drivers/hv/channel_mgmt.c

   1 /*
   2  * Copyright (c) 2009, Microsoft Corporation.
   3  *
   4  * This program is free software; you can redistribute it and/or modify it
   5  * under the terms and conditions of the GNU General Public License,
   6  * version 2, as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope it will be useful, but WITHOUT
   9  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  10  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  11  * more details.
  12  *
  13  * You should have received a copy of the GNU General Public License along with
  14  * this program; if not, write to the Free Software Foundation, Inc., 59 Temple
  15  * Place - Suite 330, Boston, MA 02111-1307 USA.
  16  *
  17  * Authors:
  18  *   Haiyang Zhang <haiyangz@microsoft.com>
  19  *   Hank Janssen  <hjanssen@microsoft.com>
  20  */
  21 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  22
  23 #include <linux/kernel.h>
  24 #include <linux/interrupt.h>
  25 #include <linux/sched.h>
  26 #include <linux/wait.h>
  27 #include <linux/mm.h>
  28 #include <linux/slab.h>
  29 #include <linux/list.h>
  30 #include <linux/module.h>
  31 #include <linux/completion.h>
  32 #include <linux/delay.h>
  33 #include <linux/hyperv.h>
  34 #include <asm/mshyperv.h>
  35
  36 #include "hyperv_vmbus.h"
  37
  38 static void init_vp_index(struct vmbus_channel *channel, u16 dev_type);
  39
  40 static const struct vmbus_device vmbus_devs[] = {
  41         /* IDE */
  42         { .dev_type = HV_IDE,
  43           HV_IDE_GUID,
  44           .perf_device = true,
  45         },
  46
  47         /* SCSI */
  48         { .dev_type = HV_SCSI,
  49           HV_SCSI_GUID,
  50           .perf_device = true,
  51         },
  52
  53         /* Fibre Channel */
  54         { .dev_type = HV_FC,
  55           HV_SYNTHFC_GUID,
  56           .perf_device = true,
  57         },
  58
  59         /* Synthetic NIC */
  60         { .dev_type = HV_NIC,
  61           HV_NIC_GUID,
  62           .perf_device = true,
  63         },
  64
  65         /* Network Direct */
  66         { .dev_type = HV_ND,
  67           HV_ND_GUID,
  68           .perf_device = true,
  69         },
  70
  71         /* PCIE */
  72         { .dev_type = HV_PCIE,
  73           HV_PCIE_GUID,
  74           .perf_device = true,
  75         },
  76
  77         /* Synthetic Frame Buffer */
  78         { .dev_type = HV_FB,
  79           HV_SYNTHVID_GUID,
  80           .perf_device = false,
  81         },
  82
  83         /* Synthetic Keyboard */
  84         { .dev_type = HV_KBD,
  85           HV_KBD_GUID,
  86           .perf_device = false,
  87         },
  88
  89         /* Synthetic MOUSE */
  90         { .dev_type = HV_MOUSE,
  91           HV_MOUSE_GUID,
  92           .perf_device = false,
  93         },
  94
  95         /* KVP */
  96         { .dev_type = HV_KVP,
  97           HV_KVP_GUID,
  98           .perf_device = false,
  99         },
 100
 101         /* Time Synch */
 102         { .dev_type = HV_TS,
 103           HV_TS_GUID,
 104           .perf_device = false,
 105         },
 106
 107         /* Heartbeat */
 108         { .dev_type = HV_HB,
 109           HV_HEART_BEAT_GUID,
 110           .perf_device = false,
 111         },
 112
 113         /* Shutdown */
 114         { .dev_type = HV_SHUTDOWN,
 115           HV_SHUTDOWN_GUID,
 116           .perf_device = false,
 117         },
 118
 119         /* File copy */
 120         { .dev_type = HV_FCOPY,
 121           HV_FCOPY_GUID,
 122           .perf_device = false,
 123         },
 124
 125         /* Backup */
 126         { .dev_type = HV_BACKUP,
 127           HV_VSS_GUID,
 128           .perf_device = false,
 129         },
 130
 131         /* Dynamic Memory */
 132         { .dev_type = HV_DM,
 133           HV_DM_GUID,
 134           .perf_device = false,
 135         },
 136
 137         /* Unknown GUID */
 138         { .dev_type = HV_UNKNOWN,
 139           .perf_device = false,
 140         },
 141 };
 142
 143 static const struct {
 144         uuid_le guid;
 145 } vmbus_unsupported_devs[] = {
 146         { HV_AVMA1_GUID },
 147         { HV_AVMA2_GUID },
 148         { HV_RDV_GUID   },
 149 };
 150
 151 /*
 152  * The rescinded channel may be blocked waiting for a response from the host;
 153  * take care of that.
 154  */
 155 static void vmbus_rescind_cleanup(struct vmbus_channel *channel)
 156 {
 157         struct vmbus_channel_msginfo *msginfo;
 158         unsigned long flags;
 159
 160
 161         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
 162
 163         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
 164                                 msglistentry) {
 165
 166                 if (msginfo->waiting_channel == channel) {
 167                         complete(&msginfo->waitevent);
 168                         break;
 169                 }
 170         }
 171         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
 172 }
 173
 174 static bool is_unsupported_vmbus_devs(const uuid_le *guid)
 175 {
 176         int i;
 177
 178         for (i = 0; i < ARRAY_SIZE(vmbus_unsupported_devs); i++)
 179                 if (!uuid_le_cmp(*guid, vmbus_unsupported_devs[i].guid))
 180                         return true;
 181         return false;
 182 }
 183
 184 static u16 hv_get_dev_type(const struct vmbus_channel *channel)
 185 {
 186         const uuid_le *guid = &channel->offermsg.offer.if_type;
 187         u16 i;
 188
 189         if (is_hvsock_channel(channel) || is_unsupported_vmbus_devs(guid))
 190                 return HV_UNKNOWN;
 191
 192         for (i = HV_IDE; i < HV_UNKNOWN; i++) {
 193                 if (!uuid_le_cmp(*guid, vmbus_devs[i].guid))
 194                         return i;
 195         }
 196         pr_info("Unknown GUID: %pUl\n", guid);
 197         return i;
 198 }
 199
 200 /**
 201  * vmbus_prep_negotiate_resp() - Create default response for Hyper-V Negotiate message
 202  * @icmsghdrp: Pointer to msg header structure
 203  * @icmsg_negotiate: Pointer to negotiate message structure
 204  * @buf: Raw buffer channel data
 205  *
 206  * @icmsghdrp is of type &struct icmsg_hdr.
