Linux Audio

Check our new training course

Loading...
Note: File does not exist in v3.15.
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * Copyright (c) Microsoft Corporation.
   4 *
   5 * Author:
   6 *   Jake Oshins <jakeo@microsoft.com>
   7 *
   8 * This driver acts as a paravirtual front-end for PCI Express root buses.
   9 * When a PCI Express function (either an entire device or an SR-IOV
  10 * Virtual Function) is being passed through to the VM, this driver exposes
  11 * a new bus to the guest VM.  This is modeled as a root PCI bus because
  12 * no bridges are being exposed to the VM.  In fact, with a "Generation 2"
  13 * VM within Hyper-V, there may seem to be no PCI bus at all in the VM
  14 * until a device as been exposed using this driver.
  15 *
  16 * Each root PCI bus has its own PCI domain, which is called "Segment" in
  17 * the PCI Firmware Specifications.  Thus while each device passed through
  18 * to the VM using this front-end will appear at "device 0", the domain will
  19 * be unique.  Typically, each bus will have one PCI function on it, though
  20 * this driver does support more than one.
  21 *
  22 * In order to map the interrupts from the device through to the guest VM,
  23 * this driver also implements an IRQ Domain, which handles interrupts (either
  24 * MSI or MSI-X) associated with the functions on the bus.  As interrupts are
  25 * set up, torn down, or reaffined, this driver communicates with the
  26 * underlying hypervisor to adjust the mappings in the I/O MMU so that each
  27 * interrupt will be delivered to the correct virtual processor at the right
  28 * vector.  This driver does not support level-triggered (line-based)
  29 * interrupts, and will report that the Interrupt Line register in the
  30 * function's configuration space is zero.
  31 *
  32 * The rest of this driver mostly maps PCI concepts onto underlying Hyper-V
  33 * facilities.  For instance, the configuration space of a function exposed
  34 * by Hyper-V is mapped into a single page of memory space, and the
  35 * read and write handlers for config space must be aware of this mechanism.
  36 * Similarly, device setup and teardown involves messages sent to and from
  37 * the PCI back-end driver in Hyper-V.
  38 */
  39
  40#include <linux/kernel.h>
  41#include <linux/module.h>
  42#include <linux/pci.h>
  43#include <linux/pci-ecam.h>
  44#include <linux/delay.h>
  45#include <linux/semaphore.h>
  46#include <linux/irq.h>
  47#include <linux/msi.h>
  48#include <linux/hyperv.h>
  49#include <linux/refcount.h>
  50#include <linux/irqdomain.h>
  51#include <linux/acpi.h>
  52#include <asm/mshyperv.h>
  53
  54/*
  55 * Protocol versions. The low word is the minor version, the high word the
  56 * major version.
  57 */
  58
  59#define PCI_MAKE_VERSION(major, minor) ((u32)(((major) << 16) | (minor)))
  60#define PCI_MAJOR_VERSION(version) ((u32)(version) >> 16)
  61#define PCI_MINOR_VERSION(version) ((u32)(version) & 0xff)
  62
  63enum pci_protocol_version_t {
  64	PCI_PROTOCOL_VERSION_1_1 = PCI_MAKE_VERSION(1, 1),	/* Win10 */
  65	PCI_PROTOCOL_VERSION_1_2 = PCI_MAKE_VERSION(1, 2),	/* RS1 */
  66	PCI_PROTOCOL_VERSION_1_3 = PCI_MAKE_VERSION(1, 3),	/* Vibranium */
  67	PCI_PROTOCOL_VERSION_1_4 = PCI_MAKE_VERSION(1, 4),	/* WS2022 */
  68};
  69
  70#define CPU_AFFINITY_ALL	-1ULL
  71
  72/*
  73 * Supported protocol versions in the order of probing - highest go
  74 * first.
  75 */
  76static enum pci_protocol_version_t pci_protocol_versions[] = {
  77	PCI_PROTOCOL_VERSION_1_4,
  78	PCI_PROTOCOL_VERSION_1_3,
  79	PCI_PROTOCOL_VERSION_1_2,
  80	PCI_PROTOCOL_VERSION_1_1,
  81};
  82
  83#define PCI_CONFIG_MMIO_LENGTH	0x2000
  84#define CFG_PAGE_OFFSET 0x1000
  85#define CFG_PAGE_SIZE (PCI_CONFIG_MMIO_LENGTH - CFG_PAGE_OFFSET)
  86
  87#define MAX_SUPPORTED_MSI_MESSAGES 0x400
  88
  89#define STATUS_REVISION_MISMATCH 0xC0000059
  90
  91/* space for 32bit serial number as string */
  92#define SLOT_NAME_SIZE 11
  93
  94/*
  95 * Size of requestor for VMbus; the value is based on the observation
  96 * that having more than one request outstanding is 'rare', and so 64
  97 * should be generous in ensuring that we don't ever run out.
  98 */
  99#define HV_PCI_RQSTOR_SIZE 64
 100
 101/*
 102 * Message Types
 103 */
 104
 105enum pci_message_type {
 106	/*
 107	 * Version 1.1
 108	 */
 109	PCI_MESSAGE_BASE                = 0x42490000,
 110	PCI_BUS_RELATIONS               = PCI_MESSAGE_BASE + 0,
 111	PCI_QUERY_BUS_RELATIONS         = PCI_MESSAGE_BASE + 1,
 112	PCI_POWER_STATE_CHANGE          = PCI_MESSAGE_BASE + 4,
 113	PCI_QUERY_RESOURCE_REQUIREMENTS = PCI_MESSAGE_BASE + 5,
 114	PCI_QUERY_RESOURCE_RESOURCES    = PCI_MESSAGE_BASE + 6,
 115	PCI_BUS_D0ENTRY                 = PCI_MESSAGE_BASE + 7,
 116	PCI_BUS_D0EXIT                  = PCI_MESSAGE_BASE + 8,
 117	PCI_READ_BLOCK                  = PCI_MESSAGE_BASE + 9,
 118	PCI_WRITE_BLOCK                 = PCI_MESSAGE_BASE + 0xA,
 119	PCI_EJECT                       = PCI_MESSAGE_BASE + 0xB,
 120	PCI_QUERY_STOP                  = PCI_MESSAGE_BASE + 0xC,
 121	PCI_REENABLE                    = PCI_MESSAGE_BASE + 0xD,
 122	PCI_QUERY_STOP_FAILED           = PCI_MESSAGE_BASE + 0xE,
 123	PCI_EJECTION_COMPLETE           = PCI_MESSAGE_BASE + 0xF,
 124	PCI_RESOURCES_ASSIGNED          = PCI_MESSAGE_BASE + 0x10,
 125	PCI_RESOURCES_RELEASED          = PCI_MESSAGE_BASE + 0x11,
 126	PCI_INVALIDATE_BLOCK            = PCI_MESSAGE_BASE + 0x12,
 127	PCI_QUERY_PROTOCOL_VERSION      = PCI_MESSAGE_BASE + 0x13,
 128	PCI_CREATE_INTERRUPT_MESSAGE    = PCI_MESSAGE_BASE + 0x14,
 129	PCI_DELETE_INTERRUPT_MESSAGE    = PCI_MESSAGE_BASE + 0x15,
 130	PCI_RESOURCES_ASSIGNED2		= PCI_MESSAGE_BASE + 0x16,
 131	PCI_CREATE_INTERRUPT_MESSAGE2	= PCI_MESSAGE_BASE + 0x17,
 132	PCI_DELETE_INTERRUPT_MESSAGE2	= PCI_MESSAGE_BASE + 0x18, /* unused */
 133	PCI_BUS_RELATIONS2		= PCI_MESSAGE_BASE + 0x19,
 134	PCI_RESOURCES_ASSIGNED3         = PCI_MESSAGE_BASE + 0x1A,
 135	PCI_CREATE_INTERRUPT_MESSAGE3   = PCI_MESSAGE_BASE + 0x1B,
 136	PCI_MESSAGE_MAXIMUM
 137};
 138
 139/*
 140 * Structures defining the virtual PCI Express protocol.
 141 */
 142
 143union pci_version {
 144	struct {
 145		u16 minor_version;
 146		u16 major_version;
 147	} parts;
 148	u32 version;
 149} __packed;
 150
 151/*
 152 * Function numbers are 8-bits wide on Express, as interpreted through ARI,
 153 * which is all this driver does.  This representation is the one used in
 154 * Windows, which is what is expected when sending this back and forth with
 155 * the Hyper-V parent partition.
 156 */
 157union win_slot_encoding {
 158	struct {
 159		u32	dev:5;
 160		u32	func:3;
 161		u32	reserved:24;
 162	} bits;
 163	u32 slot;
 164} __packed;
 165
 166/*
 167 * Pretty much as defined in the PCI Specifications.
 168 */
 169struct pci_function_description {
 170	u16	v_id;	/* vendor ID */
 171	u16	d_id;	/* device ID */
 172	u8	rev;
 173	u8	prog_intf;
 174	u8	subclass;
 175	u8	base_class;
 176	u32	subsystem_id;
 177	union win_slot_encoding win_slot;
 178	u32	ser;	/* serial number */
 179} __packed;
 180
 181enum pci_device_description_flags {
 182	HV_PCI_DEVICE_FLAG_NONE			= 0x0,
 183	HV_PCI_DEVICE_FLAG_NUMA_AFFINITY	= 0x1,
 184};
 185
 186struct pci_function_description2 {
 187	u16	v_id;	/* vendor ID */
 188	u16	d_id;	/* device ID */
 189	u8	rev;
 190	u8	prog_intf;
 191	u8	subclass;
 192	u8	base_class;
 193	u32	subsystem_id;
 194	union	win_slot_encoding win_slot;
 195	u32	ser;	/* serial number */
 196	u32	flags;
 197	u16	virtual_numa_node;
 198	u16	reserved;
 199} __packed;
 200
 201/**
 202 * struct hv_msi_desc
 203 * @vector:		IDT entry
 204 * @delivery_mode:	As defined in Intel's Programmer's
 205 *			Reference Manual, Volume 3, Chapter 8.
 206 * @vector_count:	Number of contiguous entries in the
 207 *			Interrupt Descriptor Table that are
 208 *			occupied by this Message-Signaled
 209 *			Interrupt. For "MSI", as first defined
 210 *			in PCI 2.2, this can be between 1 and
 211 *			32. For "MSI-X," as first defined in PCI
 212 *			3.0, this must be 1, as each MSI-X table
 213 *			entry would have its own descriptor.
 214 * @reserved:		Empty space
 215 * @cpu_mask:		All the target virtual processors.
 216 */
 217struct hv_msi_desc {
 218	u8	vector;
 219	u8	delivery_mode;
 220	u16	vector_count;
 221	u32	reserved;
 222	u64	cpu_mask;
 223} __packed;
 224
 225/**
 226 * struct hv_msi_desc2 - 1.2 version of hv_msi_desc
 227 * @vector:		IDT entry
 228 * @delivery_mode:	As defined in Intel's Programmer's
 229 *			Reference Manual, Volume 3, Chapter 8.
 230 * @vector_count:	Number of contiguous entries in the
 231 *			Interrupt Descriptor Table that are
 232 *			occupied by this Message-Signaled
 233 *			Interrupt. For "MSI", as first defined
 234 *			in PCI 2.2, this can be between 1 and
 235 *			32. For "MSI-X," as first defined in PCI
 236 *			3.0, this must be 1, as each MSI-X table
 237 *			entry would have its own descriptor.
 238 * @processor_count:	number of bits enabled in array.
 239 * @processor_array:	All the target virtual processors.
 240 */
 241struct hv_msi_desc2 {
 242	u8	vector;
 243	u8	delivery_mode;
 244	u16	vector_count;
 245	u16	processor_count;
 246	u16	processor_array[32];
 247} __packed;
 248
 249/*
 250 * struct hv_msi_desc3 - 1.3 version of hv_msi_desc
 251 *	Everything is the same as in 'hv_msi_desc2' except that the size of the
 252 *	'vector' field is larger to support bigger vector values. For ex: LPI
 253 *	vectors on ARM.
 254 */
 255struct hv_msi_desc3 {
 256	u32	vector;
 257	u8	delivery_mode;
 258	u8	reserved;
 259	u16	vector_count;
 260	u16	processor_count;
 261	u16	processor_array[32];
 262} __packed;
 263
 264/**
 265 * struct tran_int_desc
 266 * @reserved:		unused, padding
 267 * @vector_count:	same as in hv_msi_desc
 268 * @data:		This is the "data payload" value that is
 269 *			written by the device when it generates
 270 *			a message-signaled interrupt, either MSI
 271 *			or MSI-X.
 272 * @address:		This is the address to which the data
 273 *			payload is written on interrupt
 274 *			generation.
 275 */
 276struct tran_int_desc {
 277	u16	reserved;
 278	u16	vector_count;
 279	u32	data;
 280	u64	address;
 281} __packed;
 282
 283/*
 284 * A generic message format for virtual PCI.
 285 * Specific message formats are defined later in the file.
 286 */
 287
 288struct pci_message {
 289	u32 type;
 290} __packed;
 291
 292struct pci_child_message {
 293	struct pci_message message_type;
 294	union win_slot_encoding wslot;
 295} __packed;
 296
 297struct pci_incoming_message {
 298	struct vmpacket_descriptor hdr;
 299	struct pci_message message_type;
 300} __packed;
 301
 302struct pci_response {
 303	struct vmpacket_descriptor hdr;
 304	s32 status;			/* negative values are failures */
 305} __packed;
 306
 307struct pci_packet {
 308	void (*completion_func)(void *context, struct pci_response *resp,
 309				int resp_packet_size);
 310	void *compl_ctxt;
 311
 312	struct pci_message message[];
 313};
 314
 315/*
 316 * Specific message types supporting the PCI protocol.
 317 */
 318
 319/*
 320 * Version negotiation message. Sent from the guest to the host.
 321 * The guest is free to try different versions until the host
 322 * accepts the version.
 323 *
 324 * pci_version: The protocol version requested.
 325 * is_last_attempt: If TRUE, this is the last version guest will request.
 326 * reservedz: Reserved field, set to zero.
 327 */
 328
 329struct pci_version_request {
 330	struct pci_message message_type;
 331	u32 protocol_version;
 332} __packed;
 333
 334/*
 335 * Bus D0 Entry.  This is sent from the guest to the host when the virtual
 336 * bus (PCI Express port) is ready for action.
 337 */
 338
 339struct pci_bus_d0_entry {
 340	struct pci_message message_type;
 341	u32 reserved;
 342	u64 mmio_base;
 343} __packed;
 344
 345struct pci_bus_relations {
 346	struct pci_incoming_message incoming;
 347	u32 device_count;
 348	struct pci_function_description func[];
 349} __packed;
 350
 351struct pci_bus_relations2 {
 352	struct pci_incoming_message incoming;
 353	u32 device_count;
 354	struct pci_function_description2 func[];
 355} __packed;
 356
 357struct pci_q_res_req_response {
 358	struct vmpacket_descriptor hdr;
 359	s32 status;			/* negative values are failures */
 360	u32 probed_bar[PCI_STD_NUM_BARS];
 361} __packed;
 362
 363struct pci_set_power {
 364	struct pci_message message_type;
 365	union win_slot_encoding wslot;
 366	u32 power_state;		/* In Windows terms */
 367	u32 reserved;
 368} __packed;
 369
 370struct pci_set_power_response {
 371	struct vmpacket_descriptor hdr;
 372	s32 status;			/* negative values are failures */
 373	union win_slot_encoding wslot;
 374	u32 resultant_state;		/* In Windows terms */
 375	u32 reserved;
 376} __packed;
 377
 378struct pci_resources_assigned {
 379	struct pci_message message_type;
 380	union win_slot_encoding wslot;
 381	u8 memory_range[0x14][6];	/* not used here */
 382	u32 msi_descriptors;
 383	u32 reserved[4];
 384} __packed;
 385
 386struct pci_resources_assigned2 {
 387	struct pci_message message_type;
 388	union win_slot_encoding wslot;
 389	u8 memory_range[0x14][6];	/* not used here */
 390	u32 msi_descriptor_count;
 391	u8 reserved[70];
 392} __packed;
 393
 394struct pci_create_interrupt {
 395	struct pci_message message_type;
 396	union win_slot_encoding wslot;
 397	struct hv_msi_desc int_desc;
 398} __packed;
 399
 400struct pci_create_int_response {
 401	struct pci_response response;
 402	u32 reserved;
 403	struct tran_int_desc int_desc;
 404} __packed;
 405
 406struct pci_create_interrupt2 {
 407	struct pci_message message_type;
 408	union win_slot_encoding wslot;
 409	struct hv_msi_desc2 int_desc;
 410} __packed;
 411
 412struct pci_create_interrupt3 {
 413	struct pci_message message_type;
 414	union win_slot_encoding wslot;
 415	struct hv_msi_desc3 int_desc;
 416} __packed;
 417
 418struct pci_delete_interrupt {
 419	struct pci_message message_type;
 420	union win_slot_encoding wslot;
 421	struct tran_int_desc int_desc;
 422} __packed;
 423
 424/*
 425 * Note: the VM must pass a valid block id, wslot and bytes_requested.
 426 */
 427struct pci_read_block {
 428	struct pci_message message_type;
 429	u32 block_id;
 430	union win_slot_encoding wslot;
 431	u32 bytes_requested;
 432} __packed;
 433
 434struct pci_read_block_response {
 435	struct vmpacket_descriptor hdr;
 436	u32 status;
 437	u8 bytes[HV_CONFIG_BLOCK_SIZE_MAX];
 438} __packed;
 439
 440/*
 441 * Note: the VM must pass a valid block id, wslot and byte_count.
 442 */
 443struct pci_write_block {
 444	struct pci_message message_type;
 445	u32 block_id;
 446	union win_slot_encoding wslot;
 447	u32 byte_count;
 448	u8 bytes[HV_CONFIG_BLOCK_SIZE_MAX];
 449} __packed;
 450
 451struct pci_dev_inval_block {
 452	struct pci_incoming_message incoming;
 453	union win_slot_encoding wslot;
 454	u64 block_mask;
 455} __packed;
 456
 457struct pci_dev_incoming {
 458	struct pci_incoming_message incoming;
 459	union win_slot_encoding wslot;
 460} __packed;
 461
 462struct pci_eject_response {
 463	struct pci_message message_type;
 464	union win_slot_encoding wslot;
 465	u32 status;
 466} __packed;
 467
 468static int pci_ring_size = (4 * PAGE_SIZE);
 469
 470/*
 471 * Driver specific state.
 472 */
 473
 474enum hv_pcibus_state {
 475	hv_pcibus_init = 0,
 476	hv_pcibus_probed,
 477	hv_pcibus_installed,
 478	hv_pcibus_removing,
 479	hv_pcibus_maximum
 480};
 481
 482struct hv_pcibus_device {
 483#ifdef CONFIG_X86
 484	struct pci_sysdata sysdata;
 485#elif defined(CONFIG_ARM64)
 486	struct pci_config_window sysdata;
 487#endif
 488	struct pci_host_bridge *bridge;
 489	struct fwnode_handle *fwnode;
 490	/* Protocol version negotiated with the host */
 491	enum pci_protocol_version_t protocol_version;
 492
 493	struct mutex state_lock;
 494	enum hv_pcibus_state state;
 495
 496	struct hv_device *hdev;
 497	resource_size_t low_mmio_space;
 498	resource_size_t high_mmio_space;
 499	struct resource *mem_config;
 500	struct resource *low_mmio_res;
 501	struct resource *high_mmio_res;
 502	struct completion *survey_event;
 503	struct pci_bus *pci_bus;
 504	spinlock_t config_lock;	/* Avoid two threads writing index page */
 505	spinlock_t device_list_lock;	/* Protect lists below */
 506	void __iomem *cfg_addr;
 507
 508	struct list_head children;
 509	struct list_head dr_list;
 510
 511	struct msi_domain_info msi_info;
 512	struct irq_domain *irq_domain;
 513
 514	struct workqueue_struct *wq;
 515
 516	/* Highest slot of child device with resources allocated */
 517	int wslot_res_allocated;
 518	bool use_calls; /* Use hypercalls to access mmio cfg space */
 519};
 520
 521/*
 522 * Tracks "Device Relations" messages from the host, which must be both
 523 * processed in order and deferred so that they don't run in the context
 524 * of the incoming packet callback.
