drivers/hv/vmbus_drv.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (c) 2009, Microsoft Corporation.
   4  *
   5  * Authors:
   6  *   Haiyang Zhang <haiyangz@microsoft.com>
   7  *   Hank Janssen  <hjanssen@microsoft.com>
   8  *   K. Y. Srinivasan <kys@microsoft.com>
   9  */
  10 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  11
  12 #include <linux/init.h>
  13 #include <linux/module.h>
  14 #include <linux/device.h>
  15 #include <linux/interrupt.h>
  16 #include <linux/sysctl.h>
  17 #include <linux/slab.h>
  18 #include <linux/acpi.h>
  19 #include <linux/completion.h>
  20 #include <linux/hyperv.h>
  21 #include <linux/kernel_stat.h>
  22 #include <linux/clockchips.h>
  23 #include <linux/cpu.h>
  24 #include <linux/sched/task_stack.h>
  25
  26 #include <linux/delay.h>
  27 #include <linux/notifier.h>
  28 #include <linux/panic_notifier.h>
  29 #include <linux/ptrace.h>
  30 #include <linux/screen_info.h>
  31 #include <linux/kdebug.h>
  32 #include <linux/efi.h>
  33 #include <linux/random.h>
  34 #include <linux/kernel.h>
  35 #include <linux/syscore_ops.h>
  36 #include <clocksource/hyperv_timer.h>
  37 #include "hyperv_vmbus.h"
  38
  39 struct vmbus_dynid {
  40         struct list_head node;
  41         struct hv_vmbus_device_id id;
  42 };
  43
  44 static struct acpi_device  *hv_acpi_dev;
  45
  46 static struct completion probe_event;
  47
  48 static int hyperv_cpuhp_online;
  49
  50 static void *hv_panic_page;
  51
  52 static long __percpu *vmbus_evt;
  53
  54 /* Values parsed from ACPI DSDT */
  55 int vmbus_irq;
  56 int vmbus_interrupt;
  57
  58 /*
  59  * Boolean to control whether to report panic messages over Hyper-V.
  60  *
  61  * It can be set via /proc/sys/kernel/hyperv_record_panic_msg
  62  */
  63 static int sysctl_record_panic_msg = 1;
  64
  65 static int hyperv_report_reg(void)
  66 {
  67         return !sysctl_record_panic_msg || !hv_panic_page;
  68 }
  69
  70 static int hyperv_panic_event(struct notifier_block *nb, unsigned long val,
  71                               void *args)
  72 {
  73         struct pt_regs *regs;
  74
  75         vmbus_initiate_unload(true);
  76
  77         /*
  78          * Hyper-V should be notified only once about a panic.  If we will be
  79          * doing hyperv_report_panic_msg() later with kmsg data, don't do
  80          * the notification here.
  81          */
  82         if (ms_hyperv.misc_features & HV_FEATURE_GUEST_CRASH_MSR_AVAILABLE
  83             && hyperv_report_reg()) {
  84                 regs = current_pt_regs();
  85                 hyperv_report_panic(regs, val, false);
  86         }
  87         return NOTIFY_DONE;
  88 }
  89
  90 static int hyperv_die_event(struct notifier_block *nb, unsigned long val,
  91                             void *args)
  92 {
  93         struct die_args *die = args;
  94         struct pt_regs *regs = die->regs;
  95
  96         /* Don't notify Hyper-V if the die event is other than oops */
  97         if (val != DIE_OOPS)
  98                 return NOTIFY_DONE;
  99
 100         /*
 101          * Hyper-V should be notified only once about a panic.  If we will be
 102          * doing hyperv_report_panic_msg() later with kmsg data, don't do
 103          * the notification here.
 104          */
 105         if (hyperv_report_reg())
 106                 hyperv_report_panic(regs, val, true);
 107         return NOTIFY_DONE;
 108 }
 109
 110 static struct notifier_block hyperv_die_block = {
 111         .notifier_call = hyperv_die_event,
 112 };
 113 static struct notifier_block hyperv_panic_block = {
 114         .notifier_call = hyperv_panic_event,
 115 };
 116
 117 static const char *fb_mmio_name = "fb_range";
 118 static struct resource *fb_mmio;
 119 static struct resource *hyperv_mmio;
 120 static DEFINE_MUTEX(hyperv_mmio_lock);
 121
 122 static int vmbus_exists(void)
 123 {
 124         if (hv_acpi_dev == NULL)
 125                 return -ENODEV;
 126
 127         return 0;
 128 }
 129
 130 static u8 channel_monitor_group(const struct vmbus_channel *channel)
 131 {
 132         return (u8)channel->offermsg.monitorid / 32;
 133 }
 134
 135 static u8 channel_monitor_offset(const struct vmbus_channel *channel)
 136 {
 137         return (u8)channel->offermsg.monitorid % 32;
 138 }
 139
 140 static u32 channel_pending(const struct vmbus_channel *channel,
 141                            const struct hv_monitor_page *monitor_page)
 142 {
 143         u8 monitor_group = channel_monitor_group(channel);
 144
 145         return monitor_page->trigger_group[monitor_group].pending;
 146 }
 147
 148 static u32 channel_latency(const struct vmbus_channel *channel,
 149                            const struct hv_monitor_page *monitor_page)
 150 {
 151         u8 monitor_group = channel_monitor_group(channel);
 152         u8 monitor_offset = channel_monitor_offset(channel);
 153
 154         return monitor_page->latency[monitor_group][monitor_offset];
 155 }
 156
 157 static u32 channel_conn_id(struct vmbus_channel *channel,
 158                            struct hv_monitor_page *monitor_page)
 159 {
 160         u8 monitor_group = channel_monitor_group(channel);
 161         u8 monitor_offset = channel_monitor_offset(channel);
 162
 163         return monitor_page->parameter[monitor_group][monitor_offset].connectionid.u.id;
 164 }
 165
 166 static ssize_t id_show(struct device *dev, struct device_attribute *dev_attr,
 167                        char *buf)
 168 {
 169         struct hv_device *hv_dev = device_to_hv_device(dev);
 170
 171         if (!hv_dev->channel)
 172                 return -ENODEV;
 173         return sprintf(buf, "%d\n", hv_dev->channel->offermsg.child_relid);
 174 }
 175 static DEVICE_ATTR_RO(id);
 176
 177 static ssize_t state_show(struct device *dev, struct device_attribute *dev_attr,
 178                           char *buf)
 179 {
 180         struct hv_device *hv_dev = device_to_hv_device(dev);
 181
 182         if (!hv_dev->channel)
 183                 return -ENODEV;
 184         return sprintf(buf, "%d\n", hv_dev->channel->state);
 185 }
 186 static DEVICE_ATTR_RO(state);
 187
 188 static ssize_t monitor_id_show(struct device *dev,
 189                                struct device_attribute *dev_attr, char *buf)
 190 {
 191         struct hv_device *hv_dev = device_to_hv_device(dev);
 192
 193         if (!hv_dev->channel)
 194                 return -ENODEV;
 195         return sprintf(buf, "%d\n", hv_dev->channel->offermsg.monitorid);
 196 }
 197 static DEVICE_ATTR_RO(monitor_id);
 198
 199 static ssize_t class_id_show(struct device *dev,
 200                                struct device_attribute *dev_attr, char *buf)
 201 {
 202         struct hv_device *hv_dev = device_to_hv_device(dev);
 203
 204         if (!hv_dev->channel)
 205                 return -ENODEV;
 206         return sprintf(buf, "{%pUl}\n",
 207                        &hv_dev->channel->offermsg.offer.if_type);
 208 }
 209 static DEVICE_ATTR_RO(class_id);
 210
 211 static ssize_t device_id_show(struct device *dev,
 212                               struct device_attribute *dev_attr, char *buf)
 213 {
 214         struct hv_device *hv_dev = device_to_hv_device(dev);
 215
 216         if (!hv_dev->channel)
 217                 return -ENODEV;
 218         return sprintf(buf, "{%pUl}\n",
 219                        &hv_dev->channel->offermsg.offer.if_instance);
 220 }
 221 static DEVICE_ATTR_RO(device_id);
 222
 223 static ssize_t modalias_show(struct device *dev,
 224                              struct device_attribute *dev_attr, char *buf)
 225 {
 226         struct hv_device *hv_dev = device_to_hv_device(dev);
 227
 228         return sprintf(buf, "vmbus:%*phN\n", UUID_SIZE, &hv_dev->dev_type);
 229 }
 230 static DEVICE_ATTR_RO(modalias);
 231
 232 #ifdef CONFIG_NUMA
 233 static ssize_t numa_node_show(struct device *dev,
 234                               struct device_attribute *attr, char *buf)
 235 {
 236         struct hv_device *hv_dev = device_to_hv_device(dev);
 237
 238         if (!hv_dev->channel)
 239                 return -ENODEV;
 240
 241         return sprintf(buf, "%d\n", cpu_to_node(hv_dev->channel->target_cpu));
 242 }
 243 static DEVICE_ATTR_RO(numa_node);
 244 #endif
 245
 246 static ssize_t server_monitor_pending_show(struct device *dev,
 247                                            struct device_attribute *dev_attr,
 248                                            char *buf)
 249 {
 250         struct hv_device *hv_dev = device_to_hv_device(dev);
 251
 252         if (!hv_dev->channel)
 253                 return -ENODEV;
 254         return sprintf(buf, "%d\n",
 255                        channel_pending(hv_dev->channel,
 256                                        vmbus_connection.monitor_pages[0]));
 257 }
 258 static DEVICE_ATTR_RO(server_monitor_pending);
 259
 260 static ssize_t client_monitor_pending_show(struct device *dev,
 261                                            struct device_attribute *dev_attr,
 262                                            char *buf)
 263 {
 264         struct hv_device *hv_dev = device_to_hv_device(dev);
 265
 266         if (!hv_dev->channel)
 267                 return -ENODEV;
 268         return sprintf(buf, "%d\n",
 269                        channel_pending(hv_dev->channel,
 270                                        vmbus_connection.monitor_pages[1]));
 271 }
 272 static DEVICE_ATTR_RO(client_monitor_pending);
 273
 274 static ssize_t server_monitor_latency_show(struct device *dev,
 275                                            struct device_attribute *dev_attr,
 276                                            char *buf)
 277 {
 278         struct hv_device *hv_dev = device_to_hv_device(dev);
 279
 280         if (!hv_dev->channel)
 281                 return -ENODEV;
 282         return sprintf(buf, "%d\n",
 283                        channel_latency(hv_dev->channel,
 284                                        vmbus_connection.monitor_pages[0]));
 285 }
 286 static DEVICE_ATTR_RO(server_monitor_latency);
 287
 288 static ssize_t client_monitor_latency_show(struct device *dev,
 289                                            struct device_attribute *dev_attr,
 290                                            char *buf)
 291 {
 292         struct hv_device *hv_dev = device_to_hv_device(dev);
 293
 294         if (!hv_dev->channel)
 295                 return -ENODEV;
 296         return sprintf(buf, "%d\n",
 297                        channel_latency(hv_dev->channel,
 298                                        vmbus_connection.monitor_pages[1]));
 299 }
 300 static DEVICE_ATTR_RO(client_monitor_latency);
 301
 302 static ssize_t server_monitor_conn_id_show(struct device *dev,
 303                                            struct device_attribute *dev_attr,
 304                                            char *buf)
 305 {
 306         struct hv_device *hv_dev = device_to_hv_device(dev);
 307
 308         if (!hv_dev->channel)
 309                 return -ENODEV;
 310         return sprintf(buf, "%d\n",
 311                        channel_conn_id(hv_dev->channel,
 312                                        vmbus_connection.monitor_pages[0]));
 313 }
 314 static DEVICE_ATTR_RO(server_monitor_conn_id);
 315
 316 static ssize_t client_monitor_conn_id_show(struct device *dev,
 317                                            struct device_attribute *dev_attr,
 318                                            char *buf)
 319 {
 320         struct hv_device *hv_dev = device_to_hv_device(dev);
 321
 322         if (!hv_dev->channel)
 323                 return -ENODEV;
 324         return sprintf(buf, "%d\n",
 325                        channel_conn_id(hv_dev->channel,
 326                                        vmbus_connection.monitor_pages[1]));
 327 }
 328 static DEVICE_ATTR_RO(client_monitor_conn_id);
 329
 330 static ssize_t out_intr_mask_show(struct device *dev,
 331                                   struct device_attribute *dev_attr, char *buf)
 332 {
 333         struct hv_device *hv_dev = device_to_hv_device(dev);
 334         struct hv_ring_buffer_debug_info outbound;
 335         int ret;
 336
 337         if (!hv_dev->channel)
 338                 return -ENODEV;
 339
 340         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->outbound,
 341                                           &outbound);
 342         if (ret < 0)
 343                 return ret;
 344
 345         return sprintf(buf, "%d\n", outbound.current_interrupt_mask);
 346 }
 347 static DEVICE_ATTR_RO(out_intr_mask);
 348
 349 static ssize_t out_read_index_show(struct device *dev,
 350                                    struct device_attribute *dev_attr, char *buf)
 351 {
 352         struct hv_device *hv_dev = device_to_hv_device(dev);
 353         struct hv_ring_buffer_debug_info outbound;
 354         int ret;
 355
 356         if (!hv_dev->channel)
 357                 return -ENODEV;
 358
 359         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->outbound,
 360                                           &outbound);
 361         if (ret < 0)
 362                 return ret;
 363         return sprintf(buf, "%d\n", outbound.current_read_index);
 364 }
