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         channel->rescind = true;
 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         spin_lock_init(&channel->lock);
 336
 337         INIT_LIST_HEAD(&channel->sc_list);
 338         INIT_LIST_HEAD(&channel->percpu_list);
 339
 340         tasklet_init(&channel->callback_event,
 341                      vmbus_on_event, (unsigned long)channel);
 342
 343         return channel;
 344 }
 345
 346 /*
 347  * free_channel - Release the resources used by the vmbus channel object
 348  */
 349 static void free_channel(struct vmbus_channel *channel)
 350 {
 351         tasklet_kill(&channel->callback_event);
 352
 353         kobject_put(&channel->kobj);
 354 }
 355
 356 static void percpu_channel_enq(void *arg)
 357 {
 358         struct vmbus_channel *channel = arg;
 359         struct hv_per_cpu_context *hv_cpu
 360                 = this_cpu_ptr(hv_context.cpu_context);
 361
 362         list_add_tail_rcu(&channel->percpu_list, &hv_cpu->chan_list);
 363 }
 364
 365 static void percpu_channel_deq(void *arg)
 366 {
 367         struct vmbus_channel *channel = arg;
 368
 369         list_del_rcu(&channel->percpu_list);
 370 }
 371
 372
 373 static void vmbus_release_relid(u32 relid)
 374 {
 375         struct vmbus_channel_relid_released msg;
 376         int ret;
 377
 378         memset(&msg, 0, sizeof(struct vmbus_channel_relid_released));
 379         msg.child_relid = relid;
 380         msg.header.msgtype = CHANNELMSG_RELID_RELEASED;
 381         ret = vmbus_post_msg(&msg, sizeof(struct vmbus_channel_relid_released),
 382                              true);
 383
 384         trace_vmbus_release_relid(&msg, ret);
 385 }
 386
 387 void hv_process_channel_removal(u32 relid)
 388 {
 389         unsigned long flags;
 390         struct vmbus_channel *primary_channel, *channel;
 391
 392         BUG_ON(!mutex_is_locked(&vmbus_connection.channel_mutex));
 393
 394         /*
 395          * Make sure channel is valid as we may have raced.
 396          */
 397         channel = relid2channel(relid);
 398         if (!channel)
 399                 return;
 400
 401         BUG_ON(!channel->rescind);
 402         if (channel->target_cpu != get_cpu()) {
 403                 put_cpu();
 404                 smp_call_function_single(channel->target_cpu,
 405                                          percpu_channel_deq, channel, true);
 406         } else {
 407                 percpu_channel_deq(channel);
 408                 put_cpu();
 409         }
 410
 411         if (channel->primary_channel == NULL) {
 412                 list_del(&channel->listentry);
 413
 414                 primary_channel = channel;
 415         } else {
 416                 primary_channel = channel->primary_channel;
 417                 spin_lock_irqsave(&primary_channel->lock, flags);
 418                 list_del(&channel->sc_list);
 419                 primary_channel->num_sc--;
 420                 spin_unlock_irqrestore(&primary_channel->lock, flags);
 421         }
 422
 423         /*
 424          * We need to free the bit for init_vp_index() to work in the case
 425          * of sub-channel, when we reload drivers like hv_netvsc.
 426          */
 427         if (channel->affinity_policy == HV_LOCALIZED)
 428                 cpumask_clear_cpu(channel->target_cpu,
 429                                   &primary_channel->alloced_cpus_in_node);
 430
 431         vmbus_release_relid(relid);
 432
 433         free_channel(channel);
 434 }
 435
 436 void vmbus_free_channels(void)
 437 {
 438         struct vmbus_channel *channel, *tmp;
 439
 440         list_for_each_entry_safe(channel, tmp, &vmbus_connection.chn_list,
 441                 listentry) {
 442                 /* hv_process_channel_removal() needs this */
 443                 channel->rescind = true;
 444
 445                 vmbus_device_unregister(channel->device_obj);
 446         }
 447 }
 448
 449 /*
 450  * vmbus_process_offer - Process the offer by creating a channel/device
 451  * associated with this offer
 452  */
 453 static void vmbus_process_offer(struct vmbus_channel *newchannel)
 454 {
 455         struct vmbus_channel *channel;
 456         bool fnew = true;
 457         unsigned long flags;
 458         u16 dev_type;
 459         int ret;
 460
 461         /* Make sure this is a new offer */
 462         mutex_lock(&vmbus_connection.channel_mutex);
 463
 464         /*
 465          * Now that we have acquired the channel_mutex,
 466          * we can release the potentially racing rescind thread.