 525 */
 526struct hv_dr_work {
 527	struct work_struct wrk;
 528	struct hv_pcibus_device *bus;
 529};
 530
 531struct hv_pcidev_description {
 532	u16	v_id;	/* vendor ID */
 533	u16	d_id;	/* device ID */
 534	u8	rev;
 535	u8	prog_intf;
 536	u8	subclass;
 537	u8	base_class;
 538	u32	subsystem_id;
 539	union	win_slot_encoding win_slot;
 540	u32	ser;	/* serial number */
 541	u32	flags;
 542	u16	virtual_numa_node;
 543};
 544
 545struct hv_dr_state {
 546	struct list_head list_entry;
 547	u32 device_count;
 548	struct hv_pcidev_description func[] __counted_by(device_count);
 549};
 550
 551struct hv_pci_dev {
 552	/* List protected by pci_rescan_remove_lock */
 553	struct list_head list_entry;
 554	refcount_t refs;
 555	struct pci_slot *pci_slot;
 556	struct hv_pcidev_description desc;
 557	bool reported_missing;
 558	struct hv_pcibus_device *hbus;
 559	struct work_struct wrk;
 560
 561	void (*block_invalidate)(void *context, u64 block_mask);
 562	void *invalidate_context;
 563
 564	/*
 565	 * What would be observed if one wrote 0xFFFFFFFF to a BAR and then
 566	 * read it back, for each of the BAR offsets within config space.
 567	 */
 568	u32 probed_bar[PCI_STD_NUM_BARS];
 569};
 570
 571struct hv_pci_compl {
 572	struct completion host_event;
 573	s32 completion_status;
 574};
 575
 576static void hv_pci_onchannelcallback(void *context);
 577
 578#ifdef CONFIG_X86
 579#define DELIVERY_MODE	APIC_DELIVERY_MODE_FIXED
 580#define FLOW_HANDLER	handle_edge_irq
 581#define FLOW_NAME	"edge"
 582
 583static int hv_pci_irqchip_init(void)
 584{
 585	return 0;
 586}
 587
 588static struct irq_domain *hv_pci_get_root_domain(void)
 589{
 590	return x86_vector_domain;
 591}
 592
 593static unsigned int hv_msi_get_int_vector(struct irq_data *data)
 594{
 595	struct irq_cfg *cfg = irqd_cfg(data);
 596
 597	return cfg->vector;
 598}
 599
 600#define hv_msi_prepare		pci_msi_prepare
 601
 602/**
 603 * hv_arch_irq_unmask() - "Unmask" the IRQ by setting its current
 604 * affinity.
 605 * @data:	Describes the IRQ
 606 *
 607 * Build new a destination for the MSI and make a hypercall to
 608 * update the Interrupt Redirection Table. "Device Logical ID"
 609 * is built out of this PCI bus's instance GUID and the function
 610 * number of the device.
 611 */
 612static void hv_arch_irq_unmask(struct irq_data *data)
 613{
 614	struct msi_desc *msi_desc = irq_data_get_msi_desc(data);
 615	struct hv_retarget_device_interrupt *params;
 616	struct tran_int_desc *int_desc;
 617	struct hv_pcibus_device *hbus;
 618	const struct cpumask *dest;
 619	cpumask_var_t tmp;
 620	struct pci_bus *pbus;
 621	struct pci_dev *pdev;
 622	unsigned long flags;
 623	u32 var_size = 0;
 624	int cpu, nr_bank;
 625	u64 res;
 626
 627	dest = irq_data_get_effective_affinity_mask(data);
 628	pdev = msi_desc_to_pci_dev(msi_desc);
 629	pbus = pdev->bus;
 630	hbus = container_of(pbus->sysdata, struct hv_pcibus_device, sysdata);
 631	int_desc = data->chip_data;
 632	if (!int_desc) {
 633		dev_warn(&hbus->hdev->device, "%s() can not unmask irq %u\n",
 634			 __func__, data->irq);
 635		return;
 636	}
 637
 638	local_irq_save(flags);
 639
 640	params = *this_cpu_ptr(hyperv_pcpu_input_arg);
 641	memset(params, 0, sizeof(*params));
 642	params->partition_id = HV_PARTITION_ID_SELF;
 643	params->int_entry.source = HV_INTERRUPT_SOURCE_MSI;
 644	params->int_entry.msi_entry.address.as_uint32 = int_desc->address & 0xffffffff;
 645	params->int_entry.msi_entry.data.as_uint32 = int_desc->data;
 646	params->device_id = (hbus->hdev->dev_instance.b[5] << 24) |
 647			   (hbus->hdev->dev_instance.b[4] << 16) |
 648			   (hbus->hdev->dev_instance.b[7] << 8) |
 649			   (hbus->hdev->dev_instance.b[6] & 0xf8) |
 650			   PCI_FUNC(pdev->devfn);
 651	params->int_target.vector = hv_msi_get_int_vector(data);
 652
 653	if (hbus->protocol_version >= PCI_PROTOCOL_VERSION_1_2) {
 654		/*
 655		 * PCI_PROTOCOL_VERSION_1_2 supports the VP_SET version of the
 656		 * HVCALL_RETARGET_INTERRUPT hypercall, which also coincides
 657		 * with >64 VP support.
 658		 * ms_hyperv.hints & HV_X64_EX_PROCESSOR_MASKS_RECOMMENDED
 659		 * is not sufficient for this hypercall.
 660		 */
 661		params->int_target.flags |=
 662			HV_DEVICE_INTERRUPT_TARGET_PROCESSOR_SET;
 663
 664		if (!alloc_cpumask_var(&tmp, GFP_ATOMIC)) {
 665			res = 1;
 666			goto out;
 667		}
 668
 669		cpumask_and(tmp, dest, cpu_online_mask);
 670		nr_bank = cpumask_to_vpset(&params->int_target.vp_set, tmp);
 671		free_cpumask_var(tmp);
 672
 673		if (nr_bank <= 0) {
 674			res = 1;
 675			goto out;
 676		}
 677
 678		/*
 679		 * var-sized hypercall, var-size starts after vp_mask (thus
 680		 * vp_set.format does not count, but vp_set.valid_bank_mask
 681		 * does).
 682		 */
 683		var_size = 1 + nr_bank;
 684	} else {
 685		for_each_cpu_and(cpu, dest, cpu_online_mask) {
 686			params->int_target.vp_mask |=
 687				(1ULL << hv_cpu_number_to_vp_number(cpu));
 688		}
 689	}
 690
 691	res = hv_do_hypercall(HVCALL_RETARGET_INTERRUPT | (var_size << 17),
 692			      params, NULL);
 693
 694out:
 695	local_irq_restore(flags);
 696
 697	/*
 698	 * During hibernation, when a CPU is offlined, the kernel tries
 699	 * to move the interrupt to the remaining CPUs that haven't
 700	 * been offlined yet. In this case, the below hv_do_hypercall()
 701	 * always fails since the vmbus channel has been closed:
 702	 * refer to cpu_disable_common() -> fixup_irqs() ->
 703	 * irq_migrate_all_off_this_cpu() -> migrate_one_irq().
 704	 *
 705	 * Suppress the error message for hibernation because the failure
 706	 * during hibernation does not matter (at this time all the devices
 707	 * have been frozen). Note: the correct affinity info is still updated
 708	 * into the irqdata data structure in migrate_one_irq() ->
 709	 * irq_do_set_affinity(), so later when the VM resumes,
 710	 * hv_pci_restore_msi_state() is able to correctly restore the
 711	 * interrupt with the correct affinity.
 712	 */
 713	if (!hv_result_success(res) && hbus->state != hv_pcibus_removing)
 714		dev_err(&hbus->hdev->device,
 715			"%s() failed: %#llx", __func__, res);
 716}
 717#elif defined(CONFIG_ARM64)
 718/*
 719 * SPI vectors to use for vPCI; arch SPIs range is [32, 1019], but leaving a bit
 720 * of room at the start to allow for SPIs to be specified through ACPI and
 721 * starting with a power of two to satisfy power of 2 multi-MSI requirement.
 722 */
 723#define HV_PCI_MSI_SPI_START	64
 724#define HV_PCI_MSI_SPI_NR	(1020 - HV_PCI_MSI_SPI_START)
 725#define DELIVERY_MODE		0
 726#define FLOW_HANDLER		NULL
 727#define FLOW_NAME		NULL
 728#define hv_msi_prepare		NULL
 729
 730struct hv_pci_chip_data {
 731	DECLARE_BITMAP(spi_map, HV_PCI_MSI_SPI_NR);
 732	struct mutex	map_lock;
 733};
 734
 735/* Hyper-V vPCI MSI GIC IRQ domain */
 736static struct irq_domain *hv_msi_gic_irq_domain;
 737
 738/* Hyper-V PCI MSI IRQ chip */
 739static struct irq_chip hv_arm64_msi_irq_chip = {
 740	.name = "MSI",
 741	.irq_set_affinity = irq_chip_set_affinity_parent,
 742	.irq_eoi = irq_chip_eoi_parent,
 743	.irq_mask = irq_chip_mask_parent,
 744	.irq_unmask = irq_chip_unmask_parent
 745};
 746
 747static unsigned int hv_msi_get_int_vector(struct irq_data *irqd)
 748{
 749	return irqd->parent_data->hwirq;
 750}
 751
 752/*
 753 * @nr_bm_irqs:		Indicates the number of IRQs that were allocated from
 754 *			the bitmap.
 755 * @nr_dom_irqs:	Indicates the number of IRQs that were allocated from
 756 *			the parent domain.
 757 */
 758static void hv_pci_vec_irq_free(struct irq_domain *domain,
 759				unsigned int virq,
 760				unsigned int nr_bm_irqs,
 761				unsigned int nr_dom_irqs)
 762{
 763	struct hv_pci_chip_data *chip_data = domain->host_data;
 764	struct irq_data *d = irq_domain_get_irq_data(domain, virq);
 765	int first = d->hwirq - HV_PCI_MSI_SPI_START;
 766	int i;
 767
 768	mutex_lock(&chip_data->map_lock);
 769	bitmap_release_region(chip_data->spi_map,
 770			      first,
 771			      get_count_order(nr_bm_irqs));
 772	mutex_unlock(&chip_data->map_lock);
 773	for (i = 0; i < nr_dom_irqs; i++) {
 774		if (i)
 775			d = irq_domain_get_irq_data(domain, virq + i);
 776		irq_domain_reset_irq_data(d);
 777	}
 778
 779	irq_domain_free_irqs_parent(domain, virq, nr_dom_irqs);
 780}
 781
 782static void hv_pci_vec_irq_domain_free(struct irq_domain *domain,
 783				       unsigned int virq,
 784				       unsigned int nr_irqs)
 785{
 786	hv_pci_vec_irq_free(domain, virq, nr_irqs, nr_irqs);
 787}
 788
 789static int hv_pci_vec_alloc_device_irq(struct irq_domain *domain,
 790				       unsigned int nr_irqs,
 791				       irq_hw_number_t *hwirq)
 792{
 793	struct hv_pci_chip_data *chip_data = domain->host_data;
 794	int index;
 795
 796	/* Find and allocate region from the SPI bitmap */
 797	mutex_lock(&chip_data->map_lock);
 798	index = bitmap_find_free_region(chip_data->spi_map,
 799					HV_PCI_MSI_SPI_NR,
 800					get_count_order(nr_irqs));
 801	mutex_unlock(&chip_data->map_lock);
 802	if (index < 0)
 803		return -ENOSPC;
 804
 805	*hwirq = index + HV_PCI_MSI_SPI_START;
 806
 807	return 0;
 808}
 809
 810static int hv_pci_vec_irq_gic_domain_alloc(struct irq_domain *domain,
 811					   unsigned int virq,
 812					   irq_hw_number_t hwirq)
 813{
 814	struct irq_fwspec fwspec;
 815	struct irq_data *d;
 816	int ret;
 817
 818	fwspec.fwnode = domain->parent->fwnode;
 819	fwspec.param_count = 2;
 820	fwspec.param[0] = hwirq;
 821	fwspec.param[1] = IRQ_TYPE_EDGE_RISING;
 822
 823	ret = irq_domain_alloc_irqs_parent(domain, virq, 1, &fwspec);
 824	if (ret)
 825		return ret;
 826
 827	/*
 828	 * Since the interrupt specifier is not coming from ACPI or DT, the
 829	 * trigger type will need to be set explicitly. Otherwise, it will be
 830	 * set to whatever is in the GIC configuration.
 831	 */
 832	d = irq_domain_get_irq_data(domain->parent, virq);
 833
 834	return d->chip->irq_set_type(d, IRQ_TYPE_EDGE_RISING);
 835}
 836
 837static int hv_pci_vec_irq_domain_alloc(struct irq_domain *domain,
 838				       unsigned int virq, unsigned int nr_irqs,
 839				       void *args)
 840{
 841	irq_hw_number_t hwirq;
 842	unsigned int i;
 843	int ret;
 844
 845	ret = hv_pci_vec_alloc_device_irq(domain, nr_irqs, &hwirq);
 846	if (ret)
 847		return ret;
 848
 849	for (i = 0; i < nr_irqs; i++) {
 850		ret = hv_pci_vec_irq_gic_domain_alloc(domain, virq + i,
 851						      hwirq + i);
 852		if (ret) {
 853			hv_pci_vec_irq_free(domain, virq, nr_irqs, i);
 854			return ret;
 855		}
 856
 857		irq_domain_set_hwirq_and_chip(domain, virq + i,
 858					      hwirq + i,
 859					      &hv_arm64_msi_irq_chip,
 860					      domain->host_data);
 861		pr_debug("pID:%d vID:%u\n", (int)(hwirq + i), virq + i);
 862	}
 863
 864	return 0;
 865}
 866
 867/*
 868 * Pick the first cpu as the irq affinity that can be temporarily used for
 869 * composing MSI from the hypervisor. GIC will eventually set the right
 870 * affinity for the irq and the 'unmask' will retarget the interrupt to that
 871 * cpu.
 872 */
 873static int hv_pci_vec_irq_domain_activate(struct irq_domain *domain,
 874					  struct irq_data *irqd, bool reserve)
 875{
 876	int cpu = cpumask_first(cpu_present_mask);
 877
 878	irq_data_update_effective_affinity(irqd, cpumask_of(cpu));
 879
 880	return 0;
 881}
 882
 883static const struct irq_domain_ops hv_pci_domain_ops = {
 884	.alloc	= hv_pci_vec_irq_domain_alloc,
 885	.free	= hv_pci_vec_irq_domain_free,
 886	.activate = hv_pci_vec_irq_domain_activate,
 887};
 888
 889static int hv_pci_irqchip_init(void)
 890{
 891	static struct hv_pci_chip_data *chip_data;
 892	struct fwnode_handle *fn = NULL;
 893	int ret = -ENOMEM;
 894
 895	chip_data = kzalloc(sizeof(*chip_data), GFP_KERNEL);
 896	if (!chip_data)
 897		return ret;
 898
 899	mutex_init(&chip_data->map_lock);
 900	fn = irq_domain_alloc_named_fwnode("hv_vpci_arm64");
 901	if (!fn)
 902		goto free_chip;
 903
 904	/*
 905	 * IRQ domain once enabled, should not be removed since there is no
 906	 * way to ensure that all the corresponding devices are also gone and
 907	 * no interrupts will be generated.
 908	 */
 909	hv_msi_gic_irq_domain = acpi_irq_create_hierarchy(0, HV_PCI_MSI_SPI_NR,
 910							  fn, &hv_pci_domain_ops,
 911							  chip_data);
 912
 913	if (!hv_msi_gic_irq_domain) {
 914		pr_err("Failed to create Hyper-V arm64 vPCI MSI IRQ domain\n");
 915		goto free_chip;
 916	}
 917
 918	return 0;
 919
 920free_chip:
 921	kfree(chip_data);
 922	if (fn)
 923		irq_domain_free_fwnode(fn);
 924
 925	return ret;
 926}
 927
 928static struct irq_domain *hv_pci_get_root_domain(void)
 929{
 930	return hv_msi_gic_irq_domain;
 931}
 932
 933/*
 934 * SPIs are used for interrupts of PCI devices and SPIs is managed via GICD
 935 * registers which Hyper-V already supports, so no hypercall needed.
 936 */
 937static void hv_arch_irq_unmask(struct irq_data *data) { }
 938#endif /* CONFIG_ARM64 */
 939
 940/**
 941 * hv_pci_generic_compl() - Invoked for a completion packet
 942 * @context:		Set up by the sender of the packet.
 943 * @resp:		The response packet
 944 * @resp_packet_size:	Size in bytes of the packet
 945 *
 946 * This function is used to trigger an event and report status
 947 * for any message for which the completion packet contains a
 948 * status and nothing else.
 949 */
 950static void hv_pci_generic_compl(void *context, struct pci_response *resp,
 951				 int resp_packet_size)
 952{
 953	struct hv_pci_compl *comp_pkt = context;
 954
 955	comp_pkt->completion_status = resp->status;
 956	complete(&comp_pkt->host_event);
 957}
 958
 959static struct hv_pci_dev *get_pcichild_wslot(struct hv_pcibus_device *hbus,
 960						u32 wslot);
 961
 962static void get_pcichild(struct hv_pci_dev *hpdev)
 963{
 964	refcount_inc(&hpdev->refs);
 965}
 966
 967static void put_pcichild(struct hv_pci_dev *hpdev)
 968{
 969	if (refcount_dec_and_test(&hpdev->refs))
 970		kfree(hpdev);
 971}
 972
 973/*
 974 * There is no good way to get notified from vmbus_onoffer_rescind(),
 975 * so let's use polling here, since this is not a hot path.
 976 */
 977static int wait_for_response(struct hv_device *hdev,
 978			     struct completion *comp)
 979{
 980	while (true) {
 981		if (hdev->channel->rescind) {
 982			dev_warn_once(&hdev->device, "The device is gone.\n");
 983			return -ENODEV;
 984		}
 985
 986		if (wait_for_completion_timeout(comp, HZ / 10))
 987			break;
 988	}
 989
 990	return 0;
 991}
 992
 993/**
 994 * devfn_to_wslot() - Convert from Linux PCI slot to Windows
 995 * @devfn:	The Linux representation of PCI slot
 996 *
 997 * Windows uses a slightly different representation of PCI slot.
 998 *
 999 * Return: The Windows representation
1000 */
1001static u32 devfn_to_wslot(int devfn)
1002{
1003	union win_slot_encoding wslot;
1004
1005	wslot.slot = 0;
1006	wslot.bits.dev = PCI_SLOT(devfn);
1007	wslot.bits.func = PCI_FUNC(devfn);
1008
1009	return wslot.slot;
1010}
1011
1012/**
1013 * wslot_to_devfn() - Convert from Windows PCI slot to Linux
1014 * @wslot:	The Windows representation of PCI slot
1015 *
1016 * Windows uses a slightly different representation of PCI slot.
1017 *
1018 * Return: The Linux representation
1019 */
1020static int wslot_to_devfn(u32 wslot)
1021{
1022	union win_slot_encoding slot_no;
1023
1024	slot_no.slot = wslot;
1025	return PCI_DEVFN(slot_no.bits.dev, slot_no.bits.func);
1026}
1027
1028static void hv_pci_read_mmio(struct device *dev, phys_addr_t gpa, int size, u32 *val)
1029{
1030	struct hv_mmio_read_input *in;
1031	struct hv_mmio_read_output *out;
1032	u64 ret;
1033
1034	/*
1035	 * Must be called with interrupts disabled so it is safe
1036	 * to use the per-cpu input argument page.  Use it for
1037	 * both input and output.
