2 * Copyright (C) 2012 Red Hat, Inc. All rights reserved.
3 * Author: Alex Williamson <alex.williamson@redhat.com>
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License version 2 as
7 * published by the Free Software Foundation.
9 * Derived from original vfio:
10 * Copyright 2010 Cisco Systems, Inc. All rights reserved.
11 * Author: Tom Lyon, pugs@cisco.com
14 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
16 #include <linux/device.h>
17 #include <linux/eventfd.h>
18 #include <linux/file.h>
19 #include <linux/interrupt.h>
20 #include <linux/iommu.h>
21 #include <linux/module.h>
22 #include <linux/mutex.h>
23 #include <linux/notifier.h>
24 #include <linux/pci.h>
25 #include <linux/pm_runtime.h>
26 #include <linux/slab.h>
27 #include <linux/types.h>
28 #include <linux/uaccess.h>
29 #include <linux/vfio.h>
30 #include <linux/vgaarb.h>
32 #include "vfio_pci_private.h"
34 #define DRIVER_VERSION "0.2"
35 #define DRIVER_AUTHOR "Alex Williamson <alex.williamson@redhat.com>"
36 #define DRIVER_DESC "VFIO PCI - User Level meta-driver"
38 static char ids[1024] __initdata;
39 module_param_string(ids, ids, sizeof(ids), 0);
40 MODULE_PARM_DESC(ids, "Initial PCI IDs to add to the vfio driver, format is \"vendor:device[:subvendor[:subdevice[:class[:class_mask]]]]\" and multiple comma separated entries can be specified");
42 static bool nointxmask;
43 module_param_named(nointxmask, nointxmask, bool, S_IRUGO | S_IWUSR);
44 MODULE_PARM_DESC(nointxmask,
45 "Disable support for PCI 2.3 style INTx masking. If this resolves problems for specific devices, report lspci -vvvxxx to linux-pci@vger.kernel.org so the device can be fixed automatically via the broken_intx_masking flag.");
47 #ifdef CONFIG_VFIO_PCI_VGA
48 static bool disable_vga;
49 module_param(disable_vga, bool, S_IRUGO);
50 MODULE_PARM_DESC(disable_vga, "Disable VGA resource access through vfio-pci");
53 static bool disable_idle_d3;
54 module_param(disable_idle_d3, bool, S_IRUGO | S_IWUSR);
55 MODULE_PARM_DESC(disable_idle_d3,
56 "Disable using the PCI D3 low power state for idle, unused devices");
58 static DEFINE_MUTEX(driver_lock);
60 static inline bool vfio_vga_disabled(void)
62 #ifdef CONFIG_VFIO_PCI_VGA
70 * Our VGA arbiter participation is limited since we don't know anything
71 * about the device itself. However, if the device is the only VGA device
72 * downstream of a bridge and VFIO VGA support is disabled, then we can
73 * safely return legacy VGA IO and memory as not decoded since the user
74 * has no way to get to it and routing can be disabled externally at the
77 static unsigned int vfio_pci_set_vga_decode(void *opaque, bool single_vga)
79 struct vfio_pci_device *vdev = opaque;
80 struct pci_dev *tmp = NULL, *pdev = vdev->pdev;
81 unsigned char max_busnr;
84 if (single_vga || !vfio_vga_disabled() || pci_is_root_bus(pdev->bus))
85 return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM |
86 VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM;
88 max_busnr = pci_bus_max_busnr(pdev->bus);
89 decodes = VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
91 while ((tmp = pci_get_class(PCI_CLASS_DISPLAY_VGA << 8, tmp)) != NULL) {
93 pci_domain_nr(tmp->bus) != pci_domain_nr(pdev->bus) ||
94 pci_is_root_bus(tmp->bus))
