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/aperture.h>
14 #include <linux/device.h>
15 #include <linux/eventfd.h>
16 #include <linux/file.h>
17 #include <linux/interrupt.h>
18 #include <linux/iommu.h>
19 #include <linux/module.h>
20 #include <linux/mutex.h>
21 #include <linux/notifier.h>
22 #include <linux/pci.h>
23 #include <linux/pm_runtime.h>
24 #include <linux/slab.h>
25 #include <linux/types.h>
26 #include <linux/uaccess.h>
27 #include <linux/vgaarb.h>
28 #include <linux/nospec.h>
29 #include <linux/sched/mm.h>
30 #include <linux/iommufd.h>
31 #if IS_ENABLED(CONFIG_EEH)
35 #include "vfio_pci_priv.h"
38 #define DRIVER_DESC "core driver for VFIO based PCI devices"
40 static bool nointxmask;
41 static bool disable_vga;
42 static bool disable_idle_d3;
44 /* List of PF's that vfio_pci_core_sriov_configure() has been called on */
45 static DEFINE_MUTEX(vfio_pci_sriov_pfs_mutex);
46 static LIST_HEAD(vfio_pci_sriov_pfs);
48 struct vfio_pci_dummy_resource {
49 struct resource resource;
51 struct list_head res_next;
54 struct vfio_pci_vf_token {
60 struct vfio_pci_mmap_vma {
61 struct vm_area_struct *vma;
62 struct list_head vma_next;
65 static inline bool vfio_vga_disabled(void)
67 #ifdef CONFIG_VFIO_PCI_VGA
75 * Our VGA arbiter participation is limited since we don't know anything
76 * about the device itself. However, if the device is the only VGA device
77 * downstream of a bridge and VFIO VGA support is disabled, then we can
78 * safely return legacy VGA IO and memory as not decoded since the user
79 * has no way to get to it and routing can be disabled externally at the
82 static unsigned int vfio_pci_set_decode(struct pci_dev *pdev, bool single_vga)
84 struct pci_dev *tmp = NULL;
85 unsigned char max_busnr;
88 if (single_vga || !vfio_vga_disabled() || pci_is_root_bus(pdev->bus))
89 return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM |
90 VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM;
92 max_busnr = pci_bus_max_busnr(pdev->bus);
93 decodes = VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
95 while ((tmp = pci_get_class(PCI_CLASS_DISPLAY_VGA << 8, tmp)) != NULL) {
97 pci_domain_nr(tmp->bus) != pci_domain_nr(pdev->bus) ||
98 pci_is_root_bus(tmp->bus))
101 if (tmp->bus->number >= pdev->bus->number &&
102 tmp->bus->number <= max_busnr) {
104 decodes |= VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM;
112 static void vfio_pci_probe_mmaps(struct vfio_pci_core_device *vdev)
114 struct resource *res;
116 struct vfio_pci_dummy_resource *dummy_res;
118 for (i = 0; i < PCI_STD_NUM_BARS; i++) {
119 int bar = i + PCI_STD_RESOURCES;
121 res = &vdev->pdev->resource[bar];
123 if (!IS_ENABLED(CONFIG_VFIO_PCI_MMAP))
126 if (!(res->flags & IORESOURCE_MEM))
130 * The PCI core shouldn't set up a resource with a
131 * type but zero size. But there may be bugs that
132 * cause us to do that.
134 if (!resource_size(res))
137 if (resource_size(res) >= PAGE_SIZE) {
138 vdev->bar_mmap_supported[bar] = true;
142 if (!(res->start & ~PAGE_MASK)) {
144 * Add a dummy resource to reserve the remainder
145 * of the exclusive page in case that hot-add
146 * device's bar is assigned into it.
149 kzalloc(sizeof(*dummy_res), GFP_KERNEL_ACCOUNT);
150 if (dummy_res == NULL)
153 dummy_res->resource.name = "vfio sub-page reserved";
154 dummy_res->resource.start = res->end + 1;
155 dummy_res->resource.end = res->start + PAGE_SIZE - 1;
156 dummy_res->resource.flags = res->flags;
157 if (request_resource(res->parent,
158 &dummy_res->resource)) {
162 dummy_res->index = bar;
163 list_add(&dummy_res->res_next,
164 &vdev->dummy_resources_list);
165 vdev->bar_mmap_supported[bar] = true;
169 * Here we don't handle the case when the BAR is not page
170 * aligned because we can't expect the BAR will be
171 * assigned into the same location in a page in guest
172 * when we passthrough the BAR. And it's hard to access
173 * this BAR in userspace because we have no way to get
174 * the BAR's location in a page.
177 vdev->bar_mmap_supported[bar] = false;
181 struct vfio_pci_group_info;
182 static void vfio_pci_dev_set_try_reset(struct vfio_device_set *dev_set);
183 static int vfio_pci_dev_set_hot_reset(struct vfio_device_set *dev_set,
184 struct vfio_pci_group_info *groups,
185 struct iommufd_ctx *iommufd_ctx);
188 * INTx masking requires the ability to disable INTx signaling via PCI_COMMAND
189 * _and_ the ability detect when the device is asserting INTx via PCI_STATUS.
190 * If a device implements the former but not the latter we would typically
191 * expect broken_intx_masking be set and require an exclusive interrupt.
192 * However since we do have control of the device's ability to assert INTx,
193 * we can instead pretend that the device does not implement INTx, virtualizing
194 * the pin register to report zero and maintaining DisINTx set on the host.
196 static bool vfio_pci_nointx(struct pci_dev *pdev)
198 switch (pdev->vendor) {
199 case PCI_VENDOR_ID_INTEL:
200 switch (pdev->device) {
201 /* All i40e (XL710/X710/XXV710) 10/20/25/40GbE NICs */
204 case 0x1580 ... 0x1581:
205 case 0x1583 ... 0x158b:
206 case 0x37d0 ... 0x37d2:
218 static void vfio_pci_probe_power_state(struct vfio_pci_core_device *vdev)
220 struct pci_dev *pdev = vdev->pdev;
226 pci_read_config_word(pdev, pdev->pm_cap + PCI_PM_CTRL, &pmcsr);
228 vdev->needs_pm_restore = !(pmcsr & PCI_PM_CTRL_NO_SOFT_RESET);
232 * pci_set_power_state() wrapper handling devices which perform a soft reset on
233 * D3->D0 transition. Save state prior to D0/1/2->D3, stash it on the vdev,
234 * restore when returned to D0. Saved separately from pci_saved_state for use
235 * by PM capability emulation and separately from pci_dev internal saved state
236 * to avoid it being overwritten and consumed around other resets.
238 int vfio_pci_set_power_state(struct vfio_pci_core_device *vdev, pci_power_t state)
240 struct pci_dev *pdev = vdev->pdev;
241 bool needs_restore = false, needs_save = false;
244 /* Prevent changing power state for PFs with VFs enabled */
245 if (pci_num_vf(pdev) && state > PCI_D0)
248 if (vdev->needs_pm_restore) {
249 if (pdev->current_state < PCI_D3hot && state >= PCI_D3hot) {
250 pci_save_state(pdev);
254 if (pdev->current_state >= PCI_D3hot && state <= PCI_D0)
255 needs_restore = true;
258 ret = pci_set_power_state(pdev, state);
261 /* D3 might be unsupported via quirk, skip unless in D3 */
262 if (needs_save && pdev->current_state >= PCI_D3hot) {
264 * The current PCI state will be saved locally in
265 * 'pm_save' during the D3hot transition. When the
266 * device state is changed to D0 again with the current
267 * function, then pci_store_saved_state() will restore
268 * the state and will free the memory pointed by
269 * 'pm_save'. There are few cases where the PCI power
270 * state can be changed to D0 without the involvement
271 * of the driver. For these cases, free the earlier
272 * allocated memory first before overwriting 'pm_save'
273 * to prevent the memory leak.
275 kfree(vdev->pm_save);
276 vdev->pm_save = pci_store_saved_state(pdev);
277 } else if (needs_restore) {
278 pci_load_and_free_saved_state(pdev, &vdev->pm_save);
279 pci_restore_state(pdev);
286 static int vfio_pci_runtime_pm_entry(struct vfio_pci_core_device *vdev,
287 struct eventfd_ctx *efdctx)
290 * The vdev power related flags are protected with 'memory_lock'
293 vfio_pci_zap_and_down_write_memory_lock(vdev);
294 if (vdev->pm_runtime_engaged) {
295 up_write(&vdev->memory_lock);
299 vdev->pm_runtime_engaged = true;
300 vdev->pm_wake_eventfd_ctx = efdctx;
301 pm_runtime_put_noidle(&vdev->pdev->dev);
302 up_write(&vdev->memory_lock);
307 static int vfio_pci_core_pm_entry(struct vfio_device *device, u32 flags,
308 void __user *arg, size_t argsz)
310 struct vfio_pci_core_device *vdev =
311 container_of(device, struct vfio_pci_core_device, vdev);
314 ret = vfio_check_feature(flags, argsz, VFIO_DEVICE_FEATURE_SET, 0);
319 * Inside vfio_pci_runtime_pm_entry(), only the runtime PM usage count
320 * will be decremented. The pm_runtime_put() will be invoked again
321 * while returning from the ioctl and then the device can go into
322 * runtime suspended state.
324 return vfio_pci_runtime_pm_entry(vdev, NULL);
327 static int vfio_pci_core_pm_entry_with_wakeup(
328 struct vfio_device *device, u32 flags,
329 struct vfio_device_low_power_entry_with_wakeup __user *arg,
332 struct vfio_pci_core_device *vdev =
333 container_of(device, struct vfio_pci_core_device, vdev);
334 struct vfio_device_low_power_entry_with_wakeup entry;
335 struct eventfd_ctx *efdctx;
338 ret = vfio_check_feature(flags, argsz, VFIO_DEVICE_FEATURE_SET,
343 if (copy_from_user(&entry, arg, sizeof(entry)))
346 if (entry.wakeup_eventfd < 0)
349 efdctx = eventfd_ctx_fdget(entry.wakeup_eventfd);
351 return PTR_ERR(efdctx);
353 ret = vfio_pci_runtime_pm_entry(vdev, efdctx);
355 eventfd_ctx_put(efdctx);
360 static void __vfio_pci_runtime_pm_exit(struct vfio_pci_core_device *vdev)
362 if (vdev->pm_runtime_engaged) {
363 vdev->pm_runtime_engaged = false;
364 pm_runtime_get_noresume(&vdev->pdev->dev);
366 if (vdev->pm_wake_eventfd_ctx) {
367 eventfd_ctx_put(vdev->pm_wake_eventfd_ctx);
368 vdev->pm_wake_eventfd_ctx = NULL;
373 static void vfio_pci_runtime_pm_exit(struct vfio_pci_core_device *vdev)
376 * The vdev power related flags are protected with 'memory_lock'
379 down_write(&vdev->memory_lock);
380 __vfio_pci_runtime_pm_exit(vdev);
381 up_write(&vdev->memory_lock);
384 static int vfio_pci_core_pm_exit(struct vfio_device *device, u32 flags,
385 void __user *arg, size_t argsz)
387 struct vfio_pci_core_device *vdev =
388 container_of(device, struct vfio_pci_core_device, vdev);
391 ret = vfio_check_feature(flags, argsz, VFIO_DEVICE_FEATURE_SET, 0);
396 * The device is always in the active state here due to pm wrappers
397 * around ioctls. If the device had entered a low power state and
398 * pm_wake_eventfd_ctx is valid, vfio_pci_core_runtime_resume() has
399 * already signaled the eventfd and exited low power mode itself.
