drivers/hv/vmbus_drv.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (c) 2009, Microsoft Corporation.
   4  *
   5  * Authors:
   6  *   Haiyang Zhang <haiyangz@microsoft.com>
   7  *   Hank Janssen  <hjanssen@microsoft.com>
   8  *   K. Y. Srinivasan <kys@microsoft.com>
   9  */
  10 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  11
  12 #include <linux/init.h>
  13 #include <linux/module.h>
  14 #include <linux/device.h>
  15 #include <linux/interrupt.h>
  16 #include <linux/sysctl.h>
  17 #include <linux/slab.h>
  18 #include <linux/acpi.h>
  19 #include <linux/completion.h>
  20 #include <linux/hyperv.h>
  21 #include <linux/kernel_stat.h>
  22 #include <linux/clockchips.h>
  23 #include <linux/cpu.h>
  24 #include <linux/sched/isolation.h>
  25 #include <linux/sched/task_stack.h>
  26
  27 #include <linux/delay.h>
  28 #include <linux/panic_notifier.h>
  29 #include <linux/ptrace.h>
  30 #include <linux/screen_info.h>
  31 #include <linux/kdebug.h>
  32 #include <linux/efi.h>
  33 #include <linux/random.h>
  34 #include <linux/kernel.h>
  35 #include <linux/syscore_ops.h>
  36 #include <linux/dma-map-ops.h>
  37 #include <linux/pci.h>
  38 #include <clocksource/hyperv_timer.h>
  39 #include <asm/mshyperv.h>
  40 #include "hyperv_vmbus.h"
  41
  42 struct vmbus_dynid {
  43         struct list_head node;
  44         struct hv_vmbus_device_id id;
  45 };
  46
  47 static struct acpi_device  *hv_acpi_dev;
  48
  49 static int hyperv_cpuhp_online;
  50
  51 static void *hv_panic_page;
  52
  53 static long __percpu *vmbus_evt;
  54
  55 /* Values parsed from ACPI DSDT */
  56 int vmbus_irq;
  57 int vmbus_interrupt;
  58
  59 /*
  60  * Boolean to control whether to report panic messages over Hyper-V.
  61  *
  62  * It can be set via /proc/sys/kernel/hyperv_record_panic_msg
  63  */
  64 static int sysctl_record_panic_msg = 1;
  65
  66 static int hyperv_report_reg(void)
  67 {
  68         return !sysctl_record_panic_msg || !hv_panic_page;
  69 }
  70
  71 /*
  72  * The panic notifier below is responsible solely for unloading the
  73  * vmbus connection, which is necessary in a panic event.
  74  *
  75  * Notice an intrincate relation of this notifier with Hyper-V
  76  * framebuffer panic notifier exists - we need vmbus connection alive
  77  * there in order to succeed, so we need to order both with each other
  78  * [see hvfb_on_panic()] - this is done using notifiers' priorities.
  79  */
  80 static int hv_panic_vmbus_unload(struct notifier_block *nb, unsigned long val,
  81                               void *args)
  82 {
  83         vmbus_initiate_unload(true);
  84         return NOTIFY_DONE;
  85 }
  86 static struct notifier_block hyperv_panic_vmbus_unload_block = {
  87         .notifier_call  = hv_panic_vmbus_unload,
  88         .priority       = INT_MIN + 1, /* almost the latest one to execute */
  89 };
  90
  91 static int hv_die_panic_notify_crash(struct notifier_block *self,
  92                                      unsigned long val, void *args);
  93
  94 static struct notifier_block hyperv_die_report_block = {
  95         .notifier_call = hv_die_panic_notify_crash,
  96 };
  97 static struct notifier_block hyperv_panic_report_block = {
  98         .notifier_call = hv_die_panic_notify_crash,
  99 };
 100
 101 /*
 102  * The following callback works both as die and panic notifier; its
 103  * goal is to provide panic information to the hypervisor unless the
 104  * kmsg dumper is used [see hv_kmsg_dump()], which provides more
 105  * information but isn't always available.
 106  *
 107  * Notice that both the panic/die report notifiers are registered only
 108  * if we have the capability HV_FEATURE_GUEST_CRASH_MSR_AVAILABLE set.
 109  */
 110 static int hv_die_panic_notify_crash(struct notifier_block *self,
 111                                      unsigned long val, void *args)
 112 {
 113         struct pt_regs *regs;
 114         bool is_die;
 115
 116         /* Don't notify Hyper-V unless we have a die oops event or panic. */
 117         if (self == &hyperv_panic_report_block) {
 118                 is_die = false;
 119                 regs = current_pt_regs();
 120         } else { /* die event */
 121                 if (val != DIE_OOPS)
 122                         return NOTIFY_DONE;
 123
 124                 is_die = true;
 125                 regs = ((struct die_args *)args)->regs;
 126         }
 127
 128         /*
 129          * Hyper-V should be notified only once about a panic/die. If we will
 130          * be calling hv_kmsg_dump() later with kmsg data, don't do the
 131          * notification here.
 132          */
 133         if (hyperv_report_reg())
 134                 hyperv_report_panic(regs, val, is_die);
 135
 136         return NOTIFY_DONE;
 137 }
 138
 139 static const char *fb_mmio_name = "fb_range";
 140 static struct resource *fb_mmio;
 141 static struct resource *hyperv_mmio;
 142 static DEFINE_MUTEX(hyperv_mmio_lock);
 143
 144 static int vmbus_exists(void)
 145 {
 146         if (hv_acpi_dev == NULL)
 147                 return -ENODEV;
 148
 149         return 0;
 150 }
 151
 152 static u8 channel_monitor_group(const struct vmbus_channel *channel)
 153 {
 154         return (u8)channel->offermsg.monitorid / 32;
 155 }
 156
 157 static u8 channel_monitor_offset(const struct vmbus_channel *channel)
 158 {
 159         return (u8)channel->offermsg.monitorid % 32;
 160 }
 161
 162 static u32 channel_pending(const struct vmbus_channel *channel,
 163                            const struct hv_monitor_page *monitor_page)
 164 {
 165         u8 monitor_group = channel_monitor_group(channel);
 166
 167         return monitor_page->trigger_group[monitor_group].pending;
 168 }
 169
 170 static u32 channel_latency(const struct vmbus_channel *channel,
 171                            const struct hv_monitor_page *monitor_page)
 172 {
 173         u8 monitor_group = channel_monitor_group(channel);
 174         u8 monitor_offset = channel_monitor_offset(channel);
 175
 176         return monitor_page->latency[monitor_group][monitor_offset];
 177 }
 178
 179 static u32 channel_conn_id(struct vmbus_channel *channel,
 180                            struct hv_monitor_page *monitor_page)
 181 {
 182         u8 monitor_group = channel_monitor_group(channel);
 183         u8 monitor_offset = channel_monitor_offset(channel);
 184
 185         return monitor_page->parameter[monitor_group][monitor_offset].connectionid.u.id;
 186 }
 187
 188 static ssize_t id_show(struct device *dev, struct device_attribute *dev_attr,
 189                        char *buf)
 190 {
 191         struct hv_device *hv_dev = device_to_hv_device(dev);
 192
 193         if (!hv_dev->channel)
 194                 return -ENODEV;
 195         return sprintf(buf, "%d\n", hv_dev->channel->offermsg.child_relid);
 196 }
 197 static DEVICE_ATTR_RO(id);
 198
 199 static ssize_t state_show(struct device *dev, struct device_attribute *dev_attr,
 200                           char *buf)
 201 {
 202         struct hv_device *hv_dev = device_to_hv_device(dev);
 203
 204         if (!hv_dev->channel)
 205                 return -ENODEV;
 206         return sprintf(buf, "%d\n", hv_dev->channel->state);
 207 }
 208 static DEVICE_ATTR_RO(state);
 209
 210 static ssize_t monitor_id_show(struct device *dev,
 211                                struct device_attribute *dev_attr, char *buf)
 212 {
 213         struct hv_device *hv_dev = device_to_hv_device(dev);
 214
 215         if (!hv_dev->channel)
 216                 return -ENODEV;
 217         return sprintf(buf, "%d\n", hv_dev->channel->offermsg.monitorid);
 218 }
 219 static DEVICE_ATTR_RO(monitor_id);
 220
 221 static ssize_t class_id_show(struct device *dev,
 222                                struct device_attribute *dev_attr, 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 sprintf(buf, "{%pUl}\n",
 229                        &hv_dev->channel->offermsg.offer.if_type);
 230 }
 231 static DEVICE_ATTR_RO(class_id);
 232
 233 static ssize_t device_id_show(struct device *dev,
 234                               struct device_attribute *dev_attr, char *buf)
 235 {
 236         struct hv_device *hv_dev = device_to_hv_device(dev);
 237
 238         if (!hv_dev->channel)
 239                 return -ENODEV;
 240         return sprintf(buf, "{%pUl}\n",
 241                        &hv_dev->channel->offermsg.offer.if_instance);
 242 }
 243 static DEVICE_ATTR_RO(device_id);
 244
 245 static ssize_t modalias_show(struct device *dev,
 246                              struct device_attribute *dev_attr, char *buf)
 247 {
 248         struct hv_device *hv_dev = device_to_hv_device(dev);
 249
 250         return sprintf(buf, "vmbus:%*phN\n", UUID_SIZE, &hv_dev->dev_type);
 251 }
 252 static DEVICE_ATTR_RO(modalias);
 253
 254 #ifdef CONFIG_NUMA
 255 static ssize_t numa_node_show(struct device *dev,
 256                               struct device_attribute *attr, char *buf)
 257 {
 258         struct hv_device *hv_dev = device_to_hv_device(dev);
 259
 260         if (!hv_dev->channel)
 261                 return -ENODEV;
 262
 263         return sprintf(buf, "%d\n", cpu_to_node(hv_dev->channel->target_cpu));
 264 }
 265 static DEVICE_ATTR_RO(numa_node);
 266 #endif
 267
 268 static ssize_t server_monitor_pending_show(struct device *dev,
 269                                            struct device_attribute *dev_attr,
 270                                            char *buf)
 271 {
 272         struct hv_device *hv_dev = device_to_hv_device(dev);
 273
 274         if (!hv_dev->channel)
 275                 return -ENODEV;
 276         return sprintf(buf, "%d\n",
 277                        channel_pending(hv_dev->channel,
 278                                        vmbus_connection.monitor_pages[0]));
 279 }
 280 static DEVICE_ATTR_RO(server_monitor_pending);
 281
 282 static ssize_t client_monitor_pending_show(struct device *dev,
 283                                            struct device_attribute *dev_attr,
 284                                            char *buf)
 285 {
 286         struct hv_device *hv_dev = device_to_hv_device(dev);
 287
 288         if (!hv_dev->channel)
 289                 return -ENODEV;
 290         return sprintf(buf, "%d\n",
 291                        channel_pending(hv_dev->channel,
 292                                        vmbus_connection.monitor_pages[1]));
 293 }
 294 static DEVICE_ATTR_RO(client_monitor_pending);
 295
 296 static ssize_t server_monitor_latency_show(struct device *dev,
 297                                            struct device_attribute *dev_attr,
 298                                            char *buf)
 299 {
 300         struct hv_device *hv_dev = device_to_hv_device(dev);
 301
 302         if (!hv_dev->channel)
 303                 return -ENODEV;
 304         return sprintf(buf, "%d\n",
 305                        channel_latency(hv_dev->channel,
 306                                        vmbus_connection.monitor_pages[0]));
 307 }
 308 static DEVICE_ATTR_RO(server_monitor_latency);
 309
 310 static ssize_t client_monitor_latency_show(struct device *dev,
 311                                            struct device_attribute *dev_attr,
 312                                            char *buf)
 313 {
 314         struct hv_device *hv_dev = device_to_hv_device(dev);
 315
 316         if (!hv_dev->channel)
 317                 return -ENODEV;
 318         return sprintf(buf, "%d\n",
 319                        channel_latency(hv_dev->channel,
 320                                        vmbus_connection.monitor_pages[1]));
 321 }
 322 static DEVICE_ATTR_RO(client_monitor_latency);
 323
 324 static ssize_t server_monitor_conn_id_show(struct device *dev,
 325                                            struct device_attribute *dev_attr,
 326                                            char *buf)
 327 {
 328         struct hv_device *hv_dev = device_to_hv_device(dev);
 329
 330         if (!hv_dev->channel)
 331                 return -ENODEV;
 332         return sprintf(buf, "%d\n",
 333                        channel_conn_id(hv_dev->channel,
 334                                        vmbus_connection.monitor_pages[0]));
 335 }
 336 static DEVICE_ATTR_RO(server_monitor_conn_id);
 337
 338 static ssize_t client_monitor_conn_id_show(struct device *dev,
 339                                            struct device_attribute *dev_attr,
 340                                            char *buf)
 341 {
 342         struct hv_device *hv_dev = device_to_hv_device(dev);
 343
 344         if (!hv_dev->channel)
 345                 return -ENODEV;
 346         return sprintf(buf, "%d\n",
 347                        channel_conn_id(hv_dev->channel,
 348                                        vmbus_connection.monitor_pages[1]));
 349 }
 350 static DEVICE_ATTR_RO(client_monitor_conn_id);
 351
 352 static ssize_t out_intr_mask_show(struct device *dev,
 353                                   struct device_attribute *dev_attr, char *buf)
 354 {
 355         struct hv_device *hv_dev = device_to_hv_device(dev);
 356         struct hv_ring_buffer_debug_info outbound;
 357         int ret;
 358
 359         if (!hv_dev->channel)
 360                 return -ENODEV;
 361
 362         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->outbound,
 363                                           &outbound);
 364         if (ret < 0)