 365 static DEVICE_ATTR_RO(out_read_index);
 366
 367 static ssize_t out_write_index_show(struct device *dev,
 368                                     struct device_attribute *dev_attr,
 369                                     char *buf)
 370 {
 371         struct hv_device *hv_dev = device_to_hv_device(dev);
 372         struct hv_ring_buffer_debug_info outbound;
 373         int ret;
 374
 375         if (!hv_dev->channel)
 376                 return -ENODEV;
 377
 378         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->outbound,
 379                                           &outbound);
 380         if (ret < 0)
 381                 return ret;
 382         return sprintf(buf, "%d\n", outbound.current_write_index);
 383 }
 384 static DEVICE_ATTR_RO(out_write_index);
 385
 386 static ssize_t out_read_bytes_avail_show(struct device *dev,
 387                                          struct device_attribute *dev_attr,
 388                                          char *buf)
 389 {
 390         struct hv_device *hv_dev = device_to_hv_device(dev);
 391         struct hv_ring_buffer_debug_info outbound;
 392         int ret;
 393
 394         if (!hv_dev->channel)
 395                 return -ENODEV;
 396
 397         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->outbound,
 398                                           &outbound);
 399         if (ret < 0)
 400                 return ret;
 401         return sprintf(buf, "%d\n", outbound.bytes_avail_toread);
 402 }
 403 static DEVICE_ATTR_RO(out_read_bytes_avail);
 404
 405 static ssize_t out_write_bytes_avail_show(struct device *dev,
 406                                           struct device_attribute *dev_attr,
 407                                           char *buf)
 408 {
 409         struct hv_device *hv_dev = device_to_hv_device(dev);
 410         struct hv_ring_buffer_debug_info outbound;
 411         int ret;
 412
 413         if (!hv_dev->channel)
 414                 return -ENODEV;
 415
 416         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->outbound,
 417                                           &outbound);
 418         if (ret < 0)
 419                 return ret;
 420         return sprintf(buf, "%d\n", outbound.bytes_avail_towrite);
 421 }
 422 static DEVICE_ATTR_RO(out_write_bytes_avail);
 423
 424 static ssize_t in_intr_mask_show(struct device *dev,
 425                                  struct device_attribute *dev_attr, char *buf)
 426 {
 427         struct hv_device *hv_dev = device_to_hv_device(dev);
 428         struct hv_ring_buffer_debug_info inbound;
 429         int ret;
 430
 431         if (!hv_dev->channel)
 432                 return -ENODEV;
 433
 434         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->inbound, &inbound);
 435         if (ret < 0)
 436                 return ret;
 437
 438         return sprintf(buf, "%d\n", inbound.current_interrupt_mask);
 439 }
 440 static DEVICE_ATTR_RO(in_intr_mask);
 441
 442 static ssize_t in_read_index_show(struct device *dev,
 443                                   struct device_attribute *dev_attr, char *buf)
 444 {
 445         struct hv_device *hv_dev = device_to_hv_device(dev);
 446         struct hv_ring_buffer_debug_info inbound;
 447         int ret;
 448
 449         if (!hv_dev->channel)
 450                 return -ENODEV;
 451
 452         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->inbound, &inbound);
 453         if (ret < 0)
 454                 return ret;
 455
 456         return sprintf(buf, "%d\n", inbound.current_read_index);
 457 }
 458 static DEVICE_ATTR_RO(in_read_index);
 459
 460 static ssize_t in_write_index_show(struct device *dev,
 461                                    struct device_attribute *dev_attr, char *buf)
 462 {
 463         struct hv_device *hv_dev = device_to_hv_device(dev);
 464         struct hv_ring_buffer_debug_info inbound;
 465         int ret;
 466
 467         if (!hv_dev->channel)
 468                 return -ENODEV;
 469
 470         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->inbound, &inbound);
 471         if (ret < 0)
 472                 return ret;
 473
 474         return sprintf(buf, "%d\n", inbound.current_write_index);
 475 }
 476 static DEVICE_ATTR_RO(in_write_index);
 477
 478 static ssize_t in_read_bytes_avail_show(struct device *dev,
 479                                         struct device_attribute *dev_attr,
 480                                         char *buf)
 481 {
 482         struct hv_device *hv_dev = device_to_hv_device(dev);
 483         struct hv_ring_buffer_debug_info inbound;
 484         int ret;
 485
 486         if (!hv_dev->channel)
 487                 return -ENODEV;
 488
 489         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->inbound, &inbound);
 490         if (ret < 0)
 491                 return ret;
 492
 493         return sprintf(buf, "%d\n", inbound.bytes_avail_toread);
 494 }
 495 static DEVICE_ATTR_RO(in_read_bytes_avail);
 496
 497 static ssize_t in_write_bytes_avail_show(struct device *dev,
 498                                          struct device_attribute *dev_attr,
 499                                          char *buf)
 500 {
 501         struct hv_device *hv_dev = device_to_hv_device(dev);
 502         struct hv_ring_buffer_debug_info inbound;
 503         int ret;
 504
 505         if (!hv_dev->channel)
 506                 return -ENODEV;
 507
 508         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->inbound, &inbound);
 509         if (ret < 0)
 510                 return ret;
 511
 512         return sprintf(buf, "%d\n", inbound.bytes_avail_towrite);
 513 }
 514 static DEVICE_ATTR_RO(in_write_bytes_avail);
 515
 516 static ssize_t channel_vp_mapping_show(struct device *dev,
 517                                        struct device_attribute *dev_attr,
 518                                        char *buf)
 519 {
 520         struct hv_device *hv_dev = device_to_hv_device(dev);
 521         struct vmbus_channel *channel = hv_dev->channel, *cur_sc;
 522         int buf_size = PAGE_SIZE, n_written, tot_written;
 523         struct list_head *cur;
 524
 525         if (!channel)
 526                 return -ENODEV;
 527
 528         mutex_lock(&vmbus_connection.channel_mutex);
 529
 530         tot_written = snprintf(buf, buf_size, "%u:%u\n",
 531                 channel->offermsg.child_relid, channel->target_cpu);
 532
 533         list_for_each(cur, &channel->sc_list) {
 534                 if (tot_written >= buf_size - 1)
 535                         break;
 536
 537                 cur_sc = list_entry(cur, struct vmbus_channel, sc_list);
 538                 n_written = scnprintf(buf + tot_written,
 539                                      buf_size - tot_written,
 540                                      "%u:%u\n",
 541                                      cur_sc->offermsg.child_relid,
 542                                      cur_sc->target_cpu);
 543                 tot_written += n_written;
 544         }
 545
 546         mutex_unlock(&vmbus_connection.channel_mutex);
 547
 548         return tot_written;
 549 }
 550 static DEVICE_ATTR_RO(channel_vp_mapping);
 551
 552 static ssize_t vendor_show(struct device *dev,
 553                            struct device_attribute *dev_attr,
 554                            char *buf)
 555 {
 556         struct hv_device *hv_dev = device_to_hv_device(dev);
 557
 558         return sprintf(buf, "0x%x\n", hv_dev->vendor_id);
 559 }
 560 static DEVICE_ATTR_RO(vendor);
 561
 562 static ssize_t device_show(struct device *dev,
 563                            struct device_attribute *dev_attr,
 564                            char *buf)
 565 {
 566         struct hv_device *hv_dev = device_to_hv_device(dev);
 567
 568         return sprintf(buf, "0x%x\n", hv_dev->device_id);
 569 }
 570 static DEVICE_ATTR_RO(device);
 571
 572 static ssize_t driver_override_store(struct device *dev,
 573                                      struct device_attribute *attr,
 574                                      const char *buf, size_t count)
 575 {
 576         struct hv_device *hv_dev = device_to_hv_device(dev);
 577         char *driver_override, *old, *cp;
 578
 579         /* We need to keep extra room for a newline */
 580         if (count >= (PAGE_SIZE - 1))
 581                 return -EINVAL;
 582
 583         driver_override = kstrndup(buf, count, GFP_KERNEL);
 584         if (!driver_override)
 585                 return -ENOMEM;
 586
 587         cp = strchr(driver_override, '\n');
 588         if (cp)
 589                 *cp = '\0';
 590
 591         device_lock(dev);
 592         old = hv_dev->driver_override;
 593         if (strlen(driver_override)) {
 594                 hv_dev->driver_override = driver_override;
 595         } else {
 596                 kfree(driver_override);
 597                 hv_dev->driver_override = NULL;
 598         }
 599         device_unlock(dev);
 600
 601         kfree(old);
 602
 603         return count;
 604 }
 605
 606 static ssize_t driver_override_show(struct device *dev,
 607                                     struct device_attribute *attr, char *buf)
 608 {
 609         struct hv_device *hv_dev = device_to_hv_device(dev);
 610         ssize_t len;
 611
 612         device_lock(dev);
 613         len = snprintf(buf, PAGE_SIZE, "%s\n", hv_dev->driver_override);
 614         device_unlock(dev);
 615
 616         return len;
 617 }
 618 static DEVICE_ATTR_RW(driver_override);
 619
 620 /* Set up per device attributes in /sys/bus/vmbus/devices/<bus device> */
 621 static struct attribute *vmbus_dev_attrs[] = {
 622         &dev_attr_id.attr,
 623         &dev_attr_state.attr,
 624         &dev_attr_monitor_id.attr,
 625         &dev_attr_class_id.attr,
 626         &dev_attr_device_id.attr,
 627         &dev_attr_modalias.attr,
 628 #ifdef CONFIG_NUMA
 629         &dev_attr_numa_node.attr,
 630 #endif
 631         &dev_attr_server_monitor_pending.attr,
 632         &dev_attr_client_monitor_pending.attr,
 633         &dev_attr_server_monitor_latency.attr,
 634         &dev_attr_client_monitor_latency.attr,
 635         &dev_attr_server_monitor_conn_id.attr,
 636         &dev_attr_client_monitor_conn_id.attr,
 637         &dev_attr_out_intr_mask.attr,
 638         &dev_attr_out_read_index.attr,
 639         &dev_attr_out_write_index.attr,
 640         &dev_attr_out_read_bytes_avail.attr,
 641         &dev_attr_out_write_bytes_avail.attr,
 642         &dev_attr_in_intr_mask.attr,
 643         &dev_attr_in_read_index.attr,
 644         &dev_attr_in_write_index.attr,
 645         &dev_attr_in_read_bytes_avail.attr,
 646         &dev_attr_in_write_bytes_avail.attr,
 647         &dev_attr_channel_vp_mapping.attr,
 648         &dev_attr_vendor.attr,
 649         &dev_attr_device.attr,
 650         &dev_attr_driver_override.attr,
 651         NULL,
 652 };
 653
 654 /*
 655  * Device-level attribute_group callback function. Returns the permission for
 656  * each attribute, and returns 0 if an attribute is not visible.
 657  */
 658 static umode_t vmbus_dev_attr_is_visible(struct kobject *kobj,
 659                                          struct attribute *attr, int idx)
 660 {
 661         struct device *dev = kobj_to_dev(kobj);
 662         const struct hv_device *hv_dev = device_to_hv_device(dev);
 663
 664         /* Hide the monitor attributes if the monitor mechanism is not used. */
 665         if (!hv_dev->channel->offermsg.monitor_allocated &&
 666             (attr == &dev_attr_monitor_id.attr ||
 667              attr == &dev_attr_server_monitor_pending.attr ||
 668              attr == &dev_attr_client_monitor_pending.attr ||
 669              attr == &dev_attr_server_monitor_latency.attr ||
 670              attr == &dev_attr_client_monitor_latency.attr ||
 671              attr == &dev_attr_server_monitor_conn_id.attr ||
 672              attr == &dev_attr_client_monitor_conn_id.attr))
 673                 return 0;
 674
 675         return attr->mode;
 676 }
 677
 678 static const struct attribute_group vmbus_dev_group = {
 679         .attrs = vmbus_dev_attrs,
 680         .is_visible = vmbus_dev_attr_is_visible
 681 };
 682 __ATTRIBUTE_GROUPS(vmbus_dev);
 683
 684 /* Set up the attribute for /sys/bus/vmbus/hibernation */
 685 static ssize_t hibernation_show(struct bus_type *bus, char *buf)
 686 {
 687         return sprintf(buf, "%d\n", !!hv_is_hibernation_supported());
 688 }
 689
 690 static BUS_ATTR_RO(hibernation);
 691
 692 static struct attribute *vmbus_bus_attrs[] = {
 693         &bus_attr_hibernation.attr,
 694         NULL,
 695 };
 696 static const struct attribute_group vmbus_bus_group = {
 697         .attrs = vmbus_bus_attrs,
 698 };
 699 __ATTRIBUTE_GROUPS(vmbus_bus);
 700
 701 /*
 702  * vmbus_uevent - add uevent for our device
 703  *
 704  * This routine is invoked when a device is added or removed on the vmbus to
 705  * generate a uevent to udev in the userspace. The udev will then look at its
 706  * rule and the uevent generated here to load the appropriate driver
 707  *
 708  * The alias string will be of the form vmbus:guid where guid is the string
 709  * representation of the device guid (each byte of the guid will be
 710  * represented with two hex characters.