400 * pm_runtime_engaged protects the redundant call here.
402 vfio_pci_runtime_pm_exit(vdev);
407 static int vfio_pci_core_runtime_suspend(struct device *dev)
409 struct vfio_pci_core_device *vdev = dev_get_drvdata(dev);
411 down_write(&vdev->memory_lock);
413 * The user can move the device into D3hot state before invoking
414 * power management IOCTL. Move the device into D0 state here and then
415 * the pci-driver core runtime PM suspend function will move the device
416 * into the low power state. Also, for the devices which have
417 * NoSoftRst-, it will help in restoring the original state
418 * (saved locally in 'vdev->pm_save').
420 vfio_pci_set_power_state(vdev, PCI_D0);
421 up_write(&vdev->memory_lock);
424 * If INTx is enabled, then mask INTx before going into the runtime
425 * suspended state and unmask the same in the runtime resume.
426 * If INTx has already been masked by the user, then
427 * vfio_pci_intx_mask() will return false and in that case, INTx
428 * should not be unmasked in the runtime resume.
430 vdev->pm_intx_masked = ((vdev->irq_type == VFIO_PCI_INTX_IRQ_INDEX) &&
431 vfio_pci_intx_mask(vdev));
436 static int vfio_pci_core_runtime_resume(struct device *dev)
438 struct vfio_pci_core_device *vdev = dev_get_drvdata(dev);
441 * Resume with a pm_wake_eventfd_ctx signals the eventfd and exit
444 down_write(&vdev->memory_lock);
445 if (vdev->pm_wake_eventfd_ctx) {
446 eventfd_signal(vdev->pm_wake_eventfd_ctx);
447 __vfio_pci_runtime_pm_exit(vdev);
449 up_write(&vdev->memory_lock);
451 if (vdev->pm_intx_masked)
452 vfio_pci_intx_unmask(vdev);
456 #endif /* CONFIG_PM */
459 * The pci-driver core runtime PM routines always save the device state
460 * before going into suspended state. If the device is going into low power
461 * state with only with runtime PM ops, then no explicit handling is needed
462 * for the devices which have NoSoftRst-.
464 static const struct dev_pm_ops vfio_pci_core_pm_ops = {
465 SET_RUNTIME_PM_OPS(vfio_pci_core_runtime_suspend,
466 vfio_pci_core_runtime_resume,
470 int vfio_pci_core_enable(struct vfio_pci_core_device *vdev)
472 struct pci_dev *pdev = vdev->pdev;
477 if (!disable_idle_d3) {
478 ret = pm_runtime_resume_and_get(&pdev->dev);
483 /* Don't allow our initial saved state to include busmaster */
484 pci_clear_master(pdev);
486 ret = pci_enable_device(pdev);
490 /* If reset fails because of the device lock, fail this path entirely */
491 ret = pci_try_reset_function(pdev);
493 goto out_disable_device;
495 vdev->reset_works = !ret;
496 pci_save_state(pdev);
497 vdev->pci_saved_state = pci_store_saved_state(pdev);
498 if (!vdev->pci_saved_state)
499 pci_dbg(pdev, "%s: Couldn't store saved state\n", __func__);
501 if (likely(!nointxmask)) {
502 if (vfio_pci_nointx(pdev)) {
503 pci_info(pdev, "Masking broken INTx support\n");
507 vdev->pci_2_3 = pci_intx_mask_supported(pdev);
510 pci_read_config_word(pdev, PCI_COMMAND, &cmd);
511 if (vdev->pci_2_3 && (cmd & PCI_COMMAND_INTX_DISABLE)) {
512 cmd &= ~PCI_COMMAND_INTX_DISABLE;
513 pci_write_config_word(pdev, PCI_COMMAND, cmd);
516 ret = vfio_pci_zdev_open_device(vdev);
520 ret = vfio_config_init(vdev);
524 msix_pos = pdev->msix_cap;
529 pci_read_config_word(pdev, msix_pos + PCI_MSIX_FLAGS, &flags);
530 pci_read_config_dword(pdev, msix_pos + PCI_MSIX_TABLE, &table);
532 vdev->msix_bar = table & PCI_MSIX_TABLE_BIR;
533 vdev->msix_offset = table & PCI_MSIX_TABLE_OFFSET;
534 vdev->msix_size = ((flags & PCI_MSIX_FLAGS_QSIZE) + 1) * 16;
535 vdev->has_dyn_msix = pci_msix_can_alloc_dyn(pdev);
537 vdev->msix_bar = 0xFF;
538 vdev->has_dyn_msix = false;
541 if (!vfio_vga_disabled() && vfio_pci_is_vga(pdev))
542 vdev->has_vga = true;
548 vfio_pci_zdev_close_device(vdev);
550 kfree(vdev->pci_saved_state);
551 vdev->pci_saved_state = NULL;
553 pci_disable_device(pdev);
555 if (!disable_idle_d3)
556 pm_runtime_put(&pdev->dev);
559 EXPORT_SYMBOL_GPL(vfio_pci_core_enable);
561 void vfio_pci_core_disable(struct vfio_pci_core_device *vdev)
563 struct pci_dev *pdev = vdev->pdev;
564 struct vfio_pci_dummy_resource *dummy_res, *tmp;
565 struct vfio_pci_ioeventfd *ioeventfd, *ioeventfd_tmp;
568 /* For needs_reset */
569 lockdep_assert_held(&vdev->vdev.dev_set->lock);
572 * This function can be invoked while the power state is non-D0.
573 * This non-D0 power state can be with or without runtime PM.
574 * vfio_pci_runtime_pm_exit() will internally increment the usage
575 * count corresponding to pm_runtime_put() called during low power
576 * feature entry and then pm_runtime_resume() will wake up the device,
577 * if the device has already gone into the suspended state. Otherwise,
578 * the vfio_pci_set_power_state() will change the device power state
581 vfio_pci_runtime_pm_exit(vdev);
582 pm_runtime_resume(&pdev->dev);
585 * This function calls __pci_reset_function_locked() which internally
586 * can use pci_pm_reset() for the function reset. pci_pm_reset() will
587 * fail if the power state is non-D0. Also, for the devices which
588 * have NoSoftRst-, the reset function can cause the PCI config space
589 * reset without restoring the original state (saved locally in
592 vfio_pci_set_power_state(vdev, PCI_D0);
594 /* Stop the device from further DMA */
595 pci_clear_master(pdev);
597 vfio_pci_set_irqs_ioctl(vdev, VFIO_IRQ_SET_DATA_NONE |
598 VFIO_IRQ_SET_ACTION_TRIGGER,
599 vdev->irq_type, 0, 0, NULL);
601 /* Device closed, don't need mutex here */
602 list_for_each_entry_safe(ioeventfd, ioeventfd_tmp,
603 &vdev->ioeventfds_list, next) {
604 vfio_virqfd_disable(&ioeventfd->virqfd);
605 list_del(&ioeventfd->next);
608 vdev->ioeventfds_nr = 0;
610 vdev->virq_disabled = false;
612 for (i = 0; i < vdev->num_regions; i++)
613 vdev->region[i].ops->release(vdev, &vdev->region[i]);
615 vdev->num_regions = 0;
617 vdev->region = NULL; /* don't krealloc a freed pointer */
619 vfio_config_free(vdev);
621 for (i = 0; i < PCI_STD_NUM_BARS; i++) {
622 bar = i + PCI_STD_RESOURCES;
623 if (!vdev->barmap[bar])
625 pci_iounmap(pdev, vdev->barmap[bar]);
626 pci_release_selected_regions(pdev, 1 << bar);
627 vdev->barmap[bar] = NULL;
630 list_for_each_entry_safe(dummy_res, tmp,
631 &vdev->dummy_resources_list, res_next) {
632 list_del(&dummy_res->res_next);
633 release_resource(&dummy_res->resource);
637 vdev->needs_reset = true;
639 vfio_pci_zdev_close_device(vdev);
642 * If we have saved state, restore it. If we can reset the device,
643 * even better. Resetting with current state seems better than
644 * nothing, but saving and restoring current state without reset
647 if (pci_load_and_free_saved_state(pdev, &vdev->pci_saved_state)) {
648 pci_info(pdev, "%s: Couldn't reload saved state\n", __func__);
650 if (!vdev->reset_works)
653 pci_save_state(pdev);
657 * Disable INTx and MSI, presumably to avoid spurious interrupts
658 * during reset. Stolen from pci_reset_function()
660 pci_write_config_word(pdev, PCI_COMMAND, PCI_COMMAND_INTX_DISABLE);
663 * Try to get the locks ourselves to prevent a deadlock. The
664 * success of this is dependent on being able to lock the device,
665 * which is not always possible.
666 * We can not use the "try" reset interface here, which will
667 * overwrite the previously restored configuration information.