 365                 return ret;
 366
 367         return sprintf(buf, "%d\n", outbound.current_interrupt_mask);
 368 }
 369 static DEVICE_ATTR_RO(out_intr_mask);
 370
 371 static ssize_t out_read_index_show(struct device *dev,
 372                                    struct device_attribute *dev_attr, char *buf)
 373 {
 374         struct hv_device *hv_dev = device_to_hv_device(dev);
 375         struct hv_ring_buffer_debug_info outbound;
 376         int ret;
 377
 378         if (!hv_dev->channel)
 379                 return -ENODEV;
 380
 381         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->outbound,
 382                                           &outbound);
 383         if (ret < 0)
 384                 return ret;
 385         return sprintf(buf, "%d\n", outbound.current_read_index);
 386 }
 387 static DEVICE_ATTR_RO(out_read_index);
 388
 389 static ssize_t out_write_index_show(struct device *dev,
 390                                     struct device_attribute *dev_attr,
 391                                     char *buf)
 392 {
 393         struct hv_device *hv_dev = device_to_hv_device(dev);
 394         struct hv_ring_buffer_debug_info outbound;
 395         int ret;
 396
 397         if (!hv_dev->channel)
 398                 return -ENODEV;
 399
 400         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->outbound,
 401                                           &outbound);
 402         if (ret < 0)
 403                 return ret;
 404         return sprintf(buf, "%d\n", outbound.current_write_index);
 405 }
 406 static DEVICE_ATTR_RO(out_write_index);
 407
 408 static ssize_t out_read_bytes_avail_show(struct device *dev,
 409                                          struct device_attribute *dev_attr,
 410                                          char *buf)
 411 {
 412         struct hv_device *hv_dev = device_to_hv_device(dev);
 413         struct hv_ring_buffer_debug_info outbound;
 414         int ret;
 415
 416         if (!hv_dev->channel)
 417                 return -ENODEV;
 418
 419         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->outbound,
 420                                           &outbound);
 421         if (ret < 0)
 422                 return ret;
 423         return sprintf(buf, "%d\n", outbound.bytes_avail_toread);
 424 }
 425 static DEVICE_ATTR_RO(out_read_bytes_avail);
 426
 427 static ssize_t out_write_bytes_avail_show(struct device *dev,
 428                                           struct device_attribute *dev_attr,
 429                                           char *buf)
 430 {
 431         struct hv_device *hv_dev = device_to_hv_device(dev);
 432         struct hv_ring_buffer_debug_info outbound;
 433         int ret;
 434
 435         if (!hv_dev->channel)
 436                 return -ENODEV;
 437
 438         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->outbound,
 439                                           &outbound);
 440         if (ret < 0)
 441                 return ret;
 442         return sprintf(buf, "%d\n", outbound.bytes_avail_towrite);
 443 }
 444 static DEVICE_ATTR_RO(out_write_bytes_avail);
 445
 446 static ssize_t in_intr_mask_show(struct device *dev,
 447                                  struct device_attribute *dev_attr, char *buf)
 448 {
 449         struct hv_device *hv_dev = device_to_hv_device(dev);
 450         struct hv_ring_buffer_debug_info inbound;
 451         int ret;
 452
 453         if (!hv_dev->channel)
 454                 return -ENODEV;
 455
 456         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->inbound, &inbound);
 457         if (ret < 0)
 458                 return ret;
 459
 460         return sprintf(buf, "%d\n", inbound.current_interrupt_mask);
 461 }
 462 static DEVICE_ATTR_RO(in_intr_mask);
 463
 464 static ssize_t in_read_index_show(struct device *dev,
 465                                   struct device_attribute *dev_attr, char *buf)
 466 {
 467         struct hv_device *hv_dev = device_to_hv_device(dev);
 468         struct hv_ring_buffer_debug_info inbound;
 469         int ret;
 470
 471         if (!hv_dev->channel)
 472                 return -ENODEV;
 473
 474         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->inbound, &inbound);
 475         if (ret < 0)
 476                 return ret;
 477
 478         return sprintf(buf, "%d\n", inbound.current_read_index);
 479 }
 480 static DEVICE_ATTR_RO(in_read_index);
 481
 482 static ssize_t in_write_index_show(struct device *dev,
 483                                    struct device_attribute *dev_attr, char *buf)
 484 {
 485         struct hv_device *hv_dev = device_to_hv_device(dev);
 486         struct hv_ring_buffer_debug_info inbound;
 487         int ret;
 488
 489         if (!hv_dev->channel)
 490                 return -ENODEV;
 491
 492         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->inbound, &inbound);
 493         if (ret < 0)
 494                 return ret;
 495
 496         return sprintf(buf, "%d\n", inbound.current_write_index);
 497 }
 498 static DEVICE_ATTR_RO(in_write_index);
 499
 500 static ssize_t in_read_bytes_avail_show(struct device *dev,
 501                                         struct device_attribute *dev_attr,
 502                                         char *buf)
 503 {
 504         struct hv_device *hv_dev = device_to_hv_device(dev);
 505         struct hv_ring_buffer_debug_info inbound;
 506         int ret;
 507
 508         if (!hv_dev->channel)
 509                 return -ENODEV;
 510
 511         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->inbound, &inbound);
 512         if (ret < 0)
 513                 return ret;
 514
 515         return sprintf(buf, "%d\n", inbound.bytes_avail_toread);
 516 }
 517 static DEVICE_ATTR_RO(in_read_bytes_avail);
 518
 519 static ssize_t in_write_bytes_avail_show(struct device *dev,
 520                                          struct device_attribute *dev_attr,
 521                                          char *buf)
 522 {
 523         struct hv_device *hv_dev = device_to_hv_device(dev);
 524         struct hv_ring_buffer_debug_info inbound;
 525         int ret;
 526
 527         if (!hv_dev->channel)
 528                 return -ENODEV;
 529
 530         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->inbound, &inbound);
 531         if (ret < 0)
 532                 return ret;
 533
 534         return sprintf(buf, "%d\n", inbound.bytes_avail_towrite);
 535 }
 536 static DEVICE_ATTR_RO(in_write_bytes_avail);
 537
 538 static ssize_t channel_vp_mapping_show(struct device *dev,
 539                                        struct device_attribute *dev_attr,
 540                                        char *buf)
 541 {
 542         struct hv_device *hv_dev = device_to_hv_device(dev);
 543         struct vmbus_channel *channel = hv_dev->channel, *cur_sc;
 544         int buf_size = PAGE_SIZE, n_written, tot_written;
 545         struct list_head *cur;
 546
 547         if (!channel)
 548                 return -ENODEV;
 549
 550         mutex_lock(&vmbus_connection.channel_mutex);
 551
 552         tot_written = snprintf(buf, buf_size, "%u:%u\n",
 553                 channel->offermsg.child_relid, channel->target_cpu);
 554
 555         list_for_each(cur, &channel->sc_list) {
 556                 if (tot_written >= buf_size - 1)
 557                         break;
 558
 559                 cur_sc = list_entry(cur, struct vmbus_channel, sc_list);
 560                 n_written = scnprintf(buf + tot_written,
 561                                      buf_size - tot_written,
 562                                      "%u:%u\n",
 563                                      cur_sc->offermsg.child_relid,
 564                                      cur_sc->target_cpu);
 565                 tot_written += n_written;
 566         }
 567
 568         mutex_unlock(&vmbus_connection.channel_mutex);
 569
 570         return tot_written;
 571 }
 572 static DEVICE_ATTR_RO(channel_vp_mapping);
 573
 574 static ssize_t vendor_show(struct device *dev,
 575                            struct device_attribute *dev_attr,
 576                            char *buf)
 577 {
 578         struct hv_device *hv_dev = device_to_hv_device(dev);
 579
 580         return sprintf(buf, "0x%x\n", hv_dev->vendor_id);
 581 }
 582 static DEVICE_ATTR_RO(vendor);
 583
 584 static ssize_t device_show(struct device *dev,
 585                            struct device_attribute *dev_attr,
 586                            char *buf)
 587 {
 588         struct hv_device *hv_dev = device_to_hv_device(dev);
 589
 590         return sprintf(buf, "0x%x\n", hv_dev->device_id);
 591 }
 592 static DEVICE_ATTR_RO(device);
 593
 594 static ssize_t driver_override_store(struct device *dev,
 595                                      struct device_attribute *attr,
 596                                      const char *buf, size_t count)
 597 {
 598         struct hv_device *hv_dev = device_to_hv_device(dev);
 599         int ret;
 600
 601         ret = driver_set_override(dev, &hv_dev->driver_override, buf, count);
 602         if (ret)
 603                 return ret;
 604
 605         return count;
 606 }
 607
 608 static ssize_t driver_override_show(struct device *dev,
 609                                     struct device_attribute *attr, char *buf)
 610 {
 611         struct hv_device *hv_dev = device_to_hv_device(dev);
 612         ssize_t len;
 613
 614         device_lock(dev);
 615         len = snprintf(buf, PAGE_SIZE, "%s\n", hv_dev->driver_override);
 616         device_unlock(dev);
 617
 618         return len;
 619 }
 620 static DEVICE_ATTR_RW(driver_override);
 621
 622 /* Set up per device attributes in /sys/bus/vmbus/devices/<bus device> */
 623 static struct attribute *vmbus_dev_attrs[] = {
 624         &dev_attr_id.attr,
 625         &dev_attr_state.attr,
 626         &dev_attr_monitor_id.attr,
 627         &dev_attr_class_id.attr,
 628         &dev_attr_device_id.attr,
 629         &dev_attr_modalias.attr,
 630 #ifdef CONFIG_NUMA
 631         &dev_attr_numa_node.attr,
 632 #endif
 633         &dev_attr_server_monitor_pending.attr,
 634         &dev_attr_client_monitor_pending.attr,
 635         &dev_attr_server_monitor_latency.attr,
 636         &dev_attr_client_monitor_latency.attr,
 637         &dev_attr_server_monitor_conn_id.attr,
 638         &dev_attr_client_monitor_conn_id.attr,
 639         &dev_attr_out_intr_mask.attr,
 640         &dev_attr_out_read_index.attr,
 641         &dev_attr_out_write_index.attr,
 642         &dev_attr_out_read_bytes_avail.attr,
 643         &dev_attr_out_write_bytes_avail.attr,
 644         &dev_attr_in_intr_mask.attr,
 645         &dev_attr_in_read_index.attr,
 646         &dev_attr_in_write_index.attr,
 647         &dev_attr_in_read_bytes_avail.attr,
 648         &dev_attr_in_write_bytes_avail.attr,
 649         &dev_attr_channel_vp_mapping.attr,
 650         &dev_attr_vendor.attr,
 651         &dev_attr_device.attr,
 652         &dev_attr_driver_override.attr,
 653         NULL,
 654 };
 655
 656 /*
 657  * Device-level attribute_group callback function. Returns the permission for
 658  * each attribute, and returns 0 if an attribute is not visible.
 659  */
 660 static umode_t vmbus_dev_attr_is_visible(struct kobject *kobj,
 661                                          struct attribute *attr, int idx)
 662 {
 663         struct device *dev = kobj_to_dev(kobj);
 664         const struct hv_device *hv_dev = device_to_hv_device(dev);
 665
 666         /* Hide the monitor attributes if the monitor mechanism is not used. */
 667         if (!hv_dev->channel->offermsg.monitor_allocated &&
 668             (attr == &dev_attr_monitor_id.attr ||
 669              attr == &dev_attr_server_monitor_pending.attr ||
 670              attr == &dev_attr_client_monitor_pending.attr ||
 671              attr == &dev_attr_server_monitor_latency.attr ||
 672              attr == &dev_attr_client_monitor_latency.attr ||
 673              attr == &dev_attr_server_monitor_conn_id.attr ||
 674              attr == &dev_attr_client_monitor_conn_id.attr))
 675                 return 0;
 676
 677         return attr->mode;
 678 }
 679
 680 static const struct attribute_group vmbus_dev_group = {
 681         .attrs = vmbus_dev_attrs,
 682         .is_visible = vmbus_dev_attr_is_visible
 683 };
 684 __ATTRIBUTE_GROUPS(vmbus_dev);
 685
 686 /* Set up the attribute for /sys/bus/vmbus/hibernation */
 687 static ssize_t hibernation_show(struct bus_type *bus, char *buf)
 688 {
 689         return sprintf(buf, "%d\n", !!hv_is_hibernation_supported());
 690 }
 691
 692 static BUS_ATTR_RO(hibernation);
 693
 694 static struct attribute *vmbus_bus_attrs[] = {
 695         &bus_attr_hibernation.attr,
 696         NULL,
 697 };
 698 static const struct attribute_group vmbus_bus_group = {
 699         .attrs = vmbus_bus_attrs,
 700 };
 701 __ATTRIBUTE_GROUPS(vmbus_bus);
 702
 703 /*
 704  * vmbus_uevent - add uevent for our device
 705  *
 706  * This routine is invoked when a device is added or removed on the vmbus to
 707  * generate a uevent to udev in the userspace. The udev will then look at its
 708  * rule and the uevent generated here to load the appropriate driver
 709  *
 710  * The alias string will be of the form vmbus:guid where guid is the string
 711  * representation of the device guid (each byte of the guid will be
 712  * represented with two hex characters.
