PCIE */
  72         { .dev_type = HV_PCIE,
  73           HV_PCIE_GUID,
  74           .perf_device = true,
  75         },
  76
  77         /* Synthetic Frame Buffer */
  78         { .dev_type = HV_FB,
  79           HV_SYNTHVID_GUID,
  80           .perf_device = false,
  81         },
  82
  83         /* Synthetic Keyboard */
  84         { .dev_type = HV_KBD,
  85           HV_KBD_GUID,
  86           .perf_device = false,
  87         },
  88
  89         /* Synthetic MOUSE */
  90         { .dev_type = HV_MOUSE,
  91           HV_MOUSE_GUID,
  92           .perf_device = false,
  93         },
  94
  95         /* KVP */
  96         { .dev_type = HV_KVP,
  97           HV_KVP_GUID,
  98           .perf_device = false,
  99         },
 100
 101         /* Time Synch */
 102         { .dev_type = HV_TS,
 103           HV_TS_GUID,
 104           .perf_device = false,
 105         },
 106
 107         /* Heartbeat */
 108         { .dev_type = HV_HB,
 109           HV_HEART_BEAT_GUID,
 110           .perf_device = false,
 111         },
 112
 113         /* Shutdown */
 114         { .dev_type = HV_SHUTDOWN,
 115           HV_SHUTDOWN_GUID,
 116           .perf_device = false,
 117         },
 118
 119         /* File copy */
 120         { .dev_type = HV_FCOPY,
 121           HV_FCOPY_GUID,
 122           .perf_device = false,
 123         },
 124
 125         /* Backup */
 126         { .dev_type = HV_BACKUP,
 127           HV_VSS_GUID,
 128           .perf_device = false,
 129         },
 130
 131         /* Dynamic Memory */
 132         { .dev_type = HV_DM,
 133           HV_DM_GUID,
 134           .perf_device = false,
 135         },
 136
 137         /* Unknown GUID */
 138         { .dev_type = HV_UNKNOWN,
 139           .perf_device = false,
 140         },
 141 };
 142
 143 static const struct {
 144         uuid_le guid;
 145 } vmbus_unsupported_devs[] = {
 146         { HV_AVMA1_GUID },
 147         { HV_AVMA2_GUID },
 148         { HV_RDV_GUID   },
 149 };
 150
 151 /*
 152  * The rescinded channel may be blocked waiting for a response from the host;
 153  * take care of that.
 154  */
 155 static void vmbus_rescind_cleanup(struct vmbus_channel *channel)
 156 {
 157         struct vmbus_channel_msginfo *msginfo;
 158         unsigned long flags;
 159
 160
 161         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
 162
 163         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
 164                                 msglistentry) {
 165
 166                 if (msginfo->waiting_channel == channel) {
 167                         complete(&msginfo->waitevent);
 168                         break;
 169                 }
 170         }
 171         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
 172 }
 173
 174 static bool is_unsupported_vmbus_devs(const uuid_le *guid)
 175 {
 176         int i;
 177
 178         for (i = 0; i < ARRAY_SIZE(vmbus_unsupported_devs); i++)
 179                 if (!uuid_le_cmp(*guid, vmbus_unsupported_devs[i].guid))
 180                         return true;
 181         return false;
 182 }
 183
 184 static u16 hv_get_dev_type(const struct vmbus_channel *channel)
 185 {
 186         const uuid_le *guid = &channel->offermsg.offer.if_type;
 187         u16 i;
 188
 189         if (is_hvsock_channel(channel) || is_unsupported_vmbus_devs(guid))
 190                 return HV_UNKNOWN;
 191
 192         for (i = HV_IDE; i < HV_UNKNOWN; i++) {
 193                 if (!uuid_le_cmp(*guid, vmbus_devs[i].guid))
 194                         return i;
 195         }
 196         pr_info("Unknown GUID: %pUl\n", guid);
 197         return i;
 198 }
 199
 200 /**
 201  * vmbus_prep_negotiate_resp() - Create default response for Hyper-V Negotiate message
 202  * @icmsghdrp: Pointer to msg header structure
 203  * @icmsg_negotiate: Pointer to negotiate message structure
 204  * @buf: Raw buffer channel data
 205  *
 206  * @icmsghdrp is of type &struct icmsg_hdr.
 207  * Set up and fill in default negotiate response message.
 208  *
 209  * The fw_version and fw_vercnt specifies the framework version that
 210  * we can support.
 211  *
 212  * The srv_version and srv_vercnt specifies the service
 213  * versions we can support.
 214  *
 215  * Versions are given in decreasing order.
 216  *
 217  * nego_fw_version and nego_srv_version store the selected protocol versions.
 218  *
 219  * Mainly used by Hyper-V drivers.
 220  */
 221 bool vmbus_prep_negotiate_resp(struct icmsg_hdr *icmsghdrp,
 222                                 u8 *buf, const int *fw_version, int fw_vercnt,
 223                                 const int *srv_version, int srv_vercnt,
 224                                 int *nego_fw_version, int *nego_srv_version)
 225 {
 226         int icframe_major, icframe_minor;
 227         int icmsg_major, icmsg_minor;
 228         int fw_major, fw_minor;
 229         int srv_major, srv_minor;
 230         int i, j;
 231         bool found_match = false;
 232         struct icmsg_negotiate *negop;
 233
 234         icmsghdrp->icmsgsize = 0x10;
 235         negop = (struct icmsg_negotiate *)&buf[
 236                 sizeof(struct vmbuspipe_hdr) +
 237                 sizeof(struct icmsg_hdr)];
 238
 239         icframe_major = negop->icframe_vercnt;
 240         icframe_minor = 0;
 241
 242         icmsg_major = negop->icmsg_vercnt;
 243         icmsg_minor = 0;
 244
 245         /*
 246          * Select the framework version number we will
 247          * support.
