1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (C) 2012 Red Hat, Inc. All rights reserved.
6 * Derived from original vfio:
7 * Copyright 2010 Cisco Systems, Inc. All rights reserved.
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/vfio.h>
27 #include <linux/vgaarb.h>
28 #include <linux/nospec.h>
29 #include <linux/sched/mm.h>
31 #include "vfio_pci_private.h"
33 #define DRIVER_VERSION "0.2"
35 #define DRIVER_DESC "VFIO PCI - User Level meta-driver"
37 static char ids[1024] __initdata;
38 module_param_string(ids, ids, sizeof(ids), 0);
39 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");
41 static bool nointxmask;
42 module_param_named(nointxmask, nointxmask, bool, S_IRUGO | S_IWUSR);
43 MODULE_PARM_DESC(nointxmask,
44 "Disable support for PCI 2.3 style INTx masking. If this resolves problems for specific devices, report lspci -vvvxxx to
[email protected] so the device can be fixed automatically via the broken_intx_masking flag.");
46 #ifdef CONFIG_VFIO_PCI_VGA
47 static bool disable_vga;
48 module_param(disable_vga, bool, S_IRUGO);
49 MODULE_PARM_DESC(disable_vga, "Disable VGA resource access through vfio-pci");
52 static bool disable_idle_d3;
53 module_param(disable_idle_d3, bool, S_IRUGO | S_IWUSR);
54 MODULE_PARM_DESC(disable_idle_d3,
55 "Disable using the PCI D3 low power state for idle, unused devices");
57 static bool enable_sriov;
59 module_param(enable_sriov, bool, 0644);
60 MODULE_PARM_DESC(enable_sriov, "Enable support for SR-IOV configuration. Enabling SR-IOV on a PF typically requires support of the userspace PF driver, enabling VFs without such support may result in non-functional VFs or PF.");
63 static bool disable_denylist;
64 module_param(disable_denylist, bool, 0444);
65 MODULE_PARM_DESC(disable_denylist, "Disable use of device denylist. Disabling the denylist allows binding to devices with known errata that may lead to exploitable stability or security issues when accessed by untrusted users.");
67 static inline bool vfio_vga_disabled(void)
69 #ifdef CONFIG_VFIO_PCI_VGA
76 static bool vfio_pci_dev_in_denylist(struct pci_dev *pdev)
78 switch (pdev->vendor) {
79 case PCI_VENDOR_ID_INTEL:
80 switch (pdev->device) {
81 case PCI_DEVICE_ID_INTEL_QAT_C3XXX:
82 case PCI_DEVICE_ID_INTEL_QAT_C3XXX_VF:
83 case PCI_DEVICE_ID_INTEL_QAT_C62X:
84 case PCI_DEVICE_ID_INTEL_QAT_C62X_VF:
85 case PCI_DEVICE_ID_INTEL_QAT_DH895XCC:
86 case PCI_DEVICE_ID_INTEL_QAT_DH895XCC_VF:
96 static bool vfio_pci_is_denylisted(struct pci_dev *pdev)
98 if (!vfio_pci_dev_in_denylist(pdev))
101 if (disable_denylist) {
103 "device denylist disabled - allowing device %04x:%04x.\n",
104 pdev->vendor, pdev->device);
108 pci_warn(pdev, "%04x:%04x exists in vfio-pci device denylist, driver probing disallowed.\n",
109 pdev->vendor, pdev->device);
115 * Our VGA arbiter participation is limited since we don't know anything
116 * about the device itself. However, if the device is the only VGA device
117 * downstream of a bridge and VFIO VGA support is disabled, then we can
118 * safely return legacy VGA IO and memory as not decoded since the user
119 * has no way to get to it and routing can be disabled externally at the
122 static unsigned int vfio_pci_set_vga_decode(void *opaque, bool single_vga)
124 struct vfio_pci_device *vdev = opaque;
125 struct pci_dev *tmp = NULL, *pdev = vdev->pdev;
126 unsigned char max_busnr;
127 unsigned int decodes;
129 if (single_vga || !vfio_vga_disabled() || pci_is_root_bus(pdev->bus))
130 return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM |
131 VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM;
133 max_busnr = pci_bus_max_busnr(pdev->bus);
134 decodes = VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
136 while ((tmp = pci_get_class(PCI_CLASS_DISPLAY_VGA << 8, tmp)) != NULL) {
138 pci_domain_nr(tmp->bus) != pci_domain_nr(pdev->bus) ||
139 pci_is_root_bus(tmp->bus))
142 if (tmp->bus->number >= pdev->bus->number &&
143 tmp->bus->number <= max_busnr) {
145 decodes |= VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM;
153 static inline bool vfio_pci_is_vga(struct pci_dev *pdev)
155 return (pdev->class >> 8) == PCI_CLASS_DISPLAY_VGA;
158 static void vfio_pci_probe_mmaps(struct vfio_pci_device *vdev)
160 struct resource *res;
162 struct vfio_pci_dummy_resource *dummy_res;
164 INIT_LIST_HEAD(&vdev->dummy_resources_list);
166 for (i = 0; i < PCI_STD_NUM_BARS; i++) {
167 int bar = i + PCI_STD_RESOURCES;
169 res = &vdev->pdev->resource[bar];
171 if (!IS_ENABLED(CONFIG_VFIO_PCI_MMAP))
174 if (!(res->flags & IORESOURCE_MEM))
178 * The PCI core shouldn't set up a resource with a
179 * type but zero size. But there may be bugs that
180 * cause us to do that.
182 if (!resource_size(res))
185 if (resource_size(res) >= PAGE_SIZE) {
186 vdev->bar_mmap_supported[bar] = true;
190 if (!(res->start & ~PAGE_MASK)) {
192 * Add a dummy resource to reserve the remainder
193 * of the exclusive page in case that hot-add
194 * device's bar is assigned into it.
196 dummy_res = kzalloc(sizeof(*dummy_res), GFP_KERNEL);
197 if (dummy_res == NULL)
200 dummy_res->resource.name = "vfio sub-page reserved";
201 dummy_res->resource.start = res->end + 1;
202 dummy_res->resource.end = res->start + PAGE_SIZE - 1;
203 dummy_res->resource.flags = res->flags;
204 if (request_resource(res->parent,
205 &dummy_res->resource)) {
209 dummy_res->index = bar;
210 list_add(&dummy_res->res_next,
211 &vdev->dummy_resources_list);
212 vdev->bar_mmap_supported[bar] = true;
216 * Here we don't handle the case when the BAR is not page
217 * aligned because we can't expect the BAR will be
218 * assigned into the same location in a page in guest
219 * when we passthrough the BAR. And it's hard to access
220 * this BAR in userspace because we have no way to get
221 * the BAR's location in a page.
224 vdev->bar_mmap_supported[bar] = false;
228 static void vfio_pci_try_bus_reset(struct vfio_pci_device *vdev);
229 static void vfio_pci_disable(struct vfio_pci_device *vdev);
230 static int vfio_pci_try_zap_and_vma_lock_cb(struct pci_dev *pdev, void *data);
233 * INTx masking requires the ability to disable INTx signaling via PCI_COMMAND
234 * _and_ the ability detect when the device is asserting INTx via PCI_STATUS.
235 * If a device implements the former but not the latter we would typically
236 * expect broken_intx_masking be set and require an exclusive interrupt.
237 * However since we do have control of the device's ability to assert INTx,
238 * we can instead pretend that the device does not implement INTx, virtualizing
239 * the pin register to report zero and maintaining DisINTx set on the host.
241 static bool vfio_pci_nointx(struct pci_dev *pdev)
243 switch (pdev->vendor) {
244 case PCI_VENDOR_ID_INTEL:
245 switch (pdev->device) {
246 /* All i40e (XL710/X710/XXV710) 10/20/25/40GbE NICs */
249 case 0x1580 ... 0x1581:
250 case 0x1583 ... 0x158b:
251 case 0x37d0 ... 0x37d2:
263 static void vfio_pci_probe_power_state(struct vfio_pci_device *vdev)
265 struct pci_dev *pdev = vdev->pdev;
271 pci_read_config_word(pdev, pdev->pm_cap + PCI_PM_CTRL, &pmcsr);
273 vdev->needs_pm_restore = !(pmcsr & PCI_PM_CTRL_NO_SOFT_RESET);
277 * pci_set_power_state() wrapper handling devices which perform a soft reset on
278 * D3->D0 transition. Save state prior to D0/1/2->D3, stash it on the vdev,
279 * restore when returned to D0. Saved separately from pci_saved_state for use
280 * by PM capability emulation and separately from pci_dev internal saved state
281 * to avoid it being overwritten and consumed around other resets.
