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/device.h>
  14#include <linux/eventfd.h>
  15#include <linux/file.h>
  16#include <linux/interrupt.h>
  17#include <linux/iommu.h>
  18#include <linux/module.h>
  19#include <linux/mutex.h>
  20#include <linux/notifier.h>
  21#include <linux/pci.h>
  22#include <linux/pm_runtime.h>
  23#include <linux/slab.h>
  24#include <linux/types.h>
  25#include <linux/uaccess.h>
  26#include <linux/vfio.h>
  27#include <linux/vgaarb.h>
  28#include <linux/nospec.h>
  29#include <linux/sched/mm.h>
  30
  31#include "vfio_pci_private.h"
  32
  33#define DRIVER_VERSION  "0.2"
  34#define DRIVER_AUTHOR   "Alex Williamson <alex.williamson@redhat.com>"
  35#define DRIVER_DESC     "VFIO PCI - User Level meta-driver"
  36
  37static char ids[1024] __initdata;
  38module_param_string(ids, ids, sizeof(ids), 0);
  39MODULE_PARM_DESC(ids, "Initial PCI IDs to add to the vfio driver, format is \"vendor:device[:subvendor[:subdevice[:class[:class_mask]]]]\" and multiple comma separated entries can be specified");
  40
  41static bool nointxmask;
  42module_param_named(nointxmask, nointxmask, bool, S_IRUGO | S_IWUSR);
  43MODULE_PARM_DESC(nointxmask,
  44		  "Disable support for PCI 2.3 style INTx masking.  If this resolves problems for specific devices, report lspci -vvvxxx to linux-pci@vger.kernel.org so the device can be fixed automatically via the broken_intx_masking flag.");
  45
  46#ifdef CONFIG_VFIO_PCI_VGA
  47static bool disable_vga;
  48module_param(disable_vga, bool, S_IRUGO);
  49MODULE_PARM_DESC(disable_vga, "Disable VGA resource access through vfio-pci");
  50#endif
  51
  52static bool disable_idle_d3;
  53module_param(disable_idle_d3, bool, S_IRUGO | S_IWUSR);
  54MODULE_PARM_DESC(disable_idle_d3,
  55		 "Disable using the PCI D3 low power state for idle, unused devices");
  56
  57static bool enable_sriov;
  58#ifdef CONFIG_PCI_IOV
  59module_param(enable_sriov, bool, 0644);
  60MODULE_PARM_DESC(enable_sriov, "Enable support for SR-IOV configuration.  Enabling SR-IOV on a PF typically requires support of the userspace PF driver, enabling VFs without such support may result in non-functional VFs or PF.");
  61#endif
  62
  63static bool disable_denylist;
  64module_param(disable_denylist, bool, 0444);
  65MODULE_PARM_DESC(disable_denylist, "Disable use of device denylist. Disabling the denylist allows binding to devices with known errata that may lead to exploitable stability or security issues when accessed by untrusted users.");
  66
  67static inline bool vfio_vga_disabled(void)
  68{
  69#ifdef CONFIG_VFIO_PCI_VGA
  70	return disable_vga;
  71#else
  72	return true;
  73#endif
  74}
  75
  76static bool vfio_pci_dev_in_denylist(struct pci_dev *pdev)
  77{
  78	switch (pdev->vendor) {
  79	case PCI_VENDOR_ID_INTEL:
  80		switch (pdev->device) {
  81		case PCI_DEVICE_ID_INTEL_QAT_C3XXX:
  82		case PCI_DEVICE_ID_INTEL_QAT_C3XXX_VF:
  83		case PCI_DEVICE_ID_INTEL_QAT_C62X:
  84		case PCI_DEVICE_ID_INTEL_QAT_C62X_VF:
  85		case PCI_DEVICE_ID_INTEL_QAT_DH895XCC:
  86		case PCI_DEVICE_ID_INTEL_QAT_DH895XCC_VF:
  87			return true;
  88		default:
  89			return false;
  90		}
  91	}
  92
  93	return false;
  94}
  95
  96static bool vfio_pci_is_denylisted(struct pci_dev *pdev)
  97{
  98	if (!vfio_pci_dev_in_denylist(pdev))
  99		return false;
 100
 101	if (disable_denylist) {
 102		pci_warn(pdev,
 103			 "device denylist disabled - allowing device %04x:%04x.\n",
 104			 pdev->vendor, pdev->device);
 105		return false;
 106	}
 107
 108	pci_warn(pdev, "%04x:%04x exists in vfio-pci device denylist, driver probing disallowed.\n",
 109		 pdev->vendor, pdev->device);
 110
 111	return true;
 112}
 113
 114/*
 115 * Our VGA arbiter participation is limited since we don't know anything
 116 * about the device itself.  However, if the device is the only VGA device
 117 * downstream of a bridge and VFIO VGA support is disabled, then we can
 118 * safely return legacy VGA IO and memory as not decoded since the user
 119 * has no way to get to it and routing can be disabled externally at the
 120 * bridge.
