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 for (i = 0; i < PCI_STD_NUM_BARS; i++) {
165 int bar = i + PCI_STD_RESOURCES;
167 res = &vdev->pdev->resource[bar];
169 if (!IS_ENABLED(CONFIG_VFIO_PCI_MMAP))
172 if (!(res->flags & IORESOURCE_MEM))
176 * The PCI core shouldn't set up a resource with a
177 * type but zero size. But there may be bugs that
178 * cause us to do that.
180 if (!resource_size(res))
183 if (resource_size(res) >= PAGE_SIZE) {
184 vdev->bar_mmap_supported[bar] = true;
188 if (!(res->start & ~PAGE_MASK)) {
190 * Add a dummy resource to reserve the remainder
191 * of the exclusive page in case that hot-add
192 * device's bar is assigned into it.
194 dummy_res = kzalloc(sizeof(*dummy_res), GFP_KERNEL);
195 if (dummy_res == NULL)
198 dummy_res->resource.name = "vfio sub-page reserved";
199 dummy_res->resource.start = res->end + 1;
200 dummy_res->resource.end = res->start + PAGE_SIZE - 1;
201 dummy_res->resource.flags = res->flags;
202 if (request_resource(res->parent,
203 &dummy_res->resource)) {
207 dummy_res->index = bar;
208 list_add(&dummy_res->res_next,
209 &vdev->dummy_resources_list);
210 vdev->bar_mmap_supported[bar] = true;
214 * Here we don't handle the case when the BAR is not page
215 * aligned because we can't expect the BAR will be
216 * assigned into the same location in a page in guest
217 * when we passthrough the BAR. And it's hard to access
218 * this BAR in userspace because we have no way to get
219 * the BAR's location in a page.
222 vdev->bar_mmap_supported[bar] = false;
226 static void vfio_pci_try_bus_reset(struct vfio_pci_device *vdev);
227 static void vfio_pci_disable(struct vfio_pci_device *vdev);
228 static int vfio_pci_try_zap_and_vma_lock_cb(struct pci_dev *pdev, void *data);
231 * INTx masking requires the ability to disable INTx signaling via PCI_COMMAND
232 * _and_ the ability detect when the device is asserting INTx via PCI_STATUS.
233 * If a device implements the former but not the latter we would typically
234 * expect broken_intx_masking be set and require an exclusive interrupt.
235 * However since we do have control of the device's ability to assert INTx,
236 * we can instead pretend that the device does not implement INTx, virtualizing
237 * the pin register to report zero and maintaining DisINTx set on the host.
239 static bool vfio_pci_nointx(struct pci_dev *pdev)
241 switch (pdev->vendor) {
242 case PCI_VENDOR_ID_INTEL:
243 switch (pdev->device) {
244 /* All i40e (XL710/X710/XXV710) 10/20/25/40GbE NICs */
247 case 0x1580 ... 0x1581:
248 case 0x1583 ... 0x158b:
249 case 0x37d0 ... 0x37d2:
261 static void vfio_pci_probe_power_state(struct vfio_pci_device *vdev)
263 struct pci_dev *pdev = vdev->pdev;
269 pci_read_config_word(pdev, pdev->pm_cap + PCI_PM_CTRL, &pmcsr);
271 vdev->needs_pm_restore = !(pmcsr & PCI_PM_CTRL_NO_SOFT_RESET);
275 * pci_set_power_state() wrapper handling devices which perform a soft reset on
276 * D3->D0 transition. Save state prior to D0/1/2->D3, stash it on the vdev,
277 * restore when returned to D0. Saved separately from pci_saved_state for use
278 * by PM capability emulation and separately from pci_dev internal saved state
279 * to avoid it being overwritten and consumed around other resets.
281 int vfio_pci_set_power_state(struct vfio_pci_device *vdev, pci_power_t state)
283 struct pci_dev *pdev = vdev->pdev;
284 bool needs_restore = false, needs_save = false;
287 if (vdev->needs_pm_restore) {
288 if (pdev->current_state < PCI_D3hot && state >= PCI_D3hot) {
289 pci_save_state(pdev);
293 if (pdev->current_state >= PCI_D3hot && state <= PCI_D0)
294 needs_restore = true;
297 ret = pci_set_power_state(pdev, state);
300 /* D3 might be unsupported via quirk, skip unless in D3 */
301 if (needs_save && pdev->current_state >= PCI_D3hot) {
302 vdev->pm_save = pci_store_saved_state(pdev);
303 } else if (needs_restore) {
304 pci_load_and_free_saved_state(pdev, &vdev->pm_save);
305 pci_restore_state(pdev);
312 static int vfio_pci_enable(struct vfio_pci_device *vdev)
314 struct pci_dev *pdev = vdev->pdev;
319 vfio_pci_set_power_state(vdev, PCI_D0);
321 /* Don't allow our initial saved state to include busmaster */
322 pci_clear_master(pdev);
324 ret = pci_enable_device(pdev);
328 /* If reset fails because of the device lock, fail this path entirely */
329 ret = pci_try_reset_function(pdev);
330 if (ret == -EAGAIN) {
331 pci_disable_device(pdev);
335 vdev->reset_works = !ret;
336 pci_save_state(pdev);
337 vdev->pci_saved_state = pci_store_saved_state(pdev);
338 if (!vdev->pci_saved_state)
339 pci_dbg(pdev, "%s: Couldn't store saved state\n", __func__);
341 if (likely(!nointxmask)) {
342 if (vfio_pci_nointx(pdev)) {
343 pci_info(pdev, "Masking broken INTx support\n");
347 vdev->pci_2_3 = pci_intx_mask_supported(pdev);
350 pci_read_config_word(pdev, PCI_COMMAND, &cmd);
351 if (vdev->pci_2_3 && (cmd & PCI_COMMAND_INTX_DISABLE)) {
352 cmd &= ~PCI_COMMAND_INTX_DISABLE;
353 pci_write_config_word(pdev, PCI_COMMAND, cmd);
356 ret = vfio_config_init(vdev);
358 kfree(vdev->pci_saved_state);
359 vdev->pci_saved_state = NULL;
360 pci_disable_device(pdev);
364 msix_pos = pdev->msix_cap;
369 pci_read_config_word(pdev, msix_pos + PCI_MSIX_FLAGS, &flags);
370 pci_read_config_dword(pdev, msix_pos + PCI_MSIX_TABLE, &table);
372 vdev->msix_bar = table & PCI_MSIX_TABLE_BIR;
373 vdev->msix_offset = table & PCI_MSIX_TABLE_OFFSET;
374 vdev->msix_size = ((flags & PCI_MSIX_FLAGS_QSIZE) + 1) * 16;
376 vdev->msix_bar = 0xFF;
378 if (!vfio_vga_disabled() && vfio_pci_is_vga(pdev))
379 vdev->has_vga = true;
381 if (vfio_pci_is_vga(pdev) &&
382 pdev->vendor == PCI_VENDOR_ID_INTEL &&
383 IS_ENABLED(CONFIG_VFIO_PCI_IGD)) {
384 ret = vfio_pci_igd_init(vdev);
385 if (ret && ret != -ENODEV) {
386 pci_warn(pdev, "Failed to setup Intel IGD regions\n");
391 vfio_pci_probe_mmaps(vdev);
396 vfio_pci_disable(vdev);
400 static void vfio_pci_disable(struct vfio_pci_device *vdev)
402 struct pci_dev *pdev = vdev->pdev;
403 struct vfio_pci_dummy_resource *dummy_res, *tmp;
404 struct vfio_pci_ioeventfd *ioeventfd, *ioeventfd_tmp;
407 /* Stop the device from further DMA */
408 pci_clear_master(pdev);
410 vfio_pci_set_irqs_ioctl(vdev, VFIO_IRQ_SET_DATA_NONE |
411 VFIO_IRQ_SET_ACTION_TRIGGER,
412 vdev->irq_type, 0, 0, NULL);
414 /* Device closed, don't need mutex here */
415 list_for_each_entry_safe(ioeventfd, ioeventfd_tmp,
416 &vdev->ioeventfds_list, next) {
417 vfio_virqfd_disable(&ioeventfd->virqfd);
418 list_del(&ioeventfd->next);
421 vdev->ioeventfds_nr = 0;
423 vdev->virq_disabled = false;
425 for (i = 0; i < vdev->num_regions; i++)
426 vdev->region[i].ops->release(vdev, &vdev->region[i]);
428 vdev->num_regions = 0;
430 vdev->region = NULL; /* don't krealloc a freed pointer */
432 vfio_config_free(vdev);
434 for (i = 0; i < PCI_STD_NUM_BARS; i++) {
435 bar = i + PCI_STD_RESOURCES;
436 if (!vdev->barmap[bar])
438 pci_iounmap(pdev, vdev->barmap[bar]);
439 pci_release_selected_regions(pdev, 1 << bar);
440 vdev->barmap[bar] = NULL;
443 list_for_each_entry_safe(dummy_res, tmp,
444 &vdev->dummy_resources_list, res_next) {
445 list_del(&dummy_res->res_next);
446 release_resource(&dummy_res->resource);
450 vdev->needs_reset = true;
453 * If we have saved state, restore it. If we can reset the device,
454 * even better. Resetting with current state seems better than
455 * nothing, but saving and restoring current state without reset
458 if (pci_load_and_free_saved_state(pdev, &vdev->pci_saved_state)) {
459 pci_info(pdev, "%s: Couldn't reload saved state\n", __func__);
461 if (!vdev->reset_works)
464 pci_save_state(pdev);
468 * Disable INTx and MSI, presumably to avoid spurious interrupts
469 * during reset. Stolen from pci_reset_function()
471 pci_write_config_word(pdev, PCI_COMMAND, PCI_COMMAND_INTX_DISABLE);
474 * Try to get the locks ourselves to prevent a deadlock. The
475 * success of this is dependent on being able to lock the device,
476 * which is not always possible.
