1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (C) 2012 Red Hat, Inc. All rights reserved.
4 * Author: Alex Williamson <alex.williamson@redhat.com>
6 * Derived from original vfio:
7 * Copyright 2010 Cisco Systems, Inc. All rights reserved.
8 * Author: Tom Lyon, pugs@cisco.com
11 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
13 #include <linux/device.h>
14 #include <linux/eventfd.h>
15 #include <linux/file.h>
16 #include <linux/interrupt.h>
17 #include <linux/iommu.h>
18 #include <linux/module.h>
19 #include <linux/mutex.h>
20 #include <linux/notifier.h>
21 #include <linux/pci.h>
22 #include <linux/pm_runtime.h>
23 #include <linux/slab.h>
24 #include <linux/types.h>
25 #include <linux/uaccess.h>
26 #include <linux/vgaarb.h>
27 #include <linux/nospec.h>
28 #include <linux/sched/mm.h>
30 #include <linux/vfio_pci_core.h>
32 #define DRIVER_AUTHOR "Alex Williamson <alex.williamson@redhat.com>"
33 #define DRIVER_DESC "core driver for VFIO based PCI devices"
35 static bool nointxmask;
36 static bool disable_vga;
37 static bool disable_idle_d3;
39 static inline bool vfio_vga_disabled(void)
41 #ifdef CONFIG_VFIO_PCI_VGA
49 * Our VGA arbiter participation is limited since we don't know anything
50 * about the device itself. However, if the device is the only VGA device
51 * downstream of a bridge and VFIO VGA support is disabled, then we can
52 * safely return legacy VGA IO and memory as not decoded since the user
53 * has no way to get to it and routing can be disabled externally at the
56 static unsigned int vfio_pci_set_decode(struct pci_dev *pdev, bool single_vga)
58 struct pci_dev *tmp = NULL;
59 unsigned char max_busnr;
62 if (single_vga || !vfio_vga_disabled() || pci_is_root_bus(pdev->bus))
63 return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM |
64 VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM;
66 max_busnr = pci_bus_max_busnr(pdev->bus);
67 decodes = VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
69 while ((tmp = pci_get_class(PCI_CLASS_DISPLAY_VGA << 8, tmp)) != NULL) {
71 pci_domain_nr(tmp->bus) != pci_domain_nr(pdev->bus) ||
72 pci_is_root_bus(tmp->bus))
75 if (tmp->bus->number >= pdev->bus->number &&
76 tmp->bus->number <= max_busnr) {
78 decodes |= VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM;
86 static void vfio_pci_probe_mmaps(struct vfio_pci_core_device *vdev)
90 struct vfio_pci_dummy_resource *dummy_res;
92 for (i = 0; i < PCI_STD_NUM_BARS; i++) {
93 int bar = i + PCI_STD_RESOURCES;
95 res = &vdev->pdev->resource[bar];
97 if (!IS_ENABLED(CONFIG_VFIO_PCI_MMAP))
100 if (!(res->flags & IORESOURCE_MEM))
104 * The PCI core shouldn't set up a resource with a
105 * type but zero size. But there may be bugs that
106 * cause us to do that.
108 if (!resource_size(res))
111 if (resource_size(res) >= PAGE_SIZE) {
112 vdev->bar_mmap_supported[bar] = true;
116 if (!(res->start & ~PAGE_MASK)) {
118 * Add a dummy resource to reserve the remainder
119 * of the exclusive page in case that hot-add
120 * device's bar is assigned into it.
122 dummy_res = kzalloc(sizeof(*dummy_res), GFP_KERNEL);
123 if (dummy_res == NULL)
126 dummy_res->resource.name = "vfio sub-page reserved";
127 dummy_res->resource.start = res->end + 1;
128 dummy_res->resource.end = res->start + PAGE_SIZE - 1;
129 dummy_res->resource.flags = res->flags;
130 if (request_resource(res->parent,
131 &dummy_res->resource)) {
135 dummy_res->index = bar;
136 list_add(&dummy_res->res_next,
137 &vdev->dummy_resources_list);
138 vdev->bar_mmap_supported[bar] = true;
142 * Here we don't handle the case when the BAR is not page
143 * aligned because we can't expect the BAR will be
144 * assigned into the same location in a page in guest
145 * when we passthrough the BAR. And it's hard to access
146 * this BAR in userspace because we have no way to get
147 * the BAR's location in a page.
150 vdev->bar_mmap_supported[bar] = false;
154 struct vfio_pci_group_info;
155 static bool vfio_pci_dev_set_try_reset(struct vfio_device_set *dev_set);
156 static int vfio_pci_dev_set_hot_reset(struct vfio_device_set *dev_set,
157 struct vfio_pci_group_info *groups);
160 * INTx masking requires the ability to disable INTx signaling via PCI_COMMAND
161 * _and_ the ability detect when the device is asserting INTx via PCI_STATUS.
162 * If a device implements the former but not the latter we would typically
163 * expect broken_intx_masking be set and require an exclusive interrupt.
164 * However since we do have control of the device's ability to assert INTx,
165 * we can instead pretend that the device does not implement INTx, virtualizing
166 * the pin register to report zero and maintaining DisINTx set on the host.
168 static bool vfio_pci_nointx(struct pci_dev *pdev)
170 switch (pdev->vendor) {
171 case PCI_VENDOR_ID_INTEL:
172 switch (pdev->device) {
173 /* All i40e (XL710/X710/XXV710) 10/20/25/40GbE NICs */
176 case 0x1580 ... 0x1581:
177 case 0x1583 ... 0x158b:
178 case 0x37d0 ... 0x37d2:
190 static void vfio_pci_probe_power_state(struct vfio_pci_core_device *vdev)
192 struct pci_dev *pdev = vdev->pdev;
198 pci_read_config_word(pdev, pdev->pm_cap + PCI_PM_CTRL, &pmcsr);
200 vdev->needs_pm_restore = !(pmcsr & PCI_PM_CTRL_NO_SOFT_RESET);
204 * pci_set_power_state() wrapper handling devices which perform a soft reset on
205 * D3->D0 transition. Save state prior to D0/1/2->D3, stash it on the vdev,
206 * restore when returned to D0. Saved separately from pci_saved_state for use
207 * by PM capability emulation and separately from pci_dev internal saved state
208 * to avoid it being overwritten and consumed around other resets.