 207  * Set up and fill in default negotiate response message.
 208  *
 209  * The fw_version and fw_vercnt specifies the framework version that
 210  * we can support.
 211  *
 212  * The srv_version and srv_vercnt specifies the service
 213  * versions we can support.
 214  *
 215  * Versions are given in decreasing order.
 216  *
 217  * nego_fw_version and nego_srv_version store the selected protocol versions.
 218  *
 219  * Mainly used by Hyper-V drivers.
 220  */
 221 bool vmbus_prep_negotiate_resp(struct icmsg_hdr *icmsghdrp,
 222                                 u8 *buf, const int *fw_version, int fw_vercnt,
 223                                 const int *srv_version, int srv_vercnt,
 224                                 int *nego_fw_version, int *nego_srv_version)
 225 {
 226         int icframe_major, icframe_minor;
 227         int icmsg_major, icmsg_minor;
 228         int fw_major, fw_minor;
 229         int srv_major, srv_minor;
 230         int i, j;
 231         bool found_match = false;
 232         struct icmsg_negotiate *negop;
 233
 234         icmsghdrp->icmsgsize = 0x10;
 235         negop = (struct icmsg_negotiate *)&buf[
 236                 sizeof(struct vmbuspipe_hdr) +
 237                 sizeof(struct icmsg_hdr)];
 238
 239         icframe_major = negop->icframe_vercnt;
 240         icframe_minor = 0;
 241
 242         icmsg_major = negop->icmsg_vercnt;
 243         icmsg_minor = 0;
 244
 245         /*
 246          * Select the framework version number we will
 247          * support.
 248          */
 249
 250         for (i = 0; i < fw_vercnt; i++) {
 251                 fw_major = (fw_version[i] >> 16);
 252                 fw_minor = (fw_version[i] & 0xFFFF);
 253
 254                 for (j = 0; j < negop->icframe_vercnt; j++) {
 255                         if ((negop->icversion_data[j].major == fw_major) &&
 256                             (negop->icversion_data[j].minor == fw_minor)) {
 257                                 icframe_major = negop->icversion_data[j].major;
 258                                 icframe_minor = negop->icversion_data[j].minor;
 259                                 found_match = true;
 260                                 break;
 261                         }
 262                 }
 263
 264                 if (found_match)
 265                         break;
 266         }
 267
 268         if (!found_match)
 269                 goto fw_error;
 270
 271         found_match = false;
 272
 273         for (i = 0; i < srv_vercnt; i++) {
 274                 srv_major = (srv_version[i] >> 16);
 275                 srv_minor = (srv_version[i] & 0xFFFF);
 276
 277                 for (j = negop->icframe_vercnt;
 278                         (j < negop->icframe_vercnt + negop->icmsg_vercnt);
 279                         j++) {
 280
 281                         if ((negop->icversion_data[j].major == srv_major) &&
 282                                 (negop->icversion_data[j].minor == srv_minor)) {
 283
 284                                 icmsg_major = negop->icversion_data[j].major;
 285                                 icmsg_minor = negop->icversion_data[j].minor;
 286                                 found_match = true;
 287                                 break;
 288                         }
 289                 }
 290
 291                 if (found_match)
 292                         break;
 293         }
 294
 295         /*
 296          * Respond with the framework and service
 297          * version numbers we can support.
 298          */
 299
 300 fw_error:
 301         if (!found_match) {
 302                 negop->icframe_vercnt = 0;
 303                 negop->icmsg_vercnt = 0;
 304         } else {
 305                 negop->icframe_vercnt = 1;
 306                 negop->icmsg_vercnt = 1;
 307         }
 308
 309         if (nego_fw_version)
 310                 *nego_fw_version = (icframe_major << 16) | icframe_minor;
 311
 312         if (nego_srv_version)
 313                 *nego_srv_version = (icmsg_major << 16) | icmsg_minor;
 314
 315         negop->icversion_data[0].major = icframe_major;
 316         negop->icversion_data[0].minor = icframe_minor;
 317         negop->icversion_data[1].major = icmsg_major;
 318         negop->icversion_data[1].minor = icmsg_minor;
 319         return found_match;
 320 }
 321
 322 EXPORT_SYMBOL_GPL(vmbus_prep_negotiate_resp);
 323
 324 /*
 325  * alloc_channel - Allocate and initialize a vmbus channel object
 326  */
 327 static struct vmbus_channel *alloc_channel(void)
 328 {
 329         struct vmbus_channel *channel;
 330
 331         channel = kzalloc(sizeof(*channel), GFP_ATOMIC);
 332         if (!channel)
 333                 return NULL;
 334
 335         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(struct vmbus_channel *channel, u32 relid)
 385 {
 386         unsigned long flags;
 387         struct vmbus_channel *primary_channel;
 388
 389         BUG_ON(!channel->rescind);
 390         BUG_ON(!mutex_is_locked(&vmbus_connection.channel_mutex));
 391
 392         if (channel->target_cpu != get_cpu()) {
 393                 put_cpu();
 394                 smp_call_function_single(channel->target_cpu,
 395                                          percpu_channel_deq, channel, true);
 396         } else {
 397                 percpu_channel_deq(channel);
 398                 put_cpu();
 399         }
 400
 401         if (channel->primary_channel == NULL) {
 402                 list_del(&channel->listentry);
 403
 404                 primary_channel = channel;
 405         } else {
 406                 primary_channel = channel->primary_channel;
 407                 spin_lock_irqsave(&primary_channel->lock, flags);
 408                 list_del(&channel->sc_list);
 409                 primary_channel->num_sc--;
 410                 spin_unlock_irqrestore(&primary_channel->lock, flags);
 411         }
 412
 413         /*
 414          * We need to free the bit for init_vp_index() to work in the case
 415          * of sub-channel, when we reload drivers like hv_netvsc.
