drivers/hv/vmbus_drv.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (c) 2009, Microsoft Corporation.
   4  *
   5  * Authors:
   6  *   Haiyang Zhang <haiyangz@microsoft.com>
   7  *   Hank Janssen  <hjanssen@microsoft.com>
   8  *   K. Y. Srinivasan <kys@microsoft.com>
   9  */
  10 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  11
  12 #include <linux/init.h>
  13 #include <linux/module.h>
  14 #include <linux/device.h>
  15 #include <linux/interrupt.h>
  16 #include <linux/sysctl.h>
  17 #include <linux/slab.h>
  18 #include <linux/acpi.h>
  19 #include <linux/completion.h>
  20 #include <linux/hyperv.h>
  21 #include <linux/kernel_stat.h>
  22 #include <linux/clockchips.h>
  23 #include <linux/cpu.h>
  24 #include <linux/sched/task_stack.h>
  25
  26 #include <linux/delay.h>
  27 #include <linux/notifier.h>
  28 #include <linux/panic_notifier.h>
  29 #include <linux/ptrace.h>
  30 #include <linux/screen_info.h>
  31 #include <linux/kdebug.h>
  32 #include <linux/efi.h>
  33 #include <linux/random.h>
  34 #include <linux/kernel.h>
  35 #include <linux/syscore_ops.h>
  36 #include <linux/dma-map-ops.h>
  37 #include <clocksource/hyperv_timer.h>
  38 #include "hyperv_vmbus.h"
  39
  40 struct vmbus_dynid {
  41         struct list_head node;
  42         struct hv_vmbus_device_id id;
  43 };
  44
  45 static struct acpi_device  *hv_acpi_dev;
  46
  47 static struct completion probe_event;
  48
  49 static int hyperv_cpuhp_online;
  50
  51 static void *hv_panic_page;
  52
  53 static long __percpu *vmbus_evt;
  54
  55 /* Values parsed from ACPI DSDT */
  56 int vmbus_irq;
  57 int vmbus_interrupt;
  58
  59 /*
  60  * Boolean to control whether to report panic messages over Hyper-V.
  61  *
  62  * It can be set via /proc/sys/kernel/hyperv_record_panic_msg
  63  */
  64 static int sysctl_record_panic_msg = 1;
  65
  66 static int hyperv_report_reg(void)
  67 {
  68         return !sysctl_record_panic_msg || !hv_panic_page;
  69 }
  70
  71 static int hyperv_panic_event(struct notifier_block *nb, unsigned long val,
  72                               void *args)
  73 {
  74         struct pt_regs *regs;
  75
  76         vmbus_initiate_unload(true);
  77
  78         /*
  79          * Hyper-V should be notified only once about a panic.  If we will be
  80          * doing hv_kmsg_dump() with kmsg data later, don't do the notification
  81          * here.
  82          */
  83         if (ms_hyperv.misc_features & HV_FEATURE_GUEST_CRASH_MSR_AVAILABLE
  84             && hyperv_report_reg()) {
  85                 regs = current_pt_regs();
  86                 hyperv_report_panic(regs, val, false);
  87         }
  88         return NOTIFY_DONE;
  89 }
  90
  91 static int hyperv_die_event(struct notifier_block *nb, unsigned long val,
  92                             void *args)
  93 {
  94         struct die_args *die = args;
  95         struct pt_regs *regs = die->regs;
  96
  97         /* Don't notify Hyper-V if the die event is other than oops */
  98         if (val != DIE_OOPS)
  99                 return NOTIFY_DONE;
 100
 101         /*
 102          * Hyper-V should be notified only once about a panic.  If we will be
 103          * doing hv_kmsg_dump() with kmsg data later, don't do the notification
 104          * here.
 105          */
 106         if (hyperv_report_reg())
 107                 hyperv_report_panic(regs, val, true);
 108         return NOTIFY_DONE;
 109 }
 110
 111 static struct notifier_block hyperv_die_block = {
 112         .notifier_call = hyperv_die_event,
 113 };
 114 static struct notifier_block hyperv_panic_block = {
 115         .notifier_call = hyperv_panic_event,
 116 };
 117
 118 static const char *fb_mmio_name = "fb_range";
 119 static struct resource *fb_mmio;
 120 static struct resource *hyperv_mmio;
 121 static DEFINE_MUTEX(hyperv_mmio_lock);
 122
 123 static int vmbus_exists(void)
 124 {
 125         if (hv_acpi_dev == NULL)
 126                 return -ENODEV;
 127
 128         return 0;
 129 }
 130
 131 static u8 channel_monitor_group(const struct vmbus_channel *channel)
 132 {
 133         return (u8)channel->offermsg.monitorid / 32;
 134 }
 135
 136 static u8 channel_monitor_offset(const struct vmbus_channel *channel)
 137 {
 138         return (u8)channel->offermsg.monitorid % 32;
 139 }
 140
 141 static u32 channel_pending(const struct vmbus_channel *channel,
 142                            const struct hv_monitor_page *monitor_page)
 143 {
 144         u8 monitor_group = channel_monitor_group(channel);
 145
 146         return monitor_page->trigger_group[monitor_group].pending;
 147 }
 148
 149 static u32 channel_latency(const struct vmbus_channel *channel,
 150                            const struct hv_monitor_page *monitor_page)
 151 {
 152         u8 monitor_group = channel_monitor_group(channel);
 153         u8 monitor_offset = channel_monitor_offset(channel);
 154
 155         return monitor_page->latency[monitor_group][monitor_offset];
 156 }
 157
 158 static u32 channel_conn_id(struct vmbus_channel *channel,
 159                            struct hv_monitor_page *monitor_page)
 160 {
 161         u8 monitor_group = channel_monitor_group(channel);
 162         u8 monitor_offset = channel_monitor_offset(channel);
 163
 164         return monitor_page->parameter[monitor_group][monitor_offset].connectionid.u.id;
 165 }
 166
 167 static ssize_t id_show(struct device *dev, struct device_attribute *dev_attr,
 168                        char *buf)
 169 {
 170         struct hv_device *hv_dev = device_to_hv_device(dev);
 171
 172         if (!hv_dev->channel)
 173                 return -ENODEV;
 174         return sprintf(buf, "%d\n", hv_dev->channel->offermsg.child_relid);
 175 }
 176 static DEVICE_ATTR_RO(id);
 177
 178 static ssize_t state_show(struct device *dev, struct device_attribute *dev_attr,
 179                           char *buf)
 180 {
 181         struct hv_device *hv_dev = device_to_hv_device(dev);
 182
 183         if (!hv_dev->channel)
 184                 return -ENODEV;
 185         return sprintf(buf, "%d\n", hv_dev->channel->state);
 186 }
 187 static DEVICE_ATTR_RO(state);
 188
 189 static ssize_t monitor_id_show(struct device *dev,
 190                                struct device_attribute *dev_attr, char *buf)
 191 {
 192         struct hv_device *hv_dev = device_to_hv_device(dev);
 193
 194         if (!hv_dev->channel)
 195                 return -ENODEV;
 196         return sprintf(buf, "%d\n", hv_dev->channel->offermsg.monitorid);
 197 }
 198 static DEVICE_ATTR_RO(monitor_id);
 199
 200 static ssize_t class_id_show(struct device *dev,
 201                                struct device_attribute *dev_attr, char *buf)
 202 {
 203         struct hv_device *hv_dev = device_to_hv_device(dev);
 204
 205         if (!hv_dev->channel)
 206                 return -ENODEV;
 207         return sprintf(buf, "{%pUl}\n",
 208                        &hv_dev->channel->offermsg.offer.if_type);
 209 }
 210 static DEVICE_ATTR_RO(class_id);
 211
 212 static ssize_t device_id_show(struct device *dev,
 213                               struct device_attribute *dev_attr, char *buf)
 214 {
 215         struct hv_device *hv_dev = device_to_hv_device(dev);
 216
 217         if (!hv_dev->channel)
 218                 return -ENODEV;
 219         return sprintf(buf, "{%pUl}\n",
 220                        &hv_dev->channel->offermsg.offer.if_instance);
 221 }
 222 static DEVICE_ATTR_RO(device_id);
 223
 224 static ssize_t modalias_show(struct device *dev,
 225                              struct device_attribute *dev_attr, char *buf)
 226 {
 227         struct hv_device *hv_dev = device_to_hv_device(dev);
 228
 229         return sprintf(buf, "vmbus:%*phN\n", UUID_SIZE, &hv_dev->dev_type);
 230 }
 231 static DEVICE_ATTR_RO(modalias);
 232
 233 #ifdef CONFIG_NUMA
 234 static ssize_t numa_node_show(struct device *dev,
 235                               struct device_attribute *attr, char *buf)
 236 {
 237         struct hv_device *hv_dev = device_to_hv_device(dev);
 238
 239         if (!hv_dev->channel)
 240                 return -ENODEV;
 241
 242         return sprintf(buf, "%d\n", cpu_to_node(hv_dev->channel->target_cpu));
 243 }
 244 static DEVICE_ATTR_RO(numa_node);
 245 #endif
 246
 247 static ssize_t server_monitor_pending_show(struct device *dev,
 248                                            struct device_attribute *dev_attr,
 249                                            char *buf)
 250 {
 251         struct hv_device *hv_dev = device_to_hv_device(dev);
 252
 253         if (!hv_dev->channel)
 254                 return -ENODEV;
 255         return sprintf(buf, "%d\n",
 256                        channel_pending(hv_dev->channel,
 257                                        vmbus_connection.monitor_pages[0]));
 258 }
 259 static DEVICE_ATTR_RO(server_monitor_pending);
 260
 261 static ssize_t client_monitor_pending_show(struct device *dev,
 262                                            struct device_attribute *dev_attr,
 263                                            char *buf)
 264 {
 265         struct hv_device *hv_dev = device_to_hv_device(dev);
 266
 267         if (!hv_dev->channel)
 268                 return -ENODEV;
 269         return sprintf(buf, "%d\n",
 270                        channel_pending(hv_dev->channel,
 271                                        vmbus_connection.monitor_pages[1]));
 272 }
 273 static DEVICE_ATTR_RO(client_monitor_pending);
 274
 275 static ssize_t server_monitor_latency_show(struct device *dev,
 276                                            struct device_attribute *dev_attr,
 277                                            char *buf)
 278 {
 279         struct hv_device *hv_dev = device_to_hv_device(dev);
 280
 281         if (!hv_dev->channel)
 282                 return -ENODEV;
 283         return sprintf(buf, "%d\n",
 284                        channel_latency(hv_dev->channel,
 285                                        vmbus_connection.monitor_pages[0]));
 286 }
 287 static DEVICE_ATTR_RO(server_monitor_latency);
 288
 289 static ssize_t client_monitor_latency_show(struct device *dev,
 290                                            struct device_attribute *dev_attr,
 291                                            char *buf)
 292 {
 293         struct hv_device *hv_dev = device_to_hv_device(dev);
 294
 295         if (!hv_dev->channel)
 296                 return -ENODEV;
 297         return sprintf(buf, "%d\n",
 298                        channel_latency(hv_dev->channel,
 299                                        vmbus_connection.monitor_pages[1]));
 300 }
 301 static DEVICE_ATTR_RO(client_monitor_latency);
 302
 303 static ssize_t server_monitor_conn_id_show(struct device *dev,
 304                                            struct device_attribute *dev_attr,
 305                                            char *buf)
 306 {
 307         struct hv_device *hv_dev = device_to_hv_device(dev);
 308
 309         if (!hv_dev->channel)
 310                 return -ENODEV;
 311         return sprintf(buf, "%d\n",
 312                        channel_conn_id(hv_dev->channel,
 313                                        vmbus_connection.monitor_pages[0]));
 314 }
 315 static DEVICE_ATTR_RO(server_monitor_conn_id);
 316
 317 static ssize_t client_monitor_conn_id_show(struct device *dev,
 318                                            struct device_attribute *dev_attr,
 319                                            char *buf)
 320 {
 321         struct hv_device *hv_dev = device_to_hv_device(dev);
 322
 323         if (!hv_dev->channel)
 324                 return -ENODEV;
 325         return sprintf(buf, "%d\n",
 326                        channel_conn_id(hv_dev->channel,
 327                                        vmbus_connection.monitor_pages[1]));
 328 }
 329 static DEVICE_ATTR_RO(client_monitor_conn_id);
 330
 331 static ssize_t out_intr_mask_show(struct device *dev,
 332                                   struct device_attribute *dev_attr, char *buf)
 333 {
 334         struct hv_device *hv_dev = device_to_hv_device(dev);
 335         struct hv_ring_buffer_debug_info outbound;
 336         int ret;
 337
 338         if (!hv_dev->channel)
 339                 return -ENODEV;
 340
 341         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->outbound,
 342                                           &outbound);
 343         if (ret < 0)
 344                 return ret;
 345
 346         return sprintf(buf, "%d\n", outbound.current_interrupt_mask);
 347 }
 348 static DEVICE_ATTR_RO(out_intr_mask);
 349
 350 static ssize_t out_read_index_show(struct device *dev,
 351                                    struct device_attribute *dev_attr, char *buf)
 352 {
 353         struct hv_device *hv_dev = device_to_hv_device(dev);
 354         struct hv_ring_buffer_debug_info outbound;
 355         int ret;
 356
 357         if (!hv_dev->channel)
 358                 return -ENODEV;
 359
 360         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->outbound,
 361                                           &outbound);
 362         if (ret < 0)
 363                 return ret;
 364         return sprintf(buf, "%d\n", outbound.current_read_index);
 365 }