 711  */
 712 static int vmbus_uevent(struct device *device, struct kobj_uevent_env *env)
 713 {
 714         struct hv_device *dev = device_to_hv_device(device);
 715         const char *format = "MODALIAS=vmbus:%*phN";
 716
 717         return add_uevent_var(env, format, UUID_SIZE, &dev->dev_type);
 718 }
 719
 720 static const struct hv_vmbus_device_id *
 721 hv_vmbus_dev_match(const struct hv_vmbus_device_id *id, const guid_t *guid)
 722 {
 723         if (id == NULL)
 724                 return NULL; /* empty device table */
 725
 726         for (; !guid_is_null(&id->guid); id++)
 727                 if (guid_equal(&id->guid, guid))
 728                         return id;
 729
 730         return NULL;
 731 }
 732
 733 static const struct hv_vmbus_device_id *
 734 hv_vmbus_dynid_match(struct hv_driver *drv, const guid_t *guid)
 735 {
 736         const struct hv_vmbus_device_id *id = NULL;
 737         struct vmbus_dynid *dynid;
 738
 739         spin_lock(&drv->dynids.lock);
 740         list_for_each_entry(dynid, &drv->dynids.list, node) {
 741                 if (guid_equal(&dynid->id.guid, guid)) {
 742                         id = &dynid->id;
 743                         break;
 744                 }
 745         }
 746         spin_unlock(&drv->dynids.lock);
 747
 748         return id;
 749 }
 750
 751 static const struct hv_vmbus_device_id vmbus_device_null;
 752
 753 /*
 754  * Return a matching hv_vmbus_device_id pointer.
 755  * If there is no match, return NULL.
 756  */
 757 static const struct hv_vmbus_device_id *hv_vmbus_get_id(struct hv_driver *drv,
 758                                                         struct hv_device *dev)
 759 {
 760         const guid_t *guid = &dev->dev_type;
 761         const struct hv_vmbus_device_id *id;
 762
 763         /* When driver_override is set, only bind to the matching driver */
 764         if (dev->driver_override && strcmp(dev->driver_override, drv->name))
 765                 return NULL;
 766
 767         /* Look at the dynamic ids first, before the static ones */
 768         id = hv_vmbus_dynid_match(drv, guid);
 769         if (!id)
 770                 id = hv_vmbus_dev_match(drv->id_table, guid);
 771
 772         /* driver_override will always match, send a dummy id */
 773         if (!id && dev->driver_override)
 774                 id = &vmbus_device_null;
 775
 776         return id;
 777 }
 778
 779 /* vmbus_add_dynid - add a new device ID to this driver and re-probe devices */
 780 static int vmbus_add_dynid(struct hv_driver *drv, guid_t *guid)
 781 {
 782         struct vmbus_dynid *dynid;
 783
 784         dynid = kzalloc(sizeof(*dynid), GFP_KERNEL);
 785         if (!dynid)
 786                 return -ENOMEM;
 787
 788         dynid->id.guid = *guid;
 789
 790         spin_lock(&drv->dynids.lock);
 791         list_add_tail(&dynid->node, &drv->dynids.list);
 792         spin_unlock(&drv->dynids.lock);
 793
 794         return driver_attach(&drv->driver);
 795 }
 796
 797 static void vmbus_free_dynids(struct hv_driver *drv)
 798 {
 799         struct vmbus_dynid *dynid, *n;
 800
 801         spin_lock(&drv->dynids.lock);
 802         list_for_each_entry_safe(dynid, n, &drv->dynids.list, node) {
 803                 list_del(&dynid->node);
 804                 kfree(dynid);
 805         }
 806         spin_unlock(&drv->dynids.lock);
 807 }
 808
 809 /*
 810  * store_new_id - sysfs frontend to vmbus_add_dynid()
 811  *
 812  * Allow GUIDs to be added to an existing driver via sysfs.
 813  */
 814 static ssize_t new_id_store(struct device_driver *driver, const char *buf,
 815                             size_t count)
 816 {
 817         struct hv_driver *drv = drv_to_hv_drv(driver);
 818         guid_t guid;
 819         ssize_t retval;
 820
 821         retval = guid_parse(buf, &guid);
 822         if (retval)
 823                 return retval;
 824
 825         if (hv_vmbus_dynid_match(drv, &guid))
 826                 return -EEXIST;
 827
 828         retval = vmbus_add_dynid(drv, &guid);
 829         if (retval)
 830                 return retval;
 831         return count;
 832 }
 833 static DRIVER_ATTR_WO(new_id);
 834
 835 /*
 836  * store_remove_id - remove a PCI device ID from this driver
 837  *
 838  * Removes a dynamic pci device ID to this driver.
 839  */
 840 static ssize_t remove_id_store(struct device_driver *driver, const char *buf,
 841                                size_t count)
 842 {
 843         struct hv_driver *drv = drv_to_hv_drv(driver);
 844         struct vmbus_dynid *dynid, *n;
 845         guid_t guid;
 846         ssize_t retval;
 847
 848         retval = guid_parse(buf, &guid);
 849         if (retval)
 850                 return retval;
 851
 852         retval = -ENODEV;
 853         spin_lock(&drv->dynids.lock);
 854         list_for_each_entry_safe(dynid, n, &drv->dynids.list, node) {
 855                 struct hv_vmbus_device_id *id = &dynid->id;
 856
 857                 if (guid_equal(&id->guid, &guid)) {
 858                         list_del(&dynid->node);
 859                         kfree(dynid);
 860                         retval = count;
 861                         break;
 862                 }
 863         }
 864         spin_unlock(&drv->dynids.lock);
 865
 866         return retval;
 867 }
 868 static DRIVER_ATTR_WO(remove_id);
 869
 870 static struct attribute *vmbus_drv_attrs[] = {
 871         &driver_attr_new_id.attr,
 872         &driver_attr_remove_id.attr,
 873         NULL,
 874 };
 875 ATTRIBUTE_GROUPS(vmbus_drv);
 876
 877
 878 /*
 879  * vmbus_match - Attempt to match the specified device to the specified driver
 880  */
 881 static int vmbus_match(struct device *device, struct device_driver *driver)
 882 {
 883         struct hv_driver *drv = drv_to_hv_drv(driver);
 884         struct hv_device *hv_dev = device_to_hv_device(device);
 885
 886         /* The hv_sock driver handles all hv_sock offers. */
 887         if (is_hvsock_channel(hv_dev->channel))
 888                 return drv->hvsock;
 889
 890         if (hv_vmbus_get_id(drv, hv_dev))
 891                 return 1;
 892
 893         return 0;
 894 }
 895
 896 /*
 897  * vmbus_probe - Add the new vmbus's child device
 898  */
 899 static int vmbus_probe(struct device *child_device)
 900 {
 901         int ret = 0;
 902         struct hv_driver *drv =
 903                         drv_to_hv_drv(child_device->driver);
 904         struct hv_device *dev = device_to_hv_device(child_device);
 905         const struct hv_vmbus_device_id *dev_id;
 906
 907         dev_id = hv_vmbus_get_id(drv, dev);
 908         if (drv->probe) {
 909                 ret = drv->probe(dev, dev_id);
 910                 if (ret != 0)
 911                         pr_err("probe failed for device %s (%d)\n",
 912                                dev_name(child_device), ret);
 913
 914         } else {
 915                 pr_err("probe not set for driver %s\n",
 916                        dev_name(child_device));
 917                 ret = -ENODEV;
 918         }
 919         return ret;
 920 }
 921
 922 /*
 923  * vmbus_remove - Remove a vmbus device
 924  */
 925 static int vmbus_remove(struct device *child_device)
 926 {
 927         struct hv_driver *drv;
 928         struct hv_device *dev = device_to_hv_device(child_device);
 929
 930         if (child_device->driver) {
 931                 drv = drv_to_hv_drv(child_device->driver);
 932                 if (drv->remove)
 933                         drv->remove(dev);
 934         }
 935
 936         return 0;
 937 }
 938
 939
 940 /*
 941  * vmbus_shutdown - Shutdown a vmbus device
 942  */
 943 static void vmbus_shutdown(struct device *child_device)
 944 {
 945         struct hv_driver *drv;
 946         struct hv_device *dev = device_to_hv_device(child_device);
 947
 948
 949         /* The device may not be attached yet */
 950         if (!child_device->driver)
 951                 return;
 952
 953         drv = drv_to_hv_drv(child_device->driver);
 954
 955         if (drv->shutdown)
 956                 drv->shutdown(dev);
 957 }
 958
 959 #ifdef CONFIG_PM_SLEEP
 960 /*
 961  * vmbus_suspend - Suspend a vmbus device
 962  */
 963 static int vmbus_suspend(struct device *child_device)
 964 {
 965         struct hv_driver *drv;
 966         struct hv_device *dev = device_to_hv_device(child_device);
 967
 968         /* The device may not be attached yet */
 969         if (!child_device->driver)
 970                 return 0;
 971
 972         drv = drv_to_hv_drv(child_device->driver);
 973         if (!drv->suspend)
 974                 return -EOPNOTSUPP;
 975
 976         return drv->suspend(dev);
 977 }
 978
 979 /*
 980  * vmbus_resume - Resume a vmbus device
 981  */
 982 static int vmbus_resume(struct device *child_device)
 983 {
 984         struct hv_driver *drv;
 985         struct hv_device *dev = device_to_hv_device(child_device);
 986
 987         /* The device may not be attached yet */
 988         if (!child_device->driver)
 989                 return 0;
 990
 991         drv = drv_to_hv_drv(child_device->driver);
 992         if (!drv->resume)
 993                 return -EOPNOTSUPP;
 994
 995         return drv->resume(dev);
 996 }
 997 #else
 998 #define vmbus_suspend NULL
 999 #define vmbus_resume NULL
1000 #endif /* CONFIG_PM_SLEEP */
1001
1002 /*
1003  * vmbus_device_release - Final callback release of the vmbus child device
1004  */
1005 static void vmbus_device_release(struct device *device)
1006 {
1007         struct hv_device *hv_dev = device_to_hv_device(device);
1008         struct vmbus_channel *channel = hv_dev->channel;
1009
1010         hv_debug_rm_dev_dir(hv_dev);
1011
1012         mutex_lock(&vmbus_connection.channel_mutex);
1013         hv_process_channel_removal(channel);
1014         mutex_unlock(&vmbus_connection.channel_mutex);
1015         kfree(hv_dev);
1016 }
1017
1018 /*
1019  * Note: we must use the "noirq" ops: see the comment before vmbus_bus_pm.
1020  *
1021  * suspend_noirq/resume_noirq are set to NULL to support Suspend-to-Idle: we
1022  * shouldn't suspend the vmbus devices upon Suspend-to-Idle, otherwise there
1023  * is no way to wake up a Generation-2 VM.
1024  *
1025  * The other 4 ops are for hibernation.
1026  */
1027
1028 static const struct dev_pm_ops vmbus_pm = {
1029         .suspend_noirq  = NULL,
1030         .resume_noirq   = NULL,
1031         .freeze_noirq   = vmbus_suspend,
1032         .thaw_noirq     = vmbus_resume,
1033         .poweroff_noirq = vmbus_suspend,
1034         .restore_noirq  = vmbus_resume,
1035 };
1036
1037 /* The one and only one */
1038 static struct bus_type  hv_bus = {
1039         .name =         "vmbus",
1040         .match =                vmbus_match,
1041         .shutdown =             vmbus_shutdown,
1042         .remove =               vmbus_remove,
1043         .probe =                vmbus_probe,
1044         .uevent =               vmbus_uevent,
1045         .dev_groups =           vmbus_dev_groups,
1046         .drv_groups =           vmbus_drv_groups,
1047         .bus_groups =           vmbus_bus_groups,
1048         .pm =                   &vmbus_pm,
1049 };
1050
1051 struct onmessage_work_context {
1052         struct work_struct work;
1053         struct {
1054                 struct hv_message_header header;
1055                 u8 payload[];
1056         } msg;
1057 };
1058
1059 static void vmbus_onmessage_work(struct work_struct *work)
1060 {
1061         struct onmessage_work_context *ctx;
1062
1063         /* Do not process messages if we're in DISCONNECTED state */
1064         if (vmbus_connection.conn_state == DISCONNECTED)
1065                 return;
1066
1067         ctx = container_of(work, struct onmessage_work_context,
1068                            work);
1069         vmbus_onmessage((struct vmbus_channel_message_header *)
1070                         &ctx->msg.payload);
1071         kfree(ctx);
1072 }
1073
1074 void vmbus_on_msg_dpc(unsigned long data)
1075 {
1076         struct hv_per_cpu_context *hv_cpu = (void *)data;
1077         void *page_addr = hv_cpu->synic_message_page;
1078         struct hv_message msg_copy, *msg = (struct hv_message *)page_addr +
1079                                   VMBUS_MESSAGE_SINT;
1080         struct vmbus_channel_message_header *hdr;
1081         enum vmbus_channel_message_type msgtype;
1082         const struct vmbus_channel_message_table_entry *entry;
1083         struct onmessage_work_context *ctx;
1084         __u8 payload_size;
1085         u32 message_type;
1086
1087         /*
1088          * 'enum vmbus_channel_message_type' is supposed to always be 'u32' as
1089          * it is being used in 'struct vmbus_channel_message_header' definition
1090          * which is supposed to match hypervisor ABI.