283 int vfio_pci_set_power_state(struct vfio_pci_device *vdev, pci_power_t state)
285 struct pci_dev *pdev = vdev->pdev;
286 bool needs_restore = false, needs_save = false;
289 if (vdev->needs_pm_restore) {
290 if (pdev->current_state < PCI_D3hot && state >= PCI_D3hot) {
291 pci_save_state(pdev);
295 if (pdev->current_state >= PCI_D3hot && state <= PCI_D0)
296 needs_restore = true;
299 ret = pci_set_power_state(pdev, state);
302 /* D3 might be unsupported via quirk, skip unless in D3 */
303 if (needs_save && pdev->current_state >= PCI_D3hot) {
304 vdev->pm_save = pci_store_saved_state(pdev);
305 } else if (needs_restore) {
306 pci_load_and_free_saved_state(pdev, &vdev->pm_save);
307 pci_restore_state(pdev);
314 static int vfio_pci_enable(struct vfio_pci_device *vdev)
316 struct pci_dev *pdev = vdev->pdev;
321 vfio_pci_set_power_state(vdev, PCI_D0);
323 /* Don't allow our initial saved state to include busmaster */
324 pci_clear_master(pdev);
326 ret = pci_enable_device(pdev);
330 /* If reset fails because of the device lock, fail this path entirely */
331 ret = pci_try_reset_function(pdev);
332 if (ret == -EAGAIN) {
333 pci_disable_device(pdev);
337 vdev->reset_works = !ret;
338 pci_save_state(pdev);
339 vdev->pci_saved_state = pci_store_saved_state(pdev);
340 if (!vdev->pci_saved_state)
341 pci_dbg(pdev, "%s: Couldn't store saved state\n", __func__);
343 if (likely(!nointxmask)) {
344 if (vfio_pci_nointx(pdev)) {
345 pci_info(pdev, "Masking broken INTx support\n");
349 vdev->pci_2_3 = pci_intx_mask_supported(pdev);
352 pci_read_config_word(pdev, PCI_COMMAND, &cmd);
353 if (vdev->pci_2_3 && (cmd & PCI_COMMAND_INTX_DISABLE)) {
354 cmd &= ~PCI_COMMAND_INTX_DISABLE;
355 pci_write_config_word(pdev, PCI_COMMAND, cmd);
358 ret = vfio_config_init(vdev);
360 kfree(vdev->pci_saved_state);
361 vdev->pci_saved_state = NULL;
362 pci_disable_device(pdev);
366 msix_pos = pdev->msix_cap;
371 pci_read_config_word(pdev, msix_pos + PCI_MSIX_FLAGS, &flags);
372 pci_read_config_dword(pdev, msix_pos + PCI_MSIX_TABLE, &table);
374 vdev->msix_bar = table & PCI_MSIX_TABLE_BIR;
375 vdev->msix_offset = table & PCI_MSIX_TABLE_OFFSET;
376 vdev->msix_size = ((flags & PCI_MSIX_FLAGS_QSIZE) + 1) * 16;
378 vdev->msix_bar = 0xFF;
380 if (!vfio_vga_disabled() && vfio_pci_is_vga(pdev))
381 vdev->has_vga = true;
384 if (vfio_pci_is_vga(pdev) &&
385 pdev->vendor == PCI_VENDOR_ID_INTEL &&
386 IS_ENABLED(CONFIG_VFIO_PCI_IGD)) {
387 ret = vfio_pci_igd_init(vdev);
388 if (ret && ret != -ENODEV) {
389 pci_warn(pdev, "Failed to setup Intel IGD regions\n");
394 if (pdev->vendor == PCI_VENDOR_ID_NVIDIA &&
395 IS_ENABLED(CONFIG_VFIO_PCI_NVLINK2)) {
396 ret = vfio_pci_nvdia_v100_nvlink2_init(vdev);
397 if (ret && ret != -ENODEV) {
398 pci_warn(pdev, "Failed to setup NVIDIA NV2 RAM region\n");
403 if (pdev->vendor == PCI_VENDOR_ID_IBM &&
404 IS_ENABLED(CONFIG_VFIO_PCI_NVLINK2)) {
405 ret = vfio_pci_ibm_npu2_init(vdev);
406 if (ret && ret != -ENODEV) {
407 pci_warn(pdev, "Failed to setup NVIDIA NV2 ATSD region\n");
412 vfio_pci_probe_mmaps(vdev);
417 vfio_pci_disable(vdev);
421 static void vfio_pci_disable(struct vfio_pci_device *vdev)
423 struct pci_dev *pdev = vdev->pdev;
424 struct vfio_pci_dummy_resource *dummy_res, *tmp;
425 struct vfio_pci_ioeventfd *ioeventfd, *ioeventfd_tmp;
428 /* Stop the device from further DMA */
429 pci_clear_master(pdev);
431 vfio_pci_set_irqs_ioctl(vdev, VFIO_IRQ_SET_DATA_NONE |
432 VFIO_IRQ_SET_ACTION_TRIGGER,
433 vdev->irq_type, 0, 0, NULL);
435 /* Device closed, don't need mutex here */
436 list_for_each_entry_safe(ioeventfd, ioeventfd_tmp,
437 &vdev->ioeventfds_list, next) {
438 vfio_virqfd_disable(&ioeventfd->virqfd);
439 list_del(&ioeventfd->next);
442 vdev->ioeventfds_nr = 0;
444 vdev->virq_disabled = false;
446 for (i = 0; i < vdev->num_regions; i++)
447 vdev->region[i].ops->release(vdev, &vdev->region[i]);
449 vdev->num_regions = 0;
451 vdev->region = NULL; /* don't krealloc a freed pointer */
453 vfio_config_free(vdev);
455 for (i = 0; i < PCI_STD_NUM_BARS; i++) {
456 bar = i + PCI_STD_RESOURCES;
457 if (!vdev->barmap[bar])
459 pci_iounmap(pdev, vdev->barmap[bar]);
460 pci_release_selected_regions(pdev, 1 << bar);
461 vdev->barmap[bar] = NULL;
464 list_for_each_entry_safe(dummy_res, tmp,
465 &vdev->dummy_resources_list, res_next) {
466 list_del(&dummy_res->res_next);
467 release_resource(&dummy_res->resource);
471 vdev->needs_reset = true;
474 * If we have saved state, restore it. If we can reset the device,
475 * even better. Resetting with current state seems better than
476 * nothing, but saving and restoring current state without reset
479 if (pci_load_and_free_saved_state(pdev, &vdev->pci_saved_state)) {
480 pci_info(pdev, "%s: Couldn't reload saved state\n", __func__);
482 if (!vdev->reset_works)
485 pci_save_state(pdev);
489 * Disable INTx and MSI, presumably to avoid spurious interrupts
490 * during reset. Stolen from pci_reset_function()
492 pci_write_config_word(pdev, PCI_COMMAND, PCI_COMMAND_INTX_DISABLE);
495 * Try to get the locks ourselves to prevent a deadlock. The
496 * success of this is dependent on being able to lock the device,
497 * which is not always possible.
498 * We can not use the "try" reset interface here, which will
499 * overwrite the previously restored configuration information.
