drivers/hv/vmbus_drv.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (c) 2009, Microsoft Corporation.
   4  *
   5  * Authors:
   6  *   Haiyang Zhang <haiyangz@microsoft.com>
   7  *   Hank Janssen  <hjanssen@microsoft.com>
   8  *   K. Y. Srinivasan <kys@microsoft.com>
   9  */
  10 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  11
  12 #include <linux/init.h>
  13 #include <linux/module.h>
  14 #include <linux/device.h>
  15 #include <linux/interrupt.h>
  16 #include <linux/sysctl.h>
  17 #include <linux/slab.h>
  18 #include <linux/acpi.h>
  19 #include <linux/completion.h>
  20 #include <linux/hyperv.h>
  21 #include <linux/kernel_stat.h>
  22 #include <linux/clockchips.h>
  23 #include <linux/cpu.h>
  24 #include <linux/sched/task_stack.h>
  25
  26 #include <linux/delay.h>
  27 #include <linux/notifier.h>
  28 #include <linux/ptrace.h>
  29 #include <linux/screen_info.h>
  30 #include <linux/kdebug.h>
  31 #include <linux/efi.h>
  32 #include <linux/random.h>
  33 #include <linux/kernel.h>
  34 #include <linux/syscore_ops.h>
  35 #include <clocksource/hyperv_timer.h>
  36 #include "hyperv_vmbus.h"
  37
  38 struct vmbus_dynid {
  39         struct list_head node;
  40         struct hv_vmbus_device_id id;
  41 };
  42
  43 static struct acpi_device  *hv_acpi_dev;
  44
  45 static struct completion probe_event;
  46
  47 static int hyperv_cpuhp_online;
  48
  49 static void *hv_panic_page;
  50
  51 static long __percpu *vmbus_evt;
  52
  53 /* Values parsed from ACPI DSDT */
  54 int vmbus_irq;
  55 int vmbus_interrupt;
  56
  57 /*
  58  * Boolean to control whether to report panic messages over Hyper-V.
  59  *
  60  * It can be set via /proc/sys/kernel/hyperv_record_panic_msg
  61  */
  62 static int sysctl_record_panic_msg = 1;
  63
  64 static int hyperv_report_reg(void)
  65 {
  66         return !sysctl_record_panic_msg || !hv_panic_page;
  67 }
  68
  69 static int hyperv_panic_event(struct notifier_block *nb, unsigned long val,
  70                               void *args)
  71 {
  72         struct pt_regs *regs;
  73
  74         vmbus_initiate_unload(true);
  75
  76         /*
  77          * Hyper-V should be notified only once about a panic.  If we will be
  78          * doing hyperv_report_panic_msg() later with kmsg data, don't do
  79          * the notification here.
  80          */
  81         if (ms_hyperv.misc_features & HV_FEATURE_GUEST_CRASH_MSR_AVAILABLE
  82             && hyperv_report_reg()) {
  83                 regs = current_pt_regs();
  84                 hyperv_report_panic(regs, val, false);
  85         }
  86         return NOTIFY_DONE;
  87 }
  88
  89 static int hyperv_die_event(struct notifier_block *nb, unsigned long val,
  90                             void *args)
  91 {
  92         struct die_args *die = args;
  93         struct pt_regs *regs = die->regs;
  94
  95         /* Don't notify Hyper-V if the die event is other than oops */
  96         if (val != DIE_OOPS)
  97                 return NOTIFY_DONE;
  98
  99         /*
 100          * Hyper-V should be notified only once about a panic.  If we will be
 101          * doing hyperv_report_panic_msg() later with kmsg data, don't do
 102          * the notification here.
 103          */
 104         if (hyperv_report_reg())
 105                 hyperv_report_panic(regs, val, true);
 106         return NOTIFY_DONE;
 107 }
 108
 109 static struct notifier_block hyperv_die_block = {
 110         .notifier_call = hyperv_die_event,
 111 };
 112 static struct notifier_block hyperv_panic_block = {
 113         .notifier_call = hyperv_panic_event,
 114 };
 115
 116 static const char *fb_mmio_name = "fb_range";
 117 static struct resource *fb_mmio;
 118 static struct resource *hyperv_mmio;
 119 static DEFINE_MUTEX(hyperv_mmio_lock);
 120
 121 static int vmbus_exists(void)
 122 {
 123         if (hv_acpi_dev == NULL)
 124                 return -ENODEV;
 125
 126         return 0;
 127 }
 128
 129 static u8 channel_monitor_group(const struct vmbus_channel *channel)
 130 {
 131         return (u8)channel->offermsg.monitorid / 32;
 132 }
 133
 134 static u8 channel_monitor_offset(const struct vmbus_channel *channel)
 135 {
 136         return (u8)channel->offermsg.monitorid % 32;
 137 }
 138
 139 static u32 channel_pending(const struct vmbus_channel *channel,
 140                            const struct hv_monitor_page *monitor_page)
 141 {
 142         u8 monitor_group = channel_monitor_group(channel);
 143
 144         return monitor_page->trigger_group[monitor_group].pending;
 145 }
 146
 147 static u32 channel_latency(const struct vmbus_channel *channel,
 148                            const struct hv_monitor_page *monitor_page)
 149 {
 150         u8 monitor_group = channel_monitor_group(channel);
 151         u8 monitor_offset = channel_monitor_offset(channel);
 152
 153         return monitor_page->latency[monitor_group][monitor_offset];
 154 }
 155
 156 static u32 channel_conn_id(struct vmbus_channel *channel,
 157                            struct hv_monitor_page *monitor_page)
 158 {
 159         u8 monitor_group = channel_monitor_group(channel);
 160         u8 monitor_offset = channel_monitor_offset(channel);
 161
 162         return monitor_page->parameter[monitor_group][monitor_offset].connectionid.u.id;
 163 }
 164
 165 static ssize_t id_show(struct device *dev, struct device_attribute *dev_attr,
 166                        char *buf)
 167 {
 168         struct hv_device *hv_dev = device_to_hv_device(dev);
 169
 170         if (!hv_dev->channel)
 171                 return -ENODEV;
 172         return sprintf(buf, "%d\n", hv_dev->channel->offermsg.child_relid);
 173 }
 174 static DEVICE_ATTR_RO(id);
 175
 176 static ssize_t state_show(struct device *dev, struct device_attribute *dev_attr,
 177                           char *buf)
 178 {
 179         struct hv_device *hv_dev = device_to_hv_device(dev);
 180
 181         if (!hv_dev->channel)
 182                 return -ENODEV;
 183         return sprintf(buf, "%d\n", hv_dev->channel->state);
 184 }
 185 static DEVICE_ATTR_RO(state);
 186
 187 static ssize_t monitor_id_show(struct device *dev,
 188                                struct device_attribute *dev_attr, char *buf)
 189 {
 190         struct hv_device *hv_dev = device_to_hv_device(dev);
 191
 192         if (!hv_dev->channel)
 193                 return -ENODEV;
 194         return sprintf(buf, "%d\n", hv_dev->channel->offermsg.monitorid);
 195 }
 196 static DEVICE_ATTR_RO(monitor_id);
 197
 198 static ssize_t class_id_show(struct device *dev,
 199                                struct device_attribute *dev_attr, char *buf)
 200 {
 201         struct hv_device *hv_dev = device_to_hv_device(dev);
 202
 203         if (!hv_dev->channel)
 204                 return -ENODEV;
 205         return sprintf(buf, "{%pUl}\n",
 206                        &hv_dev->channel->offermsg.offer.if_type);
 207 }
 208 static DEVICE_ATTR_RO(class_id);
 209
 210 static ssize_t device_id_show(struct device *dev,
 211                               struct device_attribute *dev_attr, char *buf)
 212 {
 213         struct hv_device *hv_dev = device_to_hv_device(dev);
 214
 215         if (!hv_dev->channel)
 216                 return -ENODEV;
 217         return sprintf(buf, "{%pUl}\n",
 218                        &hv_dev->channel->offermsg.offer.if_instance);
 219 }
 220 static DEVICE_ATTR_RO(device_id);
 221
 222 static ssize_t modalias_show(struct device *dev,
 223                              struct device_attribute *dev_attr, char *buf)
 224 {
 225         struct hv_device *hv_dev = device_to_hv_device(dev);
 226
 227         return sprintf(buf, "vmbus:%*phN\n", UUID_SIZE, &hv_dev->dev_type);
 228 }
 229 static DEVICE_ATTR_RO(modalias);
 230
 231 #ifdef CONFIG_NUMA
 232 static ssize_t numa_node_show(struct device *dev,
 233                               struct device_attribute *attr, char *buf)
 234 {
 235         struct hv_device *hv_dev = device_to_hv_device(dev);
 236
 237         if (!hv_dev->channel)
 238                 return -ENODEV;
 239
 240         return sprintf(buf, "%d\n", cpu_to_node(hv_dev->channel->target_cpu));
 241 }
 242 static DEVICE_ATTR_RO(numa_node);
 243 #endif
 244
 245 static ssize_t server_monitor_pending_show(struct device *dev,
 246                                            struct device_attribute *dev_attr,
 247                                            char *buf)
 248 {
 249         struct hv_device *hv_dev = device_to_hv_device(dev);
 250
 251         if (!hv_dev->channel)
 252                 return -ENODEV;
 253         return sprintf(buf, "%d\n",
 254                        channel_pending(hv_dev->channel,
 255                                        vmbus_connection.monitor_pages[0]));
 256 }
 257 static DEVICE_ATTR_RO(server_monitor_pending);
 258
 259 static ssize_t client_monitor_pending_show(struct device *dev,
 260                                            struct device_attribute *dev_attr,
 261                                            char *buf)
 262 {
 263         struct hv_device *hv_dev = device_to_hv_device(dev);
 264
 265         if (!hv_dev->channel)
 266                 return -ENODEV;
 267         return sprintf(buf, "%d\n",
 268                        channel_pending(hv_dev->channel,
 269                                        vmbus_connection.monitor_pages[1]));
 270 }
 271 static DEVICE_ATTR_RO(client_monitor_pending);
 272
 273 static ssize_t server_monitor_latency_show(struct device *dev,
 274                                            struct device_attribute *dev_attr,
 275                                            char *buf)
 276 {
 277         struct hv_device *hv_dev = device_to_hv_device(dev);
 278
 279         if (!hv_dev->channel)
 280                 return -ENODEV;
 281         return sprintf(buf, "%d\n",
 282                        channel_latency(hv_dev->channel,
 283                                        vmbus_connection.monitor_pages[0]));
 284 }
 285 static DEVICE_ATTR_RO(server_monitor_latency);
 286
 287 static ssize_t client_monitor_latency_show(struct device *dev,
 288                                            struct device_attribute *dev_attr,
 289                                            char *buf)
 290 {
 291         struct hv_device *hv_dev = device_to_hv_device(dev);
 292
 293         if (!hv_dev->channel)
 294                 return -ENODEV;
 295         return sprintf(buf, "%d\n",
 296                        channel_latency(hv_dev->channel,
 297                                        vmbus_connection.monitor_pages[1]));
 298 }
 299 static DEVICE_ATTR_RO(client_monitor_latency);
 300
 301 static ssize_t server_monitor_conn_id_show(struct device *dev,
 302                                            struct device_attribute *dev_attr,
 303                                            char *buf)
 304 {
 305         struct hv_device *hv_dev = device_to_hv_device(dev);
 306
 307         if (!hv_dev->channel)
 308                 return -ENODEV;
 309         return sprintf(buf, "%d\n",
 310                        channel_conn_id(hv_dev->channel,
 311                                        vmbus_connection.monitor_pages[0]));
 312 }
 313 static DEVICE_ATTR_RO(server_monitor_conn_id);
 314
 315 static ssize_t client_monitor_conn_id_show(struct device *dev,
 316                                            struct device_attribute *dev_attr,
 317                                            char *buf)
 318 {
 319         struct hv_device *hv_dev = device_to_hv_device(dev);
 320
 321         if (!hv_dev->channel)
 322                 return -ENODEV;
 323         return sprintf(buf, "%d\n",
 324                        channel_conn_id(hv_dev->channel,
 325                                        vmbus_connection.monitor_pages[1]));
 326 }
 327 static DEVICE_ATTR_RO(client_monitor_conn_id);
 328
 329 static ssize_t out_intr_mask_show(struct device *dev,
 330                                   struct device_attribute *dev_attr, char *buf)
 331 {
 332         struct hv_device *hv_dev = device_to_hv_device(dev);
 333         struct hv_ring_buffer_debug_info outbound;
 334         int ret;
 335
 336         if (!hv_dev->channel)
 337                 return -ENODEV;
 338
 339         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->outbound,
 340                                           &outbound);
 341         if (ret < 0)
 342                 return ret;
 343
 344         return sprintf(buf, "%d\n", outbound.current_interrupt_mask);
 345 }
 346 static DEVICE_ATTR_RO(out_intr_mask);
 347
 348 static ssize_t out_read_index_show(struct device *dev,
 349                                    struct device_attribute *dev_attr, char *buf)
 350 {
 351         struct hv_device *hv_dev = device_to_hv_device(dev);
 352         struct hv_ring_buffer_debug_info outbound;
 353         int ret;
 354
 355         if (!hv_dev->channel)
 356                 return -ENODEV;
 357
 358         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->outbound,
 359                                           &outbound);
 360         if (ret < 0)
 361                 return ret;
 362         return sprintf(buf, "%d\n", outbound.current_read_index);
 363 }
 364 static DEVICE_ATTR_RO(out_read_index);
