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