97 if (tmp->bus->number >= pdev->bus->number &&
98 tmp->bus->number <= max_busnr) {
100 decodes |= VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM;
108 static inline bool vfio_pci_is_vga(struct pci_dev *pdev)
110 return (pdev->class >> 8) == PCI_CLASS_DISPLAY_VGA;
113 static void vfio_pci_try_bus_reset(struct vfio_pci_device *vdev);
115 static int vfio_pci_enable(struct vfio_pci_device *vdev)
117 struct pci_dev *pdev = vdev->pdev;
122 pci_set_power_state(pdev, PCI_D0);
124 /* Don't allow our initial saved state to include busmaster */
125 pci_clear_master(pdev);
127 ret = pci_enable_device(pdev);
131 vdev->reset_works = (pci_reset_function(pdev) == 0);
132 pci_save_state(pdev);
133 vdev->pci_saved_state = pci_store_saved_state(pdev);
134 if (!vdev->pci_saved_state)
135 pr_debug("%s: Couldn't store %s saved state\n",
136 __func__, dev_name(&pdev->dev));
138 ret = vfio_config_init(vdev);
140 kfree(vdev->pci_saved_state);
141 vdev->pci_saved_state = NULL;
142 pci_disable_device(pdev);
146 if (likely(!nointxmask))
147 vdev->pci_2_3 = pci_intx_mask_supported(pdev);
149 pci_read_config_word(pdev, PCI_COMMAND, &cmd);
150 if (vdev->pci_2_3 && (cmd & PCI_COMMAND_INTX_DISABLE)) {
151 cmd &= ~PCI_COMMAND_INTX_DISABLE;
152 pci_write_config_word(pdev, PCI_COMMAND, cmd);
155 msix_pos = pdev->msix_cap;
160 pci_read_config_word(pdev, msix_pos + PCI_MSIX_FLAGS, &flags);
161 pci_read_config_dword(pdev, msix_pos + PCI_MSIX_TABLE, &table);
163 vdev->msix_bar = table & PCI_MSIX_TABLE_BIR;
164 vdev->msix_offset = table & PCI_MSIX_TABLE_OFFSET;
165 vdev->msix_size = ((flags & PCI_MSIX_FLAGS_QSIZE) + 1) * 16;
167 vdev->msix_bar = 0xFF;
169 if (!vfio_vga_disabled() && vfio_pci_is_vga(pdev))
170 vdev->has_vga = true;
175 static void vfio_pci_disable(struct vfio_pci_device *vdev)
177 struct pci_dev *pdev = vdev->pdev;
180 /* Stop the device from further DMA */
181 pci_clear_master(pdev);
183 vfio_pci_set_irqs_ioctl(vdev, VFIO_IRQ_SET_DATA_NONE |
184 VFIO_IRQ_SET_ACTION_TRIGGER,
185 vdev->irq_type, 0, 0, NULL);
187 vdev->virq_disabled = false;
189 vfio_config_free(vdev);
191 for (bar = PCI_STD_RESOURCES; bar <= PCI_STD_RESOURCE_END; bar++) {
192 if (!vdev->barmap[bar])
194 pci_iounmap(pdev, vdev->barmap[bar]);
195 pci_release_selected_regions(pdev, 1 << bar);
196 vdev->barmap[bar] = NULL;
199 vdev->needs_reset = true;
202 * If we have saved state, restore it. If we can reset the device,
203 * even better. Resetting with current state seems better than
204 * nothing, but saving and restoring current state without reset
207 if (pci_load_and_free_saved_state(pdev, &vdev->pci_saved_state)) {
208 pr_info("%s: Couldn't reload %s saved state\n",
209 __func__, dev_name(&pdev->dev));
211 if (!vdev->reset_works)
214 pci_save_state(pdev);
218 * Disable INTx and MSI, presumably to avoid spurious interrupts
219 * during reset. Stolen from pci_reset_function()
221 pci_write_config_word(pdev, PCI_COMMAND, PCI_COMMAND_INTX_DISABLE);
224 * Try to get the locks ourselves to prevent a deadlock. The
225 * success of this is dependent on being able to lock the device,
226 * which is not always possible.
227 * We can not use the "try" reset interface here, which will
228 * overwrite the previously restored configuration information.
