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 hv_kmsg_dump() with kmsg data later, don't do the notification
  80          * 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 hv_kmsg_dump() with kmsg data later, don't do the notification
 103          * 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 void 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
 937 /*
 938  * vmbus_shutdown - Shutdown a vmbus device
 939  */
 940 static void vmbus_shutdown(struct device *child_device)
 941 {
 942         struct hv_driver *drv;
 943         struct hv_device *dev = device_to_hv_device(child_device);
 944
 945
 946         /* The device may not be attached yet */
 947         if (!child_device->driver)
 948                 return;
 949
 950         drv = drv_to_hv_drv(child_device->driver);
 951
 952         if (drv->shutdown)
 953                 drv->shutdown(dev);
 954 }
 955
 956 #ifdef CONFIG_PM_SLEEP
 957 /*
 958  * vmbus_suspend - Suspend a vmbus device
 959  */
 960 static int vmbus_suspend(struct device *child_device)
 961 {
 962         struct hv_driver *drv;
 963         struct hv_device *dev = device_to_hv_device(child_device);
 964
 965         /* The device may not be attached yet */
 966         if (!child_device->driver)
 967                 return 0;
 968
 969         drv = drv_to_hv_drv(child_device->driver);
 970         if (!drv->suspend)
 971                 return -EOPNOTSUPP;
 972
 973         return drv->suspend(dev);
 974 }
 975
 976 /*
 977  * vmbus_resume - Resume a vmbus device
 978  */
 979 static int vmbus_resume(struct device *child_device)
 980 {
 981         struct hv_driver *drv;
 982         struct hv_device *dev = device_to_hv_device(child_device);
 983
 984         /* The device may not be attached yet */
 985         if (!child_device->driver)
 986                 return 0;
 987
 988         drv = drv_to_hv_drv(child_device->driver);
 989         if (!drv->resume)
 990                 return -EOPNOTSUPP;
 991
 992         return drv->resume(dev);
 993 }
 994 #else
 995 #define vmbus_suspend NULL
 996 #define vmbus_resume NULL
 997 #endif /* CONFIG_PM_SLEEP */
 998
 999 /*
1000  * vmbus_device_release - Final callback release of the vmbus child device
1001  */
1002 static void vmbus_device_release(struct device *device)
1003 {
1004         struct hv_device *hv_dev = device_to_hv_device(device);
1005         struct vmbus_channel *channel = hv_dev->channel;
1006
1007         hv_debug_rm_dev_dir(hv_dev);
1008
1009         mutex_lock(&vmbus_connection.channel_mutex);
1010         hv_process_channel_removal(channel);
1011         mutex_unlock(&vmbus_connection.channel_mutex);
1012         kfree(hv_dev);
1013 }
1014
1015 /*
1016  * Note: we must use the "noirq" ops: see the comment before vmbus_bus_pm.
1017  *
1018  * suspend_noirq/resume_noirq are set to NULL to support Suspend-to-Idle: we
1019  * shouldn't suspend the vmbus devices upon Suspend-to-Idle, otherwise there
1020  * is no way to wake up a Generation-2 VM.
1021  *
1022  * The other 4 ops are for hibernation.
1023  */
1024
1025 static const struct dev_pm_ops vmbus_pm = {
1026         .suspend_noirq  = NULL,
1027         .resume_noirq   = NULL,
1028         .freeze_noirq   = vmbus_suspend,
1029         .thaw_noirq     = vmbus_resume,
1030         .poweroff_noirq = vmbus_suspend,
1031         .restore_noirq  = vmbus_resume,
1032 };
1033
1034 /* The one and only one */
1035 static struct bus_type  hv_bus = {
1036         .name =         "vmbus",
1037         .match =                vmbus_match,
1038         .shutdown =             vmbus_shutdown,
1039         .remove =               vmbus_remove,
1040         .probe =                vmbus_probe,
1041         .uevent =               vmbus_uevent,
1042         .dev_groups =           vmbus_dev_groups,
1043         .drv_groups =           vmbus_drv_groups,
1044         .bus_groups =           vmbus_bus_groups,
1045         .pm =                   &vmbus_pm,
1046 };
1047
1048 struct onmessage_work_context {
1049         struct work_struct work;
1050         struct {
1051                 struct hv_message_header header;
1052                 u8 payload[];
1053         } msg;
1054 };
1055
1056 static void vmbus_onmessage_work(struct work_struct *work)
1057 {
1058         struct onmessage_work_context *ctx;
1059
1060         /* Do not process messages if we're in DISCONNECTED state */
1061         if (vmbus_connection.conn_state == DISCONNECTED)
1062                 return;
1063
1064         ctx = container_of(work, struct onmessage_work_context,
1065                            work);
1066         vmbus_onmessage((struct vmbus_channel_message_header *)
1067                         &ctx->msg.payload);
1068         kfree(ctx);
1069 }
1070
1071 void vmbus_on_msg_dpc(unsigned long data)
1072 {
1073         struct hv_per_cpu_context *hv_cpu = (void *)data;
1074         void *page_addr = hv_cpu->synic_message_page;
1075         struct hv_message msg_copy, *msg = (struct hv_message *)page_addr +
1076                                   VMBUS_MESSAGE_SINT;
1077         struct vmbus_channel_message_header *hdr;
1078         enum vmbus_channel_message_type msgtype;
1079         const struct vmbus_channel_message_table_entry *entry;
1080         struct onmessage_work_context *ctx;
1081         __u8 payload_size;
1082         u32 message_type;
1083
1084         /*
1085          * 'enum vmbus_channel_message_type' is supposed to always be 'u32' as
1086          * it is being used in 'struct vmbus_channel_message_header' definition
1087          * which is supposed to match hypervisor ABI.
1088          */
1089         BUILD_BUG_ON(sizeof(enum vmbus_channel_message_type) != sizeof(u32));
1090
1091         /*
1092          * Since the message is in memory shared with the host, an erroneous or
1093          * malicious Hyper-V could modify the message while vmbus_on_msg_dpc()
1094          * or individual message handlers are executing; to prevent this, copy
1095          * the message into private memory.
1096          */
1097         memcpy(&msg_copy, msg, sizeof(struct hv_message));
1098
1099         message_type = msg_copy.header.message_type;
1100         if (message_type == HVMSG_NONE)
1101                 /* no msg */
1102                 return;
1103
1104         hdr = (struct vmbus_channel_message_header *)msg_copy.u.payload;
1105         msgtype = hdr->msgtype;
1106
1107         trace_vmbus_on_msg_dpc(hdr);
1108
1109         if (msgtype >= CHANNELMSG_COUNT) {
1110                 WARN_ONCE(1, "unknown msgtype=%d\n", msgtype);
1111                 goto msg_handled;
1112         }
1113
1114         payload_size = msg_copy.header.payload_size;
1115         if (payload_size > HV_MESSAGE_PAYLOAD_BYTE_COUNT) {
1116                 WARN_ONCE(1, "payload size is too large (%d)\n", payload_size);
1117                 goto msg_handled;
1118         }
1119
1120         entry = &channel_message_table[msgtype];
1121
1122         if (!entry->message_handler)
1123                 goto msg_handled;
1124
1125         if (payload_size < entry->min_payload_len) {
1126                 WARN_ONCE(1, "message too short: msgtype=%d len=%d\n", msgtype, payload_size);
1127                 goto msg_handled;
1128         }
1129
1130         if (entry->handler_type == VMHT_BLOCKING) {
1131                 ctx = kmalloc(sizeof(*ctx) + payload_size, GFP_ATOMIC);
1132                 if (ctx == NULL)
1133                         return;
1134
1135                 INIT_WORK(&ctx->work, vmbus_onmessage_work);
1136                 memcpy(&ctx->msg, &msg_copy, sizeof(msg->header) + payload_size);
1137
1138                 /*
1139                  * The host can generate a rescind message while we
1140                  * may still be handling the original offer. We deal with
1141                  * this condition by relying on the synchronization provided
1142                  * by offer_in_progress and by channel_mutex.  See also the
1143                  * inline comments in vmbus_onoffer_rescind().
1144                  */
1145                 switch (msgtype) {
1146                 case CHANNELMSG_RESCIND_CHANNELOFFER:
1147                         /*
1148                          * If we are handling the rescind message;
1149                          * schedule the work on the global work queue.
1150                          *
1151                          * The OFFER message and the RESCIND message should
1152                          * not be handled by the same serialized work queue,
1153                          * because the OFFER handler may call vmbus_open(),
1154                          * which tries to open the channel by sending an
1155                          * OPEN_CHANNEL message to the host and waits for
1156                          * the host's response; however, if the host has
1157                          * rescinded the channel before it receives the
1158                          * OPEN_CHANNEL message, the host just silently
1159                          * ignores the OPEN_CHANNEL message; as a result,
1160                          * the guest's OFFER handler hangs for ever, if we
1161                          * handle the RESCIND message in the same serialized
1162                          * work queue: the RESCIND handler can not start to
1163                          * run before the OFFER handler finishes.