 365
 366 static ssize_t out_write_index_show(struct device *dev,
 367                                     struct device_attribute *dev_attr,
 368                                     char *buf)
 369 {
 370         struct hv_device *hv_dev = device_to_hv_device(dev);
 371         struct hv_ring_buffer_debug_info outbound;
 372         int ret;
 373
 374         if (!hv_dev->channel)
 375                 return -ENODEV;
 376
 377         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->outbound,
 378                                           &outbound);
 379         if (ret < 0)
 380                 return ret;
 381         return sprintf(buf, "%d\n", outbound.current_write_index);
 382 }
 383 static DEVICE_ATTR_RO(out_write_index);
 384
 385 static ssize_t out_read_bytes_avail_show(struct device *dev,
 386                                          struct device_attribute *dev_attr,
 387                                          char *buf)
 388 {
 389         struct hv_device *hv_dev = device_to_hv_device(dev);
 390         struct hv_ring_buffer_debug_info outbound;
 391         int ret;
 392
 393         if (!hv_dev->channel)
 394                 return -ENODEV;
 395
 396         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->outbound,
 397                                           &outbound);
 398         if (ret < 0)
 399                 return ret;
 400         return sprintf(buf, "%d\n", outbound.bytes_avail_toread);
 401 }
 402 static DEVICE_ATTR_RO(out_read_bytes_avail);
 403
 404 static ssize_t out_write_bytes_avail_show(struct device *dev,
 405                                           struct device_attribute *dev_attr,
 406                                           char *buf)
 407 {
 408         struct hv_device *hv_dev = device_to_hv_device(dev);
 409         struct hv_ring_buffer_debug_info outbound;
 410         int ret;
 411
 412         if (!hv_dev->channel)
 413                 return -ENODEV;
 414
 415         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->outbound,
 416                                           &outbound);
 417         if (ret < 0)
 418                 return ret;
 419         return sprintf(buf, "%d\n", outbound.bytes_avail_towrite);
 420 }
 421 static DEVICE_ATTR_RO(out_write_bytes_avail);
 422
 423 static ssize_t in_intr_mask_show(struct device *dev,
 424                                  struct device_attribute *dev_attr, char *buf)
 425 {
 426         struct hv_device *hv_dev = device_to_hv_device(dev);
 427         struct hv_ring_buffer_debug_info inbound;
 428         int ret;
 429
 430         if (!hv_dev->channel)
 431                 return -ENODEV;
 432
 433         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->inbound, &inbound);
 434         if (ret < 0)
 435                 return ret;
 436
 437         return sprintf(buf, "%d\n", inbound.current_interrupt_mask);
 438 }
 439 static DEVICE_ATTR_RO(in_intr_mask);
 440
 441 static ssize_t in_read_index_show(struct device *dev,
 442                                   struct device_attribute *dev_attr, char *buf)
 443 {
 444         struct hv_device *hv_dev = device_to_hv_device(dev);
 445         struct hv_ring_buffer_debug_info inbound;
 446         int ret;
 447
 448         if (!hv_dev->channel)
 449                 return -ENODEV;
 450
 451         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->inbound, &inbound);
 452         if (ret < 0)
 453                 return ret;
 454
 455         return sprintf(buf, "%d\n", inbound.current_read_index);
 456 }
 457 static DEVICE_ATTR_RO(in_read_index);
 458
 459 static ssize_t in_write_index_show(struct device *dev,
 460                                    struct device_attribute *dev_attr, char *buf)
 461 {
 462         struct hv_device *hv_dev = device_to_hv_device(dev);
 463         struct hv_ring_buffer_debug_info inbound;
 464         int ret;
 465
 466         if (!hv_dev->channel)
 467                 return -ENODEV;
 468
 469         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->inbound, &inbound);
 470         if (ret < 0)
 471                 return ret;
 472
 473         return sprintf(buf, "%d\n", inbound.current_write_index);
 474 }
 475 static DEVICE_ATTR_RO(in_write_index);
 476
 477 static ssize_t in_read_bytes_avail_show(struct device *dev,
 478                                         struct device_attribute *dev_attr,
 479                                         char *buf)
 480 {
 481         struct hv_device *hv_dev = device_to_hv_device(dev);
 482         struct hv_ring_buffer_debug_info inbound;
 483         int ret;
 484
 485         if (!hv_dev->channel)
 486                 return -ENODEV;
 487
 488         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->inbound, &inbound);
 489         if (ret < 0)
 490                 return ret;
 491
 492         return sprintf(buf, "%d\n", inbound.bytes_avail_toread);
 493 }
 494 static DEVICE_ATTR_RO(in_read_bytes_avail);
 495
 496 static ssize_t in_write_bytes_avail_show(struct device *dev,
 497                                          struct device_attribute *dev_attr,
 498                                          char *buf)
 499 {
 500         struct hv_device *hv_dev = device_to_hv_device(dev);
 501         struct hv_ring_buffer_debug_info inbound;
 502         int ret;
 503
 504         if (!hv_dev->channel)
 505                 return -ENODEV;
 506
 507         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->inbound, &inbound);
 508         if (ret < 0)
 509                 return ret;
 510
 511         return sprintf(buf, "%d\n", inbound.bytes_avail_towrite);
 512 }
 513 static DEVICE_ATTR_RO(in_write_bytes_avail);
 514
 515 static ssize_t channel_vp_mapping_show(struct device *dev,
 516                                        struct device_attribute *dev_attr,
 517                                        char *buf)
 518 {
 519         struct hv_device *hv_dev = device_to_hv_device(dev);
 520         struct vmbus_channel *channel = hv_dev->channel, *cur_sc;
 521         int buf_size = PAGE_SIZE, n_written, tot_written;
 522         struct list_head *cur;
 523
 524         if (!channel)
 525                 return -ENODEV;
 526
 527         mutex_lock(&vmbus_connection.channel_mutex);
 528
 529         tot_written = snprintf(buf, buf_size, "%u:%u\n",
 530                 channel->offermsg.child_relid, channel->target_cpu);
 531
 532         list_for_each(cur, &channel->sc_list) {
 533                 if (tot_written >= buf_size - 1)
 534                         break;
 535
 536                 cur_sc = list_entry(cur, struct vmbus_channel, sc_list);
 537                 n_written = scnprintf(buf + tot_written,
 538                                      buf_size - tot_written,
 539                                      "%u:%u\n",
 540                                      cur_sc->offermsg.child_relid,
 541                                      cur_sc->target_cpu);
 542                 tot_written += n_written;
 543         }
 544
 545         mutex_unlock(&vmbus_connection.channel_mutex);
 546
 547         return tot_written;
 548 }
 549 static DEVICE_ATTR_RO(channel_vp_mapping);
 550
 551 static ssize_t vendor_show(struct device *dev,
 552                            struct device_attribute *dev_attr,
 553                            char *buf)
 554 {
 555         struct hv_device *hv_dev = device_to_hv_device(dev);
 556
 557         return sprintf(buf, "0x%x\n", hv_dev->vendor_id);
 558 }
 559 static DEVICE_ATTR_RO(vendor);
 560
 561 static ssize_t device_show(struct device *dev,
 562                            struct device_attribute *dev_attr,
 563                            char *buf)
 564 {
 565         struct hv_device *hv_dev = device_to_hv_device(dev);
 566
 567         return sprintf(buf, "0x%x\n", hv_dev->device_id);
 568 }
 569 static DEVICE_ATTR_RO(device);
 570
 571 static ssize_t driver_override_store(struct device *dev,
 572                                      struct device_attribute *attr,
 573                                      const char *buf, size_t count)
 574 {
 575         struct hv_device *hv_dev = device_to_hv_device(dev);
 576         char *driver_override, *old, *cp;
 577
 578         /* We need to keep extra room for a newline */
 579         if (count >= (PAGE_SIZE - 1))
 580                 return -EINVAL;
 581
 582         driver_override = kstrndup(buf, count, GFP_KERNEL);
 583         if (!driver_override)
 584                 return -ENOMEM;
 585
 586         cp = strchr(driver_override, '\n');
 587         if (cp)
 588                 *cp = '\0';
 589
 590         device_lock(dev);
 591         old = hv_dev->driver_override;
 592         if (strlen(driver_override)) {
 593                 hv_dev->driver_override = driver_override;
 594         } else {
 595                 kfree(driver_override);
 596                 hv_dev->driver_override = NULL;
 597         }
 598         device_unlock(dev);
 599
 600         kfree(old);
 601
 602         return count;
 603 }
 604
 605 static ssize_t driver_override_show(struct device *dev,
 606                                     struct device_attribute *attr, char *buf)
 607 {
 608         struct hv_device *hv_dev = device_to_hv_device(dev);
 609         ssize_t len;
 610
 611         device_lock(dev);
 612         len = snprintf(buf, PAGE_SIZE, "%s\n", hv_dev->driver_override);
 613         device_unlock(dev);
 614
 615         return len;
 616 }
 617 static DEVICE_ATTR_RW(driver_override);
 618
 619 /* Set up per device attributes in /sys/bus/vmbus/devices/<bus device> */
 620 static struct attribute *vmbus_dev_attrs[] = {
 621         &dev_attr_id.attr,
 622         &dev_attr_state.attr,
 623         &dev_attr_monitor_id.attr,
 624         &dev_attr_class_id.attr,
 625         &dev_attr_device_id.attr,
 626         &dev_attr_modalias.attr,
 627 #ifdef CONFIG_NUMA
 628         &dev_attr_numa_node.attr,
 629 #endif
 630         &dev_attr_server_monitor_pending.attr,
 631         &dev_attr_client_monitor_pending.attr,
 632         &dev_attr_server_monitor_latency.attr,
 633         &dev_attr_client_monitor_latency.attr,
 634         &dev_attr_server_monitor_conn_id.attr,
 635         &dev_attr_client_monitor_conn_id.attr,
 636         &dev_attr_out_intr_mask.attr,
 637         &dev_attr_out_read_index.attr,
 638         &dev_attr_out_write_index.attr,
 639         &dev_attr_out_read_bytes_avail.attr,
 640         &dev_attr_out_write_bytes_avail.attr,
 641         &dev_attr_in_intr_mask.attr,
 642         &dev_attr_in_read_index.attr,
 643         &dev_attr_in_write_index.attr,
 644         &dev_attr_in_read_bytes_avail.attr,
 645         &dev_attr_in_write_bytes_avail.attr,
 646         &dev_attr_channel_vp_mapping.attr,
 647         &dev_attr_vendor.attr,
 648         &dev_attr_device.attr,
 649         &dev_attr_driver_override.attr,
 650         NULL,
 651 };
 652
 653 /*
 654  * Device-level attribute_group callback function. Returns the permission for
 655  * each attribute, and returns 0 if an attribute is not visible.
 656  */
 657 static umode_t vmbus_dev_attr_is_visible(struct kobject *kobj,
 658                                          struct attribute *attr, int idx)
 659 {
 660         struct device *dev = kobj_to_dev(kobj);
 661         const struct hv_device *hv_dev = device_to_hv_device(dev);
 662
 663         /* Hide the monitor attributes if the monitor mechanism is not used. */
 664         if (!hv_dev->channel->offermsg.monitor_allocated &&
 665             (attr == &dev_attr_monitor_id.attr ||
 666              attr == &dev_attr_server_monitor_pending.attr ||
 667              attr == &dev_attr_client_monitor_pending.attr ||
 668              attr == &dev_attr_server_monitor_latency.attr ||
 669              attr == &dev_attr_client_monitor_latency.attr ||
 670              attr == &dev_attr_server_monitor_conn_id.attr ||
 671              attr == &dev_attr_client_monitor_conn_id.attr))
 672                 return 0;
 673
 674         return attr->mode;
 675 }
 676
 677 static const struct attribute_group vmbus_dev_group = {
 678         .attrs = vmbus_dev_attrs,
 679         .is_visible = vmbus_dev_attr_is_visible
 680 };
 681 __ATTRIBUTE_GROUPS(vmbus_dev);
 682
 683 /* Set up the attribute for /sys/bus/vmbus/hibernation */
 684 static ssize_t hibernation_show(struct bus_type *bus, char *buf)
 685 {
 686         return sprintf(buf, "%d\n", !!hv_is_hibernation_supported());
 687 }
 688
 689 static BUS_ATTR_RO(hibernation);
 690
 691 static struct attribute *vmbus_bus_attrs[] = {
 692         &bus_attr_hibernation.attr,
 693         NULL,
 694 };
 695 static const struct attribute_group vmbus_bus_group = {
 696         .attrs = vmbus_bus_attrs,
 697 };
 698 __ATTRIBUTE_GROUPS(vmbus_bus);
 699
 700 /*
 701  * vmbus_uevent - add uevent for our device
 702  *
 703  * This routine is invoked when a device is added or removed on the vmbus to
 704  * generate a uevent to udev in the userspace. The udev will then look at its
 705  * rule and the uevent generated here to load the appropriate driver
 706  *
 707  * The alias string will be of the form vmbus:guid where guid is the string
 708  * representation of the device guid (each byte of the guid will be
 709  * represented with two hex characters.
