drivers/hv/vmbus_drv.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (c) 2009, Microsoft Corporation.
   4  *
   5  * Authors:
   6  *   Haiyang Zhang <haiyangz@microsoft.com>
   7  *   Hank Janssen  <hjanssen@microsoft.com>
   8  *   K. Y. Srinivasan <kys@microsoft.com>
   9  */
  10 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  11
  12 #include <linux/init.h>
  13 #include <linux/module.h>
  14 #include <linux/device.h>
  15 #include <linux/interrupt.h>
  16 #include <linux/sysctl.h>
  17 #include <linux/slab.h>
  18 #include <linux/acpi.h>
  19 #include <linux/completion.h>
  20 #include <linux/hyperv.h>
  21 #include <linux/kernel_stat.h>
  22 #include <linux/clockchips.h>
  23 #include <linux/cpu.h>
  24 #include <linux/sched/isolation.h>
  25 #include <linux/sched/task_stack.h>
  26
  27 #include <linux/delay.h>
  28 #include <linux/notifier.h>
  29 #include <linux/panic_notifier.h>
  30 #include <linux/ptrace.h>
  31 #include <linux/screen_info.h>
  32 #include <linux/kdebug.h>
  33 #include <linux/efi.h>
  34 #include <linux/random.h>
  35 #include <linux/kernel.h>
  36 #include <linux/syscore_ops.h>
  37 #include <linux/dma-map-ops.h>
  38 #include <linux/pci.h>
  39 #include <clocksource/hyperv_timer.h>
  40 #include "hyperv_vmbus.h"
  41
  42 struct vmbus_dynid {
  43         struct list_head node;
  44         struct hv_vmbus_device_id id;
  45 };
  46
  47 static struct acpi_device  *hv_acpi_dev;
  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         int ret;
 579
 580         ret = driver_set_override(dev, &hv_dev->driver_override, buf, count);
 581         if (ret)
 582                 return ret;
 583
 584         return count;
 585 }
 586
 587 static ssize_t driver_override_show(struct device *dev,
 588                                     struct device_attribute *attr, char *buf)
 589 {
 590         struct hv_device *hv_dev = device_to_hv_device(dev);
 591         ssize_t len;
 592
 593         device_lock(dev);
 594         len = snprintf(buf, PAGE_SIZE, "%s\n", hv_dev->driver_override);
 595         device_unlock(dev);
 596
 597         return len;
 598 }
 599 static DEVICE_ATTR_RW(driver_override);
 600
 601 /* Set up per device attributes in /sys/bus/vmbus/devices/<bus device> */
 602 static struct attribute *vmbus_dev_attrs[] = {
 603         &dev_attr_id.attr,
 604         &dev_attr_state.attr,
 605         &dev_attr_monitor_id.attr,
 606         &dev_attr_class_id.attr,
 607         &dev_attr_device_id.attr,
 608         &dev_attr_modalias.attr,
 609 #ifdef CONFIG_NUMA
 610         &dev_attr_numa_node.attr,
 611 #endif
 612         &dev_attr_server_monitor_pending.attr,
 613         &dev_attr_client_monitor_pending.attr,
 614         &dev_attr_server_monitor_latency.attr,
 615         &dev_attr_client_monitor_latency.attr,
 616         &dev_attr_server_monitor_conn_id.attr,
 617         &dev_attr_client_monitor_conn_id.attr,
 618         &dev_attr_out_intr_mask.attr,
 619         &dev_attr_out_read_index.attr,
 620         &dev_attr_out_write_index.attr,
 621         &dev_attr_out_read_bytes_avail.attr,
 622         &dev_attr_out_write_bytes_avail.attr,
 623         &dev_attr_in_intr_mask.attr,
 624         &dev_attr_in_read_index.attr,
 625         &dev_attr_in_write_index.attr,
 626         &dev_attr_in_read_bytes_avail.attr,
 627         &dev_attr_in_write_bytes_avail.attr,
 628         &dev_attr_channel_vp_mapping.attr,
 629         &dev_attr_vendor.attr,
 630         &dev_attr_device.attr,
 631         &dev_attr_driver_override.attr,
 632         NULL,
 633 };
 634
 635 /*
 636  * Device-level attribute_group callback function. Returns the permission for
 637  * each attribute, and returns 0 if an attribute is not visible.
 638  */
 639 static umode_t vmbus_dev_attr_is_visible(struct kobject *kobj,
 640                                          struct attribute *attr, int idx)
 641 {
 642         struct device *dev = kobj_to_dev(kobj);
 643         const struct hv_device *hv_dev = device_to_hv_device(dev);
 644
 645         /* Hide the monitor attributes if the monitor mechanism is not used. */
 646         if (!hv_dev->channel->offermsg.monitor_allocated &&
 647             (attr == &dev_attr_monitor_id.attr ||
 648              attr == &dev_attr_server_monitor_pending.attr ||
 649              attr == &dev_attr_client_monitor_pending.attr ||
 650              attr == &dev_attr_server_monitor_latency.attr ||
 651              attr == &dev_attr_client_monitor_latency.attr ||
 652              attr == &dev_attr_server_monitor_conn_id.attr ||
 653              attr == &dev_attr_client_monitor_conn_id.attr))
 654                 return 0;
 655
 656         return attr->mode;
 657 }
 658
 659 static const struct attribute_group vmbus_dev_group = {
 660         .attrs = vmbus_dev_attrs,
 661         .is_visible = vmbus_dev_attr_is_visible
 662 };
 663 __ATTRIBUTE_GROUPS(vmbus_dev);
 664
 665 /* Set up the attribute for /sys/bus/vmbus/hibernation */
 666 static ssize_t hibernation_show(struct bus_type *bus, char *buf)
 667 {
 668         return sprintf(buf, "%d\n", !!hv_is_hibernation_supported());
 669 }
 670
 671 static BUS_ATTR_RO(hibernation);
 672
 673 static struct attribute *vmbus_bus_attrs[] = {
 674         &bus_attr_hibernation.attr,
 675         NULL,
 676 };
 677 static const struct attribute_group vmbus_bus_group = {
 678         .attrs = vmbus_bus_attrs,
 679 };
 680 __ATTRIBUTE_GROUPS(vmbus_bus);
 681
 682 /*
 683  * vmbus_uevent - add uevent for our device
 684  *
 685  * This routine is invoked when a device is added or removed on the vmbus to
 686  * generate a uevent to udev in the userspace. The udev will then look at its
 687  * rule and the uevent generated here to load the appropriate driver
 688  *
 689  * The alias string will be of the form vmbus:guid where guid is the string
 690  * representation of the device guid (each byte of the guid will be
 691  * represented with two hex characters.
 692  */
 693 static int vmbus_uevent(struct device *device, struct kobj_uevent_env *env)
 694 {
 695         struct hv_device *dev = device_to_hv_device(device);
 696         const char *format = "MODALIAS=vmbus:%*phN";
 697
 698         return add_uevent_var(env, format, UUID_SIZE, &dev->dev_type);
 699 }
 700
 701 static const struct hv_vmbus_device_id *
 702 hv_vmbus_dev_match(const struct hv_vmbus_device_id *id, const guid_t *guid)
 703 {
 704         if (id == NULL)
 705                 return NULL; /* empty device table */
 706
 707         for (; !guid_is_null(&id->guid); id++)
 708                 if (guid_equal(&id->guid, guid))
 709                         return id;
 710
 711         return NULL;
 712 }
 713
 714 static const struct hv_vmbus_device_id *
 715 hv_vmbus_dynid_match(struct hv_driver *drv, const guid_t *guid)
 716 {
 717         const struct hv_vmbus_device_id *id = NULL;
 718         struct vmbus_dynid *dynid;
 719
 720         spin_lock(&drv->dynids.lock);
 721         list_for_each_entry(dynid, &drv->dynids.list, node) {
 722                 if (guid_equal(&dynid->id.guid, guid)) {
 723                         id = &dynid->id;
 724                         break;
 725                 }
 726         }
 727         spin_unlock(&drv->dynids.lock);
 728
 729         return id;
 730 }
 731
 732 static const struct hv_vmbus_device_id vmbus_device_null;
 733
 734 /*
 735  * Return a matching hv_vmbus_device_id pointer.
 736  * If there is no match, return NULL.
 737  */
 738 static const struct hv_vmbus_device_id *hv_vmbus_get_id(struct hv_driver *drv,
 739                                                         struct hv_device *dev)
 740 {
 741         const guid_t *guid = &dev->dev_type;
 742         const struct hv_vmbus_device_id *id;
 743
 744         /* When driver_override is set, only bind to the matching driver */
 745         if (dev->driver_override && strcmp(dev->driver_override, drv->name))
 746                 return NULL;
 747
 748         /* Look at the dynamic ids first, before the static ones */
 749         id = hv_vmbus_dynid_match(drv, guid);
 750         if (!id)
 751                 id = hv_vmbus_dev_match(drv->id_table, guid);
 752
 753         /* driver_override will always match, send a dummy id */
 754         if (!id && dev->driver_override)
 755                 id = &vmbus_device_null;
 756
 757         return id;
 758 }
 759
 760 /* vmbus_add_dynid - add a new device ID to this driver and re-probe devices */
 761 static int vmbus_add_dynid(struct hv_driver *drv, guid_t *guid)
 762 {
 763         struct vmbus_dynid *dynid;
 764
 765         dynid = kzalloc(sizeof(*dynid), GFP_KERNEL);
 766         if (!dynid)
 767                 return -ENOMEM;
 768
 769         dynid->id.guid = *guid;
 770
 771         spin_lock(&drv->dynids.lock);
 772         list_add_tail(&dynid->node, &drv->dynids.list);
 773         spin_unlock(&drv->dynids.lock);
 774
 775         return driver_attach(&drv->driver);
 776 }
 777
 778 static void vmbus_free_dynids(struct hv_driver *drv)
 779 {
 780         struct vmbus_dynid *dynid, *n;
 781
 782         spin_lock(&drv->dynids.lock);
 783         list_for_each_entry_safe(dynid, n, &drv->dynids.list, node) {
 784                 list_del(&dynid->node);
 785                 kfree(dynid);
 786         }
 787         spin_unlock(&drv->dynids.lock);
 788 }
 789
 790 /*
 791  * store_new_id - sysfs frontend to vmbus_add_dynid()
 792  *
 793  * Allow GUIDs to be added to an existing driver via sysfs.
 794  */
 795 static ssize_t new_id_store(struct device_driver *driver, const char *buf,
 796                             size_t count)
 797 {
 798         struct hv_driver *drv = drv_to_hv_drv(driver);
 799         guid_t guid;
 800         ssize_t retval;
 801
 802         retval = guid_parse(buf, &guid);
 803         if (retval)
 804                 return retval;
 805
 806         if (hv_vmbus_dynid_match(drv, &guid))
 807                 return -EEXIST;
 808
 809         retval = vmbus_add_dynid(drv, &guid);
 810         if (retval)
 811                 return retval;
 812         return count;
 813 }
 814 static DRIVER_ATTR_WO(new_id);
 815
 816 /*
 817  * store_remove_id - remove a PCI device ID from this driver
 818  *
 819  * Removes a dynamic pci device ID to this driver.