210 int vfio_pci_set_power_state(struct vfio_pci_core_device *vdev, pci_power_t state)
212 struct pci_dev *pdev = vdev->pdev;
213 bool needs_restore = false, needs_save = false;
216 if (vdev->needs_pm_restore) {
217 if (pdev->current_state < PCI_D3hot && state >= PCI_D3hot) {
218 pci_save_state(pdev);
222 if (pdev->current_state >= PCI_D3hot && state <= PCI_D0)
223 needs_restore = true;
226 ret = pci_set_power_state(pdev, state);
229 /* D3 might be unsupported via quirk, skip unless in D3 */
230 if (needs_save && pdev->current_state >= PCI_D3hot) {
231 vdev->pm_save = pci_store_saved_state(pdev);
232 } else if (needs_restore) {
233 pci_load_and_free_saved_state(pdev, &vdev->pm_save);
234 pci_restore_state(pdev);
241 int vfio_pci_core_enable(struct vfio_pci_core_device *vdev)
243 struct pci_dev *pdev = vdev->pdev;
248 vfio_pci_set_power_state(vdev, PCI_D0);
250 /* Don't allow our initial saved state to include busmaster */
251 pci_clear_master(pdev);
253 ret = pci_enable_device(pdev);
257 /* If reset fails because of the device lock, fail this path entirely */
258 ret = pci_try_reset_function(pdev);
259 if (ret == -EAGAIN) {
260 pci_disable_device(pdev);
264 vdev->reset_works = !ret;
265 pci_save_state(pdev);
266 vdev->pci_saved_state = pci_store_saved_state(pdev);
267 if (!vdev->pci_saved_state)
268 pci_dbg(pdev, "%s: Couldn't store saved state\n", __func__);
270 if (likely(!nointxmask)) {
271 if (vfio_pci_nointx(pdev)) {
272 pci_info(pdev, "Masking broken INTx support\n");
276 vdev->pci_2_3 = pci_intx_mask_supported(pdev);
279 pci_read_config_word(pdev, PCI_COMMAND, &cmd);
280 if (vdev->pci_2_3 && (cmd & PCI_COMMAND_INTX_DISABLE)) {
281 cmd &= ~PCI_COMMAND_INTX_DISABLE;
282 pci_write_config_word(pdev, PCI_COMMAND, cmd);
285 ret = vfio_config_init(vdev);
287 kfree(vdev->pci_saved_state);
288 vdev->pci_saved_state = NULL;
289 pci_disable_device(pdev);
293 msix_pos = pdev->msix_cap;
298 pci_read_config_word(pdev, msix_pos + PCI_MSIX_FLAGS, &flags);
299 pci_read_config_dword(pdev, msix_pos + PCI_MSIX_TABLE, &table);
301 vdev->msix_bar = table & PCI_MSIX_TABLE_BIR;
302 vdev->msix_offset = table & PCI_MSIX_TABLE_OFFSET;
303 vdev->msix_size = ((flags & PCI_MSIX_FLAGS_QSIZE) + 1) * 16;
305 vdev->msix_bar = 0xFF;
307 if (!vfio_vga_disabled() && vfio_pci_is_vga(pdev))
308 vdev->has_vga = true;
313 EXPORT_SYMBOL_GPL(vfio_pci_core_enable);
315 void vfio_pci_core_disable(struct vfio_pci_core_device *vdev)
317 struct pci_dev *pdev = vdev->pdev;
318 struct vfio_pci_dummy_resource *dummy_res, *tmp;
319 struct vfio_pci_ioeventfd *ioeventfd, *ioeventfd_tmp;
322 /* For needs_reset */
323 lockdep_assert_held(&vdev->vdev.dev_set->lock);
325 /* Stop the device from further DMA */
326 pci_clear_master(pdev);
328 vfio_pci_set_irqs_ioctl(vdev, VFIO_IRQ_SET_DATA_NONE |
329 VFIO_IRQ_SET_ACTION_TRIGGER,
330 vdev->irq_type, 0, 0, NULL);
332 /* Device closed, don't need mutex here */
333 list_for_each_entry_safe(ioeventfd, ioeventfd_tmp,
334 &vdev->ioeventfds_list, next) {
335 vfio_virqfd_disable(&ioeventfd->virqfd);
336 list_del(&ioeventfd->next);
339 vdev->ioeventfds_nr = 0;
341 vdev->virq_disabled = false;
343 for (i = 0; i < vdev->num_regions; i++)
344 vdev->region[i].ops->release(vdev, &vdev->region[i]);
346 vdev->num_regions = 0;
348 vdev->region = NULL; /* don't krealloc a freed pointer */
350 vfio_config_free(vdev);
352 for (i = 0; i < PCI_STD_NUM_BARS; i++) {
353 bar = i + PCI_STD_RESOURCES;
354 if (!vdev->barmap[bar])
356 pci_iounmap(pdev, vdev->barmap[bar]);
357 pci_release_selected_regions(pdev, 1 << bar);
358 vdev->barmap[bar] = NULL;
361 list_for_each_entry_safe(dummy_res, tmp,
362 &vdev->dummy_resources_list, res_next) {
363 list_del(&dummy_res->res_next);
364 release_resource(&dummy_res->resource);
368 vdev->needs_reset = true;
371 * If we have saved state, restore it. If we can reset the device,
372 * even better. Resetting with current state seems better than
373 * nothing, but saving and restoring current state without reset
376 if (pci_load_and_free_saved_state(pdev, &vdev->pci_saved_state)) {
377 pci_info(pdev, "%s: Couldn't reload saved state\n", __func__);
379 if (!vdev->reset_works)
382 pci_save_state(pdev);
386 * Disable INTx and MSI, presumably to avoid spurious interrupts
387 * during reset. Stolen from pci_reset_function()
389 pci_write_config_word(pdev, PCI_COMMAND, PCI_COMMAND_INTX_DISABLE);
392 * Try to get the locks ourselves to prevent a deadlock. The
393 * success of this is dependent on being able to lock the device,
394 * which is not always possible.
395 * We can not use the "try" reset interface here, which will
396 * overwrite the previously restored configuration information.
398 if (vdev->reset_works && pci_dev_trylock(pdev)) {
399 if (!__pci_reset_function_locked(pdev))
400 vdev->needs_reset = false;
401 pci_dev_unlock(pdev);
404 pci_restore_state(pdev);
406 pci_disable_device(pdev);
408 if (!vfio_pci_dev_set_try_reset(vdev->vdev.dev_set) && !disable_idle_d3)
409 vfio_pci_set_power_state(vdev, PCI_D3hot);
411 EXPORT_SYMBOL_GPL(vfio_pci_core_disable);
413 static struct vfio_pci_core_device *get_pf_vdev(struct vfio_pci_core_device *vdev)
415 struct pci_dev *physfn = pci_physfn(vdev->pdev);
416 struct vfio_device *pf_dev;
418 if (!vdev->pdev->is_virtfn)
421 pf_dev = vfio_device_get_from_dev(&physfn->dev);
425 if (pci_dev_driver(physfn) != pci_dev_driver(vdev->pdev)) {
426 vfio_device_put(pf_dev);
430 return container_of(pf_dev, struct vfio_pci_core_device, vdev);
433 static void vfio_pci_vf_token_user_add(struct vfio_pci_core_device *vdev, int val)
435 struct vfio_pci_core_device *pf_vdev = get_pf_vdev(vdev);
440 mutex_lock(&pf_vdev->vf_token->lock);
441 pf_vdev->vf_token->users += val;
442 WARN_ON(pf_vdev->vf_token->users < 0);
443 mutex_unlock(&pf_vdev->vf_token->lock);
445 vfio_device_put(&pf_vdev->vdev);
448 void vfio_pci_core_close_device(struct vfio_device *core_vdev)
450 struct vfio_pci_core_device *vdev =
451 container_of(core_vdev, struct vfio_pci_core_device, vdev);
453 vfio_pci_vf_token_user_add(vdev, -1);
454 vfio_spapr_pci_eeh_release(vdev->pdev);
455 vfio_pci_core_disable(vdev);
457 mutex_lock(&vdev->igate);
458 if (vdev->err_trigger) {
459 eventfd_ctx_put(vdev->err_trigger);
460 vdev->err_trigger = NULL;
462 if (vdev->req_trigger) {
463 eventfd_ctx_put(vdev->req_trigger);
464 vdev->req_trigger = NULL;
466 mutex_unlock(&vdev->igate);
468 EXPORT_SYMBOL_GPL(vfio_pci_core_close_device);
470 void vfio_pci_core_finish_enable(struct vfio_pci_core_device *vdev)
472 vfio_pci_probe_mmaps(vdev);
473 vfio_spapr_pci_eeh_open(vdev->pdev);
474 vfio_pci_vf_token_user_add(vdev, 1);
476 EXPORT_SYMBOL_GPL(vfio_pci_core_finish_enable);
478 static int vfio_pci_get_irq_count(struct vfio_pci_core_device *vdev, int irq_type)
480 if (irq_type == VFIO_PCI_INTX_IRQ_INDEX) {
483 if (!IS_ENABLED(CONFIG_VFIO_PCI_INTX) ||
484 vdev->nointx || vdev->pdev->is_virtfn)
487 pci_read_config_byte(vdev->pdev, PCI_INTERRUPT_PIN, &pin);
490 } else if (irq_type == VFIO_PCI_MSI_IRQ_INDEX) {
494 pos = vdev->pdev->msi_cap;
496 pci_read_config_word(vdev->pdev,
497 pos + PCI_MSI_FLAGS, &flags);
498 return 1 << ((flags & PCI_MSI_FLAGS_QMASK) >> 1);