 416          */
 417         if (channel->affinity_policy == HV_LOCALIZED)
 418                 cpumask_clear_cpu(channel->target_cpu,
 419                                   &primary_channel->alloced_cpus_in_node);
 420
 421         vmbus_release_relid(relid);
 422
 423         free_channel(channel);
 424 }
 425
 426 void vmbus_free_channels(void)
 427 {
 428         struct vmbus_channel *channel, *tmp;
 429
 430         list_for_each_entry_safe(channel, tmp, &vmbus_connection.chn_list,
 431                 listentry) {
 432                 /* hv_process_channel_removal() needs this */
 433                 channel->rescind = true;
 434
 435                 vmbus_device_unregister(channel->device_obj);
 436         }
 437 }
 438
 439 /*
 440  * vmbus_process_offer - Process the offer by creating a channel/device
 441  * associated with this offer
 442  */
 443 static void vmbus_process_offer(struct vmbus_channel *newchannel)
 444 {
 445         struct vmbus_channel *channel;
 446         bool fnew = true;
 447         unsigned long flags;
 448         u16 dev_type;
 449         int ret;
 450
 451         /* Make sure this is a new offer */
 452         mutex_lock(&vmbus_connection.channel_mutex);
 453
 454         /*
 455          * Now that we have acquired the channel_mutex,
 456          * we can release the potentially racing rescind thread.
 457          */
 458         atomic_dec(&vmbus_connection.offer_in_progress);
 459
 460         list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
 461                 if (!uuid_le_cmp(channel->offermsg.offer.if_type,
 462                         newchannel->offermsg.offer.if_type) &&
 463                         !uuid_le_cmp(channel->offermsg.offer.if_instance,
 464                                 newchannel->offermsg.offer.if_instance)) {
 465                         fnew = false;
 466                         break;
 467                 }
 468         }
 469
 470         if (fnew)
 471                 list_add_tail(&newchannel->listentry,
 472                               &vmbus_connection.chn_list);
 473
 474         mutex_unlock(&vmbus_connection.channel_mutex);
 475
 476         if (!fnew) {
 477                 /*
 478                  * Check to see if this is a sub-channel.
 479                  */
 480                 if (newchannel->offermsg.offer.sub_channel_index != 0) {
 481                         /*
 482                          * Process the sub-channel.
 483                          */
 484                         newchannel->primary_channel = channel;
 485                         spin_lock_irqsave(&channel->lock, flags);
 486                         list_add_tail(&newchannel->sc_list, &channel->sc_list);
 487                         channel->num_sc++;
 488                         spin_unlock_irqrestore(&channel->lock, flags);
 489                 } else {
 490                         goto err_free_chan;
 491                 }
 492         }
 493
 494         dev_type = hv_get_dev_type(newchannel);
 495
 496         init_vp_index(newchannel, dev_type);
 497
 498         if (newchannel->target_cpu != get_cpu()) {
 499                 put_cpu();
 500                 smp_call_function_single(newchannel->target_cpu,
 501                                          percpu_channel_enq,
 502                                          newchannel, true);
 503         } else {
 504                 percpu_channel_enq(newchannel);
 505                 put_cpu();
 506         }
 507
 508         /*
 509          * This state is used to indicate a successful open
 510          * so that when we do close the channel normally, we
 511          * can cleanup properly
 512          */
 513         newchannel->state = CHANNEL_OPEN_STATE;
 514
 515         if (!fnew) {
 516                 if (channel->sc_creation_callback != NULL)
 517                         channel->sc_creation_callback(newchannel);
 518                 return;
 519         }
 520
 521         /*
 522          * Start the process of binding this offer to the driver
 523          * We need to set the DeviceObject field before calling
 524          * vmbus_child_dev_add()
 525          */
 526         newchannel->device_obj = vmbus_device_create(
 527                 &newchannel->offermsg.offer.if_type,
 528                 &newchannel->offermsg.offer.if_instance,
 529                 newchannel);
 530         if (!newchannel->device_obj)
 531                 goto err_deq_chan;
 532
 533         newchannel->device_obj->device_id = dev_type;
 534         /*
 535          * Add the new device to the bus. This will kick off device-driver
 536          * binding which eventually invokes the device driver's AddDevice()
 537          * method.
 538          */
 539         ret = vmbus_device_register(newchannel->device_obj);
 540
 541         if (ret != 0) {
 542                 pr_err("unable to add child device object (relid %d)\n",
 543                         newchannel->offermsg.child_relid);
 544                 kfree(newchannel->device_obj);
 545                 goto err_deq_chan;
 546         }
 547
 548         newchannel->probe_done = true;
 549         return;
 550
 551 err_deq_chan:
 552         mutex_lock(&vmbus_connection.channel_mutex);
 553         list_del(&newchannel->listentry);
 554         mutex_unlock(&vmbus_connection.channel_mutex);
 555
 556         if (newchannel->target_cpu != get_cpu()) {
 557                 put_cpu();
 558                 smp_call_function_single(newchannel->target_cpu,
 559                                          percpu_channel_deq, newchannel, true);
 560         } else {
 561                 percpu_channel_deq(newchannel);
 562                 put_cpu();
 563         }
 564
 565         vmbus_release_relid(newchannel->offermsg.child_relid);
 566
 567 err_free_chan:
 568         free_channel(newchannel);
 569 }
 570
 571 /*
 572  * We use this state to statically distribute the channel interrupt load.