 713  */
 714 static int vmbus_uevent(const struct device *device, struct kobj_uevent_env *env)
 715 {
 716         const struct hv_device *dev = device_to_hv_device(device);
 717         const char *format = "MODALIAS=vmbus:%*phN";
 718
 719         return add_uevent_var(env, format, UUID_SIZE, &dev->dev_type);
 720 }
 721
 722 static const struct hv_vmbus_device_id *
 723 hv_vmbus_dev_match(const struct hv_vmbus_device_id *id, const guid_t *guid)
 724 {
 725         if (id == NULL)
 726                 return NULL; /* empty device table */
 727
 728         for (; !guid_is_null(&id->guid); id++)
 729                 if (guid_equal(&id->guid, guid))
 730                         return id;
 731
 732         return NULL;
 733 }
 734
 735 static const struct hv_vmbus_device_id *
 736 hv_vmbus_dynid_match(struct hv_driver *drv, const guid_t *guid)
 737 {
 738         const struct hv_vmbus_device_id *id = NULL;
 739         struct vmbus_dynid *dynid;
 740
 741         spin_lock(&drv->dynids.lock);
 742         list_for_each_entry(dynid, &drv->dynids.list, node) {
 743                 if (guid_equal(&dynid->id.guid, guid)) {
 744                         id = &dynid->id;
 745                         break;
 746                 }
 747         }
 748         spin_unlock(&drv->dynids.lock);
 749
 750         return id;
 751 }
 752
 753 static const struct hv_vmbus_device_id vmbus_device_null;
 754
 755 /*
 756  * Return a matching hv_vmbus_device_id pointer.
 757  * If there is no match, return NULL.
 758  */
 759 static const struct hv_vmbus_device_id *hv_vmbus_get_id(struct hv_driver *drv,
 760                                                         struct hv_device *dev)
 761 {
 762         const guid_t *guid = &dev->dev_type;
 763         const struct hv_vmbus_device_id *id;
 764
 765         /* When driver_override is set, only bind to the matching driver */
 766         if (dev->driver_override && strcmp(dev->driver_override, drv->name))
 767                 return NULL;
 768
 769         /* Look at the dynamic ids first, before the static ones */
 770         id = hv_vmbus_dynid_match(drv, guid);
 771         if (!id)
 772                 id = hv_vmbus_dev_match(drv->id_table, guid);
 773
 774         /* driver_override will always match, send a dummy id */
 775         if (!id && dev->driver_override)
 776                 id = &vmbus_device_null;
 777
 778         return id;
 779 }
 780
 781 /* vmbus_add_dynid - add a new device ID to this driver and re-probe devices */
 782 static int vmbus_add_dynid(struct hv_driver *drv, guid_t *guid)
 783 {
 784         struct vmbus_dynid *dynid;
 785
 786         dynid = kzalloc(sizeof(*dynid), GFP_KERNEL);
 787         if (!dynid)
 788                 return -ENOMEM;
 789
 790         dynid->id.guid = *guid;
 791
 792         spin_lock(&drv->dynids.lock);
 793         list_add_tail(&dynid->node, &drv->dynids.list);
 794         spin_unlock(&drv->dynids.lock);
 795
 796         return driver_attach(&drv->driver);
 797 }
 798
 799 static void vmbus_free_dynids(struct hv_driver *drv)
 800 {
 801         struct vmbus_dynid *dynid, *n;
 802
 803         spin_lock(&drv->dynids.lock);
 804         list_for_each_entry_safe(dynid, n, &drv->dynids.list, node) {
 805                 list_del(&dynid->node);
 806                 kfree(dynid);
 807         }
 808         spin_unlock(&drv->dynids.lock);
 809 }
 810
 811 /*
 812  * store_new_id - sysfs frontend to vmbus_add_dynid()
 813  *
 814  * Allow GUIDs to be added to an existing driver via sysfs.
 815  */
 816 static ssize_t new_id_store(struct device_driver *driver, const char *buf,
 817                             size_t count)
 818 {
 819         struct hv_driver *drv = drv_to_hv_drv(driver);
 820         guid_t guid;
 821         ssize_t retval;
 822
 823         retval = guid_parse(buf, &guid);
 824         if (retval)
 825                 return retval;
 826
 827         if (hv_vmbus_dynid_match(drv, &guid))
 828                 return -EEXIST;
 829
 830         retval = vmbus_add_dynid(drv, &guid);
 831         if (retval)
 832                 return retval;
 833         return count;
 834 }
 835 static DRIVER_ATTR_WO(new_id);
 836
 837 /*
 838  * store_remove_id - remove a PCI device ID from this driver
 839  *
 840  * Removes a dynamic pci device ID to this driver.
 841  */
 842 static ssize_t remove_id_store(struct device_driver *driver, const char *buf,
 843                                size_t count)
 844 {
 845         struct hv_driver *drv = drv_to_hv_drv(driver);
 846         struct vmbus_dynid *dynid, *n;
 847         guid_t guid;
 848         ssize_t retval;
 849
 850         retval = guid_parse(buf, &guid);
 851         if (retval)
 852                 return retval;
 853
 854         retval = -ENODEV;
 855         spin_lock(&drv->dynids.lock);
 856         list_for_each_entry_safe(dynid, n, &drv->dynids.list, node) {
 857                 struct hv_vmbus_device_id *id = &dynid->id;
 858
 859                 if (guid_equal(&id->guid, &guid)) {
 860                         list_del(&dynid->node);
 861                         kfree(dynid);
 862                         retval = count;
 863                         break;
 864                 }
 865         }
 866         spin_unlock(&drv->dynids.lock);
 867
 868         return retval;
 869 }
 870 static DRIVER_ATTR_WO(remove_id);
 871
 872 static struct attribute *vmbus_drv_attrs[] = {
 873         &driver_attr_new_id.attr,
 874         &driver_attr_remove_id.attr,
 875         NULL,
 876 };
 877 ATTRIBUTE_GROUPS(vmbus_drv);
 878
 879
 880 /*
 881  * vmbus_match - Attempt to match the specified device to the specified driver
 882  */
 883 static int vmbus_match(struct device *device, struct device_driver *driver)
 884 {
 885         struct hv_driver *drv = drv_to_hv_drv(driver);
 886         struct hv_device *hv_dev = device_to_hv_device(device);
 887
 888         /* The hv_sock driver handles all hv_sock offers. */
 889         if (is_hvsock_channel(hv_dev->channel))
 890                 return drv->hvsock;
 891
 892         if (hv_vmbus_get_id(drv, hv_dev))
 893                 return 1;
 894
 895         return 0;
 896 }
 897
 898 /*
 899  * vmbus_probe - Add the new vmbus's child device
 900  */
 901 static int vmbus_probe(struct device *child_device)
 902 {
 903         int ret = 0;
 904         struct hv_driver *drv =
 905                         drv_to_hv_drv(child_device->driver);
 906         struct hv_device *dev = device_to_hv_device(child_device);
 907         const struct hv_vmbus_device_id *dev_id;
 908
 909         dev_id = hv_vmbus_get_id(drv, dev);
 910         if (drv->probe) {
 911                 ret = drv->probe(dev, dev_id);
 912                 if (ret != 0)
 913                         pr_err("probe failed for device %s (%d)\n",
 914                                dev_name(child_device), ret);
 915
 916         } else {
 917                 pr_err("probe not set for driver %s\n",
 918                        dev_name(child_device));
 919                 ret = -ENODEV;
 920         }
 921         return ret;
 922 }
 923
 924 /*
 925  * vmbus_dma_configure -- Configure DMA coherence for VMbus device
 926  */
 927 static int vmbus_dma_configure(struct device *child_device)
 928 {
 929         /*
 930          * On ARM64, propagate the DMA coherence setting from the top level
 931          * VMbus ACPI device to the child VMbus device being added here.
 932          * On x86/x64 coherence is assumed and these calls have no effect.
 933          */
 934         hv_setup_dma_ops(child_device,
 935                 device_get_dma_attr(&hv_acpi_dev->dev) == DEV_DMA_COHERENT);
 936         return 0;
 937 }
 938
 939 /*
 940  * vmbus_remove - Remove a vmbus device
 941  */
 942 static void vmbus_remove(struct device *child_device)
 943 {
 944         struct hv_driver *drv;
 945         struct hv_device *dev = device_to_hv_device(child_device);
 946
 947         if (child_device->driver) {
 948                 drv = drv_to_hv_drv(child_device->driver);
 949                 if (drv->remove)
 950                         drv->remove(dev);
 951         }
 952 }
 953
 954 /*
 955  * vmbus_shutdown - Shutdown a vmbus device
 956  */
 957 static void vmbus_shutdown(struct device *child_device)
 958 {
 959         struct hv_driver *drv;
 960         struct hv_device *dev = device_to_hv_device(child_device);
 961
 962
 963         /* The device may not be attached yet */
 964         if (!child_device->driver)
 965                 return;
 966
 967         drv = drv_to_hv_drv(child_device->driver);
 968
 969         if (drv->shutdown)
 970                 drv->shutdown(dev);
 971 }
 972
 973 #ifdef CONFIG_PM_SLEEP
 974 /*
 975  * vmbus_suspend - Suspend a vmbus device
 976  */
 977 static int vmbus_suspend(struct device *child_device)
 978 {
 979         struct hv_driver *drv;
 980         struct hv_device *dev = device_to_hv_device(child_device);
 981
 982         /* The device may not be attached yet */
 983         if (!child_device->driver)
 984                 return 0;
 985
 986         drv = drv_to_hv_drv(child_device->driver);
 987         if (!drv->suspend)
 988                 return -EOPNOTSUPP;
 989
 990         return drv->suspend(dev);
 991 }
 992
 993 /*
 994  * vmbus_resume - Resume a vmbus device
 995  */
 996 static int vmbus_resume(struct device *child_device)
 997 {
 998         struct hv_driver *drv;
 999         struct hv_device *dev = device_to_hv_device(child_device);
1000
1001         /* The device may not be attached yet */
1002         if (!child_device->driver)
1003                 return 0;
1004
1005         drv = drv_to_hv_drv(child_device->driver);
1006         if (!drv->resume)
1007                 return -EOPNOTSUPP;
1008
1009         return drv->resume(dev);
1010 }
1011 #else
1012 #define vmbus_suspend NULL
1013 #define vmbus_resume NULL
1014 #endif /* CONFIG_PM_SLEEP */
1015
1016 /*
1017  * vmbus_device_release - Final callback release of the vmbus child device
1018  */
1019 static void vmbus_device_release(struct device *device)
1020 {
1021         struct hv_device *hv_dev = device_to_hv_device(device);
1022         struct vmbus_channel *channel = hv_dev->channel;
1023
1024         hv_debug_rm_dev_dir(hv_dev);
1025
1026         mutex_lock(&vmbus_connection.channel_mutex);
1027         hv_process_channel_removal(channel);
1028         mutex_unlock(&vmbus_connection.channel_mutex);
1029         kfree(hv_dev);
1030 }
1031
1032 /*
1033  * Note: we must use the "noirq" ops: see the comment before vmbus_bus_pm.
1034  *
1035  * suspend_noirq/resume_noirq are set to NULL to support Suspend-to-Idle: we
1036  * shouldn't suspend the vmbus devices upon Suspend-to-Idle, otherwise there
1037  * is no way to wake up a Generation-2 VM.
1038  *
1039  * The other 4 ops are for hibernation.
1040  */
1041
1042 static const struct dev_pm_ops vmbus_pm = {
1043         .suspend_noirq  = NULL,
1044         .resume_noirq   = NULL,
1045         .freeze_noirq   = vmbus_suspend,
1046         .thaw_noirq     = vmbus_resume,
1047         .poweroff_noirq = vmbus_suspend,
1048         .restore_noirq  = vmbus_resume,
1049 };
1050
1051 /* The one and only one */
1052 static struct bus_type  hv_bus = {
1053         .name =         "vmbus",
1054         .match =                vmbus_match,
1055         .shutdown =             vmbus_shutdown,
1056         .remove =               vmbus_remove,
1057         .probe =                vmbus_probe,
1058         .uevent =               vmbus_uevent,
1059         .dma_configure =        vmbus_dma_configure,
1060         .dev_groups =           vmbus_dev_groups,
1061         .drv_groups =           vmbus_drv_groups,
1062         .bus_groups =           vmbus_bus_groups,
1063         .pm =                   &vmbus_pm,
1064 };
1065
1066 struct onmessage_work_context {
1067         struct work_struct work;
1068         struct {
1069                 struct hv_message_header header;
1070                 u8 payload[];
1071         } msg;
1072 };
1073
1074 static void vmbus_onmessage_work(struct work_struct *work)
1075 {
1076         struct onmessage_work_context *ctx;
1077
1078         /* Do not process messages if we're in DISCONNECTED state */
1079         if (vmbus_connection.conn_state == DISCONNECTED)
1080                 return;
1081
1082         ctx = container_of(work, struct onmessage_work_context,
1083                            work);
1084         vmbus_onmessage((struct vmbus_channel_message_header *)
1085                         &ctx->msg.payload);
1086         kfree(ctx);
1087 }
1088
1089 void vmbus_on_msg_dpc(unsigned long data)
1090 {
1091         struct hv_per_cpu_context *hv_cpu = (void *)data;
1092         void *page_addr = hv_cpu->synic_message_page;
1093         struct hv_message msg_copy, *msg = (struct hv_message *)page_addr +
1094                                   VMBUS_MESSAGE_SINT;
1095         struct vmbus_channel_message_header *hdr;
1096         enum vmbus_channel_message_type msgtype;
1097         const struct vmbus_channel_message_table_entry *entry;
1098         struct onmessage_work_context *ctx;
1099         __u8 payload_size;
1100         u32 message_type;
1101
1102         /*
1103          * 'enum vmbus_channel_message_type' is supposed to always be 'u32' as
1104          * it is being used in 'struct vmbus_channel_message_header' definition
1105          * which is supposed to match hypervisor ABI.
