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