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
  35 #include "hyperv_vmbus.h"
  36
  37 static void init_vp_index(struct vmbus_channel *channel, u16 dev_type);
  38
  39 static const struct vmbus_device vmbus_devs[] = {
  40         /* IDE */
  41         { .dev_type = HV_IDE,
  42           HV_IDE_GUID,
  43           .perf_device = true,
  44         },
  45
  46         /* SCSI */
  47         { .dev_type = HV_SCSI,
  48           HV_SCSI_GUID,
  49           .perf_device = true,
  50         },
  51
  52         /* Fibre Channel */
  53         { .dev_type = HV_FC,
  54           HV_SYNTHFC_GUID,
  55           .perf_device = true,
  56         },
  57
  58         /* Synthetic NIC */
  59         { .dev_type = HV_NIC,
  60           HV_NIC_GUID,
  61           .perf_device = true,
  62         },
  63
  64         /* Network Direct */
  65         { .dev_type = HV_ND,
  66           HV_ND_GUID,
  67           .perf_device = true,
  68         },
  69
  70         /* PCIE */
  71         { .dev_type = HV_PCIE,
  72           HV_PCIE_GUID,
  73           .perf_device = false,
  74         },
  75
  76         /* Synthetic Frame Buffer */
  77         { .dev_type = HV_FB,
  78           HV_SYNTHVID_GUID,
  79           .perf_device = false,
  80         },
  81
  82         /* Synthetic Keyboard */
  83         { .dev_type = HV_KBD,
  84           HV_KBD_GUID,
  85           .perf_device = false,
  86         },
  87
  88         /* Synthetic MOUSE */
  89         { .dev_type = HV_MOUSE,
  90           HV_MOUSE_GUID,
  91           .perf_device = false,
  92         },
  93
  94         /* KVP */
  95         { .dev_type = HV_KVP,
  96           HV_KVP_GUID,
  97           .perf_device = false,
  98         },
  99
 100         /* Time Synch */
 101         { .dev_type = HV_TS,
 102           HV_TS_GUID,
 103           .perf_device = false,
 104         },
 105
 106         /* Heartbeat */
 107         { .dev_type = HV_HB,
 108           HV_HEART_BEAT_GUID,
 109           .perf_device = false,
 110         },
 111
 112         /* Shutdown */
 113         { .dev_type = HV_SHUTDOWN,
 114           HV_SHUTDOWN_GUID,
 115           .perf_device = false,
 116         },
 117
 118         /* File copy */
 119         { .dev_type = HV_FCOPY,
 120           HV_FCOPY_GUID,
 121           .perf_device = false,
 122         },
 123
 124         /* Backup */
 125         { .dev_type = HV_BACKUP,
 126           HV_VSS_GUID,
 127           .perf_device = false,
 128         },
 129
 130         /* Dynamic Memory */
 131         { .dev_type = HV_DM,
 132           HV_DM_GUID,
 133           .perf_device = false,
 134         },
 135
 136         /* Unknown GUID */
 137         { .dev_type = HV_UNKOWN,
 138           .perf_device = false,
 139         },
 140 };
 141
 142 static const struct {
 143         uuid_le guid;
 144 } vmbus_unsupported_devs[] = {
 145         { HV_AVMA1_GUID },
 146         { HV_AVMA2_GUID },
 147         { HV_RDV_GUID   },
 148 };
 149
 150 /*
 151  * The rescinded channel may be blocked waiting for a response from the host;
 152  * take care of that.
 153  */
 154 static void vmbus_rescind_cleanup(struct vmbus_channel *channel)
 155 {
 156         struct vmbus_channel_msginfo *msginfo;
 157         unsigned long flags;
 158
 159
 160         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
 161
 162         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
 163                                 msglistentry) {
 164
 165                 if (msginfo->waiting_channel == channel) {
 166                         complete(&msginfo->waitevent);
 167                         break;
 168                 }
 169         }
 170         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
 171 }
 172
 173 static bool is_unsupported_vmbus_devs(const uuid_le *guid)
 174 {
 175         int i;
 176
 177         for (i = 0; i < ARRAY_SIZE(vmbus_unsupported_devs); i++)
 178                 if (!uuid_le_cmp(*guid, vmbus_unsupported_devs[i].guid))
 179                         return true;
 180         return false;
 181 }
 182
 183 static u16 hv_get_dev_type(const struct vmbus_channel *channel)
 184 {
 185         const uuid_le *guid = &channel->offermsg.offer.if_type;
 186         u16 i;
 187
 188         if (is_hvsock_channel(channel) || is_unsupported_vmbus_devs(guid))
 189                 return HV_UNKOWN;
 190
 191         for (i = HV_IDE; i < HV_UNKOWN; i++) {
 192                 if (!uuid_le_cmp(*guid, vmbus_devs[i].guid))
 193                         return i;
 194         }
 195         pr_info("Unknown GUID: %pUl\n", guid);
 196         return i;
 197 }
 198
 199 /**
 200  * vmbus_prep_negotiate_resp() - Create default response for Hyper-V Negotiate message
 201  * @icmsghdrp: Pointer to msg header structure
 202  * @icmsg_negotiate: Pointer to negotiate message structure
 203  * @buf: Raw buffer channel data
 204  *
 205  * @icmsghdrp is of type &struct icmsg_hdr.
 206  * @negop is of type &struct icmsg_negotiate.
 207  * Set up and fill in default negotiate response message.
 208  *
 209  * The fw_version specifies the  framework version that
 210  * we can support and srv_version specifies the service
 211  * version we can support.
 212  *
 213  * Mainly used by Hyper-V drivers.
