Linux Audio

Check our new training course

Loading...
v6.8
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * Copyright (C) 2012 Red Hat, Inc.  All rights reserved.
   4 *     Author: Alex Williamson <alex.williamson@redhat.com>
   5 *
   6 * Derived from original vfio:
   7 * Copyright 2010 Cisco Systems, Inc.  All rights reserved.
   8 * Author: Tom Lyon, pugs@cisco.com
   9 */
  10
  11#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  12
  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)
  32#include <asm/eeh.h>
  33#endif
  34
  35#include "vfio_pci_priv.h"
  36
  37#define DRIVER_AUTHOR   "Alex Williamson <alex.williamson@redhat.com>"
  38#define DRIVER_DESC "core driver for VFIO based PCI devices"
  39
  40static bool nointxmask;
  41static bool disable_vga;
  42static bool disable_idle_d3;
  43
  44/* List of PF's that vfio_pci_core_sriov_configure() has been called on */
  45static DEFINE_MUTEX(vfio_pci_sriov_pfs_mutex);
  46static LIST_HEAD(vfio_pci_sriov_pfs);
  47
  48struct vfio_pci_dummy_resource {
  49	struct resource		resource;
  50	int			index;
  51	struct list_head	res_next;
  52};
  53
  54struct vfio_pci_vf_token {
  55	struct mutex		lock;
  56	uuid_t			uuid;
  57	int			users;
  58};
  59
  60struct vfio_pci_mmap_vma {
  61	struct vm_area_struct	*vma;
  62	struct list_head	vma_next;
  63};
  64
  65static inline bool vfio_vga_disabled(void)
  66{
  67#ifdef CONFIG_VFIO_PCI_VGA
  68	return disable_vga;
  69#else
  70	return true;
  71#endif
  72}
  73
  74/*
  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
  80 * bridge.
  81 */
  82static unsigned int vfio_pci_set_decode(struct pci_dev *pdev, bool single_vga)
  83{
  84	struct pci_dev *tmp = NULL;
  85	unsigned char max_busnr;
  86	unsigned int decodes;
  87
  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;
  91
  92	max_busnr = pci_bus_max_busnr(pdev->bus);
  93	decodes = VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
  94
  95	while ((tmp = pci_get_class(PCI_CLASS_DISPLAY_VGA << 8, tmp)) != NULL) {
  96		if (tmp == pdev ||
  97		    pci_domain_nr(tmp->bus) != pci_domain_nr(pdev->bus) ||
  98		    pci_is_root_bus(tmp->bus))
  99			continue;
 100
 101		if (tmp->bus->number >= pdev->bus->number &&
 102		    tmp->bus->number <= max_busnr) {
 103			pci_dev_put(tmp);
 104			decodes |= VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM;
 105			break;
 106		}
 107	}
 108
 109	return decodes;
 110}
 111
 112static void vfio_pci_probe_mmaps(struct vfio_pci_core_device *vdev)
 113{
 114	struct resource *res;
 115	int i;
 116	struct vfio_pci_dummy_resource *dummy_res;
 117
 118	for (i = 0; i < PCI_STD_NUM_BARS; i++) {
 119		int bar = i + PCI_STD_RESOURCES;
 120
 121		res = &vdev->pdev->resource[bar];
 122
 123		if (!IS_ENABLED(CONFIG_VFIO_PCI_MMAP))
 124			goto no_mmap;
 125
 126		if (!(res->flags & IORESOURCE_MEM))
 127			goto no_mmap;
 128
 129		/*
 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.
 133		 */
 134		if (!resource_size(res))
 135			goto no_mmap;
 136
 137		if (resource_size(res) >= PAGE_SIZE) {
 138			vdev->bar_mmap_supported[bar] = true;
 139			continue;
 140		}
 141
 142		if (!(res->start & ~PAGE_MASK)) {
 143			/*
 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.
 147			 */
 148			dummy_res =
 149				kzalloc(sizeof(*dummy_res), GFP_KERNEL_ACCOUNT);
 150			if (dummy_res == NULL)
 151				goto no_mmap;
 152
 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)) {
 159				kfree(dummy_res);
 160				goto no_mmap;
 161			}
 162			dummy_res->index = bar;
 163			list_add(&dummy_res->res_next,
 164					&vdev->dummy_resources_list);
 165			vdev->bar_mmap_supported[bar] = true;
 166			continue;
 167		}
 168		/*
 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.
 175		 */
 176no_mmap:
 177		vdev->bar_mmap_supported[bar] = false;
 178	}
 179}
 180
 181struct vfio_pci_group_info;
 182static void vfio_pci_dev_set_try_reset(struct vfio_device_set *dev_set);
 183static 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);
 186
 187/*
 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.
 195 */
 196static bool vfio_pci_nointx(struct pci_dev *pdev)
 197{
 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 */
 202		case 0x1572:
 203		case 0x1574:
 204		case 0x1580 ... 0x1581:
 205		case 0x1583 ... 0x158b:
 206		case 0x37d0 ... 0x37d2:
 207		/* X550 */
 208		case 0x1563:
 209			return true;
 210		default:
 211			return false;
 212		}
 213	}
 214
 215	return false;
 216}
 217
 218static void vfio_pci_probe_power_state(struct vfio_pci_core_device *vdev)
 219{
 220	struct pci_dev *pdev = vdev->pdev;
 221	u16 pmcsr;
 222
 223	if (!pdev->pm_cap)
 224		return;
 225
 226	pci_read_config_word(pdev, pdev->pm_cap + PCI_PM_CTRL, &pmcsr);
 227
 228	vdev->needs_pm_restore = !(pmcsr & PCI_PM_CTRL_NO_SOFT_RESET);
 229}
 230
 231/*
 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.
 237 */
 238int vfio_pci_set_power_state(struct vfio_pci_core_device *vdev, pci_power_t state)
 239{
 240	struct pci_dev *pdev = vdev->pdev;
 241	bool needs_restore = false, needs_save = false;
 242	int ret;
 243
 244	/* Prevent changing power state for PFs with VFs enabled */
 245	if (pci_num_vf(pdev) && state > PCI_D0)
 246		return -EBUSY;
 247
 248	if (vdev->needs_pm_restore) {
 249		if (pdev->current_state < PCI_D3hot && state >= PCI_D3hot) {
 250			pci_save_state(pdev);
 251			needs_save = true;
 252		}
 253
 254		if (pdev->current_state >= PCI_D3hot && state <= PCI_D0)
 255			needs_restore = true;
 256	}
 257
 258	ret = pci_set_power_state(pdev, state);
 259
 260	if (!ret) {
 261		/* D3 might be unsupported via quirk, skip unless in D3 */
 262		if (needs_save && pdev->current_state >= PCI_D3hot) {
 263			/*
 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.
 274			 */
 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);
 280		}
 281	}
 282
 283	return ret;
 284}
 285
 286static int vfio_pci_runtime_pm_entry(struct vfio_pci_core_device *vdev,
 287				     struct eventfd_ctx *efdctx)
 288{
 289	/*
 290	 * The vdev power related flags are protected with 'memory_lock'
 291	 * semaphore.
 292	 */
 293	vfio_pci_zap_and_down_write_memory_lock(vdev);
 294	if (vdev->pm_runtime_engaged) {
 295		up_write(&vdev->memory_lock);
 296		return -EINVAL;
 297	}
 298
 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);
 303
 304	return 0;
 305}
 306
 307static int vfio_pci_core_pm_entry(struct vfio_device *device, u32 flags,
 308				  void __user *arg, size_t argsz)
 309{
 310	struct vfio_pci_core_device *vdev =
 311		container_of(device, struct vfio_pci_core_device, vdev);
 312	int ret;
 313
 314	ret = vfio_check_feature(flags, argsz, VFIO_DEVICE_FEATURE_SET, 0);
 315	if (ret != 1)
 316		return ret;
 317
 318	/*
 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.
 323	 */
 324	return vfio_pci_runtime_pm_entry(vdev, NULL);
 325}
 326
 327static 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,
 330	size_t argsz)
 331{
 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;
 336	int ret;
 337
 338	ret = vfio_check_feature(flags, argsz, VFIO_DEVICE_FEATURE_SET,
 339				 sizeof(entry));
 340	if (ret != 1)
 341		return ret;
 342
 343	if (copy_from_user(&entry, arg, sizeof(entry)))
 344		return -EFAULT;
 345
 346	if (entry.wakeup_eventfd < 0)
 347		return -EINVAL;
 348
 349	efdctx = eventfd_ctx_fdget(entry.wakeup_eventfd);
 350	if (IS_ERR(efdctx))
 351		return PTR_ERR(efdctx);
 352
 353	ret = vfio_pci_runtime_pm_entry(vdev, efdctx);
 354	if (ret)
 355		eventfd_ctx_put(efdctx);
 356
 357	return ret;
 358}
 359
 360static void __vfio_pci_runtime_pm_exit(struct vfio_pci_core_device *vdev)
 361{
 362	if (vdev->pm_runtime_engaged) {
 363		vdev->pm_runtime_engaged = false;
 364		pm_runtime_get_noresume(&vdev->pdev->dev);
 365
 366		if (vdev->pm_wake_eventfd_ctx) {
 367			eventfd_ctx_put(vdev->pm_wake_eventfd_ctx);
 368			vdev->pm_wake_eventfd_ctx = NULL;
 369		}
 370	}
 371}
 372
 373static void vfio_pci_runtime_pm_exit(struct vfio_pci_core_device *vdev)
 374{
 375	/*
 376	 * The vdev power related flags are protected with 'memory_lock'
 377	 * semaphore.
 378	 */
 379	down_write(&vdev->memory_lock);
 380	__vfio_pci_runtime_pm_exit(vdev);
 381	up_write(&vdev->memory_lock);
 382}
 383
 384static int vfio_pci_core_pm_exit(struct vfio_device *device, u32 flags,
 385				 void __user *arg, size_t argsz)
 386{
 387	struct vfio_pci_core_device *vdev =
 388		container_of(device, struct vfio_pci_core_device, vdev);
 389	int ret;
 390
 391	ret = vfio_check_feature(flags, argsz, VFIO_DEVICE_FEATURE_SET, 0);
 392	if (ret != 1)
 393		return ret;
 394
 395	/*
 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.
 401	 */
 402	vfio_pci_runtime_pm_exit(vdev);
 403	return 0;
 404}
 405
 406#ifdef CONFIG_PM
 407static int vfio_pci_core_runtime_suspend(struct device *dev)
 408{
 409	struct vfio_pci_core_device *vdev = dev_get_drvdata(dev);
 410
 411	down_write(&vdev->memory_lock);
 412	/*
 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').
 419	 */
 420	vfio_pci_set_power_state(vdev, PCI_D0);
 421	up_write(&vdev->memory_lock);
 422
 423	/*
 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.
 429	 */
 430	vdev->pm_intx_masked = ((vdev->irq_type == VFIO_PCI_INTX_IRQ_INDEX) &&
 431				vfio_pci_intx_mask(vdev));
 432
 433	return 0;
 434}
 435
 436static int vfio_pci_core_runtime_resume(struct device *dev)
 437{
 438	struct vfio_pci_core_device *vdev = dev_get_drvdata(dev);
 439
 440	/*
 441	 * Resume with a pm_wake_eventfd_ctx signals the eventfd and exit
 442	 * low power mode.
 443	 */
 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);
 448	}
 449	up_write(&vdev->memory_lock);
 450
 451	if (vdev->pm_intx_masked)
 452		vfio_pci_intx_unmask(vdev);
 453
 454	return 0;
 455}
 456#endif /* CONFIG_PM */
 457
 458/*
 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-.
 463 */
 464static 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,
 467			   NULL)
 468};
 469
 470int vfio_pci_core_enable(struct vfio_pci_core_device *vdev)
 471{
 472	struct pci_dev *pdev = vdev->pdev;
 473	int ret;
 474	u16 cmd;
 475	u8 msix_pos;
 476
 477	if (!disable_idle_d3) {
 478		ret = pm_runtime_resume_and_get(&pdev->dev);
 479		if (ret < 0)
 480			return ret;
 481	}
 482
 483	/* Don't allow our initial saved state to include busmaster */
 484	pci_clear_master(pdev);
 485
 486	ret = pci_enable_device(pdev);
 487	if (ret)
 488		goto out_power;
 489
 490	/* If reset fails because of the device lock, fail this path entirely */
 491	ret = pci_try_reset_function(pdev);
 492	if (ret == -EAGAIN)
 493		goto out_disable_device;
 494
 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__);
 500
 501	if (likely(!nointxmask)) {
 502		if (vfio_pci_nointx(pdev)) {
 503			pci_info(pdev, "Masking broken INTx support\n");
 504			vdev->nointx = true;
 505			pci_intx(pdev, 0);
 506		} else
 507			vdev->pci_2_3 = pci_intx_mask_supported(pdev);
 508	}
 509
 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);
 514	}
 515
 516	ret = vfio_pci_zdev_open_device(vdev);
 517	if (ret)
 518		goto out_free_state;
 519
 520	ret = vfio_config_init(vdev);
 521	if (ret)
 522		goto out_free_zdev;
 523
 524	msix_pos = pdev->msix_cap;
 525	if (msix_pos) {
 526		u16 flags;
 527		u32 table;
 528
 529		pci_read_config_word(pdev, msix_pos + PCI_MSIX_FLAGS, &flags);
 530		pci_read_config_dword(pdev, msix_pos + PCI_MSIX_TABLE, &table);
 531
 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);
 536	} else {
 537		vdev->msix_bar = 0xFF;
 538		vdev->has_dyn_msix = false;
 539	}
 540
 541	if (!vfio_vga_disabled() && vfio_pci_is_vga(pdev))
 542		vdev->has_vga = true;
 543
 544
 545	return 0;
 546
 547out_free_zdev:
 548	vfio_pci_zdev_close_device(vdev);
 549out_free_state:
 550	kfree(vdev->pci_saved_state);
 551	vdev->pci_saved_state = NULL;
 552out_disable_device:
 553	pci_disable_device(pdev);
 554out_power:
 555	if (!disable_idle_d3)
 556		pm_runtime_put(&pdev->dev);
 557	return ret;
 558}
 559EXPORT_SYMBOL_GPL(vfio_pci_core_enable);
 560
 561void vfio_pci_core_disable(struct vfio_pci_core_device *vdev)
 562{
 563	struct pci_dev *pdev = vdev->pdev;
 564	struct vfio_pci_dummy_resource *dummy_res, *tmp;
 565	struct vfio_pci_ioeventfd *ioeventfd, *ioeventfd_tmp;
 566	int i, bar;
 567
 568	/* For needs_reset */
 569	lockdep_assert_held(&vdev->vdev.dev_set->lock);
 570
 571	/*
 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
 579	 * to D0.
 580	 */
 581	vfio_pci_runtime_pm_exit(vdev);
 582	pm_runtime_resume(&pdev->dev);
 583
 584	/*
 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
 590	 * 'vdev->pm_save').
 591	 */
 592	vfio_pci_set_power_state(vdev, PCI_D0);
 593
 594	/* Stop the device from further DMA */
 595	pci_clear_master(pdev);
 596
 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);
 600
 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);
 606		kfree(ioeventfd);
 607	}
 608	vdev->ioeventfds_nr = 0;
 609
 610	vdev->virq_disabled = false;
 611
 612	for (i = 0; i < vdev->num_regions; i++)
 613		vdev->region[i].ops->release(vdev, &vdev->region[i]);
 614
 615	vdev->num_regions = 0;
 616	kfree(vdev->region);
 617	vdev->region = NULL; /* don't krealloc a freed pointer */
 618
 619	vfio_config_free(vdev);
 620
 621	for (i = 0; i < PCI_STD_NUM_BARS; i++) {
 622		bar = i + PCI_STD_RESOURCES;
 623		if (!vdev->barmap[bar])
 624			continue;
 625		pci_iounmap(pdev, vdev->barmap[bar]);
 626		pci_release_selected_regions(pdev, 1 << bar);
 627		vdev->barmap[bar] = NULL;
 628	}
 629
 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);
 634		kfree(dummy_res);
 635	}
 636
 637	vdev->needs_reset = true;
 638
 639	vfio_pci_zdev_close_device(vdev);
 640
 641	/*
 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
 645	 * is just busy work.
 646	 */
 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__);
 649
 650		if (!vdev->reset_works)
 651			goto out;
 652
 653		pci_save_state(pdev);
 654	}
 655
 656	/*
 657	 * Disable INTx and MSI, presumably to avoid spurious interrupts
 658	 * during reset.  Stolen from pci_reset_function()
 659	 */
 660	pci_write_config_word(pdev, PCI_COMMAND, PCI_COMMAND_INTX_DISABLE);
 661
 662	/*
 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.
 668	 */
 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);
 673	}
 674
 675	pci_restore_state(pdev);
 676out:
 677	pci_disable_device(pdev);
 678
 679	vfio_pci_dev_set_try_reset(vdev->vdev.dev_set);
 680
 681	/* Put the pm-runtime usage counter acquired during enable */
 682	if (!disable_idle_d3)
 683		pm_runtime_put(&pdev->dev);
 684}
 685EXPORT_SYMBOL_GPL(vfio_pci_core_disable);
 686
 687void vfio_pci_core_close_device(struct vfio_device *core_vdev)
 688{
 689	struct vfio_pci_core_device *vdev =
 690		container_of(core_vdev, struct vfio_pci_core_device, vdev);
 691
 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);
 697	}
 698#if IS_ENABLED(CONFIG_EEH)
 699	eeh_dev_release(vdev->pdev);
 700#endif
 701	vfio_pci_core_disable(vdev);
 702
 703	mutex_lock(&vdev->igate);
 704	if (vdev->err_trigger) {
 705		eventfd_ctx_put(vdev->err_trigger);
 706		vdev->err_trigger = NULL;
 707	}
 708	if (vdev->req_trigger) {
 709		eventfd_ctx_put(vdev->req_trigger);
 710		vdev->req_trigger = NULL;
 711	}
 712	mutex_unlock(&vdev->igate);
 713}
 714EXPORT_SYMBOL_GPL(vfio_pci_core_close_device);
 715
 716void vfio_pci_core_finish_enable(struct vfio_pci_core_device *vdev)
 717{
 718	vfio_pci_probe_mmaps(vdev);
 719#if IS_ENABLED(CONFIG_EEH)
 720	eeh_dev_open(vdev->pdev);
 721#endif
 722
 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);
 727	}
 728}
 729EXPORT_SYMBOL_GPL(vfio_pci_core_finish_enable);
 730
 731static int vfio_pci_get_irq_count(struct vfio_pci_core_device *vdev, int irq_type)
 732{
 733	if (irq_type == VFIO_PCI_INTX_IRQ_INDEX) {
 734		u8 pin;
 735
 736		if (!IS_ENABLED(CONFIG_VFIO_PCI_INTX) ||
 737		    vdev->nointx || vdev->pdev->is_virtfn)
 738			return 0;
 739
 740		pci_read_config_byte(vdev->pdev, PCI_INTERRUPT_PIN, &pin);
 741
 742		return pin ? 1 : 0;
 743	} else if (irq_type == VFIO_PCI_MSI_IRQ_INDEX) {
 744		u8 pos;
 745		u16 flags;
 746
 747		pos = vdev->pdev->msi_cap;
 748		if (pos) {
 749			pci_read_config_word(vdev->pdev,
 750					     pos + PCI_MSI_FLAGS, &flags);
 751			return 1 << ((flags & PCI_MSI_FLAGS_QMASK) >> 1);
 752		}
 753	} else if (irq_type == VFIO_PCI_MSIX_IRQ_INDEX) {
 754		u8 pos;
 755		u16 flags;
 756
 757		pos = vdev->pdev->msix_cap;
 758		if (pos) {
 759			pci_read_config_word(vdev->pdev,
 760					     pos + PCI_MSIX_FLAGS, &flags);
 761
 762			return (flags & PCI_MSIX_FLAGS_QSIZE) + 1;
 763		}
 764	} else if (irq_type == VFIO_PCI_ERR_IRQ_INDEX) {
 765		if (pci_is_pcie(vdev->pdev))
 766			return 1;
 767	} else if (irq_type == VFIO_PCI_REQ_IRQ_INDEX) {
 768		return 1;
 769	}
 770
 771	return 0;
 772}
 773
 774static int vfio_pci_count_devs(struct pci_dev *pdev, void *data)
 775{
 776	(*(int *)data)++;
 777	return 0;
 778}
 779
 780struct vfio_pci_fill_info {
 781	struct vfio_pci_dependent_device __user *devices;
 782	struct vfio_pci_dependent_device __user *devices_end;
 783	struct vfio_device *vdev;
 784	u32 count;
 785	u32 flags;
 786};
 787
 788static int vfio_pci_fill_devs(struct pci_dev *pdev, void *data)
 789{
 790	struct vfio_pci_dependent_device info = {
 791		.segment = pci_domain_nr(pdev->bus),
 792		.bus = pdev->bus->number,
 793		.devfn = pdev->devfn,
 794	};
 795	struct vfio_pci_fill_info *fill = data;
 
 796
 797	fill->count++;
 798	if (fill->devices >= fill->devices_end)
 799		return 0;
 800
 801	if (fill->flags & VFIO_PCI_HOT_RESET_FLAG_DEV_ID) {
 802		struct iommufd_ctx *iommufd = vfio_iommufd_device_ictx(fill->vdev);
 803		struct vfio_device_set *dev_set = fill->vdev->dev_set;
 804		struct vfio_device *vdev;
 805
 806		/*
 807		 * hot-reset requires all affected devices be represented in
 808		 * the dev_set.
 809		 */
 810		vdev = vfio_find_device_in_devset(dev_set, &pdev->dev);
 811		if (!vdev) {
 812			info.devid = VFIO_PCI_DEVID_NOT_OWNED;
 813		} else {
 814			int id = vfio_iommufd_get_dev_id(vdev, iommufd);
 815
 816			if (id > 0)
 817				info.devid = id;
 818			else if (id == -ENOENT)
 819				info.devid = VFIO_PCI_DEVID_OWNED;
 820			else
 821				info.devid = VFIO_PCI_DEVID_NOT_OWNED;
 822		}
 823		/* If devid is VFIO_PCI_DEVID_NOT_OWNED, clear owned flag. */
 824		if (info.devid == VFIO_PCI_DEVID_NOT_OWNED)
 825			fill->flags &= ~VFIO_PCI_HOT_RESET_FLAG_DEV_ID_OWNED;
 826	} else {
 827		struct iommu_group *iommu_group;
 828
 829		iommu_group = iommu_group_get(&pdev->dev);
 830		if (!iommu_group)
 831			return -EPERM; /* Cannot reset non-isolated devices */
 832
 833		info.group_id = iommu_group_id(iommu_group);
 834		iommu_group_put(iommu_group);
 835	}
 836
 837	if (copy_to_user(fill->devices, &info, sizeof(info)))
 838		return -EFAULT;
 839	fill->devices++;
 
 
 