 467          */
 468         atomic_dec(&vmbus_connection.offer_in_progress);
 469
 470         list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
 471                 if (!uuid_le_cmp(channel->offermsg.offer.if_type,
 472                         newchannel->offermsg.offer.if_type) &&
 473                         !uuid_le_cmp(channel->offermsg.offer.if_instance,
 474                                 newchannel->offermsg.offer.if_instance)) {
 475                         fnew = false;
 476                         break;
 477                 }
 478         }
 479
 480         if (fnew)
 481                 list_add_tail(&newchannel->listentry,
 482                               &vmbus_connection.chn_list);
 483
 484         mutex_unlock(&vmbus_connection.channel_mutex);
 485
 486         if (!fnew) {
 487                 /*
 488                  * Check to see if this is a sub-channel.
 489                  */
 490                 if (newchannel->offermsg.offer.sub_channel_index != 0) {
 491                         /*
 492                          * Process the sub-channel.
 493                          */
 494                         newchannel->primary_channel = channel;
 495                         spin_lock_irqsave(&channel->lock, flags);
 496                         list_add_tail(&newchannel->sc_list, &channel->sc_list);
 497                         channel->num_sc++;
 498                         spin_unlock_irqrestore(&channel->lock, flags);
 499                 } else {
 500                         goto err_free_chan;
 501                 }
 502         }
 503
 504         dev_type = hv_get_dev_type(newchannel);
 505
 506         init_vp_index(newchannel, dev_type);
 507
 508         if (newchannel->target_cpu != get_cpu()) {
 509                 put_cpu();
 510                 smp_call_function_single(newchannel->target_cpu,
 511                                          percpu_channel_enq,
 512                                          newchannel, true);
 513         } else {
 514                 percpu_channel_enq(newchannel);
 515                 put_cpu();
 516         }
 517
 518         /*
 519          * This state is used to indicate a successful open
 520          * so that when we do close the channel normally, we
 521          * can cleanup properly
 522          */
 523         newchannel->state = CHANNEL_OPEN_STATE;
 524
 525         if (!fnew) {
 526                 struct hv_device *dev
 527                         = newchannel->primary_channel->device_obj;
 528
 529                 if (vmbus_add_channel_kobj(dev, newchannel)) {
 530                         atomic_dec(&vmbus_connection.offer_in_progress);
 531                         goto err_free_chan;
 532                 }
 533
 534                 if (channel->sc_creation_callback != NULL)
 535                         channel->sc_creation_callback(newchannel);
 536                 newchannel->probe_done = true;
 537                 return;
 538         }
 539
 540         /*
 541          * Start the process of binding this offer to the driver
 542          * We need to set the DeviceObject field before calling
 543          * vmbus_child_dev_add()
 544          */
 545         newchannel->device_obj = vmbus_device_create(
 546                 &newchannel->offermsg.offer.if_type,
 547                 &newchannel->offermsg.offer.if_instance,
 548                 newchannel);
 549         if (!newchannel->device_obj)
 550                 goto err_deq_chan;
 551
 552         newchannel->device_obj->device_id = dev_type;
 553         /*
 554          * Add the new device to the bus. This will kick off device-driver
 555          * binding which eventually invokes the device driver's AddDevice()
 556          * method.
 557          */
 558         ret = vmbus_device_register(newchannel->device_obj);
 559
 560         if (ret != 0) {
 561                 pr_err("unable to add child device object (relid %d)\n",
 562                         newchannel->offermsg.child_relid);
 563                 kfree(newchannel->device_obj);
 564                 goto err_deq_chan;
 565         }
 566
 567         newchannel->probe_done = true;
 568         return;
 569
 570 err_deq_chan:
 571         mutex_lock(&vmbus_connection.channel_mutex);
 572         list_del(&newchannel->listentry);
 573         mutex_unlock(&vmbus_connection.channel_mutex);
 574
 575         if (newchannel->target_cpu != get_cpu()) {
 576                 put_cpu();
 577                 smp_call_function_single(newchannel->target_cpu,
 578                                          percpu_channel_deq, newchannel, true);
 579         } else {
 580                 percpu_channel_deq(newchannel);
 581                 put_cpu();
 582         }
 583
 584         vmbus_release_relid(newchannel->offermsg.child_relid);
 585
 586 err_free_chan:
 587         free_channel(newchannel);
 588 }
 589
 590 /*
 591  * We use this state to statically distribute the channel interrupt load.
 592  */
 593 static int next_numa_node_id;
 594
 595 /*
 596  * Starting with Win8, we can statically distribute the incoming
 597  * channel interrupt load by binding a channel to VCPU.
 598  * We do this in a hierarchical fashion:
 599  * First distribute the primary channels across available NUMA nodes
 600  * and then distribute the subchannels amongst the CPUs in the NUMA
 601  * node assigned to the primary channel.
 602  *
 603  * For pre-win8 hosts or non-performance critical channels we assign the
 604  * first CPU in the first NUMA node.