1106          */
1107         BUILD_BUG_ON(sizeof(enum vmbus_channel_message_type) != sizeof(u32));
1108
1109         /*
1110          * Since the message is in memory shared with the host, an erroneous or
1111          * malicious Hyper-V could modify the message while vmbus_on_msg_dpc()
1112          * or individual message handlers are executing; to prevent this, copy
1113          * the message into private memory.
1114          */
1115         memcpy(&msg_copy, msg, sizeof(struct hv_message));
1116
1117         message_type = msg_copy.header.message_type;
1118         if (message_type == HVMSG_NONE)
1119                 /* no msg */
1120                 return;
1121
1122         hdr = (struct vmbus_channel_message_header *)msg_copy.u.payload;
1123         msgtype = hdr->msgtype;
1124
1125         trace_vmbus_on_msg_dpc(hdr);
1126
1127         if (msgtype >= CHANNELMSG_COUNT) {
1128                 WARN_ONCE(1, "unknown msgtype=%d\n", msgtype);
1129                 goto msg_handled;
1130         }
1131
1132         payload_size = msg_copy.header.payload_size;
1133         if (payload_size > HV_MESSAGE_PAYLOAD_BYTE_COUNT) {
1134                 WARN_ONCE(1, "payload size is too large (%d)\n", payload_size);
1135                 goto msg_handled;
1136         }
1137
1138         entry = &channel_message_table[msgtype];
1139
1140         if (!entry->message_handler)
1141                 goto msg_handled;
1142
1143         if (payload_size < entry->min_payload_len) {
1144                 WARN_ONCE(1, "message too short: msgtype=%d len=%d\n", msgtype, payload_size);
1145                 goto msg_handled;
1146         }
1147
1148         if (entry->handler_type == VMHT_BLOCKING) {
1149                 ctx = kmalloc(struct_size(ctx, msg.payload, payload_size), GFP_ATOMIC);
1150                 if (ctx == NULL)
1151                         return;
1152
1153                 INIT_WORK(&ctx->work, vmbus_onmessage_work);
1154                 ctx->msg.header = msg_copy.header;
1155                 memcpy(&ctx->msg.payload, msg_copy.u.payload, payload_size);
1156
1157                 /*
1158                  * The host can generate a rescind message while we
1159                  * may still be handling the original offer. We deal with
1160                  * this condition by relying on the synchronization provided
1161                  * by offer_in_progress and by channel_mutex.  See also the
1162                  * inline comments in vmbus_onoffer_rescind().
1163                  */
1164                 switch (msgtype) {
1165                 case CHANNELMSG_RESCIND_CHANNELOFFER:
1166                         /*
1167                          * If we are handling the rescind message;
1168                          * schedule the work on the global work queue.
1169                          *
1170                          * The OFFER message and the RESCIND message should
1171                          * not be handled by the same serialized work queue,
1172                          * because the OFFER handler may call vmbus_open(),
1173                          * which tries to open the channel by sending an
1174                          * OPEN_CHANNEL message to the host and waits for
1175                          * the host's response; however, if the host has
1176                          * rescinded the channel before it receives the
1177                          * OPEN_CHANNEL message, the host just silently
1178                          * ignores the OPEN_CHANNEL message; as a result,
1179                          * the guest's OFFER handler hangs for ever, if we
1180                          * handle the RESCIND message in the same serialized
1181                          * work queue: the RESCIND handler can not start to
1182                          * run before the OFFER handler finishes.
1183                          */
1184                         if (vmbus_connection.ignore_any_offer_msg)
1185                                 break;
1186                         queue_work(vmbus_connection.rescind_work_queue, &ctx->work);
1187                         break;
1188
1189                 case CHANNELMSG_OFFERCHANNEL:
1190                         /*
1191                          * The host sends the offer message of a given channel
1192                          * before sending the rescind message of the same
1193                          * channel.  These messages are sent to the guest's
1194                          * connect CPU; the guest then starts processing them
1195                          * in the tasklet handler on this CPU:
1196                          *
1197                          * VMBUS_CONNECT_CPU
1198                          *
1199                          * [vmbus_on_msg_dpc()]
1200                          * atomic_inc()  // CHANNELMSG_OFFERCHANNEL
1201                          * queue_work()
1202                          * ...
1203                          * [vmbus_on_msg_dpc()]
1204                          * schedule_work()  // CHANNELMSG_RESCIND_CHANNELOFFER
1205                          *
1206                          * We rely on the memory-ordering properties of the
1207                          * queue_work() and schedule_work() primitives, which
1208                          * guarantee that the atomic increment will be visible
1209                          * to the CPUs which will execute the offer & rescind
1210                          * works by the time these works will start execution.
1211                          */
1212                         if (vmbus_connection.ignore_any_offer_msg)
1213                                 break;
1214                         atomic_inc(&vmbus_connection.offer_in_progress);
1215                         fallthrough;
1216
1217                 default:
1218                         queue_work(vmbus_connection.work_queue, &ctx->work);
1219                 }
1220         } else
1221                 entry->message_handler(hdr);
1222
1223 msg_handled:
1224         vmbus_signal_eom(msg, message_type);
1225 }
1226
1227 #ifdef CONFIG_PM_SLEEP
1228 /*
1229  * Fake RESCIND_CHANNEL messages to clean up hv_sock channels by force for
1230  * hibernation, because hv_sock connections can not persist across hibernation.
1231  */
1232 static void vmbus_force_channel_rescinded(struct vmbus_channel *channel)
1233 {
1234         struct onmessage_work_context *ctx;
1235         struct vmbus_channel_rescind_offer *rescind;
1236
1237         WARN_ON(!is_hvsock_channel(channel));
1238
1239         /*
1240          * Allocation size is small and the allocation should really not fail,
1241          * otherwise the state of the hv_sock connections ends up in limbo.
1242          */
1243         ctx = kzalloc(sizeof(*ctx) + sizeof(*rescind),
1244                       GFP_KERNEL | __GFP_NOFAIL);
1245
1246         /*
1247          * So far, these are not really used by Linux. Just set them to the
1248          * reasonable values conforming to the definitions of the fields.
1249          */
1250         ctx->msg.header.message_type = 1;
1251         ctx->msg.header.payload_size = sizeof(*rescind);
1252
1253         /* These values are actually used by Linux. */
1254         rescind = (struct vmbus_channel_rescind_offer *)ctx->msg.payload;
1255         rescind->header.msgtype = CHANNELMSG_RESCIND_CHANNELOFFER;
1256         rescind->child_relid = channel->offermsg.child_relid;
1257
1258         INIT_WORK(&ctx->work, vmbus_onmessage_work);
1259
1260         queue_work(vmbus_connection.work_queue, &ctx->work);
1261 }
1262 #endif /* CONFIG_PM_SLEEP */
1263
1264 /*
1265  * Schedule all channels with events pending
1266  */
1267 static void vmbus_chan_sched(struct hv_per_cpu_context *hv_cpu)
1268 {
1269         unsigned long *recv_int_page;
1270         u32 maxbits, relid;
1271
1272         /*
1273          * The event page can be directly checked to get the id of
1274          * the channel that has the interrupt pending.
1275          */
1276         void *page_addr = hv_cpu->synic_event_page;
1277         union hv_synic_event_flags *event
1278                 = (union hv_synic_event_flags *)page_addr +
1279                                          VMBUS_MESSAGE_SINT;
1280
1281         maxbits = HV_EVENT_FLAGS_COUNT;
1282         recv_int_page = event->flags;
1283
1284         if (unlikely(!recv_int_page))
1285                 return;
1286
1287         for_each_set_bit(relid, recv_int_page, maxbits) {
1288                 void (*callback_fn)(void *context);
1289                 struct vmbus_channel *channel;
1290
1291                 if (!sync_test_and_clear_bit(relid, recv_int_page))
1292                         continue;
1293
1294                 /* Special case - vmbus channel protocol msg */
1295                 if (relid == 0)
1296                         continue;
1297
1298                 /*
1299                  * Pairs with the kfree_rcu() in vmbus_chan_release().
1300                  * Guarantees that the channel data structure doesn't
1301                  * get freed while the channel pointer below is being
1302                  * dereferenced.
1303                  */
1304                 rcu_read_lock();
1305
1306                 /* Find channel based on relid */
1307                 channel = relid2channel(relid);
1308                 if (channel == NULL)
1309                         goto sched_unlock_rcu;
1310
1311                 if (channel->rescind)
1312                         goto sched_unlock_rcu;
1313
1314                 /*
1315                  * Make sure that the ring buffer data structure doesn't get
1316                  * freed while we dereference the ring buffer pointer.  Test
1317                  * for the channel's onchannel_callback being NULL within a
1318                  * sched_lock critical section.  See also the inline comments
1319                  * in vmbus_reset_channel_cb().
1320                  */
1321                 spin_lock(&channel->sched_lock);
1322
1323                 callback_fn = channel->onchannel_callback;
1324                 if (unlikely(callback_fn == NULL))
1325                         goto sched_unlock;
1326
1327                 trace_vmbus_chan_sched(channel);
1328
1329                 ++channel->interrupts;
1330
1331                 switch (channel->callback_mode) {
1332                 case HV_CALL_ISR:
1333                         (*callback_fn)(channel->channel_callback_context);
1334                         break;
1335
1336                 case HV_CALL_BATCHED:
1337                         hv_begin_read(&channel->inbound);
1338                         fallthrough;
1339                 case HV_CALL_DIRECT:
1340                         tasklet_schedule(&channel->callback_event);
1341                 }
1342
1343 sched_unlock:
1344                 spin_unlock(&channel->sched_lock);
1345 sched_unlock_rcu:
1346                 rcu_read_unlock();
1347         }
1348 }
1349
1350 static void vmbus_isr(void)
1351 {
1352         struct hv_per_cpu_context *hv_cpu
1353                 = this_cpu_ptr(hv_context.cpu_context);
1354         void *page_addr;
1355         struct hv_message *msg;
1356
1357         vmbus_chan_sched(hv_cpu);
1358
1359         page_addr = hv_cpu->synic_message_page;
1360         msg = (struct hv_message *)page_addr + VMBUS_MESSAGE_SINT;
1361
1362         /* Check if there are actual msgs to be processed */
1363         if (msg->header.message_type != HVMSG_NONE) {
1364                 if (msg->header.message_type == HVMSG_TIMER_EXPIRED) {
1365                         hv_stimer0_isr();
1366                         vmbus_signal_eom(msg, HVMSG_TIMER_EXPIRED);
1367                 } else
1368                         tasklet_schedule(&hv_cpu->msg_dpc);
1369         }
1370
1371         add_interrupt_randomness(vmbus_interrupt);
1372 }
1373
1374 static irqreturn_t vmbus_percpu_isr(int irq, void *dev_id)
1375 {
1376         vmbus_isr();
1377         return IRQ_HANDLED;
1378 }
1379
1380 /*
1381  * Callback from kmsg_dump. Grab as much as possible from the end of the kmsg
1382  * buffer and call into Hyper-V to transfer the data.
1383  */
1384 static void hv_kmsg_dump(struct kmsg_dumper *dumper,
1385                          enum kmsg_dump_reason reason)
1386 {
1387         struct kmsg_dump_iter iter;
1388         size_t bytes_written;
1389
1390         /* We are only interested in panics. */
1391         if ((reason != KMSG_DUMP_PANIC) || (!sysctl_record_panic_msg))
1392                 return;
1393
1394         /*
1395          * Write dump contents to the page. No need to synchronize; panic should
1396          * be single-threaded.
1397          */
1398         kmsg_dump_rewind(&iter);
1399         kmsg_dump_get_buffer(&iter, false, hv_panic_page, HV_HYP_PAGE_SIZE,
1400                              &bytes_written);
1401         if (!bytes_written)
1402                 return;
1403         /*
1404          * P3 to contain the physical address of the panic page & P4 to
1405          * contain the size of the panic data in that page. Rest of the
1406          * registers are no-op when the NOTIFY_MSG flag is set.
1407          */
1408         hv_set_register(HV_REGISTER_CRASH_P0, 0);
1409         hv_set_register(HV_REGISTER_CRASH_P1, 0);
1410         hv_set_register(HV_REGISTER_CRASH_P2, 0);
1411         hv_set_register(HV_REGISTER_CRASH_P3, virt_to_phys(hv_panic_page));
1412         hv_set_register(HV_REGISTER_CRASH_P4, bytes_written);
1413
1414         /*
1415          * Let Hyper-V know there is crash data available along with
1416          * the panic message.
1417          */
1418         hv_set_register(HV_REGISTER_CRASH_CTL,
1419                (HV_CRASH_CTL_CRASH_NOTIFY | HV_CRASH_CTL_CRASH_NOTIFY_MSG));
1420 }
1421
1422 static struct kmsg_dumper hv_kmsg_dumper = {
1423         .dump = hv_kmsg_dump,
1424 };
1425
1426 static void hv_kmsg_dump_register(void)
1427 {
1428         int ret;
1429
1430         hv_panic_page = hv_alloc_hyperv_zeroed_page();
1431         if (!hv_panic_page) {
1432                 pr_err("Hyper-V: panic message page memory allocation failed\n");
1433                 return;
1434         }
1435
1436         ret = kmsg_dump_register(&hv_kmsg_dumper);
1437         if (ret) {
1438                 pr_err("Hyper-V: kmsg dump register error 0x%x\n", ret);
1439                 hv_free_hyperv_page((unsigned long)hv_panic_page);
1440                 hv_panic_page = NULL;
1441         }
1442 }
1443
1444 static struct ctl_table_header *hv_ctl_table_hdr;
1445
1446 /*
1447  * sysctl option to allow the user to control whether kmsg data should be
1448  * reported to Hyper-V on panic.