 840	return 0;
 841}
 842
 843struct vfio_pci_group_info {
 844	int count;
 845	struct file **files;
 846};
 847
 848static bool vfio_pci_dev_below_slot(struct pci_dev *pdev, struct pci_slot *slot)
 849{
 850	for (; pdev; pdev = pdev->bus->self)
 851		if (pdev->bus == slot->bus)
 852			return (pdev->slot == slot);
 853	return false;
 854}
 855
 856struct vfio_pci_walk_info {
 857	int (*fn)(struct pci_dev *pdev, void *data);
 858	void *data;
 859	struct pci_dev *pdev;
 860	bool slot;
 861	int ret;
 862};
 863
 864static int vfio_pci_walk_wrapper(struct pci_dev *pdev, void *data)
 865{
 866	struct vfio_pci_walk_info *walk = data;
 867
 868	if (!walk->slot || vfio_pci_dev_below_slot(pdev, walk->pdev->slot))
 869		walk->ret = walk->fn(pdev, walk->data);
 870
 871	return walk->ret;
 872}
 873
 874static int vfio_pci_for_each_slot_or_bus(struct pci_dev *pdev,
 875					 int (*fn)(struct pci_dev *,
 876						   void *data), void *data,
 877					 bool slot)
 878{
 879	struct vfio_pci_walk_info walk = {
 880		.fn = fn, .data = data, .pdev = pdev, .slot = slot, .ret = 0,
 881	};
 882
 883	pci_walk_bus(pdev->bus, vfio_pci_walk_wrapper, &walk);
 884
 885	return walk.ret;
 886}
 887
 888static int msix_mmappable_cap(struct vfio_pci_core_device *vdev,
 889			      struct vfio_info_cap *caps)
 890{
 891	struct vfio_info_cap_header header = {
 892		.id = VFIO_REGION_INFO_CAP_MSIX_MAPPABLE,
 893		.version = 1
 894	};
 895
 896	return vfio_info_add_capability(caps, &header, sizeof(header));
 897}
 898
 899int vfio_pci_core_register_dev_region(struct vfio_pci_core_device *vdev,
 900				      unsigned int type, unsigned int subtype,
 901				      const struct vfio_pci_regops *ops,
 902				      size_t size, u32 flags, void *data)
 903{
 904	struct vfio_pci_region *region;
 905
 906	region = krealloc(vdev->region,
 907			  (vdev->num_regions + 1) * sizeof(*region),
 908			  GFP_KERNEL_ACCOUNT);
 909	if (!region)
 910		return -ENOMEM;
 911
 912	vdev->region = region;
 913	vdev->region[vdev->num_regions].type = type;
 914	vdev->region[vdev->num_regions].subtype = subtype;
 915	vdev->region[vdev->num_regions].ops = ops;
 916	vdev->region[vdev->num_regions].size = size;
 917	vdev->region[vdev->num_regions].flags = flags;
 918	vdev->region[vdev->num_regions].data = data;
 919
 920	vdev->num_regions++;
 921
 922	return 0;
 923}
 924EXPORT_SYMBOL_GPL(vfio_pci_core_register_dev_region);
 925
 926static int vfio_pci_info_atomic_cap(struct vfio_pci_core_device *vdev,
 927				    struct vfio_info_cap *caps)
 928{
 929	struct vfio_device_info_cap_pci_atomic_comp cap = {
 930		.header.id = VFIO_DEVICE_INFO_CAP_PCI_ATOMIC_COMP,
 931		.header.version = 1
 932	};
 933	struct pci_dev *pdev = pci_physfn(vdev->pdev);
 934	u32 devcap2;
 935
 936	pcie_capability_read_dword(pdev, PCI_EXP_DEVCAP2, &devcap2);
 937
 938	if ((devcap2 & PCI_EXP_DEVCAP2_ATOMIC_COMP32) &&
 939	    !pci_enable_atomic_ops_to_root(pdev, PCI_EXP_DEVCAP2_ATOMIC_COMP32))
 940		cap.flags |= VFIO_PCI_ATOMIC_COMP32;
 941
 942	if ((devcap2 & PCI_EXP_DEVCAP2_ATOMIC_COMP64) &&
 943	    !pci_enable_atomic_ops_to_root(pdev, PCI_EXP_DEVCAP2_ATOMIC_COMP64))
 944		cap.flags |= VFIO_PCI_ATOMIC_COMP64;
 945
 946	if ((devcap2 & PCI_EXP_DEVCAP2_ATOMIC_COMP128) &&
 947	    !pci_enable_atomic_ops_to_root(pdev,
 948					   PCI_EXP_DEVCAP2_ATOMIC_COMP128))
 949		cap.flags |= VFIO_PCI_ATOMIC_COMP128;
 950
 951	if (!cap.flags)
 952		return -ENODEV;
 953
 954	return vfio_info_add_capability(caps, &cap.header, sizeof(cap));
 955}
 956
 957static int vfio_pci_ioctl_get_info(struct vfio_pci_core_device *vdev,
 958				   struct vfio_device_info __user *arg)
 959{
 960	unsigned long minsz = offsetofend(struct vfio_device_info, num_irqs);
 961	struct vfio_device_info info = {};
 962	struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
 
 963	int ret;
 964
 
 
 
 965	if (copy_from_user(&info, arg, minsz))
 966		return -EFAULT;
 967
 968	if (info.argsz < minsz)
 969		return -EINVAL;
 970
 971	minsz = min_t(size_t, info.argsz, sizeof(info));
 
 
 
 972
 973	info.flags = VFIO_DEVICE_FLAGS_PCI;
 974
 975	if (vdev->reset_works)
 976		info.flags |= VFIO_DEVICE_FLAGS_RESET;
 977
 978	info.num_regions = VFIO_PCI_NUM_REGIONS + vdev->num_regions;
 979	info.num_irqs = VFIO_PCI_NUM_IRQS;
 980
 981	ret = vfio_pci_info_zdev_add_caps(vdev, &caps);
 982	if (ret && ret != -ENODEV) {
 983		pci_warn(vdev->pdev,
 984			 "Failed to setup zPCI info capabilities\n");
 985		return ret;
 986	}
 987
 988	ret = vfio_pci_info_atomic_cap(vdev, &caps);
 989	if (ret && ret != -ENODEV) {
 990		pci_warn(vdev->pdev,
 991			 "Failed to setup AtomicOps info capability\n");
 992		return ret;
 993	}
 994
 995	if (caps.size) {
 996		info.flags |= VFIO_DEVICE_FLAGS_CAPS;
 997		if (info.argsz < sizeof(info) + caps.size) {
 998			info.argsz = sizeof(info) + caps.size;
 999		} else {
1000			vfio_info_cap_shift(&caps, sizeof(info));
1001			if (copy_to_user(arg + 1, caps.buf, caps.size)) {
1002				kfree(caps.buf);
1003				return -EFAULT;
1004			}
1005			info.cap_offset = sizeof(*arg);
1006		}
1007
1008		kfree(caps.buf);
1009	}
1010
1011	return copy_to_user(arg, &info, minsz) ? -EFAULT : 0;
1012}
1013
1014static int vfio_pci_ioctl_get_region_info(struct vfio_pci_core_device *vdev,
1015					  struct vfio_region_info __user *arg)
1016{
1017	unsigned long minsz = offsetofend(struct vfio_region_info, offset);
1018	struct pci_dev *pdev = vdev->pdev;
1019	struct vfio_region_info info;
1020	struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
1021	int i, ret;
1022
1023	if (copy_from_user(&info, arg, minsz))
1024		return -EFAULT;
1025
1026	if (info.argsz < minsz)
1027		return -EINVAL;
1028
1029	switch (info.index) {
1030	case VFIO_PCI_CONFIG_REGION_INDEX:
1031		info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
1032		info.size = pdev->cfg_size;
1033		info.flags = VFIO_REGION_INFO_FLAG_READ |
1034			     VFIO_REGION_INFO_FLAG_WRITE;
1035		break;
1036	case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
1037		info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
1038		info.size = pci_resource_len(pdev, info.index);
1039		if (!info.size) {
1040			info.flags = 0;
1041			break;
1042		}
1043
1044		info.flags = VFIO_REGION_INFO_FLAG_READ |
1045			     VFIO_REGION_INFO_FLAG_WRITE;
1046		if (vdev->bar_mmap_supported[info.index]) {
1047			info.flags |= VFIO_REGION_INFO_FLAG_MMAP;
1048			if (info.index == vdev->msix_bar) {
1049				ret = msix_mmappable_cap(vdev, &caps);
1050				if (ret)
1051					return ret;
1052			}
1053		}
1054
1055		break;
1056	case VFIO_PCI_ROM_REGION_INDEX: {
1057		void __iomem *io;
1058		size_t size;
1059		u16 cmd;
1060
1061		info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
1062		info.flags = 0;
1063
1064		/* Report the BAR size, not the ROM size */
1065		info.size = pci_resource_len(pdev, info.index);
1066		if (!info.size) {
1067			/* Shadow ROMs appear as PCI option ROMs */
1068			if (pdev->resource[PCI_ROM_RESOURCE].flags &
1069			    IORESOURCE_ROM_SHADOW)
1070				info.size = 0x20000;
1071			else
1072				break;
1073		}
1074
1075		/*
1076		 * Is it really there?  Enable memory decode for implicit access
1077		 * in pci_map_rom().
1078		 */
1079		cmd = vfio_pci_memory_lock_and_enable(vdev);
1080		io = pci_map_rom(pdev, &size);
1081		if (io) {
1082			info.flags = VFIO_REGION_INFO_FLAG_READ;
1083			pci_unmap_rom(pdev, io);
1084		} else {
1085			info.size = 0;
1086		}
1087		vfio_pci_memory_unlock_and_restore(vdev, cmd);
1088
1089		break;
1090	}
1091	case VFIO_PCI_VGA_REGION_INDEX:
1092		if (!vdev->has_vga)
1093			return -EINVAL;
1094
1095		info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
1096		info.size = 0xc0000;
1097		info.flags = VFIO_REGION_INFO_FLAG_READ |
1098			     VFIO_REGION_INFO_FLAG_WRITE;
1099
1100		break;
1101	default: {
1102		struct vfio_region_info_cap_type cap_type = {
1103			.header.id = VFIO_REGION_INFO_CAP_TYPE,
1104			.header.version = 1
1105		};
1106
1107		if (info.index >= VFIO_PCI_NUM_REGIONS + vdev->num_regions)
1108			return -EINVAL;
1109		info.index = array_index_nospec(
1110			info.index, VFIO_PCI_NUM_REGIONS + vdev->num_regions);
1111
1112		i = info.index - VFIO_PCI_NUM_REGIONS;
1113
1114		info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
1115		info.size = vdev->region[i].size;
1116		info.flags = vdev->region[i].flags;
1117
1118		cap_type.type = vdev->region[i].type;
1119		cap_type.subtype = vdev->region[i].subtype;
1120
1121		ret = vfio_info_add_capability(&caps, &cap_type.header,
1122					       sizeof(cap_type));
1123		if (ret)
1124			return ret;
1125
1126		if (vdev->region[i].ops->add_capability) {
1127			ret = vdev->region[i].ops->add_capability(
1128				vdev, &vdev->region[i], &caps);
1129			if (ret)
1130				return ret;
1131		}
1132	}
1133	}
1134
1135	if (caps.size) {
1136		info.flags |= VFIO_REGION_INFO_FLAG_CAPS;
1137		if (info.argsz < sizeof(info) + caps.size) {
1138			info.argsz = sizeof(info) + caps.size;
1139			info.cap_offset = 0;
1140		} else {
1141			vfio_info_cap_shift(&caps, sizeof(info));
1142			if (copy_to_user(arg + 1, caps.buf, caps.size)) {
1143				kfree(caps.buf);
1144				return -EFAULT;
1145			}
1146			info.cap_offset = sizeof(*arg);
1147		}
1148
1149		kfree(caps.buf);
1150	}
1151
1152	return copy_to_user(arg, &info, minsz) ? -EFAULT : 0;
1153}
1154
1155static int vfio_pci_ioctl_get_irq_info(struct vfio_pci_core_device *vdev,
1156				       struct vfio_irq_info __user *arg)
1157{
1158	unsigned long minsz = offsetofend(struct vfio_irq_info, count);
1159	struct vfio_irq_info info;
1160
1161	if (copy_from_user(&info, arg, minsz))
1162		return -EFAULT;
1163
1164	if (info.argsz < minsz || info.index >= VFIO_PCI_NUM_IRQS)
1165		return -EINVAL;
1166
1167	switch (info.index) {
1168	case VFIO_PCI_INTX_IRQ_INDEX ... VFIO_PCI_MSIX_IRQ_INDEX:
1169	case VFIO_PCI_REQ_IRQ_INDEX:
1170		break;
1171	case VFIO_PCI_ERR_IRQ_INDEX:
1172		if (pci_is_pcie(vdev->pdev))
1173			break;
1174		fallthrough;
1175	default:
1176		return -EINVAL;
1177	}
1178
1179	info.flags = VFIO_IRQ_INFO_EVENTFD;
1180
1181	info.count = vfio_pci_get_irq_count(vdev, info.index);
1182
1183	if (info.index == VFIO_PCI_INTX_IRQ_INDEX)
1184		info.flags |=
1185			(VFIO_IRQ_INFO_MASKABLE | VFIO_IRQ_INFO_AUTOMASKED);
1186	else if (info.index != VFIO_PCI_MSIX_IRQ_INDEX || !vdev->has_dyn_msix)
1187		info.flags |= VFIO_IRQ_INFO_NORESIZE;
1188
1189	return copy_to_user(arg, &info, minsz) ? -EFAULT : 0;
1190}
1191
1192static int vfio_pci_ioctl_set_irqs(struct vfio_pci_core_device *vdev,
1193				   struct vfio_irq_set __user *arg)
1194{
1195	unsigned long minsz = offsetofend(struct vfio_irq_set, count);
1196	struct vfio_irq_set hdr;
1197	u8 *data = NULL;
1198	int max, ret = 0;
1199	size_t data_size = 0;
1200
1201	if (copy_from_user(&hdr, arg, minsz))
1202		return -EFAULT;
1203
1204	max = vfio_pci_get_irq_count(vdev, hdr.index);
1205
1206	ret = vfio_set_irqs_validate_and_prepare(&hdr, max, VFIO_PCI_NUM_IRQS,
1207						 &data_size);
1208	if (ret)
1209		return ret;
1210
1211	if (data_size) {
1212		data = memdup_user(&arg->data, data_size);
1213		if (IS_ERR(data))
1214			return PTR_ERR(data);
1215	}
1216
1217	mutex_lock(&vdev->igate);
1218
1219	ret = vfio_pci_set_irqs_ioctl(vdev, hdr.flags, hdr.index, hdr.start,
1220				      hdr.count, data);
1221
1222	mutex_unlock(&vdev->igate);
1223	kfree(data);
1224
1225	return ret;
1226}
1227
1228static int vfio_pci_ioctl_reset(struct vfio_pci_core_device *vdev,
1229				void __user *arg)
1230{
1231	int ret;
1232
1233	if (!vdev->reset_works)
1234		return -EINVAL;
1235
1236	vfio_pci_zap_and_down_write_memory_lock(vdev);
1237
1238	/*
1239	 * This function can be invoked while the power state is non-D0. If
1240	 * pci_try_reset_function() has been called while the power state is
1241	 * non-D0, then pci_try_reset_function() will internally set the power
1242	 * state to D0 without vfio driver involvement. For the devices which
1243	 * have NoSoftRst-, the reset function can cause the PCI config space
1244	 * reset without restoring the original state (saved locally in
1245	 * 'vdev->pm_save').
1246	 */
1247	vfio_pci_set_power_state(vdev, PCI_D0);
1248
1249	ret = pci_try_reset_function(vdev->pdev);
1250	up_write(&vdev->memory_lock);
1251
1252	return ret;
1253}
1254
1255static int vfio_pci_ioctl_get_pci_hot_reset_info(
1256	struct vfio_pci_core_device *vdev,
1257	struct vfio_pci_hot_reset_info __user *arg)
1258{
1259	unsigned long minsz =
1260		offsetofend(struct vfio_pci_hot_reset_info, count);
1261	struct vfio_pci_hot_reset_info hdr;
1262	struct vfio_pci_fill_info fill = {};
 
1263	bool slot = false;
1264	int ret = 0;
1265
1266	if (copy_from_user(&hdr, arg, minsz))
1267		return -EFAULT;
1268
1269	if (hdr.argsz < minsz)
1270		return -EINVAL;
1271
1272	hdr.flags = 0;
1273
1274	/* Can we do a slot or bus reset or neither? */
1275	if (!pci_probe_reset_slot(vdev->pdev->slot))
1276		slot = true;
1277	else if (pci_probe_reset_bus(vdev->pdev->bus))
1278		return -ENODEV;
1279
1280	fill.devices = arg->devices;
1281	fill.devices_end = arg->devices +
1282			   (hdr.argsz - sizeof(hdr)) / sizeof(arg->devices[0]);
1283	fill.vdev = &vdev->vdev;
1284
1285	if (vfio_device_cdev_opened(&vdev->vdev))
1286		fill.flags |= VFIO_PCI_HOT_RESET_FLAG_DEV_ID |
1287			     VFIO_PCI_HOT_RESET_FLAG_DEV_ID_OWNED;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1288
1289	mutex_lock(&vdev->vdev.dev_set->lock);
1290	ret = vfio_pci_for_each_slot_or_bus(vdev->pdev, vfio_pci_fill_devs,
1291					    &fill, slot);
1292	mutex_unlock(&vdev->vdev.dev_set->lock);
1293	if (ret)
1294		return ret;
1295
1296	hdr.count = fill.count;
1297	hdr.flags = fill.flags;
 
 
 
 
 
 
 
1298	if (copy_to_user(arg, &hdr, minsz))
1299		return -EFAULT;
1300
1301	if (fill.count > fill.devices - arg->devices)
1302		return -ENOSPC;
1303	return 0;
 
 
 
 
 
1304}
1305
1306static int
1307vfio_pci_ioctl_pci_hot_reset_groups(struct vfio_pci_core_device *vdev,
1308				    int array_count, bool slot,
1309				    struct vfio_pci_hot_reset __user *arg)
1310{
 
 
1311	int32_t *group_fds;
1312	struct file **files;
1313	struct vfio_pci_group_info info;
 
1314	int file_idx, count = 0, ret = 0;
1315
 
 
 
 
 
 
 
 
 
 
 
 
1316	/*
1317	 * We can't let userspace give us an arbitrarily large buffer to copy,
1318	 * so verify how many we think there could be.  Note groups can have
1319	 * multiple devices so one group per device is the max.
1320	 */
1321	ret = vfio_pci_for_each_slot_or_bus(vdev->pdev, vfio_pci_count_devs,
1322					    &count, slot);
1323	if (ret)
1324		return ret;
1325
1326	if (array_count > count)
 
1327		return -EINVAL;
1328
1329	group_fds = kcalloc(array_count, sizeof(*group_fds), GFP_KERNEL);
1330	files = kcalloc(array_count, sizeof(*files), GFP_KERNEL);
1331	if (!group_fds || !files) {
1332		kfree(group_fds);
1333		kfree(files);
1334		return -ENOMEM;
1335	}
1336
1337	if (copy_from_user(group_fds, arg->group_fds,
1338			   array_count * sizeof(*group_fds))) {
1339		kfree(group_fds);
1340		kfree(files);
1341		return -EFAULT;
1342	}
1343
1344	/*
1345	 * Get the group file for each fd to ensure the group is held across
1346	 * the reset
 