 820  */
 821 static ssize_t remove_id_store(struct device_driver *driver, const char *buf,
 822                                size_t count)
 823 {
 824         struct hv_driver *drv = drv_to_hv_drv(driver);
 825         struct vmbus_dynid *dynid, *n;
 826         guid_t guid;
 827         ssize_t retval;
 828
 829         retval = guid_parse(buf, &guid);
 830         if (retval)
 831                 return retval;
 832
 833         retval = -ENODEV;
 834         spin_lock(&drv->dynids.lock);
 835         list_for_each_entry_safe(dynid, n, &drv->dynids.list, node) {
 836                 struct hv_vmbus_device_id *id = &dynid->id;
 837
 838                 if (guid_equal(&id->guid, &guid)) {
 839                         list_del(&dynid->node);
 840                         kfree(dynid);
 841                         retval = count;
 842                         break;
 843                 }
 844         }
 845         spin_unlock(&drv->dynids.lock);
 846
 847         return retval;
 848 }
 849 static DRIVER_ATTR_WO(remove_id);
 850
 851 static struct attribute *vmbus_drv_attrs[] = {
 852         &driver_attr_new_id.attr,
 853         &driver_attr_remove_id.attr,
 854         NULL,
 855 };
 856 ATTRIBUTE_GROUPS(vmbus_drv);
 857
 858
 859 /*
 860  * vmbus_match - Attempt to match the specified device to the specified driver
 861  */
 862 static int vmbus_match(struct device *device, struct device_driver *driver)
 863 {
 864         struct hv_driver *drv = drv_to_hv_drv(driver);
 865         struct hv_device *hv_dev = device_to_hv_device(device);
 866
 867         /* The hv_sock driver handles all hv_sock offers. */
 868         if (is_hvsock_channel(hv_dev->channel))
 869                 return drv->hvsock;
 870
 871         if (hv_vmbus_get_id(drv, hv_dev))
 872                 return 1;
 873
 874         return 0;
 875 }
 876
 877 /*
 878  * vmbus_probe - Add the new vmbus's child device
 879  */
 880 static int vmbus_probe(struct device *child_device)
 881 {
 882         int ret = 0;
 883         struct hv_driver *drv =
 884                         drv_to_hv_drv(child_device->driver);
 885         struct hv_device *dev = device_to_hv_device(child_device);
 886         const struct hv_vmbus_device_id *dev_id;
 887
 888         dev_id = hv_vmbus_get_id(drv, dev);
 889         if (drv->probe) {
 890                 ret = drv->probe(dev, dev_id);
 891                 if (ret != 0)
 892                         pr_err("probe failed for device %s (%d)\n",
 893                                dev_name(child_device), ret);
 894
 895         } else {
 896                 pr_err("probe not set for driver %s\n",
 897                        dev_name(child_device));
 898                 ret = -ENODEV;
 899         }
 900         return ret;
 901 }
 902
 903 /*
 904  * vmbus_dma_configure -- Configure DMA coherence for VMbus device
 905  */
 906 static int vmbus_dma_configure(struct device *child_device)
 907 {
 908         /*
 909          * On ARM64, propagate the DMA coherence setting from the top level
 910          * VMbus ACPI device to the child VMbus device being added here.
 911          * On x86/x64 coherence is assumed and these calls have no effect.
 912          */
 913         hv_setup_dma_ops(child_device,
 914                 device_get_dma_attr(&hv_acpi_dev->dev) == DEV_DMA_COHERENT);
 915         return 0;
 916 }
 917
 918 /*
 919  * vmbus_remove - Remove a vmbus device
 920  */
 921 static void vmbus_remove(struct device *child_device)
 922 {
 923         struct hv_driver *drv;
 924         struct hv_device *dev = device_to_hv_device(child_device);
 925
 926         if (child_device->driver) {
 927                 drv = drv_to_hv_drv(child_device->driver);
 928                 if (drv->remove)
 929                         drv->remove(dev);
 930         }
 931 }
 932
 933 /*
 934  * vmbus_shutdown - Shutdown a vmbus device
 935  */
 936 static void vmbus_shutdown(struct device *child_device)
 937 {
 938         struct hv_driver *drv;
 939         struct hv_device *dev = device_to_hv_device(child_device);
 940
 941
 942         /* The device may not be attached yet */
 943         if (!child_device->driver)
 944                 return;
 945
 946         drv = drv_to_hv_drv(child_device->driver);
 947
 948         if (drv->shutdown)
 949                 drv->shutdown(dev);
 950 }
 951
 952 #ifdef CONFIG_PM_SLEEP
 953 /*
 954  * vmbus_suspend - Suspend a vmbus device
 955  */
 956 static int vmbus_suspend(struct device *child_device)
 957 {
 958         struct hv_driver *drv;
 959         struct hv_device *dev = device_to_hv_device(child_device);
 960
 961         /* The device may not be attached yet */
 962         if (!child_device->driver)
 963                 return 0;
 964
 965         drv = drv_to_hv_drv(child_device->driver);
 966         if (!drv->suspend)
 967                 return -EOPNOTSUPP;
 968
 969         return drv->suspend(dev);
 970 }
 971
 972 /*
 973  * vmbus_resume - Resume a vmbus device
 974  */
 975 static int vmbus_resume(struct device *child_device)
 976 {
 977         struct hv_driver *drv;
 978         struct hv_device *dev = device_to_hv_device(child_device);
 979
 980         /* The device may not be attached yet */
 981         if (!child_device->driver)
 982                 return 0;
 983
 984         drv = drv_to_hv_drv(child_device->driver);
 985         if (!drv->resume)
 986                 return -EOPNOTSUPP;
 987
 988         return drv->resume(dev);
 989 }
 990 #else
 991 #define vmbus_suspend NULL
 992 #define vmbus_resume NULL
 993 #endif /* CONFIG_PM_SLEEP */
 994
 995 /*
 996  * vmbus_device_release - Final callback release of the vmbus child device
 997  */
 998 static void vmbus_device_release(struct device *device)
 999 {
1000         struct hv_device *hv_dev = device_to_hv_device(device);
1001         struct vmbus_channel *channel = hv_dev->channel;
1002
1003         hv_debug_rm_dev_dir(hv_dev);
1004
1005         mutex_lock(&vmbus_connection.channel_mutex);
1006         hv_process_channel_removal(channel);
1007         mutex_unlock(&vmbus_connection.channel_mutex);
1008         kfree(hv_dev);
1009 }
1010
1011 /*
1012  * Note: we must use the "noirq" ops: see the comment before vmbus_bus_pm.
1013  *
1014  * suspend_noirq/resume_noirq are set to NULL to support Suspend-to-Idle: we
1015  * shouldn't suspend the vmbus devices upon Suspend-to-Idle, otherwise there
1016  * is no way to wake up a Generation-2 VM.
1017  *
1018  * The other 4 ops are for hibernation.
1019  */
1020
1021 static const struct dev_pm_ops vmbus_pm = {
1022         .suspend_noirq  = NULL,
1023         .resume_noirq   = NULL,
1024         .freeze_noirq   = vmbus_suspend,
1025         .thaw_noirq     = vmbus_resume,
1026         .poweroff_noirq = vmbus_suspend,
1027         .restore_noirq  = vmbus_resume,
1028 };
1029
1030 /* The one and only one */
1031 static struct bus_type  hv_bus = {
1032         .name =         "vmbus",
1033         .match =                vmbus_match,
1034         .shutdown =             vmbus_shutdown,
1035         .remove =               vmbus_remove,
1036         .probe =                vmbus_probe,
1037         .uevent =               vmbus_uevent,
1038         .dma_configure =        vmbus_dma_configure,
1039         .dev_groups =           vmbus_dev_groups,
1040         .drv_groups =           vmbus_drv_groups,
1041         .bus_groups =           vmbus_bus_groups,
1042         .pm =                   &vmbus_pm,
1043 };
1044
1045 struct onmessage_work_context {
1046         struct work_struct work;
1047         struct {
1048                 struct hv_message_header header;
1049                 u8 payload[];
1050         } msg;
1051 };
1052
1053 static void vmbus_onmessage_work(struct work_struct *work)
1054 {
1055         struct onmessage_work_context *ctx;
1056
1057         /* Do not process messages if we're in DISCONNECTED state */
1058         if (vmbus_connection.conn_state == DISCONNECTED)
1059                 return;
1060
1061         ctx = container_of(work, struct onmessage_work_context,
1062                            work);
1063         vmbus_onmessage((struct vmbus_channel_message_header *)
1064                         &ctx->msg.payload);
1065         kfree(ctx);
1066 }
1067
1068 void vmbus_on_msg_dpc(unsigned long data)
1069 {
1070         struct hv_per_cpu_context *hv_cpu = (void *)data;
1071         void *page_addr = hv_cpu->synic_message_page;
1072         struct hv_message msg_copy, *msg = (struct hv_message *)page_addr +
1073                                   VMBUS_MESSAGE_SINT;
1074         struct vmbus_channel_message_header *hdr;
1075         enum vmbus_channel_message_type msgtype;
1076         const struct vmbus_channel_message_table_entry *entry;
1077         struct onmessage_work_context *ctx;
1078         __u8 payload_size;
1079         u32 message_type;
1080
1081         /*
1082          * 'enum vmbus_channel_message_type' is supposed to always be 'u32' as
1083          * it is being used in 'struct vmbus_channel_message_header' definition
1084          * which is supposed to match hypervisor ABI.
1085          */
1086         BUILD_BUG_ON(sizeof(enum vmbus_channel_message_type) != sizeof(u32));
1087
1088         /*
1089          * Since the message is in memory shared with the host, an erroneous or
1090          * malicious Hyper-V could modify the message while vmbus_on_msg_dpc()
1091          * or individual message handlers are executing; to prevent this, copy
1092          * the message into private memory.
1093          */
1094         memcpy(&msg_copy, msg, sizeof(struct hv_message));
1095
1096         message_type = msg_copy.header.message_type;
1097         if (message_type == HVMSG_NONE)
1098                 /* no msg */
1099                 return;
1100
1101         hdr = (struct vmbus_channel_message_header *)msg_copy.u.payload;
1102         msgtype = hdr->msgtype;
1103
1104         trace_vmbus_on_msg_dpc(hdr);
1105
1106         if (msgtype >= CHANNELMSG_COUNT) {
1107                 WARN_ONCE(1, "unknown msgtype=%d\n", msgtype);
1108                 goto msg_handled;
1109         }
1110
1111         payload_size = msg_copy.header.payload_size;
1112         if (payload_size > HV_MESSAGE_PAYLOAD_BYTE_COUNT) {
1113                 WARN_ONCE(1, "payload size is too large (%d)\n", payload_size);
1114                 goto msg_handled;
1115         }
1116
1117         entry = &channel_message_table[msgtype];
1118
1119         if (!entry->message_handler)
1120                 goto msg_handled;
1121
1122         if (payload_size < entry->min_payload_len) {
1123                 WARN_ONCE(1, "message too short: msgtype=%d len=%d\n", msgtype, payload_size);
1124                 goto msg_handled;
1125         }
1126
1127         if (entry->handler_type == VMHT_BLOCKING) {
1128                 ctx = kmalloc(struct_size(ctx, msg.payload, payload_size), GFP_ATOMIC);
1129                 if (ctx == NULL)
1130                         return;
1131
1132                 INIT_WORK(&ctx->work, vmbus_onmessage_work);
1133                 ctx->msg.header = msg_copy.header;
1134                 memcpy(&ctx->msg.payload, msg_copy.u.payload, payload_size);
1135
1136                 /*
1137                  * The host can generate a rescind message while we
1138                  * may still be handling the original offer. We deal with
1139                  * this condition by relying on the synchronization provided
1140                  * by offer_in_progress and by channel_mutex.  See also the
1141                  * inline comments in vmbus_onoffer_rescind().