501 if (vdev->reset_works && pci_cfg_access_trylock(pdev)) {
502 if (device_trylock(&pdev->dev)) {
503 if (!__pci_reset_function_locked(pdev))
504 vdev->needs_reset = false;
505 device_unlock(&pdev->dev);
507 pci_cfg_access_unlock(pdev);
510 pci_restore_state(pdev);
512 pci_disable_device(pdev);
514 vfio_pci_try_bus_reset(vdev);
516 if (!disable_idle_d3)
517 vfio_pci_set_power_state(vdev, PCI_D3hot);
520 static struct pci_driver vfio_pci_driver;
522 static struct vfio_pci_device *get_pf_vdev(struct vfio_pci_device *vdev,
523 struct vfio_device **pf_dev)
525 struct pci_dev *physfn = pci_physfn(vdev->pdev);
527 if (!vdev->pdev->is_virtfn)
530 *pf_dev = vfio_device_get_from_dev(&physfn->dev);
534 if (pci_dev_driver(physfn) != &vfio_pci_driver) {
535 vfio_device_put(*pf_dev);
539 return vfio_device_data(*pf_dev);
542 static void vfio_pci_vf_token_user_add(struct vfio_pci_device *vdev, int val)
544 struct vfio_device *pf_dev;
545 struct vfio_pci_device *pf_vdev = get_pf_vdev(vdev, &pf_dev);
550 mutex_lock(&pf_vdev->vf_token->lock);
551 pf_vdev->vf_token->users += val;
552 WARN_ON(pf_vdev->vf_token->users < 0);
553 mutex_unlock(&pf_vdev->vf_token->lock);
555 vfio_device_put(pf_dev);
558 static void vfio_pci_release(void *device_data)
560 struct vfio_pci_device *vdev = device_data;
562 mutex_lock(&vdev->reflck->lock);
564 if (!(--vdev->refcnt)) {
565 vfio_pci_vf_token_user_add(vdev, -1);
566 vfio_spapr_pci_eeh_release(vdev->pdev);
567 vfio_pci_disable(vdev);
569 mutex_lock(&vdev->igate);
570 if (vdev->err_trigger) {
571 eventfd_ctx_put(vdev->err_trigger);
572 vdev->err_trigger = NULL;
574 if (vdev->req_trigger) {
575 eventfd_ctx_put(vdev->req_trigger);
576 vdev->req_trigger = NULL;
578 mutex_unlock(&vdev->igate);
581 mutex_unlock(&vdev->reflck->lock);
583 module_put(THIS_MODULE);
586 static int vfio_pci_open(void *device_data)
588 struct vfio_pci_device *vdev = device_data;
591 if (!try_module_get(THIS_MODULE))
594 mutex_lock(&vdev->reflck->lock);
597 ret = vfio_pci_enable(vdev);
601 vfio_spapr_pci_eeh_open(vdev->pdev);
602 vfio_pci_vf_token_user_add(vdev, 1);
606 mutex_unlock(&vdev->reflck->lock);
608 module_put(THIS_MODULE);
612 static int vfio_pci_get_irq_count(struct vfio_pci_device *vdev, int irq_type)
614 if (irq_type == VFIO_PCI_INTX_IRQ_INDEX) {
617 if (!IS_ENABLED(CONFIG_VFIO_PCI_INTX) ||
618 vdev->nointx || vdev->pdev->is_virtfn)
621 pci_read_config_byte(vdev->pdev, PCI_INTERRUPT_PIN, &pin);
624 } else if (irq_type == VFIO_PCI_MSI_IRQ_INDEX) {
628 pos = vdev->pdev->msi_cap;
630 pci_read_config_word(vdev->pdev,
631 pos + PCI_MSI_FLAGS, &flags);
632 return 1 << ((flags & PCI_MSI_FLAGS_QMASK) >> 1);
634 } else if (irq_type == VFIO_PCI_MSIX_IRQ_INDEX) {
638 pos = vdev->pdev->msix_cap;
640 pci_read_config_word(vdev->pdev,
641 pos + PCI_MSIX_FLAGS, &flags);
643 return (flags & PCI_MSIX_FLAGS_QSIZE) + 1;
645 } else if (irq_type == VFIO_PCI_ERR_IRQ_INDEX) {
646 if (pci_is_pcie(vdev->pdev))
648 } else if (irq_type == VFIO_PCI_REQ_IRQ_INDEX) {
655 static int vfio_pci_count_devs(struct pci_dev *pdev, void *data)
661 struct vfio_pci_fill_info {
664 struct vfio_pci_dependent_device *devices;
667 static int vfio_pci_fill_devs(struct pci_dev *pdev, void *data)
669 struct vfio_pci_fill_info *fill = data;
670 struct iommu_group *iommu_group;
672 if (fill->cur == fill->max)
673 return -EAGAIN; /* Something changed, try again */
675 iommu_group = iommu_group_get(&pdev->dev);
677 return -EPERM; /* Cannot reset non-isolated devices */
679 fill->devices[fill->cur].group_id = iommu_group_id(iommu_group);
680 fill->devices[fill->cur].segment = pci_domain_nr(pdev->bus);
681 fill->devices[fill->cur].bus = pdev->bus->number;
682 fill->devices[fill->cur].devfn = pdev->devfn;
684 iommu_group_put(iommu_group);
688 struct vfio_pci_group_entry {
689 struct vfio_group *group;
693 struct vfio_pci_group_info {
695 struct vfio_pci_group_entry *groups;
698 static int vfio_pci_validate_devs(struct pci_dev *pdev, void *data)
700 struct vfio_pci_group_info *info = data;
701 struct iommu_group *group;
704 group = iommu_group_get(&pdev->dev);
708 id = iommu_group_id(group);
710 for (i = 0; i < info->count; i++)
711 if (info->groups[i].id == id)
714 iommu_group_put(group);
716 return (i == info->count) ? -EINVAL : 0;
719 static bool vfio_pci_dev_below_slot(struct pci_dev *pdev, struct pci_slot *slot)
721 for (; pdev; pdev = pdev->bus->self)
722 if (pdev->bus == slot->bus)
723 return (pdev->slot == slot);
727 struct vfio_pci_walk_info {
728 int (*fn)(struct pci_dev *, void *data);
730 struct pci_dev *pdev;
735 static int vfio_pci_walk_wrapper(struct pci_dev *pdev, void *data)
737 struct vfio_pci_walk_info *walk = data;
739 if (!walk->slot || vfio_pci_dev_below_slot(pdev, walk->pdev->slot))
740 walk->ret = walk->fn(pdev, walk->data);
745 static int vfio_pci_for_each_slot_or_bus(struct pci_dev *pdev,
746 int (*fn)(struct pci_dev *,
747 void *data), void *data,
750 struct vfio_pci_walk_info walk = {
751 .fn = fn, .data = data, .pdev = pdev, .slot = slot, .ret = 0,
754 pci_walk_bus(pdev->bus, vfio_pci_walk_wrapper, &walk);
759 static int msix_mmappable_cap(struct vfio_pci_device *vdev,
760 struct vfio_info_cap *caps)
762 struct vfio_info_cap_header header = {
763 .id = VFIO_REGION_INFO_CAP_MSIX_MAPPABLE,
767 return vfio_info_add_capability(caps, &header, sizeof(header));
770 int vfio_pci_register_dev_region(struct vfio_pci_device *vdev,
771 unsigned int type, unsigned int subtype,
772 const struct vfio_pci_regops *ops,
773 size_t size, u32 flags, void *data)
775 struct vfio_pci_region *region;
777 region = krealloc(vdev->region,
778 (vdev->num_regions + 1) * sizeof(*region),
783 vdev->region = region;
784 vdev->region[vdev->num_regions].type = type;
785 vdev->region[vdev->num_regions].subtype = subtype;
786 vdev->region[vdev->num_regions].ops = ops;
787 vdev->region[vdev->num_regions].size = size;
788 vdev->region[vdev->num_regions].flags = flags;
789 vdev->region[vdev->num_regions].data = data;
796 struct vfio_devices {
797 struct vfio_device **devices;
802 static long vfio_pci_ioctl(void *device_data,
803 unsigned int cmd, unsigned long arg)
805 struct vfio_pci_device *vdev = device_data;
808 if (cmd == VFIO_DEVICE_GET_INFO) {
809 struct vfio_device_info info;
810 struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
813 minsz = offsetofend(struct vfio_device_info, num_irqs);
815 /* For backward compatibility, cannot require this */
816 capsz = offsetofend(struct vfio_iommu_type1_info, cap_offset);
818 if (copy_from_user(&info, (void __user *)arg, minsz))
821 if (info.argsz < minsz)
824 if (info.argsz >= capsz) {
829 info.flags = VFIO_DEVICE_FLAGS_PCI;
831 if (vdev->reset_works)
832 info.flags |= VFIO_DEVICE_FLAGS_RESET;
834 info.num_regions = VFIO_PCI_NUM_REGIONS + vdev->num_regions;
835 info.num_irqs = VFIO_PCI_NUM_IRQS;
837 if (IS_ENABLED(CONFIG_VFIO_PCI_ZDEV)) {
838 int ret = vfio_pci_info_zdev_add_caps(vdev, &caps);
840 if (ret && ret != -ENODEV) {
841 pci_warn(vdev->pdev, "Failed to setup zPCI info capabilities\n");
847 info.flags |= VFIO_DEVICE_FLAGS_CAPS;
848 if (info.argsz < sizeof(info) + caps.size) {
849 info.argsz = sizeof(info) + caps.size;
851 vfio_info_cap_shift(&caps, sizeof(info));
852 if (copy_to_user((void __user *)arg +
853 sizeof(info), caps.buf,
858 info.cap_offset = sizeof(info);
864 return copy_to_user((void __user *)arg, &info, minsz) ?