 248          */
 249
 250         for (i = 0; i < fw_vercnt; i++) {
 251                 fw_major = (fw_version[i] >> 16);
 252                 fw_minor = (fw_version[i] & 0xFFFF);
 253
 254                 for (j = 0; j < negop->icframe_vercnt; j++) {
 255                         if ((negop->icversion_data[j].major == fw_major) &&
 256                             (negop->icversion_data[j].minor == fw_minor)) {
 257                                 icframe_major = negop->icversion_data[j].major;
 258                                 icframe_minor = negop->icversion_data[j].minor;
 259                                 found_match = true;
 260                                 break;
 261                         }
 262                 }
 263
 264                 if (found_match)
 265                         break;
 266         }
 267
 268         if (!found_match)
 269                 goto fw_error;
 270
 271         found_match = false;
 272
 273         for (i = 0; i < srv_vercnt; i++) {
 274                 srv_major = (srv_version[i] >> 16);
 275                 srv_minor = (srv_version[i] & 0xFFFF);
 276
 277                 for (j = negop->icframe_vercnt;
 278                         (j < negop->icframe_vercnt + negop->icmsg_vercnt);
 279                         j++) {
 280
 281                         if ((negop->icversion_data[j].major == srv_major) &&
 282                                 (negop->icversion_data[j].minor == srv_minor)) {
 283
 284                                 icmsg_major = negop->icversion_data[j].major;
 285                                 icmsg_minor = negop->icversion_data[j].minor;
 286                                 found_match = true;
 287                                 break;
 288                         }
 289                 }
 290
 291                 if (found_match)
 292                         break;
 293         }
 294
 295         /*
 296          * Respond with the framework and service
 297          * version numbers we can support.
 298          */
 299
 300 fw_error:
 301         if (!found_match) {
 302                 negop->icframe_vercnt = 0;
 303                 negop->icmsg_vercnt = 0;
 304         } else {
 305                 negop->icframe_vercnt = 1;
 306                 negop->icmsg_vercnt = 1;
 307         }
 308
 309         if (nego_fw_version)
 310                 *nego_fw_version = (icframe_major << 16) | icframe_minor;
 311
 312         if (nego_srv_version)
 313                 *nego_srv_version = (icmsg_major << 16) | icmsg_minor;
 314
 315         negop->icversion_data[0].major = icframe_major;
 316         negop->icversion_data[0].minor = icframe_minor;
 317         negop->icversion_data[1].major = icmsg_major;
 318         negop->icversion_data[1].minor = icmsg_minor;
 319         return found_match;
 320 }
 321
 322 EXPORT_SYMBOL_GPL(vmbus_prep_negotiate_resp);
 323
 324 /*
 325  * alloc_channel - Allocate and initialize a vmbus channel object
 326  */
 327 static struct vmbus_channel *alloc_channel(void)
 328 {
 329         struct vmbus_channel *channel;
 330
 331         channel = kzalloc(sizeof(*channel), GFP_ATOMIC);
 332         if (!channel)
 333                 return NULL;
 334
 335         channel->acquire_ring_lock = true;
 336         spin_lock_init(&channel->inbound_lock);
 337         spin_lock_init(&channel->lock);
 338
 339         INIT_LIST_HEAD(&channel->sc_list);
 340         INIT_LIST_HEAD(&channel->percpu_list);
 341
 342         tasklet_init(&channel->callback_event,
 343                      vmbus_on_event, (unsigned long)channel);
 344
 345         return channel;
 346 }
 347
 348 /*
 349  * free_channel - Release the resources used by the vmbus channel object
 350  */
 351 static void free_channel(struct vmbus_channel *channel)
 352 {
 353         tasklet_kill(&channel->callback_event);
 354         kfree(channel);
 355 }
 356
 357 static void percpu_channel_enq(void *arg)
 358 {
 359         struct vmbus_channel *channel = arg;
 360         struct hv_per_cpu_context *hv_cpu
 361                 = this_cpu_ptr(hv_context.cpu_context);
 362
 363         list_add_tail(&channel->percpu_list, &hv_cpu->chan_list);
 364 }
 365
 366 static void percpu_channel_deq(void *arg)
 367 {
 368         struct vmbus_channel *channel = arg;
 369
 370         list_del(&channel->percpu_list);
 371 }
 372
 373
 374 static void vmbus_release_relid(u32 relid)
 375 {
 376         struct vmbus_channel_relid_released msg;
 377
 378         memset(&msg, 0, sizeof(struct vmbus_channel_relid_released));
 379         msg.child_relid = relid;
 380         msg.header.msgtype = CHANNELMSG_RELID_RELEASED;
 381         vmbus_post_msg(&msg, sizeof(struct vmbus_channel_relid_released),
 382                        true);
 383 }
 384
 385 void hv_event_tasklet_disable(struct vmbus_channel *channel)
 386 {
 387         tasklet_disable(&channel->callback_event);
 388 }
 389
 390 void hv_event_tasklet_enable(struct vmbus_channel *channel)
 391 {
 392         tasklet_enable(&channel->callback_event);
 393
 394         /* In case there is any pending event */
 395         tasklet_schedule(&channel->callback_event);
 396 }
 397
 398 void hv_process_channel_removal(struct vmbus_channel *channel, u32 relid)
 399 {
 400         unsigned long flags;
 401         struct vmbus_channel *primary_channel;
 402
 403         BUG_ON(!channel->rescind);
 404         BUG_ON(!mutex_is_locked(&vmbus_connection.channel_mutex));
 405
 406         hv_event_tasklet_disable(channel);
 407         if (channel->target_cpu != get_cpu()) {
 408                 put_cpu();
 409                 smp_call_function_single(channel->target_cpu,
 410                                          percpu_channel_deq, channel, true);
 411         } else {
 412                 percpu_channel_deq(channel);
 413                 put_cpu();
 414         }
 415         hv_event_tasklet_enable(channel);
 416
 417         if (channel->primary_channel == NULL) {
 418                 list_del(&channel->listentry);
 419
 420                 primary_channel = channel;
 421         } else {
 422                 primary_channel = channel->primary_channel;
 423                 spin_lock_irqsave(&primary_channel->lock, flags);
 424                 list_del(&channel->sc_list);
 425                 primary_channel->num_sc--;
 426                 spin_unlock_irqrestore(&primary_channel->lock, flags);
 427         }
 428
 429         /*
 430          * We need to free the bit for init_vp_index() to work in the case
 431          * of sub-channel, when we reload drivers like hv_netvsc.