 366 static DEVICE_ATTR_RO(out_read_index);
 367
 368 static ssize_t out_write_index_show(struct device *dev,
 369                                     struct device_attribute *dev_attr,
 370                                     char *buf)
 371 {
 372         struct hv_device *hv_dev = device_to_hv_device(dev);
 373         struct hv_ring_buffer_debug_info outbound;
 374         int ret;
 375
 376         if (!hv_dev->channel)
 377                 return -ENODEV;
 378
 379         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->outbound,
 380                                           &outbound);
 381         if (ret < 0)
 382                 return ret;
 383         return sprintf(buf, "%d\n", outbound.current_write_index);
 384 }
 385 static DEVICE_ATTR_RO(out_write_index);
 386
 387 static ssize_t out_read_bytes_avail_show(struct device *dev,
 388                                          struct device_attribute *dev_attr,
 389                                          char *buf)
 390 {
 391         struct hv_device *hv_dev = device_to_hv_device(dev);
 392         struct hv_ring_buffer_debug_info outbound;
 393         int ret;
 394
 395         if (!hv_dev->channel)
 396                 return -ENODEV;
 397
 398         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->outbound,
 399                                           &outbound);
 400         if (ret < 0)
 401                 return ret;
 402         return sprintf(buf, "%d\n", outbound.bytes_avail_toread);
 403 }
 404 static DEVICE_ATTR_RO(out_read_bytes_avail);
 405
 406 static ssize_t out_write_bytes_avail_show(struct device *dev,
 407                                           struct device_attribute *dev_attr,
 408                                           char *buf)
 409 {
 410         struct hv_device *hv_dev = device_to_hv_device(dev);
 411         struct hv_ring_buffer_debug_info outbound;
 412         int ret;
 413
 414         if (!hv_dev->channel)
 415                 return -ENODEV;
 416
 417         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->outbound,
 418                                           &outbound);
 419         if (ret < 0)
 420                 return ret;
 421         return sprintf(buf, "%d\n", outbound.bytes_avail_towrite);
 422 }
 423 static DEVICE_ATTR_RO(out_write_bytes_avail);
 424
 425 static ssize_t in_intr_mask_show(struct device *dev,
 426                                  struct device_attribute *dev_attr, char *buf)
 427 {
 428         struct hv_device *hv_dev = device_to_hv_device(dev);
 429         struct hv_ring_buffer_debug_info inbound;
 430         int ret;
 431
 432         if (!hv_dev->channel)
 433                 return -ENODEV;
 434
 435         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->inbound, &inbound);
 436         if (ret < 0)
 437                 return ret;
 438
 439         return sprintf(buf, "%d\n", inbound.current_interrupt_mask);
 440 }
 441 static DEVICE_ATTR_RO(in_intr_mask);
 442
 443 static ssize_t in_read_index_show(struct device *dev,
 444                                   struct device_attribute *dev_attr, char *buf)
 445 {
 446         struct hv_device *hv_dev = device_to_hv_device(dev);
 447         struct hv_ring_buffer_debug_info inbound;
 448         int ret;
 449
 450         if (!hv_dev->channel)
 451                 return -ENODEV;
 452
 453         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->inbound, &inbound);
 454         if (ret < 0)
 455                 return ret;
 456
 457         return sprintf(buf, "%d\n", inbound.current_read_index);
 458 }
 459 static DEVICE_ATTR_RO(in_read_index);
 460
 461 static ssize_t in_write_index_show(struct device *dev,
 462                                    struct device_attribute *dev_attr, char *buf)
 463 {
 464         struct hv_device *hv_dev = device_to_hv_device(dev);
 465         struct hv_ring_buffer_debug_info inbound;
 466         int ret;
 467
 468         if (!hv_dev->channel)
 469                 return -ENODEV;
 470
 471         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->inbound, &inbound);
 472         if (ret < 0)
 473                 return ret;
 474
 475         return sprintf(buf, "%d\n", inbound.current_write_index);
 476 }
 477 static DEVICE_ATTR_RO(in_write_index);
 478
 479 static ssize_t in_read_bytes_avail_show(struct device *dev,
 480                                         struct device_attribute *dev_attr,
 481                                         char *buf)
 482 {
 483         struct hv_device *hv_dev = device_to_hv_device(dev);
 484         struct hv_ring_buffer_debug_info inbound;
 485         int ret;
 486
 487         if (!hv_dev->channel)
 488                 return -ENODEV;
 489
 490         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->inbound, &inbound);
 491         if (ret < 0)
 492                 return ret;
 493
 494         return sprintf(buf, "%d\n", inbound.bytes_avail_toread);
 495 }
 496 static DEVICE_ATTR_RO(in_read_bytes_avail);
 497
 498 static ssize_t in_write_bytes_avail_show(struct device *dev,
 499                                          struct device_attribute *dev_attr,
 500                                          char *buf)
 501 {
 502         struct hv_device *hv_dev = device_to_hv_device(dev);
 503         struct hv_ring_buffer_debug_info inbound;
 504         int ret;
 505
 506         if (!hv_dev->channel)
 507                 return -ENODEV;
 508
 509         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->inbound, &inbound);
 510         if (ret < 0)
 511                 return ret;
 512
 513         return sprintf(buf, "%d\n", inbound.bytes_avail_towrite);
 514 }
 515 static DEVICE_ATTR_RO(in_write_bytes_avail);
 516
 517 static ssize_t channel_vp_mapping_show(struct device *dev,
 518                                        struct device_attribute *dev_attr,
 519                                        char *buf)
 520 {
 521         struct hv_device *hv_dev = device_to_hv_device(dev);
 522         struct vmbus_channel *channel = hv_dev->channel, *cur_sc;
 523         int buf_size = PAGE_SIZE, n_written, tot_written;
 524         struct list_head *cur;
 525
 526         if (!channel)
 527                 return -ENODEV;
 528
 529         mutex_lock(&vmbus_connection.channel_mutex);
 530
 531         tot_written = snprintf(buf, buf_size, "%u:%u\n",
 532                 channel->offermsg.child_relid, channel->target_cpu);
 533
 534         list_for_each(cur, &channel->sc_list) {
 535                 if (tot_written >= buf_size - 1)
 536                         break;
 537
 538                 cur_sc = list_entry(cur, struct vmbus_channel, sc_list);
 539                 n_written = scnprintf(buf + tot_written,
 540                                      buf_size - tot_written,
 541                                      "%u:%u\n",
 542                                      cur_sc->offermsg.child_relid,
 543                                      cur_sc->target_cpu);
 544                 tot_written += n_written;
 545         }
 546
 547         mutex_unlock(&vmbus_connection.channel_mutex);
 548
 549         return tot_written;
 550 }
 551 static DEVICE_ATTR_RO(channel_vp_mapping);
 552
 553 static ssize_t vendor_show(struct device *dev,
 554                            struct device_attribute *dev_attr,
 555                            char *buf)
 556 {
 557         struct hv_device *hv_dev = device_to_hv_device(dev);
 558
 559         return sprintf(buf, "0x%x\n", hv_dev->vendor_id);
 560 }
 561 static DEVICE_ATTR_RO(vendor);
 562
 563 static ssize_t device_show(struct device *dev,
 564                            struct device_attribute *dev_attr,
 565                            char *buf)
 566 {
 567         struct hv_device *hv_dev = device_to_hv_device(dev);
 568
 569         return sprintf(buf, "0x%x\n", hv_dev->device_id);
 570 }
 571 static DEVICE_ATTR_RO(device);
 572
 573 static ssize_t driver_override_store(struct device *dev,
 574                                      struct device_attribute *attr,
 575                                      const char *buf, size_t count)
 576 {
 577         struct hv_device *hv_dev = device_to_hv_device(dev);
 578         char *driver_override, *old, *cp;
 579
 580         /* We need to keep extra room for a newline */
 581         if (count >= (PAGE_SIZE - 1))
 582                 return -EINVAL;
 583
 584         driver_override = kstrndup(buf, count, GFP_KERNEL);
 585         if (!driver_override)
 586                 return -ENOMEM;
 587
 588         cp = strchr(driver_override, '\n');
 589         if (cp)
 590                 *cp = '\0';
 591
 592         device_lock(dev);
 593         old = hv_dev->driver_override;
 594         if (strlen(driver_override)) {
 595                 hv_dev->driver_override = driver_override;
 596         } else {
 597                 kfree(driver_override);
 598                 hv_dev->driver_override = NULL;
 599         }
 600         device_unlock(dev);
 601
 602         kfree(old);
 603
 604         return count;
 605 }
 606
 607 static ssize_t driver_override_show(struct device *dev,
 608                                     struct device_attribute *attr, char *buf)
 609 {
 610         struct hv_device *hv_dev = device_to_hv_device(dev);
 611         ssize_t len;
 612
 613         device_lock(dev);
 614         len = snprintf(buf, PAGE_SIZE, "%s\n", hv_dev->driver_override);
 615         device_unlock(dev);
 616
 617         return len;
 618 }
 619 static DEVICE_ATTR_RW(driver_override);
 620
 621 /* Set up per device attributes in /sys/bus/vmbus/devices/<bus device> */
 622 static struct attribute *vmbus_dev_attrs[] = {
 623         &dev_attr_id.attr,
 624         &dev_attr_state.attr,
 625         &dev_attr_monitor_id.attr,
 626         &dev_attr_class_id.attr,
 627         &dev_attr_device_id.attr,
 628         &dev_attr_modalias.attr,
 629 #ifdef CONFIG_NUMA
 630         &dev_attr_numa_node.attr,
 631 #endif
 632         &dev_attr_server_monitor_pending.attr,
 633         &dev_attr_client_monitor_pending.attr,
 634         &dev_attr_server_monitor_latency.attr,
 635         &dev_attr_client_monitor_latency.attr,
 636         &dev_attr_server_monitor_conn_id.attr,
 637         &dev_attr_client_monitor_conn_id.attr,
 638         &dev_attr_out_intr_mask.attr,
 639         &dev_attr_out_read_index.attr,
 640         &dev_attr_out_write_index.attr,
 641         &dev_attr_out_read_bytes_avail.attr,
 642         &dev_attr_out_write_bytes_avail.attr,
 643         &dev_attr_in_intr_mask.attr,
 644         &dev_attr_in_read_index.attr,
 645         &dev_attr_in_write_index.attr,
 646         &dev_attr_in_read_bytes_avail.attr,
 647         &dev_attr_in_write_bytes_avail.attr,
 648         &dev_attr_channel_vp_mapping.attr,
 649         &dev_attr_vendor.attr,
 650         &dev_attr_device.attr,
 651         &dev_attr_driver_override.attr,
 652         NULL,
 653 };
 654
 655 /*
 656  * Device-level attribute_group callback function. Returns the permission for
 657  * each attribute, and returns 0 if an attribute is not visible.
 658  */
 659 static umode_t vmbus_dev_attr_is_visible(struct kobject *kobj,
 660                                          struct attribute *attr, int idx)
 661 {
 662         struct device *dev = kobj_to_dev(kobj);
 663         const struct hv_device *hv_dev = device_to_hv_device(dev);
 664
 665         /* Hide the monitor attributes if the monitor mechanism is not used. */
 666         if (!hv_dev->channel->offermsg.monitor_allocated &&
 667             (attr == &dev_attr_monitor_id.attr ||
 668              attr == &dev_attr_server_monitor_pending.attr ||
 669              attr == &dev_attr_client_monitor_pending.attr ||
 670              attr == &dev_attr_server_monitor_latency.attr ||
 671              attr == &dev_attr_client_monitor_latency.attr ||
 672              attr == &dev_attr_server_monitor_conn_id.attr ||
 673              attr == &dev_attr_client_monitor_conn_id.attr))
 674                 return 0;
 675
 676         return attr->mode;
 677 }
 678
 679 static const struct attribute_group vmbus_dev_group = {
 680         .attrs = vmbus_dev_attrs,
 681         .is_visible = vmbus_dev_attr_is_visible
 682 };
 683 __ATTRIBUTE_GROUPS(vmbus_dev);
 684
 685 /* Set up the attribute for /sys/bus/vmbus/hibernation */
 686 static ssize_t hibernation_show(struct bus_type *bus, char *buf)
 687 {
 688         return sprintf(buf, "%d\n", !!hv_is_hibernation_supported());
 689 }
 690
 691 static BUS_ATTR_RO(hibernation);
 692
 693 static struct attribute *vmbus_bus_attrs[] = {
 694         &bus_attr_hibernation.attr,
 695         NULL,
 696 };
 697 static const struct attribute_group vmbus_bus_group = {
 698         .attrs = vmbus_bus_attrs,
 699 };
 700 __ATTRIBUTE_GROUPS(vmbus_bus);
 701
 702 /*
 703  * vmbus_uevent - add uevent for our device
 704  *
 705  * This routine is invoked when a device is added or removed on the vmbus to
 706  * generate a uevent to udev in the userspace. The udev will then look at its
 707  * rule and the uevent generated here to load the appropriate driver
 708  *
 709  * The alias string will be of the form vmbus:guid where guid is the string
 710  * representation of the device guid (each byte of the guid will be
 711  * represented with two hex characters.