669 if (vdev->reset_works && pci_dev_trylock(pdev)) {
670 if (!__pci_reset_function_locked(pdev))
671 vdev->needs_reset = false;
672 pci_dev_unlock(pdev);
675 pci_restore_state(pdev);
677 pci_disable_device(pdev);
679 vfio_pci_dev_set_try_reset(vdev->vdev.dev_set);
681 /* Put the pm-runtime usage counter acquired during enable */
682 if (!disable_idle_d3)
683 pm_runtime_put(&pdev->dev);
685 EXPORT_SYMBOL_GPL(vfio_pci_core_disable);
687 void vfio_pci_core_close_device(struct vfio_device *core_vdev)
689 struct vfio_pci_core_device *vdev =
690 container_of(core_vdev, struct vfio_pci_core_device, vdev);
692 if (vdev->sriov_pf_core_dev) {
693 mutex_lock(&vdev->sriov_pf_core_dev->vf_token->lock);
694 WARN_ON(!vdev->sriov_pf_core_dev->vf_token->users);
695 vdev->sriov_pf_core_dev->vf_token->users--;
696 mutex_unlock(&vdev->sriov_pf_core_dev->vf_token->lock);
698 #if IS_ENABLED(CONFIG_EEH)
699 eeh_dev_release(vdev->pdev);
701 vfio_pci_core_disable(vdev);
703 mutex_lock(&vdev->igate);
704 if (vdev->err_trigger) {
705 eventfd_ctx_put(vdev->err_trigger);
706 vdev->err_trigger = NULL;
708 if (vdev->req_trigger) {
709 eventfd_ctx_put(vdev->req_trigger);
710 vdev->req_trigger = NULL;
712 mutex_unlock(&vdev->igate);
714 EXPORT_SYMBOL_GPL(vfio_pci_core_close_device);
716 void vfio_pci_core_finish_enable(struct vfio_pci_core_device *vdev)
718 vfio_pci_probe_mmaps(vdev);
719 #if IS_ENABLED(CONFIG_EEH)
720 eeh_dev_open(vdev->pdev);
723 if (vdev->sriov_pf_core_dev) {
724 mutex_lock(&vdev->sriov_pf_core_dev->vf_token->lock);
725 vdev->sriov_pf_core_dev->vf_token->users++;
726 mutex_unlock(&vdev->sriov_pf_core_dev->vf_token->lock);
729 EXPORT_SYMBOL_GPL(vfio_pci_core_finish_enable);
731 static int vfio_pci_get_irq_count(struct vfio_pci_core_device *vdev, int irq_type)
733 if (irq_type == VFIO_PCI_INTX_IRQ_INDEX) {
736 if (!IS_ENABLED(CONFIG_VFIO_PCI_INTX) ||
737 vdev->nointx || vdev->pdev->is_virtfn)
740 pci_read_config_byte(vdev->pdev, PCI_INTERRUPT_PIN, &pin);
743 } else if (irq_type == VFIO_PCI_MSI_IRQ_INDEX) {
747 pos = vdev->pdev->msi_cap;
749 pci_read_config_word(vdev->pdev,
750 pos + PCI_MSI_FLAGS, &flags);
751 return 1 << ((flags & PCI_MSI_FLAGS_QMASK) >> 1);
753 } else if (irq_type == VFIO_PCI_MSIX_IRQ_INDEX) {
757 pos = vdev->pdev->msix_cap;
759 pci_read_config_word(vdev->pdev,
760 pos + PCI_MSIX_FLAGS, &flags);
762 return (flags & PCI_MSIX_FLAGS_QSIZE) + 1;
764 } else if (irq_type == VFIO_PCI_ERR_IRQ_INDEX) {
765 if (pci_is_pcie(vdev->pdev))
767 } else if (irq_type == VFIO_PCI_REQ_IRQ_INDEX) {
774 static int vfio_pci_count_devs(struct pci_dev *pdev, void *data)
780 struct vfio_pci_fill_info {
781 struct vfio_device *vdev;
782 struct vfio_pci_dependent_device *devices;
788 static int vfio_pci_fill_devs(struct pci_dev *pdev, void *data)
790 struct vfio_pci_dependent_device *info;
791 struct vfio_pci_fill_info *fill = data;
793 /* The topology changed since we counted devices */
794 if (fill->count >= fill->nr_devices)
797 info = &fill->devices[fill->count++];
798 info->segment = pci_domain_nr(pdev->bus);
799 info->bus = pdev->bus->number;
800 info->devfn = pdev->devfn;
802 if (fill->flags & VFIO_PCI_HOT_RESET_FLAG_DEV_ID) {
803 struct iommufd_ctx *iommufd = vfio_iommufd_device_ictx(fill->vdev);
804 struct vfio_device_set *dev_set = fill->vdev->dev_set;
805 struct vfio_device *vdev;
808 * hot-reset requires all affected devices be represented in
811 vdev = vfio_find_device_in_devset(dev_set, &pdev->dev);
813 info->devid = VFIO_PCI_DEVID_NOT_OWNED;
815 int id = vfio_iommufd_get_dev_id(vdev, iommufd);
819 else if (id == -ENOENT)
820 info->devid = VFIO_PCI_DEVID_OWNED;
822 info->devid = VFIO_PCI_DEVID_NOT_OWNED;
824 /* If devid is VFIO_PCI_DEVID_NOT_OWNED, clear owned flag. */
825 if (info->devid == VFIO_PCI_DEVID_NOT_OWNED)
826 fill->flags &= ~VFIO_PCI_HOT_RESET_FLAG_DEV_ID_OWNED;
828 struct iommu_group *iommu_group;
830 iommu_group = iommu_group_get(&pdev->dev);
832 return -EPERM; /* Cannot reset non-isolated devices */
834 info->group_id = iommu_group_id(iommu_group);
835 iommu_group_put(iommu_group);
841 struct vfio_pci_group_info {
846 static bool vfio_pci_dev_below_slot(struct pci_dev *pdev, struct pci_slot *slot)
848 for (; pdev; pdev = pdev->bus->self)
849 if (pdev->bus == slot->bus)
850 return (pdev->slot == slot);
854 struct vfio_pci_walk_info {
855 int (*fn)(struct pci_dev *pdev, void *data);
857 struct pci_dev *pdev;
862 static int vfio_pci_walk_wrapper(struct pci_dev *pdev, void *data)
864 struct vfio_pci_walk_info *walk = data;
866 if (!walk->slot || vfio_pci_dev_below_slot(pdev, walk->pdev->slot))
867 walk->ret = walk->fn(pdev, walk->data);
872 static int vfio_pci_for_each_slot_or_bus(struct pci_dev *pdev,
873 int (*fn)(struct pci_dev *,
874 void *data), void *data,
877 struct vfio_pci_walk_info walk = {
878 .fn = fn, .data = data, .pdev = pdev, .slot = slot, .ret = 0,
881 pci_walk_bus(pdev->bus, vfio_pci_walk_wrapper, &walk);
886 static int msix_mmappable_cap(struct vfio_pci_core_device *vdev,
887 struct vfio_info_cap *caps)
889 struct vfio_info_cap_header header = {
890 .id = VFIO_REGION_INFO_CAP_MSIX_MAPPABLE,
894 return vfio_info_add_capability(caps, &header, sizeof(header));
897 int vfio_pci_core_register_dev_region(struct vfio_pci_core_device *vdev,
898 unsigned int type, unsigned int subtype,
899 const struct vfio_pci_regops *ops,
900 size_t size, u32 flags, void *data)
902 struct vfio_pci_region *region;
904 region = krealloc(vdev->region,
905 (vdev->num_regions + 1) * sizeof(*region),
910 vdev->region = region;
911 vdev->region[vdev->num_regions].type = type;
912 vdev->region[vdev->num_regions].subtype = subtype;
913 vdev->region[vdev->num_regions].ops = ops;
914 vdev->region[vdev->num_regions].size = size;
915 vdev->region[vdev->num_regions].flags = flags;
916 vdev->region[vdev->num_regions].data = data;
922 EXPORT_SYMBOL_GPL(vfio_pci_core_register_dev_region);
924 static int vfio_pci_info_atomic_cap(struct vfio_pci_core_device *vdev,
925 struct vfio_info_cap *caps)
927 struct vfio_device_info_cap_pci_atomic_comp cap = {
928 .header.id = VFIO_DEVICE_INFO_CAP_PCI_ATOMIC_COMP,
931 struct pci_dev *pdev = pci_physfn(vdev->pdev);
934 pcie_capability_read_dword(pdev, PCI_EXP_DEVCAP2, &devcap2);
936 if ((devcap2 & PCI_EXP_DEVCAP2_ATOMIC_COMP32) &&
937 !pci_enable_atomic_ops_to_root(pdev, PCI_EXP_DEVCAP2_ATOMIC_COMP32))
938 cap.flags |= VFIO_PCI_ATOMIC_COMP32;
940 if ((devcap2 & PCI_EXP_DEVCAP2_ATOMIC_COMP64) &&
941 !pci_enable_atomic_ops_to_root(pdev, PCI_EXP_DEVCAP2_ATOMIC_COMP64))
942 cap.flags |= VFIO_PCI_ATOMIC_COMP64;
944 if ((devcap2 & PCI_EXP_DEVCAP2_ATOMIC_COMP128) &&
945 !pci_enable_atomic_ops_to_root(pdev,
946 PCI_EXP_DEVCAP2_ATOMIC_COMP128))
947 cap.flags |= VFIO_PCI_ATOMIC_COMP128;
952 return vfio_info_add_capability(caps, &cap.header, sizeof(cap));
955 static int vfio_pci_ioctl_get_info(struct vfio_pci_core_device *vdev,
956 struct vfio_device_info __user *arg)
958 unsigned long minsz = offsetofend(struct vfio_device_info, num_irqs);
959 struct vfio_device_info info = {};
960 struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
963 if (copy_from_user(&info, arg, minsz))
966 if (info.argsz < minsz)
969 minsz = min_t(size_t, info.argsz, sizeof(info));
971 info.flags = VFIO_DEVICE_FLAGS_PCI;
973 if (vdev->reset_works)
974 info.flags |= VFIO_DEVICE_FLAGS_RESET;
976 info.num_regions = VFIO_PCI_NUM_REGIONS + vdev->num_regions;
977 info.num_irqs = VFIO_PCI_NUM_IRQS;
979 ret = vfio_pci_info_zdev_add_caps(vdev, &caps);
980 if (ret && ret != -ENODEV) {
982 "Failed to setup zPCI info capabilities\n");
986 ret = vfio_pci_info_atomic_cap(vdev, &caps);