 432          */
 433         if (channel->affinity_policy == HV_LOCALIZED)
 434                 cpumask_clear_cpu(channel->target_cpu,
 435                                   &primary_channel->alloced_cpus_in_node);
 436
 437         vmbus_release_relid(relid);
 438
 439         free_channel(channel);
 440 }
 441
 442 void vmbus_free_channels(void)
 443 {
 444         struct vmbus_channel *channel, *tmp;
 445
 446         mutex_lock(&vmbus_connection.channel_mutex);
 447         list_for_each_entry_safe(channel, tmp, &vmbus_connection.chn_list,
 448                 listentry) {
 449                 /* hv_process_channel_removal() needs this */
 450                 channel->rescind = true;
 451
 452                 vmbus_device_unregister(channel->device_obj);
 453         }
 454         mutex_unlock(&vmbus_connection.channel_mutex);
 455 }
 456
 457 /*
 458  * vmbus_process_offer - Process the offer by creating a channel/device
 459  * associated with this offer
 460  */
 461 static void vmbus_process_offer(struct vmbus_channel *newchannel)
 462 {
 463         struct vmbus_channel *channel;
 464         bool fnew = true;
 465         unsigned long flags;
 466         u16 dev_type;
 467         int ret;
 468
 469         /* Make sure this is a new offer */
 470         mutex_lock(&vmbus_connection.channel_mutex);
 471
 472         list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
 473                 if (!uuid_le_cmp(channel->offermsg.offer.if_type,
 474                         newchannel->offermsg.offer.if_type) &&
 475                         !uuid_le_cmp(channel->offermsg.offer.if_instance,
 476                                 newchannel->offermsg.offer.if_instance)) {
 477                         fnew = false;
 478                         break;
 479                 }
 480         }
 481
 482         if (fnew)
 483                 list_add_tail(&newchannel->listentry,
 484                               &vmbus_connection.chn_list);
 485
 486         mutex_unlock(&vmbus_connection.channel_mutex);
 487
 488         if (!fnew) {
 489                 /*
 490                  * Check to see if this is a sub-channel.
 491                  */
 492                 if (newchannel->offermsg.offer.sub_channel_index != 0) {
 493                         /*
 494                          * Process the sub-channel.
 495                          */
 496                         newchannel->primary_channel = channel;
 497                         spin_lock_irqsave(&channel->lock, flags);
 498                         list_add_tail(&newchannel->sc_list, &channel->sc_list);
 499                         channel->num_sc++;
 500                         spin_unlock_irqrestore(&channel->lock, flags);
 501                 } else
 502                         goto err_free_chan;
 503         }
 504
 505         dev_type = hv_get_dev_type(newchannel);
 506
 507         init_vp_index(newchannel, dev_type);
 508
 509         hv_event_tasklet_disable(newchannel);
 510         if (newchannel->target_cpu != get_cpu()) {
 511                 put_cpu();
 512                 smp_call_function_single(newchannel->target_cpu,
 513                                          percpu_channel_enq,
 514                                          newchannel, true);
 515         } else {
 516                 percpu_channel_enq(newchannel);
 517                 put_cpu();
 518         }
 519         hv_event_tasklet_enable(newchannel);
 520
 521         /*
 522          * This state is used to indicate a successful open
 523          * so that when we do close the channel normally, we
 524          * can cleanup properly
 525          */
 526         newchannel->state = CHANNEL_OPEN_STATE;
 527
 528         if (!fnew) {
 529                 if (channel->sc_creation_callback != NULL)
 530                         channel->sc_creation_callback(newchannel);
 531                 return;
 532         }
 533
 534         /*
 535          * Start the process of binding this offer to the driver
 536          * We need to set the DeviceObject field before calling
 537          * vmbus_child_dev_add()
 538          */
 539         newchannel->device_obj = vmbus_device_create(
 540                 &newchannel->offermsg.offer.if_type,
 541                 &newchannel->offermsg.offer.if_instance,
 542                 newchannel);
 543         if (!newchannel->device_obj)
 544                 goto err_deq_chan;
 545
 546         newchannel->device_obj->device_id = dev_type;
 547         /*
 548          * Add the new device to the bus. This will kick off device-driver
 549          * binding which eventually invokes the device driver's AddDevice()
 550          * method.
