2 * Copyright (c) Microsoft Corporation.
5 * Jake Oshins <jakeo@microsoft.com>
7 * This driver acts as a paravirtual front-end for PCI Express root buses.
8 * When a PCI Express function (either an entire device or an SR-IOV
9 * Virtual Function) is being passed through to the VM, this driver exposes
10 * a new bus to the guest VM. This is modeled as a root PCI bus because
11 * no bridges are being exposed to the VM. In fact, with a "Generation 2"
12 * VM within Hyper-V, there may seem to be no PCI bus at all in the VM
13 * until a device as been exposed using this driver.
15 * Each root PCI bus has its own PCI domain, which is called "Segment" in
16 * the PCI Firmware Specifications. Thus while each device passed through
17 * to the VM using this front-end will appear at "device 0", the domain will
18 * be unique. Typically, each bus will have one PCI function on it, though
19 * this driver does support more than one.
21 * In order to map the interrupts from the device through to the guest VM,
22 * this driver also implements an IRQ Domain, which handles interrupts (either
23 * MSI or MSI-X) associated with the functions on the bus. As interrupts are
24 * set up, torn down, or reaffined, this driver communicates with the
25 * underlying hypervisor to adjust the mappings in the I/O MMU so that each
26 * interrupt will be delivered to the correct virtual processor at the right
27 * vector. This driver does not support level-triggered (line-based)
28 * interrupts, and will report that the Interrupt Line register in the
29 * function's configuration space is zero.
31 * The rest of this driver mostly maps PCI concepts onto underlying Hyper-V
32 * facilities. For instance, the configuration space of a function exposed
33 * by Hyper-V is mapped into a single page of memory space, and the
34 * read and write handlers for config space must be aware of this mechanism.
35 * Similarly, device setup and teardown involves messages sent to and from
36 * the PCI back-end driver in Hyper-V.
38 * This program is free software; you can redistribute it and/or modify it
39 * under the terms of the GNU General Public License version 2 as published
40 * by the Free Software Foundation.
42 * This program is distributed in the hope that it will be useful, but
43 * WITHOUT ANY WARRANTY; without even the implied warranty of
44 * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
45 * NON INFRINGEMENT. See the GNU General Public License for more
50 #include <linux/kernel.h>
51 #include <linux/module.h>
52 #include <linux/pci.h>
53 #include <linux/semaphore.h>
54 #include <linux/irqdomain.h>
55 #include <asm/irqdomain.h>
57 #include <linux/msi.h>
58 #include <linux/hyperv.h>
59 #include <asm/mshyperv.h>
62 * Protocol versions. The low word is the minor version, the high word the
66 #define PCI_MAKE_VERSION(major, minor) ((u32)(((major) << 16) | (major)))
67 #define PCI_MAJOR_VERSION(version) ((u32)(version) >> 16)
68 #define PCI_MINOR_VERSION(version) ((u32)(version) & 0xff)
71 PCI_PROTOCOL_VERSION_1_1 = PCI_MAKE_VERSION(1, 1),
72 PCI_PROTOCOL_VERSION_CURRENT = PCI_PROTOCOL_VERSION_1_1
75 #define PCI_CONFIG_MMIO_LENGTH 0x2000
76 #define CFG_PAGE_OFFSET 0x1000
77 #define CFG_PAGE_SIZE (PCI_CONFIG_MMIO_LENGTH - CFG_PAGE_OFFSET)
79 #define MAX_SUPPORTED_MSI_MESSAGES 0x400
85 enum pci_message_type {
89 PCI_MESSAGE_BASE = 0x42490000,
90 PCI_BUS_RELATIONS = PCI_MESSAGE_BASE + 0,
91 PCI_QUERY_BUS_RELATIONS = PCI_MESSAGE_BASE + 1,
92 PCI_POWER_STATE_CHANGE = PCI_MESSAGE_BASE + 4,
93 PCI_QUERY_RESOURCE_REQUIREMENTS = PCI_MESSAGE_BASE + 5,
94 PCI_QUERY_RESOURCE_RESOURCES = PCI_MESSAGE_BASE + 6,
95 PCI_BUS_D0ENTRY = PCI_MESSAGE_BASE + 7,
96 PCI_BUS_D0EXIT = PCI_MESSAGE_BASE + 8,
97 PCI_READ_BLOCK = PCI_MESSAGE_BASE + 9,
98 PCI_WRITE_BLOCK = PCI_MESSAGE_BASE + 0xA,
99 PCI_EJECT = PCI_MESSAGE_BASE + 0xB,
100 PCI_QUERY_STOP = PCI_MESSAGE_BASE + 0xC,
101 PCI_REENABLE = PCI_MESSAGE_BASE + 0xD,
102 PCI_QUERY_STOP_FAILED = PCI_MESSAGE_BASE + 0xE,
103 PCI_EJECTION_COMPLETE = PCI_MESSAGE_BASE + 0xF,
104 PCI_RESOURCES_ASSIGNED = PCI_MESSAGE_BASE + 0x10,
105 PCI_RESOURCES_RELEASED = PCI_MESSAGE_BASE + 0x11,
106 PCI_INVALIDATE_BLOCK = PCI_MESSAGE_BASE + 0x12,
107 PCI_QUERY_PROTOCOL_VERSION = PCI_MESSAGE_BASE + 0x13,
108 PCI_CREATE_INTERRUPT_MESSAGE = PCI_MESSAGE_BASE + 0x14,
109 PCI_DELETE_INTERRUPT_MESSAGE = PCI_MESSAGE_BASE + 0x15,
114 * Structures defining the virtual PCI Express protocol.
126 * Function numbers are 8-bits wide on Express, as interpreted through ARI,
127 * which is all this driver does. This representation is the one used in
128 * Windows, which is what is expected when sending this back and forth with
129 * the Hyper-V parent partition.
131 union win_slot_encoding {
140 * Pretty much as defined in the PCI Specifications.
142 struct pci_function_description {
143 u16 v_id; /* vendor ID */
144 u16 d_id; /* device ID */
150 union win_slot_encoding win_slot;
151 u32 ser; /* serial number */
157 * @delivery_mode: As defined in Intel's Programmer's
158 * Reference Manual, Volume 3, Chapter 8.
159 * @vector_count: Number of contiguous entries in the
160 * Interrupt Descriptor Table that are
161 * occupied by this Message-Signaled
162 * Interrupt. For "MSI", as first defined
163 * in PCI 2.2, this can be between 1 and
164 * 32. For "MSI-X," as first defined in PCI
165 * 3.0, this must be 1, as each MSI-X table
166 * entry would have its own descriptor.
167 * @reserved: Empty space
168 * @cpu_mask: All the target virtual processors.
179 * struct tran_int_desc
180 * @reserved: unused, padding
181 * @vector_count: same as in hv_msi_desc
182 * @data: This is the "data payload" value that is
183 * written by the device when it generates
184 * a message-signaled interrupt, either MSI
186 * @address: This is the address to which the data
187 * payload is written on interrupt
190 struct tran_int_desc {
198 * A generic message format for virtual PCI.
199 * Specific message formats are defined later in the file.
206 struct pci_child_message {
208 union win_slot_encoding wslot;
211 struct pci_incoming_message {
212 struct vmpacket_descriptor hdr;
213 struct pci_message message_type;
216 struct pci_response {
217 struct vmpacket_descriptor hdr;
218 s32 status; /* negative values are failures */
222 void (*completion_func)(void *context, struct pci_response *resp,
223 int resp_packet_size);
225 struct pci_message message;
229 * Specific message types supporting the PCI protocol.
233 * Version negotiation message. Sent from the guest to the host.
234 * The guest is free to try different versions until the host
235 * accepts the version.
237 * pci_version: The protocol version requested.
238 * is_last_attempt: If TRUE, this is the last version guest will request.
239 * reservedz: Reserved field, set to zero.
242 struct pci_version_request {
243 struct pci_message message_type;
244 enum pci_message_type protocol_version;
248 * Bus D0 Entry. This is sent from the guest to the host when the virtual
249 * bus (PCI Express port) is ready for action.
252 struct pci_bus_d0_entry {
253 struct pci_message message_type;
258 struct pci_bus_relations {
259 struct pci_incoming_message incoming;
261 struct pci_function_description func[1];
264 struct pci_q_res_req_response {
265 struct vmpacket_descriptor hdr;
266 s32 status; /* negative values are failures */
270 struct pci_set_power {
271 struct pci_message message_type;
272 union win_slot_encoding wslot;
273 u32 power_state; /* In Windows terms */
277 struct pci_set_power_response {
278 struct vmpacket_descriptor hdr;
279 s32 status; /* negative values are failures */
280 union win_slot_encoding wslot;
281 u32 resultant_state; /* In Windows terms */
285 struct pci_resources_assigned {
286 struct pci_message message_type;
287 union win_slot_encoding wslot;
288 u8 memory_range[0x14][6]; /* not used here */
293 struct pci_create_interrupt {
294 struct pci_message message_type;
295 union win_slot_encoding wslot;
296 struct hv_msi_desc int_desc;
299 struct pci_create_int_response {
300 struct pci_response response;
302 struct tran_int_desc int_desc;
305 struct pci_delete_interrupt {
306 struct pci_message message_type;
307 union win_slot_encoding wslot;
308 struct tran_int_desc int_desc;
311 struct pci_dev_incoming {
312 struct pci_incoming_message incoming;
313 union win_slot_encoding wslot;
316 struct pci_eject_response {
318 union win_slot_encoding wslot;
322 static int pci_ring_size = (4 * PAGE_SIZE);
325 * Definitions or interrupt steering hypercall.