987 if (ret && ret != -ENODEV) {
989 "Failed to setup AtomicOps info capability\n");
994 info.flags |= VFIO_DEVICE_FLAGS_CAPS;
995 if (info.argsz < sizeof(info) + caps.size) {
996 info.argsz = sizeof(info) + caps.size;
998 vfio_info_cap_shift(&caps, sizeof(info));
999 if (copy_to_user(arg + 1, caps.buf, caps.size)) {
1003 info.cap_offset = sizeof(*arg);
1009 return copy_to_user(arg, &info, minsz) ? -EFAULT : 0;
1012 static int vfio_pci_ioctl_get_region_info(struct vfio_pci_core_device *vdev,
1013 struct vfio_region_info __user *arg)
1015 unsigned long minsz = offsetofend(struct vfio_region_info, offset);
1016 struct pci_dev *pdev = vdev->pdev;
1017 struct vfio_region_info info;
1018 struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
1021 if (copy_from_user(&info, arg, minsz))
1024 if (info.argsz < minsz)
1027 switch (info.index) {
1028 case VFIO_PCI_CONFIG_REGION_INDEX:
1029 info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
1030 info.size = pdev->cfg_size;
1031 info.flags = VFIO_REGION_INFO_FLAG_READ |
1032 VFIO_REGION_INFO_FLAG_WRITE;
1034 case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
1035 info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
1036 info.size = pci_resource_len(pdev, info.index);
1042 info.flags = VFIO_REGION_INFO_FLAG_READ |
1043 VFIO_REGION_INFO_FLAG_WRITE;
1044 if (vdev->bar_mmap_supported[info.index]) {
1045 info.flags |= VFIO_REGION_INFO_FLAG_MMAP;
1046 if (info.index == vdev->msix_bar) {
1047 ret = msix_mmappable_cap(vdev, &caps);
1054 case VFIO_PCI_ROM_REGION_INDEX: {
1059 info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
1062 /* Report the BAR size, not the ROM size */
1063 info.size = pci_resource_len(pdev, info.index);
1065 /* Shadow ROMs appear as PCI option ROMs */
1066 if (pdev->resource[PCI_ROM_RESOURCE].flags &
1067 IORESOURCE_ROM_SHADOW)
1068 info.size = 0x20000;
1074 * Is it really there? Enable memory decode for implicit access
1077 cmd = vfio_pci_memory_lock_and_enable(vdev);
1078 io = pci_map_rom(pdev, &size);
1080 info.flags = VFIO_REGION_INFO_FLAG_READ;
1081 pci_unmap_rom(pdev, io);
1085 vfio_pci_memory_unlock_and_restore(vdev, cmd);
1089 case VFIO_PCI_VGA_REGION_INDEX:
1093 info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
1094 info.size = 0xc0000;
1095 info.flags = VFIO_REGION_INFO_FLAG_READ |
1096 VFIO_REGION_INFO_FLAG_WRITE;
1100 struct vfio_region_info_cap_type cap_type = {
1101 .header.id = VFIO_REGION_INFO_CAP_TYPE,
1105 if (info.index >= VFIO_PCI_NUM_REGIONS + vdev->num_regions)
1107 info.index = array_index_nospec(
1108 info.index, VFIO_PCI_NUM_REGIONS + vdev->num_regions);
1110 i = info.index - VFIO_PCI_NUM_REGIONS;
1112 info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
1113 info.size = vdev->region[i].size;
1114 info.flags = vdev->region[i].flags;
1116 cap_type.type = vdev->region[i].type;
1117 cap_type.subtype = vdev->region[i].subtype;
1119 ret = vfio_info_add_capability(&caps, &cap_type.header,
1124 if (vdev->region[i].ops->add_capability) {
1125 ret = vdev->region[i].ops->add_capability(
1126 vdev, &vdev->region[i], &caps);
1134 info.flags |= VFIO_REGION_INFO_FLAG_CAPS;
1135 if (info.argsz < sizeof(info) + caps.size) {
1136 info.argsz = sizeof(info) + caps.size;
1137 info.cap_offset = 0;
1139 vfio_info_cap_shift(&caps, sizeof(info));
1140 if (copy_to_user(arg + 1, caps.buf, caps.size)) {
1144 info.cap_offset = sizeof(*arg);
1150 return copy_to_user(arg, &info, minsz) ? -EFAULT : 0;
1153 static int vfio_pci_ioctl_get_irq_info(struct vfio_pci_core_device *vdev,
1154 struct vfio_irq_info __user *arg)
1156 unsigned long minsz = offsetofend(struct vfio_irq_info, count);
1157 struct vfio_irq_info info;
1159 if (copy_from_user(&info, arg, minsz))
1162 if (info.argsz < minsz || info.index >= VFIO_PCI_NUM_IRQS)
1165 switch (info.index) {
1166 case VFIO_PCI_INTX_IRQ_INDEX ... VFIO_PCI_MSIX_IRQ_INDEX:
1167 case VFIO_PCI_REQ_IRQ_INDEX:
1169 case VFIO_PCI_ERR_IRQ_INDEX:
1170 if (pci_is_pcie(vdev->pdev))
1177 info.flags = VFIO_IRQ_INFO_EVENTFD;
1179 info.count = vfio_pci_get_irq_count(vdev, info.index);
1181 if (info.index == VFIO_PCI_INTX_IRQ_INDEX)
1183 (VFIO_IRQ_INFO_MASKABLE | VFIO_IRQ_INFO_AUTOMASKED);
1184 else if (info.index != VFIO_PCI_MSIX_IRQ_INDEX || !vdev->has_dyn_msix)
1185 info.flags |= VFIO_IRQ_INFO_NORESIZE;
1187 return copy_to_user(arg, &info, minsz) ? -EFAULT : 0;
1190 static int vfio_pci_ioctl_set_irqs(struct vfio_pci_core_device *vdev,
1191 struct vfio_irq_set __user *arg)
1193 unsigned long minsz = offsetofend(struct vfio_irq_set, count);
1194 struct vfio_irq_set hdr;
1197 size_t data_size = 0;
1199 if (copy_from_user(&hdr, arg, minsz))
1202 max = vfio_pci_get_irq_count(vdev, hdr.index);
1204 ret = vfio_set_irqs_validate_and_prepare(&hdr, max, VFIO_PCI_NUM_IRQS,
1210 data = memdup_user(&arg->data, data_size);
1212 return PTR_ERR(data);
1215 mutex_lock(&vdev->igate);
1217 ret = vfio_pci_set_irqs_ioctl(vdev, hdr.flags, hdr.index, hdr.start,
1220 mutex_unlock(&vdev->igate);
1226 static int vfio_pci_ioctl_reset(struct vfio_pci_core_device *vdev,
1231 if (!vdev->reset_works)
1234 vfio_pci_zap_and_down_write_memory_lock(vdev);
1237 * This function can be invoked while the power state is non-D0. If
1238 * pci_try_reset_function() has been called while the power state is
1239 * non-D0, then pci_try_reset_function() will internally set the power
1240 * state to D0 without vfio driver involvement. For the devices which
1241 * have NoSoftRst-, the reset function can cause the PCI config space
1242 * reset without restoring the original state (saved locally in
1245 vfio_pci_set_power_state(vdev, PCI_D0);
1247 ret = pci_try_reset_function(vdev->pdev);
1248 up_write(&vdev->memory_lock);
1253 static int vfio_pci_ioctl_get_pci_hot_reset_info(
1254 struct vfio_pci_core_device *vdev,
1255 struct vfio_pci_hot_reset_info __user *arg)
1257 unsigned long minsz =
1258 offsetofend(struct vfio_pci_hot_reset_info, count);
1259 struct vfio_pci_dependent_device *devices = NULL;
1260 struct vfio_pci_hot_reset_info hdr;
1261 struct vfio_pci_fill_info fill = {};
1265 if (copy_from_user(&hdr, arg, minsz))
1268 if (hdr.argsz < minsz)
1273 /* Can we do a slot or bus reset or neither? */
1274 if (!pci_probe_reset_slot(vdev->pdev->slot))
1276 else if (pci_probe_reset_bus(vdev->pdev->bus))
1279 ret = vfio_pci_for_each_slot_or_bus(vdev->pdev, vfio_pci_count_devs,
1284 if (WARN_ON(!count)) /* Should always be at least one */
1287 if (count > (hdr.argsz - sizeof(hdr)) / sizeof(*devices)) {
1293 devices = kcalloc(count, sizeof(*devices), GFP_KERNEL);
1297 fill.devices = devices;
1298 fill.nr_devices = count;
1299 fill.vdev = &vdev->vdev;
1301 if (vfio_device_cdev_opened(&vdev->vdev))
1302 fill.flags |= VFIO_PCI_HOT_RESET_FLAG_DEV_ID |
1303 VFIO_PCI_HOT_RESET_FLAG_DEV_ID_OWNED;
1305 mutex_lock(&vdev->vdev.dev_set->lock);
1306 ret = vfio_pci_for_each_slot_or_bus(vdev->pdev, vfio_pci_fill_devs,
1308 mutex_unlock(&vdev->vdev.dev_set->lock);
1312 if (copy_to_user(arg->devices, devices,
1313 sizeof(*devices) * fill.count)) {
1318 hdr.count = fill.count;
1319 hdr.flags = fill.flags;
1322 if (copy_to_user(arg, &hdr, minsz))
1330 vfio_pci_ioctl_pci_hot_reset_groups(struct vfio_pci_core_device *vdev,
1331 int array_count, bool slot,
1332 struct vfio_pci_hot_reset __user *arg)
1335 struct file **files;
1336 struct vfio_pci_group_info info;
1337 int file_idx, count = 0, ret = 0;
1340 * We can't let userspace give us an arbitrarily large buffer to copy,
1341 * so verify how many we think there could be. Note groups can have
1342 * multiple devices so one group per device is the max.