 214  */
 215 bool vmbus_prep_negotiate_resp(struct icmsg_hdr *icmsghdrp,
 216                                 struct icmsg_negotiate *negop, u8 *buf,
 217                                 int fw_version, int srv_version)
 218 {
 219         int icframe_major, icframe_minor;
 220         int icmsg_major, icmsg_minor;
 221         int fw_major, fw_minor;
 222         int srv_major, srv_minor;
 223         int i;
 224         bool found_match = false;
 225
 226         icmsghdrp->icmsgsize = 0x10;
 227         fw_major = (fw_version >> 16);
 228         fw_minor = (fw_version & 0xFFFF);
 229
 230         srv_major = (srv_version >> 16);
 231         srv_minor = (srv_version & 0xFFFF);
 232
 233         negop = (struct icmsg_negotiate *)&buf[
 234                 sizeof(struct vmbuspipe_hdr) +
 235                 sizeof(struct icmsg_hdr)];
 236
 237         icframe_major = negop->icframe_vercnt;
 238         icframe_minor = 0;
 239
 240         icmsg_major = negop->icmsg_vercnt;
 241         icmsg_minor = 0;
 242
 243         /*
 244          * Select the framework version number we will
 245          * support.
 246          */
 247
 248         for (i = 0; i < negop->icframe_vercnt; i++) {
 249                 if ((negop->icversion_data[i].major == fw_major) &&
 250                    (negop->icversion_data[i].minor == fw_minor)) {
 251                         icframe_major = negop->icversion_data[i].major;
 252                         icframe_minor = negop->icversion_data[i].minor;
 253                         found_match = true;
 254                 }
 255         }
 256
 257         if (!found_match)
 258                 goto fw_error;
 259
 260         found_match = false;
 261
 262         for (i = negop->icframe_vercnt;
 263                  (i < negop->icframe_vercnt + negop->icmsg_vercnt); i++) {
 264                 if ((negop->icversion_data[i].major == srv_major) &&
 265                    (negop->icversion_data[i].minor == srv_minor)) {
 266                         icmsg_major = negop->icversion_data[i].major;
 267                         icmsg_minor = negop->icversion_data[i].minor;
 268                         found_match = true;
 269                 }
 270         }
 271
 272         /*
 273          * Respond with the framework and service
 274          * version numbers we can support.
 275          */
 276
 277 fw_error:
 278         if (!found_match) {
 279                 negop->icframe_vercnt = 0;
 280                 negop->icmsg_vercnt = 0;
 281         } else {
 282                 negop->icframe_vercnt = 1;
 283                 negop->icmsg_vercnt = 1;
 284         }
 285
 286         negop->icversion_data[0].major = icframe_major;
 287         negop->icversion_data[0].minor = icframe_minor;
 288         negop->icversion_data[1].major = icmsg_major;
 289         negop->icversion_data[1].minor = icmsg_minor;
 290         return found_match;
 291 }
 292
 293 EXPORT_SYMBOL_GPL(vmbus_prep_negotiate_resp);
 294
 295 /*
 296  * alloc_channel - Allocate and initialize a vmbus channel object
 297  */
 298 static struct vmbus_channel *alloc_channel(void)
 299 {
 300         struct vmbus_channel *channel;
 301
 302         channel = kzalloc(sizeof(*channel), GFP_ATOMIC);
 303         if (!channel)
 304                 return NULL;
 305
 306         channel->acquire_ring_lock = true;
 307         spin_lock_init(&channel->inbound_lock);
 308         spin_lock_init(&channel->lock);
 309
 310         INIT_LIST_HEAD(&channel->sc_list);
 311         INIT_LIST_HEAD(&channel->percpu_list);
 312
 313         return channel;
 314 }
 315
 316 /*
 317  * free_channel - Release the resources used by the vmbus channel object
 318  */
 319 static void free_channel(struct vmbus_channel *channel)
 320 {
 321         kfree(channel);
 322 }
 323
 324 static void percpu_channel_enq(void *arg)
 325 {
 326         struct vmbus_channel *channel = arg;
 327         int cpu = smp_processor_id();
 328
 329         list_add_tail(&channel->percpu_list, &hv_context.percpu_list[cpu]);
 330 }
 331
 332 static void percpu_channel_deq(void *arg)
 333 {
 334         struct vmbus_channel *channel = arg;
 335
 336         list_del(&channel->percpu_list);
 337 }
 338
 339
 340 static void vmbus_release_relid(u32 relid)
 341 {
 342         struct vmbus_channel_relid_released msg;
 343
 344         memset(&msg, 0, sizeof(struct vmbus_channel_relid_released));
 345         msg.child_relid = relid;
 346         msg.header.msgtype = CHANNELMSG_RELID_RELEASED;
 347         vmbus_post_msg(&msg, sizeof(struct vmbus_channel_relid_released),
 348                        true);
 349 }
 350
 351 void hv_event_tasklet_disable(struct vmbus_channel *channel)
 352 {
 353         struct tasklet_struct *tasklet;
 354         tasklet = hv_context.event_dpc[channel->target_cpu];
 355         tasklet_disable(tasklet);
 356 }
 357
 358 void hv_event_tasklet_enable(struct vmbus_channel *channel)
 359 {
 360         struct tasklet_struct *tasklet;
 361         tasklet = hv_context.event_dpc[channel->target_cpu];
 362         tasklet_enable(tasklet);
 363
 364         /* In case there is any pending event */
 365         tasklet_schedule(tasklet);
 366 }
 367
 368 void hv_process_channel_removal(struct vmbus_channel *channel, u32 relid)
 369 {
 370         unsigned long flags;
 371         struct vmbus_channel *primary_channel;
 372
 373         BUG_ON(!channel->rescind);
 374         BUG_ON(!mutex_is_locked(&vmbus_connection.channel_mutex));
 375
 376         hv_event_tasklet_disable(channel);
 377         if (channel->target_cpu != get_cpu()) {
 378                 put_cpu();
 379                 smp_call_function_single(channel->target_cpu,
 380                                          percpu_channel_deq, channel, true);
 381         } else {
 382                 percpu_channel_deq(channel);
 383                 put_cpu();
 384         }
 385         hv_event_tasklet_enable(channel);
 386
 387         if (channel->primary_channel == NULL) {
 388                 list_del(&channel->listentry);
 389
 390                 primary_channel = channel;
 391         } else {
 392                 primary_channel = channel->primary_channel;
 393                 spin_lock_irqsave(&primary_channel->lock, flags);
 394                 list_del(&channel->sc_list);
 395                 primary_channel->num_sc--;
 396                 spin_unlock_irqrestore(&primary_channel->lock, flags);
 397         }
 398
 399         /*
 400          * We need to free the bit for init_vp_index() to work in the case
 401          * of sub-channel, when we reload drivers like hv_netvsc.