230 if (vdev->reset_works && pci_cfg_access_trylock(pdev)) {
231 if (device_trylock(&pdev->dev)) {
232 if (!__pci_reset_function_locked(pdev))
233 vdev->needs_reset = false;
234 device_unlock(&pdev->dev);
236 pci_cfg_access_unlock(pdev);
239 pci_restore_state(pdev);
241 pci_disable_device(pdev);
243 vfio_pci_try_bus_reset(vdev);
245 if (!disable_idle_d3)
246 pci_set_power_state(pdev, PCI_D3hot);
249 static void vfio_pci_release(void *device_data)
251 struct vfio_pci_device *vdev = device_data;
253 mutex_lock(&driver_lock);
255 if (!(--vdev->refcnt)) {
256 vfio_spapr_pci_eeh_release(vdev->pdev);
257 vfio_pci_disable(vdev);
260 mutex_unlock(&driver_lock);
262 module_put(THIS_MODULE);
265 static int vfio_pci_open(void *device_data)
267 struct vfio_pci_device *vdev = device_data;
270 if (!try_module_get(THIS_MODULE))
273 mutex_lock(&driver_lock);
276 ret = vfio_pci_enable(vdev);
280 vfio_spapr_pci_eeh_open(vdev->pdev);
284 mutex_unlock(&driver_lock);
286 module_put(THIS_MODULE);
290 static int vfio_pci_get_irq_count(struct vfio_pci_device *vdev, int irq_type)
292 if (irq_type == VFIO_PCI_INTX_IRQ_INDEX) {
294 pci_read_config_byte(vdev->pdev, PCI_INTERRUPT_PIN, &pin);
295 if (IS_ENABLED(CONFIG_VFIO_PCI_INTX) && pin)
298 } else if (irq_type == VFIO_PCI_MSI_IRQ_INDEX) {
302 pos = vdev->pdev->msi_cap;
304 pci_read_config_word(vdev->pdev,
305 pos + PCI_MSI_FLAGS, &flags);
306 return 1 << ((flags & PCI_MSI_FLAGS_QMASK) >> 1);
308 } else if (irq_type == VFIO_PCI_MSIX_IRQ_INDEX) {
312 pos = vdev->pdev->msix_cap;
314 pci_read_config_word(vdev->pdev,
315 pos + PCI_MSIX_FLAGS, &flags);
317 return (flags & PCI_MSIX_FLAGS_QSIZE) + 1;
319 } else if (irq_type == VFIO_PCI_ERR_IRQ_INDEX) {
320 if (pci_is_pcie(vdev->pdev))
322 } else if (irq_type == VFIO_PCI_REQ_IRQ_INDEX) {
329 static int vfio_pci_count_devs(struct pci_dev *pdev, void *data)
335 struct vfio_pci_fill_info {
338 struct vfio_pci_dependent_device *devices;
341 static int vfio_pci_fill_devs(struct pci_dev *pdev, void *data)
343 struct vfio_pci_fill_info *fill = data;
344 struct iommu_group *iommu_group;
346 if (fill->cur == fill->max)
347 return -EAGAIN; /* Something changed, try again */
349 iommu_group = iommu_group_get(&pdev->dev);
351 return -EPERM; /* Cannot reset non-isolated devices */
353 fill->devices[fill->cur].group_id = iommu_group_id(iommu_group);
354 fill->devices[fill->cur].segment = pci_domain_nr(pdev->bus);
355 fill->devices[fill->cur].bus = pdev->bus->number;
356 fill->devices[fill->cur].devfn = pdev->devfn;
358 iommu_group_put(iommu_group);
362 struct vfio_pci_group_entry {
363 struct vfio_group *group;
367 struct vfio_pci_group_info {
369 struct vfio_pci_group_entry *groups;
372 static int vfio_pci_validate_devs(struct pci_dev *pdev, void *data)
374 struct vfio_pci_group_info *info = data;
375 struct iommu_group *group;
378 group = iommu_group_get(&pdev->dev);
382 id = iommu_group_id(group);
384 for (i = 0; i < info->count; i++)
385 if (info->groups[i].id == id)
388 iommu_group_put(group);
390 return (i == info->count) ? -EINVAL : 0;
393 static bool vfio_pci_dev_below_slot(struct pci_dev *pdev, struct pci_slot *slot)
395 for (; pdev; pdev = pdev->bus->self)
396 if (pdev->bus == slot->bus)
397 return (pdev->slot == slot);
401 struct vfio_pci_walk_info {
402 int (*fn)(struct pci_dev *, void *data);
404 struct pci_dev *pdev;
409 static int vfio_pci_walk_wrapper(struct pci_dev *pdev, void *data)
411 struct vfio_pci_walk_info *walk = data;
413 if (!walk->slot || vfio_pci_dev_below_slot(pdev, walk->pdev->slot))
414 walk->ret = walk->fn(pdev, walk->data);
419 static int vfio_pci_for_each_slot_or_bus(struct pci_dev *pdev,
420 int (*fn)(struct pci_dev *,
421 void *data), void *data,
424 struct vfio_pci_walk_info walk = {
425 .fn = fn, .data = data, .pdev = pdev, .slot = slot, .ret = 0,
428 pci_walk_bus(pdev->bus, vfio_pci_walk_wrapper, &walk);
433 static long vfio_pci_ioctl(void *device_data,
434 unsigned int cmd, unsigned long arg)
436 struct vfio_pci_device *vdev = device_data;
439 if (cmd == VFIO_DEVICE_GET_INFO) {
440 struct vfio_device_info info;
442 minsz = offsetofend(struct vfio_device_info, num_irqs);
444 if (copy_from_user(&info, (void __user *)arg, minsz))
447 if (info.argsz < minsz)
450 info.flags = VFIO_DEVICE_FLAGS_PCI;
452 if (vdev->reset_works)
453 info.flags |= VFIO_DEVICE_FLAGS_RESET;
455 info.num_regions = VFIO_PCI_NUM_REGIONS;
456 info.num_irqs = VFIO_PCI_NUM_IRQS;
458 return copy_to_user((void __user *)arg, &info, minsz) ?