1344 ret = vfio_pci_for_each_slot_or_bus(vdev->pdev, vfio_pci_count_devs,
1349 if (array_count > count)
1352 group_fds = kcalloc(array_count, sizeof(*group_fds), GFP_KERNEL);
1353 files = kcalloc(array_count, sizeof(*files), GFP_KERNEL);
1354 if (!group_fds || !files) {
1360 if (copy_from_user(group_fds, arg->group_fds,
1361 array_count * sizeof(*group_fds))) {
1368 * Get the group file for each fd to ensure the group is held across
1371 for (file_idx = 0; file_idx < array_count; file_idx++) {
1372 struct file *file = fget(group_fds[file_idx]);
1379 /* Ensure the FD is a vfio group FD.*/
1380 if (!vfio_file_is_group(file)) {
1386 files[file_idx] = file;
1391 /* release reference to groups on error */
1393 goto hot_reset_release;
1395 info.count = array_count;
1398 ret = vfio_pci_dev_set_hot_reset(vdev->vdev.dev_set, &info, NULL);
1401 for (file_idx--; file_idx >= 0; file_idx--)
1402 fput(files[file_idx]);
1408 static int vfio_pci_ioctl_pci_hot_reset(struct vfio_pci_core_device *vdev,
1409 struct vfio_pci_hot_reset __user *arg)
1411 unsigned long minsz = offsetofend(struct vfio_pci_hot_reset, count);
1412 struct vfio_pci_hot_reset hdr;
1415 if (copy_from_user(&hdr, arg, minsz))
1418 if (hdr.argsz < minsz || hdr.flags)
1421 /* zero-length array is only for cdev opened devices */
1422 if (!!hdr.count == vfio_device_cdev_opened(&vdev->vdev))
1425 /* Can we do a slot or bus reset or neither? */
1426 if (!pci_probe_reset_slot(vdev->pdev->slot))
1428 else if (pci_probe_reset_bus(vdev->pdev->bus))
1432 return vfio_pci_ioctl_pci_hot_reset_groups(vdev, hdr.count, slot, arg);
1434 return vfio_pci_dev_set_hot_reset(vdev->vdev.dev_set, NULL,
1435 vfio_iommufd_device_ictx(&vdev->vdev));
1438 static int vfio_pci_ioctl_ioeventfd(struct vfio_pci_core_device *vdev,
1439 struct vfio_device_ioeventfd __user *arg)
1441 unsigned long minsz = offsetofend(struct vfio_device_ioeventfd, fd);
1442 struct vfio_device_ioeventfd ioeventfd;
1445 if (copy_from_user(&ioeventfd, arg, minsz))
1448 if (ioeventfd.argsz < minsz)
1451 if (ioeventfd.flags & ~VFIO_DEVICE_IOEVENTFD_SIZE_MASK)
1454 count = ioeventfd.flags & VFIO_DEVICE_IOEVENTFD_SIZE_MASK;
1456 if (hweight8(count) != 1 || ioeventfd.fd < -1)
1459 return vfio_pci_ioeventfd(vdev, ioeventfd.offset, ioeventfd.data, count,
1463 long vfio_pci_core_ioctl(struct vfio_device *core_vdev, unsigned int cmd,
1466 struct vfio_pci_core_device *vdev =
1467 container_of(core_vdev, struct vfio_pci_core_device, vdev);
1468 void __user *uarg = (void __user *)arg;
1471 case VFIO_DEVICE_GET_INFO:
1472 return vfio_pci_ioctl_get_info(vdev, uarg);
1473 case VFIO_DEVICE_GET_IRQ_INFO:
1474 return vfio_pci_ioctl_get_irq_info(vdev, uarg);
1475 case VFIO_DEVICE_GET_PCI_HOT_RESET_INFO:
1476 return vfio_pci_ioctl_get_pci_hot_reset_info(vdev, uarg);
1477 case VFIO_DEVICE_GET_REGION_INFO:
1478 return vfio_pci_ioctl_get_region_info(vdev, uarg);
1479 case VFIO_DEVICE_IOEVENTFD:
1480 return vfio_pci_ioctl_ioeventfd(vdev, uarg);
1481 case VFIO_DEVICE_PCI_HOT_RESET:
1482 return vfio_pci_ioctl_pci_hot_reset(vdev, uarg);
1483 case VFIO_DEVICE_RESET:
1484 return vfio_pci_ioctl_reset(vdev, uarg);
1485 case VFIO_DEVICE_SET_IRQS:
1486 return vfio_pci_ioctl_set_irqs(vdev, uarg);
1491 EXPORT_SYMBOL_GPL(vfio_pci_core_ioctl);
1493 static int vfio_pci_core_feature_token(struct vfio_device *device, u32 flags,
1494 uuid_t __user *arg, size_t argsz)
1496 struct vfio_pci_core_device *vdev =
1497 container_of(device, struct vfio_pci_core_device, vdev);
1501 if (!vdev->vf_token)
1504 * We do not support GET of the VF Token UUID as this could
1505 * expose the token of the previous device user.
1507 ret = vfio_check_feature(flags, argsz, VFIO_DEVICE_FEATURE_SET,
1512 if (copy_from_user(&uuid, arg, sizeof(uuid)))
1515 mutex_lock(&vdev->vf_token->lock);
1516 uuid_copy(&vdev->vf_token->uuid, &uuid);
1517 mutex_unlock(&vdev->vf_token->lock);
1521 int vfio_pci_core_ioctl_feature(struct vfio_device *device, u32 flags,
1522 void __user *arg, size_t argsz)
1524 switch (flags & VFIO_DEVICE_FEATURE_MASK) {
1525 case VFIO_DEVICE_FEATURE_LOW_POWER_ENTRY:
1526 return vfio_pci_core_pm_entry(device, flags, arg, argsz);
1527 case VFIO_DEVICE_FEATURE_LOW_POWER_ENTRY_WITH_WAKEUP:
1528 return vfio_pci_core_pm_entry_with_wakeup(device, flags,
1530 case VFIO_DEVICE_FEATURE_LOW_POWER_EXIT:
1531 return vfio_pci_core_pm_exit(device, flags, arg, argsz);
1532 case VFIO_DEVICE_FEATURE_PCI_VF_TOKEN:
1533 return vfio_pci_core_feature_token(device, flags, arg, argsz);
1538 EXPORT_SYMBOL_GPL(vfio_pci_core_ioctl_feature);
1540 static ssize_t vfio_pci_rw(struct vfio_pci_core_device *vdev, char __user *buf,
1541 size_t count, loff_t *ppos, bool iswrite)
1543 unsigned int index = VFIO_PCI_OFFSET_TO_INDEX(*ppos);
1546 if (index >= VFIO_PCI_NUM_REGIONS + vdev->num_regions)
1549 ret = pm_runtime_resume_and_get(&vdev->pdev->dev);
1551 pci_info_ratelimited(vdev->pdev, "runtime resume failed %d\n",
1557 case VFIO_PCI_CONFIG_REGION_INDEX:
1558 ret = vfio_pci_config_rw(vdev, buf, count, ppos, iswrite);
1561 case VFIO_PCI_ROM_REGION_INDEX:
1565 ret = vfio_pci_bar_rw(vdev, buf, count, ppos, false);
1568 case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
1569 ret = vfio_pci_bar_rw(vdev, buf, count, ppos, iswrite);
1572 case VFIO_PCI_VGA_REGION_INDEX:
1573 ret = vfio_pci_vga_rw(vdev, buf, count, ppos, iswrite);
1577 index -= VFIO_PCI_NUM_REGIONS;
1578 ret = vdev->region[index].ops->rw(vdev, buf,
1579 count, ppos, iswrite);
1583 pm_runtime_put(&vdev->pdev->dev);
1587 ssize_t vfio_pci_core_read(struct vfio_device *core_vdev, char __user *buf,
1588 size_t count, loff_t *ppos)
1590 struct vfio_pci_core_device *vdev =
1591 container_of(core_vdev, struct vfio_pci_core_device, vdev);
1596 return vfio_pci_rw(vdev, buf, count, ppos, false);
1598 EXPORT_SYMBOL_GPL(vfio_pci_core_read);
1600 ssize_t vfio_pci_core_write(struct vfio_device *core_vdev, const char __user *buf,
1601 size_t count, loff_t *ppos)
1603 struct vfio_pci_core_device *vdev =
1604 container_of(core_vdev, struct vfio_pci_core_device, vdev);
1609 return vfio_pci_rw(vdev, (char __user *)buf, count, ppos, true);
1611 EXPORT_SYMBOL_GPL(vfio_pci_core_write);
1613 /* Return 1 on zap and vma_lock acquired, 0 on contention (only with @try) */
1614 static int vfio_pci_zap_and_vma_lock(struct vfio_pci_core_device *vdev, bool try)
1616 struct vfio_pci_mmap_vma *mmap_vma, *tmp;
1620 * vma_lock is nested under mmap_lock for vm_ops callback paths.
1621 * The memory_lock semaphore is used by both code paths calling
1622 * into this function to zap vmas and the vm_ops.fault callback
1623 * to protect the memory enable state of the device.
1625 * When zapping vmas we need to maintain the mmap_lock => vma_lock
1626 * ordering, which requires using vma_lock to walk vma_list to
1627 * acquire an mm, then dropping vma_lock to get the mmap_lock and
1628 * reacquiring vma_lock. This logic is derived from similar
1629 * requirements in uverbs_user_mmap_disassociate().
1631 * mmap_lock must always be the top-level lock when it is taken.
1632 * Therefore we can only hold the memory_lock write lock when
1633 * vma_list is empty, as we'd need to take mmap_lock to clear
1634 * entries. vma_list can only be guaranteed empty when holding
1635 * vma_lock, thus memory_lock is nested under vma_lock.
1637 * This enables the vm_ops.fault callback to acquire vma_lock,
1638 * followed by memory_lock read lock, while already holding
1639 * mmap_lock without risk of deadlock.
1642 struct mm_struct *mm = NULL;
1645 if (!mutex_trylock(&vdev->vma_lock))
1648 mutex_lock(&vdev->vma_lock);
1650 while (!list_empty(&vdev->vma_list)) {
1651 mmap_vma = list_first_entry(&vdev->vma_list,
1652 struct vfio_pci_mmap_vma,
1654 mm = mmap_vma->vma->vm_mm;
1655 if (mmget_not_zero(mm))
1658 list_del(&mmap_vma->vma_next);
1664 mutex_unlock(&vdev->vma_lock);
1667 if (!mmap_read_trylock(mm)) {
1675 if (!mutex_trylock(&vdev->vma_lock)) {
1676 mmap_read_unlock(mm);
1681 mutex_lock(&vdev->vma_lock);
1683 list_for_each_entry_safe(mmap_vma, tmp,
1684 &vdev->vma_list, vma_next) {
1685 struct vm_area_struct *vma = mmap_vma->vma;
1687 if (vma->vm_mm != mm)
1690 list_del(&mmap_vma->vma_next);
1693 zap_vma_ptes(vma, vma->vm_start,
1694 vma->vm_end - vma->vm_start);
1696 mutex_unlock(&vdev->vma_lock);
1697 mmap_read_unlock(mm);
1702 void vfio_pci_zap_and_down_write_memory_lock(struct vfio_pci_core_device *vdev)
1704 vfio_pci_zap_and_vma_lock(vdev, false);
1705 down_write(&vdev->memory_lock);
1706 mutex_unlock(&vdev->vma_lock);
1709 u16 vfio_pci_memory_lock_and_enable(struct vfio_pci_core_device *vdev)
1713 down_write(&vdev->memory_lock);
1714 pci_read_config_word(vdev->pdev, PCI_COMMAND, &cmd);
1715 if (!(cmd & PCI_COMMAND_MEMORY))
1716 pci_write_config_word(vdev->pdev, PCI_COMMAND,
1717 cmd | PCI_COMMAND_MEMORY);
1722 void vfio_pci_memory_unlock_and_restore(struct vfio_pci_core_device *vdev, u16 cmd)
1724 pci_write_config_word(vdev->pdev, PCI_COMMAND, cmd);
1725 up_write(&vdev->memory_lock);
1728 /* Caller holds vma_lock */
1729 static int __vfio_pci_add_vma(struct vfio_pci_core_device *vdev,
1730 struct vm_area_struct *vma)
1732 struct vfio_pci_mmap_vma *mmap_vma;
1734 mmap_vma = kmalloc(sizeof(*mmap_vma), GFP_KERNEL_ACCOUNT);
1738 mmap_vma->vma = vma;
1739 list_add(&mmap_vma->vma_next, &vdev->vma_list);
1745 * Zap mmaps on open so that we can fault them in on access and therefore
1746 * our vma_list only tracks mappings accessed since last zap.
