1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (c) Microsoft Corporation.
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 */
68 #define CPU_AFFINITY_ALL -1ULL
71 * Supported protocol versions in the order of probing - highest go
74 static enum pci_protocol_version_t pci_protocol_versions[] = {
75 PCI_PROTOCOL_VERSION_1_2,
76 PCI_PROTOCOL_VERSION_1_1,
80 * Protocol version negotiated by hv_pci_protocol_negotiation().
82 static enum pci_protocol_version_t pci_protocol_version;
84 #define PCI_CONFIG_MMIO_LENGTH 0x2000
85 #define CFG_PAGE_OFFSET 0x1000
86 #define CFG_PAGE_SIZE (PCI_CONFIG_MMIO_LENGTH - CFG_PAGE_OFFSET)
88 #define MAX_SUPPORTED_MSI_MESSAGES 0x400
90 #define STATUS_REVISION_MISMATCH 0xC0000059
92 /* space for 32bit serial number as string */
93 #define SLOT_NAME_SIZE 11
99 enum pci_message_type {
103 PCI_MESSAGE_BASE = 0x42490000,
104 PCI_BUS_RELATIONS = PCI_MESSAGE_BASE + 0,
105 PCI_QUERY_BUS_RELATIONS = PCI_MESSAGE_BASE + 1,
106 PCI_POWER_STATE_CHANGE = PCI_MESSAGE_BASE + 4,
107 PCI_QUERY_RESOURCE_REQUIREMENTS = PCI_MESSAGE_BASE + 5,
108 PCI_QUERY_RESOURCE_RESOURCES = PCI_MESSAGE_BASE + 6,
109 PCI_BUS_D0ENTRY = PCI_MESSAGE_BASE + 7,
110 PCI_BUS_D0EXIT = PCI_MESSAGE_BASE + 8,
111 PCI_READ_BLOCK = PCI_MESSAGE_BASE + 9,
112 PCI_WRITE_BLOCK = PCI_MESSAGE_BASE + 0xA,
113 PCI_EJECT = PCI_MESSAGE_BASE + 0xB,
114 PCI_QUERY_STOP = PCI_MESSAGE_BASE + 0xC,
115 PCI_REENABLE = PCI_MESSAGE_BASE + 0xD,
116 PCI_QUERY_STOP_FAILED = PCI_MESSAGE_BASE + 0xE,
117 PCI_EJECTION_COMPLETE = PCI_MESSAGE_BASE + 0xF,
118 PCI_RESOURCES_ASSIGNED = PCI_MESSAGE_BASE + 0x10,
119 PCI_RESOURCES_RELEASED = PCI_MESSAGE_BASE + 0x11,
120 PCI_INVALIDATE_BLOCK = PCI_MESSAGE_BASE + 0x12,
121 PCI_QUERY_PROTOCOL_VERSION = PCI_MESSAGE_BASE + 0x13,
122 PCI_CREATE_INTERRUPT_MESSAGE = PCI_MESSAGE_BASE + 0x14,
123 PCI_DELETE_INTERRUPT_MESSAGE = PCI_MESSAGE_BASE + 0x15,
124 PCI_RESOURCES_ASSIGNED2 = PCI_MESSAGE_BASE + 0x16,
125 PCI_CREATE_INTERRUPT_MESSAGE2 = PCI_MESSAGE_BASE + 0x17,
126 PCI_DELETE_INTERRUPT_MESSAGE2 = PCI_MESSAGE_BASE + 0x18, /* unused */
131 * Structures defining the virtual PCI Express protocol.
143 * Function numbers are 8-bits wide on Express, as interpreted through ARI,
144 * which is all this driver does. This representation is the one used in
145 * Windows, which is what is expected when sending this back and forth with
146 * the Hyper-V parent partition.
148 union win_slot_encoding {
158 * Pretty much as defined in the PCI Specifications.
160 struct pci_function_description {
161 u16 v_id; /* vendor ID */
162 u16 d_id; /* device ID */
168 union win_slot_encoding win_slot;
169 u32 ser; /* serial number */
175 * @delivery_mode: As defined in Intel's Programmer's
176 * Reference Manual, Volume 3, Chapter 8.
177 * @vector_count: Number of contiguous entries in the
178 * Interrupt Descriptor Table that are
179 * occupied by this Message-Signaled
180 * Interrupt. For "MSI", as first defined
181 * in PCI 2.2, this can be between 1 and
182 * 32. For "MSI-X," as first defined in PCI
183 * 3.0, this must be 1, as each MSI-X table
184 * entry would have its own descriptor.
185 * @reserved: Empty space
186 * @cpu_mask: All the target virtual processors.
197 * struct hv_msi_desc2 - 1.2 version of hv_msi_desc
199 * @delivery_mode: As defined in Intel's Programmer's
200 * Reference Manual, Volume 3, Chapter 8.
201 * @vector_count: Number of contiguous entries in the
202 * Interrupt Descriptor Table that are
203 * occupied by this Message-Signaled
204 * Interrupt. For "MSI", as first defined
205 * in PCI 2.2, this can be between 1 and
206 * 32. For "MSI-X," as first defined in PCI
207 * 3.0, this must be 1, as each MSI-X table
208 * entry would have its own descriptor.
209 * @processor_count: number of bits enabled in array.
210 * @processor_array: All the target virtual processors.
212 struct hv_msi_desc2 {
217 u16 processor_array[32];
221 * struct tran_int_desc
222 * @reserved: unused, padding
223 * @vector_count: same as in hv_msi_desc
224 * @data: This is the "data payload" value that is
225 * written by the device when it generates
226 * a message-signaled interrupt, either MSI
228 * @address: This is the address to which the data
229 * payload is written on interrupt
232 struct tran_int_desc {
240 * A generic message format for virtual PCI.
241 * Specific message formats are defined later in the file.
248 struct pci_child_message {
249 struct pci_message message_type;
250 union win_slot_encoding wslot;
253 struct pci_incoming_message {
254 struct vmpacket_descriptor hdr;
255 struct pci_message message_type;
258 struct pci_response {
259 struct vmpacket_descriptor hdr;
260 s32 status; /* negative values are failures */
264 void (*completion_func)(void *context, struct pci_response *resp,
265 int resp_packet_size);
268 struct pci_message message[0];
272 * Specific message types supporting the PCI protocol.
276 * Version negotiation message. Sent from the guest to the host.
277 * The guest is free to try different versions until the host
278 * accepts the version.
280 * pci_version: The protocol version requested.
281 * is_last_attempt: If TRUE, this is the last version guest will request.
282 * reservedz: Reserved field, set to zero.
285 struct pci_version_request {
286 struct pci_message message_type;
287 u32 protocol_version;
291 * Bus D0 Entry. This is sent from the guest to the host when the virtual
292 * bus (PCI Express port) is ready for action.
295 struct pci_bus_d0_entry {
296 struct pci_message message_type;
301 struct pci_bus_relations {
302 struct pci_incoming_message incoming;
304 struct pci_function_description func[0];
307 struct pci_q_res_req_response {
308 struct vmpacket_descriptor hdr;
309 s32 status; /* negative values are failures */
313 struct pci_set_power {
314 struct pci_message message_type;
315 union win_slot_encoding wslot;
316 u32 power_state; /* In Windows terms */
320 struct pci_set_power_response {
321 struct vmpacket_descriptor hdr;
322 s32 status; /* negative values are failures */
323 union win_slot_encoding wslot;
324 u32 resultant_state; /* In Windows terms */
328 struct pci_resources_assigned {
329 struct pci_message message_type;
330 union win_slot_encoding wslot;
331 u8 memory_range[0x14][6]; /* not used here */
336 struct pci_resources_assigned2 {
337 struct pci_message message_type;
338 union win_slot_encoding wslot;
339 u8 memory_range[0x14][6]; /* not used here */
340 u32 msi_descriptor_count;
344 struct pci_create_interrupt {
345 struct pci_message message_type;
346 union win_slot_encoding wslot;
347 struct hv_msi_desc int_desc;
350 struct pci_create_int_response {
351 struct pci_response response;
353 struct tran_int_desc int_desc;
356 struct pci_create_interrupt2 {
357 struct pci_message message_type;
358 union win_slot_encoding wslot;
359 struct hv_msi_desc2 int_desc;
362 struct pci_delete_interrupt {
363 struct pci_message message_type;
364 union win_slot_encoding wslot;
365 struct tran_int_desc int_desc;
368 struct pci_dev_incoming {
369 struct pci_incoming_message incoming;
370 union win_slot_encoding wslot;
373 struct pci_eject_response {
374 struct pci_message message_type;
375 union win_slot_encoding wslot;
379 static int pci_ring_size = (4 * PAGE_SIZE);
382 * Definitions or interrupt steering hypercall.
384 #define HV_PARTITION_ID_SELF ((u64)-1)
385 #define HVCALL_RETARGET_INTERRUPT 0x7e
387 struct hv_interrupt_entry {
388 u32 source; /* 1 for MSI(-X) */
395 * flags for hv_device_interrupt_target.flags
397 #define HV_DEVICE_INTERRUPT_TARGET_MULTICAST 1
398 #define HV_DEVICE_INTERRUPT_TARGET_PROCESSOR_SET 2
400 struct hv_device_interrupt_target {
405 struct hv_vpset vp_set;
409 struct retarget_msi_interrupt {
410 u64 partition_id; /* use "self" */
412 struct hv_interrupt_entry int_entry;
414 struct hv_device_interrupt_target int_target;
415 } __packed __aligned(8);
418 * Driver specific state.
421 enum hv_pcibus_state {
429 struct hv_pcibus_device {
430 struct pci_sysdata sysdata;
431 enum hv_pcibus_state state;
432 refcount_t remove_lock;
433 struct hv_device *hdev;
434 resource_size_t low_mmio_space;
435 resource_size_t high_mmio_space;
436 struct resource *mem_config;
437 struct resource *low_mmio_res;
438 struct resource *high_mmio_res;
439 struct completion *survey_event;
440 struct completion remove_event;
441 struct pci_bus *pci_bus;
442 spinlock_t config_lock; /* Avoid two threads writing index page */
443 spinlock_t device_list_lock; /* Protect lists below */
444 void __iomem *cfg_addr;
446 struct list_head resources_for_children;
448 struct list_head children;
449 struct list_head dr_list;
451 struct msi_domain_info msi_info;
452 struct msi_controller msi_chip;
453 struct irq_domain *irq_domain;
455 spinlock_t retarget_msi_interrupt_lock;
457 struct workqueue_struct *wq;
459 /* hypercall arg, must not cross page boundary */
460 struct retarget_msi_interrupt retarget_msi_interrupt_params;
463 * Don't put anything here: retarget_msi_interrupt_params must be last
468 * Tracks "Device Relations" messages from the host, which must be both
469 * processed in order and deferred so that they don't run in the context
470 * of the incoming packet callback.