1091          */
1092         BUILD_BUG_ON(sizeof(enum vmbus_channel_message_type) != sizeof(u32));
1093
1094         /*
1095          * Since the message is in memory shared with the host, an erroneous or
1096          * malicious Hyper-V could modify the message while vmbus_on_msg_dpc()
1097          * or individual message handlers are executing; to prevent this, copy
1098          * the message into private memory.
1099          */
1100         memcpy(&msg_copy, msg, sizeof(struct hv_message));
1101
1102         message_type = msg_copy.header.message_type;
1103         if (message_type == HVMSG_NONE)
1104                 /* no msg */
1105                 return;
1106
1107         hdr = (struct vmbus_channel_message_header *)msg_copy.u.payload;
1108         msgtype = hdr->msgtype;
1109
1110         trace_vmbus_on_msg_dpc(hdr);
1111
1112         if (msgtype >= CHANNELMSG_COUNT) {
1113                 WARN_ONCE(1, "unknown msgtype=%d\n", msgtype);
1114                 goto msg_handled;
1115         }
1116
1117         payload_size = msg_copy.header.payload_size;
1118         if (payload_size > HV_MESSAGE_PAYLOAD_BYTE_COUNT) {
1119                 WARN_ONCE(1, "payload size is too large (%d)\n", payload_size);
1120                 goto msg_handled;
1121         }
1122
1123         entry = &channel_message_table[msgtype];
1124
1125         if (!entry->message_handler)
1126                 goto msg_handled;
1127
1128         if (payload_size < entry->min_payload_len) {
1129                 WARN_ONCE(1, "message too short: msgtype=%d len=%d\n", msgtype, payload_size);
1130                 goto msg_handled;
1131         }
1132
1133         if (entry->handler_type == VMHT_BLOCKING) {
1134                 ctx = kmalloc(sizeof(*ctx) + payload_size, GFP_ATOMIC);
1135                 if (ctx == NULL)
1136                         return;
1137
1138                 INIT_WORK(&ctx->work, vmbus_onmessage_work);
1139                 memcpy(&ctx->msg, &msg_copy, sizeof(msg->header) + payload_size);
1140
1141                 /*
1142                  * The host can generate a rescind message while we
1143                  * may still be handling the original offer. We deal with
1144                  * this condition by relying on the synchronization provided
1145                  * by offer_in_progress and by channel_mutex.  See also the
1146                  * inline comments in vmbus_onoffer_rescind().
1147                  */
1148                 switch (msgtype) {
1149                 case CHANNELMSG_RESCIND_CHANNELOFFER:
1150                         /*
1151                          * If we are handling the rescind message;
1152                          * schedule the work on the global work queue.
1153                          *
1154                          * The OFFER message and the RESCIND message should
1155                          * not be handled by the same serialized work queue,
1156                          * because the OFFER handler may call vmbus_open(),
1157                          * which tries to open the channel by sending an
1158                          * OPEN_CHANNEL message to the host and waits for
1159                          * the host's response; however, if the host has
1160                          * rescinded the channel before it receives the
1161                          * OPEN_CHANNEL message, the host just silently
1162                          * ignores the OPEN_CHANNEL message; as a result,
1163                          * the guest's OFFER handler hangs for ever, if we
1164                          * handle the RESCIND message in the same serialized
1165                          * work queue: the RESCIND handler can not start to
1166                          * run before the OFFER handler finishes.
1167                          */
1168                         schedule_work(&ctx->work);
1169                         break;
1170
1171                 case CHANNELMSG_OFFERCHANNEL:
1172                         /*
1173                          * The host sends the offer message of a given channel
1174                          * before sending the rescind message of the same
1175                          * channel.  These messages are sent to the guest's
1176                          * connect CPU; the guest then starts processing them
1177                          * in the tasklet handler on this CPU:
1178                          *
1179                          * VMBUS_CONNECT_CPU
1180                          *
1181                          * [vmbus_on_msg_dpc()]
1182                          * atomic_inc()  // CHANNELMSG_OFFERCHANNEL
1183                          * queue_work()
1184                          * ...
1185                          * [vmbus_on_msg_dpc()]
1186                          * schedule_work()  // CHANNELMSG_RESCIND_CHANNELOFFER
1187                          *
1188                          * We rely on the memory-ordering properties of the
1189                          * queue_work() and schedule_work() primitives, which
1190                          * guarantee that the atomic increment will be visible
1191                          * to the CPUs which will execute the offer & rescind
1192                          * works by the time these works will start execution.
1193                          */
1194                         atomic_inc(&vmbus_connection.offer_in_progress);
1195                         fallthrough;
1196
1197                 default:
1198                         queue_work(vmbus_connection.work_queue, &ctx->work);
1199                 }
1200         } else
1201                 entry->message_handler(hdr);
1202
1203 msg_handled:
1204         vmbus_signal_eom(msg, message_type);
1205 }
1206
1207 #ifdef CONFIG_PM_SLEEP
1208 /*
1209  * Fake RESCIND_CHANNEL messages to clean up hv_sock channels by force for
1210  * hibernation, because hv_sock connections can not persist across hibernation.
1211  */
1212 static void vmbus_force_channel_rescinded(struct vmbus_channel *channel)
1213 {
1214         struct onmessage_work_context *ctx;
1215         struct vmbus_channel_rescind_offer *rescind;
1216
1217         WARN_ON(!is_hvsock_channel(channel));
1218
1219         /*
1220          * Allocation size is small and the allocation should really not fail,
1221          * otherwise the state of the hv_sock connections ends up in limbo.
1222          */
1223         ctx = kzalloc(sizeof(*ctx) + sizeof(*rescind),
1224                       GFP_KERNEL | __GFP_NOFAIL);
1225
1226         /*
1227          * So far, these are not really used by Linux. Just set them to the
1228          * reasonable values conforming to the definitions of the fields.
1229          */
1230         ctx->msg.header.message_type = 1;
1231         ctx->msg.header.payload_size = sizeof(*rescind);
1232
1233         /* These values are actually used by Linux. */
1234         rescind = (struct vmbus_channel_rescind_offer *)ctx->msg.payload;
1235         rescind->header.msgtype = CHANNELMSG_RESCIND_CHANNELOFFER;
1236         rescind->child_relid = channel->offermsg.child_relid;
1237
1238         INIT_WORK(&ctx->work, vmbus_onmessage_work);
1239
1240         queue_work(vmbus_connection.work_queue, &ctx->work);
1241 }
1242 #endif /* CONFIG_PM_SLEEP */
1243
1244 /*
1245  * Schedule all channels with events pending
1246  */
1247 static void vmbus_chan_sched(struct hv_per_cpu_context *hv_cpu)
1248 {
1249         unsigned long *recv_int_page;
1250         u32 maxbits, relid;
1251
1252         if (vmbus_proto_version < VERSION_WIN8) {
1253                 maxbits = MAX_NUM_CHANNELS_SUPPORTED;
1254                 recv_int_page = vmbus_connection.recv_int_page;
1255         } else {
1256                 /*
1257                  * When the host is win8 and beyond, the event page
1258                  * can be directly checked to get the id of the channel
1259                  * that has the interrupt pending.
1260                  */
1261                 void *page_addr = hv_cpu->synic_event_page;
1262                 union hv_synic_event_flags *event
1263                         = (union hv_synic_event_flags *)page_addr +
1264                                                  VMBUS_MESSAGE_SINT;
1265
1266                 maxbits = HV_EVENT_FLAGS_COUNT;
1267                 recv_int_page = event->flags;
1268         }
1269
1270         if (unlikely(!recv_int_page))
1271                 return;
1272
1273         for_each_set_bit(relid, recv_int_page, maxbits) {
1274                 void (*callback_fn)(void *context);
1275                 struct vmbus_channel *channel;
1276
1277                 if (!sync_test_and_clear_bit(relid, recv_int_page))
1278                         continue;
1279
1280                 /* Special case - vmbus channel protocol msg */
1281                 if (relid == 0)
1282                         continue;
1283
1284                 /*
1285                  * Pairs with the kfree_rcu() in vmbus_chan_release().
1286                  * Guarantees that the channel data structure doesn't
1287                  * get freed while the channel pointer below is being
1288                  * dereferenced.
1289                  */
1290                 rcu_read_lock();
1291
1292                 /* Find channel based on relid */
1293                 channel = relid2channel(relid);
1294                 if (channel == NULL)
1295                         goto sched_unlock_rcu;
1296
1297                 if (channel->rescind)
1298                         goto sched_unlock_rcu;
1299
1300                 /*
1301                  * Make sure that the ring buffer data structure doesn't get
1302                  * freed while we dereference the ring buffer pointer.  Test
1303                  * for the channel's onchannel_callback being NULL within a
1304                  * sched_lock critical section.  See also the inline comments
1305                  * in vmbus_reset_channel_cb().
1306                  */
1307                 spin_lock(&channel->sched_lock);
1308
1309                 callback_fn = channel->onchannel_callback;
1310                 if (unlikely(callback_fn == NULL))
1311                         goto sched_unlock;
1312
1313                 trace_vmbus_chan_sched(channel);
1314
1315                 ++channel->interrupts;
1316
1317                 switch (channel->callback_mode) {
1318                 case HV_CALL_ISR:
1319                         (*callback_fn)(channel->channel_callback_context);
1320                         break;
1321
1322                 case HV_CALL_BATCHED:
1323                         hv_begin_read(&channel->inbound);
1324                         fallthrough;
1325                 case HV_CALL_DIRECT:
1326                         tasklet_schedule(&channel->callback_event);
1327                 }
1328
1329 sched_unlock:
1330                 spin_unlock(&channel->sched_lock);
1331 sched_unlock_rcu:
1332                 rcu_read_unlock();
1333         }
1334 }
1335
1336 static void vmbus_isr(void)
1337 {
1338         struct hv_per_cpu_context *hv_cpu
1339                 = this_cpu_ptr(hv_context.cpu_context);
1340         void *page_addr = hv_cpu->synic_event_page;
1341         struct hv_message *msg;
1342         union hv_synic_event_flags *event;
1343         bool handled = false;
1344
1345         if (unlikely(page_addr == NULL))
1346                 return;
1347
1348         event = (union hv_synic_event_flags *)page_addr +
1349                                          VMBUS_MESSAGE_SINT;
1350         /*
1351          * Check for events before checking for messages. This is the order
1352          * in which events and messages are checked in Windows guests on
1353          * Hyper-V, and the Windows team suggested we do the same.
1354          */
1355
1356         if ((vmbus_proto_version == VERSION_WS2008) ||
1357                 (vmbus_proto_version == VERSION_WIN7)) {
1358
1359                 /* Since we are a child, we only need to check bit 0 */
1360                 if (sync_test_and_clear_bit(0, event->flags))
1361                         handled = true;
1362         } else {
1363                 /*
1364                  * Our host is win8 or above. The signaling mechanism
1365                  * has changed and we can directly look at the event page.
1366                  * If bit n is set then we have an interrup on the channel
1367                  * whose id is n.
1368                  */
1369                 handled = true;
1370         }
1371
1372         if (handled)
1373                 vmbus_chan_sched(hv_cpu);
1374
1375         page_addr = hv_cpu->synic_message_page;
1376         msg = (struct hv_message *)page_addr + VMBUS_MESSAGE_SINT;
1377
1378         /* Check if there are actual msgs to be processed */
1379         if (msg->header.message_type != HVMSG_NONE) {
1380                 if (msg->header.message_type == HVMSG_TIMER_EXPIRED) {
1381                         hv_stimer0_isr();
1382                         vmbus_signal_eom(msg, HVMSG_TIMER_EXPIRED);
1383                 } else
1384                         tasklet_schedule(&hv_cpu->msg_dpc);
1385         }
1386
1387         add_interrupt_randomness(vmbus_interrupt, 0);
1388 }
1389
1390 static irqreturn_t vmbus_percpu_isr(int irq, void *dev_id)
1391 {
1392         vmbus_isr();
1393         return IRQ_HANDLED;
1394 }
1395
1396 /*
1397  * Callback from kmsg_dump. Grab as much as possible from the end of the kmsg
1398  * buffer and call into Hyper-V to transfer the data.
1399  */
1400 static void hv_kmsg_dump(struct kmsg_dumper *dumper,
1401                          enum kmsg_dump_reason reason)
1402 {
1403         struct kmsg_dump_iter iter;
1404         size_t bytes_written;
1405
1406         /* We are only interested in panics. */
1407         if ((reason != KMSG_DUMP_PANIC) || (!sysctl_record_panic_msg))
1408                 return;
1409
1410         /*
1411          * Write dump contents to the page. No need to synchronize; panic should
1412          * be single-threaded.
1413          */
1414         kmsg_dump_rewind(&iter);
1415         kmsg_dump_get_buffer(&iter, false, hv_panic_page, HV_HYP_PAGE_SIZE,
1416                              &bytes_written);
1417         if (!bytes_written)
1418                 return;
1419         /*
1420          * P3 to contain the physical address of the panic page & P4 to
1421          * contain the size of the panic data in that page. Rest of the
1422          * registers are no-op when the NOTIFY_MSG flag is set.
