1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (c) Microsoft Corporation.
6 * Jake Oshins <jakeo@microsoft.com>
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.
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.
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.
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.
40 #include <linux/kernel.h>
41 #include <linux/module.h>
42 #include <linux/pci.h>
43 #include <linux/delay.h>
44 #include <linux/semaphore.h>
45 #include <linux/irqdomain.h>
46 #include <asm/irqdomain.h>
48 #include <linux/irq.h>
49 #include <linux/msi.h>
50 #include <linux/hyperv.h>
51 #include <linux/refcount.h>
52 #include <asm/mshyperv.h>
55 * Protocol versions. The low word is the minor version, the high word the
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)
63 enum 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 */
69 #define CPU_AFFINITY_ALL -1ULL
72 * Supported protocol versions in the order of probing - highest go
75 static enum pci_protocol_version_t pci_protocol_versions[] = {
76 PCI_PROTOCOL_VERSION_1_3,
77 PCI_PROTOCOL_VERSION_1_2,
78 PCI_PROTOCOL_VERSION_1_1,
81 #define PCI_CONFIG_MMIO_LENGTH 0x2000
82 #define CFG_PAGE_OFFSET 0x1000
83 #define CFG_PAGE_SIZE (PCI_CONFIG_MMIO_LENGTH - CFG_PAGE_OFFSET)
85 #define MAX_SUPPORTED_MSI_MESSAGES 0x400
87 #define STATUS_REVISION_MISMATCH 0xC0000059
89 /* space for 32bit serial number as string */
90 #define SLOT_NAME_SIZE 11
96 enum pci_message_type {
100 PCI_MESSAGE_BASE = 0x42490000,
101 PCI_BUS_RELATIONS = PCI_MESSAGE_BASE + 0,
102 PCI_QUERY_BUS_RELATIONS = PCI_MESSAGE_BASE + 1,
103 PCI_POWER_STATE_CHANGE = PCI_MESSAGE_BASE + 4,
104 PCI_QUERY_RESOURCE_REQUIREMENTS = PCI_MESSAGE_BASE + 5,
105 PCI_QUERY_RESOURCE_RESOURCES = PCI_MESSAGE_BASE + 6,
106 PCI_BUS_D0ENTRY = PCI_MESSAGE_BASE + 7,
107 PCI_BUS_D0EXIT = PCI_MESSAGE_BASE + 8,
108 PCI_READ_BLOCK = PCI_MESSAGE_BASE + 9,
109 PCI_WRITE_BLOCK = PCI_MESSAGE_BASE + 0xA,
110 PCI_EJECT = PCI_MESSAGE_BASE + 0xB,
111 PCI_QUERY_STOP = PCI_MESSAGE_BASE + 0xC,
112 PCI_REENABLE = PCI_MESSAGE_BASE + 0xD,
113 PCI_QUERY_STOP_FAILED = PCI_MESSAGE_BASE + 0xE,
114 PCI_EJECTION_COMPLETE = PCI_MESSAGE_BASE + 0xF,
115 PCI_RESOURCES_ASSIGNED = PCI_MESSAGE_BASE + 0x10,
116 PCI_RESOURCES_RELEASED = PCI_MESSAGE_BASE + 0x11,
117 PCI_INVALIDATE_BLOCK = PCI_MESSAGE_BASE + 0x12,
118 PCI_QUERY_PROTOCOL_VERSION = PCI_MESSAGE_BASE + 0x13,
119 PCI_CREATE_INTERRUPT_MESSAGE = PCI_MESSAGE_BASE + 0x14,
120 PCI_DELETE_INTERRUPT_MESSAGE = PCI_MESSAGE_BASE + 0x15,
121 PCI_RESOURCES_ASSIGNED2 = PCI_MESSAGE_BASE + 0x16,
122 PCI_CREATE_INTERRUPT_MESSAGE2 = PCI_MESSAGE_BASE + 0x17,
123 PCI_DELETE_INTERRUPT_MESSAGE2 = PCI_MESSAGE_BASE + 0x18, /* unused */
124 PCI_BUS_RELATIONS2 = PCI_MESSAGE_BASE + 0x19,
129 * Structures defining the virtual PCI Express protocol.
141 * Function numbers are 8-bits wide on Express, as interpreted through ARI,
142 * which is all this driver does. This representation is the one used in
143 * Windows, which is what is expected when sending this back and forth with
144 * the Hyper-V parent partition.
146 union win_slot_encoding {
156 * Pretty much as defined in the PCI Specifications.
158 struct pci_function_description {
159 u16 v_id; /* vendor ID */
160 u16 d_id; /* device ID */
166 union win_slot_encoding win_slot;
167 u32 ser; /* serial number */
170 enum pci_device_description_flags {
171 HV_PCI_DEVICE_FLAG_NONE = 0x0,
172 HV_PCI_DEVICE_FLAG_NUMA_AFFINITY = 0x1,
175 struct pci_function_description2 {
176 u16 v_id; /* vendor ID */
177 u16 d_id; /* device ID */
183 union win_slot_encoding win_slot;
184 u32 ser; /* serial number */
186 u16 virtual_numa_node;
193 * @delivery_mode: As defined in Intel's Programmer's
194 * Reference Manual, Volume 3, Chapter 8.
195 * @vector_count: Number of contiguous entries in the
196 * Interrupt Descriptor Table that are
197 * occupied by this Message-Signaled
198 * Interrupt. For "MSI", as first defined
199 * in PCI 2.2, this can be between 1 and
200 * 32. For "MSI-X," as first defined in PCI
201 * 3.0, this must be 1, as each MSI-X table
202 * entry would have its own descriptor.
203 * @reserved: Empty space
204 * @cpu_mask: All the target virtual processors.
215 * struct hv_msi_desc2 - 1.2 version of hv_msi_desc
217 * @delivery_mode: As defined in Intel's Programmer's
218 * Reference Manual, Volume 3, Chapter 8.
219 * @vector_count: Number of contiguous entries in the
220 * Interrupt Descriptor Table that are
221 * occupied by this Message-Signaled
222 * Interrupt. For "MSI", as first defined
223 * in PCI 2.2, this can be between 1 and
224 * 32. For "MSI-X," as first defined in PCI
225 * 3.0, this must be 1, as each MSI-X table
226 * entry would have its own descriptor.
227 * @processor_count: number of bits enabled in array.
228 * @processor_array: All the target virtual processors.
230 struct hv_msi_desc2 {
235 u16 processor_array[32];
239 * struct tran_int_desc
240 * @reserved: unused, padding
241 * @vector_count: same as in hv_msi_desc
242 * @data: This is the "data payload" value that is
243 * written by the device when it generates
244 * a message-signaled interrupt, either MSI
246 * @address: This is the address to which the data
247 * payload is written on interrupt
250 struct tran_int_desc {
258 * A generic message format for virtual PCI.
259 * Specific message formats are defined later in the file.
266 struct pci_child_message {
267 struct pci_message message_type;
268 union win_slot_encoding wslot;
271 struct pci_incoming_message {
272 struct vmpacket_descriptor hdr;
273 struct pci_message message_type;
276 struct pci_response {
277 struct vmpacket_descriptor hdr;
278 s32 status; /* negative values are failures */
282 void (*completion_func)(void *context, struct pci_response *resp,
283 int resp_packet_size);
286 struct pci_message message[];
290 * Specific message types supporting the PCI protocol.
294 * Version negotiation message. Sent from the guest to the host.
295 * The guest is free to try different versions until the host
296 * accepts the version.
298 * pci_version: The protocol version requested.
299 * is_last_attempt: If TRUE, this is the last version guest will request.
300 * reservedz: Reserved field, set to zero.
303 struct pci_version_request {
304 struct pci_message message_type;
305 u32 protocol_version;
309 * Bus D0 Entry. This is sent from the guest to the host when the virtual
310 * bus (PCI Express port) is ready for action.
313 struct pci_bus_d0_entry {
314 struct pci_message message_type;
319 struct pci_bus_relations {
320 struct pci_incoming_message incoming;
322 struct pci_function_description func[];
325 struct pci_bus_relations2 {
326 struct pci_incoming_message incoming;
328 struct pci_function_description2 func[];
331 struct pci_q_res_req_response {
332 struct vmpacket_descriptor hdr;
333 s32 status; /* negative values are failures */
334 u32 probed_bar[PCI_STD_NUM_BARS];
337 struct pci_set_power {
338 struct pci_message message_type;
339 union win_slot_encoding wslot;
340 u32 power_state; /* In Windows terms */
344 struct pci_set_power_response {
345 struct vmpacket_descriptor hdr;
346 s32 status; /* negative values are failures */
347 union win_slot_encoding wslot;
348 u32 resultant_state; /* In Windows terms */
352 struct pci_resources_assigned {
353 struct pci_message message_type;
354 union win_slot_encoding wslot;
355 u8 memory_range[0x14][6]; /* not used here */
360 struct pci_resources_assigned2 {
361 struct pci_message message_type;
362 union win_slot_encoding wslot;
363 u8 memory_range[0x14][6]; /* not used here */
364 u32 msi_descriptor_count;
368 struct pci_create_interrupt {
369 struct pci_message message_type;
370 union win_slot_encoding wslot;
371 struct hv_msi_desc int_desc;
374 struct pci_create_int_response {
375 struct pci_response response;
377 struct tran_int_desc int_desc;
380 struct pci_create_interrupt2 {
381 struct pci_message message_type;
382 union win_slot_encoding wslot;
383 struct hv_msi_desc2 int_desc;
386 struct pci_delete_interrupt {
387 struct pci_message message_type;
388 union win_slot_encoding wslot;
389 struct tran_int_desc int_desc;
393 * Note: the VM must pass a valid block id, wslot and bytes_requested.
395 struct pci_read_block {
396 struct pci_message message_type;
398 union win_slot_encoding wslot;
402 struct pci_read_block_response {
403 struct vmpacket_descriptor hdr;
405 u8 bytes[HV_CONFIG_BLOCK_SIZE_MAX];
409 * Note: the VM must pass a valid block id, wslot and byte_count.
411 struct pci_write_block {
412 struct pci_message message_type;
414 union win_slot_encoding wslot;
416 u8 bytes[HV_CONFIG_BLOCK_SIZE_MAX];
419 struct pci_dev_inval_block {
420 struct pci_incoming_message incoming;
421 union win_slot_encoding wslot;
425 struct pci_dev_incoming {
426 struct pci_incoming_message incoming;
427 union win_slot_encoding wslot;
430 struct pci_eject_response {
431 struct pci_message message_type;
432 union win_slot_encoding wslot;
436 static int pci_ring_size = (4 * PAGE_SIZE);
439 * Driver specific state.
442 enum hv_pcibus_state {
451 struct hv_pcibus_device {
452 struct pci_sysdata sysdata;
453 /* Protocol version negotiated with the host */
454 enum pci_protocol_version_t protocol_version;
455 enum hv_pcibus_state state;
456 refcount_t remove_lock;
457 struct hv_device *hdev;
458 resource_size_t low_mmio_space;
459 resource_size_t high_mmio_space;
460 struct resource *mem_config;
461 struct resource *low_mmio_res;
462 struct resource *high_mmio_res;
463 struct completion *survey_event;
464 struct completion remove_event;
465 struct pci_bus *pci_bus;
466 spinlock_t config_lock; /* Avoid two threads writing index page */
467 spinlock_t device_list_lock; /* Protect lists below */
468 void __iomem *cfg_addr;
470 struct list_head resources_for_children;
472 struct list_head children;
473 struct list_head dr_list;
475 struct msi_domain_info msi_info;
476 struct msi_controller msi_chip;
477 struct irq_domain *irq_domain;
479 spinlock_t retarget_msi_interrupt_lock;
481 struct workqueue_struct *wq;
483 /* Highest slot of child device with resources allocated */
484 int wslot_res_allocated;
486 /* hypercall arg, must not cross page boundary */
487 struct hv_retarget_device_interrupt retarget_msi_interrupt_params;
490 * Don't put anything here: retarget_msi_interrupt_params must be last
495 * Tracks "Device Relations" messages from the host, which must be both
496 * processed in order and deferred so that they don't run in the context
497 * of the incoming packet callback.
500 struct work_struct wrk;
501 struct hv_pcibus_device *bus;
504 struct hv_pcidev_description {
505 u16 v_id; /* vendor ID */
506 u16 d_id; /* device ID */
512 union win_slot_encoding win_slot;
513 u32 ser; /* serial number */
515 u16 virtual_numa_node;
519 struct list_head list_entry;
521 struct hv_pcidev_description func[];
524 enum hv_pcichild_state {
525 hv_pcichild_init = 0,
526 hv_pcichild_requirements,
527 hv_pcichild_resourced,
528 hv_pcichild_ejecting,
533 /* List protected by pci_rescan_remove_lock */
534 struct list_head list_entry;
536 enum hv_pcichild_state state;
537 struct pci_slot *pci_slot;
538 struct hv_pcidev_description desc;
539 bool reported_missing;
540 struct hv_pcibus_device *hbus;
541 struct work_struct wrk;
543 void (*block_invalidate)(void *context, u64 block_mask);
544 void *invalidate_context;
547 * What would be observed if one wrote 0xFFFFFFFF to a BAR and then
548 * read it back, for each of the BAR offsets within config space.