 710  */
 711 static int vmbus_uevent(struct device *device, struct kobj_uevent_env *env)
 712 {
 713         struct hv_device *dev = device_to_hv_device(device);
 714         const char *format = "MODALIAS=vmbus:%*phN";
 715
 716         return add_uevent_var(env, format, UUID_SIZE, &dev->dev_type);
 717 }
 718
 719 static const struct hv_vmbus_device_id *
 720 hv_vmbus_dev_match(const struct hv_vmbus_device_id *id, const guid_t *guid)
 721 {
 722         if (id == NULL)
 723                 return NULL; /* empty device table */
 724
 725         for (; !guid_is_null(&id->guid); id++)
 726                 if (guid_equal(&id->guid, guid))
 727                         return id;
 728
 729         return NULL;
 730 }
 731
 732 static const struct hv_vmbus_device_id *
 733 hv_vmbus_dynid_match(struct hv_driver *drv, const guid_t *guid)
 734 {
 735         const struct hv_vmbus_device_id *id = NULL;
 736         struct vmbus_dynid *dynid;
 737
 738         spin_lock(&drv->dynids.lock);
 739         list_for_each_entry(dynid, &drv->dynids.list, node) {
 740                 if (guid_equal(&dynid->id.guid, guid)) {
 741                         id = &dynid->id;
 742                         break;
 743                 }
 744         }
 745         spin_unlock(&drv->dynids.lock);
 746
 747         return id;
 748 }
 749
 750 static const struct hv_vmbus_device_id vmbus_device_null;
 751
 752 /*
 753  * Return a matching hv_vmbus_device_id pointer.
 754  * If there is no match, return NULL.
 755  */
 756 static const struct hv_vmbus_device_id *hv_vmbus_get_id(struct hv_driver *drv,
 757                                                         struct hv_device *dev)
 758 {
 759         const guid_t *guid = &dev->dev_type;
 760         const struct hv_vmbus_device_id *id;
 761
 762         /* When driver_override is set, only bind to the matching driver */
 763         if (dev->driver_override && strcmp(dev->driver_override, drv->name))
 764                 return NULL;
 765
 766         /* Look at the dynamic ids first, before the static ones */
 767         id = hv_vmbus_dynid_match(drv, guid);
 768         if (!id)
 769                 id = hv_vmbus_dev_match(drv->id_table, guid);
 770
 771         /* driver_override will always match, send a dummy id */
 772         if (!id && dev->driver_override)
 773                 id = &vmbus_device_null;
 774
 775         return id;
 776 }
 777
 778 /* vmbus_add_dynid - add a new device ID to this driver and re-probe devices */
 779 static int vmbus_add_dynid(struct hv_driver *drv, guid_t *guid)
 780 {
 781         struct vmbus_dynid *dynid;
 782
 783         dynid = kzalloc(sizeof(*dynid), GFP_KERNEL);
 784         if (!dynid)
 785                 return -ENOMEM;
 786
 787         dynid->id.guid = *guid;
 788
 789         spin_lock(&drv->dynids.lock);
 790         list_add_tail(&dynid->node, &drv->dynids.list);
 791         spin_unlock(&drv->dynids.lock);
 792
 793         return driver_attach(&drv->driver);
 794 }
 795
 796 static void vmbus_free_dynids(struct hv_driver *drv)
 797 {
 798         struct vmbus_dynid *dynid, *n;
 799
 800         spin_lock(&drv->dynids.lock);
 801         list_for_each_entry_safe(dynid, n, &drv->dynids.list, node) {
 802                 list_del(&dynid->node);
 803                 kfree(dynid);
 804         }
 805         spin_unlock(&drv->dynids.lock);
 806 }
 807
 808 /*
 809  * store_new_id - sysfs frontend to vmbus_add_dynid()
 810  *
 811  * Allow GUIDs to be added to an existing driver via sysfs.
 812  */
 813 static ssize_t new_id_store(struct device_driver *driver, const char *buf,
 814                             size_t count)
 815 {
 816         struct hv_driver *drv = drv_to_hv_drv(driver);
 817         guid_t guid;
 818         ssize_t retval;
 819
 820         retval = guid_parse(buf, &guid);
 821         if (retval)
 822                 return retval;
 823
 824         if (hv_vmbus_dynid_match(drv, &guid))
 825                 return -EEXIST;
 826
 827         retval = vmbus_add_dynid(drv, &guid);
 828         if (retval)
 829                 return retval;
 830         return count;
 831 }
 832 static DRIVER_ATTR_WO(new_id);
 833
 834 /*
 835  * store_remove_id - remove a PCI device ID from this driver
 836  *
 837  * Removes a dynamic pci device ID to this driver.
 838  */
 839 static ssize_t remove_id_store(struct device_driver *driver, const char *buf,
 840                                size_t count)
 841 {
 842         struct hv_driver *drv = drv_to_hv_drv(driver);
 843         struct vmbus_dynid *dynid, *n;
 844         guid_t guid;
 845         ssize_t retval;
 846
 847         retval = guid_parse(buf, &guid);
 848         if (retval)
 849                 return retval;
 850
 851         retval = -ENODEV;
 852         spin_lock(&drv->dynids.lock);
 853         list_for_each_entry_safe(dynid, n, &drv->dynids.list, node) {
 854                 struct hv_vmbus_device_id *id = &dynid->id;
 855
 856                 if (guid_equal(&id->guid, &guid)) {
 857                         list_del(&dynid->node);
 858                         kfree(dynid);
 859                         retval = count;
 860                         break;
 861                 }
 862         }
 863         spin_unlock(&drv->dynids.lock);
 864
 865         return retval;
 866 }
 867 static DRIVER_ATTR_WO(remove_id);
 868
 869 static struct attribute *vmbus_drv_attrs[] = {
 870         &driver_attr_new_id.attr,
 871         &driver_attr_remove_id.attr,
 872         NULL,
 873 };
 874 ATTRIBUTE_GROUPS(vmbus_drv);
 875
 876
 877 /*
 878  * vmbus_match - Attempt to match the specified device to the specified driver
 879  */
 880 static int vmbus_match(struct device *device, struct device_driver *driver)
 881 {
 882         struct hv_driver *drv = drv_to_hv_drv(driver);
 883         struct hv_device *hv_dev = device_to_hv_device(device);
 884
 885         /* The hv_sock driver handles all hv_sock offers. */
 886         if (is_hvsock_channel(hv_dev->channel))
 887                 return drv->hvsock;
 888
 889         if (hv_vmbus_get_id(drv, hv_dev))
 890                 return 1;
 891
 892         return 0;
 893 }
 894
 895 /*
 896  * vmbus_probe - Add the new vmbus's child device
 897  */
 898 static int vmbus_probe(struct device *child_device)
 899 {
 900         int ret = 0;
 901         struct hv_driver *drv =
 902                         drv_to_hv_drv(child_device->driver);
 903         struct hv_device *dev = device_to_hv_device(child_device);
 904         const struct hv_vmbus_device_id *dev_id;
 905
 906         dev_id = hv_vmbus_get_id(drv, dev);
 907         if (drv->probe) {
 908                 ret = drv->probe(dev, dev_id);
 909                 if (ret != 0)
 910                         pr_err("probe failed for device %s (%d)\n",
 911                                dev_name(child_device), ret);
 912
 913         } else {
 914                 pr_err("probe not set for driver %s\n",
 915                        dev_name(child_device));
 916                 ret = -ENODEV;
 917         }
 918         return ret;
 919 }
 920
 921 /*
 922  * vmbus_remove - Remove a vmbus device
 923  */
 924 static int vmbus_remove(struct device *child_device)
 925 {
 926         struct hv_driver *drv;
 927         struct hv_device *dev = device_to_hv_device(child_device);
 928
 929         if (child_device->driver) {
 930                 drv = drv_to_hv_drv(child_device->driver);
 931                 if (drv->remove)
 932                         drv->remove(dev);
 933         }
 934
 935         return 0;
 936 }
 937
 938
 939 /*
 940  * vmbus_shutdown - Shutdown a vmbus device
 941  */
 942 static void vmbus_shutdown(struct device *child_device)
 943 {
 944         struct hv_driver *drv;
 945         struct hv_device *dev = device_to_hv_device(child_device);
 946
 947
 948         /* The device may not be attached yet */
 949         if (!child_device->driver)
 950                 return;
 951
 952         drv = drv_to_hv_drv(child_device->driver);
 953
 954         if (drv->shutdown)
 955                 drv->shutdown(dev);
 956 }
 957
 958 #ifdef CONFIG_PM_SLEEP
 959 /*
 960  * vmbus_suspend - Suspend a vmbus device
 961  */
 962 static int vmbus_suspend(struct device *child_device)
 963 {
 964         struct hv_driver *drv;
 965         struct hv_device *dev = device_to_hv_device(child_device);
 966
 967         /* The device may not be attached yet */
 968         if (!child_device->driver)
 969                 return 0;
 970
 971         drv = drv_to_hv_drv(child_device->driver);
 972         if (!drv->suspend)
 973                 return -EOPNOTSUPP;
 974
 975         return drv->suspend(dev);
 976 }
 977
 978 /*
 979  * vmbus_resume - Resume a vmbus device
 980  */
 981 static int vmbus_resume(struct device *child_device)
 982 {
 983         struct hv_driver *drv;
 984         struct hv_device *dev = device_to_hv_device(child_device);
 985
 986         /* The device may not be attached yet */
 987         if (!child_device->driver)
 988                 return 0;
 989
 990         drv = drv_to_hv_drv(child_device->driver);
 991         if (!drv->resume)
 992                 return -EOPNOTSUPP;
 993
 994         return drv->resume(dev);
 995 }
 996 #else
 997 #define vmbus_suspend NULL
 998 #define vmbus_resume NULL
 999 #endif /* CONFIG_PM_SLEEP */
1000
1001 /*
1002  * vmbus_device_release - Final callback release of the vmbus child device
1003  */
1004 static void vmbus_device_release(struct device *device)
1005 {
1006         struct hv_device *hv_dev = device_to_hv_device(device);
1007         struct vmbus_channel *channel = hv_dev->channel;
1008
1009         hv_debug_rm_dev_dir(hv_dev);
1010
1011         mutex_lock(&vmbus_connection.channel_mutex);
1012         hv_process_channel_removal(channel);
1013         mutex_unlock(&vmbus_connection.channel_mutex);
1014         kfree(hv_dev);
1015 }
1016
1017 /*
1018  * Note: we must use the "noirq" ops: see the comment before vmbus_bus_pm.
1019  *
1020  * suspend_noirq/resume_noirq are set to NULL to support Suspend-to-Idle: we
1021  * shouldn't suspend the vmbus devices upon Suspend-to-Idle, otherwise there
1022  * is no way to wake up a Generation-2 VM.
1023  *
1024  * The other 4 ops are for hibernation.
1025  */
1026
1027 static const struct dev_pm_ops vmbus_pm = {
1028         .suspend_noirq  = NULL,
1029         .resume_noirq   = NULL,
1030         .freeze_noirq   = vmbus_suspend,
1031         .thaw_noirq     = vmbus_resume,
1032         .poweroff_noirq = vmbus_suspend,
1033         .restore_noirq  = vmbus_resume,
1034 };
1035
1036 /* The one and only one */
1037 static struct bus_type  hv_bus = {
1038         .name =         "vmbus",
1039         .match =                vmbus_match,
1040         .shutdown =             vmbus_shutdown,
1041         .remove =               vmbus_remove,
1042         .probe =                vmbus_probe,
1043         .uevent =               vmbus_uevent,
1044         .dev_groups =           vmbus_dev_groups,
1045         .drv_groups =           vmbus_drv_groups,
1046         .bus_groups =           vmbus_bus_groups,
1047         .pm =                   &vmbus_pm,
1048 };
1049
1050 struct onmessage_work_context {
1051         struct work_struct work;
1052         struct {
1053                 struct hv_message_header header;
1054                 u8 payload[];
1055         } msg;
1056 };
1057
1058 static void vmbus_onmessage_work(struct work_struct *work)
1059 {
1060         struct onmessage_work_context *ctx;
1061
1062         /* Do not process messages if we're in DISCONNECTED state */
1063         if (vmbus_connection.conn_state == DISCONNECTED)
1064                 return;
1065
1066         ctx = container_of(work, struct onmessage_work_context,
1067                            work);
1068         vmbus_onmessage((struct vmbus_channel_message_header *)
1069                         &ctx->msg.payload);
1070         kfree(ctx);
1071 }
1072
1073 void vmbus_on_msg_dpc(unsigned long data)
1074 {
1075         struct hv_per_cpu_context *hv_cpu = (void *)data;
1076         void *page_addr = hv_cpu->synic_message_page;
1077         struct hv_message msg_copy, *msg = (struct hv_message *)page_addr +
1078                                   VMBUS_MESSAGE_SINT;
1079         struct vmbus_channel_message_header *hdr;
1080         enum vmbus_channel_message_type msgtype;
1081         const struct vmbus_channel_message_table_entry *entry;
1082         struct onmessage_work_context *ctx;
1083         __u8 payload_size;
1084         u32 message_type;
1085
1086         /*
1087          * 'enum vmbus_channel_message_type' is supposed to always be 'u32' as
1088          * it is being used in 'struct vmbus_channel_message_header' definition
1089          * which is supposed to match hypervisor ABI.