1038	 */
1039	in = *this_cpu_ptr(hyperv_pcpu_input_arg);
1040	out = *this_cpu_ptr(hyperv_pcpu_input_arg) + sizeof(*in);
1041	in->gpa = gpa;
1042	in->size = size;
1043
1044	ret = hv_do_hypercall(HVCALL_MMIO_READ, in, out);
1045	if (hv_result_success(ret)) {
1046		switch (size) {
1047		case 1:
1048			*val = *(u8 *)(out->data);
1049			break;
1050		case 2:
1051			*val = *(u16 *)(out->data);
1052			break;
1053		default:
1054			*val = *(u32 *)(out->data);
1055			break;
1056		}
1057	} else
1058		dev_err(dev, "MMIO read hypercall error %llx addr %llx size %d\n",
1059				ret, gpa, size);
1060}
1061
1062static void hv_pci_write_mmio(struct device *dev, phys_addr_t gpa, int size, u32 val)
1063{
1064	struct hv_mmio_write_input *in;
1065	u64 ret;
1066
1067	/*
1068	 * Must be called with interrupts disabled so it is safe
1069	 * to use the per-cpu input argument memory.
1070	 */
1071	in = *this_cpu_ptr(hyperv_pcpu_input_arg);
1072	in->gpa = gpa;
1073	in->size = size;
1074	switch (size) {
1075	case 1:
1076		*(u8 *)(in->data) = val;
1077		break;
1078	case 2:
1079		*(u16 *)(in->data) = val;
1080		break;
1081	default:
1082		*(u32 *)(in->data) = val;
1083		break;
1084	}
1085
1086	ret = hv_do_hypercall(HVCALL_MMIO_WRITE, in, NULL);
1087	if (!hv_result_success(ret))
1088		dev_err(dev, "MMIO write hypercall error %llx addr %llx size %d\n",
1089				ret, gpa, size);
1090}
1091
1092/*
1093 * PCI Configuration Space for these root PCI buses is implemented as a pair
1094 * of pages in memory-mapped I/O space.  Writing to the first page chooses
1095 * the PCI function being written or read.  Once the first page has been
1096 * written to, the following page maps in the entire configuration space of
1097 * the function.
1098 */
1099
1100/**
1101 * _hv_pcifront_read_config() - Internal PCI config read
1102 * @hpdev:	The PCI driver's representation of the device
1103 * @where:	Offset within config space
1104 * @size:	Size of the transfer
1105 * @val:	Pointer to the buffer receiving the data
1106 */
1107static void _hv_pcifront_read_config(struct hv_pci_dev *hpdev, int where,
1108				     int size, u32 *val)
1109{
1110	struct hv_pcibus_device *hbus = hpdev->hbus;
1111	struct device *dev = &hbus->hdev->device;
1112	int offset = where + CFG_PAGE_OFFSET;
1113	unsigned long flags;
1114
1115	/*
1116	 * If the attempt is to read the IDs or the ROM BAR, simulate that.
1117	 */
1118	if (where + size <= PCI_COMMAND) {
1119		memcpy(val, ((u8 *)&hpdev->desc.v_id) + where, size);
1120	} else if (where >= PCI_CLASS_REVISION && where + size <=
1121		   PCI_CACHE_LINE_SIZE) {
1122		memcpy(val, ((u8 *)&hpdev->desc.rev) + where -
1123		       PCI_CLASS_REVISION, size);
1124	} else if (where >= PCI_SUBSYSTEM_VENDOR_ID && where + size <=
1125		   PCI_ROM_ADDRESS) {
1126		memcpy(val, (u8 *)&hpdev->desc.subsystem_id + where -
1127		       PCI_SUBSYSTEM_VENDOR_ID, size);
1128	} else if (where >= PCI_ROM_ADDRESS && where + size <=
1129		   PCI_CAPABILITY_LIST) {
1130		/* ROM BARs are unimplemented */
1131		*val = 0;
1132	} else if (where >= PCI_INTERRUPT_LINE && where + size <=
1133		   PCI_INTERRUPT_PIN) {
1134		/*
1135		 * Interrupt Line and Interrupt PIN are hard-wired to zero
1136		 * because this front-end only supports message-signaled
1137		 * interrupts.
1138		 */
1139		*val = 0;
1140	} else if (where + size <= CFG_PAGE_SIZE) {
1141
1142		spin_lock_irqsave(&hbus->config_lock, flags);
1143		if (hbus->use_calls) {
1144			phys_addr_t addr = hbus->mem_config->start + offset;
1145
1146			hv_pci_write_mmio(dev, hbus->mem_config->start, 4,
1147						hpdev->desc.win_slot.slot);
1148			hv_pci_read_mmio(dev, addr, size, val);
1149		} else {
1150			void __iomem *addr = hbus->cfg_addr + offset;
1151
1152			/* Choose the function to be read. (See comment above) */
1153			writel(hpdev->desc.win_slot.slot, hbus->cfg_addr);
1154			/* Make sure the function was chosen before reading. */
1155			mb();
1156			/* Read from that function's config space. */
1157			switch (size) {
1158			case 1:
1159				*val = readb(addr);
1160				break;
1161			case 2:
1162				*val = readw(addr);
1163				break;
1164			default:
1165				*val = readl(addr);
1166				break;
1167			}
1168			/*
1169			 * Make sure the read was done before we release the
1170			 * spinlock allowing consecutive reads/writes.
1171			 */
1172			mb();
1173		}
1174		spin_unlock_irqrestore(&hbus->config_lock, flags);
1175	} else {
1176		dev_err(dev, "Attempt to read beyond a function's config space.\n");
1177	}
1178}
1179
1180static u16 hv_pcifront_get_vendor_id(struct hv_pci_dev *hpdev)
1181{
1182	struct hv_pcibus_device *hbus = hpdev->hbus;
1183	struct device *dev = &hbus->hdev->device;
1184	u32 val;
1185	u16 ret;
1186	unsigned long flags;
1187
1188	spin_lock_irqsave(&hbus->config_lock, flags);
1189
1190	if (hbus->use_calls) {
1191		phys_addr_t addr = hbus->mem_config->start +
1192					 CFG_PAGE_OFFSET + PCI_VENDOR_ID;
1193
1194		hv_pci_write_mmio(dev, hbus->mem_config->start, 4,
1195					hpdev->desc.win_slot.slot);
1196		hv_pci_read_mmio(dev, addr, 2, &val);
1197		ret = val;  /* Truncates to 16 bits */
1198	} else {
1199		void __iomem *addr = hbus->cfg_addr + CFG_PAGE_OFFSET +
1200					     PCI_VENDOR_ID;
1201		/* Choose the function to be read. (See comment above) */
1202		writel(hpdev->desc.win_slot.slot, hbus->cfg_addr);
1203		/* Make sure the function was chosen before we start reading. */
1204		mb();
1205		/* Read from that function's config space. */
1206		ret = readw(addr);
1207		/*
1208		 * mb() is not required here, because the
1209		 * spin_unlock_irqrestore() is a barrier.
1210		 */
1211	}
1212
1213	spin_unlock_irqrestore(&hbus->config_lock, flags);
1214
1215	return ret;
1216}
1217
1218/**
1219 * _hv_pcifront_write_config() - Internal PCI config write
1220 * @hpdev:	The PCI driver's representation of the device
1221 * @where:	Offset within config space
1222 * @size:	Size of the transfer
1223 * @val:	The data being transferred
1224 */
1225static void _hv_pcifront_write_config(struct hv_pci_dev *hpdev, int where,
1226				      int size, u32 val)
1227{
1228	struct hv_pcibus_device *hbus = hpdev->hbus;
1229	struct device *dev = &hbus->hdev->device;
1230	int offset = where + CFG_PAGE_OFFSET;
1231	unsigned long flags;
1232
1233	if (where >= PCI_SUBSYSTEM_VENDOR_ID &&
1234	    where + size <= PCI_CAPABILITY_LIST) {
1235		/* SSIDs and ROM BARs are read-only */
1236	} else if (where >= PCI_COMMAND && where + size <= CFG_PAGE_SIZE) {
1237		spin_lock_irqsave(&hbus->config_lock, flags);
1238
1239		if (hbus->use_calls) {
1240			phys_addr_t addr = hbus->mem_config->start + offset;
1241
1242			hv_pci_write_mmio(dev, hbus->mem_config->start, 4,
1243						hpdev->desc.win_slot.slot);
1244			hv_pci_write_mmio(dev, addr, size, val);
1245		} else {
1246			void __iomem *addr = hbus->cfg_addr + offset;
1247
1248			/* Choose the function to write. (See comment above) */
1249			writel(hpdev->desc.win_slot.slot, hbus->cfg_addr);
1250			/* Make sure the function was chosen before writing. */
1251			wmb();
1252			/* Write to that function's config space. */
1253			switch (size) {
1254			case 1:
1255				writeb(val, addr);
1256				break;
1257			case 2:
1258				writew(val, addr);
1259				break;
1260			default:
1261				writel(val, addr);
1262				break;
1263			}
1264			/*
1265			 * Make sure the write was done before we release the
1266			 * spinlock allowing consecutive reads/writes.
1267			 */
1268			mb();
1269		}
1270		spin_unlock_irqrestore(&hbus->config_lock, flags);
1271	} else {
1272		dev_err(dev, "Attempt to write beyond a function's config space.\n");
1273	}
1274}
1275
1276/**
1277 * hv_pcifront_read_config() - Read configuration space
1278 * @bus: PCI Bus structure
1279 * @devfn: Device/function
1280 * @where: Offset from base
1281 * @size: Byte/word/dword
1282 * @val: Value to be read
1283 *
1284 * Return: PCIBIOS_SUCCESSFUL on success
1285 *	   PCIBIOS_DEVICE_NOT_FOUND on failure
1286 */
1287static int hv_pcifront_read_config(struct pci_bus *bus, unsigned int devfn,
1288				   int where, int size, u32 *val)
1289{
1290	struct hv_pcibus_device *hbus =
1291		container_of(bus->sysdata, struct hv_pcibus_device, sysdata);
1292	struct hv_pci_dev *hpdev;
1293
1294	hpdev = get_pcichild_wslot(hbus, devfn_to_wslot(devfn));
1295	if (!hpdev)
1296		return PCIBIOS_DEVICE_NOT_FOUND;
1297
1298	_hv_pcifront_read_config(hpdev, where, size, val);
1299
1300	put_pcichild(hpdev);
1301	return PCIBIOS_SUCCESSFUL;
1302}
1303
1304/**
1305 * hv_pcifront_write_config() - Write configuration space
1306 * @bus: PCI Bus structure
1307 * @devfn: Device/function
1308 * @where: Offset from base
1309 * @size: Byte/word/dword
1310 * @val: Value to be written to device
1311 *
1312 * Return: PCIBIOS_SUCCESSFUL on success
1313 *	   PCIBIOS_DEVICE_NOT_FOUND on failure
1314 */
1315static int hv_pcifront_write_config(struct pci_bus *bus, unsigned int devfn,
1316				    int where, int size, u32 val)
1317{
1318	struct hv_pcibus_device *hbus =
1319	    container_of(bus->sysdata, struct hv_pcibus_device, sysdata);
1320	struct hv_pci_dev *hpdev;
1321
1322	hpdev = get_pcichild_wslot(hbus, devfn_to_wslot(devfn));
1323	if (!hpdev)
1324		return PCIBIOS_DEVICE_NOT_FOUND;
1325
1326	_hv_pcifront_write_config(hpdev, where, size, val);
1327
1328	put_pcichild(hpdev);
1329	return PCIBIOS_SUCCESSFUL;
1330}
1331
1332/* PCIe operations */
1333static struct pci_ops hv_pcifront_ops = {
1334	.read  = hv_pcifront_read_config,
1335	.write = hv_pcifront_write_config,
1336};
1337
1338/*
1339 * Paravirtual backchannel
1340 *
1341 * Hyper-V SR-IOV provides a backchannel mechanism in software for
1342 * communication between a VF driver and a PF driver.  These
1343 * "configuration blocks" are similar in concept to PCI configuration space,
1344 * but instead of doing reads and writes in 32-bit chunks through a very slow
1345 * path, packets of up to 128 bytes can be sent or received asynchronously.
1346 *
1347 * Nearly every SR-IOV device contains just such a communications channel in
1348 * hardware, so using this one in software is usually optional.  Using the
1349 * software channel, however, allows driver implementers to leverage software
1350 * tools that fuzz the communications channel looking for vulnerabilities.
1351 *
1352 * The usage model for these packets puts the responsibility for reading or
1353 * writing on the VF driver.  The VF driver sends a read or a write packet,
1354 * indicating which "block" is being referred to by number.
1355 *
1356 * If the PF driver wishes to initiate communication, it can "invalidate" one or
1357 * more of the first 64 blocks.  This invalidation is delivered via a callback
1358 * supplied by the VF driver by this driver.
1359 *
1360 * No protocol is implied, except that supplied by the PF and VF drivers.
1361 */
1362
1363struct hv_read_config_compl {
1364	struct hv_pci_compl comp_pkt;
1365	void *buf;
1366	unsigned int len;
1367	unsigned int bytes_returned;
1368};
1369
1370/**
1371 * hv_pci_read_config_compl() - Invoked when a response packet
1372 * for a read config block operation arrives.
1373 * @context:		Identifies the read config operation
1374 * @resp:		The response packet itself
1375 * @resp_packet_size:	Size in bytes of the response packet
1376 */
1377static void hv_pci_read_config_compl(void *context, struct pci_response *resp,
1378				     int resp_packet_size)
1379{
1380	struct hv_read_config_compl *comp = context;
1381	struct pci_read_block_response *read_resp =
1382		(struct pci_read_block_response *)resp;
1383	unsigned int data_len, hdr_len;
1384
1385	hdr_len = offsetof(struct pci_read_block_response, bytes);
1386	if (resp_packet_size < hdr_len) {
1387		comp->comp_pkt.completion_status = -1;
1388		goto out;
1389	}
1390
1391	data_len = resp_packet_size - hdr_len;
1392	if (data_len > 0 && read_resp->status == 0) {
1393		comp->bytes_returned = min(comp->len, data_len);
1394		memcpy(comp->buf, read_resp->bytes, comp->bytes_returned);
1395	} else {
1396		comp->bytes_returned = 0;
1397	}
1398
1399	comp->comp_pkt.completion_status = read_resp->status;
1400out:
1401	complete(&comp->comp_pkt.host_event);
1402}
1403
1404/**
1405 * hv_read_config_block() - Sends a read config block request to
1406 * the back-end driver running in the Hyper-V parent partition.
1407 * @pdev:		The PCI driver's representation for this device.
1408 * @buf:		Buffer into which the config block will be copied.
1409 * @len:		Size in bytes of buf.
1410 * @block_id:		Identifies the config block which has been requested.
1411 * @bytes_returned:	Size which came back from the back-end driver.
1412 *
1413 * Return: 0 on success, -errno on failure
1414 */
1415static int hv_read_config_block(struct pci_dev *pdev, void *buf,
1416				unsigned int len, unsigned int block_id,
1417				unsigned int *bytes_returned)
1418{
1419	struct hv_pcibus_device *hbus =
1420		container_of(pdev->bus->sysdata, struct hv_pcibus_device,
1421			     sysdata);
1422	struct {
1423		struct pci_packet pkt;
1424		char buf[sizeof(struct pci_read_block)];
1425	} pkt;
1426	struct hv_read_config_compl comp_pkt;
1427	struct pci_read_block *read_blk;
1428	int ret;
1429
1430	if (len == 0 || len > HV_CONFIG_BLOCK_SIZE_MAX)
1431		return -EINVAL;
1432
1433	init_completion(&comp_pkt.comp_pkt.host_event);
1434	comp_pkt.buf = buf;
1435	comp_pkt.len = len;
1436
1437	memset(&pkt, 0, sizeof(pkt));
1438	pkt.pkt.completion_func = hv_pci_read_config_compl;
1439	pkt.pkt.compl_ctxt = &comp_pkt;
1440	read_blk = (struct pci_read_block *)&pkt.pkt.message;
1441	read_blk->message_type.type = PCI_READ_BLOCK;
1442	read_blk->wslot.slot = devfn_to_wslot(pdev->devfn);
1443	read_blk->block_id = block_id;
1444	read_blk->bytes_requested = len;
1445
1446	ret = vmbus_sendpacket(hbus->hdev->channel, read_blk,
1447			       sizeof(*read_blk), (unsigned long)&pkt.pkt,
1448			       VM_PKT_DATA_INBAND,
1449			       VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
1450	if (ret)
1451		return ret;
1452
1453	ret = wait_for_response(hbus->hdev, &comp_pkt.comp_pkt.host_event);
1454	if (ret)
1455		return ret;
1456
1457	if (comp_pkt.comp_pkt.completion_status != 0 ||
1458	    comp_pkt.bytes_returned == 0) {
1459		dev_err(&hbus->hdev->device,
1460			"Read Config Block failed: 0x%x, bytes_returned=%d\n",
1461			comp_pkt.comp_pkt.completion_status,
1462			comp_pkt.bytes_returned);
1463		return -EIO;
1464	}
1465
1466	*bytes_returned = comp_pkt.bytes_returned;
1467	return 0;
1468}
1469
1470/**
1471 * hv_pci_write_config_compl() - Invoked when a response packet for a write
1472 * config block operation arrives.
1473 * @context:		Identifies the write config operation
1474 * @resp:		The response packet itself
1475 * @resp_packet_size:	Size in bytes of the response packet
1476 */
1477static void hv_pci_write_config_compl(void *context, struct pci_response *resp,
1478				      int resp_packet_size)
1479{
1480	struct hv_pci_compl *comp_pkt = context;
1481
1482	comp_pkt->completion_status = resp->status;
1483	complete(&comp_pkt->host_event);
1484}
1485
1486/**
1487 * hv_write_config_block() - Sends a write config block request to the
1488 * back-end driver running in the Hyper-V parent partition.
1489 * @pdev:		The PCI driver's representation for this device.
1490 * @buf:		Buffer from which the config block will	be copied.
1491 * @len:		Size in bytes of buf.
1492 * @block_id:		Identifies the config block which is being written.
1493 *
1494 * Return: 0 on success, -errno on failure
1495 */
1496static int hv_write_config_block(struct pci_dev *pdev, void *buf,
1497				unsigned int len, unsigned int block_id)
1498{
1499	struct hv_pcibus_device *hbus =
1500		container_of(pdev->bus->sysdata, struct hv_pcibus_device,
1501			     sysdata);
1502	struct {
1503		struct pci_packet pkt;
1504		char buf[sizeof(struct pci_write_block)];
1505		u32 reserved;
1506	} pkt;
1507	struct hv_pci_compl comp_pkt;
1508	struct pci_write_block *write_blk;
1509	u32 pkt_size;
1510	int ret;
1511
1512	if (len == 0 || len > HV_CONFIG_BLOCK_SIZE_MAX)
1513		return -EINVAL;
1514
1515	init_completion(&comp_pkt.host_event);
1516
1517	memset(&pkt, 0, sizeof(pkt));
1518	pkt.pkt.completion_func = hv_pci_write_config_compl;
1519	pkt.pkt.compl_ctxt = &comp_pkt;
1520	write_blk = (struct pci_write_block *)&pkt.pkt.message;
1521	write_blk->message_type.type = PCI_WRITE_BLOCK;
1522	write_blk->wslot.slot = devfn_to_wslot(pdev->devfn);
1523	write_blk->block_id = block_id;
1524	write_blk->byte_count = len;
1525	memcpy(write_blk->bytes, buf, len);
1526	pkt_size = offsetof(struct pci_write_block, bytes) + len;
1527	/*
1528	 * This quirk is required on some hosts shipped around 2018, because
1529	 * these hosts don't check the pkt_size correctly (new hosts have been
1530	 * fixed since early 2019). The quirk is also safe on very old hosts
1531	 * and new hosts, because, on them, what really matters is the length
1532	 * specified in write_blk->byte_count.
1533	 */
1534	pkt_size += sizeof(pkt.reserved);
1535
1536	ret = vmbus_sendpacket(hbus->hdev->channel, write_blk, pkt_size,
1537			       (unsigned long)&pkt.pkt, VM_PKT_DATA_INBAND,
1538			       VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
1539	if (ret)
1540		return ret;
1541
1542	ret = wait_for_response(hbus->hdev, &comp_pkt.host_event);
1543	if (ret)
1544		return ret;
1545
1546	if (comp_pkt.completion_status != 0) {
1547		dev_err(&hbus->hdev->device,
1548			"Write Config Block failed: 0x%x\n",
1549			comp_pkt.completion_status);
1550		return -EIO;
1551	}
1552
1553	return 0;
1554}
1555
1556/**
1557 * hv_register_block_invalidate() - Invoked when a config block invalidation
1558 * arrives from the back-end driver.