1164                          */
1165                         schedule_work(&ctx->work);
1166                         break;
1167
1168                 case CHANNELMSG_OFFERCHANNEL:
1169                         /*
1170                          * The host sends the offer message of a given channel
1171                          * before sending the rescind message of the same
1172                          * channel.  These messages are sent to the guest's
1173                          * connect CPU; the guest then starts processing them
1174                          * in the tasklet handler on this CPU:
1175                          *
1176                          * VMBUS_CONNECT_CPU
1177                          *
1178                          * [vmbus_on_msg_dpc()]
1179                          * atomic_inc()  // CHANNELMSG_OFFERCHANNEL
1180                          * queue_work()
1181                          * ...
1182                          * [vmbus_on_msg_dpc()]
1183                          * schedule_work()  // CHANNELMSG_RESCIND_CHANNELOFFER
1184                          *
1185                          * We rely on the memory-ordering properties of the
1186                          * queue_work() and schedule_work() primitives, which
1187                          * guarantee that the atomic increment will be visible
1188                          * to the CPUs which will execute the offer & rescind
1189                          * works by the time these works will start execution.
1190                          */
1191                         atomic_inc(&vmbus_connection.offer_in_progress);
1192                         fallthrough;
1193
1194                 default:
1195                         queue_work(vmbus_connection.work_queue, &ctx->work);
1196                 }
1197         } else
1198                 entry->message_handler(hdr);
1199
1200 msg_handled:
1201         vmbus_signal_eom(msg, message_type);
1202 }
1203
1204 #ifdef CONFIG_PM_SLEEP
1205 /*
1206  * Fake RESCIND_CHANNEL messages to clean up hv_sock channels by force for
1207  * hibernation, because hv_sock connections can not persist across hibernation.
1208  */
1209 static void vmbus_force_channel_rescinded(struct vmbus_channel *channel)
1210 {
1211         struct onmessage_work_context *ctx;
1212         struct vmbus_channel_rescind_offer *rescind;
1213
1214         WARN_ON(!is_hvsock_channel(channel));
1215
1216         /*
1217          * Allocation size is small and the allocation should really not fail,
1218          * otherwise the state of the hv_sock connections ends up in limbo.
1219          */
1220         ctx = kzalloc(sizeof(*ctx) + sizeof(*rescind),
1221                       GFP_KERNEL | __GFP_NOFAIL);
1222
1223         /*
1224          * So far, these are not really used by Linux. Just set them to the
1225          * reasonable values conforming to the definitions of the fields.
1226          */
1227         ctx->msg.header.message_type = 1;
1228         ctx->msg.header.payload_size = sizeof(*rescind);
1229
1230         /* These values are actually used by Linux. */
1231         rescind = (struct vmbus_channel_rescind_offer *)ctx->msg.payload;
1232         rescind->header.msgtype = CHANNELMSG_RESCIND_CHANNELOFFER;
1233         rescind->child_relid = channel->offermsg.child_relid;
1234
1235         INIT_WORK(&ctx->work, vmbus_onmessage_work);
1236
1237         queue_work(vmbus_connection.work_queue, &ctx->work);
1238 }
1239 #endif /* CONFIG_PM_SLEEP */
1240
1241 /*
1242  * Schedule all channels with events pending
1243  */
1244 static void vmbus_chan_sched(struct hv_per_cpu_context *hv_cpu)
1245 {
1246         unsigned long *recv_int_page;
1247         u32 maxbits, relid;
1248
1249         if (vmbus_proto_version < VERSION_WIN8) {
1250                 maxbits = MAX_NUM_CHANNELS_SUPPORTED;
1251                 recv_int_page = vmbus_connection.recv_int_page;
1252         } else {
1253                 /*
1254                  * When the host is win8 and beyond, the event page
1255                  * can be directly checked to get the id of the channel
1256                  * that has the interrupt pending.
1257                  */
1258                 void *page_addr = hv_cpu->synic_event_page;
1259                 union hv_synic_event_flags *event
1260                         = (union hv_synic_event_flags *)page_addr +
1261                                                  VMBUS_MESSAGE_SINT;
1262
1263                 maxbits = HV_EVENT_FLAGS_COUNT;
1264                 recv_int_page = event->flags;
1265         }
1266
1267         if (unlikely(!recv_int_page))
1268                 return;
1269
1270         for_each_set_bit(relid, recv_int_page, maxbits) {
1271                 void (*callback_fn)(void *context);
1272                 struct vmbus_channel *channel;
1273
1274                 if (!sync_test_and_clear_bit(relid, recv_int_page))
1275                         continue;
1276
1277                 /* Special case - vmbus channel protocol msg */
1278                 if (relid == 0)
1279                         continue;
1280
1281                 /*
1282                  * Pairs with the kfree_rcu() in vmbus_chan_release().
1283                  * Guarantees that the channel data structure doesn't
1284                  * get freed while the channel pointer below is being
1285                  * dereferenced.
1286                  */
1287                 rcu_read_lock();
1288
1289                 /* Find channel based on relid */
1290                 channel = relid2channel(relid);
1291                 if (channel == NULL)
1292                         goto sched_unlock_rcu;
1293
1294                 if (channel->rescind)
1295                         goto sched_unlock_rcu;
1296
1297                 /*
1298                  * Make sure that the ring buffer data structure doesn't get
1299                  * freed while we dereference the ring buffer pointer.  Test
1300                  * for the channel's onchannel_callback being NULL within a
1301                  * sched_lock critical section.  See also the inline comments
1302                  * in vmbus_reset_channel_cb().
1303                  */
1304                 spin_lock(&channel->sched_lock);
1305
1306                 callback_fn = channel->onchannel_callback;
1307                 if (unlikely(callback_fn == NULL))
1308                         goto sched_unlock;
1309
1310                 trace_vmbus_chan_sched(channel);
1311
1312                 ++channel->interrupts;
1313
1314                 switch (channel->callback_mode) {
1315                 case HV_CALL_ISR:
1316                         (*callback_fn)(channel->channel_callback_context);
1317                         break;
1318
1319                 case HV_CALL_BATCHED:
1320                         hv_begin_read(&channel->inbound);
1321                         fallthrough;
1322                 case HV_CALL_DIRECT:
1323                         tasklet_schedule(&channel->callback_event);
1324                 }
1325
1326 sched_unlock:
1327                 spin_unlock(&channel->sched_lock);
1328 sched_unlock_rcu:
1329                 rcu_read_unlock();
1330         }
1331 }
1332
1333 static void vmbus_isr(void)
1334 {
1335         struct hv_per_cpu_context *hv_cpu
1336                 = this_cpu_ptr(hv_context.cpu_context);
1337         void *page_addr = hv_cpu->synic_event_page;
1338         struct hv_message *msg;
1339         union hv_synic_event_flags *event;
1340         bool handled = false;
1341
1342         if (unlikely(page_addr == NULL))
1343                 return;
1344
1345         event = (union hv_synic_event_flags *)page_addr +
1346                                          VMBUS_MESSAGE_SINT;
1347         /*
1348          * Check for events before checking for messages. This is the order
1349          * in which events and messages are checked in Windows guests on
1350          * Hyper-V, and the Windows team suggested we do the same.
1351          */
1352
1353         if ((vmbus_proto_version == VERSION_WS2008) ||
1354                 (vmbus_proto_version == VERSION_WIN7)) {
1355
1356                 /* Since we are a child, we only need to check bit 0 */
1357                 if (sync_test_and_clear_bit(0, event->flags))
1358                         handled = true;
1359         } else {
1360                 /*
1361                  * Our host is win8 or above. The signaling mechanism
1362                  * has changed and we can directly look at the event page.
1363                  * If bit n is set then we have an interrup on the channel
1364                  * whose id is n.
1365                  */
1366                 handled = true;
1367         }
1368
1369         if (handled)
1370                 vmbus_chan_sched(hv_cpu);
1371
1372         page_addr = hv_cpu->synic_message_page;
1373         msg = (struct hv_message *)page_addr + VMBUS_MESSAGE_SINT;
1374
1375         /* Check if there are actual msgs to be processed */
1376         if (msg->header.message_type != HVMSG_NONE) {
1377                 if (msg->header.message_type == HVMSG_TIMER_EXPIRED) {
1378                         hv_stimer0_isr();
1379                         vmbus_signal_eom(msg, HVMSG_TIMER_EXPIRED);
1380                 } else
1381                         tasklet_schedule(&hv_cpu->msg_dpc);
1382         }
1383
1384         add_interrupt_randomness(vmbus_interrupt);
1385 }
1386
1387 static irqreturn_t vmbus_percpu_isr(int irq, void *dev_id)
1388 {
1389         vmbus_isr();
1390         return IRQ_HANDLED;
1391 }
1392
1393 /*
1394  * Callback from kmsg_dump. Grab as much as possible from the end of the kmsg
1395  * buffer and call into Hyper-V to transfer the data.
1396  */
1397 static void hv_kmsg_dump(struct kmsg_dumper *dumper,
1398                          enum kmsg_dump_reason reason)
1399 {
1400         struct kmsg_dump_iter iter;
1401         size_t bytes_written;
1402
1403         /* We are only interested in panics. */
1404         if ((reason != KMSG_DUMP_PANIC) || (!sysctl_record_panic_msg))
1405                 return;
1406
1407         /*
1408          * Write dump contents to the page. No need to synchronize; panic should
1409          * be single-threaded.
