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