327 #define HV_PARTITION_ID_SELF ((u64)-1)
328 #define HVCALL_RETARGET_INTERRUPT 0x7e
330 struct retarget_msi_interrupt {
331 u64 partition_id; /* use "self" */
333 u32 source; /* 1 for MSI(-X) */
344 * Driver specific state.
347 enum hv_pcibus_state {
354 struct hv_pcibus_device {
355 struct pci_sysdata sysdata;
356 enum hv_pcibus_state state;
357 atomic_t remove_lock;
358 struct hv_device *hdev;
359 resource_size_t low_mmio_space;
360 resource_size_t high_mmio_space;
361 struct resource *mem_config;
362 struct resource *low_mmio_res;
363 struct resource *high_mmio_res;
364 struct completion *survey_event;
365 struct completion remove_event;
366 struct pci_bus *pci_bus;
367 spinlock_t config_lock; /* Avoid two threads writing index page */
368 spinlock_t device_list_lock; /* Protect lists below */
369 void __iomem *cfg_addr;
371 struct semaphore enum_sem;
372 struct list_head resources_for_children;
374 struct list_head children;
375 struct list_head dr_list;
376 struct work_struct wrk;
378 struct msi_domain_info msi_info;
379 struct msi_controller msi_chip;
380 struct irq_domain *irq_domain;
384 * Tracks "Device Relations" messages from the host, which must be both
385 * processed in order and deferred so that they don't run in the context
386 * of the incoming packet callback.
389 struct work_struct wrk;
390 struct hv_pcibus_device *bus;
394 struct list_head list_entry;
396 struct pci_function_description func[1];
399 enum hv_pcichild_state {
400 hv_pcichild_init = 0,
401 hv_pcichild_requirements,
402 hv_pcichild_resourced,
403 hv_pcichild_ejecting,
407 enum hv_pcidev_ref_reason {
408 hv_pcidev_ref_invalid = 0,
409 hv_pcidev_ref_initial,
410 hv_pcidev_ref_by_slot,
411 hv_pcidev_ref_packet,
413 hv_pcidev_ref_childlist,
419 /* List protected by pci_rescan_remove_lock */
420 struct list_head list_entry;
422 enum hv_pcichild_state state;
423 struct pci_function_description desc;
424 bool reported_missing;
425 struct hv_pcibus_device *hbus;
426 struct work_struct wrk;
429 * What would be observed if one wrote 0xFFFFFFFF to a BAR and then
430 * read it back, for each of the BAR offsets within config space.
435 struct hv_pci_compl {
436 struct completion host_event;
437 s32 completion_status;
441 * hv_pci_generic_compl() - Invoked for a completion packet
442 * @context: Set up by the sender of the packet.
443 * @resp: The response packet
444 * @resp_packet_size: Size in bytes of the packet
446 * This function is used to trigger an event and report status
447 * for any message for which the completion packet contains a
448 * status and nothing else.
452 hv_pci_generic_compl(void *context, struct pci_response *resp,
453 int resp_packet_size)
455 struct hv_pci_compl *comp_pkt = context;
457 if (resp_packet_size >= offsetofend(struct pci_response, status))
458 comp_pkt->completion_status = resp->status;
459 complete(&comp_pkt->host_event);
462 static struct hv_pci_dev *get_pcichild_wslot(struct hv_pcibus_device *hbus,
464 static void get_pcichild(struct hv_pci_dev *hv_pcidev,
465 enum hv_pcidev_ref_reason reason);
466 static void put_pcichild(struct hv_pci_dev *hv_pcidev,
467 enum hv_pcidev_ref_reason reason);
469 static void get_hvpcibus(struct hv_pcibus_device *hv_pcibus);
470 static void put_hvpcibus(struct hv_pcibus_device *hv_pcibus);
473 * devfn_to_wslot() - Convert from Linux PCI slot to Windows
474 * @devfn: The Linux representation of PCI slot
476 * Windows uses a slightly different representation of PCI slot.
478 * Return: The Windows representation
480 static u32 devfn_to_wslot(int devfn)
482 union win_slot_encoding wslot;
485 wslot.bits.func = PCI_SLOT(devfn) | (PCI_FUNC(devfn) << 5);
491 * wslot_to_devfn() - Convert from Windows PCI slot to Linux
492 * @wslot: The Windows representation of PCI slot
494 * Windows uses a slightly different representation of PCI slot.
496 * Return: The Linux representation
498 static int wslot_to_devfn(u32 wslot)
500 union win_slot_encoding slot_no;
502 slot_no.slot = wslot;
503 return PCI_DEVFN(0, slot_no.bits.func);
507 * PCI Configuration Space for these root PCI buses is implemented as a pair
508 * of pages in memory-mapped I/O space. Writing to the first page chooses
509 * the PCI function being written or read. Once the first page has been
510 * written to, the following page maps in the entire configuration space of
515 * _hv_pcifront_read_config() - Internal PCI config read
516 * @hpdev: The PCI driver's representation of the device
517 * @where: Offset within config space
518 * @size: Size of the transfer
519 * @val: Pointer to the buffer receiving the data
521 static void _hv_pcifront_read_config(struct hv_pci_dev *hpdev, int where,
525 void __iomem *addr = hpdev->hbus->cfg_addr + CFG_PAGE_OFFSET + where;
528 * If the attempt is to read the IDs or the ROM BAR, simulate that.
530 if (where + size <= PCI_COMMAND) {
531 memcpy(val, ((u8 *)&hpdev->desc.v_id) + where, size);
532 } else if (where >= PCI_CLASS_REVISION && where + size <=
533 PCI_CACHE_LINE_SIZE) {
534 memcpy(val, ((u8 *)&hpdev->desc.rev) + where -
535 PCI_CLASS_REVISION, size);
536 } else if (where >= PCI_SUBSYSTEM_VENDOR_ID && where + size <=
538 memcpy(val, (u8 *)&hpdev->desc.subsystem_id + where -
539 PCI_SUBSYSTEM_VENDOR_ID, size);
540 } else if (where >= PCI_ROM_ADDRESS && where + size <=
541 PCI_CAPABILITY_LIST) {
542 /* ROM BARs are unimplemented */
544 } else if (where >= PCI_INTERRUPT_LINE && where + size <=
547 * Interrupt Line and Interrupt PIN are hard-wired to zero
548 * because this front-end only supports message-signaled
552 } else if (where + size <= CFG_PAGE_SIZE) {
553 spin_lock_irqsave(&hpdev->hbus->config_lock, flags);
554 /* Choose the function to be read. (See comment above) */
555 writel(hpdev->desc.win_slot.slot, hpdev->hbus->cfg_addr);
556 /* Read from that function's config space. */
568 spin_unlock_irqrestore(&hpdev->hbus->config_lock, flags);
570 dev_err(&hpdev->hbus->hdev->device,
571 "Attempt to read beyond a function's config space.\n");
576 * _hv_pcifront_write_config() - Internal PCI config write
577 * @hpdev: The PCI driver's representation of the device
578 * @where: Offset within config space
579 * @size: Size of the transfer
580 * @val: The data being transferred
582 static void _hv_pcifront_write_config(struct hv_pci_dev *hpdev, int where,
586 void __iomem *addr = hpdev->hbus->cfg_addr + CFG_PAGE_OFFSET + where;
588 if (where >= PCI_SUBSYSTEM_VENDOR_ID &&
589 where + size <= PCI_CAPABILITY_LIST) {
590 /* SSIDs and ROM BARs are read-only */
591 } else if (where >= PCI_COMMAND && where + size <= CFG_PAGE_SIZE) {
592 spin_lock_irqsave(&hpdev->hbus->config_lock, flags);
593 /* Choose the function to be written. (See comment above) */
594 writel(hpdev->desc.win_slot.slot, hpdev->hbus->cfg_addr);
595 /* Write to that function's config space. */
607 spin_unlock_irqrestore(&hpdev->hbus->config_lock, flags);
609 dev_err(&hpdev->hbus->hdev->device,
610 "Attempt to write beyond a function's config space.\n");
615 * hv_pcifront_read_config() - Read configuration space
616 * @bus: PCI Bus structure
617 * @devfn: Device/function
618 * @where: Offset from base
619 * @size: Byte/word/dword
620 * @val: Value to be read
622 * Return: PCIBIOS_SUCCESSFUL on success
623 * PCIBIOS_DEVICE_NOT_FOUND on failure
625 static int hv_pcifront_read_config(struct pci_bus *bus, unsigned int devfn,
626 int where, int size, u32 *val)
628 struct hv_pcibus_device *hbus =
629 container_of(bus->sysdata, struct hv_pcibus_device, sysdata);
630 struct hv_pci_dev *hpdev;
632 hpdev = get_pcichild_wslot(hbus, devfn_to_wslot(devfn));
634 return PCIBIOS_DEVICE_NOT_FOUND;
636 _hv_pcifront_read_config(hpdev, where, size, val);
638 put_pcichild(hpdev, hv_pcidev_ref_by_slot);
639 return PCIBIOS_SUCCESSFUL;
643 * hv_pcifront_write_config() - Write configuration space
644 * @bus: PCI Bus structure
645 * @devfn: Device/function
646 * @where: Offset from base
647 * @size: Byte/word/dword
648 * @val: Value to be written to device
650 * Return: PCIBIOS_SUCCESSFUL on success
651 * PCIBIOS_DEVICE_NOT_FOUND on failure
653 static int hv_pcifront_write_config(struct pci_bus *bus, unsigned int devfn,
654 int where, int size, u32 val)
656 struct hv_pcibus_device *hbus =
657 container_of(bus->sysdata, struct hv_pcibus_device, sysdata);
658 struct hv_pci_dev *hpdev;
660 hpdev = get_pcichild_wslot(hbus, devfn_to_wslot(devfn));
662 return PCIBIOS_DEVICE_NOT_FOUND;
664 _hv_pcifront_write_config(hpdev, where, size, val);
666 put_pcichild(hpdev, hv_pcidev_ref_by_slot);
667 return PCIBIOS_SUCCESSFUL;
670 /* PCIe operations */
671 static struct pci_ops hv_pcifront_ops = {
672 .read = hv_pcifront_read_config,
673 .write = hv_pcifront_write_config,
676 /* Interrupt management hooks */
677 static void hv_int_desc_free(struct hv_pci_dev *hpdev,
678 struct tran_int_desc *int_desc)
680 struct pci_delete_interrupt *int_pkt;
682 struct pci_packet pkt;
683 u8 buffer[sizeof(struct pci_delete_interrupt) -
684 sizeof(struct pci_message)];
687 memset(&ctxt, 0, sizeof(ctxt));
688 int_pkt = (struct pci_delete_interrupt *)&ctxt.pkt.message;
689 int_pkt->message_type.message_type =
690 PCI_DELETE_INTERRUPT_MESSAGE;
691 int_pkt->wslot.slot = hpdev->desc.win_slot.slot;
692 int_pkt->int_desc = *int_desc;
693 vmbus_sendpacket(hpdev->hbus->hdev->channel, int_pkt, sizeof(*int_pkt),
694 (unsigned long)&ctxt.pkt, VM_PKT_DATA_INBAND, 0);
699 * hv_msi_free() - Free the MSI.