473 struct work_struct wrk;
474 struct hv_pcibus_device *bus;
478 struct list_head list_entry;
480 struct pci_function_description func[0];
483 enum hv_pcichild_state {
484 hv_pcichild_init = 0,
485 hv_pcichild_requirements,
486 hv_pcichild_resourced,
487 hv_pcichild_ejecting,
492 /* List protected by pci_rescan_remove_lock */
493 struct list_head list_entry;
495 enum hv_pcichild_state state;
496 struct pci_slot *pci_slot;
497 struct pci_function_description desc;
498 bool reported_missing;
499 struct hv_pcibus_device *hbus;
500 struct work_struct wrk;
503 * What would be observed if one wrote 0xFFFFFFFF to a BAR and then
504 * read it back, for each of the BAR offsets within config space.
509 struct hv_pci_compl {
510 struct completion host_event;
511 s32 completion_status;
514 static void hv_pci_onchannelcallback(void *context);
517 * hv_pci_generic_compl() - Invoked for a completion packet
518 * @context: Set up by the sender of the packet.
519 * @resp: The response packet
520 * @resp_packet_size: Size in bytes of the packet
522 * This function is used to trigger an event and report status
523 * for any message for which the completion packet contains a
524 * status and nothing else.
526 static void hv_pci_generic_compl(void *context, struct pci_response *resp,
527 int resp_packet_size)
529 struct hv_pci_compl *comp_pkt = context;
531 if (resp_packet_size >= offsetofend(struct pci_response, status))
532 comp_pkt->completion_status = resp->status;
534 comp_pkt->completion_status = -1;
536 complete(&comp_pkt->host_event);
539 static struct hv_pci_dev *get_pcichild_wslot(struct hv_pcibus_device *hbus,
542 static void get_pcichild(struct hv_pci_dev *hpdev)
544 refcount_inc(&hpdev->refs);
547 static void put_pcichild(struct hv_pci_dev *hpdev)
549 if (refcount_dec_and_test(&hpdev->refs))
553 static void get_hvpcibus(struct hv_pcibus_device *hv_pcibus);
554 static void put_hvpcibus(struct hv_pcibus_device *hv_pcibus);
557 * There is no good way to get notified from vmbus_onoffer_rescind(),
558 * so let's use polling here, since this is not a hot path.
560 static int wait_for_response(struct hv_device *hdev,
561 struct completion *comp)
564 if (hdev->channel->rescind) {
565 dev_warn_once(&hdev->device, "The device is gone.\n");
569 if (wait_for_completion_timeout(comp, HZ / 10))
577 * devfn_to_wslot() - Convert from Linux PCI slot to Windows
578 * @devfn: The Linux representation of PCI slot
580 * Windows uses a slightly different representation of PCI slot.
582 * Return: The Windows representation
584 static u32 devfn_to_wslot(int devfn)
586 union win_slot_encoding wslot;
589 wslot.bits.dev = PCI_SLOT(devfn);
590 wslot.bits.func = PCI_FUNC(devfn);
596 * wslot_to_devfn() - Convert from Windows PCI slot to Linux
597 * @wslot: The Windows representation of PCI slot
599 * Windows uses a slightly different representation of PCI slot.
601 * Return: The Linux representation
603 static int wslot_to_devfn(u32 wslot)
605 union win_slot_encoding slot_no;
607 slot_no.slot = wslot;
608 return PCI_DEVFN(slot_no.bits.dev, slot_no.bits.func);
612 * PCI Configuration Space for these root PCI buses is implemented as a pair
613 * of pages in memory-mapped I/O space. Writing to the first page chooses
614 * the PCI function being written or read. Once the first page has been
615 * written to, the following page maps in the entire configuration space of
620 * _hv_pcifront_read_config() - Internal PCI config read
621 * @hpdev: The PCI driver's representation of the device
622 * @where: Offset within config space
623 * @size: Size of the transfer
624 * @val: Pointer to the buffer receiving the data
626 static void _hv_pcifront_read_config(struct hv_pci_dev *hpdev, int where,
630 void __iomem *addr = hpdev->hbus->cfg_addr + CFG_PAGE_OFFSET + where;
633 * If the attempt is to read the IDs or the ROM BAR, simulate that.
635 if (where + size <= PCI_COMMAND) {
636 memcpy(val, ((u8 *)&hpdev->desc.v_id) + where, size);
637 } else if (where >= PCI_CLASS_REVISION && where + size <=
638 PCI_CACHE_LINE_SIZE) {
639 memcpy(val, ((u8 *)&hpdev->desc.rev) + where -
640 PCI_CLASS_REVISION, size);
641 } else if (where >= PCI_SUBSYSTEM_VENDOR_ID && where + size <=
643 memcpy(val, (u8 *)&hpdev->desc.subsystem_id + where -
644 PCI_SUBSYSTEM_VENDOR_ID, size);
645 } else if (where >= PCI_ROM_ADDRESS && where + size <=
646 PCI_CAPABILITY_LIST) {
647 /* ROM BARs are unimplemented */
649 } else if (where >= PCI_INTERRUPT_LINE && where + size <=
652 * Interrupt Line and Interrupt PIN are hard-wired to zero
653 * because this front-end only supports message-signaled
657 } else if (where + size <= CFG_PAGE_SIZE) {
658 spin_lock_irqsave(&hpdev->hbus->config_lock, flags);
659 /* Choose the function to be read. (See comment above) */
660 writel(hpdev->desc.win_slot.slot, hpdev->hbus->cfg_addr);
661 /* Make sure the function was chosen before we start reading. */
663 /* Read from that function's config space. */
676 * Make sure the read was done before we release the spinlock
677 * allowing consecutive reads/writes.
680 spin_unlock_irqrestore(&hpdev->hbus->config_lock, flags);
682 dev_err(&hpdev->hbus->hdev->device,
683 "Attempt to read beyond a function's config space.\n");
687 static u16 hv_pcifront_get_vendor_id(struct hv_pci_dev *hpdev)
691 void __iomem *addr = hpdev->hbus->cfg_addr + CFG_PAGE_OFFSET +
694 spin_lock_irqsave(&hpdev->hbus->config_lock, flags);
696 /* Choose the function to be read. (See comment above) */
697 writel(hpdev->desc.win_slot.slot, hpdev->hbus->cfg_addr);
698 /* Make sure the function was chosen before we start reading. */
700 /* Read from that function's config space. */
703 * mb() is not required here, because the spin_unlock_irqrestore()
707 spin_unlock_irqrestore(&hpdev->hbus->config_lock, flags);
713 * _hv_pcifront_write_config() - Internal PCI config write
714 * @hpdev: The PCI driver's representation of the device
715 * @where: Offset within config space
716 * @size: Size of the transfer
717 * @val: The data being transferred
719 static void _hv_pcifront_write_config(struct hv_pci_dev *hpdev, int where,
723 void __iomem *addr = hpdev->hbus->cfg_addr + CFG_PAGE_OFFSET + where;
725 if (where >= PCI_SUBSYSTEM_VENDOR_ID &&
726 where + size <= PCI_CAPABILITY_LIST) {
727 /* SSIDs and ROM BARs are read-only */
728 } else if (where >= PCI_COMMAND && where + size <= CFG_PAGE_SIZE) {
729 spin_lock_irqsave(&hpdev->hbus->config_lock, flags);
730 /* Choose the function to be written. (See comment above) */
731 writel(hpdev->desc.win_slot.slot, hpdev->hbus->cfg_addr);
732 /* Make sure the function was chosen before we start writing. */
734 /* Write to that function's config space. */
747 * Make sure the write was done before we release the spinlock
748 * allowing consecutive reads/writes.
751 spin_unlock_irqrestore(&hpdev->hbus->config_lock, flags);
753 dev_err(&hpdev->hbus->hdev->device,
754 "Attempt to write beyond a function's config space.\n");
759 * hv_pcifront_read_config() - Read configuration space
760 * @bus: PCI Bus structure
761 * @devfn: Device/function
762 * @where: Offset from base
763 * @size: Byte/word/dword
764 * @val: Value to be read
766 * Return: PCIBIOS_SUCCESSFUL on success
767 * PCIBIOS_DEVICE_NOT_FOUND on failure
769 static int hv_pcifront_read_config(struct pci_bus *bus, unsigned int devfn,
770 int where, int size, u32 *val)
772 struct hv_pcibus_device *hbus =
773 container_of(bus->sysdata, struct hv_pcibus_device, sysdata);
774 struct hv_pci_dev *hpdev;
776 hpdev = get_pcichild_wslot(hbus, devfn_to_wslot(devfn));
778 return PCIBIOS_DEVICE_NOT_FOUND;
780 _hv_pcifront_read_config(hpdev, where, size, val);
783 return PCIBIOS_SUCCESSFUL;
787 * hv_pcifront_write_config() - Write configuration space
788 * @bus: PCI Bus structure
789 * @devfn: Device/function
790 * @where: Offset from base
791 * @size: Byte/word/dword
792 * @val: Value to be written to device
794 * Return: PCIBIOS_SUCCESSFUL on success
795 * PCIBIOS_DEVICE_NOT_FOUND on failure
797 static int hv_pcifront_write_config(struct pci_bus *bus, unsigned int devfn,
798 int where, int size, u32 val)
800 struct hv_pcibus_device *hbus =
801 container_of(bus->sysdata, struct hv_pcibus_device, sysdata);
802 struct hv_pci_dev *hpdev;
804 hpdev = get_pcichild_wslot(hbus, devfn_to_wslot(devfn));
806 return PCIBIOS_DEVICE_NOT_FOUND;
808 _hv_pcifront_write_config(hpdev, where, size, val);
811 return PCIBIOS_SUCCESSFUL;
814 /* PCIe operations */
815 static struct pci_ops hv_pcifront_ops = {
816 .read = hv_pcifront_read_config,
817 .write = hv_pcifront_write_config,
820 /* Interrupt management hooks */
821 static void hv_int_desc_free(struct hv_pci_dev *hpdev,
822 struct tran_int_desc *int_desc)
824 struct pci_delete_interrupt *int_pkt;
826 struct pci_packet pkt;
827 u8 buffer[sizeof(struct pci_delete_interrupt)];
830 memset(&ctxt, 0, sizeof(ctxt));
831 int_pkt = (struct pci_delete_interrupt *)&ctxt.pkt.message;
832 int_pkt->message_type.type =
833 PCI_DELETE_INTERRUPT_MESSAGE;
834 int_pkt->wslot.slot = hpdev->desc.win_slot.slot;
835 int_pkt->int_desc = *int_desc;
836 vmbus_sendpacket(hpdev->hbus->hdev->channel, int_pkt, sizeof(*int_pkt),
837 (unsigned long)&ctxt.pkt, VM_PKT_DATA_INBAND, 0);
842 * hv_msi_free() - Free the MSI.
