FILE
vmbus_drv.c
File preview with syntax highlighting.
linux-6.19.14/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/platform_device.h> 16#include <linux/interrupt.h> 17#include <linux/sysctl.h> 18#include <linux/slab.h> 19#include <linux/acpi.h> 20#include <linux/completion.h> 21#include <linux/hyperv.h> 22#include <linux/kernel_stat.h> 23#include <linux/of_address.h> 24#include <linux/clockchips.h> 25#include <linux/cpu.h> 26#include <linux/sched/isolation.h> 27#include <linux/sched/task_stack.h> 28#include <linux/smpboot.h> 29 30#include <linux/delay.h> 31#include <linux/panic_notifier.h> 32#include <linux/ptrace.h> 33#include <linux/screen_info.h> 34#include <linux/efi.h> 35#include <linux/random.h> 36#include <linux/kernel.h> 37#include <linux/syscore_ops.h> 38#include <linux/dma-map-ops.h> 39#include <linux/pci.h> 40#include <linux/export.h> 41#include <clocksource/hyperv_timer.h> 42#include <asm/mshyperv.h> 43#include "hyperv_vmbus.h" 44 45struct vmbus_dynid { 46 struct list_head node; 47 struct hv_vmbus_device_id id; 48}; 49 50/* VMBus Root Device */ 51static struct device *vmbus_root_device; 52 53static int hyperv_cpuhp_online; 54 55static long __percpu *vmbus_evt; 56 57/* Values parsed from ACPI DSDT */ 58int vmbus_irq; 59int vmbus_interrupt; 60 61/* 62 * If the Confidential VMBus is used, the data on the "wire" is not 63 * visible to either the host or the hypervisor. 64 */ 65static bool is_confidential; 66 67bool vmbus_is_confidential(void) 68{ 69 return is_confidential; 70} 71EXPORT_SYMBOL_GPL(vmbus_is_confidential); 72 73/* 74 * The panic notifier below is responsible solely for unloading the 75 * vmbus connection, which is necessary in a panic event. 76 * 77 * Notice an intrincate relation of this notifier with Hyper-V 78 * framebuffer panic notifier exists - we need vmbus connection alive 79 * there in order to succeed, so we need to order both with each other 80 * [see hvfb_on_panic()] - this is done using notifiers' priorities. 81 */ 82static int hv_panic_vmbus_unload(struct notifier_block *nb, unsigned long val, 83 void *args) 84{ 85 vmbus_initiate_unload(true); 86 return NOTIFY_DONE; 87} 88static struct notifier_block hyperv_panic_vmbus_unload_block = { 89 .notifier_call = hv_panic_vmbus_unload, 90 .priority = INT_MIN + 1, /* almost the latest one to execute */ 91}; 92 93static const char *fb_mmio_name = "fb_range"; 94static struct resource *fb_mmio; 95static struct resource *hyperv_mmio; 96static DEFINE_MUTEX(hyperv_mmio_lock); 97 98struct device *hv_get_vmbus_root_device(void) 99{ 100 return vmbus_root_device; 101} 102EXPORT_SYMBOL_GPL(hv_get_vmbus_root_device); 103 104static int vmbus_exists(void) 105{ 106 if (vmbus_root_device == NULL) 107 return -ENODEV; 108 109 return 0; 110} 111 112static u8 channel_monitor_group(const struct vmbus_channel *channel) 113{ 114 return (u8)channel->offermsg.monitorid / 32; 115} 116 117static u8 channel_monitor_offset(const struct vmbus_channel *channel) 118{ 119 return (u8)channel->offermsg.monitorid % 32; 120} 121 122static u32 channel_pending(const struct vmbus_channel *channel, 123 const struct hv_monitor_page *monitor_page) 124{ 125 u8 monitor_group = channel_monitor_group(channel); 126 127 return monitor_page->trigger_group[monitor_group].pending; 128} 129 130static u32 channel_latency(const struct vmbus_channel *channel, 131 const struct hv_monitor_page *monitor_page) 132{ 133 u8 monitor_group = channel_monitor_group(channel); 134 u8 monitor_offset = channel_monitor_offset(channel); 135 136 return monitor_page->latency[monitor_group][monitor_offset]; 137} 138 139static u32 channel_conn_id(struct vmbus_channel *channel, 140 struct hv_monitor_page *monitor_page) 141{ 142 u8 monitor_group = channel_monitor_group(channel); 143 u8 monitor_offset = channel_monitor_offset(channel); 144 145 return monitor_page->parameter[monitor_group][monitor_offset].connectionid.u.id; 146} 147 148static ssize_t id_show(struct device *dev, struct device_attribute *dev_attr, 149 char *buf) 150{ 151 struct hv_device *hv_dev = device_to_hv_device(dev); 152 153 if (!hv_dev->channel) 154 return -ENODEV; 155 return sysfs_emit(buf, "%d\n", hv_dev->channel->offermsg.child_relid); 156} 157static DEVICE_ATTR_RO(id); 158 159static ssize_t state_show(struct device *dev, struct device_attribute *dev_attr, 160 char *buf) 161{ 162 struct hv_device *hv_dev = device_to_hv_device(dev); 163 164 if (!hv_dev->channel) 165 return -ENODEV; 166 return sysfs_emit(buf, "%d\n", hv_dev->channel->state); 167} 168static DEVICE_ATTR_RO(state); 169 170static ssize_t monitor_id_show(struct device *dev, 171 struct device_attribute *dev_attr, char *buf) 172{ 173 struct hv_device *hv_dev = device_to_hv_device(dev); 174 175 if (!hv_dev->channel) 176 return -ENODEV; 177 return sysfs_emit(buf, "%d\n", hv_dev->channel->offermsg.monitorid); 178} 179static DEVICE_ATTR_RO(monitor_id); 180 181static ssize_t class_id_show(struct device *dev, 182 struct device_attribute *dev_attr, char *buf) 183{ 184 struct hv_device *hv_dev = device_to_hv_device(dev); 185 186 if (!hv_dev->channel) 187 return -ENODEV; 188 return sysfs_emit(buf, "{%pUl}\n", 189 &hv_dev->channel->offermsg.offer.if_type); 190} 191static DEVICE_ATTR_RO(class_id); 192 193static ssize_t device_id_show(struct device *dev, 194 struct device_attribute *dev_attr, char *buf) 195{ 196 struct hv_device *hv_dev = device_to_hv_device(dev); 197 198 if (!hv_dev->channel) 199 return -ENODEV; 200 return sysfs_emit(buf, "{%pUl}\n", 201 &hv_dev->channel->offermsg.offer.if_instance); 202} 203static DEVICE_ATTR_RO(device_id); 204 205static ssize_t modalias_show(struct device *dev, 206 struct device_attribute *dev_attr, char *buf) 207{ 208 struct hv_device *hv_dev = device_to_hv_device(dev); 209 210 return sysfs_emit(buf, "vmbus:%*phN\n", UUID_SIZE, &hv_dev->dev_type); 211} 212static DEVICE_ATTR_RO(modalias); 213 214#ifdef CONFIG_NUMA 215static ssize_t numa_node_show(struct device *dev, 216 struct device_attribute *attr, char *buf) 217{ 218 struct hv_device *hv_dev = device_to_hv_device(dev); 219 220 if (!hv_dev->channel) 221 return -ENODEV; 222 223 return sysfs_emit(buf, "%d\n", cpu_to_node(hv_dev->channel->target_cpu)); 224} 225static DEVICE_ATTR_RO(numa_node); 226#endif 227 228static ssize_t server_monitor_pending_show(struct device *dev, 229 struct device_attribute *dev_attr, 230 char *buf) 231{ 232 struct hv_device *hv_dev = device_to_hv_device(dev); 233 234 if (!hv_dev->channel) 235 return -ENODEV; 236 return sysfs_emit(buf, "%d\n", channel_pending(hv_dev->channel, 237 vmbus_connection.monitor_pages[0])); 238} 239static DEVICE_ATTR_RO(server_monitor_pending); 240 241static ssize_t client_monitor_pending_show(struct device *dev, 242 struct device_attribute *dev_attr, 243 char *buf) 244{ 245 struct hv_device *hv_dev = device_to_hv_device(dev); 246 247 if (!hv_dev->channel) 248 return -ENODEV; 249 return sysfs_emit(buf, "%d\n", channel_pending(hv_dev->channel, 250 vmbus_connection.monitor_pages[1])); 251} 252static DEVICE_ATTR_RO(client_monitor_pending); 253 254static ssize_t server_monitor_latency_show(struct device *dev, 255 struct device_attribute *dev_attr, 256 char *buf) 257{ 258 struct hv_device *hv_dev = device_to_hv_device(dev); 259 260 if (!hv_dev->channel) 261 return -ENODEV; 262 return sysfs_emit(buf, "%d\n", channel_latency(hv_dev->channel, 263 vmbus_connection.monitor_pages[0])); 264} 265static DEVICE_ATTR_RO(server_monitor_latency); 266 267static ssize_t client_monitor_latency_show(struct device *dev, 268 struct device_attribute *dev_attr, 269 char *buf) 270{ 271 struct hv_device *hv_dev = device_to_hv_device(dev); 272 273 if (!hv_dev->channel) 274 return -ENODEV; 275 return sysfs_emit(buf, "%d\n", channel_latency(hv_dev->channel, 276 vmbus_connection.monitor_pages[1])); 277} 278static DEVICE_ATTR_RO(client_monitor_latency); 279 280static ssize_t server_monitor_conn_id_show(struct device *dev, 281 struct device_attribute *dev_attr, 282 char *buf) 283{ 284 struct hv_device *hv_dev = device_to_hv_device(dev); 285 286 if (!hv_dev->channel) 287 return -ENODEV; 288 return sysfs_emit(buf, "%d\n", channel_conn_id(hv_dev->channel, 289 vmbus_connection.monitor_pages[0])); 290} 291static DEVICE_ATTR_RO(server_monitor_conn_id); 292 293static ssize_t client_monitor_conn_id_show(struct device *dev, 294 struct device_attribute *dev_attr, 295 char *buf) 296{ 297 struct hv_device *hv_dev = device_to_hv_device(dev); 298 299 if (!hv_dev->channel) 300 return -ENODEV; 301 return sysfs_emit(buf, "%d\n", channel_conn_id(hv_dev->channel, 302 vmbus_connection.monitor_pages[1])); 303} 304static DEVICE_ATTR_RO(client_monitor_conn_id); 305 306static ssize_t out_intr_mask_show(struct device *dev, 307 struct device_attribute *dev_attr, char *buf) 308{ 309 struct hv_device *hv_dev = device_to_hv_device(dev); 310 struct hv_ring_buffer_debug_info outbound; 311 int ret; 312 313 if (!hv_dev->channel) 314 return -ENODEV; 315 316 ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->outbound, 317 &outbound); 318 if (ret < 0) 319 return ret; 320 321 return sysfs_emit(buf, "%d\n", outbound.current_interrupt_mask); 322} 323static DEVICE_ATTR_RO(out_intr_mask); 324 325static ssize_t out_read_index_show(struct device *dev, 326 struct device_attribute *dev_attr, char *buf) 327{ 328 struct hv_device *hv_dev = device_to_hv_device(dev); 329 struct hv_ring_buffer_debug_info outbound; 330 int ret; 331 332 if (!hv_dev->channel) 333 return -ENODEV; 334 335 ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->outbound, 336 &outbound); 337 if (ret < 0) 338 return ret; 339 return sysfs_emit(buf, "%u\n", outbound.current_read_index); 340} 341static DEVICE_ATTR_RO(out_read_index); 342 343static ssize_t out_write_index_show(struct device *dev, 344 struct device_attribute *dev_attr, 345 char *buf) 346{ 347 struct hv_device *hv_dev = device_to_hv_device(dev); 348 struct hv_ring_buffer_debug_info outbound; 349 int ret; 350 351 if (!hv_dev->channel) 352 return -ENODEV; 353 354 ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->outbound, 355 &outbound); 356 if (ret < 0) 357 return ret; 358 return sysfs_emit(buf, "%u\n", outbound.current_write_index); 359} 360static DEVICE_ATTR_RO(out_write_index); 361 362static ssize_t out_read_bytes_avail_show(struct device *dev, 363 struct device_attribute *dev_attr, 364 char *buf) 365{ 366 struct hv_device *hv_dev = device_to_hv_device(dev); 367 struct hv_ring_buffer_debug_info outbound; 368 int ret; 369 370 if (!hv_dev->channel) 371 return -ENODEV; 372 373 ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->outbound, 374 &outbound); 375 if (ret < 0) 376 return ret; 377 return sysfs_emit(buf, "%d\n", outbound.bytes_avail_toread); 378} 379static DEVICE_ATTR_RO(out_read_bytes_avail); 380 381static ssize_t out_write_bytes_avail_show(struct device *dev, 382 struct device_attribute *dev_attr, 383 char *buf) 384{ 385 struct hv_device *hv_dev = device_to_hv_device(dev); 386 struct hv_ring_buffer_debug_info outbound; 387 int ret; 388 389 if (!hv_dev->channel) 390 return -ENODEV; 391 392 ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->outbound, 393 &outbound); 394 if (ret < 0) 395 return ret; 396 return sysfs_emit(buf, "%d\n", outbound.bytes_avail_towrite); 397} 398static DEVICE_ATTR_RO(out_write_bytes_avail); 399 400static ssize_t in_intr_mask_show(struct device *dev, 401 struct device_attribute *dev_attr, char *buf) 402{ 403 struct hv_device *hv_dev = device_to_hv_device(dev); 404 struct hv_ring_buffer_debug_info inbound; 405 int ret; 406 407 if (!hv_dev->channel) 408 return -ENODEV; 409 410 ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->inbound, &inbound); 