700 * @domain: The interrupt domain pointer
701 * @info: Extra MSI-related context
702 * @irq: Identifies the IRQ.
704 * The Hyper-V parent partition and hypervisor are tracking the
705 * messages that are in use, keeping the interrupt redirection
706 * table up to date. This callback sends a message that frees
707 * the IRT entry and related tracking nonsense.
709 static void hv_msi_free(struct irq_domain *domain, struct msi_domain_info *info,
712 struct hv_pcibus_device *hbus;
713 struct hv_pci_dev *hpdev;
714 struct pci_dev *pdev;
715 struct tran_int_desc *int_desc;
716 struct irq_data *irq_data = irq_domain_get_irq_data(domain, irq);
717 struct msi_desc *msi = irq_data_get_msi_desc(irq_data);
719 pdev = msi_desc_to_pci_dev(msi);
721 hpdev = get_pcichild_wslot(hbus, devfn_to_wslot(pdev->devfn));
725 int_desc = irq_data_get_irq_chip_data(irq_data);
727 irq_data->chip_data = NULL;
728 hv_int_desc_free(hpdev, int_desc);
731 put_pcichild(hpdev, hv_pcidev_ref_by_slot);
734 static int hv_set_affinity(struct irq_data *data, const struct cpumask *dest,
737 struct irq_data *parent = data->parent_data;
739 return parent->chip->irq_set_affinity(parent, dest, force);
742 void hv_irq_mask(struct irq_data *data)
744 pci_msi_mask_irq(data);
748 * hv_irq_unmask() - "Unmask" the IRQ by setting its current
750 * @data: Describes the IRQ
752 * Build new a destination for the MSI and make a hypercall to
753 * update the Interrupt Redirection Table. "Device Logical ID"
754 * is built out of this PCI bus's instance GUID and the function
755 * number of the device.
757 void hv_irq_unmask(struct irq_data *data)
759 struct msi_desc *msi_desc = irq_data_get_msi_desc(data);
760 struct irq_cfg *cfg = irqd_cfg(data);
761 struct retarget_msi_interrupt params;
762 struct hv_pcibus_device *hbus;
763 struct cpumask *dest;
764 struct pci_bus *pbus;
765 struct pci_dev *pdev;
768 dest = irq_data_get_affinity_mask(data);
769 pdev = msi_desc_to_pci_dev(msi_desc);
771 hbus = container_of(pbus->sysdata, struct hv_pcibus_device, sysdata);
773 memset(¶ms, 0, sizeof(params));
774 params.partition_id = HV_PARTITION_ID_SELF;
775 params.source = 1; /* MSI(-X) */
776 params.address = msi_desc->msg.address_lo;
777 params.data = msi_desc->msg.data;
778 params.device_id = (hbus->hdev->dev_instance.b[5] << 24) |
779 (hbus->hdev->dev_instance.b[4] << 16) |
780 (hbus->hdev->dev_instance.b[7] << 8) |
781 (hbus->hdev->dev_instance.b[6] & 0xf8) |
782 PCI_FUNC(pdev->devfn);
783 params.vector = cfg->vector;
785 for_each_cpu_and(cpu, dest, cpu_online_mask)
786 params.vp_mask |= (1ULL << vmbus_cpu_number_to_vp_number(cpu));
788 hv_do_hypercall(HVCALL_RETARGET_INTERRUPT, ¶ms, NULL);
790 pci_msi_unmask_irq(data);
793 struct compose_comp_ctxt {
794 struct hv_pci_compl comp_pkt;
795 struct tran_int_desc int_desc;
798 static void hv_pci_compose_compl(void *context, struct pci_response *resp,
799 int resp_packet_size)
801 struct compose_comp_ctxt *comp_pkt = context;
802 struct pci_create_int_response *int_resp =
803 (struct pci_create_int_response *)resp;
805 comp_pkt->comp_pkt.completion_status = resp->status;
806 comp_pkt->int_desc = int_resp->int_desc;
807 complete(&comp_pkt->comp_pkt.host_event);
811 * hv_compose_msi_msg() - Supplies a valid MSI address/data
812 * @data: Everything about this MSI
813 * @msg: Buffer that is filled in by this function
815 * This function unpacks the IRQ looking for target CPU set, IDT
816 * vector and mode and sends a message to the parent partition
817 * asking for a mapping for that tuple in this partition. The
818 * response supplies a data value and address to which that data
819 * should be written to trigger that interrupt.
821 static void hv_compose_msi_msg(struct irq_data *data, struct msi_msg *msg)
823 struct irq_cfg *cfg = irqd_cfg(data);
824 struct hv_pcibus_device *hbus;
825 struct hv_pci_dev *hpdev;
826 struct pci_bus *pbus;
827 struct pci_dev *pdev;
828 struct pci_create_interrupt *int_pkt;
829 struct compose_comp_ctxt comp;
830 struct tran_int_desc *int_desc;
831 struct cpumask *affinity;
833 struct pci_packet pkt;
834 u8 buffer[sizeof(struct pci_create_interrupt) -
835 sizeof(struct pci_message)];
840 pdev = msi_desc_to_pci_dev(irq_data_get_msi_desc(data));
842 hbus = container_of(pbus->sysdata, struct hv_pcibus_device, sysdata);
843 hpdev = get_pcichild_wslot(hbus, devfn_to_wslot(pdev->devfn));
845 goto return_null_message;
847 /* Free any previous message that might have already been composed. */
848 if (data->chip_data) {
849 int_desc = data->chip_data;
850 data->chip_data = NULL;
851 hv_int_desc_free(hpdev, int_desc);
854 int_desc = kzalloc(sizeof(*int_desc), GFP_KERNEL);
858 memset(&ctxt, 0, sizeof(ctxt));
859 init_completion(&comp.comp_pkt.host_event);
860 ctxt.pkt.completion_func = hv_pci_compose_compl;
861 ctxt.pkt.compl_ctxt = ∁
862 int_pkt = (struct pci_create_interrupt *)&ctxt.pkt.message;
863 int_pkt->message_type.message_type = PCI_CREATE_INTERRUPT_MESSAGE;
864 int_pkt->wslot.slot = hpdev->desc.win_slot.slot;
865 int_pkt->int_desc.vector = cfg->vector;
866 int_pkt->int_desc.vector_count = 1;
867 int_pkt->int_desc.delivery_mode =
868 (apic->irq_delivery_mode == dest_LowestPrio) ? 1 : 0;
871 * This bit doesn't have to work on machines with more than 64
872 * processors because Hyper-V only supports 64 in a guest.