550 u32 probed_bar[PCI_STD_NUM_BARS];
553 struct hv_pci_compl {
554 struct completion host_event;
555 s32 completion_status;
558 static void hv_pci_onchannelcallback(void *context);
561 * hv_pci_generic_compl() - Invoked for a completion packet
562 * @context: Set up by the sender of the packet.
563 * @resp: The response packet
564 * @resp_packet_size: Size in bytes of the packet
566 * This function is used to trigger an event and report status
567 * for any message for which the completion packet contains a
568 * status and nothing else.
570 static void hv_pci_generic_compl(void *context, struct pci_response *resp,
571 int resp_packet_size)
573 struct hv_pci_compl *comp_pkt = context;
575 if (resp_packet_size >= offsetofend(struct pci_response, status))
576 comp_pkt->completion_status = resp->status;
578 comp_pkt->completion_status = -1;
580 complete(&comp_pkt->host_event);
583 static struct hv_pci_dev *get_pcichild_wslot(struct hv_pcibus_device *hbus,
586 static void get_pcichild(struct hv_pci_dev *hpdev)
588 refcount_inc(&hpdev->refs);
591 static void put_pcichild(struct hv_pci_dev *hpdev)
593 if (refcount_dec_and_test(&hpdev->refs))
597 static void get_hvpcibus(struct hv_pcibus_device *hv_pcibus);
598 static void put_hvpcibus(struct hv_pcibus_device *hv_pcibus);
601 * There is no good way to get notified from vmbus_onoffer_rescind(),
602 * so let's use polling here, since this is not a hot path.
604 static int wait_for_response(struct hv_device *hdev,
605 struct completion *comp)
608 if (hdev->channel->rescind) {
609 dev_warn_once(&hdev->device, "The device is gone.\n");
613 if (wait_for_completion_timeout(comp, HZ / 10))
621 * devfn_to_wslot() - Convert from Linux PCI slot to Windows
622 * @devfn: The Linux representation of PCI slot
624 * Windows uses a slightly different representation of PCI slot.
626 * Return: The Windows representation
628 static u32 devfn_to_wslot(int devfn)
630 union win_slot_encoding wslot;
633 wslot.bits.dev = PCI_SLOT(devfn);
634 wslot.bits.func = PCI_FUNC(devfn);
640 * wslot_to_devfn() - Convert from Windows PCI slot to Linux
641 * @wslot: The Windows representation of PCI slot
643 * Windows uses a slightly different representation of PCI slot.
645 * Return: The Linux representation
647 static int wslot_to_devfn(u32 wslot)
649 union win_slot_encoding slot_no;
651 slot_no.slot = wslot;
652 return PCI_DEVFN(slot_no.bits.dev, slot_no.bits.func);
656 * PCI Configuration Space for these root PCI buses is implemented as a pair
657 * of pages in memory-mapped I/O space. Writing to the first page chooses
658 * the PCI function being written or read. Once the first page has been
659 * written to, the following page maps in the entire configuration space of
664 * _hv_pcifront_read_config() - Internal PCI config read
665 * @hpdev: The PCI driver's representation of the device
666 * @where: Offset within config space
667 * @size: Size of the transfer
668 * @val: Pointer to the buffer receiving the data
670 static void _hv_pcifront_read_config(struct hv_pci_dev *hpdev, int where,
674 void __iomem *addr = hpdev->hbus->cfg_addr + CFG_PAGE_OFFSET + where;
677 * If the attempt is to read the IDs or the ROM BAR, simulate that.
679 if (where + size <= PCI_COMMAND) {
680 memcpy(val, ((u8 *)&hpdev->desc.v_id) + where, size);
681 } else if (where >= PCI_CLASS_REVISION && where + size <=
682 PCI_CACHE_LINE_SIZE) {
683 memcpy(val, ((u8 *)&hpdev->desc.rev) + where -
684 PCI_CLASS_REVISION, size);
685 } else if (where >= PCI_SUBSYSTEM_VENDOR_ID && where + size <=
687 memcpy(val, (u8 *)&hpdev->desc.subsystem_id + where -
688 PCI_SUBSYSTEM_VENDOR_ID, size);
689 } else if (where >= PCI_ROM_ADDRESS && where + size <=
690 PCI_CAPABILITY_LIST) {
691 /* ROM BARs are unimplemented */
693 } else if (where >= PCI_INTERRUPT_LINE && where + size <=
696 * Interrupt Line and Interrupt PIN are hard-wired to zero
697 * because this front-end only supports message-signaled
701 } else if (where + size <= CFG_PAGE_SIZE) {
702 spin_lock_irqsave(&hpdev->hbus->config_lock, flags);
703 /* Choose the function to be read. (See comment above) */
704 writel(hpdev->desc.win_slot.slot, hpdev->hbus->cfg_addr);
705 /* Make sure the function was chosen before we start reading. */
707 /* Read from that function's config space. */
720 * Make sure the read was done before we release the spinlock
721 * allowing consecutive reads/writes.
724 spin_unlock_irqrestore(&hpdev->hbus->config_lock, flags);
726 dev_err(&hpdev->hbus->hdev->device,
727 "Attempt to read beyond a function's config space.\n");
731 static u16 hv_pcifront_get_vendor_id(struct hv_pci_dev *hpdev)
735 void __iomem *addr = hpdev->hbus->cfg_addr + CFG_PAGE_OFFSET +
738 spin_lock_irqsave(&hpdev->hbus->config_lock, flags);
740 /* Choose the function to be read. (See comment above) */
741 writel(hpdev->desc.win_slot.slot, hpdev->hbus->cfg_addr);
742 /* Make sure the function was chosen before we start reading. */
744 /* Read from that function's config space. */
747 * mb() is not required here, because the spin_unlock_irqrestore()
751 spin_unlock_irqrestore(&hpdev->hbus->config_lock, flags);
757 * _hv_pcifront_write_config() - Internal PCI config write
758 * @hpdev: The PCI driver's representation of the device
759 * @where: Offset within config space
760 * @size: Size of the transfer
761 * @val: The data being transferred
763 static void _hv_pcifront_write_config(struct hv_pci_dev *hpdev, int where,
767 void __iomem *addr = hpdev->hbus->cfg_addr + CFG_PAGE_OFFSET + where;
769 if (where >= PCI_SUBSYSTEM_VENDOR_ID &&
770 where + size <= PCI_CAPABILITY_LIST) {
771 /* SSIDs and ROM BARs are read-only */
772 } else if (where >= PCI_COMMAND && where + size <= CFG_PAGE_SIZE) {
773 spin_lock_irqsave(&hpdev->hbus->config_lock, flags);
774 /* Choose the function to be written. (See comment above) */
775 writel(hpdev->desc.win_slot.slot, hpdev->hbus->cfg_addr);
776 /* Make sure the function was chosen before we start writing. */
778 /* Write to that function's config space. */
791 * Make sure the write was done before we release the spinlock
792 * allowing consecutive reads/writes.
795 spin_unlock_irqrestore(&hpdev->hbus->config_lock, flags);
797 dev_err(&hpdev->hbus->hdev->device,
798 "Attempt to write beyond a function's config space.\n");
803 * hv_pcifront_read_config() - Read configuration space
804 * @bus: PCI Bus structure
805 * @devfn: Device/function
806 * @where: Offset from base
807 * @size: Byte/word/dword
808 * @val: Value to be read
810 * Return: PCIBIOS_SUCCESSFUL on success
811 * PCIBIOS_DEVICE_NOT_FOUND on failure
813 static int hv_pcifront_read_config(struct pci_bus *bus, unsigned int devfn,
814 int where, int size, u32 *val)
816 struct hv_pcibus_device *hbus =
817 container_of(bus->sysdata, struct hv_pcibus_device, sysdata);
818 struct hv_pci_dev *hpdev;
820 hpdev = get_pcichild_wslot(hbus, devfn_to_wslot(devfn));
822 return PCIBIOS_DEVICE_NOT_FOUND;
824 _hv_pcifront_read_config(hpdev, where, size, val);
827 return PCIBIOS_SUCCESSFUL;
831 * hv_pcifront_write_config() - Write configuration space
832 * @bus: PCI Bus structure
833 * @devfn: Device/function
834 * @where: Offset from base
835 * @size: Byte/word/dword
836 * @val: Value to be written to device
838 * Return: PCIBIOS_SUCCESSFUL on success
839 * PCIBIOS_DEVICE_NOT_FOUND on failure
841 static int hv_pcifront_write_config(struct pci_bus *bus, unsigned int devfn,
842 int where, int size, u32 val)
844 struct hv_pcibus_device *hbus =
845 container_of(bus->sysdata, struct hv_pcibus_device, sysdata);
846 struct hv_pci_dev *hpdev;
848 hpdev = get_pcichild_wslot(hbus, devfn_to_wslot(devfn));
850 return PCIBIOS_DEVICE_NOT_FOUND;
852 _hv_pcifront_write_config(hpdev, where, size, val);
855 return PCIBIOS_SUCCESSFUL;
858 /* PCIe operations */
859 static struct pci_ops hv_pcifront_ops = {
860 .read = hv_pcifront_read_config,
861 .write = hv_pcifront_write_config,
865 * Paravirtual backchannel
867 * Hyper-V SR-IOV provides a backchannel mechanism in software for
868 * communication between a VF driver and a PF driver. These
869 * "configuration blocks" are similar in concept to PCI configuration space,
870 * but instead of doing reads and writes in 32-bit chunks through a very slow
871 * path, packets of up to 128 bytes can be sent or received asynchronously.
873 * Nearly every SR-IOV device contains just such a communications channel in
874 * hardware, so using this one in software is usually optional. Using the
875 * software channel, however, allows driver implementers to leverage software
876 * tools that fuzz the communications channel looking for vulnerabilities.
878 * The usage model for these packets puts the responsibility for reading or
879 * writing on the VF driver. The VF driver sends a read or a write packet,
880 * indicating which "block" is being referred to by number.
882 * If the PF driver wishes to initiate communication, it can "invalidate" one or
883 * more of the first 64 blocks. This invalidation is delivered via a callback
884 * supplied by the VF driver by this driver.
886 * No protocol is implied, except that supplied by the PF and VF drivers.
889 struct hv_read_config_compl {
890 struct hv_pci_compl comp_pkt;
893 unsigned int bytes_returned;
897 * hv_pci_read_config_compl() - Invoked when a response packet
898 * for a read config block operation arrives.
899 * @context: Identifies the read config operation
900 * @resp: The response packet itself
901 * @resp_packet_size: Size in bytes of the response packet
903 static void hv_pci_read_config_compl(void *context, struct pci_response *resp,
904 int resp_packet_size)
906 struct hv_read_config_compl *comp = context;
907 struct pci_read_block_response *read_resp =
908 (struct pci_read_block_response *)resp;
909 unsigned int data_len, hdr_len;
911 hdr_len = offsetof(struct pci_read_block_response, bytes);
912 if (resp_packet_size < hdr_len) {
913 comp->comp_pkt.completion_status = -1;
917 data_len = resp_packet_size - hdr_len;
918 if (data_len > 0 && read_resp->status == 0) {
919 comp->bytes_returned = min(comp->len, data_len);
920 memcpy(comp->buf, read_resp->bytes, comp->bytes_returned);
922 comp->bytes_returned = 0;
925 comp->comp_pkt.completion_status = read_resp->status;
927 complete(&comp->comp_pkt.host_event);
931 * hv_read_config_block() - Sends a read config block request to
932 * the back-end driver running in the Hyper-V parent partition.
933 * @pdev: The PCI driver's representation for this device.
934 * @buf: Buffer into which the config block will be copied.
935 * @len: Size in bytes of buf.
936 * @block_id: Identifies the config block which has been requested.
937 * @bytes_returned: Size which came back from the back-end driver.
939 * Return: 0 on success, -errno on failure
941 int hv_read_config_block(struct pci_dev *pdev, void *buf, unsigned int len,
942 unsigned int block_id, unsigned int *bytes_returned)
944 struct hv_pcibus_device *hbus =
945 container_of(pdev->bus->sysdata, struct hv_pcibus_device,
948 struct pci_packet pkt;
949 char buf[sizeof(struct pci_read_block)];
951 struct hv_read_config_compl comp_pkt;
952 struct pci_read_block *read_blk;
955 if (len == 0 || len > HV_CONFIG_BLOCK_SIZE_MAX)
958 init_completion(&comp_pkt.comp_pkt.host_event);
962 memset(&pkt, 0, sizeof(pkt));
963 pkt.pkt.completion_func = hv_pci_read_config_compl;
964 pkt.pkt.compl_ctxt = &comp_pkt;
965 read_blk = (struct pci_read_block *)&pkt.pkt.message;
966 read_blk->message_type.type = PCI_READ_BLOCK;
967 read_blk->wslot.slot = devfn_to_wslot(pdev->devfn);
968 read_blk->block_id = block_id;
969 read_blk->bytes_requested = len;
971 ret = vmbus_sendpacket(hbus->hdev->channel, read_blk,
972 sizeof(*read_blk), (unsigned long)&pkt.pkt,
974 VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
978 ret = wait_for_response(hbus->hdev, &comp_pkt.comp_pkt.host_event);
982 if (comp_pkt.comp_pkt.completion_status != 0 ||
983 comp_pkt.bytes_returned == 0) {
984 dev_err(&hbus->hdev->device,
985 "Read Config Block failed: 0x%x, bytes_returned=%d\n",
986 comp_pkt.comp_pkt.completion_status,
987 comp_pkt.bytes_returned);
991 *bytes_returned = comp_pkt.bytes_returned;
996 * hv_pci_write_config_compl() - Invoked when a response packet for a write
997 * config block operation arrives.