1347	 */
1348	for (file_idx = 0; file_idx < array_count; file_idx++) {
1349		struct file *file = fget(group_fds[file_idx]);
1350
1351		if (!file) {
1352			ret = -EBADF;
1353			break;
1354		}
1355
1356		/* Ensure the FD is a vfio group FD.*/
1357		if (!vfio_file_is_group(file)) {
1358			fput(file);
1359			ret = -EINVAL;
1360			break;
1361		}
1362
1363		files[file_idx] = file;
1364	}
1365
1366	kfree(group_fds);
1367
1368	/* release reference to groups on error */
1369	if (ret)
1370		goto hot_reset_release;
1371
1372	info.count = array_count;
1373	info.files = files;
1374
1375	ret = vfio_pci_dev_set_hot_reset(vdev->vdev.dev_set, &info, NULL);
1376
1377hot_reset_release:
1378	for (file_idx--; file_idx >= 0; file_idx--)
1379		fput(files[file_idx]);
1380
1381	kfree(files);
1382	return ret;
1383}
1384
1385static int vfio_pci_ioctl_pci_hot_reset(struct vfio_pci_core_device *vdev,
1386					struct vfio_pci_hot_reset __user *arg)
1387{
1388	unsigned long minsz = offsetofend(struct vfio_pci_hot_reset, count);
1389	struct vfio_pci_hot_reset hdr;
1390	bool slot = false;
1391
1392	if (copy_from_user(&hdr, arg, minsz))
1393		return -EFAULT;
1394
1395	if (hdr.argsz < minsz || hdr.flags)
1396		return -EINVAL;
1397
1398	/* zero-length array is only for cdev opened devices */
1399	if (!!hdr.count == vfio_device_cdev_opened(&vdev->vdev))
1400		return -EINVAL;
1401
1402	/* Can we do a slot or bus reset or neither? */
1403	if (!pci_probe_reset_slot(vdev->pdev->slot))
1404		slot = true;
1405	else if (pci_probe_reset_bus(vdev->pdev->bus))
1406		return -ENODEV;
1407
1408	if (hdr.count)
1409		return vfio_pci_ioctl_pci_hot_reset_groups(vdev, hdr.count, slot, arg);
1410
1411	return vfio_pci_dev_set_hot_reset(vdev->vdev.dev_set, NULL,
1412					  vfio_iommufd_device_ictx(&vdev->vdev));
1413}
1414
1415static int vfio_pci_ioctl_ioeventfd(struct vfio_pci_core_device *vdev,
1416				    struct vfio_device_ioeventfd __user *arg)
1417{
1418	unsigned long minsz = offsetofend(struct vfio_device_ioeventfd, fd);
1419	struct vfio_device_ioeventfd ioeventfd;
1420	int count;
1421
1422	if (copy_from_user(&ioeventfd, arg, minsz))
1423		return -EFAULT;
1424
1425	if (ioeventfd.argsz < minsz)
1426		return -EINVAL;
1427
1428	if (ioeventfd.flags & ~VFIO_DEVICE_IOEVENTFD_SIZE_MASK)
1429		return -EINVAL;
1430
1431	count = ioeventfd.flags & VFIO_DEVICE_IOEVENTFD_SIZE_MASK;
1432
1433	if (hweight8(count) != 1 || ioeventfd.fd < -1)
1434		return -EINVAL;
1435
1436	return vfio_pci_ioeventfd(vdev, ioeventfd.offset, ioeventfd.data, count,
1437				  ioeventfd.fd);
1438}
1439
1440long vfio_pci_core_ioctl(struct vfio_device *core_vdev, unsigned int cmd,
1441			 unsigned long arg)
1442{
1443	struct vfio_pci_core_device *vdev =
1444		container_of(core_vdev, struct vfio_pci_core_device, vdev);
1445	void __user *uarg = (void __user *)arg;
1446
1447	switch (cmd) {
1448	case VFIO_DEVICE_GET_INFO:
1449		return vfio_pci_ioctl_get_info(vdev, uarg);
1450	case VFIO_DEVICE_GET_IRQ_INFO:
1451		return vfio_pci_ioctl_get_irq_info(vdev, uarg);
1452	case VFIO_DEVICE_GET_PCI_HOT_RESET_INFO:
1453		return vfio_pci_ioctl_get_pci_hot_reset_info(vdev, uarg);
1454	case VFIO_DEVICE_GET_REGION_INFO:
1455		return vfio_pci_ioctl_get_region_info(vdev, uarg);
1456	case VFIO_DEVICE_IOEVENTFD:
1457		return vfio_pci_ioctl_ioeventfd(vdev, uarg);
1458	case VFIO_DEVICE_PCI_HOT_RESET:
1459		return vfio_pci_ioctl_pci_hot_reset(vdev, uarg);
1460	case VFIO_DEVICE_RESET:
1461		return vfio_pci_ioctl_reset(vdev, uarg);
1462	case VFIO_DEVICE_SET_IRQS:
1463		return vfio_pci_ioctl_set_irqs(vdev, uarg);
1464	default:
1465		return -ENOTTY;
1466	}
1467}
1468EXPORT_SYMBOL_GPL(vfio_pci_core_ioctl);
1469
1470static int vfio_pci_core_feature_token(struct vfio_device *device, u32 flags,
1471				       uuid_t __user *arg, size_t argsz)
1472{
1473	struct vfio_pci_core_device *vdev =
1474		container_of(device, struct vfio_pci_core_device, vdev);
1475	uuid_t uuid;
1476	int ret;
1477
1478	if (!vdev->vf_token)
1479		return -ENOTTY;
1480	/*
1481	 * We do not support GET of the VF Token UUID as this could
1482	 * expose the token of the previous device user.
1483	 */
1484	ret = vfio_check_feature(flags, argsz, VFIO_DEVICE_FEATURE_SET,
1485				 sizeof(uuid));
1486	if (ret != 1)
1487		return ret;
1488
1489	if (copy_from_user(&uuid, arg, sizeof(uuid)))
1490		return -EFAULT;
1491
1492	mutex_lock(&vdev->vf_token->lock);
1493	uuid_copy(&vdev->vf_token->uuid, &uuid);
1494	mutex_unlock(&vdev->vf_token->lock);
1495	return 0;
1496}
1497
1498int vfio_pci_core_ioctl_feature(struct vfio_device *device, u32 flags,
1499				void __user *arg, size_t argsz)
1500{
1501	switch (flags & VFIO_DEVICE_FEATURE_MASK) {
1502	case VFIO_DEVICE_FEATURE_LOW_POWER_ENTRY:
1503		return vfio_pci_core_pm_entry(device, flags, arg, argsz);
1504	case VFIO_DEVICE_FEATURE_LOW_POWER_ENTRY_WITH_WAKEUP:
1505		return vfio_pci_core_pm_entry_with_wakeup(device, flags,
1506							  arg, argsz);
1507	case VFIO_DEVICE_FEATURE_LOW_POWER_EXIT:
1508		return vfio_pci_core_pm_exit(device, flags, arg, argsz);
1509	case VFIO_DEVICE_FEATURE_PCI_VF_TOKEN:
1510		return vfio_pci_core_feature_token(device, flags, arg, argsz);
1511	default:
1512		return -ENOTTY;
1513	}
1514}
1515EXPORT_SYMBOL_GPL(vfio_pci_core_ioctl_feature);
1516
1517static ssize_t vfio_pci_rw(struct vfio_pci_core_device *vdev, char __user *buf,
1518			   size_t count, loff_t *ppos, bool iswrite)
1519{
1520	unsigned int index = VFIO_PCI_OFFSET_TO_INDEX(*ppos);
1521	int ret;
1522
1523	if (index >= VFIO_PCI_NUM_REGIONS + vdev->num_regions)
1524		return -EINVAL;
1525
1526	ret = pm_runtime_resume_and_get(&vdev->pdev->dev);
1527	if (ret) {
1528		pci_info_ratelimited(vdev->pdev, "runtime resume failed %d\n",
1529				     ret);
1530		return -EIO;
1531	}
1532
1533	switch (index) {
1534	case VFIO_PCI_CONFIG_REGION_INDEX:
1535		ret = vfio_pci_config_rw(vdev, buf, count, ppos, iswrite);
1536		break;
1537
1538	case VFIO_PCI_ROM_REGION_INDEX:
1539		if (iswrite)
1540			ret = -EINVAL;
1541		else
1542			ret = vfio_pci_bar_rw(vdev, buf, count, ppos, false);
1543		break;
1544
1545	case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
1546		ret = vfio_pci_bar_rw(vdev, buf, count, ppos, iswrite);
1547		break;
1548
1549	case VFIO_PCI_VGA_REGION_INDEX:
1550		ret = vfio_pci_vga_rw(vdev, buf, count, ppos, iswrite);
1551		break;
1552
1553	default:
1554		index -= VFIO_PCI_NUM_REGIONS;
1555		ret = vdev->region[index].ops->rw(vdev, buf,
1556						   count, ppos, iswrite);
1557		break;
1558	}
1559
1560	pm_runtime_put(&vdev->pdev->dev);
1561	return ret;
1562}
1563
1564ssize_t vfio_pci_core_read(struct vfio_device *core_vdev, char __user *buf,
1565		size_t count, loff_t *ppos)
1566{
1567	struct vfio_pci_core_device *vdev =
1568		container_of(core_vdev, struct vfio_pci_core_device, vdev);
1569
1570	if (!count)
1571		return 0;
1572
1573	return vfio_pci_rw(vdev, buf, count, ppos, false);
1574}
1575EXPORT_SYMBOL_GPL(vfio_pci_core_read);
1576
1577ssize_t vfio_pci_core_write(struct vfio_device *core_vdev, const char __user *buf,
1578		size_t count, loff_t *ppos)
1579{
1580	struct vfio_pci_core_device *vdev =
1581		container_of(core_vdev, struct vfio_pci_core_device, vdev);
1582
1583	if (!count)
1584		return 0;
1585
1586	return vfio_pci_rw(vdev, (char __user *)buf, count, ppos, true);
1587}
1588EXPORT_SYMBOL_GPL(vfio_pci_core_write);
1589
1590/* Return 1 on zap and vma_lock acquired, 0 on contention (only with @try) */
1591static int vfio_pci_zap_and_vma_lock(struct vfio_pci_core_device *vdev, bool try)
1592{
1593	struct vfio_pci_mmap_vma *mmap_vma, *tmp;
1594
1595	/*
1596	 * Lock ordering:
1597	 * vma_lock is nested under mmap_lock for vm_ops callback paths.
1598	 * The memory_lock semaphore is used by both code paths calling
1599	 * into this function to zap vmas and the vm_ops.fault callback
1600	 * to protect the memory enable state of the device.
1601	 *
1602	 * When zapping vmas we need to maintain the mmap_lock => vma_lock
1603	 * ordering, which requires using vma_lock to walk vma_list to
1604	 * acquire an mm, then dropping vma_lock to get the mmap_lock and
1605	 * reacquiring vma_lock.  This logic is derived from similar
1606	 * requirements in uverbs_user_mmap_disassociate().
1607	 *
1608	 * mmap_lock must always be the top-level lock when it is taken.
1609	 * Therefore we can only hold the memory_lock write lock when
1610	 * vma_list is empty, as we'd need to take mmap_lock to clear
1611	 * entries.  vma_list can only be guaranteed empty when holding
1612	 * vma_lock, thus memory_lock is nested under vma_lock.
1613	 *
1614	 * This enables the vm_ops.fault callback to acquire vma_lock,
1615	 * followed by memory_lock read lock, while already holding
1616	 * mmap_lock without risk of deadlock.
1617	 */
1618	while (1) {
1619		struct mm_struct *mm = NULL;
1620
1621		if (try) {
1622			if (!mutex_trylock(&vdev->vma_lock))
1623				return 0;
1624		} else {
1625			mutex_lock(&vdev->vma_lock);
1626		}
1627		while (!list_empty(&vdev->vma_list)) {
1628			mmap_vma = list_first_entry(&vdev->vma_list,
1629						    struct vfio_pci_mmap_vma,
1630						    vma_next);
1631			mm = mmap_vma->vma->vm_mm;
1632			if (mmget_not_zero(mm))
1633				break;
1634
1635			list_del(&mmap_vma->vma_next);
1636			kfree(mmap_vma);
1637			mm = NULL;
1638		}
1639		if (!mm)
1640			return 1;
1641		mutex_unlock(&vdev->vma_lock);
1642
1643		if (try) {
1644			if (!mmap_read_trylock(mm)) {
1645				mmput(mm);
1646				return 0;
1647			}
1648		} else {
1649			mmap_read_lock(mm);
1650		}
1651		if (try) {
1652			if (!mutex_trylock(&vdev->vma_lock)) {
1653				mmap_read_unlock(mm);
1654				mmput(mm);
1655				return 0;
1656			}
1657		} else {
1658			mutex_lock(&vdev->vma_lock);
1659		}
1660		list_for_each_entry_safe(mmap_vma, tmp,
1661					 &vdev->vma_list, vma_next) {
1662			struct vm_area_struct *vma = mmap_vma->vma;
1663
1664			if (vma->vm_mm != mm)
1665				continue;
1666
1667			list_del(&mmap_vma->vma_next);
1668			kfree(mmap_vma);
1669
1670			zap_vma_ptes(vma, vma->vm_start,
1671				     vma->vm_end - vma->vm_start);
1672		}
1673		mutex_unlock(&vdev->vma_lock);
1674		mmap_read_unlock(mm);
1675		mmput(mm);
1676	}
1677}
1678
1679void vfio_pci_zap_and_down_write_memory_lock(struct vfio_pci_core_device *vdev)
1680{
1681	vfio_pci_zap_and_vma_lock(vdev, false);
1682	down_write(&vdev->memory_lock);
1683	mutex_unlock(&vdev->vma_lock);
1684}
1685
1686u16 vfio_pci_memory_lock_and_enable(struct vfio_pci_core_device *vdev)
1687{
1688	u16 cmd;
1689
1690	down_write(&vdev->memory_lock);
1691	pci_read_config_word(vdev->pdev, PCI_COMMAND, &cmd);
1692	if (!(cmd & PCI_COMMAND_MEMORY))
1693		pci_write_config_word(vdev->pdev, PCI_COMMAND,
1694				      cmd | PCI_COMMAND_MEMORY);
1695
1696	return cmd;
1697}
1698
1699void vfio_pci_memory_unlock_and_restore(struct vfio_pci_core_device *vdev, u16 cmd)
1700{
1701	pci_write_config_word(vdev->pdev, PCI_COMMAND, cmd);
1702	up_write(&vdev->memory_lock);
1703}
1704
1705/* Caller holds vma_lock */
1706static int __vfio_pci_add_vma(struct vfio_pci_core_device *vdev,
1707			      struct vm_area_struct *vma)
1708{
1709	struct vfio_pci_mmap_vma *mmap_vma;
1710
1711	mmap_vma = kmalloc(sizeof(*mmap_vma), GFP_KERNEL_ACCOUNT);
1712	if (!mmap_vma)
1713		return -ENOMEM;
1714
1715	mmap_vma->vma = vma;
1716	list_add(&mmap_vma->vma_next, &vdev->vma_list);
1717
1718	return 0;
1719}
1720
1721/*
1722 * Zap mmaps on open so that we can fault them in on access and therefore
1723 * our vma_list only tracks mappings accessed since last zap.
1724 */
1725static void vfio_pci_mmap_open(struct vm_area_struct *vma)
1726{
1727	zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
1728}
1729
1730static void vfio_pci_mmap_close(struct vm_area_struct *vma)
1731{
1732	struct vfio_pci_core_device *vdev = vma->vm_private_data;
1733	struct vfio_pci_mmap_vma *mmap_vma;
1734
1735	mutex_lock(&vdev->vma_lock);
1736	list_for_each_entry(mmap_vma, &vdev->vma_list, vma_next) {
1737		if (mmap_vma->vma == vma) {
1738			list_del(&mmap_vma->vma_next);
1739			kfree(mmap_vma);
1740			break;
1741		}
1742	}
1743	mutex_unlock(&vdev->vma_lock);
1744}
1745
1746static vm_fault_t vfio_pci_mmap_fault(struct vm_fault *vmf)
1747{
1748	struct vm_area_struct *vma = vmf->vma;
1749	struct vfio_pci_core_device *vdev = vma->vm_private_data;
1750	struct vfio_pci_mmap_vma *mmap_vma;
1751	vm_fault_t ret = VM_FAULT_NOPAGE;
1752
1753	mutex_lock(&vdev->vma_lock);
1754	down_read(&vdev->memory_lock);
1755
1756	/*
1757	 * Memory region cannot be accessed if the low power feature is engaged
1758	 * or memory access is disabled.
1759	 */
1760	if (vdev->pm_runtime_engaged || !__vfio_pci_memory_enabled(vdev)) {
1761		ret = VM_FAULT_SIGBUS;
1762		goto up_out;
1763	}
1764
1765	/*
1766	 * We populate the whole vma on fault, so we need to test whether
1767	 * the vma has already been mapped, such as for concurrent faults
1768	 * to the same vma.  io_remap_pfn_range() will trigger a BUG_ON if
1769	 * we ask it to fill the same range again.
1770	 */
1771	list_for_each_entry(mmap_vma, &vdev->vma_list, vma_next) {
1772		if (mmap_vma->vma == vma)
1773			goto up_out;
1774	}
1775
1776	if (io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
1777			       vma->vm_end - vma->vm_start,
1778			       vma->vm_page_prot)) {
1779		ret = VM_FAULT_SIGBUS;
1780		zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
1781		goto up_out;
1782	}
1783
1784	if (__vfio_pci_add_vma(vdev, vma)) {
1785		ret = VM_FAULT_OOM;
1786		zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
1787	}
1788
1789up_out:
1790	up_read(&vdev->memory_lock);
1791	mutex_unlock(&vdev->vma_lock);
1792	return ret;
1793}
1794
1795static const struct vm_operations_struct vfio_pci_mmap_ops = {
1796	.open = vfio_pci_mmap_open,
1797	.close = vfio_pci_mmap_close,
1798	.fault = vfio_pci_mmap_fault,
1799};
1800
1801int vfio_pci_core_mmap(struct vfio_device *core_vdev, struct vm_area_struct *vma)
1802{
1803	struct vfio_pci_core_device *vdev =
1804		container_of(core_vdev, struct vfio_pci_core_device, vdev);
1805	struct pci_dev *pdev = vdev->pdev;
1806	unsigned int index;
1807	u64 phys_len, req_len, pgoff, req_start;
1808	int ret;
1809
1810	index = vma->vm_pgoff >> (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT);
1811
1812	if (index >= VFIO_PCI_NUM_REGIONS + vdev->num_regions)
1813		return -EINVAL;
1814	if (vma->vm_end < vma->vm_start)
1815		return -EINVAL;
1816	if ((vma->vm_flags & VM_SHARED) == 0)
1817		return -EINVAL;
1818	if (index >= VFIO_PCI_NUM_REGIONS) {
1819		int regnum = index - VFIO_PCI_NUM_REGIONS;
1820		struct vfio_pci_region *region = vdev->region + regnum;
1821
1822		if (region->ops && region->ops->mmap &&
1823		    (region->flags & VFIO_REGION_INFO_FLAG_MMAP))
1824			return region->ops->mmap(vdev, region, vma);
1825		return -EINVAL;
1826	}
1827	if (index >= VFIO_PCI_ROM_REGION_INDEX)
1828		return -EINVAL;
1829	if (!vdev->bar_mmap_supported[index])
1830		return -EINVAL;
1831
1832	phys_len = PAGE_ALIGN(pci_resource_len(pdev, index));
1833	req_len = vma->vm_end - vma->vm_start;
1834	pgoff = vma->vm_pgoff &
1835		((1U << (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT)) - 1);
1836	req_start = pgoff << PAGE_SHIFT;
1837
1838	if (req_start + req_len > phys_len)
1839		return -EINVAL;
1840
1841	/*
1842	 * Even though we don't make use of the barmap for the mmap,
1843	 * we need to request the region and the barmap tracks that.
1844	 */
1845	if (!vdev->barmap[index]) {
1846		ret = pci_request_selected_regions(pdev,
1847						   1 << index, "vfio-pci");
1848		if (ret)
1849			return ret;
1850
1851		vdev->barmap[index] = pci_iomap(pdev, index, 0);
1852		if (!vdev->barmap[index]) {
1853			pci_release_selected_regions(pdev, 1 << index);
1854			return -ENOMEM;
1855		}
1856	}
1857
1858	vma->vm_private_data = vdev;
1859	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1860	vma->vm_pgoff = (pci_resource_start(pdev, index) >> PAGE_SHIFT) + pgoff;
1861
1862	/*
1863	 * See remap_pfn_range(), called from vfio_pci_fault() but we can't
1864	 * change vm_flags within the fault handler.  Set them now.
1865	 */
1866	vm_flags_set(vma, VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP);
1867	vma->vm_ops = &vfio_pci_mmap_ops;
1868
1869	return 0;
1870}
1871EXPORT_SYMBOL_GPL(vfio_pci_core_mmap);
1872
1873void vfio_pci_core_request(struct vfio_device *core_vdev, unsigned int count)
1874{
1875	struct vfio_pci_core_device *vdev =
1876		container_of(core_vdev, struct vfio_pci_core_device, vdev);
1877	struct pci_dev *pdev = vdev->pdev;
1878
1879	mutex_lock(&vdev->igate);
1880
1881	if (vdev->req_trigger) {
1882		if (!(count % 10))
1883			pci_notice_ratelimited(pdev,
1884				"Relaying device request to user (#%u)\n",
1885				count);
1886		eventfd_signal(vdev->req_trigger);
1887	} else if (count == 0) {
1888		pci_warn(pdev,
1889			"No device request channel registered, blocked until released by user\n");
1890	}
1891
1892	mutex_unlock(&vdev->igate);
1893}
1894EXPORT_SYMBOL_GPL(vfio_pci_core_request);
1895
1896static int vfio_pci_validate_vf_token(struct vfio_pci_core_device *vdev,
1897				      bool vf_token, uuid_t *uuid)
1898{
1899	/*
1900	 * There's always some degree of trust or collaboration between SR-IOV
1901	 * PF and VFs, even if just that the PF hosts the SR-IOV capability and
1902	 * can disrupt VFs with a reset, but often the PF has more explicit
1903	 * access to deny service to the VF or access data passed through the
1904	 * VF.  We therefore require an opt-in via a shared VF token (UUID) to
1905	 * represent this trust.  This both prevents that a VF driver might
1906	 * assume the PF driver is a trusted, in-kernel driver, and also that
1907	 * a PF driver might be replaced with a rogue driver, unknown to in-use
1908	 * VF drivers.
1909	 *
1910	 * Therefore when presented with a VF, if the PF is a vfio device and
1911	 * it is bound to the vfio-pci driver, the user needs to provide a VF
1912	 * token to access the device, in the form of appending a vf_token to
1913	 * the device name, for example:
1914	 *
1915	 * "0000:04:10.0 vf_token=bd8d9d2b-5a5f-4f5a-a211-f591514ba1f3"
1916	 *
1917	 * When presented with a PF which has VFs in use, the user must also
1918	 * provide the current VF token to prove collaboration with existing
1919	 * VF users.  If VFs are not in use, the VF token provided for the PF
1920	 * device will act to set the VF token.
1921	 *
1922	 * If the VF token is provided but unused, an error is generated.
1923	 */
1924	if (vdev->pdev->is_virtfn) {
1925		struct vfio_pci_core_device *pf_vdev = vdev->sriov_pf_core_dev;
1926		bool match;
1927
1928		if (!pf_vdev) {
1929			if (!vf_token)
1930				return 0; /* PF is not vfio-pci, no VF token */
1931
1932			pci_info_ratelimited(vdev->pdev,
1933				"VF token incorrectly provided, PF not bound to vfio-pci\n");
1934			return -EINVAL;
1935		}
1936
1937		if (!vf_token) {
1938			pci_info_ratelimited(vdev->pdev,
1939				"VF token required to access device\n");
1940			return -EACCES;
1941		}
1942
1943		mutex_lock(&pf_vdev->vf_token->lock);
1944		match = uuid_equal(uuid, &pf_vdev->vf_token->uuid);
1945		mutex_unlock(&pf_vdev->vf_token->lock);
1946
1947		if (!match) {
1948			pci_info_ratelimited(vdev->pdev,
1949				"Incorrect VF token provided for device\n");
1950			return -EACCES;
1951		}
1952	} else if (vdev->vf_token) {
1953		mutex_lock(&vdev->vf_token->lock);
1954		if (vdev->vf_token->users) {
1955			if (!vf_token) {
1956				mutex_unlock(&vdev->vf_token->lock);
1957				pci_info_ratelimited(vdev->pdev,
1958					"VF token required to access device\n");
1959				return -EACCES;
1960			}
1961
1962			if (!uuid_equal(uuid, &vdev->vf_token->uuid)) {
1963				mutex_unlock(&vdev->vf_token->lock);
1964				pci_info_ratelimited(vdev->pdev,
1965					"Incorrect VF token provided for device\n");
1966				return -EACCES;
1967			}
1968		} else if (vf_token) {
1969			uuid_copy(&vdev->vf_token->uuid, uuid);
1970		}
1971
1972		mutex_unlock(&vdev->vf_token->lock);
1973	} else if (vf_token) {
1974		pci_info_ratelimited(vdev->pdev,
1975			"VF token incorrectly provided, not a PF or VF\n");
1976		return -EINVAL;
1977	}
1978
1979	return 0;
1980}
1981
1982#define VF_TOKEN_ARG "vf_token="
1983
1984int vfio_pci_core_match(struct vfio_device *core_vdev, char *buf)
1985{
1986	struct vfio_pci_core_device *vdev =
1987		container_of(core_vdev, struct vfio_pci_core_device, vdev);
1988	bool vf_token = false;
1989	uuid_t uuid;
1990	int ret;
1991
1992	if (strncmp(pci_name(vdev->pdev), buf, strlen(pci_name(vdev->pdev))))
1993		return 0; /* No match */
1994
1995	if (strlen(buf) > strlen(pci_name(vdev->pdev))) {
1996		buf += strlen(pci_name(vdev->pdev));
1997
1998		if (*buf != ' ')
1999			return 0; /* No match: non-whitespace after name */
2000
2001		while (*buf) {
2002			if (*buf == ' ') {
2003				buf++;
2004				continue;
2005			}
2006
2007			if (!vf_token && !strncmp(buf, VF_TOKEN_ARG,
2008						  strlen(VF_TOKEN_ARG))) {
2009				buf += strlen(VF_TOKEN_ARG);
2010
2011				if (strlen(buf) < UUID_STRING_LEN)
2012					return -EINVAL;
2013
2014				ret = uuid_parse(buf, &uuid);
2015				if (ret)
2016					return ret;
2017
2018				vf_token = true;
2019				buf += UUID_STRING_LEN;
2020			} else {
2021				/* Unknown/duplicate option */
2022				return -EINVAL;
2023			}
2024		}
2025	}
2026
2027	ret = vfio_pci_validate_vf_token(vdev, vf_token, &uuid);
2028	if (ret)
2029		return ret;
2030
2031	return 1; /* Match */
2032}
2033EXPORT_SYMBOL_GPL(vfio_pci_core_match);
2034
2035static int vfio_pci_bus_notifier(struct notifier_block *nb,
2036				 unsigned long action, void *data)
2037{
2038	struct vfio_pci_core_device *vdev = container_of(nb,
2039						    struct vfio_pci_core_device, nb);
2040	struct device *dev = data;
2041	struct pci_dev *pdev = to_pci_dev(dev);
2042	struct pci_dev *physfn = pci_physfn(pdev);
2043
2044	if (action == BUS_NOTIFY_ADD_DEVICE &&
2045	    pdev->is_virtfn && physfn == vdev->pdev) {
2046		pci_info(vdev->pdev, "Captured SR-IOV VF %s driver_override\n",
2047			 pci_name(pdev));
2048		pdev->driver_override = kasprintf(GFP_KERNEL, "%s",
2049						  vdev->vdev.ops->name);
2050	} else if (action == BUS_NOTIFY_BOUND_DRIVER &&
2051		   pdev->is_virtfn && physfn == vdev->pdev) {
2052		struct pci_driver *drv = pci_dev_driver(pdev);
2053
2054		if (drv && drv != pci_dev_driver(vdev->pdev))
2055			pci_warn(vdev->pdev,
2056				 "VF %s bound to driver %s while PF bound to driver %s\n",
2057				 pci_name(pdev), drv->name,
2058				 pci_dev_driver(vdev->pdev)->name);
2059	}
2060
2061	return 0;
2062}
2063
2064static int vfio_pci_vf_init(struct vfio_pci_core_device *vdev)
2065{
2066	struct pci_dev *pdev = vdev->pdev;
2067	struct vfio_pci_core_device *cur;
2068	struct pci_dev *physfn;
2069	int ret;
2070
2071	if (pdev->is_virtfn) {
2072		/*
2073		 * If this VF was created by our vfio_pci_core_sriov_configure()
2074		 * then we can find the PF vfio_pci_core_device now, and due to
2075		 * the locking in pci_disable_sriov() it cannot change until
2076		 * this VF device driver is removed.
2077		 */
2078		physfn = pci_physfn(vdev->pdev);
2079		mutex_lock(&vfio_pci_sriov_pfs_mutex);
2080		list_for_each_entry(cur, &vfio_pci_sriov_pfs, sriov_pfs_item) {
2081			if (cur->pdev == physfn) {
2082				vdev->sriov_pf_core_dev = cur;
2083				break;
2084			}
2085		}
2086		mutex_unlock(&vfio_pci_sriov_pfs_mutex);
2087		return 0;
2088	}
2089
2090	/* Not a SRIOV PF */
2091	if (!pdev->is_physfn)
2092		return 0;
2093
2094	vdev->vf_token = kzalloc(sizeof(*vdev->vf_token), GFP_KERNEL);
2095	if (!vdev->vf_token)
2096		return -ENOMEM;
2097
2098	mutex_init(&vdev->vf_token->lock);
2099	uuid_gen(&vdev->vf_token->uuid);
2100
2101	vdev->nb.notifier_call = vfio_pci_bus_notifier;
2102	ret = bus_register_notifier(&pci_bus_type, &vdev->nb);
2103	if (ret) {
2104		kfree(vdev->vf_token);
2105		return ret;
2106	}
2107	return 0;
2108}
2109
2110static void vfio_pci_vf_uninit(struct vfio_pci_core_device *vdev)
2111{
2112	if (!vdev->vf_token)
2113		return;
2114
2115	bus_unregister_notifier(&pci_bus_type, &vdev->nb);
2116	WARN_ON(vdev->vf_token->users);
2117	mutex_destroy(&vdev->vf_token->lock);
2118	kfree(vdev->vf_token);
2119}
2120
2121static int vfio_pci_vga_init(struct vfio_pci_core_device *vdev)
2122{
2123	struct pci_dev *pdev = vdev->pdev;
2124	int ret;
2125
2126	if (!vfio_pci_is_vga(pdev))
2127		return 0;
2128
2129	ret = aperture_remove_conflicting_pci_devices(pdev, vdev->vdev.ops->name);
2130	if (ret)
2131		return ret;
2132
2133	ret = vga_client_register(pdev, vfio_pci_set_decode);
2134	if (ret)
2135		return ret;
2136	vga_set_legacy_decoding(pdev, vfio_pci_set_decode(pdev, false));
2137	return 0;
2138}
2139
2140static void vfio_pci_vga_uninit(struct vfio_pci_core_device *vdev)
2141{
2142	struct pci_dev *pdev = vdev->pdev;
2143
2144	if (!vfio_pci_is_vga(pdev))
2145		return;
2146	vga_client_unregister(pdev);
2147	vga_set_legacy_decoding(pdev, VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM |
2148					      VGA_RSRC_LEGACY_IO |
2149					      VGA_RSRC_LEGACY_MEM);
2150}
2151
2152int vfio_pci_core_init_dev(struct vfio_device *core_vdev)
2153{
2154	struct vfio_pci_core_device *vdev =
2155		container_of(core_vdev, struct vfio_pci_core_device, vdev);
2156
2157	vdev->pdev = to_pci_dev(core_vdev->dev);
2158	vdev->irq_type = VFIO_PCI_NUM_IRQS;
2159	mutex_init(&vdev->igate);
2160	spin_lock_init(&vdev->irqlock);
2161	mutex_init(&vdev->ioeventfds_lock);
2162	INIT_LIST_HEAD(&vdev->dummy_resources_list);
2163	INIT_LIST_HEAD(&vdev->ioeventfds_list);
2164	mutex_init(&vdev->vma_lock);
2165	INIT_LIST_HEAD(&vdev->vma_list);
2166	INIT_LIST_HEAD(&vdev->sriov_pfs_item);
2167	init_rwsem(&vdev->memory_lock);
2168	xa_init(&vdev->ctx);
2169
2170	return 0;
2171}
2172EXPORT_SYMBOL_GPL(vfio_pci_core_init_dev);
2173
2174void vfio_pci_core_release_dev(struct vfio_device *core_vdev)
2175{
2176	struct vfio_pci_core_device *vdev =
2177		container_of(core_vdev, struct vfio_pci_core_device, vdev);
2178
2179	mutex_destroy(&vdev->igate);
2180	mutex_destroy(&vdev->ioeventfds_lock);
2181	mutex_destroy(&vdev->vma_lock);
2182	kfree(vdev->region);
2183	kfree(vdev->pm_save);
2184}
2185EXPORT_SYMBOL_GPL(vfio_pci_core_release_dev);
2186
2187int vfio_pci_core_register_device(struct vfio_pci_core_device *vdev)
2188{
2189	struct pci_dev *pdev = vdev->pdev;
2190	struct device *dev = &pdev->dev;
2191	int ret;
2192
2193	/* Drivers must set the vfio_pci_core_device to their drvdata */
2194	if (WARN_ON(vdev != dev_get_drvdata(dev)))
2195		return -EINVAL;
2196
2197	if (pdev->hdr_type != PCI_HEADER_TYPE_NORMAL)
2198		return -EINVAL;
2199
2200	if (vdev->vdev.mig_ops) {
2201		if (!(vdev->vdev.mig_ops->migration_get_state &&
2202		      vdev->vdev.mig_ops->migration_set_state &&
2203		      vdev->vdev.mig_ops->migration_get_data_size) ||
2204		    !(vdev->vdev.migration_flags & VFIO_MIGRATION_STOP_COPY))
2205			return -EINVAL;
2206	}
2207
2208	if (vdev->vdev.log_ops && !(vdev->vdev.log_ops->log_start &&
2209	    vdev->vdev.log_ops->log_stop &&
2210	    vdev->vdev.log_ops->log_read_and_clear))
2211		return -EINVAL;
2212
2213	/*
2214	 * Prevent binding to PFs with VFs enabled, the VFs might be in use
2215	 * by the host or other users.  We cannot capture the VFs if they
2216	 * already exist, nor can we track VF users.  Disabling SR-IOV here
2217	 * would initiate removing the VFs, which would unbind the driver,
2218	 * which is prone to blocking if that VF is also in use by vfio-pci.
2219	 * Just reject these PFs and let the user sort it out.
2220	 */
2221	if (pci_num_vf(pdev)) {
2222		pci_warn(pdev, "Cannot bind to PF with SR-IOV enabled\n");
2223		return -EBUSY;
2224	}
2225
2226	if (pci_is_root_bus(pdev->bus)) {
2227		ret = vfio_assign_device_set(&vdev->vdev, vdev);
2228	} else if (!pci_probe_reset_slot(pdev->slot)) {
2229		ret = vfio_assign_device_set(&vdev->vdev, pdev->slot);
2230	} else {
2231		/*
2232		 * If there is no slot reset support for this device, the whole
2233		 * bus needs to be grouped together to support bus-wide resets.
2234		 */
2235		ret = vfio_assign_device_set(&vdev->vdev, pdev->bus);
2236	}
2237
2238	if (ret)
2239		return ret;
2240	ret = vfio_pci_vf_init(vdev);
2241	if (ret)
2242		return ret;
2243	ret = vfio_pci_vga_init(vdev);
2244	if (ret)
2245		goto out_vf;
2246
2247	vfio_pci_probe_power_state(vdev);
2248
2249	/*
2250	 * pci-core sets the device power state to an unknown value at
2251	 * bootup and after being removed from a driver.  The only
2252	 * transition it allows from this unknown state is to D0, which
2253	 * typically happens when a driver calls pci_enable_device().
2254	 * We're not ready to enable the device yet, but we do want to
2255	 * be able to get to D3.  Therefore first do a D0 transition
2256	 * before enabling runtime PM.
2257	 */
2258	vfio_pci_set_power_state(vdev, PCI_D0);
2259
2260	dev->driver->pm = &vfio_pci_core_pm_ops;
2261	pm_runtime_allow(dev);
2262	if (!disable_idle_d3)
2263		pm_runtime_put(dev);
2264
2265	ret = vfio_register_group_dev(&vdev->vdev);
2266	if (ret)
2267		goto out_power;
2268	return 0;
2269
2270out_power:
2271	if (!disable_idle_d3)
2272		pm_runtime_get_noresume(dev);
2273
2274	pm_runtime_forbid(dev);
2275out_vf:
2276	vfio_pci_vf_uninit(vdev);
2277	return ret;
2278}
2279EXPORT_SYMBOL_GPL(vfio_pci_core_register_device);
2280
2281void vfio_pci_core_unregister_device(struct vfio_pci_core_device *vdev)
2282{
2283	vfio_pci_core_sriov_configure(vdev, 0);
2284
2285	vfio_unregister_group_dev(&vdev->vdev);
2286
2287	vfio_pci_vf_uninit(vdev);
2288	vfio_pci_vga_uninit(vdev);
2289
2290	if (!disable_idle_d3)
2291		pm_runtime_get_noresume(&vdev->pdev->dev);
2292
2293	pm_runtime_forbid(&vdev->pdev->dev);
2294}
2295EXPORT_SYMBOL_GPL(vfio_pci_core_unregister_device);
2296
2297pci_ers_result_t vfio_pci_core_aer_err_detected(struct pci_dev *pdev,
2298						pci_channel_state_t state)
2299{
2300	struct vfio_pci_core_device *vdev = dev_get_drvdata(&pdev->dev);
2301
2302	mutex_lock(&vdev->igate);
2303
2304	if (vdev->err_trigger)
2305		eventfd_signal(vdev->err_trigger);
2306
2307	mutex_unlock(&vdev->igate);
2308
2309	return PCI_ERS_RESULT_CAN_RECOVER;
2310}
2311EXPORT_SYMBOL_GPL(vfio_pci_core_aer_err_detected);
2312
2313int vfio_pci_core_sriov_configure(struct vfio_pci_core_device *vdev,
2314				  int nr_virtfn)
2315{
2316	struct pci_dev *pdev = vdev->pdev;
2317	int ret = 0;
2318
2319	device_lock_assert(&pdev->dev);
2320
2321	if (nr_virtfn) {
2322		mutex_lock(&vfio_pci_sriov_pfs_mutex);
2323		/*
2324		 * The thread that adds the vdev to the list is the only thread
2325		 * that gets to call pci_enable_sriov() and we will only allow
2326		 * it to be called once without going through
2327		 * pci_disable_sriov()
2328		 */
2329		if (!list_empty(&vdev->sriov_pfs_item)) {
2330			ret = -EINVAL;
2331			goto out_unlock;
2332		}
2333		list_add_tail(&vdev->sriov_pfs_item, &vfio_pci_sriov_pfs);
2334		mutex_unlock(&vfio_pci_sriov_pfs_mutex);
2335
2336		/*
2337		 * The PF power state should always be higher than the VF power
2338		 * state. The PF can be in low power state either with runtime
2339		 * power management (when there is no user) or PCI_PM_CTRL
2340		 * register write by the user. If PF is in the low power state,
2341		 * then change the power state to D0 first before enabling
2342		 * SR-IOV. Also, this function can be called at any time, and
2343		 * userspace PCI_PM_CTRL write can race against this code path,
2344		 * so protect the same with 'memory_lock'.
2345		 */
2346		ret = pm_runtime_resume_and_get(&pdev->dev);
2347		if (ret)
2348			goto out_del;
2349
2350		down_write(&vdev->memory_lock);
2351		vfio_pci_set_power_state(vdev, PCI_D0);
2352		ret = pci_enable_sriov(pdev, nr_virtfn);
2353		up_write(&vdev->memory_lock);
2354		if (ret) {
2355			pm_runtime_put(&pdev->dev);
2356			goto out_del;
2357		}
2358		return nr_virtfn;
2359	}
2360
2361	if (pci_num_vf(pdev)) {
2362		pci_disable_sriov(pdev);
2363		pm_runtime_put(&pdev->dev);
2364	}
2365
2366out_del:
2367	mutex_lock(&vfio_pci_sriov_pfs_mutex);
2368	list_del_init(&vdev->sriov_pfs_item);
2369out_unlock:
2370	mutex_unlock(&vfio_pci_sriov_pfs_mutex);
2371	return ret;
2372}
2373EXPORT_SYMBOL_GPL(vfio_pci_core_sriov_configure);
2374
2375const struct pci_error_handlers vfio_pci_core_err_handlers = {
2376	.error_detected = vfio_pci_core_aer_err_detected,
2377};
2378EXPORT_SYMBOL_GPL(vfio_pci_core_err_handlers);
2379
2380static bool vfio_dev_in_groups(struct vfio_device *vdev,
2381			       struct vfio_pci_group_info *groups)
2382{
2383	unsigned int i;
2384
2385	if (!groups)
2386		return false;
2387
2388	for (i = 0; i < groups->count; i++)
2389		if (vfio_file_has_dev(groups->files[i], vdev))
2390			return true;
2391	return false;
2392}
2393
2394static int vfio_pci_is_device_in_set(struct pci_dev *pdev, void *data)
2395{
2396	struct vfio_device_set *dev_set = data;
 