 605  */
 606 static void init_vp_index(struct vmbus_channel *channel, u16 dev_type)
 607 {
 608         u32 cur_cpu;
 609         bool perf_chn = vmbus_devs[dev_type].perf_device;
 610         struct vmbus_channel *primary = channel->primary_channel;
 611         int next_node;
 612         struct cpumask available_mask;
 613         struct cpumask *alloced_mask;
 614
 615         if ((vmbus_proto_version == VERSION_WS2008) ||
 616             (vmbus_proto_version == VERSION_WIN7) || (!perf_chn)) {
 617                 /*
 618                  * Prior to win8, all channel interrupts are
 619                  * delivered on cpu 0.
 620                  * Also if the channel is not a performance critical
 621                  * channel, bind it to cpu 0.
 622                  */
 623                 channel->numa_node = 0;
 624                 channel->target_cpu = 0;
 625                 channel->target_vp = hv_cpu_number_to_vp_number(0);
 626                 return;
 627         }
 628
 629         /*
 630          * Based on the channel affinity policy, we will assign the NUMA
 631          * nodes.
 632          */
 633
 634         if ((channel->affinity_policy == HV_BALANCED) || (!primary)) {
 635                 while (true) {
 636                         next_node = next_numa_node_id++;
 637                         if (next_node == nr_node_ids) {
 638                                 next_node = next_numa_node_id = 0;
 639                                 continue;
 640                         }
 641                         if (cpumask_empty(cpumask_of_node(next_node)))
 642                                 continue;
 643                         break;
 644                 }
 645                 channel->numa_node = next_node;
 646                 primary = channel;
 647         }
 648         alloced_mask = &hv_context.hv_numa_map[primary->numa_node];
 649
 650         if (cpumask_weight(alloced_mask) ==
 651             cpumask_weight(cpumask_of_node(primary->numa_node))) {
 652                 /*
 653                  * We have cycled through all the CPUs in the node;
 654                  * reset the alloced map.
 655                  */
 656                 cpumask_clear(alloced_mask);
 657         }
 658
 659         cpumask_xor(&available_mask, alloced_mask,
 660                     cpumask_of_node(primary->numa_node));
 661
 662         cur_cpu = -1;
 663
 664         if (primary->affinity_policy == HV_LOCALIZED) {
 665                 /*
 666                  * Normally Hyper-V host doesn't create more subchannels
 667                  * than there are VCPUs on the node but it is possible when not
 668                  * all present VCPUs on the node are initialized by guest.
 669                  * Clear the alloced_cpus_in_node to start over.
 670                  */
 671                 if (cpumask_equal(&primary->alloced_cpus_in_node,
 672                                   cpumask_of_node(primary->numa_node)))
 673                         cpumask_clear(&primary->alloced_cpus_in_node);
 674         }
 675
 676         while (true) {
 677                 cur_cpu = cpumask_next(cur_cpu, &available_mask);
 678                 if (cur_cpu >= nr_cpu_ids) {
 679                         cur_cpu = -1;
 680                         cpumask_copy(&available_mask,
 681                                      cpumask_of_node(primary->numa_node));
 682                         continue;
 683                 }
 684
 685                 if (primary->affinity_policy == HV_LOCALIZED) {
 686                         /*
 687                          * NOTE: in the case of sub-channel, we clear the
 688                          * sub-channel related bit(s) in
 689                          * primary->alloced_cpus_in_node in
 690                          * hv_process_channel_removal(), so when we
 691                          * reload drivers like hv_netvsc in SMP guest, here
 692                          * we're able to re-allocate
 693                          * bit from primary->alloced_cpus_in_node.
 694                          */
 695                         if (!cpumask_test_cpu(cur_cpu,
 696                                               &primary->alloced_cpus_in_node)) {
 697                                 cpumask_set_cpu(cur_cpu,
 698                                                 &primary->alloced_cpus_in_node);
 699                                 cpumask_set_cpu(cur_cpu, alloced_mask);
 700                                 break;
 701                         }
 702                 } else {
 703                         cpumask_set_cpu(cur_cpu, alloced_mask);
 704                         break;
 705                 }
 706         }
 707
 708         channel->target_cpu = cur_cpu;
 709         channel->target_vp = hv_cpu_number_to_vp_number(cur_cpu);
 710 }
 711
 712 static void vmbus_wait_for_unload(void)
 713 {
 714         int cpu;
 715         void *page_addr;
 716         struct hv_message *msg;
 717         struct vmbus_channel_message_header *hdr;
 718         u32 message_type;
 719
 720         /*
 721          * CHANNELMSG_UNLOAD_RESPONSE is always delivered to the CPU which was
 722          * used for initial contact or to CPU0 depending on host version. When
 723          * we're crashing on a different CPU let's hope that IRQ handler on
 724          * the cpu which receives CHANNELMSG_UNLOAD_RESPONSE is still
 725          * functional and vmbus_unload_response() will complete
 726          * vmbus_connection.unload_event. If not, the last thing we can do is
 727          * read message pages for all CPUs directly.