1142                  */
1143                 switch (msgtype) {
1144                 case CHANNELMSG_RESCIND_CHANNELOFFER:
1145                         /*
1146                          * If we are handling the rescind message;
1147                          * schedule the work on the global work queue.
1148                          *
1149                          * The OFFER message and the RESCIND message should
1150                          * not be handled by the same serialized work queue,
1151                          * because the OFFER handler may call vmbus_open(),
1152                          * which tries to open the channel by sending an
1153                          * OPEN_CHANNEL message to the host and waits for
1154                          * the host's response; however, if the host has
1155                          * rescinded the channel before it receives the
1156                          * OPEN_CHANNEL message, the host just silently
1157                          * ignores the OPEN_CHANNEL message; as a result,
1158                          * the guest's OFFER handler hangs for ever, if we
1159                          * handle the RESCIND message in the same serialized
1160                          * work queue: the RESCIND handler can not start to
1161                          * run before the OFFER handler finishes.
1162                          */
1163                         if (vmbus_connection.ignore_any_offer_msg)
1164                                 break;
1165                         queue_work(vmbus_connection.rescind_work_queue, &ctx->work);
1166                         break;
1167
1168                 case CHANNELMSG_OFFERCHANNEL:
1169                         /*
1170                          * The host sends the offer message of a given channel
1171                          * before sending the rescind message of the same
1172                          * channel.  These messages are sent to the guest's
1173                          * connect CPU; the guest then starts processing them
1174                          * in the tasklet handler on this CPU:
1175                          *
1176                          * VMBUS_CONNECT_CPU
1177                          *
1178                          * [vmbus_on_msg_dpc()]
1179                          * atomic_inc()  // CHANNELMSG_OFFERCHANNEL
1180                          * queue_work()
1181                          * ...
1182                          * [vmbus_on_msg_dpc()]
1183                          * schedule_work()  // CHANNELMSG_RESCIND_CHANNELOFFER
1184                          *
1185                          * We rely on the memory-ordering properties of the
1186                          * queue_work() and schedule_work() primitives, which
1187                          * guarantee that the atomic increment will be visible
1188                          * to the CPUs which will execute the offer & rescind
1189                          * works by the time these works will start execution.
1190                          */
1191                         if (vmbus_connection.ignore_any_offer_msg)
1192                                 break;
1193                         atomic_inc(&vmbus_connection.offer_in_progress);
1194                         fallthrough;
1195
1196                 default:
1197                         queue_work(vmbus_connection.work_queue, &ctx->work);
1198                 }
1199         } else
1200                 entry->message_handler(hdr);
1201
1202 msg_handled:
1203         vmbus_signal_eom(msg, message_type);
1204 }
1205
1206 #ifdef CONFIG_PM_SLEEP
1207 /*
1208  * Fake RESCIND_CHANNEL messages to clean up hv_sock channels by force for
1209  * hibernation, because hv_sock connections can not persist across hibernation.
1210  */
1211 static void vmbus_force_channel_rescinded(struct vmbus_channel *channel)
1212 {
1213         struct onmessage_work_context *ctx;
1214         struct vmbus_channel_rescind_offer *rescind;
1215
1216         WARN_ON(!is_hvsock_channel(channel));
1217
1218         /*
1219          * Allocation size is small and the allocation should really not fail,
1220          * otherwise the state of the hv_sock connections ends up in limbo.
1221          */
1222         ctx = kzalloc(sizeof(*ctx) + sizeof(*rescind),
1223                       GFP_KERNEL | __GFP_NOFAIL);
1224
1225         /*
1226          * So far, these are not really used by Linux. Just set them to the
1227          * reasonable values conforming to the definitions of the fields.
1228          */
1229         ctx->msg.header.message_type = 1;
1230         ctx->msg.header.payload_size = sizeof(*rescind);
1231
1232         /* These values are actually used by Linux. */
1233         rescind = (struct vmbus_channel_rescind_offer *)ctx->msg.payload;
1234         rescind->header.msgtype = CHANNELMSG_RESCIND_CHANNELOFFER;
1235         rescind->child_relid = channel->offermsg.child_relid;
1236
1237         INIT_WORK(&ctx->work, vmbus_onmessage_work);
1238
1239         queue_work(vmbus_connection.work_queue, &ctx->work);
1240 }
1241 #endif /* CONFIG_PM_SLEEP */
1242
1243 /*
1244  * Schedule all channels with events pending
1245  */
1246 static void vmbus_chan_sched(struct hv_per_cpu_context *hv_cpu)
1247 {
1248         unsigned long *recv_int_page;
1249         u32 maxbits, relid;
1250
1251         /*
1252          * The event page can be directly checked to get the id of
1253          * the channel that has the interrupt pending.
1254          */
1255         void *page_addr = hv_cpu->synic_event_page;
1256         union hv_synic_event_flags *event
1257                 = (union hv_synic_event_flags *)page_addr +
1258                                          VMBUS_MESSAGE_SINT;
1259
1260         maxbits = HV_EVENT_FLAGS_COUNT;
1261         recv_int_page = event->flags;
1262
1263         if (unlikely(!recv_int_page))
1264                 return;
1265
1266         for_each_set_bit(relid, recv_int_page, maxbits) {
1267                 void (*callback_fn)(void *context);
1268                 struct vmbus_channel *channel;
1269
1270                 if (!sync_test_and_clear_bit(relid, recv_int_page))
1271                         continue;
1272
1273                 /* Special case - vmbus channel protocol msg */
1274                 if (relid == 0)
1275                         continue;
1276
1277                 /*
1278                  * Pairs with the kfree_rcu() in vmbus_chan_release().
1279                  * Guarantees that the channel data structure doesn't
1280                  * get freed while the channel pointer below is being
1281                  * dereferenced.
1282                  */
1283                 rcu_read_lock();
1284
1285                 /* Find channel based on relid */
1286                 channel = relid2channel(relid);
1287                 if (channel == NULL)
1288                         goto sched_unlock_rcu;
1289
1290                 if (channel->rescind)
1291                         goto sched_unlock_rcu;
1292
1293                 /*
1294                  * Make sure that the ring buffer data structure doesn't get
1295                  * freed while we dereference the ring buffer pointer.  Test
1296                  * for the channel's onchannel_callback being NULL within a
1297                  * sched_lock critical section.  See also the inline comments
1298                  * in vmbus_reset_channel_cb().
1299                  */
1300                 spin_lock(&channel->sched_lock);
1301
1302                 callback_fn = channel->onchannel_callback;
1303                 if (unlikely(callback_fn == NULL))
1304                         goto sched_unlock;
1305
1306                 trace_vmbus_chan_sched(channel);
1307
1308                 ++channel->interrupts;
1309
1310                 switch (channel->callback_mode) {
1311                 case HV_CALL_ISR:
1312                         (*callback_fn)(channel->channel_callback_context);
1313                         break;
1314
1315                 case HV_CALL_BATCHED:
1316                         hv_begin_read(&channel->inbound);
1317                         fallthrough;
1318                 case HV_CALL_DIRECT:
1319                         tasklet_schedule(&channel->callback_event);
1320                 }
1321
1322 sched_unlock:
1323                 spin_unlock(&channel->sched_lock);
1324 sched_unlock_rcu:
1325                 rcu_read_unlock();
1326         }
1327 }
1328
1329 static void vmbus_isr(void)
1330 {
1331         struct hv_per_cpu_context *hv_cpu
1332                 = this_cpu_ptr(hv_context.cpu_context);
1333         void *page_addr;
1334         struct hv_message *msg;
1335
1336         vmbus_chan_sched(hv_cpu);
1337
1338         page_addr = hv_cpu->synic_message_page;
1339         msg = (struct hv_message *)page_addr + VMBUS_MESSAGE_SINT;
1340
1341         /* Check if there are actual msgs to be processed */
1342         if (msg->header.message_type != HVMSG_NONE) {
1343                 if (msg->header.message_type == HVMSG_TIMER_EXPIRED) {
1344                         hv_stimer0_isr();
1345                         vmbus_signal_eom(msg, HVMSG_TIMER_EXPIRED);
1346                 } else
1347                         tasklet_schedule(&hv_cpu->msg_dpc);
1348         }
1349
1350         add_interrupt_randomness(vmbus_interrupt);
1351 }
1352
1353 static irqreturn_t vmbus_percpu_isr(int irq, void *dev_id)
1354 {
1355         vmbus_isr();
1356         return IRQ_HANDLED;
1357 }
1358
1359 /*
1360  * Callback from kmsg_dump. Grab as much as possible from the end of the kmsg
1361  * buffer and call into Hyper-V to transfer the data.
1362  */
1363 static void hv_kmsg_dump(struct kmsg_dumper *dumper,
1364                          enum kmsg_dump_reason reason)
1365 {
1366         struct kmsg_dump_iter iter;
1367         size_t bytes_written;
1368
1369         /* We are only interested in panics. */
1370         if ((reason != KMSG_DUMP_PANIC) || (!sysctl_record_panic_msg))
1371                 return;
1372
1373         /*
1374          * Write dump contents to the page. No need to synchronize; panic should
1375          * be single-threaded.
1376          */
1377         kmsg_dump_rewind(&iter);
1378         kmsg_dump_get_buffer(&iter, false, hv_panic_page, HV_HYP_PAGE_SIZE,
1379                              &bytes_written);
1380         if (!bytes_written)
1381                 return;
1382         /*
1383          * P3 to contain the physical address of the panic page & P4 to
1384          * contain the size of the panic data in that page. Rest of the
1385          * registers are no-op when the NOTIFY_MSG flag is set.
1386          */
1387         hv_set_register(HV_REGISTER_CRASH_P0, 0);
1388         hv_set_register(HV_REGISTER_CRASH_P1, 0);
1389         hv_set_register(HV_REGISTER_CRASH_P2, 0);
1390         hv_set_register(HV_REGISTER_CRASH_P3, virt_to_phys(hv_panic_page));
1391         hv_set_register(HV_REGISTER_CRASH_P4, bytes_written);
1392
1393         /*
1394          * Let Hyper-V know there is crash data available along with
1395          * the panic message.