461 } else if (cmd == VFIO_DEVICE_GET_REGION_INFO) {
462 struct pci_dev *pdev = vdev->pdev;
463 struct vfio_region_info info;
465 minsz = offsetofend(struct vfio_region_info, offset);
467 if (copy_from_user(&info, (void __user *)arg, minsz))
470 if (info.argsz < minsz)
473 switch (info.index) {
474 case VFIO_PCI_CONFIG_REGION_INDEX:
475 info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
476 info.size = pdev->cfg_size;
477 info.flags = VFIO_REGION_INFO_FLAG_READ |
478 VFIO_REGION_INFO_FLAG_WRITE;
480 case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
481 info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
482 info.size = pci_resource_len(pdev, info.index);
488 info.flags = VFIO_REGION_INFO_FLAG_READ |
489 VFIO_REGION_INFO_FLAG_WRITE;
490 if (IS_ENABLED(CONFIG_VFIO_PCI_MMAP) &&
491 pci_resource_flags(pdev, info.index) &
492 IORESOURCE_MEM && info.size >= PAGE_SIZE)
493 info.flags |= VFIO_REGION_INFO_FLAG_MMAP;
495 case VFIO_PCI_ROM_REGION_INDEX:
501 info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
504 /* Report the BAR size, not the ROM size */
505 info.size = pci_resource_len(pdev, info.index);
510 * Is it really there? Enable memory decode for
511 * implicit access in pci_map_rom().
513 pci_read_config_word(pdev, PCI_COMMAND, &orig_cmd);
514 pci_write_config_word(pdev, PCI_COMMAND,
515 orig_cmd | PCI_COMMAND_MEMORY);
517 io = pci_map_rom(pdev, &size);
519 info.flags = VFIO_REGION_INFO_FLAG_READ;
520 pci_unmap_rom(pdev, io);
525 pci_write_config_word(pdev, PCI_COMMAND, orig_cmd);
528 case VFIO_PCI_VGA_REGION_INDEX:
532 info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
534 info.flags = VFIO_REGION_INFO_FLAG_READ |
535 VFIO_REGION_INFO_FLAG_WRITE;
542 return copy_to_user((void __user *)arg, &info, minsz) ?
545 } else if (cmd == VFIO_DEVICE_GET_IRQ_INFO) {
546 struct vfio_irq_info info;
548 minsz = offsetofend(struct vfio_irq_info, count);
550 if (copy_from_user(&info, (void __user *)arg, minsz))
553 if (info.argsz < minsz || info.index >= VFIO_PCI_NUM_IRQS)
556 switch (info.index) {
557 case VFIO_PCI_INTX_IRQ_INDEX ... VFIO_PCI_MSIX_IRQ_INDEX:
558 case VFIO_PCI_REQ_IRQ_INDEX:
560 case VFIO_PCI_ERR_IRQ_INDEX:
561 if (pci_is_pcie(vdev->pdev))
563 /* pass thru to return error */
568 info.flags = VFIO_IRQ_INFO_EVENTFD;
570 info.count = vfio_pci_get_irq_count(vdev, info.index);
572 if (info.index == VFIO_PCI_INTX_IRQ_INDEX)
573 info.flags |= (VFIO_IRQ_INFO_MASKABLE |
574 VFIO_IRQ_INFO_AUTOMASKED);
576 info.flags |= VFIO_IRQ_INFO_NORESIZE;
578 return copy_to_user((void __user *)arg, &info, minsz) ?