 728          */
 729         while (1) {
 730                 if (completion_done(&vmbus_connection.unload_event))
 731                         break;
 732
 733                 for_each_online_cpu(cpu) {
 734                         struct hv_per_cpu_context *hv_cpu
 735                                 = per_cpu_ptr(hv_context.cpu_context, cpu);
 736
 737                         page_addr = hv_cpu->synic_message_page;
 738                         msg = (struct hv_message *)page_addr
 739                                 + VMBUS_MESSAGE_SINT;
 740
 741                         message_type = READ_ONCE(msg->header.message_type);
 742                         if (message_type == HVMSG_NONE)
 743                                 continue;
 744
 745                         hdr = (struct vmbus_channel_message_header *)
 746                                 msg->u.payload;
 747
 748                         if (hdr->msgtype == CHANNELMSG_UNLOAD_RESPONSE)
 749                                 complete(&vmbus_connection.unload_event);
 750
 751                         vmbus_signal_eom(msg, message_type);
 752                 }
 753
 754                 mdelay(10);
 755         }
 756
 757         /*
 758          * We're crashing and already got the UNLOAD_RESPONSE, cleanup all
 759          * maybe-pending messages on all CPUs to be able to receive new
 760          * messages after we reconnect.
 761          */
 762         for_each_online_cpu(cpu) {
 763                 struct hv_per_cpu_context *hv_cpu
 764                         = per_cpu_ptr(hv_context.cpu_context, cpu);
 765
 766                 page_addr = hv_cpu->synic_message_page;
 767                 msg = (struct hv_message *)page_addr + VMBUS_MESSAGE_SINT;
 768                 msg->header.message_type = HVMSG_NONE;
 769         }
 770 }
 771
 772 /*
 773  * vmbus_unload_response - Handler for the unload response.
 774  */
 775 static void vmbus_unload_response(struct vmbus_channel_message_header *hdr)
 776 {
 777         /*
 778          * This is a global event; just wakeup the waiting thread.
 779          * Once we successfully unload, we can cleanup the monitor state.
 780          */
 781         complete(&vmbus_connection.unload_event);
 782 }
 783
 784 void vmbus_initiate_unload(bool crash)
 785 {
 786         struct vmbus_channel_message_header hdr;
 787
 788         /* Pre-Win2012R2 hosts don't support reconnect */
 789         if (vmbus_proto_version < VERSION_WIN8_1)
 790                 return;
 791
 792         init_completion(&vmbus_connection.unload_event);
 793         memset(&hdr, 0, sizeof(struct vmbus_channel_message_header));
 794         hdr.msgtype = CHANNELMSG_UNLOAD;
 795         vmbus_post_msg(&hdr, sizeof(struct vmbus_channel_message_header),
 796                        !crash);
 797
 798         /*
 799          * vmbus_initiate_unload() is also called on crash and the crash can be
 800          * happening in an interrupt context, where scheduling is impossible.
 801          */
 802         if (!crash)
 803                 wait_for_completion(&vmbus_connection.unload_event);
 804         else
 805                 vmbus_wait_for_unload();
 806 }
 807
 808 /*
 809  * vmbus_onoffer - Handler for channel offers from vmbus in parent partition.
 810  *
 811  */
 812 static void vmbus_onoffer(struct vmbus_channel_message_header *hdr)
 813 {
 814         struct vmbus_channel_offer_channel *offer;
 815         struct vmbus_channel *newchannel;
 816
 817         offer = (struct vmbus_channel_offer_channel *)hdr;
 818
 819         trace_vmbus_onoffer(offer);
 820
 821         /* Allocate the channel object and save this offer. */
 822         newchannel = alloc_channel();
 823         if (!newchannel) {
 824                 vmbus_release_relid(offer->child_relid);
 825                 atomic_dec(&vmbus_connection.offer_in_progress);
 826                 pr_err("Unable to allocate channel object\n");
 827                 return;
 828         }
 829
 830         /*
 831          * Setup state for signalling the host.