867 } else if (cmd == VFIO_DEVICE_GET_REGION_INFO) {
868 struct pci_dev *pdev = vdev->pdev;
869 struct vfio_region_info info;
870 struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
873 minsz = offsetofend(struct vfio_region_info, offset);
875 if (copy_from_user(&info, (void __user *)arg, minsz))
878 if (info.argsz < minsz)
881 switch (info.index) {
882 case VFIO_PCI_CONFIG_REGION_INDEX:
883 info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
884 info.size = pdev->cfg_size;
885 info.flags = VFIO_REGION_INFO_FLAG_READ |
886 VFIO_REGION_INFO_FLAG_WRITE;
888 case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
889 info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
890 info.size = pci_resource_len(pdev, info.index);
896 info.flags = VFIO_REGION_INFO_FLAG_READ |
897 VFIO_REGION_INFO_FLAG_WRITE;
898 if (vdev->bar_mmap_supported[info.index]) {
899 info.flags |= VFIO_REGION_INFO_FLAG_MMAP;
900 if (info.index == vdev->msix_bar) {
901 ret = msix_mmappable_cap(vdev, &caps);
908 case VFIO_PCI_ROM_REGION_INDEX:
914 info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
917 /* Report the BAR size, not the ROM size */
918 info.size = pci_resource_len(pdev, info.index);
920 /* Shadow ROMs appear as PCI option ROMs */
921 if (pdev->resource[PCI_ROM_RESOURCE].flags &
922 IORESOURCE_ROM_SHADOW)
929 * Is it really there? Enable memory decode for
930 * implicit access in pci_map_rom().
932 cmd = vfio_pci_memory_lock_and_enable(vdev);
933 io = pci_map_rom(pdev, &size);
935 info.flags = VFIO_REGION_INFO_FLAG_READ;
936 pci_unmap_rom(pdev, io);
940 vfio_pci_memory_unlock_and_restore(vdev, cmd);
944 case VFIO_PCI_VGA_REGION_INDEX:
948 info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
950 info.flags = VFIO_REGION_INFO_FLAG_READ |
951 VFIO_REGION_INFO_FLAG_WRITE;
956 struct vfio_region_info_cap_type cap_type = {
957 .header.id = VFIO_REGION_INFO_CAP_TYPE,
958 .header.version = 1 };
961 VFIO_PCI_NUM_REGIONS + vdev->num_regions)
963 info.index = array_index_nospec(info.index,
964 VFIO_PCI_NUM_REGIONS +
967 i = info.index - VFIO_PCI_NUM_REGIONS;
969 info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
970 info.size = vdev->region[i].size;
971 info.flags = vdev->region[i].flags;
973 cap_type.type = vdev->region[i].type;
974 cap_type.subtype = vdev->region[i].subtype;
976 ret = vfio_info_add_capability(&caps, &cap_type.header,
981 if (vdev->region[i].ops->add_capability) {
982 ret = vdev->region[i].ops->add_capability(vdev,
983 &vdev->region[i], &caps);
991 info.flags |= VFIO_REGION_INFO_FLAG_CAPS;
992 if (info.argsz < sizeof(info) + caps.size) {
993 info.argsz = sizeof(info) + caps.size;
996 vfio_info_cap_shift(&caps, sizeof(info));
997 if (copy_to_user((void __user *)arg +
998 sizeof(info), caps.buf,
1003 info.cap_offset = sizeof(info);
1009 return copy_to_user((void __user *)arg, &info, minsz) ?
1012 } else if (cmd == VFIO_DEVICE_GET_IRQ_INFO) {
1013 struct vfio_irq_info info;
1015 minsz = offsetofend(struct vfio_irq_info, count);
1017 if (copy_from_user(&info, (void __user *)arg, minsz))
1020 if (info.argsz < minsz || info.index >= VFIO_PCI_NUM_IRQS)
1023 switch (info.index) {
1024 case VFIO_PCI_INTX_IRQ_INDEX ... VFIO_PCI_MSIX_IRQ_INDEX:
1025 case VFIO_PCI_REQ_IRQ_INDEX:
1027 case VFIO_PCI_ERR_IRQ_INDEX:
1028 if (pci_is_pcie(vdev->pdev))
1035 info.flags = VFIO_IRQ_INFO_EVENTFD;
1037 info.count = vfio_pci_get_irq_count(vdev, info.index);
1039 if (info.index == VFIO_PCI_INTX_IRQ_INDEX)
1040 info.flags |= (VFIO_IRQ_INFO_MASKABLE |
1041 VFIO_IRQ_INFO_AUTOMASKED);
1043 info.flags |= VFIO_IRQ_INFO_NORESIZE;
1045 return copy_to_user((void __user *)arg, &info, minsz) ?
1048 } else if (cmd == VFIO_DEVICE_SET_IRQS) {
1049 struct vfio_irq_set hdr;
1052 size_t data_size = 0;
1054 minsz = offsetofend(struct vfio_irq_set, count);
1056 if (copy_from_user(&hdr, (void __user *)arg, minsz))
1059 max = vfio_pci_get_irq_count(vdev, hdr.index);
1061 ret = vfio_set_irqs_validate_and_prepare(&hdr, max,
1062 VFIO_PCI_NUM_IRQS, &data_size);
1067 data = memdup_user((void __user *)(arg + minsz),
1070 return PTR_ERR(data);
1073 mutex_lock(&vdev->igate);
1075 ret = vfio_pci_set_irqs_ioctl(vdev, hdr.flags, hdr.index,
1076 hdr.start, hdr.count, data);
1078 mutex_unlock(&vdev->igate);
1083 } else if (cmd == VFIO_DEVICE_RESET) {
1086 if (!vdev->reset_works)
1089 vfio_pci_zap_and_down_write_memory_lock(vdev);
1090 ret = pci_try_reset_function(vdev->pdev);
1091 up_write(&vdev->memory_lock);
1095 } else if (cmd == VFIO_DEVICE_GET_PCI_HOT_RESET_INFO) {
1096 struct vfio_pci_hot_reset_info hdr;
1097 struct vfio_pci_fill_info fill = { 0 };
1098 struct vfio_pci_dependent_device *devices = NULL;
1102 minsz = offsetofend(struct vfio_pci_hot_reset_info, count);
1104 if (copy_from_user(&hdr, (void __user *)arg, minsz))
1107 if (hdr.argsz < minsz)
1112 /* Can we do a slot or bus reset or neither? */
1113 if (!pci_probe_reset_slot(vdev->pdev->slot))
1115 else if (pci_probe_reset_bus(vdev->pdev->bus))
1118 /* How many devices are affected? */
1119 ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1120 vfio_pci_count_devs,
1125 WARN_ON(!fill.max); /* Should always be at least one */
1128 * If there's enough space, fill it now, otherwise return
1129 * -ENOSPC and the number of devices affected.
1131 if (hdr.argsz < sizeof(hdr) + (fill.max * sizeof(*devices))) {
1133 hdr.count = fill.max;
1134 goto reset_info_exit;
1137 devices = kcalloc(fill.max, sizeof(*devices), GFP_KERNEL);
1141 fill.devices = devices;
1143 ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1148 * If a device was removed between counting and filling,
1149 * we may come up short of fill.max. If a device was
1150 * added, we'll have a return of -EAGAIN above.
1153 hdr.count = fill.cur;
1156 if (copy_to_user((void __user *)arg, &hdr, minsz))
1160 if (copy_to_user((void __user *)(arg + minsz), devices,
1161 hdr.count * sizeof(*devices)))
1168 } else if (cmd == VFIO_DEVICE_PCI_HOT_RESET) {
1169 struct vfio_pci_hot_reset hdr;
1171 struct vfio_pci_group_entry *groups;
1172 struct vfio_pci_group_info info;
1173 struct vfio_devices devs = { .cur_index = 0 };
1175 int i, group_idx, mem_idx = 0, count = 0, ret = 0;
1177 minsz = offsetofend(struct vfio_pci_hot_reset, count);
1179 if (copy_from_user(&hdr, (void __user *)arg, minsz))
1182 if (hdr.argsz < minsz || hdr.flags)
1185 /* Can we do a slot or bus reset or neither? */
1186 if (!pci_probe_reset_slot(vdev->pdev->slot))
1188 else if (pci_probe_reset_bus(vdev->pdev->bus))
1192 * We can't let userspace give us an arbitrarily large
1193 * buffer to copy, so verify how many we think there
1194 * could be. Note groups can have multiple devices so
1195 * one group per device is the max.
1197 ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1198 vfio_pci_count_devs,
1203 /* Somewhere between 1 and count is OK */
1204 if (!hdr.count || hdr.count > count)
1207 group_fds = kcalloc(hdr.count, sizeof(*group_fds), GFP_KERNEL);
1208 groups = kcalloc(hdr.count, sizeof(*groups), GFP_KERNEL);
1209 if (!group_fds || !groups) {
1215 if (copy_from_user(group_fds, (void __user *)(arg + minsz),
1216 hdr.count * sizeof(*group_fds))) {
1223 * For each group_fd, get the group through the vfio external
1224 * user interface and store the group and iommu ID. This
1225 * ensures the group is held across the reset.