1410          */
1411         kmsg_dump_rewind(&iter);
1412         kmsg_dump_get_buffer(&iter, false, hv_panic_page, HV_HYP_PAGE_SIZE,
1413                              &bytes_written);
1414         if (!bytes_written)
1415                 return;
1416         /*
1417          * P3 to contain the physical address of the panic page & P4 to
1418          * contain the size of the panic data in that page. Rest of the
1419          * registers are no-op when the NOTIFY_MSG flag is set.
1420          */
1421         hv_set_register(HV_REGISTER_CRASH_P0, 0);
1422         hv_set_register(HV_REGISTER_CRASH_P1, 0);
1423         hv_set_register(HV_REGISTER_CRASH_P2, 0);
1424         hv_set_register(HV_REGISTER_CRASH_P3, virt_to_phys(hv_panic_page));
1425         hv_set_register(HV_REGISTER_CRASH_P4, bytes_written);
1426
1427         /*
1428          * Let Hyper-V know there is crash data available along with
1429          * the panic message.
1430          */
1431         hv_set_register(HV_REGISTER_CRASH_CTL,
1432                (HV_CRASH_CTL_CRASH_NOTIFY | HV_CRASH_CTL_CRASH_NOTIFY_MSG));
1433 }
1434
1435 static struct kmsg_dumper hv_kmsg_dumper = {
1436         .dump = hv_kmsg_dump,
1437 };
1438
1439 static void hv_kmsg_dump_register(void)
1440 {
1441         int ret;
1442
1443         hv_panic_page = hv_alloc_hyperv_zeroed_page();
1444         if (!hv_panic_page) {
1445                 pr_err("Hyper-V: panic message page memory allocation failed\n");
1446                 return;
1447         }
1448
1449         ret = kmsg_dump_register(&hv_kmsg_dumper);
1450         if (ret) {
1451                 pr_err("Hyper-V: kmsg dump register error 0x%x\n", ret);
1452                 hv_free_hyperv_page((unsigned long)hv_panic_page);
1453                 hv_panic_page = NULL;
1454         }
1455 }
1456
1457 static struct ctl_table_header *hv_ctl_table_hdr;
1458
1459 /*
1460  * sysctl option to allow the user to control whether kmsg data should be
1461  * reported to Hyper-V on panic.
1462  */
1463 static struct ctl_table hv_ctl_table[] = {
1464         {
1465                 .procname       = "hyperv_record_panic_msg",
1466                 .data           = &sysctl_record_panic_msg,
1467                 .maxlen         = sizeof(int),
1468                 .mode           = 0644,
1469                 .proc_handler   = proc_dointvec_minmax,
1470                 .extra1         = SYSCTL_ZERO,
1471                 .extra2         = SYSCTL_ONE
1472         },
1473         {}
1474 };
1475
1476 static struct ctl_table hv_root_table[] = {
1477         {
1478                 .procname       = "kernel",
1479                 .mode           = 0555,
1480                 .child          = hv_ctl_table
1481         },
1482         {}
1483 };
1484
1485 /*
1486  * vmbus_bus_init -Main vmbus driver initialization routine.
1487  *
1488  * Here, we
1489  *      - initialize the vmbus driver context
1490  *      - invoke the vmbus hv main init routine
1491  *      - retrieve the channel offers
1492  */
1493 static int vmbus_bus_init(void)
1494 {
1495         int ret;
1496
1497         ret = hv_init();
1498         if (ret != 0) {
1499                 pr_err("Unable to initialize the hypervisor - 0x%x\n", ret);
1500                 return ret;
1501         }
1502
1503         ret = bus_register(&hv_bus);
1504         if (ret)
1505                 return ret;
1506
1507         /*
1508          * VMbus interrupts are best modeled as per-cpu interrupts. If
1509          * on an architecture with support for per-cpu IRQs (e.g. ARM64),
1510          * allocate a per-cpu IRQ using standard Linux kernel functionality.
1511          * If not on such an architecture (e.g., x86/x64), then rely on
1512          * code in the arch-specific portion of the code tree to connect
1513          * the VMbus interrupt handler.
1514          */
1515
1516         if (vmbus_irq == -1) {
1517                 hv_setup_vmbus_handler(vmbus_isr);
1518         } else {
1519                 vmbus_evt = alloc_percpu(long);
1520                 ret = request_percpu_irq(vmbus_irq, vmbus_percpu_isr,
1521                                 "Hyper-V VMbus", vmbus_evt);
1522                 if (ret) {
1523                         pr_err("Can't request Hyper-V VMbus IRQ %d, Err %d",
1524                                         vmbus_irq, ret);
1525                         free_percpu(vmbus_evt);
1526                         goto err_setup;
1527                 }
1528         }
1529
1530         ret = hv_synic_alloc();
1531         if (ret)
1532                 goto err_alloc;
1533
1534         /*
1535          * Initialize the per-cpu interrupt state and stimer state.
1536          * Then connect to the host.
1537          */
1538         ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "hyperv/vmbus:online",
1539                                 hv_synic_init, hv_synic_cleanup);
1540         if (ret < 0)
1541                 goto err_cpuhp;
1542         hyperv_cpuhp_online = ret;
1543
1544         ret = vmbus_connect();
1545         if (ret)
1546                 goto err_connect;
1547
1548         if (hv_is_isolation_supported())
1549                 sysctl_record_panic_msg = 0;
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                  * Panic message recording (sysctl_record_panic_msg)
1558                  * is enabled by default in non-isolated guests and
1559                  * disabled by default in isolated guests; the panic
1560                  * message recording won't be available in isolated
1561                  * guests should the following registration fail.
1562                  */
1563                 hv_ctl_table_hdr = register_sysctl_table(hv_root_table);
1564                 if (!hv_ctl_table_hdr)
1565                         pr_err("Hyper-V: sysctl table register error");
1566
1567                 /*
1568                  * Register for panic kmsg callback only if the right
1569                  * capability is supported by the hypervisor.
1570                  */
1571                 hyperv_crash_ctl = hv_get_register(HV_REGISTER_CRASH_CTL);
1572                 if (hyperv_crash_ctl & HV_CRASH_CTL_CRASH_NOTIFY_MSG)
1573                         hv_kmsg_dump_register();
1574
1575                 register_die_notifier(&hyperv_die_block);
1576         }
1577
1578         /*
1579          * Always register the panic notifier because we need to unload
1580          * the VMbus channel connection to prevent any VMbus
1581          * activity after the VM panics.
1582          */
1583         atomic_notifier_chain_register(&panic_notifier_list,
1584                                &hyperv_panic_block);
1585
1586         vmbus_request_offers();
1587
1588         return 0;
1589
1590 err_connect:
1591         cpuhp_remove_state(hyperv_cpuhp_online);
1592 err_cpuhp:
1593         hv_synic_free();
1594 err_alloc:
1595         if (vmbus_irq == -1) {
1596                 hv_remove_vmbus_handler();
1597         } else {
1598                 free_percpu_irq(vmbus_irq, vmbus_evt);
1599                 free_percpu(vmbus_evt);
1600         }
1601 err_setup:
1602         bus_unregister(&hv_bus);
1603         unregister_sysctl_table(hv_ctl_table_hdr);
1604         hv_ctl_table_hdr = NULL;
1605         return ret;
1606 }
1607
1608 /**
1609  * __vmbus_child_driver_register() - Register a vmbus's driver
1610  * @hv_driver: Pointer to driver structure you want to register
1611  * @owner: owner module of the drv
1612  * @mod_name: module name string
1613  *
1614  * Registers the given driver with Linux through the 'driver_register()' call
1615  * and sets up the hyper-v vmbus handling for this driver.
1616  * It will return the state of the 'driver_register()' call.
1617  *
1618  */
1619 int __vmbus_driver_register(struct hv_driver *hv_driver, struct module *owner, const char *mod_name)
1620 {
1621         int ret;
1622
1623         pr_info("registering driver %s\n", hv_driver->name);
1624
1625         ret = vmbus_exists();
1626         if (ret < 0)
1627                 return ret;
1628
1629         hv_driver->driver.name = hv_driver->name;
1630         hv_driver->driver.owner = owner;
1631         hv_driver->driver.mod_name = mod_name;
1632         hv_driver->driver.bus = &hv_bus;
1633
1634         spin_lock_init(&hv_driver->dynids.lock);
1635         INIT_LIST_HEAD(&hv_driver->dynids.list);
1636
1637         ret = driver_register(&hv_driver->driver);
1638
1639         return ret;
1640 }
1641 EXPORT_SYMBOL_GPL(__vmbus_driver_register);
1642
1643 /**
1644  * vmbus_driver_unregister() - Unregister a vmbus's driver
1645  * @hv_driver: Pointer to driver structure you want to
1646  *             un-register
1647  *
1648  * Un-register the given driver that was previous registered with a call to
1649  * vmbus_driver_register()
1650  */
1651 void vmbus_driver_unregister(struct hv_driver *hv_driver)
1652 {
1653         pr_info("unregistering driver %s\n", hv_driver->name);
1654
1655         if (!vmbus_exists()) {
1656                 driver_unregister(&hv_driver->driver);
1657                 vmbus_free_dynids(hv_driver);
1658         }
1659 }
1660 EXPORT_SYMBOL_GPL(vmbus_driver_unregister);
1661
1662
1663 /*
1664  * Called when last reference to channel is gone.