998 * @context: Identifies the write config operation
999 * @resp: The response packet itself
1000 * @resp_packet_size: Size in bytes of the response packet
1002 static void hv_pci_write_config_compl(void *context, struct pci_response *resp,
1003 int resp_packet_size)
1005 struct hv_pci_compl *comp_pkt = context;
1007 comp_pkt->completion_status = resp->status;
1008 complete(&comp_pkt->host_event);
1012 * hv_write_config_block() - Sends a write config block request to the
1013 * back-end driver running in the Hyper-V parent partition.
1014 * @pdev: The PCI driver's representation for this device.
1015 * @buf: Buffer from which the config block will be copied.
1016 * @len: Size in bytes of buf.
1017 * @block_id: Identifies the config block which is being written.
1019 * Return: 0 on success, -errno on failure
1021 int hv_write_config_block(struct pci_dev *pdev, void *buf, unsigned int len,
1022 unsigned int block_id)
1024 struct hv_pcibus_device *hbus =
1025 container_of(pdev->bus->sysdata, struct hv_pcibus_device,
1028 struct pci_packet pkt;
1029 char buf[sizeof(struct pci_write_block)];
1032 struct hv_pci_compl comp_pkt;
1033 struct pci_write_block *write_blk;
1037 if (len == 0 || len > HV_CONFIG_BLOCK_SIZE_MAX)
1040 init_completion(&comp_pkt.host_event);
1042 memset(&pkt, 0, sizeof(pkt));
1043 pkt.pkt.completion_func = hv_pci_write_config_compl;
1044 pkt.pkt.compl_ctxt = &comp_pkt;
1045 write_blk = (struct pci_write_block *)&pkt.pkt.message;
1046 write_blk->message_type.type = PCI_WRITE_BLOCK;
1047 write_blk->wslot.slot = devfn_to_wslot(pdev->devfn);
1048 write_blk->block_id = block_id;
1049 write_blk->byte_count = len;
1050 memcpy(write_blk->bytes, buf, len);
1051 pkt_size = offsetof(struct pci_write_block, bytes) + len;
1053 * This quirk is required on some hosts shipped around 2018, because
1054 * these hosts don't check the pkt_size correctly (new hosts have been
1055 * fixed since early 2019). The quirk is also safe on very old hosts
1056 * and new hosts, because, on them, what really matters is the length
1057 * specified in write_blk->byte_count.
1059 pkt_size += sizeof(pkt.reserved);
1061 ret = vmbus_sendpacket(hbus->hdev->channel, write_blk, pkt_size,
1062 (unsigned long)&pkt.pkt, VM_PKT_DATA_INBAND,
1063 VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
1067 ret = wait_for_response(hbus->hdev, &comp_pkt.host_event);
1071 if (comp_pkt.completion_status != 0) {
1072 dev_err(&hbus->hdev->device,
1073 "Write Config Block failed: 0x%x\n",
1074 comp_pkt.completion_status);
1082 * hv_register_block_invalidate() - Invoked when a config block invalidation
1083 * arrives from the back-end driver.
1084 * @pdev: The PCI driver's representation for this device.
1085 * @context: Identifies the device.
1086 * @block_invalidate: Identifies all of the blocks being invalidated.
1088 * Return: 0 on success, -errno on failure
1090 int hv_register_block_invalidate(struct pci_dev *pdev, void *context,
1091 void (*block_invalidate)(void *context,
1094 struct hv_pcibus_device *hbus =
1095 container_of(pdev->bus->sysdata, struct hv_pcibus_device,
1097 struct hv_pci_dev *hpdev;
1099 hpdev = get_pcichild_wslot(hbus, devfn_to_wslot(pdev->devfn));
1103 hpdev->block_invalidate = block_invalidate;
1104 hpdev->invalidate_context = context;
1106 put_pcichild(hpdev);
1111 /* Interrupt management hooks */
1112 static void hv_int_desc_free(struct hv_pci_dev *hpdev,
1113 struct tran_int_desc *int_desc)
1115 struct pci_delete_interrupt *int_pkt;
1117 struct pci_packet pkt;
1118 u8 buffer[sizeof(struct pci_delete_interrupt)];
1121 memset(&ctxt, 0, sizeof(ctxt));
1122 int_pkt = (struct pci_delete_interrupt *)&ctxt.pkt.message;
1123 int_pkt->message_type.type =
1124 PCI_DELETE_INTERRUPT_MESSAGE;
1125 int_pkt->wslot.slot = hpdev->desc.win_slot.slot;
1126 int_pkt->int_desc = *int_desc;
1127 vmbus_sendpacket(hpdev->hbus->hdev->channel, int_pkt, sizeof(*int_pkt),
1128 (unsigned long)&ctxt.pkt, VM_PKT_DATA_INBAND, 0);
1133 * hv_msi_free() - Free the MSI.
1134 * @domain: The interrupt domain pointer
1135 * @info: Extra MSI-related context
1136 * @irq: Identifies the IRQ.
1138 * The Hyper-V parent partition and hypervisor are tracking the
1139 * messages that are in use, keeping the interrupt redirection
1140 * table up to date. This callback sends a message that frees
1141 * the IRT entry and related tracking nonsense.
1143 static void hv_msi_free(struct irq_domain *domain, struct msi_domain_info *info,
1146 struct hv_pcibus_device *hbus;
1147 struct hv_pci_dev *hpdev;
1148 struct pci_dev *pdev;
1149 struct tran_int_desc *int_desc;
1150 struct irq_data *irq_data = irq_domain_get_irq_data(domain, irq);
1151 struct msi_desc *msi = irq_data_get_msi_desc(irq_data);
1153 pdev = msi_desc_to_pci_dev(msi);
1155 int_desc = irq_data_get_irq_chip_data(irq_data);
1159 irq_data->chip_data = NULL;
1160 hpdev = get_pcichild_wslot(hbus, devfn_to_wslot(pdev->devfn));
1166 hv_int_desc_free(hpdev, int_desc);
1167 put_pcichild(hpdev);
1170 static int hv_set_affinity(struct irq_data *data, const struct cpumask *dest,
1173 struct irq_data *parent = data->parent_data;
1175 return parent->chip->irq_set_affinity(parent, dest, force);
1178 static void hv_irq_mask(struct irq_data *data)
1180 pci_msi_mask_irq(data);
1184 * hv_irq_unmask() - "Unmask" the IRQ by setting its current
1186 * @data: Describes the IRQ
1188 * Build new a destination for the MSI and make a hypercall to
1189 * update the Interrupt Redirection Table. "Device Logical ID"
1190 * is built out of this PCI bus's instance GUID and the function
1191 * number of the device.
1193 static void hv_irq_unmask(struct irq_data *data)
1195 struct msi_desc *msi_desc = irq_data_get_msi_desc(data);
1196 struct irq_cfg *cfg = irqd_cfg(data);
1197 struct hv_retarget_device_interrupt *params;
1198 struct hv_pcibus_device *hbus;
1199 struct cpumask *dest;
1201 struct pci_bus *pbus;
1202 struct pci_dev *pdev;
1203 unsigned long flags;
1208 dest = irq_data_get_effective_affinity_mask(data);
1209 pdev = msi_desc_to_pci_dev(msi_desc);
1211 hbus = container_of(pbus->sysdata, struct hv_pcibus_device, sysdata);
1213 spin_lock_irqsave(&hbus->retarget_msi_interrupt_lock, flags);
1215 params = &hbus->retarget_msi_interrupt_params;
1216 memset(params, 0, sizeof(*params));
1217 params->partition_id = HV_PARTITION_ID_SELF;
1218 params->int_entry.source = 1; /* MSI(-X) */
1219 hv_set_msi_entry_from_desc(¶ms->int_entry.msi_entry, msi_desc);
1220 params->device_id = (hbus->hdev->dev_instance.b[5] << 24) |
1221 (hbus->hdev->dev_instance.b[4] << 16) |
1222 (hbus->hdev->dev_instance.b[7] << 8) |
1223 (hbus->hdev->dev_instance.b[6] & 0xf8) |
1224 PCI_FUNC(pdev->devfn);
1225 params->int_target.vector = cfg->vector;
1228 * Honoring apic->irq_delivery_mode set to dest_Fixed by
1229 * setting the HV_DEVICE_INTERRUPT_TARGET_MULTICAST flag results in a
1230 * spurious interrupt storm. Not doing so does not seem to have a
1231 * negative effect (yet?).
1234 if (hbus->protocol_version >= PCI_PROTOCOL_VERSION_1_2) {
1236 * PCI_PROTOCOL_VERSION_1_2 supports the VP_SET version of the
1237 * HVCALL_RETARGET_INTERRUPT hypercall, which also coincides
1238 * with >64 VP support.
1239 * ms_hyperv.hints & HV_X64_EX_PROCESSOR_MASKS_RECOMMENDED
1240 * is not sufficient for this hypercall.
1242 params->int_target.flags |=
1243 HV_DEVICE_INTERRUPT_TARGET_PROCESSOR_SET;
1245 if (!alloc_cpumask_var(&tmp, GFP_ATOMIC)) {
1250 cpumask_and(tmp, dest, cpu_online_mask);
1251 nr_bank = cpumask_to_vpset(¶ms->int_target.vp_set, tmp);
1252 free_cpumask_var(tmp);
1260 * var-sized hypercall, var-size starts after vp_mask (thus
1261 * vp_set.format does not count, but vp_set.valid_bank_mask
1264 var_size = 1 + nr_bank;
1266 for_each_cpu_and(cpu, dest, cpu_online_mask) {
1267 params->int_target.vp_mask |=
1268 (1ULL << hv_cpu_number_to_vp_number(cpu));
1272 res = hv_do_hypercall(HVCALL_RETARGET_INTERRUPT | (var_size << 17),
1276 spin_unlock_irqrestore(&hbus->retarget_msi_interrupt_lock, flags);
1279 dev_err(&hbus->hdev->device,
1280 "%s() failed: %#llx", __func__, res);
1284 pci_msi_unmask_irq(data);
1287 struct compose_comp_ctxt {
1288 struct hv_pci_compl comp_pkt;
1289 struct tran_int_desc int_desc;
1292 static void hv_pci_compose_compl(void *context, struct pci_response *resp,
1293 int resp_packet_size)
1295 struct compose_comp_ctxt *comp_pkt = context;
1296 struct pci_create_int_response *int_resp =
1297 (struct pci_create_int_response *)resp;
1299 comp_pkt->comp_pkt.completion_status = resp->status;
1300 comp_pkt->int_desc = int_resp->int_desc;
1301 complete(&comp_pkt->comp_pkt.host_event);
1304 static u32 hv_compose_msi_req_v1(
1305 struct pci_create_interrupt *int_pkt, struct cpumask *affinity,
1306 u32 slot, u8 vector)
1308 int_pkt->message_type.type = PCI_CREATE_INTERRUPT_MESSAGE;
1309 int_pkt->wslot.slot = slot;
1310 int_pkt->int_desc.vector = vector;
1311 int_pkt->int_desc.vector_count = 1;
1312 int_pkt->int_desc.delivery_mode = dest_Fixed;
1315 * Create MSI w/ dummy vCPU set, overwritten by subsequent retarget in
1318 int_pkt->int_desc.cpu_mask = CPU_AFFINITY_ALL;
1320 return sizeof(*int_pkt);
1323 static u32 hv_compose_msi_req_v2(
1324 struct pci_create_interrupt2 *int_pkt, struct cpumask *affinity,
1325 u32 slot, u8 vector)
1329 int_pkt->message_type.type = PCI_CREATE_INTERRUPT_MESSAGE2;
1330 int_pkt->wslot.slot = slot;
1331 int_pkt->int_desc.vector = vector;
1332 int_pkt->int_desc.vector_count = 1;
1333 int_pkt->int_desc.delivery_mode = dest_Fixed;
1336 * Create MSI w/ dummy vCPU set targeting just one vCPU, overwritten
1337 * by subsequent retarget in hv_irq_unmask().
1339 cpu = cpumask_first_and(affinity, cpu_online_mask);
1340 int_pkt->int_desc.processor_array[0] =
1341 hv_cpu_number_to_vp_number(cpu);
1342 int_pkt->int_desc.processor_count = 1;
1344 return sizeof(*int_pkt);
1348 * hv_compose_msi_msg() - Supplies a valid MSI address/data
1349 * @data: Everything about this MSI
1350 * @msg: Buffer that is filled in by this function
1352 * This function unpacks the IRQ looking for target CPU set, IDT
1353 * vector and mode and sends a message to the parent partition
1354 * asking for a mapping for that tuple in this partition. The
1355 * response supplies a data value and address to which that data
1356 * should be written to trigger that interrupt.