2397
2398	return vfio_find_device_in_devset(dev_set, &pdev->dev) ? 0 : -ENODEV;
 
 
 
2399}
2400
2401/*
2402 * vfio-core considers a group to be viable and will create a vfio_device even
2403 * if some devices are bound to drivers like pci-stub or pcieport. Here we
2404 * require all PCI devices to be inside our dev_set since that ensures they stay
2405 * put and that every driver controlling the device can co-ordinate with the
2406 * device reset.
2407 *
2408 * Returns the pci_dev to pass to pci_reset_bus() if every PCI device to be
2409 * reset is inside the dev_set, and pci_reset_bus() can succeed. NULL otherwise.
2410 */
2411static struct pci_dev *
2412vfio_pci_dev_set_resettable(struct vfio_device_set *dev_set)
2413{
2414	struct pci_dev *pdev;
2415
2416	lockdep_assert_held(&dev_set->lock);
2417
2418	/*
2419	 * By definition all PCI devices in the dev_set share the same PCI
2420	 * reset, so any pci_dev will have the same outcomes for
2421	 * pci_probe_reset_*() and pci_reset_bus().
2422	 */
2423	pdev = list_first_entry(&dev_set->device_list,
2424				struct vfio_pci_core_device,
2425				vdev.dev_set_list)->pdev;
2426
2427	/* pci_reset_bus() is supported */
2428	if (pci_probe_reset_slot(pdev->slot) && pci_probe_reset_bus(pdev->bus))
2429		return NULL;
2430
2431	if (vfio_pci_for_each_slot_or_bus(pdev, vfio_pci_is_device_in_set,
2432					  dev_set,
2433					  !pci_probe_reset_slot(pdev->slot)))
2434		return NULL;
2435	return pdev;
2436}
2437
2438static int vfio_pci_dev_set_pm_runtime_get(struct vfio_device_set *dev_set)
2439{
2440	struct vfio_pci_core_device *cur;
2441	int ret;
2442
2443	list_for_each_entry(cur, &dev_set->device_list, vdev.dev_set_list) {
2444		ret = pm_runtime_resume_and_get(&cur->pdev->dev);
2445		if (ret)
2446			goto unwind;
2447	}
2448
2449	return 0;
2450
2451unwind:
2452	list_for_each_entry_continue_reverse(cur, &dev_set->device_list,
2453					     vdev.dev_set_list)
2454		pm_runtime_put(&cur->pdev->dev);
2455
2456	return ret;
2457}
2458
2459/*
2460 * We need to get memory_lock for each device, but devices can share mmap_lock,
2461 * therefore we need to zap and hold the vma_lock for each device, and only then
2462 * get each memory_lock.
2463 */
2464static int vfio_pci_dev_set_hot_reset(struct vfio_device_set *dev_set,
2465				      struct vfio_pci_group_info *groups,
2466				      struct iommufd_ctx *iommufd_ctx)
2467{
2468	struct vfio_pci_core_device *cur_mem;
2469	struct vfio_pci_core_device *cur_vma;
2470	struct vfio_pci_core_device *cur;
2471	struct pci_dev *pdev;
2472	bool is_mem = true;
2473	int ret;
2474
2475	mutex_lock(&dev_set->lock);
2476	cur_mem = list_first_entry(&dev_set->device_list,
2477				   struct vfio_pci_core_device,
2478				   vdev.dev_set_list);
2479
2480	pdev = vfio_pci_dev_set_resettable(dev_set);
2481	if (!pdev) {
2482		ret = -EINVAL;
2483		goto err_unlock;
2484	}
2485
2486	/*
2487	 * Some of the devices in the dev_set can be in the runtime suspended
2488	 * state. Increment the usage count for all the devices in the dev_set
2489	 * before reset and decrement the same after reset.
2490	 */
2491	ret = vfio_pci_dev_set_pm_runtime_get(dev_set);
2492	if (ret)
2493		goto err_unlock;
2494
2495	list_for_each_entry(cur_vma, &dev_set->device_list, vdev.dev_set_list) {
2496		bool owned;
2497
2498		/*
2499		 * Test whether all the affected devices can be reset by the
2500		 * user.
2501		 *
2502		 * If called from a group opened device and the user provides
2503		 * a set of groups, all the devices in the dev_set should be
2504		 * contained by the set of groups provided by the user.
2505		 *
2506		 * If called from a cdev opened device and the user provides
2507		 * a zero-length array, all the devices in the dev_set must
2508		 * be bound to the same iommufd_ctx as the input iommufd_ctx.
2509		 * If there is any device that has not been bound to any
2510		 * iommufd_ctx yet, check if its iommu_group has any device
2511		 * bound to the input iommufd_ctx.  Such devices can be
2512		 * considered owned by the input iommufd_ctx as the device
2513		 * cannot be owned by another iommufd_ctx when its iommu_group
2514		 * is owned.
2515		 *
2516		 * Otherwise, reset is not allowed.
2517		 */
2518		if (iommufd_ctx) {
2519			int devid = vfio_iommufd_get_dev_id(&cur_vma->vdev,
2520							    iommufd_ctx);
2521
2522			owned = (devid > 0 || devid == -ENOENT);
2523		} else {
2524			owned = vfio_dev_in_groups(&cur_vma->vdev, groups);
2525		}
2526
2527		if (!owned) {
2528			ret = -EINVAL;
2529			goto err_undo;
2530		}
2531
2532		/*
2533		 * Locking multiple devices is prone to deadlock, runaway and
2534		 * unwind if we hit contention.
2535		 */
2536		if (!vfio_pci_zap_and_vma_lock(cur_vma, true)) {
2537			ret = -EBUSY;
2538			goto err_undo;
2539		}
2540	}
2541	cur_vma = NULL;
2542
2543	list_for_each_entry(cur_mem, &dev_set->device_list, vdev.dev_set_list) {
2544		if (!down_write_trylock(&cur_mem->memory_lock)) {
2545			ret = -EBUSY;
2546			goto err_undo;
2547		}
2548		mutex_unlock(&cur_mem->vma_lock);
2549	}
2550	cur_mem = NULL;
2551
2552	/*
2553	 * The pci_reset_bus() will reset all the devices in the bus.
2554	 * The power state can be non-D0 for some of the devices in the bus.
2555	 * For these devices, the pci_reset_bus() will internally set
2556	 * the power state to D0 without vfio driver involvement.
2557	 * For the devices which have NoSoftRst-, the reset function can
2558	 * cause the PCI config space reset without restoring the original
2559	 * state (saved locally in 'vdev->pm_save').
2560	 */
2561	list_for_each_entry(cur, &dev_set->device_list, vdev.dev_set_list)
2562		vfio_pci_set_power_state(cur, PCI_D0);
2563
2564	ret = pci_reset_bus(pdev);
2565
2566err_undo:
2567	list_for_each_entry(cur, &dev_set->device_list, vdev.dev_set_list) {
2568		if (cur == cur_mem)
2569			is_mem = false;
2570		if (cur == cur_vma)
2571			break;
2572		if (is_mem)
2573			up_write(&cur->memory_lock);
2574		else
2575			mutex_unlock(&cur->vma_lock);
2576	}
2577
2578	list_for_each_entry(cur, &dev_set->device_list, vdev.dev_set_list)
2579		pm_runtime_put(&cur->pdev->dev);
2580err_unlock:
2581	mutex_unlock(&dev_set->lock);
2582	return ret;
2583}
2584
2585static bool vfio_pci_dev_set_needs_reset(struct vfio_device_set *dev_set)
2586{
2587	struct vfio_pci_core_device *cur;
2588	bool needs_reset = false;
2589
2590	/* No other VFIO device in the set can be open. */
2591	if (vfio_device_set_open_count(dev_set) > 1)
2592		return false;
2593
2594	list_for_each_entry(cur, &dev_set->device_list, vdev.dev_set_list)
2595		needs_reset |= cur->needs_reset;
2596	return needs_reset;
2597}
2598
2599/*
2600 * If a bus or slot reset is available for the provided dev_set and:
2601 *  - All of the devices affected by that bus or slot reset are unused
2602 *  - At least one of the affected devices is marked dirty via
2603 *    needs_reset (such as by lack of FLR support)
2604 * Then attempt to perform that bus or slot reset.
2605 */
2606static void vfio_pci_dev_set_try_reset(struct vfio_device_set *dev_set)
2607{
2608	struct vfio_pci_core_device *cur;
2609	struct pci_dev *pdev;
2610	bool reset_done = false;
2611
2612	if (!vfio_pci_dev_set_needs_reset(dev_set))
2613		return;
2614
2615	pdev = vfio_pci_dev_set_resettable(dev_set);
2616	if (!pdev)
2617		return;
2618
2619	/*
2620	 * Some of the devices in the bus can be in the runtime suspended
2621	 * state. Increment the usage count for all the devices in the dev_set
2622	 * before reset and decrement the same after reset.
2623	 */
2624	if (!disable_idle_d3 && vfio_pci_dev_set_pm_runtime_get(dev_set))
2625		return;
2626
2627	if (!pci_reset_bus(pdev))
2628		reset_done = true;
2629
2630	list_for_each_entry(cur, &dev_set->device_list, vdev.dev_set_list) {
2631		if (reset_done)
2632			cur->needs_reset = false;
2633
2634		if (!disable_idle_d3)
2635			pm_runtime_put(&cur->pdev->dev);
2636	}
2637}
2638
2639void vfio_pci_core_set_params(bool is_nointxmask, bool is_disable_vga,
2640			      bool is_disable_idle_d3)
2641{
2642	nointxmask = is_nointxmask;
2643	disable_vga = is_disable_vga;
2644	disable_idle_d3 = is_disable_idle_d3;
2645}
2646EXPORT_SYMBOL_GPL(vfio_pci_core_set_params);
2647
2648static void vfio_pci_core_cleanup(void)
2649{
2650	vfio_pci_uninit_perm_bits();
2651}
2652
2653static int __init vfio_pci_core_init(void)
2654{
2655	/* Allocate shared config space permission data used by all devices */
2656	return vfio_pci_init_perm_bits();
2657}
2658
2659module_init(vfio_pci_core_init);
2660module_exit(vfio_pci_core_cleanup);
2661
2662MODULE_LICENSE("GPL v2");
2663MODULE_AUTHOR(DRIVER_AUTHOR);
2664MODULE_DESCRIPTION(DRIVER_DESC);
v6.2
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * Copyright (C) 2012 Red Hat, Inc.  All rights reserved.
   4 *     Author: Alex Williamson <alex.williamson@redhat.com>
   5 *
   6 * Derived from original vfio:
   7 * Copyright 2010 Cisco Systems, Inc.  All rights reserved.
   8 * Author: Tom Lyon, pugs@cisco.com
   9 */
  10
  11#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  12
  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#if IS_ENABLED(CONFIG_EEH)
  31#include <asm/eeh.h>
  32#endif
  33
  34#include "vfio_pci_priv.h"
  35
  36#define DRIVER_AUTHOR   "Alex Williamson <alex.williamson@redhat.com>"
  37#define DRIVER_DESC "core driver for VFIO based PCI devices"
  38
  39static bool nointxmask;
  40static bool disable_vga;
  41static bool disable_idle_d3;
  42
  43/* List of PF's that vfio_pci_core_sriov_configure() has been called on */
  44static DEFINE_MUTEX(vfio_pci_sriov_pfs_mutex);
  45static LIST_HEAD(vfio_pci_sriov_pfs);
  46
  47struct vfio_pci_dummy_resource {
  48	struct resource		resource;
  49	int			index;
  50	struct list_head	res_next;
  51};
  52
  53struct vfio_pci_vf_token {
  54	struct mutex		lock;
  55	uuid_t			uuid;
  56	int			users;
  57};
  58
  59struct vfio_pci_mmap_vma {
  60	struct vm_area_struct	*vma;
  61	struct list_head	vma_next;
  62};
  63
  64static inline bool vfio_vga_disabled(void)
  65{
  66#ifdef CONFIG_VFIO_PCI_VGA
  67	return disable_vga;
  68#else
  69	return true;
  70#endif
  71}
  72
  73/*
  74 * Our VGA arbiter participation is limited since we don't know anything
  75 * about the device itself.  However, if the device is the only VGA device
  76 * downstream of a bridge and VFIO VGA support is disabled, then we can
  77 * safely return legacy VGA IO and memory as not decoded since the user
  78 * has no way to get to it and routing can be disabled externally at the
  79 * bridge.
  80 */
  81static unsigned int vfio_pci_set_decode(struct pci_dev *pdev, bool single_vga)
  82{
  83	struct pci_dev *tmp = NULL;
  84	unsigned char max_busnr;
  85	unsigned int decodes;
  86
  87	if (single_vga || !vfio_vga_disabled() || pci_is_root_bus(pdev->bus))
  88		return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM |
  89		       VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM;
  90
  91	max_busnr = pci_bus_max_busnr(pdev->bus);
  92	decodes = VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
  93
  94	while ((tmp = pci_get_class(PCI_CLASS_DISPLAY_VGA << 8, tmp)) != NULL) {
  95		if (tmp == pdev ||
  96		    pci_domain_nr(tmp->bus) != pci_domain_nr(pdev->bus) ||
  97		    pci_is_root_bus(tmp->bus))
  98			continue;
  99
 100		if (tmp->bus->number >= pdev->bus->number &&
 101		    tmp->bus->number <= max_busnr) {
 102			pci_dev_put(tmp);
 103			decodes |= VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM;
 104			break;
 105		}
 106	}
 107
 108	return decodes;
 109}
 110
 111static void vfio_pci_probe_mmaps(struct vfio_pci_core_device *vdev)
 112{
 113	struct resource *res;
 114	int i;
 115	struct vfio_pci_dummy_resource *dummy_res;
 116
 117	for (i = 0; i < PCI_STD_NUM_BARS; i++) {
 118		int bar = i + PCI_STD_RESOURCES;
 119
 120		res = &vdev->pdev->resource[bar];
 121
 122		if (!IS_ENABLED(CONFIG_VFIO_PCI_MMAP))
 123			goto no_mmap;
 124
 125		if (!(res->flags & IORESOURCE_MEM))
 126			goto no_mmap;
 127
 128		/*
 129		 * The PCI core shouldn't set up a resource with a
 130		 * type but zero size. But there may be bugs that
 131		 * cause us to do that.
 132		 */
 133		if (!resource_size(res))
 134			goto no_mmap;
 135
 136		if (resource_size(res) >= PAGE_SIZE) {
 137			vdev->bar_mmap_supported[bar] = true;
 138			continue;
 139		}
 140
 141		if (!(res->start & ~PAGE_MASK)) {
 142			/*
 143			 * Add a dummy resource to reserve the remainder
 144			 * of the exclusive page in case that hot-add
 145			 * device's bar is assigned into it.
 146			 */
 147			dummy_res = kzalloc(sizeof(*dummy_res), GFP_KERNEL);
 