1665  */
1666 static void vmbus_chan_release(struct kobject *kobj)
1667 {
1668         struct vmbus_channel *channel
1669                 = container_of(kobj, struct vmbus_channel, kobj);
1670
1671         kfree_rcu(channel, rcu);
1672 }
1673
1674 struct vmbus_chan_attribute {
1675         struct attribute attr;
1676         ssize_t (*show)(struct vmbus_channel *chan, char *buf);
1677         ssize_t (*store)(struct vmbus_channel *chan,
1678                          const char *buf, size_t count);
1679 };
1680 #define VMBUS_CHAN_ATTR(_name, _mode, _show, _store) \
1681         struct vmbus_chan_attribute chan_attr_##_name \
1682                 = __ATTR(_name, _mode, _show, _store)
1683 #define VMBUS_CHAN_ATTR_RW(_name) \
1684         struct vmbus_chan_attribute chan_attr_##_name = __ATTR_RW(_name)
1685 #define VMBUS_CHAN_ATTR_RO(_name) \
1686         struct vmbus_chan_attribute chan_attr_##_name = __ATTR_RO(_name)
1687 #define VMBUS_CHAN_ATTR_WO(_name) \
1688         struct vmbus_chan_attribute chan_attr_##_name = __ATTR_WO(_name)
1689
1690 static ssize_t vmbus_chan_attr_show(struct kobject *kobj,
1691                                     struct attribute *attr, char *buf)
1692 {
1693         const struct vmbus_chan_attribute *attribute
1694                 = container_of(attr, struct vmbus_chan_attribute, attr);
1695         struct vmbus_channel *chan
1696                 = container_of(kobj, struct vmbus_channel, kobj);
1697
1698         if (!attribute->show)
1699                 return -EIO;
1700
1701         return attribute->show(chan, buf);
1702 }
1703
1704 static ssize_t vmbus_chan_attr_store(struct kobject *kobj,
1705                                      struct attribute *attr, const char *buf,
1706                                      size_t count)
1707 {
1708         const struct vmbus_chan_attribute *attribute
1709                 = container_of(attr, struct vmbus_chan_attribute, attr);
1710         struct vmbus_channel *chan
1711                 = container_of(kobj, struct vmbus_channel, kobj);
1712
1713         if (!attribute->store)
1714                 return -EIO;
1715
1716         return attribute->store(chan, buf, count);
1717 }
1718
1719 static const struct sysfs_ops vmbus_chan_sysfs_ops = {
1720         .show = vmbus_chan_attr_show,
1721         .store = vmbus_chan_attr_store,
1722 };
1723
1724 static ssize_t out_mask_show(struct vmbus_channel *channel, char *buf)
1725 {
1726         struct hv_ring_buffer_info *rbi = &channel->outbound;
1727         ssize_t ret;
1728
1729         mutex_lock(&rbi->ring_buffer_mutex);
1730         if (!rbi->ring_buffer) {
1731                 mutex_unlock(&rbi->ring_buffer_mutex);
1732                 return -EINVAL;
1733         }
1734
1735         ret = sprintf(buf, "%u\n", rbi->ring_buffer->interrupt_mask);
1736         mutex_unlock(&rbi->ring_buffer_mutex);
1737         return ret;
1738 }
1739 static VMBUS_CHAN_ATTR_RO(out_mask);
1740
1741 static ssize_t in_mask_show(struct vmbus_channel *channel, char *buf)
1742 {
1743         struct hv_ring_buffer_info *rbi = &channel->inbound;
1744         ssize_t ret;
1745
1746         mutex_lock(&rbi->ring_buffer_mutex);
1747         if (!rbi->ring_buffer) {
1748                 mutex_unlock(&rbi->ring_buffer_mutex);
1749                 return -EINVAL;
1750         }
1751
1752         ret = sprintf(buf, "%u\n", rbi->ring_buffer->interrupt_mask);
1753         mutex_unlock(&rbi->ring_buffer_mutex);
1754         return ret;
1755 }
1756 static VMBUS_CHAN_ATTR_RO(in_mask);
1757
1758 static ssize_t read_avail_show(struct vmbus_channel *channel, char *buf)
1759 {
1760         struct hv_ring_buffer_info *rbi = &channel->inbound;
1761         ssize_t ret;
1762
1763         mutex_lock(&rbi->ring_buffer_mutex);
1764         if (!rbi->ring_buffer) {
1765                 mutex_unlock(&rbi->ring_buffer_mutex);
1766                 return -EINVAL;
1767         }
1768
1769         ret = sprintf(buf, "%u\n", hv_get_bytes_to_read(rbi));
1770         mutex_unlock(&rbi->ring_buffer_mutex);
1771         return ret;
1772 }
1773 static VMBUS_CHAN_ATTR_RO(read_avail);
1774
1775 static ssize_t write_avail_show(struct vmbus_channel *channel, char *buf)
1776 {
1777         struct hv_ring_buffer_info *rbi = &channel->outbound;
1778         ssize_t ret;
1779
1780         mutex_lock(&rbi->ring_buffer_mutex);
1781         if (!rbi->ring_buffer) {
1782                 mutex_unlock(&rbi->ring_buffer_mutex);
1783                 return -EINVAL;
1784         }
1785
1786         ret = sprintf(buf, "%u\n", hv_get_bytes_to_write(rbi));
1787         mutex_unlock(&rbi->ring_buffer_mutex);
1788         return ret;
1789 }
1790 static VMBUS_CHAN_ATTR_RO(write_avail);
1791
1792 static ssize_t target_cpu_show(struct vmbus_channel *channel, char *buf)
1793 {
1794         return sprintf(buf, "%u\n", channel->target_cpu);
1795 }
1796 static ssize_t target_cpu_store(struct vmbus_channel *channel,
1797                                 const char *buf, size_t count)
1798 {
1799         u32 target_cpu, origin_cpu;
1800         ssize_t ret = count;
1801
1802         if (vmbus_proto_version < VERSION_WIN10_V4_1)
1803                 return -EIO;
1804
1805         if (sscanf(buf, "%uu", &target_cpu) != 1)
1806                 return -EIO;
1807
1808         /* Validate target_cpu for the cpumask_test_cpu() operation below. */
1809         if (target_cpu >= nr_cpumask_bits)
1810                 return -EINVAL;
1811
1812         /* No CPUs should come up or down during this. */
1813         cpus_read_lock();
1814
1815         if (!cpu_online(target_cpu)) {
1816                 cpus_read_unlock();
1817                 return -EINVAL;
1818         }
1819
1820         /*
1821          * Synchronizes target_cpu_store() and channel closure:
1822          *
1823          * { Initially: state = CHANNEL_OPENED }
1824          *
1825          * CPU1                         CPU2
1826          *
1827          * [target_cpu_store()]         [vmbus_disconnect_ring()]
1828          *
1829          * LOCK channel_mutex           LOCK channel_mutex
1830          * LOAD r1 = state              LOAD r2 = state
1831          * IF (r1 == CHANNEL_OPENED)    IF (r2 == CHANNEL_OPENED)
1832          *   SEND MODIFYCHANNEL           STORE state = CHANNEL_OPEN
1833          *   [...]                        SEND CLOSECHANNEL
1834          * UNLOCK channel_mutex         UNLOCK channel_mutex
1835          *
1836          * Forbids: r1 == r2 == CHANNEL_OPENED (i.e., CPU1's LOCK precedes
1837          *              CPU2's LOCK) && CPU2's SEND precedes CPU1's SEND
1838          *
1839          * Note.  The host processes the channel messages "sequentially", in
1840          * the order in which they are received on a per-partition basis.
1841          */
1842         mutex_lock(&vmbus_connection.channel_mutex);
1843
1844         /*
1845          * Hyper-V will ignore MODIFYCHANNEL messages for "non-open" channels;
1846          * avoid sending the message and fail here for such channels.
1847          */
1848         if (channel->state != CHANNEL_OPENED_STATE) {
1849                 ret = -EIO;
1850                 goto cpu_store_unlock;
1851         }
1852
1853         origin_cpu = channel->target_cpu;
1854         if (target_cpu == origin_cpu)
1855                 goto cpu_store_unlock;
1856
1857         if (vmbus_send_modifychannel(channel,
1858                                      hv_cpu_number_to_vp_number(target_cpu))) {
1859                 ret = -EIO;
1860                 goto cpu_store_unlock;
1861         }
1862
1863         /*
1864          * For version before VERSION_WIN10_V5_3, the following warning holds:
1865          *
1866          * Warning.  At this point, there is *no* guarantee that the host will
1867          * have successfully processed the vmbus_send_modifychannel() request.
