vfio/pci/vfio_pci_core.c

0001 // SPDX-License-Identifier: GPL-2.0-only
0002 /*
0003  * Copyright (C) 2012 Red Hat, Inc.  All rights reserved.
0004  *     Author: Alex Williamson <alex.williamson@redhat.com>
0005  *
0006  * Derived from original vfio:
0007  * Copyright 2010 Cisco Systems, Inc.  All rights reserved.
0008  * Author: Tom Lyon, pugs@cisco.com
0009  */
0010
0011 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
0012
0013 #include <linux/aperture.h>
0014 #include <linux/device.h>
0015 #include <linux/eventfd.h>
0016 #include <linux/file.h>
0017 #include <linux/interrupt.h>
0018 #include <linux/iommu.h>
0019 #include <linux/module.h>
0020 #include <linux/mutex.h>
0021 #include <linux/notifier.h>
0022 #include <linux/pci.h>
0023 #include <linux/pm_runtime.h>
0024 #include <linux/slab.h>
0025 #include <linux/types.h>
0026 #include <linux/uaccess.h>
0027 #include <linux/vgaarb.h>
0028 #include <linux/nospec.h>
0029 #include <linux/sched/mm.h>
0030
0031 #include <linux/vfio_pci_core.h>
0032
0033 #define DRIVER_AUTHOR   "Alex Williamson <alex.williamson@redhat.com>"
0034 #define DRIVER_DESC "core driver for VFIO based PCI devices"
0035
0036 static bool nointxmask;
0037 static bool disable_vga;
0038 static bool disable_idle_d3;
0039
0040 /* List of PF's that vfio_pci_core_sriov_configure() has been called on */
0041 static DEFINE_MUTEX(vfio_pci_sriov_pfs_mutex);
0042 static LIST_HEAD(vfio_pci_sriov_pfs);
0043
0044 static inline bool vfio_vga_disabled(void)
0045 {
0046 #ifdef CONFIG_VFIO_PCI_VGA
0047     return disable_vga;
0048 #else
0049     return true;
0050 #endif
0051 }
0052
0053 /*
0054  * Our VGA arbiter participation is limited since we don't know anything
0055  * about the device itself.  However, if the device is the only VGA device
0056  * downstream of a bridge and VFIO VGA support is disabled, then we can
0057  * safely return legacy VGA IO and memory as not decoded since the user
0058  * has no way to get to it and routing can be disabled externally at the
0059  * bridge.
0060  */
0061 static unsigned int vfio_pci_set_decode(struct pci_dev *pdev, bool single_vga)
0062 {
0063     struct pci_dev *tmp = NULL;
0064     unsigned char max_busnr;
0065     unsigned int decodes;
0066
0067     if (single_vga || !vfio_vga_disabled() || pci_is_root_bus(pdev->bus))
0068         return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM |
0069                VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM;
0070
0071     max_busnr = pci_bus_max_busnr(pdev->bus);
0072     decodes = VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
0073
0074     while ((tmp = pci_get_class(PCI_CLASS_DISPLAY_VGA << 8, tmp)) != NULL) {
0075         if (tmp == pdev ||
0076             pci_domain_nr(tmp->bus) != pci_domain_nr(pdev->bus) ||
0077             pci_is_root_bus(tmp->bus))
0078             continue;
0079
0080         if (tmp->bus->number >= pdev->bus->number &&
0081             tmp->bus->number <= max_busnr) {
0082             pci_dev_put(tmp);
0083             decodes |= VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM;
0084             break;
0085         }
0086     }
0087
0088     return decodes;
0089 }
0090
0091 static void vfio_pci_probe_mmaps(struct vfio_pci_core_device *vdev)
0092 {
0093     struct resource *res;
0094     int i;
0095     struct vfio_pci_dummy_resource *dummy_res;
0096
0097     for (i = 0; i < PCI_STD_NUM_BARS; i++) {
0098         int bar = i + PCI_STD_RESOURCES;
0099
0100         res = &vdev->pdev->resource[bar];
0101
0102         if (!IS_ENABLED(CONFIG_VFIO_PCI_MMAP))
0103             goto no_mmap;
0104
0105         if (!(res->flags & IORESOURCE_MEM))
0106             goto no_mmap;
0107
0108         /*
0109          * The PCI core shouldn't set up a resource with a
0110          * type but zero size. But there may be bugs that
0111          * cause us to do that.
0112          */
0113         if (!resource_size(res))
0114             goto no_mmap;
0115
0116         if (resource_size(res) >= PAGE_SIZE) {
0117             vdev->bar_mmap_supported[bar] = true;
0118             continue;
0119         }
0120
0121         if (!(res->start & ~PAGE_MASK)) {
0122             /*
0123              * Add a dummy resource to reserve the remainder
0124              * of the exclusive page in case that hot-add
0125              * device's bar is assigned into it.
0126              */
0127             dummy_res = kzalloc(sizeof(*dummy_res), GFP_KERNEL);
0128             if (dummy_res == NULL)
0129                 goto no_mmap;
0130
0131             dummy_res->resource.name = "vfio sub-page reserved";
0132             dummy_res->resource.start = res->end + 1;
0133             dummy_res->resource.end = res->start + PAGE_SIZE - 1;
0134             dummy_res->resource.flags = res->flags;
0135             if (request_resource(res->parent,
0136                         &dummy_res->resource)) {
0137                 kfree(dummy_res);
0138                 goto no_mmap;
0139             }
0140             dummy_res->index = bar;
0141             list_add(&dummy_res->res_next,
0142                     &vdev->dummy_resources_list);
0143             vdev->bar_mmap_supported[bar] = true;
0144             continue;
0145         }
0146         /*
0147          * Here we don't handle the case when the BAR is not page
0148          * aligned because we can't expect the BAR will be
0149          * assigned into the same location in a page in guest
0150          * when we passthrough the BAR. And it's hard to access
0151          * this BAR in userspace because we have no way to get
0152          * the BAR's location in a page.
0153          */
0154 no_mmap:
0155         vdev->bar_mmap_supported[bar] = false;
0156     }
0157 }
0158
0159 struct vfio_pci_group_info;
0160 static void vfio_pci_dev_set_try_reset(struct vfio_device_set *dev_set);
0161 static int vfio_pci_dev_set_hot_reset(struct vfio_device_set *dev_set,
0162                       struct vfio_pci_group_info *groups);
0163
0164 /*
0165  * INTx masking requires the ability to disable INTx signaling via PCI_COMMAND
0166  * _and_ the ability detect when the device is asserting INTx via PCI_STATUS.
0167  * If a device implements the former but not the latter we would typically
0168  * expect broken_intx_masking be set and require an exclusive interrupt.
0169  * However since we do have control of the device's ability to assert INTx,
0170  * we can instead pretend that the device does not implement INTx, virtualizing
0171  * the pin register to report zero and maintaining DisINTx set on the host.
0172  */
0173 static bool vfio_pci_nointx(struct pci_dev *pdev)
0174 {
0175     switch (pdev->vendor) {
0176     case PCI_VENDOR_ID_INTEL:
0177         switch (pdev->device) {
0178         /* All i40e (XL710/X710/XXV710) 10/20/25/40GbE NICs */
0179         case 0x1572:
0180         case 0x1574:
0181         case 0x1580 ... 0x1581:
0182         case 0x1583 ... 0x158b:
0183         case 0x37d0 ... 0x37d2:
0184         /* X550 */
0185         case 0x1563:
0186             return true;
0187         default:
0188             return false;
0189         }
0190     }
0191
0192     return false;
0193 }
0194
0195 static void vfio_pci_probe_power_state(struct vfio_pci_core_device *vdev)
0196 {
0197     struct pci_dev *pdev = vdev->pdev;
0198     u16 pmcsr;
0199
0200     if (!pdev->pm_cap)
0201         return;
0202
0203     pci_read_config_word(pdev, pdev->pm_cap + PCI_PM_CTRL, &pmcsr);
0204
0205     vdev->needs_pm_restore = !(pmcsr & PCI_PM_CTRL_NO_SOFT_RESET);
0206 }
0207
0208 /*
0209  * pci_set_power_state() wrapper handling devices which perform a soft reset on
0210  * D3->D0 transition.  Save state prior to D0/1/2->D3, stash it on the vdev,
0211  * restore when returned to D0.  Saved separately from pci_saved_state for use
0212  * by PM capability emulation and separately from pci_dev internal saved state
0213  * to avoid it being overwritten and consumed around other resets.
0214  */
0215 int vfio_pci_set_power_state(struct vfio_pci_core_device *vdev, pci_power_t state)
0216 {
0217     struct pci_dev *pdev = vdev->pdev;
0218     bool needs_restore = false, needs_save = false;
0219     int ret;
0220
0221     /* Prevent changing power state for PFs with VFs enabled */
0222     if (pci_num_vf(pdev) && state > PCI_D0)
0223         return -EBUSY;
0224
0225     if (vdev->needs_pm_restore) {
0226         if (pdev->current_state < PCI_D3hot && state >= PCI_D3hot) {
0227             pci_save_state(pdev);
0228             needs_save = true;
0229         }
0230
0231         if (pdev->current_state >= PCI_D3hot && state <= PCI_D0)
0232             needs_restore = true;
0233     }
0234
0235     ret = pci_set_power_state(pdev, state);
0236
0237     if (!ret) {
0238         /* D3 might be unsupported via quirk, skip unless in D3 */
0239         if (needs_save && pdev->current_state >= PCI_D3hot) {
0240             /*
0241              * The current PCI state will be saved locally in
0242              * 'pm_save' during the D3hot transition. When the
0243              * device state is changed to D0 again with the current
0244              * function, then pci_store_saved_state() will restore
0245              * the state and will free the memory pointed by
0246              * 'pm_save'. There are few cases where the PCI power
0247              * state can be changed to D0 without the involvement
0248              * of the driver. For these cases, free the earlier
0249              * allocated memory first before overwriting 'pm_save'
0250              * to prevent the memory leak.
0251              */
0252             kfree(vdev->pm_save);
0253             vdev->pm_save = pci_store_saved_state(pdev);
0254         } else if (needs_restore) {
0255             pci_load_and_free_saved_state(pdev, &vdev->pm_save);
0256             pci_restore_state(pdev);
0257         }
0258     }
0259
0260     return ret;
0261 }
0262
0263 /*
0264  * The dev_pm_ops needs to be provided to make pci-driver runtime PM working,
0265  * so use structure without any callbacks.
0266  *
0267  * The pci-driver core runtime PM routines always save the device state
0268  * before going into suspended state. If the device is going into low power
0269  * state with only with runtime PM ops, then no explicit handling is needed
0270  * for the devices which have NoSoftRst-.