1090          */
1091         BUILD_BUG_ON(sizeof(enum vmbus_channel_message_type) != sizeof(u32));
1092
1093         /*
1094          * Since the message is in memory shared with the host, an erroneous or
1095          * malicious Hyper-V could modify the message while vmbus_on_msg_dpc()
1096          * or individual message handlers are executing; to prevent this, copy
1097          * the message into private memory.
1098          */
1099         memcpy(&msg_copy, msg, sizeof(struct hv_message));
1100
1101         message_type = msg_copy.header.message_type;
1102         if (message_type == HVMSG_NONE)
1103                 /* no msg */
1104                 return;
1105
1106         hdr = (struct vmbus_channel_message_header *)msg_copy.u.payload;
1107         msgtype = hdr->msgtype;
1108
1109         trace_vmbus_on_msg_dpc(hdr);
1110
1111         if (msgtype >= CHANNELMSG_COUNT) {
1112                 WARN_ONCE(1, "unknown msgtype=%d\n", msgtype);
1113                 goto msg_handled;
1114         }
1115
1116         payload_size = msg_copy.header.payload_size;
1117         if (payload_size > HV_MESSAGE_PAYLOAD_BYTE_COUNT) {
1118                 WARN_ONCE(1, "payload size is too large (%d)\n", payload_size);
1119                 goto msg_handled;
1120         }
1121
1122         entry = &channel_message_table[msgtype];
1123
1124         if (!entry->message_handler)
1125                 goto msg_handled;
1126
1127         if (payload_size < entry->min_payload_len) {
1128                 WARN_ONCE(1, "message too short: msgtype=%d len=%d\n", msgtype, payload_size);
1129                 goto msg_handled;
1130         }
1131
1132         if (entry->handler_type == VMHT_BLOCKING) {
1133                 ctx = kmalloc(sizeof(*ctx) + payload_size, GFP_ATOMIC);
1134                 if (ctx == NULL)
1135                         return;
1136
1137                 INIT_WORK(&ctx->work, vmbus_onmessage_work);
1138                 memcpy(&ctx->msg, &msg_copy, sizeof(msg->header) + payload_size);
1139
1140                 /*
1141                  * The host can generate a rescind message while we
1142                  * may still be handling the original offer. We deal with
1143                  * this condition by relying on the synchronization provided
1144                  * by offer_in_progress and by channel_mutex.  See also the
1145                  * inline comments in vmbus_onoffer_rescind().
1146                  */
1147                 switch (msgtype) {
1148                 case CHANNELMSG_RESCIND_CHANNELOFFER:
1149                         /*
1150                          * If we are handling the rescind message;
1151                          * schedule the work on the global work queue.
1152                          *
1153                          * The OFFER message and the RESCIND message should
1154                          * not be handled by the same serialized work queue,
1155                          * because the OFFER handler may call vmbus_open(),
1156                          * which tries to open the channel by sending an
1157                          * OPEN_CHANNEL message to the host and waits for
1158                          * the host's response; however, if the host has
1159                          * rescinded the channel before it receives the
1160                          * OPEN_CHANNEL message, the host just silently
1161                          * ignores the OPEN_CHANNEL message; as a result,
1162                          * the guest's OFFER handler hangs for ever, if we
1163                          * handle the RESCIND message in the same serialized
1164                          * work queue: the RESCIND handler can not start to
1165                          * run before the OFFER handler finishes.
1166                          */
1167                         schedule_work(&ctx->work);
1168                         break;
1169
1170                 case CHANNELMSG_OFFERCHANNEL:
1171                         /*
1172                          * The host sends the offer message of a given channel
1173                          * before sending the rescind message of the same
1174                          * channel.  These messages are sent to the guest's
1175                          * connect CPU; the guest then starts processing them
1176                          * in the tasklet handler on this CPU:
1177                          *
1178                          * VMBUS_CONNECT_CPU
1179                          *
1180                          * [vmbus_on_msg_dpc()]
1181                          * atomic_inc()  // CHANNELMSG_OFFERCHANNEL
1182                          * queue_work()
1183                          * ...
1184                          * [vmbus_on_msg_dpc()]
1185                          * schedule_work()  // CHANNELMSG_RESCIND_CHANNELOFFER
1186                          *
1187                          * We rely on the memory-ordering properties of the
1188                          * queue_work() and schedule_work() primitives, which
1189                          * guarantee that the atomic increment will be visible
1190                          * to the CPUs which will execute the offer & rescind
1191                          * works by the time these works will start execution.
1192                          */
1193                         atomic_inc(&vmbus_connection.offer_in_progress);
1194                         fallthrough;
1195
1196                 default:
1197                         queue_work(vmbus_connection.work_queue, &ctx->work);
1198                 }
1199         } else
1200                 entry->message_handler(hdr);
1201
1202 msg_handled:
1203         vmbus_signal_eom(msg, message_type);
1204 }
1205
1206 #ifdef CONFIG_PM_SLEEP
1207 /*
1208  * Fake RESCIND_CHANNEL messages to clean up hv_sock channels by force for
1209  * hibernation, because hv_sock connections can not persist across hibernation.
1210  */
1211 static void vmbus_force_channel_rescinded(struct vmbus_channel *channel)
1212 {
1213         struct onmessage_work_context *ctx;
1214         struct vmbus_channel_rescind_offer *rescind;
1215
1216         WARN_ON(!is_hvsock_channel(channel));
1217
1218         /*
1219          * Allocation size is small and the allocation should really not fail,
1220          * otherwise the state of the hv_sock connections ends up in limbo.
1221          */
1222         ctx = kzalloc(sizeof(*ctx) + sizeof(*rescind),
1223                       GFP_KERNEL | __GFP_NOFAIL);
1224
1225         /*
1226          * So far, these are not really used by Linux. Just set them to the
1227          * reasonable values conforming to the definitions of the fields.
1228          */
1229         ctx->msg.header.message_type = 1;
1230         ctx->msg.header.payload_size = sizeof(*rescind);
1231
1232         /* These values are actually used by Linux. */
1233         rescind = (struct vmbus_channel_rescind_offer *)ctx->msg.payload;
1234         rescind->header.msgtype = CHANNELMSG_RESCIND_CHANNELOFFER;
1235         rescind->child_relid = channel->offermsg.child_relid;
1236
1237         INIT_WORK(&ctx->work, vmbus_onmessage_work);
1238
1239         queue_work(vmbus_connection.work_queue, &ctx->work);
1240 }
1241 #endif /* CONFIG_PM_SLEEP */
1242
1243 /*
1244  * Schedule all channels with events pending
1245  */
1246 static void vmbus_chan_sched(struct hv_per_cpu_context *hv_cpu)
1247 {
1248         unsigned long *recv_int_page;
1249         u32 maxbits, relid;
1250
1251         if (vmbus_proto_version < VERSION_WIN8) {
1252                 maxbits = MAX_NUM_CHANNELS_SUPPORTED;
1253                 recv_int_page = vmbus_connection.recv_int_page;
1254         } else {
1255                 /*
1256                  * When the host is win8 and beyond, the event page
1257                  * can be directly checked to get the id of the channel
1258                  * that has the interrupt pending.
1259                  */
1260                 void *page_addr = hv_cpu->synic_event_page;
1261                 union hv_synic_event_flags *event
1262                         = (union hv_synic_event_flags *)page_addr +
1263                                                  VMBUS_MESSAGE_SINT;
1264
1265                 maxbits = HV_EVENT_FLAGS_COUNT;
1266                 recv_int_page = event->flags;
1267         }
1268
1269         if (unlikely(!recv_int_page))
1270                 return;
1271
1272         for_each_set_bit(relid, recv_int_page, maxbits) {
1273                 void (*callback_fn)(void *context);
1274                 struct vmbus_channel *channel;
1275
1276                 if (!sync_test_and_clear_bit(relid, recv_int_page))
1277                         continue;
1278
1279                 /* Special case - vmbus channel protocol msg */
1280                 if (relid == 0)
1281                         continue;
1282
1283                 /*
1284                  * Pairs with the kfree_rcu() in vmbus_chan_release().
1285                  * Guarantees that the channel data structure doesn't
1286                  * get freed while the channel pointer below is being
1287                  * dereferenced.
1288                  */
1289                 rcu_read_lock();
1290
1291                 /* Find channel based on relid */
1292                 channel = relid2channel(relid);
1293                 if (channel == NULL)
1294                         goto sched_unlock_rcu;
1295
1296                 if (channel->rescind)
1297                         goto sched_unlock_rcu;
1298
1299                 /*
1300                  * Make sure that the ring buffer data structure doesn't get
1301                  * freed while we dereference the ring buffer pointer.  Test
1302                  * for the channel's onchannel_callback being NULL within a
1303                  * sched_lock critical section.  See also the inline comments
1304                  * in vmbus_reset_channel_cb().
1305                  */
1306                 spin_lock(&channel->sched_lock);
1307
1308                 callback_fn = channel->onchannel_callback;
1309                 if (unlikely(callback_fn == NULL))
1310                         goto sched_unlock;
1311
1312                 trace_vmbus_chan_sched(channel);
1313
1314                 ++channel->interrupts;
1315
1316                 switch (channel->callback_mode) {
1317                 case HV_CALL_ISR:
1318                         (*callback_fn)(channel->channel_callback_context);
1319                         break;
1320
1321                 case HV_CALL_BATCHED:
1322                         hv_begin_read(&channel->inbound);
1323                         fallthrough;
1324                 case HV_CALL_DIRECT:
1325                         tasklet_schedule(&channel->callback_event);
1326                 }
1327
1328 sched_unlock:
1329                 spin_unlock(&channel->sched_lock);
1330 sched_unlock_rcu:
1331                 rcu_read_unlock();
1332         }
1333 }
1334
1335 static void vmbus_isr(void)
1336 {
1337         struct hv_per_cpu_context *hv_cpu
1338                 = this_cpu_ptr(hv_context.cpu_context);
1339         void *page_addr = hv_cpu->synic_event_page;
1340         struct hv_message *msg;
1341         union hv_synic_event_flags *event;
1342         bool handled = false;
1343
1344         if (unlikely(page_addr == NULL))
1345                 return;
1346
1347         event = (union hv_synic_event_flags *)page_addr +
1348                                          VMBUS_MESSAGE_SINT;
1349         /*
1350          * Check for events before checking for messages. This is the order
1351          * in which events and messages are checked in Windows guests on
1352          * Hyper-V, and the Windows team suggested we do the same.
1353          */
1354
1355         if ((vmbus_proto_version == VERSION_WS2008) ||
1356                 (vmbus_proto_version == VERSION_WIN7)) {
1357
1358                 /* Since we are a child, we only need to check bit 0 */
1359                 if (sync_test_and_clear_bit(0, event->flags))
1360                         handled = true;
1361         } else {
1362                 /*
1363                  * Our host is win8 or above. The signaling mechanism
1364                  * has changed and we can directly look at the event page.
1365                  * If bit n is set then we have an interrup on the channel
1366                  * whose id is n.
1367                  */
1368                 handled = true;
1369         }
1370
1371         if (handled)
1372                 vmbus_chan_sched(hv_cpu);
1373
1374         page_addr = hv_cpu->synic_message_page;
1375         msg = (struct hv_message *)page_addr + VMBUS_MESSAGE_SINT;
1376
1377         /* Check if there are actual msgs to be processed */
1378         if (msg->header.message_type != HVMSG_NONE) {
1379                 if (msg->header.message_type == HVMSG_TIMER_EXPIRED) {
1380                         hv_stimer0_isr();
1381                         vmbus_signal_eom(msg, HVMSG_TIMER_EXPIRED);
1382                 } else
1383                         tasklet_schedule(&hv_cpu->msg_dpc);
1384         }
1385
1386         add_interrupt_randomness(vmbus_interrupt, 0);
1387 }
1388
1389 static irqreturn_t vmbus_percpu_isr(int irq, void *dev_id)
1390 {
1391         vmbus_isr();
1392         return IRQ_HANDLED;
1393 }
1394
1395 /*
1396  * Callback from kmsg_dump. Grab as much as possible from the end of the kmsg
1397  * buffer and call into Hyper-V to transfer the data.
1398  */
1399 static void hv_kmsg_dump(struct kmsg_dumper *dumper,
1400                          enum kmsg_dump_reason reason)
1401 {
1402         struct kmsg_dump_iter iter;
1403         size_t bytes_written;
1404
1405         /* We are only interested in panics. */
1406         if ((reason != KMSG_DUMP_PANIC) || (!sysctl_record_panic_msg))
1407                 return;
1408
1409         /*
1410          * Write dump contents to the page. No need to synchronize; panic should
1411          * be single-threaded.
1412          */
1413         kmsg_dump_rewind(&iter);
1414         kmsg_dump_get_buffer(&iter, false, hv_panic_page, HV_HYP_PAGE_SIZE,
1415                              &bytes_written);
1416         if (!bytes_written)
1417                 return;
1418         /*
1419          * P3 to contain the physical address of the panic page & P4 to
1420          * contain the size of the panic data in that page. Rest of the
1421          * registers are no-op when the NOTIFY_MSG flag is set.