 573  */
 574 static int next_numa_node_id;
 575
 576 /*
 577  * Starting with Win8, we can statically distribute the incoming
 578  * channel interrupt load by binding a channel to VCPU.
 579  * We do this in a hierarchical fashion:
 580  * First distribute the primary channels across available NUMA nodes
 581  * and then distribute the subchannels amongst the CPUs in the NUMA
 582  * node assigned to the primary channel.
 583  *
 584  * For pre-win8 hosts or non-performance critical channels we assign the
 585  * first CPU in the first NUMA node.
 586  */
 587 static void init_vp_index(struct vmbus_channel *channel, u16 dev_type)
 588 {
 589         u32 cur_cpu;
 590         bool perf_chn = vmbus_devs[dev_type].perf_device;
 591         struct vmbus_channel *primary = channel->primary_channel;
 592         int next_node;
 593         struct cpumask available_mask;
 594         struct cpumask *alloced_mask;
 595
 596         if ((vmbus_proto_version == VERSION_WS2008) ||
 597             (vmbus_proto_version == VERSION_WIN7) || (!perf_chn)) {
 598                 /*
 599                  * Prior to win8, all channel interrupts are
 600                  * delivered on cpu 0.
 601                  * Also if the channel is not a performance critical
 602                  * channel, bind it to cpu 0.
 603                  */
 604                 channel->numa_node = 0;
 605                 channel->target_cpu = 0;
 606                 channel->target_vp = hv_cpu_number_to_vp_number(0);
 607                 return;
 608         }
 609
 610         /*
 611          * Based on the channel affinity policy, we will assign the NUMA
 612          * nodes.
 613          */
 614
 615         if ((channel->affinity_policy == HV_BALANCED) || (!primary)) {
 616                 while (true) {
 617                         next_node = next_numa_node_id++;
 618                         if (next_node == nr_node_ids) {
 619                                 next_node = next_numa_node_id = 0;
 620                                 continue;
 621                         }
 622                         if (cpumask_empty(cpumask_of_node(next_node)))
 623                                 continue;
 624                         break;
 625                 }
 626                 channel->numa_node = next_node;
 627                 primary = channel;
 628         }
 629         alloced_mask = &hv_context.hv_numa_map[primary->numa_node];
 630
 631         if (cpumask_weight(alloced_mask) ==
 632             cpumask_weight(cpumask_of_node(primary->numa_node))) {
 633                 /*
 634                  * We have cycled through all the CPUs in the node;
 635                  * reset the alloced map.
 636                  */
 637                 cpumask_clear(alloced_mask);
 638         }
 639
 640         cpumask_xor(&available_mask, alloced_mask,
 641                     cpumask_of_node(primary->numa_node));
 642
 643         cur_cpu = -1;
 644
 645         if (primary->affinity_policy == HV_LOCALIZED) {
 646                 /*
 647                  * Normally Hyper-V host doesn't create more subchannels
 648                  * than there are VCPUs on the node but it is possible when not
 649                  * all present VCPUs on the node are initialized by guest.
 650                  * Clear the alloced_cpus_in_node to start over.
 651                  */
 652                 if (cpumask_equal(&primary->alloced_cpus_in_node,
 653                                   cpumask_of_node(primary->numa_node)))
 654                         cpumask_clear(&primary->alloced_cpus_in_node);
 655         }
 656
 657         while (true) {
 658                 cur_cpu = cpumask_next(cur_cpu, &available_mask);
 659                 if (cur_cpu >= nr_cpu_ids) {
 660                         cur_cpu = -1;
 661                         cpumask_copy(&available_mask,
 662                                      cpumask_of_node(primary->numa_node));
 663                         continue;
 664                 }
 665
 666                 if (primary->affinity_policy == HV_LOCALIZED) {
 667                         /*
 668                          * NOTE: in the case of sub-channel, we clear the
 669                          * sub-channel related bit(s) in
 670                          * primary->alloced_cpus_in_node in
 671                          * hv_process_channel_removal(), so when we
 672                          * reload drivers like hv_netvsc in SMP guest, here
 673                          * we're able to re-allocate
 674                          * bit from primary->alloced_cpus_in_node.
 675                          */
 676                         if (!cpumask_test_cpu(cur_cpu,
 677                                               &primary->alloced_cpus_in_node)) {
 678                                 cpumask_set_cpu(cur_cpu,
 679                                                 &primary->alloced_cpus_in_node);
 680                                 cpumask_set_cpu(cur_cpu, alloced_mask);
 681                                 break;
 682                         }
 683                 } else {
 684                         cpumask_set_cpu(cur_cpu, alloced_mask);
 685                         break;
 686                 }
 687         }
 688
 689         channel->target_cpu = cur_cpu;
 690         channel->target_vp = hv_cpu_number_to_vp_number(cur_cpu);
 691 }
 692
 693 static void vmbus_wait_for_unload(void)
 694 {
 695         int cpu;
 696         void *page_addr;
 697         struct hv_message *msg;
 698         struct vmbus_channel_message_header *hdr;
 699         u32 message_type;
 700
 701         /*
 702          * CHANNELMSG_UNLOAD_RESPONSE is always delivered to the CPU which was
 703          * used for initial contact or to CPU0 depending on host version. When
 704          * we're crashing on a different CPU let's hope that IRQ handler on
 705          * the cpu which receives CHANNELMSG_UNLOAD_RESPONSE is still
 706          * functional and vmbus_unload_response() will complete
 707          * vmbus_connection.unload_event. If not, the last thing we can do is
 708          * read message pages for all CPUs directly.