843 * @domain: The interrupt domain pointer
844 * @info: Extra MSI-related context
845 * @irq: Identifies the IRQ.
847 * The Hyper-V parent partition and hypervisor are tracking the
848 * messages that are in use, keeping the interrupt redirection
849 * table up to date. This callback sends a message that frees
850 * the IRT entry and related tracking nonsense.
852 static void hv_msi_free(struct irq_domain *domain, struct msi_domain_info *info,
855 struct hv_pcibus_device *hbus;
856 struct hv_pci_dev *hpdev;
857 struct pci_dev *pdev;
858 struct tran_int_desc *int_desc;
859 struct irq_data *irq_data = irq_domain_get_irq_data(domain, irq);
860 struct msi_desc *msi = irq_data_get_msi_desc(irq_data);
862 pdev = msi_desc_to_pci_dev(msi);
864 int_desc = irq_data_get_irq_chip_data(irq_data);
868 irq_data->chip_data = NULL;
869 hpdev = get_pcichild_wslot(hbus, devfn_to_wslot(pdev->devfn));
875 hv_int_desc_free(hpdev, int_desc);
879 static int hv_set_affinity(struct irq_data *data, const struct cpumask *dest,
882 struct irq_data *parent = data->parent_data;
884 return parent->chip->irq_set_affinity(parent, dest, force);
887 static void hv_irq_mask(struct irq_data *data)
889 pci_msi_mask_irq(data);
893 * hv_irq_unmask() - "Unmask" the IRQ by setting its current
895 * @data: Describes the IRQ
897 * Build new a destination for the MSI and make a hypercall to
898 * update the Interrupt Redirection Table. "Device Logical ID"
899 * is built out of this PCI bus's instance GUID and the function
900 * number of the device.
902 static void hv_irq_unmask(struct irq_data *data)
904 struct msi_desc *msi_desc = irq_data_get_msi_desc(data);
905 struct irq_cfg *cfg = irqd_cfg(data);
906 struct retarget_msi_interrupt *params;
907 struct hv_pcibus_device *hbus;
908 struct cpumask *dest;
910 struct pci_bus *pbus;
911 struct pci_dev *pdev;
917 dest = irq_data_get_effective_affinity_mask(data);
918 pdev = msi_desc_to_pci_dev(msi_desc);
920 hbus = container_of(pbus->sysdata, struct hv_pcibus_device, sysdata);
922 spin_lock_irqsave(&hbus->retarget_msi_interrupt_lock, flags);
924 params = &hbus->retarget_msi_interrupt_params;
925 memset(params, 0, sizeof(*params));
926 params->partition_id = HV_PARTITION_ID_SELF;
927 params->int_entry.source = 1; /* MSI(-X) */
928 params->int_entry.address = msi_desc->msg.address_lo;
929 params->int_entry.data = msi_desc->msg.data;
930 params->device_id = (hbus->hdev->dev_instance.b[5] << 24) |
931 (hbus->hdev->dev_instance.b[4] << 16) |
932 (hbus->hdev->dev_instance.b[7] << 8) |
933 (hbus->hdev->dev_instance.b[6] & 0xf8) |
934 PCI_FUNC(pdev->devfn);
935 params->int_target.vector = cfg->vector;
938 * Honoring apic->irq_delivery_mode set to dest_Fixed by
939 * setting the HV_DEVICE_INTERRUPT_TARGET_MULTICAST flag results in a
940 * spurious interrupt storm. Not doing so does not seem to have a
941 * negative effect (yet?).
944 if (pci_protocol_version >= PCI_PROTOCOL_VERSION_1_2) {
946 * PCI_PROTOCOL_VERSION_1_2 supports the VP_SET version of the
947 * HVCALL_RETARGET_INTERRUPT hypercall, which also coincides
948 * with >64 VP support.
949 * ms_hyperv.hints & HV_X64_EX_PROCESSOR_MASKS_RECOMMENDED
950 * is not sufficient for this hypercall.
952 params->int_target.flags |=
953 HV_DEVICE_INTERRUPT_TARGET_PROCESSOR_SET;
955 if (!alloc_cpumask_var(&tmp, GFP_ATOMIC)) {
960 cpumask_and(tmp, dest, cpu_online_mask);
961 nr_bank = cpumask_to_vpset(¶ms->int_target.vp_set, tmp);
962 free_cpumask_var(tmp);
970 * var-sized hypercall, var-size starts after vp_mask (thus
971 * vp_set.format does not count, but vp_set.valid_bank_mask
974 var_size = 1 + nr_bank;
976 for_each_cpu_and(cpu, dest, cpu_online_mask) {
977 params->int_target.vp_mask |=
978 (1ULL << hv_cpu_number_to_vp_number(cpu));
982 res = hv_do_hypercall(HVCALL_RETARGET_INTERRUPT | (var_size << 17),
986 spin_unlock_irqrestore(&hbus->retarget_msi_interrupt_lock, flags);
989 dev_err(&hbus->hdev->device,
990 "%s() failed: %#llx", __func__, res);
994 pci_msi_unmask_irq(data);
997 struct compose_comp_ctxt {
998 struct hv_pci_compl comp_pkt;
999 struct tran_int_desc int_desc;
1002 static void hv_pci_compose_compl(void *context, struct pci_response *resp,
1003 int resp_packet_size)
1005 struct compose_comp_ctxt *comp_pkt = context;
1006 struct pci_create_int_response *int_resp =
1007 (struct pci_create_int_response *)resp;
1009 comp_pkt->comp_pkt.completion_status = resp->status;
1010 comp_pkt->int_desc = int_resp->int_desc;
1011 complete(&comp_pkt->comp_pkt.host_event);
1014 static u32 hv_compose_msi_req_v1(
1015 struct pci_create_interrupt *int_pkt, struct cpumask *affinity,
1016 u32 slot, u8 vector)
1018 int_pkt->message_type.type = PCI_CREATE_INTERRUPT_MESSAGE;
1019 int_pkt->wslot.slot = slot;
1020 int_pkt->int_desc.vector = vector;
1021 int_pkt->int_desc.vector_count = 1;
1022 int_pkt->int_desc.delivery_mode = dest_Fixed;
1025 * Create MSI w/ dummy vCPU set, overwritten by subsequent retarget in
1028 int_pkt->int_desc.cpu_mask = CPU_AFFINITY_ALL;
1030 return sizeof(*int_pkt);
1033 static u32 hv_compose_msi_req_v2(
1034 struct pci_create_interrupt2 *int_pkt, struct cpumask *affinity,
1035 u32 slot, u8 vector)
1039 int_pkt->message_type.type = PCI_CREATE_INTERRUPT_MESSAGE2;
1040 int_pkt->wslot.slot = slot;
1041 int_pkt->int_desc.vector = vector;
1042 int_pkt->int_desc.vector_count = 1;
1043 int_pkt->int_desc.delivery_mode = dest_Fixed;
1046 * Create MSI w/ dummy vCPU set targeting just one vCPU, overwritten
1047 * by subsequent retarget in hv_irq_unmask().
1049 cpu = cpumask_first_and(affinity, cpu_online_mask);
1050 int_pkt->int_desc.processor_array[0] =
1051 hv_cpu_number_to_vp_number(cpu);
1052 int_pkt->int_desc.processor_count = 1;
1054 return sizeof(*int_pkt);
1058 * hv_compose_msi_msg() - Supplies a valid MSI address/data
1059 * @data: Everything about this MSI
1060 * @msg: Buffer that is filled in by this function
1062 * This function unpacks the IRQ looking for target CPU set, IDT
1063 * vector and mode and sends a message to the parent partition
1064 * asking for a mapping for that tuple in this partition. The
1065 * response supplies a data value and address to which that data
1066 * should be written to trigger that interrupt.
1068 static void hv_compose_msi_msg(struct irq_data *data, struct msi_msg *msg)
1070 struct irq_cfg *cfg = irqd_cfg(data);
1071 struct hv_pcibus_device *hbus;
1072 struct hv_pci_dev *hpdev;
1073 struct pci_bus *pbus;
1074 struct pci_dev *pdev;
1075 struct cpumask *dest;
1076 unsigned long flags;
1077 struct compose_comp_ctxt comp;
1078 struct tran_int_desc *int_desc;
1080 struct pci_packet pci_pkt;
1082 struct pci_create_interrupt v1;
1083 struct pci_create_interrupt2 v2;
1090 pdev = msi_desc_to_pci_dev(irq_data_get_msi_desc(data));
1091 dest = irq_data_get_effective_affinity_mask(data);
1093 hbus = container_of(pbus->sysdata, struct hv_pcibus_device, sysdata);
1094 hpdev = get_pcichild_wslot(hbus, devfn_to_wslot(pdev->devfn));
1096 goto return_null_message;
1098 /* Free any previous message that might have already been composed. */
1099 if (data->chip_data) {
1100 int_desc = data->chip_data;
1101 data->chip_data = NULL;
1102 hv_int_desc_free(hpdev, int_desc);
1105 int_desc = kzalloc(sizeof(*int_desc), GFP_ATOMIC);
1107 goto drop_reference;
1109 memset(&ctxt, 0, sizeof(ctxt));
1110 init_completion(&comp.comp_pkt.host_event);
1111 ctxt.pci_pkt.completion_func = hv_pci_compose_compl;
1112 ctxt.pci_pkt.compl_ctxt = ∁
1114 switch (pci_protocol_version) {
1115 case PCI_PROTOCOL_VERSION_1_1:
1116 size = hv_compose_msi_req_v1(&ctxt.int_pkts.v1,
1118 hpdev->desc.win_slot.slot,
1122 case PCI_PROTOCOL_VERSION_1_2:
1123 size = hv_compose_msi_req_v2(&ctxt.int_pkts.v2,
1125 hpdev->desc.win_slot.slot,
1130 /* As we only negotiate protocol versions known to this driver,
1131 * this path should never hit. However, this is it not a hot
1132 * path so we print a message to aid future updates.