1748 static void vfio_pci_mmap_open(struct vm_area_struct *vma)
1750 zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
1753 static void vfio_pci_mmap_close(struct vm_area_struct *vma)
1755 struct vfio_pci_core_device *vdev = vma->vm_private_data;
1756 struct vfio_pci_mmap_vma *mmap_vma;
1758 mutex_lock(&vdev->vma_lock);
1759 list_for_each_entry(mmap_vma, &vdev->vma_list, vma_next) {
1760 if (mmap_vma->vma == vma) {
1761 list_del(&mmap_vma->vma_next);
1766 mutex_unlock(&vdev->vma_lock);
1769 static vm_fault_t vfio_pci_mmap_fault(struct vm_fault *vmf)
1771 struct vm_area_struct *vma = vmf->vma;
1772 struct vfio_pci_core_device *vdev = vma->vm_private_data;
1773 struct vfio_pci_mmap_vma *mmap_vma;
1774 vm_fault_t ret = VM_FAULT_NOPAGE;
1776 mutex_lock(&vdev->vma_lock);
1777 down_read(&vdev->memory_lock);
1780 * Memory region cannot be accessed if the low power feature is engaged
1781 * or memory access is disabled.
1783 if (vdev->pm_runtime_engaged || !__vfio_pci_memory_enabled(vdev)) {
1784 ret = VM_FAULT_SIGBUS;
1789 * We populate the whole vma on fault, so we need to test whether
1790 * the vma has already been mapped, such as for concurrent faults
1791 * to the same vma. io_remap_pfn_range() will trigger a BUG_ON if
1792 * we ask it to fill the same range again.
1794 list_for_each_entry(mmap_vma, &vdev->vma_list, vma_next) {
1795 if (mmap_vma->vma == vma)
1799 if (io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
1800 vma->vm_end - vma->vm_start,
1801 vma->vm_page_prot)) {
1802 ret = VM_FAULT_SIGBUS;
1803 zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
1807 if (__vfio_pci_add_vma(vdev, vma)) {
1809 zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
1813 up_read(&vdev->memory_lock);
1814 mutex_unlock(&vdev->vma_lock);
1818 static const struct vm_operations_struct vfio_pci_mmap_ops = {
1819 .open = vfio_pci_mmap_open,
1820 .close = vfio_pci_mmap_close,
1821 .fault = vfio_pci_mmap_fault,
1824 int vfio_pci_core_mmap(struct vfio_device *core_vdev, struct vm_area_struct *vma)
1826 struct vfio_pci_core_device *vdev =
1827 container_of(core_vdev, struct vfio_pci_core_device, vdev);
1828 struct pci_dev *pdev = vdev->pdev;
1830 u64 phys_len, req_len, pgoff, req_start;
1833 index = vma->vm_pgoff >> (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT);
1835 if (index >= VFIO_PCI_NUM_REGIONS + vdev->num_regions)
1837 if (vma->vm_end < vma->vm_start)
1839 if ((vma->vm_flags & VM_SHARED) == 0)
1841 if (index >= VFIO_PCI_NUM_REGIONS) {
1842 int regnum = index - VFIO_PCI_NUM_REGIONS;
1843 struct vfio_pci_region *region = vdev->region + regnum;
1845 if (region->ops && region->ops->mmap &&
1846 (region->flags & VFIO_REGION_INFO_FLAG_MMAP))
1847 return region->ops->mmap(vdev, region, vma);
1850 if (index >= VFIO_PCI_ROM_REGION_INDEX)
1852 if (!vdev->bar_mmap_supported[index])
1855 phys_len = PAGE_ALIGN(pci_resource_len(pdev, index));
1856 req_len = vma->vm_end - vma->vm_start;
1857 pgoff = vma->vm_pgoff &
1858 ((1U << (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT)) - 1);
1859 req_start = pgoff << PAGE_SHIFT;
1861 if (req_start + req_len > phys_len)
1865 * Even though we don't make use of the barmap for the mmap,
1866 * we need to request the region and the barmap tracks that.
1868 if (!vdev->barmap[index]) {
1869 ret = pci_request_selected_regions(pdev,
1870 1 << index, "vfio-pci");
1874 vdev->barmap[index] = pci_iomap(pdev, index, 0);
1875 if (!vdev->barmap[index]) {
1876 pci_release_selected_regions(pdev, 1 << index);
1881 vma->vm_private_data = vdev;
1882 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1883 vma->vm_pgoff = (pci_resource_start(pdev, index) >> PAGE_SHIFT) + pgoff;
1886 * See remap_pfn_range(), called from vfio_pci_fault() but we can't
1887 * change vm_flags within the fault handler. Set them now.
1889 * VM_ALLOW_ANY_UNCACHED: The VMA flag is implemented for ARM64,
1890 * allowing KVM stage 2 device mapping attributes to use Normal-NC
1891 * rather than DEVICE_nGnRE, which allows guest mappings
1892 * supporting write-combining attributes (WC). ARM does not
1893 * architecturally guarantee this is safe, and indeed some MMIO
1894 * regions like the GICv2 VCPU interface can trigger uncontained
1895 * faults if Normal-NC is used.
1897 * To safely use VFIO in KVM the platform must guarantee full
1898 * safety in the guest where no action taken against a MMIO
1899 * mapping can trigger an uncontained failure. The assumption is
1900 * that most VFIO PCI platforms support this for both mapping types,
1901 * at least in common flows, based on some expectations of how
1902 * PCI IP is integrated. Hence VM_ALLOW_ANY_UNCACHED is set in
1905 vm_flags_set(vma, VM_ALLOW_ANY_UNCACHED | VM_IO | VM_PFNMAP |
1906 VM_DONTEXPAND | VM_DONTDUMP);
1907 vma->vm_ops = &vfio_pci_mmap_ops;
1911 EXPORT_SYMBOL_GPL(vfio_pci_core_mmap);
1913 void vfio_pci_core_request(struct vfio_device *core_vdev, unsigned int count)
1915 struct vfio_pci_core_device *vdev =
1916 container_of(core_vdev, struct vfio_pci_core_device, vdev);
1917 struct pci_dev *pdev = vdev->pdev;
1919 mutex_lock(&vdev->igate);
1921 if (vdev->req_trigger) {
1923 pci_notice_ratelimited(pdev,
1924 "Relaying device request to user (#%u)\n",
1926 eventfd_signal(vdev->req_trigger);
1927 } else if (count == 0) {
1929 "No device request channel registered, blocked until released by user\n");
1932 mutex_unlock(&vdev->igate);
1934 EXPORT_SYMBOL_GPL(vfio_pci_core_request);
1936 static int vfio_pci_validate_vf_token(struct vfio_pci_core_device *vdev,
1937 bool vf_token, uuid_t *uuid)
1940 * There's always some degree of trust or collaboration between SR-IOV
1941 * PF and VFs, even if just that the PF hosts the SR-IOV capability and
1942 * can disrupt VFs with a reset, but often the PF has more explicit
1943 * access to deny service to the VF or access data passed through the
1944 * VF. We therefore require an opt-in via a shared VF token (UUID) to
1945 * represent this trust. This both prevents that a VF driver might
1946 * assume the PF driver is a trusted, in-kernel driver, and also that
1947 * a PF driver might be replaced with a rogue driver, unknown to in-use
1950 * Therefore when presented with a VF, if the PF is a vfio device and
1951 * it is bound to the vfio-pci driver, the user needs to provide a VF
1952 * token to access the device, in the form of appending a vf_token to
1953 * the device name, for example:
1955 * "0000:04:10.0 vf_token=bd8d9d2b-5a5f-4f5a-a211-f591514ba1f3"
1957 * When presented with a PF which has VFs in use, the user must also
1958 * provide the current VF token to prove collaboration with existing
1959 * VF users. If VFs are not in use, the VF token provided for the PF
1960 * device will act to set the VF token.
1962 * If the VF token is provided but unused, an error is generated.