 121 */
 122static unsigned int vfio_pci_set_vga_decode(void *opaque, bool single_vga)
 123{
 124	struct vfio_pci_device *vdev = opaque;
 125	struct pci_dev *tmp = NULL, *pdev = vdev->pdev;
 126	unsigned char max_busnr;
 127	unsigned int decodes;
 128
 129	if (single_vga || !vfio_vga_disabled() || pci_is_root_bus(pdev->bus))
 130		return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM |
 131		       VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM;
 132
 133	max_busnr = pci_bus_max_busnr(pdev->bus);
 134	decodes = VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
 135
 136	while ((tmp = pci_get_class(PCI_CLASS_DISPLAY_VGA << 8, tmp)) != NULL) {
 137		if (tmp == pdev ||
 138		    pci_domain_nr(tmp->bus) != pci_domain_nr(pdev->bus) ||
 139		    pci_is_root_bus(tmp->bus))
 140			continue;
 141
 142		if (tmp->bus->number >= pdev->bus->number &&
 143		    tmp->bus->number <= max_busnr) {
 144			pci_dev_put(tmp);
 145			decodes |= VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM;
 146			break;
 147		}
 148	}
 149
 150	return decodes;
 151}
 152
 153static inline bool vfio_pci_is_vga(struct pci_dev *pdev)
 154{
 155	return (pdev->class >> 8) == PCI_CLASS_DISPLAY_VGA;
 156}
 157
 158static void vfio_pci_probe_mmaps(struct vfio_pci_device *vdev)
 159{
 160	struct resource *res;
 161	int i;
 162	struct vfio_pci_dummy_resource *dummy_res;
 163
 164	INIT_LIST_HEAD(&vdev->dummy_resources_list);
 165
 166	for (i = 0; i < PCI_STD_NUM_BARS; i++) {
 167		int bar = i + PCI_STD_RESOURCES;
 168
 169		res = &vdev->pdev->resource[bar];
 170
 171		if (!IS_ENABLED(CONFIG_VFIO_PCI_MMAP))
 172			goto no_mmap;
 173
 174		if (!(res->flags & IORESOURCE_MEM))
 175			goto no_mmap;
 176
 177		/*
 178		 * The PCI core shouldn't set up a resource with a
 179		 * type but zero size. But there may be bugs that
 180		 * cause us to do that.
 181		 */
 182		if (!resource_size(res))
 183			goto no_mmap;
 184
 185		if (resource_size(res) >= PAGE_SIZE) {
 186			vdev->bar_mmap_supported[bar] = true;
 187			continue;
 188		}
 189
 190		if (!(res->start & ~PAGE_MASK)) {
 191			/*
 192			 * Add a dummy resource to reserve the remainder
 193			 * of the exclusive page in case that hot-add
 194			 * device's bar is assigned into it.