1868          * See the header comment of vmbus_send_modifychannel() for more info.
1869          *
1870          * Lags in the processing of the above vmbus_send_modifychannel() can
1871          * result in missed interrupts if the "old" target CPU is taken offline
1872          * before Hyper-V starts sending interrupts to the "new" target CPU.
1873          * But apart from this offlining scenario, the code tolerates such
1874          * lags.  It will function correctly even if a channel interrupt comes
1875          * in on a CPU that is different from the channel target_cpu value.
1876          */
1877
1878         channel->target_cpu = target_cpu;
1879
1880         /* See init_vp_index(). */
1881         if (hv_is_perf_channel(channel))
1882                 hv_update_alloced_cpus(origin_cpu, target_cpu);
1883
1884         /* Currently set only for storvsc channels. */
1885         if (channel->change_target_cpu_callback) {
1886                 (*channel->change_target_cpu_callback)(channel,
1887                                 origin_cpu, target_cpu);
1888         }
1889
1890 cpu_store_unlock:
1891         mutex_unlock(&vmbus_connection.channel_mutex);
1892         cpus_read_unlock();
1893         return ret;
1894 }
1895 static VMBUS_CHAN_ATTR(cpu, 0644, target_cpu_show, target_cpu_store);
1896
1897 static ssize_t channel_pending_show(struct vmbus_channel *channel,
1898                                     char *buf)
1899 {
1900         return sprintf(buf, "%d\n",
1901                        channel_pending(channel,
1902                                        vmbus_connection.monitor_pages[1]));
1903 }
1904 static VMBUS_CHAN_ATTR(pending, 0444, channel_pending_show, NULL);
1905
1906 static ssize_t channel_latency_show(struct vmbus_channel *channel,
1907                                     char *buf)
1908 {
1909         return sprintf(buf, "%d\n",
1910                        channel_latency(channel,
1911                                        vmbus_connection.monitor_pages[1]));
1912 }
1913 static VMBUS_CHAN_ATTR(latency, 0444, channel_latency_show, NULL);
1914
1915 static ssize_t channel_interrupts_show(struct vmbus_channel *channel, char *buf)
1916 {
1917         return sprintf(buf, "%llu\n", channel->interrupts);
1918 }
1919 static VMBUS_CHAN_ATTR(interrupts, 0444, channel_interrupts_show, NULL);
1920
1921 static ssize_t channel_events_show(struct vmbus_channel *channel, char *buf)
1922 {
1923         return sprintf(buf, "%llu\n", channel->sig_events);
1924 }
1925 static VMBUS_CHAN_ATTR(events, 0444, channel_events_show, NULL);
1926
1927 static ssize_t channel_intr_in_full_show(struct vmbus_channel *channel,
1928                                          char *buf)
1929 {
1930         return sprintf(buf, "%llu\n",
1931                        (unsigned long long)channel->intr_in_full);
1932 }
1933 static VMBUS_CHAN_ATTR(intr_in_full, 0444, channel_intr_in_full_show, NULL);
1934
1935 static ssize_t channel_intr_out_empty_show(struct vmbus_channel *channel,
1936                                            char *buf)
1937 {
1938         return sprintf(buf, "%llu\n",
1939                        (unsigned long long)channel->intr_out_empty);
1940 }
1941 static VMBUS_CHAN_ATTR(intr_out_empty, 0444, channel_intr_out_empty_show, NULL);
1942
1943 static ssize_t channel_out_full_first_show(struct vmbus_channel *channel,
1944                                            char *buf)
1945 {
1946         return sprintf(buf, "%llu\n",
1947                        (unsigned long long)channel->out_full_first);
1948 }
1949 static VMBUS_CHAN_ATTR(out_full_first, 0444, channel_out_full_first_show, NULL);
1950
1951 static ssize_t channel_out_full_total_show(struct vmbus_channel *channel,
1952                                            char *buf)
1953 {
1954         return sprintf(buf, "%llu\n",
1955                        (unsigned long long)channel->out_full_total);
1956 }
1957 static VMBUS_CHAN_ATTR(out_full_total, 0444, channel_out_full_total_show, NULL);
1958
1959 static ssize_t subchannel_monitor_id_show(struct vmbus_channel *channel,
1960                                           char *buf)
1961 {
1962         return sprintf(buf, "%u\n", channel->offermsg.monitorid);
1963 }
1964 static VMBUS_CHAN_ATTR(monitor_id, 0444, subchannel_monitor_id_show, NULL);
1965
1966 static ssize_t subchannel_id_show(struct vmbus_channel *channel,
1967                                   char *buf)
1968 {
1969         return sprintf(buf, "%u\n",
1970                        channel->offermsg.offer.sub_channel_index);
1971 }
1972 static VMBUS_CHAN_ATTR_RO(subchannel_id);
1973
1974 static struct attribute *vmbus_chan_attrs[] = {
1975         &chan_attr_out_mask.attr,
1976         &chan_attr_in_mask.attr,
1977         &chan_attr_read_avail.attr,
1978         &chan_attr_write_avail.attr,
1979         &chan_attr_cpu.attr,
1980         &chan_attr_pending.attr,
1981         &chan_attr_latency.attr,
1982         &chan_attr_interrupts.attr,
1983         &chan_attr_events.attr,
1984         &chan_attr_intr_in_full.attr,
1985         &chan_attr_intr_out_empty.attr,
1986         &chan_attr_out_full_first.attr,
1987         &chan_attr_out_full_total.attr,
1988         &chan_attr_monitor_id.attr,
1989         &chan_attr_subchannel_id.attr,
1990         NULL
1991 };
1992
1993 /*
1994  * Channel-level attribute_group callback function. Returns the permission for
1995  * each attribute, and returns 0 if an attribute is not visible.
1996  */
1997 static umode_t vmbus_chan_attr_is_visible(struct kobject *kobj,
1998                                           struct attribute *attr, int idx)
1999 {
2000         const struct vmbus_channel *channel =
2001                 container_of(kobj, struct vmbus_channel, kobj);
2002
2003         /* Hide the monitor attributes if the monitor mechanism is not used. */
2004         if (!channel->offermsg.monitor_allocated &&
2005             (attr == &chan_attr_pending.attr ||
2006              attr == &chan_attr_latency.attr ||
2007              attr == &chan_attr_monitor_id.attr))
2008                 return 0;
2009
2010         return attr->mode;
2011 }
2012
2013 static struct attribute_group vmbus_chan_group = {
2014         .attrs = vmbus_chan_attrs,
2015         .is_visible = vmbus_chan_attr_is_visible
2016 };
2017
2018 static struct kobj_type vmbus_chan_ktype = {
2019         .sysfs_ops = &vmbus_chan_sysfs_ops,
2020         .release = vmbus_chan_release,
2021 };
2022
2023 /*
2024  * vmbus_add_channel_kobj - setup a sub-directory under device/channels
2025  */
2026 int vmbus_add_channel_kobj(struct hv_device *dev, struct vmbus_channel *channel)
2027 {
2028         const struct device *device = &dev->device;
2029         struct kobject *kobj = &channel->kobj;
2030         u32 relid = channel->offermsg.child_relid;
2031         int ret;
2032
2033         kobj->kset = dev->channels_kset;
2034         ret = kobject_init_and_add(kobj, &vmbus_chan_ktype, NULL,
2035                                    "%u", relid);
2036         if (ret) {
2037                 kobject_put(kobj);
2038                 return ret;
2039         }
2040
2041         ret = sysfs_create_group(kobj, &vmbus_chan_group);
2042
2043         if (ret) {
2044                 /*
2045                  * The calling functions' error handling paths will cleanup the
2046                  * empty channel directory.
2047                  */
2048                 kobject_put(kobj);
2049                 dev_err(device, "Unable to set up channel sysfs files\n");
2050                 return ret;
2051         }
2052
2053         kobject_uevent(kobj, KOBJ_ADD);
2054
2055         return 0;
2056 }
2057
2058 /*
2059  * vmbus_remove_channel_attr_group - remove the channel's attribute group
2060  */
2061 void vmbus_remove_channel_attr_group(struct vmbus_channel *channel)
2062 {
2063         sysfs_remove_group(&channel->kobj, &vmbus_chan_group);
2064 }
2065
2066 /*
2067  * vmbus_device_create - Creates and registers a new child device
2068  * on the vmbus.