1559 * @pdev:		The PCI driver's representation for this device.
1560 * @context:		Identifies the device.
1561 * @block_invalidate:	Identifies all of the blocks being invalidated.
1562 *
1563 * Return: 0 on success, -errno on failure
1564 */
1565static int hv_register_block_invalidate(struct pci_dev *pdev, void *context,
1566					void (*block_invalidate)(void *context,
1567								 u64 block_mask))
1568{
1569	struct hv_pcibus_device *hbus =
1570		container_of(pdev->bus->sysdata, struct hv_pcibus_device,
1571			     sysdata);
1572	struct hv_pci_dev *hpdev;
1573
1574	hpdev = get_pcichild_wslot(hbus, devfn_to_wslot(pdev->devfn));
1575	if (!hpdev)
1576		return -ENODEV;
1577
1578	hpdev->block_invalidate = block_invalidate;
1579	hpdev->invalidate_context = context;
1580
1581	put_pcichild(hpdev);
1582	return 0;
1583
1584}
1585
1586/* Interrupt management hooks */
1587static void hv_int_desc_free(struct hv_pci_dev *hpdev,
1588			     struct tran_int_desc *int_desc)
1589{
1590	struct pci_delete_interrupt *int_pkt;
1591	struct {
1592		struct pci_packet pkt;
1593		u8 buffer[sizeof(struct pci_delete_interrupt)];
1594	} ctxt;
1595
1596	if (!int_desc->vector_count) {
1597		kfree(int_desc);
1598		return;
1599	}
1600	memset(&ctxt, 0, sizeof(ctxt));
1601	int_pkt = (struct pci_delete_interrupt *)&ctxt.pkt.message;
1602	int_pkt->message_type.type =
1603		PCI_DELETE_INTERRUPT_MESSAGE;
1604	int_pkt->wslot.slot = hpdev->desc.win_slot.slot;
1605	int_pkt->int_desc = *int_desc;
1606	vmbus_sendpacket(hpdev->hbus->hdev->channel, int_pkt, sizeof(*int_pkt),
1607			 0, VM_PKT_DATA_INBAND, 0);
1608	kfree(int_desc);
1609}
1610
1611/**
1612 * hv_msi_free() - Free the MSI.
1613 * @domain:	The interrupt domain pointer
1614 * @info:	Extra MSI-related context
1615 * @irq:	Identifies the IRQ.
1616 *
1617 * The Hyper-V parent partition and hypervisor are tracking the
1618 * messages that are in use, keeping the interrupt redirection
1619 * table up to date.  This callback sends a message that frees
1620 * the IRT entry and related tracking nonsense.
1621 */
1622static void hv_msi_free(struct irq_domain *domain, struct msi_domain_info *info,
1623			unsigned int irq)
1624{
1625	struct hv_pcibus_device *hbus;
1626	struct hv_pci_dev *hpdev;
1627	struct pci_dev *pdev;
1628	struct tran_int_desc *int_desc;
1629	struct irq_data *irq_data = irq_domain_get_irq_data(domain, irq);
1630	struct msi_desc *msi = irq_data_get_msi_desc(irq_data);
1631
1632	pdev = msi_desc_to_pci_dev(msi);
1633	hbus = info->data;
1634	int_desc = irq_data_get_irq_chip_data(irq_data);
1635	if (!int_desc)
1636		return;
1637
1638	irq_data->chip_data = NULL;
1639	hpdev = get_pcichild_wslot(hbus, devfn_to_wslot(pdev->devfn));
1640	if (!hpdev) {
1641		kfree(int_desc);
1642		return;
1643	}
1644
1645	hv_int_desc_free(hpdev, int_desc);
1646	put_pcichild(hpdev);
1647}
1648
1649static void hv_irq_mask(struct irq_data *data)
1650{
1651	pci_msi_mask_irq(data);
1652	if (data->parent_data->chip->irq_mask)
1653		irq_chip_mask_parent(data);
1654}
1655
1656static void hv_irq_unmask(struct irq_data *data)
1657{
1658	hv_arch_irq_unmask(data);
1659
1660	if (data->parent_data->chip->irq_unmask)
1661		irq_chip_unmask_parent(data);
1662	pci_msi_unmask_irq(data);
1663}
1664
1665struct compose_comp_ctxt {
1666	struct hv_pci_compl comp_pkt;
1667	struct tran_int_desc int_desc;
1668};
1669
1670static void hv_pci_compose_compl(void *context, struct pci_response *resp,
1671				 int resp_packet_size)
1672{
1673	struct compose_comp_ctxt *comp_pkt = context;
1674	struct pci_create_int_response *int_resp =
1675		(struct pci_create_int_response *)resp;
1676
1677	if (resp_packet_size < sizeof(*int_resp)) {
1678		comp_pkt->comp_pkt.completion_status = -1;
1679		goto out;
1680	}
1681	comp_pkt->comp_pkt.completion_status = resp->status;
1682	comp_pkt->int_desc = int_resp->int_desc;
1683out:
1684	complete(&comp_pkt->comp_pkt.host_event);
1685}
1686
1687static u32 hv_compose_msi_req_v1(
1688	struct pci_create_interrupt *int_pkt,
1689	u32 slot, u8 vector, u16 vector_count)
1690{
1691	int_pkt->message_type.type = PCI_CREATE_INTERRUPT_MESSAGE;
1692	int_pkt->wslot.slot = slot;
1693	int_pkt->int_desc.vector = vector;
1694	int_pkt->int_desc.vector_count = vector_count;
1695	int_pkt->int_desc.delivery_mode = DELIVERY_MODE;
1696
1697	/*
1698	 * Create MSI w/ dummy vCPU set, overwritten by subsequent retarget in
1699	 * hv_irq_unmask().
1700	 */
1701	int_pkt->int_desc.cpu_mask = CPU_AFFINITY_ALL;
1702
1703	return sizeof(*int_pkt);
1704}
1705
1706/*
1707 * The vCPU selected by hv_compose_multi_msi_req_get_cpu() and
1708 * hv_compose_msi_req_get_cpu() is a "dummy" vCPU because the final vCPU to be
1709 * interrupted is specified later in hv_irq_unmask() and communicated to Hyper-V
1710 * via the HVCALL_RETARGET_INTERRUPT hypercall. But the choice of dummy vCPU is
1711 * not irrelevant because Hyper-V chooses the physical CPU to handle the
1712 * interrupts based on the vCPU specified in message sent to the vPCI VSP in
1713 * hv_compose_msi_msg(). Hyper-V's choice of pCPU is not visible to the guest,
1714 * but assigning too many vPCI device interrupts to the same pCPU can cause a
1715 * performance bottleneck. So we spread out the dummy vCPUs to influence Hyper-V
1716 * to spread out the pCPUs that it selects.
1717 *
1718 * For the single-MSI and MSI-X cases, it's OK for hv_compose_msi_req_get_cpu()
1719 * to always return the same dummy vCPU, because a second call to
1720 * hv_compose_msi_msg() contains the "real" vCPU, causing Hyper-V to choose a
1721 * new pCPU for the interrupt. But for the multi-MSI case, the second call to
1722 * hv_compose_msi_msg() exits without sending a message to the vPCI VSP, so the
1723 * original dummy vCPU is used. This dummy vCPU must be round-robin'ed so that
1724 * the pCPUs are spread out. All interrupts for a multi-MSI device end up using
1725 * the same pCPU, even though the vCPUs will be spread out by later calls
1726 * to hv_irq_unmask(), but that is the best we can do now.
1727 *
1728 * With Hyper-V in Nov 2022, the HVCALL_RETARGET_INTERRUPT hypercall does *not*
1729 * cause Hyper-V to reselect the pCPU based on the specified vCPU. Such an
1730 * enhancement is planned for a future version. With that enhancement, the
1731 * dummy vCPU selection won't matter, and interrupts for the same multi-MSI
1732 * device will be spread across multiple pCPUs.
1733 */
1734
1735/*
1736 * Create MSI w/ dummy vCPU set targeting just one vCPU, overwritten
1737 * by subsequent retarget in hv_irq_unmask().
1738 */
1739static int hv_compose_msi_req_get_cpu(const struct cpumask *affinity)
1740{
1741	return cpumask_first_and(affinity, cpu_online_mask);
1742}
1743
1744/*
1745 * Make sure the dummy vCPU values for multi-MSI don't all point to vCPU0.
1746 */
1747static int hv_compose_multi_msi_req_get_cpu(void)
1748{
1749	static DEFINE_SPINLOCK(multi_msi_cpu_lock);
1750
1751	/* -1 means starting with CPU 0 */
1752	static int cpu_next = -1;
1753
1754	unsigned long flags;
1755	int cpu;
1756
1757	spin_lock_irqsave(&multi_msi_cpu_lock, flags);
1758
1759	cpu_next = cpumask_next_wrap(cpu_next, cpu_online_mask, nr_cpu_ids,
1760				     false);
1761	cpu = cpu_next;
1762
1763	spin_unlock_irqrestore(&multi_msi_cpu_lock, flags);
1764
1765	return cpu;
1766}
1767
1768static u32 hv_compose_msi_req_v2(
1769	struct pci_create_interrupt2 *int_pkt, int cpu,
1770	u32 slot, u8 vector, u16 vector_count)
1771{
1772	int_pkt->message_type.type = PCI_CREATE_INTERRUPT_MESSAGE2;
1773	int_pkt->wslot.slot = slot;
1774	int_pkt->int_desc.vector = vector;
1775	int_pkt->int_desc.vector_count = vector_count;
1776	int_pkt->int_desc.delivery_mode = DELIVERY_MODE;
1777	int_pkt->int_desc.processor_array[0] =
1778		hv_cpu_number_to_vp_number(cpu);
1779	int_pkt->int_desc.processor_count = 1;
1780
1781	return sizeof(*int_pkt);
1782}
1783
1784static u32 hv_compose_msi_req_v3(
1785	struct pci_create_interrupt3 *int_pkt, int cpu,
1786	u32 slot, u32 vector, u16 vector_count)
1787{
1788	int_pkt->message_type.type = PCI_CREATE_INTERRUPT_MESSAGE3;
1789	int_pkt->wslot.slot = slot;
1790	int_pkt->int_desc.vector = vector;
1791	int_pkt->int_desc.reserved = 0;
1792	int_pkt->int_desc.vector_count = vector_count;
1793	int_pkt->int_desc.delivery_mode = DELIVERY_MODE;
1794	int_pkt->int_desc.processor_array[0] =
1795		hv_cpu_number_to_vp_number(cpu);
1796	int_pkt->int_desc.processor_count = 1;
1797
1798	return sizeof(*int_pkt);
1799}
1800
1801/**
1802 * hv_compose_msi_msg() - Supplies a valid MSI address/data
1803 * @data:	Everything about this MSI
1804 * @msg:	Buffer that is filled in by this function
1805 *
1806 * This function unpacks the IRQ looking for target CPU set, IDT
1807 * vector and mode and sends a message to the parent partition
1808 * asking for a mapping for that tuple in this partition.  The
1809 * response supplies a data value and address to which that data
1810 * should be written to trigger that interrupt.
1811 */
1812static void hv_compose_msi_msg(struct irq_data *data, struct msi_msg *msg)
1813{
1814	struct hv_pcibus_device *hbus;
1815	struct vmbus_channel *channel;
1816	struct hv_pci_dev *hpdev;
1817	struct pci_bus *pbus;
1818	struct pci_dev *pdev;
1819	const struct cpumask *dest;
1820	struct compose_comp_ctxt comp;
1821	struct tran_int_desc *int_desc;
1822	struct msi_desc *msi_desc;
1823	/*
1824	 * vector_count should be u16: see hv_msi_desc, hv_msi_desc2
1825	 * and hv_msi_desc3. vector must be u32: see hv_msi_desc3.
1826	 */
1827	u16 vector_count;
1828	u32 vector;
1829	struct {
1830		struct pci_packet pci_pkt;
1831		union {
1832			struct pci_create_interrupt v1;
1833			struct pci_create_interrupt2 v2;
1834			struct pci_create_interrupt3 v3;
1835		} int_pkts;
1836	} __packed ctxt;
1837	bool multi_msi;
1838	u64 trans_id;
1839	u32 size;
1840	int ret;
1841	int cpu;
1842
1843	msi_desc  = irq_data_get_msi_desc(data);
1844	multi_msi = !msi_desc->pci.msi_attrib.is_msix &&
1845		    msi_desc->nvec_used > 1;
1846
1847	/* Reuse the previous allocation */
1848	if (data->chip_data && multi_msi) {
1849		int_desc = data->chip_data;
1850		msg->address_hi = int_desc->address >> 32;
1851		msg->address_lo = int_desc->address & 0xffffffff;
1852		msg->data = int_desc->data;
1853		return;
1854	}
1855
1856	pdev = msi_desc_to_pci_dev(msi_desc);
1857	dest = irq_data_get_effective_affinity_mask(data);
1858	pbus = pdev->bus;
1859	hbus = container_of(pbus->sysdata, struct hv_pcibus_device, sysdata);
1860	channel = hbus->hdev->channel;
1861	hpdev = get_pcichild_wslot(hbus, devfn_to_wslot(pdev->devfn));
1862	if (!hpdev)
1863		goto return_null_message;
1864
1865	/* Free any previous message that might have already been composed. */
1866	if (data->chip_data && !multi_msi) {
1867		int_desc = data->chip_data;
1868		data->chip_data = NULL;
1869		hv_int_desc_free(hpdev, int_desc);
1870	}
1871
1872	int_desc = kzalloc(sizeof(*int_desc), GFP_ATOMIC);
1873	if (!int_desc)
1874		goto drop_reference;
1875
1876	if (multi_msi) {
1877		/*
1878		 * If this is not the first MSI of Multi MSI, we already have
1879		 * a mapping.  Can exit early.
1880		 */
1881		if (msi_desc->irq != data->irq) {
1882			data->chip_data = int_desc;
1883			int_desc->address = msi_desc->msg.address_lo |
1884					    (u64)msi_desc->msg.address_hi << 32;
1885			int_desc->data = msi_desc->msg.data +
1886					 (data->irq - msi_desc->irq);
1887			msg->address_hi = msi_desc->msg.address_hi;
1888			msg->address_lo = msi_desc->msg.address_lo;
1889			msg->data = int_desc->data;
1890			put_pcichild(hpdev);
1891			return;
1892		}
1893		/*
1894		 * The vector we select here is a dummy value.  The correct
1895		 * value gets sent to the hypervisor in unmask().  This needs
1896		 * to be aligned with the count, and also not zero.  Multi-msi
1897		 * is powers of 2 up to 32, so 32 will always work here.
1898		 */
1899		vector = 32;
1900		vector_count = msi_desc->nvec_used;
1901		cpu = hv_compose_multi_msi_req_get_cpu();
1902	} else {
1903		vector = hv_msi_get_int_vector(data);
1904		vector_count = 1;
1905		cpu = hv_compose_msi_req_get_cpu(dest);
1906	}
1907
1908	/*
1909	 * hv_compose_msi_req_v1 and v2 are for x86 only, meaning 'vector'
1910	 * can't exceed u8. Cast 'vector' down to u8 for v1/v2 explicitly
1911	 * for better readability.
1912	 */
1913	memset(&ctxt, 0, sizeof(ctxt));
1914	init_completion(&comp.comp_pkt.host_event);
1915	ctxt.pci_pkt.completion_func = hv_pci_compose_compl;
1916	ctxt.pci_pkt.compl_ctxt = &comp;
1917
1918	switch (hbus->protocol_version) {
1919	case PCI_PROTOCOL_VERSION_1_1:
1920		size = hv_compose_msi_req_v1(&ctxt.int_pkts.v1,
1921					hpdev->desc.win_slot.slot,
1922					(u8)vector,
1923					vector_count);
1924		break;
1925
1926	case PCI_PROTOCOL_VERSION_1_2:
1927	case PCI_PROTOCOL_VERSION_1_3:
1928		size = hv_compose_msi_req_v2(&ctxt.int_pkts.v2,
1929					cpu,
1930					hpdev->desc.win_slot.slot,
1931					(u8)vector,
1932					vector_count);
1933		break;
1934
1935	case PCI_PROTOCOL_VERSION_1_4:
1936		size = hv_compose_msi_req_v3(&ctxt.int_pkts.v3,
1937					cpu,
1938					hpdev->desc.win_slot.slot,
1939					vector,
1940					vector_count);
1941		break;
1942
1943	default:
1944		/* As we only negotiate protocol versions known to this driver,
1945		 * this path should never hit. However, this is it not a hot
1946		 * path so we print a message to aid future updates.
1947		 */
1948		dev_err(&hbus->hdev->device,
1949			"Unexpected vPCI protocol, update driver.");
1950		goto free_int_desc;
1951	}
1952
1953	ret = vmbus_sendpacket_getid(hpdev->hbus->hdev->channel, &ctxt.int_pkts,
1954				     size, (unsigned long)&ctxt.pci_pkt,
1955				     &trans_id, VM_PKT_DATA_INBAND,
1956				     VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
1957	if (ret) {
1958		dev_err(&hbus->hdev->device,
1959			"Sending request for interrupt failed: 0x%x",
1960			comp.comp_pkt.completion_status);
1961		goto free_int_desc;
1962	}
1963
1964	/*
1965	 * Prevents hv_pci_onchannelcallback() from running concurrently
1966	 * in the tasklet.
1967	 */
1968	tasklet_disable_in_atomic(&channel->callback_event);
1969
1970	/*
1971	 * Since this function is called with IRQ locks held, can't
1972	 * do normal wait for completion; instead poll.
1973	 */
1974	while (!try_wait_for_completion(&comp.comp_pkt.host_event)) {
1975		unsigned long flags;
1976
1977		/* 0xFFFF means an invalid PCI VENDOR ID. */
1978		if (hv_pcifront_get_vendor_id(hpdev) == 0xFFFF) {
1979			dev_err_once(&hbus->hdev->device,
1980				     "the device has gone\n");
1981			goto enable_tasklet;
1982		}
1983
1984		/*
1985		 * Make sure that the ring buffer data structure doesn't get
1986		 * freed while we dereference the ring buffer pointer.  Test
1987		 * for the channel's onchannel_callback being NULL within a
1988		 * sched_lock critical section.  See also the inline comments
1989		 * in vmbus_reset_channel_cb().
1990		 */
1991		spin_lock_irqsave(&channel->sched_lock, flags);
1992		if (unlikely(channel->onchannel_callback == NULL)) {
1993			spin_unlock_irqrestore(&channel->sched_lock, flags);
1994			goto enable_tasklet;
1995		}
1996		hv_pci_onchannelcallback(hbus);
1997		spin_unlock_irqrestore(&channel->sched_lock, flags);
1998
1999		udelay(100);
2000	}
2001
2002	tasklet_enable(&channel->callback_event);
2003
2004	if (comp.comp_pkt.completion_status < 0) {
2005		dev_err(&hbus->hdev->device,
2006			"Request for interrupt failed: 0x%x",
2007			comp.comp_pkt.completion_status);
2008		goto free_int_desc;
2009	}
2010
2011	/*
2012	 * Record the assignment so that this can be unwound later. Using
2013	 * irq_set_chip_data() here would be appropriate, but the lock it takes
2014	 * is already held.
2015	 */
2016	*int_desc = comp.int_desc;
2017	data->chip_data = int_desc;
2018
2019	/* Pass up the result. */
2020	msg->address_hi = comp.int_desc.address >> 32;
2021	msg->address_lo = comp.int_desc.address & 0xffffffff;
2022	msg->data = comp.int_desc.data;
2023
2024	put_pcichild(hpdev);
2025	return;
2026
2027enable_tasklet:
2028	tasklet_enable(&channel->callback_event);
2029	/*
2030	 * The completion packet on the stack becomes invalid after 'return';
2031	 * remove the ID from the VMbus requestor if the identifier is still
2032	 * mapped to/associated with the packet.  (The identifier could have
2033	 * been 're-used', i.e., already removed and (re-)mapped.)
2034	 *
2035	 * Cf. hv_pci_onchannelcallback().