500 } else if (irq_type == VFIO_PCI_MSIX_IRQ_INDEX) {
504 pos = vdev->pdev->msix_cap;
506 pci_read_config_word(vdev->pdev,
507 pos + PCI_MSIX_FLAGS, &flags);
509 return (flags & PCI_MSIX_FLAGS_QSIZE) + 1;
511 } else if (irq_type == VFIO_PCI_ERR_IRQ_INDEX) {
512 if (pci_is_pcie(vdev->pdev))
514 } else if (irq_type == VFIO_PCI_REQ_IRQ_INDEX) {
521 static int vfio_pci_count_devs(struct pci_dev *pdev, void *data)
527 struct vfio_pci_fill_info {
530 struct vfio_pci_dependent_device *devices;
533 static int vfio_pci_fill_devs(struct pci_dev *pdev, void *data)
535 struct vfio_pci_fill_info *fill = data;
536 struct iommu_group *iommu_group;
538 if (fill->cur == fill->max)
539 return -EAGAIN; /* Something changed, try again */
541 iommu_group = iommu_group_get(&pdev->dev);
543 return -EPERM; /* Cannot reset non-isolated devices */
545 fill->devices[fill->cur].group_id = iommu_group_id(iommu_group);
546 fill->devices[fill->cur].segment = pci_domain_nr(pdev->bus);
547 fill->devices[fill->cur].bus = pdev->bus->number;
548 fill->devices[fill->cur].devfn = pdev->devfn;
550 iommu_group_put(iommu_group);
554 struct vfio_pci_group_info {
556 struct vfio_group **groups;
559 static bool vfio_pci_dev_below_slot(struct pci_dev *pdev, struct pci_slot *slot)
561 for (; pdev; pdev = pdev->bus->self)
562 if (pdev->bus == slot->bus)
563 return (pdev->slot == slot);
567 struct vfio_pci_walk_info {
568 int (*fn)(struct pci_dev *, void *data);
570 struct pci_dev *pdev;
575 static int vfio_pci_walk_wrapper(struct pci_dev *pdev, void *data)
577 struct vfio_pci_walk_info *walk = data;
579 if (!walk->slot || vfio_pci_dev_below_slot(pdev, walk->pdev->slot))
580 walk->ret = walk->fn(pdev, walk->data);
585 static int vfio_pci_for_each_slot_or_bus(struct pci_dev *pdev,
586 int (*fn)(struct pci_dev *,
587 void *data), void *data,
590 struct vfio_pci_walk_info walk = {
591 .fn = fn, .data = data, .pdev = pdev, .slot = slot, .ret = 0,
594 pci_walk_bus(pdev->bus, vfio_pci_walk_wrapper, &walk);
599 static int msix_mmappable_cap(struct vfio_pci_core_device *vdev,
600 struct vfio_info_cap *caps)
602 struct vfio_info_cap_header header = {
603 .id = VFIO_REGION_INFO_CAP_MSIX_MAPPABLE,
607 return vfio_info_add_capability(caps, &header, sizeof(header));
610 int vfio_pci_register_dev_region(struct vfio_pci_core_device *vdev,
611 unsigned int type, unsigned int subtype,
612 const struct vfio_pci_regops *ops,
613 size_t size, u32 flags, void *data)
615 struct vfio_pci_region *region;
617 region = krealloc(vdev->region,
618 (vdev->num_regions + 1) * sizeof(*region),
623 vdev->region = region;
624 vdev->region[vdev->num_regions].type = type;
625 vdev->region[vdev->num_regions].subtype = subtype;
626 vdev->region[vdev->num_regions].ops = ops;
627 vdev->region[vdev->num_regions].size = size;
628 vdev->region[vdev->num_regions].flags = flags;
629 vdev->region[vdev->num_regions].data = data;
635 EXPORT_SYMBOL_GPL(vfio_pci_register_dev_region);
637 long vfio_pci_core_ioctl(struct vfio_device *core_vdev, unsigned int cmd,
640 struct vfio_pci_core_device *vdev =
641 container_of(core_vdev, struct vfio_pci_core_device, vdev);
644 if (cmd == VFIO_DEVICE_GET_INFO) {
645 struct vfio_device_info info;
646 struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
650 minsz = offsetofend(struct vfio_device_info, num_irqs);
652 /* For backward compatibility, cannot require this */
653 capsz = offsetofend(struct vfio_iommu_type1_info, cap_offset);
655 if (copy_from_user(&info, (void __user *)arg, minsz))
658 if (info.argsz < minsz)
661 if (info.argsz >= capsz) {
666 info.flags = VFIO_DEVICE_FLAGS_PCI;
668 if (vdev->reset_works)
669 info.flags |= VFIO_DEVICE_FLAGS_RESET;
671 info.num_regions = VFIO_PCI_NUM_REGIONS + vdev->num_regions;
672 info.num_irqs = VFIO_PCI_NUM_IRQS;
674 ret = vfio_pci_info_zdev_add_caps(vdev, &caps);
675 if (ret && ret != -ENODEV) {
676 pci_warn(vdev->pdev, "Failed to setup zPCI info capabilities\n");
681 info.flags |= VFIO_DEVICE_FLAGS_CAPS;
682 if (info.argsz < sizeof(info) + caps.size) {
683 info.argsz = sizeof(info) + caps.size;
685 vfio_info_cap_shift(&caps, sizeof(info));
686 if (copy_to_user((void __user *)arg +
687 sizeof(info), caps.buf,
692 info.cap_offset = sizeof(info);
698 return copy_to_user((void __user *)arg, &info, minsz) ?
701 } else if (cmd == VFIO_DEVICE_GET_REGION_INFO) {
702 struct pci_dev *pdev = vdev->pdev;
703 struct vfio_region_info info;
704 struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
707 minsz = offsetofend(struct vfio_region_info, offset);
709 if (copy_from_user(&info, (void __user *)arg, minsz))
712 if (info.argsz < minsz)
715 switch (info.index) {
716 case VFIO_PCI_CONFIG_REGION_INDEX:
717 info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
718 info.size = pdev->cfg_size;
719 info.flags = VFIO_REGION_INFO_FLAG_READ |
720 VFIO_REGION_INFO_FLAG_WRITE;
722 case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
723 info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
724 info.size = pci_resource_len(pdev, info.index);
730 info.flags = VFIO_REGION_INFO_FLAG_READ |
731 VFIO_REGION_INFO_FLAG_WRITE;
732 if (vdev->bar_mmap_supported[info.index]) {
733 info.flags |= VFIO_REGION_INFO_FLAG_MMAP;
734 if (info.index == vdev->msix_bar) {
735 ret = msix_mmappable_cap(vdev, &caps);
742 case VFIO_PCI_ROM_REGION_INDEX:
748 info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
751 /* Report the BAR size, not the ROM size */
752 info.size = pci_resource_len(pdev, info.index);
754 /* Shadow ROMs appear as PCI option ROMs */
755 if (pdev->resource[PCI_ROM_RESOURCE].flags &
756 IORESOURCE_ROM_SHADOW)
763 * Is it really there? Enable memory decode for
764 * implicit access in pci_map_rom().
766 cmd = vfio_pci_memory_lock_and_enable(vdev);
767 io = pci_map_rom(pdev, &size);
769 info.flags = VFIO_REGION_INFO_FLAG_READ;
770 pci_unmap_rom(pdev, io);
774 vfio_pci_memory_unlock_and_restore(vdev, cmd);
778 case VFIO_PCI_VGA_REGION_INDEX:
782 info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
784 info.flags = VFIO_REGION_INFO_FLAG_READ |
785 VFIO_REGION_INFO_FLAG_WRITE;
790 struct vfio_region_info_cap_type cap_type = {
791 .header.id = VFIO_REGION_INFO_CAP_TYPE,
792 .header.version = 1 };
795 VFIO_PCI_NUM_REGIONS + vdev->num_regions)
797 info.index = array_index_nospec(info.index,
798 VFIO_PCI_NUM_REGIONS +
801 i = info.index - VFIO_PCI_NUM_REGIONS;
803 info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
804 info.size = vdev->region[i].size;
805 info.flags = vdev->region[i].flags;
807 cap_type.type = vdev->region[i].type;
808 cap_type.subtype = vdev->region[i].subtype;
810 ret = vfio_info_add_capability(&caps, &cap_type.header,
815 if (vdev->region[i].ops->add_capability) {
816 ret = vdev->region[i].ops->add_capability(vdev,
817 &vdev->region[i], &caps);
825 info.flags |= VFIO_REGION_INFO_FLAG_CAPS;
826 if (info.argsz < sizeof(info) + caps.size) {
827 info.argsz = sizeof(info) + caps.size;
830 vfio_info_cap_shift(&caps, sizeof(info));
831 if (copy_to_user((void __user *)arg +
832 sizeof(info), caps.buf,
837 info.cap_offset = sizeof(info);
843 return copy_to_user((void __user *)arg, &info, minsz) ?