 402          */
 403         if (channel->affinity_policy == HV_LOCALIZED)
 404                 cpumask_clear_cpu(channel->target_cpu,
 405                                   &primary_channel->alloced_cpus_in_node);
 406
 407         vmbus_release_relid(relid);
 408
 409         free_channel(channel);
 410 }
 411
 412 void vmbus_free_channels(void)
 413 {
 414         struct vmbus_channel *channel, *tmp;
 415
 416         mutex_lock(&vmbus_connection.channel_mutex);
 417         list_for_each_entry_safe(channel, tmp, &vmbus_connection.chn_list,
 418                 listentry) {
 419                 /* hv_process_channel_removal() needs this */
 420                 channel->rescind = true;
 421
 422                 vmbus_device_unregister(channel->device_obj);
 423         }
 424         mutex_unlock(&vmbus_connection.channel_mutex);
 425 }
 426
 427 /*
 428  * vmbus_process_offer - Process the offer by creating a channel/device
 429  * associated with this offer
 430  */
 431 static void vmbus_process_offer(struct vmbus_channel *newchannel)
 432 {
 433         struct vmbus_channel *channel;
 434         bool fnew = true;
 435         unsigned long flags;
 436         u16 dev_type;
 437         int ret;
 438
 439         /* Make sure this is a new offer */
 440         mutex_lock(&vmbus_connection.channel_mutex);
 441
 442         list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
 443                 if (!uuid_le_cmp(channel->offermsg.offer.if_type,
 444                         newchannel->offermsg.offer.if_type) &&
 445                         !uuid_le_cmp(channel->offermsg.offer.if_instance,
 446                                 newchannel->offermsg.offer.if_instance)) {
 447                         fnew = false;
 448                         break;
 449                 }
 450         }
 451
 452         if (fnew)
 453                 list_add_tail(&newchannel->listentry,
 454                               &vmbus_connection.chn_list);
 455
 456         mutex_unlock(&vmbus_connection.channel_mutex);
 457
 458         if (!fnew) {
 459                 /*
 460                  * Check to see if this is a sub-channel.
 461                  */
 462                 if (newchannel->offermsg.offer.sub_channel_index != 0) {
 463                         /*
 464                          * Process the sub-channel.
 465                          */
 466                         newchannel->primary_channel = channel;
 467                         spin_lock_irqsave(&channel->lock, flags);
 468                         list_add_tail(&newchannel->sc_list, &channel->sc_list);
 469                         channel->num_sc++;
 470                         spin_unlock_irqrestore(&channel->lock, flags);
 471                 } else
 472                         goto err_free_chan;
 473         }
 474
 475         dev_type = hv_get_dev_type(newchannel);
 476
 477         init_vp_index(newchannel, dev_type);
 478
 479         hv_event_tasklet_disable(newchannel);
 480         if (newchannel->target_cpu != get_cpu()) {
 481                 put_cpu();
 482                 smp_call_function_single(newchannel->target_cpu,
 483                                          percpu_channel_enq,
 484                                          newchannel, true);
 485         } else {
 486                 percpu_channel_enq(newchannel);
 487                 put_cpu();
 488         }
 489         hv_event_tasklet_enable(newchannel);
 490
 491         /*
 492          * This state is used to indicate a successful open
 493          * so that when we do close the channel normally, we
 494          * can cleanup properly
 495          */
 496         newchannel->state = CHANNEL_OPEN_STATE;
 497
 498         if (!fnew) {
 499                 if (channel->sc_creation_callback != NULL)
 500                         channel->sc_creation_callback(newchannel);
 501                 return;
 502         }
 503
 504         /*
 505          * Start the process of binding this offer to the driver
 506          * We need to set the DeviceObject field before calling
 507          * vmbus_child_dev_add()
 508          */
 509         newchannel->device_obj = vmbus_device_create(
 510                 &newchannel->offermsg.offer.if_type,
 511                 &newchannel->offermsg.offer.if_instance,
 512                 newchannel);
 513         if (!newchannel->device_obj)
 514                 goto err_deq_chan;
 515
 516         newchannel->device_obj->device_id = dev_type;
 517         /*
 518          * Add the new device to the bus. This will kick off device-driver
 519          * binding which eventually invokes the device driver's AddDevice()
 520          * method.
 521          */
 522         mutex_lock(&vmbus_connection.channel_mutex);
 523         ret = vmbus_device_register(newchannel->device_obj);
 524         mutex_unlock(&vmbus_connection.channel_mutex);
 525
 526         if (ret != 0) {
 527                 pr_err("unable to add child device object (relid %d)\n",
 528                         newchannel->offermsg.child_relid);
 529                 kfree(newchannel->device_obj);
 530                 goto err_deq_chan;
 531         }
 532         return;
 533
 534 err_deq_chan:
 535         mutex_lock(&vmbus_connection.channel_mutex);
 536         list_del(&newchannel->listentry);
 537         mutex_unlock(&vmbus_connection.channel_mutex);
 538
 539         hv_event_tasklet_disable(newchannel);
 540         if (newchannel->target_cpu != get_cpu()) {
 541                 put_cpu();
 542                 smp_call_function_single(newchannel->target_cpu,
 543                                          percpu_channel_deq, newchannel, true);
 544         } else {
 545                 percpu_channel_deq(newchannel);
 546                 put_cpu();
 547         }
 548         hv_event_tasklet_enable(newchannel);
 549
 550         vmbus_release_relid(newchannel->offermsg.child_relid);
 551
 552 err_free_chan:
 553         free_channel(newchannel);
 554 }
 555
 556 /*
 557  * We use this state to statically distribute the channel interrupt load.