1358 static void hv_compose_msi_msg(struct irq_data *data, struct msi_msg *msg)
1360 struct irq_cfg *cfg = irqd_cfg(data);
1361 struct hv_pcibus_device *hbus;
1362 struct hv_pci_dev *hpdev;
1363 struct pci_bus *pbus;
1364 struct pci_dev *pdev;
1365 struct cpumask *dest;
1366 unsigned long flags;
1367 struct compose_comp_ctxt comp;
1368 struct tran_int_desc *int_desc;
1370 struct pci_packet pci_pkt;
1372 struct pci_create_interrupt v1;
1373 struct pci_create_interrupt2 v2;
1380 pdev = msi_desc_to_pci_dev(irq_data_get_msi_desc(data));
1381 dest = irq_data_get_effective_affinity_mask(data);
1383 hbus = container_of(pbus->sysdata, struct hv_pcibus_device, sysdata);
1384 hpdev = get_pcichild_wslot(hbus, devfn_to_wslot(pdev->devfn));
1386 goto return_null_message;
1388 /* Free any previous message that might have already been composed. */
1389 if (data->chip_data) {
1390 int_desc = data->chip_data;
1391 data->chip_data = NULL;
1392 hv_int_desc_free(hpdev, int_desc);
1395 int_desc = kzalloc(sizeof(*int_desc), GFP_ATOMIC);
1397 goto drop_reference;
1399 memset(&ctxt, 0, sizeof(ctxt));
1400 init_completion(&comp.comp_pkt.host_event);
1401 ctxt.pci_pkt.completion_func = hv_pci_compose_compl;
1402 ctxt.pci_pkt.compl_ctxt = ∁
1404 switch (hbus->protocol_version) {
1405 case PCI_PROTOCOL_VERSION_1_1:
1406 size = hv_compose_msi_req_v1(&ctxt.int_pkts.v1,
1408 hpdev->desc.win_slot.slot,
1412 case PCI_PROTOCOL_VERSION_1_2:
1413 case PCI_PROTOCOL_VERSION_1_3:
1414 size = hv_compose_msi_req_v2(&ctxt.int_pkts.v2,
1416 hpdev->desc.win_slot.slot,
1421 /* As we only negotiate protocol versions known to this driver,
1422 * this path should never hit. However, this is it not a hot
1423 * path so we print a message to aid future updates.
1425 dev_err(&hbus->hdev->device,
1426 "Unexpected vPCI protocol, update driver.");
1430 ret = vmbus_sendpacket(hpdev->hbus->hdev->channel, &ctxt.int_pkts,
1431 size, (unsigned long)&ctxt.pci_pkt,
1433 VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
1435 dev_err(&hbus->hdev->device,
1436 "Sending request for interrupt failed: 0x%x",
1437 comp.comp_pkt.completion_status);
1442 * Since this function is called with IRQ locks held, can't
1443 * do normal wait for completion; instead poll.
1445 while (!try_wait_for_completion(&comp.comp_pkt.host_event)) {
1446 /* 0xFFFF means an invalid PCI VENDOR ID. */
1447 if (hv_pcifront_get_vendor_id(hpdev) == 0xFFFF) {
1448 dev_err_once(&hbus->hdev->device,
1449 "the device has gone\n");
1454 * When the higher level interrupt code calls us with
1455 * interrupt disabled, we must poll the channel by calling
1456 * the channel callback directly when channel->target_cpu is
1457 * the current CPU. When the higher level interrupt code
1458 * calls us with interrupt enabled, let's add the
1459 * local_irq_save()/restore() to avoid race:
1460 * hv_pci_onchannelcallback() can also run in tasklet.
1462 local_irq_save(flags);
1464 if (hbus->hdev->channel->target_cpu == smp_processor_id())
1465 hv_pci_onchannelcallback(hbus);
1467 local_irq_restore(flags);
1469 if (hpdev->state == hv_pcichild_ejecting) {
1470 dev_err_once(&hbus->hdev->device,
1471 "the device is being ejected\n");
1478 if (comp.comp_pkt.completion_status < 0) {
1479 dev_err(&hbus->hdev->device,
1480 "Request for interrupt failed: 0x%x",
1481 comp.comp_pkt.completion_status);
1486 * Record the assignment so that this can be unwound later. Using
1487 * irq_set_chip_data() here would be appropriate, but the lock it takes
1490 *int_desc = comp.int_desc;
1491 data->chip_data = int_desc;
1493 /* Pass up the result. */
1494 msg->address_hi = comp.int_desc.address >> 32;
1495 msg->address_lo = comp.int_desc.address & 0xffffffff;
1496 msg->data = comp.int_desc.data;
1498 put_pcichild(hpdev);
1504 put_pcichild(hpdev);
1505 return_null_message:
1506 msg->address_hi = 0;
1507 msg->address_lo = 0;
1511 /* HW Interrupt Chip Descriptor */
1512 static struct irq_chip hv_msi_irq_chip = {
1513 .name = "Hyper-V PCIe MSI",
1514 .irq_compose_msi_msg = hv_compose_msi_msg,
1515 .irq_set_affinity = hv_set_affinity,
1516 .irq_ack = irq_chip_ack_parent,
1517 .irq_mask = hv_irq_mask,
1518 .irq_unmask = hv_irq_unmask,
1521 static irq_hw_number_t hv_msi_domain_ops_get_hwirq(struct msi_domain_info *info,
1522 msi_alloc_info_t *arg)
1524 return arg->msi_hwirq;
1527 static struct msi_domain_ops hv_msi_ops = {
1528 .get_hwirq = hv_msi_domain_ops_get_hwirq,
1529 .msi_prepare = pci_msi_prepare,
1530 .set_desc = pci_msi_set_desc,
1531 .msi_free = hv_msi_free,
1535 * hv_pcie_init_irq_domain() - Initialize IRQ domain
1536 * @hbus: The root PCI bus
1538 * This function creates an IRQ domain which will be used for
1539 * interrupts from devices that have been passed through. These
1540 * devices only support MSI and MSI-X, not line-based interrupts
1541 * or simulations of line-based interrupts through PCIe's
1542 * fabric-layer messages. Because interrupts are remapped, we
1543 * can support multi-message MSI here.
1545 * Return: '0' on success and error value on failure
1547 static int hv_pcie_init_irq_domain(struct hv_pcibus_device *hbus)
1549 hbus->msi_info.chip = &hv_msi_irq_chip;
1550 hbus->msi_info.ops = &hv_msi_ops;
1551 hbus->msi_info.flags = (MSI_FLAG_USE_DEF_DOM_OPS |
1552 MSI_FLAG_USE_DEF_CHIP_OPS | MSI_FLAG_MULTI_PCI_MSI |
1554 hbus->msi_info.handler = handle_edge_irq;
1555 hbus->msi_info.handler_name = "edge";
1556 hbus->msi_info.data = hbus;
1557 hbus->irq_domain = pci_msi_create_irq_domain(hbus->sysdata.fwnode,
1560 if (!hbus->irq_domain) {
1561 dev_err(&hbus->hdev->device,
1562 "Failed to build an MSI IRQ domain\n");
1570 * get_bar_size() - Get the address space consumed by a BAR
1571 * @bar_val: Value that a BAR returned after -1 was written
1574 * This function returns the size of the BAR, rounded up to 1
1575 * page. It has to be rounded up because the hypervisor's page
1576 * table entry that maps the BAR into the VM can't specify an
1577 * offset within a page. The invariant is that the hypervisor
1578 * must place any BARs of smaller than page length at the
1579 * beginning of a page.
1581 * Return: Size in bytes of the consumed MMIO space.
1583 static u64 get_bar_size(u64 bar_val)
1585 return round_up((1 + ~(bar_val & PCI_BASE_ADDRESS_MEM_MASK)),
1590 * survey_child_resources() - Total all MMIO requirements
1591 * @hbus: Root PCI bus, as understood by this driver
1593 static void survey_child_resources(struct hv_pcibus_device *hbus)
1595 struct hv_pci_dev *hpdev;
1596 resource_size_t bar_size = 0;
1597 unsigned long flags;
1598 struct completion *event;
1602 /* If nobody is waiting on the answer, don't compute it. */
1603 event = xchg(&hbus->survey_event, NULL);
1607 /* If the answer has already been computed, go with it. */
1608 if (hbus->low_mmio_space || hbus->high_mmio_space) {
1613 spin_lock_irqsave(&hbus->device_list_lock, flags);
1616 * Due to an interesting quirk of the PCI spec, all memory regions
1617 * for a child device are a power of 2 in size and aligned in memory,
1618 * so it's sufficient to just add them up without tracking alignment.
1620 list_for_each_entry(hpdev, &hbus->children, list_entry) {
1621 for (i = 0; i < PCI_STD_NUM_BARS; i++) {
1622 if (hpdev->probed_bar[i] & PCI_BASE_ADDRESS_SPACE_IO)
1623 dev_err(&hbus->hdev->device,
1624 "There's an I/O BAR in this list!\n");
1626 if (hpdev->probed_bar[i] != 0) {
1628 * A probed BAR has all the upper bits set that
1632 bar_val = hpdev->probed_bar[i];
1633 if (bar_val & PCI_BASE_ADDRESS_MEM_TYPE_64)
1635 ((u64)hpdev->probed_bar[++i] << 32);
1637 bar_val |= 0xffffffff00000000ULL;
1639 bar_size = get_bar_size(bar_val);
1641 if (bar_val & PCI_BASE_ADDRESS_MEM_TYPE_64)
1642 hbus->high_mmio_space += bar_size;
1644 hbus->low_mmio_space += bar_size;
1649 spin_unlock_irqrestore(&hbus->device_list_lock, flags);
1654 * prepopulate_bars() - Fill in BARs with defaults
1655 * @hbus: Root PCI bus, as understood by this driver
1657 * The core PCI driver code seems much, much happier if the BARs
1658 * for a device have values upon first scan. So fill them in.
1659 * The algorithm below works down from large sizes to small,
1660 * attempting to pack the assignments optimally. The assumption,
1661 * enforced in other parts of the code, is that the beginning of
1662 * the memory-mapped I/O space will be aligned on the largest
1665 static void prepopulate_bars(struct hv_pcibus_device *hbus)
1667 resource_size_t high_size = 0;
1668 resource_size_t low_size = 0;
1669 resource_size_t high_base = 0;
1670 resource_size_t low_base = 0;
1671 resource_size_t bar_size;
1672 struct hv_pci_dev *hpdev;
1673 unsigned long flags;
1679 if (hbus->low_mmio_space) {
1680 low_size = 1ULL << (63 - __builtin_clzll(hbus->low_mmio_space));
1681 low_base = hbus->low_mmio_res->start;
1684 if (hbus->high_mmio_space) {
1686 (63 - __builtin_clzll(hbus->high_mmio_space));
1687 high_base = hbus->high_mmio_res->start;
1690 spin_lock_irqsave(&hbus->device_list_lock, flags);
1693 * Clear the memory enable bit, in case it's already set. This occurs
1694 * in the suspend path of hibernation, where the device is suspended,
1695 * resumed and suspended again: see hibernation_snapshot() and
1696 * hibernation_platform_enter().
1698 * If the memory enable bit is already set, Hyper-V sliently ignores
1699 * the below BAR updates, and the related PCI device driver can not
1700 * work, because reading from the device register(s) always returns
1703 list_for_each_entry(hpdev, &hbus->children, list_entry) {
1704 _hv_pcifront_read_config(hpdev, PCI_COMMAND, 2, &command);
1705 command &= ~PCI_COMMAND_MEMORY;
1706 _hv_pcifront_write_config(hpdev, PCI_COMMAND, 2, command);
1709 /* Pick addresses for the BARs. */
1711 list_for_each_entry(hpdev, &hbus->children, list_entry) {
1712 for (i = 0; i < PCI_STD_NUM_BARS; i++) {
1713 bar_val = hpdev->probed_bar[i];
1716 high = bar_val & PCI_BASE_ADDRESS_MEM_TYPE_64;
1719 ((u64)hpdev->probed_bar[i + 1]
1722 bar_val |= 0xffffffffULL << 32;
1724 bar_size = get_bar_size(bar_val);
1726 if (high_size != bar_size) {
1730 _hv_pcifront_write_config(hpdev,
1731 PCI_BASE_ADDRESS_0 + (4 * i),
1733 (u32)(high_base & 0xffffff00));
1735 _hv_pcifront_write_config(hpdev,
1736 PCI_BASE_ADDRESS_0 + (4 * i),
1737 4, (u32)(high_base >> 32));
1738 high_base += bar_size;
1740 if (low_size != bar_size)
1742 _hv_pcifront_write_config(hpdev,
1743 PCI_BASE_ADDRESS_0 + (4 * i),
1745 (u32)(low_base & 0xffffff00));
1746 low_base += bar_size;
1749 if (high_size <= 1 && low_size <= 1) {
1750 /* Set the memory enable bit. */
1751 _hv_pcifront_read_config(hpdev, PCI_COMMAND, 2,
1753 command |= PCI_COMMAND_MEMORY;
1754 _hv_pcifront_write_config(hpdev, PCI_COMMAND, 2,
1762 } while (high_size || low_size);
1764 spin_unlock_irqrestore(&hbus->device_list_lock, flags);
1768 * Assign entries in sysfs pci slot directory.