 712  */
 713 static int vmbus_uevent(struct device *device, struct kobj_uevent_env *env)
 714 {
 715         struct hv_device *dev = device_to_hv_device(device);
 716         const char *format = "MODALIAS=vmbus:%*phN";
 717
 718         return add_uevent_var(env, format, UUID_SIZE, &dev->dev_type);
 719 }
 720
 721 static const struct hv_vmbus_device_id *
 722 hv_vmbus_dev_match(const struct hv_vmbus_device_id *id, const guid_t *guid)
 723 {
 724         if (id == NULL)
 725                 return NULL; /* empty device table */
 726
 727         for (; !guid_is_null(&id->guid); id++)
 728                 if (guid_equal(&id->guid, guid))
 729                         return id;
 730
 731         return NULL;
 732 }
 733
 734 static const struct hv_vmbus_device_id *
 735 hv_vmbus_dynid_match(struct hv_driver *drv, const guid_t *guid)
 736 {
 737         const struct hv_vmbus_device_id *id = NULL;
 738         struct vmbus_dynid *dynid;
 739
 740         spin_lock(&drv->dynids.lock);
 741         list_for_each_entry(dynid, &drv->dynids.list, node) {
 742                 if (guid_equal(&dynid->id.guid, guid)) {
 743                         id = &dynid->id;
 744                         break;
 745                 }
 746         }
 747         spin_unlock(&drv->dynids.lock);
 748
 749         return id;
 750 }
 751
 752 static const struct hv_vmbus_device_id vmbus_device_null;
 753
 754 /*
 755  * Return a matching hv_vmbus_device_id pointer.
 756  * If there is no match, return NULL.
 757  */
 758 static const struct hv_vmbus_device_id *hv_vmbus_get_id(struct hv_driver *drv,
 759                                                         struct hv_device *dev)
 760 {
 761         const guid_t *guid = &dev->dev_type;
 762         const struct hv_vmbus_device_id *id;
 763
 764         /* When driver_override is set, only bind to the matching driver */
 765         if (dev->driver_override && strcmp(dev->driver_override, drv->name))
 766                 return NULL;
 767
 768         /* Look at the dynamic ids first, before the static ones */
 769         id = hv_vmbus_dynid_match(drv, guid);
 770         if (!id)
 771                 id = hv_vmbus_dev_match(drv->id_table, guid);
 772
 773         /* driver_override will always match, send a dummy id */
 774         if (!id && dev->driver_override)
 775                 id = &vmbus_device_null;
 776
 777         return id;
 778 }
 779
 780 /* vmbus_add_dynid - add a new device ID to this driver and re-probe devices */
 781 static int vmbus_add_dynid(struct hv_driver *drv, guid_t *guid)
 782 {
 783         struct vmbus_dynid *dynid;
 784
 785         dynid = kzalloc(sizeof(*dynid), GFP_KERNEL);
 786         if (!dynid)
 787                 return -ENOMEM;
 788
 789         dynid->id.guid = *guid;
 790
 791         spin_lock(&drv->dynids.lock);
 792         list_add_tail(&dynid->node, &drv->dynids.list);
 793         spin_unlock(&drv->dynids.lock);
 794
 795         return driver_attach(&drv->driver);
 796 }
 797
 798 static void vmbus_free_dynids(struct hv_driver *drv)
 799 {
 800         struct vmbus_dynid *dynid, *n;
 801
 802         spin_lock(&drv->dynids.lock);
 803         list_for_each_entry_safe(dynid, n, &drv->dynids.list, node) {
 804                 list_del(&dynid->node);
 805                 kfree(dynid);
 806         }
 807         spin_unlock(&drv->dynids.lock);
 808 }
 809
 810 /*
 811  * store_new_id - sysfs frontend to vmbus_add_dynid()
 812  *
 813  * Allow GUIDs to be added to an existing driver via sysfs.
 814  */
 815 static ssize_t new_id_store(struct device_driver *driver, const char *buf,
 816                             size_t count)
 817 {
 818         struct hv_driver *drv = drv_to_hv_drv(driver);
 819         guid_t guid;
 820         ssize_t retval;
 821
 822         retval = guid_parse(buf, &guid);
 823         if (retval)
 824                 return retval;
 825
 826         if (hv_vmbus_dynid_match(drv, &guid))
 827                 return -EEXIST;
 828
 829         retval = vmbus_add_dynid(drv, &guid);
 830         if (retval)
 831                 return retval;
 832         return count;
 833 }
 834 static DRIVER_ATTR_WO(new_id);
 835
 836 /*
 837  * store_remove_id - remove a PCI device ID from this driver
 838  *
 839  * Removes a dynamic pci device ID to this driver.
 840  */
 841 static ssize_t remove_id_store(struct device_driver *driver, const char *buf,
 842                                size_t count)
 843 {
 844         struct hv_driver *drv = drv_to_hv_drv(driver);
 845         struct vmbus_dynid *dynid, *n;
 846         guid_t guid;
 847         ssize_t retval;
 848
 849         retval = guid_parse(buf, &guid);
 850         if (retval)
 851                 return retval;
 852
 853         retval = -ENODEV;
 854         spin_lock(&drv->dynids.lock);
 855         list_for_each_entry_safe(dynid, n, &drv->dynids.list, node) {
 856                 struct hv_vmbus_device_id *id = &dynid->id;
 857
 858                 if (guid_equal(&id->guid, &guid)) {
 859                         list_del(&dynid->node);
 860                         kfree(dynid);
 861                         retval = count;
 862                         break;
 863                 }
 864         }
 865         spin_unlock(&drv->dynids.lock);
 866
 867         return retval;
 868 }
 869 static DRIVER_ATTR_WO(remove_id);
 870
 871 static struct attribute *vmbus_drv_attrs[] = {
 872         &driver_attr_new_id.attr,
 873         &driver_attr_remove_id.attr,
 874         NULL,
 875 };
 876 ATTRIBUTE_GROUPS(vmbus_drv);
 877
 878
 879 /*
 880  * vmbus_match - Attempt to match the specified device to the specified driver
 881  */
 882 static int vmbus_match(struct device *device, struct device_driver *driver)
 883 {
 884         struct hv_driver *drv = drv_to_hv_drv(driver);
 885         struct hv_device *hv_dev = device_to_hv_device(device);
 886
 887         /* The hv_sock driver handles all hv_sock offers. */
 888         if (is_hvsock_channel(hv_dev->channel))
 889                 return drv->hvsock;
 890
 891         if (hv_vmbus_get_id(drv, hv_dev))
 892                 return 1;
 893
 894         return 0;
 895 }
 896
 897 /*
 898  * vmbus_probe - Add the new vmbus's child device
 899  */
 900 static int vmbus_probe(struct device *child_device)
 901 {
 902         int ret = 0;
 903         struct hv_driver *drv =
 904                         drv_to_hv_drv(child_device->driver);
 905         struct hv_device *dev = device_to_hv_device(child_device);
 906         const struct hv_vmbus_device_id *dev_id;
 907
 908         dev_id = hv_vmbus_get_id(drv, dev);
 909         if (drv->probe) {
 910                 ret = drv->probe(dev, dev_id);
 911                 if (ret != 0)
 912                         pr_err("probe failed for device %s (%d)\n",
 913                                dev_name(child_device), ret);
 914
 915         } else {
 916                 pr_err("probe not set for driver %s\n",
 917                        dev_name(child_device));
 918                 ret = -ENODEV;
 919         }
 920         return ret;
 921 }
 922
 923 /*
 924  * vmbus_dma_configure -- Configure DMA coherence for VMbus device
 925  */
 926 static int vmbus_dma_configure(struct device *child_device)
 927 {
 928         /*
 929          * On ARM64, propagate the DMA coherence setting from the top level
 930          * VMbus ACPI device to the child VMbus device being added here.
 931          * On x86/x64 coherence is assumed and these calls have no effect.
 932          */
 933         hv_setup_dma_ops(child_device,
 934                 device_get_dma_attr(&hv_acpi_dev->dev) == DEV_DMA_COHERENT);
 935         return 0;
 936 }
 937
 938 /*
 939  * vmbus_remove - Remove a vmbus device
 940  */
 941 static void vmbus_remove(struct device *child_device)
 942 {
 943         struct hv_driver *drv;
 944         struct hv_device *dev = device_to_hv_device(child_device);
 945
 946         if (child_device->driver) {
 947                 drv = drv_to_hv_drv(child_device->driver);
 948                 if (drv->remove)
 949                         drv->remove(dev);
 950         }
 951 }
 952
 953 /*
 954  * vmbus_shutdown - Shutdown a vmbus device
 955  */
 956 static void vmbus_shutdown(struct device *child_device)
 957 {
 958         struct hv_driver *drv;
 959         struct hv_device *dev = device_to_hv_device(child_device);
 960
 961
 962         /* The device may not be attached yet */
 963         if (!child_device->driver)
 964                 return;
 965
 966         drv = drv_to_hv_drv(child_device->driver);
 967
 968         if (drv->shutdown)
 969                 drv->shutdown(dev);
 970 }
 971
 972 #ifdef CONFIG_PM_SLEEP
 973 /*
 974  * vmbus_suspend - Suspend a vmbus device
 975  */
 976 static int vmbus_suspend(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->suspend)
 987                 return -EOPNOTSUPP;
 988
 989         return drv->suspend(dev);
 990 }
 991
 992 /*
 993  * vmbus_resume - Resume a vmbus device
 994  */
 995 static int vmbus_resume(struct device *child_device)
 996 {
 997         struct hv_driver *drv;
 998         struct hv_device *dev = device_to_hv_device(child_device);
 999
1000         /* The device may not be attached yet */
1001         if (!child_device->driver)
1002                 return 0;
1003
1004         drv = drv_to_hv_drv(child_device->driver);
1005         if (!drv->resume)
1006                 return -EOPNOTSUPP;
1007
1008         return drv->resume(dev);
1009 }
1010 #else
1011 #define vmbus_suspend NULL
1012 #define vmbus_resume NULL
1013 #endif /* CONFIG_PM_SLEEP */
1014
1015 /*
1016  * vmbus_device_release - Final callback release of the vmbus child device
1017  */
1018 static void vmbus_device_release(struct device *device)
1019 {
1020         struct hv_device *hv_dev = device_to_hv_device(device);
1021         struct vmbus_channel *channel = hv_dev->channel;
1022
1023         hv_debug_rm_dev_dir(hv_dev);
1024
1025         mutex_lock(&vmbus_connection.channel_mutex);
1026         hv_process_channel_removal(channel);
1027         mutex_unlock(&vmbus_connection.channel_mutex);
1028         kfree(hv_dev);
1029 }
1030
1031 /*
1032  * Note: we must use the "noirq" ops: see the comment before vmbus_bus_pm.
1033  *
1034  * suspend_noirq/resume_noirq are set to NULL to support Suspend-to-Idle: we
1035  * shouldn't suspend the vmbus devices upon Suspend-to-Idle, otherwise there
1036  * is no way to wake up a Generation-2 VM.
1037  *
1038  * The other 4 ops are for hibernation.
1039  */
1040
1041 static const struct dev_pm_ops vmbus_pm = {
1042         .suspend_noirq  = NULL,
1043         .resume_noirq   = NULL,
1044         .freeze_noirq   = vmbus_suspend,
1045         .thaw_noirq     = vmbus_resume,
1046         .poweroff_noirq = vmbus_suspend,
1047         .restore_noirq  = vmbus_resume,
1048 };
1049
1050 /* The one and only one */
1051 static struct bus_type  hv_bus = {
1052         .name =         "vmbus",
1053         .match =                vmbus_match,
1054         .shutdown =             vmbus_shutdown,
1055         .remove =               vmbus_remove,
1056         .probe =                vmbus_probe,
1057         .uevent =               vmbus_uevent,
1058         .dma_configure =        vmbus_dma_configure,
1059         .dev_groups =           vmbus_dev_groups,
1060         .drv_groups =           vmbus_drv_groups,
1061         .bus_groups =           vmbus_bus_groups,
1062         .pm =                   &vmbus_pm,
1063 };
1064
1065 struct onmessage_work_context {
1066         struct work_struct work;
1067         struct {
1068                 struct hv_message_header header;
1069                 u8 payload[];
1070         } msg;
1071 };
1072
1073 static void vmbus_onmessage_work(struct work_struct *work)
1074 {
1075         struct onmessage_work_context *ctx;
1076
1077         /* Do not process messages if we're in DISCONNECTED state */
1078         if (vmbus_connection.conn_state == DISCONNECTED)
1079                 return;
1080
1081         ctx = container_of(work, struct onmessage_work_context,
1082                            work);
1083         vmbus_onmessage((struct vmbus_channel_message_header *)
1084                         &ctx->msg.payload);
1085         kfree(ctx);
1086 }
1087
1088 void vmbus_on_msg_dpc(unsigned long data)
1089 {
1090         struct hv_per_cpu_context *hv_cpu = (void *)data;
1091         void *page_addr = hv_cpu->synic_message_page;
1092         struct hv_message msg_copy, *msg = (struct hv_message *)page_addr +
1093                                   VMBUS_MESSAGE_SINT;
1094         struct vmbus_channel_message_header *hdr;
1095         enum vmbus_channel_message_type msgtype;
1096         const struct vmbus_channel_message_table_entry *entry;
1097         struct onmessage_work_context *ctx;
1098         __u8 payload_size;
1099         u32 message_type;
1100
1101         /*
1102          * 'enum vmbus_channel_message_type' is supposed to always be 'u32' as
1103          * it is being used in 'struct vmbus_channel_message_header' definition
1104          * which is supposed to match hypervisor ABI.