477 * We can not use the "try" reset interface here, which will
478 * overwrite the previously restored configuration information.
480 if (vdev->reset_works && pci_dev_trylock(pdev)) {
481 if (!__pci_reset_function_locked(pdev))
482 vdev->needs_reset = false;
483 pci_dev_unlock(pdev);
486 pci_restore_state(pdev);
488 pci_disable_device(pdev);
490 vfio_pci_try_bus_reset(vdev);
492 if (!disable_idle_d3)
493 vfio_pci_set_power_state(vdev, PCI_D3hot);
496 static struct pci_driver vfio_pci_driver;
498 static struct vfio_pci_device *get_pf_vdev(struct vfio_pci_device *vdev)
500 struct pci_dev *physfn = pci_physfn(vdev->pdev);
501 struct vfio_device *pf_dev;
503 if (!vdev->pdev->is_virtfn)
506 pf_dev = vfio_device_get_from_dev(&physfn->dev);
510 if (pci_dev_driver(physfn) != &vfio_pci_driver) {
511 vfio_device_put(pf_dev);
515 return container_of(pf_dev, struct vfio_pci_device, vdev);
518 static void vfio_pci_vf_token_user_add(struct vfio_pci_device *vdev, int val)
520 struct vfio_pci_device *pf_vdev = get_pf_vdev(vdev);
525 mutex_lock(&pf_vdev->vf_token->lock);
526 pf_vdev->vf_token->users += val;
527 WARN_ON(pf_vdev->vf_token->users < 0);
528 mutex_unlock(&pf_vdev->vf_token->lock);
530 vfio_device_put(&pf_vdev->vdev);
533 static void vfio_pci_release(struct vfio_device *core_vdev)
535 struct vfio_pci_device *vdev =
536 container_of(core_vdev, struct vfio_pci_device, vdev);
538 mutex_lock(&vdev->reflck->lock);
540 if (!(--vdev->refcnt)) {
541 vfio_pci_vf_token_user_add(vdev, -1);
542 vfio_spapr_pci_eeh_release(vdev->pdev);
543 vfio_pci_disable(vdev);
545 mutex_lock(&vdev->igate);
546 if (vdev->err_trigger) {
547 eventfd_ctx_put(vdev->err_trigger);
548 vdev->err_trigger = NULL;
550 if (vdev->req_trigger) {
551 eventfd_ctx_put(vdev->req_trigger);
552 vdev->req_trigger = NULL;
554 mutex_unlock(&vdev->igate);
557 mutex_unlock(&vdev->reflck->lock);
560 static int vfio_pci_open(struct vfio_device *core_vdev)
562 struct vfio_pci_device *vdev =
563 container_of(core_vdev, struct vfio_pci_device, vdev);
566 mutex_lock(&vdev->reflck->lock);
569 ret = vfio_pci_enable(vdev);
573 vfio_spapr_pci_eeh_open(vdev->pdev);
574 vfio_pci_vf_token_user_add(vdev, 1);
578 mutex_unlock(&vdev->reflck->lock);
582 static int vfio_pci_get_irq_count(struct vfio_pci_device *vdev, int irq_type)
584 if (irq_type == VFIO_PCI_INTX_IRQ_INDEX) {
587 if (!IS_ENABLED(CONFIG_VFIO_PCI_INTX) ||
588 vdev->nointx || vdev->pdev->is_virtfn)
591 pci_read_config_byte(vdev->pdev, PCI_INTERRUPT_PIN, &pin);
594 } else if (irq_type == VFIO_PCI_MSI_IRQ_INDEX) {
598 pos = vdev->pdev->msi_cap;
600 pci_read_config_word(vdev->pdev,
601 pos + PCI_MSI_FLAGS, &flags);
602 return 1 << ((flags & PCI_MSI_FLAGS_QMASK) >> 1);
604 } else if (irq_type == VFIO_PCI_MSIX_IRQ_INDEX) {
608 pos = vdev->pdev->msix_cap;
610 pci_read_config_word(vdev->pdev,
611 pos + PCI_MSIX_FLAGS, &flags);
613 return (flags & PCI_MSIX_FLAGS_QSIZE) + 1;
615 } else if (irq_type == VFIO_PCI_ERR_IRQ_INDEX) {
616 if (pci_is_pcie(vdev->pdev))
618 } else if (irq_type == VFIO_PCI_REQ_IRQ_INDEX) {
625 static int vfio_pci_count_devs(struct pci_dev *pdev, void *data)
631 struct vfio_pci_fill_info {
634 struct vfio_pci_dependent_device *devices;
637 static int vfio_pci_fill_devs(struct pci_dev *pdev, void *data)
639 struct vfio_pci_fill_info *fill = data;
640 struct iommu_group *iommu_group;
642 if (fill->cur == fill->max)
643 return -EAGAIN; /* Something changed, try again */
645 iommu_group = iommu_group_get(&pdev->dev);
647 return -EPERM; /* Cannot reset non-isolated devices */
649 fill->devices[fill->cur].group_id = iommu_group_id(iommu_group);
650 fill->devices[fill->cur].segment = pci_domain_nr(pdev->bus);
651 fill->devices[fill->cur].bus = pdev->bus->number;
652 fill->devices[fill->cur].devfn = pdev->devfn;
654 iommu_group_put(iommu_group);
658 struct vfio_pci_group_entry {
659 struct vfio_group *group;
663 struct vfio_pci_group_info {
665 struct vfio_pci_group_entry *groups;
668 static int vfio_pci_validate_devs(struct pci_dev *pdev, void *data)
670 struct vfio_pci_group_info *info = data;
671 struct iommu_group *group;
674 group = iommu_group_get(&pdev->dev);
678 id = iommu_group_id(group);
680 for (i = 0; i < info->count; i++)
681 if (info->groups[i].id == id)
684 iommu_group_put(group);
686 return (i == info->count) ? -EINVAL : 0;
689 static bool vfio_pci_dev_below_slot(struct pci_dev *pdev, struct pci_slot *slot)
691 for (; pdev; pdev = pdev->bus->self)
692 if (pdev->bus == slot->bus)
693 return (pdev->slot == slot);
697 struct vfio_pci_walk_info {
698 int (*fn)(struct pci_dev *, void *data);
700 struct pci_dev *pdev;
705 static int vfio_pci_walk_wrapper(struct pci_dev *pdev, void *data)
707 struct vfio_pci_walk_info *walk = data;
709 if (!walk->slot || vfio_pci_dev_below_slot(pdev, walk->pdev->slot))
710 walk->ret = walk->fn(pdev, walk->data);
715 static int vfio_pci_for_each_slot_or_bus(struct pci_dev *pdev,
716 int (*fn)(struct pci_dev *,
717 void *data), void *data,
720 struct vfio_pci_walk_info walk = {
721 .fn = fn, .data = data, .pdev = pdev, .slot = slot, .ret = 0,
724 pci_walk_bus(pdev->bus, vfio_pci_walk_wrapper, &walk);
729 static int msix_mmappable_cap(struct vfio_pci_device *vdev,
730 struct vfio_info_cap *caps)
732 struct vfio_info_cap_header header = {
733 .id = VFIO_REGION_INFO_CAP_MSIX_MAPPABLE,
737 return vfio_info_add_capability(caps, &header, sizeof(header));
740 int vfio_pci_register_dev_region(struct vfio_pci_device *vdev,
741 unsigned int type, unsigned int subtype,
742 const struct vfio_pci_regops *ops,