1422          */
1423         hv_set_register(HV_REGISTER_CRASH_P0, 0);
1424         hv_set_register(HV_REGISTER_CRASH_P1, 0);
1425         hv_set_register(HV_REGISTER_CRASH_P2, 0);
1426         hv_set_register(HV_REGISTER_CRASH_P3, virt_to_phys(hv_panic_page));
1427         hv_set_register(HV_REGISTER_CRASH_P4, bytes_written);
1428
1429         /*
1430          * Let Hyper-V know there is crash data available along with
1431          * the panic message.
1432          */
1433         hv_set_register(HV_REGISTER_CRASH_CTL,
1434                (HV_CRASH_CTL_CRASH_NOTIFY | HV_CRASH_CTL_CRASH_NOTIFY_MSG));
1435 }
1436
1437 static struct kmsg_dumper hv_kmsg_dumper = {
1438         .dump = hv_kmsg_dump,
1439 };
1440
1441 static void hv_kmsg_dump_register(void)
1442 {
1443         int ret;
1444
1445         hv_panic_page = hv_alloc_hyperv_zeroed_page();
1446         if (!hv_panic_page) {
1447                 pr_err("Hyper-V: panic message page memory allocation failed\n");
1448                 return;
1449         }
1450
1451         ret = kmsg_dump_register(&hv_kmsg_dumper);
1452         if (ret) {
1453                 pr_err("Hyper-V: kmsg dump register error 0x%x\n", ret);
1454                 hv_free_hyperv_page((unsigned long)hv_panic_page);
1455                 hv_panic_page = NULL;
1456         }
1457 }
1458
1459 static struct ctl_table_header *hv_ctl_table_hdr;
1460
1461 /*
1462  * sysctl option to allow the user to control whether kmsg data should be
1463  * reported to Hyper-V on panic.
1464  */
1465 static struct ctl_table hv_ctl_table[] = {
1466         {
1467                 .procname       = "hyperv_record_panic_msg",
1468                 .data           = &sysctl_record_panic_msg,
1469                 .maxlen         = sizeof(int),
1470                 .mode           = 0644,
1471                 .proc_handler   = proc_dointvec_minmax,
1472                 .extra1         = SYSCTL_ZERO,
1473                 .extra2         = SYSCTL_ONE
1474         },
1475         {}
1476 };
1477
1478 static struct ctl_table hv_root_table[] = {
1479         {
1480                 .procname       = "kernel",
1481                 .mode           = 0555,
1482                 .child          = hv_ctl_table
1483         },
1484         {}
1485 };
1486
1487 /*
1488  * vmbus_bus_init -Main vmbus driver initialization routine.
1489  *
1490  * Here, we
1491  *      - initialize the vmbus driver context
1492  *      - invoke the vmbus hv main init routine
1493  *      - retrieve the channel offers
1494  */
1495 static int vmbus_bus_init(void)
1496 {
1497         int ret;
1498
1499         ret = hv_init();
1500         if (ret != 0) {
1501                 pr_err("Unable to initialize the hypervisor - 0x%x\n", ret);
1502                 return ret;
1503         }
1504
1505         ret = bus_register(&hv_bus);
1506         if (ret)
1507                 return ret;
1508
1509         /*
1510          * VMbus interrupts are best modeled as per-cpu interrupts. If
1511          * on an architecture with support for per-cpu IRQs (e.g. ARM64),
1512          * allocate a per-cpu IRQ using standard Linux kernel functionality.
1513          * If not on such an architecture (e.g., x86/x64), then rely on
1514          * code in the arch-specific portion of the code tree to connect
1515          * the VMbus interrupt handler.
1516          */
1517
1518         if (vmbus_irq == -1) {
1519                 hv_setup_vmbus_handler(vmbus_isr);
1520         } else {
1521                 vmbus_evt = alloc_percpu(long);
1522                 ret = request_percpu_irq(vmbus_irq, vmbus_percpu_isr,
1523                                 "Hyper-V VMbus", vmbus_evt);
1524                 if (ret) {
1525                         pr_err("Can't request Hyper-V VMbus IRQ %d, Err %d",
1526                                         vmbus_irq, ret);
1527                         free_percpu(vmbus_evt);
1528                         goto err_setup;
1529                 }
1530         }
1531
1532         ret = hv_synic_alloc();
1533         if (ret)
1534                 goto err_alloc;
1535
1536         /*
1537          * Initialize the per-cpu interrupt state and stimer state.
1538          * Then connect to the host.
1539          */
1540         ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "hyperv/vmbus:online",
1541                                 hv_synic_init, hv_synic_cleanup);
1542         if (ret < 0)
1543                 goto err_cpuhp;
1544         hyperv_cpuhp_online = ret;
1545
1546         ret = vmbus_connect();
1547         if (ret)
1548                 goto err_connect;
1549
1550         /*
1551          * Only register if the crash MSRs are available
1552          */
1553         if (ms_hyperv.misc_features & HV_FEATURE_GUEST_CRASH_MSR_AVAILABLE) {
1554                 u64 hyperv_crash_ctl;
1555                 /*
1556                  * Sysctl registration is not fatal, since by default
1557                  * reporting is enabled.
1558                  */
1559                 hv_ctl_table_hdr = register_sysctl_table(hv_root_table);
1560                 if (!hv_ctl_table_hdr)
1561                         pr_err("Hyper-V: sysctl table register error");
1562
1563                 /*
1564                  * Register for panic kmsg callback only if the right
1565                  * capability is supported by the hypervisor.
1566                  */
1567                 hyperv_crash_ctl = hv_get_register(HV_REGISTER_CRASH_CTL);
1568                 if (hyperv_crash_ctl & HV_CRASH_CTL_CRASH_NOTIFY_MSG)
1569                         hv_kmsg_dump_register();
1570
1571                 register_die_notifier(&hyperv_die_block);
1572         }
1573
1574         /*
1575          * Always register the panic notifier because we need to unload
1576          * the VMbus channel connection to prevent any VMbus
1577          * activity after the VM panics.
1578          */
1579         atomic_notifier_chain_register(&panic_notifier_list,
1580                                &hyperv_panic_block);
1581
1582         vmbus_request_offers();
1583
1584         return 0;
1585
1586 err_connect:
1587         cpuhp_remove_state(hyperv_cpuhp_online);
1588 err_cpuhp:
1589         hv_synic_free();
1590 err_alloc:
1591         if (vmbus_irq == -1) {
1592                 hv_remove_vmbus_handler();
1593         } else {
1594                 free_percpu_irq(vmbus_irq, vmbus_evt);
1595                 free_percpu(vmbus_evt);
1596         }
1597 err_setup:
1598         bus_unregister(&hv_bus);
1599         unregister_sysctl_table(hv_ctl_table_hdr);
1600         hv_ctl_table_hdr = NULL;
1601         return ret;
1602 }
1603
1604 /**
1605  * __vmbus_child_driver_register() - Register a vmbus's driver
1606  * @hv_driver: Pointer to driver structure you want to register
1607  * @owner: owner module of the drv
1608  * @mod_name: module name string
1609  *
1610  * Registers the given driver with Linux through the 'driver_register()' call
1611  * and sets up the hyper-v vmbus handling for this driver.
1612  * It will return the state of the 'driver_register()' call.
1613  *
1614  */
1615 int __vmbus_driver_register(struct hv_driver *hv_driver, struct module *owner, const char *mod_name)
1616 {
1617         int ret;
1618
1619         pr_info("registering driver %s\n", hv_driver->name);
1620
1621         ret = vmbus_exists();
1622         if (ret < 0)
1623                 return ret;
1624
1625         hv_driver->driver.name = hv_driver->name;
1626         hv_driver->driver.owner = owner;
1627         hv_driver->driver.mod_name = mod_name;
1628         hv_driver->driver.bus = &hv_bus;
1629
1630         spin_lock_init(&hv_driver->dynids.lock);
1631         INIT_LIST_HEAD(&hv_driver->dynids.list);
1632
1633         ret = driver_register(&hv_driver->driver);
1634
1635         return ret;
1636 }
1637 EXPORT_SYMBOL_GPL(__vmbus_driver_register);
1638
1639 /**
1640  * vmbus_driver_unregister() - Unregister a vmbus's driver
1641  * @hv_driver: Pointer to driver structure you want to
1642  *             un-register
1643  *
1644  * Un-register the given driver that was previous registered with a call to
1645  * vmbus_driver_register()
1646  */
1647 void vmbus_driver_unregister(struct hv_driver *hv_driver)
1648 {
1649         pr_info("unregistering driver %s\n", hv_driver->name);
1650
1651         if (!vmbus_exists()) {
1652                 driver_unregister(&hv_driver->driver);
1653                 vmbus_free_dynids(hv_driver);
1654         }
1655 }
1656 EXPORT_SYMBOL_GPL(vmbus_driver_unregister);
1657
1658
1659 /*
1660  * Called when last reference to channel is gone.
1661  */
1662 static void vmbus_chan_release(struct kobject *kobj)
1663 {
1664         struct vmbus_channel *channel
1665                 = container_of(kobj, struct vmbus_channel, kobj);
1666
1667         kfree_rcu(channel, rcu);
1668 }
1669
1670 struct vmbus_chan_attribute {
1671         struct attribute attr;
1672         ssize_t (*show)(struct vmbus_channel *chan, char *buf);
1673         ssize_t (*store)(struct vmbus_channel *chan,
1674                          const char *buf, size_t count);
1675 };
1676 #define VMBUS_CHAN_ATTR(_name, _mode, _show, _store) \
1677         struct vmbus_chan_attribute chan_attr_##_name \
1678                 = __ATTR(_name, _mode, _show, _store)
1679 #define VMBUS_CHAN_ATTR_RW(_name) \
1680         struct vmbus_chan_attribute chan_attr_##_name = __ATTR_RW(_name)
1681 #define VMBUS_CHAN_ATTR_RO(_name) \
1682         struct vmbus_chan_attribute chan_attr_##_name = __ATTR_RO(_name)
1683 #define VMBUS_CHAN_ATTR_WO(_name) \
1684         struct vmbus_chan_attribute chan_attr_##_name = __ATTR_WO(_name)
1685
1686 static ssize_t vmbus_chan_attr_show(struct kobject *kobj,
1687                                     struct attribute *attr, char *buf)
1688 {
1689         const struct vmbus_chan_attribute *attribute
1690                 = container_of(attr, struct vmbus_chan_attribute, attr);
1691         struct vmbus_channel *chan
1692                 = container_of(kobj, struct vmbus_channel, kobj);
1693
1694         if (!attribute->show)
1695                 return -EIO;
1696
1697         return attribute->show(chan, buf);
1698 }
1699
1700 static ssize_t vmbus_chan_attr_store(struct kobject *kobj,
1701                                      struct attribute *attr, const char *buf,
1702                                      size_t count)
1703 {
1704         const struct vmbus_chan_attribute *attribute
1705                 = container_of(attr, struct vmbus_chan_attribute, attr);
1706         struct vmbus_channel *chan
1707                 = container_of(kobj, struct vmbus_channel, kobj);
1708
1709         if (!attribute->store)
1710                 return -EIO;
1711
1712         return attribute->store(chan, buf, count);
1713 }
1714
1715 static const struct sysfs_ops vmbus_chan_sysfs_ops = {
1716         .show = vmbus_chan_attr_show,
1717         .store = vmbus_chan_attr_store,
1718 };
1719
1720 static ssize_t out_mask_show(struct vmbus_channel *channel, char *buf)
1721 {
1722         struct hv_ring_buffer_info *rbi = &channel->outbound;
1723         ssize_t ret;
1724
1725         mutex_lock(&rbi->ring_buffer_mutex);
1726         if (!rbi->ring_buffer) {
1727                 mutex_unlock(&rbi->ring_buffer_mutex);
1728                 return -EINVAL;
1729         }
1730
1731         ret = sprintf(buf, "%u\n", rbi->ring_buffer->interrupt_mask);
1732         mutex_unlock(&rbi->ring_buffer_mutex);
1733         return ret;
1734 }
1735 static VMBUS_CHAN_ATTR_RO(out_mask);
1736
1737 static ssize_t in_mask_show(struct vmbus_channel *channel, char *buf)
1738 {
1739         struct hv_ring_buffer_info *rbi = &channel->inbound;
1740         ssize_t ret;
1741
1742         mutex_lock(&rbi->ring_buffer_mutex);
1743         if (!rbi->ring_buffer) {
1744                 mutex_unlock(&rbi->ring_buffer_mutex);
1745                 return -EINVAL;
1746         }
1747
1748         ret = sprintf(buf, "%u\n", rbi->ring_buffer->interrupt_mask);
1749         mutex_unlock(&rbi->ring_buffer_mutex);
1750         return ret;
1751 }
1752 static VMBUS_CHAN_ATTR_RO(in_mask);
1753
1754 static ssize_t read_avail_show(struct vmbus_channel *channel, char *buf)
1755 {
1756         struct hv_ring_buffer_info *rbi = &channel->inbound;
1757         ssize_t ret;
1758
1759         mutex_lock(&rbi->ring_buffer_mutex);
1760         if (!rbi->ring_buffer) {
1761                 mutex_unlock(&rbi->ring_buffer_mutex);
1762                 return -EINVAL;
1763         }
1764
1765         ret = sprintf(buf, "%u\n", hv_get_bytes_to_read(rbi));
1766         mutex_unlock(&rbi->ring_buffer_mutex);
1767         return ret;
1768 }
1769 static VMBUS_CHAN_ATTR_RO(read_avail);
1770
1771 static ssize_t write_avail_show(struct vmbus_channel *channel, char *buf)
1772 {
1773         struct hv_ring_buffer_info *rbi = &channel->outbound;
1774         ssize_t ret;
1775
1776         mutex_lock(&rbi->ring_buffer_mutex);
1777         if (!rbi->ring_buffer) {
1778                 mutex_unlock(&rbi->ring_buffer_mutex);
1779                 return -EINVAL;
1780         }
1781
1782         ret = sprintf(buf, "%u\n", hv_get_bytes_to_write(rbi));
1783         mutex_unlock(&rbi->ring_buffer_mutex);
1784         return ret;
1785 }
1786 static VMBUS_CHAN_ATTR_RO(write_avail);
1787
1788 static ssize_t target_cpu_show(struct vmbus_channel *channel, char *buf)
1789 {
1790         return sprintf(buf, "%u\n", channel->target_cpu);
1791 }
1792 static ssize_t target_cpu_store(struct vmbus_channel *channel,
1793                                 const char *buf, size_t count)
1794 {
1795         u32 target_cpu, origin_cpu;
1796         ssize_t ret = count;
1797
1798         if (vmbus_proto_version < VERSION_WIN10_V4_1)
1799                 return -EIO;
1800
1801         if (sscanf(buf, "%uu", &target_cpu) != 1)
1802                 return -EIO;
1803
1804         /* Validate target_cpu for the cpumask_test_cpu() operation below. */
1805         if (target_cpu >= nr_cpumask_bits)
1806                 return -EINVAL;
1807
1808         /* No CPUs should come up or down during this. */
1809         cpus_read_lock();
1810
1811         if (!cpu_online(target_cpu)) {
1812                 cpus_read_unlock();
1813                 return -EINVAL;
1814         }
1815
1816         /*
1817          * Synchronizes target_cpu_store() and channel closure:
1818          *
1819          * { Initially: state = CHANNEL_OPENED }
1820          *
1821          * CPU1                         CPU2
1822          *
1823          * [target_cpu_store()]         [vmbus_disconnect_ring()]
1824          *
1825          * LOCK channel_mutex           LOCK channel_mutex
1826          * LOAD r1 = state              LOAD r2 = state
1827          * IF (r1 == CHANNEL_OPENED)    IF (r2 == CHANNEL_OPENED)
1828          *   SEND MODIFYCHANNEL           STORE state = CHANNEL_OPEN
1829          *   [...]                        SEND CLOSECHANNEL
1830          * UNLOCK channel_mutex         UNLOCK channel_mutex
1831          *
1832          * Forbids: r1 == r2 == CHANNEL_OPENED (i.e., CPU1's LOCK precedes
1833          *              CPU2's LOCK) && CPU2's SEND precedes CPU1's SEND
1834          *
1835          * Note.  The host processes the channel messages "sequentially", in
1836          * the order in which they are received on a per-partition basis.