1964 if (vdev->pdev->is_virtfn) {
1965 struct vfio_pci_core_device *pf_vdev = vdev->sriov_pf_core_dev;
1970 return 0; /* PF is not vfio-pci, no VF token */
1972 pci_info_ratelimited(vdev->pdev,
1973 "VF token incorrectly provided, PF not bound to vfio-pci\n");
1978 pci_info_ratelimited(vdev->pdev,
1979 "VF token required to access device\n");
1983 mutex_lock(&pf_vdev->vf_token->lock);
1984 match = uuid_equal(uuid, &pf_vdev->vf_token->uuid);
1985 mutex_unlock(&pf_vdev->vf_token->lock);
1988 pci_info_ratelimited(vdev->pdev,
1989 "Incorrect VF token provided for device\n");
1992 } else if (vdev->vf_token) {
1993 mutex_lock(&vdev->vf_token->lock);
1994 if (vdev->vf_token->users) {
1996 mutex_unlock(&vdev->vf_token->lock);
1997 pci_info_ratelimited(vdev->pdev,
1998 "VF token required to access device\n");
2002 if (!uuid_equal(uuid, &vdev->vf_token->uuid)) {
2003 mutex_unlock(&vdev->vf_token->lock);
2004 pci_info_ratelimited(vdev->pdev,
2005 "Incorrect VF token provided for device\n");
2008 } else if (vf_token) {
2009 uuid_copy(&vdev->vf_token->uuid, uuid);
2012 mutex_unlock(&vdev->vf_token->lock);
2013 } else if (vf_token) {
2014 pci_info_ratelimited(vdev->pdev,
2015 "VF token incorrectly provided, not a PF or VF\n");
2022 #define VF_TOKEN_ARG "vf_token="
2024 int vfio_pci_core_match(struct vfio_device *core_vdev, char *buf)
2026 struct vfio_pci_core_device *vdev =
2027 container_of(core_vdev, struct vfio_pci_core_device, vdev);
2028 bool vf_token = false;
2032 if (strncmp(pci_name(vdev->pdev), buf, strlen(pci_name(vdev->pdev))))
2033 return 0; /* No match */
2035 if (strlen(buf) > strlen(pci_name(vdev->pdev))) {
2036 buf += strlen(pci_name(vdev->pdev));
2039 return 0; /* No match: non-whitespace after name */
2047 if (!vf_token && !strncmp(buf, VF_TOKEN_ARG,
2048 strlen(VF_TOKEN_ARG))) {
2049 buf += strlen(VF_TOKEN_ARG);
2051 if (strlen(buf) < UUID_STRING_LEN)
2054 ret = uuid_parse(buf, &uuid);
2059 buf += UUID_STRING_LEN;
2061 /* Unknown/duplicate option */
2067 ret = vfio_pci_validate_vf_token(vdev, vf_token, &uuid);
2071 return 1; /* Match */
2073 EXPORT_SYMBOL_GPL(vfio_pci_core_match);
2075 static int vfio_pci_bus_notifier(struct notifier_block *nb,
2076 unsigned long action, void *data)
2078 struct vfio_pci_core_device *vdev = container_of(nb,
2079 struct vfio_pci_core_device, nb);
2080 struct device *dev = data;
2081 struct pci_dev *pdev = to_pci_dev(dev);
2082 struct pci_dev *physfn = pci_physfn(pdev);
2084 if (action == BUS_NOTIFY_ADD_DEVICE &&
2085 pdev->is_virtfn && physfn == vdev->pdev) {
2086 pci_info(vdev->pdev, "Captured SR-IOV VF %s driver_override\n",
2088 pdev->driver_override = kasprintf(GFP_KERNEL, "%s",
2089 vdev->vdev.ops->name);
2090 WARN_ON(!pdev->driver_override);
2091 } else if (action == BUS_NOTIFY_BOUND_DRIVER &&
2092 pdev->is_virtfn && physfn == vdev->pdev) {
2093 struct pci_driver *drv = pci_dev_driver(pdev);
2095 if (drv && drv != pci_dev_driver(vdev->pdev))
2096 pci_warn(vdev->pdev,
2097 "VF %s bound to driver %s while PF bound to driver %s\n",
2098 pci_name(pdev), drv->name,
2099 pci_dev_driver(vdev->pdev)->name);
2105 static int vfio_pci_vf_init(struct vfio_pci_core_device *vdev)
2107 struct pci_dev *pdev = vdev->pdev;
2108 struct vfio_pci_core_device *cur;
2109 struct pci_dev *physfn;
2112 if (pdev->is_virtfn) {
2114 * If this VF was created by our vfio_pci_core_sriov_configure()
2115 * then we can find the PF vfio_pci_core_device now, and due to
2116 * the locking in pci_disable_sriov() it cannot change until
2117 * this VF device driver is removed.
2119 physfn = pci_physfn(vdev->pdev);
2120 mutex_lock(&vfio_pci_sriov_pfs_mutex);
2121 list_for_each_entry(cur, &vfio_pci_sriov_pfs, sriov_pfs_item) {
2122 if (cur->pdev == physfn) {
2123 vdev->sriov_pf_core_dev = cur;
2127 mutex_unlock(&vfio_pci_sriov_pfs_mutex);
2131 /* Not a SRIOV PF */
2132 if (!pdev->is_physfn)
2135 vdev->vf_token = kzalloc(sizeof(*vdev->vf_token), GFP_KERNEL);
2136 if (!vdev->vf_token)
2139 mutex_init(&vdev->vf_token->lock);
2140 uuid_gen(&vdev->vf_token->uuid);
2142 vdev->nb.notifier_call = vfio_pci_bus_notifier;
2143 ret = bus_register_notifier(&pci_bus_type, &vdev->nb);
2145 kfree(vdev->vf_token);
2151 static void vfio_pci_vf_uninit(struct vfio_pci_core_device *vdev)
2153 if (!vdev->vf_token)
2156 bus_unregister_notifier(&pci_bus_type, &vdev->nb);
2157 WARN_ON(vdev->vf_token->users);
2158 mutex_destroy(&vdev->vf_token->lock);
2159 kfree(vdev->vf_token);
2162 static int vfio_pci_vga_init(struct vfio_pci_core_device *vdev)
2164 struct pci_dev *pdev = vdev->pdev;
2167 if (!vfio_pci_is_vga(pdev))
2170 ret = aperture_remove_conflicting_pci_devices(pdev, vdev->vdev.ops->name);
2174 ret = vga_client_register(pdev, vfio_pci_set_decode);
2177 vga_set_legacy_decoding(pdev, vfio_pci_set_decode(pdev, false));
2181 static void vfio_pci_vga_uninit(struct vfio_pci_core_device *vdev)
2183 struct pci_dev *pdev = vdev->pdev;
2185 if (!vfio_pci_is_vga(pdev))
2187 vga_client_unregister(pdev);
2188 vga_set_legacy_decoding(pdev, VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM |
2189 VGA_RSRC_LEGACY_IO |
2190 VGA_RSRC_LEGACY_MEM);
2193 int vfio_pci_core_init_dev(struct vfio_device *core_vdev)
2195 struct vfio_pci_core_device *vdev =
2196 container_of(core_vdev, struct vfio_pci_core_device, vdev);
2198 vdev->pdev = to_pci_dev(core_vdev->dev);
2199 vdev->irq_type = VFIO_PCI_NUM_IRQS;
2200 mutex_init(&vdev->igate);
2201 spin_lock_init(&vdev->irqlock);
2202 mutex_init(&vdev->ioeventfds_lock);
2203 INIT_LIST_HEAD(&vdev->dummy_resources_list);
2204 INIT_LIST_HEAD(&vdev->ioeventfds_list);
2205 mutex_init(&vdev->vma_lock);
2206 INIT_LIST_HEAD(&vdev->vma_list);
2207 INIT_LIST_HEAD(&vdev->sriov_pfs_item);
2208 init_rwsem(&vdev->memory_lock);
2209 xa_init(&vdev->ctx);
2213 EXPORT_SYMBOL_GPL(vfio_pci_core_init_dev);
2215 void vfio_pci_core_release_dev(struct vfio_device *core_vdev)
2217 struct vfio_pci_core_device *vdev =
2218 container_of(core_vdev, struct vfio_pci_core_device, vdev);
2220 mutex_destroy(&vdev->igate);
2221 mutex_destroy(&vdev->ioeventfds_lock);
2222 mutex_destroy(&vdev->vma_lock);
2223 kfree(vdev->region);
2224 kfree(vdev->pm_save);
2226 EXPORT_SYMBOL_GPL(vfio_pci_core_release_dev);
2228 int vfio_pci_core_register_device(struct vfio_pci_core_device *vdev)
2230 struct pci_dev *pdev = vdev->pdev;
2231 struct device *dev = &pdev->dev;
2234 /* Drivers must set the vfio_pci_core_device to their drvdata */
2235 if (WARN_ON(vdev != dev_get_drvdata(dev)))
2238 if (pdev->hdr_type != PCI_HEADER_TYPE_NORMAL)
2241 if (vdev->vdev.mig_ops) {
2242 if (!(vdev->vdev.mig_ops->migration_get_state &&
2243 vdev->vdev.mig_ops->migration_set_state &&
2244 vdev->vdev.mig_ops->migration_get_data_size) ||
2245 !(vdev->vdev.migration_flags & VFIO_MIGRATION_STOP_COPY))
2249 if (vdev->vdev.log_ops && !(vdev->vdev.log_ops->log_start &&
2250 vdev->vdev.log_ops->log_stop &&
2251 vdev->vdev.log_ops->log_read_and_clear))
2255 * Prevent binding to PFs with VFs enabled, the VFs might be in use
2256 * by the host or other users. We cannot capture the VFs if they
2257 * already exist, nor can we track VF users. Disabling SR-IOV here
2258 * would initiate removing the VFs, which would unbind the driver,
2259 * which is prone to blocking if that VF is also in use by vfio-pci.
2260 * Just reject these PFs and let the user sort it out.
2262 if (pci_num_vf(pdev)) {
2263 pci_warn(pdev, "Cannot bind to PF with SR-IOV enabled\n");
2267 if (pci_is_root_bus(pdev->bus)) {
2268 ret = vfio_assign_device_set(&vdev->vdev, vdev);
2269 } else if (!pci_probe_reset_slot(pdev->slot)) {
2270 ret = vfio_assign_device_set(&vdev->vdev, pdev->slot);
2273 * If there is no slot reset support for this device, the whole
2274 * bus needs to be grouped together to support bus-wide resets.
2276 ret = vfio_assign_device_set(&vdev->vdev, pdev->bus);
2281 ret = vfio_pci_vf_init(vdev);
2284 ret = vfio_pci_vga_init(vdev);
2288 vfio_pci_probe_power_state(vdev);
2291 * pci-core sets the device power state to an unknown value at
2292 * bootup and after being removed from a driver. The only
2293 * transition it allows from this unknown state is to D0, which
2294 * typically happens when a driver calls pci_enable_device().
2295 * We're not ready to enable the device yet, but we do want to
2296 * be able to get to D3. Therefore first do a D0 transition
2297 * before enabling runtime PM.
2299 vfio_pci_set_power_state(vdev, PCI_D0);
2301 dev->driver->pm = &vfio_pci_core_pm_ops;
2302 pm_runtime_allow(dev);
2303 if (!disable_idle_d3)
2304 pm_runtime_put(dev);
2306 ret = vfio_register_group_dev(&vdev->vdev);
2312 if (!disable_idle_d3)
2313 pm_runtime_get_noresume(dev);
2315 pm_runtime_forbid(dev);
2317 vfio_pci_vf_uninit(vdev);
2320 EXPORT_SYMBOL_GPL(vfio_pci_core_register_device);
2322 void vfio_pci_core_unregister_device(struct vfio_pci_core_device *vdev)
2324 vfio_pci_core_sriov_configure(vdev, 0);
2326 vfio_unregister_group_dev(&vdev->vdev);
2328 vfio_pci_vf_uninit(vdev);
2329 vfio_pci_vga_uninit(vdev);
2331 if (!disable_idle_d3)
2332 pm_runtime_get_noresume(&vdev->pdev->dev);
2334 pm_runtime_forbid(&vdev->pdev->dev);
2336 EXPORT_SYMBOL_GPL(vfio_pci_core_unregister_device);
2338 pci_ers_result_t vfio_pci_core_aer_err_detected(struct pci_dev *pdev,
2339 pci_channel_state_t state)
2341 struct vfio_pci_core_device *vdev = dev_get_drvdata(&pdev->dev);
2343 mutex_lock(&vdev->igate);
2345 if (vdev->err_trigger)
2346 eventfd_signal(vdev->err_trigger);
2348 mutex_unlock(&vdev->igate);
2350 return PCI_ERS_RESULT_CAN_RECOVER;
2352 EXPORT_SYMBOL_GPL(vfio_pci_core_aer_err_detected);
2354 int vfio_pci_core_sriov_configure(struct vfio_pci_core_device *vdev,
2357 struct pci_dev *pdev = vdev->pdev;
2360 device_lock_assert(&pdev->dev);
2363 mutex_lock(&vfio_pci_sriov_pfs_mutex);
2365 * The thread that adds the vdev to the list is the only thread
2366 * that gets to call pci_enable_sriov() and we will only allow
2367 * it to be called once without going through
2368 * pci_disable_sriov()
2370 if (!list_empty(&vdev->sriov_pfs_item)) {
2374 list_add_tail(&vdev->sriov_pfs_item, &vfio_pci_sriov_pfs);
2375 mutex_unlock(&vfio_pci_sriov_pfs_mutex);
2378 * The PF power state should always be higher than the VF power
2379 * state. The PF can be in low power state either with runtime
2380 * power management (when there is no user) or PCI_PM_CTRL
2381 * register write by the user. If PF is in the low power state,
2382 * then change the power state to D0 first before enabling
2383 * SR-IOV. Also, this function can be called at any time, and
2384 * userspace PCI_PM_CTRL write can race against this code path,
2385 * so protect the same with 'memory_lock'.