1227 for (group_idx = 0; group_idx < hdr.count; group_idx++) {
1228 struct vfio_group *group;
1229 struct fd f = fdget(group_fds[group_idx]);
1235 group = vfio_group_get_external_user(f.file);
1237 if (IS_ERR(group)) {
1238 ret = PTR_ERR(group);
1242 groups[group_idx].group = group;
1243 groups[group_idx].id =
1244 vfio_external_user_iommu_id(group);
1249 /* release reference to groups on error */
1251 goto hot_reset_release;
1253 info.count = hdr.count;
1254 info.groups = groups;
1257 * Test whether all the affected devices are contained
1258 * by the set of groups provided by the user.
1260 ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1261 vfio_pci_validate_devs,
1264 goto hot_reset_release;
1266 devs.max_index = count;
1267 devs.devices = kcalloc(count, sizeof(struct vfio_device *),
1269 if (!devs.devices) {
1271 goto hot_reset_release;
1275 * We need to get memory_lock for each device, but devices
1276 * can share mmap_lock, therefore we need to zap and hold
1277 * the vma_lock for each device, and only then get each
1280 ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1281 vfio_pci_try_zap_and_vma_lock_cb,
1284 goto hot_reset_release;
1286 for (; mem_idx < devs.cur_index; mem_idx++) {
1287 struct vfio_pci_device *tmp;
1289 tmp = vfio_device_data(devs.devices[mem_idx]);
1291 ret = down_write_trylock(&tmp->memory_lock);
1294 goto hot_reset_release;
1296 mutex_unlock(&tmp->vma_lock);
1299 /* User has access, do the reset */
1300 ret = pci_reset_bus(vdev->pdev);
1303 for (i = 0; i < devs.cur_index; i++) {
1304 struct vfio_device *device;
1305 struct vfio_pci_device *tmp;
1307 device = devs.devices[i];
1308 tmp = vfio_device_data(device);
1311 up_write(&tmp->memory_lock);
1313 mutex_unlock(&tmp->vma_lock);
1314 vfio_device_put(device);
1316 kfree(devs.devices);
1318 for (group_idx--; group_idx >= 0; group_idx--)
1319 vfio_group_put_external_user(groups[group_idx].group);
1323 } else if (cmd == VFIO_DEVICE_IOEVENTFD) {
1324 struct vfio_device_ioeventfd ioeventfd;
1327 minsz = offsetofend(struct vfio_device_ioeventfd, fd);
1329 if (copy_from_user(&ioeventfd, (void __user *)arg, minsz))
1332 if (ioeventfd.argsz < minsz)
1335 if (ioeventfd.flags & ~VFIO_DEVICE_IOEVENTFD_SIZE_MASK)
1338 count = ioeventfd.flags & VFIO_DEVICE_IOEVENTFD_SIZE_MASK;
1340 if (hweight8(count) != 1 || ioeventfd.fd < -1)
1343 return vfio_pci_ioeventfd(vdev, ioeventfd.offset,
1344 ioeventfd.data, count, ioeventfd.fd);
1345 } else if (cmd == VFIO_DEVICE_FEATURE) {
1346 struct vfio_device_feature feature;
1349 minsz = offsetofend(struct vfio_device_feature, flags);
1351 if (copy_from_user(&feature, (void __user *)arg, minsz))
1354 if (feature.argsz < minsz)
1357 /* Check unknown flags */
1358 if (feature.flags & ~(VFIO_DEVICE_FEATURE_MASK |
1359 VFIO_DEVICE_FEATURE_SET |
1360 VFIO_DEVICE_FEATURE_GET |
1361 VFIO_DEVICE_FEATURE_PROBE))
1364 /* GET & SET are mutually exclusive except with PROBE */
1365 if (!(feature.flags & VFIO_DEVICE_FEATURE_PROBE) &&
1366 (feature.flags & VFIO_DEVICE_FEATURE_SET) &&
1367 (feature.flags & VFIO_DEVICE_FEATURE_GET))
1370 switch (feature.flags & VFIO_DEVICE_FEATURE_MASK) {
1371 case VFIO_DEVICE_FEATURE_PCI_VF_TOKEN:
1372 if (!vdev->vf_token)
1376 * We do not support GET of the VF Token UUID as this
1377 * could expose the token of the previous device user.
1379 if (feature.flags & VFIO_DEVICE_FEATURE_GET)
1382 if (feature.flags & VFIO_DEVICE_FEATURE_PROBE)
1385 /* Don't SET unless told to do so */
1386 if (!(feature.flags & VFIO_DEVICE_FEATURE_SET))
1389 if (feature.argsz < minsz + sizeof(uuid))
1392 if (copy_from_user(&uuid, (void __user *)(arg + minsz),
1396 mutex_lock(&vdev->vf_token->lock);
1397 uuid_copy(&vdev->vf_token->uuid, &uuid);
1398 mutex_unlock(&vdev->vf_token->lock);
1409 static ssize_t vfio_pci_rw(void *device_data, char __user *buf,
1410 size_t count, loff_t *ppos, bool iswrite)
1412 unsigned int index = VFIO_PCI_OFFSET_TO_INDEX(*ppos);
1413 struct vfio_pci_device *vdev = device_data;
1415 if (index >= VFIO_PCI_NUM_REGIONS + vdev->num_regions)
1419 case VFIO_PCI_CONFIG_REGION_INDEX:
1420 return vfio_pci_config_rw(vdev, buf, count, ppos, iswrite);
1422 case VFIO_PCI_ROM_REGION_INDEX:
1425 return vfio_pci_bar_rw(vdev, buf, count, ppos, false);
1427 case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
1428 return vfio_pci_bar_rw(vdev, buf, count, ppos, iswrite);
1430 case VFIO_PCI_VGA_REGION_INDEX:
1431 return vfio_pci_vga_rw(vdev, buf, count, ppos, iswrite);
1433 index -= VFIO_PCI_NUM_REGIONS;
1434 return vdev->region[index].ops->rw(vdev, buf,
1435 count, ppos, iswrite);
1441 static ssize_t vfio_pci_read(void *device_data, char __user *buf,
1442 size_t count, loff_t *ppos)
1447 return vfio_pci_rw(device_data, buf, count, ppos, false);
1450 static ssize_t vfio_pci_write(void *device_data, const char __user *buf,
1451 size_t count, loff_t *ppos)
1456 return vfio_pci_rw(device_data, (char __user *)buf, count, ppos, true);
1459 /* Return 1 on zap and vma_lock acquired, 0 on contention (only with @try) */
1460 static int vfio_pci_zap_and_vma_lock(struct vfio_pci_device *vdev, bool try)
1462 struct vfio_pci_mmap_vma *mmap_vma, *tmp;
1466 * vma_lock is nested under mmap_lock for vm_ops callback paths.
1467 * The memory_lock semaphore is used by both code paths calling
1468 * into this function to zap vmas and the vm_ops.fault callback
1469 * to protect the memory enable state of the device.
1471 * When zapping vmas we need to maintain the mmap_lock => vma_lock
1472 * ordering, which requires using vma_lock to walk vma_list to
1473 * acquire an mm, then dropping vma_lock to get the mmap_lock and
1474 * reacquiring vma_lock. This logic is derived from similar
1475 * requirements in uverbs_user_mmap_disassociate().
1477 * mmap_lock must always be the top-level lock when it is taken.
1478 * Therefore we can only hold the memory_lock write lock when
1479 * vma_list is empty, as we'd need to take mmap_lock to clear
1480 * entries. vma_list can only be guaranteed empty when holding
1481 * vma_lock, thus memory_lock is nested under vma_lock.
1483 * This enables the vm_ops.fault callback to acquire vma_lock,
1484 * followed by memory_lock read lock, while already holding
1485 * mmap_lock without risk of deadlock.