1105          */
1106         BUILD_BUG_ON(sizeof(enum vmbus_channel_message_type) != sizeof(u32));
1107
1108         /*
1109          * Since the message is in memory shared with the host, an erroneous or
1110          * malicious Hyper-V could modify the message while vmbus_on_msg_dpc()
1111          * or individual message handlers are executing; to prevent this, copy
1112          * the message into private memory.
1113          */
1114         memcpy(&msg_copy, msg, sizeof(struct hv_message));
1115
1116         message_type = msg_copy.header.message_type;
1117         if (message_type == HVMSG_NONE)
1118                 /* no msg */
1119                 return;
1120
1121         hdr = (struct vmbus_channel_message_header *)msg_copy.u.payload;
1122         msgtype = hdr->msgtype;
1123
1124         trace_vmbus_on_msg_dpc(hdr);
1125
1126         if (msgtype >= CHANNELMSG_COUNT) {
1127                 WARN_ONCE(1, "unknown msgtype=%d\n", msgtype);
1128                 goto msg_handled;
1129         }
1130
1131         payload_size = msg_copy.header.payload_size;
1132         if (payload_size > HV_MESSAGE_PAYLOAD_BYTE_COUNT) {
1133                 WARN_ONCE(1, "payload size is too large (%d)\n", payload_size);
1134                 goto msg_handled;
1135         }
1136
1137         entry = &channel_message_table[msgtype];
1138
1139         if (!entry->message_handler)
1140                 goto msg_handled;
1141
1142         if (payload_size < entry->min_payload_len) {
1143                 WARN_ONCE(1, "message too short: msgtype=%d len=%d\n", msgtype, payload_size);
1144                 goto msg_handled;
1145         }
1146
1147         if (entry->handler_type == VMHT_BLOCKING) {
1148                 ctx = kmalloc(struct_size(ctx, msg.payload, payload_size), GFP_ATOMIC);
1149                 if (ctx == NULL)
1150                         return;
1151
1152                 INIT_WORK(&ctx->work, vmbus_onmessage_work);
1153                 memcpy(&ctx->msg, &msg_copy, sizeof(msg->header) + payload_size);
1154
1155                 /*
1156                  * The host can generate a rescind message while we
1157                  * may still be handling the original offer. We deal with
1158                  * this condition by relying on the synchronization provided
1159                  * by offer_in_progress and by channel_mutex.  See also the
1160                  * inline comments in vmbus_onoffer_rescind().
1161                  */
1162                 switch (msgtype) {
1163                 case CHANNELMSG_RESCIND_CHANNELOFFER:
1164                         /*
1165                          * If we are handling the rescind message;
1166                          * schedule the work on the global work queue.
1167                          *
1168                          * The OFFER message and the RESCIND message should
1169                          * not be handled by the same serialized work queue,
1170                          * because the OFFER handler may call vmbus_open(),
1171                          * which tries to open the channel by sending an
1172                          * OPEN_CHANNEL message to the host and waits for
1173                          * the host's response; however, if the host has
1174                          * rescinded the channel before it receives the
1175                          * OPEN_CHANNEL message, the host just silently
1176                          * ignores the OPEN_CHANNEL message; as a result,
1177                          * the guest's OFFER handler hangs for ever, if we
1178                          * handle the RESCIND message in the same serialized
1179                          * work queue: the RESCIND handler can not start to
1180                          * run before the OFFER handler finishes.
1181                          */
1182                         schedule_work(&ctx->work);
1183                         break;
1184
1185                 case CHANNELMSG_OFFERCHANNEL:
1186                         /*
1187                          * The host sends the offer message of a given channel
1188                          * before sending the rescind message of the same
1189                          * channel.  These messages are sent to the guest's
1190                          * connect CPU; the guest then starts processing them
1191                          * in the tasklet handler on this CPU:
1192                          *
1193                          * VMBUS_CONNECT_CPU
1194                          *
1195                          * [vmbus_on_msg_dpc()]
1196                          * atomic_inc()  // CHANNELMSG_OFFERCHANNEL
1197                          * queue_work()
1198                          * ...
1199                          * [vmbus_on_msg_dpc()]
1200                          * schedule_work()  // CHANNELMSG_RESCIND_CHANNELOFFER
1201                          *
1202                          * We rely on the memory-ordering properties of the
1203                          * queue_work() and schedule_work() primitives, which
1204                          * guarantee that the atomic increment will be visible
1205                          * to the CPUs which will execute the offer & rescind
1206                          * works by the time these works will start execution.
1207                          */
1208                         atomic_inc(&vmbus_connection.offer_in_progress);
1209                         fallthrough;
1210
1211                 default:
1212                         queue_work(vmbus_connection.work_queue, &ctx->work);
1213                 }
1214         } else
1215                 entry->message_handler(hdr);
1216
1217 msg_handled:
1218         vmbus_signal_eom(msg, message_type);
1219 }
1220
1221 #ifdef CONFIG_PM_SLEEP
1222 /*
1223  * Fake RESCIND_CHANNEL messages to clean up hv_sock channels by force for
1224  * hibernation, because hv_sock connections can not persist across hibernation.
1225  */
1226 static void vmbus_force_channel_rescinded(struct vmbus_channel *channel)
1227 {
1228         struct onmessage_work_context *ctx;
1229         struct vmbus_channel_rescind_offer *rescind;
1230
1231         WARN_ON(!is_hvsock_channel(channel));
1232
1233         /*
1234          * Allocation size is small and the allocation should really not fail,
1235          * otherwise the state of the hv_sock connections ends up in limbo.
1236          */
1237         ctx = kzalloc(sizeof(*ctx) + sizeof(*rescind),
1238                       GFP_KERNEL | __GFP_NOFAIL);
1239
1240         /*
1241          * So far, these are not really used by Linux. Just set them to the
1242          * reasonable values conforming to the definitions of the fields.
1243          */
1244         ctx->msg.header.message_type = 1;
1245         ctx->msg.header.payload_size = sizeof(*rescind);
1246
1247         /* These values are actually used by Linux. */
1248         rescind = (struct vmbus_channel_rescind_offer *)ctx->msg.payload;
1249         rescind->header.msgtype = CHANNELMSG_RESCIND_CHANNELOFFER;
1250         rescind->child_relid = channel->offermsg.child_relid;
1251
1252         INIT_WORK(&ctx->work, vmbus_onmessage_work);
1253
1254         queue_work(vmbus_connection.work_queue, &ctx->work);
1255 }
1256 #endif /* CONFIG_PM_SLEEP */
1257
1258 /*
1259  * Schedule all channels with events pending
1260  */
1261 static void vmbus_chan_sched(struct hv_per_cpu_context *hv_cpu)
1262 {
1263         unsigned long *recv_int_page;
1264         u32 maxbits, relid;
1265
1266         /*
1267          * The event page can be directly checked to get the id of
1268          * the channel that has the interrupt pending.
1269          */
1270         void *page_addr = hv_cpu->synic_event_page;
1271         union hv_synic_event_flags *event
1272                 = (union hv_synic_event_flags *)page_addr +
1273                                          VMBUS_MESSAGE_SINT;
1274
1275         maxbits = HV_EVENT_FLAGS_COUNT;
1276         recv_int_page = event->flags;
1277
1278         if (unlikely(!recv_int_page))
1279                 return;
1280
1281         for_each_set_bit(relid, recv_int_page, maxbits) {
1282                 void (*callback_fn)(void *context);
1283                 struct vmbus_channel *channel;
1284
1285                 if (!sync_test_and_clear_bit(relid, recv_int_page))
1286                         continue;
1287
1288                 /* Special case - vmbus channel protocol msg */
1289                 if (relid == 0)
1290                         continue;
1291
1292                 /*
1293                  * Pairs with the kfree_rcu() in vmbus_chan_release().
1294                  * Guarantees that the channel data structure doesn't
1295                  * get freed while the channel pointer below is being
1296                  * dereferenced.
1297                  */
1298                 rcu_read_lock();
1299
1300                 /* Find channel based on relid */
1301                 channel = relid2channel(relid);
1302                 if (channel == NULL)
1303                         goto sched_unlock_rcu;
1304
1305                 if (channel->rescind)
1306                         goto sched_unlock_rcu;
1307
1308                 /*
1309                  * Make sure that the ring buffer data structure doesn't get
1310                  * freed while we dereference the ring buffer pointer.  Test
1311                  * for the channel's onchannel_callback being NULL within a
1312                  * sched_lock critical section.  See also the inline comments
1313                  * in vmbus_reset_channel_cb().
1314                  */
1315                 spin_lock(&channel->sched_lock);
1316
1317                 callback_fn = channel->onchannel_callback;
1318                 if (unlikely(callback_fn == NULL))
1319                         goto sched_unlock;
1320
1321                 trace_vmbus_chan_sched(channel);
1322
1323                 ++channel->interrupts;
1324
1325                 switch (channel->callback_mode) {
1326                 case HV_CALL_ISR:
1327                         (*callback_fn)(channel->channel_callback_context);
1328                         break;
1329
1330                 case HV_CALL_BATCHED:
1331                         hv_begin_read(&channel->inbound);
1332                         fallthrough;
1333                 case HV_CALL_DIRECT:
1334                         tasklet_schedule(&channel->callback_event);
1335                 }
1336
1337 sched_unlock:
1338                 spin_unlock(&channel->sched_lock);
1339 sched_unlock_rcu:
1340                 rcu_read_unlock();
1341         }
1342 }
1343
1344 static void vmbus_isr(void)
1345 {
1346         struct hv_per_cpu_context *hv_cpu
1347                 = this_cpu_ptr(hv_context.cpu_context);
1348         void *page_addr;
1349         struct hv_message *msg;
1350
1351         vmbus_chan_sched(hv_cpu);
1352
1353         page_addr = hv_cpu->synic_message_page;
1354         msg = (struct hv_message *)page_addr + VMBUS_MESSAGE_SINT;
1355
1356         /* Check if there are actual msgs to be processed */
1357         if (msg->header.message_type != HVMSG_NONE) {
1358                 if (msg->header.message_type == HVMSG_TIMER_EXPIRED) {
1359                         hv_stimer0_isr();
1360                         vmbus_signal_eom(msg, HVMSG_TIMER_EXPIRED);
1361                 } else
1362                         tasklet_schedule(&hv_cpu->msg_dpc);
1363         }
1364
1365         add_interrupt_randomness(vmbus_interrupt);
1366 }
1367
1368 static irqreturn_t vmbus_percpu_isr(int irq, void *dev_id)
1369 {
1370         vmbus_isr();
1371         return IRQ_HANDLED;
1372 }
1373
1374 /*
1375  * Callback from kmsg_dump. Grab as much as possible from the end of the kmsg
1376  * buffer and call into Hyper-V to transfer the data.
1377  */
1378 static void hv_kmsg_dump(struct kmsg_dumper *dumper,
1379                          enum kmsg_dump_reason reason)
1380 {
1381         struct kmsg_dump_iter iter;
1382         size_t bytes_written;
1383
1384         /* We are only interested in panics. */
1385         if ((reason != KMSG_DUMP_PANIC) || (!sysctl_record_panic_msg))
1386                 return;
1387
1388         /*
1389          * Write dump contents to the page. No need to synchronize; panic should
1390          * be single-threaded.
1391          */
1392         kmsg_dump_rewind(&iter);
1393         kmsg_dump_get_buffer(&iter, false, hv_panic_page, HV_HYP_PAGE_SIZE,
1394                              &bytes_written);
1395         if (!bytes_written)
1396                 return;
1397         /*
1398          * P3 to contain the physical address of the panic page & P4 to
1399          * contain the size of the panic data in that page. Rest of the
1400          * registers are no-op when the NOTIFY_MSG flag is set.
1401          */
1402         hv_set_register(HV_REGISTER_CRASH_P0, 0);
1403         hv_set_register(HV_REGISTER_CRASH_P1, 0);
1404         hv_set_register(HV_REGISTER_CRASH_P2, 0);
1405         hv_set_register(HV_REGISTER_CRASH_P3, virt_to_phys(hv_panic_page));
1406         hv_set_register(HV_REGISTER_CRASH_P4, bytes_written);
1407
1408         /*
1409          * Let Hyper-V know there is crash data available along with
1410          * the panic message.
1411          */
1412         hv_set_register(HV_REGISTER_CRASH_CTL,
1413                (HV_CRASH_CTL_CRASH_NOTIFY | HV_CRASH_CTL_CRASH_NOTIFY_MSG));
1414 }
1415
1416 static struct kmsg_dumper hv_kmsg_dumper = {
1417         .dump = hv_kmsg_dump,
1418 };
1419
1420 static void hv_kmsg_dump_register(void)
1421 {
1422         int ret;
1423
1424         hv_panic_page = hv_alloc_hyperv_zeroed_page();
1425         if (!hv_panic_page) {
1426                 pr_err("Hyper-V: panic message page memory allocation failed\n");
1427                 return;
1428         }
1429
1430         ret = kmsg_dump_register(&hv_kmsg_dumper);
1431         if (ret) {
1432                 pr_err("Hyper-V: kmsg dump register error 0x%x\n", ret);
1433                 hv_free_hyperv_page((unsigned long)hv_panic_page);
1434                 hv_panic_page = NULL;
1435         }
1436 }
1437
1438 static struct ctl_table_header *hv_ctl_table_hdr;
1439
1440 /*
1441  * sysctl option to allow the user to control whether kmsg data should be
1442  * reported to Hyper-V on panic.