 709          */
 710         while (1) {
 711                 if (completion_done(&vmbus_connection.unload_event))
 712                         break;
 713
 714                 for_each_online_cpu(cpu) {
 715                         struct hv_per_cpu_context *hv_cpu
 716                                 = per_cpu_ptr(hv_context.cpu_context, cpu);
 717
 718                         page_addr = hv_cpu->synic_message_page;
 719                         msg = (struct hv_message *)page_addr
 720                                 + VMBUS_MESSAGE_SINT;
 721
 722                         message_type = READ_ONCE(msg->header.message_type);
 723                         if (message_type == HVMSG_NONE)
 724                                 continue;
 725
 726                         hdr = (struct vmbus_channel_message_header *)
 727                                 msg->u.payload;
 728
 729                         if (hdr->msgtype == CHANNELMSG_UNLOAD_RESPONSE)
 730                                 complete(&vmbus_connection.unload_event);
 731
 732                         vmbus_signal_eom(msg, message_type);
 733                 }
 734
 735                 mdelay(10);
 736         }
 737
 738         /*
 739          * We're crashing and already got the UNLOAD_RESPONSE, cleanup all
 740          * maybe-pending messages on all CPUs to be able to receive new
 741          * messages after we reconnect.
 742          */
 743         for_each_online_cpu(cpu) {
 744                 struct hv_per_cpu_context *hv_cpu
 745                         = per_cpu_ptr(hv_context.cpu_context, cpu);
 746
 747                 page_addr = hv_cpu->synic_message_page;
 748                 msg = (struct hv_message *)page_addr + VMBUS_MESSAGE_SINT;
 749                 msg->header.message_type = HVMSG_NONE;
 750         }
 751 }
 752
 753 /*
 754  * vmbus_unload_response - Handler for the unload response.
 755  */
 756 static void vmbus_unload_response(struct vmbus_channel_message_header *hdr)
 757 {
 758         /*
 759          * This is a global event; just wakeup the waiting thread.
 760          * Once we successfully unload, we can cleanup the monitor state.
 761          */
 762         complete(&vmbus_connection.unload_event);
 763 }
 764
 765 void vmbus_initiate_unload(bool crash)
 766 {
 767         struct vmbus_channel_message_header hdr;
 768
 769         /* Pre-Win2012R2 hosts don't support reconnect */
 770         if (vmbus_proto_version < VERSION_WIN8_1)
 771                 return;
 772
 773         init_completion(&vmbus_connection.unload_event);
 774         memset(&hdr, 0, sizeof(struct vmbus_channel_message_header));
 775         hdr.msgtype = CHANNELMSG_UNLOAD;
 776         vmbus_post_msg(&hdr, sizeof(struct vmbus_channel_message_header),
 777                        !crash);
 778
 779         /*
 780          * vmbus_initiate_unload() is also called on crash and the crash can be
 781          * happening in an interrupt context, where scheduling is impossible.
 782          */
 783         if (!crash)
 784                 wait_for_completion(&vmbus_connection.unload_event);
 785         else
 786                 vmbus_wait_for_unload();
 787 }
 788
 789 /*
 790  * vmbus_onoffer - Handler for channel offers from vmbus in parent partition.
 791  *
 792  */
 793 static void vmbus_onoffer(struct vmbus_channel_message_header *hdr)
 794 {
 795         struct vmbus_channel_offer_channel *offer;
 796         struct vmbus_channel *newchannel;
 797
 798         offer = (struct vmbus_channel_offer_channel *)hdr;
 799
 800         /* Allocate the channel object and save this offer. */
 801         newchannel = alloc_channel();
 802         if (!newchannel) {
 803                 vmbus_release_relid(offer->child_relid);
 804                 atomic_dec(&vmbus_connection.offer_in_progress);
 805                 pr_err("Unable to allocate channel object\n");
 806                 return;
 807         }
 808
 809         /*
 810          * Setup state for signalling the host.
 811          */
 812         newchannel->sig_event = VMBUS_EVENT_CONNECTION_ID;
 813
 814         if (vmbus_proto_version != VERSION_WS2008) {
 815                 newchannel->is_dedicated_interrupt =
 816                                 (offer->is_dedicated_interrupt != 0);
 817                 newchannel->sig_event = offer->connection_id;
 818         }
 819
 820         memcpy(&newchannel->offermsg, offer,
 821                sizeof(struct vmbus_channel_offer_channel));
 822         newchannel->monitor_grp = (u8)offer->monitorid / 32;
 823         newchannel->monitor_bit = (u8)offer->monitorid % 32;
 824
 825         vmbus_process_offer(newchannel);
 826 }
 827
 828 /*
 829  * vmbus_onoffer_rescind - Rescind offer handler.
 830  *
 831  * We queue a work item to process this offer synchronously
 832  */
 833 static void vmbus_onoffer_rescind(struct vmbus_channel_message_header *hdr)
 834 {
 835         struct vmbus_channel_rescind_offer *rescind;
 836         struct vmbus_channel *channel;
 837         unsigned long flags;
 838         struct device *dev;
 839
 840         rescind = (struct vmbus_channel_rescind_offer *)hdr;
 841
 842         /*
 843          * The offer msg and the corresponding rescind msg
 844          * from the host are guranteed to be ordered -
 845          * offer comes in first and then the rescind.
 846          * Since we process these events in work elements,
 847          * and with preemption, we may end up processing
 848          * the events out of order. Given that we handle these
 849          * work elements on the same CPU, this is possible only
 850          * in the case of preemption. In any case wait here
 851          * until the offer processing has moved beyond the
 852          * point where the channel is discoverable.