 551          */
 552         mutex_lock(&vmbus_connection.channel_mutex);
 553         ret = vmbus_device_register(newchannel->device_obj);
 554         mutex_unlock(&vmbus_connection.channel_mutex);
 555
 556         if (ret != 0) {
 557                 pr_err("unable to add child device object (relid %d)\n",
 558                         newchannel->offermsg.child_relid);
 559                 kfree(newchannel->device_obj);
 560                 goto err_deq_chan;
 561         }
 562         return;
 563
 564 err_deq_chan:
 565         mutex_lock(&vmbus_connection.channel_mutex);
 566         list_del(&newchannel->listentry);
 567         mutex_unlock(&vmbus_connection.channel_mutex);
 568
 569         hv_event_tasklet_disable(newchannel);
 570         if (newchannel->target_cpu != get_cpu()) {
 571                 put_cpu();
 572                 smp_call_function_single(newchannel->target_cpu,
 573                                          percpu_channel_deq, newchannel, true);
 574         } else {
 575                 percpu_channel_deq(newchannel);
 576                 put_cpu();
 577         }
 578         hv_event_tasklet_enable(newchannel);
 579
 580         vmbus_release_relid(newchannel->offermsg.child_relid);
 581
 582 err_free_chan:
 583         free_channel(newchannel);
 584 }
 585
 586 /*
 587  * We use this state to statically distribute the channel interrupt load.
 588  */
 589 static int next_numa_node_id;
 590
 591 /*
 592  * Starting with Win8, we can statically distribute the incoming
 593  * channel interrupt load by binding a channel to VCPU.
 594  * We do this in a hierarchical fashion:
 595  * First distribute the primary channels across available NUMA nodes
 596  * and then distribute the subchannels amongst the CPUs in the NUMA
 597  * node assigned to the primary channel.
 598  *
 599  * For pre-win8 hosts or non-performance critical channels we assign the
 600  * first CPU in the first NUMA node.
 601  */
 602 static void init_vp_index(struct vmbus_channel *channel, u16 dev_type)
 603 {
 604         u32 cur_cpu;
 605         bool perf_chn = vmbus_devs[dev_type].perf_device;
 606         struct vmbus_channel *primary = channel->primary_channel;
 607         int next_node;
 608         struct cpumask available_mask;
 609         struct cpumask *alloced_mask;
 610
 611         if ((vmbus_proto_version == VERSION_WS2008) ||
 612             (vmbus_proto_version == VERSION_WIN7) || (!perf_chn)) {
 613                 /*
 614                  * Prior to win8, all channel interrupts are
 615                  * delivered on cpu 0.
 616                  * Also if the channel is not a performance critical
 617                  * channel, bind it to cpu 0.
 618                  */
 619                 channel->numa_node = 0;
 620                 channel->target_cpu = 0;
 621                 channel->target_vp = hv_context.vp_index[0];
 622                 return;
 623         }
 624
 625         /*
 626          * Based on the channel affinity policy, we will assign the NUMA
 627          * nodes.
 628          */
 629
 630         if ((channel->affinity_policy == HV_BALANCED) || (!primary)) {
 631                 while (true) {
 632                         next_node = next_numa_node_id++;
 633                         if (next_node == nr_node_ids) {
 634                                 next_node = next_numa_node_id = 0;
 635                                 continue;
 636                         }
 637                         if (cpumask_empty(cpumask_of_node(next_node)))
 638                                 continue;
 639                         break;
 640                 }
 641                 channel->numa_node = next_node;
 642                 primary = channel;
 643         }
 644         alloced_mask = &hv_context.hv_numa_map[primary->numa_node];
 645
 646         if (cpumask_weight(alloced_mask) ==
 647             cpumask_weight(cpumask_of_node(primary->numa_node))) {
 648                 /*
 649                  * We have cycled through all the CPUs in the node;
 650                  * reset the alloced map.
 651                  */
 652                 cpumask_clear(alloced_mask);
 653         }
 654
 655         cpumask_xor(&available_mask, alloced_mask,
 656                     cpumask_of_node(primary->numa_node));
 657
 658         cur_cpu = -1;
 659
 660         if (primary->affinity_policy == HV_LOCALIZED) {
 661                 /*
 662                  * Normally Hyper-V host doesn't create more subchannels
 663                  * than there are VCPUs on the node but it is possible when not
 664                  * all present VCPUs on the node are initialized by guest.
 665                  * Clear the alloced_cpus_in_node to start over.
 666                  */
 667                 if (cpumask_equal(&primary->alloced_cpus_in_node,
 668                                   cpumask_of_node(primary->numa_node)))
 669                         cpumask_clear(&primary->alloced_cpus_in_node);
 670         }
 671
 672         while (true) {
 673                 cur_cpu = cpumask_next(cur_cpu, &available_mask);
 674                 if (cur_cpu >= nr_cpu_ids) {
 675                         cur_cpu = -1;
 676                         cpumask_copy(&available_mask,
 677                                      cpumask_of_node(primary->numa_node));
 678                         continue;
 679                 }
 680
 681                 if (primary->affinity_policy == HV_LOCALIZED) {
 682                         /*
 683                          * NOTE: in the case of sub-channel, we clear the
 684                          * sub-channel related bit(s) in
 685                          * primary->alloced_cpus_in_node in
 686                          * hv_process_channel_removal(), so when we
 687                          * reload drivers like hv_netvsc in SMP guest, here
 688                          * we're able to re-allocate
 689                          * bit from primary->alloced_cpus_in_node.