581 } else if (cmd == VFIO_DEVICE_SET_IRQS) {
582 struct vfio_irq_set hdr;
587 minsz = offsetofend(struct vfio_irq_set, count);
589 if (copy_from_user(&hdr, (void __user *)arg, minsz))
592 if (hdr.argsz < minsz || hdr.index >= VFIO_PCI_NUM_IRQS ||
593 hdr.count >= (U32_MAX - hdr.start) ||
594 hdr.flags & ~(VFIO_IRQ_SET_DATA_TYPE_MASK |
595 VFIO_IRQ_SET_ACTION_TYPE_MASK))
598 max = vfio_pci_get_irq_count(vdev, hdr.index);
599 if (hdr.start >= max || hdr.start + hdr.count > max)
602 switch (hdr.flags & VFIO_IRQ_SET_DATA_TYPE_MASK) {
603 case VFIO_IRQ_SET_DATA_NONE:
606 case VFIO_IRQ_SET_DATA_BOOL:
607 size = sizeof(uint8_t);
609 case VFIO_IRQ_SET_DATA_EVENTFD:
610 size = sizeof(int32_t);
617 if (hdr.argsz - minsz < hdr.count * size)
620 data = memdup_user((void __user *)(arg + minsz),
623 return PTR_ERR(data);
626 mutex_lock(&vdev->igate);
628 ret = vfio_pci_set_irqs_ioctl(vdev, hdr.flags, hdr.index,
629 hdr.start, hdr.count, data);
631 mutex_unlock(&vdev->igate);
636 } else if (cmd == VFIO_DEVICE_RESET) {
637 return vdev->reset_works ?
638 pci_try_reset_function(vdev->pdev) : -EINVAL;
640 } else if (cmd == VFIO_DEVICE_GET_PCI_HOT_RESET_INFO) {
641 struct vfio_pci_hot_reset_info hdr;
642 struct vfio_pci_fill_info fill = { 0 };
643 struct vfio_pci_dependent_device *devices = NULL;
647 minsz = offsetofend(struct vfio_pci_hot_reset_info, count);
649 if (copy_from_user(&hdr, (void __user *)arg, minsz))
652 if (hdr.argsz < minsz)
657 /* Can we do a slot or bus reset or neither? */
658 if (!pci_probe_reset_slot(vdev->pdev->slot))
660 else if (pci_probe_reset_bus(vdev->pdev->bus))
663 /* How many devices are affected? */
664 ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
670 WARN_ON(!fill.max); /* Should always be at least one */
673 * If there's enough space, fill it now, otherwise return
674 * -ENOSPC and the number of devices affected.
676 if (hdr.argsz < sizeof(hdr) + (fill.max * sizeof(*devices))) {
678 hdr.count = fill.max;
679 goto reset_info_exit;
682 devices = kcalloc(fill.max, sizeof(*devices), GFP_KERNEL);
686 fill.devices = devices;
688 ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
693 * If a device was removed between counting and filling,
694 * we may come up short of fill.max. If a device was
695 * added, we'll have a return of -EAGAIN above.
698 hdr.count = fill.cur;
701 if (copy_to_user((void __user *)arg, &hdr, minsz))
705 if (copy_to_user((void __user *)(arg + minsz), devices,
706 hdr.count * sizeof(*devices)))
713 } else if (cmd == VFIO_DEVICE_PCI_HOT_RESET) {
714 struct vfio_pci_hot_reset hdr;
716 struct vfio_pci_group_entry *groups;
717 struct vfio_pci_group_info info;
719 int i, count = 0, ret = 0;
721 minsz = offsetofend(struct vfio_pci_hot_reset, count);
723 if (copy_from_user(&hdr, (void __user *)arg, minsz))
726 if (hdr.argsz < minsz || hdr.flags)
729 /* Can we do a slot or bus reset or neither? */
730 if (!pci_probe_reset_slot(vdev->pdev->slot))
732 else if (pci_probe_reset_bus(vdev->pdev->bus))
736 * We can't let userspace give us an arbitrarily large
737 * buffer to copy, so verify how many we think there
738 * could be. Note groups can have multiple devices so
739 * one group per device is the max.
741 ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
747 /* Somewhere between 1 and count is OK */
748 if (!hdr.count || hdr.count > count)
751 group_fds = kcalloc(hdr.count, sizeof(*group_fds), GFP_KERNEL);
752 groups = kcalloc(hdr.count, sizeof(*groups), GFP_KERNEL);
753 if (!group_fds || !groups) {
759 if (copy_from_user(group_fds, (void __user *)(arg + minsz),
760 hdr.count * sizeof(*group_fds))) {
767 * For each group_fd, get the group through the vfio external
768 * user interface and store the group and iommu ID. This
769 * ensures the group is held across the reset.