1443  */
1444 static struct ctl_table hv_ctl_table[] = {
1445         {
1446                 .procname       = "hyperv_record_panic_msg",
1447                 .data           = &sysctl_record_panic_msg,
1448                 .maxlen         = sizeof(int),
1449                 .mode           = 0644,
1450                 .proc_handler   = proc_dointvec_minmax,
1451                 .extra1         = SYSCTL_ZERO,
1452                 .extra2         = SYSCTL_ONE
1453         },
1454         {}
1455 };
1456
1457 static struct ctl_table hv_root_table[] = {
1458         {
1459                 .procname       = "kernel",
1460                 .mode           = 0555,
1461                 .child          = hv_ctl_table
1462         },
1463         {}
1464 };
1465
1466 /*
1467  * vmbus_bus_init -Main vmbus driver initialization routine.
1468  *
1469  * Here, we
1470  *      - initialize the vmbus driver context
1471  *      - invoke the vmbus hv main init routine
1472  *      - retrieve the channel offers
1473  */
1474 static int vmbus_bus_init(void)
1475 {
1476         int ret;
1477
1478         ret = hv_init();
1479         if (ret != 0) {
1480                 pr_err("Unable to initialize the hypervisor - 0x%x\n", ret);
1481                 return ret;
1482         }
1483
1484         ret = bus_register(&hv_bus);
1485         if (ret)
1486                 return ret;
1487
1488         /*
1489          * VMbus interrupts are best modeled as per-cpu interrupts. If
1490          * on an architecture with support for per-cpu IRQs (e.g. ARM64),
1491          * allocate a per-cpu IRQ using standard Linux kernel functionality.
1492          * If not on such an architecture (e.g., x86/x64), then rely on
1493          * code in the arch-specific portion of the code tree to connect
1494          * the VMbus interrupt handler.
1495          */
1496
1497         if (vmbus_irq == -1) {
1498                 hv_setup_vmbus_handler(vmbus_isr);
1499         } else {
1500                 vmbus_evt = alloc_percpu(long);
1501                 ret = request_percpu_irq(vmbus_irq, vmbus_percpu_isr,
1502                                 "Hyper-V VMbus", vmbus_evt);
1503                 if (ret) {
1504                         pr_err("Can't request Hyper-V VMbus IRQ %d, Err %d",
1505                                         vmbus_irq, ret);
1506                         free_percpu(vmbus_evt);
1507                         goto err_setup;
1508                 }
1509         }
1510
1511         ret = hv_synic_alloc();
1512         if (ret)
1513                 goto err_alloc;
1514
1515         /*
1516          * Initialize the per-cpu interrupt state and stimer state.
1517          * Then connect to the host.
1518          */
1519         ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "hyperv/vmbus:online",
1520                                 hv_synic_init, hv_synic_cleanup);
1521         if (ret < 0)
1522                 goto err_cpuhp;
1523         hyperv_cpuhp_online = ret;
1524
1525         ret = vmbus_connect();
1526         if (ret)
1527                 goto err_connect;
1528
1529         if (hv_is_isolation_supported())
1530                 sysctl_record_panic_msg = 0;
1531
1532         /*
1533          * Only register if the crash MSRs are available
1534          */
1535         if (ms_hyperv.misc_features & HV_FEATURE_GUEST_CRASH_MSR_AVAILABLE) {
1536                 u64 hyperv_crash_ctl;
1537                 /*
1538                  * Panic message recording (sysctl_record_panic_msg)
1539                  * is enabled by default in non-isolated guests and
1540                  * disabled by default in isolated guests; the panic
1541                  * message recording won't be available in isolated
1542                  * guests should the following registration fail.
1543                  */
1544                 hv_ctl_table_hdr = register_sysctl_table(hv_root_table);
1545                 if (!hv_ctl_table_hdr)
1546                         pr_err("Hyper-V: sysctl table register error");
1547
1548                 /*
1549                  * Register for panic kmsg callback only if the right
1550                  * capability is supported by the hypervisor.
1551                  */
1552                 hyperv_crash_ctl = hv_get_register(HV_REGISTER_CRASH_CTL);
1553                 if (hyperv_crash_ctl & HV_CRASH_CTL_CRASH_NOTIFY_MSG)
1554                         hv_kmsg_dump_register();
1555
1556                 register_die_notifier(&hyperv_die_block);
1557         }
1558
1559         /*
1560          * Always register the panic notifier because we need to unload
1561          * the VMbus channel connection to prevent any VMbus
1562          * activity after the VM panics.
1563          */
1564         atomic_notifier_chain_register(&panic_notifier_list,
1565                                &hyperv_panic_block);
1566
1567         vmbus_request_offers();
1568
1569         return 0;
1570
1571 err_connect:
1572         cpuhp_remove_state(hyperv_cpuhp_online);
1573 err_cpuhp:
1574         hv_synic_free();
1575 err_alloc:
1576         if (vmbus_irq == -1) {
1577                 hv_remove_vmbus_handler();
1578         } else {
1579                 free_percpu_irq(vmbus_irq, vmbus_evt);
1580                 free_percpu(vmbus_evt);
1581         }
1582 err_setup:
1583         bus_unregister(&hv_bus);
1584         unregister_sysctl_table(hv_ctl_table_hdr);
1585         hv_ctl_table_hdr = NULL;
1586         return ret;
1587 }
1588
1589 /**
1590  * __vmbus_child_driver_register() - Register a vmbus's driver
1591  * @hv_driver: Pointer to driver structure you want to register
1592  * @owner: owner module of the drv
1593  * @mod_name: module name string
1594  *
1595  * Registers the given driver with Linux through the 'driver_register()' call
1596  * and sets up the hyper-v vmbus handling for this driver.
1597  * It will return the state of the 'driver_register()' call.
1598  *
1599  */
1600 int __vmbus_driver_register(struct hv_driver *hv_driver, struct module *owner, const char *mod_name)
1601 {
1602         int ret;
1603
1604         pr_info("registering driver %s\n", hv_driver->name);
1605
1606         ret = vmbus_exists();
1607         if (ret < 0)
1608                 return ret;
1609
1610         hv_driver->driver.name = hv_driver->name;
1611         hv_driver->driver.owner = owner;
1612         hv_driver->driver.mod_name = mod_name;
1613         hv_driver->driver.bus = &hv_bus;
1614
1615         spin_lock_init(&hv_driver->dynids.lock);
1616         INIT_LIST_HEAD(&hv_driver->dynids.list);
1617
1618         ret = driver_register(&hv_driver->driver);
1619
1620         return ret;
1621 }
1622 EXPORT_SYMBOL_GPL(__vmbus_driver_register);
1623
1624 /**
1625  * vmbus_driver_unregister() - Unregister a vmbus's driver
1626  * @hv_driver: Pointer to driver structure you want to
1627  *             un-register
1628  *
1629  * Un-register the given driver that was previous registered with a call to
1630  * vmbus_driver_register()
1631  */
1632 void vmbus_driver_unregister(struct hv_driver *hv_driver)
1633 {
1634         pr_info("unregistering driver %s\n", hv_driver->name);
1635
1636         if (!vmbus_exists()) {
1637                 driver_unregister(&hv_driver->driver);
1638                 vmbus_free_dynids(hv_driver);
1639         }
1640 }
1641 EXPORT_SYMBOL_GPL(vmbus_driver_unregister);
1642
1643
1644 /*
1645  * Called when last reference to channel is gone.
1646  */
1647 static void vmbus_chan_release(struct kobject *kobj)
1648 {
1649         struct vmbus_channel *channel
1650                 = container_of(kobj, struct vmbus_channel, kobj);
1651
1652         kfree_rcu(channel, rcu);
1653 }
1654
1655 struct vmbus_chan_attribute {
1656         struct attribute attr;
1657         ssize_t (*show)(struct vmbus_channel *chan, char *buf);
1658         ssize_t (*store)(struct vmbus_channel *chan,
1659                          const char *buf, size_t count);
1660 };
1661 #define VMBUS_CHAN_ATTR(_name, _mode, _show, _store) \
1662         struct vmbus_chan_attribute chan_attr_##_name \
1663                 = __ATTR(_name, _mode, _show, _store)
1664 #define VMBUS_CHAN_ATTR_RW(_name) \
1665         struct vmbus_chan_attribute chan_attr_##_name = __ATTR_RW(_name)
1666 #define VMBUS_CHAN_ATTR_RO(_name) \
1667         struct vmbus_chan_attribute chan_attr_##_name = __ATTR_RO(_name)
1668 #define VMBUS_CHAN_ATTR_WO(_name) \
1669         struct vmbus_chan_attribute chan_attr_##_name = __ATTR_WO(_name)
1670
1671 static ssize_t vmbus_chan_attr_show(struct kobject *kobj,
1672                                     struct attribute *attr, char *buf)
1673 {
1674         const struct vmbus_chan_attribute *attribute
1675                 = container_of(attr, struct vmbus_chan_attribute, attr);
1676         struct vmbus_channel *chan
1677                 = container_of(kobj, struct vmbus_channel, kobj);
1678
1679         if (!attribute->show)
1680                 return -EIO;
1681
1682         return attribute->show(chan, buf);
1683 }
1684
1685 static ssize_t vmbus_chan_attr_store(struct kobject *kobj,
1686                                      struct attribute *attr, const char *buf,
1687                                      size_t count)
1688 {
1689         const struct vmbus_chan_attribute *attribute
1690                 = container_of(attr, struct vmbus_chan_attribute, attr);
1691         struct vmbus_channel *chan
1692                 = container_of(kobj, struct vmbus_channel, kobj);
1693
1694         if (!attribute->store)
1695                 return -EIO;
1696
1697         return attribute->store(chan, buf, count);
1698 }
1699
1700 static const struct sysfs_ops vmbus_chan_sysfs_ops = {
1701         .show = vmbus_chan_attr_show,
1702         .store = vmbus_chan_attr_store,
1703 };
1704
1705 static ssize_t out_mask_show(struct vmbus_channel *channel, char *buf)
1706 {
1707         struct hv_ring_buffer_info *rbi = &channel->outbound;
1708         ssize_t ret;
1709
1710         mutex_lock(&rbi->ring_buffer_mutex);
1711         if (!rbi->ring_buffer) {
1712                 mutex_unlock(&rbi->ring_buffer_mutex);
1713                 return -EINVAL;
1714         }
1715
1716         ret = sprintf(buf, "%u\n", rbi->ring_buffer->interrupt_mask);
1717         mutex_unlock(&rbi->ring_buffer_mutex);
1718         return ret;
1719 }
1720 static VMBUS_CHAN_ATTR_RO(out_mask);
1721
1722 static ssize_t in_mask_show(struct vmbus_channel *channel, char *buf)
1723 {
1724         struct hv_ring_buffer_info *rbi = &channel->inbound;
1725         ssize_t ret;
1726
1727         mutex_lock(&rbi->ring_buffer_mutex);
1728         if (!rbi->ring_buffer) {
1729                 mutex_unlock(&rbi->ring_buffer_mutex);
1730                 return -EINVAL;
1731         }
1732
1733         ret = sprintf(buf, "%u\n", rbi->ring_buffer->interrupt_mask);
1734         mutex_unlock(&rbi->ring_buffer_mutex);
1735         return ret;
1736 }
1737 static VMBUS_CHAN_ATTR_RO(in_mask);
1738
1739 static ssize_t read_avail_show(struct vmbus_channel *channel, char *buf)
1740 {
1741         struct hv_ring_buffer_info *rbi = &channel->inbound;
1742         ssize_t ret;
1743
1744         mutex_lock(&rbi->ring_buffer_mutex);
1745         if (!rbi->ring_buffer) {
1746                 mutex_unlock(&rbi->ring_buffer_mutex);
1747                 return -EINVAL;
1748         }
1749
1750         ret = sprintf(buf, "%u\n", hv_get_bytes_to_read(rbi));
1751         mutex_unlock(&rbi->ring_buffer_mutex);
1752         return ret;
1753 }
1754 static VMBUS_CHAN_ATTR_RO(read_avail);
1755
1756 static ssize_t write_avail_show(struct vmbus_channel *channel, char *buf)
1757 {
1758         struct hv_ring_buffer_info *rbi = &channel->outbound;
1759         ssize_t ret;
1760
1761         mutex_lock(&rbi->ring_buffer_mutex);
1762         if (!rbi->ring_buffer) {
1763                 mutex_unlock(&rbi->ring_buffer_mutex);
1764                 return -EINVAL;
1765         }
1766
1767         ret = sprintf(buf, "%u\n", hv_get_bytes_to_write(rbi));
1768         mutex_unlock(&rbi->ring_buffer_mutex);
1769         return ret;
1770 }
1771 static VMBUS_CHAN_ATTR_RO(write_avail);
1772
1773 static ssize_t target_cpu_show(struct vmbus_channel *channel, char *buf)
1774 {
1775         return sprintf(buf, "%u\n", channel->target_cpu);
1776 }
1777 static ssize_t target_cpu_store(struct vmbus_channel *channel,
1778                                 const char *buf, size_t count)
1779 {
1780         u32 target_cpu, origin_cpu;
1781         ssize_t ret = count;
1782
1783         if (vmbus_proto_version < VERSION_WIN10_V4_1)
1784                 return -EIO;
1785
1786         if (sscanf(buf, "%uu", &target_cpu) != 1)
1787                 return -EIO;
1788
1789         /* Validate target_cpu for the cpumask_test_cpu() operation below. */
1790         if (target_cpu >= nr_cpumask_bits)
1791                 return -EINVAL;
1792
1793         /* No CPUs should come up or down during this. */
1794         cpus_read_lock();
1795
1796         if (!cpu_online(target_cpu)) {
1797                 cpus_read_unlock();
1798                 return -EINVAL;
1799         }
1800
1801         /*
1802          * Synchronizes target_cpu_store() and channel closure:
1803          *
1804          * { Initially: state = CHANNEL_OPENED }
1805          *
1806          * CPU1                         CPU2
1807          *
1808          * [target_cpu_store()]         [vmbus_disconnect_ring()]
1809          *
1810          * LOCK channel_mutex           LOCK channel_mutex
1811          * LOAD r1 = state              LOAD r2 = state
1812          * IF (r1 == CHANNEL_OPENED)    IF (r2 == CHANNEL_OPENED)
1813          *   SEND MODIFYCHANNEL           STORE state = CHANNEL_OPEN
1814          *   [...]                        SEND CLOSECHANNEL
1815          * UNLOCK channel_mutex         UNLOCK channel_mutex
1816          *
1817          * Forbids: r1 == r2 == CHANNEL_OPENED (i.e., CPU1's LOCK precedes
1818          *              CPU2's LOCK) && CPU2's SEND precedes CPU1's SEND
1819          *
1820          * Note.  The host processes the channel messages "sequentially", in
1821          * the order in which they are received on a per-partition basis.