 853          */
 854
 855         while (atomic_read(&vmbus_connection.offer_in_progress) != 0) {
 856                 /*
 857                  * We wait here until any channel offer is currently
 858                  * being processed.
 859                  */
 860                 msleep(1);
 861         }
 862
 863         mutex_lock(&vmbus_connection.channel_mutex);
 864         channel = relid2channel(rescind->child_relid);
 865         mutex_unlock(&vmbus_connection.channel_mutex);
 866
 867         if (channel == NULL) {
 868                 /*
 869                  * We failed in processing the offer message;
 870                  * we would have cleaned up the relid in that
 871                  * failure path.
 872                  */
 873                 return;
 874         }
 875
 876         spin_lock_irqsave(&channel->lock, flags);
 877         channel->rescind = true;
 878         spin_unlock_irqrestore(&channel->lock, flags);
 879
 880         /*
 881          * Now that we have posted the rescind state, perform
 882          * rescind related cleanup.
 883          */
 884         vmbus_rescind_cleanup(channel);
 885
 886         /*
 887          * Now wait for offer handling to complete.
 888          */
 889         while (READ_ONCE(channel->probe_done) == false) {
 890                 /*
 891                  * We wait here until any channel offer is currently
 892                  * being processed.
 893                  */
 894                 msleep(1);
 895         }
 896
 897         /*
 898          * At this point, the rescind handling can proceed safely.
 899          */
 900
 901         if (channel->device_obj) {
 902                 if (channel->chn_rescind_callback) {
 903                         channel->chn_rescind_callback(channel);
 904                         return;
 905                 }
 906                 /*
 907                  * We will have to unregister this device from the
 908                  * driver core.
 909                  */
 910                 dev = get_device(&channel->device_obj->device);
 911                 if (dev) {
 912                         vmbus_device_unregister(channel->device_obj);
 913                         put_device(dev);
 914                 }
 915         }
 916         if (channel->primary_channel != NULL) {
 917                 /*
 918                  * Sub-channel is being rescinded. Following is the channel
 919                  * close sequence when initiated from the driveri (refer to
 920                  * vmbus_close() for details):
 921                  * 1. Close all sub-channels first
 922                  * 2. Then close the primary channel.
 923                  */
 924                 if (channel->state == CHANNEL_OPEN_STATE) {
 925                         /*
 926                          * The channel is currently not open;
 927                          * it is safe for us to cleanup the channel.
 928                          */
 929                         mutex_lock(&vmbus_connection.channel_mutex);
 930                         hv_process_channel_removal(channel,
 931                                                 channel->offermsg.child_relid);
 932                         mutex_unlock(&vmbus_connection.channel_mutex);
 933                 }
 934         }
 935 }
 936
 937 void vmbus_hvsock_device_unregister(struct vmbus_channel *channel)
 938 {
 939         BUG_ON(!is_hvsock_channel(channel));
 940
 941         channel->rescind = true;
 942         vmbus_device_unregister(channel->device_obj);
 943 }
 944 EXPORT_SYMBOL_GPL(vmbus_hvsock_device_unregister);
 945
 946
 947 /*
 948  * vmbus_onoffers_delivered -
 949  * This is invoked when all offers have been delivered.
 950  *
 951  * Nothing to do here.
 952  */
 953 static void vmbus_onoffers_delivered(
 954                         struct vmbus_channel_message_header *hdr)
 955 {
 956 }
 957
 958 /*
 959  * vmbus_onopen_result - Open result handler.
 960  *
 961  * This is invoked when we received a response to our channel open request.
 962  * Find the matching request, copy the response and signal the requesting
 963  * thread.
 964  */
 965 static void vmbus_onopen_result(struct vmbus_channel_message_header *hdr)
 966 {
 967         struct vmbus_channel_open_result *result;
 968         struct vmbus_channel_msginfo *msginfo;
 969         struct vmbus_channel_message_header *requestheader;
 970         struct vmbus_channel_open_channel *openmsg;
 971         unsigned long flags;
 972
 973         result = (struct vmbus_channel_open_result *)hdr;
 974
 975         /*
 976          * Find the open msg, copy the result and signal/unblock the wait event
 977          */
 978         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
 979
 980         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
 981                                 msglistentry) {
 982                 requestheader =
 983                         (struct vmbus_channel_message_header *)msginfo->msg;
 984
 985                 if (requestheader->msgtype == CHANNELMSG_OPENCHANNEL) {
 986                         openmsg =
 987                         (struct vmbus_channel_open_channel *)msginfo->msg;
 988                         if (openmsg->child_relid == result->child_relid &&
 989                             openmsg->openid == result->openid) {
 990                                 memcpy(&msginfo->response.open_result,
 991                                        result,
 992                                        sizeof(
 993                                         struct vmbus_channel_open_result));
 994                                 complete(&msginfo->waitevent);
 995                                 break;
 996                         }
 997                 }
 998         }
 999         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1000 }
1001
1002 /*
1003  * vmbus_ongpadl_created - GPADL created handler.
1004  *
1005  * This is invoked when we received a response to our gpadl create request.
1006  * Find the matching request, copy the response and signal the requesting
1007  * thread.