0271  */
0272 static const struct dev_pm_ops vfio_pci_core_pm_ops = { };
0273
0274 int vfio_pci_core_enable(struct vfio_pci_core_device *vdev)
0275 {
0276     struct pci_dev *pdev = vdev->pdev;
0277     int ret;
0278     u16 cmd;
0279     u8 msix_pos;
0280
0281     if (!disable_idle_d3) {
0282         ret = pm_runtime_resume_and_get(&pdev->dev);
0283         if (ret < 0)
0284             return ret;
0285     }
0286
0287     /* Don't allow our initial saved state to include busmaster */
0288     pci_clear_master(pdev);
0289
0290     ret = pci_enable_device(pdev);
0291     if (ret)
0292         goto out_power;
0293
0294     /* If reset fails because of the device lock, fail this path entirely */
0295     ret = pci_try_reset_function(pdev);
0296     if (ret == -EAGAIN)
0297         goto out_disable_device;
0298
0299     vdev->reset_works = !ret;
0300     pci_save_state(pdev);
0301     vdev->pci_saved_state = pci_store_saved_state(pdev);
0302     if (!vdev->pci_saved_state)
0303         pci_dbg(pdev, "%s: Couldn't store saved state\n", __func__);
0304
0305     if (likely(!nointxmask)) {
0306         if (vfio_pci_nointx(pdev)) {
0307             pci_info(pdev, "Masking broken INTx support\n");
0308             vdev->nointx = true;
0309             pci_intx(pdev, 0);
0310         } else
0311             vdev->pci_2_3 = pci_intx_mask_supported(pdev);
0312     }
0313
0314     pci_read_config_word(pdev, PCI_COMMAND, &cmd);
0315     if (vdev->pci_2_3 && (cmd & PCI_COMMAND_INTX_DISABLE)) {
0316         cmd &= ~PCI_COMMAND_INTX_DISABLE;
0317         pci_write_config_word(pdev, PCI_COMMAND, cmd);
0318     }
0319
0320     ret = vfio_pci_zdev_open_device(vdev);
0321     if (ret)
0322         goto out_free_state;
0323
0324     ret = vfio_config_init(vdev);
0325     if (ret)
0326         goto out_free_zdev;
0327
0328     msix_pos = pdev->msix_cap;
0329     if (msix_pos) {
0330         u16 flags;
0331         u32 table;
0332
0333         pci_read_config_word(pdev, msix_pos + PCI_MSIX_FLAGS, &flags);
0334         pci_read_config_dword(pdev, msix_pos + PCI_MSIX_TABLE, &table);
0335
0336         vdev->msix_bar = table & PCI_MSIX_TABLE_BIR;
0337         vdev->msix_offset = table & PCI_MSIX_TABLE_OFFSET;
0338         vdev->msix_size = ((flags & PCI_MSIX_FLAGS_QSIZE) + 1) * 16;
0339     } else
0340         vdev->msix_bar = 0xFF;
0341
0342     if (!vfio_vga_disabled() && vfio_pci_is_vga(pdev))
0343         vdev->has_vga = true;
0344
0345
0346     return 0;
0347
0348 out_free_zdev:
0349     vfio_pci_zdev_close_device(vdev);
0350 out_free_state:
0351     kfree(vdev->pci_saved_state);
0352     vdev->pci_saved_state = NULL;
0353 out_disable_device:
0354     pci_disable_device(pdev);
0355 out_power:
0356     if (!disable_idle_d3)
0357         pm_runtime_put(&pdev->dev);
0358     return ret;
0359 }
0360 EXPORT_SYMBOL_GPL(vfio_pci_core_enable);
0361
0362 void vfio_pci_core_disable(struct vfio_pci_core_device *vdev)
0363 {
0364     struct pci_dev *pdev = vdev->pdev;
0365     struct vfio_pci_dummy_resource *dummy_res, *tmp;
0366     struct vfio_pci_ioeventfd *ioeventfd, *ioeventfd_tmp;
0367     int i, bar;
0368
0369     /* For needs_reset */
0370     lockdep_assert_held(&vdev->vdev.dev_set->lock);
0371
0372     /*
0373      * This function can be invoked while the power state is non-D0.
0374      * This function calls __pci_reset_function_locked() which internally
0375      * can use pci_pm_reset() for the function reset. pci_pm_reset() will
0376      * fail if the power state is non-D0. Also, for the devices which
0377      * have NoSoftRst-, the reset function can cause the PCI config space
0378      * reset without restoring the original state (saved locally in
0379      * 'vdev->pm_save').
0380      */
0381     vfio_pci_set_power_state(vdev, PCI_D0);
0382
0383     /* Stop the device from further DMA */
0384     pci_clear_master(pdev);
0385
0386     vfio_pci_set_irqs_ioctl(vdev, VFIO_IRQ_SET_DATA_NONE |
0387                 VFIO_IRQ_SET_ACTION_TRIGGER,
0388                 vdev->irq_type, 0, 0, NULL);
0389
0390     /* Device closed, don't need mutex here */
0391     list_for_each_entry_safe(ioeventfd, ioeventfd_tmp,
0392                  &vdev->ioeventfds_list, next) {
0393         vfio_virqfd_disable(&ioeventfd->virqfd);
0394         list_del(&ioeventfd->next);
0395         kfree(ioeventfd);
0396     }
0397     vdev->ioeventfds_nr = 0;
0398
0399     vdev->virq_disabled = false;
0400
0401     for (i = 0; i < vdev->num_regions; i++)
0402         vdev->region[i].ops->release(vdev, &vdev->region[i]);
0403
0404     vdev->num_regions = 0;
0405     kfree(vdev->region);
0406     vdev->region = NULL; /* don't krealloc a freed pointer */
0407
0408     vfio_config_free(vdev);
0409
0410     for (i = 0; i < PCI_STD_NUM_BARS; i++) {
0411         bar = i + PCI_STD_RESOURCES;
0412         if (!vdev->barmap[bar])
0413             continue;
0414         pci_iounmap(pdev, vdev->barmap[bar]);
0415         pci_release_selected_regions(pdev, 1 << bar);
0416         vdev->barmap[bar] = NULL;
0417     }
0418
0419     list_for_each_entry_safe(dummy_res, tmp,
0420                  &vdev->dummy_resources_list, res_next) {
0421         list_del(&dummy_res->res_next);
0422         release_resource(&dummy_res->resource);
0423         kfree(dummy_res);
0424     }
0425
0426     vdev->needs_reset = true;
0427
0428     vfio_pci_zdev_close_device(vdev);
0429
0430     /*
0431      * If we have saved state, restore it.  If we can reset the device,
0432      * even better.  Resetting with current state seems better than
0433      * nothing, but saving and restoring current state without reset
0434      * is just busy work.
0435      */
0436     if (pci_load_and_free_saved_state(pdev, &vdev->pci_saved_state)) {
0437         pci_info(pdev, "%s: Couldn't reload saved state\n", __func__);
0438
0439         if (!vdev->reset_works)
0440             goto out;
0441
0442         pci_save_state(pdev);
0443     }
0444
0445     /*
0446      * Disable INTx and MSI, presumably to avoid spurious interrupts
0447      * during reset.  Stolen from pci_reset_function()
0448      */
0449     pci_write_config_word(pdev, PCI_COMMAND, PCI_COMMAND_INTX_DISABLE);
0450
0451     /*
0452      * Try to get the locks ourselves to prevent a deadlock. The
0453      * success of this is dependent on being able to lock the device,
0454      * which is not always possible.
0455      * We can not use the "try" reset interface here, which will
0456      * overwrite the previously restored configuration information.
0457      */
0458     if (vdev->reset_works && pci_dev_trylock(pdev)) {
0459         if (!__pci_reset_function_locked(pdev))
0460             vdev->needs_reset = false;
0461         pci_dev_unlock(pdev);
0462     }
0463
0464     pci_restore_state(pdev);
0465 out:
0466     pci_disable_device(pdev);
0467
0468     vfio_pci_dev_set_try_reset(vdev->vdev.dev_set);
0469
0470     /* Put the pm-runtime usage counter acquired during enable */
0471     if (!disable_idle_d3)
0472         pm_runtime_put(&pdev->dev);
0473 }
0474 EXPORT_SYMBOL_GPL(vfio_pci_core_disable);
0475
0476 void vfio_pci_core_close_device(struct vfio_device *core_vdev)
0477 {
0478     struct vfio_pci_core_device *vdev =
0479         container_of(core_vdev, struct vfio_pci_core_device, vdev);
0480
0481     if (vdev->sriov_pf_core_dev) {
0482         mutex_lock(&vdev->sriov_pf_core_dev->vf_token->lock);
0483         WARN_ON(!vdev->sriov_pf_core_dev->vf_token->users);
0484         vdev->sriov_pf_core_dev->vf_token->users--;
0485         mutex_unlock(&vdev->sriov_pf_core_dev->vf_token->lock);
0486     }
0487     vfio_spapr_pci_eeh_release(vdev->pdev);
0488     vfio_pci_core_disable(vdev);
0489
0490     mutex_lock(&vdev->igate);
0491     if (vdev->err_trigger) {
0492         eventfd_ctx_put(vdev->err_trigger);
0493         vdev->err_trigger = NULL;
0494     }
0495     if (vdev->req_trigger) {
0496         eventfd_ctx_put(vdev->req_trigger);
0497         vdev->req_trigger = NULL;
0498     }
0499     mutex_unlock(&vdev->igate);
0500 }
0501 EXPORT_SYMBOL_GPL(vfio_pci_core_close_device);
0502
0503 void vfio_pci_core_finish_enable(struct vfio_pci_core_device *vdev)
0504 {
0505     vfio_pci_probe_mmaps(vdev);
0506     vfio_spapr_pci_eeh_open(vdev->pdev);
0507
0508     if (vdev->sriov_pf_core_dev) {
0509         mutex_lock(&vdev->sriov_pf_core_dev->vf_token->lock);
0510         vdev->sriov_pf_core_dev->vf_token->users++;
0511         mutex_unlock(&vdev->sriov_pf_core_dev->vf_token->lock);
0512     }
0513 }
0514 EXPORT_SYMBOL_GPL(vfio_pci_core_finish_enable);
0515
0516 static int vfio_pci_get_irq_count(struct vfio_pci_core_device *vdev, int irq_type)
0517 {
0518     if (irq_type == VFIO_PCI_INTX_IRQ_INDEX) {
0519         u8 pin;
0520
0521         if (!IS_ENABLED(CONFIG_VFIO_PCI_INTX) ||
0522             vdev->nointx || vdev->pdev->is_virtfn)
0523             return 0;
0524
0525         pci_read_config_byte(vdev->pdev, PCI_INTERRUPT_PIN, &pin);
0526
0527         return pin ? 1 : 0;
0528     } else if (irq_type == VFIO_PCI_MSI_IRQ_INDEX) {
0529         u8 pos;
0530         u16 flags;
0531
0532         pos = vdev->pdev->msi_cap;
0533         if (pos) {
0534             pci_read_config_word(vdev->pdev,
0535                          pos + PCI_MSI_FLAGS, &flags);
0536             return 1 << ((flags & PCI_MSI_FLAGS_QMASK) >> 1);
0537         }
0538     } else if (irq_type == VFIO_PCI_MSIX_IRQ_INDEX) {
0539         u8 pos;
0540         u16 flags;
0541
0542         pos = vdev->pdev->msix_cap;
0543         if (pos) {
0544             pci_read_config_word(vdev->pdev,
0545                          pos + PCI_MSIX_FLAGS, &flags);
0546
0547             return (flags & PCI_MSIX_FLAGS_QSIZE) + 1;
0548         }
0549     } else if (irq_type == VFIO_PCI_ERR_IRQ_INDEX) {
0550         if (pci_is_pcie(vdev->pdev))
0551             return 1;
0552     } else if (irq_type == VFIO_PCI_REQ_IRQ_INDEX) {
0553         return 1;
0554     }
0555
0556     return 0;
0557 }
0558
0559 static int vfio_pci_count_devs(struct pci_dev *pdev, void *data)
0560 {
0561     (*(int *)data)++;
0562     return 0;
0563 }
0564
0565 struct vfio_pci_fill_info {
0566     int max;
0567     int cur;
0568     struct vfio_pci_dependent_device *devices;
0569 };
0570
0571 static int vfio_pci_fill_devs(struct pci_dev *pdev, void *data)
0572 {
0573     struct vfio_pci_fill_info *fill = data;
0574     struct iommu_group *iommu_group;
0575
0576     if (fill->cur == fill->max)
0577         return -EAGAIN; /* Something changed, try again */
0578
0579     iommu_group = iommu_group_get(&pdev->dev);
0580     if (!iommu_group)
0581         return -EPERM; /* Cannot reset non-isolated devices */
0582
0583     fill->devices[fill->cur].group_id = iommu_group_id(iommu_group);
0584     fill->devices[fill->cur].segment = pci_domain_nr(pdev->bus);