 148			if (dummy_res == NULL)
 149				goto no_mmap;
 150
 151			dummy_res->resource.name = "vfio sub-page reserved";
 152			dummy_res->resource.start = res->end + 1;
 153			dummy_res->resource.end = res->start + PAGE_SIZE - 1;
 154			dummy_res->resource.flags = res->flags;
 155			if (request_resource(res->parent,
 156						&dummy_res->resource)) {
 157				kfree(dummy_res);
 158				goto no_mmap;
 159			}
 160			dummy_res->index = bar;
 161			list_add(&dummy_res->res_next,
 162					&vdev->dummy_resources_list);
 163			vdev->bar_mmap_supported[bar] = true;
 164			continue;
 165		}
 166		/*
 167		 * Here we don't handle the case when the BAR is not page
 168		 * aligned because we can't expect the BAR will be
 169		 * assigned into the same location in a page in guest
 170		 * when we passthrough the BAR. And it's hard to access
 171		 * this BAR in userspace because we have no way to get
 172		 * the BAR's location in a page.
 173		 */
 174no_mmap:
 175		vdev->bar_mmap_supported[bar] = false;
 176	}
 177}
 178
 179struct vfio_pci_group_info;
 180static void vfio_pci_dev_set_try_reset(struct vfio_device_set *dev_set);
 181static int vfio_pci_dev_set_hot_reset(struct vfio_device_set *dev_set,
 182				      struct vfio_pci_group_info *groups);
 
 183
 184/*
 185 * INTx masking requires the ability to disable INTx signaling via PCI_COMMAND
 186 * _and_ the ability detect when the device is asserting INTx via PCI_STATUS.
 187 * If a device implements the former but not the latter we would typically
 188 * expect broken_intx_masking be set and require an exclusive interrupt.
 189 * However since we do have control of the device's ability to assert INTx,
 190 * we can instead pretend that the device does not implement INTx, virtualizing
 191 * the pin register to report zero and maintaining DisINTx set on the host.
 192 */
 193static bool vfio_pci_nointx(struct pci_dev *pdev)
 194{
 195	switch (pdev->vendor) {
 196	case PCI_VENDOR_ID_INTEL:
 197		switch (pdev->device) {
 198		/* All i40e (XL710/X710/XXV710) 10/20/25/40GbE NICs */
 199		case 0x1572:
 200		case 0x1574:
 201		case 0x1580 ... 0x1581:
 202		case 0x1583 ... 0x158b:
 203		case 0x37d0 ... 0x37d2:
 204		/* X550 */
 205		case 0x1563:
 206			return true;
 207		default:
 208			return false;
 209		}
 210	}
 211
 212	return false;
 213}
 214
 215static void vfio_pci_probe_power_state(struct vfio_pci_core_device *vdev)
 216{
 217	struct pci_dev *pdev = vdev->pdev;
 218	u16 pmcsr;
 219
 220	if (!pdev->pm_cap)
 221		return;
 222
 223	pci_read_config_word(pdev, pdev->pm_cap + PCI_PM_CTRL, &pmcsr);
 224
 225	vdev->needs_pm_restore = !(pmcsr & PCI_PM_CTRL_NO_SOFT_RESET);
 226}
 227
 228/*
 229 * pci_set_power_state() wrapper handling devices which perform a soft reset on
 230 * D3->D0 transition.  Save state prior to D0/1/2->D3, stash it on the vdev,
 231 * restore when returned to D0.  Saved separately from pci_saved_state for use
 232 * by PM capability emulation and separately from pci_dev internal saved state
 233 * to avoid it being overwritten and consumed around other resets.
 234 */
 235int vfio_pci_set_power_state(struct vfio_pci_core_device *vdev, pci_power_t state)
 236{
 237	struct pci_dev *pdev = vdev->pdev;
 238	bool needs_restore = false, needs_save = false;
 239	int ret;
 240
 241	/* Prevent changing power state for PFs with VFs enabled */
 242	if (pci_num_vf(pdev) && state > PCI_D0)
 243		return -EBUSY;
 244
 245	if (vdev->needs_pm_restore) {
 246		if (pdev->current_state < PCI_D3hot && state >= PCI_D3hot) {
 247			pci_save_state(pdev);
 248			needs_save = true;
 249		}
 250
 251		if (pdev->current_state >= PCI_D3hot && state <= PCI_D0)
 252			needs_restore = true;
 253	}
 254
 255	ret = pci_set_power_state(pdev, state);
 256
 257	if (!ret) {
 258		/* D3 might be unsupported via quirk, skip unless in D3 */
 259		if (needs_save && pdev->current_state >= PCI_D3hot) {
 260			/*
 261			 * The current PCI state will be saved locally in
 262			 * 'pm_save' during the D3hot transition. When the
 263			 * device state is changed to D0 again with the current
 264			 * function, then pci_store_saved_state() will restore
 265			 * the state and will free the memory pointed by
 266			 * 'pm_save'. There are few cases where the PCI power
 267			 * state can be changed to D0 without the involvement
 268			 * of the driver. For these cases, free the earlier
 269			 * allocated memory first before overwriting 'pm_save'
 270			 * to prevent the memory leak.
 271			 */
 272			kfree(vdev->pm_save);
 273			vdev->pm_save = pci_store_saved_state(pdev);
 274		} else if (needs_restore) {
 275			pci_load_and_free_saved_state(pdev, &vdev->pm_save);
 276			pci_restore_state(pdev);
 277		}
 278	}
 279
 280	return ret;
 281}
 282
 283static int vfio_pci_runtime_pm_entry(struct vfio_pci_core_device *vdev,
 284				     struct eventfd_ctx *efdctx)
 285{
 286	/*
 287	 * The vdev power related flags are protected with 'memory_lock'
 288	 * semaphore.
 289	 */
 290	vfio_pci_zap_and_down_write_memory_lock(vdev);
 291	if (vdev->pm_runtime_engaged) {
 292		up_write(&vdev->memory_lock);
 293		return -EINVAL;
 294	}
 295
 296	vdev->pm_runtime_engaged = true;
 297	vdev->pm_wake_eventfd_ctx = efdctx;
 298	pm_runtime_put_noidle(&vdev->pdev->dev);
 299	up_write(&vdev->memory_lock);
 300
 301	return 0;
 302}
 303
 304static int vfio_pci_core_pm_entry(struct vfio_device *device, u32 flags,
 305				  void __user *arg, size_t argsz)
 306{
 307	struct vfio_pci_core_device *vdev =
 308		container_of(device, struct vfio_pci_core_device, vdev);
 309	int ret;
 310
 311	ret = vfio_check_feature(flags, argsz, VFIO_DEVICE_FEATURE_SET, 0);
 312	if (ret != 1)
 313		return ret;
 314
 315	/*
 316	 * Inside vfio_pci_runtime_pm_entry(), only the runtime PM usage count
 317	 * will be decremented. The pm_runtime_put() will be invoked again
 318	 * while returning from the ioctl and then the device can go into
 319	 * runtime suspended state.
 320	 */
 321	return vfio_pci_runtime_pm_entry(vdev, NULL);
 322}
 323
 324static int vfio_pci_core_pm_entry_with_wakeup(
 325	struct vfio_device *device, u32 flags,
 326	struct vfio_device_low_power_entry_with_wakeup __user *arg,
 327	size_t argsz)
 328{
 329	struct vfio_pci_core_device *vdev =
 330		container_of(device, struct vfio_pci_core_device, vdev);
 331	struct vfio_device_low_power_entry_with_wakeup entry;
 332	struct eventfd_ctx *efdctx;
 333	int ret;
 334
 335	ret = vfio_check_feature(flags, argsz, VFIO_DEVICE_FEATURE_SET,
 336				 sizeof(entry));
 337	if (ret != 1)
 338		return ret;
 339
 340	if (copy_from_user(&entry, arg, sizeof(entry)))
 341		return -EFAULT;
 342
 343	if (entry.wakeup_eventfd < 0)
 344		return -EINVAL;
 345
 346	efdctx = eventfd_ctx_fdget(entry.wakeup_eventfd);
 347	if (IS_ERR(efdctx))
 348		return PTR_ERR(efdctx);
 349
 350	ret = vfio_pci_runtime_pm_entry(vdev, efdctx);
 351	if (ret)
 352		eventfd_ctx_put(efdctx);
 353
 354	return ret;
 355}
 356
 357static void __vfio_pci_runtime_pm_exit(struct vfio_pci_core_device *vdev)
 358{
 359	if (vdev->pm_runtime_engaged) {
 360		vdev->pm_runtime_engaged = false;
 361		pm_runtime_get_noresume(&vdev->pdev->dev);
 362
 363		if (vdev->pm_wake_eventfd_ctx) {
 364			eventfd_ctx_put(vdev->pm_wake_eventfd_ctx);
 365			vdev->pm_wake_eventfd_ctx = NULL;
 366		}
 367	}
 368}
 369
 370static void vfio_pci_runtime_pm_exit(struct vfio_pci_core_device *vdev)
 371{
 372	/*
 373	 * The vdev power related flags are protected with 'memory_lock'
 374	 * semaphore.
 375	 */
 376	down_write(&vdev->memory_lock);
 377	__vfio_pci_runtime_pm_exit(vdev);
 378	up_write(&vdev->memory_lock);
 379}
 380
 381static int vfio_pci_core_pm_exit(struct vfio_device *device, u32 flags,
 382				 void __user *arg, size_t argsz)
 383{
 384	struct vfio_pci_core_device *vdev =
 385		container_of(device, struct vfio_pci_core_device, vdev);
 386	int ret;
 387
 388	ret = vfio_check_feature(flags, argsz, VFIO_DEVICE_FEATURE_SET, 0);
 389	if (ret != 1)
 390		return ret;
 391
 392	/*
 393	 * The device is always in the active state here due to pm wrappers
 394	 * around ioctls. If the device had entered a low power state and
 395	 * pm_wake_eventfd_ctx is valid, vfio_pci_core_runtime_resume() has
 396	 * already signaled the eventfd and exited low power mode itself.
 397	 * pm_runtime_engaged protects the redundant call here.
 398	 */
 399	vfio_pci_runtime_pm_exit(vdev);
 400	return 0;
 401}
 402
 403#ifdef CONFIG_PM
 404static int vfio_pci_core_runtime_suspend(struct device *dev)
 405{
 406	struct vfio_pci_core_device *vdev = dev_get_drvdata(dev);
 407
 408	down_write(&vdev->memory_lock);
 409	/*
 410	 * The user can move the device into D3hot state before invoking
 411	 * power management IOCTL. Move the device into D0 state here and then
 412	 * the pci-driver core runtime PM suspend function will move the device
 413	 * into the low power state. Also, for the devices which have
 414	 * NoSoftRst-, it will help in restoring the original state
 415	 * (saved locally in 'vdev->pm_save').
 416	 */
 417	vfio_pci_set_power_state(vdev, PCI_D0);
 418	up_write(&vdev->memory_lock);
 419
 420	/*
 421	 * If INTx is enabled, then mask INTx before going into the runtime
 422	 * suspended state and unmask the same in the runtime resume.
 423	 * If INTx has already been masked by the user, then
 424	 * vfio_pci_intx_mask() will return false and in that case, INTx
 425	 * should not be unmasked in the runtime resume.
 426	 */
 427	vdev->pm_intx_masked = ((vdev->irq_type == VFIO_PCI_INTX_IRQ_INDEX) &&
 428				vfio_pci_intx_mask(vdev));
 429
 430	return 0;
 431}
 432
 433static int vfio_pci_core_runtime_resume(struct device *dev)
 434{
 435	struct vfio_pci_core_device *vdev = dev_get_drvdata(dev);
 436
 437	/*
 438	 * Resume with a pm_wake_eventfd_ctx signals the eventfd and exit
 439	 * low power mode.
 440	 */
 441	down_write(&vdev->memory_lock);
 442	if (vdev->pm_wake_eventfd_ctx) {
 443		eventfd_signal(vdev->pm_wake_eventfd_ctx, 1);
 444		__vfio_pci_runtime_pm_exit(vdev);
 445	}
 446	up_write(&vdev->memory_lock);
 447
 448	if (vdev->pm_intx_masked)
 449		vfio_pci_intx_unmask(vdev);
 450
 451	return 0;
 452}
 453#endif /* CONFIG_PM */
 454
 455/*
 456 * The pci-driver core runtime PM routines always save the device state
 457 * before going into suspended state. If the device is going into low power
 458 * state with only with runtime PM ops, then no explicit handling is needed
 459 * for the devices which have NoSoftRst-.
 460 */
 461static const struct dev_pm_ops vfio_pci_core_pm_ops = {
 462	SET_RUNTIME_PM_OPS(vfio_pci_core_runtime_suspend,
 463			   vfio_pci_core_runtime_resume,
 464			   NULL)
 465};
 466
 467int vfio_pci_core_enable(struct vfio_pci_core_device *vdev)
 468{
 469	struct pci_dev *pdev = vdev->pdev;
 470	int ret;
 471	u16 cmd;
 472	u8 msix_pos;
 473
 474	if (!disable_idle_d3) {
 475		ret = pm_runtime_resume_and_get(&pdev->dev);
 476		if (ret < 0)
 477			return ret;
 478	}
 479
 480	/* Don't allow our initial saved state to include busmaster */
 481	pci_clear_master(pdev);
 482
 483	ret = pci_enable_device(pdev);
 484	if (ret)
 485		goto out_power;
 486
 487	/* If reset fails because of the device lock, fail this path entirely */
 488	ret = pci_try_reset_function(pdev);
 489	if (ret == -EAGAIN)
 490		goto out_disable_device;
 491
 492	vdev->reset_works = !ret;
 493	pci_save_state(pdev);
 494	vdev->pci_saved_state = pci_store_saved_state(pdev);
 495	if (!vdev->pci_saved_state)
 496		pci_dbg(pdev, "%s: Couldn't store saved state\n", __func__);
 497
 498	if (likely(!nointxmask)) {
 499		if (vfio_pci_nointx(pdev)) {
 500			pci_info(pdev, "Masking broken INTx support\n");
 501			vdev->nointx = true;
 502			pci_intx(pdev, 0);
 503		} else
 504			vdev->pci_2_3 = pci_intx_mask_supported(pdev);
 505	}
 506
 507	pci_read_config_word(pdev, PCI_COMMAND, &cmd);
 508	if (vdev->pci_2_3 && (cmd & PCI_COMMAND_INTX_DISABLE)) {
 509		cmd &= ~PCI_COMMAND_INTX_DISABLE;
 510		pci_write_config_word(pdev, PCI_COMMAND, cmd);
 511	}
 512
 513	ret = vfio_pci_zdev_open_device(vdev);
 514	if (ret)
 515		goto out_free_state;
 516
 517	ret = vfio_config_init(vdev);
 518	if (ret)
 519		goto out_free_zdev;
 520
 521	msix_pos = pdev->msix_cap;
 522	if (msix_pos) {
 523		u16 flags;
 524		u32 table;
 525
 526		pci_read_config_word(pdev, msix_pos + PCI_MSIX_FLAGS, &flags);
 527		pci_read_config_dword(pdev, msix_pos + PCI_MSIX_TABLE, &table);
 528
 529		vdev->msix_bar = table & PCI_MSIX_TABLE_BIR;
 530		vdev->msix_offset = table & PCI_MSIX_TABLE_OFFSET;
 531		vdev->msix_size = ((flags & PCI_MSIX_FLAGS_QSIZE) + 1) * 16;
 532	} else
 
 533		vdev->msix_bar = 0xFF;
 
 
 534
 535	if (!vfio_vga_disabled() && vfio_pci_is_vga(pdev))
 536		vdev->has_vga = true;
 537
 538
 539	return 0;
 540
 541out_free_zdev:
 542	vfio_pci_zdev_close_device(vdev);
 543out_free_state:
 544	kfree(vdev->pci_saved_state);
 545	vdev->pci_saved_state = NULL;
 546out_disable_device:
 547	pci_disable_device(pdev);
 548out_power:
 549	if (!disable_idle_d3)
 550		pm_runtime_put(&pdev->dev);
 551	return ret;
 552}
 553EXPORT_SYMBOL_GPL(vfio_pci_core_enable);
 554
 555void vfio_pci_core_disable(struct vfio_pci_core_device *vdev)
 556{
 557	struct pci_dev *pdev = vdev->pdev;
 558	struct vfio_pci_dummy_resource *dummy_res, *tmp;
 559	struct vfio_pci_ioeventfd *ioeventfd, *ioeventfd_tmp;
 560	int i, bar;
 561
 562	/* For needs_reset */
 563	lockdep_assert_held(&vdev->vdev.dev_set->lock);
 564
 565	/*
 566	 * This function can be invoked while the power state is non-D0.
 567	 * This non-D0 power state can be with or without runtime PM.
 568	 * vfio_pci_runtime_pm_exit() will internally increment the usage
 569	 * count corresponding to pm_runtime_put() called during low power
 570	 * feature entry and then pm_runtime_resume() will wake up the device,
 571	 * if the device has already gone into the suspended state. Otherwise,
 572	 * the vfio_pci_set_power_state() will change the device power state
 573	 * to D0.
 574	 */
 575	vfio_pci_runtime_pm_exit(vdev);
 576	pm_runtime_resume(&pdev->dev);
 577
 578	/*
 579	 * This function calls __pci_reset_function_locked() which internally
 580	 * can use pci_pm_reset() for the function reset. pci_pm_reset() will
 581	 * fail if the power state is non-D0. Also, for the devices which
 582	 * have NoSoftRst-, the reset function can cause the PCI config space
 583	 * reset without restoring the original state (saved locally in
 584	 * 'vdev->pm_save').
 585	 */
 586	vfio_pci_set_power_state(vdev, PCI_D0);
 587
 588	/* Stop the device from further DMA */
 589	pci_clear_master(pdev);
 590
 591	vfio_pci_set_irqs_ioctl(vdev, VFIO_IRQ_SET_DATA_NONE |
 592				VFIO_IRQ_SET_ACTION_TRIGGER,
 593				vdev->irq_type, 0, 0, NULL);
 594
 595	/* Device closed, don't need mutex here */
 596	list_for_each_entry_safe(ioeventfd, ioeventfd_tmp,
 597				 &vdev->ioeventfds_list, next) {
 598		vfio_virqfd_disable(&ioeventfd->virqfd);
 599		list_del(&ioeventfd->next);
 600		kfree(ioeventfd);
 601	}
 602	vdev->ioeventfds_nr = 0;
 603
 604	vdev->virq_disabled = false;
 605
 606	for (i = 0; i < vdev->num_regions; i++)
 607		vdev->region[i].ops->release(vdev, &vdev->region[i]);
 608
 609	vdev->num_regions = 0;
 610	kfree(vdev->region);
 611	vdev->region = NULL; /* don't krealloc a freed pointer */
 612
 613	vfio_config_free(vdev);
 614
 615	for (i = 0; i < PCI_STD_NUM_BARS; i++) {
 616		bar = i + PCI_STD_RESOURCES;
 617		if (!vdev->barmap[bar])
 618			continue;
 619		pci_iounmap(pdev, vdev->barmap[bar]);
 620		pci_release_selected_regions(pdev, 1 << bar);
 621		vdev->barmap[bar] = NULL;
 622	}
 623
 624	list_for_each_entry_safe(dummy_res, tmp,
 625				 &vdev->dummy_resources_list, res_next) {
 626		list_del(&dummy_res->res_next);
 627		release_resource(&dummy_res->resource);
 628		kfree(dummy_res);
 629	}
 630
 631	vdev->needs_reset = true;
 632
 633	vfio_pci_zdev_close_device(vdev);
 634
 635	/*
 636	 * If we have saved state, restore it.  If we can reset the device,
 637	 * even better.  Resetting with current state seems better than
 638	 * nothing, but saving and restoring current state without reset
 639	 * is just busy work.
 640	 */
 641	if (pci_load_and_free_saved_state(pdev, &vdev->pci_saved_state)) {
 642		pci_info(pdev, "%s: Couldn't reload saved state\n", __func__);
 643
 644		if (!vdev->reset_works)
 645			goto out;
 646
 647		pci_save_state(pdev);
 648	}
 649
 650	/*
 651	 * Disable INTx and MSI, presumably to avoid spurious interrupts
 652	 * during reset.  Stolen from pci_reset_function()
 653	 */
 654	pci_write_config_word(pdev, PCI_COMMAND, PCI_COMMAND_INTX_DISABLE);
 655
 656	/*
 657	 * Try to get the locks ourselves to prevent a deadlock. The
 658	 * success of this is dependent on being able to lock the device,
 659	 * which is not always possible.
 660	 * We can not use the "try" reset interface here, which will
 661	 * overwrite the previously restored configuration information.
 662	 */
 663	if (vdev->reset_works && pci_dev_trylock(pdev)) {
 664		if (!__pci_reset_function_locked(pdev))
 665			vdev->needs_reset = false;
 666		pci_dev_unlock(pdev);
 667	}
 668
 669	pci_restore_state(pdev);
 670out:
 671	pci_disable_device(pdev);
 672
 673	vfio_pci_dev_set_try_reset(vdev->vdev.dev_set);
 674
 675	/* Put the pm-runtime usage counter acquired during enable */
 676	if (!disable_idle_d3)
 677		pm_runtime_put(&pdev->dev);
 678}
 679EXPORT_SYMBOL_GPL(vfio_pci_core_disable);
 680
 681void vfio_pci_core_close_device(struct vfio_device *core_vdev)
 682{
 683	struct vfio_pci_core_device *vdev =
 684		container_of(core_vdev, struct vfio_pci_core_device, vdev);
 685
 686	if (vdev->sriov_pf_core_dev) {
 687		mutex_lock(&vdev->sriov_pf_core_dev->vf_token->lock);
 688		WARN_ON(!vdev->sriov_pf_core_dev->vf_token->users);
 689		vdev->sriov_pf_core_dev->vf_token->users--;
 690		mutex_unlock(&vdev->sriov_pf_core_dev->vf_token->lock);
 691	}
 692#if IS_ENABLED(CONFIG_EEH)
 693	eeh_dev_release(vdev->pdev);
 694#endif
 695	vfio_pci_core_disable(vdev);
 696
 697	mutex_lock(&vdev->igate);
 698	if (vdev->err_trigger) {
 699		eventfd_ctx_put(vdev->err_trigger);
 700		vdev->err_trigger = NULL;
 701	}
 702	if (vdev->req_trigger) {
 703		eventfd_ctx_put(vdev->req_trigger);
 704		vdev->req_trigger = NULL;
 705	}
 706	mutex_unlock(&vdev->igate);
 707}
 708EXPORT_SYMBOL_GPL(vfio_pci_core_close_device);
 709
 710void vfio_pci_core_finish_enable(struct vfio_pci_core_device *vdev)
 711{
 712	vfio_pci_probe_mmaps(vdev);
 713#if IS_ENABLED(CONFIG_EEH)
 714	eeh_dev_open(vdev->pdev);
 715#endif
 716
 717	if (vdev->sriov_pf_core_dev) {
 718		mutex_lock(&vdev->sriov_pf_core_dev->vf_token->lock);
 719		vdev->sriov_pf_core_dev->vf_token->users++;
 720		mutex_unlock(&vdev->sriov_pf_core_dev->vf_token->lock);
 721	}
 722}
 723EXPORT_SYMBOL_GPL(vfio_pci_core_finish_enable);
 724
 725static int vfio_pci_get_irq_count(struct vfio_pci_core_device *vdev, int irq_type)
 726{
 727	if (irq_type == VFIO_PCI_INTX_IRQ_INDEX) {
 728		u8 pin;
 729
 730		if (!IS_ENABLED(CONFIG_VFIO_PCI_INTX) ||
 731		    vdev->nointx || vdev->pdev->is_virtfn)
 732			return 0;
 733
 734		pci_read_config_byte(vdev->pdev, PCI_INTERRUPT_PIN, &pin);
 735
 736		return pin ? 1 : 0;
 737	} else if (irq_type == VFIO_PCI_MSI_IRQ_INDEX) {
 738		u8 pos;
 739		u16 flags;
 740
 741		pos = vdev->pdev->msi_cap;
 742		if (pos) {
 743			pci_read_config_word(vdev->pdev,
 744					     pos + PCI_MSI_FLAGS, &flags);
 745			return 1 << ((flags & PCI_MSI_FLAGS_QMASK) >> 1);
 746		}
 747	} else if (irq_type == VFIO_PCI_MSIX_IRQ_INDEX) {
 748		u8 pos;
 749		u16 flags;
 750
 751		pos = vdev->pdev->msix_cap;
 752		if (pos) {
 753			pci_read_config_word(vdev->pdev,
 754					     pos + PCI_MSIX_FLAGS, &flags);
 755
 756			return (flags & PCI_MSIX_FLAGS_QSIZE) + 1;
 757		}
 758	} else if (irq_type == VFIO_PCI_ERR_IRQ_INDEX) {
 759		if (pci_is_pcie(vdev->pdev))
 760			return 1;
 761	} else if (irq_type == VFIO_PCI_REQ_IRQ_INDEX) {
 762		return 1;
 763	}
 764
 765	return 0;
 766}
 767
 768static int vfio_pci_count_devs(struct pci_dev *pdev, void *data)
 769{
 770	(*(int *)data)++;
 771	return 0;
 772}
 773
 774struct vfio_pci_fill_info {
 775	int max;
 776	int cur;
 777	struct vfio_pci_dependent_device *devices;
 