743 size_t size, u32 flags, void *data)
745 struct vfio_pci_region *region;
747 region = krealloc(vdev->region,
748 (vdev->num_regions + 1) * sizeof(*region),
753 vdev->region = region;
754 vdev->region[vdev->num_regions].type = type;
755 vdev->region[vdev->num_regions].subtype = subtype;
756 vdev->region[vdev->num_regions].ops = ops;
757 vdev->region[vdev->num_regions].size = size;
758 vdev->region[vdev->num_regions].flags = flags;
759 vdev->region[vdev->num_regions].data = data;
766 struct vfio_devices {
767 struct vfio_pci_device **devices;
772 static long vfio_pci_ioctl(struct vfio_device *core_vdev,
773 unsigned int cmd, unsigned long arg)
775 struct vfio_pci_device *vdev =
776 container_of(core_vdev, struct vfio_pci_device, vdev);
779 if (cmd == VFIO_DEVICE_GET_INFO) {
780 struct vfio_device_info info;
781 struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
785 minsz = offsetofend(struct vfio_device_info, num_irqs);
787 /* For backward compatibility, cannot require this */
788 capsz = offsetofend(struct vfio_iommu_type1_info, cap_offset);
790 if (copy_from_user(&info, (void __user *)arg, minsz))
793 if (info.argsz < minsz)
796 if (info.argsz >= capsz) {
801 info.flags = VFIO_DEVICE_FLAGS_PCI;
803 if (vdev->reset_works)
804 info.flags |= VFIO_DEVICE_FLAGS_RESET;
806 info.num_regions = VFIO_PCI_NUM_REGIONS + vdev->num_regions;
807 info.num_irqs = VFIO_PCI_NUM_IRQS;
809 ret = vfio_pci_info_zdev_add_caps(vdev, &caps);
810 if (ret && ret != -ENODEV) {
811 pci_warn(vdev->pdev, "Failed to setup zPCI info capabilities\n");
816 info.flags |= VFIO_DEVICE_FLAGS_CAPS;
817 if (info.argsz < sizeof(info) + caps.size) {
818 info.argsz = sizeof(info) + caps.size;
820 vfio_info_cap_shift(&caps, sizeof(info));
821 if (copy_to_user((void __user *)arg +
822 sizeof(info), caps.buf,
827 info.cap_offset = sizeof(info);
833 return copy_to_user((void __user *)arg, &info, minsz) ?
836 } else if (cmd == VFIO_DEVICE_GET_REGION_INFO) {
837 struct pci_dev *pdev = vdev->pdev;
838 struct vfio_region_info info;
839 struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
842 minsz = offsetofend(struct vfio_region_info, offset);
844 if (copy_from_user(&info, (void __user *)arg, minsz))
847 if (info.argsz < minsz)
850 switch (info.index) {
851 case VFIO_PCI_CONFIG_REGION_INDEX:
852 info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
853 info.size = pdev->cfg_size;
854 info.flags = VFIO_REGION_INFO_FLAG_READ |
855 VFIO_REGION_INFO_FLAG_WRITE;
857 case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
858 info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
859 info.size = pci_resource_len(pdev, info.index);
865 info.flags = VFIO_REGION_INFO_FLAG_READ |
866 VFIO_REGION_INFO_FLAG_WRITE;
867 if (vdev->bar_mmap_supported[info.index]) {
868 info.flags |= VFIO_REGION_INFO_FLAG_MMAP;
869 if (info.index == vdev->msix_bar) {
870 ret = msix_mmappable_cap(vdev, &caps);
877 case VFIO_PCI_ROM_REGION_INDEX:
883 info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
886 /* Report the BAR size, not the ROM size */
887 info.size = pci_resource_len(pdev, info.index);
889 /* Shadow ROMs appear as PCI option ROMs */
890 if (pdev->resource[PCI_ROM_RESOURCE].flags &
891 IORESOURCE_ROM_SHADOW)
898 * Is it really there? Enable memory decode for
899 * implicit access in pci_map_rom().
901 cmd = vfio_pci_memory_lock_and_enable(vdev);
902 io = pci_map_rom(pdev, &size);
904 info.flags = VFIO_REGION_INFO_FLAG_READ;
905 pci_unmap_rom(pdev, io);
909 vfio_pci_memory_unlock_and_restore(vdev, cmd);
913 case VFIO_PCI_VGA_REGION_INDEX:
917 info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
919 info.flags = VFIO_REGION_INFO_FLAG_READ |
920 VFIO_REGION_INFO_FLAG_WRITE;
925 struct vfio_region_info_cap_type cap_type = {
926 .header.id = VFIO_REGION_INFO_CAP_TYPE,
927 .header.version = 1 };
930 VFIO_PCI_NUM_REGIONS + vdev->num_regions)
932 info.index = array_index_nospec(info.index,
933 VFIO_PCI_NUM_REGIONS +
936 i = info.index - VFIO_PCI_NUM_REGIONS;
938 info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
939 info.size = vdev->region[i].size;
940 info.flags = vdev->region[i].flags;
942 cap_type.type = vdev->region[i].type;
943 cap_type.subtype = vdev->region[i].subtype;
945 ret = vfio_info_add_capability(&caps, &cap_type.header,
950 if (vdev->region[i].ops->add_capability) {
951 ret = vdev->region[i].ops->add_capability(vdev,
952 &vdev->region[i], &caps);
960 info.flags |= VFIO_REGION_INFO_FLAG_CAPS;
961 if (info.argsz < sizeof(info) + caps.size) {
962 info.argsz = sizeof(info) + caps.size;
965 vfio_info_cap_shift(&caps, sizeof(info));
966 if (copy_to_user((void __user *)arg +
967 sizeof(info), caps.buf,
972 info.cap_offset = sizeof(info);
978 return copy_to_user((void __user *)arg, &info, minsz) ?
981 } else if (cmd == VFIO_DEVICE_GET_IRQ_INFO) {
982 struct vfio_irq_info info;
984 minsz = offsetofend(struct vfio_irq_info, count);
986 if (copy_from_user(&info, (void __user *)arg, minsz))
989 if (info.argsz < minsz || info.index >= VFIO_PCI_NUM_IRQS)
992 switch (info.index) {
993 case VFIO_PCI_INTX_IRQ_INDEX ... VFIO_PCI_MSIX_IRQ_INDEX:
994 case VFIO_PCI_REQ_IRQ_INDEX:
996 case VFIO_PCI_ERR_IRQ_INDEX:
997 if (pci_is_pcie(vdev->pdev))
1004 info.flags = VFIO_IRQ_INFO_EVENTFD;
1006 info.count = vfio_pci_get_irq_count(vdev, info.index);
1008 if (info.index == VFIO_PCI_INTX_IRQ_INDEX)
1009 info.flags |= (VFIO_IRQ_INFO_MASKABLE |
1010 VFIO_IRQ_INFO_AUTOMASKED);
1012 info.flags |= VFIO_IRQ_INFO_NORESIZE;
1014 return copy_to_user((void __user *)arg, &info, minsz) ?