771 for (i = 0; i < hdr.count; i++) {
772 struct vfio_group *group;
773 struct fd f = fdget(group_fds[i]);
779 group = vfio_group_get_external_user(f.file);
782 ret = PTR_ERR(group);
786 groups[i].group = group;
787 groups[i].id = vfio_external_user_iommu_id(group);
792 /* release reference to groups on error */
794 goto hot_reset_release;
796 info.count = hdr.count;
797 info.groups = groups;
800 * Test whether all the affected devices are contained
801 * by the set of groups provided by the user.
803 ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
804 vfio_pci_validate_devs,
807 /* User has access, do the reset */
808 ret = slot ? pci_try_reset_slot(vdev->pdev->slot) :
809 pci_try_reset_bus(vdev->pdev->bus);
812 for (i--; i >= 0; i--)
813 vfio_group_put_external_user(groups[i].group);
822 static ssize_t vfio_pci_rw(void *device_data, char __user *buf,
823 size_t count, loff_t *ppos, bool iswrite)
825 unsigned int index = VFIO_PCI_OFFSET_TO_INDEX(*ppos);
826 struct vfio_pci_device *vdev = device_data;
828 if (index >= VFIO_PCI_NUM_REGIONS)
832 case VFIO_PCI_CONFIG_REGION_INDEX:
833 return vfio_pci_config_rw(vdev, buf, count, ppos, iswrite);
835 case VFIO_PCI_ROM_REGION_INDEX:
838 return vfio_pci_bar_rw(vdev, buf, count, ppos, false);
840 case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
841 return vfio_pci_bar_rw(vdev, buf, count, ppos, iswrite);
843 case VFIO_PCI_VGA_REGION_INDEX:
844 return vfio_pci_vga_rw(vdev, buf, count, ppos, iswrite);
850 static ssize_t vfio_pci_read(void *device_data, char __user *buf,
851 size_t count, loff_t *ppos)
856 return vfio_pci_rw(device_data, buf, count, ppos, false);
859 static ssize_t vfio_pci_write(void *device_data, const char __user *buf,
860 size_t count, loff_t *ppos)
865 return vfio_pci_rw(device_data, (char __user *)buf, count, ppos, true);
868 static int vfio_pci_mmap(void *device_data, struct vm_area_struct *vma)
870 struct vfio_pci_device *vdev = device_data;
871 struct pci_dev *pdev = vdev->pdev;
873 u64 phys_len, req_len, pgoff, req_start;
876 index = vma->vm_pgoff >> (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT);
878 if (vma->vm_end < vma->vm_start)
880 if ((vma->vm_flags & VM_SHARED) == 0)
882 if (index >= VFIO_PCI_ROM_REGION_INDEX)
884 if (!(pci_resource_flags(pdev, index) & IORESOURCE_MEM))
887 phys_len = pci_resource_len(pdev, index);
888 req_len = vma->vm_end - vma->vm_start;
889 pgoff = vma->vm_pgoff &
890 ((1U << (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT)) - 1);
891 req_start = pgoff << PAGE_SHIFT;
893 if (phys_len < PAGE_SIZE || req_start + req_len > phys_len)
896 if (index == vdev->msix_bar) {
898 * Disallow mmaps overlapping the MSI-X table; users don't
899 * get to touch this directly. We could find somewhere
900 * else to map the overlap, but page granularity is only
901 * a recommendation, not a requirement, so the user needs
902 * to know which bits are real. Requiring them to mmap
903 * around the table makes that clear.
906 /* If neither entirely above nor below, then it overlaps */
907 if (!(req_start >= vdev->msix_offset + vdev->msix_size ||
908 req_start + req_len <= vdev->msix_offset))
913 * Even though we don't make use of the barmap for the mmap,
914 * we need to request the region and the barmap tracks that.