0585     fill->devices[fill->cur].bus = pdev->bus->number;
0586     fill->devices[fill->cur].devfn = pdev->devfn;
0587     fill->cur++;
0588     iommu_group_put(iommu_group);
0589     return 0;
0590 }
0591
0592 struct vfio_pci_group_info {
0593     int count;
0594     struct file **files;
0595 };
0596
0597 static bool vfio_pci_dev_below_slot(struct pci_dev *pdev, struct pci_slot *slot)
0598 {
0599     for (; pdev; pdev = pdev->bus->self)
0600         if (pdev->bus == slot->bus)
0601             return (pdev->slot == slot);
0602     return false;
0603 }
0604
0605 struct vfio_pci_walk_info {
0606     int (*fn)(struct pci_dev *pdev, void *data);
0607     void *data;
0608     struct pci_dev *pdev;
0609     bool slot;
0610     int ret;
0611 };
0612
0613 static int vfio_pci_walk_wrapper(struct pci_dev *pdev, void *data)
0614 {
0615     struct vfio_pci_walk_info *walk = data;
0616
0617     if (!walk->slot || vfio_pci_dev_below_slot(pdev, walk->pdev->slot))
0618         walk->ret = walk->fn(pdev, walk->data);
0619
0620     return walk->ret;
0621 }
0622
0623 static int vfio_pci_for_each_slot_or_bus(struct pci_dev *pdev,
0624                      int (*fn)(struct pci_dev *,
0625                            void *data), void *data,
0626                      bool slot)
0627 {
0628     struct vfio_pci_walk_info walk = {
0629         .fn = fn, .data = data, .pdev = pdev, .slot = slot, .ret = 0,
0630     };
0631
0632     pci_walk_bus(pdev->bus, vfio_pci_walk_wrapper, &walk);
0633
0634     return walk.ret;
0635 }
0636
0637 static int msix_mmappable_cap(struct vfio_pci_core_device *vdev,
0638                   struct vfio_info_cap *caps)
0639 {
0640     struct vfio_info_cap_header header = {
0641         .id = VFIO_REGION_INFO_CAP_MSIX_MAPPABLE,
0642         .version = 1
0643     };
0644
0645     return vfio_info_add_capability(caps, &header, sizeof(header));
0646 }
0647
0648 int vfio_pci_register_dev_region(struct vfio_pci_core_device *vdev,
0649                  unsigned int type, unsigned int subtype,
0650                  const struct vfio_pci_regops *ops,
0651                  size_t size, u32 flags, void *data)
0652 {
0653     struct vfio_pci_region *region;
0654
0655     region = krealloc(vdev->region,
0656               (vdev->num_regions + 1) * sizeof(*region),
0657               GFP_KERNEL);
0658     if (!region)
0659         return -ENOMEM;
0660
0661     vdev->region = region;
0662     vdev->region[vdev->num_regions].type = type;
0663     vdev->region[vdev->num_regions].subtype = subtype;
0664     vdev->region[vdev->num_regions].ops = ops;
0665     vdev->region[vdev->num_regions].size = size;
0666     vdev->region[vdev->num_regions].flags = flags;
0667     vdev->region[vdev->num_regions].data = data;
0668
0669     vdev->num_regions++;
0670
0671     return 0;
0672 }
0673 EXPORT_SYMBOL_GPL(vfio_pci_register_dev_region);
0674
0675 long vfio_pci_core_ioctl(struct vfio_device *core_vdev, unsigned int cmd,
0676         unsigned long arg)
0677 {
0678     struct vfio_pci_core_device *vdev =
0679         container_of(core_vdev, struct vfio_pci_core_device, vdev);
0680     unsigned long minsz;
0681
0682     if (cmd == VFIO_DEVICE_GET_INFO) {
0683         struct vfio_device_info info;
0684         struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
0685         unsigned long capsz;
0686         int ret;
0687
0688         minsz = offsetofend(struct vfio_device_info, num_irqs);
0689
0690         /* For backward compatibility, cannot require this */
0691         capsz = offsetofend(struct vfio_iommu_type1_info, cap_offset);
0692
0693         if (copy_from_user(&info, (void __user *)arg, minsz))
0694             return -EFAULT;
0695
0696         if (info.argsz < minsz)
0697             return -EINVAL;
0698
0699         if (info.argsz >= capsz) {
0700             minsz = capsz;
0701             info.cap_offset = 0;
0702         }
0703
0704         info.flags = VFIO_DEVICE_FLAGS_PCI;
0705
0706         if (vdev->reset_works)
0707             info.flags |= VFIO_DEVICE_FLAGS_RESET;
0708
0709         info.num_regions = VFIO_PCI_NUM_REGIONS + vdev->num_regions;
0710         info.num_irqs = VFIO_PCI_NUM_IRQS;
0711
0712         ret = vfio_pci_info_zdev_add_caps(vdev, &caps);
0713         if (ret && ret != -ENODEV) {
0714             pci_warn(vdev->pdev, "Failed to setup zPCI info capabilities\n");
0715             return ret;
0716         }
0717
0718         if (caps.size) {
0719             info.flags |= VFIO_DEVICE_FLAGS_CAPS;
0720             if (info.argsz < sizeof(info) + caps.size) {
0721                 info.argsz = sizeof(info) + caps.size;
0722             } else {
0723                 vfio_info_cap_shift(&caps, sizeof(info));
0724                 if (copy_to_user((void __user *)arg +
0725                           sizeof(info), caps.buf,
0726                           caps.size)) {
0727                     kfree(caps.buf);
0728                     return -EFAULT;
0729                 }
0730                 info.cap_offset = sizeof(info);
0731             }
0732
0733             kfree(caps.buf);
0734         }
0735
0736         return copy_to_user((void __user *)arg, &info, minsz) ?
0737             -EFAULT : 0;
0738
0739     } else if (cmd == VFIO_DEVICE_GET_REGION_INFO) {
0740         struct pci_dev *pdev = vdev->pdev;
0741         struct vfio_region_info info;
0742         struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
0743         int i, ret;
0744
0745         minsz = offsetofend(struct vfio_region_info, offset);
0746
0747         if (copy_from_user(&info, (void __user *)arg, minsz))
0748             return -EFAULT;
0749
0750         if (info.argsz < minsz)
0751             return -EINVAL;
0752
0753         switch (info.index) {
0754         case VFIO_PCI_CONFIG_REGION_INDEX:
0755             info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
0756             info.size = pdev->cfg_size;
0757             info.flags = VFIO_REGION_INFO_FLAG_READ |
0758                      VFIO_REGION_INFO_FLAG_WRITE;
0759             break;
0760         case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
0761             info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
0762             info.size = pci_resource_len(pdev, info.index);
0763             if (!info.size) {
0764                 info.flags = 0;
0765                 break;
0766             }
0767
0768             info.flags = VFIO_REGION_INFO_FLAG_READ |
0769                      VFIO_REGION_INFO_FLAG_WRITE;
0770             if (vdev->bar_mmap_supported[info.index]) {
0771                 info.flags |= VFIO_REGION_INFO_FLAG_MMAP;
0772                 if (info.index == vdev->msix_bar) {
0773                     ret = msix_mmappable_cap(vdev, &caps);
0774                     if (ret)
0775                         return ret;
0776                 }
0777             }
0778
0779             break;
0780         case VFIO_PCI_ROM_REGION_INDEX:
0781         {
0782             void __iomem *io;
0783             size_t size;
0784             u16 cmd;
0785
0786             info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
0787             info.flags = 0;
0788
0789             /* Report the BAR size, not the ROM size */
0790             info.size = pci_resource_len(pdev, info.index);
0791             if (!info.size) {
0792                 /* Shadow ROMs appear as PCI option ROMs */
0793                 if (pdev->resource[PCI_ROM_RESOURCE].flags &
0794                             IORESOURCE_ROM_SHADOW)
0795                     info.size = 0x20000;
0796                 else
0797                     break;
0798             }
0799
0800             /*
0801              * Is it really there?  Enable memory decode for
0802              * implicit access in pci_map_rom().
0803              */
0804             cmd = vfio_pci_memory_lock_and_enable(vdev);
0805             io = pci_map_rom(pdev, &size);
0806             if (io) {
0807                 info.flags = VFIO_REGION_INFO_FLAG_READ;
0808                 pci_unmap_rom(pdev, io);
0809             } else {
0810                 info.size = 0;
0811             }
0812             vfio_pci_memory_unlock_and_restore(vdev, cmd);
0813
0814             break;
0815         }
0816         case VFIO_PCI_VGA_REGION_INDEX:
0817             if (!vdev->has_vga)
0818                 return -EINVAL;
0819
0820             info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
0821             info.size = 0xc0000;
0822             info.flags = VFIO_REGION_INFO_FLAG_READ |
0823                      VFIO_REGION_INFO_FLAG_WRITE;
0824
0825             break;
0826         default:
0827         {
0828             struct vfio_region_info_cap_type cap_type = {
0829                     .header.id = VFIO_REGION_INFO_CAP_TYPE,
0830                     .header.version = 1 };
0831
0832             if (info.index >=
0833                 VFIO_PCI_NUM_REGIONS + vdev->num_regions)
0834                 return -EINVAL;
0835             info.index = array_index_nospec(info.index,
0836                             VFIO_PCI_NUM_REGIONS +
0837                             vdev->num_regions);
0838
0839             i = info.index - VFIO_PCI_NUM_REGIONS;
0840
0841             info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
0842             info.size = vdev->region[i].size;
0843             info.flags = vdev->region[i].flags;
0844
0845             cap_type.type = vdev->region[i].type;
0846             cap_type.subtype = vdev->region[i].subtype;
0847
0848             ret = vfio_info_add_capability(&caps, &cap_type.header,
0849                                sizeof(cap_type));
0850             if (ret)
0851                 return ret;
0852
0853             if (vdev->region[i].ops->add_capability) {
0854                 ret = vdev->region[i].ops->add_capability(vdev,
0855                         &vdev->region[i], &caps);
0856                 if (ret)
0857                     return ret;
0858             }
0859         }
0860         }
0861
0862         if (caps.size) {
0863             info.flags |= VFIO_REGION_INFO_FLAG_CAPS;
0864             if (info.argsz < sizeof(info) + caps.size) {
0865                 info.argsz = sizeof(info) + caps.size;
0866                 info.cap_offset = 0;
0867             } else {
0868                 vfio_info_cap_shift(&caps, sizeof(info));
0869                 if (copy_to_user((void __user *)arg +
0870                           sizeof(info), caps.buf,
0871                           caps.size)) {
0872                     kfree(caps.buf);
0873                     return -EFAULT;
0874                 }
0875                 info.cap_offset = sizeof(info);
0876             }
0877
0878             kfree(caps.buf);
0879         }
0880
0881         return copy_to_user((void __user *)arg, &info, minsz) ?