 558  */
 559 static int next_numa_node_id;
 560
 561 /*
 562  * Starting with Win8, we can statically distribute the incoming
 563  * channel interrupt load by binding a channel to VCPU.
 564  * We do this in a hierarchical fashion:
 565  * First distribute the primary channels across available NUMA nodes
 566  * and then distribute the subchannels amongst the CPUs in the NUMA
 567  * node assigned to the primary channel.
 568  *
 569  * For pre-win8 hosts or non-performance critical channels we assign the
 570  * first CPU in the first NUMA node.
 571  */
 572 static void init_vp_index(struct vmbus_channel *channel, u16 dev_type)
 573 {
 574         u32 cur_cpu;
 575         bool perf_chn = vmbus_devs[dev_type].perf_device;
 576         struct vmbus_channel *primary = channel->primary_channel;
 577         int next_node;
 578         struct cpumask available_mask;
 579         struct cpumask *alloced_mask;
 580
 581         if ((vmbus_proto_version == VERSION_WS2008) ||
 582             (vmbus_proto_version == VERSION_WIN7) || (!perf_chn)) {
 583                 /*
 584                  * Prior to win8, all channel interrupts are
 585                  * delivered on cpu 0.
 586                  * Also if the channel is not a performance critical
 587                  * channel, bind it to cpu 0.
 588                  */
 589                 channel->numa_node = 0;
 590                 channel->target_cpu = 0;
 591                 channel->target_vp = hv_context.vp_index[0];
 592                 return;
 593         }
 594
 595         /*
 596          * Based on the channel affinity policy, we will assign the NUMA
 597          * nodes.
 598          */
 599
 600         if ((channel->affinity_policy == HV_BALANCED) || (!primary)) {
 601                 while (true) {
 602                         next_node = next_numa_node_id++;
 603                         if (next_node == nr_node_ids) {
 604                                 next_node = next_numa_node_id = 0;
 605                                 continue;
 606                         }
 607                         if (cpumask_empty(cpumask_of_node(next_node)))
 608                                 continue;
 609                         break;
 610                 }
 611                 channel->numa_node = next_node;
 612                 primary = channel;
 613         }
 614         alloced_mask = &hv_context.hv_numa_map[primary->numa_node];
 615
 616         if (cpumask_weight(alloced_mask) ==
 617             cpumask_weight(cpumask_of_node(primary->numa_node))) {
 618                 /*
 619                  * We have cycled through all the CPUs in the node;
 620                  * reset the alloced map.
 621                  */
 622                 cpumask_clear(alloced_mask);
 623         }
 624
 625         cpumask_xor(&available_mask, alloced_mask,
 626                     cpumask_of_node(primary->numa_node));
 627
 628         cur_cpu = -1;
 629
 630         if (primary->affinity_policy == HV_LOCALIZED) {
 631                 /*
 632                  * Normally Hyper-V host doesn't create more subchannels
 633                  * than there are VCPUs on the node but it is possible when not
 634                  * all present VCPUs on the node are initialized by guest.
 635                  * Clear the alloced_cpus_in_node to start over.
 636                  */
 637                 if (cpumask_equal(&primary->alloced_cpus_in_node,
 638                                   cpumask_of_node(primary->numa_node)))
 639                         cpumask_clear(&primary->alloced_cpus_in_node);
 640         }
 641
 642         while (true) {
 643                 cur_cpu = cpumask_next(cur_cpu, &available_mask);
 644                 if (cur_cpu >= nr_cpu_ids) {
 645                         cur_cpu = -1;
 646                         cpumask_copy(&available_mask,
 647                                      cpumask_of_node(primary->numa_node));
 648                         continue;
 649                 }
 650
 651                 if (primary->affinity_policy == HV_LOCALIZED) {
 652                         /*
 653                          * NOTE: in the case of sub-channel, we clear the
 654                          * sub-channel related bit(s) in
 655                          * primary->alloced_cpus_in_node in
 656                          * hv_process_channel_removal(), so when we
 657                          * reload drivers like hv_netvsc in SMP guest, here
 658                          * we're able to re-allocate
 659                          * bit from primary->alloced_cpus_in_node.
 660                          */
 661                         if (!cpumask_test_cpu(cur_cpu,
 662                                               &primary->alloced_cpus_in_node)) {
 663                                 cpumask_set_cpu(cur_cpu,
 664                                                 &primary->alloced_cpus_in_node);
 665                                 cpumask_set_cpu(cur_cpu, alloced_mask);
 666                                 break;
 667                         }
 668                 } else {
 669                         cpumask_set_cpu(cur_cpu, alloced_mask);
 670                         break;
 671                 }
 672         }
 673
 674         channel->target_cpu = cur_cpu;
 675         channel->target_vp = hv_context.vp_index[cur_cpu];
 676 }
 677
 678 #define UNLOAD_DELAY_UNIT_MS    10              /* 10 milliseconds */
 679 #define UNLOAD_WAIT_MS          (100*1000)      /* 100 seconds */
 680 #define UNLOAD_WAIT_LOOPS       (UNLOAD_WAIT_MS/UNLOAD_DELAY_UNIT_MS)
 681 #define UNLOAD_MSG_MS           (5*1000)        /* Every 5 seconds */
 682 #define UNLOAD_MSG_LOOPS        (UNLOAD_MSG_MS/UNLOAD_DELAY_UNIT_MS)
 683
 684 static void vmbus_wait_for_unload(void)
 685 {
 686         int cpu;
 687         void *page_addr;
 688         struct hv_message *msg;
 689         struct vmbus_channel_message_header *hdr;
 690         u32 message_type, i;
 691
 692         /*
 693          * CHANNELMSG_UNLOAD_RESPONSE is always delivered to the CPU which was
 694          * used for initial contact or to CPU0 depending on host version. When
 695          * we're crashing on a different CPU let's hope that IRQ handler on
 696          * the cpu which receives CHANNELMSG_UNLOAD_RESPONSE is still
 697          * functional and vmbus_unload_response() will complete
 698          * vmbus_connection.unload_event. If not, the last thing we can do is
 699          * read message pages for all CPUs directly.