916 if (!vdev->barmap[index]) {
917 ret = pci_request_selected_regions(pdev,
918 1 << index, "vfio-pci");
922 vdev->barmap[index] = pci_iomap(pdev, index, 0);
923 if (!vdev->barmap[index]) {
924 pci_release_selected_regions(pdev, 1 << index);
929 vma->vm_private_data = vdev;
930 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
931 vma->vm_pgoff = (pci_resource_start(pdev, index) >> PAGE_SHIFT) + pgoff;
933 return remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
934 req_len, vma->vm_page_prot);
937 static void vfio_pci_request(void *device_data, unsigned int count)
939 struct vfio_pci_device *vdev = device_data;
941 mutex_lock(&vdev->igate);
943 if (vdev->req_trigger) {
945 dev_notice_ratelimited(&vdev->pdev->dev,
946 "Relaying device request to user (#%u)\n",
948 eventfd_signal(vdev->req_trigger, 1);
949 } else if (count == 0) {
950 dev_warn(&vdev->pdev->dev,
951 "No device request channel registered, blocked until released by user\n");
954 mutex_unlock(&vdev->igate);
957 static const struct vfio_device_ops vfio_pci_ops = {
959 .open = vfio_pci_open,
960 .release = vfio_pci_release,
961 .ioctl = vfio_pci_ioctl,
962 .read = vfio_pci_read,
963 .write = vfio_pci_write,
964 .mmap = vfio_pci_mmap,
965 .request = vfio_pci_request,
968 static int vfio_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
970 struct vfio_pci_device *vdev;
971 struct iommu_group *group;
974 if (pdev->hdr_type != PCI_HEADER_TYPE_NORMAL)
977 group = iommu_group_get(&pdev->dev);
981 vdev = kzalloc(sizeof(*vdev), GFP_KERNEL);
983 iommu_group_put(group);
988 vdev->irq_type = VFIO_PCI_NUM_IRQS;
989 mutex_init(&vdev->igate);
990 spin_lock_init(&vdev->irqlock);
992 ret = vfio_add_group_dev(&pdev->dev, &vfio_pci_ops, vdev);
994 iommu_group_put(group);
999 if (vfio_pci_is_vga(pdev)) {
1000 vga_client_register(pdev, vdev, NULL, vfio_pci_set_vga_decode);
1001 vga_set_legacy_decoding(pdev,
1002 vfio_pci_set_vga_decode(vdev, false));
1005 if (!disable_idle_d3) {
1007 * pci-core sets the device power state to an unknown value at
1008 * bootup and after being removed from a driver. The only
1009 * transition it allows from this unknown state is to D0, which
1010 * typically happens when a driver calls pci_enable_device().
1011 * We're not ready to enable the device yet, but we do want to
1012 * be able to get to D3. Therefore first do a D0 transition
1013 * before going to D3.
1015 pci_set_power_state(pdev, PCI_D0);
1016 pci_set_power_state(pdev, PCI_D3hot);
1022 static void vfio_pci_remove(struct pci_dev *pdev)
1024 struct vfio_pci_device *vdev;
1026 vdev = vfio_del_group_dev(&pdev->dev);
1030 iommu_group_put(pdev->dev.iommu_group);
1033 if (vfio_pci_is_vga(pdev)) {
1034 vga_client_register(pdev, NULL, NULL, NULL);
1035 vga_set_legacy_decoding(pdev,
1036 VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM |
1037 VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM);
1040 if (!disable_idle_d3)
1041 pci_set_power_state(pdev, PCI_D0);
1044 static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
1045 pci_channel_state_t state)
1047 struct vfio_pci_device *vdev;
1048 struct vfio_device *device;
1050 device = vfio_device_get_from_dev(&pdev->dev);
1052 return PCI_ERS_RESULT_DISCONNECT;
1054 vdev = vfio_device_data(device);
1056 vfio_device_put(device);
1057 return PCI_ERS_RESULT_DISCONNECT;
1060 mutex_lock(&vdev->igate);
1062 if (vdev->err_trigger)
1063 eventfd_signal(vdev->err_trigger, 1);
1065 mutex_unlock(&vdev->igate);
1067 vfio_device_put(device);
1069 return PCI_ERS_RESULT_CAN_RECOVER;
1072 static const struct pci_error_handlers vfio_err_handlers = {
1073 .error_detected = vfio_pci_aer_err_detected,
1076 static struct pci_driver vfio_pci_driver = {
1078 .id_table = NULL, /* only dynamic ids */
1079 .probe = vfio_pci_probe,
1080 .remove = vfio_pci_remove,
1081 .err_handler = &vfio_err_handlers,
1084 struct vfio_devices {
1085 struct vfio_device **devices;
1090 static int vfio_pci_get_devs(struct pci_dev *pdev, void *data)
1092 struct vfio_devices *devs = data;
1093 struct vfio_device *device;
1095 if (devs->cur_index == devs->max_index)
1098 device = vfio_device_get_from_dev(&pdev->dev);
1102 if (pci_dev_driver(pdev) != &vfio_pci_driver) {
1103 vfio_device_put(device);
1107 devs->devices[devs->cur_index++] = device;
1112 * Attempt to do a bus/slot reset if there are devices affected by a reset for
1113 * this device that are needs_reset and all of the affected devices are unused
1114 * (!refcnt). Callers are required to hold driver_lock when calling this to
1115 * prevent device opens and concurrent bus reset attempts. We prevent device
1116 * unbinds by acquiring and holding a reference to the vfio_device.