1423          */
1424         hv_set_register(HV_REGISTER_CRASH_P0, 0);
1425         hv_set_register(HV_REGISTER_CRASH_P1, 0);
1426         hv_set_register(HV_REGISTER_CRASH_P2, 0);
1427         hv_set_register(HV_REGISTER_CRASH_P3, virt_to_phys(hv_panic_page));
1428         hv_set_register(HV_REGISTER_CRASH_P4, bytes_written);
1429
1430         /*
1431          * Let Hyper-V know there is crash data available along with
1432          * the panic message.
1433          */
1434         hv_set_register(HV_REGISTER_CRASH_CTL,
1435                (HV_CRASH_CTL_CRASH_NOTIFY | HV_CRASH_CTL_CRASH_NOTIFY_MSG));
1436 }
1437
1438 static struct kmsg_dumper hv_kmsg_dumper = {
1439         .dump = hv_kmsg_dump,
1440 };
1441
1442 static void hv_kmsg_dump_register(void)
1443 {
1444         int ret;
1445
1446         hv_panic_page = hv_alloc_hyperv_zeroed_page();
1447         if (!hv_panic_page) {
1448                 pr_err("Hyper-V: panic message page memory allocation failed\n");
1449                 return;
1450         }
1451
1452         ret = kmsg_dump_register(&hv_kmsg_dumper);
1453         if (ret) {
1454                 pr_err("Hyper-V: kmsg dump register error 0x%x\n", ret);
1455                 hv_free_hyperv_page((unsigned long)hv_panic_page);
1456                 hv_panic_page = NULL;
1457         }
1458 }
1459
1460 static struct ctl_table_header *hv_ctl_table_hdr;
1461
1462 /*
1463  * sysctl option to allow the user to control whether kmsg data should be
1464  * reported to Hyper-V on panic.
1465  */
1466 static struct ctl_table hv_ctl_table[] = {
1467         {
1468                 .procname       = "hyperv_record_panic_msg",
1469                 .data           = &sysctl_record_panic_msg,
1470                 .maxlen         = sizeof(int),
1471                 .mode           = 0644,
1472                 .proc_handler   = proc_dointvec_minmax,
1473                 .extra1         = SYSCTL_ZERO,
1474                 .extra2         = SYSCTL_ONE
1475         },
1476         {}
1477 };
1478
1479 static struct ctl_table hv_root_table[] = {
1480         {
1481                 .procname       = "kernel",
1482                 .mode           = 0555,
1483                 .child          = hv_ctl_table
1484         },
1485         {}
1486 };
1487
1488 /*
1489  * vmbus_bus_init -Main vmbus driver initialization routine.
1490  *
1491  * Here, we
1492  *      - initialize the vmbus driver context
1493  *      - invoke the vmbus hv main init routine
1494  *      - retrieve the channel offers
1495  */
1496 static int vmbus_bus_init(void)
1497 {
1498         int ret;
1499
1500         ret = hv_init();
1501         if (ret != 0) {
1502                 pr_err("Unable to initialize the hypervisor - 0x%x\n", ret);
1503                 return ret;
1504         }
1505
1506         ret = bus_register(&hv_bus);
1507         if (ret)
1508                 return ret;
1509
1510         /*
1511          * VMbus interrupts are best modeled as per-cpu interrupts. If
1512          * on an architecture with support for per-cpu IRQs (e.g. ARM64),
1513          * allocate a per-cpu IRQ using standard Linux kernel functionality.
1514          * If not on such an architecture (e.g., x86/x64), then rely on
1515          * code in the arch-specific portion of the code tree to connect
1516          * the VMbus interrupt handler.
1517          */
1518
1519         if (vmbus_irq == -1) {
1520                 hv_setup_vmbus_handler(vmbus_isr);
1521         } else {
1522                 vmbus_evt = alloc_percpu(long);
1523                 ret = request_percpu_irq(vmbus_irq, vmbus_percpu_isr,
1524                                 "Hyper-V VMbus", vmbus_evt);
1525                 if (ret) {
1526                         pr_err("Can't request Hyper-V VMbus IRQ %d, Err %d",
1527                                         vmbus_irq, ret);
1528                         free_percpu(vmbus_evt);
1529                         goto err_setup;
1530                 }
1531         }
1532
1533         ret = hv_synic_alloc();
1534         if (ret)
1535                 goto err_alloc;
1536
1537         /*
1538          * Initialize the per-cpu interrupt state and stimer state.
1539          * Then connect to the host.
1540          */
1541         ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "hyperv/vmbus:online",
1542                                 hv_synic_init, hv_synic_cleanup);
1543         if (ret < 0)
1544                 goto err_cpuhp;
1545         hyperv_cpuhp_online = ret;
1546
1547         ret = vmbus_connect();
1548         if (ret)
1549                 goto err_connect;
1550
1551         /*
1552          * Only register if the crash MSRs are available
1553          */
1554         if (ms_hyperv.misc_features & HV_FEATURE_GUEST_CRASH_MSR_AVAILABLE) {
1555                 u64 hyperv_crash_ctl;
1556                 /*
1557                  * Sysctl registration is not fatal, since by default
1558                  * reporting is enabled.
1559                  */
1560                 hv_ctl_table_hdr = register_sysctl_table(hv_root_table);
1561                 if (!hv_ctl_table_hdr)
1562                         pr_err("Hyper-V: sysctl table register error");
1563
1564                 /*
1565                  * Register for panic kmsg callback only if the right
1566                  * capability is supported by the hypervisor.
1567                  */
1568                 hyperv_crash_ctl = hv_get_register(HV_REGISTER_CRASH_CTL);
1569                 if (hyperv_crash_ctl & HV_CRASH_CTL_CRASH_NOTIFY_MSG)
1570                         hv_kmsg_dump_register();
1571
1572                 register_die_notifier(&hyperv_die_block);
1573         }
1574
1575         /*
1576          * Always register the panic notifier because we need to unload
1577          * the VMbus channel connection to prevent any VMbus
1578          * activity after the VM panics.
1579          */
1580         atomic_notifier_chain_register(&panic_notifier_list,
1581                                &hyperv_panic_block);
1582
1583         vmbus_request_offers();
1584
1585         return 0;
1586
1587 err_connect:
1588         cpuhp_remove_state(hyperv_cpuhp_online);
1589 err_cpuhp:
1590         hv_synic_free();
1591 err_alloc:
1592         if (vmbus_irq == -1) {
1593                 hv_remove_vmbus_handler();
1594         } else {
1595                 free_percpu_irq(vmbus_irq, vmbus_evt);
1596                 free_percpu(vmbus_evt);
1597         }
1598 err_setup:
1599         bus_unregister(&hv_bus);
1600         unregister_sysctl_table(hv_ctl_table_hdr);
1601         hv_ctl_table_hdr = NULL;
1602         return ret;
1603 }
1604
1605 /**
1606  * __vmbus_child_driver_register() - Register a vmbus's driver
1607  * @hv_driver: Pointer to driver structure you want to register
1608  * @owner: owner module of the drv
1609  * @mod_name: module name string
1610  *
1611  * Registers the given driver with Linux through the 'driver_register()' call
1612  * and sets up the hyper-v vmbus handling for this driver.
1613  * It will return the state of the 'driver_register()' call.
1614  *
1615  */
1616 int __vmbus_driver_register(struct hv_driver *hv_driver, struct module *owner, const char *mod_name)
1617 {
1618         int ret;
1619
1620         pr_info("registering driver %s\n", hv_driver->name);
1621
1622         ret = vmbus_exists();
1623         if (ret < 0)
1624                 return ret;
1625
1626         hv_driver->driver.name = hv_driver->name;
1627         hv_driver->driver.owner = owner;
1628         hv_driver->driver.mod_name = mod_name;
1629         hv_driver->driver.bus = &hv_bus;
1630
1631         spin_lock_init(&hv_driver->dynids.lock);
1632         INIT_LIST_HEAD(&hv_driver->dynids.list);
1633
1634         ret = driver_register(&hv_driver->driver);
1635
1636         return ret;
1637 }
1638 EXPORT_SYMBOL_GPL(__vmbus_driver_register);
1639
1640 /**
1641  * vmbus_driver_unregister() - Unregister a vmbus's driver
1642  * @hv_driver: Pointer to driver structure you want to
1643  *             un-register
1644  *
1645  * Un-register the given driver that was previous registered with a call to
1646  * vmbus_driver_register()
1647  */
1648 void vmbus_driver_unregister(struct hv_driver *hv_driver)
1649 {
1650         pr_info("unregistering driver %s\n", hv_driver->name);
1651
1652         if (!vmbus_exists()) {
1653                 driver_unregister(&hv_driver->driver);
1654                 vmbus_free_dynids(hv_driver);
1655         }
1656 }
1657 EXPORT_SYMBOL_GPL(vmbus_driver_unregister);
1658
1659
1660 /*
1661  * Called when last reference to channel is gone.
1662  */
1663 static void vmbus_chan_release(struct kobject *kobj)
1664 {
1665         struct vmbus_channel *channel
1666                 = container_of(kobj, struct vmbus_channel, kobj);
1667
1668         kfree_rcu(channel, rcu);
1669 }
1670
1671 struct vmbus_chan_attribute {
1672         struct attribute attr;
1673         ssize_t (*show)(struct vmbus_channel *chan, char *buf);
1674         ssize_t (*store)(struct vmbus_channel *chan,
1675                          const char *buf, size_t count);
1676 };
1677 #define VMBUS_CHAN_ATTR(_name, _mode, _show, _store) \
1678         struct vmbus_chan_attribute chan_attr_##_name \
1679                 = __ATTR(_name, _mode, _show, _store)
1680 #define VMBUS_CHAN_ATTR_RW(_name) \
1681         struct vmbus_chan_attribute chan_attr_##_name = __ATTR_RW(_name)
1682 #define VMBUS_CHAN_ATTR_RO(_name) \
1683         struct vmbus_chan_attribute chan_attr_##_name = __ATTR_RO(_name)
1684 #define VMBUS_CHAN_ATTR_WO(_name) \
1685         struct vmbus_chan_attribute chan_attr_##_name = __ATTR_WO(_name)
1686
1687 static ssize_t vmbus_chan_attr_show(struct kobject *kobj,
1688                                     struct attribute *attr, char *buf)
1689 {
1690         const struct vmbus_chan_attribute *attribute
1691                 = container_of(attr, struct vmbus_chan_attribute, attr);
1692         struct vmbus_channel *chan
1693                 = container_of(kobj, struct vmbus_channel, kobj);
1694
1695         if (!attribute->show)
1696                 return -EIO;
1697
1698         return attribute->show(chan, buf);
1699 }
1700
1701 static ssize_t vmbus_chan_attr_store(struct kobject *kobj,
1702                                      struct attribute *attr, const char *buf,
1703                                      size_t count)
1704 {
1705         const struct vmbus_chan_attribute *attribute
1706                 = container_of(attr, struct vmbus_chan_attribute, attr);
1707         struct vmbus_channel *chan
1708                 = container_of(kobj, struct vmbus_channel, kobj);
1709
1710         if (!attribute->store)
1711                 return -EIO;
1712
1713         return attribute->store(chan, buf, count);
1714 }
1715
1716 static const struct sysfs_ops vmbus_chan_sysfs_ops = {
1717         .show = vmbus_chan_attr_show,
1718         .store = vmbus_chan_attr_store,
1719 };
1720
1721 static ssize_t out_mask_show(struct vmbus_channel *channel, char *buf)
1722 {
1723         struct hv_ring_buffer_info *rbi = &channel->outbound;
1724         ssize_t ret;
1725
1726         mutex_lock(&rbi->ring_buffer_mutex);
1727         if (!rbi->ring_buffer) {
1728                 mutex_unlock(&rbi->ring_buffer_mutex);
1729                 return -EINVAL;
1730         }
1731
1732         ret = sprintf(buf, "%u\n", rbi->ring_buffer->interrupt_mask);
1733         mutex_unlock(&rbi->ring_buffer_mutex);
1734         return ret;
1735 }
1736 static VMBUS_CHAN_ATTR_RO(out_mask);
1737
1738 static ssize_t in_mask_show(struct vmbus_channel *channel, char *buf)
1739 {
1740         struct hv_ring_buffer_info *rbi = &channel->inbound;
1741         ssize_t ret;
1742
1743         mutex_lock(&rbi->ring_buffer_mutex);
1744         if (!rbi->ring_buffer) {
1745                 mutex_unlock(&rbi->ring_buffer_mutex);
1746                 return -EINVAL;
1747         }
1748
1749         ret = sprintf(buf, "%u\n", rbi->ring_buffer->interrupt_mask);
1750         mutex_unlock(&rbi->ring_buffer_mutex);
1751         return ret;
1752 }
1753 static VMBUS_CHAN_ATTR_RO(in_mask);
1754
1755 static ssize_t read_avail_show(struct vmbus_channel *channel, char *buf)
1756 {
1757         struct hv_ring_buffer_info *rbi = &channel->inbound;
1758         ssize_t ret;
1759
1760         mutex_lock(&rbi->ring_buffer_mutex);
1761         if (!rbi->ring_buffer) {
1762                 mutex_unlock(&rbi->ring_buffer_mutex);
1763                 return -EINVAL;
1764         }
1765
1766         ret = sprintf(buf, "%u\n", hv_get_bytes_to_read(rbi));
1767         mutex_unlock(&rbi->ring_buffer_mutex);
1768         return ret;
1769 }
1770 static VMBUS_CHAN_ATTR_RO(read_avail);
1771
1772 static ssize_t write_avail_show(struct vmbus_channel *channel, char *buf)
1773 {
1774         struct hv_ring_buffer_info *rbi = &channel->outbound;
1775         ssize_t ret;
1776
1777         mutex_lock(&rbi->ring_buffer_mutex);
1778         if (!rbi->ring_buffer) {
1779                 mutex_unlock(&rbi->ring_buffer_mutex);
1780                 return -EINVAL;
1781         }
1782
1783         ret = sprintf(buf, "%u\n", hv_get_bytes_to_write(rbi));
1784         mutex_unlock(&rbi->ring_buffer_mutex);
1785         return ret;
1786 }
1787 static VMBUS_CHAN_ATTR_RO(write_avail);
1788
1789 static ssize_t target_cpu_show(struct vmbus_channel *channel, char *buf)
1790 {
1791         return sprintf(buf, "%u\n", channel->target_cpu);
1792 }
1793 static ssize_t target_cpu_store(struct vmbus_channel *channel,
1794                                 const char *buf, size_t count)
1795 {
1796         u32 target_cpu, origin_cpu;
1797         ssize_t ret = count;
1798
1799         if (vmbus_proto_version < VERSION_WIN10_V4_1)
1800                 return -EIO;
1801
1802         if (sscanf(buf, "%uu", &target_cpu) != 1)
1803                 return -EIO;
1804
1805         /* Validate target_cpu for the cpumask_test_cpu() operation below. */
1806         if (target_cpu >= nr_cpumask_bits)
1807                 return -EINVAL;
1808
1809         /* No CPUs should come up or down during this. */
1810         cpus_read_lock();
1811
1812         if (!cpu_online(target_cpu)) {
1813                 cpus_read_unlock();
1814                 return -EINVAL;
1815         }
1816
1817         /*
1818          * Synchronizes target_cpu_store() and channel closure:
1819          *
1820          * { Initially: state = CHANNEL_OPENED }
1821          *
1822          * CPU1                         CPU2
1823          *
1824          * [target_cpu_store()]         [vmbus_disconnect_ring()]
1825          *
1826          * LOCK channel_mutex           LOCK channel_mutex
1827          * LOAD r1 = state              LOAD r2 = state
1828          * IF (r1 == CHANNEL_OPENED)    IF (r2 == CHANNEL_OPENED)
1829          *   SEND MODIFYCHANNEL           STORE state = CHANNEL_OPEN
1830          *   [...]                        SEND CLOSECHANNEL
1831          * UNLOCK channel_mutex         UNLOCK channel_mutex
1832          *
1833          * Forbids: r1 == r2 == CHANNEL_OPENED (i.e., CPU1's LOCK precedes
1834          *              CPU2's LOCK) && CPU2's SEND precedes CPU1's SEND
1835          *
1836          * Note.  The host processes the channel messages "sequentially", in
1837          * the order in which they are received on a per-partition basis.