846 } else if (cmd == VFIO_DEVICE_GET_IRQ_INFO) {
847 struct vfio_irq_info info;
849 minsz = offsetofend(struct vfio_irq_info, count);
851 if (copy_from_user(&info, (void __user *)arg, minsz))
854 if (info.argsz < minsz || info.index >= VFIO_PCI_NUM_IRQS)
857 switch (info.index) {
858 case VFIO_PCI_INTX_IRQ_INDEX ... VFIO_PCI_MSIX_IRQ_INDEX:
859 case VFIO_PCI_REQ_IRQ_INDEX:
861 case VFIO_PCI_ERR_IRQ_INDEX:
862 if (pci_is_pcie(vdev->pdev))
869 info.flags = VFIO_IRQ_INFO_EVENTFD;
871 info.count = vfio_pci_get_irq_count(vdev, info.index);
873 if (info.index == VFIO_PCI_INTX_IRQ_INDEX)
874 info.flags |= (VFIO_IRQ_INFO_MASKABLE |
875 VFIO_IRQ_INFO_AUTOMASKED);
877 info.flags |= VFIO_IRQ_INFO_NORESIZE;
879 return copy_to_user((void __user *)arg, &info, minsz) ?
882 } else if (cmd == VFIO_DEVICE_SET_IRQS) {
883 struct vfio_irq_set hdr;
886 size_t data_size = 0;
888 minsz = offsetofend(struct vfio_irq_set, count);
890 if (copy_from_user(&hdr, (void __user *)arg, minsz))
893 max = vfio_pci_get_irq_count(vdev, hdr.index);
895 ret = vfio_set_irqs_validate_and_prepare(&hdr, max,
896 VFIO_PCI_NUM_IRQS, &data_size);
901 data = memdup_user((void __user *)(arg + minsz),
904 return PTR_ERR(data);
907 mutex_lock(&vdev->igate);
909 ret = vfio_pci_set_irqs_ioctl(vdev, hdr.flags, hdr.index,
910 hdr.start, hdr.count, data);
912 mutex_unlock(&vdev->igate);
917 } else if (cmd == VFIO_DEVICE_RESET) {
920 if (!vdev->reset_works)
923 vfio_pci_zap_and_down_write_memory_lock(vdev);
924 ret = pci_try_reset_function(vdev->pdev);
925 up_write(&vdev->memory_lock);
929 } else if (cmd == VFIO_DEVICE_GET_PCI_HOT_RESET_INFO) {
930 struct vfio_pci_hot_reset_info hdr;
931 struct vfio_pci_fill_info fill = { 0 };
932 struct vfio_pci_dependent_device *devices = NULL;
936 minsz = offsetofend(struct vfio_pci_hot_reset_info, count);
938 if (copy_from_user(&hdr, (void __user *)arg, minsz))
941 if (hdr.argsz < minsz)
946 /* Can we do a slot or bus reset or neither? */
947 if (!pci_probe_reset_slot(vdev->pdev->slot))
949 else if (pci_probe_reset_bus(vdev->pdev->bus))
952 /* How many devices are affected? */
953 ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
959 WARN_ON(!fill.max); /* Should always be at least one */
962 * If there's enough space, fill it now, otherwise return
963 * -ENOSPC and the number of devices affected.
965 if (hdr.argsz < sizeof(hdr) + (fill.max * sizeof(*devices))) {
967 hdr.count = fill.max;
968 goto reset_info_exit;
971 devices = kcalloc(fill.max, sizeof(*devices), GFP_KERNEL);
975 fill.devices = devices;
977 ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
982 * If a device was removed between counting and filling,
983 * we may come up short of fill.max. If a device was
984 * added, we'll have a return of -EAGAIN above.
987 hdr.count = fill.cur;
990 if (copy_to_user((void __user *)arg, &hdr, minsz))
994 if (copy_to_user((void __user *)(arg + minsz), devices,
995 hdr.count * sizeof(*devices)))
1002 } else if (cmd == VFIO_DEVICE_PCI_HOT_RESET) {
1003 struct vfio_pci_hot_reset hdr;
1005 struct vfio_group **groups;
1006 struct vfio_pci_group_info info;
1008 int group_idx, count = 0, ret = 0;
1010 minsz = offsetofend(struct vfio_pci_hot_reset, count);
1012 if (copy_from_user(&hdr, (void __user *)arg, minsz))
1015 if (hdr.argsz < minsz || hdr.flags)
1018 /* Can we do a slot or bus reset or neither? */
1019 if (!pci_probe_reset_slot(vdev->pdev->slot))
1021 else if (pci_probe_reset_bus(vdev->pdev->bus))
1025 * We can't let userspace give us an arbitrarily large
1026 * buffer to copy, so verify how many we think there
1027 * could be. Note groups can have multiple devices so
1028 * one group per device is the max.
1030 ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1031 vfio_pci_count_devs,
1036 /* Somewhere between 1 and count is OK */
1037 if (!hdr.count || hdr.count > count)
1040 group_fds = kcalloc(hdr.count, sizeof(*group_fds), GFP_KERNEL);
1041 groups = kcalloc(hdr.count, sizeof(*groups), GFP_KERNEL);
1042 if (!group_fds || !groups) {
1048 if (copy_from_user(group_fds, (void __user *)(arg + minsz),
1049 hdr.count * sizeof(*group_fds))) {
1056 * For each group_fd, get the group through the vfio external
1057 * user interface and store the group and iommu ID. This
1058 * ensures the group is held across the reset.
1060 for (group_idx = 0; group_idx < hdr.count; group_idx++) {
1061 struct vfio_group *group;
1062 struct fd f = fdget(group_fds[group_idx]);
1068 group = vfio_group_get_external_user(f.file);
1070 if (IS_ERR(group)) {
1071 ret = PTR_ERR(group);
1075 groups[group_idx] = group;
1080 /* release reference to groups on error */
1082 goto hot_reset_release;
1084 info.count = hdr.count;
1085 info.groups = groups;
1087 ret = vfio_pci_dev_set_hot_reset(vdev->vdev.dev_set, &info);
1090 for (group_idx--; group_idx >= 0; group_idx--)
1091 vfio_group_put_external_user(groups[group_idx]);
1095 } else if (cmd == VFIO_DEVICE_IOEVENTFD) {
1096 struct vfio_device_ioeventfd ioeventfd;
1099 minsz = offsetofend(struct vfio_device_ioeventfd, fd);
1101 if (copy_from_user(&ioeventfd, (void __user *)arg, minsz))
1104 if (ioeventfd.argsz < minsz)
1107 if (ioeventfd.flags & ~VFIO_DEVICE_IOEVENTFD_SIZE_MASK)
1110 count = ioeventfd.flags & VFIO_DEVICE_IOEVENTFD_SIZE_MASK;
1112 if (hweight8(count) != 1 || ioeventfd.fd < -1)
1115 return vfio_pci_ioeventfd(vdev, ioeventfd.offset,
1116 ioeventfd.data, count, ioeventfd.fd);
1117 } else if (cmd == VFIO_DEVICE_FEATURE) {
1118 struct vfio_device_feature feature;
1121 minsz = offsetofend(struct vfio_device_feature, flags);
1123 if (copy_from_user(&feature, (void __user *)arg, minsz))
1126 if (feature.argsz < minsz)
1129 /* Check unknown flags */
1130 if (feature.flags & ~(VFIO_DEVICE_FEATURE_MASK |
1131 VFIO_DEVICE_FEATURE_SET |
1132 VFIO_DEVICE_FEATURE_GET |
1133 VFIO_DEVICE_FEATURE_PROBE))
1136 /* GET & SET are mutually exclusive except with PROBE */
1137 if (!(feature.flags & VFIO_DEVICE_FEATURE_PROBE) &&
1138 (feature.flags & VFIO_DEVICE_FEATURE_SET) &&
1139 (feature.flags & VFIO_DEVICE_FEATURE_GET))
1142 switch (feature.flags & VFIO_DEVICE_FEATURE_MASK) {
1143 case VFIO_DEVICE_FEATURE_PCI_VF_TOKEN:
1144 if (!vdev->vf_token)
1148 * We do not support GET of the VF Token UUID as this
1149 * could expose the token of the previous device user.