411 if (ret < 0) 412 return ret; 413 414 return sysfs_emit(buf, "%d\n", inbound.current_interrupt_mask); 415} 416static DEVICE_ATTR_RO(in_intr_mask); 417 418static ssize_t in_read_index_show(struct device *dev, 419 struct device_attribute *dev_attr, char *buf) 420{ 421 struct hv_device *hv_dev = device_to_hv_device(dev); 422 struct hv_ring_buffer_debug_info inbound; 423 int ret; 424 425 if (!hv_dev->channel) 426 return -ENODEV; 427 428 ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->inbound, &inbound); 429 if (ret < 0) 430 return ret; 431 432 return sysfs_emit(buf, "%d\n", inbound.current_read_index); 433} 434static DEVICE_ATTR_RO(in_read_index); 435 436static ssize_t in_write_index_show(struct device *dev, 437 struct device_attribute *dev_attr, char *buf) 438{ 439 struct hv_device *hv_dev = device_to_hv_device(dev); 440 struct hv_ring_buffer_debug_info inbound; 441 int ret; 442 443 if (!hv_dev->channel) 444 return -ENODEV; 445 446 ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->inbound, &inbound); 447 if (ret < 0) 448 return ret; 449 450 return sysfs_emit(buf, "%d\n", inbound.current_write_index); 451} 452static DEVICE_ATTR_RO(in_write_index); 453 454static ssize_t in_read_bytes_avail_show(struct device *dev, 455 struct device_attribute *dev_attr, 456 char *buf) 457{ 458 struct hv_device *hv_dev = device_to_hv_device(dev); 459 struct hv_ring_buffer_debug_info inbound; 460 int ret; 461 462 if (!hv_dev->channel) 463 return -ENODEV; 464 465 ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->inbound, &inbound); 466 if (ret < 0) 467 return ret; 468 469 return sysfs_emit(buf, "%d\n", inbound.bytes_avail_toread); 470} 471static DEVICE_ATTR_RO(in_read_bytes_avail); 472 473static ssize_t in_write_bytes_avail_show(struct device *dev, 474 struct device_attribute *dev_attr, 475 char *buf) 476{ 477 struct hv_device *hv_dev = device_to_hv_device(dev); 478 struct hv_ring_buffer_debug_info inbound; 479 int ret; 480 481 if (!hv_dev->channel) 482 return -ENODEV; 483 484 ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->inbound, &inbound); 485 if (ret < 0) 486 return ret; 487 488 return sysfs_emit(buf, "%d\n", inbound.bytes_avail_towrite); 489} 490static DEVICE_ATTR_RO(in_write_bytes_avail); 491 492static ssize_t channel_vp_mapping_show(struct device *dev, 493 struct device_attribute *dev_attr, 494 char *buf) 495{ 496 struct hv_device *hv_dev = device_to_hv_device(dev); 497 struct vmbus_channel *channel = hv_dev->channel, *cur_sc; 498 int n_written; 499 struct list_head *cur; 500 501 if (!channel) 502 return -ENODEV; 503 504 mutex_lock(&vmbus_connection.channel_mutex); 505 506 n_written = sysfs_emit(buf, "%u:%u\n", 507 channel->offermsg.child_relid, 508 channel->target_cpu); 509 510 list_for_each(cur, &channel->sc_list) { 511 512 cur_sc = list_entry(cur, struct vmbus_channel, sc_list); 513 n_written += sysfs_emit_at(buf, n_written, "%u:%u\n", 514 cur_sc->offermsg.child_relid, 515 cur_sc->target_cpu); 516 } 517 518 mutex_unlock(&vmbus_connection.channel_mutex); 519 520 return n_written; 521} 522static DEVICE_ATTR_RO(channel_vp_mapping); 523 524static ssize_t vendor_show(struct device *dev, 525 struct device_attribute *dev_attr, 526 char *buf) 527{ 528 struct hv_device *hv_dev = device_to_hv_device(dev); 529 530 return sysfs_emit(buf, "0x%x\n", hv_dev->vendor_id); 531} 532static DEVICE_ATTR_RO(vendor); 533 534static ssize_t device_show(struct device *dev, 535 struct device_attribute *dev_attr, 536 char *buf) 537{ 538 struct hv_device *hv_dev = device_to_hv_device(dev); 539 540 return sysfs_emit(buf, "0x%x\n", hv_dev->device_id); 541} 542static DEVICE_ATTR_RO(device); 543 544static ssize_t driver_override_store(struct device *dev, 545 struct device_attribute *attr, 546 const char *buf, size_t count) 547{ 548 struct hv_device *hv_dev = device_to_hv_device(dev); 549 int ret; 550 551 ret = driver_set_override(dev, &hv_dev->driver_override, buf, count); 552 if (ret) 553 return ret; 554 555 return count; 556} 557 558static ssize_t driver_override_show(struct device *dev, 559 struct device_attribute *attr, char *buf) 560{ 561 struct hv_device *hv_dev = device_to_hv_device(dev); 562 ssize_t len; 563 564 device_lock(dev); 565 len = sysfs_emit(buf, "%s\n", hv_dev->driver_override); 566 device_unlock(dev); 567 568 return len; 569} 570static DEVICE_ATTR_RW(driver_override); 571 572/* Set up per device attributes in /sys/bus/vmbus/devices/<bus device> */ 573static struct attribute *vmbus_dev_attrs[] = { 574 &dev_attr_id.attr, 575 &dev_attr_state.attr, 576 &dev_attr_monitor_id.attr, 577 &dev_attr_class_id.attr, 578 &dev_attr_device_id.attr, 579 &dev_attr_modalias.attr, 580#ifdef CONFIG_NUMA 581 &dev_attr_numa_node.attr, 582#endif 583 &dev_attr_server_monitor_pending.attr, 584 &dev_attr_client_monitor_pending.attr, 585 &dev_attr_server_monitor_latency.attr, 586 &dev_attr_client_monitor_latency.attr, 587 &dev_attr_server_monitor_conn_id.attr, 588 &dev_attr_client_monitor_conn_id.attr, 589 &dev_attr_out_intr_mask.attr, 590 &dev_attr_out_read_index.attr, 591 &dev_attr_out_write_index.attr, 592 &dev_attr_out_read_bytes_avail.attr, 593 &dev_attr_out_write_bytes_avail.attr, 594 &dev_attr_in_intr_mask.attr, 595 &dev_attr_in_read_index.attr, 596 &dev_attr_in_write_index.attr, 597 &dev_attr_in_read_bytes_avail.attr, 598 &dev_attr_in_write_bytes_avail.attr, 599 &dev_attr_channel_vp_mapping.attr, 600 &dev_attr_vendor.attr, 601 &dev_attr_device.attr, 602 &dev_attr_driver_override.attr, 603 NULL, 604}; 605 606/* 607 * Device-level attribute_group callback function. Returns the permission for 608 * each attribute, and returns 0 if an attribute is not visible. 609 */ 610static umode_t vmbus_dev_attr_is_visible(struct kobject *kobj, 611 struct attribute *attr, int idx) 612{ 613 struct device *dev = kobj_to_dev(kobj); 614 const struct hv_device *hv_dev = device_to_hv_device(dev); 615 616 /* Hide the monitor attributes if the monitor mechanism is not used. */ 617 if (!hv_dev->channel->offermsg.monitor_allocated && 618 (attr == &dev_attr_monitor_id.attr || 619 attr == &dev_attr_server_monitor_pending.attr || 620 attr == &dev_attr_client_monitor_pending.attr || 621 attr == &dev_attr_server_monitor_latency.attr || 622 attr == &dev_attr_client_monitor_latency.attr || 623 attr == &dev_attr_server_monitor_conn_id.attr || 624 attr == &dev_attr_client_monitor_conn_id.attr)) 625 return 0; 626 627 return attr->mode; 628} 629 630static const struct attribute_group vmbus_dev_group = { 631 .attrs = vmbus_dev_attrs, 632 .is_visible = vmbus_dev_attr_is_visible 633}; 634__ATTRIBUTE_GROUPS(vmbus_dev); 635 636/* Set up the attribute for /sys/bus/vmbus/hibernation */ 637static ssize_t hibernation_show(const struct bus_type *bus, char *buf) 638{ 639 return sprintf(buf, "%d\n", !!hv_is_hibernation_supported()); 640} 641 642static BUS_ATTR_RO(hibernation); 643 644static struct attribute *vmbus_bus_attrs[] = { 645 &bus_attr_hibernation.attr, 646 NULL, 647}; 648static const struct attribute_group vmbus_bus_group = { 649 .attrs = vmbus_bus_attrs, 650}; 651__ATTRIBUTE_GROUPS(vmbus_bus); 652 653/* 654 * vmbus_uevent - add uevent for our device 655 * 656 * This routine is invoked when a device is added or removed on the vmbus to 657 * generate a uevent to udev in the userspace. The udev will then look at its 658 * rule and the uevent generated here to load the appropriate driver 659 * 660 * The alias string will be of the form vmbus:guid where guid is the string 661 * representation of the device guid (each byte of the guid will be 662 * represented with two hex characters. 663 */ 664static int vmbus_uevent(const struct device *device, struct kobj_uevent_env *env) 665{ 666 const struct hv_device *dev = device_to_hv_device(device); 667 const char *format = "MODALIAS=vmbus:%*phN"; 668 669 return add_uevent_var(env, format, UUID_SIZE, &dev->dev_type); 670} 671 672static const struct hv_vmbus_device_id * 673hv_vmbus_dev_match(const struct hv_vmbus_device_id *id, const guid_t *guid) 674{ 675 if (id == NULL) 676 return NULL; /* empty device table */ 677 678 for (; !guid_is_null(&id->guid); id++) 679 if (guid_equal(&id->guid, guid)) 680 return id; 681 682 return NULL; 683} 684 685static const struct hv_vmbus_device_id * 686hv_vmbus_dynid_match(struct hv_driver *drv, const guid_t *guid) 687{ 688 const struct hv_vmbus_device_id *id = NULL; 689 struct vmbus_dynid *dynid; 690 691 spin_lock(&drv->dynids.lock); 692 list_for_each_entry(dynid, &drv->dynids.list, node) { 693 if (guid_equal(&dynid->id.guid, guid)) { 694 id = &dynid->id; 695 break; 696 } 697 } 698 spin_unlock(&drv->dynids.lock); 699 700 return id; 701} 702 703static const struct hv_vmbus_device_id vmbus_device_null; 704 705/* 706 * Return a matching hv_vmbus_device_id pointer. 707 * If there is no match, return NULL. 708 */ 709static const struct hv_vmbus_device_id *hv_vmbus_get_id(const struct hv_driver *drv, 710 struct hv_device *dev) 711{ 712 const guid_t *guid = &dev->dev_type; 713 const struct hv_vmbus_device_id *id; 714 715 /* When driver_override is set, only bind to the matching driver */ 716 if (dev->driver_override && strcmp(dev->driver_override, drv->name)) 717 return NULL; 718 719 /* Look at the dynamic ids first, before the static ones */ 720 id = hv_vmbus_dynid_match((struct hv_driver *)drv, guid); 721 if (!id) 722 id = hv_vmbus_dev_match(drv->id_table, guid); 723 724 /* driver_override will always match, send a dummy id */ 725 if (!id && dev->driver_override) 726 id = &vmbus_device_null; 727 728 return id; 729} 730 731/* vmbus_add_dynid - add a new device ID to this driver and re-probe devices 732 * 733 * This function can race with vmbus_device_register(). This function is 734 * typically running on a user thread in response to writing to the "new_id" 735 * sysfs entry for a driver. vmbus_device_register() is running on a 736 * workqueue thread in response to the Hyper-V host offering a device to the 737 * guest. This function calls driver_attach(), which looks for an existing 738 * device matching the new id, and attaches the driver to which the new id 739 * has been assigned. vmbus_device_register() calls device_register(), which 740 * looks for a driver that matches the device being registered. If both 741 * operations are running simultaneously, the device driver probe function runs 742 * on whichever thread establishes the linkage between the driver and device. 743 * 744 * In most cases, it doesn't matter which thread runs the driver probe 745 * function. But if vmbus_device_register() does not find a matching driver, 746 * it proceeds to create the "channels" subdirectory and numbered per-channel 747 * subdirectory in sysfs. While that multi-step creation is in progress, this 748 * function could run the driver probe function. If the probe function checks 749 * for, or operates on, entries in the "channels" subdirectory, including by 750 * calling hv_create_ring_sysfs(), the operation may or may not succeed 751 * depending on the race. The race can't create a kernel failure in VMBus 752 * or device subsystem code, but probe functions in VMBus drivers doing such 753 * operations must be prepared for the failure case. 754 */ 755static int vmbus_add_dynid(struct hv_driver *drv, guid_t *guid) 756{ 757 struct vmbus_dynid *dynid; 758 759 dynid = kzalloc(sizeof(*dynid), GFP_KERNEL); 760 if (!dynid) 761 return -ENOMEM; 762 763 dynid->id.guid = *guid; 764 765 spin_lock(&drv->dynids.lock); 766 list_add_tail(&dynid->node, &drv->dynids.list); 767 spin_unlock(&drv->dynids.lock); 768 769 return driver_attach(&drv->driver); 770} 771 772static void vmbus_free_dynids(struct hv_driver *drv) 773{ 774 struct vmbus_dynid *dynid, *n; 775 776 spin_lock(&drv->dynids.lock); 777 list_for_each_entry_safe(dynid, n, &drv->dynids.list, node) { 778 list_del(&dynid->node); 779 kfree(dynid); 780 } 781 spin_unlock(&drv->dynids.lock); 782} 783 784/* 785 * store_new_id - sysfs frontend to vmbus_add_dynid() 786 * 787 * Allow GUIDs to be added to an existing driver via sysfs. 