1770 * Note that this function does not need to lock the children list
1771 * because it is called from pci_devices_present_work which
1772 * is serialized with hv_eject_device_work because they are on the
1773 * same ordered workqueue. Therefore hbus->children list will not change
1774 * even when pci_create_slot sleeps.
1776 static void hv_pci_assign_slots(struct hv_pcibus_device *hbus)
1778 struct hv_pci_dev *hpdev;
1779 char name[SLOT_NAME_SIZE];
1782 list_for_each_entry(hpdev, &hbus->children, list_entry) {
1783 if (hpdev->pci_slot)
1786 slot_nr = PCI_SLOT(wslot_to_devfn(hpdev->desc.win_slot.slot));
1787 snprintf(name, SLOT_NAME_SIZE, "%u", hpdev->desc.ser);
1788 hpdev->pci_slot = pci_create_slot(hbus->pci_bus, slot_nr,
1790 if (IS_ERR(hpdev->pci_slot)) {
1791 pr_warn("pci_create slot %s failed\n", name);
1792 hpdev->pci_slot = NULL;
1798 * Remove entries in sysfs pci slot directory.
1800 static void hv_pci_remove_slots(struct hv_pcibus_device *hbus)
1802 struct hv_pci_dev *hpdev;
1804 list_for_each_entry(hpdev, &hbus->children, list_entry) {
1805 if (!hpdev->pci_slot)
1807 pci_destroy_slot(hpdev->pci_slot);
1808 hpdev->pci_slot = NULL;
1813 * Set NUMA node for the devices on the bus
1815 static void hv_pci_assign_numa_node(struct hv_pcibus_device *hbus)
1817 struct pci_dev *dev;
1818 struct pci_bus *bus = hbus->pci_bus;
1819 struct hv_pci_dev *hv_dev;
1821 list_for_each_entry(dev, &bus->devices, bus_list) {
1822 hv_dev = get_pcichild_wslot(hbus, devfn_to_wslot(dev->devfn));
1826 if (hv_dev->desc.flags & HV_PCI_DEVICE_FLAG_NUMA_AFFINITY)
1827 set_dev_node(&dev->dev, hv_dev->desc.virtual_numa_node);
1829 put_pcichild(hv_dev);
1834 * create_root_hv_pci_bus() - Expose a new root PCI bus
1835 * @hbus: Root PCI bus, as understood by this driver
1837 * Return: 0 on success, -errno on failure
1839 static int create_root_hv_pci_bus(struct hv_pcibus_device *hbus)
1841 /* Register the device */
1842 hbus->pci_bus = pci_create_root_bus(&hbus->hdev->device,
1843 0, /* bus number is always zero */
1846 &hbus->resources_for_children);
1850 hbus->pci_bus->msi = &hbus->msi_chip;
1851 hbus->pci_bus->msi->dev = &hbus->hdev->device;
1853 pci_lock_rescan_remove();
1854 pci_scan_child_bus(hbus->pci_bus);
1855 hv_pci_assign_numa_node(hbus);
1856 pci_bus_assign_resources(hbus->pci_bus);
1857 hv_pci_assign_slots(hbus);
1858 pci_bus_add_devices(hbus->pci_bus);
1859 pci_unlock_rescan_remove();
1860 hbus->state = hv_pcibus_installed;
1864 struct q_res_req_compl {
1865 struct completion host_event;
1866 struct hv_pci_dev *hpdev;
1870 * q_resource_requirements() - Query Resource Requirements
1871 * @context: The completion context.
1872 * @resp: The response that came from the host.
1873 * @resp_packet_size: The size in bytes of resp.
1875 * This function is invoked on completion of a Query Resource
1876 * Requirements packet.
1878 static void q_resource_requirements(void *context, struct pci_response *resp,
1879 int resp_packet_size)
1881 struct q_res_req_compl *completion = context;
1882 struct pci_q_res_req_response *q_res_req =
1883 (struct pci_q_res_req_response *)resp;
1886 if (resp->status < 0) {
1887 dev_err(&completion->hpdev->hbus->hdev->device,
1888 "query resource requirements failed: %x\n",
1891 for (i = 0; i < PCI_STD_NUM_BARS; i++) {
1892 completion->hpdev->probed_bar[i] =
1893 q_res_req->probed_bar[i];
1897 complete(&completion->host_event);
1901 * new_pcichild_device() - Create a new child device
1902 * @hbus: The internal struct tracking this root PCI bus.
1903 * @desc: The information supplied so far from the host
1906 * This function creates the tracking structure for a new child
1907 * device and kicks off the process of figuring out what it is.
1909 * Return: Pointer to the new tracking struct
1911 static struct hv_pci_dev *new_pcichild_device(struct hv_pcibus_device *hbus,
1912 struct hv_pcidev_description *desc)
1914 struct hv_pci_dev *hpdev;
1915 struct pci_child_message *res_req;
1916 struct q_res_req_compl comp_pkt;
1918 struct pci_packet init_packet;
1919 u8 buffer[sizeof(struct pci_child_message)];
1921 unsigned long flags;
1924 hpdev = kzalloc(sizeof(*hpdev), GFP_KERNEL);
1930 memset(&pkt, 0, sizeof(pkt));
1931 init_completion(&comp_pkt.host_event);
1932 comp_pkt.hpdev = hpdev;
1933 pkt.init_packet.compl_ctxt = &comp_pkt;
1934 pkt.init_packet.completion_func = q_resource_requirements;
1935 res_req = (struct pci_child_message *)&pkt.init_packet.message;
1936 res_req->message_type.type = PCI_QUERY_RESOURCE_REQUIREMENTS;
1937 res_req->wslot.slot = desc->win_slot.slot;
1939 ret = vmbus_sendpacket(hbus->hdev->channel, res_req,
1940 sizeof(struct pci_child_message),
1941 (unsigned long)&pkt.init_packet,
1943 VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
1947 if (wait_for_response(hbus->hdev, &comp_pkt.host_event))
1950 hpdev->desc = *desc;
1951 refcount_set(&hpdev->refs, 1);
1952 get_pcichild(hpdev);
1953 spin_lock_irqsave(&hbus->device_list_lock, flags);
1955 list_add_tail(&hpdev->list_entry, &hbus->children);
1956 spin_unlock_irqrestore(&hbus->device_list_lock, flags);
1965 * get_pcichild_wslot() - Find device from slot
1966 * @hbus: Root PCI bus, as understood by this driver
1967 * @wslot: Location on the bus
1969 * This function looks up a PCI device and returns the internal
1970 * representation of it. It acquires a reference on it, so that
1971 * the device won't be deleted while somebody is using it. The
1972 * caller is responsible for calling put_pcichild() to release
1975 * Return: Internal representation of a PCI device
1977 static struct hv_pci_dev *get_pcichild_wslot(struct hv_pcibus_device *hbus,
1980 unsigned long flags;
1981 struct hv_pci_dev *iter, *hpdev = NULL;
1983 spin_lock_irqsave(&hbus->device_list_lock, flags);
1984 list_for_each_entry(iter, &hbus->children, list_entry) {
1985 if (iter->desc.win_slot.slot == wslot) {
1987 get_pcichild(hpdev);
1991 spin_unlock_irqrestore(&hbus->device_list_lock, flags);
1997 * pci_devices_present_work() - Handle new list of child devices
1998 * @work: Work struct embedded in struct hv_dr_work
2000 * "Bus Relations" is the Windows term for "children of this
2001 * bus." The terminology is preserved here for people trying to
2002 * debug the interaction between Hyper-V and Linux. This
2003 * function is called when the parent partition reports a list
2004 * of functions that should be observed under this PCI Express
2007 * This function updates the list, and must tolerate being
2008 * called multiple times with the same information. The typical
2009 * number of child devices is one, with very atypical cases
2010 * involving three or four, so the algorithms used here can be
2011 * simple and inefficient.
2013 * It must also treat the omission of a previously observed device as
2014 * notification that the device no longer exists.
2016 * Note that this function is serialized with hv_eject_device_work(),
2017 * because both are pushed to the ordered workqueue hbus->wq.
2019 static void pci_devices_present_work(struct work_struct *work)
2023 struct hv_pcidev_description *new_desc;
2024 struct hv_pci_dev *hpdev;
2025 struct hv_pcibus_device *hbus;
2026 struct list_head removed;
2027 struct hv_dr_work *dr_wrk;
2028 struct hv_dr_state *dr = NULL;
2029 unsigned long flags;
2031 dr_wrk = container_of(work, struct hv_dr_work, wrk);
2035 INIT_LIST_HEAD(&removed);
2037 /* Pull this off the queue and process it if it was the last one. */
2038 spin_lock_irqsave(&hbus->device_list_lock, flags);
2039 while (!list_empty(&hbus->dr_list)) {
2040 dr = list_first_entry(&hbus->dr_list, struct hv_dr_state,
2042 list_del(&dr->list_entry);
2044 /* Throw this away if the list still has stuff in it. */
2045 if (!list_empty(&hbus->dr_list)) {
2050 spin_unlock_irqrestore(&hbus->device_list_lock, flags);
2057 /* First, mark all existing children as reported missing. */
2058 spin_lock_irqsave(&hbus->device_list_lock, flags);
2059 list_for_each_entry(hpdev, &hbus->children, list_entry) {
2060 hpdev->reported_missing = true;
2062 spin_unlock_irqrestore(&hbus->device_list_lock, flags);
2064 /* Next, add back any reported devices. */
2065 for (child_no = 0; child_no < dr->device_count; child_no++) {
2067 new_desc = &dr->func[child_no];
2069 spin_lock_irqsave(&hbus->device_list_lock, flags);
2070 list_for_each_entry(hpdev, &hbus->children, list_entry) {
2071 if ((hpdev->desc.win_slot.slot == new_desc->win_slot.slot) &&
2072 (hpdev->desc.v_id == new_desc->v_id) &&
2073 (hpdev->desc.d_id == new_desc->d_id) &&
2074 (hpdev->desc.ser == new_desc->ser)) {
2075 hpdev->reported_missing = false;
2079 spin_unlock_irqrestore(&hbus->device_list_lock, flags);
2082 hpdev = new_pcichild_device(hbus, new_desc);
2084 dev_err(&hbus->hdev->device,
2085 "couldn't record a child device.\n");
2089 /* Move missing children to a list on the stack. */
2090 spin_lock_irqsave(&hbus->device_list_lock, flags);
2093 list_for_each_entry(hpdev, &hbus->children, list_entry) {
2094 if (hpdev->reported_missing) {
2096 put_pcichild(hpdev);
2097 list_move_tail(&hpdev->list_entry, &removed);
2102 spin_unlock_irqrestore(&hbus->device_list_lock, flags);
2104 /* Delete everything that should no longer exist. */
2105 while (!list_empty(&removed)) {
2106 hpdev = list_first_entry(&removed, struct hv_pci_dev,
2108 list_del(&hpdev->list_entry);
2110 if (hpdev->pci_slot)
2111 pci_destroy_slot(hpdev->pci_slot);
2113 put_pcichild(hpdev);
2116 switch (hbus->state) {
2117 case hv_pcibus_installed:
2119 * Tell the core to rescan bus
2120 * because there may have been changes.
2122 pci_lock_rescan_remove();
2123 pci_scan_child_bus(hbus->pci_bus);
2124 hv_pci_assign_numa_node(hbus);
2125 hv_pci_assign_slots(hbus);
2126 pci_unlock_rescan_remove();
2129 case hv_pcibus_init:
2130 case hv_pcibus_probed:
2131 survey_child_resources(hbus);
2143 * hv_pci_start_relations_work() - Queue work to start device discovery
2144 * @hbus: Root PCI bus, as understood by this driver
2145 * @dr: The list of children returned from host
2147 * Return: 0 on success, -errno on failure
2149 static int hv_pci_start_relations_work(struct hv_pcibus_device *hbus,
2150 struct hv_dr_state *dr)
2152 struct hv_dr_work *dr_wrk;
2153 unsigned long flags;
2156 if (hbus->state == hv_pcibus_removing) {
2157 dev_info(&hbus->hdev->device,
2158 "PCI VMBus BUS_RELATIONS: ignored\n");
2162 dr_wrk = kzalloc(sizeof(*dr_wrk), GFP_NOWAIT);
2166 INIT_WORK(&dr_wrk->wrk, pci_devices_present_work);
2169 spin_lock_irqsave(&hbus->device_list_lock, flags);
2171 * If pending_dr is true, we have already queued a work,
2172 * which will see the new dr. Otherwise, we need to
2175 pending_dr = !list_empty(&hbus->dr_list);
2176 list_add_tail(&dr->list_entry, &hbus->dr_list);
2177 spin_unlock_irqrestore(&hbus->device_list_lock, flags);
2183 queue_work(hbus->wq, &dr_wrk->wrk);
2190 * hv_pci_devices_present() - Handle list of new children
2191 * @hbus: Root PCI bus, as understood by this driver
2192 * @relations: Packet from host listing children
2194 * Process a new list of devices on the bus. The list of devices is
2195 * discovered by VSP and sent to us via VSP message PCI_BUS_RELATIONS,
2196 * whenever a new list of devices for this bus appears.