1017 } else if (cmd == VFIO_DEVICE_SET_IRQS) {
1018 struct vfio_irq_set hdr;
1021 size_t data_size = 0;
1023 minsz = offsetofend(struct vfio_irq_set, count);
1025 if (copy_from_user(&hdr, (void __user *)arg, minsz))
1028 max = vfio_pci_get_irq_count(vdev, hdr.index);
1030 ret = vfio_set_irqs_validate_and_prepare(&hdr, max,
1031 VFIO_PCI_NUM_IRQS, &data_size);
1036 data = memdup_user((void __user *)(arg + minsz),
1039 return PTR_ERR(data);
1042 mutex_lock(&vdev->igate);
1044 ret = vfio_pci_set_irqs_ioctl(vdev, hdr.flags, hdr.index,
1045 hdr.start, hdr.count, data);
1047 mutex_unlock(&vdev->igate);
1052 } else if (cmd == VFIO_DEVICE_RESET) {
1055 if (!vdev->reset_works)
1058 vfio_pci_zap_and_down_write_memory_lock(vdev);
1059 ret = pci_try_reset_function(vdev->pdev);
1060 up_write(&vdev->memory_lock);
1064 } else if (cmd == VFIO_DEVICE_GET_PCI_HOT_RESET_INFO) {
1065 struct vfio_pci_hot_reset_info hdr;
1066 struct vfio_pci_fill_info fill = { 0 };
1067 struct vfio_pci_dependent_device *devices = NULL;
1071 minsz = offsetofend(struct vfio_pci_hot_reset_info, count);
1073 if (copy_from_user(&hdr, (void __user *)arg, minsz))
1076 if (hdr.argsz < minsz)
1081 /* Can we do a slot or bus reset or neither? */
1082 if (!pci_probe_reset_slot(vdev->pdev->slot))
1084 else if (pci_probe_reset_bus(vdev->pdev->bus))
1087 /* How many devices are affected? */
1088 ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1089 vfio_pci_count_devs,
1094 WARN_ON(!fill.max); /* Should always be at least one */
1097 * If there's enough space, fill it now, otherwise return
1098 * -ENOSPC and the number of devices affected.
1100 if (hdr.argsz < sizeof(hdr) + (fill.max * sizeof(*devices))) {
1102 hdr.count = fill.max;
1103 goto reset_info_exit;
1106 devices = kcalloc(fill.max, sizeof(*devices), GFP_KERNEL);
1110 fill.devices = devices;
1112 ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1117 * If a device was removed between counting and filling,
1118 * we may come up short of fill.max. If a device was
1119 * added, we'll have a return of -EAGAIN above.
1122 hdr.count = fill.cur;
1125 if (copy_to_user((void __user *)arg, &hdr, minsz))
1129 if (copy_to_user((void __user *)(arg + minsz), devices,
1130 hdr.count * sizeof(*devices)))
1137 } else if (cmd == VFIO_DEVICE_PCI_HOT_RESET) {
1138 struct vfio_pci_hot_reset hdr;
1140 struct vfio_pci_group_entry *groups;
1141 struct vfio_pci_group_info info;
1142 struct vfio_devices devs = { .cur_index = 0 };
1144 int i, group_idx, mem_idx = 0, count = 0, ret = 0;
1146 minsz = offsetofend(struct vfio_pci_hot_reset, count);
1148 if (copy_from_user(&hdr, (void __user *)arg, minsz))
1151 if (hdr.argsz < minsz || hdr.flags)
1154 /* Can we do a slot or bus reset or neither? */
1155 if (!pci_probe_reset_slot(vdev->pdev->slot))
1157 else if (pci_probe_reset_bus(vdev->pdev->bus))
1161 * We can't let userspace give us an arbitrarily large
1162 * buffer to copy, so verify how many we think there
1163 * could be. Note groups can have multiple devices so
1164 * one group per device is the max.
1166 ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1167 vfio_pci_count_devs,
1172 /* Somewhere between 1 and count is OK */
1173 if (!hdr.count || hdr.count > count)
1176 group_fds = kcalloc(hdr.count, sizeof(*group_fds), GFP_KERNEL);
1177 groups = kcalloc(hdr.count, sizeof(*groups), GFP_KERNEL);
1178 if (!group_fds || !groups) {
1184 if (copy_from_user(group_fds, (void __user *)(arg + minsz),
1185 hdr.count * sizeof(*group_fds))) {
1192 * For each group_fd, get the group through the vfio external
1193 * user interface and store the group and iommu ID. This
1194 * ensures the group is held across the reset.
1196 for (group_idx = 0; group_idx < hdr.count; group_idx++) {
1197 struct vfio_group *group;
1198 struct fd f = fdget(group_fds[group_idx]);
1204 group = vfio_group_get_external_user(f.file);
1206 if (IS_ERR(group)) {
1207 ret = PTR_ERR(group);
1211 groups[group_idx].group = group;
1212 groups[group_idx].id =
1213 vfio_external_user_iommu_id(group);
1218 /* release reference to groups on error */
1220 goto hot_reset_release;
1222 info.count = hdr.count;
1223 info.groups = groups;
1226 * Test whether all the affected devices are contained
1227 * by the set of groups provided by the user.
1229 ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1230 vfio_pci_validate_devs,
1233 goto hot_reset_release;
1235 devs.max_index = count;
1236 devs.devices = kcalloc(count, sizeof(struct vfio_device *),
1238 if (!devs.devices) {
1240 goto hot_reset_release;
1244 * We need to get memory_lock for each device, but devices
1245 * can share mmap_lock, therefore we need to zap and hold
1246 * the vma_lock for each device, and only then get each
1249 ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1250 vfio_pci_try_zap_and_vma_lock_cb,
1253 goto hot_reset_release;
1255 for (; mem_idx < devs.cur_index; mem_idx++) {
1256 struct vfio_pci_device *tmp = devs.devices[mem_idx];
1258 ret = down_write_trylock(&tmp->memory_lock);
1261 goto hot_reset_release;
1263 mutex_unlock(&tmp->vma_lock);
1266 /* User has access, do the reset */
1267 ret = pci_reset_bus(vdev->pdev);
1270 for (i = 0; i < devs.cur_index; i++) {
1271 struct vfio_pci_device *tmp = devs.devices[i];
1274 up_write(&tmp->memory_lock);
1276 mutex_unlock(&tmp->vma_lock);
1277 vfio_device_put(&tmp->vdev);
1279 kfree(devs.devices);
1281 for (group_idx--; group_idx >= 0; group_idx--)
1282 vfio_group_put_external_user(groups[group_idx].group);
1286 } else if (cmd == VFIO_DEVICE_IOEVENTFD) {
1287 struct vfio_device_ioeventfd ioeventfd;
1290 minsz = offsetofend(struct vfio_device_ioeventfd, fd);
1292 if (copy_from_user(&ioeventfd, (void __user *)arg, minsz))
1295 if (ioeventfd.argsz < minsz)
1298 if (ioeventfd.flags & ~VFIO_DEVICE_IOEVENTFD_SIZE_MASK)
1301 count = ioeventfd.flags & VFIO_DEVICE_IOEVENTFD_SIZE_MASK;
1303 if (hweight8(count) != 1 || ioeventfd.fd < -1)
1306 return vfio_pci_ioeventfd(vdev, ioeventfd.offset,
1307 ioeventfd.data, count, ioeventfd.fd);
1308 } else if (cmd == VFIO_DEVICE_FEATURE) {
1309 struct vfio_device_feature feature;
1312 minsz = offsetofend(struct vfio_device_feature, flags);
1314 if (copy_from_user(&feature, (void __user *)arg, minsz))
1317 if (feature.argsz < minsz)
1320 /* Check unknown flags */
1321 if (feature.flags & ~(VFIO_DEVICE_FEATURE_MASK |
1322 VFIO_DEVICE_FEATURE_SET |
1323 VFIO_DEVICE_FEATURE_GET |
1324 VFIO_DEVICE_FEATURE_PROBE))
1327 /* GET & SET are mutually exclusive except with PROBE */
1328 if (!(feature.flags & VFIO_DEVICE_FEATURE_PROBE) &&
1329 (feature.flags & VFIO_DEVICE_FEATURE_SET) &&
1330 (feature.flags & VFIO_DEVICE_FEATURE_GET))
1333 switch (feature.flags & VFIO_DEVICE_FEATURE_MASK) {
1334 case VFIO_DEVICE_FEATURE_PCI_VF_TOKEN:
1335 if (!vdev->vf_token)
1339 * We do not support GET of the VF Token UUID as this
1340 * could expose the token of the previous device user.