 700          *
 701          * Wait up to 100 seconds since an Azure host must writeback any dirty
 702          * data in its disk cache before the VMbus UNLOAD request will
 703          * complete. This flushing has been empirically observed to take up
 704          * to 50 seconds in cases with a lot of dirty data, so allow additional
 705          * leeway and for inaccuracies in mdelay(). But eventually time out so
 706          * that the panic path can't get hung forever in case the response
 707          * message isn't seen.
 708          */
 709         for (i = 1; i <= UNLOAD_WAIT_LOOPS; i++) {
 710                 if (completion_done(&vmbus_connection.unload_event))
 711                         goto completed;
 712
 713                 for_each_online_cpu(cpu) {
 714                         page_addr = hv_context.synic_message_page[cpu];
 715                         msg = (struct hv_message *)page_addr +
 716                                 VMBUS_MESSAGE_SINT;
 717
 718                         message_type = READ_ONCE(msg->header.message_type);
 719                         if (message_type == HVMSG_NONE)
 720                                 continue;
 721
 722                         hdr = (struct vmbus_channel_message_header *)
 723                                 msg->u.payload;
 724
 725                         if (hdr->msgtype == CHANNELMSG_UNLOAD_RESPONSE)
 726                                 complete(&vmbus_connection.unload_event);
 727
 728                         vmbus_signal_eom(msg, message_type);
 729                 }
 730
 731                 /*
 732                  * Give a notice periodically so someone watching the
 733                  * serial output won't think it is completely hung.
 734                  */
 735                 if (!(i % UNLOAD_MSG_LOOPS))
 736                         pr_notice("Waiting for VMBus UNLOAD to complete\n");
 737
 738                 mdelay(UNLOAD_DELAY_UNIT_MS);
 739         }
 740         pr_err("Continuing even though VMBus UNLOAD did not complete\n");
 741
 742 completed:
 743         /*
 744          * We're crashing and already got the UNLOAD_RESPONSE, cleanup all
 745          * maybe-pending messages on all CPUs to be able to receive new
 746          * messages after we reconnect.
 747          */
 748         for_each_online_cpu(cpu) {
 749                 page_addr = hv_context.synic_message_page[cpu];
 750                 msg = (struct hv_message *)page_addr + VMBUS_MESSAGE_SINT;
 751                 msg->header.message_type = HVMSG_NONE;
 752         }
 753 }
 754
 755 /*
 756  * vmbus_unload_response - Handler for the unload response.
 757  */
 758 static void vmbus_unload_response(struct vmbus_channel_message_header *hdr)
 759 {
 760         /*
 761          * This is a global event; just wakeup the waiting thread.
 762          * Once we successfully unload, we can cleanup the monitor state.
 763          */
 764         complete(&vmbus_connection.unload_event);
 765 }
 766
 767 void vmbus_initiate_unload(bool crash)
 768 {
 769         struct vmbus_channel_message_header hdr;
 770
 771         /* Pre-Win2012R2 hosts don't support reconnect */
 772         if (vmbus_proto_version < VERSION_WIN8_1)
 773                 return;
 774
 775         init_completion(&vmbus_connection.unload_event);
 776         memset(&hdr, 0, sizeof(struct vmbus_channel_message_header));
 777         hdr.msgtype = CHANNELMSG_UNLOAD;
 778         vmbus_post_msg(&hdr, sizeof(struct vmbus_channel_message_header),
 779                        !crash);
 780
 781         /*
 782          * vmbus_initiate_unload() is also called on crash and the crash can be
 783          * happening in an interrupt context, where scheduling is impossible.
 784          */
 785         if (!crash)
 786                 wait_for_completion(&vmbus_connection.unload_event);
 787         else
 788                 vmbus_wait_for_unload();
 789 }
 790
 791 /*
 792  * vmbus_onoffer - Handler for channel offers from vmbus in parent partition.
 793  *
 794  */
 795 static void vmbus_onoffer(struct vmbus_channel_message_header *hdr)
 796 {
 797         struct vmbus_channel_offer_channel *offer;
 798         struct vmbus_channel *newchannel;
 799
 800         offer = (struct vmbus_channel_offer_channel *)hdr;
 801
 802         /* Allocate the channel object and save this offer. */
 803         newchannel = alloc_channel();
 804         if (!newchannel) {
 805                 vmbus_release_relid(offer->child_relid);
 806                 pr_err("Unable to allocate channel object\n");
 807                 return;
 808         }
 809
 810         /*
 811          * By default we setup state to enable batched
 812          * reading. A specific service can choose to
 813          * disable this prior to opening the channel.
 814          */
 815         newchannel->batched_reading = true;
 816
 817         /*
 818          * Setup state for signalling the host.