2198 static void hv_pci_devices_present(struct hv_pcibus_device *hbus,
2199 struct pci_bus_relations *relations)
2201 struct hv_dr_state *dr;
2204 dr = kzalloc(struct_size(dr, func, relations->device_count),
2209 dr->device_count = relations->device_count;
2210 for (i = 0; i < dr->device_count; i++) {
2211 dr->func[i].v_id = relations->func[i].v_id;
2212 dr->func[i].d_id = relations->func[i].d_id;
2213 dr->func[i].rev = relations->func[i].rev;
2214 dr->func[i].prog_intf = relations->func[i].prog_intf;
2215 dr->func[i].subclass = relations->func[i].subclass;
2216 dr->func[i].base_class = relations->func[i].base_class;
2217 dr->func[i].subsystem_id = relations->func[i].subsystem_id;
2218 dr->func[i].win_slot = relations->func[i].win_slot;
2219 dr->func[i].ser = relations->func[i].ser;
2222 if (hv_pci_start_relations_work(hbus, dr))
2227 * hv_pci_devices_present2() - Handle list of new children
2228 * @hbus: Root PCI bus, as understood by this driver
2229 * @relations: Packet from host listing children
2231 * This function is the v2 version of hv_pci_devices_present()
2233 static void hv_pci_devices_present2(struct hv_pcibus_device *hbus,
2234 struct pci_bus_relations2 *relations)
2236 struct hv_dr_state *dr;
2239 dr = kzalloc(struct_size(dr, func, relations->device_count),
2244 dr->device_count = relations->device_count;
2245 for (i = 0; i < dr->device_count; i++) {
2246 dr->func[i].v_id = relations->func[i].v_id;
2247 dr->func[i].d_id = relations->func[i].d_id;
2248 dr->func[i].rev = relations->func[i].rev;
2249 dr->func[i].prog_intf = relations->func[i].prog_intf;
2250 dr->func[i].subclass = relations->func[i].subclass;
2251 dr->func[i].base_class = relations->func[i].base_class;
2252 dr->func[i].subsystem_id = relations->func[i].subsystem_id;
2253 dr->func[i].win_slot = relations->func[i].win_slot;
2254 dr->func[i].ser = relations->func[i].ser;
2255 dr->func[i].flags = relations->func[i].flags;
2256 dr->func[i].virtual_numa_node =
2257 relations->func[i].virtual_numa_node;
2260 if (hv_pci_start_relations_work(hbus, dr))
2265 * hv_eject_device_work() - Asynchronously handles ejection
2266 * @work: Work struct embedded in internal device struct
2268 * This function handles ejecting a device. Windows will
2269 * attempt to gracefully eject a device, waiting 60 seconds to
2270 * hear back from the guest OS that this completed successfully.
2271 * If this timer expires, the device will be forcibly removed.
2273 static void hv_eject_device_work(struct work_struct *work)
2275 struct pci_eject_response *ejct_pkt;
2276 struct hv_pcibus_device *hbus;
2277 struct hv_pci_dev *hpdev;
2278 struct pci_dev *pdev;
2279 unsigned long flags;
2282 struct pci_packet pkt;
2283 u8 buffer[sizeof(struct pci_eject_response)];
2286 hpdev = container_of(work, struct hv_pci_dev, wrk);
2289 WARN_ON(hpdev->state != hv_pcichild_ejecting);
2292 * Ejection can come before or after the PCI bus has been set up, so
2293 * attempt to find it and tear down the bus state, if it exists. This
2294 * must be done without constructs like pci_domain_nr(hbus->pci_bus)
2295 * because hbus->pci_bus may not exist yet.
2297 wslot = wslot_to_devfn(hpdev->desc.win_slot.slot);
2298 pdev = pci_get_domain_bus_and_slot(hbus->sysdata.domain, 0, wslot);
2300 pci_lock_rescan_remove();
2301 pci_stop_and_remove_bus_device(pdev);
2303 pci_unlock_rescan_remove();
2306 spin_lock_irqsave(&hbus->device_list_lock, flags);
2307 list_del(&hpdev->list_entry);
2308 spin_unlock_irqrestore(&hbus->device_list_lock, flags);
2310 if (hpdev->pci_slot)
2311 pci_destroy_slot(hpdev->pci_slot);
2313 memset(&ctxt, 0, sizeof(ctxt));
2314 ejct_pkt = (struct pci_eject_response *)&ctxt.pkt.message;
2315 ejct_pkt->message_type.type = PCI_EJECTION_COMPLETE;
2316 ejct_pkt->wslot.slot = hpdev->desc.win_slot.slot;
2317 vmbus_sendpacket(hbus->hdev->channel, ejct_pkt,
2318 sizeof(*ejct_pkt), (unsigned long)&ctxt.pkt,
2319 VM_PKT_DATA_INBAND, 0);
2321 /* For the get_pcichild() in hv_pci_eject_device() */
2322 put_pcichild(hpdev);
2323 /* For the two refs got in new_pcichild_device() */
2324 put_pcichild(hpdev);
2325 put_pcichild(hpdev);
2326 /* hpdev has been freed. Do not use it any more. */
2332 * hv_pci_eject_device() - Handles device ejection
2333 * @hpdev: Internal device tracking struct
2335 * This function is invoked when an ejection packet arrives. It
2336 * just schedules work so that we don't re-enter the packet
2337 * delivery code handling the ejection.
2339 static void hv_pci_eject_device(struct hv_pci_dev *hpdev)
2341 struct hv_pcibus_device *hbus = hpdev->hbus;
2342 struct hv_device *hdev = hbus->hdev;
2344 if (hbus->state == hv_pcibus_removing) {
2345 dev_info(&hdev->device, "PCI VMBus EJECT: ignored\n");
2349 hpdev->state = hv_pcichild_ejecting;
2350 get_pcichild(hpdev);
2351 INIT_WORK(&hpdev->wrk, hv_eject_device_work);
2353 queue_work(hbus->wq, &hpdev->wrk);
2357 * hv_pci_onchannelcallback() - Handles incoming packets
2358 * @context: Internal bus tracking struct
2360 * This function is invoked whenever the host sends a packet to
2361 * this channel (which is private to this root PCI bus).
2363 static void hv_pci_onchannelcallback(void *context)
2365 const int packet_size = 0x100;
2367 struct hv_pcibus_device *hbus = context;
2370 struct vmpacket_descriptor *desc;
2371 unsigned char *buffer;
2372 int bufferlen = packet_size;
2373 struct pci_packet *comp_packet;
2374 struct pci_response *response;
2375 struct pci_incoming_message *new_message;
2376 struct pci_bus_relations *bus_rel;
2377 struct pci_bus_relations2 *bus_rel2;
2378 struct pci_dev_inval_block *inval;
2379 struct pci_dev_incoming *dev_message;
2380 struct hv_pci_dev *hpdev;
2382 buffer = kmalloc(bufferlen, GFP_ATOMIC);
2387 ret = vmbus_recvpacket_raw(hbus->hdev->channel, buffer,
2388 bufferlen, &bytes_recvd, &req_id);
2390 if (ret == -ENOBUFS) {
2392 /* Handle large packet */
2393 bufferlen = bytes_recvd;
2394 buffer = kmalloc(bytes_recvd, GFP_ATOMIC);
2400 /* Zero length indicates there are no more packets. */
2401 if (ret || !bytes_recvd)
2405 * All incoming packets must be at least as large as a
2408 if (bytes_recvd <= sizeof(struct pci_response))
2410 desc = (struct vmpacket_descriptor *)buffer;
2412 switch (desc->type) {
2416 * The host is trusted, and thus it's safe to interpret
2417 * this transaction ID as a pointer.
2419 comp_packet = (struct pci_packet *)req_id;
2420 response = (struct pci_response *)buffer;
2421 comp_packet->completion_func(comp_packet->compl_ctxt,
2426 case VM_PKT_DATA_INBAND:
2428 new_message = (struct pci_incoming_message *)buffer;
2429 switch (new_message->message_type.type) {
2430 case PCI_BUS_RELATIONS:
2432 bus_rel = (struct pci_bus_relations *)buffer;
2434 struct_size(bus_rel, func,
2435 bus_rel->device_count)) {
2436 dev_err(&hbus->hdev->device,
2437 "bus relations too small\n");
2441 hv_pci_devices_present(hbus, bus_rel);
2444 case PCI_BUS_RELATIONS2:
2446 bus_rel2 = (struct pci_bus_relations2 *)buffer;
2448 struct_size(bus_rel2, func,
2449 bus_rel2->device_count)) {
2450 dev_err(&hbus->hdev->device,
2451 "bus relations v2 too small\n");
2455 hv_pci_devices_present2(hbus, bus_rel2);
2460 dev_message = (struct pci_dev_incoming *)buffer;
2461 hpdev = get_pcichild_wslot(hbus,
2462 dev_message->wslot.slot);
2464 hv_pci_eject_device(hpdev);
2465 put_pcichild(hpdev);
2469 case PCI_INVALIDATE_BLOCK:
2471 inval = (struct pci_dev_inval_block *)buffer;
2472 hpdev = get_pcichild_wslot(hbus,
2475 if (hpdev->block_invalidate) {
2476 hpdev->block_invalidate(
2477 hpdev->invalidate_context,
2480 put_pcichild(hpdev);
2485 dev_warn(&hbus->hdev->device,
2486 "Unimplemented protocol message %x\n",
2487 new_message->message_type.type);
2493 dev_err(&hbus->hdev->device,
2494 "unhandled packet type %d, tid %llx len %d\n",
2495 desc->type, req_id, bytes_recvd);
2504 * hv_pci_protocol_negotiation() - Set up protocol
2505 * @hdev: VMBus's tracking struct for this root PCI bus
2507 * This driver is intended to support running on Windows 10
2508 * (server) and later versions. It will not run on earlier
2509 * versions, as they assume that many of the operations which
2510 * Linux needs accomplished with a spinlock held were done via
2511 * asynchronous messaging via VMBus. Windows 10 increases the
2512 * surface area of PCI emulation so that these actions can take
2513 * place by suspending a virtual processor for their duration.
2515 * This function negotiates the channel protocol version,
2516 * failing if the host doesn't support the necessary protocol
2519 static int hv_pci_protocol_negotiation(struct hv_device *hdev,
2520 enum pci_protocol_version_t version[],
2523 struct hv_pcibus_device *hbus = hv_get_drvdata(hdev);
2524 struct pci_version_request *version_req;
2525 struct hv_pci_compl comp_pkt;
2526 struct pci_packet *pkt;
2531 * Initiate the handshake with the host and negotiate
2532 * a version that the host can support. We start with the
2533 * highest version number and go down if the host cannot
2536 pkt = kzalloc(sizeof(*pkt) + sizeof(*version_req), GFP_KERNEL);
2540 init_completion(&comp_pkt.host_event);
2541 pkt->completion_func = hv_pci_generic_compl;
2542 pkt->compl_ctxt = &comp_pkt;
2543 version_req = (struct pci_version_request *)&pkt->message;
2544 version_req->message_type.type = PCI_QUERY_PROTOCOL_VERSION;
2546 for (i = 0; i < num_version; i++) {
2547 version_req->protocol_version = version[i];
2548 ret = vmbus_sendpacket(hdev->channel, version_req,
2549 sizeof(struct pci_version_request),
2550 (unsigned long)pkt, VM_PKT_DATA_INBAND,
2551 VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
2553 ret = wait_for_response(hdev, &comp_pkt.host_event);
2556 dev_err(&hdev->device,
2557 "PCI Pass-through VSP failed to request version: %d",
2562 if (comp_pkt.completion_status >= 0) {
2563 hbus->protocol_version = version[i];
2564 dev_info(&hdev->device,
2565 "PCI VMBus probing: Using version %#x\n",
2566 hbus->protocol_version);
2570 if (comp_pkt.completion_status != STATUS_REVISION_MISMATCH) {
2571 dev_err(&hdev->device,
2572 "PCI Pass-through VSP failed version request: %#x",
2573 comp_pkt.completion_status);
2578 reinit_completion(&comp_pkt.host_event);
2581 dev_err(&hdev->device,
2582 "PCI pass-through VSP failed to find supported version");
2591 * hv_pci_free_bridge_windows() - Release memory regions for the
2593 * @hbus: Root PCI bus, as understood by this driver
2595 static void hv_pci_free_bridge_windows(struct hv_pcibus_device *hbus)
2598 * Set the resources back to the way they looked when they
2599 * were allocated by setting IORESOURCE_BUSY again.