 
 778};
 779
 780static int vfio_pci_fill_devs(struct pci_dev *pdev, void *data)
 781{
 
 
 
 
 
 782	struct vfio_pci_fill_info *fill = data;
 783	struct iommu_group *iommu_group;
 784
 785	if (fill->cur == fill->max)
 786		return -EAGAIN; /* Something changed, try again */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 787
 788	iommu_group = iommu_group_get(&pdev->dev);
 789	if (!iommu_group)
 790		return -EPERM; /* Cannot reset non-isolated devices */
 791
 792	fill->devices[fill->cur].group_id = iommu_group_id(iommu_group);
 793	fill->devices[fill->cur].segment = pci_domain_nr(pdev->bus);
 794	fill->devices[fill->cur].bus = pdev->bus->number;
 795	fill->devices[fill->cur].devfn = pdev->devfn;
 796	fill->cur++;
 797	iommu_group_put(iommu_group);
 798	return 0;
 799}
 800
 801struct vfio_pci_group_info {
 802	int count;
 803	struct file **files;
 804};
 805
 806static bool vfio_pci_dev_below_slot(struct pci_dev *pdev, struct pci_slot *slot)
 807{
 808	for (; pdev; pdev = pdev->bus->self)
 809		if (pdev->bus == slot->bus)
 810			return (pdev->slot == slot);
 811	return false;
 812}
 813
 814struct vfio_pci_walk_info {
 815	int (*fn)(struct pci_dev *pdev, void *data);
 816	void *data;
 817	struct pci_dev *pdev;
 818	bool slot;
 819	int ret;
 820};
 821
 822static int vfio_pci_walk_wrapper(struct pci_dev *pdev, void *data)
 823{
 824	struct vfio_pci_walk_info *walk = data;
 825
 826	if (!walk->slot || vfio_pci_dev_below_slot(pdev, walk->pdev->slot))
 827		walk->ret = walk->fn(pdev, walk->data);
 828
 829	return walk->ret;
 830}
 831
 832static int vfio_pci_for_each_slot_or_bus(struct pci_dev *pdev,
 833					 int (*fn)(struct pci_dev *,
 834						   void *data), void *data,
 835					 bool slot)
 836{
 837	struct vfio_pci_walk_info walk = {
 838		.fn = fn, .data = data, .pdev = pdev, .slot = slot, .ret = 0,
 839	};
 840
 841	pci_walk_bus(pdev->bus, vfio_pci_walk_wrapper, &walk);
 842
 843	return walk.ret;
 844}
 845
 846static int msix_mmappable_cap(struct vfio_pci_core_device *vdev,
 847			      struct vfio_info_cap *caps)
 848{
 849	struct vfio_info_cap_header header = {
 850		.id = VFIO_REGION_INFO_CAP_MSIX_MAPPABLE,
 851		.version = 1
 852	};
 853
 854	return vfio_info_add_capability(caps, &header, sizeof(header));
 855}
 856
 857int vfio_pci_core_register_dev_region(struct vfio_pci_core_device *vdev,
 858				      unsigned int type, unsigned int subtype,
 859				      const struct vfio_pci_regops *ops,
 860				      size_t size, u32 flags, void *data)
 861{
 862	struct vfio_pci_region *region;
 863
 864	region = krealloc(vdev->region,
 865			  (vdev->num_regions + 1) * sizeof(*region),
 866			  GFP_KERNEL);
 867	if (!region)
 868		return -ENOMEM;
 869
 870	vdev->region = region;
 871	vdev->region[vdev->num_regions].type = type;
 872	vdev->region[vdev->num_regions].subtype = subtype;
 873	vdev->region[vdev->num_regions].ops = ops;
 874	vdev->region[vdev->num_regions].size = size;
 875	vdev->region[vdev->num_regions].flags = flags;
 876	vdev->region[vdev->num_regions].data = data;
 877
 878	vdev->num_regions++;
 879
 880	return 0;
 881}
 882EXPORT_SYMBOL_GPL(vfio_pci_core_register_dev_region);
 883
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 884static int vfio_pci_ioctl_get_info(struct vfio_pci_core_device *vdev,
 885				   struct vfio_device_info __user *arg)
 886{
 887	unsigned long minsz = offsetofend(struct vfio_device_info, num_irqs);
 888	struct vfio_device_info info;
 889	struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
 890	unsigned long capsz;
 891	int ret;
 892
 893	/* For backward compatibility, cannot require this */
 894	capsz = offsetofend(struct vfio_iommu_type1_info, cap_offset);
 895
 896	if (copy_from_user(&info, arg, minsz))
 897		return -EFAULT;
 898
 899	if (info.argsz < minsz)
 900		return -EINVAL;
 901
 902	if (info.argsz >= capsz) {
 903		minsz = capsz;
 904		info.cap_offset = 0;
 905	}
 906
 907	info.flags = VFIO_DEVICE_FLAGS_PCI;
 908
 909	if (vdev->reset_works)
 910		info.flags |= VFIO_DEVICE_FLAGS_RESET;
 911
 912	info.num_regions = VFIO_PCI_NUM_REGIONS + vdev->num_regions;
 913	info.num_irqs = VFIO_PCI_NUM_IRQS;
 914
 915	ret = vfio_pci_info_zdev_add_caps(vdev, &caps);
 916	if (ret && ret != -ENODEV) {
 917		pci_warn(vdev->pdev,
 918			 "Failed to setup zPCI info capabilities\n");
 919		return ret;
 920	}
 921
 
 
 
 
 
 
 
 922	if (caps.size) {
 923		info.flags |= VFIO_DEVICE_FLAGS_CAPS;
 924		if (info.argsz < sizeof(info) + caps.size) {
 925			info.argsz = sizeof(info) + caps.size;
 926		} else {
 927			vfio_info_cap_shift(&caps, sizeof(info));
 928			if (copy_to_user(arg + 1, caps.buf, caps.size)) {
 929				kfree(caps.buf);
 930				return -EFAULT;
 931			}
 932			info.cap_offset = sizeof(*arg);
 933		}
 934
 935		kfree(caps.buf);
 936	}
 937
 938	return copy_to_user(arg, &info, minsz) ? -EFAULT : 0;
 939}
 940
 941static int vfio_pci_ioctl_get_region_info(struct vfio_pci_core_device *vdev,
 942					  struct vfio_region_info __user *arg)
 943{
 944	unsigned long minsz = offsetofend(struct vfio_region_info, offset);
 945	struct pci_dev *pdev = vdev->pdev;
 946	struct vfio_region_info info;
 947	struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
 948	int i, ret;
 949
 950	if (copy_from_user(&info, arg, minsz))
 951		return -EFAULT;
 952
 953	if (info.argsz < minsz)
 954		return -EINVAL;
 955
 956	switch (info.index) {
 957	case VFIO_PCI_CONFIG_REGION_INDEX:
 958		info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
 959		info.size = pdev->cfg_size;
 960		info.flags = VFIO_REGION_INFO_FLAG_READ |
 961			     VFIO_REGION_INFO_FLAG_WRITE;
 962		break;
 963	case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
 964		info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
 965		info.size = pci_resource_len(pdev, info.index);
 966		if (!info.size) {
 967			info.flags = 0;
 968			break;
 969		}
 970
 971		info.flags = VFIO_REGION_INFO_FLAG_READ |
 972			     VFIO_REGION_INFO_FLAG_WRITE;
 973		if (vdev->bar_mmap_supported[info.index]) {
 974			info.flags |= VFIO_REGION_INFO_FLAG_MMAP;
 975			if (info.index == vdev->msix_bar) {
 976				ret = msix_mmappable_cap(vdev, &caps);
 977				if (ret)
 978					return ret;
 979			}
 980		}
 981
 982		break;
 983	case VFIO_PCI_ROM_REGION_INDEX: {
 984		void __iomem *io;
 985		size_t size;
 986		u16 cmd;
 987
 988		info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
 989		info.flags = 0;
 990
 991		/* Report the BAR size, not the ROM size */
 992		info.size = pci_resource_len(pdev, info.index);
 993		if (!info.size) {
 994			/* Shadow ROMs appear as PCI option ROMs */
 995			if (pdev->resource[PCI_ROM_RESOURCE].flags &
 996			    IORESOURCE_ROM_SHADOW)
 997				info.size = 0x20000;
 998			else
 999				break;
1000		}
1001
1002		/*
1003		 * Is it really there?  Enable memory decode for implicit access
1004		 * in pci_map_rom().
1005		 */
1006		cmd = vfio_pci_memory_lock_and_enable(vdev);
1007		io = pci_map_rom(pdev, &size);
1008		if (io) {
1009			info.flags = VFIO_REGION_INFO_FLAG_READ;
1010			pci_unmap_rom(pdev, io);
1011		} else {
1012			info.size = 0;
1013		}
1014		vfio_pci_memory_unlock_and_restore(vdev, cmd);
1015
1016		break;
1017	}
1018	case VFIO_PCI_VGA_REGION_INDEX:
1019		if (!vdev->has_vga)
1020			return -EINVAL;
1021
1022		info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
1023		info.size = 0xc0000;
1024		info.flags = VFIO_REGION_INFO_FLAG_READ |
1025			     VFIO_REGION_INFO_FLAG_WRITE;
1026
1027		break;
1028	default: {
1029		struct vfio_region_info_cap_type cap_type = {
1030			.header.id = VFIO_REGION_INFO_CAP_TYPE,
1031			.header.version = 1
1032		};
1033
1034		if (info.index >= VFIO_PCI_NUM_REGIONS + vdev->num_regions)
1035			return -EINVAL;
1036		info.index = array_index_nospec(
1037			info.index, VFIO_PCI_NUM_REGIONS + vdev->num_regions);
1038
1039		i = info.index - VFIO_PCI_NUM_REGIONS;
1040
1041		info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
1042		info.size = vdev->region[i].size;
1043		info.flags = vdev->region[i].flags;
1044
1045		cap_type.type = vdev->region[i].type;
1046		cap_type.subtype = vdev->region[i].subtype;
1047
1048		ret = vfio_info_add_capability(&caps, &cap_type.header,
1049					       sizeof(cap_type));
1050		if (ret)
1051			return ret;
1052
1053		if (vdev->region[i].ops->add_capability) {
1054			ret = vdev->region[i].ops->add_capability(
1055				vdev, &vdev->region[i], &caps);
1056			if (ret)
1057				return ret;
1058		}
1059	}
1060	}
1061
1062	if (caps.size) {
1063		info.flags |= VFIO_REGION_INFO_FLAG_CAPS;
1064		if (info.argsz < sizeof(info) + caps.size) {
1065			info.argsz = sizeof(info) + caps.size;
1066			info.cap_offset = 0;
1067		} else {
1068			vfio_info_cap_shift(&caps, sizeof(info));
1069			if (copy_to_user(arg + 1, caps.buf, caps.size)) {
1070				kfree(caps.buf);
1071				return -EFAULT;
1072			}
1073			info.cap_offset = sizeof(*arg);
1074		}
1075
1076		kfree(caps.buf);
1077	}
1078
1079	return copy_to_user(arg, &info, minsz) ? -EFAULT : 0;
1080}
1081
1082static int vfio_pci_ioctl_get_irq_info(struct vfio_pci_core_device *vdev,
1083				       struct vfio_irq_info __user *arg)
1084{
1085	unsigned long minsz = offsetofend(struct vfio_irq_info, count);
1086	struct vfio_irq_info info;
1087
1088	if (copy_from_user(&info, arg, minsz))
1089		return -EFAULT;
1090
1091	if (info.argsz < minsz || info.index >= VFIO_PCI_NUM_IRQS)
1092		return -EINVAL;
1093
1094	switch (info.index) {
1095	case VFIO_PCI_INTX_IRQ_INDEX ... VFIO_PCI_MSIX_IRQ_INDEX:
1096	case VFIO_PCI_REQ_IRQ_INDEX:
1097		break;
1098	case VFIO_PCI_ERR_IRQ_INDEX:
1099		if (pci_is_pcie(vdev->pdev))
1100			break;
1101		fallthrough;
1102	default:
1103		return -EINVAL;
1104	}
1105
1106	info.flags = VFIO_IRQ_INFO_EVENTFD;
1107
1108	info.count = vfio_pci_get_irq_count(vdev, info.index);
1109
1110	if (info.index == VFIO_PCI_INTX_IRQ_INDEX)
1111		info.flags |=
1112			(VFIO_IRQ_INFO_MASKABLE | VFIO_IRQ_INFO_AUTOMASKED);
1113	else
1114		info.flags |= VFIO_IRQ_INFO_NORESIZE;
1115
1116	return copy_to_user(arg, &info, minsz) ? -EFAULT : 0;
1117}
1118
1119static int vfio_pci_ioctl_set_irqs(struct vfio_pci_core_device *vdev,
1120				   struct vfio_irq_set __user *arg)
1121{
1122	unsigned long minsz = offsetofend(struct vfio_irq_set, count);
1123	struct vfio_irq_set hdr;
1124	u8 *data = NULL;
1125	int max, ret = 0;
1126	size_t data_size = 0;
1127
1128	if (copy_from_user(&hdr, arg, minsz))
1129		return -EFAULT;
1130
1131	max = vfio_pci_get_irq_count(vdev, hdr.index);
1132
1133	ret = vfio_set_irqs_validate_and_prepare(&hdr, max, VFIO_PCI_NUM_IRQS,
1134						 &data_size);
1135	if (ret)
1136		return ret;
1137
1138	if (data_size) {
1139		data = memdup_user(&arg->data, data_size);
1140		if (IS_ERR(data))
1141			return PTR_ERR(data);
1142	}
1143
1144	mutex_lock(&vdev->igate);
1145
1146	ret = vfio_pci_set_irqs_ioctl(vdev, hdr.flags, hdr.index, hdr.start,
1147				      hdr.count, data);
1148
1149	mutex_unlock(&vdev->igate);
1150	kfree(data);
1151
1152	return ret;
1153}
1154
1155static int vfio_pci_ioctl_reset(struct vfio_pci_core_device *vdev,
1156				void __user *arg)
1157{
1158	int ret;
1159
1160	if (!vdev->reset_works)
1161		return -EINVAL;
1162
1163	vfio_pci_zap_and_down_write_memory_lock(vdev);
1164
1165	/*
1166	 * This function can be invoked while the power state is non-D0. If
1167	 * pci_try_reset_function() has been called while the power state is
1168	 * non-D0, then pci_try_reset_function() will internally set the power
1169	 * state to D0 without vfio driver involvement. For the devices which
1170	 * have NoSoftRst-, the reset function can cause the PCI config space
1171	 * reset without restoring the original state (saved locally in
1172	 * 'vdev->pm_save').
1173	 */
1174	vfio_pci_set_power_state(vdev, PCI_D0);
1175
1176	ret = pci_try_reset_function(vdev->pdev);
1177	up_write(&vdev->memory_lock);
1178
1179	return ret;
1180}
1181
1182static int vfio_pci_ioctl_get_pci_hot_reset_info(
1183	struct vfio_pci_core_device *vdev,
1184	struct vfio_pci_hot_reset_info __user *arg)
1185{
1186	unsigned long minsz =
1187		offsetofend(struct vfio_pci_hot_reset_info, count);
1188	struct vfio_pci_hot_reset_info hdr;
1189	struct vfio_pci_fill_info fill = { 0 };
1190	struct vfio_pci_dependent_device *devices = NULL;
1191	bool slot = false;
1192	int ret = 0;
1193
1194	if (copy_from_user(&hdr, arg, minsz))
1195		return -EFAULT;
1196
1197	if (hdr.argsz < minsz)
1198		return -EINVAL;
1199
1200	hdr.flags = 0;
1201
1202	/* Can we do a slot or bus reset or neither? */
1203	if (!pci_probe_reset_slot(vdev->pdev->slot))
1204		slot = true;
1205	else if (pci_probe_reset_bus(vdev->pdev->bus))
1206		return -ENODEV;
1207
1208	/* How many devices are affected? */
1209	ret = vfio_pci_for_each_slot_or_bus(vdev->pdev, vfio_pci_count_devs,
1210					    &fill.max, slot);
1211	if (ret)
1212		return ret;
1213
1214	WARN_ON(!fill.max); /* Should always be at least one */
1215
1216	/*
1217	 * If there's enough space, fill it now, otherwise return -ENOSPC and
1218	 * the number of devices affected.
1219	 */
1220	if (hdr.argsz < sizeof(hdr) + (fill.max * sizeof(*devices))) {
1221		ret = -ENOSPC;
1222		hdr.count = fill.max;
1223		goto reset_info_exit;
1224	}
1225
1226	devices = kcalloc(fill.max, sizeof(*devices), GFP_KERNEL);
1227	if (!devices)
1228		return -ENOMEM;
1229
1230	fill.devices = devices;
1231
 
1232	ret = vfio_pci_for_each_slot_or_bus(vdev->pdev, vfio_pci_fill_devs,
1233					    &fill, slot);
 
 
 
1234
1235	/*
1236	 * If a device was removed between counting and filling, we may come up
1237	 * short of fill.max.  If a device was added, we'll have a return of
1238	 * -EAGAIN above.
1239	 */
1240	if (!ret)
1241		hdr.count = fill.cur;
1242
1243reset_info_exit:
1244	if (copy_to_user(arg, &hdr, minsz))
1245		ret = -EFAULT;
1246
1247	if (!ret) {
1248		if (copy_to_user(&arg->devices, devices,
1249				 hdr.count * sizeof(*devices)))
1250			ret = -EFAULT;
1251	}
1252
1253	kfree(devices);
1254	return ret;
1255}
1256
1257static int vfio_pci_ioctl_pci_hot_reset(struct vfio_pci_core_device *vdev,
1258					struct vfio_pci_hot_reset __user *arg)
 
