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