1396          */
1397         hv_set_register(HV_REGISTER_CRASH_CTL,
1398                (HV_CRASH_CTL_CRASH_NOTIFY | HV_CRASH_CTL_CRASH_NOTIFY_MSG));
1399 }
1400
1401 static struct kmsg_dumper hv_kmsg_dumper = {
1402         .dump = hv_kmsg_dump,
1403 };
1404
1405 static void hv_kmsg_dump_register(void)
1406 {
1407         int ret;
1408
1409         hv_panic_page = hv_alloc_hyperv_zeroed_page();
1410         if (!hv_panic_page) {
1411                 pr_err("Hyper-V: panic message page memory allocation failed\n");
1412                 return;
1413         }
1414
1415         ret = kmsg_dump_register(&hv_kmsg_dumper);
1416         if (ret) {
1417                 pr_err("Hyper-V: kmsg dump register error 0x%x\n", ret);
1418                 hv_free_hyperv_page((unsigned long)hv_panic_page);
1419                 hv_panic_page = NULL;
1420         }
1421 }
1422
1423 static struct ctl_table_header *hv_ctl_table_hdr;
1424
1425 /*
1426  * sysctl option to allow the user to control whether kmsg data should be
1427  * reported to Hyper-V on panic.
1428  */
1429 static struct ctl_table hv_ctl_table[] = {
1430         {
1431                 .procname       = "hyperv_record_panic_msg",
1432                 .data           = &sysctl_record_panic_msg,
1433                 .maxlen         = sizeof(int),
1434                 .mode           = 0644,
1435                 .proc_handler   = proc_dointvec_minmax,
1436                 .extra1         = SYSCTL_ZERO,
1437                 .extra2         = SYSCTL_ONE
1438         },
1439         {}
1440 };
1441
1442 static struct ctl_table hv_root_table[] = {
1443         {
1444                 .procname       = "kernel",
1445                 .mode           = 0555,
1446                 .child          = hv_ctl_table
1447         },
1448         {}
1449 };
1450
1451 /*
1452  * vmbus_bus_init -Main vmbus driver initialization routine.
1453  *
1454  * Here, we
1455  *      - initialize the vmbus driver context
1456  *      - invoke the vmbus hv main init routine
1457  *      - retrieve the channel offers
1458  */
1459 static int vmbus_bus_init(void)
1460 {
1461         int ret;
1462
1463         ret = hv_init();
1464         if (ret != 0) {
1465                 pr_err("Unable to initialize the hypervisor - 0x%x\n", ret);
1466                 return ret;
1467         }
1468
1469         ret = bus_register(&hv_bus);
1470         if (ret)
1471                 return ret;
1472
1473         /*
1474          * VMbus interrupts are best modeled as per-cpu interrupts. If
1475          * on an architecture with support for per-cpu IRQs (e.g. ARM64),
1476          * allocate a per-cpu IRQ using standard Linux kernel functionality.
1477          * If not on such an architecture (e.g., x86/x64), then rely on
1478          * code in the arch-specific portion of the code tree to connect
1479          * the VMbus interrupt handler.
1480          */
1481
1482         if (vmbus_irq == -1) {
1483                 hv_setup_vmbus_handler(vmbus_isr);
1484         } else {
1485                 vmbus_evt = alloc_percpu(long);
1486                 ret = request_percpu_irq(vmbus_irq, vmbus_percpu_isr,
1487                                 "Hyper-V VMbus", vmbus_evt);
1488                 if (ret) {
1489                         pr_err("Can't request Hyper-V VMbus IRQ %d, Err %d",
1490                                         vmbus_irq, ret);
1491                         free_percpu(vmbus_evt);
1492                         goto err_setup;
1493                 }
1494         }
1495
1496         ret = hv_synic_alloc();
1497         if (ret)
1498                 goto err_alloc;
1499
1500         /*
1501          * Initialize the per-cpu interrupt state and stimer state.
1502          * Then connect to the host.
1503          */
1504         ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "hyperv/vmbus:online",
1505                                 hv_synic_init, hv_synic_cleanup);
1506         if (ret < 0)
1507                 goto err_cpuhp;
1508         hyperv_cpuhp_online = ret;
1509
1510         ret = vmbus_connect();
1511         if (ret)
1512                 goto err_connect;
1513
1514         if (hv_is_isolation_supported())
1515                 sysctl_record_panic_msg = 0;
1516
1517         /*
1518          * Only register if the crash MSRs are available
1519          */
1520         if (ms_hyperv.misc_features & HV_FEATURE_GUEST_CRASH_MSR_AVAILABLE) {
1521                 u64 hyperv_crash_ctl;
1522                 /*
1523                  * Panic message recording (sysctl_record_panic_msg)
1524                  * is enabled by default in non-isolated guests and
1525                  * disabled by default in isolated guests; the panic
1526                  * message recording won't be available in isolated
1527                  * guests should the following registration fail.
1528                  */
1529                 hv_ctl_table_hdr = register_sysctl_table(hv_root_table);
1530                 if (!hv_ctl_table_hdr)
1531                         pr_err("Hyper-V: sysctl table register error");
1532
1533                 /*
1534                  * Register for panic kmsg callback only if the right
1535                  * capability is supported by the hypervisor.
1536                  */
1537                 hyperv_crash_ctl = hv_get_register(HV_REGISTER_CRASH_CTL);
1538                 if (hyperv_crash_ctl & HV_CRASH_CTL_CRASH_NOTIFY_MSG)
1539                         hv_kmsg_dump_register();
1540
1541                 register_die_notifier(&hyperv_die_block);
1542         }
1543
1544         /*
1545          * Always register the panic notifier because we need to unload
1546          * the VMbus channel connection to prevent any VMbus
1547          * activity after the VM panics.
1548          */
1549         atomic_notifier_chain_register(&panic_notifier_list,
1550                                &hyperv_panic_block);
1551
1552         vmbus_request_offers();
1553
1554         return 0;
1555
1556 err_connect:
1557         cpuhp_remove_state(hyperv_cpuhp_online);
1558 err_cpuhp:
1559         hv_synic_free();
1560 err_alloc:
1561         if (vmbus_irq == -1) {
1562                 hv_remove_vmbus_handler();
1563         } else {
1564                 free_percpu_irq(vmbus_irq, vmbus_evt);
1565                 free_percpu(vmbus_evt);
1566         }
1567 err_setup:
1568         bus_unregister(&hv_bus);
1569         unregister_sysctl_table(hv_ctl_table_hdr);
1570         hv_ctl_table_hdr = NULL;
1571         return ret;
1572 }
1573
1574 /**
1575  * __vmbus_driver_register() - Register a vmbus's driver
1576  * @hv_driver: Pointer to driver structure you want to register
1577  * @owner: owner module of the drv
1578  * @mod_name: module name string
1579  *
1580  * Registers the given driver with Linux through the 'driver_register()' call
1581  * and sets up the hyper-v vmbus handling for this driver.
1582  * It will return the state of the 'driver_register()' call.
1583  *
1584  */
1585 int __vmbus_driver_register(struct hv_driver *hv_driver, struct module *owner, const char *mod_name)
1586 {
1587         int ret;
1588
1589         pr_info("registering driver %s\n", hv_driver->name);
1590
1591         ret = vmbus_exists();
1592         if (ret < 0)
1593                 return ret;
1594
1595         hv_driver->driver.name = hv_driver->name;
1596         hv_driver->driver.owner = owner;
1597         hv_driver->driver.mod_name = mod_name;
1598         hv_driver->driver.bus = &hv_bus;
1599
1600         spin_lock_init(&hv_driver->dynids.lock);
1601         INIT_LIST_HEAD(&hv_driver->dynids.list);
1602
1603         ret = driver_register(&hv_driver->driver);
1604
1605         return ret;
1606 }
1607 EXPORT_SYMBOL_GPL(__vmbus_driver_register);
1608
1609 /**
1610  * vmbus_driver_unregister() - Unregister a vmbus's driver
1611  * @hv_driver: Pointer to driver structure you want to
1612  *             un-register
1613  *
1614  * Un-register the given driver that was previous registered with a call to
1615  * vmbus_driver_register()
1616  */
1617 void vmbus_driver_unregister(struct hv_driver *hv_driver)
1618 {
1619         pr_info("unregistering driver %s\n", hv_driver->name);
1620
1621         if (!vmbus_exists()) {
1622                 driver_unregister(&hv_driver->driver);
1623                 vmbus_free_dynids(hv_driver);
1624         }
1625 }
1626 EXPORT_SYMBOL_GPL(vmbus_driver_unregister);
1627
1628
1629 /*
1630  * Called when last reference to channel is gone.