1151 if (feature.flags & VFIO_DEVICE_FEATURE_GET)
1154 if (feature.flags & VFIO_DEVICE_FEATURE_PROBE)
1157 /* Don't SET unless told to do so */
1158 if (!(feature.flags & VFIO_DEVICE_FEATURE_SET))
1161 if (feature.argsz < minsz + sizeof(uuid))
1164 if (copy_from_user(&uuid, (void __user *)(arg + minsz),
1168 mutex_lock(&vdev->vf_token->lock);
1169 uuid_copy(&vdev->vf_token->uuid, &uuid);
1170 mutex_unlock(&vdev->vf_token->lock);
1180 EXPORT_SYMBOL_GPL(vfio_pci_core_ioctl);
1182 static ssize_t vfio_pci_rw(struct vfio_pci_core_device *vdev, char __user *buf,
1183 size_t count, loff_t *ppos, bool iswrite)
1185 unsigned int index = VFIO_PCI_OFFSET_TO_INDEX(*ppos);
1187 if (index >= VFIO_PCI_NUM_REGIONS + vdev->num_regions)
1191 case VFIO_PCI_CONFIG_REGION_INDEX:
1192 return vfio_pci_config_rw(vdev, buf, count, ppos, iswrite);
1194 case VFIO_PCI_ROM_REGION_INDEX:
1197 return vfio_pci_bar_rw(vdev, buf, count, ppos, false);
1199 case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
1200 return vfio_pci_bar_rw(vdev, buf, count, ppos, iswrite);
1202 case VFIO_PCI_VGA_REGION_INDEX:
1203 return vfio_pci_vga_rw(vdev, buf, count, ppos, iswrite);
1205 index -= VFIO_PCI_NUM_REGIONS;
1206 return vdev->region[index].ops->rw(vdev, buf,
1207 count, ppos, iswrite);
1213 ssize_t vfio_pci_core_read(struct vfio_device *core_vdev, char __user *buf,
1214 size_t count, loff_t *ppos)
1216 struct vfio_pci_core_device *vdev =
1217 container_of(core_vdev, struct vfio_pci_core_device, vdev);
1222 return vfio_pci_rw(vdev, buf, count, ppos, false);
1224 EXPORT_SYMBOL_GPL(vfio_pci_core_read);
1226 ssize_t vfio_pci_core_write(struct vfio_device *core_vdev, const char __user *buf,
1227 size_t count, loff_t *ppos)
1229 struct vfio_pci_core_device *vdev =
1230 container_of(core_vdev, struct vfio_pci_core_device, vdev);
1235 return vfio_pci_rw(vdev, (char __user *)buf, count, ppos, true);
1237 EXPORT_SYMBOL_GPL(vfio_pci_core_write);
1239 /* Return 1 on zap and vma_lock acquired, 0 on contention (only with @try) */
1240 static int vfio_pci_zap_and_vma_lock(struct vfio_pci_core_device *vdev, bool try)
1242 struct vfio_pci_mmap_vma *mmap_vma, *tmp;
1246 * vma_lock is nested under mmap_lock for vm_ops callback paths.
1247 * The memory_lock semaphore is used by both code paths calling
1248 * into this function to zap vmas and the vm_ops.fault callback
1249 * to protect the memory enable state of the device.
1251 * When zapping vmas we need to maintain the mmap_lock => vma_lock
1252 * ordering, which requires using vma_lock to walk vma_list to
1253 * acquire an mm, then dropping vma_lock to get the mmap_lock and
1254 * reacquiring vma_lock. This logic is derived from similar
1255 * requirements in uverbs_user_mmap_disassociate().
1257 * mmap_lock must always be the top-level lock when it is taken.
1258 * Therefore we can only hold the memory_lock write lock when
1259 * vma_list is empty, as we'd need to take mmap_lock to clear
1260 * entries. vma_list can only be guaranteed empty when holding
1261 * vma_lock, thus memory_lock is nested under vma_lock.
1263 * This enables the vm_ops.fault callback to acquire vma_lock,
1264 * followed by memory_lock read lock, while already holding
1265 * mmap_lock without risk of deadlock.
1268 struct mm_struct *mm = NULL;
1271 if (!mutex_trylock(&vdev->vma_lock))
1274 mutex_lock(&vdev->vma_lock);
1276 while (!list_empty(&vdev->vma_list)) {
1277 mmap_vma = list_first_entry(&vdev->vma_list,
1278 struct vfio_pci_mmap_vma,
1280 mm = mmap_vma->vma->vm_mm;
1281 if (mmget_not_zero(mm))
1284 list_del(&mmap_vma->vma_next);
1290 mutex_unlock(&vdev->vma_lock);
1293 if (!mmap_read_trylock(mm)) {
1301 if (!mutex_trylock(&vdev->vma_lock)) {
1302 mmap_read_unlock(mm);
1307 mutex_lock(&vdev->vma_lock);
1309 list_for_each_entry_safe(mmap_vma, tmp,
1310 &vdev->vma_list, vma_next) {
1311 struct vm_area_struct *vma = mmap_vma->vma;
1313 if (vma->vm_mm != mm)
1316 list_del(&mmap_vma->vma_next);
1319 zap_vma_ptes(vma, vma->vm_start,
1320 vma->vm_end - vma->vm_start);
1322 mutex_unlock(&vdev->vma_lock);
1323 mmap_read_unlock(mm);
1328 void vfio_pci_zap_and_down_write_memory_lock(struct vfio_pci_core_device *vdev)
1330 vfio_pci_zap_and_vma_lock(vdev, false);
1331 down_write(&vdev->memory_lock);
1332 mutex_unlock(&vdev->vma_lock);
1335 u16 vfio_pci_memory_lock_and_enable(struct vfio_pci_core_device *vdev)
1339 down_write(&vdev->memory_lock);
1340 pci_read_config_word(vdev->pdev, PCI_COMMAND, &cmd);
1341 if (!(cmd & PCI_COMMAND_MEMORY))
1342 pci_write_config_word(vdev->pdev, PCI_COMMAND,
1343 cmd | PCI_COMMAND_MEMORY);
1348 void vfio_pci_memory_unlock_and_restore(struct vfio_pci_core_device *vdev, u16 cmd)
1350 pci_write_config_word(vdev->pdev, PCI_COMMAND, cmd);
1351 up_write(&vdev->memory_lock);
1354 /* Caller holds vma_lock */
1355 static int __vfio_pci_add_vma(struct vfio_pci_core_device *vdev,
1356 struct vm_area_struct *vma)
1358 struct vfio_pci_mmap_vma *mmap_vma;
1360 mmap_vma = kmalloc(sizeof(*mmap_vma), GFP_KERNEL);
1364 mmap_vma->vma = vma;
1365 list_add(&mmap_vma->vma_next, &vdev->vma_list);
1371 * Zap mmaps on open so that we can fault them in on access and therefore
1372 * our vma_list only tracks mappings accessed since last zap.