1838          */
1839         mutex_lock(&vmbus_connection.channel_mutex);
1840
1841         /*
1842          * Hyper-V will ignore MODIFYCHANNEL messages for "non-open" channels;
1843          * avoid sending the message and fail here for such channels.
1844          */
1845         if (channel->state != CHANNEL_OPENED_STATE) {
1846                 ret = -EIO;
1847                 goto cpu_store_unlock;
1848         }
1849
1850         origin_cpu = channel->target_cpu;
1851         if (target_cpu == origin_cpu)
1852                 goto cpu_store_unlock;
1853
1854         if (vmbus_send_modifychannel(channel,
1855                                      hv_cpu_number_to_vp_number(target_cpu))) {
1856                 ret = -EIO;
1857                 goto cpu_store_unlock;
1858         }
1859
1860         /*
1861          * For version before VERSION_WIN10_V5_3, the following warning holds:
1862          *
1863          * Warning.  At this point, there is *no* guarantee that the host will
1864          * have successfully processed the vmbus_send_modifychannel() request.
1865          * See the header comment of vmbus_send_modifychannel() for more info.
1866          *
1867          * Lags in the processing of the above vmbus_send_modifychannel() can
1868          * result in missed interrupts if the "old" target CPU is taken offline
1869          * before Hyper-V starts sending interrupts to the "new" target CPU.
1870          * But apart from this offlining scenario, the code tolerates such
1871          * lags.  It will function correctly even if a channel interrupt comes
1872          * in on a CPU that is different from the channel target_cpu value.
1873          */
1874
1875         channel->target_cpu = target_cpu;
1876
1877         /* See init_vp_index(). */
1878         if (hv_is_perf_channel(channel))
1879                 hv_update_alloced_cpus(origin_cpu, target_cpu);
1880
1881         /* Currently set only for storvsc channels. */
1882         if (channel->change_target_cpu_callback) {
1883                 (*channel->change_target_cpu_callback)(channel,
1884                                 origin_cpu, target_cpu);
1885         }
1886
1887 cpu_store_unlock:
1888         mutex_unlock(&vmbus_connection.channel_mutex);
1889         cpus_read_unlock();
1890         return ret;
1891 }
1892 static VMBUS_CHAN_ATTR(cpu, 0644, target_cpu_show, target_cpu_store);
1893
1894 static ssize_t channel_pending_show(struct vmbus_channel *channel,
1895                                     char *buf)
1896 {
1897         return sprintf(buf, "%d\n",
1898                        channel_pending(channel,
1899                                        vmbus_connection.monitor_pages[1]));
1900 }
1901 static VMBUS_CHAN_ATTR(pending, 0444, channel_pending_show, NULL);
1902
1903 static ssize_t channel_latency_show(struct vmbus_channel *channel,
1904                                     char *buf)
1905 {
1906         return sprintf(buf, "%d\n",
1907                        channel_latency(channel,
1908                                        vmbus_connection.monitor_pages[1]));
1909 }
1910 static VMBUS_CHAN_ATTR(latency, 0444, channel_latency_show, NULL);
1911
1912 static ssize_t channel_interrupts_show(struct vmbus_channel *channel, char *buf)
1913 {
1914         return sprintf(buf, "%llu\n", channel->interrupts);
1915 }
1916 static VMBUS_CHAN_ATTR(interrupts, 0444, channel_interrupts_show, NULL);
1917
1918 static ssize_t channel_events_show(struct vmbus_channel *channel, char *buf)
1919 {
1920         return sprintf(buf, "%llu\n", channel->sig_events);
1921 }
1922 static VMBUS_CHAN_ATTR(events, 0444, channel_events_show, NULL);
1923
1924 static ssize_t channel_intr_in_full_show(struct vmbus_channel *channel,
1925                                          char *buf)
1926 {
1927         return sprintf(buf, "%llu\n",
1928                        (unsigned long long)channel->intr_in_full);
1929 }
1930 static VMBUS_CHAN_ATTR(intr_in_full, 0444, channel_intr_in_full_show, NULL);
1931
1932 static ssize_t channel_intr_out_empty_show(struct vmbus_channel *channel,
1933                                            char *buf)
1934 {
1935         return sprintf(buf, "%llu\n",
1936                        (unsigned long long)channel->intr_out_empty);
1937 }
1938 static VMBUS_CHAN_ATTR(intr_out_empty, 0444, channel_intr_out_empty_show, NULL);
1939
1940 static ssize_t channel_out_full_first_show(struct vmbus_channel *channel,
1941                                            char *buf)
1942 {
1943         return sprintf(buf, "%llu\n",
1944                        (unsigned long long)channel->out_full_first);
1945 }
1946 static VMBUS_CHAN_ATTR(out_full_first, 0444, channel_out_full_first_show, NULL);
1947
1948 static ssize_t channel_out_full_total_show(struct vmbus_channel *channel,
1949                                            char *buf)
1950 {
1951         return sprintf(buf, "%llu\n",
1952                        (unsigned long long)channel->out_full_total);
1953 }
1954 static VMBUS_CHAN_ATTR(out_full_total, 0444, channel_out_full_total_show, NULL);
1955
1956 static ssize_t subchannel_monitor_id_show(struct vmbus_channel *channel,
1957                                           char *buf)
1958 {
1959         return sprintf(buf, "%u\n", channel->offermsg.monitorid);
1960 }
1961 static VMBUS_CHAN_ATTR(monitor_id, 0444, subchannel_monitor_id_show, NULL);
1962
1963 static ssize_t subchannel_id_show(struct vmbus_channel *channel,
1964                                   char *buf)
1965 {
1966         return sprintf(buf, "%u\n",
1967                        channel->offermsg.offer.sub_channel_index);
1968 }
1969 static VMBUS_CHAN_ATTR_RO(subchannel_id);
1970
1971 static struct attribute *vmbus_chan_attrs[] = {
1972         &chan_attr_out_mask.attr,
1973         &chan_attr_in_mask.attr,
1974         &chan_attr_read_avail.attr,
1975         &chan_attr_write_avail.attr,
1976         &chan_attr_cpu.attr,
1977         &chan_attr_pending.attr,
1978         &chan_attr_latency.attr,
1979         &chan_attr_interrupts.attr,
1980         &chan_attr_events.attr,
1981         &chan_attr_intr_in_full.attr,
1982         &chan_attr_intr_out_empty.attr,
1983         &chan_attr_out_full_first.attr,
1984         &chan_attr_out_full_total.attr,
1985         &chan_attr_monitor_id.attr,
1986         &chan_attr_subchannel_id.attr,
1987         NULL
1988 };
1989
1990 /*
1991  * Channel-level attribute_group callback function. Returns the permission for
1992  * each attribute, and returns 0 if an attribute is not visible.
1993  */
1994 static umode_t vmbus_chan_attr_is_visible(struct kobject *kobj,
1995                                           struct attribute *attr, int idx)
1996 {
1997         const struct vmbus_channel *channel =
1998                 container_of(kobj, struct vmbus_channel, kobj);
1999
2000         /* Hide the monitor attributes if the monitor mechanism is not used. */
2001         if (!channel->offermsg.monitor_allocated &&
2002             (attr == &chan_attr_pending.attr ||
2003              attr == &chan_attr_latency.attr ||
2004              attr == &chan_attr_monitor_id.attr))
2005                 return 0;
2006
2007         return attr->mode;
2008 }
2009
2010 static struct attribute_group vmbus_chan_group = {
2011         .attrs = vmbus_chan_attrs,
2012         .is_visible = vmbus_chan_attr_is_visible
2013 };
2014
2015 static struct kobj_type vmbus_chan_ktype = {
2016         .sysfs_ops = &vmbus_chan_sysfs_ops,
2017         .release = vmbus_chan_release,
2018 };
2019
2020 /*
2021  * vmbus_add_channel_kobj - setup a sub-directory under device/channels
2022  */
2023 int vmbus_add_channel_kobj(struct hv_device *dev, struct vmbus_channel *channel)
2024 {
2025         const struct device *device = &dev->device;
2026         struct kobject *kobj = &channel->kobj;
2027         u32 relid = channel->offermsg.child_relid;
2028         int ret;
2029
2030         kobj->kset = dev->channels_kset;
2031         ret = kobject_init_and_add(kobj, &vmbus_chan_ktype, NULL,
2032                                    "%u", relid);
2033         if (ret)
2034                 return ret;
2035
2036         ret = sysfs_create_group(kobj, &vmbus_chan_group);
2037
2038         if (ret) {
2039                 /*
2040                  * The calling functions' error handling paths will cleanup the
2041                  * empty channel directory.
2042                  */
2043                 dev_err(device, "Unable to set up channel sysfs files\n");
2044                 return ret;
2045         }
2046
2047         kobject_uevent(kobj, KOBJ_ADD);
2048
2049         return 0;
2050 }
2051
2052 /*
2053  * vmbus_remove_channel_attr_group - remove the channel's attribute group
2054  */
2055 void vmbus_remove_channel_attr_group(struct vmbus_channel *channel)
2056 {
2057         sysfs_remove_group(&channel->kobj, &vmbus_chan_group);
2058 }
2059
2060 /*
2061  * vmbus_device_create - Creates and registers a new child device
2062  * on the vmbus.