1449  */
1450 static struct ctl_table hv_ctl_table[] = {
1451         {
1452                 .procname       = "hyperv_record_panic_msg",
1453                 .data           = &sysctl_record_panic_msg,
1454                 .maxlen         = sizeof(int),
1455                 .mode           = 0644,
1456                 .proc_handler   = proc_dointvec_minmax,
1457                 .extra1         = SYSCTL_ZERO,
1458                 .extra2         = SYSCTL_ONE
1459         },
1460         {}
1461 };
1462
1463 static struct ctl_table hv_root_table[] = {
1464         {
1465                 .procname       = "kernel",
1466                 .mode           = 0555,
1467                 .child          = hv_ctl_table
1468         },
1469         {}
1470 };
1471
1472 /*
1473  * vmbus_bus_init -Main vmbus driver initialization routine.
1474  *
1475  * Here, we
1476  *      - initialize the vmbus driver context
1477  *      - invoke the vmbus hv main init routine
1478  *      - retrieve the channel offers
1479  */
1480 static int vmbus_bus_init(void)
1481 {
1482         int ret;
1483
1484         ret = hv_init();
1485         if (ret != 0) {
1486                 pr_err("Unable to initialize the hypervisor - 0x%x\n", ret);
1487                 return ret;
1488         }
1489
1490         ret = bus_register(&hv_bus);
1491         if (ret)
1492                 return ret;
1493
1494         /*
1495          * VMbus interrupts are best modeled as per-cpu interrupts. If
1496          * on an architecture with support for per-cpu IRQs (e.g. ARM64),
1497          * allocate a per-cpu IRQ using standard Linux kernel functionality.
1498          * If not on such an architecture (e.g., x86/x64), then rely on
1499          * code in the arch-specific portion of the code tree to connect
1500          * the VMbus interrupt handler.
1501          */
1502
1503         if (vmbus_irq == -1) {
1504                 hv_setup_vmbus_handler(vmbus_isr);
1505         } else {
1506                 vmbus_evt = alloc_percpu(long);
1507                 ret = request_percpu_irq(vmbus_irq, vmbus_percpu_isr,
1508                                 "Hyper-V VMbus", vmbus_evt);
1509                 if (ret) {
1510                         pr_err("Can't request Hyper-V VMbus IRQ %d, Err %d",
1511                                         vmbus_irq, ret);
1512                         free_percpu(vmbus_evt);
1513                         goto err_setup;
1514                 }
1515         }
1516
1517         ret = hv_synic_alloc();
1518         if (ret)
1519                 goto err_alloc;
1520
1521         /*
1522          * Initialize the per-cpu interrupt state and stimer state.
1523          * Then connect to the host.
1524          */
1525         ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "hyperv/vmbus:online",
1526                                 hv_synic_init, hv_synic_cleanup);
1527         if (ret < 0)
1528                 goto err_cpuhp;
1529         hyperv_cpuhp_online = ret;
1530
1531         ret = vmbus_connect();
1532         if (ret)
1533                 goto err_connect;
1534
1535         if (hv_is_isolation_supported())
1536                 sysctl_record_panic_msg = 0;
1537
1538         /*
1539          * Only register if the crash MSRs are available
1540          */
1541         if (ms_hyperv.misc_features & HV_FEATURE_GUEST_CRASH_MSR_AVAILABLE) {
1542                 u64 hyperv_crash_ctl;
1543                 /*
1544                  * Panic message recording (sysctl_record_panic_msg)
1545                  * is enabled by default in non-isolated guests and
1546                  * disabled by default in isolated guests; the panic
1547                  * message recording won't be available in isolated
1548                  * guests should the following registration fail.
1549                  */
1550                 hv_ctl_table_hdr = register_sysctl_table(hv_root_table);
1551                 if (!hv_ctl_table_hdr)
1552                         pr_err("Hyper-V: sysctl table register error");
1553
1554                 /*
1555                  * Register for panic kmsg callback only if the right
1556                  * capability is supported by the hypervisor.
1557                  */
1558                 hyperv_crash_ctl = hv_get_register(HV_REGISTER_CRASH_CTL);
1559                 if (hyperv_crash_ctl & HV_CRASH_CTL_CRASH_NOTIFY_MSG)
1560                         hv_kmsg_dump_register();
1561
1562                 register_die_notifier(&hyperv_die_report_block);
1563                 atomic_notifier_chain_register(&panic_notifier_list,
1564                                                 &hyperv_panic_report_block);
1565         }
1566
1567         /*
1568          * Always register the vmbus unload panic notifier because we
1569          * need to shut the VMbus channel connection on panic.
1570          */
1571         atomic_notifier_chain_register(&panic_notifier_list,
1572                                &hyperv_panic_vmbus_unload_block);
1573
1574         vmbus_request_offers();
1575
1576         return 0;
1577
1578 err_connect:
1579         cpuhp_remove_state(hyperv_cpuhp_online);
1580 err_cpuhp:
1581         hv_synic_free();
1582 err_alloc:
1583         if (vmbus_irq == -1) {
1584                 hv_remove_vmbus_handler();
1585         } else {
1586                 free_percpu_irq(vmbus_irq, vmbus_evt);
1587                 free_percpu(vmbus_evt);
1588         }
1589 err_setup:
1590         bus_unregister(&hv_bus);
1591         unregister_sysctl_table(hv_ctl_table_hdr);
1592         hv_ctl_table_hdr = NULL;
1593         return ret;
1594 }
1595
1596 /**
1597  * __vmbus_driver_register() - Register a vmbus's driver
1598  * @hv_driver: Pointer to driver structure you want to register
1599  * @owner: owner module of the drv
1600  * @mod_name: module name string
1601  *
1602  * Registers the given driver with Linux through the 'driver_register()' call
1603  * and sets up the hyper-v vmbus handling for this driver.
1604  * It will return the state of the 'driver_register()' call.
1605  *
1606  */
1607 int __vmbus_driver_register(struct hv_driver *hv_driver, struct module *owner, const char *mod_name)
1608 {
1609         int ret;
1610
1611         pr_info("registering driver %s\n", hv_driver->name);
1612
1613         ret = vmbus_exists();
1614         if (ret < 0)
1615                 return ret;
1616
1617         hv_driver->driver.name = hv_driver->name;
1618         hv_driver->driver.owner = owner;
1619         hv_driver->driver.mod_name = mod_name;
1620         hv_driver->driver.bus = &hv_bus;
1621
1622         spin_lock_init(&hv_driver->dynids.lock);
1623         INIT_LIST_HEAD(&hv_driver->dynids.list);
1624
1625         ret = driver_register(&hv_driver->driver);
1626
1627         return ret;
1628 }
1629 EXPORT_SYMBOL_GPL(__vmbus_driver_register);
1630
1631 /**
1632  * vmbus_driver_unregister() - Unregister a vmbus's driver
1633  * @hv_driver: Pointer to driver structure you want to
1634  *             un-register
1635  *
1636  * Un-register the given driver that was previous registered with a call to
1637  * vmbus_driver_register()
1638  */
1639 void vmbus_driver_unregister(struct hv_driver *hv_driver)
1640 {
1641         pr_info("unregistering driver %s\n", hv_driver->name);
1642
1643         if (!vmbus_exists()) {
1644                 driver_unregister(&hv_driver->driver);
1645                 vmbus_free_dynids(hv_driver);
1646         }
1647 }
1648 EXPORT_SYMBOL_GPL(vmbus_driver_unregister);
1649
1650
1651 /*
1652  * Called when last reference to channel is gone.
1653  */
1654 static void vmbus_chan_release(struct kobject *kobj)
1655 {
1656         struct vmbus_channel *channel
1657                 = container_of(kobj, struct vmbus_channel, kobj);
1658
1659         kfree_rcu(channel, rcu);
1660 }
1661
1662 struct vmbus_chan_attribute {
1663         struct attribute attr;
1664         ssize_t (*show)(struct vmbus_channel *chan, char *buf);
1665         ssize_t (*store)(struct vmbus_channel *chan,
1666                          const char *buf, size_t count);
1667 };
1668 #define VMBUS_CHAN_ATTR(_name, _mode, _show, _store) \
1669         struct vmbus_chan_attribute chan_attr_##_name \
1670                 = __ATTR(_name, _mode, _show, _store)
1671 #define VMBUS_CHAN_ATTR_RW(_name) \
1672         struct vmbus_chan_attribute chan_attr_##_name = __ATTR_RW(_name)
1673 #define VMBUS_CHAN_ATTR_RO(_name) \
1674         struct vmbus_chan_attribute chan_attr_##_name = __ATTR_RO(_name)
1675 #define VMBUS_CHAN_ATTR_WO(_name) \
1676         struct vmbus_chan_attribute chan_attr_##_name = __ATTR_WO(_name)
1677
1678 static ssize_t vmbus_chan_attr_show(struct kobject *kobj,
1679                                     struct attribute *attr, char *buf)
1680 {
1681         const struct vmbus_chan_attribute *attribute
1682                 = container_of(attr, struct vmbus_chan_attribute, attr);
1683         struct vmbus_channel *chan
1684                 = container_of(kobj, struct vmbus_channel, kobj);
1685
1686         if (!attribute->show)
1687                 return -EIO;
1688
1689         return attribute->show(chan, buf);
1690 }
1691
1692 static ssize_t vmbus_chan_attr_store(struct kobject *kobj,
1693                                      struct attribute *attr, const char *buf,
1694                                      size_t count)
1695 {
1696         const struct vmbus_chan_attribute *attribute
1697                 = container_of(attr, struct vmbus_chan_attribute, attr);
1698         struct vmbus_channel *chan
1699                 = container_of(kobj, struct vmbus_channel, kobj);
1700
1701         if (!attribute->store)
1702                 return -EIO;
1703
1704         return attribute->store(chan, buf, count);
1705 }
1706
1707 static const struct sysfs_ops vmbus_chan_sysfs_ops = {
1708         .show = vmbus_chan_attr_show,
1709         .store = vmbus_chan_attr_store,
1710 };
1711
1712 static ssize_t out_mask_show(struct vmbus_channel *channel, char *buf)
1713 {
1714         struct hv_ring_buffer_info *rbi = &channel->outbound;
1715         ssize_t ret;
1716
1717         mutex_lock(&rbi->ring_buffer_mutex);
1718         if (!rbi->ring_buffer) {
1719                 mutex_unlock(&rbi->ring_buffer_mutex);
1720                 return -EINVAL;
1721         }
1722
1723         ret = sprintf(buf, "%u\n", rbi->ring_buffer->interrupt_mask);
1724         mutex_unlock(&rbi->ring_buffer_mutex);
1725         return ret;
1726 }
1727 static VMBUS_CHAN_ATTR_RO(out_mask);
1728
1729 static ssize_t in_mask_show(struct vmbus_channel *channel, char *buf)
1730 {
1731         struct hv_ring_buffer_info *rbi = &channel->inbound;
1732         ssize_t ret;
1733
1734         mutex_lock(&rbi->ring_buffer_mutex);
1735         if (!rbi->ring_buffer) {
1736                 mutex_unlock(&rbi->ring_buffer_mutex);
1737                 return -EINVAL;
1738         }
1739
1740         ret = sprintf(buf, "%u\n", rbi->ring_buffer->interrupt_mask);
1741         mutex_unlock(&rbi->ring_buffer_mutex);
1742         return ret;
1743 }
1744 static VMBUS_CHAN_ATTR_RO(in_mask);
1745
1746 static ssize_t read_avail_show(struct vmbus_channel *channel, char *buf)
1747 {
1748         struct hv_ring_buffer_info *rbi = &channel->inbound;
1749         ssize_t ret;
1750
1751         mutex_lock(&rbi->ring_buffer_mutex);
1752         if (!rbi->ring_buffer) {
1753                 mutex_unlock(&rbi->ring_buffer_mutex);
1754                 return -EINVAL;
1755         }
1756
1757         ret = sprintf(buf, "%u\n", hv_get_bytes_to_read(rbi));
1758         mutex_unlock(&rbi->ring_buffer_mutex);
1759         return ret;
1760 }
1761 static VMBUS_CHAN_ATTR_RO(read_avail);
1762
1763 static ssize_t write_avail_show(struct vmbus_channel *channel, char *buf)
1764 {
1765         struct hv_ring_buffer_info *rbi = &channel->outbound;
1766         ssize_t ret;
1767
1768         mutex_lock(&rbi->ring_buffer_mutex);
1769         if (!rbi->ring_buffer) {
1770                 mutex_unlock(&rbi->ring_buffer_mutex);
1771                 return -EINVAL;
1772         }
1773
1774         ret = sprintf(buf, "%u\n", hv_get_bytes_to_write(rbi));
1775         mutex_unlock(&rbi->ring_buffer_mutex);
1776         return ret;
1777 }
1778 static VMBUS_CHAN_ATTR_RO(write_avail);
1779
1780 static ssize_t target_cpu_show(struct vmbus_channel *channel, char *buf)
1781 {
1782         return sprintf(buf, "%u\n", channel->target_cpu);
1783 }
1784 static ssize_t target_cpu_store(struct vmbus_channel *channel,
1785                                 const char *buf, size_t count)
1786 {
1787         u32 target_cpu, origin_cpu;
1788         ssize_t ret = count;
1789
1790         if (vmbus_proto_version < VERSION_WIN10_V4_1)
1791                 return -EIO;
1792
1793         if (sscanf(buf, "%uu", &target_cpu) != 1)
1794                 return -EIO;
1795
1796         /* Validate target_cpu for the cpumask_test_cpu() operation below. */
1797         if (target_cpu >= nr_cpumask_bits)
1798                 return -EINVAL;
1799
1800         if (!cpumask_test_cpu(target_cpu, housekeeping_cpumask(HK_TYPE_MANAGED_IRQ)))
1801                 return -EINVAL;
1802
1803         /* No CPUs should come up or down during this. */
1804         cpus_read_lock();
1805
1806         if (!cpu_online(target_cpu)) {
1807                 cpus_read_unlock();
1808                 return -EINVAL;
1809         }
1810
1811         /*
1812          * Synchronizes target_cpu_store() and channel closure:
1813          *
1814          * { Initially: state = CHANNEL_OPENED }
1815          *
1816          * CPU1                         CPU2
1817          *
1818          * [target_cpu_store()]         [vmbus_disconnect_ring()]
1819          *
1820          * LOCK channel_mutex           LOCK channel_mutex
1821          * LOAD r1 = state              LOAD r2 = state
1822          * IF (r1 == CHANNEL_OPENED)    IF (r2 == CHANNEL_OPENED)
1823          *   SEND MODIFYCHANNEL           STORE state = CHANNEL_OPEN
1824          *   [...]                        SEND CLOSECHANNEL
1825          * UNLOCK channel_mutex         UNLOCK channel_mutex
1826          *
1827          * Forbids: r1 == r2 == CHANNEL_OPENED (i.e., CPU1's LOCK precedes
1828          *              CPU2's LOCK) && CPU2's SEND precedes CPU1's SEND
1829          *
1830          * Note.  The host processes the channel messages "sequentially", in
1831          * the order in which they are received on a per-partition basis.