2602 if (hbus->low_mmio_space && hbus->low_mmio_res) {
2603 hbus->low_mmio_res->flags |= IORESOURCE_BUSY;
2604 vmbus_free_mmio(hbus->low_mmio_res->start,
2605 resource_size(hbus->low_mmio_res));
2608 if (hbus->high_mmio_space && hbus->high_mmio_res) {
2609 hbus->high_mmio_res->flags |= IORESOURCE_BUSY;
2610 vmbus_free_mmio(hbus->high_mmio_res->start,
2611 resource_size(hbus->high_mmio_res));
2616 * hv_pci_allocate_bridge_windows() - Allocate memory regions
2618 * @hbus: Root PCI bus, as understood by this driver
2620 * This function calls vmbus_allocate_mmio(), which is itself a
2621 * bit of a compromise. Ideally, we might change the pnp layer
2622 * in the kernel such that it comprehends either PCI devices
2623 * which are "grandchildren of ACPI," with some intermediate bus
2624 * node (in this case, VMBus) or change it such that it
2625 * understands VMBus. The pnp layer, however, has been declared
2626 * deprecated, and not subject to change.
2628 * The workaround, implemented here, is to ask VMBus to allocate
2629 * MMIO space for this bus. VMBus itself knows which ranges are
2630 * appropriate by looking at its own ACPI objects. Then, after
2631 * these ranges are claimed, they're modified to look like they
2632 * would have looked if the ACPI and pnp code had allocated
2633 * bridge windows. These descriptors have to exist in this form
2634 * in order to satisfy the code which will get invoked when the
2635 * endpoint PCI function driver calls request_mem_region() or
2636 * request_mem_region_exclusive().
2638 * Return: 0 on success, -errno on failure
2640 static int hv_pci_allocate_bridge_windows(struct hv_pcibus_device *hbus)
2642 resource_size_t align;
2645 if (hbus->low_mmio_space) {
2646 align = 1ULL << (63 - __builtin_clzll(hbus->low_mmio_space));
2647 ret = vmbus_allocate_mmio(&hbus->low_mmio_res, hbus->hdev, 0,
2648 (u64)(u32)0xffffffff,
2649 hbus->low_mmio_space,
2652 dev_err(&hbus->hdev->device,
2653 "Need %#llx of low MMIO space. Consider reconfiguring the VM.\n",
2654 hbus->low_mmio_space);
2658 /* Modify this resource to become a bridge window. */
2659 hbus->low_mmio_res->flags |= IORESOURCE_WINDOW;
2660 hbus->low_mmio_res->flags &= ~IORESOURCE_BUSY;
2661 pci_add_resource(&hbus->resources_for_children,
2662 hbus->low_mmio_res);
2665 if (hbus->high_mmio_space) {
2666 align = 1ULL << (63 - __builtin_clzll(hbus->high_mmio_space));
2667 ret = vmbus_allocate_mmio(&hbus->high_mmio_res, hbus->hdev,
2669 hbus->high_mmio_space, align,
2672 dev_err(&hbus->hdev->device,
2673 "Need %#llx of high MMIO space. Consider reconfiguring the VM.\n",
2674 hbus->high_mmio_space);
2675 goto release_low_mmio;
2678 /* Modify this resource to become a bridge window. */
2679 hbus->high_mmio_res->flags |= IORESOURCE_WINDOW;
2680 hbus->high_mmio_res->flags &= ~IORESOURCE_BUSY;
2681 pci_add_resource(&hbus->resources_for_children,
2682 hbus->high_mmio_res);
2688 if (hbus->low_mmio_res) {
2689 vmbus_free_mmio(hbus->low_mmio_res->start,
2690 resource_size(hbus->low_mmio_res));
2697 * hv_allocate_config_window() - Find MMIO space for PCI Config
2698 * @hbus: Root PCI bus, as understood by this driver
2700 * This function claims memory-mapped I/O space for accessing
2701 * configuration space for the functions on this bus.
2703 * Return: 0 on success, -errno on failure
2705 static int hv_allocate_config_window(struct hv_pcibus_device *hbus)
2710 * Set up a region of MMIO space to use for accessing configuration
2713 ret = vmbus_allocate_mmio(&hbus->mem_config, hbus->hdev, 0, -1,
2714 PCI_CONFIG_MMIO_LENGTH, 0x1000, false);
2719 * vmbus_allocate_mmio() gets used for allocating both device endpoint
2720 * resource claims (those which cannot be overlapped) and the ranges
2721 * which are valid for the children of this bus, which are intended
2722 * to be overlapped by those children. Set the flag on this claim
2723 * meaning that this region can't be overlapped.
2726 hbus->mem_config->flags |= IORESOURCE_BUSY;
2731 static void hv_free_config_window(struct hv_pcibus_device *hbus)
2733 vmbus_free_mmio(hbus->mem_config->start, PCI_CONFIG_MMIO_LENGTH);
2736 static int hv_pci_bus_exit(struct hv_device *hdev, bool keep_devs);
2739 * hv_pci_enter_d0() - Bring the "bus" into the D0 power state
2740 * @hdev: VMBus's tracking struct for this root PCI bus
2742 * Return: 0 on success, -errno on failure
2744 static int hv_pci_enter_d0(struct hv_device *hdev)
2746 struct hv_pcibus_device *hbus = hv_get_drvdata(hdev);
2747 struct pci_bus_d0_entry *d0_entry;
2748 struct hv_pci_compl comp_pkt;
2749 struct pci_packet *pkt;
2755 * Tell the host that the bus is ready to use, and moved into the
2756 * powered-on state. This includes telling the host which region
2757 * of memory-mapped I/O space has been chosen for configuration space
2760 pkt = kzalloc(sizeof(*pkt) + sizeof(*d0_entry), GFP_KERNEL);
2764 init_completion(&comp_pkt.host_event);
2765 pkt->completion_func = hv_pci_generic_compl;
2766 pkt->compl_ctxt = &comp_pkt;
2767 d0_entry = (struct pci_bus_d0_entry *)&pkt->message;
2768 d0_entry->message_type.type = PCI_BUS_D0ENTRY;
2769 d0_entry->mmio_base = hbus->mem_config->start;
2771 ret = vmbus_sendpacket(hdev->channel, d0_entry, sizeof(*d0_entry),
2772 (unsigned long)pkt, VM_PKT_DATA_INBAND,
2773 VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
2775 ret = wait_for_response(hdev, &comp_pkt.host_event);
2781 * In certain case (Kdump) the pci device of interest was
2782 * not cleanly shut down and resource is still held on host
2783 * side, the host could return invalid device status.
2784 * We need to explicitly request host to release the resource
2785 * and try to enter D0 again.
2787 if (comp_pkt.completion_status < 0 && retry) {
2790 dev_err(&hdev->device, "Retrying D0 Entry\n");
2793 * Hv_pci_bus_exit() calls hv_send_resource_released()
2794 * to free up resources of its child devices.
2795 * In the kdump kernel we need to set the
2796 * wslot_res_allocated to 255 so it scans all child
2797 * devices to release resources allocated in the
2798 * normal kernel before panic happened.
2800 hbus->wslot_res_allocated = 255;
2802 ret = hv_pci_bus_exit(hdev, true);
2806 goto enter_d0_retry;
2808 dev_err(&hdev->device,
2809 "Retrying D0 failed with ret %d\n", ret);
2812 if (comp_pkt.completion_status < 0) {
2813 dev_err(&hdev->device,
2814 "PCI Pass-through VSP failed D0 Entry with status %x\n",
2815 comp_pkt.completion_status);
2828 * hv_pci_query_relations() - Ask host to send list of child
2830 * @hdev: VMBus's tracking struct for this root PCI bus
2832 * Return: 0 on success, -errno on failure
2834 static int hv_pci_query_relations(struct hv_device *hdev)
2836 struct hv_pcibus_device *hbus = hv_get_drvdata(hdev);
2837 struct pci_message message;
2838 struct completion comp;
2841 /* Ask the host to send along the list of child devices */
2842 init_completion(&comp);
2843 if (cmpxchg(&hbus->survey_event, NULL, &comp))
2846 memset(&message, 0, sizeof(message));
2847 message.type = PCI_QUERY_BUS_RELATIONS;
2849 ret = vmbus_sendpacket(hdev->channel, &message, sizeof(message),
2850 0, VM_PKT_DATA_INBAND, 0);
2852 ret = wait_for_response(hdev, &comp);
2858 * hv_send_resources_allocated() - Report local resource choices
2859 * @hdev: VMBus's tracking struct for this root PCI bus
2861 * The host OS is expecting to be sent a request as a message
2862 * which contains all the resources that the device will use.
2863 * The response contains those same resources, "translated"
2864 * which is to say, the values which should be used by the
2865 * hardware, when it delivers an interrupt. (MMIO resources are
2866 * used in local terms.) This is nice for Windows, and lines up
2867 * with the FDO/PDO split, which doesn't exist in Linux. Linux
2868 * is deeply expecting to scan an emulated PCI configuration
2869 * space. So this message is sent here only to drive the state
2870 * machine on the host forward.
2872 * Return: 0 on success, -errno on failure
2874 static int hv_send_resources_allocated(struct hv_device *hdev)
2876 struct hv_pcibus_device *hbus = hv_get_drvdata(hdev);
2877 struct pci_resources_assigned *res_assigned;
2878 struct pci_resources_assigned2 *res_assigned2;
2879 struct hv_pci_compl comp_pkt;
2880 struct hv_pci_dev *hpdev;
2881 struct pci_packet *pkt;
2886 size_res = (hbus->protocol_version < PCI_PROTOCOL_VERSION_1_2)
2887 ? sizeof(*res_assigned) : sizeof(*res_assigned2);
2889 pkt = kmalloc(sizeof(*pkt) + size_res, GFP_KERNEL);
2895 for (wslot = 0; wslot < 256; wslot++) {
2896 hpdev = get_pcichild_wslot(hbus, wslot);
2900 memset(pkt, 0, sizeof(*pkt) + size_res);
2901 init_completion(&comp_pkt.host_event);
2902 pkt->completion_func = hv_pci_generic_compl;
2903 pkt->compl_ctxt = &comp_pkt;
2905 if (hbus->protocol_version < PCI_PROTOCOL_VERSION_1_2) {
2907 (struct pci_resources_assigned *)&pkt->message;
2908 res_assigned->message_type.type =
2909 PCI_RESOURCES_ASSIGNED;
2910 res_assigned->wslot.slot = hpdev->desc.win_slot.slot;
2913 (struct pci_resources_assigned2 *)&pkt->message;
2914 res_assigned2->message_type.type =
2915 PCI_RESOURCES_ASSIGNED2;
2916 res_assigned2->wslot.slot = hpdev->desc.win_slot.slot;
2918 put_pcichild(hpdev);
2920 ret = vmbus_sendpacket(hdev->channel, &pkt->message,
2921 size_res, (unsigned long)pkt,
2923 VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
2925 ret = wait_for_response(hdev, &comp_pkt.host_event);
2929 if (comp_pkt.completion_status < 0) {
2931 dev_err(&hdev->device,
2932 "resource allocated returned 0x%x",
2933 comp_pkt.completion_status);
2937 hbus->wslot_res_allocated = wslot;
2945 * hv_send_resources_released() - Report local resources
2947 * @hdev: VMBus's tracking struct for this root PCI bus
2949 * Return: 0 on success, -errno on failure
2951 static int hv_send_resources_released(struct hv_device *hdev)
2953 struct hv_pcibus_device *hbus = hv_get_drvdata(hdev);
2954 struct pci_child_message pkt;
2955 struct hv_pci_dev *hpdev;
2959 for (wslot = hbus->wslot_res_allocated; wslot >= 0; wslot--) {
2960 hpdev = get_pcichild_wslot(hbus, wslot);
2964 memset(&pkt, 0, sizeof(pkt));
2965 pkt.message_type.type = PCI_RESOURCES_RELEASED;
2966 pkt.wslot.slot = hpdev->desc.win_slot.slot;
2968 put_pcichild(hpdev);
2970 ret = vmbus_sendpacket(hdev->channel, &pkt, sizeof(pkt), 0,
2971 VM_PKT_DATA_INBAND, 0);
2975 hbus->wslot_res_allocated = wslot - 1;
2978 hbus->wslot_res_allocated = -1;
2983 static void get_hvpcibus(struct hv_pcibus_device *hbus)
2985 refcount_inc(&hbus->remove_lock);
2988 static void put_hvpcibus(struct hv_pcibus_device *hbus)
2990 if (refcount_dec_and_test(&hbus->remove_lock))
2991 complete(&hbus->remove_event);
2994 #define HVPCI_DOM_MAP_SIZE (64 * 1024)
2995 static DECLARE_BITMAP(hvpci_dom_map, HVPCI_DOM_MAP_SIZE);
2998 * PCI domain number 0 is used by emulated devices on Gen1 VMs, so define 0
2999 * as invalid for passthrough PCI devices of this driver.