788 */ 789static ssize_t new_id_store(struct device_driver *driver, const char *buf, 790 size_t count) 791{ 792 struct hv_driver *drv = drv_to_hv_drv(driver); 793 guid_t guid; 794 ssize_t retval; 795 796 retval = guid_parse(buf, &guid); 797 if (retval) 798 return retval; 799 800 if (hv_vmbus_dynid_match(drv, &guid)) 801 return -EEXIST; 802 803 retval = vmbus_add_dynid(drv, &guid); 804 if (retval) 805 return retval; 806 return count; 807} 808static DRIVER_ATTR_WO(new_id); 809 810/* 811 * store_remove_id - remove a PCI device ID from this driver 812 * 813 * Removes a dynamic pci device ID to this driver. 814 */ 815static ssize_t remove_id_store(struct device_driver *driver, const char *buf, 816 size_t count) 817{ 818 struct hv_driver *drv = drv_to_hv_drv(driver); 819 struct vmbus_dynid *dynid, *n; 820 guid_t guid; 821 ssize_t retval; 822 823 retval = guid_parse(buf, &guid); 824 if (retval) 825 return retval; 826 827 retval = -ENODEV; 828 spin_lock(&drv->dynids.lock); 829 list_for_each_entry_safe(dynid, n, &drv->dynids.list, node) { 830 struct hv_vmbus_device_id *id = &dynid->id; 831 832 if (guid_equal(&id->guid, &guid)) { 833 list_del(&dynid->node); 834 kfree(dynid); 835 retval = count; 836 break; 837 } 838 } 839 spin_unlock(&drv->dynids.lock); 840 841 return retval; 842} 843static DRIVER_ATTR_WO(remove_id); 844 845static struct attribute *vmbus_drv_attrs[] = { 846 &driver_attr_new_id.attr, 847 &driver_attr_remove_id.attr, 848 NULL, 849}; 850ATTRIBUTE_GROUPS(vmbus_drv); 851 852 853/* 854 * vmbus_match - Attempt to match the specified device to the specified driver 855 */ 856static int vmbus_match(struct device *device, const struct device_driver *driver) 857{ 858 const struct hv_driver *drv = drv_to_hv_drv(driver); 859 struct hv_device *hv_dev = device_to_hv_device(device); 860 861 /* The hv_sock driver handles all hv_sock offers. */ 862 if (is_hvsock_channel(hv_dev->channel)) 863 return drv->hvsock; 864 865 if (hv_vmbus_get_id(drv, hv_dev)) 866 return 1; 867 868 return 0; 869} 870 871/* 872 * vmbus_probe - Add the new vmbus's child device 873 */ 874static int vmbus_probe(struct device *child_device) 875{ 876 int ret = 0; 877 struct hv_driver *drv = 878 drv_to_hv_drv(child_device->driver); 879 struct hv_device *dev = device_to_hv_device(child_device); 880 const struct hv_vmbus_device_id *dev_id; 881 882 dev_id = hv_vmbus_get_id(drv, dev); 883 if (drv->probe) { 884 ret = drv->probe(dev, dev_id); 885 if (ret != 0) 886 pr_err("probe failed for device %s (%d)\n", 887 dev_name(child_device), ret); 888 889 } else { 890 pr_err("probe not set for driver %s\n", 891 dev_name(child_device)); 892 ret = -ENODEV; 893 } 894 return ret; 895} 896 897/* 898 * vmbus_dma_configure -- Configure DMA coherence for VMbus device 899 */ 900static int vmbus_dma_configure(struct device *child_device) 901{ 902 /* 903 * On ARM64, propagate the DMA coherence setting from the top level 904 * VMbus ACPI device to the child VMbus device being added here. 905 * On x86/x64 coherence is assumed and these calls have no effect. 906 */ 907 hv_setup_dma_ops(child_device, 908 device_get_dma_attr(vmbus_root_device) == DEV_DMA_COHERENT); 909 return 0; 910} 911 912/* 913 * vmbus_remove - Remove a vmbus device 914 */ 915static void vmbus_remove(struct device *child_device) 916{ 917 struct hv_driver *drv; 918 struct hv_device *dev = device_to_hv_device(child_device); 919 920 if (child_device->driver) { 921 drv = drv_to_hv_drv(child_device->driver); 922 if (drv->remove) 923 drv->remove(dev); 924 } 925} 926 927/* 928 * vmbus_shutdown - Shutdown a vmbus device 929 */ 930static void vmbus_shutdown(struct device *child_device) 931{ 932 struct hv_driver *drv; 933 struct hv_device *dev = device_to_hv_device(child_device); 934 935 936 /* The device may not be attached yet */ 937 if (!child_device->driver) 938 return; 939 940 drv = drv_to_hv_drv(child_device->driver); 941 942 if (drv->shutdown) 943 drv->shutdown(dev); 944} 945 946#ifdef CONFIG_PM_SLEEP 947/* 948 * vmbus_suspend - Suspend a vmbus device 949 */ 950static int vmbus_suspend(struct device *child_device) 951{ 952 struct hv_driver *drv; 953 struct hv_device *dev = device_to_hv_device(child_device); 954 955 /* The device may not be attached yet */ 956 if (!child_device->driver) 957 return 0; 958 959 drv = drv_to_hv_drv(child_device->driver); 960 if (!drv->suspend) 961 return -EOPNOTSUPP; 962 963 return drv->suspend(dev); 964} 965 966/* 967 * vmbus_resume - Resume a vmbus device 968 */ 969static int vmbus_resume(struct device *child_device) 970{ 971 struct hv_driver *drv; 972 struct hv_device *dev = device_to_hv_device(child_device); 973 974 /* The device may not be attached yet */ 975 if (!child_device->driver) 976 return 0; 977 978 drv = drv_to_hv_drv(child_device->driver); 979 if (!drv->resume) 980 return -EOPNOTSUPP; 981 982 return drv->resume(dev); 983} 984#else 985#define vmbus_suspend NULL 986#define vmbus_resume NULL 987#endif /* CONFIG_PM_SLEEP */ 988 989/* 990 * vmbus_device_release - Final callback release of the vmbus child device 991 */ 992static void vmbus_device_release(struct device *device) 993{ 994 struct hv_device *hv_dev = device_to_hv_device(device); 995 struct vmbus_channel *channel = hv_dev->channel; 996 997 hv_debug_rm_dev_dir(hv_dev); 998 999 mutex_lock(&vmbus_connection.channel_mutex); 1000 hv_process_channel_removal(channel); 1001 mutex_unlock(&vmbus_connection.channel_mutex); 1002 kfree(hv_dev); 1003} 1004 1005/* 1006 * Note: we must use the "noirq" ops: see the comment before vmbus_bus_pm. 1007 * 1008 * suspend_noirq/resume_noirq are set to NULL to support Suspend-to-Idle: we 1009 * shouldn't suspend the vmbus devices upon Suspend-to-Idle, otherwise there 1010 * is no way to wake up a Generation-2 VM. 1011 * 1012 * The other 4 ops are for hibernation. 1013 */ 1014 1015static const struct dev_pm_ops vmbus_pm = { 1016 .suspend_noirq = NULL, 1017 .resume_noirq = NULL, 1018 .freeze_noirq = vmbus_suspend, 1019 .thaw_noirq = vmbus_resume, 1020 .poweroff_noirq = vmbus_suspend, 1021 .restore_noirq = vmbus_resume, 1022}; 1023 1024/* The one and only one */ 1025static const struct bus_type hv_bus = { 1026 .name = "vmbus", 1027 .match = vmbus_match, 1028 .shutdown = vmbus_shutdown, 1029 .remove = vmbus_remove, 1030 .probe = vmbus_probe, 1031 .uevent = vmbus_uevent, 1032 .dma_configure = vmbus_dma_configure, 1033 .dev_groups = vmbus_dev_groups, 1034 .drv_groups = vmbus_drv_groups, 1035 .bus_groups = vmbus_bus_groups, 1036 .pm = &vmbus_pm, 1037}; 1038 1039struct onmessage_work_context { 1040 struct work_struct work; 1041 struct { 1042 struct hv_message_header header; 1043 u8 payload[]; 1044 } msg; 1045}; 1046 1047static void vmbus_onmessage_work(struct work_struct *work) 1048{ 1049 struct onmessage_work_context *ctx; 1050 1051 /* Do not process messages if we're in DISCONNECTED state */ 1052 if (vmbus_connection.conn_state == DISCONNECTED) 1053 return; 1054 1055 ctx = container_of(work, struct onmessage_work_context, 1056 work); 1057 vmbus_onmessage((struct vmbus_channel_message_header *) 1058 &ctx->msg.payload); 1059 kfree(ctx); 1060} 1061 1062static void __vmbus_on_msg_dpc(void *message_page_addr) 1063{ 1064 struct hv_message msg_copy, *msg; 1065 struct vmbus_channel_message_header *hdr; 1066 enum vmbus_channel_message_type msgtype; 1067 const struct vmbus_channel_message_table_entry *entry; 1068 struct onmessage_work_context *ctx; 1069 __u8 payload_size; 1070 u32 message_type; 1071 1072 if (!message_page_addr) 1073 return; 1074 msg = (struct hv_message *)message_page_addr + VMBUS_MESSAGE_SINT; 1075 1076 /* 1077 * 'enum vmbus_channel_message_type' is supposed to always be 'u32' as 1078 * it is being used in 'struct vmbus_channel_message_header' definition 1079 * which is supposed to match hypervisor ABI. 1080 */ 1081 BUILD_BUG_ON(sizeof(enum vmbus_channel_message_type) != sizeof(u32)); 1082 1083 /* 1084 * Since the message is in memory shared with the host, an erroneous or 1085 * malicious Hyper-V could modify the message while vmbus_on_msg_dpc() 1086 * or individual message handlers are executing; to prevent this, copy 1087 * the message into private memory. 1088 */ 1089 memcpy(&msg_copy, msg, sizeof(struct hv_message)); 1090 1091 message_type = msg_copy.header.message_type; 1092 if (message_type == HVMSG_NONE) 1093 /* no msg */ 1094 return; 1095 1096 hdr = (struct vmbus_channel_message_header *)msg_copy.u.payload; 1097 msgtype = hdr->msgtype; 1098 1099 trace_vmbus_on_msg_dpc(hdr); 1100 1101 if (msgtype >= CHANNELMSG_COUNT) { 1102 WARN_ONCE(1, "unknown msgtype=%d\n", msgtype); 1103 goto msg_handled; 1104 } 1105 1106 payload_size = msg_copy.header.payload_size; 1107 if (payload_size > HV_MESSAGE_PAYLOAD_BYTE_COUNT) { 1108 WARN_ONCE(1, "payload size is too large (%d)\n", payload_size); 1109 goto msg_handled; 1110 } 1111 1112 entry = &channel_message_table[msgtype]; 1113 1114 if (!entry->message_handler) 1115 goto msg_handled; 1116 1117 if (payload_size < entry->min_payload_len) { 1118 WARN_ONCE(1, "message too short: msgtype=%d len=%d\n", msgtype, payload_size); 1119 goto msg_handled; 1120 } 1121 1122 if (entry->handler_type == VMHT_BLOCKING) { 1123 ctx = kmalloc(struct_size(ctx, msg.payload, payload_size), GFP_ATOMIC); 1124 if (ctx == NULL) 1125 return; 1126 1127 INIT_WORK(&ctx->work, vmbus_onmessage_work); 1128 ctx->msg.header = msg_copy.header; 1129 memcpy(&ctx->msg.payload, msg_copy.u.payload, payload_size); 1130 1131 /* 1132 * The host can generate a rescind message while we 1133 * may still be handling the original offer. We deal with 1134 * this condition by relying on the synchronization provided 1135 * by offer_in_progress and by channel_mutex. See also the 1136 * inline comments in vmbus_onoffer_rescind(). 1137 */ 1138 switch (msgtype) { 1139 case CHANNELMSG_RESCIND_CHANNELOFFER: 1140 /* 1141 * If we are handling the rescind message; 1142 * schedule the work on the global work queue. 1143 * 1144 * The OFFER message and the RESCIND message should 1145 * not be handled by the same serialized work queue, 1146 * because the OFFER handler may call vmbus_open(), 1147 * which tries to open the channel by sending an 1148 * OPEN_CHANNEL message to the host and waits for 1149 * the host's response; however, if the host has 1150 * rescinded the channel before it receives the 1151 * OPEN_CHANNEL message, the host just silently 1152 * ignores the OPEN_CHANNEL message; as a result, 1153 * the guest's OFFER handler hangs for ever, if we 1154 * handle the RESCIND message in the same serialized 1155 * work queue: the RESCIND handler can not start to 1156 * run before the OFFER handler finishes. 1157 */ 1158 if (vmbus_connection.ignore_any_offer_msg) 1159 break; 1160 queue_work(vmbus_connection.rescind_work_queue, &ctx->work); 1161 break; 1162 1163 case CHANNELMSG_OFFERCHANNEL: 1164 /* 1165 * The host sends the offer message of a given channel 1166 * before sending the rescind message of the same 1167 * channel. These messages are sent to the guest's 1168 * connect CPU; the guest then starts processing them 1169 * in the tasklet handler on this CPU: 1170 * 1171 * VMBUS_CONNECT_CPU 1172 * 1173 * [vmbus_on_msg_dpc()] 1174 * atomic_inc() // CHANNELMSG_OFFERCHANNEL 1175 * queue_work() 1176 * ... 