1134 dev_err(&hbus->hdev->device,
1135 "Unexpected vPCI protocol, update driver.");
1139 ret = vmbus_sendpacket(hpdev->hbus->hdev->channel, &ctxt.int_pkts,
1140 size, (unsigned long)&ctxt.pci_pkt,
1142 VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
1144 dev_err(&hbus->hdev->device,
1145 "Sending request for interrupt failed: 0x%x",
1146 comp.comp_pkt.completion_status);
1151 * Since this function is called with IRQ locks held, can't
1152 * do normal wait for completion; instead poll.
1154 while (!try_wait_for_completion(&comp.comp_pkt.host_event)) {
1155 /* 0xFFFF means an invalid PCI VENDOR ID. */
1156 if (hv_pcifront_get_vendor_id(hpdev) == 0xFFFF) {
1157 dev_err_once(&hbus->hdev->device,
1158 "the device has gone\n");
1163 * When the higher level interrupt code calls us with
1164 * interrupt disabled, we must poll the channel by calling
1165 * the channel callback directly when channel->target_cpu is
1166 * the current CPU. When the higher level interrupt code
1167 * calls us with interrupt enabled, let's add the
1168 * local_irq_save()/restore() to avoid race:
1169 * hv_pci_onchannelcallback() can also run in tasklet.
1171 local_irq_save(flags);
1173 if (hbus->hdev->channel->target_cpu == smp_processor_id())
1174 hv_pci_onchannelcallback(hbus);
1176 local_irq_restore(flags);
1178 if (hpdev->state == hv_pcichild_ejecting) {
1179 dev_err_once(&hbus->hdev->device,
1180 "the device is being ejected\n");
1187 if (comp.comp_pkt.completion_status < 0) {
1188 dev_err(&hbus->hdev->device,
1189 "Request for interrupt failed: 0x%x",
1190 comp.comp_pkt.completion_status);
1195 * Record the assignment so that this can be unwound later. Using
1196 * irq_set_chip_data() here would be appropriate, but the lock it takes
1199 *int_desc = comp.int_desc;
1200 data->chip_data = int_desc;
1202 /* Pass up the result. */
1203 msg->address_hi = comp.int_desc.address >> 32;
1204 msg->address_lo = comp.int_desc.address & 0xffffffff;
1205 msg->data = comp.int_desc.data;
1207 put_pcichild(hpdev);
1213 put_pcichild(hpdev);
1214 return_null_message:
1215 msg->address_hi = 0;
1216 msg->address_lo = 0;
1220 /* HW Interrupt Chip Descriptor */
1221 static struct irq_chip hv_msi_irq_chip = {
1222 .name = "Hyper-V PCIe MSI",
1223 .irq_compose_msi_msg = hv_compose_msi_msg,
1224 .irq_set_affinity = hv_set_affinity,
1225 .irq_ack = irq_chip_ack_parent,
1226 .irq_mask = hv_irq_mask,
1227 .irq_unmask = hv_irq_unmask,
1230 static irq_hw_number_t hv_msi_domain_ops_get_hwirq(struct msi_domain_info *info,
1231 msi_alloc_info_t *arg)
1233 return arg->msi_hwirq;
1236 static struct msi_domain_ops hv_msi_ops = {
1237 .get_hwirq = hv_msi_domain_ops_get_hwirq,
1238 .msi_prepare = pci_msi_prepare,
1239 .set_desc = pci_msi_set_desc,
1240 .msi_free = hv_msi_free,
1244 * hv_pcie_init_irq_domain() - Initialize IRQ domain
1245 * @hbus: The root PCI bus
1247 * This function creates an IRQ domain which will be used for
1248 * interrupts from devices that have been passed through. These
1249 * devices only support MSI and MSI-X, not line-based interrupts
1250 * or simulations of line-based interrupts through PCIe's
1251 * fabric-layer messages. Because interrupts are remapped, we
1252 * can support multi-message MSI here.
1254 * Return: '0' on success and error value on failure
1256 static int hv_pcie_init_irq_domain(struct hv_pcibus_device *hbus)
1258 hbus->msi_info.chip = &hv_msi_irq_chip;
1259 hbus->msi_info.ops = &hv_msi_ops;
1260 hbus->msi_info.flags = (MSI_FLAG_USE_DEF_DOM_OPS |
1261 MSI_FLAG_USE_DEF_CHIP_OPS | MSI_FLAG_MULTI_PCI_MSI |
1263 hbus->msi_info.handler = handle_edge_irq;
1264 hbus->msi_info.handler_name = "edge";
1265 hbus->msi_info.data = hbus;
1266 hbus->irq_domain = pci_msi_create_irq_domain(hbus->sysdata.fwnode,
1269 if (!hbus->irq_domain) {
1270 dev_err(&hbus->hdev->device,
1271 "Failed to build an MSI IRQ domain\n");
1279 * get_bar_size() - Get the address space consumed by a BAR
1280 * @bar_val: Value that a BAR returned after -1 was written
1283 * This function returns the size of the BAR, rounded up to 1
1284 * page. It has to be rounded up because the hypervisor's page
1285 * table entry that maps the BAR into the VM can't specify an
1286 * offset within a page. The invariant is that the hypervisor
1287 * must place any BARs of smaller than page length at the
1288 * beginning of a page.
1290 * Return: Size in bytes of the consumed MMIO space.
1292 static u64 get_bar_size(u64 bar_val)
1294 return round_up((1 + ~(bar_val & PCI_BASE_ADDRESS_MEM_MASK)),
1299 * survey_child_resources() - Total all MMIO requirements
1300 * @hbus: Root PCI bus, as understood by this driver
1302 static void survey_child_resources(struct hv_pcibus_device *hbus)
1304 struct hv_pci_dev *hpdev;
1305 resource_size_t bar_size = 0;
1306 unsigned long flags;
1307 struct completion *event;
1311 /* If nobody is waiting on the answer, don't compute it. */
1312 event = xchg(&hbus->survey_event, NULL);
1316 /* If the answer has already been computed, go with it. */
1317 if (hbus->low_mmio_space || hbus->high_mmio_space) {
1322 spin_lock_irqsave(&hbus->device_list_lock, flags);
1325 * Due to an interesting quirk of the PCI spec, all memory regions
1326 * for a child device are a power of 2 in size and aligned in memory,
1327 * so it's sufficient to just add them up without tracking alignment.
1329 list_for_each_entry(hpdev, &hbus->children, list_entry) {
1330 for (i = 0; i < 6; i++) {
1331 if (hpdev->probed_bar[i] & PCI_BASE_ADDRESS_SPACE_IO)
1332 dev_err(&hbus->hdev->device,
1333 "There's an I/O BAR in this list!\n");
1335 if (hpdev->probed_bar[i] != 0) {
1337 * A probed BAR has all the upper bits set that
1341 bar_val = hpdev->probed_bar[i];
1342 if (bar_val & PCI_BASE_ADDRESS_MEM_TYPE_64)
1344 ((u64)hpdev->probed_bar[++i] << 32);
1346 bar_val |= 0xffffffff00000000ULL;
1348 bar_size = get_bar_size(bar_val);
1350 if (bar_val & PCI_BASE_ADDRESS_MEM_TYPE_64)
1351 hbus->high_mmio_space += bar_size;
1353 hbus->low_mmio_space += bar_size;
1358 spin_unlock_irqrestore(&hbus->device_list_lock, flags);
1363 * prepopulate_bars() - Fill in BARs with defaults
1364 * @hbus: Root PCI bus, as understood by this driver
1366 * The core PCI driver code seems much, much happier if the BARs
1367 * for a device have values upon first scan. So fill them in.
1368 * The algorithm below works down from large sizes to small,
1369 * attempting to pack the assignments optimally. The assumption,
1370 * enforced in other parts of the code, is that the beginning of
1371 * the memory-mapped I/O space will be aligned on the largest
1374 static void prepopulate_bars(struct hv_pcibus_device *hbus)
1376 resource_size_t high_size = 0;
1377 resource_size_t low_size = 0;
1378 resource_size_t high_base = 0;
1379 resource_size_t low_base = 0;
1380 resource_size_t bar_size;
1381 struct hv_pci_dev *hpdev;
1382 unsigned long flags;
1388 if (hbus->low_mmio_space) {
1389 low_size = 1ULL << (63 - __builtin_clzll(hbus->low_mmio_space));
1390 low_base = hbus->low_mmio_res->start;
1393 if (hbus->high_mmio_space) {
1395 (63 - __builtin_clzll(hbus->high_mmio_space));
1396 high_base = hbus->high_mmio_res->start;
1399 spin_lock_irqsave(&hbus->device_list_lock, flags);
1401 /* Pick addresses for the BARs. */
1403 list_for_each_entry(hpdev, &hbus->children, list_entry) {
1404 for (i = 0; i < 6; i++) {
1405 bar_val = hpdev->probed_bar[i];
1408 high = bar_val & PCI_BASE_ADDRESS_MEM_TYPE_64;
1411 ((u64)hpdev->probed_bar[i + 1]
1414 bar_val |= 0xffffffffULL << 32;
1416 bar_size = get_bar_size(bar_val);
1418 if (high_size != bar_size) {
1422 _hv_pcifront_write_config(hpdev,
1423 PCI_BASE_ADDRESS_0 + (4 * i),
1425 (u32)(high_base & 0xffffff00));
1427 _hv_pcifront_write_config(hpdev,
1428 PCI_BASE_ADDRESS_0 + (4 * i),
1429 4, (u32)(high_base >> 32));
1430 high_base += bar_size;
1432 if (low_size != bar_size)
1434 _hv_pcifront_write_config(hpdev,
1435 PCI_BASE_ADDRESS_0 + (4 * i),
1437 (u32)(low_base & 0xffffff00));
1438 low_base += bar_size;
1441 if (high_size <= 1 && low_size <= 1) {
1442 /* Set the memory enable bit. */
1443 _hv_pcifront_read_config(hpdev, PCI_COMMAND, 2,
1445 command |= PCI_COMMAND_MEMORY;
1446 _hv_pcifront_write_config(hpdev, PCI_COMMAND, 2,
1454 } while (high_size || low_size);
1456 spin_unlock_irqrestore(&hbus->device_list_lock, flags);
1460 * Assign entries in sysfs pci slot directory.