 195			 */
 196			dummy_res = kzalloc(sizeof(*dummy_res), GFP_KERNEL);
 197			if (dummy_res == NULL)
 198				goto no_mmap;
 199
 200			dummy_res->resource.name = "vfio sub-page reserved";
 201			dummy_res->resource.start = res->end + 1;
 202			dummy_res->resource.end = res->start + PAGE_SIZE - 1;
 203			dummy_res->resource.flags = res->flags;
 204			if (request_resource(res->parent,
 205						&dummy_res->resource)) {
 206				kfree(dummy_res);
 207				goto no_mmap;
 208			}
 209			dummy_res->index = bar;
 210			list_add(&dummy_res->res_next,
 211					&vdev->dummy_resources_list);
 212			vdev->bar_mmap_supported[bar] = true;
 213			continue;
 214		}
 215		/*
 216		 * Here we don't handle the case when the BAR is not page
 217		 * aligned because we can't expect the BAR will be
 218		 * assigned into the same location in a page in guest
 219		 * when we passthrough the BAR. And it's hard to access
 220		 * this BAR in userspace because we have no way to get
 221		 * the BAR's location in a page.
 222		 */
 223no_mmap:
 224		vdev->bar_mmap_supported[bar] = false;
 225	}
 226}
 227
 228static void vfio_pci_try_bus_reset(struct vfio_pci_device *vdev);
 229static void vfio_pci_disable(struct vfio_pci_device *vdev);
 230static int vfio_pci_try_zap_and_vma_lock_cb(struct pci_dev *pdev, void *data);
 231
 232/*
 233 * INTx masking requires the ability to disable INTx signaling via PCI_COMMAND
 234 * _and_ the ability detect when the device is asserting INTx via PCI_STATUS.
 235 * If a device implements the former but not the latter we would typically
 236 * expect broken_intx_masking be set and require an exclusive interrupt.
 237 * However since we do have control of the device's ability to assert INTx,
 238 * we can instead pretend that the device does not implement INTx, virtualizing
 239 * the pin register to report zero and maintaining DisINTx set on the host.
 240 */
 241static bool vfio_pci_nointx(struct pci_dev *pdev)
 242{
 243	switch (pdev->vendor) {
 244	case PCI_VENDOR_ID_INTEL:
 245		switch (pdev->device) {
 246		/* All i40e (XL710/X710/XXV710) 10/20/25/40GbE NICs */
 247		case 0x1572:
 248		case 0x1574:
 249		case 0x1580 ... 0x1581:
 250		case 0x1583 ... 0x158b:
 251		case 0x37d0 ... 0x37d2:
 252		/* X550 */
 253		case 0x1563:
 254			return true;
 255		default:
 256			return false;
 257		}
 258	}
 259
 260	return false;
 261}
 262
 263static void vfio_pci_probe_power_state(struct vfio_pci_device *vdev)
 264{
 265	struct pci_dev *pdev = vdev->pdev;
 266	u16 pmcsr;
 267
 268	if (!pdev->pm_cap)
 269		return;
 270
 271	pci_read_config_word(pdev, pdev->pm_cap + PCI_PM_CTRL, &pmcsr);
 272
 273	vdev->needs_pm_restore = !(pmcsr & PCI_PM_CTRL_NO_SOFT_RESET);
 274}
 275
 276/*
 277 * pci_set_power_state() wrapper handling devices which perform a soft reset on
 278 * D3->D0 transition.  Save state prior to D0/1/2->D3, stash it on the vdev,
 279 * restore when returned to D0.  Saved separately from pci_saved_state for use
 280 * by PM capability emulation and separately from pci_dev internal saved state
 281 * to avoid it being overwritten and consumed around other resets.