1631  */
1632 static void vmbus_chan_release(struct kobject *kobj)
1633 {
1634         struct vmbus_channel *channel
1635                 = container_of(kobj, struct vmbus_channel, kobj);
1636
1637         kfree_rcu(channel, rcu);
1638 }
1639
1640 struct vmbus_chan_attribute {
1641         struct attribute attr;
1642         ssize_t (*show)(struct vmbus_channel *chan, char *buf);
1643         ssize_t (*store)(struct vmbus_channel *chan,
1644                          const char *buf, size_t count);
1645 };
1646 #define VMBUS_CHAN_ATTR(_name, _mode, _show, _store) \
1647         struct vmbus_chan_attribute chan_attr_##_name \
1648                 = __ATTR(_name, _mode, _show, _store)
1649 #define VMBUS_CHAN_ATTR_RW(_name) \
1650         struct vmbus_chan_attribute chan_attr_##_name = __ATTR_RW(_name)
1651 #define VMBUS_CHAN_ATTR_RO(_name) \
1652         struct vmbus_chan_attribute chan_attr_##_name = __ATTR_RO(_name)
1653 #define VMBUS_CHAN_ATTR_WO(_name) \
1654         struct vmbus_chan_attribute chan_attr_##_name = __ATTR_WO(_name)
1655
1656 static ssize_t vmbus_chan_attr_show(struct kobject *kobj,
1657                                     struct attribute *attr, char *buf)
1658 {
1659         const struct vmbus_chan_attribute *attribute
1660                 = container_of(attr, struct vmbus_chan_attribute, attr);
1661         struct vmbus_channel *chan
1662                 = container_of(kobj, struct vmbus_channel, kobj);
1663
1664         if (!attribute->show)
1665                 return -EIO;
1666
1667         return attribute->show(chan, buf);
1668 }
1669
1670 static ssize_t vmbus_chan_attr_store(struct kobject *kobj,
1671                                      struct attribute *attr, const char *buf,
1672                                      size_t count)
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->store)
1680                 return -EIO;
1681
1682         return attribute->store(chan, buf, count);
1683 }
1684
1685 static const struct sysfs_ops vmbus_chan_sysfs_ops = {
1686         .show = vmbus_chan_attr_show,
1687         .store = vmbus_chan_attr_store,
1688 };
1689
1690 static ssize_t out_mask_show(struct vmbus_channel *channel, char *buf)
1691 {
1692         struct hv_ring_buffer_info *rbi = &channel->outbound;
1693         ssize_t ret;
1694
1695         mutex_lock(&rbi->ring_buffer_mutex);
1696         if (!rbi->ring_buffer) {
1697                 mutex_unlock(&rbi->ring_buffer_mutex);
1698                 return -EINVAL;
1699         }
1700
1701         ret = sprintf(buf, "%u\n", rbi->ring_buffer->interrupt_mask);
1702         mutex_unlock(&rbi->ring_buffer_mutex);
1703         return ret;
1704 }
1705 static VMBUS_CHAN_ATTR_RO(out_mask);
1706
1707 static ssize_t in_mask_show(struct vmbus_channel *channel, char *buf)
1708 {
1709         struct hv_ring_buffer_info *rbi = &channel->inbound;
1710         ssize_t ret;
1711
1712         mutex_lock(&rbi->ring_buffer_mutex);
1713         if (!rbi->ring_buffer) {
1714                 mutex_unlock(&rbi->ring_buffer_mutex);
1715                 return -EINVAL;
1716         }
1717
1718         ret = sprintf(buf, "%u\n", rbi->ring_buffer->interrupt_mask);
1719         mutex_unlock(&rbi->ring_buffer_mutex);
1720         return ret;
1721 }
1722 static VMBUS_CHAN_ATTR_RO(in_mask);
1723
1724 static ssize_t read_avail_show(struct vmbus_channel *channel, char *buf)
1725 {
1726         struct hv_ring_buffer_info *rbi = &channel->inbound;
1727         ssize_t ret;
1728
1729         mutex_lock(&rbi->ring_buffer_mutex);
1730         if (!rbi->ring_buffer) {
1731                 mutex_unlock(&rbi->ring_buffer_mutex);
1732                 return -EINVAL;
1733         }
1734
1735         ret = sprintf(buf, "%u\n", hv_get_bytes_to_read(rbi));
1736         mutex_unlock(&rbi->ring_buffer_mutex);
1737         return ret;
1738 }
1739 static VMBUS_CHAN_ATTR_RO(read_avail);
1740
1741 static ssize_t write_avail_show(struct vmbus_channel *channel, char *buf)
1742 {
1743         struct hv_ring_buffer_info *rbi = &channel->outbound;
1744         ssize_t ret;
1745
1746         mutex_lock(&rbi->ring_buffer_mutex);
1747         if (!rbi->ring_buffer) {
1748                 mutex_unlock(&rbi->ring_buffer_mutex);
1749                 return -EINVAL;
1750         }
1751
1752         ret = sprintf(buf, "%u\n", hv_get_bytes_to_write(rbi));
1753         mutex_unlock(&rbi->ring_buffer_mutex);
1754         return ret;
1755 }
1756 static VMBUS_CHAN_ATTR_RO(write_avail);
1757
1758 static ssize_t target_cpu_show(struct vmbus_channel *channel, char *buf)
1759 {
1760         return sprintf(buf, "%u\n", channel->target_cpu);
1761 }
1762 static ssize_t target_cpu_store(struct vmbus_channel *channel,
1763                                 const char *buf, size_t count)
1764 {
1765         u32 target_cpu, origin_cpu;
1766         ssize_t ret = count;
1767
1768         if (vmbus_proto_version < VERSION_WIN10_V4_1)
1769                 return -EIO;
1770
1771         if (sscanf(buf, "%uu", &target_cpu) != 1)
1772                 return -EIO;
1773
1774         /* Validate target_cpu for the cpumask_test_cpu() operation below. */
1775         if (target_cpu >= nr_cpumask_bits)
1776                 return -EINVAL;
1777
1778         if (!cpumask_test_cpu(target_cpu, housekeeping_cpumask(HK_TYPE_MANAGED_IRQ)))
1779                 return -EINVAL;
1780
1781         /* No CPUs should come up or down during this. */
1782         cpus_read_lock();
1783
1784         if (!cpu_online(target_cpu)) {
1785                 cpus_read_unlock();
1786                 return -EINVAL;
1787         }
1788
1789         /*
1790          * Synchronizes target_cpu_store() and channel closure:
1791          *
1792          * { Initially: state = CHANNEL_OPENED }
1793          *
1794          * CPU1                         CPU2
1795          *
1796          * [target_cpu_store()]         [vmbus_disconnect_ring()]
1797          *
1798          * LOCK channel_mutex           LOCK channel_mutex
1799          * LOAD r1 = state              LOAD r2 = state
1800          * IF (r1 == CHANNEL_OPENED)    IF (r2 == CHANNEL_OPENED)
1801          *   SEND MODIFYCHANNEL           STORE state = CHANNEL_OPEN
1802          *   [...]                        SEND CLOSECHANNEL
1803          * UNLOCK channel_mutex         UNLOCK channel_mutex
1804          *
1805          * Forbids: r1 == r2 == CHANNEL_OPENED (i.e., CPU1's LOCK precedes
1806          *              CPU2's LOCK) && CPU2's SEND precedes CPU1's SEND
1807          *
1808          * Note.  The host processes the channel messages "sequentially", in
1809          * the order in which they are received on a per-partition basis.
1810          */
1811         mutex_lock(&vmbus_connection.channel_mutex);
1812
1813         /*
1814          * Hyper-V will ignore MODIFYCHANNEL messages for "non-open" channels;
1815          * avoid sending the message and fail here for such channels.
1816          */
1817         if (channel->state != CHANNEL_OPENED_STATE) {
1818                 ret = -EIO;
1819                 goto cpu_store_unlock;
1820         }
1821
1822         origin_cpu = channel->target_cpu;
1823         if (target_cpu == origin_cpu)
1824                 goto cpu_store_unlock;
1825
1826         if (vmbus_send_modifychannel(channel,
1827                                      hv_cpu_number_to_vp_number(target_cpu))) {
1828                 ret = -EIO;
1829                 goto cpu_store_unlock;
1830         }
1831
1832         /*
1833          * For version before VERSION_WIN10_V5_3, the following warning holds:
1834          *
1835          * Warning.  At this point, there is *no* guarantee that the host will
1836          * have successfully processed the vmbus_send_modifychannel() request.
1837          * See the header comment of vmbus_send_modifychannel() for more info.
1838          *
1839          * Lags in the processing of the above vmbus_send_modifychannel() can
1840          * result in missed interrupts if the "old" target CPU is taken offline
1841          * before Hyper-V starts sending interrupts to the "new" target CPU.
1842          * But apart from this offlining scenario, the code tolerates such
1843          * lags.  It will function correctly even if a channel interrupt comes
1844          * in on a CPU that is different from the channel target_cpu value.
1845          */
1846
1847         channel->target_cpu = target_cpu;
1848
1849         /* See init_vp_index(). */
1850         if (hv_is_perf_channel(channel))
1851                 hv_update_allocated_cpus(origin_cpu, target_cpu);
1852
1853         /* Currently set only for storvsc channels. */
1854         if (channel->change_target_cpu_callback) {
1855                 (*channel->change_target_cpu_callback)(channel,
1856                                 origin_cpu, target_cpu);
1857         }
1858
1859 cpu_store_unlock:
1860         mutex_unlock(&vmbus_connection.channel_mutex);
1861         cpus_read_unlock();
1862         return ret;
1863 }
1864 static VMBUS_CHAN_ATTR(cpu, 0644, target_cpu_show, target_cpu_store);
1865
1866 static ssize_t channel_pending_show(struct vmbus_channel *channel,
1867                                     char *buf)
1868 {
1869         return sprintf(buf, "%d\n",
1870                        channel_pending(channel,
1871                                        vmbus_connection.monitor_pages[1]));
1872 }
1873 static VMBUS_CHAN_ATTR(pending, 0444, channel_pending_show, NULL);
1874
1875 static ssize_t channel_latency_show(struct vmbus_channel *channel,
1876                                     char *buf)
1877 {
1878         return sprintf(buf, "%d\n",
1879                        channel_latency(channel,
1880                                        vmbus_connection.monitor_pages[1]));
1881 }
1882 static VMBUS_CHAN_ATTR(latency, 0444, channel_latency_show, NULL);
1883
1884 static ssize_t channel_interrupts_show(struct vmbus_channel *channel, char *buf)
1885 {
1886         return sprintf(buf, "%llu\n", channel->interrupts);
1887 }
1888 static VMBUS_CHAN_ATTR(interrupts, 0444, channel_interrupts_show, NULL);
1889
1890 static ssize_t channel_events_show(struct vmbus_channel *channel, char *buf)
1891 {
1892         return sprintf(buf, "%llu\n", channel->sig_events);
1893 }
1894 static VMBUS_CHAN_ATTR(events, 0444, channel_events_show, NULL);
1895
1896 static ssize_t channel_intr_in_full_show(struct vmbus_channel *channel,
1897                                          char *buf)
1898 {
1899         return sprintf(buf, "%llu\n",
1900                        (unsigned long long)channel->intr_in_full);
1901 }
1902 static VMBUS_CHAN_ATTR(intr_in_full, 0444, channel_intr_in_full_show, NULL);
1903
1904 static ssize_t channel_intr_out_empty_show(struct vmbus_channel *channel,
1905                                            char *buf)
1906 {
1907         return sprintf(buf, "%llu\n",
1908                        (unsigned long long)channel->intr_out_empty);
1909 }
1910 static VMBUS_CHAN_ATTR(intr_out_empty, 0444, channel_intr_out_empty_show, NULL);
1911
1912 static ssize_t channel_out_full_first_show(struct vmbus_channel *channel,
1913                                            char *buf)
1914 {
1915         return sprintf(buf, "%llu\n",
1916                        (unsigned long long)channel->out_full_first);
1917 }
1918 static VMBUS_CHAN_ATTR(out_full_first, 0444, channel_out_full_first_show, NULL);
1919
1920 static ssize_t channel_out_full_total_show(struct vmbus_channel *channel,
1921                                            char *buf)
1922 {
1923         return sprintf(buf, "%llu\n",
1924                        (unsigned long long)channel->out_full_total);
1925 }
1926 static VMBUS_CHAN_ATTR(out_full_total, 0444, channel_out_full_total_show, NULL);
1927
1928 static ssize_t subchannel_monitor_id_show(struct vmbus_channel *channel,
1929                                           char *buf)
1930 {
1931         return sprintf(buf, "%u\n", channel->offermsg.monitorid);
1932 }
1933 static VMBUS_CHAN_ATTR(monitor_id, 0444, subchannel_monitor_id_show, NULL);
1934
1935 static ssize_t subchannel_id_show(struct vmbus_channel *channel,
1936                                   char *buf)
1937 {
1938         return sprintf(buf, "%u\n",
1939                        channel->offermsg.offer.sub_channel_index);
1940 }
1941 static VMBUS_CHAN_ATTR_RO(subchannel_id);
1942
1943 static struct attribute *vmbus_chan_attrs[] = {
1944         &chan_attr_out_mask.attr,
1945         &chan_attr_in_mask.attr,
1946         &chan_attr_read_avail.attr,
1947         &chan_attr_write_avail.attr,
1948         &chan_attr_cpu.attr,
1949         &chan_attr_pending.attr,
1950         &chan_attr_latency.attr,
1951         &chan_attr_interrupts.attr,
1952         &chan_attr_events.attr,
1953         &chan_attr_intr_in_full.attr,
1954         &chan_attr_intr_out_empty.attr,
1955         &chan_attr_out_full_first.attr,
1956         &chan_attr_out_full_total.attr,
1957         &chan_attr_monitor_id.attr,
1958         &chan_attr_subchannel_id.attr,
1959         NULL
1960 };
1961
1962 /*
1963  * Channel-level attribute_group callback function. Returns the permission for
1964  * each attribute, and returns 0 if an attribute is not visible.