874 affinity = irq_data_get_affinity_mask(data);
875 for_each_cpu_and(cpu, affinity, cpu_online_mask) {
876 int_pkt->int_desc.cpu_mask |=
877 (1ULL << vmbus_cpu_number_to_vp_number(cpu));
880 ret = vmbus_sendpacket(hpdev->hbus->hdev->channel, int_pkt,
881 sizeof(*int_pkt), (unsigned long)&ctxt.pkt,
883 VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
885 wait_for_completion(&comp.comp_pkt.host_event);
887 if (comp.comp_pkt.completion_status < 0) {
888 dev_err(&hbus->hdev->device,
889 "Request for interrupt failed: 0x%x",
890 comp.comp_pkt.completion_status);
895 * Record the assignment so that this can be unwound later. Using
896 * irq_set_chip_data() here would be appropriate, but the lock it takes
899 *int_desc = comp.int_desc;
900 data->chip_data = int_desc;
902 /* Pass up the result. */
903 msg->address_hi = comp.int_desc.address >> 32;
904 msg->address_lo = comp.int_desc.address & 0xffffffff;
905 msg->data = comp.int_desc.data;
907 put_pcichild(hpdev, hv_pcidev_ref_by_slot);
913 put_pcichild(hpdev, hv_pcidev_ref_by_slot);
920 /* HW Interrupt Chip Descriptor */
921 static struct irq_chip hv_msi_irq_chip = {
922 .name = "Hyper-V PCIe MSI",
923 .irq_compose_msi_msg = hv_compose_msi_msg,
924 .irq_set_affinity = hv_set_affinity,
925 .irq_ack = irq_chip_ack_parent,
926 .irq_mask = hv_irq_mask,
927 .irq_unmask = hv_irq_unmask,
930 static irq_hw_number_t hv_msi_domain_ops_get_hwirq(struct msi_domain_info *info,
931 msi_alloc_info_t *arg)
933 return arg->msi_hwirq;
936 static struct msi_domain_ops hv_msi_ops = {
937 .get_hwirq = hv_msi_domain_ops_get_hwirq,
938 .msi_prepare = pci_msi_prepare,
939 .set_desc = pci_msi_set_desc,
940 .msi_free = hv_msi_free,
944 * hv_pcie_init_irq_domain() - Initialize IRQ domain
945 * @hbus: The root PCI bus
947 * This function creates an IRQ domain which will be used for
948 * interrupts from devices that have been passed through. These
949 * devices only support MSI and MSI-X, not line-based interrupts
950 * or simulations of line-based interrupts through PCIe's
951 * fabric-layer messages. Because interrupts are remapped, we
952 * can support multi-message MSI here.
954 * Return: '0' on success and error value on failure
956 static int hv_pcie_init_irq_domain(struct hv_pcibus_device *hbus)
958 hbus->msi_info.chip = &hv_msi_irq_chip;
959 hbus->msi_info.ops = &hv_msi_ops;
960 hbus->msi_info.flags = (MSI_FLAG_USE_DEF_DOM_OPS |
961 MSI_FLAG_USE_DEF_CHIP_OPS | MSI_FLAG_MULTI_PCI_MSI |
963 hbus->msi_info.handler = handle_edge_irq;
964 hbus->msi_info.handler_name = "edge";
965 hbus->msi_info.data = hbus;
966 hbus->irq_domain = pci_msi_create_irq_domain(hbus->sysdata.fwnode,
969 if (!hbus->irq_domain) {
970 dev_err(&hbus->hdev->device,
971 "Failed to build an MSI IRQ domain\n");
979 * get_bar_size() - Get the address space consumed by a BAR
980 * @bar_val: Value that a BAR returned after -1 was written
983 * This function returns the size of the BAR, rounded up to 1
984 * page. It has to be rounded up because the hypervisor's page
985 * table entry that maps the BAR into the VM can't specify an
986 * offset within a page. The invariant is that the hypervisor
987 * must place any BARs of smaller than page length at the
988 * beginning of a page.
990 * Return: Size in bytes of the consumed MMIO space.
992 static u64 get_bar_size(u64 bar_val)
994 return round_up((1 + ~(bar_val & PCI_BASE_ADDRESS_MEM_MASK)),
999 * survey_child_resources() - Total all MMIO requirements
1000 * @hbus: Root PCI bus, as understood by this driver
1002 static void survey_child_resources(struct hv_pcibus_device *hbus)
1004 struct list_head *iter;
1005 struct hv_pci_dev *hpdev;
1006 resource_size_t bar_size = 0;
1007 unsigned long flags;
1008 struct completion *event;
1012 /* If nobody is waiting on the answer, don't compute it. */
1013 event = xchg(&hbus->survey_event, NULL);
1017 /* If the answer has already been computed, go with it. */
1018 if (hbus->low_mmio_space || hbus->high_mmio_space) {
1023 spin_lock_irqsave(&hbus->device_list_lock, flags);
1026 * Due to an interesting quirk of the PCI spec, all memory regions
1027 * for a child device are a power of 2 in size and aligned in memory,
1028 * so it's sufficient to just add them up without tracking alignment.
1030 list_for_each(iter, &hbus->children) {
1031 hpdev = container_of(iter, struct hv_pci_dev, list_entry);
1032 for (i = 0; i < 6; i++) {
1033 if (hpdev->probed_bar[i] & PCI_BASE_ADDRESS_SPACE_IO)
1034 dev_err(&hbus->hdev->device,
1035 "There's an I/O BAR in this list!\n");
1037 if (hpdev->probed_bar[i] != 0) {
1039 * A probed BAR has all the upper bits set that
1043 bar_val = hpdev->probed_bar[i];
1044 if (bar_val & PCI_BASE_ADDRESS_MEM_TYPE_64)
1046 ((u64)hpdev->probed_bar[++i] << 32);
1048 bar_val |= 0xffffffff00000000ULL;
1050 bar_size = get_bar_size(bar_val);
1052 if (bar_val & PCI_BASE_ADDRESS_MEM_TYPE_64)
1053 hbus->high_mmio_space += bar_size;
1055 hbus->low_mmio_space += bar_size;
1060 spin_unlock_irqrestore(&hbus->device_list_lock, flags);
1065 * prepopulate_bars() - Fill in BARs with defaults
1066 * @hbus: Root PCI bus, as understood by this driver
1068 * The core PCI driver code seems much, much happier if the BARs
1069 * for a device have values upon first scan. So fill them in.
1070 * The algorithm below works down from large sizes to small,
1071 * attempting to pack the assignments optimally. The assumption,
1072 * enforced in other parts of the code, is that the beginning of
1073 * the memory-mapped I/O space will be aligned on the largest
1076 static void prepopulate_bars(struct hv_pcibus_device *hbus)
1078 resource_size_t high_size = 0;
1079 resource_size_t low_size = 0;
1080 resource_size_t high_base = 0;
1081 resource_size_t low_base = 0;
1082 resource_size_t bar_size;
1083 struct hv_pci_dev *hpdev;
1084 struct list_head *iter;
1085 unsigned long flags;
1091 if (hbus->low_mmio_space) {
1092 low_size = 1ULL << (63 - __builtin_clzll(hbus->low_mmio_space));
1093 low_base = hbus->low_mmio_res->start;
1096 if (hbus->high_mmio_space) {
1098 (63 - __builtin_clzll(hbus->high_mmio_space));
1099 high_base = hbus->high_mmio_res->start;
1102 spin_lock_irqsave(&hbus->device_list_lock, flags);
1104 /* Pick addresses for the BARs. */
1106 list_for_each(iter, &hbus->children) {
1107 hpdev = container_of(iter, struct hv_pci_dev,
1109 for (i = 0; i < 6; i++) {
1110 bar_val = hpdev->probed_bar[i];
1113 high = bar_val & PCI_BASE_ADDRESS_MEM_TYPE_64;
1116 ((u64)hpdev->probed_bar[i + 1]
1119 bar_val |= 0xffffffffULL << 32;
1121 bar_size = get_bar_size(bar_val);
1123 if (high_size != bar_size) {
1127 _hv_pcifront_write_config(hpdev,
1128 PCI_BASE_ADDRESS_0 + (4 * i),
1130 (u32)(high_base & 0xffffff00));
1132 _hv_pcifront_write_config(hpdev,
1133 PCI_BASE_ADDRESS_0 + (4 * i),
1134 4, (u32)(high_base >> 32));
1135 high_base += bar_size;
1137 if (low_size != bar_size)
1139 _hv_pcifront_write_config(hpdev,
1140 PCI_BASE_ADDRESS_0 + (4 * i),
1142 (u32)(low_base & 0xffffff00));
1143 low_base += bar_size;
1146 if (high_size <= 1 && low_size <= 1) {
1147 /* Set the memory enable bit. */
1148 _hv_pcifront_read_config(hpdev, PCI_COMMAND, 2,
1150 command |= PCI_COMMAND_MEMORY;
1151 _hv_pcifront_write_config(hpdev, PCI_COMMAND, 2,
1159 } while (high_size || low_size);
1161 spin_unlock_irqrestore(&hbus->device_list_lock, flags);
1165 * create_root_hv_pci_bus() - Expose a new root PCI bus
1166 * @hbus: Root PCI bus, as understood by this driver
1168 * Return: 0 on success, -errno on failure
1170 static int create_root_hv_pci_bus(struct hv_pcibus_device *hbus)
1172 /* Register the device */
1173 hbus->pci_bus = pci_create_root_bus(&hbus->hdev->device,
1174 0, /* bus number is always zero */
1177 &hbus->resources_for_children);
1181 hbus->pci_bus->msi = &hbus->msi_chip;
1182 hbus->pci_bus->msi->dev = &hbus->hdev->device;
1184 pci_scan_child_bus(hbus->pci_bus);
1185 pci_bus_assign_resources(hbus->pci_bus);
1186 pci_bus_add_devices(hbus->pci_bus);
1187 hbus->state = hv_pcibus_installed;
1191 struct q_res_req_compl {
1192 struct completion host_event;
1193 struct hv_pci_dev *hpdev;
1197 * q_resource_requirements() - Query Resource Requirements
1198 * @context: The completion context.