1832          */
1833         mutex_lock(&vmbus_connection.channel_mutex);
1834
1835         /*
1836          * Hyper-V will ignore MODIFYCHANNEL messages for "non-open" channels;
1837          * avoid sending the message and fail here for such channels.
1838          */
1839         if (channel->state != CHANNEL_OPENED_STATE) {
1840                 ret = -EIO;
1841                 goto cpu_store_unlock;
1842         }
1843
1844         origin_cpu = channel->target_cpu;
1845         if (target_cpu == origin_cpu)
1846                 goto cpu_store_unlock;
1847
1848         if (vmbus_send_modifychannel(channel,
1849                                      hv_cpu_number_to_vp_number(target_cpu))) {
1850                 ret = -EIO;
1851                 goto cpu_store_unlock;
1852         }
1853
1854         /*
1855          * For version before VERSION_WIN10_V5_3, the following warning holds:
1856          *
1857          * Warning.  At this point, there is *no* guarantee that the host will
1858          * have successfully processed the vmbus_send_modifychannel() request.
1859          * See the header comment of vmbus_send_modifychannel() for more info.
1860          *
1861          * Lags in the processing of the above vmbus_send_modifychannel() can
1862          * result in missed interrupts if the "old" target CPU is taken offline
1863          * before Hyper-V starts sending interrupts to the "new" target CPU.
1864          * But apart from this offlining scenario, the code tolerates such
1865          * lags.  It will function correctly even if a channel interrupt comes
1866          * in on a CPU that is different from the channel target_cpu value.
1867          */
1868
1869         channel->target_cpu = target_cpu;
1870
1871         /* See init_vp_index(). */
1872         if (hv_is_perf_channel(channel))
1873                 hv_update_allocated_cpus(origin_cpu, target_cpu);
1874
1875         /* Currently set only for storvsc channels. */
1876         if (channel->change_target_cpu_callback) {
1877                 (*channel->change_target_cpu_callback)(channel,
1878                                 origin_cpu, target_cpu);
1879         }
1880
1881 cpu_store_unlock:
1882         mutex_unlock(&vmbus_connection.channel_mutex);
1883         cpus_read_unlock();
1884         return ret;
1885 }
1886 static VMBUS_CHAN_ATTR(cpu, 0644, target_cpu_show, target_cpu_store);
1887
1888 static ssize_t channel_pending_show(struct vmbus_channel *channel,
1889                                     char *buf)
1890 {
1891         return sprintf(buf, "%d\n",
1892                        channel_pending(channel,
1893                                        vmbus_connection.monitor_pages[1]));
1894 }
1895 static VMBUS_CHAN_ATTR(pending, 0444, channel_pending_show, NULL);
1896
1897 static ssize_t channel_latency_show(struct vmbus_channel *channel,
1898                                     char *buf)
1899 {
1900         return sprintf(buf, "%d\n",
1901                        channel_latency(channel,
1902                                        vmbus_connection.monitor_pages[1]));
1903 }
1904 static VMBUS_CHAN_ATTR(latency, 0444, channel_latency_show, NULL);
1905
1906 static ssize_t channel_interrupts_show(struct vmbus_channel *channel, char *buf)
1907 {
1908         return sprintf(buf, "%llu\n", channel->interrupts);
1909 }
1910 static VMBUS_CHAN_ATTR(interrupts, 0444, channel_interrupts_show, NULL);
1911
1912 static ssize_t channel_events_show(struct vmbus_channel *channel, char *buf)
1913 {
1914         return sprintf(buf, "%llu\n", channel->sig_events);
1915 }
1916 static VMBUS_CHAN_ATTR(events, 0444, channel_events_show, NULL);
1917
1918 static ssize_t channel_intr_in_full_show(struct vmbus_channel *channel,
1919                                          char *buf)
1920 {
1921         return sprintf(buf, "%llu\n",
1922                        (unsigned long long)channel->intr_in_full);
1923 }
1924 static VMBUS_CHAN_ATTR(intr_in_full, 0444, channel_intr_in_full_show, NULL);
1925
1926 static ssize_t channel_intr_out_empty_show(struct vmbus_channel *channel,
1927                                            char *buf)
1928 {
1929         return sprintf(buf, "%llu\n",
1930                        (unsigned long long)channel->intr_out_empty);
1931 }
1932 static VMBUS_CHAN_ATTR(intr_out_empty, 0444, channel_intr_out_empty_show, NULL);
1933
1934 static ssize_t channel_out_full_first_show(struct vmbus_channel *channel,
1935                                            char *buf)
1936 {
1937         return sprintf(buf, "%llu\n",
1938                        (unsigned long long)channel->out_full_first);
1939 }
1940 static VMBUS_CHAN_ATTR(out_full_first, 0444, channel_out_full_first_show, NULL);
1941
1942 static ssize_t channel_out_full_total_show(struct vmbus_channel *channel,
1943                                            char *buf)
1944 {
1945         return sprintf(buf, "%llu\n",
1946                        (unsigned long long)channel->out_full_total);
1947 }
1948 static VMBUS_CHAN_ATTR(out_full_total, 0444, channel_out_full_total_show, NULL);
1949
1950 static ssize_t subchannel_monitor_id_show(struct vmbus_channel *channel,
1951                                           char *buf)
1952 {
1953         return sprintf(buf, "%u\n", channel->offermsg.monitorid);
1954 }
1955 static VMBUS_CHAN_ATTR(monitor_id, 0444, subchannel_monitor_id_show, NULL);
1956
1957 static ssize_t subchannel_id_show(struct vmbus_channel *channel,
1958                                   char *buf)
1959 {
1960         return sprintf(buf, "%u\n",
1961                        channel->offermsg.offer.sub_channel_index);
1962 }
1963 static VMBUS_CHAN_ATTR_RO(subchannel_id);
1964
1965 static struct attribute *vmbus_chan_attrs[] = {
1966         &chan_attr_out_mask.attr,
1967         &chan_attr_in_mask.attr,
1968         &chan_attr_read_avail.attr,
1969         &chan_attr_write_avail.attr,
1970         &chan_attr_cpu.attr,
1971         &chan_attr_pending.attr,
1972         &chan_attr_latency.attr,
1973         &chan_attr_interrupts.attr,
1974         &chan_attr_events.attr,
1975         &chan_attr_intr_in_full.attr,
1976         &chan_attr_intr_out_empty.attr,
1977         &chan_attr_out_full_first.attr,
1978         &chan_attr_out_full_total.attr,
1979         &chan_attr_monitor_id.attr,
1980         &chan_attr_subchannel_id.attr,
1981         NULL
1982 };
1983
1984 /*
1985  * Channel-level attribute_group callback function. Returns the permission for
1986  * each attribute, and returns 0 if an attribute is not visible.
1987  */
1988 static umode_t vmbus_chan_attr_is_visible(struct kobject *kobj,
1989                                           struct attribute *attr, int idx)
1990 {
1991         const struct vmbus_channel *channel =
1992                 container_of(kobj, struct vmbus_channel, kobj);
1993
1994         /* Hide the monitor attributes if the monitor mechanism is not used. */
1995         if (!channel->offermsg.monitor_allocated &&
1996             (attr == &chan_attr_pending.attr ||
1997              attr == &chan_attr_latency.attr ||
1998              attr == &chan_attr_monitor_id.attr))
1999                 return 0;
2000
2001         return attr->mode;
2002 }
2003
2004 static struct attribute_group vmbus_chan_group = {
2005         .attrs = vmbus_chan_attrs,
2006         .is_visible = vmbus_chan_attr_is_visible
2007 };
2008
2009 static struct kobj_type vmbus_chan_ktype = {
2010         .sysfs_ops = &vmbus_chan_sysfs_ops,
2011         .release = vmbus_chan_release,
2012 };
2013
2014 /*
2015  * vmbus_add_channel_kobj - setup a sub-directory under device/channels
2016  */
2017 int vmbus_add_channel_kobj(struct hv_device *dev, struct vmbus_channel *channel)
2018 {
2019         const struct device *device = &dev->device;
2020         struct kobject *kobj = &channel->kobj;
2021         u32 relid = channel->offermsg.child_relid;
2022         int ret;
2023
2024         kobj->kset = dev->channels_kset;
2025         ret = kobject_init_and_add(kobj, &vmbus_chan_ktype, NULL,
2026                                    "%u", relid);
2027         if (ret) {
2028                 kobject_put(kobj);
2029                 return ret;
2030         }
2031
2032         ret = sysfs_create_group(kobj, &vmbus_chan_group);
2033
2034         if (ret) {
2035                 /*
2036                  * The calling functions' error handling paths will cleanup the
2037                  * empty channel directory.
2038                  */
2039                 kobject_put(kobj);
2040                 dev_err(device, "Unable to set up channel sysfs files\n");
2041                 return ret;
2042         }
2043
2044         kobject_uevent(kobj, KOBJ_ADD);
2045
2046         return 0;
2047 }
2048
2049 /*
2050  * vmbus_remove_channel_attr_group - remove the channel's attribute group
2051  */
2052 void vmbus_remove_channel_attr_group(struct vmbus_channel *channel)
2053 {
2054         sysfs_remove_group(&channel->kobj, &vmbus_chan_group);
2055 }
2056
2057 /*
2058  * vmbus_device_create - Creates and registers a new child device
2059  * on the vmbus.
2060  */
2061 struct hv_device *vmbus_device_create(const guid_t *type,
2062                                       const guid_t *instance,
2063                                       struct vmbus_channel *channel)
2064 {
2065         struct hv_device *child_device_obj;
2066
2067         child_device_obj = kzalloc(sizeof(struct hv_device), GFP_KERNEL);
2068         if (!child_device_obj) {
2069                 pr_err("Unable to allocate device object for child device\n");
2070                 return NULL;
2071         }
2072
2073         child_device_obj->channel = channel;
2074         guid_copy(&child_device_obj->dev_type, type);
2075         guid_copy(&child_device_obj->dev_instance, instance);
2076         child_device_obj->vendor_id = PCI_VENDOR_ID_MICROSOFT;
2077
2078         return child_device_obj;
2079 }
2080
2081 /*
2082  * vmbus_device_register - Register the child device
2083  */
2084 int vmbus_device_register(struct hv_device *child_device_obj)
2085 {
2086         struct kobject *kobj = &child_device_obj->device.kobj;
2087         int ret;
2088
2089         dev_set_name(&child_device_obj->device, "%pUl",
2090                      &child_device_obj->channel->offermsg.offer.if_instance);
2091
2092         child_device_obj->device.bus = &hv_bus;
2093         child_device_obj->device.parent = &hv_acpi_dev->dev;
2094         child_device_obj->device.release = vmbus_device_release;
2095
2096         child_device_obj->device.dma_parms = &child_device_obj->dma_parms;
2097         child_device_obj->device.dma_mask = &child_device_obj->dma_mask;
2098         dma_set_mask(&child_device_obj->device, DMA_BIT_MASK(64));
2099
2100         /*
2101          * Register with the LDM. This will kick off the driver/device
2102          * binding...which will eventually call vmbus_match() and vmbus_probe()
2103          */
2104         ret = device_register(&child_device_obj->device);
2105         if (ret) {
2106                 pr_err("Unable to register child device\n");
2107                 put_device(&child_device_obj->device);
2108                 return ret;
2109         }
2110
2111         child_device_obj->channels_kset = kset_create_and_add("channels",
2112                                                               NULL, kobj);
2113         if (!child_device_obj->channels_kset) {
2114                 ret = -ENOMEM;
2115                 goto err_dev_unregister;
2116         }
2117
2118         ret = vmbus_add_channel_kobj(child_device_obj,
2119                                      child_device_obj->channel);
2120         if (ret) {
2121                 pr_err("Unable to register primary channeln");
2122                 goto err_kset_unregister;
2123         }
2124         hv_debug_add_dev_dir(child_device_obj);
2125
2126         return 0;
2127
2128 err_kset_unregister:
2129         kset_unregister(child_device_obj->channels_kset);
2130
2131 err_dev_unregister:
2132         device_unregister(&child_device_obj->device);
2133         return ret;
2134 }
2135
2136 /*
2137  * vmbus_device_unregister - Remove the specified child device
2138  * from the vmbus.