 282 */
 283int vfio_pci_set_power_state(struct vfio_pci_device *vdev, pci_power_t state)
 284{
 285	struct pci_dev *pdev = vdev->pdev;
 286	bool needs_restore = false, needs_save = false;
 287	int ret;
 288
 289	if (vdev->needs_pm_restore) {
 290		if (pdev->current_state < PCI_D3hot && state >= PCI_D3hot) {
 291			pci_save_state(pdev);
 292			needs_save = true;
 293		}
 294
 295		if (pdev->current_state >= PCI_D3hot && state <= PCI_D0)
 296			needs_restore = true;
 297	}
 298
 299	ret = pci_set_power_state(pdev, state);
 300
 301	if (!ret) {
 302		/* D3 might be unsupported via quirk, skip unless in D3 */
 303		if (needs_save && pdev->current_state >= PCI_D3hot) {
 304			vdev->pm_save = pci_store_saved_state(pdev);
 305		} else if (needs_restore) {
 306			pci_load_and_free_saved_state(pdev, &vdev->pm_save);
 307			pci_restore_state(pdev);
 308		}
 309	}
 310
 311	return ret;
 312}
 313
 314static int vfio_pci_enable(struct vfio_pci_device *vdev)
 315{
 316	struct pci_dev *pdev = vdev->pdev;
 317	int ret;
 318	u16 cmd;
 319	u8 msix_pos;
 320
 321	vfio_pci_set_power_state(vdev, PCI_D0);
 322
 323	/* Don't allow our initial saved state to include busmaster */
 324	pci_clear_master(pdev);
 325
 326	ret = pci_enable_device(pdev);
 327	if (ret)
 328		return ret;
 329
 330	/* If reset fails because of the device lock, fail this path entirely */
 331	ret = pci_try_reset_function(pdev);
 332	if (ret == -EAGAIN) {
 333		pci_disable_device(pdev);
 334		return ret;
 335	}
 336
 337	vdev->reset_works = !ret;
 338	pci_save_state(pdev);
 339	vdev->pci_saved_state = pci_store_saved_state(pdev);
 340	if (!vdev->pci_saved_state)
 341		pci_dbg(pdev, "%s: Couldn't store saved state\n", __func__);
 
 342
 343	if (likely(!nointxmask)) {
 344		if (vfio_pci_nointx(pdev)) {
 345			pci_info(pdev, "Masking broken INTx support\n");
 346			vdev->nointx = true;
 347			pci_intx(pdev, 0);
 348		} else
 349			vdev->pci_2_3 = pci_intx_mask_supported(pdev);
 350	}
 351
 
 
 
 352	pci_read_config_word(pdev, PCI_COMMAND, &cmd);
 353	if (vdev->pci_2_3 && (cmd & PCI_COMMAND_INTX_DISABLE)) {
 354		cmd &= ~PCI_COMMAND_INTX_DISABLE;
 355		pci_write_config_word(pdev, PCI_COMMAND, cmd);
 356	}
 357
 358	ret = vfio_config_init(vdev);
 359	if (ret) {
 360		kfree(vdev->pci_saved_state);
 361		vdev->pci_saved_state = NULL;
 362		pci_disable_device(pdev);
 363		return ret;
 364	}
 365
 366	msix_pos = pdev->msix_cap;
 367	if (msix_pos) {
 368		u16 flags;
 369		u32 table;
 370
 371		pci_read_config_word(pdev, msix_pos + PCI_MSIX_FLAGS, &flags);
 372		pci_read_config_dword(pdev, msix_pos + PCI_MSIX_TABLE, &table);
 373
 374		vdev->msix_bar = table & PCI_MSIX_TABLE_BIR;
 375		vdev->msix_offset = table & PCI_MSIX_TABLE_OFFSET;
 376		vdev->msix_size = ((flags & PCI_MSIX_FLAGS_QSIZE) + 1) * 16;
 377	} else
 378		vdev->msix_bar = 0xFF;
 379
 380	if (!vfio_vga_disabled() && vfio_pci_is_vga(pdev))
 
 381		vdev->has_vga = true;
 382
 383
 384	if (vfio_pci_is_vga(pdev) &&
 385	    pdev->vendor == PCI_VENDOR_ID_INTEL &&
 386	    IS_ENABLED(CONFIG_VFIO_PCI_IGD)) {
 387		ret = vfio_pci_igd_init(vdev);
 388		if (ret) {
 389			pci_warn(pdev, "Failed to setup Intel IGD regions\n");
 390			goto disable_exit;
 391		}