1199 * @resp: The response that came from the host.
1200 * @resp_packet_size: The size in bytes of resp.
1202 * This function is invoked on completion of a Query Resource
1203 * Requirements packet.
1205 static void q_resource_requirements(void *context, struct pci_response *resp,
1206 int resp_packet_size)
1208 struct q_res_req_compl *completion = context;
1209 struct pci_q_res_req_response *q_res_req =
1210 (struct pci_q_res_req_response *)resp;
1213 if (resp->status < 0) {
1214 dev_err(&completion->hpdev->hbus->hdev->device,
1215 "query resource requirements failed: %x\n",
1218 for (i = 0; i < 6; i++) {
1219 completion->hpdev->probed_bar[i] =
1220 q_res_req->probed_bar[i];
1224 complete(&completion->host_event);
1227 static void get_pcichild(struct hv_pci_dev *hpdev,
1228 enum hv_pcidev_ref_reason reason)
1230 atomic_inc(&hpdev->refs);
1233 static void put_pcichild(struct hv_pci_dev *hpdev,
1234 enum hv_pcidev_ref_reason reason)
1236 if (atomic_dec_and_test(&hpdev->refs))
1241 * new_pcichild_device() - Create a new child device
1242 * @hbus: The internal struct tracking this root PCI bus.
1243 * @desc: The information supplied so far from the host
1246 * This function creates the tracking structure for a new child
1247 * device and kicks off the process of figuring out what it is.
1249 * Return: Pointer to the new tracking struct
1251 static struct hv_pci_dev *new_pcichild_device(struct hv_pcibus_device *hbus,
1252 struct pci_function_description *desc)
1254 struct hv_pci_dev *hpdev;
1255 struct pci_child_message *res_req;
1256 struct q_res_req_compl comp_pkt;
1258 struct pci_packet init_packet;
1261 unsigned long flags;
1264 hpdev = kzalloc(sizeof(*hpdev), GFP_ATOMIC);
1270 memset(&pkt, 0, sizeof(pkt));
1271 init_completion(&comp_pkt.host_event);
1272 comp_pkt.hpdev = hpdev;
1273 pkt.init_packet.compl_ctxt = &comp_pkt;
1274 pkt.init_packet.completion_func = q_resource_requirements;
1275 res_req = (struct pci_child_message *)&pkt.init_packet.message;
1276 res_req->message_type = PCI_QUERY_RESOURCE_REQUIREMENTS;
1277 res_req->wslot.slot = desc->win_slot.slot;
1279 ret = vmbus_sendpacket(hbus->hdev->channel, res_req,
1280 sizeof(struct pci_child_message),
1281 (unsigned long)&pkt.init_packet,
1283 VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
1287 wait_for_completion(&comp_pkt.host_event);
1289 hpdev->desc = *desc;
1290 get_pcichild(hpdev, hv_pcidev_ref_initial);
1291 get_pcichild(hpdev, hv_pcidev_ref_childlist);
1292 spin_lock_irqsave(&hbus->device_list_lock, flags);
1293 list_add_tail(&hpdev->list_entry, &hbus->children);
1294 spin_unlock_irqrestore(&hbus->device_list_lock, flags);
1303 * get_pcichild_wslot() - Find device from slot
1304 * @hbus: Root PCI bus, as understood by this driver
1305 * @wslot: Location on the bus
1307 * This function looks up a PCI device and returns the internal
1308 * representation of it. It acquires a reference on it, so that
1309 * the device won't be deleted while somebody is using it. The
1310 * caller is responsible for calling put_pcichild() to release
1313 * Return: Internal representation of a PCI device
1315 static struct hv_pci_dev *get_pcichild_wslot(struct hv_pcibus_device *hbus,
1318 unsigned long flags;
1319 struct hv_pci_dev *iter, *hpdev = NULL;
1321 spin_lock_irqsave(&hbus->device_list_lock, flags);
1322 list_for_each_entry(iter, &hbus->children, list_entry) {
1323 if (iter->desc.win_slot.slot == wslot) {
1325 get_pcichild(hpdev, hv_pcidev_ref_by_slot);
1329 spin_unlock_irqrestore(&hbus->device_list_lock, flags);
1335 * pci_devices_present_work() - Handle new list of child devices
1336 * @work: Work struct embedded in struct hv_dr_work
1338 * "Bus Relations" is the Windows term for "children of this
1339 * bus." The terminology is preserved here for people trying to
1340 * debug the interaction between Hyper-V and Linux. This
1341 * function is called when the parent partition reports a list
1342 * of functions that should be observed under this PCI Express
1345 * This function updates the list, and must tolerate being
1346 * called multiple times with the same information. The typical
1347 * number of child devices is one, with very atypical cases
1348 * involving three or four, so the algorithms used here can be
1349 * simple and inefficient.
1351 * It must also treat the omission of a previously observed device as
1352 * notification that the device no longer exists.
1354 * Note that this function is a work item, and it may not be
1355 * invoked in the order that it was queued. Back to back
1356 * updates of the list of present devices may involve queuing
1357 * multiple work items, and this one may run before ones that
1358 * were sent later. As such, this function only does something
1359 * if is the last one in the queue.
1361 static void pci_devices_present_work(struct work_struct *work)
1365 struct list_head *iter;
1366 struct pci_function_description *new_desc;
1367 struct hv_pci_dev *hpdev;
1368 struct hv_pcibus_device *hbus;
1369 struct list_head removed;
1370 struct hv_dr_work *dr_wrk;
1371 struct hv_dr_state *dr = NULL;
1372 unsigned long flags;
1374 dr_wrk = container_of(work, struct hv_dr_work, wrk);
1378 INIT_LIST_HEAD(&removed);
1380 if (down_interruptible(&hbus->enum_sem)) {
1385 /* Pull this off the queue and process it if it was the last one. */
1386 spin_lock_irqsave(&hbus->device_list_lock, flags);
1387 while (!list_empty(&hbus->dr_list)) {
1388 dr = list_first_entry(&hbus->dr_list, struct hv_dr_state,
1390 list_del(&dr->list_entry);
1392 /* Throw this away if the list still has stuff in it. */
1393 if (!list_empty(&hbus->dr_list)) {
1398 spin_unlock_irqrestore(&hbus->device_list_lock, flags);
1401 up(&hbus->enum_sem);
1406 /* First, mark all existing children as reported missing. */
1407 spin_lock_irqsave(&hbus->device_list_lock, flags);
1408 list_for_each(iter, &hbus->children) {
1409 hpdev = container_of(iter, struct hv_pci_dev,
1411 hpdev->reported_missing = true;
1413 spin_unlock_irqrestore(&hbus->device_list_lock, flags);
1415 /* Next, add back any reported devices. */
1416 for (child_no = 0; child_no < dr->device_count; child_no++) {
1418 new_desc = &dr->func[child_no];
1420 spin_lock_irqsave(&hbus->device_list_lock, flags);
1421 list_for_each(iter, &hbus->children) {
1422 hpdev = container_of(iter, struct hv_pci_dev,
1424 if ((hpdev->desc.win_slot.slot ==
1425 new_desc->win_slot.slot) &&
1426 (hpdev->desc.v_id == new_desc->v_id) &&
1427 (hpdev->desc.d_id == new_desc->d_id) &&
1428 (hpdev->desc.ser == new_desc->ser)) {
1429 hpdev->reported_missing = false;
1433 spin_unlock_irqrestore(&hbus->device_list_lock, flags);
1436 hpdev = new_pcichild_device(hbus, new_desc);
1438 dev_err(&hbus->hdev->device,
1439 "couldn't record a child device.\n");
1443 /* Move missing children to a list on the stack. */
1444 spin_lock_irqsave(&hbus->device_list_lock, flags);
1447 list_for_each(iter, &hbus->children) {
1448 hpdev = container_of(iter, struct hv_pci_dev,
1450 if (hpdev->reported_missing) {
1452 put_pcichild(hpdev, hv_pcidev_ref_childlist);
1453 list_del(&hpdev->list_entry);
1454 list_add_tail(&hpdev->list_entry, &removed);
1459 spin_unlock_irqrestore(&hbus->device_list_lock, flags);
1461 /* Delete everything that should no longer exist. */
1462 while (!list_empty(&removed)) {
1463 hpdev = list_first_entry(&removed, struct hv_pci_dev,
1465 list_del(&hpdev->list_entry);
1466 put_pcichild(hpdev, hv_pcidev_ref_initial);
1469 /* Tell the core to rescan bus because there may have been changes. */
1470 if (hbus->state == hv_pcibus_installed) {
1471 pci_lock_rescan_remove();
1472 pci_scan_child_bus(hbus->pci_bus);
1473 pci_unlock_rescan_remove();
1475 survey_child_resources(hbus);
1478 up(&hbus->enum_sem);
1484 * hv_pci_devices_present() - Handles list of new children
1485 * @hbus: Root PCI bus, as understood by this driver
1486 * @relations: Packet from host listing children
1488 * This function is invoked whenever a new list of devices for
1491 static void hv_pci_devices_present(struct hv_pcibus_device *hbus,
1492 struct pci_bus_relations *relations)
1494 struct hv_dr_state *dr;
1495 struct hv_dr_work *dr_wrk;
1496 unsigned long flags;
1498 dr_wrk = kzalloc(sizeof(*dr_wrk), GFP_NOWAIT);
1502 dr = kzalloc(offsetof(struct hv_dr_state, func) +
1503 (sizeof(struct pci_function_description) *
1504 (relations->device_count)), GFP_NOWAIT);
1510 INIT_WORK(&dr_wrk->wrk, pci_devices_present_work);
1512 dr->device_count = relations->device_count;
1513 if (dr->device_count != 0) {
1514 memcpy(dr->func, relations->func,
1515 sizeof(struct pci_function_description) *
1519 spin_lock_irqsave(&hbus->device_list_lock, flags);
1520 list_add_tail(&dr->list_entry, &hbus->dr_list);
1521 spin_unlock_irqrestore(&hbus->device_list_lock, flags);
1524 schedule_work(&dr_wrk->wrk);
1528 * hv_eject_device_work() - Asynchronously handles ejection
1529 * @work: Work struct embedded in internal device struct
1531 * This function handles ejecting a device. Windows will
1532 * attempt to gracefully eject a device, waiting 60 seconds to
1533 * hear back from the guest OS that this completed successfully.