2063  */
2064 struct hv_device *vmbus_device_create(const guid_t *type,
2065                                       const guid_t *instance,
2066                                       struct vmbus_channel *channel)
2067 {
2068         struct hv_device *child_device_obj;
2069
2070         child_device_obj = kzalloc(sizeof(struct hv_device), GFP_KERNEL);
2071         if (!child_device_obj) {
2072                 pr_err("Unable to allocate device object for child device\n");
2073                 return NULL;
2074         }
2075
2076         child_device_obj->channel = channel;
2077         guid_copy(&child_device_obj->dev_type, type);
2078         guid_copy(&child_device_obj->dev_instance, instance);
2079         child_device_obj->vendor_id = 0x1414; /* MSFT vendor ID */
2080
2081         return child_device_obj;
2082 }
2083
2084 /*
2085  * vmbus_device_register - Register the child device
2086  */
2087 int vmbus_device_register(struct hv_device *child_device_obj)
2088 {
2089         struct kobject *kobj = &child_device_obj->device.kobj;
2090         int ret;
2091
2092         dev_set_name(&child_device_obj->device, "%pUl",
2093                      &child_device_obj->channel->offermsg.offer.if_instance);
2094
2095         child_device_obj->device.bus = &hv_bus;
2096         child_device_obj->device.parent = &hv_acpi_dev->dev;
2097         child_device_obj->device.release = vmbus_device_release;
2098
2099         /*
2100          * Register with the LDM. This will kick off the driver/device
2101          * binding...which will eventually call vmbus_match() and vmbus_probe()
2102          */
2103         ret = device_register(&child_device_obj->device);
2104         if (ret) {
2105                 pr_err("Unable to register child device\n");
2106                 return ret;
2107         }
2108
2109         child_device_obj->channels_kset = kset_create_and_add("channels",
2110                                                               NULL, kobj);
2111         if (!child_device_obj->channels_kset) {
2112                 ret = -ENOMEM;
2113                 goto err_dev_unregister;
2114         }
2115
2116         ret = vmbus_add_channel_kobj(child_device_obj,
2117                                      child_device_obj->channel);
2118         if (ret) {
2119                 pr_err("Unable to register primary channeln");
2120                 goto err_kset_unregister;
2121         }
2122         hv_debug_add_dev_dir(child_device_obj);
2123
2124         return 0;
2125
2126 err_kset_unregister:
2127         kset_unregister(child_device_obj->channels_kset);
2128
2129 err_dev_unregister:
2130         device_unregister(&child_device_obj->device);
2131         return ret;
2132 }
2133
2134 /*
2135  * vmbus_device_unregister - Remove the specified child device
2136  * from the vmbus.
2137  */
2138 void vmbus_device_unregister(struct hv_device *device_obj)
2139 {
2140         pr_debug("child device %s unregistered\n",
2141                 dev_name(&device_obj->device));
2142
2143         kset_unregister(device_obj->channels_kset);
2144
2145         /*
2146          * Kick off the process of unregistering the device.
2147          * This will call vmbus_remove() and eventually vmbus_device_release()
2148          */
2149         device_unregister(&device_obj->device);
2150 }
2151
2152
2153 /*
2154  * VMBUS is an acpi enumerated device. Get the information we
2155  * need from DSDT.
2156  */
2157 #define VTPM_BASE_ADDRESS 0xfed40000
2158 static acpi_status vmbus_walk_resources(struct acpi_resource *res, void *ctx)
2159 {
2160         resource_size_t start = 0;
2161         resource_size_t end = 0;
2162         struct resource *new_res;
2163         struct resource **old_res = &hyperv_mmio;
2164         struct resource **prev_res = NULL;
2165         struct resource r;
2166
2167         switch (res->type) {
2168
2169         /*
2170          * "Address" descriptors are for bus windows. Ignore
2171          * "memory" descriptors, which are for registers on
2172          * devices.
2173          */
2174         case ACPI_RESOURCE_TYPE_ADDRESS32:
2175                 start = res->data.address32.address.minimum;
2176                 end = res->data.address32.address.maximum;
2177                 break;
2178
2179         case ACPI_RESOURCE_TYPE_ADDRESS64:
2180                 start = res->data.address64.address.minimum;
2181                 end = res->data.address64.address.maximum;
2182                 break;
2183
2184         /*
2185          * The IRQ information is needed only on ARM64, which Hyper-V
2186          * sets up in the extended format. IRQ information is present
2187          * on x86/x64 in the non-extended format but it is not used by
2188          * Linux. So don't bother checking for the non-extended format.
2189          */
2190         case ACPI_RESOURCE_TYPE_EXTENDED_IRQ:
2191                 if (!acpi_dev_resource_interrupt(res, 0, &r)) {
2192                         pr_err("Unable to parse Hyper-V ACPI interrupt\n");
2193                         return AE_ERROR;
2194                 }
2195                 /* ARM64 INTID for VMbus */
2196                 vmbus_interrupt = res->data.extended_irq.interrupts[0];
2197                 /* Linux IRQ number */
2198                 vmbus_irq = r.start;
2199                 return AE_OK;
2200
2201         default:
2202                 /* Unused resource type */
2203                 return AE_OK;
2204
2205         }
2206         /*
2207          * Ignore ranges that are below 1MB, as they're not
2208          * necessary or useful here.
2209          */
2210         if (end < 0x100000)
2211                 return AE_OK;
2212
2213         new_res = kzalloc(sizeof(*new_res), GFP_ATOMIC);
2214         if (!new_res)
2215                 return AE_NO_MEMORY;
2216
2217         /* If this range overlaps the virtual TPM, truncate it. */
2218         if (end > VTPM_BASE_ADDRESS && start < VTPM_BASE_ADDRESS)
2219                 end = VTPM_BASE_ADDRESS;
2220
2221         new_res->name = "hyperv mmio";
2222         new_res->flags = IORESOURCE_MEM;
2223         new_res->start = start;
2224         new_res->end = end;
2225
2226         /*
2227          * If two ranges are adjacent, merge them.
2228          */
2229         do {
2230                 if (!*old_res) {
2231                         *old_res = new_res;
2232                         break;
2233                 }
2234
2235                 if (((*old_res)->end + 1) == new_res->start) {
2236                         (*old_res)->end = new_res->end;
2237                         kfree(new_res);
2238                         break;
2239                 }
2240
2241                 if ((*old_res)->start == new_res->end + 1) {
2242                         (*old_res)->start = new_res->start;
2243                         kfree(new_res);
2244                         break;
2245                 }
2246
2247                 if ((*old_res)->start > new_res->end) {
2248                         new_res->sibling = *old_res;
2249                         if (prev_res)
2250                                 (*prev_res)->sibling = new_res;
2251                         *old_res = new_res;
2252                         break;
2253                 }
2254
2255                 prev_res = old_res;
2256                 old_res = &(*old_res)->sibling;
2257
2258         } while (1);
2259
2260         return AE_OK;
2261 }
2262
2263 static int vmbus_acpi_remove(struct acpi_device *device)
2264 {
2265         struct resource *cur_res;
2266         struct resource *next_res;
2267
2268         if (hyperv_mmio) {
2269                 if (fb_mmio) {
2270                         __release_region(hyperv_mmio, fb_mmio->start,
2271                                          resource_size(fb_mmio));
2272                         fb_mmio = NULL;
2273                 }
2274
2275                 for (cur_res = hyperv_mmio; cur_res; cur_res = next_res) {
2276                         next_res = cur_res->sibling;
2277                         kfree(cur_res);
2278                 }
2279         }
2280
2281         return 0;
2282 }
2283
2284 static void vmbus_reserve_fb(void)
2285 {
2286         int size;
2287         /*
2288          * Make a claim for the frame buffer in the resource tree under the
2289          * first node, which will be the one below 4GB.  The length seems to
2290          * be underreported, particularly in a Generation 1 VM.  So start out
2291          * reserving a larger area and make it smaller until it succeeds.
2292          */
2293
2294         if (screen_info.lfb_base) {
2295                 if (efi_enabled(EFI_BOOT))
2296                         size = max_t(__u32, screen_info.lfb_size, 0x800000);
2297                 else
2298                         size = max_t(__u32, screen_info.lfb_size, 0x4000000);
2299
2300                 for (; !fb_mmio && (size >= 0x100000); size >>= 1) {
2301                         fb_mmio = __request_region(hyperv_mmio,
2302                                                    screen_info.lfb_base, size,
2303                                                    fb_mmio_name, 0);
2304                 }
2305         }
2306 }
2307
2308 /**
2309  * vmbus_allocate_mmio() - Pick a memory-mapped I/O range.
2310  * @new:                If successful, supplied a pointer to the
2311  *                      allocated MMIO space.
2312  * @device_obj:         Identifies the caller
2313  * @min:                Minimum guest physical address of the
2314  *                      allocation
2315  * @max:                Maximum guest physical address
2316  * @size:               Size of the range to be allocated
2317  * @align:              Alignment of the range to be allocated
2318  * @fb_overlap_ok:      Whether this allocation can be allowed
2319  *                      to overlap the video frame buffer.
2320  *
2321  * This function walks the resources granted to VMBus by the
2322  * _CRS object in the ACPI namespace underneath the parent
2323  * "bridge" whether that's a root PCI bus in the Generation 1
2324  * case or a Module Device in the Generation 2 case.  It then
2325  * attempts to allocate from the global MMIO pool in a way that
2326  * matches the constraints supplied in these parameters and by
2327  * that _CRS.
2328  *
2329  * Return: 0 on success, -errno on failure
2330  */
2331 int vmbus_allocate_mmio(struct resource **new, struct hv_device *device_obj,
2332                         resource_size_t min, resource_size_t max,
2333                         resource_size_t size, resource_size_t align,
2334                         bool fb_overlap_ok)
2335 {
2336         struct resource *iter, *shadow;
2337         resource_size_t range_min, range_max, start;
2338         const char *dev_n = dev_name(&device_obj->device);
2339         int retval;
2340
2341         retval = -ENXIO;
2342         mutex_lock(&hyperv_mmio_lock);
2343
2344         /*
2345          * If overlaps with frame buffers are allowed, then first attempt to
2346          * make the allocation from within the reserved region.  Because it
2347          * is already reserved, no shadow allocation is necessary.
2348          */
2349         if (fb_overlap_ok && fb_mmio && !(min > fb_mmio->end) &&
2350             !(max < fb_mmio->start)) {
2351
2352                 range_min = fb_mmio->start;
2353                 range_max = fb_mmio->end;
2354                 start = (range_min + align - 1) & ~(align - 1);
2355                 for (; start + size - 1 <= range_max; start += align) {
2356                         *new = request_mem_region_exclusive(start, size, dev_n);
2357                         if (*new) {
2358                                 retval = 0;
2359                                 goto exit;
2360                         }
2361                 }
2362         }
2363
2364         for (iter = hyperv_mmio; iter; iter = iter->sibling) {
2365                 if ((iter->start >= max) || (iter->end <= min))
2366                         continue;
2367
2368                 range_min = iter->start;
2369                 range_max = iter->end;
2370                 start = (range_min + align - 1) & ~(align - 1);
2371                 for (; start + size - 1 <= range_max; start += align) {
2372                         shadow = __request_region(iter, start, size, NULL,
2373                                                   IORESOURCE_BUSY);
2374                         if (!shadow)
2375                                 continue;
2376
2377                         *new = request_mem_region_exclusive(start, size, dev_n);
2378                         if (*new) {
2379                                 shadow->name = (char *)*new;
2380                                 retval = 0;
2381                                 goto exit;
2382                         }
2383
2384                         __release_region(iter, start, size);
2385                 }
2386         }
2387
2388 exit:
2389         mutex_unlock(&hyperv_mmio_lock);
2390         return retval;
2391 }
2392 EXPORT_SYMBOL_GPL(vmbus_allocate_mmio);
2393
2394 /**
2395  * vmbus_free_mmio() - Free a memory-mapped I/O range.
2396  * @start:              Base address of region to release.
2397  * @size:               Size of the range to be allocated
2398  *
2399  * This function releases anything requested by
2400  * vmbus_mmio_allocate().
2401  */
2402 void vmbus_free_mmio(resource_size_t start, resource_size_t size)
2403 {
2404         struct resource *iter;
2405
2406         mutex_lock(&hyperv_mmio_lock);
2407         for (iter = hyperv_mmio; iter; iter = iter->sibling) {
2408                 if ((iter->start >= start + size) || (iter->end <= start))
2409                         continue;
2410
2411                 __release_region(iter, start, size);
2412         }
2413         release_mem_region(start, size);
2414         mutex_unlock(&hyperv_mmio_lock);
2415
2416 }
2417 EXPORT_SYMBOL_GPL(vmbus_free_mmio);
2418
2419 static int vmbus_acpi_add(struct acpi_device *device)
2420 {
2421         acpi_status result;
2422         int ret_val = -ENODEV;
2423         struct acpi_device *ancestor;
2424
2425         hv_acpi_dev = device;
2426
2427         result = acpi_walk_resources(device->handle, METHOD_NAME__CRS,
2428                                         vmbus_walk_resources, NULL);
2429
2430         if (ACPI_FAILURE(result))
2431                 goto acpi_walk_err;
2432         /*
2433          * Some ancestor of the vmbus acpi device (Gen1 or Gen2
2434          * firmware) is the VMOD that has the mmio ranges. Get that.
2435          */
2436         for (ancestor = device->parent; ancestor; ancestor = ancestor->parent) {
2437                 result = acpi_walk_resources(ancestor->handle, METHOD_NAME__CRS,
2438                                              vmbus_walk_resources, NULL);
2439
2440                 if (ACPI_FAILURE(result))
2441                         continue;
2442                 if (hyperv_mmio) {
2443                         vmbus_reserve_fb();
2444                         break;
2445                 }
2446         }
2447         ret_val = 0;
2448
2449 acpi_walk_err:
2450         complete(&probe_event);
2451         if (ret_val)
2452                 vmbus_acpi_remove(device);
2453         return ret_val;
2454 }
2455
2456 #ifdef CONFIG_PM_SLEEP
2457 static int vmbus_bus_suspend(struct device *dev)
2458 {
2459         struct vmbus_channel *channel, *sc;
2460
2461         while (atomic_read(&vmbus_connection.offer_in_progress) != 0) {
2462                 /*
2463                  * We wait here until the completion of any channel
2464                  * offers that are currently in progress.