2069  */
2070 struct hv_device *vmbus_device_create(const guid_t *type,
2071                                       const guid_t *instance,
2072                                       struct vmbus_channel *channel)
2073 {
2074         struct hv_device *child_device_obj;
2075
2076         child_device_obj = kzalloc(sizeof(struct hv_device), GFP_KERNEL);
2077         if (!child_device_obj) {
2078                 pr_err("Unable to allocate device object for child device\n");
2079                 return NULL;
2080         }
2081
2082         child_device_obj->channel = channel;
2083         guid_copy(&child_device_obj->dev_type, type);
2084         guid_copy(&child_device_obj->dev_instance, instance);
2085         child_device_obj->vendor_id = 0x1414; /* MSFT vendor ID */
2086
2087         return child_device_obj;
2088 }
2089
2090 /*
2091  * vmbus_device_register - Register the child device
2092  */
2093 int vmbus_device_register(struct hv_device *child_device_obj)
2094 {
2095         struct kobject *kobj = &child_device_obj->device.kobj;
2096         int ret;
2097
2098         dev_set_name(&child_device_obj->device, "%pUl",
2099                      &child_device_obj->channel->offermsg.offer.if_instance);
2100
2101         child_device_obj->device.bus = &hv_bus;
2102         child_device_obj->device.parent = &hv_acpi_dev->dev;
2103         child_device_obj->device.release = vmbus_device_release;
2104
2105         /*
2106          * Register with the LDM. This will kick off the driver/device
2107          * binding...which will eventually call vmbus_match() and vmbus_probe()
2108          */
2109         ret = device_register(&child_device_obj->device);
2110         if (ret) {
2111                 pr_err("Unable to register child device\n");
2112                 return ret;
2113         }
2114
2115         child_device_obj->channels_kset = kset_create_and_add("channels",
2116                                                               NULL, kobj);
2117         if (!child_device_obj->channels_kset) {
2118                 ret = -ENOMEM;
2119                 goto err_dev_unregister;
2120         }
2121
2122         ret = vmbus_add_channel_kobj(child_device_obj,
2123                                      child_device_obj->channel);
2124         if (ret) {
2125                 pr_err("Unable to register primary channeln");
2126                 goto err_kset_unregister;
2127         }
2128         hv_debug_add_dev_dir(child_device_obj);
2129
2130         return 0;
2131
2132 err_kset_unregister:
2133         kset_unregister(child_device_obj->channels_kset);
2134
2135 err_dev_unregister:
2136         device_unregister(&child_device_obj->device);
2137         return ret;
2138 }
2139
2140 /*
2141  * vmbus_device_unregister - Remove the specified child device
2142  * from the vmbus.
2143  */
2144 void vmbus_device_unregister(struct hv_device *device_obj)
2145 {
2146         pr_debug("child device %s unregistered\n",
2147                 dev_name(&device_obj->device));
2148
2149         kset_unregister(device_obj->channels_kset);
2150
2151         /*
2152          * Kick off the process of unregistering the device.
2153          * This will call vmbus_remove() and eventually vmbus_device_release()
2154          */
2155         device_unregister(&device_obj->device);
2156 }
2157
2158
2159 /*
2160  * VMBUS is an acpi enumerated device. Get the information we
2161  * need from DSDT.
2162  */
2163 #define VTPM_BASE_ADDRESS 0xfed40000
2164 static acpi_status vmbus_walk_resources(struct acpi_resource *res, void *ctx)
2165 {
2166         resource_size_t start = 0;
2167         resource_size_t end = 0;
2168         struct resource *new_res;
2169         struct resource **old_res = &hyperv_mmio;
2170         struct resource **prev_res = NULL;
2171         struct resource r;
2172
2173         switch (res->type) {
2174
2175         /*
2176          * "Address" descriptors are for bus windows. Ignore
2177          * "memory" descriptors, which are for registers on
2178          * devices.
2179          */
2180         case ACPI_RESOURCE_TYPE_ADDRESS32:
2181                 start = res->data.address32.address.minimum;
2182                 end = res->data.address32.address.maximum;
2183                 break;
2184
2185         case ACPI_RESOURCE_TYPE_ADDRESS64:
2186                 start = res->data.address64.address.minimum;
2187                 end = res->data.address64.address.maximum;
2188                 break;
2189
2190         /*
2191          * The IRQ information is needed only on ARM64, which Hyper-V
2192          * sets up in the extended format. IRQ information is present
2193          * on x86/x64 in the non-extended format but it is not used by
2194          * Linux. So don't bother checking for the non-extended format.
2195          */
2196         case ACPI_RESOURCE_TYPE_EXTENDED_IRQ:
2197                 if (!acpi_dev_resource_interrupt(res, 0, &r)) {
2198                         pr_err("Unable to parse Hyper-V ACPI interrupt\n");
2199                         return AE_ERROR;
2200                 }
2201                 /* ARM64 INTID for VMbus */
2202                 vmbus_interrupt = res->data.extended_irq.interrupts[0];
2203                 /* Linux IRQ number */
2204                 vmbus_irq = r.start;
2205                 return AE_OK;
2206
2207         default:
2208                 /* Unused resource type */
2209                 return AE_OK;
2210
2211         }
2212         /*
2213          * Ignore ranges that are below 1MB, as they're not
2214          * necessary or useful here.
2215          */
2216         if (end < 0x100000)
2217                 return AE_OK;
2218
2219         new_res = kzalloc(sizeof(*new_res), GFP_ATOMIC);
2220         if (!new_res)
2221                 return AE_NO_MEMORY;
2222
2223         /* If this range overlaps the virtual TPM, truncate it. */
2224         if (end > VTPM_BASE_ADDRESS && start < VTPM_BASE_ADDRESS)
2225                 end = VTPM_BASE_ADDRESS;
2226
2227         new_res->name = "hyperv mmio";
2228         new_res->flags = IORESOURCE_MEM;
2229         new_res->start = start;
2230         new_res->end = end;
2231
2232         /*
2233          * If two ranges are adjacent, merge them.
2234          */
2235         do {
2236                 if (!*old_res) {
2237                         *old_res = new_res;
2238                         break;
2239                 }
2240
2241                 if (((*old_res)->end + 1) == new_res->start) {
2242                         (*old_res)->end = new_res->end;
2243                         kfree(new_res);
2244                         break;
2245                 }
2246
2247                 if ((*old_res)->start == new_res->end + 1) {
2248                         (*old_res)->start = new_res->start;
2249                         kfree(new_res);
2250                         break;
2251                 }
2252
2253                 if ((*old_res)->start > new_res->end) {
2254                         new_res->sibling = *old_res;
2255                         if (prev_res)
2256                                 (*prev_res)->sibling = new_res;
2257                         *old_res = new_res;
2258                         break;
2259                 }
2260
2261                 prev_res = old_res;
2262                 old_res = &(*old_res)->sibling;
2263
2264         } while (1);
2265
2266         return AE_OK;
2267 }
2268
2269 static int vmbus_acpi_remove(struct acpi_device *device)
2270 {
2271         struct resource *cur_res;
2272         struct resource *next_res;
2273
2274         if (hyperv_mmio) {
2275                 if (fb_mmio) {
2276                         __release_region(hyperv_mmio, fb_mmio->start,
2277                                          resource_size(fb_mmio));
2278                         fb_mmio = NULL;
2279                 }
2280
2281                 for (cur_res = hyperv_mmio; cur_res; cur_res = next_res) {
2282                         next_res = cur_res->sibling;
2283                         kfree(cur_res);
2284                 }
2285         }
2286
2287         return 0;
2288 }
2289
2290 static void vmbus_reserve_fb(void)
2291 {
2292         int size;
2293         /*
2294          * Make a claim for the frame buffer in the resource tree under the
2295          * first node, which will be the one below 4GB.  The length seems to
2296          * be underreported, particularly in a Generation 1 VM.  So start out
2297          * reserving a larger area and make it smaller until it succeeds.
2298          */
2299
2300         if (screen_info.lfb_base) {
2301                 if (efi_enabled(EFI_BOOT))
2302                         size = max_t(__u32, screen_info.lfb_size, 0x800000);
2303                 else
2304                         size = max_t(__u32, screen_info.lfb_size, 0x4000000);
2305
2306                 for (; !fb_mmio && (size >= 0x100000); size >>= 1) {
2307                         fb_mmio = __request_region(hyperv_mmio,
2308                                                    screen_info.lfb_base, size,
2309                                                    fb_mmio_name, 0);
2310                 }
2311         }
2312 }
2313
2314 /**
2315  * vmbus_allocate_mmio() - Pick a memory-mapped I/O range.
2316  * @new:                If successful, supplied a pointer to the
2317  *                      allocated MMIO space.
2318  * @device_obj:         Identifies the caller
2319  * @min:                Minimum guest physical address of the
2320  *                      allocation
2321  * @max:                Maximum guest physical address
2322  * @size:               Size of the range to be allocated
2323  * @align:              Alignment of the range to be allocated
2324  * @fb_overlap_ok:      Whether this allocation can be allowed
2325  *                      to overlap the video frame buffer.
2326  *
2327  * This function walks the resources granted to VMBus by the
2328  * _CRS object in the ACPI namespace underneath the parent
2329  * "bridge" whether that's a root PCI bus in the Generation 1
2330  * case or a Module Device in the Generation 2 case.  It then
2331  * attempts to allocate from the global MMIO pool in a way that
2332  * matches the constraints supplied in these parameters and by
2333  * that _CRS.
2334  *
2335  * Return: 0 on success, -errno on failure
2336  */
2337 int vmbus_allocate_mmio(struct resource **new, struct hv_device *device_obj,
2338                         resource_size_t min, resource_size_t max,
2339                         resource_size_t size, resource_size_t align,
2340                         bool fb_overlap_ok)
2341 {
2342         struct resource *iter, *shadow;
2343         resource_size_t range_min, range_max, start, end;
2344         const char *dev_n = dev_name(&device_obj->device);
2345         int retval;
2346
2347         retval = -ENXIO;
2348         mutex_lock(&hyperv_mmio_lock);
2349
2350         /*
2351          * If overlaps with frame buffers are allowed, then first attempt to
2352          * make the allocation from within the reserved region.  Because it
2353          * is already reserved, no shadow allocation is necessary.