2036	 */
2037	vmbus_request_addr_match(channel, trans_id, (unsigned long)&ctxt.pci_pkt);
2038free_int_desc:
2039	kfree(int_desc);
2040drop_reference:
2041	put_pcichild(hpdev);
2042return_null_message:
2043	msg->address_hi = 0;
2044	msg->address_lo = 0;
2045	msg->data = 0;
2046}
2047
2048/* HW Interrupt Chip Descriptor */
2049static struct irq_chip hv_msi_irq_chip = {
2050	.name			= "Hyper-V PCIe MSI",
2051	.irq_compose_msi_msg	= hv_compose_msi_msg,
2052	.irq_set_affinity	= irq_chip_set_affinity_parent,
2053#ifdef CONFIG_X86
2054	.irq_ack		= irq_chip_ack_parent,
2055#elif defined(CONFIG_ARM64)
2056	.irq_eoi		= irq_chip_eoi_parent,
2057#endif
2058	.irq_mask		= hv_irq_mask,
2059	.irq_unmask		= hv_irq_unmask,
2060};
2061
2062static struct msi_domain_ops hv_msi_ops = {
2063	.msi_prepare	= hv_msi_prepare,
2064	.msi_free	= hv_msi_free,
2065};
2066
2067/**
2068 * hv_pcie_init_irq_domain() - Initialize IRQ domain
2069 * @hbus:	The root PCI bus
2070 *
2071 * This function creates an IRQ domain which will be used for
2072 * interrupts from devices that have been passed through.  These
2073 * devices only support MSI and MSI-X, not line-based interrupts
2074 * or simulations of line-based interrupts through PCIe's
2075 * fabric-layer messages.  Because interrupts are remapped, we
2076 * can support multi-message MSI here.
2077 *
2078 * Return: '0' on success and error value on failure
2079 */
2080static int hv_pcie_init_irq_domain(struct hv_pcibus_device *hbus)
2081{
2082	hbus->msi_info.chip = &hv_msi_irq_chip;
2083	hbus->msi_info.ops = &hv_msi_ops;
2084	hbus->msi_info.flags = (MSI_FLAG_USE_DEF_DOM_OPS |
2085		MSI_FLAG_USE_DEF_CHIP_OPS | MSI_FLAG_MULTI_PCI_MSI |
2086		MSI_FLAG_PCI_MSIX);
2087	hbus->msi_info.handler = FLOW_HANDLER;
2088	hbus->msi_info.handler_name = FLOW_NAME;
2089	hbus->msi_info.data = hbus;
2090	hbus->irq_domain = pci_msi_create_irq_domain(hbus->fwnode,
2091						     &hbus->msi_info,
2092						     hv_pci_get_root_domain());
2093	if (!hbus->irq_domain) {
2094		dev_err(&hbus->hdev->device,
2095			"Failed to build an MSI IRQ domain\n");
2096		return -ENODEV;
2097	}
2098
2099	dev_set_msi_domain(&hbus->bridge->dev, hbus->irq_domain);
2100
2101	return 0;
2102}
2103
2104/**
2105 * get_bar_size() - Get the address space consumed by a BAR
2106 * @bar_val:	Value that a BAR returned after -1 was written
2107 *              to it.
2108 *
2109 * This function returns the size of the BAR, rounded up to 1
2110 * page.  It has to be rounded up because the hypervisor's page
2111 * table entry that maps the BAR into the VM can't specify an
2112 * offset within a page.  The invariant is that the hypervisor
2113 * must place any BARs of smaller than page length at the
2114 * beginning of a page.
2115 *
2116 * Return:	Size in bytes of the consumed MMIO space.
2117 */
2118static u64 get_bar_size(u64 bar_val)
2119{
2120	return round_up((1 + ~(bar_val & PCI_BASE_ADDRESS_MEM_MASK)),
2121			PAGE_SIZE);
2122}
2123
2124/**
2125 * survey_child_resources() - Total all MMIO requirements
2126 * @hbus:	Root PCI bus, as understood by this driver
2127 */
2128static void survey_child_resources(struct hv_pcibus_device *hbus)
2129{
2130	struct hv_pci_dev *hpdev;
2131	resource_size_t bar_size = 0;
2132	unsigned long flags;
2133	struct completion *event;
2134	u64 bar_val;
2135	int i;
2136
2137	/* If nobody is waiting on the answer, don't compute it. */
2138	event = xchg(&hbus->survey_event, NULL);
2139	if (!event)
2140		return;
2141
2142	/* If the answer has already been computed, go with it. */
2143	if (hbus->low_mmio_space || hbus->high_mmio_space) {
2144		complete(event);
2145		return;
2146	}
2147
2148	spin_lock_irqsave(&hbus->device_list_lock, flags);
2149
2150	/*
2151	 * Due to an interesting quirk of the PCI spec, all memory regions
2152	 * for a child device are a power of 2 in size and aligned in memory,
2153	 * so it's sufficient to just add them up without tracking alignment.
2154	 */
2155	list_for_each_entry(hpdev, &hbus->children, list_entry) {
2156		for (i = 0; i < PCI_STD_NUM_BARS; i++) {
2157			if (hpdev->probed_bar[i] & PCI_BASE_ADDRESS_SPACE_IO)
2158				dev_err(&hbus->hdev->device,
2159					"There's an I/O BAR in this list!\n");
2160
2161			if (hpdev->probed_bar[i] != 0) {
2162				/*
2163				 * A probed BAR has all the upper bits set that
2164				 * can be changed.
2165				 */
2166
2167				bar_val = hpdev->probed_bar[i];
2168				if (bar_val & PCI_BASE_ADDRESS_MEM_TYPE_64)
2169					bar_val |=
2170					((u64)hpdev->probed_bar[++i] << 32);
2171				else
2172					bar_val |= 0xffffffff00000000ULL;
2173
2174				bar_size = get_bar_size(bar_val);
2175
2176				if (bar_val & PCI_BASE_ADDRESS_MEM_TYPE_64)
2177					hbus->high_mmio_space += bar_size;
2178				else
2179					hbus->low_mmio_space += bar_size;
2180			}
2181		}
2182	}
2183
2184	spin_unlock_irqrestore(&hbus->device_list_lock, flags);
2185	complete(event);
2186}
2187
2188/**
2189 * prepopulate_bars() - Fill in BARs with defaults
2190 * @hbus:	Root PCI bus, as understood by this driver
2191 *
2192 * The core PCI driver code seems much, much happier if the BARs
2193 * for a device have values upon first scan. So fill them in.
2194 * The algorithm below works down from large sizes to small,
2195 * attempting to pack the assignments optimally. The assumption,
2196 * enforced in other parts of the code, is that the beginning of
2197 * the memory-mapped I/O space will be aligned on the largest
2198 * BAR size.
2199 */
2200static void prepopulate_bars(struct hv_pcibus_device *hbus)
2201{
2202	resource_size_t high_size = 0;
2203	resource_size_t low_size = 0;
2204	resource_size_t high_base = 0;
2205	resource_size_t low_base = 0;
2206	resource_size_t bar_size;
2207	struct hv_pci_dev *hpdev;
2208	unsigned long flags;
2209	u64 bar_val;
2210	u32 command;
2211	bool high;
2212	int i;
2213
2214	if (hbus->low_mmio_space) {
2215		low_size = 1ULL << (63 - __builtin_clzll(hbus->low_mmio_space));
2216		low_base = hbus->low_mmio_res->start;
2217	}
2218
2219	if (hbus->high_mmio_space) {
2220		high_size = 1ULL <<
2221			(63 - __builtin_clzll(hbus->high_mmio_space));
2222		high_base = hbus->high_mmio_res->start;
2223	}
2224
2225	spin_lock_irqsave(&hbus->device_list_lock, flags);
2226
2227	/*
2228	 * Clear the memory enable bit, in case it's already set. This occurs
2229	 * in the suspend path of hibernation, where the device is suspended,
2230	 * resumed and suspended again: see hibernation_snapshot() and
2231	 * hibernation_platform_enter().
2232	 *
2233	 * If the memory enable bit is already set, Hyper-V silently ignores
2234	 * the below BAR updates, and the related PCI device driver can not
2235	 * work, because reading from the device register(s) always returns
2236	 * 0xFFFFFFFF (PCI_ERROR_RESPONSE).
2237	 */
2238	list_for_each_entry(hpdev, &hbus->children, list_entry) {
2239		_hv_pcifront_read_config(hpdev, PCI_COMMAND, 2, &command);
2240		command &= ~PCI_COMMAND_MEMORY;
2241		_hv_pcifront_write_config(hpdev, PCI_COMMAND, 2, command);
2242	}
2243
2244	/* Pick addresses for the BARs. */
2245	do {
2246		list_for_each_entry(hpdev, &hbus->children, list_entry) {
2247			for (i = 0; i < PCI_STD_NUM_BARS; i++) {
2248				bar_val = hpdev->probed_bar[i];
2249				if (bar_val == 0)
2250					continue;
2251				high = bar_val & PCI_BASE_ADDRESS_MEM_TYPE_64;
2252				if (high) {
2253					bar_val |=
2254						((u64)hpdev->probed_bar[i + 1]
2255						 << 32);
2256				} else {
2257					bar_val |= 0xffffffffULL << 32;
2258				}
2259				bar_size = get_bar_size(bar_val);
2260				if (high) {
2261					if (high_size != bar_size) {
2262						i++;
2263						continue;
2264					}
2265					_hv_pcifront_write_config(hpdev,
2266						PCI_BASE_ADDRESS_0 + (4 * i),
2267						4,
2268						(u32)(high_base & 0xffffff00));
2269					i++;
2270					_hv_pcifront_write_config(hpdev,
2271						PCI_BASE_ADDRESS_0 + (4 * i),
2272						4, (u32)(high_base >> 32));
2273					high_base += bar_size;
2274				} else {
2275					if (low_size != bar_size)
2276						continue;
2277					_hv_pcifront_write_config(hpdev,
2278						PCI_BASE_ADDRESS_0 + (4 * i),
2279						4,
2280						(u32)(low_base & 0xffffff00));
2281					low_base += bar_size;
2282				}
2283			}
2284			if (high_size <= 1 && low_size <= 1) {
2285				/*
2286				 * No need to set the PCI_COMMAND_MEMORY bit as
2287				 * the core PCI driver doesn't require the bit
2288				 * to be pre-set. Actually here we intentionally
2289				 * keep the bit off so that the PCI BAR probing
2290				 * in the core PCI driver doesn't cause Hyper-V
2291				 * to unnecessarily unmap/map the virtual BARs
2292				 * from/to the physical BARs multiple times.
2293				 * This reduces the VM boot time significantly
2294				 * if the BAR sizes are huge.
2295				 */
2296				break;
2297			}
2298		}
2299
2300		high_size >>= 1;
2301		low_size >>= 1;
2302	}  while (high_size || low_size);
2303
2304	spin_unlock_irqrestore(&hbus->device_list_lock, flags);
2305}
2306
2307/*
2308 * Assign entries in sysfs pci slot directory.
2309 *
2310 * Note that this function does not need to lock the children list
2311 * because it is called from pci_devices_present_work which
2312 * is serialized with hv_eject_device_work because they are on the
2313 * same ordered workqueue. Therefore hbus->children list will not change
2314 * even when pci_create_slot sleeps.
2315 */
2316static void hv_pci_assign_slots(struct hv_pcibus_device *hbus)
2317{
2318	struct hv_pci_dev *hpdev;
2319	char name[SLOT_NAME_SIZE];
2320	int slot_nr;
2321
2322	list_for_each_entry(hpdev, &hbus->children, list_entry) {
2323		if (hpdev->pci_slot)
2324			continue;
2325
2326		slot_nr = PCI_SLOT(wslot_to_devfn(hpdev->desc.win_slot.slot));
2327		snprintf(name, SLOT_NAME_SIZE, "%u", hpdev->desc.ser);
2328		hpdev->pci_slot = pci_create_slot(hbus->bridge->bus, slot_nr,
2329					  name, NULL);
2330		if (IS_ERR(hpdev->pci_slot)) {
2331			pr_warn("pci_create slot %s failed\n", name);
2332			hpdev->pci_slot = NULL;
2333		}
2334	}
2335}
2336
2337/*
2338 * Remove entries in sysfs pci slot directory.
2339 */
2340static void hv_pci_remove_slots(struct hv_pcibus_device *hbus)
2341{
2342	struct hv_pci_dev *hpdev;
2343
2344	list_for_each_entry(hpdev, &hbus->children, list_entry) {
2345		if (!hpdev->pci_slot)
2346			continue;
2347		pci_destroy_slot(hpdev->pci_slot);
2348		hpdev->pci_slot = NULL;
2349	}
2350}
2351
2352/*
2353 * Set NUMA node for the devices on the bus
2354 */
2355static void hv_pci_assign_numa_node(struct hv_pcibus_device *hbus)
2356{
2357	struct pci_dev *dev;
2358	struct pci_bus *bus = hbus->bridge->bus;
2359	struct hv_pci_dev *hv_dev;
2360
2361	list_for_each_entry(dev, &bus->devices, bus_list) {
2362		hv_dev = get_pcichild_wslot(hbus, devfn_to_wslot(dev->devfn));
2363		if (!hv_dev)
2364			continue;
2365
2366		if (hv_dev->desc.flags & HV_PCI_DEVICE_FLAG_NUMA_AFFINITY &&
2367		    hv_dev->desc.virtual_numa_node < num_possible_nodes())
2368			/*
2369			 * The kernel may boot with some NUMA nodes offline
2370			 * (e.g. in a KDUMP kernel) or with NUMA disabled via
2371			 * "numa=off". In those cases, adjust the host provided
2372			 * NUMA node to a valid NUMA node used by the kernel.
2373			 */
2374			set_dev_node(&dev->dev,
2375				     numa_map_to_online_node(
2376					     hv_dev->desc.virtual_numa_node));
2377
2378		put_pcichild(hv_dev);
2379	}
2380}
2381
2382/**
2383 * create_root_hv_pci_bus() - Expose a new root PCI bus
2384 * @hbus:	Root PCI bus, as understood by this driver
2385 *
2386 * Return: 0 on success, -errno on failure
2387 */
2388static int create_root_hv_pci_bus(struct hv_pcibus_device *hbus)
2389{
2390	int error;
2391	struct pci_host_bridge *bridge = hbus->bridge;
2392
2393	bridge->dev.parent = &hbus->hdev->device;
2394	bridge->sysdata = &hbus->sysdata;
2395	bridge->ops = &hv_pcifront_ops;
2396
2397	error = pci_scan_root_bus_bridge(bridge);
2398	if (error)
2399		return error;
2400
2401	pci_lock_rescan_remove();
2402	hv_pci_assign_numa_node(hbus);
2403	pci_bus_assign_resources(bridge->bus);
2404	hv_pci_assign_slots(hbus);
2405	pci_bus_add_devices(bridge->bus);
2406	pci_unlock_rescan_remove();
2407	hbus->state = hv_pcibus_installed;
2408	return 0;
2409}
2410
2411struct q_res_req_compl {
2412	struct completion host_event;
2413	struct hv_pci_dev *hpdev;
2414};
2415
2416/**
2417 * q_resource_requirements() - Query Resource Requirements
2418 * @context:		The completion context.
2419 * @resp:		The response that came from the host.
2420 * @resp_packet_size:	The size in bytes of resp.
2421 *
2422 * This function is invoked on completion of a Query Resource
2423 * Requirements packet.
2424 */
2425static void q_resource_requirements(void *context, struct pci_response *resp,
2426				    int resp_packet_size)
2427{
2428	struct q_res_req_compl *completion = context;
2429	struct pci_q_res_req_response *q_res_req =
2430		(struct pci_q_res_req_response *)resp;
2431	s32 status;
2432	int i;
2433
2434	status = (resp_packet_size < sizeof(*q_res_req)) ? -1 : resp->status;
2435	if (status < 0) {
2436		dev_err(&completion->hpdev->hbus->hdev->device,
2437			"query resource requirements failed: %x\n",
2438			status);
2439	} else {
2440		for (i = 0; i < PCI_STD_NUM_BARS; i++) {
2441			completion->hpdev->probed_bar[i] =
2442				q_res_req->probed_bar[i];
2443		}
2444	}
2445
2446	complete(&completion->host_event);
2447}
2448
2449/**
2450 * new_pcichild_device() - Create a new child device
2451 * @hbus:	The internal struct tracking this root PCI bus.
2452 * @desc:	The information supplied so far from the host
2453 *              about the device.
2454 *
2455 * This function creates the tracking structure for a new child
2456 * device and kicks off the process of figuring out what it is.
2457 *
2458 * Return: Pointer to the new tracking struct
2459 */
2460static struct hv_pci_dev *new_pcichild_device(struct hv_pcibus_device *hbus,
2461		struct hv_pcidev_description *desc)
2462{
2463	struct hv_pci_dev *hpdev;
2464	struct pci_child_message *res_req;
2465	struct q_res_req_compl comp_pkt;
2466	struct {
2467		struct pci_packet init_packet;
2468		u8 buffer[sizeof(struct pci_child_message)];
2469	} pkt;
2470	unsigned long flags;
2471	int ret;
2472
2473	hpdev = kzalloc(sizeof(*hpdev), GFP_KERNEL);
2474	if (!hpdev)
2475		return NULL;
2476
2477	hpdev->hbus = hbus;
2478
2479	memset(&pkt, 0, sizeof(pkt));
2480	init_completion(&comp_pkt.host_event);
2481	comp_pkt.hpdev = hpdev;
2482	pkt.init_packet.compl_ctxt = &comp_pkt;
2483	pkt.init_packet.completion_func = q_resource_requirements;
2484	res_req = (struct pci_child_message *)&pkt.init_packet.message;
2485	res_req->message_type.type = PCI_QUERY_RESOURCE_REQUIREMENTS;
2486	res_req->wslot.slot = desc->win_slot.slot;
2487
2488	ret = vmbus_sendpacket(hbus->hdev->channel, res_req,
2489			       sizeof(struct pci_child_message),
2490			       (unsigned long)&pkt.init_packet,
2491			       VM_PKT_DATA_INBAND,
2492			       VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
2493	if (ret)
2494		goto error;
2495
2496	if (wait_for_response(hbus->hdev, &comp_pkt.host_event))
2497		goto error;
2498
2499	hpdev->desc = *desc;
2500	refcount_set(&hpdev->refs, 1);
2501	get_pcichild(hpdev);
2502	spin_lock_irqsave(&hbus->device_list_lock, flags);
2503
2504	list_add_tail(&hpdev->list_entry, &hbus->children);
2505	spin_unlock_irqrestore(&hbus->device_list_lock, flags);
2506	return hpdev;
2507
2508error:
2509	kfree(hpdev);
2510	return NULL;
2511}
2512
2513/**
2514 * get_pcichild_wslot() - Find device from slot
2515 * @hbus:	Root PCI bus, as understood by this driver
2516 * @wslot:	Location on the bus
2517 *
2518 * This function looks up a PCI device and returns the internal
2519 * representation of it.  It acquires a reference on it, so that
2520 * the device won't be deleted while somebody is using it.  The
2521 * caller is responsible for calling put_pcichild() to release
2522 * this reference.
2523 *
2524 * Return:	Internal representation of a PCI device
2525 */
2526static struct hv_pci_dev *get_pcichild_wslot(struct hv_pcibus_device *hbus,
2527					     u32 wslot)
2528{
2529	unsigned long flags;
2530	struct hv_pci_dev *iter, *hpdev = NULL;
2531
2532	spin_lock_irqsave(&hbus->device_list_lock, flags);
2533	list_for_each_entry(iter, &hbus->children, list_entry) {
2534		if (iter->desc.win_slot.slot == wslot) {
2535			hpdev = iter;
2536			get_pcichild(hpdev);
2537			break;
2538		}
2539	}
2540	spin_unlock_irqrestore(&hbus->device_list_lock, flags);
2541
2542	return hpdev;
2543}
2544
2545/**
2546 * pci_devices_present_work() - Handle new list of child devices
2547 * @work:	Work struct embedded in struct hv_dr_work
2548 *
2549 * "Bus Relations" is the Windows term for "children of this
2550 * bus."  The terminology is preserved here for people trying to
2551 * debug the interaction between Hyper-V and Linux.  This
2552 * function is called when the parent partition reports a list
2553 * of functions that should be observed under this PCI Express
2554 * port (bus).