1488 struct mm_struct *mm = NULL;
1491 if (!mutex_trylock(&vdev->vma_lock))
1494 mutex_lock(&vdev->vma_lock);
1496 while (!list_empty(&vdev->vma_list)) {
1497 mmap_vma = list_first_entry(&vdev->vma_list,
1498 struct vfio_pci_mmap_vma,
1500 mm = mmap_vma->vma->vm_mm;
1501 if (mmget_not_zero(mm))
1504 list_del(&mmap_vma->vma_next);
1510 mutex_unlock(&vdev->vma_lock);
1513 if (!mmap_read_trylock(mm)) {
1521 if (!mutex_trylock(&vdev->vma_lock)) {
1522 mmap_read_unlock(mm);
1527 mutex_lock(&vdev->vma_lock);
1529 list_for_each_entry_safe(mmap_vma, tmp,
1530 &vdev->vma_list, vma_next) {
1531 struct vm_area_struct *vma = mmap_vma->vma;
1533 if (vma->vm_mm != mm)
1536 list_del(&mmap_vma->vma_next);
1539 zap_vma_ptes(vma, vma->vm_start,
1540 vma->vm_end - vma->vm_start);
1542 mutex_unlock(&vdev->vma_lock);
1543 mmap_read_unlock(mm);
1548 void vfio_pci_zap_and_down_write_memory_lock(struct vfio_pci_device *vdev)
1550 vfio_pci_zap_and_vma_lock(vdev, false);
1551 down_write(&vdev->memory_lock);
1552 mutex_unlock(&vdev->vma_lock);
1555 u16 vfio_pci_memory_lock_and_enable(struct vfio_pci_device *vdev)
1559 down_write(&vdev->memory_lock);
1560 pci_read_config_word(vdev->pdev, PCI_COMMAND, &cmd);
1561 if (!(cmd & PCI_COMMAND_MEMORY))
1562 pci_write_config_word(vdev->pdev, PCI_COMMAND,
1563 cmd | PCI_COMMAND_MEMORY);
1568 void vfio_pci_memory_unlock_and_restore(struct vfio_pci_device *vdev, u16 cmd)
1570 pci_write_config_word(vdev->pdev, PCI_COMMAND, cmd);
1571 up_write(&vdev->memory_lock);
1574 /* Caller holds vma_lock */
1575 static int __vfio_pci_add_vma(struct vfio_pci_device *vdev,
1576 struct vm_area_struct *vma)
1578 struct vfio_pci_mmap_vma *mmap_vma;
1580 mmap_vma = kmalloc(sizeof(*mmap_vma), GFP_KERNEL);
1584 mmap_vma->vma = vma;
1585 list_add(&mmap_vma->vma_next, &vdev->vma_list);
1591 * Zap mmaps on open so that we can fault them in on access and therefore
1592 * our vma_list only tracks mappings accessed since last zap.
1594 static void vfio_pci_mmap_open(struct vm_area_struct *vma)
1596 zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
1599 static void vfio_pci_mmap_close(struct vm_area_struct *vma)
1601 struct vfio_pci_device *vdev = vma->vm_private_data;
1602 struct vfio_pci_mmap_vma *mmap_vma;
1604 mutex_lock(&vdev->vma_lock);
1605 list_for_each_entry(mmap_vma, &vdev->vma_list, vma_next) {
1606 if (mmap_vma->vma == vma) {
1607 list_del(&mmap_vma->vma_next);
1612 mutex_unlock(&vdev->vma_lock);
1615 static vm_fault_t vfio_pci_mmap_fault(struct vm_fault *vmf)
1617 struct vm_area_struct *vma = vmf->vma;
1618 struct vfio_pci_device *vdev = vma->vm_private_data;
1619 vm_fault_t ret = VM_FAULT_NOPAGE;
1621 mutex_lock(&vdev->vma_lock);
1622 down_read(&vdev->memory_lock);
1624 if (!__vfio_pci_memory_enabled(vdev)) {
1625 ret = VM_FAULT_SIGBUS;
1626 mutex_unlock(&vdev->vma_lock);
1630 if (__vfio_pci_add_vma(vdev, vma)) {
1632 mutex_unlock(&vdev->vma_lock);
1636 mutex_unlock(&vdev->vma_lock);
1638 if (remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
1639 vma->vm_end - vma->vm_start, vma->vm_page_prot))
1640 ret = VM_FAULT_SIGBUS;
1643 up_read(&vdev->memory_lock);
1647 static const struct vm_operations_struct vfio_pci_mmap_ops = {
1648 .open = vfio_pci_mmap_open,
1649 .close = vfio_pci_mmap_close,
1650 .fault = vfio_pci_mmap_fault,
1653 static int vfio_pci_mmap(void *device_data, struct vm_area_struct *vma)
1655 struct vfio_pci_device *vdev = device_data;
1656 struct pci_dev *pdev = vdev->pdev;
1658 u64 phys_len, req_len, pgoff, req_start;
1661 index = vma->vm_pgoff >> (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT);
1663 if (vma->vm_end < vma->vm_start)
1665 if ((vma->vm_flags & VM_SHARED) == 0)
1667 if (index >= VFIO_PCI_NUM_REGIONS) {
1668 int regnum = index - VFIO_PCI_NUM_REGIONS;
1669 struct vfio_pci_region *region = vdev->region + regnum;
1671 if (region && region->ops && region->ops->mmap &&
1672 (region->flags & VFIO_REGION_INFO_FLAG_MMAP))
1673 return region->ops->mmap(vdev, region, vma);
1676 if (index >= VFIO_PCI_ROM_REGION_INDEX)
1678 if (!vdev->bar_mmap_supported[index])
1681 phys_len = PAGE_ALIGN(pci_resource_len(pdev, index));
1682 req_len = vma->vm_end - vma->vm_start;
1683 pgoff = vma->vm_pgoff &
1684 ((1U << (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT)) - 1);
1685 req_start = pgoff << PAGE_SHIFT;
1687 if (req_start + req_len > phys_len)
1691 * Even though we don't make use of the barmap for the mmap,
1692 * we need to request the region and the barmap tracks that.
1694 if (!vdev->barmap[index]) {
1695 ret = pci_request_selected_regions(pdev,
1696 1 << index, "vfio-pci");
1700 vdev->barmap[index] = pci_iomap(pdev, index, 0);
1701 if (!vdev->barmap[index]) {
1702 pci_release_selected_regions(pdev, 1 << index);
1707 vma->vm_private_data = vdev;
1708 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1709 vma->vm_pgoff = (pci_resource_start(pdev, index) >> PAGE_SHIFT) + pgoff;
1712 * See remap_pfn_range(), called from vfio_pci_fault() but we can't
1713 * change vm_flags within the fault handler. Set them now.
1715 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1716 vma->vm_ops = &vfio_pci_mmap_ops;
1721 static void vfio_pci_request(void *device_data, unsigned int count)
1723 struct vfio_pci_device *vdev = device_data;
1724 struct pci_dev *pdev = vdev->pdev;
1726 mutex_lock(&vdev->igate);
1728 if (vdev->req_trigger) {
1730 pci_notice_ratelimited(pdev,
1731 "Relaying device request to user (#%u)\n",
1733 eventfd_signal(vdev->req_trigger, 1);
1734 } else if (count == 0) {
1736 "No device request channel registered, blocked until released by user\n");
1739 mutex_unlock(&vdev->igate);
1742 static int vfio_pci_validate_vf_token(struct vfio_pci_device *vdev,
1743 bool vf_token, uuid_t *uuid)
1746 * There's always some degree of trust or collaboration between SR-IOV
1747 * PF and VFs, even if just that the PF hosts the SR-IOV capability and
1748 * can disrupt VFs with a reset, but often the PF has more explicit
1749 * access to deny service to the VF or access data passed through the
1750 * VF. We therefore require an opt-in via a shared VF token (UUID) to
1751 * represent this trust. This both prevents that a VF driver might
1752 * assume the PF driver is a trusted, in-kernel driver, and also that
1753 * a PF driver might be replaced with a rogue driver, unknown to in-use
1756 * Therefore when presented with a VF, if the PF is a vfio device and
1757 * it is bound to the vfio-pci driver, the user needs to provide a VF
1758 * token to access the device, in the form of appending a vf_token to
1759 * the device name, for example:
1761 * "0000:04:10.0 vf_token=bd8d9d2b-5a5f-4f5a-a211-f591514ba1f3"
1763 * When presented with a PF which has VFs in use, the user must also
1764 * provide the current VF token to prove collaboration with existing
1765 * VF users. If VFs are not in use, the VF token provided for the PF
1766 * device will act to set the VF token.
1768 * If the VF token is provided but unused, an error is generated.