1534 * If this timer expires, the device will be forcibly removed.
1536 static void hv_eject_device_work(struct work_struct *work)
1538 struct pci_eject_response *ejct_pkt;
1539 struct hv_pci_dev *hpdev;
1540 struct pci_dev *pdev;
1541 unsigned long flags;
1544 struct pci_packet pkt;
1545 u8 buffer[sizeof(struct pci_eject_response) -
1546 sizeof(struct pci_message)];
1549 hpdev = container_of(work, struct hv_pci_dev, wrk);
1551 if (hpdev->state != hv_pcichild_ejecting) {
1552 put_pcichild(hpdev, hv_pcidev_ref_pnp);
1557 * Ejection can come before or after the PCI bus has been set up, so
1558 * attempt to find it and tear down the bus state, if it exists. This
1559 * must be done without constructs like pci_domain_nr(hbus->pci_bus)
1560 * because hbus->pci_bus may not exist yet.
1562 wslot = wslot_to_devfn(hpdev->desc.win_slot.slot);
1563 pdev = pci_get_domain_bus_and_slot(hpdev->hbus->sysdata.domain, 0,
1566 pci_stop_and_remove_bus_device(pdev);
1570 memset(&ctxt, 0, sizeof(ctxt));
1571 ejct_pkt = (struct pci_eject_response *)&ctxt.pkt.message;
1572 ejct_pkt->message_type = PCI_EJECTION_COMPLETE;
1573 ejct_pkt->wslot.slot = hpdev->desc.win_slot.slot;
1574 vmbus_sendpacket(hpdev->hbus->hdev->channel, ejct_pkt,
1575 sizeof(*ejct_pkt), (unsigned long)&ctxt.pkt,
1576 VM_PKT_DATA_INBAND, 0);
1578 spin_lock_irqsave(&hpdev->hbus->device_list_lock, flags);
1579 list_del(&hpdev->list_entry);
1580 spin_unlock_irqrestore(&hpdev->hbus->device_list_lock, flags);
1582 put_pcichild(hpdev, hv_pcidev_ref_childlist);
1583 put_pcichild(hpdev, hv_pcidev_ref_pnp);
1584 put_hvpcibus(hpdev->hbus);
1588 * hv_pci_eject_device() - Handles device ejection
1589 * @hpdev: Internal device tracking struct
1591 * This function is invoked when an ejection packet arrives. It
1592 * just schedules work so that we don't re-enter the packet
1593 * delivery code handling the ejection.
1595 static void hv_pci_eject_device(struct hv_pci_dev *hpdev)
1597 hpdev->state = hv_pcichild_ejecting;
1598 get_pcichild(hpdev, hv_pcidev_ref_pnp);
1599 INIT_WORK(&hpdev->wrk, hv_eject_device_work);
1600 get_hvpcibus(hpdev->hbus);
1601 schedule_work(&hpdev->wrk);
1605 * hv_pci_onchannelcallback() - Handles incoming packets
1606 * @context: Internal bus tracking struct
1608 * This function is invoked whenever the host sends a packet to
1609 * this channel (which is private to this root PCI bus).
1611 static void hv_pci_onchannelcallback(void *context)
1613 const int packet_size = 0x100;
1615 struct hv_pcibus_device *hbus = context;
1618 struct vmpacket_descriptor *desc;
1619 unsigned char *buffer;
1620 int bufferlen = packet_size;
1621 struct pci_packet *comp_packet;
1622 struct pci_response *response;
1623 struct pci_incoming_message *new_message;
1624 struct pci_bus_relations *bus_rel;
1625 struct pci_dev_incoming *dev_message;
1626 struct hv_pci_dev *hpdev;
1628 buffer = kmalloc(bufferlen, GFP_ATOMIC);
1633 ret = vmbus_recvpacket_raw(hbus->hdev->channel, buffer,
1634 bufferlen, &bytes_recvd, &req_id);
1636 if (ret == -ENOBUFS) {
1638 /* Handle large packet */
1639 bufferlen = bytes_recvd;
1640 buffer = kmalloc(bytes_recvd, GFP_ATOMIC);
1647 * All incoming packets must be at least as large as a
1650 if (bytes_recvd <= sizeof(struct pci_response)) {
1654 desc = (struct vmpacket_descriptor *)buffer;
1656 switch (desc->type) {
1660 * The host is trusted, and thus it's safe to interpret
1661 * this transaction ID as a pointer.
1663 comp_packet = (struct pci_packet *)req_id;
1664 response = (struct pci_response *)buffer;
1665 comp_packet->completion_func(comp_packet->compl_ctxt,
1671 case VM_PKT_DATA_INBAND:
1673 new_message = (struct pci_incoming_message *)buffer;
1674 switch (new_message->message_type.message_type) {
1675 case PCI_BUS_RELATIONS:
1677 bus_rel = (struct pci_bus_relations *)buffer;
1679 offsetof(struct pci_bus_relations, func) +
1680 (sizeof(struct pci_function_description) *
1681 (bus_rel->device_count))) {
1682 dev_err(&hbus->hdev->device,
1683 "bus relations too small\n");
1687 hv_pci_devices_present(hbus, bus_rel);
1692 dev_message = (struct pci_dev_incoming *)buffer;
1693 hpdev = get_pcichild_wslot(hbus,
1694 dev_message->wslot.slot);
1696 hv_pci_eject_device(hpdev);
1698 hv_pcidev_ref_by_slot);
1703 dev_warn(&hbus->hdev->device,
1704 "Unimplemented protocol message %x\n",
1705 new_message->message_type.message_type);
1711 dev_err(&hbus->hdev->device,
1712 "unhandled packet type %d, tid %llx len %d\n",
1713 desc->type, req_id, bytes_recvd);
1721 * hv_pci_protocol_negotiation() - Set up protocol
1722 * @hdev: VMBus's tracking struct for this root PCI bus
1724 * This driver is intended to support running on Windows 10
1725 * (server) and later versions. It will not run on earlier
1726 * versions, as they assume that many of the operations which
1727 * Linux needs accomplished with a spinlock held were done via
1728 * asynchronous messaging via VMBus. Windows 10 increases the
1729 * surface area of PCI emulation so that these actions can take
1730 * place by suspending a virtual processor for their duration.
1732 * This function negotiates the channel protocol version,
1733 * failing if the host doesn't support the necessary protocol
1736 static int hv_pci_protocol_negotiation(struct hv_device *hdev)
1738 struct pci_version_request *version_req;
1739 struct hv_pci_compl comp_pkt;
1740 struct pci_packet *pkt;
1744 * Initiate the handshake with the host and negotiate
1745 * a version that the host can support. We start with the
1746 * highest version number and go down if the host cannot
1749 pkt = kzalloc(sizeof(*pkt) + sizeof(*version_req), GFP_KERNEL);
1753 init_completion(&comp_pkt.host_event);
1754 pkt->completion_func = hv_pci_generic_compl;
1755 pkt->compl_ctxt = &comp_pkt;
1756 version_req = (struct pci_version_request *)&pkt->message;
1757 version_req->message_type.message_type = PCI_QUERY_PROTOCOL_VERSION;
1758 version_req->protocol_version = PCI_PROTOCOL_VERSION_CURRENT;
1760 ret = vmbus_sendpacket(hdev->channel, version_req,
1761 sizeof(struct pci_version_request),
1762 (unsigned long)pkt, VM_PKT_DATA_INBAND,
1763 VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
1767 wait_for_completion(&comp_pkt.host_event);
1769 if (comp_pkt.completion_status < 0) {
1770 dev_err(&hdev->device,
1771 "PCI Pass-through VSP failed version request %x\n",
1772 comp_pkt.completion_status);
1785 * hv_pci_free_bridge_windows() - Release memory regions for the
1787 * @hbus: Root PCI bus, as understood by this driver
1789 static void hv_pci_free_bridge_windows(struct hv_pcibus_device *hbus)
1792 * Set the resources back to the way they looked when they
1793 * were allocated by setting IORESOURCE_BUSY again.