 392	}
 393
 394	if (pdev->vendor == PCI_VENDOR_ID_NVIDIA &&
 395	    IS_ENABLED(CONFIG_VFIO_PCI_NVLINK2)) {
 396		ret = vfio_pci_nvdia_v100_nvlink2_init(vdev);
 397		if (ret && ret != -ENODEV) {
 398			pci_warn(pdev, "Failed to setup NVIDIA NV2 RAM region\n");
 399			goto disable_exit;
 400		}
 401	}
 402
 403	if (pdev->vendor == PCI_VENDOR_ID_IBM &&
 404	    IS_ENABLED(CONFIG_VFIO_PCI_NVLINK2)) {
 405		ret = vfio_pci_ibm_npu2_init(vdev);
 406		if (ret && ret != -ENODEV) {
 407			pci_warn(pdev, "Failed to setup NVIDIA NV2 ATSD region\n");
 408			goto disable_exit;
 409		}
 410	}
 411
 412	vfio_pci_probe_mmaps(vdev);
 413
 414	return 0;
 415
 416disable_exit:
 417	vfio_pci_disable(vdev);
 418	return ret;
 419}
 420
 421static void vfio_pci_disable(struct vfio_pci_device *vdev)
 422{
 423	struct pci_dev *pdev = vdev->pdev;
 424	struct vfio_pci_dummy_resource *dummy_res, *tmp;
 425	struct vfio_pci_ioeventfd *ioeventfd, *ioeventfd_tmp;
 426	int i, bar;
 427
 428	/* Stop the device from further DMA */
 429	pci_clear_master(pdev);
 430
 431	vfio_pci_set_irqs_ioctl(vdev, VFIO_IRQ_SET_DATA_NONE |
 432				VFIO_IRQ_SET_ACTION_TRIGGER,
 433				vdev->irq_type, 0, 0, NULL);
 434
 435	/* Device closed, don't need mutex here */
 436	list_for_each_entry_safe(ioeventfd, ioeventfd_tmp,
 437				 &vdev->ioeventfds_list, next) {
 438		vfio_virqfd_disable(&ioeventfd->virqfd);
 439		list_del(&ioeventfd->next);
 440		kfree(ioeventfd);
 441	}
 442	vdev->ioeventfds_nr = 0;
 443
 444	vdev->virq_disabled = false;
 445
 446	for (i = 0; i < vdev->num_regions; i++)
 447		vdev->region[i].ops->release(vdev, &vdev->region[i]);
 448
 449	vdev->num_regions = 0;
 450	kfree(vdev->region);
 451	vdev->region = NULL; /* don't krealloc a freed pointer */
 452
 453	vfio_config_free(vdev);
 454
 455	for (i = 0; i < PCI_STD_NUM_BARS; i++) {
 456		bar = i + PCI_STD_RESOURCES;
 457		if (!vdev->barmap[bar])
 458			continue;
 459		pci_iounmap(pdev, vdev->barmap[bar]);
 460		pci_release_selected_regions(pdev, 1 << bar);
 461		vdev->barmap[bar] = NULL;
 462	}
 463
 464	list_for_each_entry_safe(dummy_res, tmp,
 465				 &vdev->dummy_resources_list, res_next) {
 466		list_del(&dummy_res->res_next);
 467		release_resource(&dummy_res->resource);
 468		kfree(dummy_res);
 469	}
 470
 471	vdev->needs_reset = true;
 472
 473	/*
 474	 * If we have saved state, restore it.  If we can reset the device,
 475	 * even better.  Resetting with current state seems better than
 476	 * nothing, but saving and restoring current state without reset
 477	 * is just busy work.
 478	 */
 479	if (pci_load_and_free_saved_state(pdev, &vdev->pci_saved_state)) {
 480		pci_info(pdev, "%s: Couldn't reload saved state\n", __func__);
 
 481
 482		if (!vdev->reset_works)
 483			goto out;
 484
 485		pci_save_state(pdev);
 486	}
 487
 488	/*
 489	 * Disable INTx and MSI, presumably to avoid spurious interrupts
 490	 * during reset.  Stolen from pci_reset_function()
 491	 */
 492	pci_write_config_word(pdev, PCI_COMMAND, PCI_COMMAND_INTX_DISABLE);
 493
 494	/*
 495	 * Try to get the locks ourselves to prevent a deadlock. The
 496	 * success of this is dependent on being able to lock the device,
 497	 * which is not always possible.