1342 if (feature.flags & VFIO_DEVICE_FEATURE_GET)
1345 if (feature.flags & VFIO_DEVICE_FEATURE_PROBE)
1348 /* Don't SET unless told to do so */
1349 if (!(feature.flags & VFIO_DEVICE_FEATURE_SET))
1352 if (feature.argsz < minsz + sizeof(uuid))
1355 if (copy_from_user(&uuid, (void __user *)(arg + minsz),
1359 mutex_lock(&vdev->vf_token->lock);
1360 uuid_copy(&vdev->vf_token->uuid, &uuid);
1361 mutex_unlock(&vdev->vf_token->lock);
1372 static ssize_t vfio_pci_rw(struct vfio_pci_device *vdev, char __user *buf,
1373 size_t count, loff_t *ppos, bool iswrite)
1375 unsigned int index = VFIO_PCI_OFFSET_TO_INDEX(*ppos);
1377 if (index >= VFIO_PCI_NUM_REGIONS + vdev->num_regions)
1381 case VFIO_PCI_CONFIG_REGION_INDEX:
1382 return vfio_pci_config_rw(vdev, buf, count, ppos, iswrite);
1384 case VFIO_PCI_ROM_REGION_INDEX:
1387 return vfio_pci_bar_rw(vdev, buf, count, ppos, false);
1389 case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
1390 return vfio_pci_bar_rw(vdev, buf, count, ppos, iswrite);
1392 case VFIO_PCI_VGA_REGION_INDEX:
1393 return vfio_pci_vga_rw(vdev, buf, count, ppos, iswrite);
1395 index -= VFIO_PCI_NUM_REGIONS;
1396 return vdev->region[index].ops->rw(vdev, buf,
1397 count, ppos, iswrite);
1403 static ssize_t vfio_pci_read(struct vfio_device *core_vdev, char __user *buf,
1404 size_t count, loff_t *ppos)
1406 struct vfio_pci_device *vdev =
1407 container_of(core_vdev, struct vfio_pci_device, vdev);
1412 return vfio_pci_rw(vdev, buf, count, ppos, false);
1415 static ssize_t vfio_pci_write(struct vfio_device *core_vdev, const char __user *buf,
1416 size_t count, loff_t *ppos)
1418 struct vfio_pci_device *vdev =
1419 container_of(core_vdev, struct vfio_pci_device, vdev);
1424 return vfio_pci_rw(vdev, (char __user *)buf, count, ppos, true);
1427 /* Return 1 on zap and vma_lock acquired, 0 on contention (only with @try) */
1428 static int vfio_pci_zap_and_vma_lock(struct vfio_pci_device *vdev, bool try)
1430 struct vfio_pci_mmap_vma *mmap_vma, *tmp;
1434 * vma_lock is nested under mmap_lock for vm_ops callback paths.
1435 * The memory_lock semaphore is used by both code paths calling
1436 * into this function to zap vmas and the vm_ops.fault callback
1437 * to protect the memory enable state of the device.
1439 * When zapping vmas we need to maintain the mmap_lock => vma_lock
1440 * ordering, which requires using vma_lock to walk vma_list to
1441 * acquire an mm, then dropping vma_lock to get the mmap_lock and
1442 * reacquiring vma_lock. This logic is derived from similar
1443 * requirements in uverbs_user_mmap_disassociate().
1445 * mmap_lock must always be the top-level lock when it is taken.
1446 * Therefore we can only hold the memory_lock write lock when
1447 * vma_list is empty, as we'd need to take mmap_lock to clear
1448 * entries. vma_list can only be guaranteed empty when holding
1449 * vma_lock, thus memory_lock is nested under vma_lock.
1451 * This enables the vm_ops.fault callback to acquire vma_lock,
1452 * followed by memory_lock read lock, while already holding
1453 * mmap_lock without risk of deadlock.
1456 struct mm_struct *mm = NULL;
1459 if (!mutex_trylock(&vdev->vma_lock))
1462 mutex_lock(&vdev->vma_lock);
1464 while (!list_empty(&vdev->vma_list)) {
1465 mmap_vma = list_first_entry(&vdev->vma_list,
1466 struct vfio_pci_mmap_vma,
1468 mm = mmap_vma->vma->vm_mm;
1469 if (mmget_not_zero(mm))
1472 list_del(&mmap_vma->vma_next);
1478 mutex_unlock(&vdev->vma_lock);
1481 if (!mmap_read_trylock(mm)) {
1489 if (!mutex_trylock(&vdev->vma_lock)) {
1490 mmap_read_unlock(mm);
1495 mutex_lock(&vdev->vma_lock);
1497 list_for_each_entry_safe(mmap_vma, tmp,
1498 &vdev->vma_list, vma_next) {
1499 struct vm_area_struct *vma = mmap_vma->vma;
1501 if (vma->vm_mm != mm)
1504 list_del(&mmap_vma->vma_next);
1507 zap_vma_ptes(vma, vma->vm_start,
1508 vma->vm_end - vma->vm_start);
1510 mutex_unlock(&vdev->vma_lock);
1511 mmap_read_unlock(mm);
1516 void vfio_pci_zap_and_down_write_memory_lock(struct vfio_pci_device *vdev)
1518 vfio_pci_zap_and_vma_lock(vdev, false);
1519 down_write(&vdev->memory_lock);
1520 mutex_unlock(&vdev->vma_lock);
1523 u16 vfio_pci_memory_lock_and_enable(struct vfio_pci_device *vdev)
1527 down_write(&vdev->memory_lock);
1528 pci_read_config_word(vdev->pdev, PCI_COMMAND, &cmd);
1529 if (!(cmd & PCI_COMMAND_MEMORY))
1530 pci_write_config_word(vdev->pdev, PCI_COMMAND,
1531 cmd | PCI_COMMAND_MEMORY);
1536 void vfio_pci_memory_unlock_and_restore(struct vfio_pci_device *vdev, u16 cmd)
1538 pci_write_config_word(vdev->pdev, PCI_COMMAND, cmd);
1539 up_write(&vdev->memory_lock);
1542 /* Caller holds vma_lock */
1543 static int __vfio_pci_add_vma(struct vfio_pci_device *vdev,
1544 struct vm_area_struct *vma)
1546 struct vfio_pci_mmap_vma *mmap_vma;
1548 mmap_vma = kmalloc(sizeof(*mmap_vma), GFP_KERNEL);
1552 mmap_vma->vma = vma;
1553 list_add(&mmap_vma->vma_next, &vdev->vma_list);
1559 * Zap mmaps on open so that we can fault them in on access and therefore
1560 * our vma_list only tracks mappings accessed since last zap.
1562 static void vfio_pci_mmap_open(struct vm_area_struct *vma)
1564 zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
1567 static void vfio_pci_mmap_close(struct vm_area_struct *vma)
1569 struct vfio_pci_device *vdev = vma->vm_private_data;
1570 struct vfio_pci_mmap_vma *mmap_vma;
1572 mutex_lock(&vdev->vma_lock);
1573 list_for_each_entry(mmap_vma, &vdev->vma_list, vma_next) {
1574 if (mmap_vma->vma == vma) {
1575 list_del(&mmap_vma->vma_next);
1580 mutex_unlock(&vdev->vma_lock);
1583 static vm_fault_t vfio_pci_mmap_fault(struct vm_fault *vmf)
1585 struct vm_area_struct *vma = vmf->vma;
1586 struct vfio_pci_device *vdev = vma->vm_private_data;
1587 struct vfio_pci_mmap_vma *mmap_vma;
1588 vm_fault_t ret = VM_FAULT_NOPAGE;
1590 mutex_lock(&vdev->vma_lock);
1591 down_read(&vdev->memory_lock);
1593 if (!__vfio_pci_memory_enabled(vdev)) {
1594 ret = VM_FAULT_SIGBUS;
1599 * We populate the whole vma on fault, so we need to test whether
1600 * the vma has already been mapped, such as for concurrent faults
1601 * to the same vma. io_remap_pfn_range() will trigger a BUG_ON if
1602 * we ask it to fill the same range again.