1796 if (hbus->low_mmio_space && hbus->low_mmio_res) {
1797 hbus->low_mmio_res->flags |= IORESOURCE_BUSY;
1798 release_mem_region(hbus->low_mmio_res->start,
1799 resource_size(hbus->low_mmio_res));
1802 if (hbus->high_mmio_space && hbus->high_mmio_res) {
1803 hbus->high_mmio_res->flags |= IORESOURCE_BUSY;
1804 release_mem_region(hbus->high_mmio_res->start,
1805 resource_size(hbus->high_mmio_res));
1810 * hv_pci_allocate_bridge_windows() - Allocate memory regions
1812 * @hbus: Root PCI bus, as understood by this driver
1814 * This function calls vmbus_allocate_mmio(), which is itself a
1815 * bit of a compromise. Ideally, we might change the pnp layer
1816 * in the kernel such that it comprehends either PCI devices
1817 * which are "grandchildren of ACPI," with some intermediate bus
1818 * node (in this case, VMBus) or change it such that it
1819 * understands VMBus. The pnp layer, however, has been declared
1820 * deprecated, and not subject to change.
1822 * The workaround, implemented here, is to ask VMBus to allocate
1823 * MMIO space for this bus. VMBus itself knows which ranges are
1824 * appropriate by looking at its own ACPI objects. Then, after
1825 * these ranges are claimed, they're modified to look like they
1826 * would have looked if the ACPI and pnp code had allocated
1827 * bridge windows. These descriptors have to exist in this form
1828 * in order to satisfy the code which will get invoked when the
1829 * endpoint PCI function driver calls request_mem_region() or
1830 * request_mem_region_exclusive().
1832 * Return: 0 on success, -errno on failure
1834 static int hv_pci_allocate_bridge_windows(struct hv_pcibus_device *hbus)
1836 resource_size_t align;
1839 if (hbus->low_mmio_space) {
1840 align = 1ULL << (63 - __builtin_clzll(hbus->low_mmio_space));
1841 ret = vmbus_allocate_mmio(&hbus->low_mmio_res, hbus->hdev, 0,
1842 (u64)(u32)0xffffffff,
1843 hbus->low_mmio_space,
1846 dev_err(&hbus->hdev->device,
1847 "Need %#llx of low MMIO space. Consider reconfiguring the VM.\n",
1848 hbus->low_mmio_space);
1852 /* Modify this resource to become a bridge window. */
1853 hbus->low_mmio_res->flags |= IORESOURCE_WINDOW;
1854 hbus->low_mmio_res->flags &= ~IORESOURCE_BUSY;
1855 pci_add_resource(&hbus->resources_for_children,
1856 hbus->low_mmio_res);
1859 if (hbus->high_mmio_space) {
1860 align = 1ULL << (63 - __builtin_clzll(hbus->high_mmio_space));
1861 ret = vmbus_allocate_mmio(&hbus->high_mmio_res, hbus->hdev,
1863 hbus->high_mmio_space, align,
1866 dev_err(&hbus->hdev->device,
1867 "Need %#llx of high MMIO space. Consider reconfiguring the VM.\n",
1868 hbus->high_mmio_space);
1869 goto release_low_mmio;
1872 /* Modify this resource to become a bridge window. */
1873 hbus->high_mmio_res->flags |= IORESOURCE_WINDOW;
1874 hbus->high_mmio_res->flags &= ~IORESOURCE_BUSY;
1875 pci_add_resource(&hbus->resources_for_children,
1876 hbus->high_mmio_res);
1882 if (hbus->low_mmio_res) {
1883 release_mem_region(hbus->low_mmio_res->start,
1884 resource_size(hbus->low_mmio_res));
1891 * hv_allocate_config_window() - Find MMIO space for PCI Config
1892 * @hbus: Root PCI bus, as understood by this driver
1894 * This function claims memory-mapped I/O space for accessing
1895 * configuration space for the functions on this bus.
1897 * Return: 0 on success, -errno on failure
1899 static int hv_allocate_config_window(struct hv_pcibus_device *hbus)
1904 * Set up a region of MMIO space to use for accessing configuration
1907 ret = vmbus_allocate_mmio(&hbus->mem_config, hbus->hdev, 0, -1,
1908 PCI_CONFIG_MMIO_LENGTH, 0x1000, false);
1913 * vmbus_allocate_mmio() gets used for allocating both device endpoint
1914 * resource claims (those which cannot be overlapped) and the ranges
1915 * which are valid for the children of this bus, which are intended
1916 * to be overlapped by those children. Set the flag on this claim
1917 * meaning that this region can't be overlapped.
1920 hbus->mem_config->flags |= IORESOURCE_BUSY;
1925 static void hv_free_config_window(struct hv_pcibus_device *hbus)
1927 release_mem_region(hbus->mem_config->start, PCI_CONFIG_MMIO_LENGTH);
1931 * hv_pci_enter_d0() - Bring the "bus" into the D0 power state
1932 * @hdev: VMBus's tracking struct for this root PCI bus
1934 * Return: 0 on success, -errno on failure
1936 static int hv_pci_enter_d0(struct hv_device *hdev)
1938 struct hv_pcibus_device *hbus = hv_get_drvdata(hdev);
1939 struct pci_bus_d0_entry *d0_entry;
1940 struct hv_pci_compl comp_pkt;
1941 struct pci_packet *pkt;
1945 * Tell the host that the bus is ready to use, and moved into the
1946 * powered-on state. This includes telling the host which region
1947 * of memory-mapped I/O space has been chosen for configuration space
1950 pkt = kzalloc(sizeof(*pkt) + sizeof(*d0_entry), GFP_KERNEL);
1954 init_completion(&comp_pkt.host_event);
1955 pkt->completion_func = hv_pci_generic_compl;
1956 pkt->compl_ctxt = &comp_pkt;
1957 d0_entry = (struct pci_bus_d0_entry *)&pkt->message;
1958 d0_entry->message_type.message_type = PCI_BUS_D0ENTRY;
1959 d0_entry->mmio_base = hbus->mem_config->start;
1961 ret = vmbus_sendpacket(hdev->channel, d0_entry, sizeof(*d0_entry),
1962 (unsigned long)pkt, VM_PKT_DATA_INBAND,
1963 VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
1967 wait_for_completion(&comp_pkt.host_event);
1969 if (comp_pkt.completion_status < 0) {
1970 dev_err(&hdev->device,
1971 "PCI Pass-through VSP failed D0 Entry with status %x\n",
1972 comp_pkt.completion_status);
1985 * hv_pci_query_relations() - Ask host to send list of child
1987 * @hdev: VMBus's tracking struct for this root PCI bus
1989 * Return: 0 on success, -errno on failure
1991 static int hv_pci_query_relations(struct hv_device *hdev)
1993 struct hv_pcibus_device *hbus = hv_get_drvdata(hdev);
1994 struct pci_message message;
1995 struct completion comp;
1998 /* Ask the host to send along the list of child devices */
1999 init_completion(&comp);
2000 if (cmpxchg(&hbus->survey_event, NULL, &comp))
2003 memset(&message, 0, sizeof(message));
2004 message.message_type = PCI_QUERY_BUS_RELATIONS;
2006 ret = vmbus_sendpacket(hdev->channel, &message, sizeof(message),
2007 0, VM_PKT_DATA_INBAND, 0);
2011 wait_for_completion(&comp);
2016 * hv_send_resources_allocated() - Report local resource choices
2017 * @hdev: VMBus's tracking struct for this root PCI bus
2019 * The host OS is expecting to be sent a request as a message
2020 * which contains all the resources that the device will use.
2021 * The response contains those same resources, "translated"
2022 * which is to say, the values which should be used by the
2023 * hardware, when it delivers an interrupt. (MMIO resources are
2024 * used in local terms.) This is nice for Windows, and lines up
2025 * with the FDO/PDO split, which doesn't exist in Linux. Linux
2026 * is deeply expecting to scan an emulated PCI configuration
2027 * space. So this message is sent here only to drive the state
2028 * machine on the host forward.