 832          */
 833         newchannel->sig_event = VMBUS_EVENT_CONNECTION_ID;
 834
 835         if (vmbus_proto_version != VERSION_WS2008) {
 836                 newchannel->is_dedicated_interrupt =
 837                                 (offer->is_dedicated_interrupt != 0);
 838                 newchannel->sig_event = offer->connection_id;
 839         }
 840
 841         memcpy(&newchannel->offermsg, offer,
 842                sizeof(struct vmbus_channel_offer_channel));
 843         newchannel->monitor_grp = (u8)offer->monitorid / 32;
 844         newchannel->monitor_bit = (u8)offer->monitorid % 32;
 845
 846         vmbus_process_offer(newchannel);
 847 }
 848
 849 /*
 850  * vmbus_onoffer_rescind - Rescind offer handler.
 851  *
 852  * We queue a work item to process this offer synchronously
 853  */
 854 static void vmbus_onoffer_rescind(struct vmbus_channel_message_header *hdr)
 855 {
 856         struct vmbus_channel_rescind_offer *rescind;
 857         struct vmbus_channel *channel;
 858         struct device *dev;
 859
 860         rescind = (struct vmbus_channel_rescind_offer *)hdr;
 861
 862         trace_vmbus_onoffer_rescind(rescind);
 863
 864         /*
 865          * The offer msg and the corresponding rescind msg
 866          * from the host are guranteed to be ordered -
 867          * offer comes in first and then the rescind.
 868          * Since we process these events in work elements,
 869          * and with preemption, we may end up processing
 870          * the events out of order. Given that we handle these
 871          * work elements on the same CPU, this is possible only
 872          * in the case of preemption. In any case wait here
 873          * until the offer processing has moved beyond the
 874          * point where the channel is discoverable.
 875          */
 876
 877         while (atomic_read(&vmbus_connection.offer_in_progress) != 0) {
 878                 /*
 879                  * We wait here until any channel offer is currently
 880                  * being processed.
 881                  */
 882                 msleep(1);
 883         }
 884
 885         mutex_lock(&vmbus_connection.channel_mutex);
 886         channel = relid2channel(rescind->child_relid);
 887         mutex_unlock(&vmbus_connection.channel_mutex);
 888
 889         if (channel == NULL) {
 890                 /*
 891                  * We failed in processing the offer message;
 892                  * we would have cleaned up the relid in that
 893                  * failure path.
 894                  */
 895                 return;
 896         }
 897
 898         /*
 899          * Now wait for offer handling to complete.
 900          */
 901         while (READ_ONCE(channel->probe_done) == false) {
 902                 /*
 903                  * We wait here until any channel offer is currently
 904                  * being processed.
 905                  */
 906                 msleep(1);
 907         }
 908
 909         /*
 910          * At this point, the rescind handling can proceed safely.
 911          */
 912
 913         if (channel->device_obj) {
 914                 if (channel->chn_rescind_callback) {
 915                         channel->chn_rescind_callback(channel);
 916                         vmbus_rescind_cleanup(channel);
 917                         return;
 918                 }
 919                 /*
 920                  * We will have to unregister this device from the
 921                  * driver core.
 922                  */
 923                 dev = get_device(&channel->device_obj->device);
 924                 if (dev) {
 925                         vmbus_rescind_cleanup(channel);
 926                         vmbus_device_unregister(channel->device_obj);
 927                         put_device(dev);
 928                 }
 929         }
 930         if (channel->primary_channel != NULL) {
 931                 /*
 932                  * Sub-channel is being rescinded. Following is the channel
 933                  * close sequence when initiated from the driveri (refer to
 934                  * vmbus_close() for details):
 935                  * 1. Close all sub-channels first
 936                  * 2. Then close the primary channel.
 937                  */
 938                 mutex_lock(&vmbus_connection.channel_mutex);
 939                 vmbus_rescind_cleanup(channel);
 940                 if (channel->state == CHANNEL_OPEN_STATE) {
 941                         /*
 942                          * The channel is currently not open;
 943                          * it is safe for us to cleanup the channel.
 944                          */
 945                         hv_process_channel_removal(rescind->child_relid);
 946                 }
 947                 mutex_unlock(&vmbus_connection.channel_mutex);
 948         }
 949 }
 950
 951 void vmbus_hvsock_device_unregister(struct vmbus_channel *channel)
 952 {
 953         BUG_ON(!is_hvsock_channel(channel));
 954
 955         /* We always get a rescind msg when a connection is closed. */
 956         while (!READ_ONCE(channel->probe_done) || !READ_ONCE(channel->rescind))
 957                 msleep(1);
 958
 959         vmbus_device_unregister(channel->device_obj);
 960 }
 961 EXPORT_SYMBOL_GPL(vmbus_hvsock_device_unregister);
 962
 963
 964 /*
 965  * vmbus_onoffers_delivered -
 966  * This is invoked when all offers have been delivered.
 967  *
 968  * Nothing to do here.
 969  */
 970 static void vmbus_onoffers_delivered(
 971                         struct vmbus_channel_message_header *hdr)
 972 {
 973 }
 974
 975 /*
 976  * vmbus_onopen_result - Open result handler.