2555 *
2556 * This function updates the list, and must tolerate being
2557 * called multiple times with the same information.  The typical
2558 * number of child devices is one, with very atypical cases
2559 * involving three or four, so the algorithms used here can be
2560 * simple and inefficient.
2561 *
2562 * It must also treat the omission of a previously observed device as
2563 * notification that the device no longer exists.
2564 *
2565 * Note that this function is serialized with hv_eject_device_work(),
2566 * because both are pushed to the ordered workqueue hbus->wq.
2567 */
2568static void pci_devices_present_work(struct work_struct *work)
2569{
2570	u32 child_no;
2571	bool found;
2572	struct hv_pcidev_description *new_desc;
2573	struct hv_pci_dev *hpdev;
2574	struct hv_pcibus_device *hbus;
2575	struct list_head removed;
2576	struct hv_dr_work *dr_wrk;
2577	struct hv_dr_state *dr = NULL;
2578	unsigned long flags;
2579
2580	dr_wrk = container_of(work, struct hv_dr_work, wrk);
2581	hbus = dr_wrk->bus;
2582	kfree(dr_wrk);
2583
2584	INIT_LIST_HEAD(&removed);
2585
2586	/* Pull this off the queue and process it if it was the last one. */
2587	spin_lock_irqsave(&hbus->device_list_lock, flags);
2588	while (!list_empty(&hbus->dr_list)) {
2589		dr = list_first_entry(&hbus->dr_list, struct hv_dr_state,
2590				      list_entry);
2591		list_del(&dr->list_entry);
2592
2593		/* Throw this away if the list still has stuff in it. */
2594		if (!list_empty(&hbus->dr_list)) {
2595			kfree(dr);
2596			continue;
2597		}
2598	}
2599	spin_unlock_irqrestore(&hbus->device_list_lock, flags);
2600
2601	if (!dr)
2602		return;
2603
2604	mutex_lock(&hbus->state_lock);
2605
2606	/* First, mark all existing children as reported missing. */
2607	spin_lock_irqsave(&hbus->device_list_lock, flags);
2608	list_for_each_entry(hpdev, &hbus->children, list_entry) {
2609		hpdev->reported_missing = true;
2610	}
2611	spin_unlock_irqrestore(&hbus->device_list_lock, flags);
2612
2613	/* Next, add back any reported devices. */
2614	for (child_no = 0; child_no < dr->device_count; child_no++) {
2615		found = false;
2616		new_desc = &dr->func[child_no];
2617
2618		spin_lock_irqsave(&hbus->device_list_lock, flags);
2619		list_for_each_entry(hpdev, &hbus->children, list_entry) {
2620			if ((hpdev->desc.win_slot.slot == new_desc->win_slot.slot) &&
2621			    (hpdev->desc.v_id == new_desc->v_id) &&
2622			    (hpdev->desc.d_id == new_desc->d_id) &&
2623			    (hpdev->desc.ser == new_desc->ser)) {
2624				hpdev->reported_missing = false;
2625				found = true;
2626			}
2627		}
2628		spin_unlock_irqrestore(&hbus->device_list_lock, flags);
2629
2630		if (!found) {
2631			hpdev = new_pcichild_device(hbus, new_desc);
2632			if (!hpdev)
2633				dev_err(&hbus->hdev->device,
2634					"couldn't record a child device.\n");
2635		}
2636	}
2637
2638	/* Move missing children to a list on the stack. */
2639	spin_lock_irqsave(&hbus->device_list_lock, flags);
2640	do {
2641		found = false;
2642		list_for_each_entry(hpdev, &hbus->children, list_entry) {
2643			if (hpdev->reported_missing) {
2644				found = true;
2645				put_pcichild(hpdev);
2646				list_move_tail(&hpdev->list_entry, &removed);
2647				break;
2648			}
2649		}
2650	} while (found);
2651	spin_unlock_irqrestore(&hbus->device_list_lock, flags);
2652
2653	/* Delete everything that should no longer exist. */
2654	while (!list_empty(&removed)) {
2655		hpdev = list_first_entry(&removed, struct hv_pci_dev,
2656					 list_entry);
2657		list_del(&hpdev->list_entry);
2658
2659		if (hpdev->pci_slot)
2660			pci_destroy_slot(hpdev->pci_slot);
2661
2662		put_pcichild(hpdev);
2663	}
2664
2665	switch (hbus->state) {
2666	case hv_pcibus_installed:
2667		/*
2668		 * Tell the core to rescan bus
2669		 * because there may have been changes.
2670		 */
2671		pci_lock_rescan_remove();
2672		pci_scan_child_bus(hbus->bridge->bus);
2673		hv_pci_assign_numa_node(hbus);
2674		hv_pci_assign_slots(hbus);
2675		pci_unlock_rescan_remove();
2676		break;
2677
2678	case hv_pcibus_init:
2679	case hv_pcibus_probed:
2680		survey_child_resources(hbus);
2681		break;
2682
2683	default:
2684		break;
2685	}
2686
2687	mutex_unlock(&hbus->state_lock);
2688
2689	kfree(dr);
2690}
2691
2692/**
2693 * hv_pci_start_relations_work() - Queue work to start device discovery
2694 * @hbus:	Root PCI bus, as understood by this driver
2695 * @dr:		The list of children returned from host
2696 *
2697 * Return:  0 on success, -errno on failure
2698 */
2699static int hv_pci_start_relations_work(struct hv_pcibus_device *hbus,
2700				       struct hv_dr_state *dr)
2701{
2702	struct hv_dr_work *dr_wrk;
2703	unsigned long flags;
2704	bool pending_dr;
2705
2706	if (hbus->state == hv_pcibus_removing) {
2707		dev_info(&hbus->hdev->device,
2708			 "PCI VMBus BUS_RELATIONS: ignored\n");
2709		return -ENOENT;
2710	}
2711
2712	dr_wrk = kzalloc(sizeof(*dr_wrk), GFP_NOWAIT);
2713	if (!dr_wrk)
2714		return -ENOMEM;
2715
2716	INIT_WORK(&dr_wrk->wrk, pci_devices_present_work);
2717	dr_wrk->bus = hbus;
2718
2719	spin_lock_irqsave(&hbus->device_list_lock, flags);
2720	/*
2721	 * If pending_dr is true, we have already queued a work,
2722	 * which will see the new dr. Otherwise, we need to
2723	 * queue a new work.
2724	 */
2725	pending_dr = !list_empty(&hbus->dr_list);
2726	list_add_tail(&dr->list_entry, &hbus->dr_list);
2727	spin_unlock_irqrestore(&hbus->device_list_lock, flags);
2728
2729	if (pending_dr)
2730		kfree(dr_wrk);
2731	else
2732		queue_work(hbus->wq, &dr_wrk->wrk);
2733
2734	return 0;
2735}
2736
2737/**
2738 * hv_pci_devices_present() - Handle list of new children
2739 * @hbus:      Root PCI bus, as understood by this driver
2740 * @relations: Packet from host listing children
2741 *
2742 * Process a new list of devices on the bus. The list of devices is
2743 * discovered by VSP and sent to us via VSP message PCI_BUS_RELATIONS,
2744 * whenever a new list of devices for this bus appears.
2745 */
2746static void hv_pci_devices_present(struct hv_pcibus_device *hbus,
2747				   struct pci_bus_relations *relations)
2748{
2749	struct hv_dr_state *dr;
2750	int i;
2751
2752	dr = kzalloc(struct_size(dr, func, relations->device_count),
2753		     GFP_NOWAIT);
2754	if (!dr)
2755		return;
2756
2757	dr->device_count = relations->device_count;
2758	for (i = 0; i < dr->device_count; i++) {
2759		dr->func[i].v_id = relations->func[i].v_id;
2760		dr->func[i].d_id = relations->func[i].d_id;
2761		dr->func[i].rev = relations->func[i].rev;
2762		dr->func[i].prog_intf = relations->func[i].prog_intf;
2763		dr->func[i].subclass = relations->func[i].subclass;
2764		dr->func[i].base_class = relations->func[i].base_class;
2765		dr->func[i].subsystem_id = relations->func[i].subsystem_id;
2766		dr->func[i].win_slot = relations->func[i].win_slot;
2767		dr->func[i].ser = relations->func[i].ser;
2768	}
2769
2770	if (hv_pci_start_relations_work(hbus, dr))
2771		kfree(dr);
2772}
2773
2774/**
2775 * hv_pci_devices_present2() - Handle list of new children
2776 * @hbus:	Root PCI bus, as understood by this driver
2777 * @relations:	Packet from host listing children
2778 *
2779 * This function is the v2 version of hv_pci_devices_present()
2780 */
2781static void hv_pci_devices_present2(struct hv_pcibus_device *hbus,
2782				    struct pci_bus_relations2 *relations)
2783{
2784	struct hv_dr_state *dr;
2785	int i;
2786
2787	dr = kzalloc(struct_size(dr, func, relations->device_count),
2788		     GFP_NOWAIT);
2789	if (!dr)
2790		return;
2791
2792	dr->device_count = relations->device_count;
2793	for (i = 0; i < dr->device_count; i++) {
2794		dr->func[i].v_id = relations->func[i].v_id;
2795		dr->func[i].d_id = relations->func[i].d_id;
2796		dr->func[i].rev = relations->func[i].rev;
2797		dr->func[i].prog_intf = relations->func[i].prog_intf;
2798		dr->func[i].subclass = relations->func[i].subclass;
2799		dr->func[i].base_class = relations->func[i].base_class;
2800		dr->func[i].subsystem_id = relations->func[i].subsystem_id;
2801		dr->func[i].win_slot = relations->func[i].win_slot;
2802		dr->func[i].ser = relations->func[i].ser;
2803		dr->func[i].flags = relations->func[i].flags;
2804		dr->func[i].virtual_numa_node =
2805			relations->func[i].virtual_numa_node;
2806	}
2807
2808	if (hv_pci_start_relations_work(hbus, dr))
2809		kfree(dr);
2810}
2811
2812/**
2813 * hv_eject_device_work() - Asynchronously handles ejection
2814 * @work:	Work struct embedded in internal device struct
2815 *
2816 * This function handles ejecting a device.  Windows will
2817 * attempt to gracefully eject a device, waiting 60 seconds to
2818 * hear back from the guest OS that this completed successfully.
2819 * If this timer expires, the device will be forcibly removed.
2820 */
2821static void hv_eject_device_work(struct work_struct *work)
2822{
2823	struct pci_eject_response *ejct_pkt;
2824	struct hv_pcibus_device *hbus;
2825	struct hv_pci_dev *hpdev;
2826	struct pci_dev *pdev;
2827	unsigned long flags;
2828	int wslot;
2829	struct {
2830		struct pci_packet pkt;
2831		u8 buffer[sizeof(struct pci_eject_response)];
2832	} ctxt;
2833
2834	hpdev = container_of(work, struct hv_pci_dev, wrk);
2835	hbus = hpdev->hbus;
2836
2837	mutex_lock(&hbus->state_lock);
2838
2839	/*
2840	 * Ejection can come before or after the PCI bus has been set up, so
2841	 * attempt to find it and tear down the bus state, if it exists.  This
2842	 * must be done without constructs like pci_domain_nr(hbus->bridge->bus)
2843	 * because hbus->bridge->bus may not exist yet.
2844	 */
2845	wslot = wslot_to_devfn(hpdev->desc.win_slot.slot);
2846	pdev = pci_get_domain_bus_and_slot(hbus->bridge->domain_nr, 0, wslot);
2847	if (pdev) {
2848		pci_lock_rescan_remove();
2849		pci_stop_and_remove_bus_device(pdev);
2850		pci_dev_put(pdev);
2851		pci_unlock_rescan_remove();
2852	}
2853
2854	spin_lock_irqsave(&hbus->device_list_lock, flags);
2855	list_del(&hpdev->list_entry);
2856	spin_unlock_irqrestore(&hbus->device_list_lock, flags);
2857
2858	if (hpdev->pci_slot)
2859		pci_destroy_slot(hpdev->pci_slot);
2860
2861	memset(&ctxt, 0, sizeof(ctxt));
2862	ejct_pkt = (struct pci_eject_response *)&ctxt.pkt.message;
2863	ejct_pkt->message_type.type = PCI_EJECTION_COMPLETE;
2864	ejct_pkt->wslot.slot = hpdev->desc.win_slot.slot;
2865	vmbus_sendpacket(hbus->hdev->channel, ejct_pkt,
2866			 sizeof(*ejct_pkt), 0,
2867			 VM_PKT_DATA_INBAND, 0);
2868
2869	/* For the get_pcichild() in hv_pci_eject_device() */
2870	put_pcichild(hpdev);
2871	/* For the two refs got in new_pcichild_device() */
2872	put_pcichild(hpdev);
2873	put_pcichild(hpdev);
2874	/* hpdev has been freed. Do not use it any more. */
2875
2876	mutex_unlock(&hbus->state_lock);
2877}
2878
2879/**
2880 * hv_pci_eject_device() - Handles device ejection
2881 * @hpdev:	Internal device tracking struct
2882 *
2883 * This function is invoked when an ejection packet arrives.  It
2884 * just schedules work so that we don't re-enter the packet
2885 * delivery code handling the ejection.
2886 */
2887static void hv_pci_eject_device(struct hv_pci_dev *hpdev)
2888{
2889	struct hv_pcibus_device *hbus = hpdev->hbus;
2890	struct hv_device *hdev = hbus->hdev;
2891
2892	if (hbus->state == hv_pcibus_removing) {
2893		dev_info(&hdev->device, "PCI VMBus EJECT: ignored\n");
2894		return;
2895	}
2896
2897	get_pcichild(hpdev);
2898	INIT_WORK(&hpdev->wrk, hv_eject_device_work);
2899	queue_work(hbus->wq, &hpdev->wrk);
2900}
2901
2902/**
2903 * hv_pci_onchannelcallback() - Handles incoming packets
2904 * @context:	Internal bus tracking struct
2905 *
2906 * This function is invoked whenever the host sends a packet to
2907 * this channel (which is private to this root PCI bus).
2908 */
2909static void hv_pci_onchannelcallback(void *context)
2910{
2911	const int packet_size = 0x100;
2912	int ret;
2913	struct hv_pcibus_device *hbus = context;
2914	struct vmbus_channel *chan = hbus->hdev->channel;
2915	u32 bytes_recvd;
2916	u64 req_id, req_addr;
2917	struct vmpacket_descriptor *desc;
2918	unsigned char *buffer;
2919	int bufferlen = packet_size;
2920	struct pci_packet *comp_packet;
2921	struct pci_response *response;
2922	struct pci_incoming_message *new_message;
2923	struct pci_bus_relations *bus_rel;
2924	struct pci_bus_relations2 *bus_rel2;
2925	struct pci_dev_inval_block *inval;
2926	struct pci_dev_incoming *dev_message;
2927	struct hv_pci_dev *hpdev;
2928	unsigned long flags;
2929
2930	buffer = kmalloc(bufferlen, GFP_ATOMIC);
2931	if (!buffer)
2932		return;
2933
2934	while (1) {
2935		ret = vmbus_recvpacket_raw(chan, buffer, bufferlen,
2936					   &bytes_recvd, &req_id);
2937
2938		if (ret == -ENOBUFS) {
2939			kfree(buffer);
2940			/* Handle large packet */
2941			bufferlen = bytes_recvd;
2942			buffer = kmalloc(bytes_recvd, GFP_ATOMIC);
2943			if (!buffer)
2944				return;
2945			continue;
2946		}
2947
2948		/* Zero length indicates there are no more packets. */
2949		if (ret || !bytes_recvd)
2950			break;
2951
2952		/*
2953		 * All incoming packets must be at least as large as a
2954		 * response.
2955		 */
2956		if (bytes_recvd <= sizeof(struct pci_response))
2957			continue;
2958		desc = (struct vmpacket_descriptor *)buffer;
2959
2960		switch (desc->type) {
2961		case VM_PKT_COMP:
2962
2963			lock_requestor(chan, flags);
2964			req_addr = __vmbus_request_addr_match(chan, req_id,
2965							      VMBUS_RQST_ADDR_ANY);
2966			if (req_addr == VMBUS_RQST_ERROR) {
2967				unlock_requestor(chan, flags);
2968				dev_err(&hbus->hdev->device,
2969					"Invalid transaction ID %llx\n",
2970					req_id);
2971				break;
2972			}
2973			comp_packet = (struct pci_packet *)req_addr;
2974			response = (struct pci_response *)buffer;
2975			/*
2976			 * Call ->completion_func() within the critical section to make
2977			 * sure that the packet pointer is still valid during the call:
2978			 * here 'valid' means that there's a task still waiting for the
2979			 * completion, and that the packet data is still on the waiting
2980			 * task's stack.  Cf. hv_compose_msi_msg().
2981			 */
2982			comp_packet->completion_func(comp_packet->compl_ctxt,
2983						     response,
2984						     bytes_recvd);
2985			unlock_requestor(chan, flags);
2986			break;
2987
2988		case VM_PKT_DATA_INBAND:
2989
2990			new_message = (struct pci_incoming_message *)buffer;
2991			switch (new_message->message_type.type) {
2992			case PCI_BUS_RELATIONS:
2993
2994				bus_rel = (struct pci_bus_relations *)buffer;
2995				if (bytes_recvd < sizeof(*bus_rel) ||
2996				    bytes_recvd <
2997					struct_size(bus_rel, func,
2998						    bus_rel->device_count)) {
2999					dev_err(&hbus->hdev->device,
3000						"bus relations too small\n");
3001					break;
3002				}
3003
3004				hv_pci_devices_present(hbus, bus_rel);
3005				break;
3006
3007			case PCI_BUS_RELATIONS2:
3008
3009				bus_rel2 = (struct pci_bus_relations2 *)buffer;
3010				if (bytes_recvd < sizeof(*bus_rel2) ||
3011				    bytes_recvd <
3012					struct_size(bus_rel2, func,
3013						    bus_rel2->device_count)) {
3014					dev_err(&hbus->hdev->device,
3015						"bus relations v2 too small\n");
3016					break;
3017				}
3018
3019				hv_pci_devices_present2(hbus, bus_rel2);
3020				break;
3021
3022			case PCI_EJECT:
3023
3024				dev_message = (struct pci_dev_incoming *)buffer;
3025				if (bytes_recvd < sizeof(*dev_message)) {
3026					dev_err(&hbus->hdev->device,
3027						"eject message too small\n");
3028					break;
3029				}
3030				hpdev = get_pcichild_wslot(hbus,
3031						      dev_message->wslot.slot);
3032				if (hpdev) {
3033					hv_pci_eject_device(hpdev);
3034					put_pcichild(hpdev);
3035				}
3036				break;
3037
3038			case PCI_INVALIDATE_BLOCK:
3039
3040				inval = (struct pci_dev_inval_block *)buffer;
3041				if (bytes_recvd < sizeof(*inval)) {
3042					dev_err(&hbus->hdev->device,
3043						"invalidate message too small\n");
3044					break;
3045				}
3046				hpdev = get_pcichild_wslot(hbus,
3047							   inval->wslot.slot);
3048				if (hpdev) {
3049					if (hpdev->block_invalidate) {
3050						hpdev->block_invalidate(
3051						    hpdev->invalidate_context,
3052						    inval->block_mask);
3053					}
3054					put_pcichild(hpdev);
3055				}
3056				break;
3057
3058			default:
3059				dev_warn(&hbus->hdev->device,
3060					"Unimplemented protocol message %x\n",
3061					new_message->message_type.type);
3062				break;
3063			}
3064			break;
3065
3066		default:
3067			dev_err(&hbus->hdev->device,
3068				"unhandled packet type %d, tid %llx len %d\n",
3069				desc->type, req_id, bytes_recvd);
3070			break;
3071		}
3072	}
3073
3074	kfree(buffer);
3075}
3076
3077/**
3078 * hv_pci_protocol_negotiation() - Set up protocol
3079 * @hdev:		VMBus's tracking struct for this root PCI bus.
3080 * @version:		Array of supported channel protocol versions in
3081 *			the order of probing - highest go first.