 690                          */
 691                         if (!cpumask_test_cpu(cur_cpu,
 692                                               &primary->alloced_cpus_in_node)) {
 693                                 cpumask_set_cpu(cur_cpu,
 694                                                 &primary->alloced_cpus_in_node);
 695                                 cpumask_set_cpu(cur_cpu, alloced_mask);
 696                                 break;
 697                         }
 698                 } else {
 699                         cpumask_set_cpu(cur_cpu, alloced_mask);
 700                         break;
 701                 }
 702         }
 703
 704         channel->target_cpu = cur_cpu;
 705         channel->target_vp = hv_context.vp_index[cur_cpu];
 706 }
 707
 708 static void vmbus_wait_for_unload(void)
 709 {
 710         int cpu;
 711         void *page_addr;
 712         struct hv_message *msg;
 713         struct vmbus_channel_message_header *hdr;
 714         u32 message_type;
 715
 716         /*
 717          * CHANNELMSG_UNLOAD_RESPONSE is always delivered to the CPU which was
 718          * used for initial contact or to CPU0 depending on host version. When
 719          * we're crashing on a different CPU let's hope that IRQ handler on
 720          * the cpu which receives CHANNELMSG_UNLOAD_RESPONSE is still
 721          * functional and vmbus_unload_response() will complete
 722          * vmbus_connection.unload_event. If not, the last thing we can do is
 723          * read message pages for all CPUs directly.
 724          */
 725         while (1) {
 726                 if (completion_done(&vmbus_connection.unload_event))
 727                         break;
 728
 729                 for_each_online_cpu(cpu) {
 730                         struct hv_per_cpu_context *hv_cpu
 731                                 = per_cpu_ptr(hv_context.cpu_context, cpu);
 732
 733                         page_addr = hv_cpu->synic_message_page;
 734                         msg = (struct hv_message *)page_addr
 735                                 + VMBUS_MESSAGE_SINT;
 736
 737                         message_type = READ_ONCE(msg->header.message_type);
 738                         if (message_type == HVMSG_NONE)
 739                                 continue;
 740
 741                         hdr = (struct vmbus_channel_message_header *)
 742                                 msg->u.payload;
 743
 744                         if (hdr->msgtype == CHANNELMSG_UNLOAD_RESPONSE)
 745                                 complete(&vmbus_connection.unload_event);
 746
 747                         vmbus_signal_eom(msg, message_type);
 748                 }
 749
 750                 mdelay(10);
 751         }
 752
 753         /*
 754          * We're crashing and already got the UNLOAD_RESPONSE, cleanup all
 755          * maybe-pending messages on all CPUs to be able to receive new
 756          * messages after we reconnect.
 757          */
 758         for_each_online_cpu(cpu) {
 759                 struct hv_per_cpu_context *hv_cpu
 760                         = per_cpu_ptr(hv_context.cpu_context, cpu);
 761
 762                 page_addr = hv_cpu->synic_message_page;
 763                 msg = (struct hv_message *)page_addr + VMBUS_MESSAGE_SINT;
 764                 msg->header.message_type = HVMSG_NONE;
 765         }
 766 }
 767
 768 /*
 769  * vmbus_unload_response - Handler for the unload response.
 770  */
 771 static void vmbus_unload_response(struct vmbus_channel_message_header *hdr)
 772 {
 773         /*
 774          * This is a global event; just wakeup the waiting thread.
 775          * Once we successfully unload, we can cleanup the monitor state.
 776          */
 777         complete(&vmbus_connection.unload_event);
 778 }
 779
 780 void vmbus_initiate_unload(bool crash)
 781 {
 782         struct vmbus_channel_message_header hdr;
 783
 784         /* Pre-Win2012R2 hosts don't support reconnect */
 785         if (vmbus_proto_version < VERSION_WIN8_1)
 786                 return;
 787
 788         init_completion(&vmbus_connection.unload_event);
 789         memset(&hdr, 0, sizeof(struct vmbus_channel_message_header));
 790         hdr.msgtype = CHANNELMSG_UNLOAD;
 791         vmbus_post_msg(&hdr, sizeof(struct vmbus_channel_message_header),
 792                        !crash);
 793
 794         /*
 795          * vmbus_initiate_unload() is also called on crash and the crash can be
 796          * happening in an interrupt context, where scheduling is impossible.
 797          */
 798         if (!crash)
 799                 wait_for_completion(&vmbus_connection.unload_event);
 800         else
 801                 vmbus_wait_for_unload();
 802 }
 803
 804 /*
 805  * vmbus_onoffer - Handler for channel offers from vmbus in parent partition.
 806  *
 807  */
 808 static void vmbus_onoffer(struct vmbus_channel_message_header *hdr)
 809 {
 810         struct vmbus_channel_offer_channel *offer;
 811         struct vmbus_channel *newchannel;
 812
 813         offer = (struct vmbus_channel_offer_channel *)hdr;
 814
 815         /* Allocate the channel object and save this offer. */
 816         newchannel = alloc_channel();
 817         if (!newchannel) {
 818                 pr_err("Unable to allocate channel object\n");
 819                 return;
 820         }
 821
 822         /*
 823          * By default we setup state to enable batched
 824          * reading. A specific service can choose to
 825          * disable this prior to opening the channel.
 826          */
 827         newchannel->batched_reading = true;
 828
 829         /*
 830          * Setup state for signalling the host.
 831          */
 832         newchannel->sig_event = (struct hv_input_signal_event *)
 833                                 (ALIGN((unsigned long)
 834                                 &newchannel->sig_buf,
 835                                 HV_HYPERCALL_PARAM_ALIGN));
 836
 837         newchannel->sig_event->connectionid.asu32 = 0;
 838         newchannel->sig_event->connectionid.u.id = VMBUS_EVENT_CONNECTION_ID;
 839         newchannel->sig_event->flag_number = 0;
 840         newchannel->sig_event->rsvdz = 0;
 841
 842         if (vmbus_proto_version != VERSION_WS2008) {
 843                 newchannel->is_dedicated_interrupt =
 844                                 (offer->is_dedicated_interrupt != 0);
 845                 newchannel->sig_event->connectionid.u.id =
 846                                 offer->connection_id;
 847         }
 848
 849         memcpy(&newchannel->offermsg, offer,
 850                sizeof(struct vmbus_channel_offer_channel));
 851         newchannel->monitor_grp = (u8)offer->monitorid / 32;
 852         newchannel->monitor_bit = (u8)offer->monitorid % 32;
 853
 854         vmbus_process_offer(newchannel);
 855 }
 856
 857 /*
 858  * vmbus_onoffer_rescind - Rescind offer handler.