2139  */
2140 void vmbus_device_unregister(struct hv_device *device_obj)
2141 {
2142         pr_debug("child device %s unregistered\n",
2143                 dev_name(&device_obj->device));
2144
2145         kset_unregister(device_obj->channels_kset);
2146
2147         /*
2148          * Kick off the process of unregistering the device.
2149          * This will call vmbus_remove() and eventually vmbus_device_release()
2150          */
2151         device_unregister(&device_obj->device);
2152 }
2153
2154
2155 /*
2156  * VMBUS is an acpi enumerated device. Get the information we
2157  * need from DSDT.
2158  */
2159 #define VTPM_BASE_ADDRESS 0xfed40000
2160 static acpi_status vmbus_walk_resources(struct acpi_resource *res, void *ctx)
2161 {
2162         resource_size_t start = 0;
2163         resource_size_t end = 0;
2164         struct resource *new_res;
2165         struct resource **old_res = &hyperv_mmio;
2166         struct resource **prev_res = NULL;
2167         struct resource r;
2168
2169         switch (res->type) {
2170
2171         /*
2172          * "Address" descriptors are for bus windows. Ignore
2173          * "memory" descriptors, which are for registers on
2174          * devices.
2175          */
2176         case ACPI_RESOURCE_TYPE_ADDRESS32:
2177                 start = res->data.address32.address.minimum;
2178                 end = res->data.address32.address.maximum;
2179                 break;
2180
2181         case ACPI_RESOURCE_TYPE_ADDRESS64:
2182                 start = res->data.address64.address.minimum;
2183                 end = res->data.address64.address.maximum;
2184                 break;
2185
2186         /*
2187          * The IRQ information is needed only on ARM64, which Hyper-V
2188          * sets up in the extended format. IRQ information is present
2189          * on x86/x64 in the non-extended format but it is not used by
2190          * Linux. So don't bother checking for the non-extended format.
2191          */
2192         case ACPI_RESOURCE_TYPE_EXTENDED_IRQ:
2193                 if (!acpi_dev_resource_interrupt(res, 0, &r)) {
2194                         pr_err("Unable to parse Hyper-V ACPI interrupt\n");
2195                         return AE_ERROR;
2196                 }
2197                 /* ARM64 INTID for VMbus */
2198                 vmbus_interrupt = res->data.extended_irq.interrupts[0];
2199                 /* Linux IRQ number */
2200                 vmbus_irq = r.start;
2201                 return AE_OK;
2202
2203         default:
2204                 /* Unused resource type */
2205                 return AE_OK;
2206
2207         }
2208         /*
2209          * Ignore ranges that are below 1MB, as they're not
2210          * necessary or useful here.
2211          */
2212         if (end < 0x100000)
2213                 return AE_OK;
2214
2215         new_res = kzalloc(sizeof(*new_res), GFP_ATOMIC);
2216         if (!new_res)
2217                 return AE_NO_MEMORY;
2218
2219         /* If this range overlaps the virtual TPM, truncate it. */
2220         if (end > VTPM_BASE_ADDRESS && start < VTPM_BASE_ADDRESS)
2221                 end = VTPM_BASE_ADDRESS;
2222
2223         new_res->name = "hyperv mmio";
2224         new_res->flags = IORESOURCE_MEM;
2225         new_res->start = start;
2226         new_res->end = end;
2227
2228         /*
2229          * If two ranges are adjacent, merge them.
2230          */
2231         do {
2232                 if (!*old_res) {
2233                         *old_res = new_res;
2234                         break;
2235                 }
2236
2237                 if (((*old_res)->end + 1) == new_res->start) {
2238                         (*old_res)->end = new_res->end;
2239                         kfree(new_res);
2240                         break;
2241                 }
2242
2243                 if ((*old_res)->start == new_res->end + 1) {
2244                         (*old_res)->start = new_res->start;
2245                         kfree(new_res);
2246                         break;
2247                 }
2248
2249                 if ((*old_res)->start > new_res->end) {
2250                         new_res->sibling = *old_res;
2251                         if (prev_res)
2252                                 (*prev_res)->sibling = new_res;
2253                         *old_res = new_res;
2254                         break;
2255                 }
2256
2257                 prev_res = old_res;
2258                 old_res = &(*old_res)->sibling;
2259
2260         } while (1);
2261
2262         return AE_OK;
2263 }
2264
2265 static void vmbus_acpi_remove(struct acpi_device *device)
2266 {
2267         struct resource *cur_res;
2268         struct resource *next_res;
2269
2270         if (hyperv_mmio) {
2271                 if (fb_mmio) {
2272                         __release_region(hyperv_mmio, fb_mmio->start,
2273                                          resource_size(fb_mmio));
2274                         fb_mmio = NULL;
2275                 }
2276
2277                 for (cur_res = hyperv_mmio; cur_res; cur_res = next_res) {
2278                         next_res = cur_res->sibling;
2279                         kfree(cur_res);
2280                 }
2281         }
2282 }
2283
2284 static void vmbus_reserve_fb(void)
2285 {
2286         resource_size_t start = 0, size;
2287         struct pci_dev *pdev;
2288
2289         if (efi_enabled(EFI_BOOT)) {
2290                 /* Gen2 VM: get FB base from EFI framebuffer */
2291                 start = screen_info.lfb_base;
2292                 size = max_t(__u32, screen_info.lfb_size, 0x800000);
2293         } else {
2294                 /* Gen1 VM: get FB base from PCI */
2295                 pdev = pci_get_device(PCI_VENDOR_ID_MICROSOFT,
2296                                       PCI_DEVICE_ID_HYPERV_VIDEO, NULL);
2297                 if (!pdev)
2298                         return;
2299
2300                 if (pdev->resource[0].flags & IORESOURCE_MEM) {
2301                         start = pci_resource_start(pdev, 0);
2302                         size = pci_resource_len(pdev, 0);
2303                 }
2304
2305                 /*
2306                  * Release the PCI device so hyperv_drm or hyperv_fb driver can
2307                  * grab it later.
2308                  */
2309                 pci_dev_put(pdev);
2310         }
2311
2312         if (!start)
2313                 return;
2314
2315         /*
2316          * Make a claim for the frame buffer in the resource tree under the
2317          * first node, which will be the one below 4GB.  The length seems to
2318          * be underreported, particularly in a Generation 1 VM.  So start out
2319          * reserving a larger area and make it smaller until it succeeds.
2320          */
2321         for (; !fb_mmio && (size >= 0x100000); size >>= 1)
2322                 fb_mmio = __request_region(hyperv_mmio, start, size, fb_mmio_name, 0);
2323 }
2324
2325 /**
2326  * vmbus_allocate_mmio() - Pick a memory-mapped I/O range.
2327  * @new:                If successful, supplied a pointer to the
2328  *                      allocated MMIO space.
2329  * @device_obj:         Identifies the caller
2330  * @min:                Minimum guest physical address of the
2331  *                      allocation
2332  * @max:                Maximum guest physical address
2333  * @size:               Size of the range to be allocated
2334  * @align:              Alignment of the range to be allocated
2335  * @fb_overlap_ok:      Whether this allocation can be allowed
2336  *                      to overlap the video frame buffer.
2337  *
2338  * This function walks the resources granted to VMBus by the
2339  * _CRS object in the ACPI namespace underneath the parent
2340  * "bridge" whether that's a root PCI bus in the Generation 1
2341  * case or a Module Device in the Generation 2 case.  It then
2342  * attempts to allocate from the global MMIO pool in a way that
2343  * matches the constraints supplied in these parameters and by
2344  * that _CRS.
2345  *
2346  * Return: 0 on success, -errno on failure
2347  */
2348 int vmbus_allocate_mmio(struct resource **new, struct hv_device *device_obj,
2349                         resource_size_t min, resource_size_t max,
2350                         resource_size_t size, resource_size_t align,
2351                         bool fb_overlap_ok)
2352 {
2353         struct resource *iter, *shadow;
2354         resource_size_t range_min, range_max, start, end;
2355         const char *dev_n = dev_name(&device_obj->device);
2356         int retval;
2357
2358         retval = -ENXIO;
2359         mutex_lock(&hyperv_mmio_lock);
2360
2361         /*
2362          * If overlaps with frame buffers are allowed, then first attempt to
2363          * make the allocation from within the reserved region.  Because it
2364          * is already reserved, no shadow allocation is necessary.
2365          */
2366         if (fb_overlap_ok && fb_mmio && !(min > fb_mmio->end) &&
2367             !(max < fb_mmio->start)) {
2368
2369                 range_min = fb_mmio->start;
2370                 range_max = fb_mmio->end;
2371                 start = (range_min + align - 1) & ~(align - 1);
2372                 for (; start + size - 1 <= range_max; start += align) {
2373                         *new = request_mem_region_exclusive(start, size, dev_n);
2374                         if (*new) {
2375                                 retval = 0;
2376                                 goto exit;
2377                         }
2378                 }
2379         }
2380
2381         for (iter = hyperv_mmio; iter; iter = iter->sibling) {
2382                 if ((iter->start >= max) || (iter->end <= min))
2383                         continue;
2384
2385                 range_min = iter->start;
2386                 range_max = iter->end;
2387                 start = (range_min + align - 1) & ~(align - 1);
2388                 for (; start + size - 1 <= range_max; start += align) {
2389                         end = start + size - 1;
2390
2391                         /* Skip the whole fb_mmio region if not fb_overlap_ok */
2392                         if (!fb_overlap_ok && fb_mmio &&
2393                             (((start >= fb_mmio->start) && (start <= fb_mmio->end)) ||
2394                              ((end >= fb_mmio->start) && (end <= fb_mmio->end))))
2395                                 continue;
2396
2397                         shadow = __request_region(iter, start, size, NULL,
2398                                                   IORESOURCE_BUSY);
2399                         if (!shadow)
2400                                 continue;
2401
2402                         *new = request_mem_region_exclusive(start, size, dev_n);
2403                         if (*new) {
2404                                 shadow->name = (char *)*new;
2405                                 retval = 0;
2406                                 goto exit;
2407                         }
2408
2409                         __release_region(iter, start, size);
2410                 }
2411         }
2412
2413 exit:
2414         mutex_unlock(&hyperv_mmio_lock);
2415         return retval;
2416 }
2417 EXPORT_SYMBOL_GPL(vmbus_allocate_mmio);
2418
2419 /**
2420  * vmbus_free_mmio() - Free a memory-mapped I/O range.
2421  * @start:              Base address of region to release.
2422  * @size:               Size of the range to be allocated
2423  *
2424  * This function releases anything requested by
2425  * vmbus_mmio_allocate().
2426  */
2427 void vmbus_free_mmio(resource_size_t start, resource_size_t size)
2428 {
2429         struct resource *iter;
2430
2431         mutex_lock(&hyperv_mmio_lock);
2432         for (iter = hyperv_mmio; iter; iter = iter->sibling) {
2433                 if ((iter->start >= start + size) || (iter->end <= start))
2434                         continue;
2435
2436                 __release_region(iter, start, size);
2437         }
2438         release_mem_region(start, size);
2439         mutex_unlock(&hyperv_mmio_lock);
2440
2441 }
2442 EXPORT_SYMBOL_GPL(vmbus_free_mmio);
2443
2444 static int vmbus_acpi_add(struct acpi_device *device)
2445 {
2446         acpi_status result;
2447         int ret_val = -ENODEV;
2448         struct acpi_device *ancestor;
2449
2450         hv_acpi_dev = device;
2451
2452         /*
2453          * Older versions of Hyper-V for ARM64 fail to include the _CCA
2454          * method on the top level VMbus device in the DSDT. But devices
2455          * are hardware coherent in all current Hyper-V use cases, so fix
2456          * up the ACPI device to behave as if _CCA is present and indicates
2457          * hardware coherence.
2458          */
2459         ACPI_COMPANION_SET(&device->dev, device);
2460         if (IS_ENABLED(CONFIG_ACPI_CCA_REQUIRED) &&
2461             device_get_dma_attr(&device->dev) == DEV_DMA_NOT_SUPPORTED) {
2462                 pr_info("No ACPI _CCA found; assuming coherent device I/O\n");
2463                 device->flags.cca_seen = true;
2464                 device->flags.coherent_dma = true;
2465         }
2466
2467         result = acpi_walk_resources(device->handle, METHOD_NAME__CRS,
2468                                         vmbus_walk_resources, NULL);
2469
2470         if (ACPI_FAILURE(result))
2471                 goto acpi_walk_err;
2472         /*
2473          * Some ancestor of the vmbus acpi device (Gen1 or Gen2
2474          * firmware) is the VMOD that has the mmio ranges. Get that.