3082 * @num_version:	Number of elements in the version array.
3083 *
3084 * This driver is intended to support running on Windows 10
3085 * (server) and later versions. It will not run on earlier
3086 * versions, as they assume that many of the operations which
3087 * Linux needs accomplished with a spinlock held were done via
3088 * asynchronous messaging via VMBus.  Windows 10 increases the
3089 * surface area of PCI emulation so that these actions can take
3090 * place by suspending a virtual processor for their duration.
3091 *
3092 * This function negotiates the channel protocol version,
3093 * failing if the host doesn't support the necessary protocol
3094 * level.
3095 */
3096static int hv_pci_protocol_negotiation(struct hv_device *hdev,
3097				       enum pci_protocol_version_t version[],
3098				       int num_version)
3099{
3100	struct hv_pcibus_device *hbus = hv_get_drvdata(hdev);
3101	struct pci_version_request *version_req;
3102	struct hv_pci_compl comp_pkt;
3103	struct pci_packet *pkt;
3104	int ret;
3105	int i;
3106
3107	/*
3108	 * Initiate the handshake with the host and negotiate
3109	 * a version that the host can support. We start with the
3110	 * highest version number and go down if the host cannot
3111	 * support it.
3112	 */
3113	pkt = kzalloc(sizeof(*pkt) + sizeof(*version_req), GFP_KERNEL);
3114	if (!pkt)
3115		return -ENOMEM;
3116
3117	init_completion(&comp_pkt.host_event);
3118	pkt->completion_func = hv_pci_generic_compl;
3119	pkt->compl_ctxt = &comp_pkt;
3120	version_req = (struct pci_version_request *)&pkt->message;
3121	version_req->message_type.type = PCI_QUERY_PROTOCOL_VERSION;
3122
3123	for (i = 0; i < num_version; i++) {
3124		version_req->protocol_version = version[i];
3125		ret = vmbus_sendpacket(hdev->channel, version_req,
3126				sizeof(struct pci_version_request),
3127				(unsigned long)pkt, VM_PKT_DATA_INBAND,
3128				VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
3129		if (!ret)
3130			ret = wait_for_response(hdev, &comp_pkt.host_event);
3131
3132		if (ret) {
3133			dev_err(&hdev->device,
3134				"PCI Pass-through VSP failed to request version: %d",
3135				ret);
3136			goto exit;
3137		}
3138
3139		if (comp_pkt.completion_status >= 0) {
3140			hbus->protocol_version = version[i];
3141			dev_info(&hdev->device,
3142				"PCI VMBus probing: Using version %#x\n",
3143				hbus->protocol_version);
3144			goto exit;
3145		}
3146
3147		if (comp_pkt.completion_status != STATUS_REVISION_MISMATCH) {
3148			dev_err(&hdev->device,
3149				"PCI Pass-through VSP failed version request: %#x",
3150				comp_pkt.completion_status);
3151			ret = -EPROTO;
3152			goto exit;
3153		}
3154
3155		reinit_completion(&comp_pkt.host_event);
3156	}
3157
3158	dev_err(&hdev->device,
3159		"PCI pass-through VSP failed to find supported version");
3160	ret = -EPROTO;
3161
3162exit:
3163	kfree(pkt);
3164	return ret;
3165}
3166
3167/**
3168 * hv_pci_free_bridge_windows() - Release memory regions for the
3169 * bus
3170 * @hbus:	Root PCI bus, as understood by this driver
3171 */
3172static void hv_pci_free_bridge_windows(struct hv_pcibus_device *hbus)
3173{
3174	/*
3175	 * Set the resources back to the way they looked when they
3176	 * were allocated by setting IORESOURCE_BUSY again.
3177	 */
3178
3179	if (hbus->low_mmio_space && hbus->low_mmio_res) {
3180		hbus->low_mmio_res->flags |= IORESOURCE_BUSY;
3181		vmbus_free_mmio(hbus->low_mmio_res->start,
3182				resource_size(hbus->low_mmio_res));
3183	}
3184
3185	if (hbus->high_mmio_space && hbus->high_mmio_res) {
3186		hbus->high_mmio_res->flags |= IORESOURCE_BUSY;
3187		vmbus_free_mmio(hbus->high_mmio_res->start,
3188				resource_size(hbus->high_mmio_res));
3189	}
3190}
3191
3192/**
3193 * hv_pci_allocate_bridge_windows() - Allocate memory regions
3194 * for the bus
3195 * @hbus:	Root PCI bus, as understood by this driver
3196 *
3197 * This function calls vmbus_allocate_mmio(), which is itself a
3198 * bit of a compromise.  Ideally, we might change the pnp layer
3199 * in the kernel such that it comprehends either PCI devices
3200 * which are "grandchildren of ACPI," with some intermediate bus
3201 * node (in this case, VMBus) or change it such that it
3202 * understands VMBus.  The pnp layer, however, has been declared
3203 * deprecated, and not subject to change.
3204 *
3205 * The workaround, implemented here, is to ask VMBus to allocate
3206 * MMIO space for this bus.  VMBus itself knows which ranges are
3207 * appropriate by looking at its own ACPI objects.  Then, after
3208 * these ranges are claimed, they're modified to look like they
3209 * would have looked if the ACPI and pnp code had allocated
3210 * bridge windows.  These descriptors have to exist in this form
3211 * in order to satisfy the code which will get invoked when the
3212 * endpoint PCI function driver calls request_mem_region() or
3213 * request_mem_region_exclusive().
3214 *
3215 * Return: 0 on success, -errno on failure
3216 */
3217static int hv_pci_allocate_bridge_windows(struct hv_pcibus_device *hbus)
3218{
3219	resource_size_t align;
3220	int ret;
3221
3222	if (hbus->low_mmio_space) {
3223		align = 1ULL << (63 - __builtin_clzll(hbus->low_mmio_space));
3224		ret = vmbus_allocate_mmio(&hbus->low_mmio_res, hbus->hdev, 0,
3225					  (u64)(u32)0xffffffff,
3226					  hbus->low_mmio_space,
3227					  align, false);
3228		if (ret) {
3229			dev_err(&hbus->hdev->device,
3230				"Need %#llx of low MMIO space. Consider reconfiguring the VM.\n",
3231				hbus->low_mmio_space);
3232			return ret;
3233		}
3234
3235		/* Modify this resource to become a bridge window. */
3236		hbus->low_mmio_res->flags |= IORESOURCE_WINDOW;
3237		hbus->low_mmio_res->flags &= ~IORESOURCE_BUSY;
3238		pci_add_resource(&hbus->bridge->windows, hbus->low_mmio_res);
3239	}
3240
3241	if (hbus->high_mmio_space) {
3242		align = 1ULL << (63 - __builtin_clzll(hbus->high_mmio_space));
3243		ret = vmbus_allocate_mmio(&hbus->high_mmio_res, hbus->hdev,
3244					  0x100000000, -1,
3245					  hbus->high_mmio_space, align,
3246					  false);
3247		if (ret) {
3248			dev_err(&hbus->hdev->device,
3249				"Need %#llx of high MMIO space. Consider reconfiguring the VM.\n",
3250				hbus->high_mmio_space);
3251			goto release_low_mmio;
3252		}
3253
3254		/* Modify this resource to become a bridge window. */
3255		hbus->high_mmio_res->flags |= IORESOURCE_WINDOW;
3256		hbus->high_mmio_res->flags &= ~IORESOURCE_BUSY;
3257		pci_add_resource(&hbus->bridge->windows, hbus->high_mmio_res);
3258	}
3259
3260	return 0;
3261
3262release_low_mmio:
3263	if (hbus->low_mmio_res) {
3264		vmbus_free_mmio(hbus->low_mmio_res->start,
3265				resource_size(hbus->low_mmio_res));
3266	}
3267
3268	return ret;
3269}
3270
3271/**
3272 * hv_allocate_config_window() - Find MMIO space for PCI Config
3273 * @hbus:	Root PCI bus, as understood by this driver
3274 *
3275 * This function claims memory-mapped I/O space for accessing
3276 * configuration space for the functions on this bus.
3277 *
3278 * Return: 0 on success, -errno on failure
3279 */
3280static int hv_allocate_config_window(struct hv_pcibus_device *hbus)
3281{
3282	int ret;
3283
3284	/*
3285	 * Set up a region of MMIO space to use for accessing configuration
3286	 * space.
3287	 */
3288	ret = vmbus_allocate_mmio(&hbus->mem_config, hbus->hdev, 0, -1,
3289				  PCI_CONFIG_MMIO_LENGTH, 0x1000, false);
3290	if (ret)
3291		return ret;
3292
3293	/*
3294	 * vmbus_allocate_mmio() gets used for allocating both device endpoint
3295	 * resource claims (those which cannot be overlapped) and the ranges
3296	 * which are valid for the children of this bus, which are intended
3297	 * to be overlapped by those children.  Set the flag on this claim
3298	 * meaning that this region can't be overlapped.
3299	 */
3300
3301	hbus->mem_config->flags |= IORESOURCE_BUSY;
3302
3303	return 0;
3304}
3305
3306static void hv_free_config_window(struct hv_pcibus_device *hbus)
3307{
3308	vmbus_free_mmio(hbus->mem_config->start, PCI_CONFIG_MMIO_LENGTH);
3309}
3310
3311static int hv_pci_bus_exit(struct hv_device *hdev, bool keep_devs);
3312
3313/**
3314 * hv_pci_enter_d0() - Bring the "bus" into the D0 power state
3315 * @hdev:	VMBus's tracking struct for this root PCI bus
3316 *
3317 * Return: 0 on success, -errno on failure
3318 */
3319static int hv_pci_enter_d0(struct hv_device *hdev)
3320{
3321	struct hv_pcibus_device *hbus = hv_get_drvdata(hdev);
3322	struct pci_bus_d0_entry *d0_entry;
3323	struct hv_pci_compl comp_pkt;
3324	struct pci_packet *pkt;
3325	bool retry = true;
3326	int ret;
3327
3328enter_d0_retry:
3329	/*
3330	 * Tell the host that the bus is ready to use, and moved into the
3331	 * powered-on state.  This includes telling the host which region
3332	 * of memory-mapped I/O space has been chosen for configuration space
3333	 * access.
3334	 */
3335	pkt = kzalloc(sizeof(*pkt) + sizeof(*d0_entry), GFP_KERNEL);
3336	if (!pkt)
3337		return -ENOMEM;
3338
3339	init_completion(&comp_pkt.host_event);
3340	pkt->completion_func = hv_pci_generic_compl;
3341	pkt->compl_ctxt = &comp_pkt;
3342	d0_entry = (struct pci_bus_d0_entry *)&pkt->message;
3343	d0_entry->message_type.type = PCI_BUS_D0ENTRY;
3344	d0_entry->mmio_base = hbus->mem_config->start;
3345
3346	ret = vmbus_sendpacket(hdev->channel, d0_entry, sizeof(*d0_entry),
3347			       (unsigned long)pkt, VM_PKT_DATA_INBAND,
3348			       VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
3349	if (!ret)
3350		ret = wait_for_response(hdev, &comp_pkt.host_event);
3351
3352	if (ret)
3353		goto exit;
3354
3355	/*
3356	 * In certain case (Kdump) the pci device of interest was
3357	 * not cleanly shut down and resource is still held on host
3358	 * side, the host could return invalid device status.
3359	 * We need to explicitly request host to release the resource
3360	 * and try to enter D0 again.
3361	 */
3362	if (comp_pkt.completion_status < 0 && retry) {
3363		retry = false;
3364
3365		dev_err(&hdev->device, "Retrying D0 Entry\n");
3366
3367		/*
3368		 * Hv_pci_bus_exit() calls hv_send_resource_released()
3369		 * to free up resources of its child devices.
3370		 * In the kdump kernel we need to set the
3371		 * wslot_res_allocated to 255 so it scans all child
3372		 * devices to release resources allocated in the
3373		 * normal kernel before panic happened.
3374		 */
3375		hbus->wslot_res_allocated = 255;
3376
3377		ret = hv_pci_bus_exit(hdev, true);
3378
3379		if (ret == 0) {
3380			kfree(pkt);
3381			goto enter_d0_retry;
3382		}
3383		dev_err(&hdev->device,
3384			"Retrying D0 failed with ret %d\n", ret);
3385	}
3386
3387	if (comp_pkt.completion_status < 0) {
3388		dev_err(&hdev->device,
3389			"PCI Pass-through VSP failed D0 Entry with status %x\n",
3390			comp_pkt.completion_status);
3391		ret = -EPROTO;
3392		goto exit;
3393	}
3394
3395	ret = 0;
3396
3397exit:
3398	kfree(pkt);
3399	return ret;
3400}
3401
3402/**
3403 * hv_pci_query_relations() - Ask host to send list of child
3404 * devices
3405 * @hdev:	VMBus's tracking struct for this root PCI bus
3406 *
3407 * Return: 0 on success, -errno on failure
3408 */
3409static int hv_pci_query_relations(struct hv_device *hdev)
3410{
3411	struct hv_pcibus_device *hbus = hv_get_drvdata(hdev);
3412	struct pci_message message;
3413	struct completion comp;
3414	int ret;
3415
3416	/* Ask the host to send along the list of child devices */
3417	init_completion(&comp);
3418	if (cmpxchg(&hbus->survey_event, NULL, &comp))
3419		return -ENOTEMPTY;
3420
3421	memset(&message, 0, sizeof(message));
3422	message.type = PCI_QUERY_BUS_RELATIONS;
3423
3424	ret = vmbus_sendpacket(hdev->channel, &message, sizeof(message),
3425			       0, VM_PKT_DATA_INBAND, 0);
3426	if (!ret)
3427		ret = wait_for_response(hdev, &comp);
3428
3429	/*
3430	 * In the case of fast device addition/removal, it's possible that
3431	 * vmbus_sendpacket() or wait_for_response() returns -ENODEV but we
3432	 * already got a PCI_BUS_RELATIONS* message from the host and the
3433	 * channel callback already scheduled a work to hbus->wq, which can be
3434	 * running pci_devices_present_work() -> survey_child_resources() ->
3435	 * complete(&hbus->survey_event), even after hv_pci_query_relations()
3436	 * exits and the stack variable 'comp' is no longer valid; as a result,
3437	 * a hang or a page fault may happen when the complete() calls
3438	 * raw_spin_lock_irqsave(). Flush hbus->wq before we exit from
3439	 * hv_pci_query_relations() to avoid the issues. Note: if 'ret' is
3440	 * -ENODEV, there can't be any more work item scheduled to hbus->wq
3441	 * after the flush_workqueue(): see vmbus_onoffer_rescind() ->
3442	 * vmbus_reset_channel_cb(), vmbus_rescind_cleanup() ->
3443	 * channel->rescind = true.
3444	 */
3445	flush_workqueue(hbus->wq);
3446
3447	return ret;
3448}
3449
3450/**
3451 * hv_send_resources_allocated() - Report local resource choices
3452 * @hdev:	VMBus's tracking struct for this root PCI bus
3453 *
3454 * The host OS is expecting to be sent a request as a message
3455 * which contains all the resources that the device will use.
3456 * The response contains those same resources, "translated"
3457 * which is to say, the values which should be used by the
3458 * hardware, when it delivers an interrupt.  (MMIO resources are
3459 * used in local terms.)  This is nice for Windows, and lines up
3460 * with the FDO/PDO split, which doesn't exist in Linux.  Linux
3461 * is deeply expecting to scan an emulated PCI configuration
3462 * space.  So this message is sent here only to drive the state
3463 * machine on the host forward.
3464 *
3465 * Return: 0 on success, -errno on failure
3466 */
3467static int hv_send_resources_allocated(struct hv_device *hdev)
3468{
3469	struct hv_pcibus_device *hbus = hv_get_drvdata(hdev);
3470	struct pci_resources_assigned *res_assigned;
3471	struct pci_resources_assigned2 *res_assigned2;
3472	struct hv_pci_compl comp_pkt;
3473	struct hv_pci_dev *hpdev;
3474	struct pci_packet *pkt;
3475	size_t size_res;
3476	int wslot;
3477	int ret;
3478
3479	size_res = (hbus->protocol_version < PCI_PROTOCOL_VERSION_1_2)
3480			? sizeof(*res_assigned) : sizeof(*res_assigned2);
3481
3482	pkt = kmalloc(sizeof(*pkt) + size_res, GFP_KERNEL);
3483	if (!pkt)
3484		return -ENOMEM;
3485
3486	ret = 0;
3487
3488	for (wslot = 0; wslot < 256; wslot++) {
3489		hpdev = get_pcichild_wslot(hbus, wslot);
3490		if (!hpdev)
3491			continue;
3492
3493		memset(pkt, 0, sizeof(*pkt) + size_res);
3494		init_completion(&comp_pkt.host_event);
3495		pkt->completion_func = hv_pci_generic_compl;
3496		pkt->compl_ctxt = &comp_pkt;
3497
3498		if (hbus->protocol_version < PCI_PROTOCOL_VERSION_1_2) {
3499			res_assigned =
3500				(struct pci_resources_assigned *)&pkt->message;
3501			res_assigned->message_type.type =
3502				PCI_RESOURCES_ASSIGNED;
3503			res_assigned->wslot.slot = hpdev->desc.win_slot.slot;
3504		} else {
3505			res_assigned2 =
3506				(struct pci_resources_assigned2 *)&pkt->message;
3507			res_assigned2->message_type.type =
3508				PCI_RESOURCES_ASSIGNED2;
3509			res_assigned2->wslot.slot = hpdev->desc.win_slot.slot;
3510		}
3511		put_pcichild(hpdev);
3512
3513		ret = vmbus_sendpacket(hdev->channel, &pkt->message,
3514				size_res, (unsigned long)pkt,
3515				VM_PKT_DATA_INBAND,
3516				VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
3517		if (!ret)
3518			ret = wait_for_response(hdev, &comp_pkt.host_event);
3519		if (ret)
3520			break;
3521
3522		if (comp_pkt.completion_status < 0) {
3523			ret = -EPROTO;
3524			dev_err(&hdev->device,
3525				"resource allocated returned 0x%x",
3526				comp_pkt.completion_status);
3527			break;
3528		}
3529
3530		hbus->wslot_res_allocated = wslot;
3531	}
3532
3533	kfree(pkt);
3534	return ret;
3535}
3536
3537/**
3538 * hv_send_resources_released() - Report local resources
3539 * released
3540 * @hdev:	VMBus's tracking struct for this root PCI bus
3541 *
3542 * Return: 0 on success, -errno on failure
3543 */
3544static int hv_send_resources_released(struct hv_device *hdev)
3545{
3546	struct hv_pcibus_device *hbus = hv_get_drvdata(hdev);
3547	struct pci_child_message pkt;
3548	struct hv_pci_dev *hpdev;
3549	int wslot;
3550	int ret;
3551
3552	for (wslot = hbus->wslot_res_allocated; wslot >= 0; wslot--) {
3553		hpdev = get_pcichild_wslot(hbus, wslot);
3554		if (!hpdev)
3555			continue;
3556
3557		memset(&pkt, 0, sizeof(pkt));
3558		pkt.message_type.type = PCI_RESOURCES_RELEASED;
3559		pkt.wslot.slot = hpdev->desc.win_slot.slot;
3560
3561		put_pcichild(hpdev);
3562
3563		ret = vmbus_sendpacket(hdev->channel, &pkt, sizeof(pkt), 0,
3564				       VM_PKT_DATA_INBAND, 0);
3565		if (ret)
3566			return ret;
3567
3568		hbus->wslot_res_allocated = wslot - 1;
3569	}
3570
3571	hbus->wslot_res_allocated = -1;
3572
3573	return 0;
3574}
3575
3576#define HVPCI_DOM_MAP_SIZE (64 * 1024)
3577static DECLARE_BITMAP(hvpci_dom_map, HVPCI_DOM_MAP_SIZE);
3578
3579/*
3580 * PCI domain number 0 is used by emulated devices on Gen1 VMs, so define 0
3581 * as invalid for passthrough PCI devices of this driver.
3582 */
3583#define HVPCI_DOM_INVALID 0
3584
3585/**
3586 * hv_get_dom_num() - Get a valid PCI domain number
3587 * Check if the PCI domain number is in use, and return another number if
3588 * it is in use.