1965  */
1966 static umode_t vmbus_chan_attr_is_visible(struct kobject *kobj,
1967                                           struct attribute *attr, int idx)
1968 {
1969         const struct vmbus_channel *channel =
1970                 container_of(kobj, struct vmbus_channel, kobj);
1971
1972         /* Hide the monitor attributes if the monitor mechanism is not used. */
1973         if (!channel->offermsg.monitor_allocated &&
1974             (attr == &chan_attr_pending.attr ||
1975              attr == &chan_attr_latency.attr ||
1976              attr == &chan_attr_monitor_id.attr))
1977                 return 0;
1978
1979         return attr->mode;
1980 }
1981
1982 static struct attribute_group vmbus_chan_group = {
1983         .attrs = vmbus_chan_attrs,
1984         .is_visible = vmbus_chan_attr_is_visible
1985 };
1986
1987 static struct kobj_type vmbus_chan_ktype = {
1988         .sysfs_ops = &vmbus_chan_sysfs_ops,
1989         .release = vmbus_chan_release,
1990 };
1991
1992 /*
1993  * vmbus_add_channel_kobj - setup a sub-directory under device/channels
1994  */
1995 int vmbus_add_channel_kobj(struct hv_device *dev, struct vmbus_channel *channel)
1996 {
1997         const struct device *device = &dev->device;
1998         struct kobject *kobj = &channel->kobj;
1999         u32 relid = channel->offermsg.child_relid;
2000         int ret;
2001
2002         kobj->kset = dev->channels_kset;
2003         ret = kobject_init_and_add(kobj, &vmbus_chan_ktype, NULL,
2004                                    "%u", relid);
2005         if (ret) {
2006                 kobject_put(kobj);
2007                 return ret;
2008         }
2009
2010         ret = sysfs_create_group(kobj, &vmbus_chan_group);
2011
2012         if (ret) {
2013                 /*
2014                  * The calling functions' error handling paths will cleanup the
2015                  * empty channel directory.
2016                  */
2017                 kobject_put(kobj);
2018                 dev_err(device, "Unable to set up channel sysfs files\n");
2019                 return ret;
2020         }
2021
2022         kobject_uevent(kobj, KOBJ_ADD);
2023
2024         return 0;
2025 }
2026
2027 /*
2028  * vmbus_remove_channel_attr_group - remove the channel's attribute group
2029  */
2030 void vmbus_remove_channel_attr_group(struct vmbus_channel *channel)
2031 {
2032         sysfs_remove_group(&channel->kobj, &vmbus_chan_group);
2033 }
2034
2035 /*
2036  * vmbus_device_create - Creates and registers a new child device
2037  * on the vmbus.
2038  */
2039 struct hv_device *vmbus_device_create(const guid_t *type,
2040                                       const guid_t *instance,
2041                                       struct vmbus_channel *channel)
2042 {
2043         struct hv_device *child_device_obj;
2044
2045         child_device_obj = kzalloc(sizeof(struct hv_device), GFP_KERNEL);
2046         if (!child_device_obj) {
2047                 pr_err("Unable to allocate device object for child device\n");
2048                 return NULL;
2049         }
2050
2051         child_device_obj->channel = channel;
2052         guid_copy(&child_device_obj->dev_type, type);
2053         guid_copy(&child_device_obj->dev_instance, instance);
2054         child_device_obj->vendor_id = PCI_VENDOR_ID_MICROSOFT;
2055
2056         return child_device_obj;
2057 }
2058
2059 /*
2060  * vmbus_device_register - Register the child device
2061  */
2062 int vmbus_device_register(struct hv_device *child_device_obj)
2063 {
2064         struct kobject *kobj = &child_device_obj->device.kobj;
2065         int ret;
2066
2067         dev_set_name(&child_device_obj->device, "%pUl",
2068                      &child_device_obj->channel->offermsg.offer.if_instance);
2069
2070         child_device_obj->device.bus = &hv_bus;
2071         child_device_obj->device.parent = &hv_acpi_dev->dev;
2072         child_device_obj->device.release = vmbus_device_release;
2073
2074         child_device_obj->device.dma_parms = &child_device_obj->dma_parms;
2075         child_device_obj->device.dma_mask = &child_device_obj->dma_mask;
2076         dma_set_mask(&child_device_obj->device, DMA_BIT_MASK(64));
2077
2078         /*
2079          * Register with the LDM. This will kick off the driver/device
2080          * binding...which will eventually call vmbus_match() and vmbus_probe()
2081          */
2082         ret = device_register(&child_device_obj->device);
2083         if (ret) {
2084                 pr_err("Unable to register child device\n");
2085                 return ret;
2086         }
2087
2088         child_device_obj->channels_kset = kset_create_and_add("channels",
2089                                                               NULL, kobj);
2090         if (!child_device_obj->channels_kset) {
2091                 ret = -ENOMEM;
2092                 goto err_dev_unregister;
2093         }
2094
2095         ret = vmbus_add_channel_kobj(child_device_obj,
2096                                      child_device_obj->channel);
2097         if (ret) {
2098                 pr_err("Unable to register primary channeln");
2099                 goto err_kset_unregister;
2100         }
2101         hv_debug_add_dev_dir(child_device_obj);
2102
2103         return 0;
2104
2105 err_kset_unregister:
2106         kset_unregister(child_device_obj->channels_kset);
2107
2108 err_dev_unregister:
2109         device_unregister(&child_device_obj->device);
2110         return ret;
2111 }
2112
2113 /*
2114  * vmbus_device_unregister - Remove the specified child device
2115  * from the vmbus.
2116  */
2117 void vmbus_device_unregister(struct hv_device *device_obj)
2118 {
2119         pr_debug("child device %s unregistered\n",
2120                 dev_name(&device_obj->device));
2121
2122         kset_unregister(device_obj->channels_kset);
2123
2124         /*
2125          * Kick off the process of unregistering the device.
2126          * This will call vmbus_remove() and eventually vmbus_device_release()
2127          */
2128         device_unregister(&device_obj->device);
2129 }
2130
2131
2132 /*
2133  * VMBUS is an acpi enumerated device. Get the information we
2134  * need from DSDT.
2135  */
2136 #define VTPM_BASE_ADDRESS 0xfed40000
2137 static acpi_status vmbus_walk_resources(struct acpi_resource *res, void *ctx)
2138 {
2139         resource_size_t start = 0;
2140         resource_size_t end = 0;
2141         struct resource *new_res;
2142         struct resource **old_res = &hyperv_mmio;
2143         struct resource **prev_res = NULL;
2144         struct resource r;
2145
2146         switch (res->type) {
2147
2148         /*
2149          * "Address" descriptors are for bus windows. Ignore
2150          * "memory" descriptors, which are for registers on
2151          * devices.
2152          */
2153         case ACPI_RESOURCE_TYPE_ADDRESS32:
2154                 start = res->data.address32.address.minimum;
2155                 end = res->data.address32.address.maximum;
2156                 break;
2157
2158         case ACPI_RESOURCE_TYPE_ADDRESS64:
2159                 start = res->data.address64.address.minimum;
2160                 end = res->data.address64.address.maximum;
2161                 break;
2162
2163         /*
2164          * The IRQ information is needed only on ARM64, which Hyper-V
2165          * sets up in the extended format. IRQ information is present
2166          * on x86/x64 in the non-extended format but it is not used by
2167          * Linux. So don't bother checking for the non-extended format.
2168          */
2169         case ACPI_RESOURCE_TYPE_EXTENDED_IRQ:
2170                 if (!acpi_dev_resource_interrupt(res, 0, &r)) {
2171                         pr_err("Unable to parse Hyper-V ACPI interrupt\n");
2172                         return AE_ERROR;
2173                 }
2174                 /* ARM64 INTID for VMbus */
2175                 vmbus_interrupt = res->data.extended_irq.interrupts[0];
2176                 /* Linux IRQ number */
2177                 vmbus_irq = r.start;
2178                 return AE_OK;
2179
2180         default:
2181                 /* Unused resource type */
2182                 return AE_OK;
2183
2184         }
2185         /*
2186          * Ignore ranges that are below 1MB, as they're not
2187          * necessary or useful here.
2188          */
2189         if (end < 0x100000)
2190                 return AE_OK;
2191
2192         new_res = kzalloc(sizeof(*new_res), GFP_ATOMIC);
2193         if (!new_res)
2194                 return AE_NO_MEMORY;
2195
2196         /* If this range overlaps the virtual TPM, truncate it. */
2197         if (end > VTPM_BASE_ADDRESS && start < VTPM_BASE_ADDRESS)
2198                 end = VTPM_BASE_ADDRESS;
2199
2200         new_res->name = "hyperv mmio";
2201         new_res->flags = IORESOURCE_MEM;
2202         new_res->start = start;
2203         new_res->end = end;
2204
2205         /*
2206          * If two ranges are adjacent, merge them.
2207          */
2208         do {
2209                 if (!*old_res) {
2210                         *old_res = new_res;
2211                         break;
2212                 }
2213
2214                 if (((*old_res)->end + 1) == new_res->start) {
2215                         (*old_res)->end = new_res->end;
2216                         kfree(new_res);
2217                         break;
2218                 }
2219
2220                 if ((*old_res)->start == new_res->end + 1) {
2221                         (*old_res)->start = new_res->start;
2222                         kfree(new_res);
2223                         break;
2224                 }
2225
2226                 if ((*old_res)->start > new_res->end) {
2227                         new_res->sibling = *old_res;
2228                         if (prev_res)
2229                                 (*prev_res)->sibling = new_res;
2230                         *old_res = new_res;
2231                         break;
2232                 }
2233
2234                 prev_res = old_res;
2235                 old_res = &(*old_res)->sibling;
2236
2237         } while (1);
2238
2239         return AE_OK;
2240 }
2241
2242 static int vmbus_acpi_remove(struct acpi_device *device)
2243 {
2244         struct resource *cur_res;
2245         struct resource *next_res;
2246
2247         if (hyperv_mmio) {
2248                 if (fb_mmio) {
2249                         __release_region(hyperv_mmio, fb_mmio->start,
2250                                          resource_size(fb_mmio));
2251                         fb_mmio = NULL;
2252                 }
2253
2254                 for (cur_res = hyperv_mmio; cur_res; cur_res = next_res) {
2255                         next_res = cur_res->sibling;
2256                         kfree(cur_res);
2257                 }
2258         }
2259
2260         return 0;
2261 }
2262
2263 static void vmbus_reserve_fb(void)
2264 {
2265         resource_size_t start = 0, size;
2266         struct pci_dev *pdev;
2267
2268         if (efi_enabled(EFI_BOOT)) {
2269                 /* Gen2 VM: get FB base from EFI framebuffer */
2270                 start = screen_info.lfb_base;
2271                 size = max_t(__u32, screen_info.lfb_size, 0x800000);
2272         } else {
2273                 /* Gen1 VM: get FB base from PCI */
2274                 pdev = pci_get_device(PCI_VENDOR_ID_MICROSOFT,
2275                                       PCI_DEVICE_ID_HYPERV_VIDEO, NULL);
2276                 if (!pdev)
2277                         return;
2278
2279                 if (pdev->resource[0].flags & IORESOURCE_MEM) {
2280                         start = pci_resource_start(pdev, 0);
2281                         size = pci_resource_len(pdev, 0);
2282                 }
2283
2284                 /*
2285                  * Release the PCI device so hyperv_drm or hyperv_fb driver can
2286                  * grab it later.