1822          */
1823         mutex_lock(&vmbus_connection.channel_mutex);
1824
1825         /*
1826          * Hyper-V will ignore MODIFYCHANNEL messages for "non-open" channels;
1827          * avoid sending the message and fail here for such channels.
1828          */
1829         if (channel->state != CHANNEL_OPENED_STATE) {
1830                 ret = -EIO;
1831                 goto cpu_store_unlock;
1832         }
1833
1834         origin_cpu = channel->target_cpu;
1835         if (target_cpu == origin_cpu)
1836                 goto cpu_store_unlock;
1837
1838         if (vmbus_send_modifychannel(channel,
1839                                      hv_cpu_number_to_vp_number(target_cpu))) {
1840                 ret = -EIO;
1841                 goto cpu_store_unlock;
1842         }
1843
1844         /*
1845          * For version before VERSION_WIN10_V5_3, the following warning holds:
1846          *
1847          * Warning.  At this point, there is *no* guarantee that the host will
1848          * have successfully processed the vmbus_send_modifychannel() request.
1849          * See the header comment of vmbus_send_modifychannel() for more info.
1850          *
1851          * Lags in the processing of the above vmbus_send_modifychannel() can
1852          * result in missed interrupts if the "old" target CPU is taken offline
1853          * before Hyper-V starts sending interrupts to the "new" target CPU.
1854          * But apart from this offlining scenario, the code tolerates such
1855          * lags.  It will function correctly even if a channel interrupt comes
1856          * in on a CPU that is different from the channel target_cpu value.
1857          */
1858
1859         channel->target_cpu = target_cpu;
1860
1861         /* See init_vp_index(). */
1862         if (hv_is_perf_channel(channel))
1863                 hv_update_allocated_cpus(origin_cpu, target_cpu);
1864
1865         /* Currently set only for storvsc channels. */
1866         if (channel->change_target_cpu_callback) {
1867                 (*channel->change_target_cpu_callback)(channel,
1868                                 origin_cpu, target_cpu);
1869         }
1870
1871 cpu_store_unlock:
1872         mutex_unlock(&vmbus_connection.channel_mutex);
1873         cpus_read_unlock();
1874         return ret;
1875 }
1876 static VMBUS_CHAN_ATTR(cpu, 0644, target_cpu_show, target_cpu_store);
1877
1878 static ssize_t channel_pending_show(struct vmbus_channel *channel,
1879                                     char *buf)
1880 {
1881         return sprintf(buf, "%d\n",
1882                        channel_pending(channel,
1883                                        vmbus_connection.monitor_pages[1]));
1884 }
1885 static VMBUS_CHAN_ATTR(pending, 0444, channel_pending_show, NULL);
1886
1887 static ssize_t channel_latency_show(struct vmbus_channel *channel,
1888                                     char *buf)
1889 {
1890         return sprintf(buf, "%d\n",
1891                        channel_latency(channel,
1892                                        vmbus_connection.monitor_pages[1]));
1893 }
1894 static VMBUS_CHAN_ATTR(latency, 0444, channel_latency_show, NULL);
1895
1896 static ssize_t channel_interrupts_show(struct vmbus_channel *channel, char *buf)
1897 {
1898         return sprintf(buf, "%llu\n", channel->interrupts);
1899 }
1900 static VMBUS_CHAN_ATTR(interrupts, 0444, channel_interrupts_show, NULL);
1901
1902 static ssize_t channel_events_show(struct vmbus_channel *channel, char *buf)
1903 {
1904         return sprintf(buf, "%llu\n", channel->sig_events);
1905 }
1906 static VMBUS_CHAN_ATTR(events, 0444, channel_events_show, NULL);
1907
1908 static ssize_t channel_intr_in_full_show(struct vmbus_channel *channel,
1909                                          char *buf)
1910 {
1911         return sprintf(buf, "%llu\n",
1912                        (unsigned long long)channel->intr_in_full);
1913 }
1914 static VMBUS_CHAN_ATTR(intr_in_full, 0444, channel_intr_in_full_show, NULL);
1915
1916 static ssize_t channel_intr_out_empty_show(struct vmbus_channel *channel,
1917                                            char *buf)
1918 {
1919         return sprintf(buf, "%llu\n",
1920                        (unsigned long long)channel->intr_out_empty);
1921 }
1922 static VMBUS_CHAN_ATTR(intr_out_empty, 0444, channel_intr_out_empty_show, NULL);
1923
1924 static ssize_t channel_out_full_first_show(struct vmbus_channel *channel,
1925                                            char *buf)
1926 {
1927         return sprintf(buf, "%llu\n",
1928                        (unsigned long long)channel->out_full_first);
1929 }
1930 static VMBUS_CHAN_ATTR(out_full_first, 0444, channel_out_full_first_show, NULL);
1931
1932 static ssize_t channel_out_full_total_show(struct vmbus_channel *channel,
1933                                            char *buf)
1934 {
1935         return sprintf(buf, "%llu\n",
1936                        (unsigned long long)channel->out_full_total);
1937 }
1938 static VMBUS_CHAN_ATTR(out_full_total, 0444, channel_out_full_total_show, NULL);
1939
1940 static ssize_t subchannel_monitor_id_show(struct vmbus_channel *channel,
1941                                           char *buf)
1942 {
1943         return sprintf(buf, "%u\n", channel->offermsg.monitorid);
1944 }
1945 static VMBUS_CHAN_ATTR(monitor_id, 0444, subchannel_monitor_id_show, NULL);
1946
1947 static ssize_t subchannel_id_show(struct vmbus_channel *channel,
1948                                   char *buf)
1949 {
1950         return sprintf(buf, "%u\n",
1951                        channel->offermsg.offer.sub_channel_index);
1952 }
1953 static VMBUS_CHAN_ATTR_RO(subchannel_id);
1954
1955 static struct attribute *vmbus_chan_attrs[] = {
1956         &chan_attr_out_mask.attr,
1957         &chan_attr_in_mask.attr,
1958         &chan_attr_read_avail.attr,
1959         &chan_attr_write_avail.attr,
1960         &chan_attr_cpu.attr,
1961         &chan_attr_pending.attr,
1962         &chan_attr_latency.attr,
1963         &chan_attr_interrupts.attr,
1964         &chan_attr_events.attr,
1965         &chan_attr_intr_in_full.attr,
1966         &chan_attr_intr_out_empty.attr,
1967         &chan_attr_out_full_first.attr,
1968         &chan_attr_out_full_total.attr,
1969         &chan_attr_monitor_id.attr,
1970         &chan_attr_subchannel_id.attr,
1971         NULL
1972 };
1973
1974 /*
1975  * Channel-level attribute_group callback function. Returns the permission for
1976  * each attribute, and returns 0 if an attribute is not visible.
1977  */
1978 static umode_t vmbus_chan_attr_is_visible(struct kobject *kobj,
1979                                           struct attribute *attr, int idx)
1980 {
1981         const struct vmbus_channel *channel =
1982                 container_of(kobj, struct vmbus_channel, kobj);
1983
1984         /* Hide the monitor attributes if the monitor mechanism is not used. */
1985         if (!channel->offermsg.monitor_allocated &&
1986             (attr == &chan_attr_pending.attr ||
1987              attr == &chan_attr_latency.attr ||
1988              attr == &chan_attr_monitor_id.attr))
1989                 return 0;
1990
1991         return attr->mode;
1992 }
1993
1994 static struct attribute_group vmbus_chan_group = {
1995         .attrs = vmbus_chan_attrs,
1996         .is_visible = vmbus_chan_attr_is_visible
1997 };
1998
1999 static struct kobj_type vmbus_chan_ktype = {
2000         .sysfs_ops = &vmbus_chan_sysfs_ops,
2001         .release = vmbus_chan_release,
2002 };
2003
2004 /*
2005  * vmbus_add_channel_kobj - setup a sub-directory under device/channels
2006  */
2007 int vmbus_add_channel_kobj(struct hv_device *dev, struct vmbus_channel *channel)
2008 {
2009         const struct device *device = &dev->device;
2010         struct kobject *kobj = &channel->kobj;
2011         u32 relid = channel->offermsg.child_relid;
2012         int ret;
2013
2014         kobj->kset = dev->channels_kset;
2015         ret = kobject_init_and_add(kobj, &vmbus_chan_ktype, NULL,
2016                                    "%u", relid);
2017         if (ret) {
2018                 kobject_put(kobj);
2019                 return ret;
2020         }
2021
2022         ret = sysfs_create_group(kobj, &vmbus_chan_group);
2023
2024         if (ret) {
2025                 /*
2026                  * The calling functions' error handling paths will cleanup the
2027                  * empty channel directory.
2028                  */
2029                 kobject_put(kobj);
2030                 dev_err(device, "Unable to set up channel sysfs files\n");
2031                 return ret;
2032         }
2033
2034         kobject_uevent(kobj, KOBJ_ADD);
2035
2036         return 0;
2037 }
2038
2039 /*
2040  * vmbus_remove_channel_attr_group - remove the channel's attribute group
2041  */
2042 void vmbus_remove_channel_attr_group(struct vmbus_channel *channel)
2043 {
2044         sysfs_remove_group(&channel->kobj, &vmbus_chan_group);
2045 }
2046
2047 /*
2048  * vmbus_device_create - Creates and registers a new child device
2049  * on the vmbus.
2050  */
2051 struct hv_device *vmbus_device_create(const guid_t *type,
2052                                       const guid_t *instance,
2053                                       struct vmbus_channel *channel)
2054 {
2055         struct hv_device *child_device_obj;
2056
2057         child_device_obj = kzalloc(sizeof(struct hv_device), GFP_KERNEL);
2058         if (!child_device_obj) {
2059                 pr_err("Unable to allocate device object for child device\n");
2060                 return NULL;
2061         }
2062
2063         child_device_obj->channel = channel;
2064         guid_copy(&child_device_obj->dev_type, type);
2065         guid_copy(&child_device_obj->dev_instance, instance);
2066         child_device_obj->vendor_id = 0x1414; /* MSFT vendor ID */
2067
2068         return child_device_obj;
2069 }
2070
2071 /*
2072  * vmbus_device_register - Register the child device
2073  */
2074 int vmbus_device_register(struct hv_device *child_device_obj)
2075 {
2076         struct kobject *kobj = &child_device_obj->device.kobj;
2077         int ret;
2078
2079         dev_set_name(&child_device_obj->device, "%pUl",
2080                      &child_device_obj->channel->offermsg.offer.if_instance);
2081
2082         child_device_obj->device.bus = &hv_bus;
2083         child_device_obj->device.parent = &hv_acpi_dev->dev;
2084         child_device_obj->device.release = vmbus_device_release;
2085
2086         child_device_obj->device.dma_parms = &child_device_obj->dma_parms;
2087         child_device_obj->device.dma_mask = &child_device_obj->dma_mask;
2088         dma_set_mask(&child_device_obj->device, DMA_BIT_MASK(64));
2089
2090         /*
2091          * Register with the LDM. This will kick off the driver/device
2092          * binding...which will eventually call vmbus_match() and vmbus_probe()
2093          */
2094         ret = device_register(&child_device_obj->device);
2095         if (ret) {
2096                 pr_err("Unable to register child device\n");
2097                 return ret;
2098         }
2099
2100         child_device_obj->channels_kset = kset_create_and_add("channels",
2101                                                               NULL, kobj);
2102         if (!child_device_obj->channels_kset) {
2103                 ret = -ENOMEM;
2104                 goto err_dev_unregister;
2105         }
2106
2107         ret = vmbus_add_channel_kobj(child_device_obj,
2108                                      child_device_obj->channel);
2109         if (ret) {
2110                 pr_err("Unable to register primary channeln");
2111                 goto err_kset_unregister;
2112         }
2113         hv_debug_add_dev_dir(child_device_obj);
2114
2115         return 0;
2116
2117 err_kset_unregister:
2118         kset_unregister(child_device_obj->channels_kset);
2119
2120 err_dev_unregister:
2121         device_unregister(&child_device_obj->device);
2122         return ret;
2123 }
2124
2125 /*
2126  * vmbus_device_unregister - Remove the specified child device
2127  * from the vmbus.