1837          */
1838         mutex_lock(&vmbus_connection.channel_mutex);
1839
1840         /*
1841          * Hyper-V will ignore MODIFYCHANNEL messages for "non-open" channels;
1842          * avoid sending the message and fail here for such channels.
1843          */
1844         if (channel->state != CHANNEL_OPENED_STATE) {
1845                 ret = -EIO;
1846                 goto cpu_store_unlock;
1847         }
1848
1849         origin_cpu = channel->target_cpu;
1850         if (target_cpu == origin_cpu)
1851                 goto cpu_store_unlock;
1852
1853         if (vmbus_send_modifychannel(channel,
1854                                      hv_cpu_number_to_vp_number(target_cpu))) {
1855                 ret = -EIO;
1856                 goto cpu_store_unlock;
1857         }
1858
1859         /*
1860          * For version before VERSION_WIN10_V5_3, the following warning holds:
1861          *
1862          * Warning.  At this point, there is *no* guarantee that the host will
1863          * have successfully processed the vmbus_send_modifychannel() request.
1864          * See the header comment of vmbus_send_modifychannel() for more info.
1865          *
1866          * Lags in the processing of the above vmbus_send_modifychannel() can
1867          * result in missed interrupts if the "old" target CPU is taken offline
1868          * before Hyper-V starts sending interrupts to the "new" target CPU.
1869          * But apart from this offlining scenario, the code tolerates such
1870          * lags.  It will function correctly even if a channel interrupt comes
1871          * in on a CPU that is different from the channel target_cpu value.
1872          */
1873
1874         channel->target_cpu = target_cpu;
1875
1876         /* See init_vp_index(). */
1877         if (hv_is_perf_channel(channel))
1878                 hv_update_alloced_cpus(origin_cpu, target_cpu);
1879
1880         /* Currently set only for storvsc channels. */
1881         if (channel->change_target_cpu_callback) {
1882                 (*channel->change_target_cpu_callback)(channel,
1883                                 origin_cpu, target_cpu);
1884         }
1885
1886 cpu_store_unlock:
1887         mutex_unlock(&vmbus_connection.channel_mutex);
1888         cpus_read_unlock();
1889         return ret;
1890 }
1891 static VMBUS_CHAN_ATTR(cpu, 0644, target_cpu_show, target_cpu_store);
1892
1893 static ssize_t channel_pending_show(struct vmbus_channel *channel,
1894                                     char *buf)
1895 {
1896         return sprintf(buf, "%d\n",
1897                        channel_pending(channel,
1898                                        vmbus_connection.monitor_pages[1]));
1899 }
1900 static VMBUS_CHAN_ATTR(pending, 0444, channel_pending_show, NULL);
1901
1902 static ssize_t channel_latency_show(struct vmbus_channel *channel,
1903                                     char *buf)
1904 {
1905         return sprintf(buf, "%d\n",
1906                        channel_latency(channel,
1907                                        vmbus_connection.monitor_pages[1]));
1908 }
1909 static VMBUS_CHAN_ATTR(latency, 0444, channel_latency_show, NULL);
1910
1911 static ssize_t channel_interrupts_show(struct vmbus_channel *channel, char *buf)
1912 {
1913         return sprintf(buf, "%llu\n", channel->interrupts);
1914 }
1915 static VMBUS_CHAN_ATTR(interrupts, 0444, channel_interrupts_show, NULL);
1916
1917 static ssize_t channel_events_show(struct vmbus_channel *channel, char *buf)
1918 {
1919         return sprintf(buf, "%llu\n", channel->sig_events);
1920 }
1921 static VMBUS_CHAN_ATTR(events, 0444, channel_events_show, NULL);
1922
1923 static ssize_t channel_intr_in_full_show(struct vmbus_channel *channel,
1924                                          char *buf)
1925 {
1926         return sprintf(buf, "%llu\n",
1927                        (unsigned long long)channel->intr_in_full);
1928 }
1929 static VMBUS_CHAN_ATTR(intr_in_full, 0444, channel_intr_in_full_show, NULL);
1930
1931 static ssize_t channel_intr_out_empty_show(struct vmbus_channel *channel,
1932                                            char *buf)
1933 {
1934         return sprintf(buf, "%llu\n",
1935                        (unsigned long long)channel->intr_out_empty);
1936 }
1937 static VMBUS_CHAN_ATTR(intr_out_empty, 0444, channel_intr_out_empty_show, NULL);
1938
1939 static ssize_t channel_out_full_first_show(struct vmbus_channel *channel,
1940                                            char *buf)
1941 {
1942         return sprintf(buf, "%llu\n",
1943                        (unsigned long long)channel->out_full_first);
1944 }
1945 static VMBUS_CHAN_ATTR(out_full_first, 0444, channel_out_full_first_show, NULL);
1946
1947 static ssize_t channel_out_full_total_show(struct vmbus_channel *channel,
1948                                            char *buf)
1949 {
1950         return sprintf(buf, "%llu\n",
1951                        (unsigned long long)channel->out_full_total);
1952 }
1953 static VMBUS_CHAN_ATTR(out_full_total, 0444, channel_out_full_total_show, NULL);
1954
1955 static ssize_t subchannel_monitor_id_show(struct vmbus_channel *channel,
1956                                           char *buf)
1957 {
1958         return sprintf(buf, "%u\n", channel->offermsg.monitorid);
1959 }
1960 static VMBUS_CHAN_ATTR(monitor_id, 0444, subchannel_monitor_id_show, NULL);
1961
1962 static ssize_t subchannel_id_show(struct vmbus_channel *channel,
1963                                   char *buf)
1964 {
1965         return sprintf(buf, "%u\n",
1966                        channel->offermsg.offer.sub_channel_index);
1967 }
1968 static VMBUS_CHAN_ATTR_RO(subchannel_id);
1969
1970 static struct attribute *vmbus_chan_attrs[] = {
1971         &chan_attr_out_mask.attr,
1972         &chan_attr_in_mask.attr,
1973         &chan_attr_read_avail.attr,
1974         &chan_attr_write_avail.attr,
1975         &chan_attr_cpu.attr,
1976         &chan_attr_pending.attr,
1977         &chan_attr_latency.attr,
1978         &chan_attr_interrupts.attr,
1979         &chan_attr_events.attr,
1980         &chan_attr_intr_in_full.attr,
1981         &chan_attr_intr_out_empty.attr,
1982         &chan_attr_out_full_first.attr,
1983         &chan_attr_out_full_total.attr,
1984         &chan_attr_monitor_id.attr,
1985         &chan_attr_subchannel_id.attr,
1986         NULL
1987 };
1988
1989 /*
1990  * Channel-level attribute_group callback function. Returns the permission for
1991  * each attribute, and returns 0 if an attribute is not visible.
1992  */
1993 static umode_t vmbus_chan_attr_is_visible(struct kobject *kobj,
1994                                           struct attribute *attr, int idx)
1995 {
1996         const struct vmbus_channel *channel =
1997                 container_of(kobj, struct vmbus_channel, kobj);
1998
1999         /* Hide the monitor attributes if the monitor mechanism is not used. */
2000         if (!channel->offermsg.monitor_allocated &&
2001             (attr == &chan_attr_pending.attr ||
2002              attr == &chan_attr_latency.attr ||
2003              attr == &chan_attr_monitor_id.attr))
2004                 return 0;
2005
2006         return attr->mode;
2007 }
2008
2009 static struct attribute_group vmbus_chan_group = {
2010         .attrs = vmbus_chan_attrs,
2011         .is_visible = vmbus_chan_attr_is_visible
2012 };
2013
2014 static struct kobj_type vmbus_chan_ktype = {
2015         .sysfs_ops = &vmbus_chan_sysfs_ops,
2016         .release = vmbus_chan_release,
2017 };
2018
2019 /*
2020  * vmbus_add_channel_kobj - setup a sub-directory under device/channels
2021  */
2022 int vmbus_add_channel_kobj(struct hv_device *dev, struct vmbus_channel *channel)
2023 {
2024         const struct device *device = &dev->device;
2025         struct kobject *kobj = &channel->kobj;
2026         u32 relid = channel->offermsg.child_relid;
2027         int ret;
2028
2029         kobj->kset = dev->channels_kset;
2030         ret = kobject_init_and_add(kobj, &vmbus_chan_ktype, NULL,
2031                                    "%u", relid);
2032         if (ret)
2033                 return ret;
2034
2035         ret = sysfs_create_group(kobj, &vmbus_chan_group);
2036
2037         if (ret) {
2038                 /*
2039                  * The calling functions' error handling paths will cleanup the
2040                  * empty channel directory.
2041                  */
2042                 dev_err(device, "Unable to set up channel sysfs files\n");
2043                 return ret;
2044         }
2045
2046         kobject_uevent(kobj, KOBJ_ADD);
2047
2048         return 0;
2049 }
2050
2051 /*
2052  * vmbus_remove_channel_attr_group - remove the channel's attribute group
2053  */
2054 void vmbus_remove_channel_attr_group(struct vmbus_channel *channel)
2055 {
2056         sysfs_remove_group(&channel->kobj, &vmbus_chan_group);
2057 }
2058
2059 /*
2060  * vmbus_device_create - Creates and registers a new child device
2061  * on the vmbus.