2030 * Return: 0 on success, -errno on failure
2032 static int hv_send_resources_allocated(struct hv_device *hdev)
2034 struct hv_pcibus_device *hbus = hv_get_drvdata(hdev);
2035 struct pci_resources_assigned *res_assigned;
2036 struct hv_pci_compl comp_pkt;
2037 struct hv_pci_dev *hpdev;
2038 struct pci_packet *pkt;
2042 pkt = kmalloc(sizeof(*pkt) + sizeof(*res_assigned), GFP_KERNEL);
2048 for (wslot = 0; wslot < 256; wslot++) {
2049 hpdev = get_pcichild_wslot(hbus, wslot);
2053 memset(pkt, 0, sizeof(*pkt) + sizeof(*res_assigned));
2054 init_completion(&comp_pkt.host_event);
2055 pkt->completion_func = hv_pci_generic_compl;
2056 pkt->compl_ctxt = &comp_pkt;
2057 pkt->message.message_type = PCI_RESOURCES_ASSIGNED;
2058 res_assigned = (struct pci_resources_assigned *)&pkt->message;
2059 res_assigned->wslot.slot = hpdev->desc.win_slot.slot;
2061 put_pcichild(hpdev, hv_pcidev_ref_by_slot);
2063 ret = vmbus_sendpacket(
2064 hdev->channel, &pkt->message,
2065 sizeof(*res_assigned),
2068 VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
2072 wait_for_completion(&comp_pkt.host_event);
2074 if (comp_pkt.completion_status < 0) {
2076 dev_err(&hdev->device,
2077 "resource allocated returned 0x%x",
2078 comp_pkt.completion_status);
2088 * hv_send_resources_released() - Report local resources
2090 * @hdev: VMBus's tracking struct for this root PCI bus
2092 * Return: 0 on success, -errno on failure
2094 static int hv_send_resources_released(struct hv_device *hdev)
2096 struct hv_pcibus_device *hbus = hv_get_drvdata(hdev);
2097 struct pci_child_message pkt;
2098 struct hv_pci_dev *hpdev;
2102 for (wslot = 0; wslot < 256; wslot++) {
2103 hpdev = get_pcichild_wslot(hbus, wslot);
2107 memset(&pkt, 0, sizeof(pkt));
2108 pkt.message_type = PCI_RESOURCES_RELEASED;
2109 pkt.wslot.slot = hpdev->desc.win_slot.slot;
2111 put_pcichild(hpdev, hv_pcidev_ref_by_slot);
2113 ret = vmbus_sendpacket(hdev->channel, &pkt, sizeof(pkt), 0,
2114 VM_PKT_DATA_INBAND, 0);
2122 static void get_hvpcibus(struct hv_pcibus_device *hbus)
2124 atomic_inc(&hbus->remove_lock);
2127 static void put_hvpcibus(struct hv_pcibus_device *hbus)
2129 if (atomic_dec_and_test(&hbus->remove_lock))
2130 complete(&hbus->remove_event);
2134 * hv_pci_probe() - New VMBus channel probe, for a root PCI bus
2135 * @hdev: VMBus's tracking struct for this root PCI bus
2136 * @dev_id: Identifies the device itself
2138 * Return: 0 on success, -errno on failure
2140 static int hv_pci_probe(struct hv_device *hdev,
2141 const struct hv_vmbus_device_id *dev_id)
2143 struct hv_pcibus_device *hbus;
2146 hbus = kzalloc(sizeof(*hbus), GFP_KERNEL);
2151 * The PCI bus "domain" is what is called "segment" in ACPI and
2152 * other specs. Pull it from the instance ID, to get something
2153 * unique. Bytes 8 and 9 are what is used in Windows guests, so
2154 * do the same thing for consistency. Note that, since this code
2155 * only runs in a Hyper-V VM, Hyper-V can (and does) guarantee
2156 * that (1) the only domain in use for something that looks like
2157 * a physical PCI bus (which is actually emulated by the
2158 * hypervisor) is domain 0 and (2) there will be no overlap
2159 * between domains derived from these instance IDs in the same
2162 hbus->sysdata.domain = hdev->dev_instance.b[9] |
2163 hdev->dev_instance.b[8] << 8;
2166 atomic_inc(&hbus->remove_lock);
2167 INIT_LIST_HEAD(&hbus->children);
2168 INIT_LIST_HEAD(&hbus->dr_list);
2169 INIT_LIST_HEAD(&hbus->resources_for_children);
2170 spin_lock_init(&hbus->config_lock);
2171 spin_lock_init(&hbus->device_list_lock);
2172 sema_init(&hbus->enum_sem, 1);
2173 init_completion(&hbus->remove_event);
2175 ret = vmbus_open(hdev->channel, pci_ring_size, pci_ring_size, NULL, 0,
2176 hv_pci_onchannelcallback, hbus);
2180 hv_set_drvdata(hdev, hbus);
2182 ret = hv_pci_protocol_negotiation(hdev);
2186 ret = hv_allocate_config_window(hbus);
2190 hbus->cfg_addr = ioremap(hbus->mem_config->start,
2191 PCI_CONFIG_MMIO_LENGTH);
2192 if (!hbus->cfg_addr) {
2193 dev_err(&hdev->device,
2194 "Unable to map a virtual address for config space\n");
2199 hbus->sysdata.fwnode = irq_domain_alloc_fwnode(hbus);
2200 if (!hbus->sysdata.fwnode) {
2205 ret = hv_pcie_init_irq_domain(hbus);
2209 ret = hv_pci_query_relations(hdev);
2211 goto free_irq_domain;
2213 ret = hv_pci_enter_d0(hdev);
2215 goto free_irq_domain;
2217 ret = hv_pci_allocate_bridge_windows(hbus);
2219 goto free_irq_domain;
2221 ret = hv_send_resources_allocated(hdev);
2225 prepopulate_bars(hbus);
2227 hbus->state = hv_pcibus_probed;
2229 ret = create_root_hv_pci_bus(hbus);
2236 hv_pci_free_bridge_windows(hbus);
2238 irq_domain_remove(hbus->irq_domain);
2240 irq_domain_free_fwnode(hbus->sysdata.fwnode);
2242 iounmap(hbus->cfg_addr);
2244 hv_free_config_window(hbus);
2246 vmbus_close(hdev->channel);
2253 * hv_pci_remove() - Remove routine for this VMBus channel
2254 * @hdev: VMBus's tracking struct for this root PCI bus
2256 * Return: 0 on success, -errno on failure
2258 static int hv_pci_remove(struct hv_device *hdev)
2261 struct hv_pcibus_device *hbus;
2263 struct pci_packet teardown_packet;
2266 struct pci_bus_relations relations;
2267 struct hv_pci_compl comp_pkt;
2269 hbus = hv_get_drvdata(hdev);
2271 ret = hv_send_resources_released(hdev);
2273 dev_err(&hdev->device,
2274 "Couldn't send resources released packet(s)\n");
2276 memset(&pkt.teardown_packet, 0, sizeof(pkt.teardown_packet));
2277 init_completion(&comp_pkt.host_event);
2278 pkt.teardown_packet.completion_func = hv_pci_generic_compl;
2279 pkt.teardown_packet.compl_ctxt = &comp_pkt;
2280 pkt.teardown_packet.message.message_type = PCI_BUS_D0EXIT;
2282 ret = vmbus_sendpacket(hdev->channel, &pkt.teardown_packet.message,
2283 sizeof(struct pci_message),
2284 (unsigned long)&pkt.teardown_packet,
2286 VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
2288 wait_for_completion_timeout(&comp_pkt.host_event, 10 * HZ);
2290 if (hbus->state == hv_pcibus_installed) {
2291 /* Remove the bus from PCI's point of view. */
2292 pci_lock_rescan_remove();
2293 pci_stop_root_bus(hbus->pci_bus);
2294 pci_remove_root_bus(hbus->pci_bus);
2295 pci_unlock_rescan_remove();
2298 vmbus_close(hdev->channel);
2300 /* Delete any children which might still exist. */
2301 memset(&relations, 0, sizeof(relations));
2302 hv_pci_devices_present(hbus, &relations);
2304 iounmap(hbus->cfg_addr);
2305 hv_free_config_window(hbus);
2306 pci_free_resource_list(&hbus->resources_for_children);
2307 hv_pci_free_bridge_windows(hbus);
2308 irq_domain_remove(hbus->irq_domain);
2309 irq_domain_free_fwnode(hbus->sysdata.fwnode);
2311 wait_for_completion(&hbus->remove_event);
2316 static const struct hv_vmbus_device_id hv_pci_id_table[] = {
2317 /* PCI Pass-through Class ID */
2318 /* 44C4F61D-4444-4400-9D52-802E27EDE19F */
2323 MODULE_DEVICE_TABLE(vmbus, hv_pci_id_table);
2325 static struct hv_driver hv_pci_drv = {
2327 .id_table = hv_pci_id_table,
2328 .probe = hv_pci_probe,
2329 .remove = hv_pci_remove,
2332 static void __exit exit_hv_pci_drv(void)
2334 vmbus_driver_unregister(&hv_pci_drv);
2337 static int __init init_hv_pci_drv(void)
2339 return vmbus_driver_register(&hv_pci_drv);
2342 module_init(init_hv_pci_drv);
2343 module_exit(exit_hv_pci_drv);
2345 MODULE_DESCRIPTION("Hyper-V PCI");
2346 MODULE_LICENSE("GPL v2");