1604 list_for_each_entry(mmap_vma, &vdev->vma_list, vma_next) {
1605 if (mmap_vma->vma == vma)
1609 if (io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
1610 vma->vm_end - vma->vm_start,
1611 vma->vm_page_prot)) {
1612 ret = VM_FAULT_SIGBUS;
1613 zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
1617 if (__vfio_pci_add_vma(vdev, vma)) {
1619 zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
1623 up_read(&vdev->memory_lock);
1624 mutex_unlock(&vdev->vma_lock);
1628 static const struct vm_operations_struct vfio_pci_mmap_ops = {
1629 .open = vfio_pci_mmap_open,
1630 .close = vfio_pci_mmap_close,
1631 .fault = vfio_pci_mmap_fault,
1634 static int vfio_pci_mmap(struct vfio_device *core_vdev, struct vm_area_struct *vma)
1636 struct vfio_pci_device *vdev =
1637 container_of(core_vdev, struct vfio_pci_device, vdev);
1638 struct pci_dev *pdev = vdev->pdev;
1640 u64 phys_len, req_len, pgoff, req_start;
1643 index = vma->vm_pgoff >> (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT);
1645 if (index >= VFIO_PCI_NUM_REGIONS + vdev->num_regions)
1647 if (vma->vm_end < vma->vm_start)
1649 if ((vma->vm_flags & VM_SHARED) == 0)
1651 if (index >= VFIO_PCI_NUM_REGIONS) {
1652 int regnum = index - VFIO_PCI_NUM_REGIONS;
1653 struct vfio_pci_region *region = vdev->region + regnum;
1655 if (region->ops && region->ops->mmap &&
1656 (region->flags & VFIO_REGION_INFO_FLAG_MMAP))
1657 return region->ops->mmap(vdev, region, vma);
1660 if (index >= VFIO_PCI_ROM_REGION_INDEX)
1662 if (!vdev->bar_mmap_supported[index])
1665 phys_len = PAGE_ALIGN(pci_resource_len(pdev, index));
1666 req_len = vma->vm_end - vma->vm_start;
1667 pgoff = vma->vm_pgoff &
1668 ((1U << (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT)) - 1);
1669 req_start = pgoff << PAGE_SHIFT;
1671 if (req_start + req_len > phys_len)
1675 * Even though we don't make use of the barmap for the mmap,
1676 * we need to request the region and the barmap tracks that.
1678 if (!vdev->barmap[index]) {
1679 ret = pci_request_selected_regions(pdev,
1680 1 << index, "vfio-pci");
1684 vdev->barmap[index] = pci_iomap(pdev, index, 0);
1685 if (!vdev->barmap[index]) {
1686 pci_release_selected_regions(pdev, 1 << index);
1691 vma->vm_private_data = vdev;
1692 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1693 vma->vm_pgoff = (pci_resource_start(pdev, index) >> PAGE_SHIFT) + pgoff;
1696 * See remap_pfn_range(), called from vfio_pci_fault() but we can't
1697 * change vm_flags within the fault handler. Set them now.
1699 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1700 vma->vm_ops = &vfio_pci_mmap_ops;
1705 static void vfio_pci_request(struct vfio_device *core_vdev, unsigned int count)
1707 struct vfio_pci_device *vdev =
1708 container_of(core_vdev, struct vfio_pci_device, vdev);
1709 struct pci_dev *pdev = vdev->pdev;
1711 mutex_lock(&vdev->igate);
1713 if (vdev->req_trigger) {
1715 pci_notice_ratelimited(pdev,
1716 "Relaying device request to user (#%u)\n",
1718 eventfd_signal(vdev->req_trigger, 1);
1719 } else if (count == 0) {
1721 "No device request channel registered, blocked until released by user\n");
1724 mutex_unlock(&vdev->igate);
1727 static int vfio_pci_validate_vf_token(struct vfio_pci_device *vdev,
1728 bool vf_token, uuid_t *uuid)
1731 * There's always some degree of trust or collaboration between SR-IOV
1732 * PF and VFs, even if just that the PF hosts the SR-IOV capability and
1733 * can disrupt VFs with a reset, but often the PF has more explicit
1734 * access to deny service to the VF or access data passed through the
1735 * VF. We therefore require an opt-in via a shared VF token (UUID) to
1736 * represent this trust. This both prevents that a VF driver might
1737 * assume the PF driver is a trusted, in-kernel driver, and also that
1738 * a PF driver might be replaced with a rogue driver, unknown to in-use
1741 * Therefore when presented with a VF, if the PF is a vfio device and
1742 * it is bound to the vfio-pci driver, the user needs to provide a VF
1743 * token to access the device, in the form of appending a vf_token to
1744 * the device name, for example:
1746 * "0000:04:10.0 vf_token=bd8d9d2b-5a5f-4f5a-a211-f591514ba1f3"
1748 * When presented with a PF which has VFs in use, the user must also
1749 * provide the current VF token to prove collaboration with existing
1750 * VF users. If VFs are not in use, the VF token provided for the PF
1751 * device will act to set the VF token.
1753 * If the VF token is provided but unused, an error is generated.
1755 if (!vdev->pdev->is_virtfn && !vdev->vf_token && !vf_token)
1756 return 0; /* No VF token provided or required */
1758 if (vdev->pdev->is_virtfn) {
1759 struct vfio_pci_device *pf_vdev = get_pf_vdev(vdev);
1764 return 0; /* PF is not vfio-pci, no VF token */
1766 pci_info_ratelimited(vdev->pdev,
1767 "VF token incorrectly provided, PF not bound to vfio-pci\n");
1772 vfio_device_put(&pf_vdev->vdev);
1773 pci_info_ratelimited(vdev->pdev,
1774 "VF token required to access device\n");
1778 mutex_lock(&pf_vdev->vf_token->lock);
1779 match = uuid_equal(uuid, &pf_vdev->vf_token->uuid);
1780 mutex_unlock(&pf_vdev->vf_token->lock);
1782 vfio_device_put(&pf_vdev->vdev);
1785 pci_info_ratelimited(vdev->pdev,
1786 "Incorrect VF token provided for device\n");
1789 } else if (vdev->vf_token) {
1790 mutex_lock(&vdev->vf_token->lock);
1791 if (vdev->vf_token->users) {
1793 mutex_unlock(&vdev->vf_token->lock);
1794 pci_info_ratelimited(vdev->pdev,
1795 "VF token required to access device\n");
1799 if (!uuid_equal(uuid, &vdev->vf_token->uuid)) {
1800 mutex_unlock(&vdev->vf_token->lock);
1801 pci_info_ratelimited(vdev->pdev,
1802 "Incorrect VF token provided for device\n");
1805 } else if (vf_token) {
1806 uuid_copy(&vdev->vf_token->uuid, uuid);
1809 mutex_unlock(&vdev->vf_token->lock);
1810 } else if (vf_token) {
1811 pci_info_ratelimited(vdev->pdev,
1812 "VF token incorrectly provided, not a PF or VF\n");
1819 #define VF_TOKEN_ARG "vf_token="
1821 static int vfio_pci_match(struct vfio_device *core_vdev, char *buf)
1823 struct vfio_pci_device *vdev =
1824 container_of(core_vdev, struct vfio_pci_device, vdev);
1825 bool vf_token = false;
1829 if (strncmp(pci_name(vdev->pdev), buf, strlen(pci_name(vdev->pdev))))
1830 return 0; /* No match */
1832 if (strlen(buf) > strlen(pci_name(vdev->pdev))) {
1833 buf += strlen(pci_name(vdev->pdev));
1836 return 0; /* No match: non-whitespace after name */
1844 if (!vf_token && !strncmp(buf, VF_TOKEN_ARG,
1845 strlen(VF_TOKEN_ARG))) {
1846 buf += strlen(VF_TOKEN_ARG);
1848 if (strlen(buf) < UUID_STRING_LEN)