0882             -EFAULT : 0;
0883
0884     } else if (cmd == VFIO_DEVICE_GET_IRQ_INFO) {
0885         struct vfio_irq_info info;
0886
0887         minsz = offsetofend(struct vfio_irq_info, count);
0888
0889         if (copy_from_user(&info, (void __user *)arg, minsz))
0890             return -EFAULT;
0891
0892         if (info.argsz < minsz || info.index >= VFIO_PCI_NUM_IRQS)
0893             return -EINVAL;
0894
0895         switch (info.index) {
0896         case VFIO_PCI_INTX_IRQ_INDEX ... VFIO_PCI_MSIX_IRQ_INDEX:
0897         case VFIO_PCI_REQ_IRQ_INDEX:
0898             break;
0899         case VFIO_PCI_ERR_IRQ_INDEX:
0900             if (pci_is_pcie(vdev->pdev))
0901                 break;
0902             fallthrough;
0903         default:
0904             return -EINVAL;
0905         }
0906
0907         info.flags = VFIO_IRQ_INFO_EVENTFD;
0908
0909         info.count = vfio_pci_get_irq_count(vdev, info.index);
0910
0911         if (info.index == VFIO_PCI_INTX_IRQ_INDEX)
0912             info.flags |= (VFIO_IRQ_INFO_MASKABLE |
0913                        VFIO_IRQ_INFO_AUTOMASKED);
0914         else
0915             info.flags |= VFIO_IRQ_INFO_NORESIZE;
0916
0917         return copy_to_user((void __user *)arg, &info, minsz) ?
0918             -EFAULT : 0;
0919
0920     } else if (cmd == VFIO_DEVICE_SET_IRQS) {
0921         struct vfio_irq_set hdr;
0922         u8 *data = NULL;
0923         int max, ret = 0;
0924         size_t data_size = 0;
0925
0926         minsz = offsetofend(struct vfio_irq_set, count);
0927
0928         if (copy_from_user(&hdr, (void __user *)arg, minsz))
0929             return -EFAULT;
0930
0931         max = vfio_pci_get_irq_count(vdev, hdr.index);
0932
0933         ret = vfio_set_irqs_validate_and_prepare(&hdr, max,
0934                          VFIO_PCI_NUM_IRQS, &data_size);
0935         if (ret)
0936             return ret;
0937
0938         if (data_size) {
0939             data = memdup_user((void __user *)(arg + minsz),
0940                         data_size);
0941             if (IS_ERR(data))
0942                 return PTR_ERR(data);
0943         }
0944
0945         mutex_lock(&vdev->igate);
0946
0947         ret = vfio_pci_set_irqs_ioctl(vdev, hdr.flags, hdr.index,
0948                           hdr.start, hdr.count, data);
0949
0950         mutex_unlock(&vdev->igate);
0951         kfree(data);
0952
0953         return ret;
0954
0955     } else if (cmd == VFIO_DEVICE_RESET) {
0956         int ret;
0957
0958         if (!vdev->reset_works)
0959             return -EINVAL;
0960
0961         vfio_pci_zap_and_down_write_memory_lock(vdev);
0962
0963         /*
0964          * This function can be invoked while the power state is non-D0.
0965          * If pci_try_reset_function() has been called while the power
0966          * state is non-D0, then pci_try_reset_function() will
0967          * internally set the power state to D0 without vfio driver
0968          * involvement. For the devices which have NoSoftRst-, the
0969          * reset function can cause the PCI config space reset without
0970          * restoring the original state (saved locally in
0971          * 'vdev->pm_save').
0972          */
0973         vfio_pci_set_power_state(vdev, PCI_D0);
0974
0975         ret = pci_try_reset_function(vdev->pdev);
0976         up_write(&vdev->memory_lock);
0977
0978         return ret;
0979
0980     } else if (cmd == VFIO_DEVICE_GET_PCI_HOT_RESET_INFO) {
0981         struct vfio_pci_hot_reset_info hdr;
0982         struct vfio_pci_fill_info fill = { 0 };
0983         struct vfio_pci_dependent_device *devices = NULL;
0984         bool slot = false;
0985         int ret = 0;
0986
0987         minsz = offsetofend(struct vfio_pci_hot_reset_info, count);
0988
0989         if (copy_from_user(&hdr, (void __user *)arg, minsz))
0990             return -EFAULT;
0991
0992         if (hdr.argsz < minsz)
0993             return -EINVAL;
0994
0995         hdr.flags = 0;
0996
0997         /* Can we do a slot or bus reset or neither? */
0998         if (!pci_probe_reset_slot(vdev->pdev->slot))
0999             slot = true;
1000         else if (pci_probe_reset_bus(vdev->pdev->bus))
1001             return -ENODEV;
1002
1003         /* How many devices are affected? */
1004         ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1005                             vfio_pci_count_devs,
1006                             &fill.max, slot);
1007         if (ret)
1008             return ret;
1009
1010         WARN_ON(!fill.max); /* Should always be at least one */
1011
1012         /*
1013          * If there's enough space, fill it now, otherwise return
1014          * -ENOSPC and the number of devices affected.
1015          */
1016         if (hdr.argsz < sizeof(hdr) + (fill.max * sizeof(*devices))) {
1017             ret = -ENOSPC;
1018             hdr.count = fill.max;
1019             goto reset_info_exit;
1020         }
1021
1022         devices = kcalloc(fill.max, sizeof(*devices), GFP_KERNEL);
1023         if (!devices)
1024             return -ENOMEM;
1025
1026         fill.devices = devices;
1027
1028         ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1029                             vfio_pci_fill_devs,
1030                             &fill, slot);
1031
1032         /*
1033          * If a device was removed between counting and filling,
1034          * we may come up short of fill.max.  If a device was
1035          * added, we'll have a return of -EAGAIN above.
1036          */
1037         if (!ret)
1038             hdr.count = fill.cur;
1039
1040 reset_info_exit:
1041         if (copy_to_user((void __user *)arg, &hdr, minsz))
1042             ret = -EFAULT;
1043
1044         if (!ret) {
1045             if (copy_to_user((void __user *)(arg + minsz), devices,
1046                      hdr.count * sizeof(*devices)))
1047                 ret = -EFAULT;
1048         }
1049
1050         kfree(devices);
1051         return ret;
1052
1053     } else if (cmd == VFIO_DEVICE_PCI_HOT_RESET) {
1054         struct vfio_pci_hot_reset hdr;
1055         int32_t *group_fds;
1056         struct file **files;
1057         struct vfio_pci_group_info info;
1058         bool slot = false;
1059         int file_idx, count = 0, ret = 0;
1060
1061         minsz = offsetofend(struct vfio_pci_hot_reset, count);
1062
1063         if (copy_from_user(&hdr, (void __user *)arg, minsz))
1064             return -EFAULT;
1065
1066         if (hdr.argsz < minsz || hdr.flags)
1067             return -EINVAL;
1068
1069         /* Can we do a slot or bus reset or neither? */
1070         if (!pci_probe_reset_slot(vdev->pdev->slot))
1071             slot = true;
1072         else if (pci_probe_reset_bus(vdev->pdev->bus))
1073             return -ENODEV;
1074
1075         /*
1076          * We can't let userspace give us an arbitrarily large
1077          * buffer to copy, so verify how many we think there
1078          * could be.  Note groups can have multiple devices so
1079          * one group per device is the max.
1080          */
1081         ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1082                             vfio_pci_count_devs,
1083                             &count, slot);
1084         if (ret)
1085             return ret;
1086
1087         /* Somewhere between 1 and count is OK */
1088         if (!hdr.count || hdr.count > count)
1089             return -EINVAL;
1090
1091         group_fds = kcalloc(hdr.count, sizeof(*group_fds), GFP_KERNEL);
1092         files = kcalloc(hdr.count, sizeof(*files), GFP_KERNEL);
1093         if (!group_fds || !files) {
1094             kfree(group_fds);
1095             kfree(files);
1096             return -ENOMEM;
1097         }
1098
1099         if (copy_from_user(group_fds, (void __user *)(arg + minsz),
1100                    hdr.count * sizeof(*group_fds))) {
1101             kfree(group_fds);
1102             kfree(files);
1103             return -EFAULT;
1104         }
1105
1106         /*
1107          * For each group_fd, get the group through the vfio external
1108          * user interface and store the group and iommu ID.  This
1109          * ensures the group is held across the reset.
1110          */
1111         for (file_idx = 0; file_idx < hdr.count; file_idx++) {
1112             struct file *file = fget(group_fds[file_idx]);
1113
1114             if (!file) {
1115                 ret = -EBADF;
1116                 break;
1117             }
1118
1119             /* Ensure the FD is a vfio group FD.*/
1120             if (!vfio_file_iommu_group(file)) {
1121                 fput(file);
1122                 ret = -EINVAL;
1123                 break;
1124             }
1125
1126             files[file_idx] = file;
1127         }
1128
1129         kfree(group_fds);
1130
1131         /* release reference to groups on error */
1132         if (ret)
1133             goto hot_reset_release;
1134
1135         info.count = hdr.count;
1136         info.files = files;
1137
1138         ret = vfio_pci_dev_set_hot_reset(vdev->vdev.dev_set, &info);
1139
1140 hot_reset_release:
1141         for (file_idx--; file_idx >= 0; file_idx--)
1142             fput(files[file_idx]);
1143
1144         kfree(files);
1145         return ret;
1146     } else if (cmd == VFIO_DEVICE_IOEVENTFD) {
1147         struct vfio_device_ioeventfd ioeventfd;
1148         int count;
1149
1150         minsz = offsetofend(struct vfio_device_ioeventfd, fd);
1151
1152         if (copy_from_user(&ioeventfd, (void __user *)arg, minsz))
1153             return -EFAULT;
1154
1155         if (ioeventfd.argsz < minsz)
1156             return -EINVAL;
1157
1158         if (ioeventfd.flags & ~VFIO_DEVICE_IOEVENTFD_SIZE_MASK)
1159             return -EINVAL;
1160
1161         count = ioeventfd.flags & VFIO_DEVICE_IOEVENTFD_SIZE_MASK;
1162
1163         if (hweight8(count) != 1 || ioeventfd.fd < -1)
1164             return -EINVAL;
1165
1166         return vfio_pci_ioeventfd(vdev, ioeventfd.offset,
1167                       ioeventfd.data, count, ioeventfd.fd);
1168     }
1169     return -ENOTTY;
1170 }
1171 EXPORT_SYMBOL_GPL(vfio_pci_core_ioctl);
1172
1173 static int vfio_pci_core_feature_token(struct vfio_device *device, u32 flags,
1174                        void __user *arg, size_t argsz)
1175 {
1176     struct vfio_pci_core_device *vdev =
1177         container_of(device, struct vfio_pci_core_device, vdev);
1178     uuid_t uuid;
1179     int ret;
1180
1181     if (!vdev->vf_token)
1182         return -ENOTTY;
1183     /*
1184      * We do not support GET of the VF Token UUID as this could
1185      * expose the token of the previous device user.