1177 * [vmbus_on_msg_dpc()] 1178 * schedule_work() // CHANNELMSG_RESCIND_CHANNELOFFER 1179 * 1180 * We rely on the memory-ordering properties of the 1181 * queue_work() and schedule_work() primitives, which 1182 * guarantee that the atomic increment will be visible 1183 * to the CPUs which will execute the offer & rescind 1184 * works by the time these works will start execution. 1185 */ 1186 if (vmbus_connection.ignore_any_offer_msg) 1187 break; 1188 atomic_inc(&vmbus_connection.offer_in_progress); 1189 fallthrough; 1190 1191 default: 1192 queue_work(vmbus_connection.work_queue, &ctx->work); 1193 } 1194 } else 1195 entry->message_handler(hdr); 1196 1197msg_handled: 1198 vmbus_signal_eom(msg, message_type); 1199} 1200 1201void vmbus_on_msg_dpc(unsigned long data) 1202{ 1203 struct hv_per_cpu_context *hv_cpu = (void *)data; 1204 1205 __vmbus_on_msg_dpc(hv_cpu->hyp_synic_message_page); 1206 __vmbus_on_msg_dpc(hv_cpu->para_synic_message_page); 1207} 1208 1209#ifdef CONFIG_PM_SLEEP 1210/* 1211 * Fake RESCIND_CHANNEL messages to clean up hv_sock channels by force for 1212 * hibernation, because hv_sock connections can not persist across hibernation. 1213 */ 1214static void vmbus_force_channel_rescinded(struct vmbus_channel *channel) 1215{ 1216 struct onmessage_work_context *ctx; 1217 struct vmbus_channel_rescind_offer *rescind; 1218 1219 WARN_ON(!is_hvsock_channel(channel)); 1220 1221 /* 1222 * Allocation size is small and the allocation should really not fail, 1223 * otherwise the state of the hv_sock connections ends up in limbo. 1224 */ 1225 ctx = kzalloc(sizeof(*ctx) + sizeof(*rescind), 1226 GFP_KERNEL | __GFP_NOFAIL); 1227 1228 /* 1229 * So far, these are not really used by Linux. Just set them to the 1230 * reasonable values conforming to the definitions of the fields. 1231 */ 1232 ctx->msg.header.message_type = 1; 1233 ctx->msg.header.payload_size = sizeof(*rescind); 1234 1235 /* These values are actually used by Linux. */ 1236 rescind = (struct vmbus_channel_rescind_offer *)ctx->msg.payload; 1237 rescind->header.msgtype = CHANNELMSG_RESCIND_CHANNELOFFER; 1238 rescind->child_relid = channel->offermsg.child_relid; 1239 1240 INIT_WORK(&ctx->work, vmbus_onmessage_work); 1241 1242 queue_work(vmbus_connection.work_queue, &ctx->work); 1243} 1244#endif /* CONFIG_PM_SLEEP */ 1245 1246/* 1247 * Schedule all channels with events pending. 1248 * The event page can be directly checked to get the id of 1249 * the channel that has the interrupt pending. 1250 */ 1251static void vmbus_chan_sched(void *event_page_addr) 1252{ 1253 unsigned long *recv_int_page; 1254 u32 maxbits, relid; 1255 union hv_synic_event_flags *event; 1256 1257 if (!event_page_addr) 1258 return; 1259 event = (union hv_synic_event_flags *)event_page_addr + VMBUS_MESSAGE_SINT; 1260 1261 maxbits = HV_EVENT_FLAGS_COUNT; 1262 recv_int_page = event->flags; 1263 1264 if (unlikely(!recv_int_page)) 1265 return; 1266 1267 /* 1268 * Suggested-by: Michael Kelley <mhklinux@outlook.com> 1269 * One possible optimization would be to keep track of the largest relID that's in use, 1270 * and only scan up to that relID. 1271 */ 1272 for_each_set_bit(relid, recv_int_page, maxbits) { 1273 void (*callback_fn)(void *context); 1274 struct vmbus_channel *channel; 1275 1276 if (!sync_test_and_clear_bit(relid, recv_int_page)) 1277 continue; 1278 1279 /* Special case - vmbus channel protocol msg */ 1280 if (relid == 0) 1281 continue; 1282 1283 /* 1284 * Pairs with the kfree_rcu() in vmbus_chan_release(). 1285 * Guarantees that the channel data structure doesn't 1286 * get freed while the channel pointer below is being 1287 * dereferenced. 1288 */ 1289 rcu_read_lock(); 1290 1291 /* Find channel based on relid */ 1292 channel = relid2channel(relid); 1293 if (channel == NULL) 1294 goto sched_unlock_rcu; 1295 1296 if (channel->rescind) 1297 goto sched_unlock_rcu; 1298 1299 /* 1300 * Make sure that the ring buffer data structure doesn't get 1301 * freed while we dereference the ring buffer pointer. Test 1302 * for the channel's onchannel_callback being NULL within a 1303 * sched_lock critical section. See also the inline comments 1304 * in vmbus_reset_channel_cb(). 1305 */ 1306 spin_lock(&channel->sched_lock); 1307 1308 callback_fn = channel->onchannel_callback; 1309 if (unlikely(callback_fn == NULL)) 1310 goto sched_unlock; 1311 1312 trace_vmbus_chan_sched(channel); 1313 1314 ++channel->interrupts; 1315 1316 switch (channel->callback_mode) { 1317 case HV_CALL_ISR: 1318 (*callback_fn)(channel->channel_callback_context); 1319 break; 1320 1321 case HV_CALL_BATCHED: 1322 hv_begin_read(&channel->inbound); 1323 fallthrough; 1324 case HV_CALL_DIRECT: 1325 tasklet_schedule(&channel->callback_event); 1326 } 1327 1328sched_unlock: 1329 spin_unlock(&channel->sched_lock); 1330sched_unlock_rcu: 1331 rcu_read_unlock(); 1332 } 1333} 1334 1335static void vmbus_message_sched(struct hv_per_cpu_context *hv_cpu, void *message_page_addr) 1336{ 1337 struct hv_message *msg; 1338 1339 if (!message_page_addr) 1340 return; 1341 msg = (struct hv_message *)message_page_addr + VMBUS_MESSAGE_SINT; 1342 1343 /* Check if there are actual msgs to be processed */ 1344 if (msg->header.message_type != HVMSG_NONE) { 1345 if (msg->header.message_type == HVMSG_TIMER_EXPIRED) { 1346 hv_stimer0_isr(); 1347 vmbus_signal_eom(msg, HVMSG_TIMER_EXPIRED); 1348 } else { 1349 tasklet_schedule(&hv_cpu->msg_dpc); 1350 } 1351 } 1352} 1353 1354static void __vmbus_isr(void) 1355{ 1356 struct hv_per_cpu_context *hv_cpu 1357 = this_cpu_ptr(hv_context.cpu_context); 1358 1359 vmbus_chan_sched(hv_cpu->hyp_synic_event_page); 1360 vmbus_chan_sched(hv_cpu->para_synic_event_page); 1361 1362 vmbus_message_sched(hv_cpu, hv_cpu->hyp_synic_message_page); 1363 vmbus_message_sched(hv_cpu, hv_cpu->para_synic_message_page); 1364 1365 add_interrupt_randomness(vmbus_interrupt); 1366} 1367 1368static DEFINE_PER_CPU(bool, vmbus_irq_pending); 1369static DEFINE_PER_CPU(struct task_struct *, vmbus_irqd); 1370 1371static void vmbus_irqd_wake(void) 1372{ 1373 struct task_struct *tsk = __this_cpu_read(vmbus_irqd); 1374 1375 __this_cpu_write(vmbus_irq_pending, true); 1376 wake_up_process(tsk); 1377} 1378 1379static void vmbus_irqd_setup(unsigned int cpu) 1380{ 1381 sched_set_fifo(current); 1382} 1383 1384static int vmbus_irqd_should_run(unsigned int cpu) 1385{ 1386 return __this_cpu_read(vmbus_irq_pending); 1387} 1388 1389static void run_vmbus_irqd(unsigned int cpu) 1390{ 1391 __this_cpu_write(vmbus_irq_pending, false); 1392 __vmbus_isr(); 1393} 1394 1395static bool vmbus_irq_initialized; 1396 1397static struct smp_hotplug_thread vmbus_irq_threads = { 1398 .store = &vmbus_irqd, 1399 .setup = vmbus_irqd_setup, 1400 .thread_should_run = vmbus_irqd_should_run, 1401 .thread_fn = run_vmbus_irqd, 1402 .thread_comm = "vmbus_irq/%u", 1403}; 1404 1405void vmbus_isr(void) 1406{ 1407 if (IS_ENABLED(CONFIG_PREEMPT_RT)) { 1408 vmbus_irqd_wake(); 1409 } else { 1410 lockdep_hardirq_threaded(); 1411 __vmbus_isr(); 1412 } 1413} 1414EXPORT_SYMBOL_FOR_MODULES(vmbus_isr, "mshv_vtl"); 1415 1416static irqreturn_t vmbus_percpu_isr(int irq, void *dev_id) 1417{ 1418 vmbus_isr(); 1419 return IRQ_HANDLED; 1420} 1421 1422static void vmbus_percpu_work(struct work_struct *work) 1423{ 1424 unsigned int cpu = smp_processor_id(); 1425 1426 hv_synic_init(cpu); 1427} 1428 1429static int vmbus_alloc_synic_and_connect(void) 1430{ 1431 int ret, cpu; 1432 struct work_struct __percpu *works; 1433 int hyperv_cpuhp_online; 1434 1435 ret = hv_synic_alloc(); 1436 if (ret < 0) 1437 goto err_alloc; 1438 1439 works = alloc_percpu(struct work_struct); 1440 if (!works) { 1441 ret = -ENOMEM; 1442 goto err_alloc; 1443 } 1444 1445 /* 1446 * Initialize the per-cpu interrupt state and stimer state. 1447 * Then connect to the host. 1448 */ 1449 cpus_read_lock(); 1450 for_each_online_cpu(cpu) { 1451 struct work_struct *work = per_cpu_ptr(works, cpu); 1452 1453 INIT_WORK(work, vmbus_percpu_work); 1454 schedule_work_on(cpu, work); 1455 } 1456 1457 for_each_online_cpu(cpu) 1458 flush_work(per_cpu_ptr(works, cpu)); 1459 1460 /* Register the callbacks for possible CPU online/offline'ing */ 1461 ret = cpuhp_setup_state_nocalls_cpuslocked(CPUHP_AP_ONLINE_DYN, "hyperv/vmbus:online", 1462 hv_synic_init, hv_synic_cleanup); 1463 cpus_read_unlock(); 1464 free_percpu(works); 1465 if (ret < 0) 1466 goto err_alloc; 1467 hyperv_cpuhp_online = ret; 1468 1469 ret = vmbus_connect(); 1470 if (ret) 1471 goto err_connect; 1472 return 0; 1473 1474err_connect: 1475 cpuhp_remove_state(hyperv_cpuhp_online); 1476 return -ENODEV; 1477err_alloc: 1478 hv_synic_free(); 1479 return -ENOMEM; 1480} 1481 1482/* 1483 * vmbus_bus_init -Main vmbus driver initialization routine. 1484 * 1485 * Here, we 1486 * - initialize the vmbus driver context 1487 * - invoke the vmbus hv main init routine 1488 * - retrieve the channel offers 1489 */ 1490static int vmbus_bus_init(void) 1491{ 1492 int ret; 1493 1494 ret = hv_init(); 1495 if (ret != 0) { 1496 pr_err("Unable to initialize the hypervisor - 0x%x\n", ret); 1497 return ret; 1498 } 1499 1500 ret = bus_register(&hv_bus); 1501 if (ret) 1502 return ret; 1503 1504 /* 1505 * VMbus interrupts are best modeled as per-cpu interrupts. If 1506 * on an architecture with support for per-cpu IRQs (e.g. ARM64), 1507 * allocate a per-cpu IRQ using standard Linux kernel functionality. 1508 * If not on such an architecture (e.g., x86/x64), then rely on 1509 * code in the arch-specific portion of the code tree to connect 1510 * the VMbus interrupt handler. 1511 */ 1512 1513 if (IS_ENABLED(CONFIG_PREEMPT_RT) && !vmbus_irq_initialized) { 1514 ret = smpboot_register_percpu_thread(&vmbus_irq_threads); 1515 if (ret) 1516 goto err_kthread; 1517 vmbus_irq_initialized = true; 1518 } 1519 1520 if (vmbus_irq == -1) { 1521 hv_setup_vmbus_handler(vmbus_isr); 1522 } else { 1523 vmbus_evt = alloc_percpu(long); 1524 ret = request_percpu_irq(vmbus_irq, vmbus_percpu_isr, 1525 "Hyper-V VMbus", vmbus_evt); 1526 if (ret) { 1527 pr_err("Can't request Hyper-V VMbus IRQ %d, Err %d", 1528 vmbus_irq, ret); 1529 free_percpu(vmbus_evt); 1530 goto err_setup; 1531 } 1532 } 1533 1534 /* 1535 * Cache the value as getting it involves a VM exit on x86(_64), and 1536 * doing that on each VP while initializing SynIC's wastes time. 1537 */ 1538 is_confidential = ms_hyperv.confidential_vmbus_available; 1539 if (is_confidential) 1540 pr_info("Establishing connection to the confidential VMBus\n"); 1541 hv_para_set_sint_proxy(!is_confidential); 1542 ret = vmbus_alloc_synic_and_connect(); 1543 if (ret) 1544 goto err_connect; 1545 1546 /* 1547 * Always register the vmbus unload panic notifier because we 1548 * need to shut the VMbus channel connection on panic. 1549 */ 1550 atomic_notifier_chain_register(&panic_notifier_list, 1551 &hyperv_panic_vmbus_unload_block); 1552 1553 vmbus_request_offers(); 1554 1555 return 0; 1556 1557err_connect: 1558 if (vmbus_irq == -1) { 1559 hv_remove_vmbus_handler(); 1560 } else { 1561 free_percpu_irq(vmbus_irq, vmbus_evt); 1562 free_percpu(vmbus_evt); 1563 } 1564err_setup: 1565 if (IS_ENABLED(CONFIG_PREEMPT_RT) && vmbus_irq_initialized) { 1566 smpboot_unregister_percpu_thread(&vmbus_irq_threads); 1567 vmbus_irq_initialized = false; 1568 } 1569err_kthread: 1570 bus_unregister(&hv_bus); 1571 return ret; 1572} 1573 1574/** 1575 * __vmbus_driver_register() - Register a vmbus's driver 1576 * @hv_driver: Pointer to driver structure you want to register 1577 * @owner: owner module of the drv 1578 * @mod_name: module name string 1579 * 1580 * Registers the given driver with Linux through the 'driver_register()' call 1581 * and sets up the hyper-v vmbus handling for this driver. 