 859  *
 860  * We queue a work item to process this offer synchronously
 861  */
 862 static void vmbus_onoffer_rescind(struct vmbus_channel_message_header *hdr)
 863 {
 864         struct vmbus_channel_rescind_offer *rescind;
 865         struct vmbus_channel *channel;
 866         unsigned long flags;
 867         struct device *dev;
 868
 869         rescind = (struct vmbus_channel_rescind_offer *)hdr;
 870
 871         mutex_lock(&vmbus_connection.channel_mutex);
 872         channel = relid2channel(rescind->child_relid);
 873
 874         if (channel == NULL) {
 875                 /*
 876                  * This is very impossible, because in
 877                  * vmbus_process_offer(), we have already invoked
 878                  * vmbus_release_relid() on error.
 879                  */
 880                 goto out;
 881         }
 882
 883         spin_lock_irqsave(&channel->lock, flags);
 884         channel->rescind = true;
 885         spin_unlock_irqrestore(&channel->lock, flags);
 886
 887         vmbus_rescind_cleanup(channel);
 888
 889         if (channel->device_obj) {
 890                 if (channel->chn_rescind_callback) {
 891                         channel->chn_rescind_callback(channel);
 892                         goto out;
 893                 }
 894                 /*
 895                  * We will have to unregister this device from the
 896                  * driver core.
 897                  */
 898                 dev = get_device(&channel->device_obj->device);
 899                 if (dev) {
 900                         vmbus_device_unregister(channel->device_obj);
 901                         put_device(dev);
 902                 }
 903         } else {
 904                 hv_process_channel_removal(channel,
 905                         channel->offermsg.child_relid);
 906         }
 907
 908 out:
 909         mutex_unlock(&vmbus_connection.channel_mutex);
 910 }
 911
 912 void vmbus_hvsock_device_unregister(struct vmbus_channel *channel)
 913 {
 914         mutex_lock(&vmbus_connection.channel_mutex);
 915
 916         BUG_ON(!is_hvsock_channel(channel));
 917
 918         channel->rescind = true;
 919         vmbus_device_unregister(channel->device_obj);
 920
 921         mutex_unlock(&vmbus_connection.channel_mutex);
 922 }
 923 EXPORT_SYMBOL_GPL(vmbus_hvsock_device_unregister);
 924
 925
 926 /*
 927  * vmbus_onoffers_delivered -
 928  * This is invoked when all offers have been delivered.
 929  *
 930  * Nothing to do here.
 931  */
 932 static void vmbus_onoffers_delivered(
 933                         struct vmbus_channel_message_header *hdr)
 934 {
 935 }
 936
 937 /*
 938  * vmbus_onopen_result - Open result handler.
 939  *
 940  * This is invoked when we received a response to our channel open request.
 941  * Find the matching request, copy the response and signal the requesting
 942  * thread.
 943  */
 944 static void vmbus_onopen_result(struct vmbus_channel_message_header *hdr)
 945 {
 946         struct vmbus_channel_open_result *result;
 947         struct vmbus_channel_msginfo *msginfo;
 948         struct vmbus_channel_message_header *requestheader;
 949         struct vmbus_channel_open_channel *openmsg;
 950         unsigned long flags;
 951
 952         result = (struct vmbus_channel_open_result *)hdr;
 953
 954         /*
 955          * Find the open msg, copy the result and signal/unblock the wait event
 956          */
 957         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
 958
 959         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
 960                                 msglistentry) {
 961                 requestheader =
 962                         (struct vmbus_channel_message_header *)msginfo->msg;
 963
 964                 if (requestheader->msgtype == CHANNELMSG_OPENCHANNEL) {
 965                         openmsg =
 966                         (struct vmbus_channel_open_channel *)msginfo->msg;
 967                         if (openmsg->child_relid == result->child_relid &&
 968                             openmsg->openid == result->openid) {
 969                                 memcpy(&msginfo->response.open_result,
 970                                        result,
 971                                        sizeof(
 972                                         struct vmbus_channel_open_result));
 973                                 complete(&msginfo->waitevent);
 974                                 break;
 975                         }
 976                 }
 977         }
 978         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
 979 }
 980
 981 /*
 982  * vmbus_ongpadl_created - GPADL created handler.
 983  *
 984  * This is invoked when we received a response to our gpadl create request.
 985  * Find the matching request, copy the response and signal the requesting
 986  * thread.