1186      */
1187     ret = vfio_check_feature(flags, argsz, VFIO_DEVICE_FEATURE_SET,
1188                  sizeof(uuid));
1189     if (ret != 1)
1190         return ret;
1191
1192     if (copy_from_user(&uuid, arg, sizeof(uuid)))
1193         return -EFAULT;
1194
1195     mutex_lock(&vdev->vf_token->lock);
1196     uuid_copy(&vdev->vf_token->uuid, &uuid);
1197     mutex_unlock(&vdev->vf_token->lock);
1198     return 0;
1199 }
1200
1201 int vfio_pci_core_ioctl_feature(struct vfio_device *device, u32 flags,
1202                 void __user *arg, size_t argsz)
1203 {
1204     switch (flags & VFIO_DEVICE_FEATURE_MASK) {
1205     case VFIO_DEVICE_FEATURE_PCI_VF_TOKEN:
1206         return vfio_pci_core_feature_token(device, flags, arg, argsz);
1207     default:
1208         return -ENOTTY;
1209     }
1210 }
1211 EXPORT_SYMBOL_GPL(vfio_pci_core_ioctl_feature);
1212
1213 static ssize_t vfio_pci_rw(struct vfio_pci_core_device *vdev, char __user *buf,
1214                size_t count, loff_t *ppos, bool iswrite)
1215 {
1216     unsigned int index = VFIO_PCI_OFFSET_TO_INDEX(*ppos);
1217
1218     if (index >= VFIO_PCI_NUM_REGIONS + vdev->num_regions)
1219         return -EINVAL;
1220
1221     switch (index) {
1222     case VFIO_PCI_CONFIG_REGION_INDEX:
1223         return vfio_pci_config_rw(vdev, buf, count, ppos, iswrite);
1224
1225     case VFIO_PCI_ROM_REGION_INDEX:
1226         if (iswrite)
1227             return -EINVAL;
1228         return vfio_pci_bar_rw(vdev, buf, count, ppos, false);
1229
1230     case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
1231         return vfio_pci_bar_rw(vdev, buf, count, ppos, iswrite);
1232
1233     case VFIO_PCI_VGA_REGION_INDEX:
1234         return vfio_pci_vga_rw(vdev, buf, count, ppos, iswrite);
1235     default:
1236         index -= VFIO_PCI_NUM_REGIONS;
1237         return vdev->region[index].ops->rw(vdev, buf,
1238                            count, ppos, iswrite);
1239     }
1240
1241     return -EINVAL;
1242 }
1243
1244 ssize_t vfio_pci_core_read(struct vfio_device *core_vdev, char __user *buf,
1245         size_t count, loff_t *ppos)
1246 {
1247     struct vfio_pci_core_device *vdev =
1248         container_of(core_vdev, struct vfio_pci_core_device, vdev);
1249
1250     if (!count)
1251         return 0;
1252
1253     return vfio_pci_rw(vdev, buf, count, ppos, false);
1254 }
1255 EXPORT_SYMBOL_GPL(vfio_pci_core_read);
1256
1257 ssize_t vfio_pci_core_write(struct vfio_device *core_vdev, const char __user *buf,
1258         size_t count, loff_t *ppos)
1259 {
1260     struct vfio_pci_core_device *vdev =
1261         container_of(core_vdev, struct vfio_pci_core_device, vdev);
1262
1263     if (!count)
1264         return 0;
1265
1266     return vfio_pci_rw(vdev, (char __user *)buf, count, ppos, true);
1267 }
1268 EXPORT_SYMBOL_GPL(vfio_pci_core_write);
1269
1270 /* Return 1 on zap and vma_lock acquired, 0 on contention (only with @try) */
1271 static int vfio_pci_zap_and_vma_lock(struct vfio_pci_core_device *vdev, bool try)
1272 {
1273     struct vfio_pci_mmap_vma *mmap_vma, *tmp;
1274
1275     /*
1276      * Lock ordering:
1277      * vma_lock is nested under mmap_lock for vm_ops callback paths.
1278      * The memory_lock semaphore is used by both code paths calling
1279      * into this function to zap vmas and the vm_ops.fault callback
1280      * to protect the memory enable state of the device.
1281      *
1282      * When zapping vmas we need to maintain the mmap_lock => vma_lock
1283      * ordering, which requires using vma_lock to walk vma_list to
1284      * acquire an mm, then dropping vma_lock to get the mmap_lock and
1285      * reacquiring vma_lock.  This logic is derived from similar
1286      * requirements in uverbs_user_mmap_disassociate().
1287      *
1288      * mmap_lock must always be the top-level lock when it is taken.
1289      * Therefore we can only hold the memory_lock write lock when
1290      * vma_list is empty, as we'd need to take mmap_lock to clear
1291      * entries.  vma_list can only be guaranteed empty when holding
1292      * vma_lock, thus memory_lock is nested under vma_lock.
1293      *
1294      * This enables the vm_ops.fault callback to acquire vma_lock,
1295      * followed by memory_lock read lock, while already holding
1296      * mmap_lock without risk of deadlock.
1297      */
1298     while (1) {
1299         struct mm_struct *mm = NULL;
1300
1301         if (try) {
1302             if (!mutex_trylock(&vdev->vma_lock))
1303                 return 0;
1304         } else {
1305             mutex_lock(&vdev->vma_lock);
1306         }
1307         while (!list_empty(&vdev->vma_list)) {
1308             mmap_vma = list_first_entry(&vdev->vma_list,
1309                             struct vfio_pci_mmap_vma,
1310                             vma_next);
1311             mm = mmap_vma->vma->vm_mm;
1312             if (mmget_not_zero(mm))
1313                 break;
1314
1315             list_del(&mmap_vma->vma_next);
1316             kfree(mmap_vma);
1317             mm = NULL;
1318         }
1319         if (!mm)
1320             return 1;
1321         mutex_unlock(&vdev->vma_lock);
1322
1323         if (try) {
1324             if (!mmap_read_trylock(mm)) {
1325                 mmput(mm);
1326                 return 0;
1327             }
1328         } else {
1329             mmap_read_lock(mm);
1330         }
1331         if (try) {
1332             if (!mutex_trylock(&vdev->vma_lock)) {
1333                 mmap_read_unlock(mm);
1334                 mmput(mm);
1335                 return 0;
1336             }
1337         } else {
1338             mutex_lock(&vdev->vma_lock);
1339         }
1340         list_for_each_entry_safe(mmap_vma, tmp,
1341                      &vdev->vma_list, vma_next) {
1342             struct vm_area_struct *vma = mmap_vma->vma;
1343
1344             if (vma->vm_mm != mm)
1345                 continue;
1346
1347             list_del(&mmap_vma->vma_next);
1348             kfree(mmap_vma);
1349
1350             zap_vma_ptes(vma, vma->vm_start,
1351                      vma->vm_end - vma->vm_start);
1352         }
1353         mutex_unlock(&vdev->vma_lock);
1354         mmap_read_unlock(mm);
1355         mmput(mm);
1356     }
1357 }
1358
1359 void vfio_pci_zap_and_down_write_memory_lock(struct vfio_pci_core_device *vdev)
1360 {
1361     vfio_pci_zap_and_vma_lock(vdev, false);
1362     down_write(&vdev->memory_lock);
1363     mutex_unlock(&vdev->vma_lock);
1364 }
1365
1366 u16 vfio_pci_memory_lock_and_enable(struct vfio_pci_core_device *vdev)
1367 {
1368     u16 cmd;
1369
1370     down_write(&vdev->memory_lock);
1371     pci_read_config_word(vdev->pdev, PCI_COMMAND, &cmd);
1372     if (!(cmd & PCI_COMMAND_MEMORY))
1373         pci_write_config_word(vdev->pdev, PCI_COMMAND,
1374                       cmd | PCI_COMMAND_MEMORY);
1375
1376     return cmd;
1377 }
1378
1379 void vfio_pci_memory_unlock_and_restore(struct vfio_pci_core_device *vdev, u16 cmd)
1380 {
1381     pci_write_config_word(vdev->pdev, PCI_COMMAND, cmd);
1382     up_write(&vdev->memory_lock);
1383 }
1384
1385 /* Caller holds vma_lock */
1386 static int __vfio_pci_add_vma(struct vfio_pci_core_device *vdev,
1387                   struct vm_area_struct *vma)
1388 {
1389     struct vfio_pci_mmap_vma *mmap_vma;
1390
1391     mmap_vma = kmalloc(sizeof(*mmap_vma), GFP_KERNEL);
1392     if (!mmap_vma)
1393         return -ENOMEM;
1394
1395     mmap_vma->vma = vma;
1396     list_add(&mmap_vma->vma_next, &vdev->vma_list);
1397
1398     return 0;
1399 }
1400
1401 /*
1402  * Zap mmaps on open so that we can fault them in on access and therefore
1403  * our vma_list only tracks mappings accessed since last zap.
1404  */
1405 static void vfio_pci_mmap_open(struct vm_area_struct *vma)
1406 {
1407     zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
1408 }
1409
1410 static void vfio_pci_mmap_close(struct vm_area_struct *vma)
1411 {
1412     struct vfio_pci_core_device *vdev = vma->vm_private_data;
1413     struct vfio_pci_mmap_vma *mmap_vma;
1414
1415     mutex_lock(&vdev->vma_lock);
1416     list_for_each_entry(mmap_vma, &vdev->vma_list, vma_next) {
1417         if (mmap_vma->vma == vma) {
1418             list_del(&mmap_vma->vma_next);
1419             kfree(mmap_vma);
1420             break;
1421         }
1422     }
1423     mutex_unlock(&vdev->vma_lock);
1424 }
1425
1426 static vm_fault_t vfio_pci_mmap_fault(struct vm_fault *vmf)
1427 {
1428     struct vm_area_struct *vma = vmf->vma;
1429     struct vfio_pci_core_device *vdev = vma->vm_private_data;
1430     struct vfio_pci_mmap_vma *mmap_vma;
1431     vm_fault_t ret = VM_FAULT_NOPAGE;
1432
1433     mutex_lock(&vdev->vma_lock);
1434     down_read(&vdev->memory_lock);
1435
1436     if (!__vfio_pci_memory_enabled(vdev)) {
1437         ret = VM_FAULT_SIGBUS;
1438         goto up_out;
1439     }
1440
1441     /*
1442      * We populate the whole vma on fault, so we need to test whether
1443      * the vma has already been mapped, such as for concurrent faults
1444      * to the same vma.  io_remap_pfn_range() will trigger a BUG_ON if
1445      * we ask it to fill the same range again.