2354          */
2355         if (fb_overlap_ok && fb_mmio && !(min > fb_mmio->end) &&
2356             !(max < fb_mmio->start)) {
2357
2358                 range_min = fb_mmio->start;
2359                 range_max = fb_mmio->end;
2360                 start = (range_min + align - 1) & ~(align - 1);
2361                 for (; start + size - 1 <= range_max; start += align) {
2362                         *new = request_mem_region_exclusive(start, size, dev_n);
2363                         if (*new) {
2364                                 retval = 0;
2365                                 goto exit;
2366                         }
2367                 }
2368         }
2369
2370         for (iter = hyperv_mmio; iter; iter = iter->sibling) {
2371                 if ((iter->start >= max) || (iter->end <= min))
2372                         continue;
2373
2374                 range_min = iter->start;
2375                 range_max = iter->end;
2376                 start = (range_min + align - 1) & ~(align - 1);
2377                 for (; start + size - 1 <= range_max; start += align) {
2378                         end = start + size - 1;
2379
2380                         /* Skip the whole fb_mmio region if not fb_overlap_ok */
2381                         if (!fb_overlap_ok && fb_mmio &&
2382                             (((start >= fb_mmio->start) && (start <= fb_mmio->end)) ||
2383                              ((end >= fb_mmio->start) && (end <= fb_mmio->end))))
2384                                 continue;
2385
2386                         shadow = __request_region(iter, start, size, NULL,
2387                                                   IORESOURCE_BUSY);
2388                         if (!shadow)
2389                                 continue;
2390
2391                         *new = request_mem_region_exclusive(start, size, dev_n);
2392                         if (*new) {
2393                                 shadow->name = (char *)*new;
2394                                 retval = 0;
2395                                 goto exit;
2396                         }
2397
2398                         __release_region(iter, start, size);
2399                 }
2400         }
2401
2402 exit:
2403         mutex_unlock(&hyperv_mmio_lock);
2404         return retval;
2405 }
2406 EXPORT_SYMBOL_GPL(vmbus_allocate_mmio);
2407
2408 /**
2409  * vmbus_free_mmio() - Free a memory-mapped I/O range.
2410  * @start:              Base address of region to release.
2411  * @size:               Size of the range to be allocated
2412  *
2413  * This function releases anything requested by
2414  * vmbus_mmio_allocate().
2415  */
2416 void vmbus_free_mmio(resource_size_t start, resource_size_t size)
2417 {
2418         struct resource *iter;
2419
2420         mutex_lock(&hyperv_mmio_lock);
2421         for (iter = hyperv_mmio; iter; iter = iter->sibling) {
2422                 if ((iter->start >= start + size) || (iter->end <= start))
2423                         continue;
2424
2425                 __release_region(iter, start, size);
2426         }
2427         release_mem_region(start, size);
2428         mutex_unlock(&hyperv_mmio_lock);
2429
2430 }
2431 EXPORT_SYMBOL_GPL(vmbus_free_mmio);
2432
2433 static int vmbus_acpi_add(struct acpi_device *device)
2434 {
2435         acpi_status result;
2436         int ret_val = -ENODEV;
2437         struct acpi_device *ancestor;
2438
2439         hv_acpi_dev = device;
2440
2441         result = acpi_walk_resources(device->handle, METHOD_NAME__CRS,
2442                                         vmbus_walk_resources, NULL);
2443
2444         if (ACPI_FAILURE(result))
2445                 goto acpi_walk_err;
2446         /*
2447          * Some ancestor of the vmbus acpi device (Gen1 or Gen2
2448          * firmware) is the VMOD that has the mmio ranges. Get that.
2449          */
2450         for (ancestor = device->parent; ancestor; ancestor = ancestor->parent) {
2451                 result = acpi_walk_resources(ancestor->handle, METHOD_NAME__CRS,
2452                                              vmbus_walk_resources, NULL);
2453
2454                 if (ACPI_FAILURE(result))
2455                         continue;
2456                 if (hyperv_mmio) {
2457                         vmbus_reserve_fb();
2458                         break;
2459                 }
2460         }
2461         ret_val = 0;
2462
2463 acpi_walk_err:
2464         complete(&probe_event);
2465         if (ret_val)
2466                 vmbus_acpi_remove(device);
2467         return ret_val;
2468 }
2469
2470 #ifdef CONFIG_PM_SLEEP
2471 static int vmbus_bus_suspend(struct device *dev)
2472 {
2473         struct vmbus_channel *channel, *sc;
2474
2475         while (atomic_read(&vmbus_connection.offer_in_progress) != 0) {
2476                 /*
2477                  * We wait here until the completion of any channel
2478                  * offers that are currently in progress.
2479                  */
2480                 usleep_range(1000, 2000);
2481         }
2482
2483         mutex_lock(&vmbus_connection.channel_mutex);
2484         list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
2485                 if (!is_hvsock_channel(channel))
2486                         continue;
2487
2488                 vmbus_force_channel_rescinded(channel);
2489         }
2490         mutex_unlock(&vmbus_connection.channel_mutex);
2491
2492         /*
2493          * Wait until all the sub-channels and hv_sock channels have been
2494          * cleaned up. Sub-channels should be destroyed upon suspend, otherwise
2495          * they would conflict with the new sub-channels that will be created
2496          * in the resume path. hv_sock channels should also be destroyed, but
2497          * a hv_sock channel of an established hv_sock connection can not be
2498          * really destroyed since it may still be referenced by the userspace
2499          * application, so we just force the hv_sock channel to be rescinded
2500          * by vmbus_force_channel_rescinded(), and the userspace application
2501          * will thoroughly destroy the channel after hibernation.
2502          *
2503          * Note: the counter nr_chan_close_on_suspend may never go above 0 if
2504          * the VM has no sub-channel and hv_sock channel, e.g. a 1-vCPU VM.
2505          */
2506         if (atomic_read(&vmbus_connection.nr_chan_close_on_suspend) > 0)
2507                 wait_for_completion(&vmbus_connection.ready_for_suspend_event);
2508
2509         if (atomic_read(&vmbus_connection.nr_chan_fixup_on_resume) != 0) {
2510                 pr_err("Can not suspend due to a previous failed resuming\n");
2511                 return -EBUSY;
2512         }
2513
2514         mutex_lock(&vmbus_connection.channel_mutex);
2515
2516         list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
2517                 /*
2518                  * Remove the channel from the array of channels and invalidate
2519                  * the channel's relid.  Upon resume, vmbus_onoffer() will fix
2520                  * up the relid (and other fields, if necessary) and add the
2521                  * channel back to the array.
2522                  */
2523                 vmbus_channel_unmap_relid(channel);
2524                 channel->offermsg.child_relid = INVALID_RELID;
2525
2526                 if (is_hvsock_channel(channel)) {
2527                         if (!channel->rescind) {
2528                                 pr_err("hv_sock channel not rescinded!\n");
2529                                 WARN_ON_ONCE(1);
2530                         }
2531                         continue;
2532                 }
2533
2534                 list_for_each_entry(sc, &channel->sc_list, sc_list) {
2535                         pr_err("Sub-channel not deleted!\n");
2536                         WARN_ON_ONCE(1);
2537                 }
2538
2539                 atomic_inc(&vmbus_connection.nr_chan_fixup_on_resume);
2540         }
2541
2542         mutex_unlock(&vmbus_connection.channel_mutex);
2543
2544         vmbus_initiate_unload(false);
2545
2546         /* Reset the event for the next resume. */
2547         reinit_completion(&vmbus_connection.ready_for_resume_event);
2548
2549         return 0;
2550 }
2551
2552 static int vmbus_bus_resume(struct device *dev)
2553 {
2554         struct vmbus_channel_msginfo *msginfo;
2555         size_t msgsize;
2556         int ret;
2557
2558         /*
2559          * We only use the 'vmbus_proto_version', which was in use before
2560          * hibernation, to re-negotiate with the host.