 987  */
 988 static void vmbus_ongpadl_created(struct vmbus_channel_message_header *hdr)
 989 {
 990         struct vmbus_channel_gpadl_created *gpadlcreated;
 991         struct vmbus_channel_msginfo *msginfo;
 992         struct vmbus_channel_message_header *requestheader;
 993         struct vmbus_channel_gpadl_header *gpadlheader;
 994         unsigned long flags;
 995
 996         gpadlcreated = (struct vmbus_channel_gpadl_created *)hdr;
 997
 998         /*
 999          * Find the establish msg, copy the result and signal/unblock the wait
1000          * event
1001          */
1002         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1003
1004         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1005                                 msglistentry) {
1006                 requestheader =
1007                         (struct vmbus_channel_message_header *)msginfo->msg;
1008
1009                 if (requestheader->msgtype == CHANNELMSG_GPADL_HEADER) {
1010                         gpadlheader =
1011                         (struct vmbus_channel_gpadl_header *)requestheader;
1012
1013                         if ((gpadlcreated->child_relid ==
1014                              gpadlheader->child_relid) &&
1015                             (gpadlcreated->gpadl == gpadlheader->gpadl)) {
1016                                 memcpy(&msginfo->response.gpadl_created,
1017                                        gpadlcreated,
1018                                        sizeof(
1019                                         struct vmbus_channel_gpadl_created));
1020                                 complete(&msginfo->waitevent);
1021                                 break;
1022                         }
1023                 }
1024         }
1025         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1026 }
1027
1028 /*
1029  * vmbus_ongpadl_torndown - GPADL torndown handler.
1030  *
1031  * This is invoked when we received a response to our gpadl teardown request.
1032  * Find the matching request, copy the response and signal the requesting
1033  * thread.
1034  */
1035 static void vmbus_ongpadl_torndown(
1036                         struct vmbus_channel_message_header *hdr)
1037 {
1038         struct vmbus_channel_gpadl_torndown *gpadl_torndown;
1039         struct vmbus_channel_msginfo *msginfo;
1040         struct vmbus_channel_message_header *requestheader;
1041         struct vmbus_channel_gpadl_teardown *gpadl_teardown;
1042         unsigned long flags;
1043
1044         gpadl_torndown = (struct vmbus_channel_gpadl_torndown *)hdr;
1045
1046         /*
1047          * Find the open msg, copy the result and signal/unblock the wait event
1048          */
1049         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1050
1051         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1052                                 msglistentry) {
1053                 requestheader =
1054                         (struct vmbus_channel_message_header *)msginfo->msg;
1055
1056                 if (requestheader->msgtype == CHANNELMSG_GPADL_TEARDOWN) {
1057                         gpadl_teardown =
1058                         (struct vmbus_channel_gpadl_teardown *)requestheader;
1059
1060                         if (gpadl_torndown->gpadl == gpadl_teardown->gpadl) {
1061                                 memcpy(&msginfo->response.gpadl_torndown,
1062                                        gpadl_torndown,
1063                                        sizeof(
1064                                         struct vmbus_channel_gpadl_torndown));
1065                                 complete(&msginfo->waitevent);
1066                                 break;
1067                         }
1068                 }
1069         }
1070         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1071 }
1072
1073 /*
1074  * vmbus_onversion_response - Version response handler
1075  *
1076  * This is invoked when we received a response to our initiate contact request.
1077  * Find the matching request, copy the response and signal the requesting
1078  * thread.
1079  */
1080 static void vmbus_onversion_response(
1081                 struct vmbus_channel_message_header *hdr)
1082 {
1083         struct vmbus_channel_msginfo *msginfo;
1084         struct vmbus_channel_message_header *requestheader;
1085         struct vmbus_channel_version_response *version_response;
1086         unsigned long flags;
1087
1088         version_response = (struct vmbus_channel_version_response *)hdr;
1089         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1090
1091         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1092                                 msglistentry) {
1093                 requestheader =
1094                         (struct vmbus_channel_message_header *)msginfo->msg;
1095
1096                 if (requestheader->msgtype ==
1097                     CHANNELMSG_INITIATE_CONTACT) {
1098                         memcpy(&msginfo->response.version_response,
1099                               version_response,
1100                               sizeof(struct vmbus_channel_version_response));
1101                         complete(&msginfo->waitevent);
1102                 }
1103         }
1104         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1105 }
1106
1107 /* Channel message dispatch table */
1108 struct vmbus_channel_message_table_entry
1109         channel_message_table[CHANNELMSG_COUNT] = {
1110         {CHANNELMSG_INVALID,                    0, NULL},
1111         {CHANNELMSG_OFFERCHANNEL,               0, vmbus_onoffer},
1112         {CHANNELMSG_RESCIND_CHANNELOFFER,       0, vmbus_onoffer_rescind},
1113         {CHANNELMSG_REQUESTOFFERS,              0, NULL},
1114         {CHANNELMSG_ALLOFFERS_DELIVERED,        1, vmbus_onoffers_delivered},
1115         {CHANNELMSG_OPENCHANNEL,                0, NULL},
1116         {CHANNELMSG_OPENCHANNEL_RESULT,         1, vmbus_onopen_result},
1117         {CHANNELMSG_CLOSECHANNEL,               0, NULL},
1118         {CHANNELMSG_GPADL_HEADER,               0, NULL},
1119         {CHANNELMSG_GPADL_BODY,                 0, NULL},
1120         {CHANNELMSG_GPADL_CREATED,              1, vmbus_ongpadl_created},
1121         {CHANNELMSG_GPADL_TEARDOWN,             0, NULL},
1122         {CHANNELMSG_GPADL_TORNDOWN,             1, vmbus_ongpadl_torndown},
1123         {CHANNELMSG_RELID_RELEASED,             0, NULL},
1124         {CHANNELMSG_INITIATE_CONTACT,           0, NULL},
1125         {CHANNELMSG_VERSION_RESPONSE,           1, vmbus_onversion_response},
1126         {CHANNELMSG_UNLOAD,                     0, NULL},
1127         {CHANNELMSG_UNLOAD_RESPONSE,            1, vmbus_unload_response},
1128         {CHANNELMSG_18,                         0, NULL},
1129         {CHANNELMSG_19,                         0, NULL},
1130         {CHANNELMSG_20,                         0, NULL},
1131         {CHANNELMSG_TL_CONNECT_REQUEST,         0, NULL},
1132 };
1133
1134 /*
1135  * vmbus_onmessage - Handler for channel protocol messages.
