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