 977  *
 978  * This is invoked when we received a response to our channel open request.
 979  * Find the matching request, copy the response and signal the requesting
 980  * thread.
 981  */
 982 static void vmbus_onopen_result(struct vmbus_channel_message_header *hdr)
 983 {
 984         struct vmbus_channel_open_result *result;
 985         struct vmbus_channel_msginfo *msginfo;
 986         struct vmbus_channel_message_header *requestheader;
 987         struct vmbus_channel_open_channel *openmsg;
 988         unsigned long flags;
 989
 990         result = (struct vmbus_channel_open_result *)hdr;
 991
 992         trace_vmbus_onopen_result(result);
 993
 994         /*
 995          * Find the open msg, copy the result and signal/unblock the wait event
 996          */
 997         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
 998
 999         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1000                                 msglistentry) {
1001                 requestheader =
1002                         (struct vmbus_channel_message_header *)msginfo->msg;
1003
1004                 if (requestheader->msgtype == CHANNELMSG_OPENCHANNEL) {
1005                         openmsg =
1006                         (struct vmbus_channel_open_channel *)msginfo->msg;
1007                         if (openmsg->child_relid == result->child_relid &&
1008                             openmsg->openid == result->openid) {
1009                                 memcpy(&msginfo->response.open_result,
1010                                        result,
1011                                        sizeof(
1012                                         struct vmbus_channel_open_result));
1013                                 complete(&msginfo->waitevent);
1014                                 break;
1015                         }
1016                 }
1017         }
1018         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1019 }
1020
1021 /*
1022  * vmbus_ongpadl_created - GPADL created handler.
1023  *
1024  * This is invoked when we received a response to our gpadl create request.
1025  * Find the matching request, copy the response and signal the requesting
1026  * thread.
1027  */
1028 static void vmbus_ongpadl_created(struct vmbus_channel_message_header *hdr)
1029 {
1030         struct vmbus_channel_gpadl_created *gpadlcreated;
1031         struct vmbus_channel_msginfo *msginfo;
1032         struct vmbus_channel_message_header *requestheader;
1033         struct vmbus_channel_gpadl_header *gpadlheader;
1034         unsigned long flags;
1035
1036         gpadlcreated = (struct vmbus_channel_gpadl_created *)hdr;
1037
1038         trace_vmbus_ongpadl_created(gpadlcreated);
1039
1040         /*
1041          * Find the establish msg, copy the result and signal/unblock the wait
1042          * event
1043          */
1044         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1045
1046         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1047                                 msglistentry) {
1048                 requestheader =
1049                         (struct vmbus_channel_message_header *)msginfo->msg;
1050
1051                 if (requestheader->msgtype == CHANNELMSG_GPADL_HEADER) {
1052                         gpadlheader =
1053                         (struct vmbus_channel_gpadl_header *)requestheader;
1054
1055                         if ((gpadlcreated->child_relid ==
1056                              gpadlheader->child_relid) &&
1057                             (gpadlcreated->gpadl == gpadlheader->gpadl)) {
1058                                 memcpy(&msginfo->response.gpadl_created,
1059                                        gpadlcreated,
1060                                        sizeof(
1061                                         struct vmbus_channel_gpadl_created));
1062                                 complete(&msginfo->waitevent);
1063                                 break;
1064                         }
1065                 }
1066         }
1067         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1068 }
1069
1070 /*
1071  * vmbus_ongpadl_torndown - GPADL torndown handler.
1072  *
1073  * This is invoked when we received a response to our gpadl teardown request.
1074  * Find the matching request, copy the response and signal the requesting
1075  * thread.
1076  */
1077 static void vmbus_ongpadl_torndown(
1078                         struct vmbus_channel_message_header *hdr)
1079 {
1080         struct vmbus_channel_gpadl_torndown *gpadl_torndown;
1081         struct vmbus_channel_msginfo *msginfo;
1082         struct vmbus_channel_message_header *requestheader;
1083         struct vmbus_channel_gpadl_teardown *gpadl_teardown;
1084         unsigned long flags;
1085
1086         gpadl_torndown = (struct vmbus_channel_gpadl_torndown *)hdr;
1087
1088         trace_vmbus_ongpadl_torndown(gpadl_torndown);
1089
1090         /*
1091          * Find the open msg, copy the result and signal/unblock the wait event
1092          */
1093         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1094
1095         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1096                                 msglistentry) {
1097                 requestheader =
1098                         (struct vmbus_channel_message_header *)msginfo->msg;
1099
1100                 if (requestheader->msgtype == CHANNELMSG_GPADL_TEARDOWN) {
1101                         gpadl_teardown =
1102                         (struct vmbus_channel_gpadl_teardown *)requestheader;
1103
1104                         if (gpadl_torndown->gpadl == gpadl_teardown->gpadl) {
1105                                 memcpy(&msginfo->response.gpadl_torndown,
1106                                        gpadl_torndown,
1107                                        sizeof(
1108                                         struct vmbus_channel_gpadl_torndown));
1109                                 complete(&msginfo->waitevent);
1110                                 break;
1111                         }
1112                 }
1113         }
1114         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1115 }
1116
1117 /*
1118  * vmbus_onversion_response - Version response handler
1119  *
1120  * This is invoked when we received a response to our initiate contact request.