2128  */
2129 void vmbus_device_unregister(struct hv_device *device_obj)
2130 {
2131         pr_debug("child device %s unregistered\n",
2132                 dev_name(&device_obj->device));
2133
2134         kset_unregister(device_obj->channels_kset);
2135
2136         /*
2137          * Kick off the process of unregistering the device.
2138          * This will call vmbus_remove() and eventually vmbus_device_release()
2139          */
2140         device_unregister(&device_obj->device);
2141 }
2142
2143
2144 /*
2145  * VMBUS is an acpi enumerated device. Get the information we
2146  * need from DSDT.
2147  */
2148 #define VTPM_BASE_ADDRESS 0xfed40000
2149 static acpi_status vmbus_walk_resources(struct acpi_resource *res, void *ctx)
2150 {
2151         resource_size_t start = 0;
2152         resource_size_t end = 0;
2153         struct resource *new_res;
2154         struct resource **old_res = &hyperv_mmio;
2155         struct resource **prev_res = NULL;
2156         struct resource r;
2157
2158         switch (res->type) {
2159
2160         /*
2161          * "Address" descriptors are for bus windows. Ignore
2162          * "memory" descriptors, which are for registers on
2163          * devices.
2164          */
2165         case ACPI_RESOURCE_TYPE_ADDRESS32:
2166                 start = res->data.address32.address.minimum;
2167                 end = res->data.address32.address.maximum;
2168                 break;
2169
2170         case ACPI_RESOURCE_TYPE_ADDRESS64:
2171                 start = res->data.address64.address.minimum;
2172                 end = res->data.address64.address.maximum;
2173                 break;
2174
2175         /*
2176          * The IRQ information is needed only on ARM64, which Hyper-V
2177          * sets up in the extended format. IRQ information is present
2178          * on x86/x64 in the non-extended format but it is not used by
2179          * Linux. So don't bother checking for the non-extended format.
2180          */
2181         case ACPI_RESOURCE_TYPE_EXTENDED_IRQ:
2182                 if (!acpi_dev_resource_interrupt(res, 0, &r)) {
2183                         pr_err("Unable to parse Hyper-V ACPI interrupt\n");
2184                         return AE_ERROR;
2185                 }
2186                 /* ARM64 INTID for VMbus */
2187                 vmbus_interrupt = res->data.extended_irq.interrupts[0];
2188                 /* Linux IRQ number */
2189                 vmbus_irq = r.start;
2190                 return AE_OK;
2191
2192         default:
2193                 /* Unused resource type */
2194                 return AE_OK;
2195
2196         }
2197         /*
2198          * Ignore ranges that are below 1MB, as they're not
2199          * necessary or useful here.
2200          */
2201         if (end < 0x100000)
2202                 return AE_OK;
2203
2204         new_res = kzalloc(sizeof(*new_res), GFP_ATOMIC);
2205         if (!new_res)
2206                 return AE_NO_MEMORY;
2207
2208         /* If this range overlaps the virtual TPM, truncate it. */
2209         if (end > VTPM_BASE_ADDRESS && start < VTPM_BASE_ADDRESS)
2210                 end = VTPM_BASE_ADDRESS;
2211
2212         new_res->name = "hyperv mmio";
2213         new_res->flags = IORESOURCE_MEM;
2214         new_res->start = start;
2215         new_res->end = end;
2216
2217         /*
2218          * If two ranges are adjacent, merge them.
2219          */
2220         do {
2221                 if (!*old_res) {
2222                         *old_res = new_res;
2223                         break;
2224                 }
2225
2226                 if (((*old_res)->end + 1) == new_res->start) {
2227                         (*old_res)->end = new_res->end;
2228                         kfree(new_res);
2229                         break;
2230                 }
2231
2232                 if ((*old_res)->start == new_res->end + 1) {
2233                         (*old_res)->start = new_res->start;
2234                         kfree(new_res);
2235                         break;
2236                 }
2237
2238                 if ((*old_res)->start > new_res->end) {
2239                         new_res->sibling = *old_res;
2240                         if (prev_res)
2241                                 (*prev_res)->sibling = new_res;
2242                         *old_res = new_res;
2243                         break;
2244                 }
2245
2246                 prev_res = old_res;
2247                 old_res = &(*old_res)->sibling;
2248
2249         } while (1);
2250
2251         return AE_OK;
2252 }
2253
2254 static int vmbus_acpi_remove(struct acpi_device *device)
2255 {
2256         struct resource *cur_res;
2257         struct resource *next_res;
2258
2259         if (hyperv_mmio) {
2260                 if (fb_mmio) {
2261                         __release_region(hyperv_mmio, fb_mmio->start,
2262                                          resource_size(fb_mmio));
2263                         fb_mmio = NULL;
2264                 }
2265
2266                 for (cur_res = hyperv_mmio; cur_res; cur_res = next_res) {
2267                         next_res = cur_res->sibling;
2268                         kfree(cur_res);
2269                 }
2270         }
2271
2272         return 0;
2273 }
2274
2275 static void vmbus_reserve_fb(void)
2276 {
2277         int size;
2278         /*
2279          * Make a claim for the frame buffer in the resource tree under the
2280          * first node, which will be the one below 4GB.  The length seems to
2281          * be underreported, particularly in a Generation 1 VM.  So start out
2282          * reserving a larger area and make it smaller until it succeeds.
2283          */
2284
2285         if (screen_info.lfb_base) {
2286                 if (efi_enabled(EFI_BOOT))
2287                         size = max_t(__u32, screen_info.lfb_size, 0x800000);
2288                 else
2289                         size = max_t(__u32, screen_info.lfb_size, 0x4000000);
2290
2291                 for (; !fb_mmio && (size >= 0x100000); size >>= 1) {
2292                         fb_mmio = __request_region(hyperv_mmio,
2293                                                    screen_info.lfb_base, size,
2294                                                    fb_mmio_name, 0);
2295                 }
2296         }
2297 }
2298
2299 /**
2300  * vmbus_allocate_mmio() - Pick a memory-mapped I/O range.
2301  * @new:                If successful, supplied a pointer to the
2302  *                      allocated MMIO space.
2303  * @device_obj:         Identifies the caller
2304  * @min:                Minimum guest physical address of the
2305  *                      allocation
2306  * @max:                Maximum guest physical address
2307  * @size:               Size of the range to be allocated
2308  * @align:              Alignment of the range to be allocated
2309  * @fb_overlap_ok:      Whether this allocation can be allowed
2310  *                      to overlap the video frame buffer.
2311  *
2312  * This function walks the resources granted to VMBus by the
2313  * _CRS object in the ACPI namespace underneath the parent
2314  * "bridge" whether that's a root PCI bus in the Generation 1
2315  * case or a Module Device in the Generation 2 case.  It then
2316  * attempts to allocate from the global MMIO pool in a way that
2317  * matches the constraints supplied in these parameters and by
2318  * that _CRS.
2319  *
2320  * Return: 0 on success, -errno on failure
2321  */
2322 int vmbus_allocate_mmio(struct resource **new, struct hv_device *device_obj,
2323                         resource_size_t min, resource_size_t max,
2324                         resource_size_t size, resource_size_t align,
2325                         bool fb_overlap_ok)
2326 {
2327         struct resource *iter, *shadow;
2328         resource_size_t range_min, range_max, start;
2329         const char *dev_n = dev_name(&device_obj->device);
2330         int retval;
2331
2332         retval = -ENXIO;
2333         mutex_lock(&hyperv_mmio_lock);
2334
2335         /*
2336          * If overlaps with frame buffers are allowed, then first attempt to
2337          * make the allocation from within the reserved region.  Because it
2338          * is already reserved, no shadow allocation is necessary.
2339          */
2340         if (fb_overlap_ok && fb_mmio && !(min > fb_mmio->end) &&
2341             !(max < fb_mmio->start)) {
2342
2343                 range_min = fb_mmio->start;
2344                 range_max = fb_mmio->end;
2345                 start = (range_min + align - 1) & ~(align - 1);
2346                 for (; start + size - 1 <= range_max; start += align) {
2347                         *new = request_mem_region_exclusive(start, size, dev_n);
2348                         if (*new) {
2349                                 retval = 0;
2350                                 goto exit;
2351                         }
2352                 }
2353         }
2354
2355         for (iter = hyperv_mmio; iter; iter = iter->sibling) {
2356                 if ((iter->start >= max) || (iter->end <= min))
2357                         continue;
2358
2359                 range_min = iter->start;
2360                 range_max = iter->end;
2361                 start = (range_min + align - 1) & ~(align - 1);
2362                 for (; start + size - 1 <= range_max; start += align) {
2363                         shadow = __request_region(iter, start, size, NULL,
2364                                                   IORESOURCE_BUSY);
2365                         if (!shadow)
2366                                 continue;
2367
2368                         *new = request_mem_region_exclusive(start, size, dev_n);
2369                         if (*new) {
2370                                 shadow->name = (char *)*new;
2371                                 retval = 0;
2372                                 goto exit;
2373                         }
2374
2375                         __release_region(iter, start, size);
2376                 }
2377         }
2378
2379 exit:
2380         mutex_unlock(&hyperv_mmio_lock);
2381         return retval;
2382 }
2383 EXPORT_SYMBOL_GPL(vmbus_allocate_mmio);
2384
2385 /**
2386  * vmbus_free_mmio() - Free a memory-mapped I/O range.
2387  * @start:              Base address of region to release.
2388  * @size:               Size of the range to be allocated
2389  *
2390  * This function releases anything requested by
2391  * vmbus_mmio_allocate().
2392  */
2393 void vmbus_free_mmio(resource_size_t start, resource_size_t size)
2394 {
2395         struct resource *iter;
2396
2397         mutex_lock(&hyperv_mmio_lock);
2398         for (iter = hyperv_mmio; iter; iter = iter->sibling) {
2399                 if ((iter->start >= start + size) || (iter->end <= start))
2400                         continue;
2401
2402                 __release_region(iter, start, size);
2403         }
2404         release_mem_region(start, size);
2405         mutex_unlock(&hyperv_mmio_lock);
2406
2407 }
2408 EXPORT_SYMBOL_GPL(vmbus_free_mmio);
2409
2410 static int vmbus_acpi_add(struct acpi_device *device)
2411 {
2412         acpi_status result;
2413         int ret_val = -ENODEV;
2414         struct acpi_device *ancestor;
2415
2416         hv_acpi_dev = device;
2417
2418         /*
2419          * Older versions of Hyper-V for ARM64 fail to include the _CCA
2420          * method on the top level VMbus device in the DSDT. But devices
2421          * are hardware coherent in all current Hyper-V use cases, so fix
2422          * up the ACPI device to behave as if _CCA is present and indicates
2423          * hardware coherence.
2424          */
2425         ACPI_COMPANION_SET(&device->dev, device);
2426         if (IS_ENABLED(CONFIG_ACPI_CCA_REQUIRED) &&
2427             device_get_dma_attr(&device->dev) == DEV_DMA_NOT_SUPPORTED) {
2428                 pr_info("No ACPI _CCA found; assuming coherent device I/O\n");
2429                 device->flags.cca_seen = true;
2430                 device->flags.coherent_dma = true;
2431         }
2432
2433         result = acpi_walk_resources(device->handle, METHOD_NAME__CRS,
2434                                         vmbus_walk_resources, NULL);
2435
2436         if (ACPI_FAILURE(result))
2437                 goto acpi_walk_err;
2438         /*
2439          * Some ancestor of the vmbus acpi device (Gen1 or Gen2
2440          * firmware) is the VMOD that has the mmio ranges. Get that.
2441          */
2442         for (ancestor = device->parent; ancestor; ancestor = ancestor->parent) {
2443                 result = acpi_walk_resources(ancestor->handle, METHOD_NAME__CRS,
2444                                              vmbus_walk_resources, NULL);
2445
2446                 if (ACPI_FAILURE(result))
2447                         continue;
2448                 if (hyperv_mmio) {
2449                         vmbus_reserve_fb();
2450                         break;
2451                 }
2452         }
2453         ret_val = 0;
2454
2455 acpi_walk_err:
2456         complete(&probe_event);
2457         if (ret_val)
2458                 vmbus_acpi_remove(device);
2459         return ret_val;
2460 }
2461
2462 #ifdef CONFIG_PM_SLEEP
2463 static int vmbus_bus_suspend(struct device *dev)
2464 {
2465         struct vmbus_channel *channel, *sc;
2466
2467         while (atomic_read(&vmbus_connection.offer_in_progress) != 0) {
2468                 /*
2469                  * We wait here until the completion of any channel
2470                  * offers that are currently in progress.