1374 static void vfio_pci_mmap_open(struct vm_area_struct *vma)
1376 zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
1379 static void vfio_pci_mmap_close(struct vm_area_struct *vma)
1381 struct vfio_pci_core_device *vdev = vma->vm_private_data;
1382 struct vfio_pci_mmap_vma *mmap_vma;
1384 mutex_lock(&vdev->vma_lock);
1385 list_for_each_entry(mmap_vma, &vdev->vma_list, vma_next) {
1386 if (mmap_vma->vma == vma) {
1387 list_del(&mmap_vma->vma_next);
1392 mutex_unlock(&vdev->vma_lock);
1395 static vm_fault_t vfio_pci_mmap_fault(struct vm_fault *vmf)
1397 struct vm_area_struct *vma = vmf->vma;
1398 struct vfio_pci_core_device *vdev = vma->vm_private_data;
1399 struct vfio_pci_mmap_vma *mmap_vma;
1400 vm_fault_t ret = VM_FAULT_NOPAGE;
1402 mutex_lock(&vdev->vma_lock);
1403 down_read(&vdev->memory_lock);
1405 if (!__vfio_pci_memory_enabled(vdev)) {
1406 ret = VM_FAULT_SIGBUS;
1411 * We populate the whole vma on fault, so we need to test whether
1412 * the vma has already been mapped, such as for concurrent faults
1413 * to the same vma. io_remap_pfn_range() will trigger a BUG_ON if
1414 * we ask it to fill the same range again.
1416 list_for_each_entry(mmap_vma, &vdev->vma_list, vma_next) {
1417 if (mmap_vma->vma == vma)
1421 if (io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
1422 vma->vm_end - vma->vm_start,
1423 vma->vm_page_prot)) {
1424 ret = VM_FAULT_SIGBUS;
1425 zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
1429 if (__vfio_pci_add_vma(vdev, vma)) {
1431 zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
1435 up_read(&vdev->memory_lock);
1436 mutex_unlock(&vdev->vma_lock);
1440 static const struct vm_operations_struct vfio_pci_mmap_ops = {
1441 .open = vfio_pci_mmap_open,
1442 .close = vfio_pci_mmap_close,
1443 .fault = vfio_pci_mmap_fault,
1446 int vfio_pci_core_mmap(struct vfio_device *core_vdev, struct vm_area_struct *vma)
1448 struct vfio_pci_core_device *vdev =
1449 container_of(core_vdev, struct vfio_pci_core_device, vdev);
1450 struct pci_dev *pdev = vdev->pdev;
1452 u64 phys_len, req_len, pgoff, req_start;
1455 index = vma->vm_pgoff >> (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT);
1457 if (index >= VFIO_PCI_NUM_REGIONS + vdev->num_regions)
1459 if (vma->vm_end < vma->vm_start)
1461 if ((vma->vm_flags & VM_SHARED) == 0)
1463 if (index >= VFIO_PCI_NUM_REGIONS) {
1464 int regnum = index - VFIO_PCI_NUM_REGIONS;
1465 struct vfio_pci_region *region = vdev->region + regnum;
1467 if (region->ops && region->ops->mmap &&
1468 (region->flags & VFIO_REGION_INFO_FLAG_MMAP))
1469 return region->ops->mmap(vdev, region, vma);
1472 if (index >= VFIO_PCI_ROM_REGION_INDEX)
1474 if (!vdev->bar_mmap_supported[index])
1477 phys_len = PAGE_ALIGN(pci_resource_len(pdev, index));
1478 req_len = vma->vm_end - vma->vm_start;
1479 pgoff = vma->vm_pgoff &
1480 ((1U << (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT)) - 1);
1481 req_start = pgoff << PAGE_SHIFT;
1483 if (req_start + req_len > phys_len)
1487 * Even though we don't make use of the barmap for the mmap,
1488 * we need to request the region and the barmap tracks that.
1490 if (!vdev->barmap[index]) {
1491 ret = pci_request_selected_regions(pdev,
1492 1 << index, "vfio-pci");
1496 vdev->barmap[index] = pci_iomap(pdev, index, 0);
1497 if (!vdev->barmap[index]) {
1498 pci_release_selected_regions(pdev, 1 << index);
1503 vma->vm_private_data = vdev;
1504 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1505 vma->vm_pgoff = (pci_resource_start(pdev, index) >> PAGE_SHIFT) + pgoff;
1508 * See remap_pfn_range(), called from vfio_pci_fault() but we can't
1509 * change vm_flags within the fault handler. Set them now.
1511 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1512 vma->vm_ops = &vfio_pci_mmap_ops;
1516 EXPORT_SYMBOL_GPL(vfio_pci_core_mmap);
1518 void vfio_pci_core_request(struct vfio_device *core_vdev, unsigned int count)
1520 struct vfio_pci_core_device *vdev =
1521 container_of(core_vdev, struct vfio_pci_core_device, vdev);
1522 struct pci_dev *pdev = vdev->pdev;
1524 mutex_lock(&vdev->igate);
1526 if (vdev->req_trigger) {
1528 pci_notice_ratelimited(pdev,
1529 "Relaying device request to user (#%u)\n",
1531 eventfd_signal(vdev->req_trigger, 1);
1532 } else if (count == 0) {
1534 "No device request channel registered, blocked until released by user\n");
1537 mutex_unlock(&vdev->igate);
1539 EXPORT_SYMBOL_GPL(vfio_pci_core_request);
1541 static int vfio_pci_validate_vf_token(struct vfio_pci_core_device *vdev,
1542 bool vf_token, uuid_t *uuid)
1545 * There's always some degree of trust or collaboration between SR-IOV
1546 * PF and VFs, even if just that the PF hosts the SR-IOV capability and
1547 * can disrupt VFs with a reset, but often the PF has more explicit
1548 * access to deny service to the VF or access data passed through the
1549 * VF. We therefore require an opt-in via a shared VF token (UUID) to
1550 * represent this trust. This both prevents that a VF driver might
1551 * assume the PF driver is a trusted, in-kernel driver, and also that
1552 * a PF driver might be replaced with a rogue driver, unknown to in-use
1555 * Therefore when presented with a VF, if the PF is a vfio device and
1556 * it is bound to the vfio-pci driver, the user needs to provide a VF
1557 * token to access the device, in the form of appending a vf_token to
1558 * the device name, for example:
1560 * "0000:04:10.0 vf_token=bd8d9d2b-5a5f-4f5a-a211-f591514ba1f3"
1562 * When presented with a PF which has VFs in use, the user must also
1563 * provide the current VF token to prove collaboration with existing
1564 * VF users. If VFs are not in use, the VF token provided for the PF
1565 * device will act to set the VF token.
1567 * If the VF token is provided but unused, an error is generated.