3589 *
3590 * @dom: Requested domain number
3591 *
3592 * return: domain number on success, HVPCI_DOM_INVALID on failure
3593 */
3594static u16 hv_get_dom_num(u16 dom)
3595{
3596	unsigned int i;
3597
3598	if (test_and_set_bit(dom, hvpci_dom_map) == 0)
3599		return dom;
3600
3601	for_each_clear_bit(i, hvpci_dom_map, HVPCI_DOM_MAP_SIZE) {
3602		if (test_and_set_bit(i, hvpci_dom_map) == 0)
3603			return i;
3604	}
3605
3606	return HVPCI_DOM_INVALID;
3607}
3608
3609/**
3610 * hv_put_dom_num() - Mark the PCI domain number as free
3611 * @dom: Domain number to be freed
3612 */
3613static void hv_put_dom_num(u16 dom)
3614{
3615	clear_bit(dom, hvpci_dom_map);
3616}
3617
3618/**
3619 * hv_pci_probe() - New VMBus channel probe, for a root PCI bus
3620 * @hdev:	VMBus's tracking struct for this root PCI bus
3621 * @dev_id:	Identifies the device itself
3622 *
3623 * Return: 0 on success, -errno on failure
3624 */
3625static int hv_pci_probe(struct hv_device *hdev,
3626			const struct hv_vmbus_device_id *dev_id)
3627{
3628	struct pci_host_bridge *bridge;
3629	struct hv_pcibus_device *hbus;
3630	u16 dom_req, dom;
3631	char *name;
3632	int ret;
3633
3634	bridge = devm_pci_alloc_host_bridge(&hdev->device, 0);
3635	if (!bridge)
3636		return -ENOMEM;
3637
3638	hbus = kzalloc(sizeof(*hbus), GFP_KERNEL);
3639	if (!hbus)
3640		return -ENOMEM;
3641
3642	hbus->bridge = bridge;
3643	mutex_init(&hbus->state_lock);
3644	hbus->state = hv_pcibus_init;
3645	hbus->wslot_res_allocated = -1;
3646
3647	/*
3648	 * The PCI bus "domain" is what is called "segment" in ACPI and other
3649	 * specs. Pull it from the instance ID, to get something usually
3650	 * unique. In rare cases of collision, we will find out another number
3651	 * not in use.
3652	 *
3653	 * Note that, since this code only runs in a Hyper-V VM, Hyper-V
3654	 * together with this guest driver can guarantee that (1) The only
3655	 * domain used by Gen1 VMs for something that looks like a physical
3656	 * PCI bus (which is actually emulated by the hypervisor) is domain 0.
3657	 * (2) There will be no overlap between domains (after fixing possible
3658	 * collisions) in the same VM.
3659	 */
3660	dom_req = hdev->dev_instance.b[5] << 8 | hdev->dev_instance.b[4];
3661	dom = hv_get_dom_num(dom_req);
3662
3663	if (dom == HVPCI_DOM_INVALID) {
3664		dev_err(&hdev->device,
3665			"Unable to use dom# 0x%x or other numbers", dom_req);
3666		ret = -EINVAL;
3667		goto free_bus;
3668	}
3669
3670	if (dom != dom_req)
3671		dev_info(&hdev->device,
3672			 "PCI dom# 0x%x has collision, using 0x%x",
3673			 dom_req, dom);
3674
3675	hbus->bridge->domain_nr = dom;
3676#ifdef CONFIG_X86
3677	hbus->sysdata.domain = dom;
3678	hbus->use_calls = !!(ms_hyperv.hints & HV_X64_USE_MMIO_HYPERCALLS);
3679#elif defined(CONFIG_ARM64)
3680	/*
3681	 * Set the PCI bus parent to be the corresponding VMbus
3682	 * device. Then the VMbus device will be assigned as the
3683	 * ACPI companion in pcibios_root_bridge_prepare() and
3684	 * pci_dma_configure() will propagate device coherence
3685	 * information to devices created on the bus.
3686	 */
3687	hbus->sysdata.parent = hdev->device.parent;
3688	hbus->use_calls = false;
3689#endif
3690
3691	hbus->hdev = hdev;
3692	INIT_LIST_HEAD(&hbus->children);
3693	INIT_LIST_HEAD(&hbus->dr_list);
3694	spin_lock_init(&hbus->config_lock);
3695	spin_lock_init(&hbus->device_list_lock);
3696	hbus->wq = alloc_ordered_workqueue("hv_pci_%x", 0,
3697					   hbus->bridge->domain_nr);
3698	if (!hbus->wq) {
3699		ret = -ENOMEM;
3700		goto free_dom;
3701	}
3702
3703	hdev->channel->next_request_id_callback = vmbus_next_request_id;
3704	hdev->channel->request_addr_callback = vmbus_request_addr;
3705	hdev->channel->rqstor_size = HV_PCI_RQSTOR_SIZE;
3706
3707	ret = vmbus_open(hdev->channel, pci_ring_size, pci_ring_size, NULL, 0,
3708			 hv_pci_onchannelcallback, hbus);
3709	if (ret)
3710		goto destroy_wq;
3711
3712	hv_set_drvdata(hdev, hbus);
3713
3714	ret = hv_pci_protocol_negotiation(hdev, pci_protocol_versions,
3715					  ARRAY_SIZE(pci_protocol_versions));
3716	if (ret)
3717		goto close;
3718
3719	ret = hv_allocate_config_window(hbus);
3720	if (ret)
3721		goto close;
3722
3723	hbus->cfg_addr = ioremap(hbus->mem_config->start,
3724				 PCI_CONFIG_MMIO_LENGTH);
3725	if (!hbus->cfg_addr) {
3726		dev_err(&hdev->device,
3727			"Unable to map a virtual address for config space\n");
3728		ret = -ENOMEM;
3729		goto free_config;
3730	}
3731
3732	name = kasprintf(GFP_KERNEL, "%pUL", &hdev->dev_instance);
3733	if (!name) {
3734		ret = -ENOMEM;
3735		goto unmap;
3736	}
3737
3738	hbus->fwnode = irq_domain_alloc_named_fwnode(name);
3739	kfree(name);
3740	if (!hbus->fwnode) {
3741		ret = -ENOMEM;
3742		goto unmap;
3743	}
3744
3745	ret = hv_pcie_init_irq_domain(hbus);
3746	if (ret)
3747		goto free_fwnode;
3748
3749	ret = hv_pci_query_relations(hdev);
3750	if (ret)
3751		goto free_irq_domain;
3752
3753	mutex_lock(&hbus->state_lock);
3754
3755	ret = hv_pci_enter_d0(hdev);
3756	if (ret)
3757		goto release_state_lock;
3758
3759	ret = hv_pci_allocate_bridge_windows(hbus);
3760	if (ret)
3761		goto exit_d0;
3762
3763	ret = hv_send_resources_allocated(hdev);
3764	if (ret)
3765		goto free_windows;
3766
3767	prepopulate_bars(hbus);
3768
3769	hbus->state = hv_pcibus_probed;
3770
3771	ret = create_root_hv_pci_bus(hbus);
3772	if (ret)
3773		goto free_windows;
3774
3775	mutex_unlock(&hbus->state_lock);
3776	return 0;
3777
3778free_windows:
3779	hv_pci_free_bridge_windows(hbus);
3780exit_d0:
3781	(void) hv_pci_bus_exit(hdev, true);
3782release_state_lock:
3783	mutex_unlock(&hbus->state_lock);
3784free_irq_domain:
3785	irq_domain_remove(hbus->irq_domain);
3786free_fwnode:
3787	irq_domain_free_fwnode(hbus->fwnode);
3788unmap:
3789	iounmap(hbus->cfg_addr);
3790free_config:
3791	hv_free_config_window(hbus);
3792close:
3793	vmbus_close(hdev->channel);
3794destroy_wq:
3795	destroy_workqueue(hbus->wq);
3796free_dom:
3797	hv_put_dom_num(hbus->bridge->domain_nr);
3798free_bus:
3799	kfree(hbus);
3800	return ret;
3801}
3802
3803static int hv_pci_bus_exit(struct hv_device *hdev, bool keep_devs)
3804{
3805	struct hv_pcibus_device *hbus = hv_get_drvdata(hdev);
3806	struct vmbus_channel *chan = hdev->channel;
3807	struct {
3808		struct pci_packet teardown_packet;
3809		u8 buffer[sizeof(struct pci_message)];
3810	} pkt;
3811	struct hv_pci_compl comp_pkt;
3812	struct hv_pci_dev *hpdev, *tmp;
3813	unsigned long flags;
3814	u64 trans_id;
3815	int ret;
3816
3817	/*
3818	 * After the host sends the RESCIND_CHANNEL message, it doesn't
3819	 * access the per-channel ringbuffer any longer.
3820	 */
3821	if (chan->rescind)
3822		return 0;
3823
3824	if (!keep_devs) {
3825		struct list_head removed;
3826
3827		/* Move all present children to the list on stack */
3828		INIT_LIST_HEAD(&removed);
3829		spin_lock_irqsave(&hbus->device_list_lock, flags);
3830		list_for_each_entry_safe(hpdev, tmp, &hbus->children, list_entry)
3831			list_move_tail(&hpdev->list_entry, &removed);
3832		spin_unlock_irqrestore(&hbus->device_list_lock, flags);
3833
3834		/* Remove all children in the list */
3835		list_for_each_entry_safe(hpdev, tmp, &removed, list_entry) {
3836			list_del(&hpdev->list_entry);
3837			if (hpdev->pci_slot)
3838				pci_destroy_slot(hpdev->pci_slot);
3839			/* For the two refs got in new_pcichild_device() */
3840			put_pcichild(hpdev);
3841			put_pcichild(hpdev);
3842		}
3843	}
3844
3845	ret = hv_send_resources_released(hdev);
3846	if (ret) {
3847		dev_err(&hdev->device,
3848			"Couldn't send resources released packet(s)\n");
3849		return ret;
3850	}
3851
3852	memset(&pkt.teardown_packet, 0, sizeof(pkt.teardown_packet));
3853	init_completion(&comp_pkt.host_event);
3854	pkt.teardown_packet.completion_func = hv_pci_generic_compl;
3855	pkt.teardown_packet.compl_ctxt = &comp_pkt;
3856	pkt.teardown_packet.message[0].type = PCI_BUS_D0EXIT;
3857
3858	ret = vmbus_sendpacket_getid(chan, &pkt.teardown_packet.message,
3859				     sizeof(struct pci_message),
3860				     (unsigned long)&pkt.teardown_packet,
3861				     &trans_id, VM_PKT_DATA_INBAND,
3862				     VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
3863	if (ret)
3864		return ret;
3865
3866	if (wait_for_completion_timeout(&comp_pkt.host_event, 10 * HZ) == 0) {
3867		/*
3868		 * The completion packet on the stack becomes invalid after
3869		 * 'return'; remove the ID from the VMbus requestor if the
3870		 * identifier is still mapped to/associated with the packet.
3871		 *
3872		 * Cf. hv_pci_onchannelcallback().
3873		 */
3874		vmbus_request_addr_match(chan, trans_id,
3875					 (unsigned long)&pkt.teardown_packet);
3876		return -ETIMEDOUT;
3877	}
3878
3879	return 0;
3880}
3881
3882/**
3883 * hv_pci_remove() - Remove routine for this VMBus channel
3884 * @hdev:	VMBus's tracking struct for this root PCI bus
3885 */
3886static void hv_pci_remove(struct hv_device *hdev)
3887{
3888	struct hv_pcibus_device *hbus;
3889
3890	hbus = hv_get_drvdata(hdev);
3891	if (hbus->state == hv_pcibus_installed) {
3892		tasklet_disable(&hdev->channel->callback_event);
3893		hbus->state = hv_pcibus_removing;
3894		tasklet_enable(&hdev->channel->callback_event);
3895		destroy_workqueue(hbus->wq);
3896		hbus->wq = NULL;
3897		/*
3898		 * At this point, no work is running or can be scheduled
3899		 * on hbus-wq. We can't race with hv_pci_devices_present()
3900		 * or hv_pci_eject_device(), it's safe to proceed.
3901		 */
3902
3903		/* Remove the bus from PCI's point of view. */
3904		pci_lock_rescan_remove();
3905		pci_stop_root_bus(hbus->bridge->bus);
3906		hv_pci_remove_slots(hbus);
3907		pci_remove_root_bus(hbus->bridge->bus);
3908		pci_unlock_rescan_remove();
3909	}
3910
3911	hv_pci_bus_exit(hdev, false);
3912
3913	vmbus_close(hdev->channel);
3914
3915	iounmap(hbus->cfg_addr);
3916	hv_free_config_window(hbus);
3917	hv_pci_free_bridge_windows(hbus);
3918	irq_domain_remove(hbus->irq_domain);
3919	irq_domain_free_fwnode(hbus->fwnode);
3920
3921	hv_put_dom_num(hbus->bridge->domain_nr);
3922
3923	kfree(hbus);
3924}
3925
3926static int hv_pci_suspend(struct hv_device *hdev)
3927{
3928	struct hv_pcibus_device *hbus = hv_get_drvdata(hdev);
3929	enum hv_pcibus_state old_state;
3930	int ret;
3931
3932	/*
3933	 * hv_pci_suspend() must make sure there are no pending work items
3934	 * before calling vmbus_close(), since it runs in a process context
3935	 * as a callback in dpm_suspend().  When it starts to run, the channel
3936	 * callback hv_pci_onchannelcallback(), which runs in a tasklet
3937	 * context, can be still running concurrently and scheduling new work
3938	 * items onto hbus->wq in hv_pci_devices_present() and
3939	 * hv_pci_eject_device(), and the work item handlers can access the
3940	 * vmbus channel, which can be being closed by hv_pci_suspend(), e.g.
3941	 * the work item handler pci_devices_present_work() ->
3942	 * new_pcichild_device() writes to the vmbus channel.
3943	 *
3944	 * To eliminate the race, hv_pci_suspend() disables the channel
3945	 * callback tasklet, sets hbus->state to hv_pcibus_removing, and
3946	 * re-enables the tasklet. This way, when hv_pci_suspend() proceeds,
3947	 * it knows that no new work item can be scheduled, and then it flushes
3948	 * hbus->wq and safely closes the vmbus channel.
3949	 */
3950	tasklet_disable(&hdev->channel->callback_event);
3951
3952	/* Change the hbus state to prevent new work items. */
3953	old_state = hbus->state;
3954	if (hbus->state == hv_pcibus_installed)
3955		hbus->state = hv_pcibus_removing;
3956
3957	tasklet_enable(&hdev->channel->callback_event);
3958
3959	if (old_state != hv_pcibus_installed)
3960		return -EINVAL;
3961
3962	flush_workqueue(hbus->wq);
3963
3964	ret = hv_pci_bus_exit(hdev, true);
3965	if (ret)
3966		return ret;
3967
3968	vmbus_close(hdev->channel);
3969
3970	return 0;
3971}
3972
3973static int hv_pci_restore_msi_msg(struct pci_dev *pdev, void *arg)
3974{
3975	struct irq_data *irq_data;
3976	struct msi_desc *entry;
3977	int ret = 0;
3978
3979	if (!pdev->msi_enabled && !pdev->msix_enabled)
3980		return 0;
3981
3982	msi_lock_descs(&pdev->dev);
3983	msi_for_each_desc(entry, &pdev->dev, MSI_DESC_ASSOCIATED) {
3984		irq_data = irq_get_irq_data(entry->irq);
3985		if (WARN_ON_ONCE(!irq_data)) {
3986			ret = -EINVAL;
3987			break;
3988		}
3989
3990		hv_compose_msi_msg(irq_data, &entry->msg);
3991	}
3992	msi_unlock_descs(&pdev->dev);
3993
3994	return ret;
3995}
3996
3997/*
3998 * Upon resume, pci_restore_msi_state() -> ... ->  __pci_write_msi_msg()
3999 * directly writes the MSI/MSI-X registers via MMIO, but since Hyper-V
4000 * doesn't trap and emulate the MMIO accesses, here hv_compose_msi_msg()
4001 * must be used to ask Hyper-V to re-create the IOMMU Interrupt Remapping
4002 * Table entries.
4003 */
4004static void hv_pci_restore_msi_state(struct hv_pcibus_device *hbus)
4005{
4006	pci_walk_bus(hbus->bridge->bus, hv_pci_restore_msi_msg, NULL);
4007}
4008
4009static int hv_pci_resume(struct hv_device *hdev)
4010{
4011	struct hv_pcibus_device *hbus = hv_get_drvdata(hdev);
4012	enum pci_protocol_version_t version[1];
4013	int ret;
4014
4015	hbus->state = hv_pcibus_init;
4016
4017	hdev->channel->next_request_id_callback = vmbus_next_request_id;
4018	hdev->channel->request_addr_callback = vmbus_request_addr;
4019	hdev->channel->rqstor_size = HV_PCI_RQSTOR_SIZE;
4020
4021	ret = vmbus_open(hdev->channel, pci_ring_size, pci_ring_size, NULL, 0,
4022			 hv_pci_onchannelcallback, hbus);
4023	if (ret)
4024		return ret;
4025
4026	/* Only use the version that was in use before hibernation. */
4027	version[0] = hbus->protocol_version;
4028	ret = hv_pci_protocol_negotiation(hdev, version, 1);
4029	if (ret)
4030		goto out;
4031
4032	ret = hv_pci_query_relations(hdev);
4033	if (ret)
4034		goto out;
4035
4036	mutex_lock(&hbus->state_lock);
4037
4038	ret = hv_pci_enter_d0(hdev);
4039	if (ret)
4040		goto release_state_lock;
4041
4042	ret = hv_send_resources_allocated(hdev);
4043	if (ret)
4044		goto release_state_lock;
4045
4046	prepopulate_bars(hbus);
4047
4048	hv_pci_restore_msi_state(hbus);
4049
4050	hbus->state = hv_pcibus_installed;
4051	mutex_unlock(&hbus->state_lock);
4052	return 0;
4053
4054release_state_lock:
4055	mutex_unlock(&hbus->state_lock);
4056out:
4057	vmbus_close(hdev->channel);
4058	return ret;
4059}
4060
4061static const struct hv_vmbus_device_id hv_pci_id_table[] = {
4062	/* PCI Pass-through Class ID */
4063	/* 44C4F61D-4444-4400-9D52-802E27EDE19F */
4064	{ HV_PCIE_GUID, },
4065	{ },
4066};
4067
4068MODULE_DEVICE_TABLE(vmbus, hv_pci_id_table);
4069
4070static struct hv_driver hv_pci_drv = {
4071	.name		= "hv_pci",
4072	.id_table	= hv_pci_id_table,
4073	.probe		= hv_pci_probe,
4074	.remove		= hv_pci_remove,
4075	.suspend	= hv_pci_suspend,
4076	.resume		= hv_pci_resume,
4077};
4078
4079static void __exit exit_hv_pci_drv(void)
4080{
4081	vmbus_driver_unregister(&hv_pci_drv);
4082
4083	hvpci_block_ops.read_block = NULL;
4084	hvpci_block_ops.write_block = NULL;
4085	hvpci_block_ops.reg_blk_invalidate = NULL;
4086}
4087
4088static int __init init_hv_pci_drv(void)
4089{
4090	int ret;
4091
4092	if (!hv_is_hyperv_initialized())
4093		return -ENODEV;
4094
4095	ret = hv_pci_irqchip_init();
4096	if (ret)
4097		return ret;
4098
4099	/* Set the invalid domain number's bit, so it will not be used */
4100	set_bit(HVPCI_DOM_INVALID, hvpci_dom_map);
4101
4102	/* Initialize PCI block r/w interface */
4103	hvpci_block_ops.read_block = hv_read_config_block;
4104	hvpci_block_ops.write_block = hv_write_config_block;
4105	hvpci_block_ops.reg_blk_invalidate = hv_register_block_invalidate;
4106
4107	return vmbus_driver_register(&hv_pci_drv);
4108}
4109
4110module_init(init_hv_pci_drv);
4111module_exit(exit_hv_pci_drv);
4112
4113MODULE_DESCRIPTION("Hyper-V PCI");
4114MODULE_LICENSE("GPL v2");