 819          */
 820         newchannel->sig_event = (struct hv_input_signal_event *)
 821                                 (ALIGN((unsigned long)
 822                                 &newchannel->sig_buf,
 823                                 HV_HYPERCALL_PARAM_ALIGN));
 824
 825         newchannel->sig_event->connectionid.asu32 = 0;
 826         newchannel->sig_event->connectionid.u.id = VMBUS_EVENT_CONNECTION_ID;
 827         newchannel->sig_event->flag_number = 0;
 828         newchannel->sig_event->rsvdz = 0;
 829
 830         if (vmbus_proto_version != VERSION_WS2008) {
 831                 newchannel->is_dedicated_interrupt =
 832                                 (offer->is_dedicated_interrupt != 0);
 833                 newchannel->sig_event->connectionid.u.id =
 834                                 offer->connection_id;
 835         }
 836
 837         memcpy(&newchannel->offermsg, offer,
 838                sizeof(struct vmbus_channel_offer_channel));
 839         newchannel->monitor_grp = (u8)offer->monitorid / 32;
 840         newchannel->monitor_bit = (u8)offer->monitorid % 32;
 841
 842         vmbus_process_offer(newchannel);
 843 }
 844
 845 /*
 846  * vmbus_onoffer_rescind - Rescind offer handler.
 847  *
 848  * We queue a work item to process this offer synchronously
 849  */
 850 static void vmbus_onoffer_rescind(struct vmbus_channel_message_header *hdr)
 851 {
 852         struct vmbus_channel_rescind_offer *rescind;
 853         struct vmbus_channel *channel;
 854         unsigned long flags;
 855         struct device *dev;
 856
 857         rescind = (struct vmbus_channel_rescind_offer *)hdr;
 858
 859         mutex_lock(&vmbus_connection.channel_mutex);
 860         channel = relid2channel(rescind->child_relid);
 861
 862         if (channel == NULL) {
 863                 /*
 864                  * This is very impossible, because in
 865                  * vmbus_process_offer(), we have already invoked
 866                  * vmbus_release_relid() on error.
 867                  */
 868                 goto out;
 869         }
 870
 871         spin_lock_irqsave(&channel->lock, flags);
 872         channel->rescind = true;
 873         spin_unlock_irqrestore(&channel->lock, flags);
 874
 875         vmbus_rescind_cleanup(channel);
 876
 877         if (channel->device_obj) {
 878                 if (channel->chn_rescind_callback) {
 879                         channel->chn_rescind_callback(channel);
 880                         goto out;
 881                 }
 882                 /*
 883                  * We will have to unregister this device from the
 884                  * driver core.
 885                  */
 886                 dev = get_device(&channel->device_obj->device);
 887                 if (dev) {
 888                         vmbus_device_unregister(channel->device_obj);
 889                         put_device(dev);
 890                 }
 891         } else {
 892                 hv_process_channel_removal(channel,
 893                         channel->offermsg.child_relid);
 894         }
 895
 896 out:
 897         mutex_unlock(&vmbus_connection.channel_mutex);
 898 }
 899
 900 void vmbus_hvsock_device_unregister(struct vmbus_channel *channel)
 901 {
 902         mutex_lock(&vmbus_connection.channel_mutex);
 903
 904         BUG_ON(!is_hvsock_channel(channel));
 905
 906         channel->rescind = true;
 907         vmbus_device_unregister(channel->device_obj);
 908
 909         mutex_unlock(&vmbus_connection.channel_mutex);
 910 }
 911 EXPORT_SYMBOL_GPL(vmbus_hvsock_device_unregister);
 912
 913
 914 /*
 915  * vmbus_onoffers_delivered -
 916  * This is invoked when all offers have been delivered.
 917  *
 918  * Nothing to do here.
 919  */
 920 static void vmbus_onoffers_delivered(
 921                         struct vmbus_channel_message_header *hdr)
 922 {
 923 }
 924
 925 /*
 926  * vmbus_onopen_result - Open result handler.
 927  *
 928  * This is invoked when we received a response to our channel open request.
 929  * Find the matching request, copy the response and signal the requesting
 930  * thread.
 931  */
 932 static void vmbus_onopen_result(struct vmbus_channel_message_header *hdr)
 933 {
 934         struct vmbus_channel_open_result *result;
 935         struct vmbus_channel_msginfo *msginfo;
 936         struct vmbus_channel_message_header *requestheader;
 937         struct vmbus_channel_open_channel *openmsg;
 938         unsigned long flags;
 939
 940         result = (struct vmbus_channel_open_result *)hdr;
 941
 942         /*
 943          * Find the open msg, copy the result and signal/unblock the wait event
 944          */
 945         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
 946
 947         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
 948                                 msglistentry) {
 949                 requestheader =
 950                         (struct vmbus_channel_message_header *)msginfo->msg;
 951
 952                 if (requestheader->msgtype == CHANNELMSG_OPENCHANNEL) {
 953                         openmsg =
 954                         (struct vmbus_channel_open_channel *)msginfo->msg;
 955                         if (openmsg->child_relid == result->child_relid &&
 956                             openmsg->openid == result->openid) {
 957                                 memcpy(&msginfo->response.open_result,
 958                                        result,
 959                                        sizeof(
 960                                         struct vmbus_channel_open_result));
 961                                 complete(&msginfo->waitevent);
 962                                 break;
 963                         }
 964                 }
 965         }
 966         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
 967 }
 968
 969 /*
 970  * vmbus_ongpadl_created - GPADL created handler.
 971  *
 972  * This is invoked when we received a response to our gpadl create request.
 973  * Find the matching request, copy the response and signal the requesting
 974  * thread.