1770 if (!vdev->pdev->is_virtfn && !vdev->vf_token && !vf_token)
1771 return 0; /* No VF token provided or required */
1773 if (vdev->pdev->is_virtfn) {
1774 struct vfio_device *pf_dev;
1775 struct vfio_pci_device *pf_vdev = get_pf_vdev(vdev, &pf_dev);
1780 return 0; /* PF is not vfio-pci, no VF token */
1782 pci_info_ratelimited(vdev->pdev,
1783 "VF token incorrectly provided, PF not bound to vfio-pci\n");
1788 vfio_device_put(pf_dev);
1789 pci_info_ratelimited(vdev->pdev,
1790 "VF token required to access device\n");
1794 mutex_lock(&pf_vdev->vf_token->lock);
1795 match = uuid_equal(uuid, &pf_vdev->vf_token->uuid);
1796 mutex_unlock(&pf_vdev->vf_token->lock);
1798 vfio_device_put(pf_dev);
1801 pci_info_ratelimited(vdev->pdev,
1802 "Incorrect VF token provided for device\n");
1805 } else if (vdev->vf_token) {
1806 mutex_lock(&vdev->vf_token->lock);
1807 if (vdev->vf_token->users) {
1809 mutex_unlock(&vdev->vf_token->lock);
1810 pci_info_ratelimited(vdev->pdev,
1811 "VF token required to access device\n");
1815 if (!uuid_equal(uuid, &vdev->vf_token->uuid)) {
1816 mutex_unlock(&vdev->vf_token->lock);
1817 pci_info_ratelimited(vdev->pdev,
1818 "Incorrect VF token provided for device\n");
1821 } else if (vf_token) {
1822 uuid_copy(&vdev->vf_token->uuid, uuid);
1825 mutex_unlock(&vdev->vf_token->lock);
1826 } else if (vf_token) {
1827 pci_info_ratelimited(vdev->pdev,
1828 "VF token incorrectly provided, not a PF or VF\n");
1835 #define VF_TOKEN_ARG "vf_token="
1837 static int vfio_pci_match(void *device_data, char *buf)
1839 struct vfio_pci_device *vdev = device_data;
1840 bool vf_token = false;
1844 if (strncmp(pci_name(vdev->pdev), buf, strlen(pci_name(vdev->pdev))))
1845 return 0; /* No match */
1847 if (strlen(buf) > strlen(pci_name(vdev->pdev))) {
1848 buf += strlen(pci_name(vdev->pdev));
1851 return 0; /* No match: non-whitespace after name */
1859 if (!vf_token && !strncmp(buf, VF_TOKEN_ARG,
1860 strlen(VF_TOKEN_ARG))) {
1861 buf += strlen(VF_TOKEN_ARG);
1863 if (strlen(buf) < UUID_STRING_LEN)
1866 ret = uuid_parse(buf, &uuid);
1871 buf += UUID_STRING_LEN;
1873 /* Unknown/duplicate option */
1879 ret = vfio_pci_validate_vf_token(vdev, vf_token, &uuid);
1883 return 1; /* Match */
1886 static const struct vfio_device_ops vfio_pci_ops = {
1888 .open = vfio_pci_open,
1889 .release = vfio_pci_release,
1890 .ioctl = vfio_pci_ioctl,
1891 .read = vfio_pci_read,
1892 .write = vfio_pci_write,
1893 .mmap = vfio_pci_mmap,
1894 .request = vfio_pci_request,
1895 .match = vfio_pci_match,
1898 static int vfio_pci_reflck_attach(struct vfio_pci_device *vdev);
1899 static void vfio_pci_reflck_put(struct vfio_pci_reflck *reflck);
1901 static int vfio_pci_bus_notifier(struct notifier_block *nb,
1902 unsigned long action, void *data)
1904 struct vfio_pci_device *vdev = container_of(nb,
1905 struct vfio_pci_device, nb);
1906 struct device *dev = data;
1907 struct pci_dev *pdev = to_pci_dev(dev);
1908 struct pci_dev *physfn = pci_physfn(pdev);
1910 if (action == BUS_NOTIFY_ADD_DEVICE &&
1911 pdev->is_virtfn && physfn == vdev->pdev) {
1912 pci_info(vdev->pdev, "Captured SR-IOV VF %s driver_override\n",
1914 pdev->driver_override = kasprintf(GFP_KERNEL, "%s",
1916 } else if (action == BUS_NOTIFY_BOUND_DRIVER &&
1917 pdev->is_virtfn && physfn == vdev->pdev) {
1918 struct pci_driver *drv = pci_dev_driver(pdev);
1920 if (drv && drv != &vfio_pci_driver)
1921 pci_warn(vdev->pdev,
1922 "VF %s bound to driver %s while PF bound to vfio-pci\n",
1923 pci_name(pdev), drv->name);
1929 static int vfio_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
1931 struct vfio_pci_device *vdev;
1932 struct iommu_group *group;
1935 if (vfio_pci_is_denylisted(pdev))
1938 if (pdev->hdr_type != PCI_HEADER_TYPE_NORMAL)
1942 * Prevent binding to PFs with VFs enabled, the VFs might be in use
1943 * by the host or other users. We cannot capture the VFs if they
1944 * already exist, nor can we track VF users. Disabling SR-IOV here
1945 * would initiate removing the VFs, which would unbind the driver,
1946 * which is prone to blocking if that VF is also in use by vfio-pci.
1947 * Just reject these PFs and let the user sort it out.
1949 if (pci_num_vf(pdev)) {
1950 pci_warn(pdev, "Cannot bind to PF with SR-IOV enabled\n");
1954 group = vfio_iommu_group_get(&pdev->dev);
1958 vdev = kzalloc(sizeof(*vdev), GFP_KERNEL);
1965 vdev->irq_type = VFIO_PCI_NUM_IRQS;
1966 mutex_init(&vdev->igate);
1967 spin_lock_init(&vdev->irqlock);
1968 mutex_init(&vdev->ioeventfds_lock);
1969 INIT_LIST_HEAD(&vdev->ioeventfds_list);
1970 mutex_init(&vdev->vma_lock);
1971 INIT_LIST_HEAD(&vdev->vma_list);
1972 init_rwsem(&vdev->memory_lock);
1974 ret = vfio_add_group_dev(&pdev->dev, &vfio_pci_ops, vdev);
1978 ret = vfio_pci_reflck_attach(vdev);
1980 goto out_del_group_dev;
1982 if (pdev->is_physfn) {
1983 vdev->vf_token = kzalloc(sizeof(*vdev->vf_token), GFP_KERNEL);
1984 if (!vdev->vf_token) {
1989 mutex_init(&vdev->vf_token->lock);
1990 uuid_gen(&vdev->vf_token->uuid);
1992 vdev->nb.notifier_call = vfio_pci_bus_notifier;
1993 ret = bus_register_notifier(&pci_bus_type, &vdev->nb);
1998 if (vfio_pci_is_vga(pdev)) {
1999 vga_client_register(pdev, vdev, NULL, vfio_pci_set_vga_decode);
2000 vga_set_legacy_decoding(pdev,
2001 vfio_pci_set_vga_decode(vdev, false));
2004 vfio_pci_probe_power_state(vdev);
2006 if (!disable_idle_d3) {
2008 * pci-core sets the device power state to an unknown value at
2009 * bootup and after being removed from a driver. The only
2010 * transition it allows from this unknown state is to D0, which
2011 * typically happens when a driver calls pci_enable_device().
2012 * We're not ready to enable the device yet, but we do want to
2013 * be able to get to D3. Therefore first do a D0 transition
2014 * before going to D3.