2471                  */
2472                 usleep_range(1000, 2000);
2473         }
2474
2475         mutex_lock(&vmbus_connection.channel_mutex);
2476         list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
2477                 if (!is_hvsock_channel(channel))
2478                         continue;
2479
2480                 vmbus_force_channel_rescinded(channel);
2481         }
2482         mutex_unlock(&vmbus_connection.channel_mutex);
2483
2484         /*
2485          * Wait until all the sub-channels and hv_sock channels have been
2486          * cleaned up. Sub-channels should be destroyed upon suspend, otherwise
2487          * they would conflict with the new sub-channels that will be created
2488          * in the resume path. hv_sock channels should also be destroyed, but
2489          * a hv_sock channel of an established hv_sock connection can not be
2490          * really destroyed since it may still be referenced by the userspace
2491          * application, so we just force the hv_sock channel to be rescinded
2492          * by vmbus_force_channel_rescinded(), and the userspace application
2493          * will thoroughly destroy the channel after hibernation.
2494          *
2495          * Note: the counter nr_chan_close_on_suspend may never go above 0 if
2496          * the VM has no sub-channel and hv_sock channel, e.g. a 1-vCPU VM.
2497          */
2498         if (atomic_read(&vmbus_connection.nr_chan_close_on_suspend) > 0)
2499                 wait_for_completion(&vmbus_connection.ready_for_suspend_event);
2500
2501         if (atomic_read(&vmbus_connection.nr_chan_fixup_on_resume) != 0) {
2502                 pr_err("Can not suspend due to a previous failed resuming\n");
2503                 return -EBUSY;
2504         }
2505
2506         mutex_lock(&vmbus_connection.channel_mutex);
2507
2508         list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
2509                 /*
2510                  * Remove the channel from the array of channels and invalidate
2511                  * the channel's relid.  Upon resume, vmbus_onoffer() will fix
2512                  * up the relid (and other fields, if necessary) and add the
2513                  * channel back to the array.
2514                  */
2515                 vmbus_channel_unmap_relid(channel);
2516                 channel->offermsg.child_relid = INVALID_RELID;
2517
2518                 if (is_hvsock_channel(channel)) {
2519                         if (!channel->rescind) {
2520                                 pr_err("hv_sock channel not rescinded!\n");
2521                                 WARN_ON_ONCE(1);
2522                         }
2523                         continue;
2524                 }
2525
2526                 list_for_each_entry(sc, &channel->sc_list, sc_list) {
2527                         pr_err("Sub-channel not deleted!\n");
2528                         WARN_ON_ONCE(1);
2529                 }
2530
2531                 atomic_inc(&vmbus_connection.nr_chan_fixup_on_resume);
2532         }
2533
2534         mutex_unlock(&vmbus_connection.channel_mutex);
2535
2536         vmbus_initiate_unload(false);
2537
2538         /* Reset the event for the next resume. */
2539         reinit_completion(&vmbus_connection.ready_for_resume_event);
2540
2541         return 0;
2542 }
2543
2544 static int vmbus_bus_resume(struct device *dev)
2545 {
2546         struct vmbus_channel_msginfo *msginfo;
2547         size_t msgsize;
2548         int ret;
2549
2550         /*
2551          * We only use the 'vmbus_proto_version', which was in use before
2552          * hibernation, to re-negotiate with the host.
2553          */
2554         if (!vmbus_proto_version) {
2555                 pr_err("Invalid proto version = 0x%x\n", vmbus_proto_version);
2556                 return -EINVAL;
2557         }
2558
2559         msgsize = sizeof(*msginfo) +
2560                   sizeof(struct vmbus_channel_initiate_contact);
2561
2562         msginfo = kzalloc(msgsize, GFP_KERNEL);
2563
2564         if (msginfo == NULL)
2565                 return -ENOMEM;
2566
2567         ret = vmbus_negotiate_version(msginfo, vmbus_proto_version);
2568
2569         kfree(msginfo);
2570
2571         if (ret != 0)
2572                 return ret;
2573
2574         WARN_ON(atomic_read(&vmbus_connection.nr_chan_fixup_on_resume) == 0);
2575
2576         vmbus_request_offers();
2577
2578         if (wait_for_completion_timeout(
2579                 &vmbus_connection.ready_for_resume_event, 10 * HZ) == 0)
2580                 pr_err("Some vmbus device is missing after suspending?\n");
2581
2582         /* Reset the event for the next suspend. */
2583         reinit_completion(&vmbus_connection.ready_for_suspend_event);
2584
2585         return 0;
2586 }
2587 #else
2588 #define vmbus_bus_suspend NULL
2589 #define vmbus_bus_resume NULL
2590 #endif /* CONFIG_PM_SLEEP */
2591
2592 static const struct acpi_device_id vmbus_acpi_device_ids[] = {
2593         {"VMBUS", 0},
2594         {"VMBus", 0},
2595         {"", 0},
2596 };
2597 MODULE_DEVICE_TABLE(acpi, vmbus_acpi_device_ids);
2598
2599 /*
2600  * Note: we must use the "no_irq" ops, otherwise hibernation can not work with
2601  * PCI device assignment, because "pci_dev_pm_ops" uses the "noirq" ops: in
2602  * the resume path, the pci "noirq" restore op runs before "non-noirq" op (see
2603  * resume_target_kernel() -> dpm_resume_start(), and hibernation_restore() ->
2604  * dpm_resume_end()). This means vmbus_bus_resume() and the pci-hyperv's
2605  * resume callback must also run via the "noirq" ops.
2606  *
2607  * Set suspend_noirq/resume_noirq to NULL for Suspend-to-Idle: see the comment
2608  * earlier in this file before vmbus_pm.
2609  */
2610
2611 static const struct dev_pm_ops vmbus_bus_pm = {
2612         .suspend_noirq  = NULL,
2613         .resume_noirq   = NULL,
2614         .freeze_noirq   = vmbus_bus_suspend,
2615         .thaw_noirq     = vmbus_bus_resume,
2616         .poweroff_noirq = vmbus_bus_suspend,
2617         .restore_noirq  = vmbus_bus_resume
2618 };
2619
2620 static struct acpi_driver vmbus_acpi_driver = {
2621         .name = "vmbus",
2622         .ids = vmbus_acpi_device_ids,
2623         .ops = {
2624                 .add = vmbus_acpi_add,
2625                 .remove = vmbus_acpi_remove,
2626         },
2627         .drv.pm = &vmbus_bus_pm,
2628 };
2629
2630 static void hv_kexec_handler(void)
2631 {
2632         hv_stimer_global_cleanup();
2633         vmbus_initiate_unload(false);
2634         /* Make sure conn_state is set as hv_synic_cleanup checks for it */
2635         mb();
2636         cpuhp_remove_state(hyperv_cpuhp_online);
2637 };
2638
2639 static void hv_crash_handler(struct pt_regs *regs)
2640 {
2641         int cpu;
2642
2643         vmbus_initiate_unload(true);
2644         /*
2645          * In crash handler we can't schedule synic cleanup for all CPUs,
2646          * doing the cleanup for current CPU only. This should be sufficient
2647          * for kdump.
2648          */
2649         cpu = smp_processor_id();
2650         hv_stimer_cleanup(cpu);
2651         hv_synic_disable_regs(cpu);
2652 };
2653
2654 static int hv_synic_suspend(void)
2655 {
2656         /*
2657          * When we reach here, all the non-boot CPUs have been offlined.
2658          * If we're in a legacy configuration where stimer Direct Mode is
2659          * not enabled, the stimers on the non-boot CPUs have been unbound
2660          * in hv_synic_cleanup() -> hv_stimer_legacy_cleanup() ->
2661          * hv_stimer_cleanup() -> clockevents_unbind_device().
2662          *
2663          * hv_synic_suspend() only runs on CPU0 with interrupts disabled.
2664          * Here we do not call hv_stimer_legacy_cleanup() on CPU0 because:
2665          * 1) it's unnecessary as interrupts remain disabled between
2666          * syscore_suspend() and syscore_resume(): see create_image() and
2667          * resume_target_kernel()
2668          * 2) the stimer on CPU0 is automatically disabled later by
2669          * syscore_suspend() -> timekeeping_suspend() -> tick_suspend() -> ...
2670          * -> clockevents_shutdown() -> ... -> hv_ce_shutdown()
2671          * 3) a warning would be triggered if we call
2672          * clockevents_unbind_device(), which may sleep, in an
2673          * interrupts-disabled context.
2674          */
2675
2676         hv_synic_disable_regs(0);
2677
2678         return 0;
2679 }
2680
2681 static void hv_synic_resume(void)
2682 {
2683         hv_synic_enable_regs(0);
2684
2685         /*
2686          * Note: we don't need to call hv_stimer_init(0), because the timer
2687          * on CPU0 is not unbound in hv_synic_suspend(), and the timer is
2688          * automatically re-enabled in timekeeping_resume().
2689          */
2690 }
2691
2692 /* The callbacks run only on CPU0, with irqs_disabled. */
2693 static struct syscore_ops hv_synic_syscore_ops = {
2694         .suspend = hv_synic_suspend,
2695         .resume = hv_synic_resume,
2696 };
2697
2698 static int __init hv_acpi_init(void)
2699 {
2700         int ret, t;
2701
2702         if (!hv_is_hyperv_initialized())
2703                 return -ENODEV;
2704
2705         if (hv_root_partition)
2706                 return 0;
2707
2708         init_completion(&probe_event);
2709
2710         /*
2711          * Get ACPI resources first.
2712          */
2713         ret = acpi_bus_register_driver(&vmbus_acpi_driver);
2714
2715         if (ret)
2716                 return ret;
2717
2718         t = wait_for_completion_timeout(&probe_event, 5*HZ);
2719         if (t == 0) {
2720                 ret = -ETIMEDOUT;
2721                 goto cleanup;
2722         }
2723
2724         /*
2725          * If we're on an architecture with a hardcoded hypervisor
2726          * vector (i.e. x86/x64), override the VMbus interrupt found
2727          * in the ACPI tables. Ensure vmbus_irq is not set since the
2728          * normal Linux IRQ mechanism is not used in this case.
2729          */
2730 #ifdef HYPERVISOR_CALLBACK_VECTOR
2731         vmbus_interrupt = HYPERVISOR_CALLBACK_VECTOR;
2732         vmbus_irq = -1;
2733 #endif
2734
2735         hv_debug_init();
2736
2737         ret = vmbus_bus_init();
2738         if (ret)
2739                 goto cleanup;
2740
2741         hv_setup_kexec_handler(hv_kexec_handler);
2742         hv_setup_crash_handler(hv_crash_handler);
2743
2744         register_syscore_ops(&hv_synic_syscore_ops);
2745
2746         return 0;
2747
2748 cleanup:
2749         acpi_bus_unregister_driver(&vmbus_acpi_driver);
2750         hv_acpi_dev = NULL;
2751         return ret;
2752 }
2753
2754 static void __exit vmbus_exit(void)
2755 {
2756         int cpu;
2757
2758         unregister_syscore_ops(&hv_synic_syscore_ops);
2759
2760         hv_remove_kexec_handler();
2761         hv_remove_crash_handler();
2762         vmbus_connection.conn_state = DISCONNECTED;
2763         hv_stimer_global_cleanup();
2764         vmbus_disconnect();
2765         if (vmbus_irq == -1) {
2766                 hv_remove_vmbus_handler();
2767         } else {
2768                 free_percpu_irq(vmbus_irq, vmbus_evt);
2769                 free_percpu(vmbus_evt);
2770         }
2771         for_each_online_cpu(cpu) {
2772                 struct hv_per_cpu_context *hv_cpu
2773                         = per_cpu_ptr(hv_context.cpu_context, cpu);
2774
2775                 tasklet_kill(&hv_cpu->msg_dpc);
2776         }
2777         hv_debug_rm_all_dir();
2778
2779         vmbus_free_channels();
2780         kfree(vmbus_connection.channels);
2781
2782         if (ms_hyperv.misc_features & HV_FEATURE_GUEST_CRASH_MSR_AVAILABLE) {
2783                 kmsg_dump_unregister(&hv_kmsg_dumper);
2784                 unregister_die_notifier(&hyperv_die_block);
2785         }
2786
2787         /*
2788          * The panic notifier is always registered, hence we should
2789          * also unconditionally unregister it here as well.
2790          */
2791         atomic_notifier_chain_unregister(&panic_notifier_list,
2792                                          &hyperv_panic_block);
2793
2794         free_page((unsigned long)hv_panic_page);
2795         unregister_sysctl_table(hv_ctl_table_hdr);
2796         hv_ctl_table_hdr = NULL;
2797         bus_unregister(&hv_bus);
2798
2799         cpuhp_remove_state(hyperv_cpuhp_online);
2800         hv_synic_free();
2801         acpi_bus_unregister_driver(&vmbus_acpi_driver);
2802 }
2803
2804
2805 MODULE_LICENSE("GPL");
2806 MODULE_DESCRIPTION("Microsoft Hyper-V VMBus Driver");
2807
2808 subsys_initcall(hv_acpi_init);
2809 module_exit(vmbus_exit);