1582 * It will return the state of the 'driver_register()' call. 1583 * 1584 */ 1585int __vmbus_driver_register(struct hv_driver *hv_driver, struct module *owner, const char *mod_name) 1586{ 1587 int ret; 1588 1589 pr_info("registering driver %s\n", hv_driver->name); 1590 1591 ret = vmbus_exists(); 1592 if (ret < 0) 1593 return ret; 1594 1595 hv_driver->driver.name = hv_driver->name; 1596 hv_driver->driver.owner = owner; 1597 hv_driver->driver.mod_name = mod_name; 1598 hv_driver->driver.bus = &hv_bus; 1599 1600 spin_lock_init(&hv_driver->dynids.lock); 1601 INIT_LIST_HEAD(&hv_driver->dynids.list); 1602 1603 ret = driver_register(&hv_driver->driver); 1604 1605 return ret; 1606} 1607EXPORT_SYMBOL_GPL(__vmbus_driver_register); 1608 1609/** 1610 * vmbus_driver_unregister() - Unregister a vmbus's driver 1611 * @hv_driver: Pointer to driver structure you want to 1612 * un-register 1613 * 1614 * Un-register the given driver that was previous registered with a call to 1615 * vmbus_driver_register() 1616 */ 1617void vmbus_driver_unregister(struct hv_driver *hv_driver) 1618{ 1619 pr_info("unregistering driver %s\n", hv_driver->name); 1620 1621 if (!vmbus_exists()) { 1622 driver_unregister(&hv_driver->driver); 1623 vmbus_free_dynids(hv_driver); 1624 } 1625} 1626EXPORT_SYMBOL_GPL(vmbus_driver_unregister); 1627 1628 1629/* 1630 * Called when last reference to channel is gone. 1631 */ 1632static void vmbus_chan_release(struct kobject *kobj) 1633{ 1634 struct vmbus_channel *channel 1635 = container_of(kobj, struct vmbus_channel, kobj); 1636 1637 kfree_rcu(channel, rcu); 1638} 1639 1640struct vmbus_chan_attribute { 1641 struct attribute attr; 1642 ssize_t (*show)(struct vmbus_channel *chan, char *buf); 1643 ssize_t (*store)(struct vmbus_channel *chan, 1644 const char *buf, size_t count); 1645}; 1646#define VMBUS_CHAN_ATTR(_name, _mode, _show, _store) \ 1647 struct vmbus_chan_attribute chan_attr_##_name \ 1648 = __ATTR(_name, _mode, _show, _store) 1649#define VMBUS_CHAN_ATTR_RW(_name) \ 1650 struct vmbus_chan_attribute chan_attr_##_name = __ATTR_RW(_name) 1651#define VMBUS_CHAN_ATTR_RO(_name) \ 1652 struct vmbus_chan_attribute chan_attr_##_name = __ATTR_RO(_name) 1653#define VMBUS_CHAN_ATTR_WO(_name) \ 1654 struct vmbus_chan_attribute chan_attr_##_name = __ATTR_WO(_name) 1655 1656static ssize_t vmbus_chan_attr_show(struct kobject *kobj, 1657 struct attribute *attr, char *buf) 1658{ 1659 const struct vmbus_chan_attribute *attribute 1660 = container_of(attr, struct vmbus_chan_attribute, attr); 1661 struct vmbus_channel *chan 1662 = container_of(kobj, struct vmbus_channel, kobj); 1663 1664 if (!attribute->show) 1665 return -EIO; 1666 1667 return attribute->show(chan, buf); 1668} 1669 1670static ssize_t vmbus_chan_attr_store(struct kobject *kobj, 1671 struct attribute *attr, const char *buf, 1672 size_t count) 1673{ 1674 const struct vmbus_chan_attribute *attribute 1675 = container_of(attr, struct vmbus_chan_attribute, attr); 1676 struct vmbus_channel *chan 1677 = container_of(kobj, struct vmbus_channel, kobj); 1678 1679 if (!attribute->store) 1680 return -EIO; 1681 1682 return attribute->store(chan, buf, count); 1683} 1684 1685static const struct sysfs_ops vmbus_chan_sysfs_ops = { 1686 .show = vmbus_chan_attr_show, 1687 .store = vmbus_chan_attr_store, 1688}; 1689 1690static ssize_t out_mask_show(struct vmbus_channel *channel, char *buf) 1691{ 1692 struct hv_ring_buffer_info *rbi = &channel->outbound; 1693 ssize_t ret; 1694 1695 mutex_lock(&rbi->ring_buffer_mutex); 1696 if (!rbi->ring_buffer) { 1697 mutex_unlock(&rbi->ring_buffer_mutex); 1698 return -EINVAL; 1699 } 1700 1701 ret = sprintf(buf, "%u\n", rbi->ring_buffer->interrupt_mask); 1702 mutex_unlock(&rbi->ring_buffer_mutex); 1703 return ret; 1704} 1705static VMBUS_CHAN_ATTR_RO(out_mask); 1706 1707static ssize_t in_mask_show(struct vmbus_channel *channel, char *buf) 1708{ 1709 struct hv_ring_buffer_info *rbi = &channel->inbound; 1710 ssize_t ret; 1711 1712 mutex_lock(&rbi->ring_buffer_mutex); 1713 if (!rbi->ring_buffer) { 1714 mutex_unlock(&rbi->ring_buffer_mutex); 1715 return -EINVAL; 1716 } 1717 1718 ret = sprintf(buf, "%u\n", rbi->ring_buffer->interrupt_mask); 1719 mutex_unlock(&rbi->ring_buffer_mutex); 1720 return ret; 1721} 1722static VMBUS_CHAN_ATTR_RO(in_mask); 1723 1724static ssize_t read_avail_show(struct vmbus_channel *channel, char *buf) 1725{ 1726 struct hv_ring_buffer_info *rbi = &channel->inbound; 1727 ssize_t ret; 1728 1729 mutex_lock(&rbi->ring_buffer_mutex); 1730 if (!rbi->ring_buffer) { 1731 mutex_unlock(&rbi->ring_buffer_mutex); 1732 return -EINVAL; 1733 } 1734 1735 ret = sprintf(buf, "%u\n", hv_get_bytes_to_read(rbi)); 1736 mutex_unlock(&rbi->ring_buffer_mutex); 1737 return ret; 1738} 1739static VMBUS_CHAN_ATTR_RO(read_avail); 1740 1741static ssize_t write_avail_show(struct vmbus_channel *channel, char *buf) 1742{ 1743 struct hv_ring_buffer_info *rbi = &channel->outbound; 1744 ssize_t ret; 1745 1746 mutex_lock(&rbi->ring_buffer_mutex); 1747 if (!rbi->ring_buffer) { 1748 mutex_unlock(&rbi->ring_buffer_mutex); 1749 return -EINVAL; 1750 } 1751 1752 ret = sprintf(buf, "%u\n", hv_get_bytes_to_write(rbi)); 1753 mutex_unlock(&rbi->ring_buffer_mutex); 1754 return ret; 1755} 1756static VMBUS_CHAN_ATTR_RO(write_avail); 1757 1758static ssize_t target_cpu_show(struct vmbus_channel *channel, char *buf) 1759{ 1760 return sprintf(buf, "%u\n", channel->target_cpu); 1761} 1762 1763int vmbus_channel_set_cpu(struct vmbus_channel *channel, u32 target_cpu) 1764{ 1765 u32 origin_cpu; 1766 int ret = 0; 1767 1768 lockdep_assert_cpus_held(); 1769 lockdep_assert_held(&vmbus_connection.channel_mutex); 1770 1771 if (vmbus_proto_version < VERSION_WIN10_V4_1) 1772 return -EIO; 1773 1774 /* Validate target_cpu for the cpumask_test_cpu() operation below. */ 1775 if (target_cpu >= nr_cpumask_bits) 1776 return -EINVAL; 1777 1778 if (!cpumask_test_cpu(target_cpu, housekeeping_cpumask(HK_TYPE_MANAGED_IRQ))) 1779 return -EINVAL; 1780 1781 if (!cpu_online(target_cpu)) 1782 return -EINVAL; 1783 1784 /* 1785 * Synchronizes vmbus_channel_set_cpu() and channel closure: 1786 * 1787 * { Initially: state = CHANNEL_OPENED } 1788 * 1789 * CPU1 CPU2 1790 * 1791 * [vmbus_channel_set_cpu()] [vmbus_disconnect_ring()] 1792 * 1793 * LOCK channel_mutex LOCK channel_mutex 1794 * LOAD r1 = state LOAD r2 = state 1795 * IF (r1 == CHANNEL_OPENED) IF (r2 == CHANNEL_OPENED) 1796 * SEND MODIFYCHANNEL STORE state = CHANNEL_OPEN 1797 * [...] SEND CLOSECHANNEL 1798 * UNLOCK channel_mutex UNLOCK channel_mutex 1799 * 1800 * Forbids: r1 == r2 == CHANNEL_OPENED (i.e., CPU1's LOCK precedes 1801 * CPU2's LOCK) && CPU2's SEND precedes CPU1's SEND 1802 * 1803 * Note. The host processes the channel messages "sequentially", in 1804 * the order in which they are received on a per-partition basis. 1805 */ 1806 1807 /* 1808 * Hyper-V will ignore MODIFYCHANNEL messages for "non-open" channels; 1809 * avoid sending the message and fail here for such channels. 1810 */ 1811 if (channel->state != CHANNEL_OPENED_STATE) { 1812 ret = -EIO; 1813 goto end; 1814 } 1815 1816 origin_cpu = channel->target_cpu; 1817 if (target_cpu == origin_cpu) 1818 goto end; 1819 1820 if (vmbus_send_modifychannel(channel, 1821 hv_cpu_number_to_vp_number(target_cpu))) { 1822 ret = -EIO; 1823 goto end; 1824 } 1825 1826 /* 1827 * For version before VERSION_WIN10_V5_3, the following warning holds: 1828 * 1829 * Warning. At this point, there is *no* guarantee that the host will 1830 * have successfully processed the vmbus_send_modifychannel() request. 1831 * See the header comment of vmbus_send_modifychannel() for more info. 1832 * 1833 * Lags in the processing of the above vmbus_send_modifychannel() can 1834 * result in missed interrupts if the "old" target CPU is taken offline 1835 * before Hyper-V starts sending interrupts to the "new" target CPU. 1836 * But apart from this offlining scenario, the code tolerates such 1837 * lags. It will function correctly even if a channel interrupt comes 1838 * in on a CPU that is different from the channel target_cpu value. 1839 */ 1840 1841 channel->target_cpu = target_cpu; 1842 1843 /* See init_vp_index(). */ 1844 if (hv_is_perf_channel(channel)) 1845 hv_update_allocated_cpus(origin_cpu, target_cpu); 1846 1847 /* Currently set only for storvsc channels. */ 1848 if (channel->change_target_cpu_callback) { 1849 (*channel->change_target_cpu_callback)(channel, 1850 origin_cpu, target_cpu); 1851 } 1852 1853end: 1854 return ret; 1855} 1856 1857static ssize_t target_cpu_store(struct vmbus_channel *channel, 1858 const char *buf, size_t count) 1859{ 1860 u32 target_cpu; 1861 ssize_t ret; 1862 1863 if (sscanf(buf, "%u", &target_cpu) != 1) 1864 return -EIO; 1865 1866 cpus_read_lock(); 1867 mutex_lock(&vmbus_connection.channel_mutex); 1868 ret = vmbus_channel_set_cpu(channel, target_cpu); 1869 mutex_unlock(&vmbus_connection.channel_mutex); 1870 cpus_read_unlock(); 1871 1872 return ret ?: count; 1873} 1874static VMBUS_CHAN_ATTR(cpu, 0644, target_cpu_show, target_cpu_store); 1875 1876static ssize_t channel_pending_show(struct vmbus_channel *channel, 1877 char *buf) 1878{ 1879 return sprintf(buf, "%d\n", 1880 channel_pending(channel, 1881 vmbus_connection.monitor_pages[1])); 1882} 1883static VMBUS_CHAN_ATTR(pending, 0444, channel_pending_show, NULL); 1884 1885static ssize_t channel_latency_show(struct vmbus_channel *channel, 1886 char *buf) 1887{ 1888 return sprintf(buf, "%d\n", 1889 channel_latency(channel, 1890 vmbus_connection.monitor_pages[1])); 1891} 1892static VMBUS_CHAN_ATTR(latency, 0444, channel_latency_show, NULL); 1893 1894static ssize_t channel_interrupts_show(struct vmbus_channel *channel, char *buf) 1895{ 1896 return sprintf(buf, "%llu\n", channel->interrupts); 1897} 1898static VMBUS_CHAN_ATTR(interrupts, 0444, channel_interrupts_show, NULL); 1899 1900static ssize_t channel_events_show(struct vmbus_channel *channel, char *buf) 1901{ 1902 return sprintf(buf, "%llu\n", channel->sig_events); 1903} 1904static VMBUS_CHAN_ATTR(events, 0444, channel_events_show, NULL); 1905 1906static ssize_t channel_intr_in_full_show(struct vmbus_channel *channel, 1907 char *buf) 1908{ 1909 return sprintf(buf, "%llu\n", 1910 (unsigned long long)channel->intr_in_full); 1911} 1912static VMBUS_CHAN_ATTR(intr_in_full, 0444, channel_intr_in_full_show, NULL); 1913 1914static ssize_t channel_intr_out_empty_show(struct vmbus_channel *channel, 1915 char *buf) 1916{ 1917 return sprintf(buf, "%llu\n", 1918 (unsigned long long)channel->intr_out_empty); 1919} 1920static VMBUS_CHAN_ATTR(intr_out_empty, 0444, channel_intr_out_empty_show, NULL); 1921 1922static ssize_t channel_out_full_first_show(struct vmbus_channel *channel, 1923 char *buf) 1924{ 1925 return sprintf(buf, "%llu\n", 1926 (unsigned long long)channel->out_full_first); 1927} 1928static VMBUS_CHAN_ATTR(out_full_first, 0444, channel_out_full_first_show, NULL); 1929 1930static ssize_t channel_out_full_total_show(struct vmbus_channel *channel, 1931 char *buf) 1932{ 1933 return sprintf(buf, "%llu\n", 1934 (unsigned long long)channel->out_full_total); 1935} 1936static VMBUS_CHAN_ATTR(out_full_total, 0444, channel_out_full_total_show, NULL); 1937 1938static ssize_t subchannel_monitor_id_show(struct vmbus_channel *channel, 1939 char *buf) 1940{ 1941 return sprintf(buf, "%u\n", channel->offermsg.monitorid); 1942} 1943static VMBUS_CHAN_ATTR(monitor_id, 0444, subchannel_monitor_id_show, NULL); 1944 1945static ssize_t subchannel_id_show(struct vmbus_channel *channel, 1946 char *buf) 1947{ 1948 return sprintf(buf, "%u\n", 1949 channel->offermsg.offer.sub_channel_index); 1950} 1951static VMBUS_CHAN_ATTR_RO(subchannel_id); 1952 1953static int hv_mmap_ring_buffer_wrapper(struct file *filp, struct kobject *kobj, 1954 const struct bin_attribute *attr, 1955 struct vm_area_struct *vma) 1956{ 1957 struct vmbus_channel *channel = container_of(kobj, struct vmbus_channel, kobj); 1958 1959 /* 1960 * hv_(create|remove)_ring_sysfs implementation ensures that mmap_ring_buffer 1961 * is not NULL. 1962 */ 1963 return channel->mmap_ring_buffer(channel, vma); 1964} 1965 1966static struct bin_attribute chan_attr_ring_buffer = { 1967 .attr = { 1968 .name = "ring", 1969 .mode = 0600, 1970 }, 1971 .mmap = hv_mmap_ring_buffer_wrapper, 1972}; 1973static struct attribute *vmbus_chan_attrs[] = { 1974 &chan_attr_out_mask.attr, 1975 &chan_attr_in_mask.attr, 1976 &chan_attr_read_avail.attr, 1977 &chan_attr_write_avail.attr, 1978 &chan_attr_cpu.attr, 1979 &chan_attr_pending.attr, 1980 &chan_attr_latency.attr, 1981 &chan_attr_interrupts.attr, 1982 &chan_attr_events.attr, 1983 &chan_attr_intr_in_full.attr, 1984 &chan_attr_intr_out_empty.attr, 1985 &chan_attr_out_full_first.attr, 1986 &chan_attr_out_full_total.attr, 1987 &chan_attr_monitor_id.attr, 1988 &chan_attr_subchannel_id.attr, 1989 NULL 1990}; 1991 1992static const struct bin_attribute *vmbus_chan_bin_attrs[] = { 1993 &chan_attr_ring_buffer, 1994 NULL 1995}; 1996 1997/* 1998 * Channel-level attribute_group callback function. Returns the permission for 1999 * each attribute, and returns 0 if an attribute is not visible. 