2387 ret = pm_runtime_resume_and_get(&pdev->dev);
2391 down_write(&vdev->memory_lock);
2392 vfio_pci_set_power_state(vdev, PCI_D0);
2393 ret = pci_enable_sriov(pdev, nr_virtfn);
2394 up_write(&vdev->memory_lock);
2396 pm_runtime_put(&pdev->dev);
2402 if (pci_num_vf(pdev)) {
2403 pci_disable_sriov(pdev);
2404 pm_runtime_put(&pdev->dev);
2408 mutex_lock(&vfio_pci_sriov_pfs_mutex);
2409 list_del_init(&vdev->sriov_pfs_item);
2411 mutex_unlock(&vfio_pci_sriov_pfs_mutex);
2414 EXPORT_SYMBOL_GPL(vfio_pci_core_sriov_configure);
2416 const struct pci_error_handlers vfio_pci_core_err_handlers = {
2417 .error_detected = vfio_pci_core_aer_err_detected,
2419 EXPORT_SYMBOL_GPL(vfio_pci_core_err_handlers);
2421 static bool vfio_dev_in_groups(struct vfio_device *vdev,
2422 struct vfio_pci_group_info *groups)
2429 for (i = 0; i < groups->count; i++)
2430 if (vfio_file_has_dev(groups->files[i], vdev))
2435 static int vfio_pci_is_device_in_set(struct pci_dev *pdev, void *data)
2437 struct vfio_device_set *dev_set = data;
2439 return vfio_find_device_in_devset(dev_set, &pdev->dev) ? 0 : -ENODEV;
2443 * vfio-core considers a group to be viable and will create a vfio_device even
2444 * if some devices are bound to drivers like pci-stub or pcieport. Here we
2445 * require all PCI devices to be inside our dev_set since that ensures they stay
2446 * put and that every driver controlling the device can co-ordinate with the
2449 * Returns the pci_dev to pass to pci_reset_bus() if every PCI device to be
2450 * reset is inside the dev_set, and pci_reset_bus() can succeed. NULL otherwise.
2452 static struct pci_dev *
2453 vfio_pci_dev_set_resettable(struct vfio_device_set *dev_set)
2455 struct pci_dev *pdev;
2457 lockdep_assert_held(&dev_set->lock);
2460 * By definition all PCI devices in the dev_set share the same PCI
2461 * reset, so any pci_dev will have the same outcomes for
2462 * pci_probe_reset_*() and pci_reset_bus().
2464 pdev = list_first_entry(&dev_set->device_list,
2465 struct vfio_pci_core_device,
2466 vdev.dev_set_list)->pdev;
2468 /* pci_reset_bus() is supported */
2469 if (pci_probe_reset_slot(pdev->slot) && pci_probe_reset_bus(pdev->bus))
2472 if (vfio_pci_for_each_slot_or_bus(pdev, vfio_pci_is_device_in_set,
2474 !pci_probe_reset_slot(pdev->slot)))
2479 static int vfio_pci_dev_set_pm_runtime_get(struct vfio_device_set *dev_set)
2481 struct vfio_pci_core_device *cur;
2484 list_for_each_entry(cur, &dev_set->device_list, vdev.dev_set_list) {
2485 ret = pm_runtime_resume_and_get(&cur->pdev->dev);
2493 list_for_each_entry_continue_reverse(cur, &dev_set->device_list,
2495 pm_runtime_put(&cur->pdev->dev);
2501 * We need to get memory_lock for each device, but devices can share mmap_lock,
2502 * therefore we need to zap and hold the vma_lock for each device, and only then
2503 * get each memory_lock.
2505 static int vfio_pci_dev_set_hot_reset(struct vfio_device_set *dev_set,
2506 struct vfio_pci_group_info *groups,
2507 struct iommufd_ctx *iommufd_ctx)
2509 struct vfio_pci_core_device *cur_mem;
2510 struct vfio_pci_core_device *cur_vma;
2511 struct vfio_pci_core_device *cur;
2512 struct pci_dev *pdev;
2516 mutex_lock(&dev_set->lock);
2517 cur_mem = list_first_entry(&dev_set->device_list,
2518 struct vfio_pci_core_device,
2521 pdev = vfio_pci_dev_set_resettable(dev_set);
2528 * Some of the devices in the dev_set can be in the runtime suspended
2529 * state. Increment the usage count for all the devices in the dev_set
2530 * before reset and decrement the same after reset.
2532 ret = vfio_pci_dev_set_pm_runtime_get(dev_set);
2536 list_for_each_entry(cur_vma, &dev_set->device_list, vdev.dev_set_list) {
2540 * Test whether all the affected devices can be reset by the
2543 * If called from a group opened device and the user provides
2544 * a set of groups, all the devices in the dev_set should be
2545 * contained by the set of groups provided by the user.
2547 * If called from a cdev opened device and the user provides
2548 * a zero-length array, all the devices in the dev_set must
2549 * be bound to the same iommufd_ctx as the input iommufd_ctx.
2550 * If there is any device that has not been bound to any
2551 * iommufd_ctx yet, check if its iommu_group has any device
2552 * bound to the input iommufd_ctx. Such devices can be
2553 * considered owned by the input iommufd_ctx as the device
2554 * cannot be owned by another iommufd_ctx when its iommu_group
2557 * Otherwise, reset is not allowed.
2560 int devid = vfio_iommufd_get_dev_id(&cur_vma->vdev,
2563 owned = (devid > 0 || devid == -ENOENT);
2565 owned = vfio_dev_in_groups(&cur_vma->vdev, groups);
2574 * Locking multiple devices is prone to deadlock, runaway and
2575 * unwind if we hit contention.
2577 if (!vfio_pci_zap_and_vma_lock(cur_vma, true)) {
2584 list_for_each_entry(cur_mem, &dev_set->device_list, vdev.dev_set_list) {
2585 if (!down_write_trylock(&cur_mem->memory_lock)) {
2589 mutex_unlock(&cur_mem->vma_lock);
2594 * The pci_reset_bus() will reset all the devices in the bus.
2595 * The power state can be non-D0 for some of the devices in the bus.
2596 * For these devices, the pci_reset_bus() will internally set
2597 * the power state to D0 without vfio driver involvement.
2598 * For the devices which have NoSoftRst-, the reset function can
2599 * cause the PCI config space reset without restoring the original
2600 * state (saved locally in 'vdev->pm_save').
2602 list_for_each_entry(cur, &dev_set->device_list, vdev.dev_set_list)
2603 vfio_pci_set_power_state(cur, PCI_D0);
2605 ret = pci_reset_bus(pdev);
2608 list_for_each_entry(cur, &dev_set->device_list, vdev.dev_set_list) {
2614 up_write(&cur->memory_lock);
2616 mutex_unlock(&cur->vma_lock);
2619 list_for_each_entry(cur, &dev_set->device_list, vdev.dev_set_list)
2620 pm_runtime_put(&cur->pdev->dev);
2622 mutex_unlock(&dev_set->lock);
2626 static bool vfio_pci_dev_set_needs_reset(struct vfio_device_set *dev_set)
2628 struct vfio_pci_core_device *cur;
2629 bool needs_reset = false;
2631 /* No other VFIO device in the set can be open. */
2632 if (vfio_device_set_open_count(dev_set) > 1)
2635 list_for_each_entry(cur, &dev_set->device_list, vdev.dev_set_list)
2636 needs_reset |= cur->needs_reset;
2641 * If a bus or slot reset is available for the provided dev_set and:
2642 * - All of the devices affected by that bus or slot reset are unused
2643 * - At least one of the affected devices is marked dirty via
2644 * needs_reset (such as by lack of FLR support)
2645 * Then attempt to perform that bus or slot reset.
2647 static void vfio_pci_dev_set_try_reset(struct vfio_device_set *dev_set)
2649 struct vfio_pci_core_device *cur;
2650 struct pci_dev *pdev;
2651 bool reset_done = false;
2653 if (!vfio_pci_dev_set_needs_reset(dev_set))
2656 pdev = vfio_pci_dev_set_resettable(dev_set);
2661 * Some of the devices in the bus can be in the runtime suspended
2662 * state. Increment the usage count for all the devices in the dev_set
2663 * before reset and decrement the same after reset.
2665 if (!disable_idle_d3 && vfio_pci_dev_set_pm_runtime_get(dev_set))
2668 if (!pci_reset_bus(pdev))
2671 list_for_each_entry(cur, &dev_set->device_list, vdev.dev_set_list) {
2673 cur->needs_reset = false;
2675 if (!disable_idle_d3)
2676 pm_runtime_put(&cur->pdev->dev);
2680 void vfio_pci_core_set_params(bool is_nointxmask, bool is_disable_vga,
2681 bool is_disable_idle_d3)
2683 nointxmask = is_nointxmask;
2684 disable_vga = is_disable_vga;
2685 disable_idle_d3 = is_disable_idle_d3;
2687 EXPORT_SYMBOL_GPL(vfio_pci_core_set_params);
2689 static void vfio_pci_core_cleanup(void)
2691 vfio_pci_uninit_perm_bits();
2694 static int __init vfio_pci_core_init(void)
2696 /* Allocate shared config space permission data used by all devices */
2697 return vfio_pci_init_perm_bits();
2700 module_init(vfio_pci_core_init);
2701 module_exit(vfio_pci_core_cleanup);
2703 MODULE_LICENSE("GPL v2");
2704 MODULE_AUTHOR(DRIVER_AUTHOR);
2705 MODULE_DESCRIPTION(DRIVER_DESC);