 
1259{
1260	unsigned long minsz = offsetofend(struct vfio_pci_hot_reset, count);
1261	struct vfio_pci_hot_reset hdr;
1262	int32_t *group_fds;
1263	struct file **files;
1264	struct vfio_pci_group_info info;
1265	bool slot = false;
1266	int file_idx, count = 0, ret = 0;
1267
1268	if (copy_from_user(&hdr, arg, minsz))
1269		return -EFAULT;
1270
1271	if (hdr.argsz < minsz || hdr.flags)
1272		return -EINVAL;
1273
1274	/* Can we do a slot or bus reset or neither? */
1275	if (!pci_probe_reset_slot(vdev->pdev->slot))
1276		slot = true;
1277	else if (pci_probe_reset_bus(vdev->pdev->bus))
1278		return -ENODEV;
1279
1280	/*
1281	 * We can't let userspace give us an arbitrarily large buffer to copy,
1282	 * so verify how many we think there could be.  Note groups can have
1283	 * multiple devices so one group per device is the max.
1284	 */
1285	ret = vfio_pci_for_each_slot_or_bus(vdev->pdev, vfio_pci_count_devs,
1286					    &count, slot);
1287	if (ret)
1288		return ret;
1289
1290	/* Somewhere between 1 and count is OK */
1291	if (!hdr.count || hdr.count > count)
1292		return -EINVAL;
1293
1294	group_fds = kcalloc(hdr.count, sizeof(*group_fds), GFP_KERNEL);
1295	files = kcalloc(hdr.count, sizeof(*files), GFP_KERNEL);
1296	if (!group_fds || !files) {
1297		kfree(group_fds);
1298		kfree(files);
1299		return -ENOMEM;
1300	}
1301
1302	if (copy_from_user(group_fds, arg->group_fds,
1303			   hdr.count * sizeof(*group_fds))) {
1304		kfree(group_fds);
1305		kfree(files);
1306		return -EFAULT;
1307	}
1308
1309	/*
1310	 * For each group_fd, get the group through the vfio external user
1311	 * interface and store the group and iommu ID.  This ensures the group
1312	 * is held across the reset.
1313	 */
1314	for (file_idx = 0; file_idx < hdr.count; file_idx++) {
1315		struct file *file = fget(group_fds[file_idx]);
1316
1317		if (!file) {
1318			ret = -EBADF;
1319			break;
1320		}
1321
1322		/* Ensure the FD is a vfio group FD.*/
1323		if (!vfio_file_is_group(file)) {
1324			fput(file);
1325			ret = -EINVAL;
1326			break;
1327		}
1328
1329		files[file_idx] = file;
1330	}
1331
1332	kfree(group_fds);
1333
1334	/* release reference to groups on error */
1335	if (ret)
1336		goto hot_reset_release;
1337
1338	info.count = hdr.count;
1339	info.files = files;
1340
1341	ret = vfio_pci_dev_set_hot_reset(vdev->vdev.dev_set, &info);
1342
1343hot_reset_release:
1344	for (file_idx--; file_idx >= 0; file_idx--)
1345		fput(files[file_idx]);
1346
1347	kfree(files);
1348	return ret;
1349}
1350
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1351static int vfio_pci_ioctl_ioeventfd(struct vfio_pci_core_device *vdev,
1352				    struct vfio_device_ioeventfd __user *arg)
1353{
1354	unsigned long minsz = offsetofend(struct vfio_device_ioeventfd, fd);
1355	struct vfio_device_ioeventfd ioeventfd;
1356	int count;
1357
1358	if (copy_from_user(&ioeventfd, arg, minsz))
1359		return -EFAULT;
1360
1361	if (ioeventfd.argsz < minsz)
1362		return -EINVAL;
1363
1364	if (ioeventfd.flags & ~VFIO_DEVICE_IOEVENTFD_SIZE_MASK)
1365		return -EINVAL;
1366
1367	count = ioeventfd.flags & VFIO_DEVICE_IOEVENTFD_SIZE_MASK;
1368
1369	if (hweight8(count) != 1 || ioeventfd.fd < -1)
1370		return -EINVAL;
1371
1372	return vfio_pci_ioeventfd(vdev, ioeventfd.offset, ioeventfd.data, count,
1373				  ioeventfd.fd);
1374}
1375
1376long vfio_pci_core_ioctl(struct vfio_device *core_vdev, unsigned int cmd,
1377			 unsigned long arg)
1378{
1379	struct vfio_pci_core_device *vdev =
1380		container_of(core_vdev, struct vfio_pci_core_device, vdev);
1381	void __user *uarg = (void __user *)arg;
1382
1383	switch (cmd) {
1384	case VFIO_DEVICE_GET_INFO:
1385		return vfio_pci_ioctl_get_info(vdev, uarg);
1386	case VFIO_DEVICE_GET_IRQ_INFO:
1387		return vfio_pci_ioctl_get_irq_info(vdev, uarg);
1388	case VFIO_DEVICE_GET_PCI_HOT_RESET_INFO:
1389		return vfio_pci_ioctl_get_pci_hot_reset_info(vdev, uarg);
1390	case VFIO_DEVICE_GET_REGION_INFO:
1391		return vfio_pci_ioctl_get_region_info(vdev, uarg);
1392	case VFIO_DEVICE_IOEVENTFD:
1393		return vfio_pci_ioctl_ioeventfd(vdev, uarg);
1394	case VFIO_DEVICE_PCI_HOT_RESET:
1395		return vfio_pci_ioctl_pci_hot_reset(vdev, uarg);
1396	case VFIO_DEVICE_RESET:
1397		return vfio_pci_ioctl_reset(vdev, uarg);
1398	case VFIO_DEVICE_SET_IRQS:
1399		return vfio_pci_ioctl_set_irqs(vdev, uarg);
1400	default:
1401		return -ENOTTY;
1402	}
1403}
1404EXPORT_SYMBOL_GPL(vfio_pci_core_ioctl);
1405
1406static int vfio_pci_core_feature_token(struct vfio_device *device, u32 flags,
1407				       uuid_t __user *arg, size_t argsz)
1408{
1409	struct vfio_pci_core_device *vdev =
1410		container_of(device, struct vfio_pci_core_device, vdev);
1411	uuid_t uuid;
1412	int ret;
1413
1414	if (!vdev->vf_token)
1415		return -ENOTTY;
1416	/*
1417	 * We do not support GET of the VF Token UUID as this could
1418	 * expose the token of the previous device user.
1419	 */
1420	ret = vfio_check_feature(flags, argsz, VFIO_DEVICE_FEATURE_SET,
1421				 sizeof(uuid));
1422	if (ret != 1)
1423		return ret;
1424
1425	if (copy_from_user(&uuid, arg, sizeof(uuid)))
1426		return -EFAULT;
1427
1428	mutex_lock(&vdev->vf_token->lock);
1429	uuid_copy(&vdev->vf_token->uuid, &uuid);
1430	mutex_unlock(&vdev->vf_token->lock);
1431	return 0;
1432}
1433
1434int vfio_pci_core_ioctl_feature(struct vfio_device *device, u32 flags,
1435				void __user *arg, size_t argsz)
1436{
1437	switch (flags & VFIO_DEVICE_FEATURE_MASK) {
1438	case VFIO_DEVICE_FEATURE_LOW_POWER_ENTRY:
1439		return vfio_pci_core_pm_entry(device, flags, arg, argsz);
1440	case VFIO_DEVICE_FEATURE_LOW_POWER_ENTRY_WITH_WAKEUP:
1441		return vfio_pci_core_pm_entry_with_wakeup(device, flags,
1442							  arg, argsz);
1443	case VFIO_DEVICE_FEATURE_LOW_POWER_EXIT:
1444		return vfio_pci_core_pm_exit(device, flags, arg, argsz);
1445	case VFIO_DEVICE_FEATURE_PCI_VF_TOKEN:
1446		return vfio_pci_core_feature_token(device, flags, arg, argsz);
1447	default:
1448		return -ENOTTY;
1449	}
1450}
1451EXPORT_SYMBOL_GPL(vfio_pci_core_ioctl_feature);
1452
1453static ssize_t vfio_pci_rw(struct vfio_pci_core_device *vdev, char __user *buf,
1454			   size_t count, loff_t *ppos, bool iswrite)
1455{
1456	unsigned int index = VFIO_PCI_OFFSET_TO_INDEX(*ppos);
1457	int ret;
1458
1459	if (index >= VFIO_PCI_NUM_REGIONS + vdev->num_regions)
1460		return -EINVAL;
1461
1462	ret = pm_runtime_resume_and_get(&vdev->pdev->dev);
1463	if (ret) {
1464		pci_info_ratelimited(vdev->pdev, "runtime resume failed %d\n",
1465				     ret);
1466		return -EIO;
1467	}
1468
1469	switch (index) {
1470	case VFIO_PCI_CONFIG_REGION_INDEX:
1471		ret = vfio_pci_config_rw(vdev, buf, count, ppos, iswrite);
1472		break;
1473
1474	case VFIO_PCI_ROM_REGION_INDEX:
1475		if (iswrite)
1476			ret = -EINVAL;
1477		else
1478			ret = vfio_pci_bar_rw(vdev, buf, count, ppos, false);
1479		break;
1480
1481	case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
1482		ret = vfio_pci_bar_rw(vdev, buf, count, ppos, iswrite);
1483		break;
1484
1485	case VFIO_PCI_VGA_REGION_INDEX:
1486		ret = vfio_pci_vga_rw(vdev, buf, count, ppos, iswrite);
1487		break;
1488
1489	default:
1490		index -= VFIO_PCI_NUM_REGIONS;
1491		ret = vdev->region[index].ops->rw(vdev, buf,
1492						   count, ppos, iswrite);
1493		break;
1494	}
1495
1496	pm_runtime_put(&vdev->pdev->dev);
1497	return ret;
1498}
1499
1500ssize_t vfio_pci_core_read(struct vfio_device *core_vdev, char __user *buf,
1501		size_t count, loff_t *ppos)
1502{
1503	struct vfio_pci_core_device *vdev =
1504		container_of(core_vdev, struct vfio_pci_core_device, vdev);
1505
1506	if (!count)
1507		return 0;
1508
1509	return vfio_pci_rw(vdev, buf, count, ppos, false);
1510}
1511EXPORT_SYMBOL_GPL(vfio_pci_core_read);
1512
1513ssize_t vfio_pci_core_write(struct vfio_device *core_vdev, const char __user *buf,
1514		size_t count, loff_t *ppos)
1515{
1516	struct vfio_pci_core_device *vdev =
1517		container_of(core_vdev, struct vfio_pci_core_device, vdev);
1518
1519	if (!count)
1520		return 0;
1521
1522	return vfio_pci_rw(vdev, (char __user *)buf, count, ppos, true);
1523}
1524EXPORT_SYMBOL_GPL(vfio_pci_core_write);
1525
1526/* Return 1 on zap and vma_lock acquired, 0 on contention (only with @try) */
1527static int vfio_pci_zap_and_vma_lock(struct vfio_pci_core_device *vdev, bool try)
1528{
1529	struct vfio_pci_mmap_vma *mmap_vma, *tmp;
1530
1531	/*
1532	 * Lock ordering:
1533	 * vma_lock is nested under mmap_lock for vm_ops callback paths.
1534	 * The memory_lock semaphore is used by both code paths calling
1535	 * into this function to zap vmas and the vm_ops.fault callback
1536	 * to protect the memory enable state of the device.
1537	 *
1538	 * When zapping vmas we need to maintain the mmap_lock => vma_lock
1539	 * ordering, which requires using vma_lock to walk vma_list to
1540	 * acquire an mm, then dropping vma_lock to get the mmap_lock and
1541	 * reacquiring vma_lock.  This logic is derived from similar
1542	 * requirements in uverbs_user_mmap_disassociate().
1543	 *
1544	 * mmap_lock must always be the top-level lock when it is taken.
1545	 * Therefore we can only hold the memory_lock write lock when
1546	 * vma_list is empty, as we'd need to take mmap_lock to clear
1547	 * entries.  vma_list can only be guaranteed empty when holding
1548	 * vma_lock, thus memory_lock is nested under vma_lock.
1549	 *
1550	 * This enables the vm_ops.fault callback to acquire vma_lock,
1551	 * followed by memory_lock read lock, while already holding
1552	 * mmap_lock without risk of deadlock.
1553	 */
1554	while (1) {
1555		struct mm_struct *mm = NULL;
1556
1557		if (try) {
1558			if (!mutex_trylock(&vdev->vma_lock))
1559				return 0;
1560		} else {
1561			mutex_lock(&vdev->vma_lock);
1562		}
1563		while (!list_empty(&vdev->vma_list)) {
1564			mmap_vma = list_first_entry(&vdev->vma_list,
1565						    struct vfio_pci_mmap_vma,
1566						    vma_next);
1567			mm = mmap_vma->vma->vm_mm;
1568			if (mmget_not_zero(mm))
1569				break;
1570
1571			list_del(&mmap_vma->vma_next);
1572			kfree(mmap_vma);
1573			mm = NULL;
1574		}
1575		if (!mm)
1576			return 1;
1577		mutex_unlock(&vdev->vma_lock);
1578
1579		if (try) {
1580			if (!mmap_read_trylock(mm)) {
1581				mmput(mm);
1582				return 0;
1583			}
1584		} else {
1585			mmap_read_lock(mm);
1586		}
1587		if (try) {
1588			if (!mutex_trylock(&vdev->vma_lock)) {
1589				mmap_read_unlock(mm);
1590				mmput(mm);
1591				return 0;
1592			}
1593		} else {
1594			mutex_lock(&vdev->vma_lock);
1595		}
1596		list_for_each_entry_safe(mmap_vma, tmp,
1597					 &vdev->vma_list, vma_next) {
1598			struct vm_area_struct *vma = mmap_vma->vma;
1599
1600			if (vma->vm_mm != mm)
1601				continue;
1602
1603			list_del(&mmap_vma->vma_next);
1604			kfree(mmap_vma);
1605
1606			zap_vma_ptes(vma, vma->vm_start,
1607				     vma->vm_end - vma->vm_start);
1608		}
1609		mutex_unlock(&vdev->vma_lock);
1610		mmap_read_unlock(mm);
1611		mmput(mm);
1612	}
1613}
1614
1615void vfio_pci_zap_and_down_write_memory_lock(struct vfio_pci_core_device *vdev)
1616{
1617	vfio_pci_zap_and_vma_lock(vdev, false);
1618	down_write(&vdev->memory_lock);
1619	mutex_unlock(&vdev->vma_lock);
1620}
1621
1622u16 vfio_pci_memory_lock_and_enable(struct vfio_pci_core_device *vdev)
1623{
1624	u16 cmd;
1625
1626	down_write(&vdev->memory_lock);
1627	pci_read_config_word(vdev->pdev, PCI_COMMAND, &cmd);
1628	if (!(cmd & PCI_COMMAND_MEMORY))
1629		pci_write_config_word(vdev->pdev, PCI_COMMAND,
1630				      cmd | PCI_COMMAND_MEMORY);
1631
1632	return cmd;
1633}
1634
1635void vfio_pci_memory_unlock_and_restore(struct vfio_pci_core_device *vdev, u16 cmd)
1636{
1637	pci_write_config_word(vdev->pdev, PCI_COMMAND, cmd);
1638	up_write(&vdev->memory_lock);
1639}
1640
1641/* Caller holds vma_lock */
1642static int __vfio_pci_add_vma(struct vfio_pci_core_device *vdev,
1643			      struct vm_area_struct *vma)
1644{
1645	struct vfio_pci_mmap_vma *mmap_vma;
1646
1647	mmap_vma = kmalloc(sizeof(*mmap_vma), GFP_KERNEL);
1648	if (!mmap_vma)
1649		return -ENOMEM;
1650
1651	mmap_vma->vma = vma;
1652	list_add(&mmap_vma->vma_next, &vdev->vma_list);
1653
1654	return 0;
1655}
1656
1657/*
1658 * Zap mmaps on open so that we can fault them in on access and therefore
1659 * our vma_list only tracks mappings accessed since last zap.
1660 */
1661static void vfio_pci_mmap_open(struct vm_area_struct *vma)
1662{
1663	zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
1664}
1665
1666static void vfio_pci_mmap_close(struct vm_area_struct *vma)
1667{
1668	struct vfio_pci_core_device *vdev = vma->vm_private_data;
1669	struct vfio_pci_mmap_vma *mmap_vma;
1670
1671	mutex_lock(&vdev->vma_lock);
1672	list_for_each_entry(mmap_vma, &vdev->vma_list, vma_next) {
1673		if (mmap_vma->vma == vma) {
1674			list_del(&mmap_vma->vma_next);
1675			kfree(mmap_vma);
1676			break;
1677		}
1678	}
1679	mutex_unlock(&vdev->vma_lock);
1680}
1681
1682static vm_fault_t vfio_pci_mmap_fault(struct vm_fault *vmf)
1683{
1684	struct vm_area_struct *vma = vmf->vma;
1685	struct vfio_pci_core_device *vdev = vma->vm_private_data;
1686	struct vfio_pci_mmap_vma *mmap_vma;
1687	vm_fault_t ret = VM_FAULT_NOPAGE;
1688
1689	mutex_lock(&vdev->vma_lock);
1690	down_read(&vdev->memory_lock);
1691
1692	/*
1693	 * Memory region cannot be accessed if the low power feature is engaged
1694	 * or memory access is disabled.
1695	 */
1696	if (vdev->pm_runtime_engaged || !__vfio_pci_memory_enabled(vdev)) {
1697		ret = VM_FAULT_SIGBUS;
1698		goto up_out;
1699	}
1700
1701	/*
1702	 * We populate the whole vma on fault, so we need to test whether
1703	 * the vma has already been mapped, such as for concurrent faults
1704	 * to the same vma.  io_remap_pfn_range() will trigger a BUG_ON if
1705	 * we ask it to fill the same range again.
1706	 */
1707	list_for_each_entry(mmap_vma, &vdev->vma_list, vma_next) {
1708		if (mmap_vma->vma == vma)
1709			goto up_out;
1710	}
1711
1712	if (io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
1713			       vma->vm_end - vma->vm_start,
1714			       vma->vm_page_prot)) {
1715		ret = VM_FAULT_SIGBUS;
1716		zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
1717		goto up_out;
1718	}
1719
1720	if (__vfio_pci_add_vma(vdev, vma)) {
1721		ret = VM_FAULT_OOM;
1722		zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
1723	}
1724
1725up_out:
1726	up_read(&vdev->memory_lock);
1727	mutex_unlock(&vdev->vma_lock);
1728	return ret;
1729}
1730
1731static const struct vm_operations_struct vfio_pci_mmap_ops = {
1732	.open = vfio_pci_mmap_open,
1733	.close = vfio_pci_mmap_close,
1734	.fault = vfio_pci_mmap_fault,
1735};
1736
1737int vfio_pci_core_mmap(struct vfio_device *core_vdev, struct vm_area_struct *vma)
1738{
1739	struct vfio_pci_core_device *vdev =
1740		container_of(core_vdev, struct vfio_pci_core_device, vdev);
1741	struct pci_dev *pdev = vdev->pdev;
1742	unsigned int index;
1743	u64 phys_len, req_len, pgoff, req_start;
1744	int ret;
1745
1746	index = vma->vm_pgoff >> (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT);
1747
1748	if (index >= VFIO_PCI_NUM_REGIONS + vdev->num_regions)
1749		return -EINVAL;
1750	if (vma->vm_end < vma->vm_start)
1751		return -EINVAL;
1752	if ((vma->vm_flags & VM_SHARED) == 0)
1753		return -EINVAL;
1754	if (index >= VFIO_PCI_NUM_REGIONS) {
1755		int regnum = index - VFIO_PCI_NUM_REGIONS;
1756		struct vfio_pci_region *region = vdev->region + regnum;
1757
1758		if (region->ops && region->ops->mmap &&
1759		    (region->flags & VFIO_REGION_INFO_FLAG_MMAP))
1760			return region->ops->mmap(vdev, region, vma);
1761		return -EINVAL;
1762	}
1763	if (index >= VFIO_PCI_ROM_REGION_INDEX)
1764		return -EINVAL;
1765	if (!vdev->bar_mmap_supported[index])
1766		return -EINVAL;
1767
1768	phys_len = PAGE_ALIGN(pci_resource_len(pdev, index));
1769	req_len = vma->vm_end - vma->vm_start;
1770	pgoff = vma->vm_pgoff &
1771		((1U << (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT)) - 1);
1772	req_start = pgoff << PAGE_SHIFT;
1773
1774	if (req_start + req_len > phys_len)
1775		return -EINVAL;
1776
1777	/*
1778	 * Even though we don't make use of the barmap for the mmap,
1779	 * we need to request the region and the barmap tracks that.
1780	 */
1781	if (!vdev->barmap[index]) {
1782		ret = pci_request_selected_regions(pdev,
1783						   1 << index, "vfio-pci");
1784		if (ret)
1785			return ret;
1786
1787		vdev->barmap[index] = pci_iomap(pdev, index, 0);
1788		if (!vdev->barmap[index]) {
1789			pci_release_selected_regions(pdev, 1 << index);
1790			return -ENOMEM;
1791		}
1792	}
1793
1794	vma->vm_private_data = vdev;
1795	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1796	vma->vm_pgoff = (pci_resource_start(pdev, index) >> PAGE_SHIFT) + pgoff;
1797
1798	/*
1799	 * See remap_pfn_range(), called from vfio_pci_fault() but we can't
1800	 * change vm_flags within the fault handler.  Set them now.
1801	 */
1802	vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1803	vma->vm_ops = &vfio_pci_mmap_ops;
1804
1805	return 0;
1806}
1807EXPORT_SYMBOL_GPL(vfio_pci_core_mmap);
1808
1809void vfio_pci_core_request(struct vfio_device *core_vdev, unsigned int count)
1810{
1811	struct vfio_pci_core_device *vdev =
1812		container_of(core_vdev, struct vfio_pci_core_device, vdev);
1813	struct pci_dev *pdev = vdev->pdev;
1814
1815	mutex_lock(&vdev->igate);
1816
1817	if (vdev->req_trigger) {
1818		if (!(count % 10))
1819			pci_notice_ratelimited(pdev,
1820				"Relaying device request to user (#%u)\n",
1821				count);
1822		eventfd_signal(vdev->req_trigger, 1);
1823	} else if (count == 0) {
1824		pci_warn(pdev,
1825			"No device request channel registered, blocked until released by user\n");
1826	}
1827
1828	mutex_unlock(&vdev->igate);
1829}
1830EXPORT_SYMBOL_GPL(vfio_pci_core_request);
1831
1832static int vfio_pci_validate_vf_token(struct vfio_pci_core_device *vdev,
1833				      bool vf_token, uuid_t *uuid)
1834{
1835	/*
1836	 * There's always some degree of trust or collaboration between SR-IOV
1837	 * PF and VFs, even if just that the PF hosts the SR-IOV capability and
1838	 * can disrupt VFs with a reset, but often the PF has more explicit
1839	 * access to deny service to the VF or access data passed through the
1840	 * VF.  We therefore require an opt-in via a shared VF token (UUID) to
1841	 * represent this trust.  This both prevents that a VF driver might
1842	 * assume the PF driver is a trusted, in-kernel driver, and also that
1843	 * a PF driver might be replaced with a rogue driver, unknown to in-use
1844	 * VF drivers.
1845	 *
1846	 * Therefore when presented with a VF, if the PF is a vfio device and
1847	 * it is bound to the vfio-pci driver, the user needs to provide a VF
1848	 * token to access the device, in the form of appending a vf_token to
1849	 * the device name, for example:
1850	 *
1851	 * "0000:04:10.0 vf_token=bd8d9d2b-5a5f-4f5a-a211-f591514ba1f3"
1852	 *
1853	 * When presented with a PF which has VFs in use, the user must also
1854	 * provide the current VF token to prove collaboration with existing
1855	 * VF users.  If VFs are not in use, the VF token provided for the PF
1856	 * device will act to set the VF token.
1857	 *
1858	 * If the VF token is provided but unused, an error is generated.
1859	 */
1860	if (vdev->pdev->is_virtfn) {
1861		struct vfio_pci_core_device *pf_vdev = vdev->sriov_pf_core_dev;
1862		bool match;
1863
1864		if (!pf_vdev) {
1865			if (!vf_token)
1866				return 0; /* PF is not vfio-pci, no VF token */
1867
1868			pci_info_ratelimited(vdev->pdev,
1869				"VF token incorrectly provided, PF not bound to vfio-pci\n");
1870			return -EINVAL;
1871		}
1872
1873		if (!vf_token) {
1874			pci_info_ratelimited(vdev->pdev,
1875				"VF token required to access device\n");
1876			return -EACCES;
1877		}
1878
1879		mutex_lock(&pf_vdev->vf_token->lock);
1880		match = uuid_equal(uuid, &pf_vdev->vf_token->uuid);
1881		mutex_unlock(&pf_vdev->vf_token->lock);
1882
1883		if (!match) {
1884			pci_info_ratelimited(vdev->pdev,
1885				"Incorrect VF token provided for device\n");
1886			return -EACCES;
1887		}
1888	} else if (vdev->vf_token) {
1889		mutex_lock(&vdev->vf_token->lock);
1890		if (vdev->vf_token->users) {
1891			if (!vf_token) {
1892				mutex_unlock(&vdev->vf_token->lock);
1893				pci_info_ratelimited(vdev->pdev,
1894					"VF token required to access device\n");
1895				return -EACCES;
1896			}
1897
1898			if (!uuid_equal(uuid, &vdev->vf_token->uuid)) {
1899				mutex_unlock(&vdev->vf_token->lock);
1900				pci_info_ratelimited(vdev->pdev,
1901					"Incorrect VF token provided for device\n");
1902				return -EACCES;
1903			}
1904		} else if (vf_token) {
1905			uuid_copy(&vdev->vf_token->uuid, uuid);
1906		}
1907
1908		mutex_unlock(&vdev->vf_token->lock);
1909	} else if (vf_token) {
1910		pci_info_ratelimited(vdev->pdev,
1911			"VF token incorrectly provided, not a PF or VF\n");
1912		return -EINVAL;
1913	}
1914
1915	return 0;
1916}
1917
1918#define VF_TOKEN_ARG "vf_token="
1919
1920int vfio_pci_core_match(struct vfio_device *core_vdev, char *buf)
1921{
1922	struct vfio_pci_core_device *vdev =
1923		container_of(core_vdev, struct vfio_pci_core_device, vdev);
1924	bool vf_token = false;
1925	uuid_t uuid;
1926	int ret;
1927
1928	if (strncmp(pci_name(vdev->pdev), buf, strlen(pci_name(vdev->pdev))))
1929		return 0; /* No match */
1930
1931	if (strlen(buf) > strlen(pci_name(vdev->pdev))) {
1932		buf += strlen(pci_name(vdev->pdev));
1933
1934		if (*buf != ' ')
1935			return 0; /* No match: non-whitespace after name */
1936
1937		while (*buf) {
1938			if (*buf == ' ') {
1939				buf++;
1940				continue;
1941			}
1942
1943			if (!vf_token && !strncmp(buf, VF_TOKEN_ARG,
1944						  strlen(VF_TOKEN_ARG))) {
1945				buf += strlen(VF_TOKEN_ARG);
1946
1947				if (strlen(buf) < UUID_STRING_LEN)
1948					return -EINVAL;
1949
1950				ret = uuid_parse(buf, &uuid);
1951				if (ret)
1952					return ret;
1953
1954				vf_token = true;
1955				buf += UUID_STRING_LEN;
1956			} else {
1957				/* Unknown/duplicate option */
1958				return -EINVAL;
1959			}
1960		}
1961	}
1962
1963	ret = vfio_pci_validate_vf_token(vdev, vf_token, &uuid);
1964	if (ret)
1965		return ret;
1966
1967	return 1; /* Match */
1968}
1969EXPORT_SYMBOL_GPL(vfio_pci_core_match);
1970
1971static int vfio_pci_bus_notifier(struct notifier_block *nb,
1972				 unsigned long action, void *data)
1973{
1974	struct vfio_pci_core_device *vdev = container_of(nb,
1975						    struct vfio_pci_core_device, nb);
1976	struct device *dev = data;
1977	struct pci_dev *pdev = to_pci_dev(dev);
1978	struct pci_dev *physfn = pci_physfn(pdev);
1979
1980	if (action == BUS_NOTIFY_ADD_DEVICE &&
1981	    pdev->is_virtfn && physfn == vdev->pdev) {
1982		pci_info(vdev->pdev, "Captured SR-IOV VF %s driver_override\n",
1983			 pci_name(pdev));
1984		pdev->driver_override = kasprintf(GFP_KERNEL, "%s",
1985						  vdev->vdev.ops->name);
1986	} else if (action == BUS_NOTIFY_BOUND_DRIVER &&
1987		   pdev->is_virtfn && physfn == vdev->pdev) {
1988		struct pci_driver *drv = pci_dev_driver(pdev);
1989
1990		if (drv && drv != pci_dev_driver(vdev->pdev))
1991			pci_warn(vdev->pdev,
1992				 "VF %s bound to driver %s while PF bound to driver %s\n",
1993				 pci_name(pdev), drv->name,
1994				 pci_dev_driver(vdev->pdev)->name);
1995	}
1996
1997	return 0;
1998}
1999
2000static int vfio_pci_vf_init(struct vfio_pci_core_device *vdev)
2001{
2002	struct pci_dev *pdev = vdev->pdev;
2003	struct vfio_pci_core_device *cur;
2004	struct pci_dev *physfn;
2005	int ret;
2006
2007	if (pdev->is_virtfn) {
2008		/*
2009		 * If this VF was created by our vfio_pci_core_sriov_configure()
2010		 * then we can find the PF vfio_pci_core_device now, and due to
2011		 * the locking in pci_disable_sriov() it cannot change until
2012		 * this VF device driver is removed.
2013		 */
2014		physfn = pci_physfn(vdev->pdev);
2015		mutex_lock(&vfio_pci_sriov_pfs_mutex);
2016		list_for_each_entry(cur, &vfio_pci_sriov_pfs, sriov_pfs_item) {
2017			if (cur->pdev == physfn) {
2018				vdev->sriov_pf_core_dev = cur;
2019				break;
2020			}
2021		}
2022		mutex_unlock(&vfio_pci_sriov_pfs_mutex);
2023		return 0;
2024	}
2025
2026	/* Not a SRIOV PF */
2027	if (!pdev->is_physfn)
2028		return 0;
2029
2030	vdev->vf_token = kzalloc(sizeof(*vdev->vf_token), GFP_KERNEL);
2031	if (!vdev->vf_token)
2032		return -ENOMEM;
2033
2034	mutex_init(&vdev->vf_token->lock);
2035	uuid_gen(&vdev->vf_token->uuid);
2036
2037	vdev->nb.notifier_call = vfio_pci_bus_notifier;
2038	ret = bus_register_notifier(&pci_bus_type, &vdev->nb);
2039	if (ret) {
2040		kfree(vdev->vf_token);
2041		return ret;
2042	}
2043	return 0;
2044}
2045
2046static void vfio_pci_vf_uninit(struct vfio_pci_core_device *vdev)
2047{
2048	if (!vdev->vf_token)
2049		return;
2050
2051	bus_unregister_notifier(&pci_bus_type, &vdev->nb);
2052	WARN_ON(vdev->vf_token->users);
2053	mutex_destroy(&vdev->vf_token->lock);
2054	kfree(vdev->vf_token);
2055}
2056
2057static int vfio_pci_vga_init(struct vfio_pci_core_device *vdev)
2058{
2059	struct pci_dev *pdev = vdev->pdev;
2060	int ret;
2061
2062	if (!vfio_pci_is_vga(pdev))
2063		return 0;
2064
2065	ret = aperture_remove_conflicting_pci_devices(pdev, vdev->vdev.ops->name);
2066	if (ret)
2067		return ret;
2068
2069	ret = vga_client_register(pdev, vfio_pci_set_decode);
2070	if (ret)
2071		return ret;
2072	vga_set_legacy_decoding(pdev, vfio_pci_set_decode(pdev, false));
2073	return 0;
2074}
2075
2076static void vfio_pci_vga_uninit(struct vfio_pci_core_device *vdev)
2077{
2078	struct pci_dev *pdev = vdev->pdev;
2079
2080	if (!vfio_pci_is_vga(pdev))
2081		return;
2082	vga_client_unregister(pdev);
2083	vga_set_legacy_decoding(pdev, VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM |
2084					      VGA_RSRC_LEGACY_IO |
2085					      VGA_RSRC_LEGACY_MEM);
2086}
2087
2088int vfio_pci_core_init_dev(struct vfio_device *core_vdev)
2089{
2090	struct vfio_pci_core_device *vdev =
2091		container_of(core_vdev, struct vfio_pci_core_device, vdev);
2092
2093	vdev->pdev = to_pci_dev(core_vdev->dev);
2094	vdev->irq_type = VFIO_PCI_NUM_IRQS;
2095	mutex_init(&vdev->igate);
2096	spin_lock_init(&vdev->irqlock);
2097	mutex_init(&vdev->ioeventfds_lock);
2098	INIT_LIST_HEAD(&vdev->dummy_resources_list);
2099	INIT_LIST_HEAD(&vdev->ioeventfds_list);
2100	mutex_init(&vdev->vma_lock);
2101	INIT_LIST_HEAD(&vdev->vma_list);
2102	INIT_LIST_HEAD(&vdev->sriov_pfs_item);
2103	init_rwsem(&vdev->memory_lock);
 