1136  *
1137  * This is invoked in the vmbus worker thread context.
1138  */
1139 void vmbus_onmessage(void *context)
1140 {
1141         struct hv_message *msg = context;
1142         struct vmbus_channel_message_header *hdr;
1143         int size;
1144
1145         hdr = (struct vmbus_channel_message_header *)msg->u.payload;
1146         size = msg->header.payload_size;
1147
1148         if (hdr->msgtype >= CHANNELMSG_COUNT) {
1149                 pr_err("Received invalid channel message type %d size %d\n",
1150                            hdr->msgtype, size);
1151                 print_hex_dump_bytes("", DUMP_PREFIX_NONE,
1152                                      (unsigned char *)msg->u.payload, size);
1153                 return;
1154         }
1155
1156         if (channel_message_table[hdr->msgtype].message_handler)
1157                 channel_message_table[hdr->msgtype].message_handler(hdr);
1158         else
1159                 pr_err("Unhandled channel message type %d\n", hdr->msgtype);
1160 }
1161
1162 /*
1163  * vmbus_request_offers - Send a request to get all our pending offers.
1164  */
1165 int vmbus_request_offers(void)
1166 {
1167         struct vmbus_channel_message_header *msg;
1168         struct vmbus_channel_msginfo *msginfo;
1169         int ret;
1170
1171         msginfo = kmalloc(sizeof(*msginfo) +
1172                           sizeof(struct vmbus_channel_message_header),
1173                           GFP_KERNEL);
1174         if (!msginfo)
1175                 return -ENOMEM;
1176
1177         msg = (struct vmbus_channel_message_header *)msginfo->msg;
1178
1179         msg->msgtype = CHANNELMSG_REQUESTOFFERS;
1180
1181
1182         ret = vmbus_post_msg(msg, sizeof(struct vmbus_channel_message_header),
1183                              true);
1184         if (ret != 0) {
1185                 pr_err("Unable to request offers - %d\n", ret);
1186
1187                 goto cleanup;
1188         }
1189
1190 cleanup:
1191         kfree(msginfo);
1192
1193         return ret;
1194 }
1195
1196 /*
1197  * Retrieve the (sub) channel on which to send an outgoing request.
1198  * When a primary channel has multiple sub-channels, we try to
1199  * distribute the load equally amongst all available channels.
1200  */
1201 struct vmbus_channel *vmbus_get_outgoing_channel(struct vmbus_channel *primary)
1202 {
1203         struct list_head *cur, *tmp;
1204         int cur_cpu;
1205         struct vmbus_channel *cur_channel;
1206         struct vmbus_channel *outgoing_channel = primary;
1207         int next_channel;
1208         int i = 1;
1209
1210         if (list_empty(&primary->sc_list))
1211                 return outgoing_channel;
1212
1213         next_channel = primary->next_oc++;
1214
1215         if (next_channel > (primary->num_sc)) {
1216                 primary->next_oc = 0;
1217                 return outgoing_channel;
1218         }
1219
1220         cur_cpu = hv_context.vp_index[get_cpu()];
1221         put_cpu();
1222         list_for_each_safe(cur, tmp, &primary->sc_list) {
1223                 cur_channel = list_entry(cur, struct vmbus_channel, sc_list);
1224                 if (cur_channel->state != CHANNEL_OPENED_STATE)
1225                         continue;
1226
1227                 if (cur_channel->target_vp == cur_cpu)
1228                         return cur_channel;
1229
1230                 if (i == next_channel)
1231                         return cur_channel;
1232
1233                 i++;
1234         }
1235
1236         return outgoing_channel;
1237 }
1238 EXPORT_SYMBOL_GPL(vmbus_get_outgoing_channel);
1239
1240 static void invoke_sc_cb(struct vmbus_channel *primary_channel)
1241 {
1242         struct list_head *cur, *tmp;
1243         struct vmbus_channel *cur_channel;
1244
1245         if (primary_channel->sc_creation_callback == NULL)
1246                 return;
1247
1248         list_for_each_safe(cur, tmp, &primary_channel->sc_list) {
1249                 cur_channel = list_entry(cur, struct vmbus_channel, sc_list);
1250
1251                 primary_channel->sc_creation_callback(cur_channel);
1252         }
1253 }
1254
1255 void vmbus_set_sc_create_callback(struct vmbus_channel *primary_channel,
1256                                 void (*sc_cr_cb)(struct vmbus_channel *new_sc))
1257 {
1258         primary_channel->sc_creation_callback = sc_cr_cb;
1259 }
1260 EXPORT_SYMBOL_GPL(vmbus_set_sc_create_callback);
1261
1262 bool vmbus_are_subchannels_present(struct vmbus_channel *primary)
1263 {
1264         bool ret;
1265
1266         ret = !list_empty(&primary->sc_list);
1267
1268         if (ret) {
1269                 /*
1270                  * Invoke the callback on sub-channel creation.
1271                  * This will present a uniform interface to the
1272                  * clients.
1273                  */
1274                 invoke_sc_cb(primary);
1275         }
1276
1277         return ret;
1278 }
1279 EXPORT_SYMBOL_GPL(vmbus_are_subchannels_present);
1280
1281 void vmbus_set_chn_rescind_callback(struct vmbus_channel *channel,
1282                 void (*chn_rescind_cb)(struct vmbus_channel *))
1283 {
1284         channel->chn_rescind_callback = chn_rescind_cb;
1285 }
1286 EXPORT_SYMBOL_GPL(vmbus_set_chn_rescind_callback);