3001 #define HVPCI_DOM_INVALID 0
3004 * hv_get_dom_num() - Get a valid PCI domain number
3005 * Check if the PCI domain number is in use, and return another number if
3008 * @dom: Requested domain number
3010 * return: domain number on success, HVPCI_DOM_INVALID on failure
3012 static u16 hv_get_dom_num(u16 dom)
3016 if (test_and_set_bit(dom, hvpci_dom_map) == 0)
3019 for_each_clear_bit(i, hvpci_dom_map, HVPCI_DOM_MAP_SIZE) {
3020 if (test_and_set_bit(i, hvpci_dom_map) == 0)
3024 return HVPCI_DOM_INVALID;
3028 * hv_put_dom_num() - Mark the PCI domain number as free
3029 * @dom: Domain number to be freed
3031 static void hv_put_dom_num(u16 dom)
3033 clear_bit(dom, hvpci_dom_map);
3037 * hv_pci_probe() - New VMBus channel probe, for a root PCI bus
3038 * @hdev: VMBus's tracking struct for this root PCI bus
3039 * @dev_id: Identifies the device itself
3041 * Return: 0 on success, -errno on failure
3043 static int hv_pci_probe(struct hv_device *hdev,
3044 const struct hv_vmbus_device_id *dev_id)
3046 struct hv_pcibus_device *hbus;
3052 * hv_pcibus_device contains the hypercall arguments for retargeting in
3053 * hv_irq_unmask(). Those must not cross a page boundary.
3055 BUILD_BUG_ON(sizeof(*hbus) > HV_HYP_PAGE_SIZE);
3058 * With the recent 59bb47985c1d ("mm, sl[aou]b: guarantee natural
3059 * alignment for kmalloc(power-of-two)"), kzalloc() is able to allocate
3060 * a 4KB buffer that is guaranteed to be 4KB-aligned. Here the size and
3061 * alignment of hbus is important because hbus's field
3062 * retarget_msi_interrupt_params must not cross a 4KB page boundary.
3064 * Here we prefer kzalloc to get_zeroed_page(), because a buffer
3065 * allocated by the latter is not tracked and scanned by kmemleak, and
3066 * hence kmemleak reports the pointer contained in the hbus buffer
3067 * (i.e. the hpdev struct, which is created in new_pcichild_device() and
3068 * is tracked by hbus->children) as memory leak (false positive).
3070 * If the kernel doesn't have 59bb47985c1d, get_zeroed_page() *must* be
3071 * used to allocate the hbus buffer and we can avoid the kmemleak false
3072 * positive by using kmemleak_alloc() and kmemleak_free() to ask
3073 * kmemleak to track and scan the hbus buffer.
3075 hbus = kzalloc(HV_HYP_PAGE_SIZE, GFP_KERNEL);
3078 hbus->state = hv_pcibus_init;
3079 hbus->wslot_res_allocated = -1;
3082 * The PCI bus "domain" is what is called "segment" in ACPI and other
3083 * specs. Pull it from the instance ID, to get something usually
3084 * unique. In rare cases of collision, we will find out another number
3087 * Note that, since this code only runs in a Hyper-V VM, Hyper-V
3088 * together with this guest driver can guarantee that (1) The only
3089 * domain used by Gen1 VMs for something that looks like a physical
3090 * PCI bus (which is actually emulated by the hypervisor) is domain 0.
3091 * (2) There will be no overlap between domains (after fixing possible
3092 * collisions) in the same VM.
3094 dom_req = hdev->dev_instance.b[5] << 8 | hdev->dev_instance.b[4];
3095 dom = hv_get_dom_num(dom_req);
3097 if (dom == HVPCI_DOM_INVALID) {
3098 dev_err(&hdev->device,
3099 "Unable to use dom# 0x%hx or other numbers", dom_req);
3105 dev_info(&hdev->device,
3106 "PCI dom# 0x%hx has collision, using 0x%hx",
3109 hbus->sysdata.domain = dom;
3112 refcount_set(&hbus->remove_lock, 1);
3113 INIT_LIST_HEAD(&hbus->children);
3114 INIT_LIST_HEAD(&hbus->dr_list);
3115 INIT_LIST_HEAD(&hbus->resources_for_children);
3116 spin_lock_init(&hbus->config_lock);
3117 spin_lock_init(&hbus->device_list_lock);
3118 spin_lock_init(&hbus->retarget_msi_interrupt_lock);
3119 init_completion(&hbus->remove_event);
3120 hbus->wq = alloc_ordered_workqueue("hv_pci_%x", 0,
3121 hbus->sysdata.domain);
3127 ret = vmbus_open(hdev->channel, pci_ring_size, pci_ring_size, NULL, 0,
3128 hv_pci_onchannelcallback, hbus);
3132 hv_set_drvdata(hdev, hbus);
3134 ret = hv_pci_protocol_negotiation(hdev, pci_protocol_versions,
3135 ARRAY_SIZE(pci_protocol_versions));
3139 ret = hv_allocate_config_window(hbus);
3143 hbus->cfg_addr = ioremap(hbus->mem_config->start,
3144 PCI_CONFIG_MMIO_LENGTH);
3145 if (!hbus->cfg_addr) {
3146 dev_err(&hdev->device,
3147 "Unable to map a virtual address for config space\n");
3152 name = kasprintf(GFP_KERNEL, "%pUL", &hdev->dev_instance);
3158 hbus->sysdata.fwnode = irq_domain_alloc_named_fwnode(name);
3160 if (!hbus->sysdata.fwnode) {
3165 ret = hv_pcie_init_irq_domain(hbus);
3169 ret = hv_pci_query_relations(hdev);
3171 goto free_irq_domain;
3173 ret = hv_pci_enter_d0(hdev);
3175 goto free_irq_domain;
3177 ret = hv_pci_allocate_bridge_windows(hbus);
3181 ret = hv_send_resources_allocated(hdev);
3185 prepopulate_bars(hbus);
3187 hbus->state = hv_pcibus_probed;
3189 ret = create_root_hv_pci_bus(hbus);
3196 hv_pci_free_bridge_windows(hbus);
3198 (void) hv_pci_bus_exit(hdev, true);
3200 irq_domain_remove(hbus->irq_domain);
3202 irq_domain_free_fwnode(hbus->sysdata.fwnode);
3204 iounmap(hbus->cfg_addr);
3206 hv_free_config_window(hbus);
3208 vmbus_close(hdev->channel);
3210 destroy_workqueue(hbus->wq);
3212 hv_put_dom_num(hbus->sysdata.domain);
3218 static int hv_pci_bus_exit(struct hv_device *hdev, bool keep_devs)
3220 struct hv_pcibus_device *hbus = hv_get_drvdata(hdev);
3222 struct pci_packet teardown_packet;
3223 u8 buffer[sizeof(struct pci_message)];
3225 struct hv_dr_state *dr;
3226 struct hv_pci_compl comp_pkt;
3230 * After the host sends the RESCIND_CHANNEL message, it doesn't
3231 * access the per-channel ringbuffer any longer.
3233 if (hdev->channel->rescind)
3237 /* Delete any children which might still exist. */
3238 dr = kzalloc(sizeof(*dr), GFP_KERNEL);
3239 if (dr && hv_pci_start_relations_work(hbus, dr))
3243 ret = hv_send_resources_released(hdev);
3245 dev_err(&hdev->device,
3246 "Couldn't send resources released packet(s)\n");
3250 memset(&pkt.teardown_packet, 0, sizeof(pkt.teardown_packet));
3251 init_completion(&comp_pkt.host_event);
3252 pkt.teardown_packet.completion_func = hv_pci_generic_compl;
3253 pkt.teardown_packet.compl_ctxt = &comp_pkt;
3254 pkt.teardown_packet.message[0].type = PCI_BUS_D0EXIT;
3256 ret = vmbus_sendpacket(hdev->channel, &pkt.teardown_packet.message,
3257 sizeof(struct pci_message),
3258 (unsigned long)&pkt.teardown_packet,
3260 VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
3264 if (wait_for_completion_timeout(&comp_pkt.host_event, 10 * HZ) == 0)
3271 * hv_pci_remove() - Remove routine for this VMBus channel
3272 * @hdev: VMBus's tracking struct for this root PCI bus
3274 * Return: 0 on success, -errno on failure
3276 static int hv_pci_remove(struct hv_device *hdev)
3278 struct hv_pcibus_device *hbus;
3281 hbus = hv_get_drvdata(hdev);
3282 if (hbus->state == hv_pcibus_installed) {
3283 /* Remove the bus from PCI's point of view. */
3284 pci_lock_rescan_remove();
3285 pci_stop_root_bus(hbus->pci_bus);
3286 hv_pci_remove_slots(hbus);
3287 pci_remove_root_bus(hbus->pci_bus);
3288 pci_unlock_rescan_remove();
3289 hbus->state = hv_pcibus_removed;
3292 ret = hv_pci_bus_exit(hdev, false);
3294 vmbus_close(hdev->channel);
3296 iounmap(hbus->cfg_addr);
3297 hv_free_config_window(hbus);
3298 pci_free_resource_list(&hbus->resources_for_children);
3299 hv_pci_free_bridge_windows(hbus);
3300 irq_domain_remove(hbus->irq_domain);
3301 irq_domain_free_fwnode(hbus->sysdata.fwnode);
3303 wait_for_completion(&hbus->remove_event);
3304 destroy_workqueue(hbus->wq);
3306 hv_put_dom_num(hbus->sysdata.domain);
3312 static int hv_pci_suspend(struct hv_device *hdev)
3314 struct hv_pcibus_device *hbus = hv_get_drvdata(hdev);
3315 enum hv_pcibus_state old_state;
3319 * hv_pci_suspend() must make sure there are no pending work items
3320 * before calling vmbus_close(), since it runs in a process context
3321 * as a callback in dpm_suspend(). When it starts to run, the channel
3322 * callback hv_pci_onchannelcallback(), which runs in a tasklet
3323 * context, can be still running concurrently and scheduling new work
3324 * items onto hbus->wq in hv_pci_devices_present() and
3325 * hv_pci_eject_device(), and the work item handlers can access the
3326 * vmbus channel, which can be being closed by hv_pci_suspend(), e.g.
3327 * the work item handler pci_devices_present_work() ->
3328 * new_pcichild_device() writes to the vmbus channel.
3330 * To eliminate the race, hv_pci_suspend() disables the channel
3331 * callback tasklet, sets hbus->state to hv_pcibus_removing, and
3332 * re-enables the tasklet. This way, when hv_pci_suspend() proceeds,
3333 * it knows that no new work item can be scheduled, and then it flushes
3334 * hbus->wq and safely closes the vmbus channel.
3336 tasklet_disable(&hdev->channel->callback_event);
3338 /* Change the hbus state to prevent new work items. */
3339 old_state = hbus->state;
3340 if (hbus->state == hv_pcibus_installed)
3341 hbus->state = hv_pcibus_removing;
3343 tasklet_enable(&hdev->channel->callback_event);
3345 if (old_state != hv_pcibus_installed)
3348 flush_workqueue(hbus->wq);
3350 ret = hv_pci_bus_exit(hdev, true);
3354 vmbus_close(hdev->channel);
3359 static int hv_pci_resume(struct hv_device *hdev)
3361 struct hv_pcibus_device *hbus = hv_get_drvdata(hdev);
3362 enum pci_protocol_version_t version[1];
3365 hbus->state = hv_pcibus_init;
3367 ret = vmbus_open(hdev->channel, pci_ring_size, pci_ring_size, NULL, 0,
3368 hv_pci_onchannelcallback, hbus);
3372 /* Only use the version that was in use before hibernation. */
3373 version[0] = hbus->protocol_version;
3374 ret = hv_pci_protocol_negotiation(hdev, version, 1);
3378 ret = hv_pci_query_relations(hdev);
3382 ret = hv_pci_enter_d0(hdev);
3386 ret = hv_send_resources_allocated(hdev);
3390 prepopulate_bars(hbus);
3392 hbus->state = hv_pcibus_installed;
3395 vmbus_close(hdev->channel);
3399 static const struct hv_vmbus_device_id hv_pci_id_table[] = {
3400 /* PCI Pass-through Class ID */
3401 /* 44C4F61D-4444-4400-9D52-802E27EDE19F */
3406 MODULE_DEVICE_TABLE(vmbus, hv_pci_id_table);
3408 static struct hv_driver hv_pci_drv = {
3410 .id_table = hv_pci_id_table,
3411 .probe = hv_pci_probe,
3412 .remove = hv_pci_remove,
3413 .suspend = hv_pci_suspend,
3414 .resume = hv_pci_resume,
3417 static void __exit exit_hv_pci_drv(void)
3419 vmbus_driver_unregister(&hv_pci_drv);
3421 hvpci_block_ops.read_block = NULL;
3422 hvpci_block_ops.write_block = NULL;
3423 hvpci_block_ops.reg_blk_invalidate = NULL;
3426 static int __init init_hv_pci_drv(void)
3428 /* Set the invalid domain number's bit, so it will not be used */
3429 set_bit(HVPCI_DOM_INVALID, hvpci_dom_map);
3431 /* Initialize PCI block r/w interface */
3432 hvpci_block_ops.read_block = hv_read_config_block;
3433 hvpci_block_ops.write_block = hv_write_config_block;
3434 hvpci_block_ops.reg_blk_invalidate = hv_register_block_invalidate;
3436 return vmbus_driver_register(&hv_pci_drv);
3439 module_init(init_hv_pci_drv);
3440 module_exit(exit_hv_pci_drv);
3442 MODULE_DESCRIPTION("Hyper-V PCI");
3443 MODULE_LICENSE("GPL v2");