2287                  */
2288                 pci_dev_put(pdev);
2289         }
2290
2291         if (!start)
2292                 return;
2293
2294         /*
2295          * Make a claim for the frame buffer in the resource tree under the
2296          * first node, which will be the one below 4GB.  The length seems to
2297          * be underreported, particularly in a Generation 1 VM.  So start out
2298          * reserving a larger area and make it smaller until it succeeds.
2299          */
2300         for (; !fb_mmio && (size >= 0x100000); size >>= 1)
2301                 fb_mmio = __request_region(hyperv_mmio, start, size, fb_mmio_name, 0);
2302 }
2303
2304 /**
2305  * vmbus_allocate_mmio() - Pick a memory-mapped I/O range.
2306  * @new:                If successful, supplied a pointer to the
2307  *                      allocated MMIO space.
2308  * @device_obj:         Identifies the caller
2309  * @min:                Minimum guest physical address of the
2310  *                      allocation
2311  * @max:                Maximum guest physical address
2312  * @size:               Size of the range to be allocated
2313  * @align:              Alignment of the range to be allocated
2314  * @fb_overlap_ok:      Whether this allocation can be allowed
2315  *                      to overlap the video frame buffer.
2316  *
2317  * This function walks the resources granted to VMBus by the
2318  * _CRS object in the ACPI namespace underneath the parent
2319  * "bridge" whether that's a root PCI bus in the Generation 1
2320  * case or a Module Device in the Generation 2 case.  It then
2321  * attempts to allocate from the global MMIO pool in a way that
2322  * matches the constraints supplied in these parameters and by
2323  * that _CRS.
2324  *
2325  * Return: 0 on success, -errno on failure
2326  */
2327 int vmbus_allocate_mmio(struct resource **new, struct hv_device *device_obj,
2328                         resource_size_t min, resource_size_t max,
2329                         resource_size_t size, resource_size_t align,
2330                         bool fb_overlap_ok)
2331 {
2332         struct resource *iter, *shadow;
2333         resource_size_t range_min, range_max, start, end;
2334         const char *dev_n = dev_name(&device_obj->device);
2335         int retval;
2336
2337         retval = -ENXIO;
2338         mutex_lock(&hyperv_mmio_lock);
2339
2340         /*
2341          * If overlaps with frame buffers are allowed, then first attempt to
2342          * make the allocation from within the reserved region.  Because it
2343          * is already reserved, no shadow allocation is necessary.
2344          */
2345         if (fb_overlap_ok && fb_mmio && !(min > fb_mmio->end) &&
2346             !(max < fb_mmio->start)) {
2347
2348                 range_min = fb_mmio->start;
2349                 range_max = fb_mmio->end;
2350                 start = (range_min + align - 1) & ~(align - 1);
2351                 for (; start + size - 1 <= range_max; start += align) {
2352                         *new = request_mem_region_exclusive(start, size, dev_n);
2353                         if (*new) {
2354                                 retval = 0;
2355                                 goto exit;
2356                         }
2357                 }
2358         }
2359
2360         for (iter = hyperv_mmio; iter; iter = iter->sibling) {
2361                 if ((iter->start >= max) || (iter->end <= min))
2362                         continue;
2363
2364                 range_min = iter->start;
2365                 range_max = iter->end;
2366                 start = (range_min + align - 1) & ~(align - 1);
2367                 for (; start + size - 1 <= range_max; start += align) {
2368                         end = start + size - 1;
2369
2370                         /* Skip the whole fb_mmio region if not fb_overlap_ok */
2371                         if (!fb_overlap_ok && fb_mmio &&
2372                             (((start >= fb_mmio->start) && (start <= fb_mmio->end)) ||
2373                              ((end >= fb_mmio->start) && (end <= fb_mmio->end))))
2374                                 continue;
2375
2376                         shadow = __request_region(iter, start, size, NULL,
2377                                                   IORESOURCE_BUSY);
2378                         if (!shadow)
2379                                 continue;
2380
2381                         *new = request_mem_region_exclusive(start, size, dev_n);
2382                         if (*new) {
2383                                 shadow->name = (char *)*new;
2384                                 retval = 0;
2385                                 goto exit;
2386                         }
2387
2388                         __release_region(iter, start, size);
2389                 }
2390         }
2391
2392 exit:
2393         mutex_unlock(&hyperv_mmio_lock);
2394         return retval;
2395 }
2396 EXPORT_SYMBOL_GPL(vmbus_allocate_mmio);
2397
2398 /**
2399  * vmbus_free_mmio() - Free a memory-mapped I/O range.
2400  * @start:              Base address of region to release.
2401  * @size:               Size of the range to be allocated
2402  *
2403  * This function releases anything requested by
2404  * vmbus_mmio_allocate().
2405  */
2406 void vmbus_free_mmio(resource_size_t start, resource_size_t size)
2407 {
2408         struct resource *iter;
2409
2410         mutex_lock(&hyperv_mmio_lock);
2411         for (iter = hyperv_mmio; iter; iter = iter->sibling) {
2412                 if ((iter->start >= start + size) || (iter->end <= start))
2413                         continue;
2414
2415                 __release_region(iter, start, size);
2416         }
2417         release_mem_region(start, size);
2418         mutex_unlock(&hyperv_mmio_lock);
2419
2420 }
2421 EXPORT_SYMBOL_GPL(vmbus_free_mmio);
2422
2423 static int vmbus_acpi_add(struct acpi_device *device)
2424 {
2425         acpi_status result;
2426         int ret_val = -ENODEV;
2427         struct acpi_device *ancestor;
2428
2429         hv_acpi_dev = device;
2430
2431         /*
2432          * Older versions of Hyper-V for ARM64 fail to include the _CCA
2433          * method on the top level VMbus device in the DSDT. But devices
2434          * are hardware coherent in all current Hyper-V use cases, so fix
2435          * up the ACPI device to behave as if _CCA is present and indicates
2436          * hardware coherence.
2437          */
2438         ACPI_COMPANION_SET(&device->dev, device);
2439         if (IS_ENABLED(CONFIG_ACPI_CCA_REQUIRED) &&
2440             device_get_dma_attr(&device->dev) == DEV_DMA_NOT_SUPPORTED) {
2441                 pr_info("No ACPI _CCA found; assuming coherent device I/O\n");
2442                 device->flags.cca_seen = true;
2443                 device->flags.coherent_dma = true;
2444         }
2445
2446         result = acpi_walk_resources(device->handle, METHOD_NAME__CRS,
2447                                         vmbus_walk_resources, NULL);
2448
2449         if (ACPI_FAILURE(result))
2450                 goto acpi_walk_err;
2451         /*
2452          * Some ancestor of the vmbus acpi device (Gen1 or Gen2
2453          * firmware) is the VMOD that has the mmio ranges. Get that.
2454          */
2455         for (ancestor = acpi_dev_parent(device); ancestor;
2456              ancestor = acpi_dev_parent(ancestor)) {
2457                 result = acpi_walk_resources(ancestor->handle, METHOD_NAME__CRS,
2458                                              vmbus_walk_resources, NULL);
2459
2460                 if (ACPI_FAILURE(result))
2461                         continue;
2462                 if (hyperv_mmio) {
2463                         vmbus_reserve_fb();
2464                         break;
2465                 }
2466         }
2467         ret_val = 0;
2468
2469 acpi_walk_err:
2470         if (ret_val)
2471                 vmbus_acpi_remove(device);
2472         return ret_val;
2473 }
2474
2475 #ifdef CONFIG_PM_SLEEP
2476 static int vmbus_bus_suspend(struct device *dev)
2477 {
2478         struct hv_per_cpu_context *hv_cpu = per_cpu_ptr(
2479                         hv_context.cpu_context, VMBUS_CONNECT_CPU);
2480         struct vmbus_channel *channel, *sc;
2481
2482         tasklet_disable(&hv_cpu->msg_dpc);
2483         vmbus_connection.ignore_any_offer_msg = true;
2484         /* The tasklet_enable() takes care of providing a memory barrier */
2485         tasklet_enable(&hv_cpu->msg_dpc);
2486
2487         /* Drain all the workqueues as we are in suspend */
2488         drain_workqueue(vmbus_connection.rescind_work_queue);
2489         drain_workqueue(vmbus_connection.work_queue);
2490         drain_workqueue(vmbus_connection.handle_primary_chan_wq);
2491         drain_workqueue(vmbus_connection.handle_sub_chan_wq);
2492
2493         mutex_lock(&vmbus_connection.channel_mutex);
2494         list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
2495                 if (!is_hvsock_channel(channel))
2496                         continue;
2497
2498                 vmbus_force_channel_rescinded(channel);
2499         }
2500         mutex_unlock(&vmbus_connection.channel_mutex);
2501
2502         /*
2503          * Wait until all the sub-channels and hv_sock channels have been
2504          * cleaned up. Sub-channels should be destroyed upon suspend, otherwise
2505          * they would conflict with the new sub-channels that will be created
2506          * in the resume path. hv_sock channels should also be destroyed, but
2507          * a hv_sock channel of an established hv_sock connection can not be
2508          * really destroyed since it may still be referenced by the userspace
2509          * application, so we just force the hv_sock channel to be rescinded
2510          * by vmbus_force_channel_rescinded(), and the userspace application
2511          * will thoroughly destroy the channel after hibernation.
2512          *
2513          * Note: the counter nr_chan_close_on_suspend may never go above 0 if
2514          * the VM has no sub-channel and hv_sock channel, e.g. a 1-vCPU VM.
2515          */
2516         if (atomic_read(&vmbus_connection.nr_chan_close_on_suspend) > 0)
2517                 wait_for_completion(&vmbus_connection.ready_for_suspend_event);
2518
2519         if (atomic_read(&vmbus_connection.nr_chan_fixup_on_resume) != 0) {
2520                 pr_err("Can not suspend due to a previous failed resuming\n");
2521                 return -EBUSY;
2522         }
2523
2524         mutex_lock(&vmbus_connection.channel_mutex);
2525
2526         list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
2527                 /*
2528                  * Remove the channel from the array of channels and invalidate
2529                  * the channel's relid.  Upon resume, vmbus_onoffer() will fix
2530                  * up the relid (and other fields, if necessary) and add the
2531                  * channel back to the array.