2104
2105	return 0;
2106}
2107EXPORT_SYMBOL_GPL(vfio_pci_core_init_dev);
2108
2109void vfio_pci_core_release_dev(struct vfio_device *core_vdev)
2110{
2111	struct vfio_pci_core_device *vdev =
2112		container_of(core_vdev, struct vfio_pci_core_device, vdev);
2113
2114	mutex_destroy(&vdev->igate);
2115	mutex_destroy(&vdev->ioeventfds_lock);
2116	mutex_destroy(&vdev->vma_lock);
2117	kfree(vdev->region);
2118	kfree(vdev->pm_save);
2119}
2120EXPORT_SYMBOL_GPL(vfio_pci_core_release_dev);
2121
2122int vfio_pci_core_register_device(struct vfio_pci_core_device *vdev)
2123{
2124	struct pci_dev *pdev = vdev->pdev;
2125	struct device *dev = &pdev->dev;
2126	int ret;
2127
2128	/* Drivers must set the vfio_pci_core_device to their drvdata */
2129	if (WARN_ON(vdev != dev_get_drvdata(dev)))
2130		return -EINVAL;
2131
2132	if (pdev->hdr_type != PCI_HEADER_TYPE_NORMAL)
2133		return -EINVAL;
2134
2135	if (vdev->vdev.mig_ops) {
2136		if (!(vdev->vdev.mig_ops->migration_get_state &&
2137		      vdev->vdev.mig_ops->migration_set_state &&
2138		      vdev->vdev.mig_ops->migration_get_data_size) ||
2139		    !(vdev->vdev.migration_flags & VFIO_MIGRATION_STOP_COPY))
2140			return -EINVAL;
2141	}
2142
2143	if (vdev->vdev.log_ops && !(vdev->vdev.log_ops->log_start &&
2144	    vdev->vdev.log_ops->log_stop &&
2145	    vdev->vdev.log_ops->log_read_and_clear))
2146		return -EINVAL;
2147
2148	/*
2149	 * Prevent binding to PFs with VFs enabled, the VFs might be in use
2150	 * by the host or other users.  We cannot capture the VFs if they
2151	 * already exist, nor can we track VF users.  Disabling SR-IOV here
2152	 * would initiate removing the VFs, which would unbind the driver,
2153	 * which is prone to blocking if that VF is also in use by vfio-pci.
2154	 * Just reject these PFs and let the user sort it out.
2155	 */
2156	if (pci_num_vf(pdev)) {
2157		pci_warn(pdev, "Cannot bind to PF with SR-IOV enabled\n");
2158		return -EBUSY;
2159	}
2160
2161	if (pci_is_root_bus(pdev->bus)) {
2162		ret = vfio_assign_device_set(&vdev->vdev, vdev);
2163	} else if (!pci_probe_reset_slot(pdev->slot)) {
2164		ret = vfio_assign_device_set(&vdev->vdev, pdev->slot);
2165	} else {
2166		/*
2167		 * If there is no slot reset support for this device, the whole
2168		 * bus needs to be grouped together to support bus-wide resets.
2169		 */
2170		ret = vfio_assign_device_set(&vdev->vdev, pdev->bus);
2171	}
2172
2173	if (ret)
2174		return ret;
2175	ret = vfio_pci_vf_init(vdev);
2176	if (ret)
2177		return ret;
2178	ret = vfio_pci_vga_init(vdev);
2179	if (ret)
2180		goto out_vf;
2181
2182	vfio_pci_probe_power_state(vdev);
2183
2184	/*
2185	 * pci-core sets the device power state to an unknown value at
2186	 * bootup and after being removed from a driver.  The only
2187	 * transition it allows from this unknown state is to D0, which
2188	 * typically happens when a driver calls pci_enable_device().
2189	 * We're not ready to enable the device yet, but we do want to
2190	 * be able to get to D3.  Therefore first do a D0 transition
2191	 * before enabling runtime PM.
2192	 */
2193	vfio_pci_set_power_state(vdev, PCI_D0);
2194
2195	dev->driver->pm = &vfio_pci_core_pm_ops;
2196	pm_runtime_allow(dev);
2197	if (!disable_idle_d3)
2198		pm_runtime_put(dev);
2199
2200	ret = vfio_register_group_dev(&vdev->vdev);
2201	if (ret)
2202		goto out_power;
2203	return 0;
2204
2205out_power:
2206	if (!disable_idle_d3)
2207		pm_runtime_get_noresume(dev);
2208
2209	pm_runtime_forbid(dev);
2210out_vf:
2211	vfio_pci_vf_uninit(vdev);
2212	return ret;
2213}
2214EXPORT_SYMBOL_GPL(vfio_pci_core_register_device);
2215
2216void vfio_pci_core_unregister_device(struct vfio_pci_core_device *vdev)
2217{
2218	vfio_pci_core_sriov_configure(vdev, 0);
2219
2220	vfio_unregister_group_dev(&vdev->vdev);
2221
2222	vfio_pci_vf_uninit(vdev);
2223	vfio_pci_vga_uninit(vdev);
2224
2225	if (!disable_idle_d3)
2226		pm_runtime_get_noresume(&vdev->pdev->dev);
2227
2228	pm_runtime_forbid(&vdev->pdev->dev);
2229}
2230EXPORT_SYMBOL_GPL(vfio_pci_core_unregister_device);
2231
2232pci_ers_result_t vfio_pci_core_aer_err_detected(struct pci_dev *pdev,
2233						pci_channel_state_t state)
2234{
2235	struct vfio_pci_core_device *vdev = dev_get_drvdata(&pdev->dev);
2236
2237	mutex_lock(&vdev->igate);
2238
2239	if (vdev->err_trigger)
2240		eventfd_signal(vdev->err_trigger, 1);
2241
2242	mutex_unlock(&vdev->igate);
2243
2244	return PCI_ERS_RESULT_CAN_RECOVER;
2245}
2246EXPORT_SYMBOL_GPL(vfio_pci_core_aer_err_detected);
2247
2248int vfio_pci_core_sriov_configure(struct vfio_pci_core_device *vdev,
2249				  int nr_virtfn)
2250{
2251	struct pci_dev *pdev = vdev->pdev;
2252	int ret = 0;
2253
2254	device_lock_assert(&pdev->dev);
2255
2256	if (nr_virtfn) {
2257		mutex_lock(&vfio_pci_sriov_pfs_mutex);
2258		/*
2259		 * The thread that adds the vdev to the list is the only thread
2260		 * that gets to call pci_enable_sriov() and we will only allow
2261		 * it to be called once without going through
2262		 * pci_disable_sriov()
2263		 */
2264		if (!list_empty(&vdev->sriov_pfs_item)) {
2265			ret = -EINVAL;
2266			goto out_unlock;
2267		}
2268		list_add_tail(&vdev->sriov_pfs_item, &vfio_pci_sriov_pfs);
2269		mutex_unlock(&vfio_pci_sriov_pfs_mutex);
2270
2271		/*
2272		 * The PF power state should always be higher than the VF power
2273		 * state. The PF can be in low power state either with runtime
2274		 * power management (when there is no user) or PCI_PM_CTRL
2275		 * register write by the user. If PF is in the low power state,
2276		 * then change the power state to D0 first before enabling
2277		 * SR-IOV. Also, this function can be called at any time, and
2278		 * userspace PCI_PM_CTRL write can race against this code path,
2279		 * so protect the same with 'memory_lock'.
2280		 */
2281		ret = pm_runtime_resume_and_get(&pdev->dev);
2282		if (ret)
2283			goto out_del;
2284
2285		down_write(&vdev->memory_lock);
2286		vfio_pci_set_power_state(vdev, PCI_D0);
2287		ret = pci_enable_sriov(pdev, nr_virtfn);
2288		up_write(&vdev->memory_lock);
2289		if (ret) {
2290			pm_runtime_put(&pdev->dev);
2291			goto out_del;
2292		}
2293		return nr_virtfn;
2294	}
2295
2296	if (pci_num_vf(pdev)) {
2297		pci_disable_sriov(pdev);
2298		pm_runtime_put(&pdev->dev);
2299	}
2300
2301out_del:
2302	mutex_lock(&vfio_pci_sriov_pfs_mutex);
2303	list_del_init(&vdev->sriov_pfs_item);
2304out_unlock:
2305	mutex_unlock(&vfio_pci_sriov_pfs_mutex);
2306	return ret;
2307}
2308EXPORT_SYMBOL_GPL(vfio_pci_core_sriov_configure);
2309
2310const struct pci_error_handlers vfio_pci_core_err_handlers = {
2311	.error_detected = vfio_pci_core_aer_err_detected,
2312};
2313EXPORT_SYMBOL_GPL(vfio_pci_core_err_handlers);
2314
2315static bool vfio_dev_in_groups(struct vfio_pci_core_device *vdev,
2316			       struct vfio_pci_group_info *groups)
2317{
2318	unsigned int i;
2319
 
 
 
2320	for (i = 0; i < groups->count; i++)
2321		if (vfio_file_has_dev(groups->files[i], &vdev->vdev))
2322			return true;
2323	return false;
2324}
2325
2326static int vfio_pci_is_device_in_set(struct pci_dev *pdev, void *data)
2327{
2328	struct vfio_device_set *dev_set = data;
2329	struct vfio_device *cur;
2330
2331	list_for_each_entry(cur, &dev_set->device_list, dev_set_list)
2332		if (cur->dev == &pdev->dev)
2333			return 0;
2334	return -EBUSY;
2335}
2336
2337/*
2338 * vfio-core considers a group to be viable and will create a vfio_device even
2339 * if some devices are bound to drivers like pci-stub or pcieport. Here we
2340 * require all PCI devices to be inside our dev_set since that ensures they stay
2341 * put and that every driver controlling the device can co-ordinate with the
2342 * device reset.
2343 *
2344 * Returns the pci_dev to pass to pci_reset_bus() if every PCI device to be
2345 * reset is inside the dev_set, and pci_reset_bus() can succeed. NULL otherwise.
2346 */
2347static struct pci_dev *
2348vfio_pci_dev_set_resettable(struct vfio_device_set *dev_set)
2349{
2350	struct pci_dev *pdev;
2351
2352	lockdep_assert_held(&dev_set->lock);
2353
2354	/*
2355	 * By definition all PCI devices in the dev_set share the same PCI
2356	 * reset, so any pci_dev will have the same outcomes for
2357	 * pci_probe_reset_*() and pci_reset_bus().
2358	 */
2359	pdev = list_first_entry(&dev_set->device_list,
2360				struct vfio_pci_core_device,
2361				vdev.dev_set_list)->pdev;
2362
2363	/* pci_reset_bus() is supported */
2364	if (pci_probe_reset_slot(pdev->slot) && pci_probe_reset_bus(pdev->bus))
2365		return NULL;
2366
2367	if (vfio_pci_for_each_slot_or_bus(pdev, vfio_pci_is_device_in_set,
2368					  dev_set,
2369					  !pci_probe_reset_slot(pdev->slot)))
2370		return NULL;
2371	return pdev;
2372}
2373
2374static int vfio_pci_dev_set_pm_runtime_get(struct vfio_device_set *dev_set)
2375{
2376	struct vfio_pci_core_device *cur;
2377	int ret;
2378
2379	list_for_each_entry(cur, &dev_set->device_list, vdev.dev_set_list) {
2380		ret = pm_runtime_resume_and_get(&cur->pdev->dev);
2381		if (ret)
2382			goto unwind;
2383	}
2384
2385	return 0;
2386
2387unwind:
2388	list_for_each_entry_continue_reverse(cur, &dev_set->device_list,
2389					     vdev.dev_set_list)
2390		pm_runtime_put(&cur->pdev->dev);
2391
2392	return ret;
2393}
2394
2395/*
2396 * We need to get memory_lock for each device, but devices can share mmap_lock,
2397 * therefore we need to zap and hold the vma_lock for each device, and only then
2398 * get each memory_lock.
2399 */
2400static int vfio_pci_dev_set_hot_reset(struct vfio_device_set *dev_set,
2401				      struct vfio_pci_group_info *groups)
 
2402{
2403	struct vfio_pci_core_device *cur_mem;
2404	struct vfio_pci_core_device *cur_vma;
2405	struct vfio_pci_core_device *cur;
2406	struct pci_dev *pdev;
2407	bool is_mem = true;
2408	int ret;
2409
2410	mutex_lock(&dev_set->lock);
2411	cur_mem = list_first_entry(&dev_set->device_list,
2412				   struct vfio_pci_core_device,
2413				   vdev.dev_set_list);
2414
2415	pdev = vfio_pci_dev_set_resettable(dev_set);
2416	if (!pdev) {
2417		ret = -EINVAL;
2418		goto err_unlock;
2419	}
2420
2421	/*
2422	 * Some of the devices in the dev_set can be in the runtime suspended
2423	 * state. Increment the usage count for all the devices in the dev_set
2424	 * before reset and decrement the same after reset.
2425	 */
2426	ret = vfio_pci_dev_set_pm_runtime_get(dev_set);
2427	if (ret)
2428		goto err_unlock;
2429
2430	list_for_each_entry(cur_vma, &dev_set->device_list, vdev.dev_set_list) {
 
 
2431		/*
2432		 * Test whether all the affected devices are contained by the
2433		 * set of groups provided by the user.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2434		 */
2435		if (!vfio_dev_in_groups(cur_vma, groups)) {
 
 
 
 
 
 
 
 
 
2436			ret = -EINVAL;
2437			goto err_undo;
2438		}
2439
2440		/*
2441		 * Locking multiple devices is prone to deadlock, runaway and
2442		 * unwind if we hit contention.
2443		 */
2444		if (!vfio_pci_zap_and_vma_lock(cur_vma, true)) {
2445			ret = -EBUSY;
2446			goto err_undo;
2447		}
2448	}
2449	cur_vma = NULL;
2450
2451	list_for_each_entry(cur_mem, &dev_set->device_list, vdev.dev_set_list) {
2452		if (!down_write_trylock(&cur_mem->memory_lock)) {
2453			ret = -EBUSY;
2454			goto err_undo;
2455		}
2456		mutex_unlock(&cur_mem->vma_lock);
2457	}
2458	cur_mem = NULL;
2459
2460	/*
2461	 * The pci_reset_bus() will reset all the devices in the bus.
2462	 * The power state can be non-D0 for some of the devices in the bus.
2463	 * For these devices, the pci_reset_bus() will internally set
2464	 * the power state to D0 without vfio driver involvement.
2465	 * For the devices which have NoSoftRst-, the reset function can
2466	 * cause the PCI config space reset without restoring the original
2467	 * state (saved locally in 'vdev->pm_save').
2468	 */
2469	list_for_each_entry(cur, &dev_set->device_list, vdev.dev_set_list)
2470		vfio_pci_set_power_state(cur, PCI_D0);
2471
2472	ret = pci_reset_bus(pdev);
2473
2474err_undo:
2475	list_for_each_entry(cur, &dev_set->device_list, vdev.dev_set_list) {
2476		if (cur == cur_mem)
2477			is_mem = false;
2478		if (cur == cur_vma)
2479			break;
2480		if (is_mem)
2481			up_write(&cur->memory_lock);
2482		else
2483			mutex_unlock(&cur->vma_lock);
2484	}
2485
2486	list_for_each_entry(cur, &dev_set->device_list, vdev.dev_set_list)
2487		pm_runtime_put(&cur->pdev->dev);
2488err_unlock:
2489	mutex_unlock(&dev_set->lock);
2490	return ret;
2491}
2492
2493static bool vfio_pci_dev_set_needs_reset(struct vfio_device_set *dev_set)
2494{
2495	struct vfio_pci_core_device *cur;
2496	bool needs_reset = false;
2497
2498	/* No other VFIO device in the set can be open. */
2499	if (vfio_device_set_open_count(dev_set) > 1)
2500		return false;
2501
2502	list_for_each_entry(cur, &dev_set->device_list, vdev.dev_set_list)
2503		needs_reset |= cur->needs_reset;
2504	return needs_reset;
2505}
2506
2507/*
2508 * If a bus or slot reset is available for the provided dev_set and:
2509 *  - All of the devices affected by that bus or slot reset are unused
2510 *  - At least one of the affected devices is marked dirty via
2511 *    needs_reset (such as by lack of FLR support)
2512 * Then attempt to perform that bus or slot reset.
2513 */
2514static void vfio_pci_dev_set_try_reset(struct vfio_device_set *dev_set)
2515{
2516	struct vfio_pci_core_device *cur;
2517	struct pci_dev *pdev;
2518	bool reset_done = false;
2519
2520	if (!vfio_pci_dev_set_needs_reset(dev_set))
2521		return;
2522
2523	pdev = vfio_pci_dev_set_resettable(dev_set);
2524	if (!pdev)
2525		return;
2526
2527	/*
2528	 * Some of the devices in the bus 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.
2531	 */
2532	if (!disable_idle_d3 && vfio_pci_dev_set_pm_runtime_get(dev_set))
2533		return;
2534
2535	if (!pci_reset_bus(pdev))
2536		reset_done = true;
2537
2538	list_for_each_entry(cur, &dev_set->device_list, vdev.dev_set_list) {
2539		if (reset_done)
2540			cur->needs_reset = false;
2541
2542		if (!disable_idle_d3)
2543			pm_runtime_put(&cur->pdev->dev);
2544	}
2545}
2546
2547void vfio_pci_core_set_params(bool is_nointxmask, bool is_disable_vga,
2548			      bool is_disable_idle_d3)
2549{
2550	nointxmask = is_nointxmask;
2551	disable_vga = is_disable_vga;
2552	disable_idle_d3 = is_disable_idle_d3;
2553}
2554EXPORT_SYMBOL_GPL(vfio_pci_core_set_params);
2555
2556static void vfio_pci_core_cleanup(void)
2557{
2558	vfio_pci_uninit_perm_bits();
2559}
2560
2561static int __init vfio_pci_core_init(void)
2562{
2563	/* Allocate shared config space permission data used by all devices */
2564	return vfio_pci_init_perm_bits();
2565}
2566
2567module_init(vfio_pci_core_init);
2568module_exit(vfio_pci_core_cleanup);
2569
2570MODULE_LICENSE("GPL v2");
2571MODULE_AUTHOR(DRIVER_AUTHOR);
2572MODULE_DESCRIPTION(DRIVER_DESC);