1121  * Find the matching request, copy the response and signal the requesting
1122  * thread.
1123  */
1124 static void vmbus_onversion_response(
1125                 struct vmbus_channel_message_header *hdr)
1126 {
1127         struct vmbus_channel_msginfo *msginfo;
1128         struct vmbus_channel_message_header *requestheader;
1129         struct vmbus_channel_version_response *version_response;
1130         unsigned long flags;
1131
1132         version_response = (struct vmbus_channel_version_response *)hdr;
1133
1134         trace_vmbus_onversion_response(version_response);
1135
1136         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1137
1138         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1139                                 msglistentry) {
1140                 requestheader =
1141                         (struct vmbus_channel_message_header *)msginfo->msg;
1142
1143                 if (requestheader->msgtype ==
1144                     CHANNELMSG_INITIATE_CONTACT) {
1145                         memcpy(&msginfo->response.version_response,
1146                               version_response,
1147                               sizeof(struct vmbus_channel_version_response));
1148                         complete(&msginfo->waitevent);
1149                 }
1150         }
1151         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1152 }
1153
1154 /* Channel message dispatch table */
1155 const struct vmbus_channel_message_table_entry
1156 channel_message_table[CHANNELMSG_COUNT] = {
1157         { CHANNELMSG_INVALID,                   0, NULL },
1158         { CHANNELMSG_OFFERCHANNEL,              0, vmbus_onoffer },
1159         { CHANNELMSG_RESCIND_CHANNELOFFER,      0, vmbus_onoffer_rescind },
1160         { CHANNELMSG_REQUESTOFFERS,             0, NULL },
1161         { CHANNELMSG_ALLOFFERS_DELIVERED,       1, vmbus_onoffers_delivered },
1162         { CHANNELMSG_OPENCHANNEL,               0, NULL },
1163         { CHANNELMSG_OPENCHANNEL_RESULT,        1, vmbus_onopen_result },
1164         { CHANNELMSG_CLOSECHANNEL,              0, NULL },
1165         { CHANNELMSG_GPADL_HEADER,              0, NULL },
1166         { CHANNELMSG_GPADL_BODY,                0, NULL },
1167         { CHANNELMSG_GPADL_CREATED,             1, vmbus_ongpadl_created },
1168         { CHANNELMSG_GPADL_TEARDOWN,            0, NULL },
1169         { CHANNELMSG_GPADL_TORNDOWN,            1, vmbus_ongpadl_torndown },
1170         { CHANNELMSG_RELID_RELEASED,            0, NULL },
1171         { CHANNELMSG_INITIATE_CONTACT,          0, NULL },
1172         { CHANNELMSG_VERSION_RESPONSE,          1, vmbus_onversion_response },
1173         { CHANNELMSG_UNLOAD,                    0, NULL },
1174         { CHANNELMSG_UNLOAD_RESPONSE,           1, vmbus_unload_response },
1175         { CHANNELMSG_18,                        0, NULL },
1176         { CHANNELMSG_19,                        0, NULL },
1177         { CHANNELMSG_20,                        0, NULL },
1178         { CHANNELMSG_TL_CONNECT_REQUEST,        0, NULL },
1179 };
1180
1181 /*
1182  * vmbus_onmessage - Handler for channel protocol messages.
1183  *
1184  * This is invoked in the vmbus worker thread context.
1185  */
1186 void vmbus_onmessage(void *context)
1187 {
1188         struct hv_message *msg = context;
1189         struct vmbus_channel_message_header *hdr;
1190         int size;
1191
1192         hdr = (struct vmbus_channel_message_header *)msg->u.payload;
1193         size = msg->header.payload_size;
1194
1195         trace_vmbus_on_message(hdr);
1196
1197         if (hdr->msgtype >= CHANNELMSG_COUNT) {
1198                 pr_err("Received invalid channel message type %d size %d\n",
1199                            hdr->msgtype, size);
1200                 print_hex_dump_bytes("", DUMP_PREFIX_NONE,
1201                                      (unsigned char *)msg->u.payload, size);
1202                 return;
1203         }
1204
1205         if (channel_message_table[hdr->msgtype].message_handler)
1206                 channel_message_table[hdr->msgtype].message_handler(hdr);
1207         else
1208                 pr_err("Unhandled channel message type %d\n", hdr->msgtype);
1209 }
1210
1211 /*
1212  * vmbus_request_offers - Send a request to get all our pending offers.