1446      */
1447     list_for_each_entry(mmap_vma, &vdev->vma_list, vma_next) {
1448         if (mmap_vma->vma == vma)
1449             goto up_out;
1450     }
1451
1452     if (io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
1453                    vma->vm_end - vma->vm_start,
1454                    vma->vm_page_prot)) {
1455         ret = VM_FAULT_SIGBUS;
1456         zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
1457         goto up_out;
1458     }
1459
1460     if (__vfio_pci_add_vma(vdev, vma)) {
1461         ret = VM_FAULT_OOM;
1462         zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
1463     }
1464
1465 up_out:
1466     up_read(&vdev->memory_lock);
1467     mutex_unlock(&vdev->vma_lock);
1468     return ret;
1469 }
1470
1471 static const struct vm_operations_struct vfio_pci_mmap_ops = {
1472     .open = vfio_pci_mmap_open,
1473     .close = vfio_pci_mmap_close,
1474     .fault = vfio_pci_mmap_fault,
1475 };
1476
1477 int vfio_pci_core_mmap(struct vfio_device *core_vdev, struct vm_area_struct *vma)
1478 {
1479     struct vfio_pci_core_device *vdev =
1480         container_of(core_vdev, struct vfio_pci_core_device, vdev);
1481     struct pci_dev *pdev = vdev->pdev;
1482     unsigned int index;
1483     u64 phys_len, req_len, pgoff, req_start;
1484     int ret;
1485
1486     index = vma->vm_pgoff >> (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT);
1487
1488     if (index >= VFIO_PCI_NUM_REGIONS + vdev->num_regions)
1489         return -EINVAL;
1490     if (vma->vm_end < vma->vm_start)
1491         return -EINVAL;
1492     if ((vma->vm_flags & VM_SHARED) == 0)
1493         return -EINVAL;
1494     if (index >= VFIO_PCI_NUM_REGIONS) {
1495         int regnum = index - VFIO_PCI_NUM_REGIONS;
1496         struct vfio_pci_region *region = vdev->region + regnum;
1497
1498         if (region->ops && region->ops->mmap &&
1499             (region->flags & VFIO_REGION_INFO_FLAG_MMAP))
1500             return region->ops->mmap(vdev, region, vma);
1501         return -EINVAL;
1502     }
1503     if (index >= VFIO_PCI_ROM_REGION_INDEX)
1504         return -EINVAL;
1505     if (!vdev->bar_mmap_supported[index])
1506         return -EINVAL;
1507
1508     phys_len = PAGE_ALIGN(pci_resource_len(pdev, index));
1509     req_len = vma->vm_end - vma->vm_start;
1510     pgoff = vma->vm_pgoff &
1511         ((1U << (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT)) - 1);
1512     req_start = pgoff << PAGE_SHIFT;
1513
1514     if (req_start + req_len > phys_len)
1515         return -EINVAL;
1516
1517     /*
1518      * Even though we don't make use of the barmap for the mmap,
1519      * we need to request the region and the barmap tracks that.
1520      */
1521     if (!vdev->barmap[index]) {
1522         ret = pci_request_selected_regions(pdev,
1523                            1 << index, "vfio-pci");
1524         if (ret)
1525             return ret;
1526
1527         vdev->barmap[index] = pci_iomap(pdev, index, 0);
1528         if (!vdev->barmap[index]) {
1529             pci_release_selected_regions(pdev, 1 << index);
1530             return -ENOMEM;
1531         }
1532     }
1533
1534     vma->vm_private_data = vdev;
1535     vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1536     vma->vm_pgoff = (pci_resource_start(pdev, index) >> PAGE_SHIFT) + pgoff;
1537
1538     /*
1539      * See remap_pfn_range(), called from vfio_pci_fault() but we can't
1540      * change vm_flags within the fault handler.  Set them now.
1541      */
1542     vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1543     vma->vm_ops = &vfio_pci_mmap_ops;
1544
1545     return 0;
1546 }
1547 EXPORT_SYMBOL_GPL(vfio_pci_core_mmap);
1548
1549 void vfio_pci_core_request(struct vfio_device *core_vdev, unsigned int count)
1550 {
1551     struct vfio_pci_core_device *vdev =
1552         container_of(core_vdev, struct vfio_pci_core_device, vdev);
1553     struct pci_dev *pdev = vdev->pdev;
1554
1555     mutex_lock(&vdev->igate);
1556
1557     if (vdev->req_trigger) {
1558         if (!(count % 10))
1559             pci_notice_ratelimited(pdev,
1560                 "Relaying device request to user (#%u)\n",
1561                 count);
1562         eventfd_signal(vdev->req_trigger, 1);
1563     } else if (count == 0) {
1564         pci_warn(pdev,
1565             "No device request channel registered, blocked until released by user\n");
1566     }
1567
1568     mutex_unlock(&vdev->igate);
1569 }
1570 EXPORT_SYMBOL_GPL(vfio_pci_core_request);
1571
1572 static int vfio_pci_validate_vf_token(struct vfio_pci_core_device *vdev,
1573                       bool vf_token, uuid_t *uuid)
1574 {
1575     /*
1576      * There's always some degree of trust or collaboration between SR-IOV
1577      * PF and VFs, even if just that the PF hosts the SR-IOV capability and
1578      * can disrupt VFs with a reset, but often the PF has more explicit
1579      * access to deny service to the VF or access data passed through the
1580      * VF.  We therefore require an opt-in via a shared VF token (UUID) to
1581      * represent this trust.  This both prevents that a VF driver might
1582      * assume the PF driver is a trusted, in-kernel driver, and also that
1583      * a PF driver might be replaced with a rogue driver, unknown to in-use
1584      * VF drivers.
1585      *
1586      * Therefore when presented with a VF, if the PF is a vfio device and
1587      * it is bound to the vfio-pci driver, the user needs to provide a VF
1588      * token to access the device, in the form of appending a vf_token to
1589      * the device name, for example:
1590      *
1591      * "0000:04:10.0 vf_token=bd8d9d2b-5a5f-4f5a-a211-f591514ba1f3"
1592      *
1593      * When presented with a PF which has VFs in use, the user must also
1594      * provide the current VF token to prove collaboration with existing
1595      * VF users.  If VFs are not in use, the VF token provided for the PF
1596      * device will act to set the VF token.
1597      *
1598      * If the VF token is provided but unused, an error is generated.
1599      */
1600     if (vdev->pdev->is_virtfn) {
1601         struct vfio_pci_core_device *pf_vdev = vdev->sriov_pf_core_dev;
1602         bool match;
1603
1604         if (!pf_vdev) {
1605             if (!vf_token)
1606                 return 0; /* PF is not vfio-pci, no VF token */
1607
1608             pci_info_ratelimited(vdev->pdev,
1609                 "VF token incorrectly provided, PF not bound to vfio-pci\n");
1610             return -EINVAL;
1611         }
1612
1613         if (!vf_token) {
1614             pci_info_ratelimited(vdev->pdev,
1615                 "VF token required to access device\n");
1616             return -EACCES;
1617         }
1618
1619         mutex_lock(&pf_vdev->vf_token->lock);
1620         match = uuid_equal(uuid, &pf_vdev->vf_token->uuid);
1621         mutex_unlock(&pf_vdev->vf_token->lock);
1622
1623         if (!match) {
1624             pci_info_ratelimited(vdev->pdev,
1625                 "Incorrect VF token provided for device\n");
1626             return -EACCES;
1627         }
1628     } else if (vdev->vf_token) {
1629         mutex_lock(&vdev->vf_token->lock);
1630         if (vdev->vf_token->users) {
1631             if (!vf_token) {
1632                 mutex_unlock(&vdev->vf_token->lock);
1633                 pci_info_ratelimited(vdev->pdev,
1634                     "VF token required to access device\n");
1635                 return -EACCES;
1636             }
1637
1638             if (!uuid_equal(uuid, &vdev->vf_token->uuid)) {
1639                 mutex_unlock(&vdev->vf_token->lock);
1640                 pci_info_ratelimited(vdev->pdev,
1641                     "Incorrect VF token provided for device\n");
1642                 return -EACCES;
1643             }
1644         } else if (vf_token) {
1645             uuid_copy(&vdev->vf_token->uuid, uuid);
1646         }
1647
1648         mutex_unlock(&vdev->vf_token->lock);
1649     } else if (vf_token) {
1650         pci_info_ratelimited(vdev->pdev,
1651             "VF token incorrectly provided, not a PF or VF\n");
1652         return -EINVAL;
1653     }
1654
1655     return 0;
1656 }
1657
1658 #define VF_TOKEN_ARG "vf_token="
1659
1660 int vfio_pci_core_match(struct vfio_device *core_vdev, char *buf)
1661 {
1662     struct vfio_pci_core_device *vdev =
1663         container_of(core_vdev, struct vfio_pci_core_device, vdev);
1664     bool vf_token = false;
1665     uuid_t uuid;
1666     int ret;
1667
1668     if (strncmp(pci_name(vdev->pdev), buf, strlen(pci_name(vdev->pdev))))
1669         return 0; /* No match */
1670
1671     if (strlen(buf) > strlen(pci_name(vdev->pdev))) {
1672         buf += strlen(pci_name(vdev->pdev));
1673
1674         if (*buf != ' ')
1675             return 0; /* No match: non-whitespace after name */
1676
1677         while (*buf) {
1678             if (*buf == ' ') {
1679                 buf++;
1680                 continue;
1681             }
1682
1683             if (!vf_token && !strncmp(buf, VF_TOKEN_ARG,
1684                           strlen(VF_TOKEN_ARG))) {
1685                 buf += strlen(VF_TOKEN_ARG);
1686
1687                 if (strlen(buf) < UUID_STRING_LEN)
1688                     return -EINVAL;
1689
1690                 ret = uuid_parse(buf, &uuid);
1691                 if (ret)
1692                     return ret;
1693
1694                 vf_token = true;
1695                 buf += UUID_STRING_LEN;
1696             } else {
1697                 /* Unknown/duplicate option */
1698                 return -EINVAL;
1699             }
1700         }
1701     }
1702
1703     ret = vfio_pci_validate_vf_token(vdev, vf_token, &uuid);
1704     if (ret)
1705         return ret;
1706
1707     return 1; /* Match */
1708 }
1709 EXPORT_SYMBOL_GPL(vfio_pci_core_match);
1710
1711 static int vfio_pci_bus_notifier(struct notifier_block *nb,
1712                  unsigned long action, void *data)
1713 {
1714     struct vfio_pci_core_device *vdev = container_of(nb,
1715                             struct vfio_pci_core_device, nb);
1716     struct device *dev = data;
1717     struct pci_dev *pdev = to_pci_dev(dev);
1718     struct pci_dev *physfn = pci_physfn(pdev);
1719
1720     if (action == BUS_NOTIFY_ADD_DEVICE &&
1721         pdev->is_virtfn && physfn == vdev->pdev) {
1722         pci_info(vdev->pdev, "Captured SR-IOV VF %s driver_override\n",
1723              pci_name(pdev));
1724         pdev->driver_override = kasprintf(GFP_KERNEL, "%s",
1725                           vdev->vdev.ops->name);
1726     } else if (action == BUS_NOTIFY_BOUND_DRIVER &&
1727            pdev->is_virtfn && physfn == vdev->pdev) {
1728         struct pci_driver *drv = pci_dev_driver(pdev);
1729
1730         if (drv && drv != pci_dev_driver(vdev->pdev))
1731             pci_warn(vdev->pdev,
1732                  "VF %s bound to driver %s while PF bound to driver %s\n",
1733                  pci_name(pdev), drv->name,
1734                  pci_dev_driver(vdev->pdev)->name);
1735     }
1736
1737     return 0;
1738 }
1739
1740 static int vfio_pci_vf_init(struct vfio_pci_core_device *vdev)
1741 {
1742     struct pci_dev *pdev = vdev->pdev;
1743     struct vfio_pci_core_device *cur;
1744     struct pci_dev *physfn;
1745     int ret;
1746
1747     if (pdev->is_virtfn) {
1748         /*
1749          * If this VF was created by our vfio_pci_core_sriov_configure()
1750          * then we can find the PF vfio_pci_core_device now, and due to
1751          * the locking in pci_disable_sriov() it cannot change until
1752          * this VF device driver is removed.