2062  */
2063 struct hv_device *vmbus_device_create(const guid_t *type,
2064                                       const guid_t *instance,
2065                                       struct vmbus_channel *channel)
2066 {
2067         struct hv_device *child_device_obj;
2068
2069         child_device_obj = kzalloc(sizeof(struct hv_device), GFP_KERNEL);
2070         if (!child_device_obj) {
2071                 pr_err("Unable to allocate device object for child device\n");
2072                 return NULL;
2073         }
2074
2075         child_device_obj->channel = channel;
2076         guid_copy(&child_device_obj->dev_type, type);
2077         guid_copy(&child_device_obj->dev_instance, instance);
2078         child_device_obj->vendor_id = 0x1414; /* MSFT vendor ID */
2079
2080         return child_device_obj;
2081 }
2082
2083 /*
2084  * vmbus_device_register - Register the child device
2085  */
2086 int vmbus_device_register(struct hv_device *child_device_obj)
2087 {
2088         struct kobject *kobj = &child_device_obj->device.kobj;
2089         int ret;
2090
2091         dev_set_name(&child_device_obj->device, "%pUl",
2092                      &child_device_obj->channel->offermsg.offer.if_instance);
2093
2094         child_device_obj->device.bus = &hv_bus;
2095         child_device_obj->device.parent = &hv_acpi_dev->dev;
2096         child_device_obj->device.release = vmbus_device_release;
2097
2098         /*
2099          * Register with the LDM. This will kick off the driver/device
2100          * binding...which will eventually call vmbus_match() and vmbus_probe()
2101          */
2102         ret = device_register(&child_device_obj->device);
2103         if (ret) {
2104                 pr_err("Unable to register child device\n");
2105                 return ret;
2106         }
2107
2108         child_device_obj->channels_kset = kset_create_and_add("channels",
2109                                                               NULL, kobj);
2110         if (!child_device_obj->channels_kset) {
2111                 ret = -ENOMEM;
2112                 goto err_dev_unregister;
2113         }
2114
2115         ret = vmbus_add_channel_kobj(child_device_obj,
2116                                      child_device_obj->channel);
2117         if (ret) {
2118                 pr_err("Unable to register primary channeln");
2119                 goto err_kset_unregister;
2120         }
2121         hv_debug_add_dev_dir(child_device_obj);
2122
2123         return 0;
2124
2125 err_kset_unregister:
2126         kset_unregister(child_device_obj->channels_kset);
2127
2128 err_dev_unregister:
2129         device_unregister(&child_device_obj->device);
2130         return ret;
2131 }
2132
2133 /*
2134  * vmbus_device_unregister - Remove the specified child device
2135  * from the vmbus.
2136  */
2137 void vmbus_device_unregister(struct hv_device *device_obj)
2138 {
2139         pr_debug("child device %s unregistered\n",
2140                 dev_name(&device_obj->device));
2141
2142         kset_unregister(device_obj->channels_kset);
2143
2144         /*
2145          * Kick off the process of unregistering the device.
2146          * This will call vmbus_remove() and eventually vmbus_device_release()
2147          */
2148         device_unregister(&device_obj->device);
2149 }
2150
2151
2152 /*
2153  * VMBUS is an acpi enumerated device. Get the information we
2154  * need from DSDT.
2155  */
2156 #define VTPM_BASE_ADDRESS 0xfed40000
2157 static acpi_status vmbus_walk_resources(struct acpi_resource *res, void *ctx)
2158 {
2159         resource_size_t start = 0;
2160         resource_size_t end = 0;
2161         struct resource *new_res;
2162         struct resource **old_res = &hyperv_mmio;
2163         struct resource **prev_res = NULL;
2164         struct resource r;
2165
2166         switch (res->type) {
2167
2168         /*
2169          * "Address" descriptors are for bus windows. Ignore
2170          * "memory" descriptors, which are for registers on
2171          * devices.
2172          */
2173         case ACPI_RESOURCE_TYPE_ADDRESS32:
2174                 start = res->data.address32.address.minimum;
2175                 end = res->data.address32.address.maximum;
2176                 break;
2177
2178         case ACPI_RESOURCE_TYPE_ADDRESS64:
2179                 start = res->data.address64.address.minimum;
2180                 end = res->data.address64.address.maximum;
2181                 break;
2182
2183         /*
2184          * The IRQ information is needed only on ARM64, which Hyper-V
2185          * sets up in the extended format. IRQ information is present
2186          * on x86/x64 in the non-extended format but it is not used by
2187          * Linux. So don't bother checking for the non-extended format.
2188          */
2189         case ACPI_RESOURCE_TYPE_EXTENDED_IRQ:
2190                 if (!acpi_dev_resource_interrupt(res, 0, &r)) {
2191                         pr_err("Unable to parse Hyper-V ACPI interrupt\n");
2192                         return AE_ERROR;
2193                 }
2194                 /* ARM64 INTID for VMbus */
2195                 vmbus_interrupt = res->data.extended_irq.interrupts[0];
2196                 /* Linux IRQ number */
2197                 vmbus_irq = r.start;
2198                 return AE_OK;
2199
2200         default:
2201                 /* Unused resource type */
2202                 return AE_OK;
2203
2204         }
2205         /*
2206          * Ignore ranges that are below 1MB, as they're not
2207          * necessary or useful here.
2208          */
2209         if (end < 0x100000)
2210                 return AE_OK;
2211
2212         new_res = kzalloc(sizeof(*new_res), GFP_ATOMIC);
2213         if (!new_res)
2214                 return AE_NO_MEMORY;
2215
2216         /* If this range overlaps the virtual TPM, truncate it. */
2217         if (end > VTPM_BASE_ADDRESS && start < VTPM_BASE_ADDRESS)
2218                 end = VTPM_BASE_ADDRESS;
2219
2220         new_res->name = "hyperv mmio";
2221         new_res->flags = IORESOURCE_MEM;
2222         new_res->start = start;
2223         new_res->end = end;
2224
2225         /*
2226          * If two ranges are adjacent, merge them.
2227          */
2228         do {
2229                 if (!*old_res) {
2230                         *old_res = new_res;
2231                         break;
2232                 }
2233
2234                 if (((*old_res)->end + 1) == new_res->start) {
2235                         (*old_res)->end = new_res->end;
2236                         kfree(new_res);
2237                         break;
2238                 }
2239
2240                 if ((*old_res)->start == new_res->end + 1) {
2241                         (*old_res)->start = new_res->start;
2242                         kfree(new_res);
2243                         break;
2244                 }
2245
2246                 if ((*old_res)->start > new_res->end) {
2247                         new_res->sibling = *old_res;
2248                         if (prev_res)
2249                                 (*prev_res)->sibling = new_res;
2250                         *old_res = new_res;
2251                         break;
2252                 }
2253
2254                 prev_res = old_res;
2255                 old_res = &(*old_res)->sibling;
2256
2257         } while (1);
2258
2259         return AE_OK;
2260 }
2261
2262 static int vmbus_acpi_remove(struct acpi_device *device)
2263 {
2264         struct resource *cur_res;
2265         struct resource *next_res;
2266
2267         if (hyperv_mmio) {
2268                 if (fb_mmio) {
2269                         __release_region(hyperv_mmio, fb_mmio->start,
2270                                          resource_size(fb_mmio));
2271                         fb_mmio = NULL;
2272                 }
2273
2274                 for (cur_res = hyperv_mmio; cur_res; cur_res = next_res) {
2275                         next_res = cur_res->sibling;
2276                         kfree(cur_res);
2277                 }
2278         }
2279
2280         return 0;
2281 }
2282
2283 static void vmbus_reserve_fb(void)
2284 {
2285         int size;
2286         /*
2287          * Make a claim for the frame buffer in the resource tree under the
2288          * first node, which will be the one below 4GB.  The length seems to
2289          * be underreported, particularly in a Generation 1 VM.  So start out
2290          * reserving a larger area and make it smaller until it succeeds.
2291          */
2292
2293         if (screen_info.lfb_base) {
2294                 if (efi_enabled(EFI_BOOT))
2295                         size = max_t(__u32, screen_info.lfb_size, 0x800000);
2296                 else
2297                         size = max_t(__u32, screen_info.lfb_size, 0x4000000);
2298
2299                 for (; !fb_mmio && (size >= 0x100000); size >>= 1) {
2300                         fb_mmio = __request_region(hyperv_mmio,
2301                                                    screen_info.lfb_base, size,
2302                                                    fb_mmio_name, 0);
2303                 }
2304         }
2305 }
2306
2307 /**
2308  * vmbus_allocate_mmio() - Pick a memory-mapped I/O range.
2309  * @new:                If successful, supplied a pointer to the
2310  *                      allocated MMIO space.
2311  * @device_obj:         Identifies the caller
2312  * @min:                Minimum guest physical address of the
2313  *                      allocation
2314  * @max:                Maximum guest physical address
2315  * @size:               Size of the range to be allocated
2316  * @align:              Alignment of the range to be allocated
2317  * @fb_overlap_ok:      Whether this allocation can be allowed
2318  *                      to overlap the video frame buffer.
2319  *
2320  * This function walks the resources granted to VMBus by the
2321  * _CRS object in the ACPI namespace underneath the parent
2322  * "bridge" whether that's a root PCI bus in the Generation 1
2323  * case or a Module Device in the Generation 2 case.  It then
2324  * attempts to allocate from the global MMIO pool in a way that
2325  * matches the constraints supplied in these parameters and by
2326  * that _CRS.
2327  *
2328  * Return: 0 on success, -errno on failure
2329  */
2330 int vmbus_allocate_mmio(struct resource **new, struct hv_device *device_obj,
2331                         resource_size_t min, resource_size_t max,
2332                         resource_size_t size, resource_size_t align,
2333                         bool fb_overlap_ok)
2334 {
2335         struct resource *iter, *shadow;
2336         resource_size_t range_min, range_max, start;
2337         const char *dev_n = dev_name(&device_obj->device);
2338         int retval;
2339
2340         retval = -ENXIO;
2341         mutex_lock(&hyperv_mmio_lock);
2342
2343         /*
2344          * If overlaps with frame buffers are allowed, then first attempt to
2345          * make the allocation from within the reserved region.  Because it
2346          * is already reserved, no shadow allocation is necessary.
2347          */
2348         if (fb_overlap_ok && fb_mmio && !(min > fb_mmio->end) &&
2349             !(max < fb_mmio->start)) {
2350
2351                 range_min = fb_mmio->start;
2352                 range_max = fb_mmio->end;
2353                 start = (range_min + align - 1) & ~(align - 1);
2354                 for (; start + size - 1 <= range_max; start += align) {
2355                         *new = request_mem_region_exclusive(start, size, dev_n);
2356                         if (*new) {
2357                                 retval = 0;
2358                                 goto exit;
2359                         }
2360                 }
2361         }
2362
2363         for (iter = hyperv_mmio; iter; iter = iter->sibling) {
2364                 if ((iter->start >= max) || (iter->end <= min))
2365                         continue;
2366
2367                 range_min = iter->start;
2368                 range_max = iter->end;
2369                 start = (range_min + align - 1) & ~(align - 1);
2370                 for (; start + size - 1 <= range_max; start += align) {
2371                         shadow = __request_region(iter, start, size, NULL,
2372                                                   IORESOURCE_BUSY);
2373                         if (!shadow)
2374                                 continue;
2375
2376                         *new = request_mem_region_exclusive(start, size, dev_n);
2377                         if (*new) {
2378                                 shadow->name = (char *)*new;
2379                                 retval = 0;
2380                                 goto exit;
2381                         }
2382
2383                         __release_region(iter, start, size);
2384                 }
2385         }
2386
2387 exit:
2388         mutex_unlock(&hyperv_mmio_lock);
2389         return retval;
2390 }
2391 EXPORT_SYMBOL_GPL(vmbus_allocate_mmio);
2392
2393 /**
2394  * vmbus_free_mmio() - Free a memory-mapped I/O range.
2395  * @start:              Base address of region to release.
2396  * @size:               Size of the range to be allocated
2397  *
2398  * This function releases anything requested by
2399  * vmbus_mmio_allocate().
2400  */
2401 void vmbus_free_mmio(resource_size_t start, resource_size_t size)
2402 {
2403         struct resource *iter;
2404
2405         mutex_lock(&hyperv_mmio_lock);
2406         for (iter = hyperv_mmio; iter; iter = iter->sibling) {
2407                 if ((iter->start >= start + size) || (iter->end <= start))
2408                         continue;
2409
2410                 __release_region(iter, start, size);
2411         }
2412         release_mem_region(start, size);
2413         mutex_unlock(&hyperv_mmio_lock);
2414
2415 }
2416 EXPORT_SYMBOL_GPL(vmbus_free_mmio);
2417
2418 static int vmbus_acpi_add(struct acpi_device *device)
2419 {
2420         acpi_status result;
2421         int ret_val = -ENODEV;
2422         struct acpi_device *ancestor;
2423
2424         hv_acpi_dev = device;
2425
2426         result = acpi_walk_resources(device->handle, METHOD_NAME__CRS,
2427                                         vmbus_walk_resources, NULL);
2428
2429         if (ACPI_FAILURE(result))
2430                 goto acpi_walk_err;
2431         /*
2432          * Some ancestor of the vmbus acpi device (Gen1 or Gen2
2433          * firmware) is the VMOD that has the mmio ranges. Get that.
2434          */
2435         for (ancestor = device->parent; ancestor; ancestor = ancestor->parent) {
2436                 result = acpi_walk_resources(ancestor->handle, METHOD_NAME__CRS,
2437                                              vmbus_walk_resources, NULL);
2438
2439                 if (ACPI_FAILURE(result))
2440                         continue;
2441                 if (hyperv_mmio) {
2442                         vmbus_reserve_fb();
2443                         break;
2444                 }
2445         }
2446         ret_val = 0;
2447
2448 acpi_walk_err:
2449         complete(&probe_event);
2450         if (ret_val)
2451                 vmbus_acpi_remove(device);
2452         return ret_val;
2453 }
2454
2455 #ifdef CONFIG_PM_SLEEP
2456 static int vmbus_bus_suspend(struct device *dev)
2457 {
2458         struct vmbus_channel *channel, *sc;
2459
2460         while (atomic_read(&vmbus_connection.offer_in_progress) != 0) {
2461                 /*
2462                  * We wait here until the completion of any channel
2463                  * offers that are currently in progress.