1213  */
1214 int vmbus_request_offers(void)
1215 {
1216         struct vmbus_channel_message_header *msg;
1217         struct vmbus_channel_msginfo *msginfo;
1218         int ret;
1219
1220         msginfo = kmalloc(sizeof(*msginfo) +
1221                           sizeof(struct vmbus_channel_message_header),
1222                           GFP_KERNEL);
1223         if (!msginfo)
1224                 return -ENOMEM;
1225
1226         msg = (struct vmbus_channel_message_header *)msginfo->msg;
1227
1228         msg->msgtype = CHANNELMSG_REQUESTOFFERS;
1229
1230         ret = vmbus_post_msg(msg, sizeof(struct vmbus_channel_message_header),
1231                              true);
1232
1233         trace_vmbus_request_offers(ret);
1234
1235         if (ret != 0) {
1236                 pr_err("Unable to request offers - %d\n", ret);
1237
1238                 goto cleanup;
1239         }
1240
1241 cleanup:
1242         kfree(msginfo);
1243
1244         return ret;
1245 }
1246
1247 /*
1248  * Retrieve the (sub) channel on which to send an outgoing request.
1249  * When a primary channel has multiple sub-channels, we try to
1250  * distribute the load equally amongst all available channels.
1251  */
1252 struct vmbus_channel *vmbus_get_outgoing_channel(struct vmbus_channel *primary)
1253 {
1254         struct list_head *cur, *tmp;
1255         int cur_cpu;
1256         struct vmbus_channel *cur_channel;
1257         struct vmbus_channel *outgoing_channel = primary;
1258         int next_channel;
1259         int i = 1;
1260
1261         if (list_empty(&primary->sc_list))
1262                 return outgoing_channel;
1263
1264         next_channel = primary->next_oc++;
1265
1266         if (next_channel > (primary->num_sc)) {
1267                 primary->next_oc = 0;
1268                 return outgoing_channel;
1269         }
1270
1271         cur_cpu = hv_cpu_number_to_vp_number(smp_processor_id());
1272         list_for_each_safe(cur, tmp, &primary->sc_list) {
1273                 cur_channel = list_entry(cur, struct vmbus_channel, sc_list);
1274                 if (cur_channel->state != CHANNEL_OPENED_STATE)
1275                         continue;
1276
1277                 if (cur_channel->target_vp == cur_cpu)
1278                         return cur_channel;
1279
1280                 if (i == next_channel)
1281                         return cur_channel;
1282
1283                 i++;
1284         }
1285
1286         return outgoing_channel;
1287 }
1288 EXPORT_SYMBOL_GPL(vmbus_get_outgoing_channel);
1289
1290 static void invoke_sc_cb(struct vmbus_channel *primary_channel)
1291 {
1292         struct list_head *cur, *tmp;
1293         struct vmbus_channel *cur_channel;
1294
1295         if (primary_channel->sc_creation_callback == NULL)
1296                 return;
1297
1298         list_for_each_safe(cur, tmp, &primary_channel->sc_list) {
1299                 cur_channel = list_entry(cur, struct vmbus_channel, sc_list);
1300
1301                 primary_channel->sc_creation_callback(cur_channel);
1302         }
1303 }
1304
1305 void vmbus_set_sc_create_callback(struct vmbus_channel *primary_channel,
1306                                 void (*sc_cr_cb)(struct vmbus_channel *new_sc))
1307 {
1308         primary_channel->sc_creation_callback = sc_cr_cb;
1309 }
1310 EXPORT_SYMBOL_GPL(vmbus_set_sc_create_callback);
1311
1312 bool vmbus_are_subchannels_present(struct vmbus_channel *primary)
1313 {
1314         bool ret;
1315
1316         ret = !list_empty(&primary->sc_list);
1317
1318         if (ret) {
1319                 /*
1320                  * Invoke the callback on sub-channel creation.
1321                  * This will present a uniform interface to the
1322                  * clients.
1323                  */
1324                 invoke_sc_cb(primary);
1325         }
1326
1327         return ret;
1328 }
1329 EXPORT_SYMBOL_GPL(vmbus_are_subchannels_present);
1330
1331 void vmbus_set_chn_rescind_callback(struct vmbus_channel *channel,
1332                 void (*chn_rescind_cb)(struct vmbus_channel *))
1333 {
1334         channel->chn_rescind_callback = chn_rescind_cb;
1335 }
1336 EXPORT_SYMBOL_GPL(vmbus_set_chn_rescind_callback);