1753          */
1754         physfn = pci_physfn(vdev->pdev);
1755         mutex_lock(&vfio_pci_sriov_pfs_mutex);
1756         list_for_each_entry(cur, &vfio_pci_sriov_pfs, sriov_pfs_item) {
1757             if (cur->pdev == physfn) {
1758                 vdev->sriov_pf_core_dev = cur;
1759                 break;
1760             }
1761         }
1762         mutex_unlock(&vfio_pci_sriov_pfs_mutex);
1763         return 0;
1764     }
1765
1766     /* Not a SRIOV PF */
1767     if (!pdev->is_physfn)
1768         return 0;
1769
1770     vdev->vf_token = kzalloc(sizeof(*vdev->vf_token), GFP_KERNEL);
1771     if (!vdev->vf_token)
1772         return -ENOMEM;
1773
1774     mutex_init(&vdev->vf_token->lock);
1775     uuid_gen(&vdev->vf_token->uuid);
1776
1777     vdev->nb.notifier_call = vfio_pci_bus_notifier;
1778     ret = bus_register_notifier(&pci_bus_type, &vdev->nb);
1779     if (ret) {
1780         kfree(vdev->vf_token);
1781         return ret;
1782     }
1783     return 0;
1784 }
1785
1786 static void vfio_pci_vf_uninit(struct vfio_pci_core_device *vdev)
1787 {
1788     if (!vdev->vf_token)
1789         return;
1790
1791     bus_unregister_notifier(&pci_bus_type, &vdev->nb);
1792     WARN_ON(vdev->vf_token->users);
1793     mutex_destroy(&vdev->vf_token->lock);
1794     kfree(vdev->vf_token);
1795 }
1796
1797 static int vfio_pci_vga_init(struct vfio_pci_core_device *vdev)
1798 {
1799     struct pci_dev *pdev = vdev->pdev;
1800     int ret;
1801
1802     if (!vfio_pci_is_vga(pdev))
1803         return 0;
1804
1805     ret = aperture_remove_conflicting_pci_devices(pdev, vdev->vdev.ops->name);
1806     if (ret)
1807         return ret;
1808
1809     ret = vga_client_register(pdev, vfio_pci_set_decode);
1810     if (ret)
1811         return ret;
1812     vga_set_legacy_decoding(pdev, vfio_pci_set_decode(pdev, false));
1813     return 0;
1814 }
1815
1816 static void vfio_pci_vga_uninit(struct vfio_pci_core_device *vdev)
1817 {
1818     struct pci_dev *pdev = vdev->pdev;
1819
1820     if (!vfio_pci_is_vga(pdev))
1821         return;
1822     vga_client_unregister(pdev);
1823     vga_set_legacy_decoding(pdev, VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM |
1824                           VGA_RSRC_LEGACY_IO |
1825                           VGA_RSRC_LEGACY_MEM);
1826 }
1827
1828 void vfio_pci_core_init_device(struct vfio_pci_core_device *vdev,
1829                    struct pci_dev *pdev,
1830                    const struct vfio_device_ops *vfio_pci_ops)
1831 {
1832     vfio_init_group_dev(&vdev->vdev, &pdev->dev, vfio_pci_ops);
1833     vdev->pdev = pdev;
1834     vdev->irq_type = VFIO_PCI_NUM_IRQS;
1835     mutex_init(&vdev->igate);
1836     spin_lock_init(&vdev->irqlock);
1837     mutex_init(&vdev->ioeventfds_lock);
1838     INIT_LIST_HEAD(&vdev->dummy_resources_list);
1839     INIT_LIST_HEAD(&vdev->ioeventfds_list);
1840     mutex_init(&vdev->vma_lock);
1841     INIT_LIST_HEAD(&vdev->vma_list);
1842     INIT_LIST_HEAD(&vdev->sriov_pfs_item);
1843     init_rwsem(&vdev->memory_lock);
1844 }
1845 EXPORT_SYMBOL_GPL(vfio_pci_core_init_device);
1846
1847 void vfio_pci_core_uninit_device(struct vfio_pci_core_device *vdev)
1848 {
1849     mutex_destroy(&vdev->igate);
1850     mutex_destroy(&vdev->ioeventfds_lock);
1851     mutex_destroy(&vdev->vma_lock);
1852     vfio_uninit_group_dev(&vdev->vdev);
1853     kfree(vdev->region);
1854     kfree(vdev->pm_save);
1855 }
1856 EXPORT_SYMBOL_GPL(vfio_pci_core_uninit_device);
1857
1858 int vfio_pci_core_register_device(struct vfio_pci_core_device *vdev)
1859 {
1860     struct pci_dev *pdev = vdev->pdev;
1861     struct device *dev = &pdev->dev;
1862     int ret;
1863
1864     /* Drivers must set the vfio_pci_core_device to their drvdata */
1865     if (WARN_ON(vdev != dev_get_drvdata(dev)))
1866         return -EINVAL;
1867
1868     if (pdev->hdr_type != PCI_HEADER_TYPE_NORMAL)
1869         return -EINVAL;
1870
1871     if (vdev->vdev.mig_ops) {
1872         if (!(vdev->vdev.mig_ops->migration_get_state &&
1873               vdev->vdev.mig_ops->migration_set_state) ||
1874             !(vdev->vdev.migration_flags & VFIO_MIGRATION_STOP_COPY))
1875             return -EINVAL;
1876     }
1877
1878     /*
1879      * Prevent binding to PFs with VFs enabled, the VFs might be in use
1880      * by the host or other users.  We cannot capture the VFs if they
1881      * already exist, nor can we track VF users.  Disabling SR-IOV here
1882      * would initiate removing the VFs, which would unbind the driver,
1883      * which is prone to blocking if that VF is also in use by vfio-pci.
1884      * Just reject these PFs and let the user sort it out.
1885      */
1886     if (pci_num_vf(pdev)) {
1887         pci_warn(pdev, "Cannot bind to PF with SR-IOV enabled\n");
1888         return -EBUSY;
1889     }
1890
1891     if (pci_is_root_bus(pdev->bus)) {
1892         ret = vfio_assign_device_set(&vdev->vdev, vdev);
1893     } else if (!pci_probe_reset_slot(pdev->slot)) {
1894         ret = vfio_assign_device_set(&vdev->vdev, pdev->slot);
1895     } else {
1896         /*
1897          * If there is no slot reset support for this device, the whole
1898          * bus needs to be grouped together to support bus-wide resets.
1899          */
1900         ret = vfio_assign_device_set(&vdev->vdev, pdev->bus);
1901     }
1902
1903     if (ret)
1904         return ret;
1905     ret = vfio_pci_vf_init(vdev);
1906     if (ret)
1907         return ret;
1908     ret = vfio_pci_vga_init(vdev);
1909     if (ret)
1910         goto out_vf;
1911
1912     vfio_pci_probe_power_state(vdev);
1913
1914     /*
1915      * pci-core sets the device power state to an unknown value at
1916      * bootup and after being removed from a driver.  The only
1917      * transition it allows from this unknown state is to D0, which
1918      * typically happens when a driver calls pci_enable_device().
1919      * We're not ready to enable the device yet, but we do want to
1920      * be able to get to D3.  Therefore first do a D0 transition
1921      * before enabling runtime PM.
1922      */
1923     vfio_pci_set_power_state(vdev, PCI_D0);
1924
1925     dev->driver->pm = &vfio_pci_core_pm_ops;
1926     pm_runtime_allow(dev);
1927     if (!disable_idle_d3)
1928         pm_runtime_put(dev);
1929
1930     ret = vfio_register_group_dev(&vdev->vdev);
1931     if (ret)
1932         goto out_power;
1933     return 0;
1934
1935 out_power:
1936     if (!disable_idle_d3)
1937         pm_runtime_get_noresume(dev);
1938
1939     pm_runtime_forbid(dev);
1940 out_vf:
1941     vfio_pci_vf_uninit(vdev);
1942     return ret;
1943 }
1944 EXPORT_SYMBOL_GPL(vfio_pci_core_register_device);
1945
1946 void vfio_pci_core_unregister_device(struct vfio_pci_core_device *vdev)
1947 {
1948     vfio_pci_core_sriov_configure(vdev, 0);
1949
1950     vfio_unregister_group_dev(&vdev->vdev);
1951
1952     vfio_pci_vf_uninit(vdev);
1953     vfio_pci_vga_uninit(vdev);
1954
1955     if (!disable_idle_d3)
1956         pm_runtime_get_noresume(&vdev->pdev->dev);
1957
1958     pm_runtime_forbid(&vdev->pdev->dev);
1959 }
1960 EXPORT_SYMBOL_GPL(vfio_pci_core_unregister_device);
1961
1962 pci_ers_result_t vfio_pci_core_aer_err_detected(struct pci_dev *pdev,
1963                         pci_channel_state_t state)
1964 {
1965     struct vfio_pci_core_device *vdev = dev_get_drvdata(&pdev->dev);
1966
1967     mutex_lock(&vdev->igate);
1968
1969     if (vdev->err_trigger)
1970         eventfd_signal(vdev->err_trigger, 1);
1971
1972     mutex_unlock(&vdev->igate);
1973
1974     return PCI_ERS_RESULT_CAN_RECOVER;
1975 }
1976 EXPORT_SYMBOL_GPL(vfio_pci_core_aer_err_detected);
1977
1978 int vfio_pci_core_sriov_configure(struct vfio_pci_core_device *vdev,
1979                   int nr_virtfn)
1980 {
1981     struct pci_dev *pdev = vdev->pdev;
1982     int ret = 0;
1983
1984     device_lock_assert(&pdev->dev);
1985
1986     if (nr_virtfn) {
1987         mutex_lock(&vfio_pci_sriov_pfs_mutex);
1988         /*
1989          * The thread that adds the vdev to the list is the only thread
1990          * that gets to call pci_enable_sriov() and we will only allow
1991          * it to be called once without going through
1992          * pci_disable_sriov()
1993          */
1994         if (!list_empty(&vdev->sriov_pfs_item)) {
1995             ret = -EINVAL;
1996             goto out_unlock;
1997         }
1998         list_add_tail(&vdev->sriov_pfs_item, &vfio_pci_sriov_pfs);
1999         mutex_unlock(&vfio_pci_sriov_pfs_mutex);
2000
2001         /*
2002          * The PF power state should always be higher than the VF power
2003          * state. The PF can be in low power state either with runtime
2004          * power management (when there is no user) or PCI_PM_CTRL
2005          * register write by the user. If PF is in the low power state,
2006          * then change the power state to D0 first before enabling
2007          * SR-IOV. Also, this function can be called at any time, and
2008          * userspace PCI_PM_CTRL write can race against this code path,
2009          * so protect the same with 'memory_lock'.