 975  */
 976 static void vmbus_ongpadl_created(struct vmbus_channel_message_header *hdr)
 977 {
 978         struct vmbus_channel_gpadl_created *gpadlcreated;
 979         struct vmbus_channel_msginfo *msginfo;
 980         struct vmbus_channel_message_header *requestheader;
 981         struct vmbus_channel_gpadl_header *gpadlheader;
 982         unsigned long flags;
 983
 984         gpadlcreated = (struct vmbus_channel_gpadl_created *)hdr;
 985
 986         /*
 987          * Find the establish msg, copy the result and signal/unblock the wait
 988          * event
 989          */
 990         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
 991
 992         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
 993                                 msglistentry) {
 994                 requestheader =
 995                         (struct vmbus_channel_message_header *)msginfo->msg;
 996
 997                 if (requestheader->msgtype == CHANNELMSG_GPADL_HEADER) {
 998                         gpadlheader =
 999                         (struct vmbus_channel_gpadl_header *)requestheader;
1000
1001                         if ((gpadlcreated->child_relid ==
1002                              gpadlheader->child_relid) &&
1003                             (gpadlcreated->gpadl == gpadlheader->gpadl)) {
1004                                 memcpy(&msginfo->response.gpadl_created,
1005                                        gpadlcreated,
1006                                        sizeof(
1007                                         struct vmbus_channel_gpadl_created));
1008                                 complete(&msginfo->waitevent);
1009                                 break;
1010                         }
1011                 }
1012         }
1013         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1014 }
1015
1016 /*
1017  * vmbus_ongpadl_torndown - GPADL torndown handler.
1018  *
1019  * This is invoked when we received a response to our gpadl teardown request.
1020  * Find the matching request, copy the response and signal the requesting
1021  * thread.
1022  */
1023 static void vmbus_ongpadl_torndown(
1024                         struct vmbus_channel_message_header *hdr)
1025 {
1026         struct vmbus_channel_gpadl_torndown *gpadl_torndown;
1027         struct vmbus_channel_msginfo *msginfo;
1028         struct vmbus_channel_message_header *requestheader;
1029         struct vmbus_channel_gpadl_teardown *gpadl_teardown;
1030         unsigned long flags;
1031
1032         gpadl_torndown = (struct vmbus_channel_gpadl_torndown *)hdr;
1033
1034         /*
1035          * Find the open msg, copy the result and signal/unblock the wait event
1036          */
1037         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1038
1039         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1040                                 msglistentry) {
1041                 requestheader =
1042                         (struct vmbus_channel_message_header *)msginfo->msg;
1043
1044                 if (requestheader->msgtype == CHANNELMSG_GPADL_TEARDOWN) {
1045                         gpadl_teardown =
1046                         (struct vmbus_channel_gpadl_teardown *)requestheader;
1047
1048                         if (gpadl_torndown->gpadl == gpadl_teardown->gpadl) {
1049                                 memcpy(&msginfo->response.gpadl_torndown,
1050                                        gpadl_torndown,
1051                                        sizeof(
1052                                         struct vmbus_channel_gpadl_torndown));
1053                                 complete(&msginfo->waitevent);
1054                                 break;
1055                         }
1056                 }
1057         }
1058         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1059 }
1060
1061 /*
1062  * vmbus_onversion_response - Version response handler
1063  *
1064  * This is invoked when we received a response to our initiate contact request.
1065  * Find the matching request, copy the response and signal the requesting
1066  * thread.
1067  */
1068 static void vmbus_onversion_response(
1069                 struct vmbus_channel_message_header *hdr)
1070 {
1071         struct vmbus_channel_msginfo *msginfo;
1072         struct vmbus_channel_message_header *requestheader;
1073         struct vmbus_channel_version_response *version_response;
1074         unsigned long flags;
1075
1076         version_response = (struct vmbus_channel_version_response *)hdr;
1077         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1078
1079         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1080                                 msglistentry) {
1081                 requestheader =
1082                         (struct vmbus_channel_message_header *)msginfo->msg;
1083
1084                 if (requestheader->msgtype ==
1085                     CHANNELMSG_INITIATE_CONTACT) {
1086                         memcpy(&msginfo->response.version_response,
1087                               version_response,
1088                               sizeof(struct vmbus_channel_version_response));
1089                         complete(&msginfo->waitevent);
1090                 }
1091         }
1092         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1093 }
1094
1095 /* Channel message dispatch table */
1096 struct vmbus_channel_message_table_entry
1097         channel_message_table[CHANNELMSG_COUNT] = {
1098         {CHANNELMSG_INVALID,                    0, NULL},
1099         {CHANNELMSG_OFFERCHANNEL,               0, vmbus_onoffer},
1100         {CHANNELMSG_RESCIND_CHANNELOFFER,       0, vmbus_onoffer_rescind},
1101         {CHANNELMSG_REQUESTOFFERS,              0, NULL},
1102         {CHANNELMSG_ALLOFFERS_DELIVERED,        1, vmbus_onoffers_delivered},
1103         {CHANNELMSG_OPENCHANNEL,                0, NULL},
1104         {CHANNELMSG_OPENCHANNEL_RESULT,         1, vmbus_onopen_result},
1105         {CHANNELMSG_CLOSECHANNEL,               0, NULL},
1106         {CHANNELMSG_GPADL_HEADER,               0, NULL},
1107         {CHANNELMSG_GPADL_BODY,                 0, NULL},
1108         {CHANNELMSG_GPADL_CREATED,              1, vmbus_ongpadl_created},
1109         {CHANNELMSG_GPADL_TEARDOWN,             0, NULL},
1110         {CHANNELMSG_GPADL_TORNDOWN,             1, vmbus_ongpadl_torndown},
1111         {CHANNELMSG_RELID_RELEASED,             0, NULL},
1112         {CHANNELMSG_INITIATE_CONTACT,           0, NULL},
1113         {CHANNELMSG_VERSION_RESPONSE,           1, vmbus_onversion_response},
1114         {CHANNELMSG_UNLOAD,                     0, NULL},
1115         {CHANNELMSG_UNLOAD_RESPONSE,            1, vmbus_unload_response},
1116         {CHANNELMSG_18,                         0, NULL},
1117         {CHANNELMSG_19,                         0, NULL},
1118         {CHANNELMSG_20,                         0, NULL},
1119         {CHANNELMSG_TL_CONNECT_REQUEST,         0, NULL},
1120 };
1121
1122 /*
1123  * vmbus_onmessage - Handler for channel protocol messages.