1462 * Note that this function does not need to lock the children list
1463 * because it is called from pci_devices_present_work which
1464 * is serialized with hv_eject_device_work because they are on the
1465 * same ordered workqueue. Therefore hbus->children list will not change
1466 * even when pci_create_slot sleeps.
1468 static void hv_pci_assign_slots(struct hv_pcibus_device *hbus)
1470 struct hv_pci_dev *hpdev;
1471 char name[SLOT_NAME_SIZE];
1474 list_for_each_entry(hpdev, &hbus->children, list_entry) {
1475 if (hpdev->pci_slot)
1478 slot_nr = PCI_SLOT(wslot_to_devfn(hpdev->desc.win_slot.slot));
1479 snprintf(name, SLOT_NAME_SIZE, "%u", hpdev->desc.ser);
1480 hpdev->pci_slot = pci_create_slot(hbus->pci_bus, slot_nr,
1482 if (IS_ERR(hpdev->pci_slot)) {
1483 pr_warn("pci_create slot %s failed\n", name);
1484 hpdev->pci_slot = NULL;
1490 * create_root_hv_pci_bus() - Expose a new root PCI bus
1491 * @hbus: Root PCI bus, as understood by this driver
1493 * Return: 0 on success, -errno on failure
1495 static int create_root_hv_pci_bus(struct hv_pcibus_device *hbus)
1497 /* Register the device */
1498 hbus->pci_bus = pci_create_root_bus(&hbus->hdev->device,
1499 0, /* bus number is always zero */
1502 &hbus->resources_for_children);
1506 hbus->pci_bus->msi = &hbus->msi_chip;
1507 hbus->pci_bus->msi->dev = &hbus->hdev->device;
1509 pci_lock_rescan_remove();
1510 pci_scan_child_bus(hbus->pci_bus);
1511 pci_bus_assign_resources(hbus->pci_bus);
1512 hv_pci_assign_slots(hbus);
1513 pci_bus_add_devices(hbus->pci_bus);
1514 pci_unlock_rescan_remove();
1515 hbus->state = hv_pcibus_installed;
1519 struct q_res_req_compl {
1520 struct completion host_event;
1521 struct hv_pci_dev *hpdev;
1525 * q_resource_requirements() - Query Resource Requirements
1526 * @context: The completion context.
1527 * @resp: The response that came from the host.
1528 * @resp_packet_size: The size in bytes of resp.
1530 * This function is invoked on completion of a Query Resource
1531 * Requirements packet.
1533 static void q_resource_requirements(void *context, struct pci_response *resp,
1534 int resp_packet_size)
1536 struct q_res_req_compl *completion = context;
1537 struct pci_q_res_req_response *q_res_req =
1538 (struct pci_q_res_req_response *)resp;
1541 if (resp->status < 0) {
1542 dev_err(&completion->hpdev->hbus->hdev->device,
1543 "query resource requirements failed: %x\n",
1546 for (i = 0; i < 6; i++) {
1547 completion->hpdev->probed_bar[i] =
1548 q_res_req->probed_bar[i];
1552 complete(&completion->host_event);
1556 * new_pcichild_device() - Create a new child device
1557 * @hbus: The internal struct tracking this root PCI bus.
1558 * @desc: The information supplied so far from the host
1561 * This function creates the tracking structure for a new child
1562 * device and kicks off the process of figuring out what it is.
1564 * Return: Pointer to the new tracking struct
1566 static struct hv_pci_dev *new_pcichild_device(struct hv_pcibus_device *hbus,
1567 struct pci_function_description *desc)
1569 struct hv_pci_dev *hpdev;
1570 struct pci_child_message *res_req;
1571 struct q_res_req_compl comp_pkt;
1573 struct pci_packet init_packet;
1574 u8 buffer[sizeof(struct pci_child_message)];
1576 unsigned long flags;
1579 hpdev = kzalloc(sizeof(*hpdev), GFP_KERNEL);
1585 memset(&pkt, 0, sizeof(pkt));
1586 init_completion(&comp_pkt.host_event);
1587 comp_pkt.hpdev = hpdev;
1588 pkt.init_packet.compl_ctxt = &comp_pkt;
1589 pkt.init_packet.completion_func = q_resource_requirements;
1590 res_req = (struct pci_child_message *)&pkt.init_packet.message;
1591 res_req->message_type.type = PCI_QUERY_RESOURCE_REQUIREMENTS;
1592 res_req->wslot.slot = desc->win_slot.slot;
1594 ret = vmbus_sendpacket(hbus->hdev->channel, res_req,
1595 sizeof(struct pci_child_message),
1596 (unsigned long)&pkt.init_packet,
1598 VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
1602 if (wait_for_response(hbus->hdev, &comp_pkt.host_event))
1605 hpdev->desc = *desc;
1606 refcount_set(&hpdev->refs, 1);
1607 get_pcichild(hpdev);
1608 spin_lock_irqsave(&hbus->device_list_lock, flags);
1610 list_add_tail(&hpdev->list_entry, &hbus->children);
1611 spin_unlock_irqrestore(&hbus->device_list_lock, flags);
1620 * get_pcichild_wslot() - Find device from slot
1621 * @hbus: Root PCI bus, as understood by this driver
1622 * @wslot: Location on the bus
1624 * This function looks up a PCI device and returns the internal
1625 * representation of it. It acquires a reference on it, so that
1626 * the device won't be deleted while somebody is using it. The
1627 * caller is responsible for calling put_pcichild() to release
1630 * Return: Internal representation of a PCI device
1632 static struct hv_pci_dev *get_pcichild_wslot(struct hv_pcibus_device *hbus,
1635 unsigned long flags;
1636 struct hv_pci_dev *iter, *hpdev = NULL;
1638 spin_lock_irqsave(&hbus->device_list_lock, flags);
1639 list_for_each_entry(iter, &hbus->children, list_entry) {
1640 if (iter->desc.win_slot.slot == wslot) {
1642 get_pcichild(hpdev);
1646 spin_unlock_irqrestore(&hbus->device_list_lock, flags);
1652 * pci_devices_present_work() - Handle new list of child devices
1653 * @work: Work struct embedded in struct hv_dr_work
1655 * "Bus Relations" is the Windows term for "children of this
1656 * bus." The terminology is preserved here for people trying to
1657 * debug the interaction between Hyper-V and Linux. This
1658 * function is called when the parent partition reports a list
1659 * of functions that should be observed under this PCI Express
1662 * This function updates the list, and must tolerate being
1663 * called multiple times with the same information. The typical
1664 * number of child devices is one, with very atypical cases
1665 * involving three or four, so the algorithms used here can be
1666 * simple and inefficient.
1668 * It must also treat the omission of a previously observed device as
1669 * notification that the device no longer exists.
1671 * Note that this function is serialized with hv_eject_device_work(),
1672 * because both are pushed to the ordered workqueue hbus->wq.
1674 static void pci_devices_present_work(struct work_struct *work)
1678 struct pci_function_description *new_desc;
1679 struct hv_pci_dev *hpdev;
1680 struct hv_pcibus_device *hbus;
1681 struct list_head removed;
1682 struct hv_dr_work *dr_wrk;
1683 struct hv_dr_state *dr = NULL;
1684 unsigned long flags;
1686 dr_wrk = container_of(work, struct hv_dr_work, wrk);
1690 INIT_LIST_HEAD(&removed);
1692 /* Pull this off the queue and process it if it was the last one. */
1693 spin_lock_irqsave(&hbus->device_list_lock, flags);
1694 while (!list_empty(&hbus->dr_list)) {
1695 dr = list_first_entry(&hbus->dr_list, struct hv_dr_state,
1697 list_del(&dr->list_entry);
1699 /* Throw this away if the list still has stuff in it. */
1700 if (!list_empty(&hbus->dr_list)) {
1705 spin_unlock_irqrestore(&hbus->device_list_lock, flags);
1712 /* First, mark all existing children as reported missing. */
1713 spin_lock_irqsave(&hbus->device_list_lock, flags);
1714 list_for_each_entry(hpdev, &hbus->children, list_entry) {
1715 hpdev->reported_missing = true;
1717 spin_unlock_irqrestore(&hbus->device_list_lock, flags);
1719 /* Next, add back any reported devices. */
1720 for (child_no = 0; child_no < dr->device_count; child_no++) {
1722 new_desc = &dr->func[child_no];
1724 spin_lock_irqsave(&hbus->device_list_lock, flags);
1725 list_for_each_entry(hpdev, &hbus->children, list_entry) {
1726 if ((hpdev->desc.win_slot.slot == new_desc->win_slot.slot) &&
1727 (hpdev->desc.v_id == new_desc->v_id) &&
1728 (hpdev->desc.d_id == new_desc->d_id) &&
1729 (hpdev->desc.ser == new_desc->ser)) {
1730 hpdev->reported_missing = false;
1734 spin_unlock_irqrestore(&hbus->device_list_lock, flags);
1737 hpdev = new_pcichild_device(hbus, new_desc);
1739 dev_err(&hbus->hdev->device,
1740 "couldn't record a child device.\n");
1744 /* Move missing children to a list on the stack. */
1745 spin_lock_irqsave(&hbus->device_list_lock, flags);
1748 list_for_each_entry(hpdev, &hbus->children, list_entry) {
1749 if (hpdev->reported_missing) {
1751 put_pcichild(hpdev);
1752 list_move_tail(&hpdev->list_entry, &removed);
1757 spin_unlock_irqrestore(&hbus->device_list_lock, flags);
1759 /* Delete everything that should no longer exist. */
1760 while (!list_empty(&removed)) {
1761 hpdev = list_first_entry(&removed, struct hv_pci_dev,
1763 list_del(&hpdev->list_entry);
1764 put_pcichild(hpdev);
1767 switch (hbus->state) {
1768 case hv_pcibus_installed:
1770 * Tell the core to rescan bus
1771 * because there may have been changes.