1569 if (!vdev->pdev->is_virtfn && !vdev->vf_token && !vf_token)
1570 return 0; /* No VF token provided or required */
1572 if (vdev->pdev->is_virtfn) {
1573 struct vfio_pci_core_device *pf_vdev = get_pf_vdev(vdev);
1578 return 0; /* PF is not vfio-pci, no VF token */
1580 pci_info_ratelimited(vdev->pdev,
1581 "VF token incorrectly provided, PF not bound to vfio-pci\n");
1586 vfio_device_put(&pf_vdev->vdev);
1587 pci_info_ratelimited(vdev->pdev,
1588 "VF token required to access device\n");
1592 mutex_lock(&pf_vdev->vf_token->lock);
1593 match = uuid_equal(uuid, &pf_vdev->vf_token->uuid);
1594 mutex_unlock(&pf_vdev->vf_token->lock);
1596 vfio_device_put(&pf_vdev->vdev);
1599 pci_info_ratelimited(vdev->pdev,
1600 "Incorrect VF token provided for device\n");
1603 } else if (vdev->vf_token) {
1604 mutex_lock(&vdev->vf_token->lock);
1605 if (vdev->vf_token->users) {
1607 mutex_unlock(&vdev->vf_token->lock);
1608 pci_info_ratelimited(vdev->pdev,
1609 "VF token required to access device\n");
1613 if (!uuid_equal(uuid, &vdev->vf_token->uuid)) {
1614 mutex_unlock(&vdev->vf_token->lock);
1615 pci_info_ratelimited(vdev->pdev,
1616 "Incorrect VF token provided for device\n");
1619 } else if (vf_token) {
1620 uuid_copy(&vdev->vf_token->uuid, uuid);
1623 mutex_unlock(&vdev->vf_token->lock);
1624 } else if (vf_token) {
1625 pci_info_ratelimited(vdev->pdev,
1626 "VF token incorrectly provided, not a PF or VF\n");
1633 #define VF_TOKEN_ARG "vf_token="
1635 int vfio_pci_core_match(struct vfio_device *core_vdev, char *buf)
1637 struct vfio_pci_core_device *vdev =
1638 container_of(core_vdev, struct vfio_pci_core_device, vdev);
1639 bool vf_token = false;
1643 if (strncmp(pci_name(vdev->pdev), buf, strlen(pci_name(vdev->pdev))))
1644 return 0; /* No match */
1646 if (strlen(buf) > strlen(pci_name(vdev->pdev))) {
1647 buf += strlen(pci_name(vdev->pdev));
1650 return 0; /* No match: non-whitespace after name */
1658 if (!vf_token && !strncmp(buf, VF_TOKEN_ARG,
1659 strlen(VF_TOKEN_ARG))) {
1660 buf += strlen(VF_TOKEN_ARG);
1662 if (strlen(buf) < UUID_STRING_LEN)
1665 ret = uuid_parse(buf, &uuid);
1670 buf += UUID_STRING_LEN;
1672 /* Unknown/duplicate option */
1678 ret = vfio_pci_validate_vf_token(vdev, vf_token, &uuid);
1682 return 1; /* Match */
1684 EXPORT_SYMBOL_GPL(vfio_pci_core_match);
1686 static int vfio_pci_bus_notifier(struct notifier_block *nb,
1687 unsigned long action, void *data)
1689 struct vfio_pci_core_device *vdev = container_of(nb,
1690 struct vfio_pci_core_device, nb);
1691 struct device *dev = data;
1692 struct pci_dev *pdev = to_pci_dev(dev);
1693 struct pci_dev *physfn = pci_physfn(pdev);
1695 if (action == BUS_NOTIFY_ADD_DEVICE &&
1696 pdev->is_virtfn && physfn == vdev->pdev) {
1697 pci_info(vdev->pdev, "Captured SR-IOV VF %s driver_override\n",
1699 pdev->driver_override = kasprintf(GFP_KERNEL, "%s",
1700 vdev->vdev.ops->name);
1701 } else if (action == BUS_NOTIFY_BOUND_DRIVER &&
1702 pdev->is_virtfn && physfn == vdev->pdev) {
1703 struct pci_driver *drv = pci_dev_driver(pdev);
1705 if (drv && drv != pci_dev_driver(vdev->pdev))
1706 pci_warn(vdev->pdev,
1707 "VF %s bound to driver %s while PF bound to driver %s\n",
1708 pci_name(pdev), drv->name,
1709 pci_dev_driver(vdev->pdev)->name);
1715 static int vfio_pci_vf_init(struct vfio_pci_core_device *vdev)
1717 struct pci_dev *pdev = vdev->pdev;
1720 if (!pdev->is_physfn)
1723 vdev->vf_token = kzalloc(sizeof(*vdev->vf_token), GFP_KERNEL);
1724 if (!vdev->vf_token)
1727 mutex_init(&vdev->vf_token->lock);
1728 uuid_gen(&vdev->vf_token->uuid);
1730 vdev->nb.notifier_call = vfio_pci_bus_notifier;
1731 ret = bus_register_notifier(&pci_bus_type, &vdev->nb);
1733 kfree(vdev->vf_token);
1739 static void vfio_pci_vf_uninit(struct vfio_pci_core_device *vdev)
1741 if (!vdev->vf_token)
1744 bus_unregister_notifier(&pci_bus_type, &vdev->nb);
1745 WARN_ON(vdev->vf_token->users);
1746 mutex_destroy(&vdev->vf_token->lock);
1747 kfree(vdev->vf_token);
1750 static int vfio_pci_vga_init(struct vfio_pci_core_device *vdev)
1752 struct pci_dev *pdev = vdev->pdev;
1755 if (!vfio_pci_is_vga(pdev))
1758 ret = vga_client_register(pdev, vfio_pci_set_decode);
1761 vga_set_legacy_decoding(pdev, vfio_pci_set_decode(pdev, false));
1765 static void vfio_pci_vga_uninit(struct vfio_pci_core_device *vdev)
1767 struct pci_dev *pdev = vdev->pdev;
1769 if (!vfio_pci_is_vga(pdev))
1771 vga_client_unregister(pdev);
1772 vga_set_legacy_decoding(pdev, VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM |
1773 VGA_RSRC_LEGACY_IO |
1774 VGA_RSRC_LEGACY_MEM);
1777 void vfio_pci_core_init_device(struct vfio_pci_core_device *vdev,
1778 struct pci_dev *pdev,
1779 const struct vfio_device_ops *vfio_pci_ops)
1781 vfio_init_group_dev(&vdev->vdev, &pdev->dev, vfio_pci_ops);
1783 vdev->irq_type = VFIO_PCI_NUM_IRQS;
1784 mutex_init(&vdev->igate);
1785 spin_lock_init(&vdev->irqlock);
1786 mutex_init(&vdev->ioeventfds_lock);
1787 INIT_LIST_HEAD(&vdev->dummy_resources_list);
1788 INIT_LIST_HEAD(&vdev->ioeventfds_list);
1789 mutex_init(&vdev->vma_lock);
1790 INIT_LIST_HEAD(&vdev->vma_list);
1791 init_rwsem(&vdev->memory_lock);
1793 EXPORT_SYMBOL_GPL(vfio_pci_core_init_device);
1795 void vfio_pci_core_uninit_device(struct vfio_pci_core_device *vdev)
1797 mutex_destroy(&vdev->igate);
1798 mutex_destroy(&vdev->ioeventfds_lock);
1799 mutex_destroy(&vdev->vma_lock);
1800 vfio_uninit_group_dev(&vdev->vdev);
1801 kfree(vdev->region);
1802 kfree(vdev->pm_save);
1804 EXPORT_SYMBOL_GPL(vfio_pci_core_uninit_device);
1806 int vfio_pci_core_register_device(struct vfio_pci_core_device *vdev)
1808 struct pci_dev *pdev = vdev->pdev;
1809 struct iommu_group *group;
1812 if (pdev->hdr_type != PCI_HEADER_TYPE_NORMAL)
1816 * Prevent binding to PFs with VFs enabled, the VFs might be in use
1817 * by the host or other users. We cannot capture the VFs if they
1818 * already exist, nor can we track VF users. Disabling SR-IOV here
1819 * would initiate removing the VFs, which would unbind the driver,
1820 * which is prone to blocking if that VF is also in use by vfio-pci.
1821 * Just reject these PFs and let the user sort it out.
1823 if (pci_num_vf(pdev)) {
1824 pci_warn(pdev, "Cannot bind to PF with SR-IOV enabled\n");
1828 group = vfio_iommu_group_get(&pdev->dev);
1832 if (pci_is_root_bus(pdev->bus)) {
1833 ret = vfio_assign_device_set(&vdev->vdev, vdev);
1834 } else if (!pci_probe_reset_slot(pdev->slot)) {
1835 ret = vfio_assign_device_set(&vdev->vdev, pdev->slot);
1838 * If there is no slot reset support for this device, the whole
1839 * bus needs to be grouped together to support bus-wide resets.
1841 ret = vfio_assign_device_set(&vdev->vdev, pdev->bus);
1846 ret = vfio_pci_vf_init(vdev);
1849 ret = vfio_pci_vga_init(vdev);
1853 vfio_pci_probe_power_state(vdev);
1855 if (!disable_idle_d3) {
1857 * pci-core sets the device power state to an unknown value at
1858 * bootup and after being removed from a driver. The only
1859 * transition it allows from this unknown state is to D0, which
1860 * typically happens when a driver calls pci_enable_device().
1861 * We're not ready to enable the device yet, but we do want to
1862 * be able to get to D3. Therefore first do a D0 transition
1863 * before going to D3.