1124  *
1125  * This is invoked in the vmbus worker thread context.
1126  */
1127 void vmbus_onmessage(void *context)
1128 {
1129         struct hv_message *msg = context;
1130         struct vmbus_channel_message_header *hdr;
1131         int size;
1132
1133         hdr = (struct vmbus_channel_message_header *)msg->u.payload;
1134         size = msg->header.payload_size;
1135
1136         if (hdr->msgtype >= CHANNELMSG_COUNT) {
1137                 pr_err("Received invalid channel message type %d size %d\n",
1138                            hdr->msgtype, size);
1139                 print_hex_dump_bytes("", DUMP_PREFIX_NONE,
1140                                      (unsigned char *)msg->u.payload, size);
1141                 return;
1142         }
1143
1144         if (channel_message_table[hdr->msgtype].message_handler)
1145                 channel_message_table[hdr->msgtype].message_handler(hdr);
1146         else
1147                 pr_err("Unhandled channel message type %d\n", hdr->msgtype);
1148 }
1149
1150 /*
1151  * vmbus_request_offers - Send a request to get all our pending offers.
1152  */
1153 int vmbus_request_offers(void)
1154 {
1155         struct vmbus_channel_message_header *msg;
1156         struct vmbus_channel_msginfo *msginfo;
1157         int ret;
1158
1159         msginfo = kmalloc(sizeof(*msginfo) +
1160                           sizeof(struct vmbus_channel_message_header),
1161                           GFP_KERNEL);
1162         if (!msginfo)
1163                 return -ENOMEM;
1164
1165         msg = (struct vmbus_channel_message_header *)msginfo->msg;
1166
1167         msg->msgtype = CHANNELMSG_REQUESTOFFERS;
1168
1169
1170         ret = vmbus_post_msg(msg, sizeof(struct vmbus_channel_message_header),
1171                              true);
1172         if (ret != 0) {
1173                 pr_err("Unable to request offers - %d\n", ret);
1174
1175                 goto cleanup;
1176         }
1177
1178 cleanup:
1179         kfree(msginfo);
1180
1181         return ret;
1182 }
1183
1184 /*
1185  * Retrieve the (sub) channel on which to send an outgoing request.
1186  * When a primary channel has multiple sub-channels, we try to
1187  * distribute the load equally amongst all available channels.
1188  */
1189 struct vmbus_channel *vmbus_get_outgoing_channel(struct vmbus_channel *primary)
1190 {
1191         struct list_head *cur, *tmp;
1192         int cur_cpu;
1193         struct vmbus_channel *cur_channel;
1194         struct vmbus_channel *outgoing_channel = primary;
1195         int next_channel;
1196         int i = 1;
1197
1198         if (list_empty(&primary->sc_list))
1199                 return outgoing_channel;
1200
1201         next_channel = primary->next_oc++;
1202
1203         if (next_channel > (primary->num_sc)) {
1204                 primary->next_oc = 0;
1205                 return outgoing_channel;
1206         }
1207
1208         cur_cpu = hv_context.vp_index[get_cpu()];
1209         put_cpu();
1210         list_for_each_safe(cur, tmp, &primary->sc_list) {
1211                 cur_channel = list_entry(cur, struct vmbus_channel, sc_list);
1212                 if (cur_channel->state != CHANNEL_OPENED_STATE)
1213                         continue;
1214
1215                 if (cur_channel->target_vp == cur_cpu)
1216                         return cur_channel;
1217
1218                 if (i == next_channel)
1219                         return cur_channel;
1220
1221                 i++;
1222         }
1223
1224         return outgoing_channel;
1225 }
1226 EXPORT_SYMBOL_GPL(vmbus_get_outgoing_channel);
1227
1228 static void invoke_sc_cb(struct vmbus_channel *primary_channel)
1229 {
1230         struct list_head *cur, *tmp;
1231         struct vmbus_channel *cur_channel;
1232
1233         if (primary_channel->sc_creation_callback == NULL)
1234                 return;
1235
1236         list_for_each_safe(cur, tmp, &primary_channel->sc_list) {
1237                 cur_channel = list_entry(cur, struct vmbus_channel, sc_list);
1238
1239                 primary_channel->sc_creation_callback(cur_channel);
1240         }
1241 }
1242
1243 void vmbus_set_sc_create_callback(struct vmbus_channel *primary_channel,
1244                                 void (*sc_cr_cb)(struct vmbus_channel *new_sc))
1245 {
1246         primary_channel->sc_creation_callback = sc_cr_cb;
1247 }
1248 EXPORT_SYMBOL_GPL(vmbus_set_sc_create_callback);
1249
1250 bool vmbus_are_subchannels_present(struct vmbus_channel *primary)
1251 {
1252         bool ret;
1253
1254         ret = !list_empty(&primary->sc_list);
1255
1256         if (ret) {
1257                 /*
1258                  * Invoke the callback on sub-channel creation.
1259                  * This will present a uniform interface to the
1260                  * clients.
1261                  */
1262                 invoke_sc_cb(primary);
1263         }
1264
1265         return ret;
1266 }
1267 EXPORT_SYMBOL_GPL(vmbus_are_subchannels_present);
1268
1269 void vmbus_set_chn_rescind_callback(struct vmbus_channel *channel,
1270                 void (*chn_rescind_cb)(struct vmbus_channel *))
1271 {
1272         channel->chn_rescind_callback = chn_rescind_cb;
1273 }
1274 EXPORT_SYMBOL_GPL(vmbus_set_chn_rescind_callback);