2010          */
2011         ret = pm_runtime_resume_and_get(&pdev->dev);
2012         if (ret)
2013             goto out_del;
2014
2015         down_write(&vdev->memory_lock);
2016         vfio_pci_set_power_state(vdev, PCI_D0);
2017         ret = pci_enable_sriov(pdev, nr_virtfn);
2018         up_write(&vdev->memory_lock);
2019         if (ret) {
2020             pm_runtime_put(&pdev->dev);
2021             goto out_del;
2022         }
2023         return nr_virtfn;
2024     }
2025
2026     if (pci_num_vf(pdev)) {
2027         pci_disable_sriov(pdev);
2028         pm_runtime_put(&pdev->dev);
2029     }
2030
2031 out_del:
2032     mutex_lock(&vfio_pci_sriov_pfs_mutex);
2033     list_del_init(&vdev->sriov_pfs_item);
2034 out_unlock:
2035     mutex_unlock(&vfio_pci_sriov_pfs_mutex);
2036     return ret;
2037 }
2038 EXPORT_SYMBOL_GPL(vfio_pci_core_sriov_configure);
2039
2040 const struct pci_error_handlers vfio_pci_core_err_handlers = {
2041     .error_detected = vfio_pci_core_aer_err_detected,
2042 };
2043 EXPORT_SYMBOL_GPL(vfio_pci_core_err_handlers);
2044
2045 static bool vfio_dev_in_groups(struct vfio_pci_core_device *vdev,
2046                    struct vfio_pci_group_info *groups)
2047 {
2048     unsigned int i;
2049
2050     for (i = 0; i < groups->count; i++)
2051         if (vfio_file_has_dev(groups->files[i], &vdev->vdev))
2052             return true;
2053     return false;
2054 }
2055
2056 static int vfio_pci_is_device_in_set(struct pci_dev *pdev, void *data)
2057 {
2058     struct vfio_device_set *dev_set = data;
2059     struct vfio_device *cur;
2060
2061     list_for_each_entry(cur, &dev_set->device_list, dev_set_list)
2062         if (cur->dev == &pdev->dev)
2063             return 0;
2064     return -EBUSY;
2065 }
2066
2067 /*
2068  * vfio-core considers a group to be viable and will create a vfio_device even
2069  * if some devices are bound to drivers like pci-stub or pcieport. Here we
2070  * require all PCI devices to be inside our dev_set since that ensures they stay
2071  * put and that every driver controlling the device can co-ordinate with the
2072  * device reset.
2073  *
2074  * Returns the pci_dev to pass to pci_reset_bus() if every PCI device to be
2075  * reset is inside the dev_set, and pci_reset_bus() can succeed. NULL otherwise.
2076  */
2077 static struct pci_dev *
2078 vfio_pci_dev_set_resettable(struct vfio_device_set *dev_set)
2079 {
2080     struct pci_dev *pdev;
2081
2082     lockdep_assert_held(&dev_set->lock);
2083
2084     /*
2085      * By definition all PCI devices in the dev_set share the same PCI
2086      * reset, so any pci_dev will have the same outcomes for
2087      * pci_probe_reset_*() and pci_reset_bus().
2088      */
2089     pdev = list_first_entry(&dev_set->device_list,
2090                 struct vfio_pci_core_device,
2091                 vdev.dev_set_list)->pdev;
2092
2093     /* pci_reset_bus() is supported */
2094     if (pci_probe_reset_slot(pdev->slot) && pci_probe_reset_bus(pdev->bus))
2095         return NULL;
2096
2097     if (vfio_pci_for_each_slot_or_bus(pdev, vfio_pci_is_device_in_set,
2098                       dev_set,
2099                       !pci_probe_reset_slot(pdev->slot)))
2100         return NULL;
2101     return pdev;
2102 }
2103
2104 static int vfio_pci_dev_set_pm_runtime_get(struct vfio_device_set *dev_set)
2105 {
2106     struct vfio_pci_core_device *cur;
2107     int ret;
2108
2109     list_for_each_entry(cur, &dev_set->device_list, vdev.dev_set_list) {
2110         ret = pm_runtime_resume_and_get(&cur->pdev->dev);
2111         if (ret)
2112             goto unwind;
2113     }
2114
2115     return 0;
2116
2117 unwind:
2118     list_for_each_entry_continue_reverse(cur, &dev_set->device_list,
2119                          vdev.dev_set_list)
2120         pm_runtime_put(&cur->pdev->dev);
2121
2122     return ret;
2123 }
2124
2125 /*
2126  * We need to get memory_lock for each device, but devices can share mmap_lock,
2127  * therefore we need to zap and hold the vma_lock for each device, and only then
2128  * get each memory_lock.
2129  */
2130 static int vfio_pci_dev_set_hot_reset(struct vfio_device_set *dev_set,
2131                       struct vfio_pci_group_info *groups)
2132 {
2133     struct vfio_pci_core_device *cur_mem;
2134     struct vfio_pci_core_device *cur_vma;
2135     struct vfio_pci_core_device *cur;
2136     struct pci_dev *pdev;
2137     bool is_mem = true;
2138     int ret;
2139
2140     mutex_lock(&dev_set->lock);
2141     cur_mem = list_first_entry(&dev_set->device_list,
2142                    struct vfio_pci_core_device,
2143                    vdev.dev_set_list);
2144
2145     pdev = vfio_pci_dev_set_resettable(dev_set);
2146     if (!pdev) {
2147         ret = -EINVAL;
2148         goto err_unlock;
2149     }
2150
2151     list_for_each_entry(cur_vma, &dev_set->device_list, vdev.dev_set_list) {
2152         /*
2153          * Test whether all the affected devices are contained by the
2154          * set of groups provided by the user.
2155          */
2156         if (!vfio_dev_in_groups(cur_vma, groups)) {
2157             ret = -EINVAL;
2158             goto err_undo;
2159         }
2160
2161         /*
2162          * Locking multiple devices is prone to deadlock, runaway and
2163          * unwind if we hit contention.
2164          */
2165         if (!vfio_pci_zap_and_vma_lock(cur_vma, true)) {
2166             ret = -EBUSY;
2167             goto err_undo;
2168         }
2169     }
2170     cur_vma = NULL;
2171
2172     list_for_each_entry(cur_mem, &dev_set->device_list, vdev.dev_set_list) {
2173         if (!down_write_trylock(&cur_mem->memory_lock)) {
2174             ret = -EBUSY;
2175             goto err_undo;
2176         }
2177         mutex_unlock(&cur_mem->vma_lock);
2178     }
2179     cur_mem = NULL;
2180
2181     /*
2182      * The pci_reset_bus() will reset all the devices in the bus.
2183      * The power state can be non-D0 for some of the devices in the bus.
2184      * For these devices, the pci_reset_bus() will internally set
2185      * the power state to D0 without vfio driver involvement.
2186      * For the devices which have NoSoftRst-, the reset function can
2187      * cause the PCI config space reset without restoring the original
2188      * state (saved locally in 'vdev->pm_save').
2189      */
2190     list_for_each_entry(cur, &dev_set->device_list, vdev.dev_set_list)
2191         vfio_pci_set_power_state(cur, PCI_D0);
2192
2193     ret = pci_reset_bus(pdev);
2194
2195 err_undo:
2196     list_for_each_entry(cur, &dev_set->device_list, vdev.dev_set_list) {
2197         if (cur == cur_mem)
2198             is_mem = false;
2199         if (cur == cur_vma)
2200             break;
2201         if (is_mem)
2202             up_write(&cur->memory_lock);
2203         else
2204             mutex_unlock(&cur->vma_lock);
2205     }
2206 err_unlock:
2207     mutex_unlock(&dev_set->lock);
2208     return ret;
2209 }
2210
2211 static bool vfio_pci_dev_set_needs_reset(struct vfio_device_set *dev_set)
2212 {
2213     struct vfio_pci_core_device *cur;
2214     bool needs_reset = false;
2215
2216     list_for_each_entry(cur, &dev_set->device_list, vdev.dev_set_list) {
2217         /* No VFIO device in the set can have an open device FD */
2218         if (cur->vdev.open_count)
2219             return false;
2220         needs_reset |= cur->needs_reset;
2221     }
2222     return needs_reset;
2223 }
2224
2225 /*
2226  * If a bus or slot reset is available for the provided dev_set and:
2227  *  - All of the devices affected by that bus or slot reset are unused
2228  *  - At least one of the affected devices is marked dirty via
2229  *    needs_reset (such as by lack of FLR support)
2230  * Then attempt to perform that bus or slot reset.
2231  */
2232 static void vfio_pci_dev_set_try_reset(struct vfio_device_set *dev_set)
2233 {
2234     struct vfio_pci_core_device *cur;
2235     struct pci_dev *pdev;
2236     bool reset_done = false;
2237
2238     if (!vfio_pci_dev_set_needs_reset(dev_set))
2239         return;
2240
2241     pdev = vfio_pci_dev_set_resettable(dev_set);
2242     if (!pdev)
2243         return;
2244
2245     /*
2246      * Some of the devices in the bus can be in the runtime suspended
2247      * state. Increment the usage count for all the devices in the dev_set
2248      * before reset and decrement the same after reset.
2249      */
2250     if (!disable_idle_d3 && vfio_pci_dev_set_pm_runtime_get(dev_set))
2251         return;
2252
2253     if (!pci_reset_bus(pdev))
2254         reset_done = true;
2255
2256     list_for_each_entry(cur, &dev_set->device_list, vdev.dev_set_list) {
2257         if (reset_done)
2258             cur->needs_reset = false;
2259
2260         if (!disable_idle_d3)
2261             pm_runtime_put(&cur->pdev->dev);
2262     }
2263 }
2264
2265 void vfio_pci_core_set_params(bool is_nointxmask, bool is_disable_vga,
2266                   bool is_disable_idle_d3)
2267 {
2268     nointxmask = is_nointxmask;
2269     disable_vga = is_disable_vga;
2270     disable_idle_d3 = is_disable_idle_d3;
2271 }
2272 EXPORT_SYMBOL_GPL(vfio_pci_core_set_params);
2273
2274 static void vfio_pci_core_cleanup(void)
2275 {
2276     vfio_pci_uninit_perm_bits();
2277 }
2278
2279 static int __init vfio_pci_core_init(void)
2280 {
2281     /* Allocate shared config space permission data used by all devices */
2282     return vfio_pci_init_perm_bits();
2283 }
2284
2285 module_init(vfio_pci_core_init);
2286 module_exit(vfio_pci_core_cleanup);
2287
2288 MODULE_LICENSE("GPL v2");
2289 MODULE_AUTHOR(DRIVER_AUTHOR);
2290 MODULE_DESCRIPTION(DRIVER_DESC);