2561          */
2562         if (!vmbus_proto_version) {
2563                 pr_err("Invalid proto version = 0x%x\n", vmbus_proto_version);
2564                 return -EINVAL;
2565         }
2566
2567         msgsize = sizeof(*msginfo) +
2568                   sizeof(struct vmbus_channel_initiate_contact);
2569
2570         msginfo = kzalloc(msgsize, GFP_KERNEL);
2571
2572         if (msginfo == NULL)
2573                 return -ENOMEM;
2574
2575         ret = vmbus_negotiate_version(msginfo, vmbus_proto_version);
2576
2577         kfree(msginfo);
2578
2579         if (ret != 0)
2580                 return ret;
2581
2582         WARN_ON(atomic_read(&vmbus_connection.nr_chan_fixup_on_resume) == 0);
2583
2584         vmbus_request_offers();
2585
2586         if (wait_for_completion_timeout(
2587                 &vmbus_connection.ready_for_resume_event, 10 * HZ) == 0)
2588                 pr_err("Some vmbus device is missing after suspending?\n");
2589
2590         /* Reset the event for the next suspend. */
2591         reinit_completion(&vmbus_connection.ready_for_suspend_event);
2592
2593         return 0;
2594 }
2595 #else
2596 #define vmbus_bus_suspend NULL
2597 #define vmbus_bus_resume NULL
2598 #endif /* CONFIG_PM_SLEEP */
2599
2600 static const struct acpi_device_id vmbus_acpi_device_ids[] = {
2601         {"VMBUS", 0},
2602         {"VMBus", 0},
2603         {"", 0},
2604 };
2605 MODULE_DEVICE_TABLE(acpi, vmbus_acpi_device_ids);
2606
2607 /*
2608  * Note: we must use the "no_irq" ops, otherwise hibernation can not work with
2609  * PCI device assignment, because "pci_dev_pm_ops" uses the "noirq" ops: in
2610  * the resume path, the pci "noirq" restore op runs before "non-noirq" op (see
2611  * resume_target_kernel() -> dpm_resume_start(), and hibernation_restore() ->
2612  * dpm_resume_end()). This means vmbus_bus_resume() and the pci-hyperv's
2613  * resume callback must also run via the "noirq" ops.
2614  *
2615  * Set suspend_noirq/resume_noirq to NULL for Suspend-to-Idle: see the comment
2616  * earlier in this file before vmbus_pm.
2617  */
2618
2619 static const struct dev_pm_ops vmbus_bus_pm = {
2620         .suspend_noirq  = NULL,
2621         .resume_noirq   = NULL,
2622         .freeze_noirq   = vmbus_bus_suspend,
2623         .thaw_noirq     = vmbus_bus_resume,
2624         .poweroff_noirq = vmbus_bus_suspend,
2625         .restore_noirq  = vmbus_bus_resume
2626 };
2627
2628 static struct acpi_driver vmbus_acpi_driver = {
2629         .name = "vmbus",
2630         .ids = vmbus_acpi_device_ids,
2631         .ops = {
2632                 .add = vmbus_acpi_add,
2633                 .remove = vmbus_acpi_remove,
2634         },
2635         .drv.pm = &vmbus_bus_pm,
2636 };
2637
2638 static void hv_kexec_handler(void)
2639 {
2640         hv_stimer_global_cleanup();
2641         vmbus_initiate_unload(false);
2642         /* Make sure conn_state is set as hv_synic_cleanup checks for it */
2643         mb();
2644         cpuhp_remove_state(hyperv_cpuhp_online);
2645 };
2646
2647 static void hv_crash_handler(struct pt_regs *regs)
2648 {
2649         int cpu;
2650
2651         vmbus_initiate_unload(true);
2652         /*
2653          * In crash handler we can't schedule synic cleanup for all CPUs,
2654          * doing the cleanup for current CPU only. This should be sufficient
2655          * for kdump.
2656          */
2657         cpu = smp_processor_id();
2658         hv_stimer_cleanup(cpu);
2659         hv_synic_disable_regs(cpu);
2660 };
2661
2662 static int hv_synic_suspend(void)
2663 {
2664         /*
2665          * When we reach here, all the non-boot CPUs have been offlined.
2666          * If we're in a legacy configuration where stimer Direct Mode is
2667          * not enabled, the stimers on the non-boot CPUs have been unbound
2668          * in hv_synic_cleanup() -> hv_stimer_legacy_cleanup() ->
2669          * hv_stimer_cleanup() -> clockevents_unbind_device().
2670          *
2671          * hv_synic_suspend() only runs on CPU0 with interrupts disabled.
2672          * Here we do not call hv_stimer_legacy_cleanup() on CPU0 because:
2673          * 1) it's unnecessary as interrupts remain disabled between
2674          * syscore_suspend() and syscore_resume(): see create_image() and
2675          * resume_target_kernel()
2676          * 2) the stimer on CPU0 is automatically disabled later by
2677          * syscore_suspend() -> timekeeping_suspend() -> tick_suspend() -> ...
2678          * -> clockevents_shutdown() -> ... -> hv_ce_shutdown()
2679          * 3) a warning would be triggered if we call
2680          * clockevents_unbind_device(), which may sleep, in an
2681          * interrupts-disabled context.
2682          */
2683
2684         hv_synic_disable_regs(0);
2685
2686         return 0;
2687 }
2688
2689 static void hv_synic_resume(void)
2690 {
2691         hv_synic_enable_regs(0);
2692
2693         /*
2694          * Note: we don't need to call hv_stimer_init(0), because the timer
2695          * on CPU0 is not unbound in hv_synic_suspend(), and the timer is
2696          * automatically re-enabled in timekeeping_resume().
2697          */
2698 }
2699
2700 /* The callbacks run only on CPU0, with irqs_disabled. */
2701 static struct syscore_ops hv_synic_syscore_ops = {
2702         .suspend = hv_synic_suspend,
2703         .resume = hv_synic_resume,
2704 };
2705
2706 static int __init hv_acpi_init(void)
2707 {
2708         int ret, t;
2709
2710         if (!hv_is_hyperv_initialized())
2711                 return -ENODEV;
2712
2713         if (hv_root_partition)
2714                 return 0;
2715
2716         init_completion(&probe_event);
2717
2718         /*
2719          * Get ACPI resources first.
2720          */
2721         ret = acpi_bus_register_driver(&vmbus_acpi_driver);
2722
2723         if (ret)
2724                 return ret;
2725
2726         t = wait_for_completion_timeout(&probe_event, 5*HZ);
2727         if (t == 0) {
2728                 ret = -ETIMEDOUT;
2729                 goto cleanup;
2730         }
2731
2732         /*
2733          * If we're on an architecture with a hardcoded hypervisor
2734          * vector (i.e. x86/x64), override the VMbus interrupt found
2735          * in the ACPI tables. Ensure vmbus_irq is not set since the
2736          * normal Linux IRQ mechanism is not used in this case.
2737          */
2738 #ifdef HYPERVISOR_CALLBACK_VECTOR
2739         vmbus_interrupt = HYPERVISOR_CALLBACK_VECTOR;
2740         vmbus_irq = -1;
2741 #endif
2742
2743         hv_debug_init();
2744
2745         ret = vmbus_bus_init();
2746         if (ret)
2747                 goto cleanup;
2748
2749         hv_setup_kexec_handler(hv_kexec_handler);
2750         hv_setup_crash_handler(hv_crash_handler);
2751
2752         register_syscore_ops(&hv_synic_syscore_ops);
2753
2754         return 0;
2755
2756 cleanup:
2757         acpi_bus_unregister_driver(&vmbus_acpi_driver);
2758         hv_acpi_dev = NULL;
2759         return ret;
2760 }
2761
2762 static void __exit vmbus_exit(void)
2763 {
2764         int cpu;
2765
2766         unregister_syscore_ops(&hv_synic_syscore_ops);
2767
2768         hv_remove_kexec_handler();
2769         hv_remove_crash_handler();
2770         vmbus_connection.conn_state = DISCONNECTED;
2771         hv_stimer_global_cleanup();
2772         vmbus_disconnect();
2773         if (vmbus_irq == -1) {
2774                 hv_remove_vmbus_handler();
2775         } else {
2776                 free_percpu_irq(vmbus_irq, vmbus_evt);
2777                 free_percpu(vmbus_evt);
2778         }
2779         for_each_online_cpu(cpu) {
2780                 struct hv_per_cpu_context *hv_cpu
2781                         = per_cpu_ptr(hv_context.cpu_context, cpu);
2782
2783                 tasklet_kill(&hv_cpu->msg_dpc);
2784         }
2785         hv_debug_rm_all_dir();
2786
2787         vmbus_free_channels();
2788         kfree(vmbus_connection.channels);
2789
2790         if (ms_hyperv.misc_features & HV_FEATURE_GUEST_CRASH_MSR_AVAILABLE) {
2791                 kmsg_dump_unregister(&hv_kmsg_dumper);
2792                 unregister_die_notifier(&hyperv_die_block);
2793         }
2794
2795         /*
2796          * The panic notifier is always registered, hence we should
2797          * also unconditionally unregister it here as well.
2798          */
2799         atomic_notifier_chain_unregister(&panic_notifier_list,
2800                                          &hyperv_panic_block);
2801
2802         free_page((unsigned long)hv_panic_page);
2803         unregister_sysctl_table(hv_ctl_table_hdr);
2804         hv_ctl_table_hdr = NULL;
2805         bus_unregister(&hv_bus);
2806
2807         cpuhp_remove_state(hyperv_cpuhp_online);
2808         hv_synic_free();
2809         acpi_bus_unregister_driver(&vmbus_acpi_driver);
2810 }
2811
2812
2813 MODULE_LICENSE("GPL");
2814 MODULE_DESCRIPTION("Microsoft Hyper-V VMBus Driver");
2815
2816 subsys_initcall(hv_acpi_init);
2817 module_exit(vmbus_exit);