2532                  */
2533                 vmbus_channel_unmap_relid(channel);
2534                 channel->offermsg.child_relid = INVALID_RELID;
2535
2536                 if (is_hvsock_channel(channel)) {
2537                         if (!channel->rescind) {
2538                                 pr_err("hv_sock channel not rescinded!\n");
2539                                 WARN_ON_ONCE(1);
2540                         }
2541                         continue;
2542                 }
2543
2544                 list_for_each_entry(sc, &channel->sc_list, sc_list) {
2545                         pr_err("Sub-channel not deleted!\n");
2546                         WARN_ON_ONCE(1);
2547                 }
2548
2549                 atomic_inc(&vmbus_connection.nr_chan_fixup_on_resume);
2550         }
2551
2552         mutex_unlock(&vmbus_connection.channel_mutex);
2553
2554         vmbus_initiate_unload(false);
2555
2556         /* Reset the event for the next resume. */
2557         reinit_completion(&vmbus_connection.ready_for_resume_event);
2558
2559         return 0;
2560 }
2561
2562 static int vmbus_bus_resume(struct device *dev)
2563 {
2564         struct vmbus_channel_msginfo *msginfo;
2565         size_t msgsize;
2566         int ret;
2567
2568         vmbus_connection.ignore_any_offer_msg = false;
2569
2570         /*
2571          * We only use the 'vmbus_proto_version', which was in use before
2572          * hibernation, to re-negotiate with the host.
2573          */
2574         if (!vmbus_proto_version) {
2575                 pr_err("Invalid proto version = 0x%x\n", vmbus_proto_version);
2576                 return -EINVAL;
2577         }
2578
2579         msgsize = sizeof(*msginfo) +
2580                   sizeof(struct vmbus_channel_initiate_contact);
2581
2582         msginfo = kzalloc(msgsize, GFP_KERNEL);
2583
2584         if (msginfo == NULL)
2585                 return -ENOMEM;
2586
2587         ret = vmbus_negotiate_version(msginfo, vmbus_proto_version);
2588
2589         kfree(msginfo);
2590
2591         if (ret != 0)
2592                 return ret;
2593
2594         WARN_ON(atomic_read(&vmbus_connection.nr_chan_fixup_on_resume) == 0);
2595
2596         vmbus_request_offers();
2597
2598         if (wait_for_completion_timeout(
2599                 &vmbus_connection.ready_for_resume_event, 10 * HZ) == 0)
2600                 pr_err("Some vmbus device is missing after suspending?\n");
2601
2602         /* Reset the event for the next suspend. */
2603         reinit_completion(&vmbus_connection.ready_for_suspend_event);
2604
2605         return 0;
2606 }
2607 #else
2608 #define vmbus_bus_suspend NULL
2609 #define vmbus_bus_resume NULL
2610 #endif /* CONFIG_PM_SLEEP */
2611
2612 static const struct acpi_device_id vmbus_acpi_device_ids[] = {
2613         {"VMBUS", 0},
2614         {"VMBus", 0},
2615         {"", 0},
2616 };
2617 MODULE_DEVICE_TABLE(acpi, vmbus_acpi_device_ids);
2618
2619 /*
2620  * Note: we must use the "no_irq" ops, otherwise hibernation can not work with
2621  * PCI device assignment, because "pci_dev_pm_ops" uses the "noirq" ops: in
2622  * the resume path, the pci "noirq" restore op runs before "non-noirq" op (see
2623  * resume_target_kernel() -> dpm_resume_start(), and hibernation_restore() ->
2624  * dpm_resume_end()). This means vmbus_bus_resume() and the pci-hyperv's
2625  * resume callback must also run via the "noirq" ops.
2626  *
2627  * Set suspend_noirq/resume_noirq to NULL for Suspend-to-Idle: see the comment
2628  * earlier in this file before vmbus_pm.
2629  */
2630
2631 static const struct dev_pm_ops vmbus_bus_pm = {
2632         .suspend_noirq  = NULL,
2633         .resume_noirq   = NULL,
2634         .freeze_noirq   = vmbus_bus_suspend,
2635         .thaw_noirq     = vmbus_bus_resume,
2636         .poweroff_noirq = vmbus_bus_suspend,
2637         .restore_noirq  = vmbus_bus_resume
2638 };
2639
2640 static struct acpi_driver vmbus_acpi_driver = {
2641         .name = "vmbus",
2642         .ids = vmbus_acpi_device_ids,
2643         .ops = {
2644                 .add = vmbus_acpi_add,
2645                 .remove = vmbus_acpi_remove,
2646         },
2647         .drv.pm = &vmbus_bus_pm,
2648         .drv.probe_type = PROBE_FORCE_SYNCHRONOUS,
2649 };
2650
2651 static void hv_kexec_handler(void)
2652 {
2653         hv_stimer_global_cleanup();
2654         vmbus_initiate_unload(false);
2655         /* Make sure conn_state is set as hv_synic_cleanup checks for it */
2656         mb();
2657         cpuhp_remove_state(hyperv_cpuhp_online);
2658 };
2659
2660 static void hv_crash_handler(struct pt_regs *regs)
2661 {
2662         int cpu;
2663
2664         vmbus_initiate_unload(true);
2665         /*
2666          * In crash handler we can't schedule synic cleanup for all CPUs,
2667          * doing the cleanup for current CPU only. This should be sufficient
2668          * for kdump.
2669          */
2670         cpu = smp_processor_id();
2671         hv_stimer_cleanup(cpu);
2672         hv_synic_disable_regs(cpu);
2673 };
2674
2675 static int hv_synic_suspend(void)
2676 {
2677         /*
2678          * When we reach here, all the non-boot CPUs have been offlined.
2679          * If we're in a legacy configuration where stimer Direct Mode is
2680          * not enabled, the stimers on the non-boot CPUs have been unbound
2681          * in hv_synic_cleanup() -> hv_stimer_legacy_cleanup() ->
2682          * hv_stimer_cleanup() -> clockevents_unbind_device().
2683          *
2684          * hv_synic_suspend() only runs on CPU0 with interrupts disabled.
2685          * Here we do not call hv_stimer_legacy_cleanup() on CPU0 because:
2686          * 1) it's unnecessary as interrupts remain disabled between
2687          * syscore_suspend() and syscore_resume(): see create_image() and
2688          * resume_target_kernel()
2689          * 2) the stimer on CPU0 is automatically disabled later by
2690          * syscore_suspend() -> timekeeping_suspend() -> tick_suspend() -> ...
2691          * -> clockevents_shutdown() -> ... -> hv_ce_shutdown()
2692          * 3) a warning would be triggered if we call
2693          * clockevents_unbind_device(), which may sleep, in an
2694          * interrupts-disabled context.
2695          */
2696
2697         hv_synic_disable_regs(0);
2698
2699         return 0;
2700 }
2701
2702 static void hv_synic_resume(void)
2703 {
2704         hv_synic_enable_regs(0);
2705
2706         /*
2707          * Note: we don't need to call hv_stimer_init(0), because the timer
2708          * on CPU0 is not unbound in hv_synic_suspend(), and the timer is
2709          * automatically re-enabled in timekeeping_resume().
2710          */
2711 }
2712
2713 /* The callbacks run only on CPU0, with irqs_disabled. */
2714 static struct syscore_ops hv_synic_syscore_ops = {
2715         .suspend = hv_synic_suspend,
2716         .resume = hv_synic_resume,
2717 };
2718
2719 static int __init hv_acpi_init(void)
2720 {
2721         int ret;
2722
2723         if (!hv_is_hyperv_initialized())
2724                 return -ENODEV;
2725
2726         if (hv_root_partition)
2727                 return 0;
2728
2729         /*
2730          * Get ACPI resources first.
2731          */
2732         ret = acpi_bus_register_driver(&vmbus_acpi_driver);
2733
2734         if (ret)
2735                 return ret;
2736
2737         if (!hv_acpi_dev) {
2738                 ret = -ENODEV;
2739                 goto cleanup;
2740         }
2741
2742         /*
2743          * If we're on an architecture with a hardcoded hypervisor
2744          * vector (i.e. x86/x64), override the VMbus interrupt found
2745          * in the ACPI tables. Ensure vmbus_irq is not set since the
2746          * normal Linux IRQ mechanism is not used in this case.
2747          */
2748 #ifdef HYPERVISOR_CALLBACK_VECTOR
2749         vmbus_interrupt = HYPERVISOR_CALLBACK_VECTOR;
2750         vmbus_irq = -1;
2751 #endif
2752
2753         hv_debug_init();
2754
2755         ret = vmbus_bus_init();
2756         if (ret)
2757                 goto cleanup;
2758
2759         hv_setup_kexec_handler(hv_kexec_handler);
2760         hv_setup_crash_handler(hv_crash_handler);
2761
2762         register_syscore_ops(&hv_synic_syscore_ops);
2763
2764         return 0;
2765
2766 cleanup:
2767         acpi_bus_unregister_driver(&vmbus_acpi_driver);
2768         hv_acpi_dev = NULL;
2769         return ret;
2770 }
2771
2772 static void __exit vmbus_exit(void)
2773 {
2774         int cpu;
2775
2776         unregister_syscore_ops(&hv_synic_syscore_ops);
2777
2778         hv_remove_kexec_handler();
2779         hv_remove_crash_handler();
2780         vmbus_connection.conn_state = DISCONNECTED;
2781         hv_stimer_global_cleanup();
2782         vmbus_disconnect();
2783         if (vmbus_irq == -1) {
2784                 hv_remove_vmbus_handler();
2785         } else {
2786                 free_percpu_irq(vmbus_irq, vmbus_evt);
2787                 free_percpu(vmbus_evt);
2788         }
2789         for_each_online_cpu(cpu) {
2790                 struct hv_per_cpu_context *hv_cpu
2791                         = per_cpu_ptr(hv_context.cpu_context, cpu);
2792
2793                 tasklet_kill(&hv_cpu->msg_dpc);
2794         }
2795         hv_debug_rm_all_dir();
2796
2797         vmbus_free_channels();
2798         kfree(vmbus_connection.channels);
2799
2800         if (ms_hyperv.misc_features & HV_FEATURE_GUEST_CRASH_MSR_AVAILABLE) {
2801                 kmsg_dump_unregister(&hv_kmsg_dumper);
2802                 unregister_die_notifier(&hyperv_die_block);
2803         }
2804
2805         /*
2806          * The panic notifier is always registered, hence we should
2807          * also unconditionally unregister it here as well.
2808          */
2809         atomic_notifier_chain_unregister(&panic_notifier_list,
2810                                          &hyperv_panic_block);
2811
2812         free_page((unsigned long)hv_panic_page);
2813         unregister_sysctl_table(hv_ctl_table_hdr);
2814         hv_ctl_table_hdr = NULL;
2815         bus_unregister(&hv_bus);
2816
2817         cpuhp_remove_state(hyperv_cpuhp_online);
2818         hv_synic_free();
2819         acpi_bus_unregister_driver(&vmbus_acpi_driver);
2820 }
2821
2822
2823 MODULE_LICENSE("GPL");
2824 MODULE_DESCRIPTION("Microsoft Hyper-V VMBus Driver");
2825
2826 subsys_initcall(hv_acpi_init);
2827 module_exit(vmbus_exit);