1008  */
1009 static void vmbus_ongpadl_created(struct vmbus_channel_message_header *hdr)
1010 {
1011         struct vmbus_channel_gpadl_created *gpadlcreated;
1012         struct vmbus_channel_msginfo *msginfo;
1013         struct vmbus_channel_message_header *requestheader;
1014         struct vmbus_channel_gpadl_header *gpadlheader;
1015         unsigned long flags;
1016
1017         gpadlcreated = (struct vmbus_channel_gpadl_created *)hdr;
1018
1019         /*
1020          * Find the establish msg, copy the result and signal/unblock the wait
1021          * event
1022          */
1023         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1024
1025         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1026                                 msglistentry) {
1027                 requestheader =
1028                         (struct vmbus_channel_message_header *)msginfo->msg;
1029
1030                 if (requestheader->msgtype == CHANNELMSG_GPADL_HEADER) {
1031                         gpadlheader =
1032                         (struct vmbus_channel_gpadl_header *)requestheader;
1033
1034                         if ((gpadlcreated->child_relid ==
1035                              gpadlheader->child_relid) &&
1036                             (gpadlcreated->gpadl == gpadlheader->gpadl)) {
1037                                 memcpy(&msginfo->response.gpadl_created,
1038                                        gpadlcreated,
1039                                        sizeof(
1040                                         struct vmbus_channel_gpadl_created));
1041                                 complete(&msginfo->waitevent);
1042                                 break;
1043                         }
1044                 }
1045         }
1046         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1047 }
1048
1049 /*
1050  * vmbus_ongpadl_torndown - GPADL torndown handler.
1051  *
1052  * This is invoked when we received a response to our gpadl teardown request.
1053  * Find the matching request, copy the response and signal the requesting
1054  * thread.
1055  */
1056 static void vmbus_ongpadl_torndown(
1057                         struct vmbus_channel_message_header *hdr)
1058 {
1059         struct vmbus_channel_gpadl_torndown *gpadl_torndown;
1060         struct vmbus_channel_msginfo *msginfo;
1061         struct vmbus_channel_message_header *requestheader;
1062         struct vmbus_channel_gpadl_teardown *gpadl_teardown;
1063         unsigned long flags;
1064
1065         gpadl_torndown = (struct vmbus_channel_gpadl_torndown *)hdr;
1066
1067         /*
1068          * Find the open msg, copy the result and signal/unblock the wait event
1069          */
1070         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1071
1072         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1073                                 msglistentry) {
1074                 requestheader =
1075                         (struct vmbus_channel_message_header *)msginfo->msg;
1076
1077                 if (requestheader->msgtype == CHANNELMSG_GPADL_TEARDOWN) {
1078                         gpadl_teardown =
1079                         (struct vmbus_channel_gpadl_teardown *)requestheader;
1080
1081                         if (gpadl_torndown->gpadl == gpadl_teardown->gpadl) {
1082                                 memcpy(&msginfo->response.gpadl_torndown,
1083                                        gpadl_torndown,
1084                                        sizeof(
1085                                         struct vmbus_channel_gpadl_torndown));
1086                                 complete(&msginfo->waitevent);
1087                                 break;
1088                         }
1089                 }
1090         }
1091         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1092 }
1093
1094 /*
1095  * vmbus_onversion_response - Version response handler
1096  *
1097  * This is invoked when we received a response to our initiate contact request.
1098  * Find the matching request, copy the response and signal the requesting
1099  * thread.
1100  */
1101 static void vmbus_onversion_response(
1102                 struct vmbus_channel_message_header *hdr)
1103 {
1104         struct vmbus_channel_msginfo *msginfo;
1105         struct vmbus_channel_message_header *requestheader;
1106         struct vmbus_channel_version_response *version_response;
1107         unsigned long flags;
1108
1109         version_response = (struct vmbus_channel_version_response *)hdr;
1110         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1111
1112         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1113                                 msglistentry) {
1114                 requestheader =
1115                         (struct vmbus_channel_message_header *)msginfo->msg;
1116
1117                 if (requestheader->msgtype ==
1118                     CHANNELMSG_INITIATE_CONTACT) {
1119                         memcpy(&msginfo->response.version_response,
1120                               version_response,
1121                               sizeof(struct vmbus_channel_version_response));
1122                         complete(&msginfo->waitevent);
1123                 }
1124         }
1125         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1126 }
1127
1128 /* Channel message dispatch table */
1129 const struct vmbus_channel_message_table_entry
1130 channel_message_table[CHANNELMSG_COUNT] = {
1131         { CHANNELMSG_INVALID,                   0, NULL },
1132         { CHANNELMSG_OFFERCHANNEL,              0, vmbus_onoffer },
1133         { CHANNELMSG_RESCIND_CHANNELOFFER,      0, vmbus_onoffer_rescind },
1134         { CHANNELMSG_REQUESTOFFERS,             0, NULL },
1135         { CHANNELMSG_ALLOFFERS_DELIVERED,       1, vmbus_onoffers_delivered },
1136         { CHANNELMSG_OPENCHANNEL,               0, NULL },
1137         { CHANNELMSG_OPENCHANNEL_RESULT,        1, vmbus_onopen_result },
1138         { CHANNELMSG_CLOSECHANNEL,              0, NULL },
1139         { CHANNELMSG_GPADL_HEADER,              0, NULL },
1140         { CHANNELMSG_GPADL_BODY,                0, NULL },
1141         { CHANNELMSG_GPADL_CREATED,             1, vmbus_ongpadl_created },
1142         { CHANNELMSG_GPADL_TEARDOWN,            0, NULL },
1143         { CHANNELMSG_GPADL_TORNDOWN,            1, vmbus_ongpadl_torndown },
1144         { CHANNELMSG_RELID_RELEASED,            0, NULL },
1145         { CHANNELMSG_INITIATE_CONTACT,          0, NULL },
1146         { CHANNELMSG_VERSION_RESPONSE,          1, vmbus_onversion_response },
1147         { CHANNELMSG_UNLOAD,                    0, NULL },
1148         { CHANNELMSG_UNLOAD_RESPONSE,           1, vmbus_unload_response },
1149         { CHANNELMSG_18,                        0, NULL },
1150         { CHANNELMSG_19,                        0, NULL },
1151         { CHANNELMSG_20,                        0, NULL },
1152         { CHANNELMSG_TL_CONNECT_REQUEST,        0, NULL },
1153 };
1154
1155 /*
1156  * vmbus_onmessage - Handler for channel protocol messages.