 498	 * We can not use the "try" reset interface here, which will
 499	 * overwrite the previously restored configuration information.
 500	 */
 501	if (vdev->reset_works && pci_cfg_access_trylock(pdev)) {
 502		if (device_trylock(&pdev->dev)) {
 503			if (!__pci_reset_function_locked(pdev))
 504				vdev->needs_reset = false;
 505			device_unlock(&pdev->dev);
 506		}
 507		pci_cfg_access_unlock(pdev);
 508	}
 509
 510	pci_restore_state(pdev);
 511out:
 512	pci_disable_device(pdev);
 513
 514	vfio_pci_try_bus_reset(vdev);
 515
 516	if (!disable_idle_d3)
 517		vfio_pci_set_power_state(vdev, PCI_D3hot);
 518}
 519
 520static struct pci_driver vfio_pci_driver;
 521
 522static struct vfio_pci_device *get_pf_vdev(struct vfio_pci_device *vdev,
 523					   struct vfio_device **pf_dev)
 524{
 525	struct pci_dev *physfn = pci_physfn(vdev->pdev);
 526
 527	if (!vdev->pdev->is_virtfn)
 528		return NULL;
 529
 530	*pf_dev = vfio_device_get_from_dev(&physfn->dev);
 531	if (!*pf_dev)
 532		return NULL;
 533
 534	if (pci_dev_driver(physfn) != &vfio_pci_driver) {
 535		vfio_device_put(*pf_dev);
 536		return NULL;
 537	}
 538
 539	return vfio_device_data(*pf_dev);
 540}
 541
 542static void vfio_pci_vf_token_user_add(struct vfio_pci_device *vdev, int val)
 543{
 544	struct vfio_device *pf_dev;
 545	struct vfio_pci_device *pf_vdev = get_pf_vdev(vdev, &pf_dev);
 546
 547	if (!pf_vdev)
 548		return;
 549
 550	mutex_lock(&pf_vdev->vf_token->lock);
 551	pf_vdev->vf_token->users += val;
 552	WARN_ON(pf_vdev->vf_token->users < 0);
 553	mutex_unlock(&pf_vdev->vf_token->lock);
 554
 555	vfio_device_put(pf_dev);
 556}
 557
 558static void vfio_pci_release(void *device_data)
 559{
 560	struct vfio_pci_device *vdev = device_data;
 561
 562	mutex_lock(&vdev->reflck->lock);
 563
 564	if (!(--vdev->refcnt)) {
 565		vfio_pci_vf_token_user_add(vdev, -1);
 566		vfio_spapr_pci_eeh_release(vdev->pdev);
 567		vfio_pci_disable(vdev);
 568
 569		mutex_lock(&vdev->igate);
 570		if (vdev->err_trigger) {
 571			eventfd_ctx_put(vdev->err_trigger);
 572			vdev->err_trigger = NULL;
 573		}
 574		if (vdev->req_trigger) {
 575			eventfd_ctx_put(vdev->req_trigger);
 576			vdev->req_trigger = NULL;
 577		}
 578		mutex_unlock(&vdev->igate);
 579	}
 580
 581	mutex_unlock(&vdev->reflck->lock);
 582
 583	module_put(THIS_MODULE);
 584}
 585
 586static int vfio_pci_open(void *device_data)
 587{
 588	struct vfio_pci_device *vdev = device_data;
 589	int ret = 0;
 590
 591	if (!try_module_get(THIS_MODULE))
 592		return -ENODEV;
 593
 594	mutex_lock(&vdev->reflck->lock);
 595
 596	if (!vdev->refcnt) {
 597		ret = vfio_pci_enable(vdev);
 598		if (ret)
 599			goto error;
 600
 601		vfio_spapr_pci_eeh_open(vdev->pdev);
 602		vfio_pci_vf_token_user_add(vdev, 1);
 603	}
 604	vdev->refcnt++;
 605error:
 606	mutex_unlock(&vdev->reflck->lock);