2016 vfio_pci_set_power_state(vdev, PCI_D0);
2017 vfio_pci_set_power_state(vdev, PCI_D3hot);
2023 kfree(vdev->vf_token);
2025 vfio_pci_reflck_put(vdev->reflck);
2027 vfio_del_group_dev(&pdev->dev);
2031 vfio_iommu_group_put(group, &pdev->dev);
2035 static void vfio_pci_remove(struct pci_dev *pdev)
2037 struct vfio_pci_device *vdev;
2039 pci_disable_sriov(pdev);
2041 vdev = vfio_del_group_dev(&pdev->dev);
2045 if (vdev->vf_token) {
2046 WARN_ON(vdev->vf_token->users);
2047 mutex_destroy(&vdev->vf_token->lock);
2048 kfree(vdev->vf_token);
2051 if (vdev->nb.notifier_call)
2052 bus_unregister_notifier(&pci_bus_type, &vdev->nb);
2054 vfio_pci_reflck_put(vdev->reflck);
2056 vfio_iommu_group_put(pdev->dev.iommu_group, &pdev->dev);
2057 kfree(vdev->region);
2058 mutex_destroy(&vdev->ioeventfds_lock);
2060 if (!disable_idle_d3)
2061 vfio_pci_set_power_state(vdev, PCI_D0);
2063 kfree(vdev->pm_save);
2066 if (vfio_pci_is_vga(pdev)) {
2067 vga_client_register(pdev, NULL, NULL, NULL);
2068 vga_set_legacy_decoding(pdev,
2069 VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM |
2070 VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM);
2074 static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
2075 pci_channel_state_t state)
2077 struct vfio_pci_device *vdev;
2078 struct vfio_device *device;
2080 device = vfio_device_get_from_dev(&pdev->dev);
2082 return PCI_ERS_RESULT_DISCONNECT;
2084 vdev = vfio_device_data(device);
2086 vfio_device_put(device);
2087 return PCI_ERS_RESULT_DISCONNECT;
2090 mutex_lock(&vdev->igate);
2092 if (vdev->err_trigger)
2093 eventfd_signal(vdev->err_trigger, 1);
2095 mutex_unlock(&vdev->igate);
2097 vfio_device_put(device);
2099 return PCI_ERS_RESULT_CAN_RECOVER;
2102 static int vfio_pci_sriov_configure(struct pci_dev *pdev, int nr_virtfn)
2104 struct vfio_pci_device *vdev;
2105 struct vfio_device *device;
2113 device = vfio_device_get_from_dev(&pdev->dev);
2117 vdev = vfio_device_data(device);
2119 vfio_device_put(device);
2124 pci_disable_sriov(pdev);
2126 ret = pci_enable_sriov(pdev, nr_virtfn);
2128 vfio_device_put(device);
2130 return ret < 0 ? ret : nr_virtfn;
2133 static const struct pci_error_handlers vfio_err_handlers = {
2134 .error_detected = vfio_pci_aer_err_detected,
2137 static struct pci_driver vfio_pci_driver = {
2139 .id_table = NULL, /* only dynamic ids */
2140 .probe = vfio_pci_probe,
2141 .remove = vfio_pci_remove,
2142 .sriov_configure = vfio_pci_sriov_configure,
2143 .err_handler = &vfio_err_handlers,
2146 static DEFINE_MUTEX(reflck_lock);
2148 static struct vfio_pci_reflck *vfio_pci_reflck_alloc(void)
2150 struct vfio_pci_reflck *reflck;
2152 reflck = kzalloc(sizeof(*reflck), GFP_KERNEL);
2154 return ERR_PTR(-ENOMEM);
2156 kref_init(&reflck->kref);
2157 mutex_init(&reflck->lock);
2162 static void vfio_pci_reflck_get(struct vfio_pci_reflck *reflck)
2164 kref_get(&reflck->kref);
2167 static int vfio_pci_reflck_find(struct pci_dev *pdev, void *data)
2169 struct vfio_pci_reflck **preflck = data;
2170 struct vfio_device *device;
2171 struct vfio_pci_device *vdev;
2173 device = vfio_device_get_from_dev(&pdev->dev);
2177 if (pci_dev_driver(pdev) != &vfio_pci_driver) {
2178 vfio_device_put(device);
2182 vdev = vfio_device_data(device);
2185 vfio_pci_reflck_get(vdev->reflck);
2186 *preflck = vdev->reflck;
2187 vfio_device_put(device);
2191 vfio_device_put(device);
2195 static int vfio_pci_reflck_attach(struct vfio_pci_device *vdev)
2197 bool slot = !pci_probe_reset_slot(vdev->pdev->slot);
2199 mutex_lock(&reflck_lock);
2201 if (pci_is_root_bus(vdev->pdev->bus) ||
2202 vfio_pci_for_each_slot_or_bus(vdev->pdev, vfio_pci_reflck_find,
2203 &vdev->reflck, slot) <= 0)
2204 vdev->reflck = vfio_pci_reflck_alloc();
2206 mutex_unlock(&reflck_lock);
2208 return PTR_ERR_OR_ZERO(vdev->reflck);
2211 static void vfio_pci_reflck_release(struct kref *kref)
2213 struct vfio_pci_reflck *reflck = container_of(kref,
2214 struct vfio_pci_reflck,
2218 mutex_unlock(&reflck_lock);
2221 static void vfio_pci_reflck_put(struct vfio_pci_reflck *reflck)
2223 kref_put_mutex(&reflck->kref, vfio_pci_reflck_release, &reflck_lock);
2226 static int vfio_pci_get_unused_devs(struct pci_dev *pdev, void *data)
2228 struct vfio_devices *devs = data;
2229 struct vfio_device *device;
2230 struct vfio_pci_device *vdev;
2232 if (devs->cur_index == devs->max_index)
2235 device = vfio_device_get_from_dev(&pdev->dev);
2239 if (pci_dev_driver(pdev) != &vfio_pci_driver) {
2240 vfio_device_put(device);
2244 vdev = vfio_device_data(device);
2246 /* Fault if the device is not unused */
2248 vfio_device_put(device);
2252 devs->devices[devs->cur_index++] = device;
2256 static int vfio_pci_try_zap_and_vma_lock_cb(struct pci_dev *pdev, void *data)
2258 struct vfio_devices *devs = data;
2259 struct vfio_device *device;
2260 struct vfio_pci_device *vdev;
2262 if (devs->cur_index == devs->max_index)
2265 device = vfio_device_get_from_dev(&pdev->dev);
2269 if (pci_dev_driver(pdev) != &vfio_pci_driver) {
2270 vfio_device_put(device);
2274 vdev = vfio_device_data(device);
2277 * Locking multiple devices is prone to deadlock, runaway and
2278 * unwind if we hit contention.
2280 if (!vfio_pci_zap_and_vma_lock(vdev, true)) {
2281 vfio_device_put(device);
2285 devs->devices[devs->cur_index++] = device;
2290 * If a bus or slot reset is available for the provided device and:
2291 * - All of the devices affected by that bus or slot reset are unused
2293 * - At least one of the affected devices is marked dirty via
2294 * needs_reset (such as by lack of FLR support)
2295 * Then attempt to perform that bus or slot reset. Callers are required
2296 * to hold vdev->reflck->lock, protecting the bus/slot reset group from
2297 * concurrent opens. A vfio_device reference is acquired for each device
2298 * to prevent unbinds during the reset operation.
2300 * NB: vfio-core considers a group to be viable even if some devices are
2301 * bound to drivers like pci-stub or pcieport. Here we require all devices
2302 * to be bound to vfio_pci since that's the only way we can be sure they
2305 static void vfio_pci_try_bus_reset(struct vfio_pci_device *vdev)
2307 struct vfio_devices devs = { .cur_index = 0 };
2308 int i = 0, ret = -EINVAL;
2310 struct vfio_pci_device *tmp;
2312 if (!pci_probe_reset_slot(vdev->pdev->slot))
2314 else if (pci_probe_reset_bus(vdev->pdev->bus))
2317 if (vfio_pci_for_each_slot_or_bus(vdev->pdev, vfio_pci_count_devs,
2322 devs.devices = kcalloc(i, sizeof(struct vfio_device *), GFP_KERNEL);
2326 if (vfio_pci_for_each_slot_or_bus(vdev->pdev,
2327 vfio_pci_get_unused_devs,
2331 /* Does at least one need a reset? */
2332 for (i = 0; i < devs.cur_index; i++) {
2333 tmp = vfio_device_data(devs.devices[i]);
2334 if (tmp->needs_reset) {
2335 ret = pci_reset_bus(vdev->pdev);
2341 for (i = 0; i < devs.cur_index; i++) {
2342 tmp = vfio_device_data(devs.devices[i]);
2345 * If reset was successful, affected devices no longer need
2346 * a reset and we should return all the collateral devices
2347 * to low power. If not successful, we either didn't reset
2348 * the bus or timed out waiting for it, so let's not touch
2352 tmp->needs_reset = false;
2354 if (tmp != vdev && !disable_idle_d3)
2355 vfio_pci_set_power_state(tmp, PCI_D3hot);
2358 vfio_device_put(devs.devices[i]);
2361 kfree(devs.devices);
2364 static void __exit vfio_pci_cleanup(void)
2366 pci_unregister_driver(&vfio_pci_driver);
2367 vfio_pci_uninit_perm_bits();
2370 static void __init vfio_pci_fill_ids(void)
2375 /* no ids passed actually */
2379 /* add ids specified in the module parameter */
2381 while ((id = strsep(&p, ","))) {
2382 unsigned int vendor, device, subvendor = PCI_ANY_ID,
2383 subdevice = PCI_ANY_ID, class = 0, class_mask = 0;
2389 fields = sscanf(id, "%x:%x:%x:%x:%x:%x",
2390 &vendor, &device, &subvendor, &subdevice,
2391 &class, &class_mask);
2394 pr_warn("invalid id string \"%s\"\n", id);
2398 rc = pci_add_dynid(&vfio_pci_driver, vendor, device,
2399 subvendor, subdevice, class, class_mask, 0);
2401 pr_warn("failed to add dynamic id [%04x:%04x[%04x:%04x]] class %#08x/%08x (%d)\n",
2402 vendor, device, subvendor, subdevice,
2403 class, class_mask, rc);
2405 pr_info("add [%04x:%04x[%04x:%04x]] class %#08x/%08x\n",
2406 vendor, device, subvendor, subdevice,
2411 static int __init vfio_pci_init(void)
2415 /* Allocate shared config space permision data used by all devices */
2416 ret = vfio_pci_init_perm_bits();
2420 /* Register and scan for devices */
2421 ret = pci_register_driver(&vfio_pci_driver);
2425 vfio_pci_fill_ids();
2427 if (disable_denylist)
2428 pr_warn("device denylist disabled.\n");
2433 vfio_pci_uninit_perm_bits();
2437 module_init(vfio_pci_init);
2438 module_exit(vfio_pci_cleanup);
2440 MODULE_VERSION(DRIVER_VERSION);
2441 MODULE_LICENSE("GPL v2");
2442 MODULE_AUTHOR(DRIVER_AUTHOR);
2443 MODULE_DESCRIPTION(DRIVER_DESC);