1118 * NB: vfio-core considers a group to be viable even if some devices are
1119 * bound to drivers like pci-stub or pcieport. Here we require all devices
1120 * to be bound to vfio_pci since that's the only way we can be sure they
1123 static void vfio_pci_try_bus_reset(struct vfio_pci_device *vdev)
1125 struct vfio_devices devs = { .cur_index = 0 };
1126 int i = 0, ret = -EINVAL;
1127 bool needs_reset = false, slot = false;
1128 struct vfio_pci_device *tmp;
1130 if (!pci_probe_reset_slot(vdev->pdev->slot))
1132 else if (pci_probe_reset_bus(vdev->pdev->bus))
1135 if (vfio_pci_for_each_slot_or_bus(vdev->pdev, vfio_pci_count_devs,
1140 devs.devices = kcalloc(i, sizeof(struct vfio_device *), GFP_KERNEL);
1144 if (vfio_pci_for_each_slot_or_bus(vdev->pdev,
1145 vfio_pci_get_devs, &devs, slot))
1148 for (i = 0; i < devs.cur_index; i++) {
1149 tmp = vfio_device_data(devs.devices[i]);
1150 if (tmp->needs_reset)
1157 ret = slot ? pci_try_reset_slot(vdev->pdev->slot) :
1158 pci_try_reset_bus(vdev->pdev->bus);
1161 for (i = 0; i < devs.cur_index; i++) {
1162 tmp = vfio_device_data(devs.devices[i]);
1164 tmp->needs_reset = false;
1166 if (!tmp->refcnt && !disable_idle_d3)
1167 pci_set_power_state(tmp->pdev, PCI_D3hot);
1169 vfio_device_put(devs.devices[i]);
1172 kfree(devs.devices);
1175 static void __exit vfio_pci_cleanup(void)
1177 pci_unregister_driver(&vfio_pci_driver);
1178 vfio_pci_uninit_perm_bits();
1181 static void __init vfio_pci_fill_ids(void)
1186 /* no ids passed actually */
1190 /* add ids specified in the module parameter */
1192 while ((id = strsep(&p, ","))) {
1193 unsigned int vendor, device, subvendor = PCI_ANY_ID,
1194 subdevice = PCI_ANY_ID, class = 0, class_mask = 0;
1200 fields = sscanf(id, "%x:%x:%x:%x:%x:%x",
1201 &vendor, &device, &subvendor, &subdevice,
1202 &class, &class_mask);
1205 pr_warn("invalid id string \"%s\"\n", id);
1209 rc = pci_add_dynid(&vfio_pci_driver, vendor, device,
1210 subvendor, subdevice, class, class_mask, 0);
1212 pr_warn("failed to add dynamic id [%04x:%04x[%04x:%04x]] class %#08x/%08x (%d)\n",
1213 vendor, device, subvendor, subdevice,
1214 class, class_mask, rc);
1216 pr_info("add [%04x:%04x[%04x:%04x]] class %#08x/%08x\n",
1217 vendor, device, subvendor, subdevice,
1222 static int __init vfio_pci_init(void)
1226 /* Allocate shared config space permision data used by all devices */
1227 ret = vfio_pci_init_perm_bits();
1231 /* Register and scan for devices */
1232 ret = pci_register_driver(&vfio_pci_driver);
1236 vfio_pci_fill_ids();
1241 vfio_pci_uninit_perm_bits();
1245 module_init(vfio_pci_init);
1246 module_exit(vfio_pci_cleanup);
1248 MODULE_VERSION(DRIVER_VERSION);
1249 MODULE_LICENSE("GPL v2");
1250 MODULE_AUTHOR(DRIVER_AUTHOR);
1251 MODULE_DESCRIPTION(DRIVER_DESC);