1851 ret = uuid_parse(buf, &uuid);
1856 buf += UUID_STRING_LEN;
1858 /* Unknown/duplicate option */
1864 ret = vfio_pci_validate_vf_token(vdev, vf_token, &uuid);
1868 return 1; /* Match */
1871 static const struct vfio_device_ops vfio_pci_ops = {
1873 .open = vfio_pci_open,
1874 .release = vfio_pci_release,
1875 .ioctl = vfio_pci_ioctl,
1876 .read = vfio_pci_read,
1877 .write = vfio_pci_write,
1878 .mmap = vfio_pci_mmap,
1879 .request = vfio_pci_request,
1880 .match = vfio_pci_match,
1883 static int vfio_pci_reflck_attach(struct vfio_pci_device *vdev);
1884 static void vfio_pci_reflck_put(struct vfio_pci_reflck *reflck);
1886 static int vfio_pci_bus_notifier(struct notifier_block *nb,
1887 unsigned long action, void *data)
1889 struct vfio_pci_device *vdev = container_of(nb,
1890 struct vfio_pci_device, nb);
1891 struct device *dev = data;
1892 struct pci_dev *pdev = to_pci_dev(dev);
1893 struct pci_dev *physfn = pci_physfn(pdev);
1895 if (action == BUS_NOTIFY_ADD_DEVICE &&
1896 pdev->is_virtfn && physfn == vdev->pdev) {
1897 pci_info(vdev->pdev, "Captured SR-IOV VF %s driver_override\n",
1899 pdev->driver_override = kasprintf(GFP_KERNEL, "%s",
1901 } else if (action == BUS_NOTIFY_BOUND_DRIVER &&
1902 pdev->is_virtfn && physfn == vdev->pdev) {
1903 struct pci_driver *drv = pci_dev_driver(pdev);
1905 if (drv && drv != &vfio_pci_driver)
1906 pci_warn(vdev->pdev,
1907 "VF %s bound to driver %s while PF bound to vfio-pci\n",
1908 pci_name(pdev), drv->name);
1914 static int vfio_pci_vf_init(struct vfio_pci_device *vdev)
1916 struct pci_dev *pdev = vdev->pdev;
1919 if (!pdev->is_physfn)
1922 vdev->vf_token = kzalloc(sizeof(*vdev->vf_token), GFP_KERNEL);
1923 if (!vdev->vf_token)
1926 mutex_init(&vdev->vf_token->lock);
1927 uuid_gen(&vdev->vf_token->uuid);
1929 vdev->nb.notifier_call = vfio_pci_bus_notifier;
1930 ret = bus_register_notifier(&pci_bus_type, &vdev->nb);
1932 kfree(vdev->vf_token);
1938 static void vfio_pci_vf_uninit(struct vfio_pci_device *vdev)
1940 if (!vdev->vf_token)
1943 bus_unregister_notifier(&pci_bus_type, &vdev->nb);
1944 WARN_ON(vdev->vf_token->users);
1945 mutex_destroy(&vdev->vf_token->lock);
1946 kfree(vdev->vf_token);
1949 static int vfio_pci_vga_init(struct vfio_pci_device *vdev)
1951 struct pci_dev *pdev = vdev->pdev;
1954 if (!vfio_pci_is_vga(pdev))
1957 ret = vga_client_register(pdev, vdev, NULL, vfio_pci_set_vga_decode);
1960 vga_set_legacy_decoding(pdev, vfio_pci_set_vga_decode(vdev, false));
1964 static void vfio_pci_vga_uninit(struct vfio_pci_device *vdev)
1966 struct pci_dev *pdev = vdev->pdev;
1968 if (!vfio_pci_is_vga(pdev))
1970 vga_client_register(pdev, NULL, NULL, NULL);
1971 vga_set_legacy_decoding(pdev, VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM |
1972 VGA_RSRC_LEGACY_IO |
1973 VGA_RSRC_LEGACY_MEM);
1976 static int vfio_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
1978 struct vfio_pci_device *vdev;
1979 struct iommu_group *group;
1982 if (vfio_pci_is_denylisted(pdev))
1985 if (pdev->hdr_type != PCI_HEADER_TYPE_NORMAL)
1989 * Prevent binding to PFs with VFs enabled, the VFs might be in use
1990 * by the host or other users. We cannot capture the VFs if they
1991 * already exist, nor can we track VF users. Disabling SR-IOV here
1992 * would initiate removing the VFs, which would unbind the driver,
1993 * which is prone to blocking if that VF is also in use by vfio-pci.
1994 * Just reject these PFs and let the user sort it out.
1996 if (pci_num_vf(pdev)) {
1997 pci_warn(pdev, "Cannot bind to PF with SR-IOV enabled\n");
2001 group = vfio_iommu_group_get(&pdev->dev);
2005 vdev = kzalloc(sizeof(*vdev), GFP_KERNEL);
2011 vfio_init_group_dev(&vdev->vdev, &pdev->dev, &vfio_pci_ops);
2013 vdev->irq_type = VFIO_PCI_NUM_IRQS;
2014 mutex_init(&vdev->igate);
2015 spin_lock_init(&vdev->irqlock);
2016 mutex_init(&vdev->ioeventfds_lock);
2017 INIT_LIST_HEAD(&vdev->dummy_resources_list);
2018 INIT_LIST_HEAD(&vdev->ioeventfds_list);
2019 mutex_init(&vdev->vma_lock);
2020 INIT_LIST_HEAD(&vdev->vma_list);
2021 init_rwsem(&vdev->memory_lock);
2023 ret = vfio_pci_reflck_attach(vdev);
2026 ret = vfio_pci_vf_init(vdev);
2029 ret = vfio_pci_vga_init(vdev);
2033 vfio_pci_probe_power_state(vdev);
2035 if (!disable_idle_d3) {
2037 * pci-core sets the device power state to an unknown value at
2038 * bootup and after being removed from a driver. The only
2039 * transition it allows from this unknown state is to D0, which
2040 * typically happens when a driver calls pci_enable_device().
2041 * We're not ready to enable the device yet, but we do want to
2042 * be able to get to D3. Therefore first do a D0 transition
2043 * before going to D3.
2045 vfio_pci_set_power_state(vdev, PCI_D0);
2046 vfio_pci_set_power_state(vdev, PCI_D3hot);
2049 ret = vfio_register_group_dev(&vdev->vdev);
2052 dev_set_drvdata(&pdev->dev, vdev);
2056 if (!disable_idle_d3)
2057 vfio_pci_set_power_state(vdev, PCI_D0);
2059 vfio_pci_vf_uninit(vdev);
2061 vfio_pci_reflck_put(vdev->reflck);
2063 kfree(vdev->pm_save);
2066 vfio_iommu_group_put(group, &pdev->dev);
2070 static void vfio_pci_remove(struct pci_dev *pdev)
2072 struct vfio_pci_device *vdev = dev_get_drvdata(&pdev->dev);
2074 pci_disable_sriov(pdev);
2076 vfio_unregister_group_dev(&vdev->vdev);
2078 vfio_pci_vf_uninit(vdev);
2079 vfio_pci_reflck_put(vdev->reflck);
2080 vfio_pci_vga_uninit(vdev);
2082 vfio_iommu_group_put(pdev->dev.iommu_group, &pdev->dev);
2084 if (!disable_idle_d3)
2085 vfio_pci_set_power_state(vdev, PCI_D0);
2087 mutex_destroy(&vdev->ioeventfds_lock);
2088 kfree(vdev->region);
2089 kfree(vdev->pm_save);
2093 static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
2094 pci_channel_state_t state)
2096 struct vfio_pci_device *vdev;
2097 struct vfio_device *device;
2099 device = vfio_device_get_from_dev(&pdev->dev);
2101 return PCI_ERS_RESULT_DISCONNECT;
2103 vdev = container_of(device, struct vfio_pci_device, vdev);
2105 mutex_lock(&vdev->igate);
2107 if (vdev->err_trigger)
2108 eventfd_signal(vdev->err_trigger, 1);
2110 mutex_unlock(&vdev->igate);
2112 vfio_device_put(device);
2114 return PCI_ERS_RESULT_CAN_RECOVER;
2117 static int vfio_pci_sriov_configure(struct pci_dev *pdev, int nr_virtfn)
2119 struct vfio_device *device;
2127 device = vfio_device_get_from_dev(&pdev->dev);
2132 pci_disable_sriov(pdev);
2134 ret = pci_enable_sriov(pdev, nr_virtfn);
2136 vfio_device_put(device);
2138 return ret < 0 ? ret : nr_virtfn;
2141 static const struct pci_error_handlers vfio_err_handlers = {
2142 .error_detected = vfio_pci_aer_err_detected,
2145 static struct pci_driver vfio_pci_driver = {