1865 vfio_pci_set_power_state(vdev, PCI_D0);
1866 vfio_pci_set_power_state(vdev, PCI_D3hot);
1869 ret = vfio_register_group_dev(&vdev->vdev);
1875 if (!disable_idle_d3)
1876 vfio_pci_set_power_state(vdev, PCI_D0);
1878 vfio_pci_vf_uninit(vdev);
1880 vfio_iommu_group_put(group, &pdev->dev);
1883 EXPORT_SYMBOL_GPL(vfio_pci_core_register_device);
1885 void vfio_pci_core_unregister_device(struct vfio_pci_core_device *vdev)
1887 struct pci_dev *pdev = vdev->pdev;
1889 pci_disable_sriov(pdev);
1891 vfio_unregister_group_dev(&vdev->vdev);
1893 vfio_pci_vf_uninit(vdev);
1894 vfio_pci_vga_uninit(vdev);
1896 vfio_iommu_group_put(pdev->dev.iommu_group, &pdev->dev);
1898 if (!disable_idle_d3)
1899 vfio_pci_set_power_state(vdev, PCI_D0);
1901 EXPORT_SYMBOL_GPL(vfio_pci_core_unregister_device);
1903 static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
1904 pci_channel_state_t state)
1906 struct vfio_pci_core_device *vdev;
1907 struct vfio_device *device;
1909 device = vfio_device_get_from_dev(&pdev->dev);
1911 return PCI_ERS_RESULT_DISCONNECT;
1913 vdev = container_of(device, struct vfio_pci_core_device, vdev);
1915 mutex_lock(&vdev->igate);
1917 if (vdev->err_trigger)
1918 eventfd_signal(vdev->err_trigger, 1);
1920 mutex_unlock(&vdev->igate);
1922 vfio_device_put(device);
1924 return PCI_ERS_RESULT_CAN_RECOVER;
1927 int vfio_pci_core_sriov_configure(struct pci_dev *pdev, int nr_virtfn)
1929 struct vfio_device *device;
1932 device = vfio_device_get_from_dev(&pdev->dev);
1937 pci_disable_sriov(pdev);
1939 ret = pci_enable_sriov(pdev, nr_virtfn);
1941 vfio_device_put(device);
1943 return ret < 0 ? ret : nr_virtfn;
1945 EXPORT_SYMBOL_GPL(vfio_pci_core_sriov_configure);
1947 const struct pci_error_handlers vfio_pci_core_err_handlers = {
1948 .error_detected = vfio_pci_aer_err_detected,
1950 EXPORT_SYMBOL_GPL(vfio_pci_core_err_handlers);
1952 static bool vfio_dev_in_groups(struct vfio_pci_core_device *vdev,
1953 struct vfio_pci_group_info *groups)
1957 for (i = 0; i < groups->count; i++)
1958 if (groups->groups[i] == vdev->vdev.group)
1963 static int vfio_pci_is_device_in_set(struct pci_dev *pdev, void *data)
1965 struct vfio_device_set *dev_set = data;
1966 struct vfio_device *cur;
1968 list_for_each_entry(cur, &dev_set->device_list, dev_set_list)
1969 if (cur->dev == &pdev->dev)
1975 * vfio-core considers a group to be viable and will create a vfio_device even
1976 * if some devices are bound to drivers like pci-stub or pcieport. Here we
1977 * require all PCI devices to be inside our dev_set since that ensures they stay
1978 * put and that every driver controlling the device can co-ordinate with the
1981 * Returns the pci_dev to pass to pci_reset_bus() if every PCI device to be
1982 * reset is inside the dev_set, and pci_reset_bus() can succeed. NULL otherwise.
1984 static struct pci_dev *
1985 vfio_pci_dev_set_resettable(struct vfio_device_set *dev_set)
1987 struct pci_dev *pdev;
1989 lockdep_assert_held(&dev_set->lock);
1992 * By definition all PCI devices in the dev_set share the same PCI
1993 * reset, so any pci_dev will have the same outcomes for
1994 * pci_probe_reset_*() and pci_reset_bus().
1996 pdev = list_first_entry(&dev_set->device_list,
1997 struct vfio_pci_core_device,
1998 vdev.dev_set_list)->pdev;
2000 /* pci_reset_bus() is supported */
2001 if (pci_probe_reset_slot(pdev->slot) && pci_probe_reset_bus(pdev->bus))
2004 if (vfio_pci_for_each_slot_or_bus(pdev, vfio_pci_is_device_in_set,
2006 !pci_probe_reset_slot(pdev->slot)))
2012 * We need to get memory_lock for each device, but devices can share mmap_lock,
2013 * therefore we need to zap and hold the vma_lock for each device, and only then
2014 * get each memory_lock.
2016 static int vfio_pci_dev_set_hot_reset(struct vfio_device_set *dev_set,
2017 struct vfio_pci_group_info *groups)
2019 struct vfio_pci_core_device *cur_mem;
2020 struct vfio_pci_core_device *cur_vma;
2021 struct vfio_pci_core_device *cur;
2022 struct pci_dev *pdev;
2026 mutex_lock(&dev_set->lock);
2027 cur_mem = list_first_entry(&dev_set->device_list,
2028 struct vfio_pci_core_device,
2031 pdev = vfio_pci_dev_set_resettable(dev_set);
2037 list_for_each_entry(cur_vma, &dev_set->device_list, vdev.dev_set_list) {
2039 * Test whether all the affected devices are contained by the
2040 * set of groups provided by the user.
2042 if (!vfio_dev_in_groups(cur_vma, groups)) {
2048 * Locking multiple devices is prone to deadlock, runaway and
2049 * unwind if we hit contention.
2051 if (!vfio_pci_zap_and_vma_lock(cur_vma, true)) {
2058 list_for_each_entry(cur_mem, &dev_set->device_list, vdev.dev_set_list) {
2059 if (!down_write_trylock(&cur_mem->memory_lock)) {
2063 mutex_unlock(&cur_mem->vma_lock);
2067 ret = pci_reset_bus(pdev);
2070 list_for_each_entry(cur, &dev_set->device_list, vdev.dev_set_list) {
2076 up_write(&cur->memory_lock);
2078 mutex_unlock(&cur->vma_lock);
2081 mutex_unlock(&dev_set->lock);
2085 static bool vfio_pci_dev_set_needs_reset(struct vfio_device_set *dev_set)
2087 struct vfio_pci_core_device *cur;
2088 bool needs_reset = false;
2090 list_for_each_entry(cur, &dev_set->device_list, vdev.dev_set_list) {
2091 /* No VFIO device in the set can have an open device FD */
2092 if (cur->vdev.open_count)
2094 needs_reset |= cur->needs_reset;
2100 * If a bus or slot reset is available for the provided dev_set and:
2101 * - All of the devices affected by that bus or slot reset are unused
2102 * - At least one of the affected devices is marked dirty via
2103 * needs_reset (such as by lack of FLR support)
2104 * Then attempt to perform that bus or slot reset.
2105 * Returns true if the dev_set was reset.
2107 static bool vfio_pci_dev_set_try_reset(struct vfio_device_set *dev_set)
2109 struct vfio_pci_core_device *cur;
2110 struct pci_dev *pdev;
2113 if (!vfio_pci_dev_set_needs_reset(dev_set))
2116 pdev = vfio_pci_dev_set_resettable(dev_set);
2120 ret = pci_reset_bus(pdev);
2124 list_for_each_entry(cur, &dev_set->device_list, vdev.dev_set_list) {
2125 cur->needs_reset = false;
2126 if (!disable_idle_d3)
2127 vfio_pci_set_power_state(cur, PCI_D3hot);
2132 void vfio_pci_core_set_params(bool is_nointxmask, bool is_disable_vga,
2133 bool is_disable_idle_d3)
2135 nointxmask = is_nointxmask;
2136 disable_vga = is_disable_vga;
2137 disable_idle_d3 = is_disable_idle_d3;
2139 EXPORT_SYMBOL_GPL(vfio_pci_core_set_params);
2141 static void vfio_pci_core_cleanup(void)
2143 vfio_pci_uninit_perm_bits();
2146 static int __init vfio_pci_core_init(void)
2148 /* Allocate shared config space permission data used by all devices */
2149 return vfio_pci_init_perm_bits();
2152 module_init(vfio_pci_core_init);
2153 module_exit(vfio_pci_core_cleanup);
2155 MODULE_LICENSE("GPL v2");
2156 MODULE_AUTHOR(DRIVER_AUTHOR);
2157 MODULE_DESCRIPTION(DRIVER_DESC);