1157  *
1158  * This is invoked in the vmbus worker thread context.
1159  */
1160 void vmbus_onmessage(void *context)
1161 {
1162         struct hv_message *msg = context;
1163         struct vmbus_channel_message_header *hdr;
1164         int size;
1165
1166         hdr = (struct vmbus_channel_message_header *)msg->u.payload;
1167         size = msg->header.payload_size;
1168
1169         if (hdr->msgtype >= CHANNELMSG_COUNT) {
1170                 pr_err("Received invalid channel message type %d size %d\n",
1171                            hdr->msgtype, size);
1172                 print_hex_dump_bytes("", DUMP_PREFIX_NONE,
1173                                      (unsigned char *)msg->u.payload, size);
1174                 return;
1175         }
1176
1177         if (channel_message_table[hdr->msgtype].message_handler)
1178                 channel_message_table[hdr->msgtype].message_handler(hdr);
1179         else
1180                 pr_err("Unhandled channel message type %d\n", hdr->msgtype);
1181 }
1182
1183 /*
1184  * vmbus_request_offers - Send a request to get all our pending offers.
1185  */
1186 int vmbus_request_offers(void)
1187 {
1188         struct vmbus_channel_message_header *msg;
1189         struct vmbus_channel_msginfo *msginfo;
1190         int ret;
1191
1192         msginfo = kmalloc(sizeof(*msginfo) +
1193                           sizeof(struct vmbus_channel_message_header),
1194                           GFP_KERNEL);
1195         if (!msginfo)
1196                 return -ENOMEM;
1197
1198         msg = (struct vmbus_channel_message_header *)msginfo->msg;
1199
1200         msg->msgtype = CHANNELMSG_REQUESTOFFERS;
1201
1202
1203         ret = vmbus_post_msg(msg, sizeof(struct vmbus_channel_message_header),
1204                              true);
1205         if (ret != 0) {
1206                 pr_err("Unable to request offers - %d\n", ret);
1207
1208                 goto cleanup;
1209         }
1210
1211 cleanup:
1212         kfree(msginfo);
1213
1214         return ret;
1215 }
1216
1217 /*
1218  * Retrieve the (sub) channel on which to send an outgoing request.
1219  * When a primary channel has multiple sub-channels, we try to
1220  * distribute the load equally amongst all available channels.
1221  */
1222 struct vmbus_channel *vmbus_get_outgoing_channel(struct vmbus_channel *primary)
1223 {
1224         struct list_head *cur, *tmp;
1225         int cur_cpu;
1226         struct vmbus_channel *cur_channel;
1227         struct vmbus_channel *outgoing_channel = primary;
1228         int next_channel;
1229         int i = 1;
1230
1231         if (list_empty(&primary->sc_list))
1232                 return outgoing_channel;
1233
1234         next_channel = primary->next_oc++;
1235
1236         if (next_channel > (primary->num_sc)) {
1237                 primary->next_oc = 0;
1238                 return outgoing_channel;
1239         }
1240
1241         cur_cpu = hv_cpu_number_to_vp_number(smp_processor_id());
1242         list_for_each_safe(cur, tmp, &primary->sc_list) {
1243                 cur_channel = list_entry(cur, struct vmbus_channel, sc_list);
1244                 if (cur_channel->state != CHANNEL_OPENED_STATE)
1245                         continue;
1246
1247                 if (cur_channel->target_vp == cur_cpu)
1248                         return cur_channel;
1249
1250                 if (i == next_channel)
1251                         return cur_channel;
1252
1253                 i++;
1254         }
1255
1256         return outgoing_channel;
1257 }
1258 EXPORT_SYMBOL_GPL(vmbus_get_outgoing_channel);
1259
1260 static void invoke_sc_cb(struct vmbus_channel *primary_channel)
1261 {
1262         struct list_head *cur, *tmp;
1263         struct vmbus_channel *cur_channel;
1264
1265         if (primary_channel->sc_creation_callback == NULL)
1266                 return;
1267
1268         list_for_each_safe(cur, tmp, &primary_channel->sc_list) {
1269                 cur_channel = list_entry(cur, struct vmbus_channel, sc_list);
1270
1271                 primary_channel->sc_creation_callback(cur_channel);
1272         }
1273 }
1274
1275 void vmbus_set_sc_create_callback(struct vmbus_channel *primary_channel,
1276                                 void (*sc_cr_cb)(struct vmbus_channel *new_sc))
1277 {
1278         primary_channel->sc_creation_callback = sc_cr_cb;
1279 }
1280 EXPORT_SYMBOL_GPL(vmbus_set_sc_create_callback);
1281
1282 bool vmbus_are_subchannels_present(struct vmbus_channel *primary)
1283 {
1284         bool ret;
1285
1286         ret = !list_empty(&primary->sc_list);
1287
1288         if (ret) {
1289                 /*
1290                  * Invoke the callback on sub-channel creation.
1291                  * This will present a uniform interface to the
1292                  * clients.
1293                  */
1294                 invoke_sc_cb(primary);
1295         }
1296
1297         return ret;
1298 }
1299 EXPORT_SYMBOL_GPL(vmbus_are_subchannels_present);
1300
1301 void vmbus_set_chn_rescind_callback(struct vmbus_channel *channel,
1302                 void (*chn_rescind_cb)(struct vmbus_channel *))
1303 {
1304         channel->chn_rescind_callback = chn_rescind_cb;
1305 }
1306 EXPORT_SYMBOL_GPL(vmbus_set_chn_rescind_callback);