2000 */ 2001static umode_t vmbus_chan_attr_is_visible(struct kobject *kobj, 2002 struct attribute *attr, int idx) 2003{ 2004 const struct vmbus_channel *channel = 2005 container_of(kobj, struct vmbus_channel, kobj); 2006 2007 /* Hide the monitor attributes if the monitor mechanism is not used. */ 2008 if (!channel->offermsg.monitor_allocated && 2009 (attr == &chan_attr_pending.attr || 2010 attr == &chan_attr_latency.attr || 2011 attr == &chan_attr_monitor_id.attr)) 2012 return 0; 2013 2014 return attr->mode; 2015} 2016 2017static umode_t vmbus_chan_bin_attr_is_visible(struct kobject *kobj, 2018 const struct bin_attribute *attr, int idx) 2019{ 2020 const struct vmbus_channel *channel = 2021 container_of(kobj, struct vmbus_channel, kobj); 2022 2023 /* Hide ring attribute if channel's ring_sysfs_visible is set to false */ 2024 if (attr == &chan_attr_ring_buffer && !channel->ring_sysfs_visible) 2025 return 0; 2026 2027 return attr->attr.mode; 2028} 2029 2030static size_t vmbus_chan_bin_size(struct kobject *kobj, 2031 const struct bin_attribute *bin_attr, int a) 2032{ 2033 const struct vmbus_channel *channel = 2034 container_of(kobj, struct vmbus_channel, kobj); 2035 2036 return channel->ringbuffer_pagecount << PAGE_SHIFT; 2037} 2038 2039static const struct attribute_group vmbus_chan_group = { 2040 .attrs = vmbus_chan_attrs, 2041 .bin_attrs = vmbus_chan_bin_attrs, 2042 .is_visible = vmbus_chan_attr_is_visible, 2043 .is_bin_visible = vmbus_chan_bin_attr_is_visible, 2044 .bin_size = vmbus_chan_bin_size, 2045}; 2046 2047static const struct kobj_type vmbus_chan_ktype = { 2048 .sysfs_ops = &vmbus_chan_sysfs_ops, 2049 .release = vmbus_chan_release, 2050}; 2051 2052/** 2053 * hv_create_ring_sysfs() - create "ring" sysfs entry corresponding to ring buffers for a channel. 2054 * @channel: Pointer to vmbus_channel structure 2055 * @hv_mmap_ring_buffer: function pointer for initializing the function to be called on mmap of 2056 * channel's "ring" sysfs node, which is for the ring buffer of that channel. 2057 * Function pointer is of below type: 2058 * int (*hv_mmap_ring_buffer)(struct vmbus_channel *channel, 2059 * struct vm_area_struct *vma)) 2060 * This has a pointer to the channel and a pointer to vm_area_struct, 2061 * used for mmap, as arguments. 2062 * 2063 * Sysfs node for ring buffer of a channel is created along with other fields, however its 2064 * visibility is disabled by default. Sysfs creation needs to be controlled when the use-case 2065 * is running. 2066 * For example, HV_NIC device is used either by uio_hv_generic or hv_netvsc at any given point of 2067 * time, and "ring" sysfs is needed only when uio_hv_generic is bound to that device. To avoid 2068 * exposing the ring buffer by default, this function is responsible to enable visibility of 2069 * ring for userspace to use. 2070 * Note: Race conditions can happen with userspace and it is not encouraged to create new 2071 * use-cases for this. This was added to maintain backward compatibility, while solving 2072 * one of the race conditions in uio_hv_generic while creating sysfs. See comments with 2073 * vmbus_add_dynid() and vmbus_device_register(). 2074 * 2075 * Returns 0 on success or error code on failure. 2076 */ 2077int hv_create_ring_sysfs(struct vmbus_channel *channel, 2078 int (*hv_mmap_ring_buffer)(struct vmbus_channel *channel, 2079 struct vm_area_struct *vma)) 2080{ 2081 struct kobject *kobj = &channel->kobj; 2082 2083 channel->mmap_ring_buffer = hv_mmap_ring_buffer; 2084 channel->ring_sysfs_visible = true; 2085 2086 return sysfs_update_group(kobj, &vmbus_chan_group); 2087} 2088EXPORT_SYMBOL_GPL(hv_create_ring_sysfs); 2089 2090/** 2091 * hv_remove_ring_sysfs() - remove ring sysfs entry corresponding to ring buffers for a channel. 2092 * @channel: Pointer to vmbus_channel structure 2093 * 2094 * Hide "ring" sysfs for a channel by changing its is_visible attribute and updating sysfs group. 2095 * 2096 * Returns 0 on success or error code on failure. 2097 */ 2098int hv_remove_ring_sysfs(struct vmbus_channel *channel) 2099{ 2100 struct kobject *kobj = &channel->kobj; 2101 int ret; 2102 2103 channel->ring_sysfs_visible = false; 2104 ret = sysfs_update_group(kobj, &vmbus_chan_group); 2105 channel->mmap_ring_buffer = NULL; 2106 return ret; 2107} 2108EXPORT_SYMBOL_GPL(hv_remove_ring_sysfs); 2109 2110/* 2111 * vmbus_add_channel_kobj - setup a sub-directory under device/channels 2112 */ 2113int vmbus_add_channel_kobj(struct hv_device *dev, struct vmbus_channel *channel) 2114{ 2115 const struct device *device = &dev->device; 2116 struct kobject *kobj = &channel->kobj; 2117 u32 relid = channel->offermsg.child_relid; 2118 int ret; 2119 2120 kobj->kset = dev->channels_kset; 2121 ret = kobject_init_and_add(kobj, &vmbus_chan_ktype, NULL, 2122 "%u", relid); 2123 if (ret) { 2124 kobject_put(kobj); 2125 return ret; 2126 } 2127 2128 ret = sysfs_create_group(kobj, &vmbus_chan_group); 2129 2130 if (ret) { 2131 /* 2132 * The calling functions' error handling paths will cleanup the 2133 * empty channel directory. 2134 */ 2135 kobject_put(kobj); 2136 dev_err(device, "Unable to set up channel sysfs files\n"); 2137 return ret; 2138 } 2139 2140 kobject_uevent(kobj, KOBJ_ADD); 2141 2142 return 0; 2143} 2144 2145/* 2146 * vmbus_remove_channel_attr_group - remove the channel's attribute group 2147 */ 2148void vmbus_remove_channel_attr_group(struct vmbus_channel *channel) 2149{ 2150 sysfs_remove_group(&channel->kobj, &vmbus_chan_group); 2151} 2152 2153/* 2154 * vmbus_device_create - Creates and registers a new child device 2155 * on the vmbus. 2156 */ 2157struct hv_device *vmbus_device_create(const guid_t *type, 2158 const guid_t *instance, 2159 struct vmbus_channel *channel) 2160{ 2161 struct hv_device *child_device_obj; 2162 2163 child_device_obj = kzalloc(sizeof(struct hv_device), GFP_KERNEL); 2164 if (!child_device_obj) { 2165 pr_err("Unable to allocate device object for child device\n"); 2166 return NULL; 2167 } 2168 2169 child_device_obj->channel = channel; 2170 guid_copy(&child_device_obj->dev_type, type); 2171 guid_copy(&child_device_obj->dev_instance, instance); 2172 child_device_obj->vendor_id = PCI_VENDOR_ID_MICROSOFT; 2173 2174 return child_device_obj; 2175} 2176 2177/* 2178 * vmbus_device_register - Register the child device 2179 */ 2180int vmbus_device_register(struct hv_device *child_device_obj) 2181{ 2182 struct kobject *kobj = &child_device_obj->device.kobj; 2183 int ret; 2184 2185 dev_set_name(&child_device_obj->device, "%pUl", 2186 &child_device_obj->channel->offermsg.offer.if_instance); 2187 2188 child_device_obj->device.bus = &hv_bus; 2189 child_device_obj->device.parent = vmbus_root_device; 2190 child_device_obj->device.release = vmbus_device_release; 2191 2192 child_device_obj->device.dma_parms = &child_device_obj->dma_parms; 2193 child_device_obj->device.dma_mask = &child_device_obj->dma_mask; 2194 dma_set_mask(&child_device_obj->device, DMA_BIT_MASK(64)); 2195 2196 /* 2197 * Register with the LDM. This will kick off the driver/device 2198 * binding...which will eventually call vmbus_match() and vmbus_probe() 2199 */ 2200 ret = device_register(&child_device_obj->device); 2201 if (ret) { 2202 pr_err("Unable to register child device\n"); 2203 put_device(&child_device_obj->device); 2204 return ret; 2205 } 2206 2207 /* 2208 * If device_register() found a driver to assign to the device, the 2209 * driver's probe function has already run at this point. If that 2210 * probe function accesses or operates on the "channels" subdirectory 2211 * in sysfs, those operations will have failed because the "channels" 2212 * subdirectory doesn't exist until the code below runs. Or if the 2213 * probe function creates a /dev entry, a user space program could 2214 * find and open the /dev entry, and then create a race by accessing 2215 * the "channels" subdirectory while the creation steps are in progress 2216 * here. The race can't result in a kernel failure, but the user space 2217 * program may get an error in accessing "channels" or its 2218 * subdirectories. See also comments with vmbus_add_dynid() about a 2219 * related race condition. 2220 */ 2221 child_device_obj->channels_kset = kset_create_and_add("channels", 2222 NULL, kobj); 2223 if (!child_device_obj->channels_kset) { 2224 ret = -ENOMEM; 2225 goto err_dev_unregister; 2226 } 2227 2228 ret = vmbus_add_channel_kobj(child_device_obj, 2229 child_device_obj->channel); 2230 if (ret) { 2231 pr_err("Unable to register primary channel\n"); 2232 goto err_kset_unregister; 2233 } 2234 hv_debug_add_dev_dir(child_device_obj); 2235 2236 return 0; 2237 2238err_kset_unregister: 2239 kset_unregister(child_device_obj->channels_kset); 2240 2241err_dev_unregister: 2242 device_unregister(&child_device_obj->device); 2243 return ret; 2244} 2245 2246/* 2247 * vmbus_device_unregister - Remove the specified child device 2248 * from the vmbus. 2249 */ 2250void vmbus_device_unregister(struct hv_device *device_obj) 2251{ 2252 pr_debug("child device %s unregistered\n", 2253 dev_name(&device_obj->device)); 2254 2255 kset_unregister(device_obj->channels_kset); 2256 2257 /* 2258 * Kick off the process of unregistering the device. 2259 * This will call vmbus_remove() and eventually vmbus_device_release() 2260 */ 2261 device_unregister(&device_obj->device); 2262} 2263EXPORT_SYMBOL_GPL(vmbus_device_unregister); 2264 2265#ifdef CONFIG_ACPI 2266/* 2267 * VMBUS is an acpi enumerated device. Get the information we 2268 * need from DSDT. 2269 */ 2270static acpi_status vmbus_walk_resources(struct acpi_resource *res, void *ctx) 2271{ 2272 resource_size_t start = 0; 2273 resource_size_t end = 0; 2274 struct resource *new_res; 2275 struct resource **old_res = &hyperv_mmio; 2276 struct resource **prev_res = NULL; 2277 struct resource r; 2278 2279 switch (res->type) { 2280 2281 /* 2282 * "Address" descriptors are for bus windows. Ignore 2283 * "memory" descriptors, which are for registers on 2284 * devices. 