1773 pci_lock_rescan_remove();
1774 pci_scan_child_bus(hbus->pci_bus);
1775 hv_pci_assign_slots(hbus);
1776 pci_unlock_rescan_remove();
1779 case hv_pcibus_init:
1780 case hv_pcibus_probed:
1781 survey_child_resources(hbus);
1793 * hv_pci_devices_present() - Handles list of new children
1794 * @hbus: Root PCI bus, as understood by this driver
1795 * @relations: Packet from host listing children
1797 * This function is invoked whenever a new list of devices for
1800 static void hv_pci_devices_present(struct hv_pcibus_device *hbus,
1801 struct pci_bus_relations *relations)
1803 struct hv_dr_state *dr;
1804 struct hv_dr_work *dr_wrk;
1805 unsigned long flags;
1808 dr_wrk = kzalloc(sizeof(*dr_wrk), GFP_NOWAIT);
1812 dr = kzalloc(offsetof(struct hv_dr_state, func) +
1813 (sizeof(struct pci_function_description) *
1814 (relations->device_count)), GFP_NOWAIT);
1820 INIT_WORK(&dr_wrk->wrk, pci_devices_present_work);
1822 dr->device_count = relations->device_count;
1823 if (dr->device_count != 0) {
1824 memcpy(dr->func, relations->func,
1825 sizeof(struct pci_function_description) *
1829 spin_lock_irqsave(&hbus->device_list_lock, flags);
1831 * If pending_dr is true, we have already queued a work,
1832 * which will see the new dr. Otherwise, we need to
1835 pending_dr = !list_empty(&hbus->dr_list);
1836 list_add_tail(&dr->list_entry, &hbus->dr_list);
1837 spin_unlock_irqrestore(&hbus->device_list_lock, flags);
1843 queue_work(hbus->wq, &dr_wrk->wrk);
1848 * hv_eject_device_work() - Asynchronously handles ejection
1849 * @work: Work struct embedded in internal device struct
1851 * This function handles ejecting a device. Windows will
1852 * attempt to gracefully eject a device, waiting 60 seconds to
1853 * hear back from the guest OS that this completed successfully.
1854 * If this timer expires, the device will be forcibly removed.
1856 static void hv_eject_device_work(struct work_struct *work)
1858 struct pci_eject_response *ejct_pkt;
1859 struct hv_pci_dev *hpdev;
1860 struct pci_dev *pdev;
1861 unsigned long flags;
1864 struct pci_packet pkt;
1865 u8 buffer[sizeof(struct pci_eject_response)];
1868 hpdev = container_of(work, struct hv_pci_dev, wrk);
1870 WARN_ON(hpdev->state != hv_pcichild_ejecting);
1873 * Ejection can come before or after the PCI bus has been set up, so
1874 * attempt to find it and tear down the bus state, if it exists. This
1875 * must be done without constructs like pci_domain_nr(hbus->pci_bus)
1876 * because hbus->pci_bus may not exist yet.
1878 wslot = wslot_to_devfn(hpdev->desc.win_slot.slot);
1879 pdev = pci_get_domain_bus_and_slot(hpdev->hbus->sysdata.domain, 0,
1882 pci_lock_rescan_remove();
1883 pci_stop_and_remove_bus_device(pdev);
1885 pci_unlock_rescan_remove();
1888 spin_lock_irqsave(&hpdev->hbus->device_list_lock, flags);
1889 list_del(&hpdev->list_entry);
1890 spin_unlock_irqrestore(&hpdev->hbus->device_list_lock, flags);
1892 if (hpdev->pci_slot)
1893 pci_destroy_slot(hpdev->pci_slot);
1895 memset(&ctxt, 0, sizeof(ctxt));
1896 ejct_pkt = (struct pci_eject_response *)&ctxt.pkt.message;
1897 ejct_pkt->message_type.type = PCI_EJECTION_COMPLETE;
1898 ejct_pkt->wslot.slot = hpdev->desc.win_slot.slot;
1899 vmbus_sendpacket(hpdev->hbus->hdev->channel, ejct_pkt,
1900 sizeof(*ejct_pkt), (unsigned long)&ctxt.pkt,
1901 VM_PKT_DATA_INBAND, 0);
1903 put_pcichild(hpdev);
1904 put_pcichild(hpdev);
1905 put_hvpcibus(hpdev->hbus);
1909 * hv_pci_eject_device() - Handles device ejection
1910 * @hpdev: Internal device tracking struct
1912 * This function is invoked when an ejection packet arrives. It
1913 * just schedules work so that we don't re-enter the packet
1914 * delivery code handling the ejection.
1916 static void hv_pci_eject_device(struct hv_pci_dev *hpdev)
1918 hpdev->state = hv_pcichild_ejecting;
1919 get_pcichild(hpdev);
1920 INIT_WORK(&hpdev->wrk, hv_eject_device_work);
1921 get_hvpcibus(hpdev->hbus);
1922 queue_work(hpdev->hbus->wq, &hpdev->wrk);
1926 * hv_pci_onchannelcallback() - Handles incoming packets
1927 * @context: Internal bus tracking struct
1929 * This function is invoked whenever the host sends a packet to
1930 * this channel (which is private to this root PCI bus).
1932 static void hv_pci_onchannelcallback(void *context)
1934 const int packet_size = 0x100;
1936 struct hv_pcibus_device *hbus = context;
1939 struct vmpacket_descriptor *desc;
1940 unsigned char *buffer;
1941 int bufferlen = packet_size;
1942 struct pci_packet *comp_packet;
1943 struct pci_response *response;
1944 struct pci_incoming_message *new_message;
1945 struct pci_bus_relations *bus_rel;
1946 struct pci_dev_incoming *dev_message;
1947 struct hv_pci_dev *hpdev;
1949 buffer = kmalloc(bufferlen, GFP_ATOMIC);
1954 ret = vmbus_recvpacket_raw(hbus->hdev->channel, buffer,
1955 bufferlen, &bytes_recvd, &req_id);
1957 if (ret == -ENOBUFS) {
1959 /* Handle large packet */
1960 bufferlen = bytes_recvd;
1961 buffer = kmalloc(bytes_recvd, GFP_ATOMIC);
1967 /* Zero length indicates there are no more packets. */
1968 if (ret || !bytes_recvd)
1972 * All incoming packets must be at least as large as a
1975 if (bytes_recvd <= sizeof(struct pci_response))
1977 desc = (struct vmpacket_descriptor *)buffer;
1979 switch (desc->type) {
1983 * The host is trusted, and thus it's safe to interpret
1984 * this transaction ID as a pointer.
1986 comp_packet = (struct pci_packet *)req_id;
1987 response = (struct pci_response *)buffer;
1988 comp_packet->completion_func(comp_packet->compl_ctxt,
1993 case VM_PKT_DATA_INBAND:
1995 new_message = (struct pci_incoming_message *)buffer;
1996 switch (new_message->message_type.type) {
1997 case PCI_BUS_RELATIONS:
1999 bus_rel = (struct pci_bus_relations *)buffer;
2001 offsetof(struct pci_bus_relations, func) +
2002 (sizeof(struct pci_function_description) *
2003 (bus_rel->device_count))) {
2004 dev_err(&hbus->hdev->device,
2005 "bus relations too small\n");
2009 hv_pci_devices_present(hbus, bus_rel);
2014 dev_message = (struct pci_dev_incoming *)buffer;
2015 hpdev = get_pcichild_wslot(hbus,
2016 dev_message->wslot.slot);
2018 hv_pci_eject_device(hpdev);
2019 put_pcichild(hpdev);
2024 dev_warn(&hbus->hdev->device,
2025 "Unimplemented protocol message %x\n",
2026 new_message->message_type.type);
2032 dev_err(&hbus->hdev->device,
2033 "unhandled packet type %d, tid %llx len %d\n",
2034 desc->type, req_id, bytes_recvd);
2043 * hv_pci_protocol_negotiation() - Set up protocol
2044 * @hdev: VMBus's tracking struct for this root PCI bus
2046 * This driver is intended to support running on Windows 10
2047 * (server) and later versions. It will not run on earlier
2048 * versions, as they assume that many of the operations which
2049 * Linux needs accomplished with a spinlock held were done via
2050 * asynchronous messaging via VMBus. Windows 10 increases the
2051 * surface area of PCI emulation so that these actions can take
2052 * place by suspending a virtual processor for their duration.