2465                  */
2466                 usleep_range(1000, 2000);
2467         }
2468
2469         mutex_lock(&vmbus_connection.channel_mutex);
2470         list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
2471                 if (!is_hvsock_channel(channel))
2472                         continue;
2473
2474                 vmbus_force_channel_rescinded(channel);
2475         }
2476         mutex_unlock(&vmbus_connection.channel_mutex);
2477
2478         /*
2479          * Wait until all the sub-channels and hv_sock channels have been
2480          * cleaned up. Sub-channels should be destroyed upon suspend, otherwise
2481          * they would conflict with the new sub-channels that will be created
2482          * in the resume path. hv_sock channels should also be destroyed, but
2483          * a hv_sock channel of an established hv_sock connection can not be
2484          * really destroyed since it may still be referenced by the userspace
2485          * application, so we just force the hv_sock channel to be rescinded
2486          * by vmbus_force_channel_rescinded(), and the userspace application
2487          * will thoroughly destroy the channel after hibernation.
2488          *
2489          * Note: the counter nr_chan_close_on_suspend may never go above 0 if
2490          * the VM has no sub-channel and hv_sock channel, e.g. a 1-vCPU VM.
2491          */
2492         if (atomic_read(&vmbus_connection.nr_chan_close_on_suspend) > 0)
2493                 wait_for_completion(&vmbus_connection.ready_for_suspend_event);
2494
2495         if (atomic_read(&vmbus_connection.nr_chan_fixup_on_resume) != 0) {
2496                 pr_err("Can not suspend due to a previous failed resuming\n");
2497                 return -EBUSY;
2498         }
2499
2500         mutex_lock(&vmbus_connection.channel_mutex);
2501
2502         list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
2503                 /*
2504                  * Remove the channel from the array of channels and invalidate
2505                  * the channel's relid.  Upon resume, vmbus_onoffer() will fix
2506                  * up the relid (and other fields, if necessary) and add the
2507                  * channel back to the array.
2508                  */
2509                 vmbus_channel_unmap_relid(channel);
2510                 channel->offermsg.child_relid = INVALID_RELID;
2511
2512                 if (is_hvsock_channel(channel)) {
2513                         if (!channel->rescind) {
2514                                 pr_err("hv_sock channel not rescinded!\n");
2515                                 WARN_ON_ONCE(1);
2516                         }
2517                         continue;
2518                 }
2519
2520                 list_for_each_entry(sc, &channel->sc_list, sc_list) {
2521                         pr_err("Sub-channel not deleted!\n");
2522                         WARN_ON_ONCE(1);
2523                 }
2524
2525                 atomic_inc(&vmbus_connection.nr_chan_fixup_on_resume);
2526         }
2527
2528         mutex_unlock(&vmbus_connection.channel_mutex);
2529
2530         vmbus_initiate_unload(false);
2531
2532         /* Reset the event for the next resume. */
2533         reinit_completion(&vmbus_connection.ready_for_resume_event);
2534
2535         return 0;
2536 }
2537
2538 static int vmbus_bus_resume(struct device *dev)
2539 {
2540         struct vmbus_channel_msginfo *msginfo;
2541         size_t msgsize;
2542         int ret;
2543
2544         /*
2545          * We only use the 'vmbus_proto_version', which was in use before
2546          * hibernation, to re-negotiate with the host.
2547          */
2548         if (!vmbus_proto_version) {
2549                 pr_err("Invalid proto version = 0x%x\n", vmbus_proto_version);
2550                 return -EINVAL;
2551         }
2552
2553         msgsize = sizeof(*msginfo) +
2554                   sizeof(struct vmbus_channel_initiate_contact);
2555
2556         msginfo = kzalloc(msgsize, GFP_KERNEL);
2557
2558         if (msginfo == NULL)
2559                 return -ENOMEM;
2560
2561         ret = vmbus_negotiate_version(msginfo, vmbus_proto_version);
2562
2563         kfree(msginfo);
2564
2565         if (ret != 0)
2566                 return ret;
2567
2568         WARN_ON(atomic_read(&vmbus_connection.nr_chan_fixup_on_resume) == 0);
2569
2570         vmbus_request_offers();
2571
2572         if (wait_for_completion_timeout(
2573                 &vmbus_connection.ready_for_resume_event, 10 * HZ) == 0)
2574                 pr_err("Some vmbus device is missing after suspending?\n");
2575
2576         /* Reset the event for the next suspend. */
2577         reinit_completion(&vmbus_connection.ready_for_suspend_event);
2578
2579         return 0;
2580 }
2581 #else
2582 #define vmbus_bus_suspend NULL
2583 #define vmbus_bus_resume NULL
2584 #endif /* CONFIG_PM_SLEEP */
2585
2586 static const struct acpi_device_id vmbus_acpi_device_ids[] = {
2587         {"VMBUS", 0},
2588         {"VMBus", 0},
2589         {"", 0},
2590 };
2591 MODULE_DEVICE_TABLE(acpi, vmbus_acpi_device_ids);
2592
2593 /*
2594  * Note: we must use the "no_irq" ops, otherwise hibernation can not work with
2595  * PCI device assignment, because "pci_dev_pm_ops" uses the "noirq" ops: in
2596  * the resume path, the pci "noirq" restore op runs before "non-noirq" op (see
2597  * resume_target_kernel() -> dpm_resume_start(), and hibernation_restore() ->
2598  * dpm_resume_end()). This means vmbus_bus_resume() and the pci-hyperv's
2599  * resume callback must also run via the "noirq" ops.
2600  *
2601  * Set suspend_noirq/resume_noirq to NULL for Suspend-to-Idle: see the comment
2602  * earlier in this file before vmbus_pm.
2603  */
2604
2605 static const struct dev_pm_ops vmbus_bus_pm = {
2606         .suspend_noirq  = NULL,
2607         .resume_noirq   = NULL,
2608         .freeze_noirq   = vmbus_bus_suspend,
2609         .thaw_noirq     = vmbus_bus_resume,
2610         .poweroff_noirq = vmbus_bus_suspend,
2611         .restore_noirq  = vmbus_bus_resume
2612 };
2613
2614 static struct acpi_driver vmbus_acpi_driver = {
2615         .name = "vmbus",
2616         .ids = vmbus_acpi_device_ids,
2617         .ops = {
2618                 .add = vmbus_acpi_add,
2619                 .remove = vmbus_acpi_remove,
2620         },
2621         .drv.pm = &vmbus_bus_pm,
2622 };
2623
2624 static void hv_kexec_handler(void)
2625 {
2626         hv_stimer_global_cleanup();
2627         vmbus_initiate_unload(false);
2628         /* Make sure conn_state is set as hv_synic_cleanup checks for it */
2629         mb();
2630         cpuhp_remove_state(hyperv_cpuhp_online);
2631 };
2632
2633 static void hv_crash_handler(struct pt_regs *regs)
2634 {
2635         int cpu;
2636
2637         vmbus_initiate_unload(true);
2638         /*
2639          * In crash handler we can't schedule synic cleanup for all CPUs,
2640          * doing the cleanup for current CPU only. This should be sufficient
2641          * for kdump.
2642          */
2643         cpu = smp_processor_id();
2644         hv_stimer_cleanup(cpu);
2645         hv_synic_disable_regs(cpu);
2646 };
2647
2648 static int hv_synic_suspend(void)
2649 {
2650         /*
2651          * When we reach here, all the non-boot CPUs have been offlined.
2652          * If we're in a legacy configuration where stimer Direct Mode is
2653          * not enabled, the stimers on the non-boot CPUs have been unbound
2654          * in hv_synic_cleanup() -> hv_stimer_legacy_cleanup() ->
2655          * hv_stimer_cleanup() -> clockevents_unbind_device().
2656          *
2657          * hv_synic_suspend() only runs on CPU0 with interrupts disabled.
2658          * Here we do not call hv_stimer_legacy_cleanup() on CPU0 because:
2659          * 1) it's unnecessary as interrupts remain disabled between
2660          * syscore_suspend() and syscore_resume(): see create_image() and
2661          * resume_target_kernel()
2662          * 2) the stimer on CPU0 is automatically disabled later by
2663          * syscore_suspend() -> timekeeping_suspend() -> tick_suspend() -> ...
2664          * -> clockevents_shutdown() -> ... -> hv_ce_shutdown()
2665          * 3) a warning would be triggered if we call
2666          * clockevents_unbind_device(), which may sleep, in an
2667          * interrupts-disabled context.
2668          */
2669
2670         hv_synic_disable_regs(0);
2671
2672         return 0;
2673 }
2674
2675 static void hv_synic_resume(void)
2676 {
2677         hv_synic_enable_regs(0);
2678
2679         /*
2680          * Note: we don't need to call hv_stimer_init(0), because the timer
2681          * on CPU0 is not unbound in hv_synic_suspend(), and the timer is
2682          * automatically re-enabled in timekeeping_resume().
2683          */
2684 }
2685
2686 /* The callbacks run only on CPU0, with irqs_disabled. */
2687 static struct syscore_ops hv_synic_syscore_ops = {
2688         .suspend = hv_synic_suspend,
2689         .resume = hv_synic_resume,
2690 };
2691
2692 static int __init hv_acpi_init(void)
2693 {
2694         int ret, t;
2695
2696         if (!hv_is_hyperv_initialized())
2697                 return -ENODEV;
2698
2699         if (hv_root_partition)
2700                 return 0;
2701
2702         init_completion(&probe_event);
2703
2704         /*
2705          * Get ACPI resources first.
2706          */
2707         ret = acpi_bus_register_driver(&vmbus_acpi_driver);
2708
2709         if (ret)
2710                 return ret;
2711
2712         t = wait_for_completion_timeout(&probe_event, 5*HZ);
2713         if (t == 0) {
2714                 ret = -ETIMEDOUT;
2715                 goto cleanup;
2716         }
2717
2718         /*
2719          * If we're on an architecture with a hardcoded hypervisor
2720          * vector (i.e. x86/x64), override the VMbus interrupt found
2721          * in the ACPI tables. Ensure vmbus_irq is not set since the
2722          * normal Linux IRQ mechanism is not used in this case.
2723          */
2724 #ifdef HYPERVISOR_CALLBACK_VECTOR
2725         vmbus_interrupt = HYPERVISOR_CALLBACK_VECTOR;
2726         vmbus_irq = -1;
2727 #endif
2728
2729         hv_debug_init();
2730
2731         ret = vmbus_bus_init();
2732         if (ret)
2733                 goto cleanup;
2734
2735         hv_setup_kexec_handler(hv_kexec_handler);
2736         hv_setup_crash_handler(hv_crash_handler);
2737
2738         register_syscore_ops(&hv_synic_syscore_ops);
2739
2740         return 0;
2741
2742 cleanup:
2743         acpi_bus_unregister_driver(&vmbus_acpi_driver);
2744         hv_acpi_dev = NULL;
2745         return ret;
2746 }
2747
2748 static void __exit vmbus_exit(void)
2749 {
2750         int cpu;
2751
2752         unregister_syscore_ops(&hv_synic_syscore_ops);
2753
2754         hv_remove_kexec_handler();
2755         hv_remove_crash_handler();
2756         vmbus_connection.conn_state = DISCONNECTED;
2757         hv_stimer_global_cleanup();
2758         vmbus_disconnect();
2759         if (vmbus_irq == -1) {
2760                 hv_remove_vmbus_handler();
2761         } else {
2762                 free_percpu_irq(vmbus_irq, vmbus_evt);
2763                 free_percpu(vmbus_evt);
2764         }
2765         for_each_online_cpu(cpu) {
2766                 struct hv_per_cpu_context *hv_cpu
2767                         = per_cpu_ptr(hv_context.cpu_context, cpu);
2768
2769                 tasklet_kill(&hv_cpu->msg_dpc);
2770         }
2771         hv_debug_rm_all_dir();
2772
2773         vmbus_free_channels();
2774         kfree(vmbus_connection.channels);
2775
2776         if (ms_hyperv.misc_features & HV_FEATURE_GUEST_CRASH_MSR_AVAILABLE) {
2777                 kmsg_dump_unregister(&hv_kmsg_dumper);
2778                 unregister_die_notifier(&hyperv_die_block);
2779                 atomic_notifier_chain_unregister(&panic_notifier_list,
2780                                                  &hyperv_panic_block);
2781         }
2782
2783         free_page((unsigned long)hv_panic_page);
2784         unregister_sysctl_table(hv_ctl_table_hdr);
2785         hv_ctl_table_hdr = NULL;
2786         bus_unregister(&hv_bus);
2787
2788         cpuhp_remove_state(hyperv_cpuhp_online);
2789         hv_synic_free();
2790         acpi_bus_unregister_driver(&vmbus_acpi_driver);
2791 }
2792
2793
2794 MODULE_LICENSE("GPL");
2795 MODULE_DESCRIPTION("Microsoft Hyper-V VMBus Driver");
2796
2797 subsys_initcall(hv_acpi_init);
2798 module_exit(vmbus_exit);