2285 */ 2286 case ACPI_RESOURCE_TYPE_ADDRESS32: 2287 start = res->data.address32.address.minimum; 2288 end = res->data.address32.address.maximum; 2289 break; 2290 2291 case ACPI_RESOURCE_TYPE_ADDRESS64: 2292 start = res->data.address64.address.minimum; 2293 end = res->data.address64.address.maximum; 2294 break; 2295 2296 /* 2297 * The IRQ information is needed only on ARM64, which Hyper-V 2298 * sets up in the extended format. IRQ information is present 2299 * on x86/x64 in the non-extended format but it is not used by 2300 * Linux. So don't bother checking for the non-extended format. 2301 */ 2302 case ACPI_RESOURCE_TYPE_EXTENDED_IRQ: 2303 if (!acpi_dev_resource_interrupt(res, 0, &r)) { 2304 pr_err("Unable to parse Hyper-V ACPI interrupt\n"); 2305 return AE_ERROR; 2306 } 2307 /* ARM64 INTID for VMbus */ 2308 vmbus_interrupt = res->data.extended_irq.interrupts[0]; 2309 /* Linux IRQ number */ 2310 vmbus_irq = r.start; 2311 return AE_OK; 2312 2313 default: 2314 /* Unused resource type */ 2315 return AE_OK; 2316 2317 } 2318 /* 2319 * Ignore ranges that are below 1MB, as they're not 2320 * necessary or useful here. 2321 */ 2322 if (end < 0x100000) 2323 return AE_OK; 2324 2325 new_res = kzalloc(sizeof(*new_res), GFP_ATOMIC); 2326 if (!new_res) 2327 return AE_NO_MEMORY; 2328 2329 /* If this range overlaps the virtual TPM, truncate it. */ 2330 if (end > VTPM_BASE_ADDRESS && start < VTPM_BASE_ADDRESS) 2331 end = VTPM_BASE_ADDRESS; 2332 2333 new_res->name = "hyperv mmio"; 2334 new_res->flags = IORESOURCE_MEM; 2335 new_res->start = start; 2336 new_res->end = end; 2337 2338 /* 2339 * If two ranges are adjacent, merge them. 2340 */ 2341 do { 2342 if (!*old_res) { 2343 *old_res = new_res; 2344 break; 2345 } 2346 2347 if (((*old_res)->end + 1) == new_res->start) { 2348 (*old_res)->end = new_res->end; 2349 kfree(new_res); 2350 break; 2351 } 2352 2353 if ((*old_res)->start == new_res->end + 1) { 2354 (*old_res)->start = new_res->start; 2355 kfree(new_res); 2356 break; 2357 } 2358 2359 if ((*old_res)->start > new_res->end) { 2360 new_res->sibling = *old_res; 2361 if (prev_res) 2362 (*prev_res)->sibling = new_res; 2363 *old_res = new_res; 2364 break; 2365 } 2366 2367 prev_res = old_res; 2368 old_res = &(*old_res)->sibling; 2369 2370 } while (1); 2371 2372 return AE_OK; 2373} 2374#endif 2375 2376static void vmbus_mmio_remove(void) 2377{ 2378 struct resource *cur_res; 2379 struct resource *next_res; 2380 2381 if (hyperv_mmio) { 2382 if (fb_mmio) { 2383 __release_region(hyperv_mmio, fb_mmio->start, 2384 resource_size(fb_mmio)); 2385 fb_mmio = NULL; 2386 } 2387 2388 for (cur_res = hyperv_mmio; cur_res; cur_res = next_res) { 2389 next_res = cur_res->sibling; 2390 kfree(cur_res); 2391 } 2392 } 2393} 2394 2395static void __maybe_unused vmbus_reserve_fb(void) 2396{ 2397 resource_size_t start = 0, size; 2398 struct pci_dev *pdev; 2399 2400 if (efi_enabled(EFI_BOOT)) { 2401 /* Gen2 VM: get FB base from EFI framebuffer */ 2402 if (IS_ENABLED(CONFIG_SYSFB)) { 2403 start = screen_info.lfb_base; 2404 size = max_t(__u32, screen_info.lfb_size, 0x800000); 2405 } 2406 } else { 2407 /* Gen1 VM: get FB base from PCI */ 2408 pdev = pci_get_device(PCI_VENDOR_ID_MICROSOFT, 2409 PCI_DEVICE_ID_HYPERV_VIDEO, NULL); 2410 if (!pdev) 2411 return; 2412 2413 if (pdev->resource[0].flags & IORESOURCE_MEM) { 2414 start = pci_resource_start(pdev, 0); 2415 size = pci_resource_len(pdev, 0); 2416 } 2417 2418 /* 2419 * Release the PCI device so hyperv_drm or hyperv_fb driver can 2420 * grab it later. 2421 */ 2422 pci_dev_put(pdev); 2423 } 2424 2425 if (!start) 2426 return; 2427 2428 /* 2429 * Make a claim for the frame buffer in the resource tree under the 2430 * first node, which will be the one below 4GB. The length seems to 2431 * be underreported, particularly in a Generation 1 VM. So start out 2432 * reserving a larger area and make it smaller until it succeeds. 2433 */ 2434 for (; !fb_mmio && (size >= 0x100000); size >>= 1) 2435 fb_mmio = __request_region(hyperv_mmio, start, size, fb_mmio_name, 0); 2436} 2437 2438/** 2439 * vmbus_allocate_mmio() - Pick a memory-mapped I/O range. 2440 * @new: If successful, supplied a pointer to the 2441 * allocated MMIO space. 2442 * @device_obj: Identifies the caller 2443 * @min: Minimum guest physical address of the 2444 * allocation 2445 * @max: Maximum guest physical address 2446 * @size: Size of the range to be allocated 2447 * @align: Alignment of the range to be allocated 2448 * @fb_overlap_ok: Whether this allocation can be allowed 2449 * to overlap the video frame buffer. 2450 * 2451 * This function walks the resources granted to VMBus by the 2452 * _CRS object in the ACPI namespace underneath the parent 2453 * "bridge" whether that's a root PCI bus in the Generation 1 2454 * case or a Module Device in the Generation 2 case. It then 2455 * attempts to allocate from the global MMIO pool in a way that 2456 * matches the constraints supplied in these parameters and by 2457 * that _CRS. 2458 * 2459 * Return: 0 on success, -errno on failure 2460 */ 2461int vmbus_allocate_mmio(struct resource **new, struct hv_device *device_obj, 2462 resource_size_t min, resource_size_t max, 2463 resource_size_t size, resource_size_t align, 2464 bool fb_overlap_ok) 2465{ 2466 struct resource *iter, *shadow; 2467 resource_size_t range_min, range_max, start, end; 2468 const char *dev_n = dev_name(&device_obj->device); 2469 int retval; 2470 2471 retval = -ENXIO; 2472 mutex_lock(&hyperv_mmio_lock); 2473 2474 /* 2475 * If overlaps with frame buffers are allowed, then first attempt to 2476 * make the allocation from within the reserved region. Because it 2477 * is already reserved, no shadow allocation is necessary. 2478 */ 2479 if (fb_overlap_ok && fb_mmio && !(min > fb_mmio->end) && 2480 !(max < fb_mmio->start)) { 2481 2482 range_min = fb_mmio->start; 2483 range_max = fb_mmio->end; 2484 start = (range_min + align - 1) & ~(align - 1); 2485 for (; start + size - 1 <= range_max; start += align) { 2486 *new = request_mem_region_exclusive(start, size, dev_n); 2487 if (*new) { 2488 retval = 0; 2489 goto exit; 2490 } 2491 } 2492 } 2493 2494 for (iter = hyperv_mmio; iter; iter = iter->sibling) { 2495 if ((iter->start >= max) || (iter->end <= min)) 2496 continue; 2497 2498 range_min = iter->start; 2499 range_max = iter->end; 2500 start = (range_min + align - 1) & ~(align - 1); 2501 for (; start + size - 1 <= range_max; start += align) { 2502 end = start + size - 1; 2503 2504 /* Skip the whole fb_mmio region if not fb_overlap_ok */ 2505 if (!fb_overlap_ok && fb_mmio && 2506 (((start >= fb_mmio->start) && (start <= fb_mmio->end)) || 2507 ((end >= fb_mmio->start) && (end <= fb_mmio->end)))) 2508 continue; 2509 2510 shadow = __request_region(iter, start, size, NULL, 2511 IORESOURCE_BUSY); 2512 if (!shadow) 2513 continue; 2514 2515 *new = request_mem_region_exclusive(start, size, dev_n); 2516 if (*new) { 2517 shadow->name = (char *)*new; 2518 retval = 0; 2519 goto exit; 2520 } 2521 2522 __release_region(iter, start, size); 2523 } 2524 } 2525 2526exit: 2527 mutex_unlock(&hyperv_mmio_lock); 2528 return retval; 2529} 2530EXPORT_SYMBOL_GPL(vmbus_allocate_mmio); 2531 2532/** 2533 * vmbus_free_mmio() - Free a memory-mapped I/O range. 2534 * @start: Base address of region to release. 2535 * @size: Size of the range to be allocated 2536 * 2537 * This function releases anything requested by 2538 * vmbus_mmio_allocate(). 2539 */ 2540void vmbus_free_mmio(resource_size_t start, resource_size_t size) 2541{ 2542 struct resource *iter; 2543 2544 mutex_lock(&hyperv_mmio_lock); 2545 2546 /* 2547 * If all bytes of the MMIO range to be released are within the 2548 * special case fb_mmio shadow region, skip releasing the shadow 2549 * region since no corresponding __request_region() was done 2550 * in vmbus_allocate_mmio(). 2551 */ 2552 if (fb_mmio && start >= fb_mmio->start && 2553 (start + size - 1 <= fb_mmio->end)) 2554 goto skip_shadow_release; 2555 2556 for (iter = hyperv_mmio; iter; iter = iter->sibling) { 2557 if ((iter->start >= start + size) || (iter->end <= start)) 2558 continue; 2559 2560 __release_region(iter, start, size); 2561 } 2562 2563skip_shadow_release: 2564 release_mem_region(start, size); 2565 mutex_unlock(&hyperv_mmio_lock); 2566 2567} 2568EXPORT_SYMBOL_GPL(vmbus_free_mmio); 2569 2570#ifdef CONFIG_ACPI 2571static int vmbus_acpi_add(struct platform_device *pdev) 2572{ 2573 acpi_status result; 2574 int ret_val = -ENODEV; 2575 struct acpi_device *ancestor; 2576 struct acpi_device *device = ACPI_COMPANION(&pdev->dev); 2577 2578 vmbus_root_device = &device->dev; 2579 2580 /* 2581 * Older versions of Hyper-V for ARM64 fail to include the _CCA 2582 * method on the top level VMbus device in the DSDT. But devices 2583 * are hardware coherent in all current Hyper-V use cases, so fix 2584 * up the ACPI device to behave as if _CCA is present and indicates 2585 * hardware coherence. 2586 */ 2587 ACPI_COMPANION_SET(&device->dev, device); 2588 if (IS_ENABLED(CONFIG_ACPI_CCA_REQUIRED) && 2589 device_get_dma_attr(&device->dev) == DEV_DMA_NOT_SUPPORTED) { 2590 pr_info("No ACPI _CCA found; assuming coherent device I/O\n"); 2591 device->flags.cca_seen = true; 2592 device->flags.coherent_dma = true; 2593 } 2594 2595 result = acpi_walk_resources(device->handle, METHOD_NAME__CRS, 2596 vmbus_walk_resources, NULL); 2597 2598 if (ACPI_FAILURE(result)) 2599 goto acpi_walk_err; 2600 /* 2601 * Some ancestor of the vmbus acpi device (Gen1 or Gen2 2602 * firmware) is the VMOD that has the mmio ranges. Get that. 2603 */ 2604 for (ancestor = acpi_dev_parent(device); 2605 ancestor && ancestor->handle != ACPI_ROOT_OBJECT; 2606 ancestor = acpi_dev_parent(ancestor)) { 2607 result = acpi_walk_resources(ancestor->handle, METHOD_NAME__CRS, 2608 vmbus_walk_resources, NULL); 2609 2610 if (ACPI_FAILURE(result)) 2611 continue; 2612 if (hyperv_mmio) { 2613 vmbus_reserve_fb(); 2614 break; 2615 } 2616 } 2617 ret_val = 0; 2618 2619acpi_walk_err: 2620 if (ret_val) 2621 vmbus_mmio_remove(); 2622 return ret_val; 2623} 2624#else 2625static int vmbus_acpi_add(struct platform_device *pdev) 2626{ 2627 return 0; 2628} 2629#endif 2630#ifndef HYPERVISOR_CALLBACK_VECTOR 2631static int vmbus_set_irq(struct platform_device *pdev) 2632{ 2633 struct irq_data *data; 2634 int irq; 2635 irq_hw_number_t hwirq; 2636 2637 irq = platform_get_irq(pdev, 0); 2638 /* platform_get_irq() may not return 0. */ 2639 if (irq < 0) 2640 return irq; 2641 2642 data = irq_get_irq_data(irq); 2643 if (!data) { 2644 pr_err("No interrupt data for VMBus virq %d\n", irq); 2645 return -ENODEV; 2646 } 2647 hwirq = irqd_to_hwirq(data); 2648 2649 vmbus_irq = irq; 2650 vmbus_interrupt = hwirq; 2651 pr_debug("VMBus virq %d, hwirq %d\n", vmbus_irq, vmbus_interrupt); 2652 2653 return 0; 2654} 2655#endif 2656 2657static int vmbus_device_add(struct platform_device *pdev) 2658{ 2659 struct resource **cur_res = &hyperv_mmio; 2660 struct of_range range; 2661 struct of_range_parser parser; 2662 struct device_node *np = pdev->dev.of_node; 2663 int ret; 2664 2665 vmbus_root_device = &pdev->dev; 2666 2667 ret = of_range_parser_init(&parser, np); 2668 if (ret) 2669 return ret; 2670 2671#ifndef HYPERVISOR_CALLBACK_VECTOR 2672 ret = vmbus_set_irq(pdev); 2673 if (ret) 2674 return ret; 2675#endif 2676 for_each_of_range(&parser, &range) { 2677 struct resource *res; 2678 2679 res = kzalloc(sizeof(*res), GFP_KERNEL); 2680 if (!res) { 2681 vmbus_mmio_remove(); 2682 return -ENOMEM; 2683 } 2684 2685 res->name = "hyperv mmio"; 2686 res->flags = range.flags; 2687 res->start = range.cpu_addr; 2688 res->end = range.cpu_addr + range.size; 2689 2690 *cur_res = res; 2691 cur_res = &res->sibling; 2692 } 2693 2694 return ret; 2695} 2696 2697static int vmbus_platform_driver_probe(struct platform_device *pdev) 2698{ 2699 if (acpi_disabled) 2700 return vmbus_device_add(pdev); 2701 else 2702 return vmbus_acpi_add(pdev); 2703} 2704 2705static void vmbus_platform_driver_remove(struct platform_device *pdev) 2706{ 2707 vmbus_mmio_remove(); 2708} 2709 2710#ifdef CONFIG_PM_SLEEP 2711static int vmbus_bus_suspend(struct device *dev) 2712{ 2713 struct hv_per_cpu_context *hv_cpu = per_cpu_ptr( 2714 hv_context.cpu_context, VMBUS_CONNECT_CPU); 2715 struct vmbus_channel *channel, *sc; 2716 2717 tasklet_disable(&hv_cpu->msg_dpc); 2718 vmbus_connection.ignore_any_offer_msg = true; 2719 /* The tasklet_enable() takes care of providing a memory barrier */ 2720 tasklet_enable(&hv_cpu->msg_dpc); 2721 2722 /* Drain all the workqueues as we are in suspend */ 2723 drain_workqueue(vmbus_connection.rescind_work_queue); 2724 drain_workqueue(vmbus_connection.work_queue); 2725 drain_workqueue(vmbus_connection.handle_primary_chan_wq); 2726 drain_workqueue(vmbus_connection.handle_sub_chan_wq); 2727 2728 mutex_lock(&vmbus_connection.channel_mutex); 2729 list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) { 2730 if (!is_hvsock_channel(channel)) 2731 continue; 2732 2733 vmbus_force_channel_rescinded(channel); 2734 } 2735 mutex_unlock(&vmbus_connection.channel_mutex); 2736 2737 /* 2738 * Wait until all the sub-channels and hv_sock channels have been 2739 * cleaned up. Sub-channels should be destroyed upon suspend, otherwise 2740 * they would conflict with the new sub-channels that will be created 2741 * in the resume path. hv_sock channels should also be destroyed, but 2742 * a hv_sock channel of an established hv_sock connection can not be 2743 * really destroyed since it may still be referenced by the userspace 2744 * application, so we just force the hv_sock channel to be rescinded 2745 * by vmbus_force_channel_rescinded(), and the userspace application 2746 * will thoroughly destroy the channel after hibernation. 2747 * 2748 * Note: the counter nr_chan_close_on_suspend may never go above 0 if 2749 * the VM has no sub-channel and hv_sock channel, e.g. a 1-vCPU VM. 2750 */ 2751 if (atomic_read(&vmbus_connection.nr_chan_close_on_suspend) > 0) 2752 wait_for_completion(&vmbus_connection.ready_for_suspend_event); 2753 2754 mutex_lock(&vmbus_connection.channel_mutex); 2755 2756 list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) { 2757 /* 2758 * Remove the channel from the array of channels and invalidate 2759 * the channel's relid. Upon resume, vmbus_onoffer() will fix 2760 * up the relid (and other fields, if necessary) and add the 2761 * channel back to the array. 2762 */ 2763 vmbus_channel_unmap_relid(channel); 2764 channel->offermsg.child_relid = INVALID_RELID; 2765 2766 if (is_hvsock_channel(channel)) { 2767 if (!channel->rescind) { 2768 pr_err("hv_sock channel not rescinded!\n"); 2769 WARN_ON_ONCE(1); 2770 } 2771 continue; 2772 } 2773 2774 list_for_each_entry(sc, &channel->sc_list, sc_list) { 2775 pr_err("Sub-channel not deleted!\n"); 2776 WARN_ON_ONCE(1); 2777 } 2778 } 2779 2780 mutex_unlock(&vmbus_connection.channel_mutex); 2781 2782 vmbus_initiate_unload(false); 2783 2784 return 0; 2785} 2786 2787static int vmbus_bus_resume(struct device *dev) 2788{ 2789 struct vmbus_channel *channel; 2790 struct vmbus_channel_msginfo *msginfo; 2791 size_t msgsize; 2792 int ret; 2793 2794 vmbus_connection.ignore_any_offer_msg = false; 2795 2796 /* 2797 * We only use the 'vmbus_proto_version', which was in use before 2798 * hibernation, to re-negotiate with the host. 2799 */ 2800 if (!vmbus_proto_version) { 2801 pr_err("Invalid proto version = 0x%x\n", vmbus_proto_version); 2802 return -EINVAL; 2803 } 2804 2805 msgsize = sizeof(*msginfo) + 2806 sizeof(struct vmbus_channel_initiate_contact); 2807 2808 msginfo = kzalloc(msgsize, GFP_KERNEL); 2809 2810 if (msginfo == NULL) 2811 return -ENOMEM; 2812 2813 ret = vmbus_negotiate_version(msginfo, vmbus_proto_version); 2814 2815 kfree(msginfo); 2816 2817 if (ret != 0) 2818 return ret; 2819 2820 vmbus_request_offers(); 2821 2822 mutex_lock(&vmbus_connection.channel_mutex); 2823 list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) { 2824 if (channel->offermsg.child_relid != INVALID_RELID) 2825 continue; 2826 2827 /* hvsock channels are not expected to be present. */ 2828 if (is_hvsock_channel(channel)) 2829 continue; 2830 2831 pr_err("channel %pUl/%pUl not present after resume.\n", 2832 &channel->offermsg.offer.if_type, 2833 &channel->offermsg.offer.if_instance); 2834 /* ToDo: Cleanup these channels here */ 2835 } 2836 mutex_unlock(&vmbus_connection.channel_mutex); 2837 2838 /* Reset the event for the next suspend. */ 2839 reinit_completion(&vmbus_connection.ready_for_suspend_event); 2840 2841 return 0; 2842} 2843#else 2844#define vmbus_bus_suspend NULL 2845#define vmbus_bus_resume NULL 2846#endif /* CONFIG_PM_SLEEP */ 2847 2848static const __maybe_unused struct of_device_id vmbus_of_match[] = { 2849 { 2850 .compatible = "microsoft,vmbus", 2851 }, 2852 { 2853 /* sentinel */ 2854 }, 2855}; 2856MODULE_DEVICE_TABLE(of, vmbus_of_match); 2857 2858static const __maybe_unused struct acpi_device_id vmbus_acpi_device_ids[] = { 2859 {"VMBUS", 0}, 2860 {"VMBus", 0}, 2861 {"", 0}, 2862}; 2863MODULE_DEVICE_TABLE(acpi, vmbus_acpi_device_ids); 2864 2865/* 2866 * Note: we must use the "no_irq" ops, otherwise hibernation can not work with 2867 * PCI device assignment, because "pci_dev_pm_ops" uses the "noirq" ops: in 2868 * the resume path, the pci "noirq" restore op runs before "non-noirq" op (see 2869 * resume_target_kernel() -> dpm_resume_start(), and hibernation_restore() -> 2870 * dpm_resume_end()). This means vmbus_bus_resume() and the pci-hyperv's 2871 * resume callback must also run via the "noirq" ops. 2872 * 2873 * Set suspend_noirq/resume_noirq to NULL for Suspend-to-Idle: see the comment 2874 * earlier in this file before vmbus_pm. 2875 */ 2876 2877static const struct dev_pm_ops vmbus_bus_pm = { 2878 .suspend_noirq = NULL, 2879 .resume_noirq = NULL, 2880 .freeze_noirq = vmbus_bus_suspend, 2881 .thaw_noirq = vmbus_bus_resume, 2882 .poweroff_noirq = vmbus_bus_suspend, 2883 .restore_noirq = vmbus_bus_resume 2884}; 2885 2886static struct platform_driver vmbus_platform_driver = { 2887 .probe = vmbus_platform_driver_probe, 2888 .remove = vmbus_platform_driver_remove, 2889 .driver = { 2890 .name = "vmbus", 2891 .acpi_match_table = ACPI_PTR(vmbus_acpi_device_ids), 2892 .of_match_table = of_match_ptr(vmbus_of_match), 2893 .pm = &vmbus_bus_pm, 2894 .probe_type = PROBE_FORCE_SYNCHRONOUS, 2895 } 2896}; 2897 2898static void hv_kexec_handler(void) 2899{ 2900 hv_stimer_global_cleanup(); 2901 vmbus_initiate_unload(false); 2902 /* Make sure conn_state is set as hv_synic_cleanup checks for it */ 2903 mb(); 2904 cpuhp_remove_state(hyperv_cpuhp_online); 2905}; 2906 2907static void hv_crash_handler(struct pt_regs *regs) 2908{ 2909 int cpu; 2910 2911 vmbus_initiate_unload(true); 2912 /* 2913 * In crash handler we can't schedule synic cleanup for all CPUs, 2914 * doing the cleanup for current CPU only. This should be sufficient 2915 * for kdump. 2916 */ 2917 cpu = smp_processor_id(); 2918 hv_stimer_cleanup(cpu); 2919 hv_hyp_synic_disable_regs(cpu); 2920}; 2921 2922static int hv_synic_suspend(void *data) 2923{ 2924 /* 2925 * When we reach here, all the non-boot CPUs have been offlined. 2926 * If we're in a legacy configuration where stimer Direct Mode is 2927 * not enabled, the stimers on the non-boot CPUs have been unbound 2928 * in hv_synic_cleanup() -> hv_stimer_legacy_cleanup() -> 2929 * hv_stimer_cleanup() -> clockevents_unbind_device(). 2930 * 2931 * hv_synic_suspend() only runs on CPU0 with interrupts disabled. 2932 * Here we do not call hv_stimer_legacy_cleanup() on CPU0 because: 2933 * 1) it's unnecessary as interrupts remain disabled between 2934 * syscore_suspend() and syscore_resume(): see create_image() and 2935 * resume_target_kernel() 2936 * 2) the stimer on CPU0 is automatically disabled later by 2937 * syscore_suspend() -> timekeeping_suspend() -> tick_suspend() -> ... 2938 * -> clockevents_shutdown() -> ... -> hv_ce_shutdown() 2939 * 3) a warning would be triggered if we call 2940 * clockevents_unbind_device(), which may sleep, in an 2941 * interrupts-disabled context. 2942 */ 2943 2944 hv_hyp_synic_disable_regs(0); 2945 2946 return 0; 2947} 2948 2949static void hv_synic_resume(void *data) 2950{ 2951 hv_hyp_synic_enable_regs(0); 2952 2953 /* 2954 * Note: we don't need to call hv_stimer_init(0), because the timer 2955 * on CPU0 is not unbound in hv_synic_suspend(), and the timer is 2956 * automatically re-enabled in timekeeping_resume(). 2957 */ 2958} 2959 2960/* The callbacks run only on CPU0, with irqs_disabled. */ 2961static const struct syscore_ops hv_synic_syscore_ops = { 2962 .suspend = hv_synic_suspend, 2963 .resume = hv_synic_resume, 2964}; 2965 2966static struct syscore hv_synic_syscore = { 2967 .ops = &hv_synic_syscore_ops, 2968}; 2969 2970static int __init hv_acpi_init(void) 2971{ 2972 int ret; 2973 2974 if (!hv_is_hyperv_initialized()) 2975 return -ENODEV; 2976 2977 if (hv_root_partition() && !hv_nested) 2978 return 0; 2979 2980 /* 2981 * Get ACPI resources first. 2982 */ 2983 ret = platform_driver_register(&vmbus_platform_driver); 2984 if (ret) 2985 return ret; 2986 2987 if (!vmbus_root_device) { 2988 ret = -ENODEV; 2989 goto cleanup; 2990 } 2991 2992 /* 2993 * If we're on an architecture with a hardcoded hypervisor 2994 * vector (i.e. x86/x64), override the VMbus interrupt found 2995 * in the ACPI tables. Ensure vmbus_irq is not set since the 2996 * normal Linux IRQ mechanism is not used in this case. 2997 */ 2998#ifdef HYPERVISOR_CALLBACK_VECTOR 2999 vmbus_interrupt = HYPERVISOR_CALLBACK_VECTOR; 3000 vmbus_irq = -1; 3001#endif 3002 3003 hv_debug_init(); 3004 3005 ret = vmbus_bus_init(); 3006 if (ret) 3007 goto cleanup; 3008 3009 hv_setup_kexec_handler(hv_kexec_handler); 3010 hv_setup_crash_handler(hv_crash_handler); 3011 3012 register_syscore(&hv_synic_syscore); 3013 3014 return 0; 3015 3016cleanup: 3017 platform_driver_unregister(&vmbus_platform_driver); 3018 vmbus_root_device = NULL; 3019 return ret; 3020} 3021 3022static void __exit vmbus_exit(void) 3023{ 3024 int cpu; 3025 3026 unregister_syscore(&hv_synic_syscore); 3027 3028 hv_remove_kexec_handler(); 3029 hv_remove_crash_handler(); 3030 vmbus_connection.conn_state = DISCONNECTED; 3031 hv_stimer_global_cleanup(); 3032 vmbus_disconnect(); 3033 if (vmbus_irq == -1) { 3034 hv_remove_vmbus_handler(); 3035 } else { 3036 free_percpu_irq(vmbus_irq, vmbus_evt); 3037 free_percpu(vmbus_evt); 3038 } 3039 if (IS_ENABLED(CONFIG_PREEMPT_RT) && vmbus_irq_initialized) { 3040 smpboot_unregister_percpu_thread(&vmbus_irq_threads); 3041 vmbus_irq_initialized = false; 3042 } 3043 for_each_online_cpu(cpu) { 3044 struct hv_per_cpu_context *hv_cpu 3045 = per_cpu_ptr(hv_context.cpu_context, cpu); 3046 3047 tasklet_kill(&hv_cpu->msg_dpc); 3048 } 3049 hv_debug_rm_all_dir(); 3050 3051 vmbus_free_channels(); 3052 kfree(vmbus_connection.channels); 3053 3054 /* 3055 * The vmbus panic notifier is always registered, hence we should 3056 * also unconditionally unregister it here as well. 3057 */ 3058 atomic_notifier_chain_unregister(&panic_notifier_list, 3059 &hyperv_panic_vmbus_unload_block); 3060 3061 bus_unregister(&hv_bus); 3062 3063 cpuhp_remove_state(hyperv_cpuhp_online); 3064 hv_synic_free(); 3065 platform_driver_unregister(&vmbus_platform_driver); 3066} 3067 3068 3069MODULE_LICENSE("GPL"); 3070MODULE_DESCRIPTION("Microsoft Hyper-V VMBus Driver"); 3071 3072subsys_initcall(hv_acpi_init); 3073module_exit(vmbus_exit);
~/
Read-only file preview.