2475          */
2476         for (ancestor = acpi_dev_parent(device); ancestor;
2477              ancestor = acpi_dev_parent(ancestor)) {
2478                 result = acpi_walk_resources(ancestor->handle, METHOD_NAME__CRS,
2479                                              vmbus_walk_resources, NULL);
2480
2481                 if (ACPI_FAILURE(result))
2482                         continue;
2483                 if (hyperv_mmio) {
2484                         vmbus_reserve_fb();
2485                         break;
2486                 }
2487         }
2488         ret_val = 0;
2489
2490 acpi_walk_err:
2491         if (ret_val)
2492                 vmbus_acpi_remove(device);
2493         return ret_val;
2494 }
2495
2496 #ifdef CONFIG_PM_SLEEP
2497 static int vmbus_bus_suspend(struct device *dev)
2498 {
2499         struct hv_per_cpu_context *hv_cpu = per_cpu_ptr(
2500                         hv_context.cpu_context, VMBUS_CONNECT_CPU);
2501         struct vmbus_channel *channel, *sc;
2502
2503         tasklet_disable(&hv_cpu->msg_dpc);
2504         vmbus_connection.ignore_any_offer_msg = true;
2505         /* The tasklet_enable() takes care of providing a memory barrier */
2506         tasklet_enable(&hv_cpu->msg_dpc);
2507
2508         /* Drain all the workqueues as we are in suspend */
2509         drain_workqueue(vmbus_connection.rescind_work_queue);
2510         drain_workqueue(vmbus_connection.work_queue);
2511         drain_workqueue(vmbus_connection.handle_primary_chan_wq);
2512         drain_workqueue(vmbus_connection.handle_sub_chan_wq);
2513
2514         mutex_lock(&vmbus_connection.channel_mutex);
2515         list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
2516                 if (!is_hvsock_channel(channel))
2517                         continue;
2518
2519                 vmbus_force_channel_rescinded(channel);
2520         }
2521         mutex_unlock(&vmbus_connection.channel_mutex);
2522
2523         /*
2524          * Wait until all the sub-channels and hv_sock channels have been
2525          * cleaned up. Sub-channels should be destroyed upon suspend, otherwise
2526          * they would conflict with the new sub-channels that will be created
2527          * in the resume path. hv_sock channels should also be destroyed, but
2528          * a hv_sock channel of an established hv_sock connection can not be
2529          * really destroyed since it may still be referenced by the userspace
2530          * application, so we just force the hv_sock channel to be rescinded
2531          * by vmbus_force_channel_rescinded(), and the userspace application
2532          * will thoroughly destroy the channel after hibernation.
2533          *
2534          * Note: the counter nr_chan_close_on_suspend may never go above 0 if
2535          * the VM has no sub-channel and hv_sock channel, e.g. a 1-vCPU VM.
2536          */
2537         if (atomic_read(&vmbus_connection.nr_chan_close_on_suspend) > 0)
2538                 wait_for_completion(&vmbus_connection.ready_for_suspend_event);
2539
2540         if (atomic_read(&vmbus_connection.nr_chan_fixup_on_resume) != 0) {
2541                 pr_err("Can not suspend due to a previous failed resuming\n");
2542                 return -EBUSY;
2543         }
2544
2545         mutex_lock(&vmbus_connection.channel_mutex);
2546
2547         list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
2548                 /*
2549                  * Remove the channel from the array of channels and invalidate
2550                  * the channel's relid.  Upon resume, vmbus_onoffer() will fix
2551                  * up the relid (and other fields, if necessary) and add the
2552                  * channel back to the array.
2553                  */
2554                 vmbus_channel_unmap_relid(channel);
2555                 channel->offermsg.child_relid = INVALID_RELID;
2556
2557                 if (is_hvsock_channel(channel)) {
2558                         if (!channel->rescind) {
2559                                 pr_err("hv_sock channel not rescinded!\n");
2560                                 WARN_ON_ONCE(1);
2561                         }
2562                         continue;
2563                 }
2564
2565                 list_for_each_entry(sc, &channel->sc_list, sc_list) {
2566                         pr_err("Sub-channel not deleted!\n");
2567                         WARN_ON_ONCE(1);
2568                 }
2569
2570                 atomic_inc(&vmbus_connection.nr_chan_fixup_on_resume);
2571         }
2572
2573         mutex_unlock(&vmbus_connection.channel_mutex);
2574
2575         vmbus_initiate_unload(false);
2576
2577         /* Reset the event for the next resume. */
2578         reinit_completion(&vmbus_connection.ready_for_resume_event);
2579
2580         return 0;
2581 }
2582
2583 static int vmbus_bus_resume(struct device *dev)
2584 {
2585         struct vmbus_channel_msginfo *msginfo;
2586         size_t msgsize;
2587         int ret;
2588
2589         vmbus_connection.ignore_any_offer_msg = false;
2590
2591         /*
2592          * We only use the 'vmbus_proto_version', which was in use before
2593          * hibernation, to re-negotiate with the host.
2594          */
2595         if (!vmbus_proto_version) {
2596                 pr_err("Invalid proto version = 0x%x\n", vmbus_proto_version);
2597                 return -EINVAL;
2598         }
2599
2600         msgsize = sizeof(*msginfo) +
2601                   sizeof(struct vmbus_channel_initiate_contact);
2602
2603         msginfo = kzalloc(msgsize, GFP_KERNEL);
2604
2605         if (msginfo == NULL)
2606                 return -ENOMEM;
2607
2608         ret = vmbus_negotiate_version(msginfo, vmbus_proto_version);
2609
2610         kfree(msginfo);
2611
2612         if (ret != 0)
2613                 return ret;
2614
2615         WARN_ON(atomic_read(&vmbus_connection.nr_chan_fixup_on_resume) == 0);
2616
2617         vmbus_request_offers();
2618
2619         if (wait_for_completion_timeout(
2620                 &vmbus_connection.ready_for_resume_event, 10 * HZ) == 0)
2621                 pr_err("Some vmbus device is missing after suspending?\n");
2622
2623         /* Reset the event for the next suspend. */
2624         reinit_completion(&vmbus_connection.ready_for_suspend_event);
2625
2626         return 0;
2627 }
2628 #else
2629 #define vmbus_bus_suspend NULL
2630 #define vmbus_bus_resume NULL
2631 #endif /* CONFIG_PM_SLEEP */
2632
2633 static const struct acpi_device_id vmbus_acpi_device_ids[] = {
2634         {"VMBUS", 0},
2635         {"VMBus", 0},
2636         {"", 0},
2637 };
2638 MODULE_DEVICE_TABLE(acpi, vmbus_acpi_device_ids);
2639
2640 /*
2641  * Note: we must use the "no_irq" ops, otherwise hibernation can not work with
2642  * PCI device assignment, because "pci_dev_pm_ops" uses the "noirq" ops: in
2643  * the resume path, the pci "noirq" restore op runs before "non-noirq" op (see
2644  * resume_target_kernel() -> dpm_resume_start(), and hibernation_restore() ->
2645  * dpm_resume_end()). This means vmbus_bus_resume() and the pci-hyperv's
2646  * resume callback must also run via the "noirq" ops.
2647  *
2648  * Set suspend_noirq/resume_noirq to NULL for Suspend-to-Idle: see the comment
2649  * earlier in this file before vmbus_pm.
2650  */
2651
2652 static const struct dev_pm_ops vmbus_bus_pm = {
2653         .suspend_noirq  = NULL,
2654         .resume_noirq   = NULL,
2655         .freeze_noirq   = vmbus_bus_suspend,
2656         .thaw_noirq     = vmbus_bus_resume,
2657         .poweroff_noirq = vmbus_bus_suspend,
2658         .restore_noirq  = vmbus_bus_resume
2659 };
2660
2661 static struct acpi_driver vmbus_acpi_driver = {
2662         .name = "vmbus",
2663         .ids = vmbus_acpi_device_ids,
2664         .ops = {
2665                 .add = vmbus_acpi_add,
2666                 .remove = vmbus_acpi_remove,
2667         },
2668         .drv.pm = &vmbus_bus_pm,
2669         .drv.probe_type = PROBE_FORCE_SYNCHRONOUS,
2670 };
2671
2672 static void hv_kexec_handler(void)
2673 {
2674         hv_stimer_global_cleanup();
2675         vmbus_initiate_unload(false);
2676         /* Make sure conn_state is set as hv_synic_cleanup checks for it */
2677         mb();
2678         cpuhp_remove_state(hyperv_cpuhp_online);
2679 };
2680
2681 static void hv_crash_handler(struct pt_regs *regs)
2682 {
2683         int cpu;
2684
2685         vmbus_initiate_unload(true);
2686         /*
2687          * In crash handler we can't schedule synic cleanup for all CPUs,
2688          * doing the cleanup for current CPU only. This should be sufficient
2689          * for kdump.
2690          */
2691         cpu = smp_processor_id();
2692         hv_stimer_cleanup(cpu);
2693         hv_synic_disable_regs(cpu);
2694 };
2695
2696 static int hv_synic_suspend(void)
2697 {
2698         /*
2699          * When we reach here, all the non-boot CPUs have been offlined.
2700          * If we're in a legacy configuration where stimer Direct Mode is
2701          * not enabled, the stimers on the non-boot CPUs have been unbound
2702          * in hv_synic_cleanup() -> hv_stimer_legacy_cleanup() ->
2703          * hv_stimer_cleanup() -> clockevents_unbind_device().
2704          *
2705          * hv_synic_suspend() only runs on CPU0 with interrupts disabled.
2706          * Here we do not call hv_stimer_legacy_cleanup() on CPU0 because:
2707          * 1) it's unnecessary as interrupts remain disabled between
2708          * syscore_suspend() and syscore_resume(): see create_image() and
2709          * resume_target_kernel()
2710          * 2) the stimer on CPU0 is automatically disabled later by
2711          * syscore_suspend() -> timekeeping_suspend() -> tick_suspend() -> ...
2712          * -> clockevents_shutdown() -> ... -> hv_ce_shutdown()
2713          * 3) a warning would be triggered if we call
2714          * clockevents_unbind_device(), which may sleep, in an
2715          * interrupts-disabled context.
2716          */
2717
2718         hv_synic_disable_regs(0);
2719
2720         return 0;
2721 }
2722
2723 static void hv_synic_resume(void)
2724 {
2725         hv_synic_enable_regs(0);
2726
2727         /*
2728          * Note: we don't need to call hv_stimer_init(0), because the timer
2729          * on CPU0 is not unbound in hv_synic_suspend(), and the timer is
2730          * automatically re-enabled in timekeeping_resume().
2731          */
2732 }
2733
2734 /* The callbacks run only on CPU0, with irqs_disabled. */
2735 static struct syscore_ops hv_synic_syscore_ops = {
2736         .suspend = hv_synic_suspend,
2737         .resume = hv_synic_resume,
2738 };
2739
2740 static int __init hv_acpi_init(void)
2741 {
2742         int ret;
2743
2744         if (!hv_is_hyperv_initialized())
2745                 return -ENODEV;
2746
2747         if (hv_root_partition && !hv_nested)
2748                 return 0;
2749
2750         /*
2751          * Get ACPI resources first.
2752          */
2753         ret = acpi_bus_register_driver(&vmbus_acpi_driver);
2754
2755         if (ret)
2756                 return ret;
2757
2758         if (!hv_acpi_dev) {
2759                 ret = -ENODEV;
2760                 goto cleanup;
2761         }
2762
2763         /*
2764          * If we're on an architecture with a hardcoded hypervisor
2765          * vector (i.e. x86/x64), override the VMbus interrupt found
2766          * in the ACPI tables. Ensure vmbus_irq is not set since the
2767          * normal Linux IRQ mechanism is not used in this case.
2768          */
2769 #ifdef HYPERVISOR_CALLBACK_VECTOR
2770         vmbus_interrupt = HYPERVISOR_CALLBACK_VECTOR;
2771         vmbus_irq = -1;
2772 #endif
2773
2774         hv_debug_init();
2775
2776         ret = vmbus_bus_init();
2777         if (ret)
2778                 goto cleanup;
2779
2780         hv_setup_kexec_handler(hv_kexec_handler);
2781         hv_setup_crash_handler(hv_crash_handler);
2782
2783         register_syscore_ops(&hv_synic_syscore_ops);
2784
2785         return 0;
2786
2787 cleanup:
2788         acpi_bus_unregister_driver(&vmbus_acpi_driver);
2789         hv_acpi_dev = NULL;
2790         return ret;
2791 }
2792
2793 static void __exit vmbus_exit(void)
2794 {
2795         int cpu;
2796
2797         unregister_syscore_ops(&hv_synic_syscore_ops);
2798
2799         hv_remove_kexec_handler();
2800         hv_remove_crash_handler();
2801         vmbus_connection.conn_state = DISCONNECTED;
2802         hv_stimer_global_cleanup();
2803         vmbus_disconnect();
2804         if (vmbus_irq == -1) {
2805                 hv_remove_vmbus_handler();
2806         } else {
2807                 free_percpu_irq(vmbus_irq, vmbus_evt);
2808                 free_percpu(vmbus_evt);
2809         }
2810         for_each_online_cpu(cpu) {
2811                 struct hv_per_cpu_context *hv_cpu
2812                         = per_cpu_ptr(hv_context.cpu_context, cpu);
2813
2814                 tasklet_kill(&hv_cpu->msg_dpc);
2815         }
2816         hv_debug_rm_all_dir();
2817
2818         vmbus_free_channels();
2819         kfree(vmbus_connection.channels);
2820
2821         if (ms_hyperv.misc_features & HV_FEATURE_GUEST_CRASH_MSR_AVAILABLE) {
2822                 kmsg_dump_unregister(&hv_kmsg_dumper);
2823                 unregister_die_notifier(&hyperv_die_report_block);
2824                 atomic_notifier_chain_unregister(&panic_notifier_list,
2825                                                 &hyperv_panic_report_block);
2826         }
2827
2828         /*
2829          * The vmbus panic notifier is always registered, hence we should
2830          * also unconditionally unregister it here as well.
2831          */
2832         atomic_notifier_chain_unregister(&panic_notifier_list,
2833                                         &hyperv_panic_vmbus_unload_block);
2834
2835         free_page((unsigned long)hv_panic_page);
2836         unregister_sysctl_table(hv_ctl_table_hdr);
2837         hv_ctl_table_hdr = NULL;
2838         bus_unregister(&hv_bus);
2839
2840         cpuhp_remove_state(hyperv_cpuhp_online);
2841         hv_synic_free();
2842         acpi_bus_unregister_driver(&vmbus_acpi_driver);
2843 }
2844
2845
2846 MODULE_LICENSE("GPL");
2847 MODULE_DESCRIPTION("Microsoft Hyper-V VMBus Driver");
2848
2849 subsys_initcall(hv_acpi_init);
2850 module_exit(vmbus_exit);