2464                  */
2465                 usleep_range(1000, 2000);
2466         }
2467
2468         mutex_lock(&vmbus_connection.channel_mutex);
2469         list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
2470                 if (!is_hvsock_channel(channel))
2471                         continue;
2472
2473                 vmbus_force_channel_rescinded(channel);
2474         }
2475         mutex_unlock(&vmbus_connection.channel_mutex);
2476
2477         /*
2478          * Wait until all the sub-channels and hv_sock channels have been
2479          * cleaned up. Sub-channels should be destroyed upon suspend, otherwise
2480          * they would conflict with the new sub-channels that will be created
2481          * in the resume path. hv_sock channels should also be destroyed, but
2482          * a hv_sock channel of an established hv_sock connection can not be
2483          * really destroyed since it may still be referenced by the userspace
2484          * application, so we just force the hv_sock channel to be rescinded
2485          * by vmbus_force_channel_rescinded(), and the userspace application
2486          * will thoroughly destroy the channel after hibernation.
2487          *
2488          * Note: the counter nr_chan_close_on_suspend may never go above 0 if
2489          * the VM has no sub-channel and hv_sock channel, e.g. a 1-vCPU VM.
2490          */
2491         if (atomic_read(&vmbus_connection.nr_chan_close_on_suspend) > 0)
2492                 wait_for_completion(&vmbus_connection.ready_for_suspend_event);
2493
2494         if (atomic_read(&vmbus_connection.nr_chan_fixup_on_resume) != 0) {
2495                 pr_err("Can not suspend due to a previous failed resuming\n");
2496                 return -EBUSY;
2497         }
2498
2499         mutex_lock(&vmbus_connection.channel_mutex);
2500
2501         list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
2502                 /*
2503                  * Remove the channel from the array of channels and invalidate
2504                  * the channel's relid.  Upon resume, vmbus_onoffer() will fix
2505                  * up the relid (and other fields, if necessary) and add the
2506                  * channel back to the array.
2507                  */
2508                 vmbus_channel_unmap_relid(channel);
2509                 channel->offermsg.child_relid = INVALID_RELID;
2510
2511                 if (is_hvsock_channel(channel)) {
2512                         if (!channel->rescind) {
2513                                 pr_err("hv_sock channel not rescinded!\n");
2514                                 WARN_ON_ONCE(1);
2515                         }
2516                         continue;
2517                 }
2518
2519                 list_for_each_entry(sc, &channel->sc_list, sc_list) {
2520                         pr_err("Sub-channel not deleted!\n");
2521                         WARN_ON_ONCE(1);
2522                 }
2523
2524                 atomic_inc(&vmbus_connection.nr_chan_fixup_on_resume);
2525         }
2526
2527         mutex_unlock(&vmbus_connection.channel_mutex);
2528
2529         vmbus_initiate_unload(false);
2530
2531         /* Reset the event for the next resume. */
2532         reinit_completion(&vmbus_connection.ready_for_resume_event);
2533
2534         return 0;
2535 }
2536
2537 static int vmbus_bus_resume(struct device *dev)
2538 {
2539         struct vmbus_channel_msginfo *msginfo;
2540         size_t msgsize;
2541         int ret;
2542
2543         /*
2544          * We only use the 'vmbus_proto_version', which was in use before
2545          * hibernation, to re-negotiate with the host.
2546          */
2547         if (!vmbus_proto_version) {
2548                 pr_err("Invalid proto version = 0x%x\n", vmbus_proto_version);
2549                 return -EINVAL;
2550         }
2551
2552         msgsize = sizeof(*msginfo) +
2553                   sizeof(struct vmbus_channel_initiate_contact);
2554
2555         msginfo = kzalloc(msgsize, GFP_KERNEL);
2556
2557         if (msginfo == NULL)
2558                 return -ENOMEM;
2559
2560         ret = vmbus_negotiate_version(msginfo, vmbus_proto_version);
2561
2562         kfree(msginfo);
2563
2564         if (ret != 0)
2565                 return ret;
2566
2567         WARN_ON(atomic_read(&vmbus_connection.nr_chan_fixup_on_resume) == 0);
2568
2569         vmbus_request_offers();
2570
2571         if (wait_for_completion_timeout(
2572                 &vmbus_connection.ready_for_resume_event, 10 * HZ) == 0)
2573                 pr_err("Some vmbus device is missing after suspending?\n");
2574
2575         /* Reset the event for the next suspend. */
2576         reinit_completion(&vmbus_connection.ready_for_suspend_event);
2577
2578         return 0;
2579 }
2580 #else
2581 #define vmbus_bus_suspend NULL
2582 #define vmbus_bus_resume NULL
2583 #endif /* CONFIG_PM_SLEEP */
2584
2585 static const struct acpi_device_id vmbus_acpi_device_ids[] = {
2586         {"VMBUS", 0},
2587         {"VMBus", 0},
2588         {"", 0},
2589 };
2590 MODULE_DEVICE_TABLE(acpi, vmbus_acpi_device_ids);
2591
2592 /*
2593  * Note: we must use the "no_irq" ops, otherwise hibernation can not work with
2594  * PCI device assignment, because "pci_dev_pm_ops" uses the "noirq" ops: in
2595  * the resume path, the pci "noirq" restore op runs before "non-noirq" op (see
2596  * resume_target_kernel() -> dpm_resume_start(), and hibernation_restore() ->
2597  * dpm_resume_end()). This means vmbus_bus_resume() and the pci-hyperv's
2598  * resume callback must also run via the "noirq" ops.
2599  *
2600  * Set suspend_noirq/resume_noirq to NULL for Suspend-to-Idle: see the comment
2601  * earlier in this file before vmbus_pm.
2602  */
2603
2604 static const struct dev_pm_ops vmbus_bus_pm = {
2605         .suspend_noirq  = NULL,
2606         .resume_noirq   = NULL,
2607         .freeze_noirq   = vmbus_bus_suspend,
2608         .thaw_noirq     = vmbus_bus_resume,
2609         .poweroff_noirq = vmbus_bus_suspend,
2610         .restore_noirq  = vmbus_bus_resume
2611 };
2612
2613 static struct acpi_driver vmbus_acpi_driver = {
2614         .name = "vmbus",
2615         .ids = vmbus_acpi_device_ids,
2616         .ops = {
2617                 .add = vmbus_acpi_add,
2618                 .remove = vmbus_acpi_remove,
2619         },
2620         .drv.pm = &vmbus_bus_pm,
2621 };
2622
2623 static void hv_kexec_handler(void)
2624 {
2625         hv_stimer_global_cleanup();
2626         vmbus_initiate_unload(false);
2627         /* Make sure conn_state is set as hv_synic_cleanup checks for it */
2628         mb();
2629         cpuhp_remove_state(hyperv_cpuhp_online);
2630 };
2631
2632 static void hv_crash_handler(struct pt_regs *regs)
2633 {
2634         int cpu;
2635
2636         vmbus_initiate_unload(true);
2637         /*
2638          * In crash handler we can't schedule synic cleanup for all CPUs,
2639          * doing the cleanup for current CPU only. This should be sufficient
2640          * for kdump.
2641          */
2642         cpu = smp_processor_id();
2643         hv_stimer_cleanup(cpu);
2644         hv_synic_disable_regs(cpu);
2645 };
2646
2647 static int hv_synic_suspend(void)
2648 {
2649         /*
2650          * When we reach here, all the non-boot CPUs have been offlined.
2651          * If we're in a legacy configuration where stimer Direct Mode is
2652          * not enabled, the stimers on the non-boot CPUs have been unbound
2653          * in hv_synic_cleanup() -> hv_stimer_legacy_cleanup() ->
2654          * hv_stimer_cleanup() -> clockevents_unbind_device().
2655          *
2656          * hv_synic_suspend() only runs on CPU0 with interrupts disabled.
2657          * Here we do not call hv_stimer_legacy_cleanup() on CPU0 because:
2658          * 1) it's unnecessary as interrupts remain disabled between
2659          * syscore_suspend() and syscore_resume(): see create_image() and
2660          * resume_target_kernel()
2661          * 2) the stimer on CPU0 is automatically disabled later by
2662          * syscore_suspend() -> timekeeping_suspend() -> tick_suspend() -> ...
2663          * -> clockevents_shutdown() -> ... -> hv_ce_shutdown()
2664          * 3) a warning would be triggered if we call
2665          * clockevents_unbind_device(), which may sleep, in an
2666          * interrupts-disabled context.
2667          */
2668
2669         hv_synic_disable_regs(0);
2670
2671         return 0;
2672 }
2673
2674 static void hv_synic_resume(void)
2675 {
2676         hv_synic_enable_regs(0);
2677
2678         /*
2679          * Note: we don't need to call hv_stimer_init(0), because the timer
2680          * on CPU0 is not unbound in hv_synic_suspend(), and the timer is
2681          * automatically re-enabled in timekeeping_resume().
2682          */
2683 }
2684
2685 /* The callbacks run only on CPU0, with irqs_disabled. */
2686 static struct syscore_ops hv_synic_syscore_ops = {
2687         .suspend = hv_synic_suspend,
2688         .resume = hv_synic_resume,
2689 };
2690
2691 static int __init hv_acpi_init(void)
2692 {
2693         int ret, t;
2694
2695         if (!hv_is_hyperv_initialized())
2696                 return -ENODEV;
2697
2698         if (hv_root_partition)
2699                 return 0;
2700
2701         init_completion(&probe_event);
2702
2703         /*
2704          * Get ACPI resources first.
2705          */
2706         ret = acpi_bus_register_driver(&vmbus_acpi_driver);
2707
2708         if (ret)
2709                 return ret;
2710
2711         t = wait_for_completion_timeout(&probe_event, 5*HZ);
2712         if (t == 0) {
2713                 ret = -ETIMEDOUT;
2714                 goto cleanup;
2715         }
2716
2717         /*
2718          * If we're on an architecture with a hardcoded hypervisor
2719          * vector (i.e. x86/x64), override the VMbus interrupt found
2720          * in the ACPI tables. Ensure vmbus_irq is not set since the
2721          * normal Linux IRQ mechanism is not used in this case.
2722          */
2723 #ifdef HYPERVISOR_CALLBACK_VECTOR
2724         vmbus_interrupt = HYPERVISOR_CALLBACK_VECTOR;
2725         vmbus_irq = -1;
2726 #endif
2727
2728         hv_debug_init();
2729
2730         ret = vmbus_bus_init();
2731         if (ret)
2732                 goto cleanup;
2733
2734         hv_setup_kexec_handler(hv_kexec_handler);
2735         hv_setup_crash_handler(hv_crash_handler);
2736
2737         register_syscore_ops(&hv_synic_syscore_ops);
2738
2739         return 0;
2740
2741 cleanup:
2742         acpi_bus_unregister_driver(&vmbus_acpi_driver);
2743         hv_acpi_dev = NULL;
2744         return ret;
2745 }
2746
2747 static void __exit vmbus_exit(void)
2748 {
2749         int cpu;
2750
2751         unregister_syscore_ops(&hv_synic_syscore_ops);
2752
2753         hv_remove_kexec_handler();
2754         hv_remove_crash_handler();
2755         vmbus_connection.conn_state = DISCONNECTED;
2756         hv_stimer_global_cleanup();
2757         vmbus_disconnect();
2758         if (vmbus_irq == -1) {
2759                 hv_remove_vmbus_handler();
2760         } else {
2761                 free_percpu_irq(vmbus_irq, vmbus_evt);
2762                 free_percpu(vmbus_evt);
2763         }
2764         for_each_online_cpu(cpu) {
2765                 struct hv_per_cpu_context *hv_cpu
2766                         = per_cpu_ptr(hv_context.cpu_context, cpu);
2767
2768                 tasklet_kill(&hv_cpu->msg_dpc);
2769         }
2770         hv_debug_rm_all_dir();
2771
2772         vmbus_free_channels();
2773         kfree(vmbus_connection.channels);
2774
2775         if (ms_hyperv.misc_features & HV_FEATURE_GUEST_CRASH_MSR_AVAILABLE) {
2776                 kmsg_dump_unregister(&hv_kmsg_dumper);
2777                 unregister_die_notifier(&hyperv_die_block);
2778                 atomic_notifier_chain_unregister(&panic_notifier_list,
2779                                                  &hyperv_panic_block);
2780         }
2781
2782         free_page((unsigned long)hv_panic_page);
2783         unregister_sysctl_table(hv_ctl_table_hdr);
2784         hv_ctl_table_hdr = NULL;
2785         bus_unregister(&hv_bus);
2786
2787         cpuhp_remove_state(hyperv_cpuhp_online);
2788         hv_synic_free();
2789         acpi_bus_unregister_driver(&vmbus_acpi_driver);
2790 }
2791
2792
2793 MODULE_LICENSE("GPL");
2794 MODULE_DESCRIPTION("Microsoft Hyper-V VMBus Driver");
2795
2796 subsys_initcall(hv_acpi_init);
2797 module_exit(vmbus_exit);