2147 .id_table = NULL, /* only dynamic ids */
2148 .probe = vfio_pci_probe,
2149 .remove = vfio_pci_remove,
2150 .sriov_configure = vfio_pci_sriov_configure,
2151 .err_handler = &vfio_err_handlers,
2154 static DEFINE_MUTEX(reflck_lock);
2156 static struct vfio_pci_reflck *vfio_pci_reflck_alloc(void)
2158 struct vfio_pci_reflck *reflck;
2160 reflck = kzalloc(sizeof(*reflck), GFP_KERNEL);
2162 return ERR_PTR(-ENOMEM);
2164 kref_init(&reflck->kref);
2165 mutex_init(&reflck->lock);
2170 static void vfio_pci_reflck_get(struct vfio_pci_reflck *reflck)
2172 kref_get(&reflck->kref);
2175 static int vfio_pci_reflck_find(struct pci_dev *pdev, void *data)
2177 struct vfio_pci_reflck **preflck = data;
2178 struct vfio_device *device;
2179 struct vfio_pci_device *vdev;
2181 device = vfio_device_get_from_dev(&pdev->dev);
2185 if (pci_dev_driver(pdev) != &vfio_pci_driver) {
2186 vfio_device_put(device);
2190 vdev = container_of(device, struct vfio_pci_device, vdev);
2193 vfio_pci_reflck_get(vdev->reflck);
2194 *preflck = vdev->reflck;
2195 vfio_device_put(device);
2199 vfio_device_put(device);
2203 static int vfio_pci_reflck_attach(struct vfio_pci_device *vdev)
2205 bool slot = !pci_probe_reset_slot(vdev->pdev->slot);
2207 mutex_lock(&reflck_lock);
2209 if (pci_is_root_bus(vdev->pdev->bus) ||
2210 vfio_pci_for_each_slot_or_bus(vdev->pdev, vfio_pci_reflck_find,
2211 &vdev->reflck, slot) <= 0)
2212 vdev->reflck = vfio_pci_reflck_alloc();
2214 mutex_unlock(&reflck_lock);
2216 return PTR_ERR_OR_ZERO(vdev->reflck);
2219 static void vfio_pci_reflck_release(struct kref *kref)
2221 struct vfio_pci_reflck *reflck = container_of(kref,
2222 struct vfio_pci_reflck,
2226 mutex_unlock(&reflck_lock);
2229 static void vfio_pci_reflck_put(struct vfio_pci_reflck *reflck)
2231 kref_put_mutex(&reflck->kref, vfio_pci_reflck_release, &reflck_lock);
2234 static int vfio_pci_get_unused_devs(struct pci_dev *pdev, void *data)
2236 struct vfio_devices *devs = data;
2237 struct vfio_device *device;
2238 struct vfio_pci_device *vdev;
2240 if (devs->cur_index == devs->max_index)
2243 device = vfio_device_get_from_dev(&pdev->dev);
2247 if (pci_dev_driver(pdev) != &vfio_pci_driver) {
2248 vfio_device_put(device);
2252 vdev = container_of(device, struct vfio_pci_device, vdev);
2254 /* Fault if the device is not unused */
2256 vfio_device_put(device);
2260 devs->devices[devs->cur_index++] = vdev;
2264 static int vfio_pci_try_zap_and_vma_lock_cb(struct pci_dev *pdev, void *data)
2266 struct vfio_devices *devs = data;
2267 struct vfio_device *device;
2268 struct vfio_pci_device *vdev;
2270 if (devs->cur_index == devs->max_index)
2273 device = vfio_device_get_from_dev(&pdev->dev);
2277 if (pci_dev_driver(pdev) != &vfio_pci_driver) {
2278 vfio_device_put(device);
2282 vdev = container_of(device, struct vfio_pci_device, vdev);
2285 * Locking multiple devices is prone to deadlock, runaway and
2286 * unwind if we hit contention.
2288 if (!vfio_pci_zap_and_vma_lock(vdev, true)) {
2289 vfio_device_put(device);
2293 devs->devices[devs->cur_index++] = vdev;
2298 * If a bus or slot reset is available for the provided device and:
2299 * - All of the devices affected by that bus or slot reset are unused
2301 * - At least one of the affected devices is marked dirty via
2302 * needs_reset (such as by lack of FLR support)
2303 * Then attempt to perform that bus or slot reset. Callers are required
2304 * to hold vdev->reflck->lock, protecting the bus/slot reset group from
2305 * concurrent opens. A vfio_device reference is acquired for each device
2306 * to prevent unbinds during the reset operation.
2308 * NB: vfio-core considers a group to be viable even if some devices are
2309 * bound to drivers like pci-stub or pcieport. Here we require all devices
2310 * to be bound to vfio_pci since that's the only way we can be sure they
2313 static void vfio_pci_try_bus_reset(struct vfio_pci_device *vdev)
2315 struct vfio_devices devs = { .cur_index = 0 };
2316 int i = 0, ret = -EINVAL;
2318 struct vfio_pci_device *tmp;
2320 if (!pci_probe_reset_slot(vdev->pdev->slot))
2322 else if (pci_probe_reset_bus(vdev->pdev->bus))
2325 if (vfio_pci_for_each_slot_or_bus(vdev->pdev, vfio_pci_count_devs,
2330 devs.devices = kcalloc(i, sizeof(struct vfio_device *), GFP_KERNEL);
2334 if (vfio_pci_for_each_slot_or_bus(vdev->pdev,
2335 vfio_pci_get_unused_devs,
2339 /* Does at least one need a reset? */
2340 for (i = 0; i < devs.cur_index; i++) {
2341 tmp = devs.devices[i];
2342 if (tmp->needs_reset) {
2343 ret = pci_reset_bus(vdev->pdev);
2349 for (i = 0; i < devs.cur_index; i++) {
2350 tmp = devs.devices[i];
2353 * If reset was successful, affected devices no longer need
2354 * a reset and we should return all the collateral devices
2355 * to low power. If not successful, we either didn't reset
2356 * the bus or timed out waiting for it, so let's not touch
2360 tmp->needs_reset = false;
2362 if (tmp != vdev && !disable_idle_d3)
2363 vfio_pci_set_power_state(tmp, PCI_D3hot);
2366 vfio_device_put(&tmp->vdev);
2369 kfree(devs.devices);
2372 static void __exit vfio_pci_cleanup(void)
2374 pci_unregister_driver(&vfio_pci_driver);
2375 vfio_pci_uninit_perm_bits();
2378 static void __init vfio_pci_fill_ids(void)
2383 /* no ids passed actually */
2387 /* add ids specified in the module parameter */
2389 while ((id = strsep(&p, ","))) {
2390 unsigned int vendor, device, subvendor = PCI_ANY_ID,
2391 subdevice = PCI_ANY_ID, class = 0, class_mask = 0;
2397 fields = sscanf(id, "%x:%x:%x:%x:%x:%x",
2398 &vendor, &device, &subvendor, &subdevice,
2399 &class, &class_mask);
2402 pr_warn("invalid id string \"%s\"\n", id);
2406 rc = pci_add_dynid(&vfio_pci_driver, vendor, device,
2407 subvendor, subdevice, class, class_mask, 0);
2409 pr_warn("failed to add dynamic id [%04x:%04x[%04x:%04x]] class %#08x/%08x (%d)\n",
2410 vendor, device, subvendor, subdevice,
2411 class, class_mask, rc);
2413 pr_info("add [%04x:%04x[%04x:%04x]] class %#08x/%08x\n",
2414 vendor, device, subvendor, subdevice,
2419 static int __init vfio_pci_init(void)
2423 /* Allocate shared config space permission data used by all devices */
2424 ret = vfio_pci_init_perm_bits();
2428 /* Register and scan for devices */
2429 ret = pci_register_driver(&vfio_pci_driver);
2433 vfio_pci_fill_ids();
2435 if (disable_denylist)
2436 pr_warn("device denylist disabled.\n");
2441 vfio_pci_uninit_perm_bits();
2445 module_init(vfio_pci_init);
2446 module_exit(vfio_pci_cleanup);
2448 MODULE_VERSION(DRIVER_VERSION);
2449 MODULE_LICENSE("GPL v2");
2450 MODULE_AUTHOR(DRIVER_AUTHOR);
2451 MODULE_DESCRIPTION(DRIVER_DESC);