2054 * This function negotiates the channel protocol version,
2055 * failing if the host doesn't support the necessary protocol
2058 static int hv_pci_protocol_negotiation(struct hv_device *hdev)
2060 struct pci_version_request *version_req;
2061 struct hv_pci_compl comp_pkt;
2062 struct pci_packet *pkt;
2067 * Initiate the handshake with the host and negotiate
2068 * a version that the host can support. We start with the
2069 * highest version number and go down if the host cannot
2072 pkt = kzalloc(sizeof(*pkt) + sizeof(*version_req), GFP_KERNEL);
2076 init_completion(&comp_pkt.host_event);
2077 pkt->completion_func = hv_pci_generic_compl;
2078 pkt->compl_ctxt = &comp_pkt;
2079 version_req = (struct pci_version_request *)&pkt->message;
2080 version_req->message_type.type = PCI_QUERY_PROTOCOL_VERSION;
2082 for (i = 0; i < ARRAY_SIZE(pci_protocol_versions); i++) {
2083 version_req->protocol_version = pci_protocol_versions[i];
2084 ret = vmbus_sendpacket(hdev->channel, version_req,
2085 sizeof(struct pci_version_request),
2086 (unsigned long)pkt, VM_PKT_DATA_INBAND,
2087 VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
2089 ret = wait_for_response(hdev, &comp_pkt.host_event);
2092 dev_err(&hdev->device,
2093 "PCI Pass-through VSP failed to request version: %d",
2098 if (comp_pkt.completion_status >= 0) {
2099 pci_protocol_version = pci_protocol_versions[i];
2100 dev_info(&hdev->device,
2101 "PCI VMBus probing: Using version %#x\n",
2102 pci_protocol_version);
2106 if (comp_pkt.completion_status != STATUS_REVISION_MISMATCH) {
2107 dev_err(&hdev->device,
2108 "PCI Pass-through VSP failed version request: %#x",
2109 comp_pkt.completion_status);
2114 reinit_completion(&comp_pkt.host_event);
2117 dev_err(&hdev->device,
2118 "PCI pass-through VSP failed to find supported version");
2127 * hv_pci_free_bridge_windows() - Release memory regions for the
2129 * @hbus: Root PCI bus, as understood by this driver
2131 static void hv_pci_free_bridge_windows(struct hv_pcibus_device *hbus)
2134 * Set the resources back to the way they looked when they
2135 * were allocated by setting IORESOURCE_BUSY again.
2138 if (hbus->low_mmio_space && hbus->low_mmio_res) {
2139 hbus->low_mmio_res->flags |= IORESOURCE_BUSY;
2140 vmbus_free_mmio(hbus->low_mmio_res->start,
2141 resource_size(hbus->low_mmio_res));
2144 if (hbus->high_mmio_space && hbus->high_mmio_res) {
2145 hbus->high_mmio_res->flags |= IORESOURCE_BUSY;
2146 vmbus_free_mmio(hbus->high_mmio_res->start,
2147 resource_size(hbus->high_mmio_res));
2152 * hv_pci_allocate_bridge_windows() - Allocate memory regions
2154 * @hbus: Root PCI bus, as understood by this driver
2156 * This function calls vmbus_allocate_mmio(), which is itself a
2157 * bit of a compromise. Ideally, we might change the pnp layer
2158 * in the kernel such that it comprehends either PCI devices
2159 * which are "grandchildren of ACPI," with some intermediate bus
2160 * node (in this case, VMBus) or change it such that it
2161 * understands VMBus. The pnp layer, however, has been declared
2162 * deprecated, and not subject to change.
2164 * The workaround, implemented here, is to ask VMBus to allocate
2165 * MMIO space for this bus. VMBus itself knows which ranges are
2166 * appropriate by looking at its own ACPI objects. Then, after
2167 * these ranges are claimed, they're modified to look like they
2168 * would have looked if the ACPI and pnp code had allocated
2169 * bridge windows. These descriptors have to exist in this form
2170 * in order to satisfy the code which will get invoked when the
2171 * endpoint PCI function driver calls request_mem_region() or
2172 * request_mem_region_exclusive().
2174 * Return: 0 on success, -errno on failure
2176 static int hv_pci_allocate_bridge_windows(struct hv_pcibus_device *hbus)
2178 resource_size_t align;
2181 if (hbus->low_mmio_space) {
2182 align = 1ULL << (63 - __builtin_clzll(hbus->low_mmio_space));
2183 ret = vmbus_allocate_mmio(&hbus->low_mmio_res, hbus->hdev, 0,
2184 (u64)(u32)0xffffffff,
2185 hbus->low_mmio_space,
2188 dev_err(&hbus->hdev->device,
2189 "Need %#llx of low MMIO space. Consider reconfiguring the VM.\n",
2190 hbus->low_mmio_space);
2194 /* Modify this resource to become a bridge window. */
2195 hbus->low_mmio_res->flags |= IORESOURCE_WINDOW;
2196 hbus->low_mmio_res->flags &= ~IORESOURCE_BUSY;
2197 pci_add_resource(&hbus->resources_for_children,
2198 hbus->low_mmio_res);
2201 if (hbus->high_mmio_space) {
2202 align = 1ULL << (63 - __builtin_clzll(hbus->high_mmio_space));
2203 ret = vmbus_allocate_mmio(&hbus->high_mmio_res, hbus->hdev,
2205 hbus->high_mmio_space, align,
2208 dev_err(&hbus->hdev->device,
2209 "Need %#llx of high MMIO space. Consider reconfiguring the VM.\n",
2210 hbus->high_mmio_space);
2211 goto release_low_mmio;
2214 /* Modify this resource to become a bridge window. */
2215 hbus->high_mmio_res->flags |= IORESOURCE_WINDOW;
2216 hbus->high_mmio_res->flags &= ~IORESOURCE_BUSY;
2217 pci_add_resource(&hbus->resources_for_children,
2218 hbus->high_mmio_res);
2224 if (hbus->low_mmio_res) {
2225 vmbus_free_mmio(hbus->low_mmio_res->start,
2226 resource_size(hbus->low_mmio_res));
2233 * hv_allocate_config_window() - Find MMIO space for PCI Config
2234 * @hbus: Root PCI bus, as understood by this driver
2236 * This function claims memory-mapped I/O space for accessing
2237 * configuration space for the functions on this bus.
2239 * Return: 0 on success, -errno on failure
2241 static int hv_allocate_config_window(struct hv_pcibus_device *hbus)
2246 * Set up a region of MMIO space to use for accessing configuration
2249 ret = vmbus_allocate_mmio(&hbus->mem_config, hbus->hdev, 0, -1,
2250 PCI_CONFIG_MMIO_LENGTH, 0x1000, false);
2255 * vmbus_allocate_mmio() gets used for allocating both device endpoint
2256 * resource claims (those which cannot be overlapped) and the ranges
2257 * which are valid for the children of this bus, which are intended
2258 * to be overlapped by those children. Set the flag on this claim
2259 * meaning that this region can't be overlapped.
2262 hbus->mem_config->flags |= IORESOURCE_BUSY;
2267 static void hv_free_config_window(struct hv_pcibus_device *hbus)
2269 vmbus_free_mmio(hbus->mem_config->start, PCI_CONFIG_MMIO_LENGTH);
2273 * hv_pci_enter_d0() - Bring the "bus" into the D0 power state
2274 * @hdev: VMBus's tracking struct for this root PCI bus
2276 * Return: 0 on success, -errno on failure
2278 static int hv_pci_enter_d0(struct hv_device *hdev)
2280 struct hv_pcibus_device *hbus = hv_get_drvdata(hdev);
2281 struct pci_bus_d0_entry *d0_entry;
2282 struct hv_pci_compl comp_pkt;
2283 struct pci_packet *pkt;
2287 * Tell the host that the bus is ready to use, and moved into the
2288 * powered-on state. This includes telling the host which region
2289 * of memory-mapped I/O space has been chosen for configuration space
2292 pkt = kzalloc(sizeof(*pkt) + sizeof(*d0_entry), GFP_KERNEL);
2296 init_completion(&comp_pkt.host_event);
2297 pkt->completion_func = hv_pci_generic_compl;
2298 pkt->compl_ctxt = &comp_pkt;
2299 d0_entry = (struct pci_bus_d0_entry *)&pkt->message;
2300 d0_entry->message_type.type = PCI_BUS_D0ENTRY;
2301 d0_entry->mmio_base = hbus->mem_config->start;
2303 ret = vmbus_sendpacket(hdev->channel, d0_entry, sizeof(*d0_entry),
2304 (unsigned long)pkt, VM_PKT_DATA_INBAND,
2305 VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
2307 ret = wait_for_response(hdev, &comp_pkt.host_event);
2312 if (comp_pkt.completion_status < 0) {
2313 dev_err(&hdev->device,
2314 "PCI Pass-through VSP failed D0 Entry with status %x\n",
2315 comp_pkt.completion_status);
2328 * hv_pci_query_relations() - Ask host to send list of child
2330 * @hdev: VMBus's tracking struct for this root PCI bus
2332 * Return: 0 on success, -errno on failure
2334 static int hv_pci_query_relations(struct hv_device *hdev)
2336 struct hv_pcibus_device *hbus = hv_get_drvdata(hdev);
2337 struct pci_message message;
2338 struct completion comp;
2341 /* Ask the host to send along the list of child devices */
2342 init_completion(&comp);
2343 if (cmpxchg(&hbus->survey_event, NULL, &comp))
2346 memset(&message, 0, sizeof(message));
2347 message.type = PCI_QUERY_BUS_RELATIONS;
2349 ret = vmbus_sendpacket(hdev->channel, &message, sizeof(message),
2350 0, VM_PKT_DATA_INBAND, 0);
2352 ret = wait_for_response(hdev, &comp);
2358 * hv_send_resources_allocated() - Report local resource choices
2359 * @hdev: VMBus's tracking struct for this root PCI bus
2361 * The host OS is expecting to be sent a request as a message
2362 * which contains all the resources that the device will use.
2363 * The response contains those same resources, "translated"
2364 * which is to say, the values which should be used by the
2365 * hardware, when it delivers an interrupt. (MMIO resources are
2366 * used in local terms.) This is nice for Windows, and lines up
2367 * with the FDO/PDO split, which doesn't exist in Linux. Linux
2368 * is deeply expecting to scan an emulated PCI configuration
2369 * space. So this message is sent here only to drive the state
2370 * machine on the host forward.
2372 * Return: 0 on success, -errno on failure
2374 static int hv_send_resources_allocated(struct hv_device *hdev)
2376 struct hv_pcibus_device *hbus = hv_get_drvdata(hdev);
2377 struct pci_resources_assigned *res_assigned;
2378 struct pci_resources_assigned2 *res_assigned2;
2379 struct hv_pci_compl comp_pkt;
2380 struct hv_pci_dev *hpdev;
2381 struct pci_packet *pkt;
2386 size_res = (pci_protocol_version < PCI_PROTOCOL_VERSION_1_2)
2387 ? sizeof(*res_assigned) : sizeof(*res_assigned2);
2389 pkt = kmalloc(sizeof(*pkt) + size_res, GFP_KERNEL);
2395 for (wslot = 0; wslot < 256; wslot++) {
2396 hpdev = get_pcichild_wslot(hbus, wslot);
2400 memset(pkt, 0, sizeof(*pkt) + size_res);
2401 init_completion(&comp_pkt.host_event);
2402 pkt->completion_func = hv_pci_generic_compl;
2403 pkt->compl_ctxt = &comp_pkt;
2405 if (pci_protocol_version < PCI_PROTOCOL_VERSION_1_2) {
2407 (struct pci_resources_assigned *)&pkt->message;
2408 res_assigned->message_type.type =
2409 PCI_RESOURCES_ASSIGNED;
2410 res_assigned->wslot.slot = hpdev->desc.win_slot.slot;
2413 (struct pci_resources_assigned2 *)&pkt->message;
2414 res_assigned2->message_type.type =
2415 PCI_RESOURCES_ASSIGNED2;
2416 res_assigned2->wslot.slot = hpdev->desc.win_slot.slot;
2418 put_pcichild(hpdev);
2420 ret = vmbus_sendpacket(hdev->channel, &pkt->message,
2421 size_res, (unsigned long)pkt,
2423 VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
2425 ret = wait_for_response(hdev, &comp_pkt.host_event);
2429 if (comp_pkt.completion_status < 0) {
2431 dev_err(&hdev->device,
2432 "resource allocated returned 0x%x",
2433 comp_pkt.completion_status);
2443 * hv_send_resources_released() - Report local resources
2445 * @hdev: VMBus's tracking struct for this root PCI bus
2447 * Return: 0 on success, -errno on failure
2449 static int hv_send_resources_released(struct hv_device *hdev)
2451 struct hv_pcibus_device *hbus = hv_get_drvdata(hdev);
2452 struct pci_child_message pkt;
2453 struct hv_pci_dev *hpdev;
2457 for (wslot = 0; wslot < 256; wslot++) {
2458 hpdev = get_pcichild_wslot(hbus, wslot);
2462 memset(&pkt, 0, sizeof(pkt));
2463 pkt.message_type.type = PCI_RESOURCES_RELEASED;
2464 pkt.wslot.slot = hpdev->desc.win_slot.slot;
2466 put_pcichild(hpdev);
2468 ret = vmbus_sendpacket(hdev->channel, &pkt, sizeof(pkt), 0,
2469 VM_PKT_DATA_INBAND, 0);
2477 static void get_hvpcibus(struct hv_pcibus_device *hbus)
2479 refcount_inc(&hbus->remove_lock);
2482 static void put_hvpcibus(struct hv_pcibus_device *hbus)
2484 if (refcount_dec_and_test(&hbus->remove_lock))
2485 complete(&hbus->remove_event);
2489 * hv_pci_probe() - New VMBus channel probe, for a root PCI bus
2490 * @hdev: VMBus's tracking struct for this root PCI bus
2491 * @dev_id: Identifies the device itself
2493 * Return: 0 on success, -errno on failure
2495 static int hv_pci_probe(struct hv_device *hdev,
2496 const struct hv_vmbus_device_id *dev_id)
2498 struct hv_pcibus_device *hbus;
2502 * hv_pcibus_device contains the hypercall arguments for retargeting in
2503 * hv_irq_unmask(). Those must not cross a page boundary.
2505 BUILD_BUG_ON(sizeof(*hbus) > PAGE_SIZE);
2507 hbus = (struct hv_pcibus_device *)get_zeroed_page(GFP_KERNEL);
2510 hbus->state = hv_pcibus_init;
2513 * The PCI bus "domain" is what is called "segment" in ACPI and
2514 * other specs. Pull it from the instance ID, to get something
2515 * unique. Bytes 8 and 9 are what is used in Windows guests, so
2516 * do the same thing for consistency. Note that, since this code
2517 * only runs in a Hyper-V VM, Hyper-V can (and does) guarantee
2518 * that (1) the only domain in use for something that looks like
2519 * a physical PCI bus (which is actually emulated by the
2520 * hypervisor) is domain 0 and (2) there will be no overlap
2521 * between domains derived from these instance IDs in the same
2524 hbus->sysdata.domain = hdev->dev_instance.b[9] |
2525 hdev->dev_instance.b[8] << 8;
2528 refcount_set(&hbus->remove_lock, 1);
2529 INIT_LIST_HEAD(&hbus->children);
2530 INIT_LIST_HEAD(&hbus->dr_list);
2531 INIT_LIST_HEAD(&hbus->resources_for_children);
2532 spin_lock_init(&hbus->config_lock);
2533 spin_lock_init(&hbus->device_list_lock);
2534 spin_lock_init(&hbus->retarget_msi_interrupt_lock);
2535 init_completion(&hbus->remove_event);
2536 hbus->wq = alloc_ordered_workqueue("hv_pci_%x", 0,
2537 hbus->sysdata.domain);
2543 ret = vmbus_open(hdev->channel, pci_ring_size, pci_ring_size, NULL, 0,
2544 hv_pci_onchannelcallback, hbus);
2548 hv_set_drvdata(hdev, hbus);
2550 ret = hv_pci_protocol_negotiation(hdev);
2554 ret = hv_allocate_config_window(hbus);
2558 hbus->cfg_addr = ioremap(hbus->mem_config->start,
2559 PCI_CONFIG_MMIO_LENGTH);
2560 if (!hbus->cfg_addr) {
2561 dev_err(&hdev->device,
2562 "Unable to map a virtual address for config space\n");
2567 hbus->sysdata.fwnode = irq_domain_alloc_fwnode(hbus);
2568 if (!hbus->sysdata.fwnode) {
2573 ret = hv_pcie_init_irq_domain(hbus);
2577 ret = hv_pci_query_relations(hdev);
2579 goto free_irq_domain;
2581 ret = hv_pci_enter_d0(hdev);
2583 goto free_irq_domain;
2585 ret = hv_pci_allocate_bridge_windows(hbus);
2587 goto free_irq_domain;
2589 ret = hv_send_resources_allocated(hdev);
2593 prepopulate_bars(hbus);
2595 hbus->state = hv_pcibus_probed;
2597 ret = create_root_hv_pci_bus(hbus);
2604 hv_pci_free_bridge_windows(hbus);
2606 irq_domain_remove(hbus->irq_domain);
2608 irq_domain_free_fwnode(hbus->sysdata.fwnode);
2610 iounmap(hbus->cfg_addr);
2612 hv_free_config_window(hbus);
2614 vmbus_close(hdev->channel);
2616 destroy_workqueue(hbus->wq);
2618 free_page((unsigned long)hbus);
2622 static void hv_pci_bus_exit(struct hv_device *hdev)
2624 struct hv_pcibus_device *hbus = hv_get_drvdata(hdev);
2626 struct pci_packet teardown_packet;
2627 u8 buffer[sizeof(struct pci_message)];
2629 struct pci_bus_relations relations;
2630 struct hv_pci_compl comp_pkt;
2634 * After the host sends the RESCIND_CHANNEL message, it doesn't
2635 * access the per-channel ringbuffer any longer.
2637 if (hdev->channel->rescind)
2640 /* Delete any children which might still exist. */
2641 memset(&relations, 0, sizeof(relations));
2642 hv_pci_devices_present(hbus, &relations);
2644 ret = hv_send_resources_released(hdev);
2646 dev_err(&hdev->device,
2647 "Couldn't send resources released packet(s)\n");
2649 memset(&pkt.teardown_packet, 0, sizeof(pkt.teardown_packet));
2650 init_completion(&comp_pkt.host_event);
2651 pkt.teardown_packet.completion_func = hv_pci_generic_compl;
2652 pkt.teardown_packet.compl_ctxt = &comp_pkt;
2653 pkt.teardown_packet.message[0].type = PCI_BUS_D0EXIT;
2655 ret = vmbus_sendpacket(hdev->channel, &pkt.teardown_packet.message,
2656 sizeof(struct pci_message),
2657 (unsigned long)&pkt.teardown_packet,
2659 VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
2661 wait_for_completion_timeout(&comp_pkt.host_event, 10 * HZ);
2665 * hv_pci_remove() - Remove routine for this VMBus channel
2666 * @hdev: VMBus's tracking struct for this root PCI bus
2668 * Return: 0 on success, -errno on failure
2670 static int hv_pci_remove(struct hv_device *hdev)
2672 struct hv_pcibus_device *hbus;
2674 hbus = hv_get_drvdata(hdev);
2675 if (hbus->state == hv_pcibus_installed) {
2676 /* Remove the bus from PCI's point of view. */
2677 pci_lock_rescan_remove();
2678 pci_stop_root_bus(hbus->pci_bus);
2679 pci_remove_root_bus(hbus->pci_bus);
2680 pci_unlock_rescan_remove();
2681 hbus->state = hv_pcibus_removed;
2684 hv_pci_bus_exit(hdev);
2686 vmbus_close(hdev->channel);
2688 iounmap(hbus->cfg_addr);
2689 hv_free_config_window(hbus);
2690 pci_free_resource_list(&hbus->resources_for_children);
2691 hv_pci_free_bridge_windows(hbus);
2692 irq_domain_remove(hbus->irq_domain);
2693 irq_domain_free_fwnode(hbus->sysdata.fwnode);
2695 wait_for_completion(&hbus->remove_event);
2696 destroy_workqueue(hbus->wq);
2697 free_page((unsigned long)hbus);
2701 static const struct hv_vmbus_device_id hv_pci_id_table[] = {
2702 /* PCI Pass-through Class ID */
2703 /* 44C4F61D-4444-4400-9D52-802E27EDE19F */
2708 MODULE_DEVICE_TABLE(vmbus, hv_pci_id_table);
2710 static struct hv_driver hv_pci_drv = {
2712 .id_table = hv_pci_id_table,
2713 .probe = hv_pci_probe,
2714 .remove = hv_pci_remove,
2717 static void __exit exit_hv_pci_drv(void)
2719 vmbus_driver_unregister(&hv_pci_drv);
2722 static int __init init_hv_pci_drv(void)
2724 return vmbus_driver_register(&hv_pci_drv);
2727 module_init(init_hv_pci_drv);
2728 module_exit(exit_hv_pci_drv);
2730 MODULE_DESCRIPTION("Hyper-V PCI");
2731 MODULE_LICENSE("GPL v2");