 607	if (ret)
 608		module_put(THIS_MODULE);
 609	return ret;
 610}
 611
 612static int vfio_pci_get_irq_count(struct vfio_pci_device *vdev, int irq_type)
 613{
 614	if (irq_type == VFIO_PCI_INTX_IRQ_INDEX) {
 615		u8 pin;
 616
 617		if (!IS_ENABLED(CONFIG_VFIO_PCI_INTX) ||
 618		    vdev->nointx || vdev->pdev->is_virtfn)
 619			return 0;
 620
 621		pci_read_config_byte(vdev->pdev, PCI_INTERRUPT_PIN, &pin);
 
 
 622
 623		return pin ? 1 : 0;
 624	} else if (irq_type == VFIO_PCI_MSI_IRQ_INDEX) {
 625		u8 pos;
 626		u16 flags;
 627
 628		pos = vdev->pdev->msi_cap;
 629		if (pos) {
 630			pci_read_config_word(vdev->pdev,
 631					     pos + PCI_MSI_FLAGS, &flags);
 632			return 1 << ((flags & PCI_MSI_FLAGS_QMASK) >> 1);
 
 633		}
 634	} else if (irq_type == VFIO_PCI_MSIX_IRQ_INDEX) {
 635		u8 pos;
 636		u16 flags;
 637
 638		pos = vdev->pdev->msix_cap;
 639		if (pos) {
 640			pci_read_config_word(vdev->pdev,
 641					     pos + PCI_MSIX_FLAGS, &flags);
 642
 643			return (flags & PCI_MSIX_FLAGS_QSIZE) + 1;
 644		}
 645	} else if (irq_type == VFIO_PCI_ERR_IRQ_INDEX) {
 646		if (pci_is_pcie(vdev->pdev))
 647			return 1;
 648	} else if (irq_type == VFIO_PCI_REQ_IRQ_INDEX) {
 649		return 1;
 650	}
 651
 652	return 0;
 653}
 654
 655static int vfio_pci_count_devs(struct pci_dev *pdev, void *data)
 656{
 657	(*(int *)data)++;
 658	return 0;
 659}
 660
 661struct vfio_pci_fill_info {
 662	int max;
 663	int cur;
 664	struct vfio_pci_dependent_device *devices;
 665};
 666
 667static int vfio_pci_fill_devs(struct pci_dev *pdev, void *data)
 668{
 669	struct vfio_pci_fill_info *fill = data;
 670	struct iommu_group *iommu_group;
 671
 672	if (fill->cur == fill->max)
 673		return -EAGAIN; /* Something changed, try again */
 674
 675	iommu_group = iommu_group_get(&pdev->dev);
 676	if (!iommu_group)
 677		return -EPERM; /* Cannot reset non-isolated devices */
 678
 679	fill->devices[fill->cur].group_id = iommu_group_id(iommu_group);
 680	fill->devices[fill->cur].segment = pci_domain_nr(pdev->bus);
 681	fill->devices[fill->cur].bus = pdev->bus->number;
 682	fill->devices[fill->cur].devfn = pdev->devfn;
 683	fill->cur++;
 684	iommu_group_put(iommu_group);
 685	return 0;
 686}
 687
 688struct vfio_pci_group_entry {
 689	struct vfio_group *group;
 690	int id;
 691};
 692
 693struct vfio_pci_group_info {
 694	int count;
 695	struct vfio_pci_group_entry *groups;
 696};
 697
 698static int vfio_pci_validate_devs(struct pci_dev *pdev, void *data)
 699{
 700	struct vfio_pci_group_info *info = data;
 701	struct iommu_group *group;
 702	int id, i;
 703
 704	group = iommu_group_get(&pdev->dev);
 705	if (!group)
 706		return -EPERM;
 707
 708	id = iommu_group_id(group);
 709
 710	for (i = 0; i < info->count; i++)
 711		if (info->groups[i].id == id)
 712			break;
 713
 714	iommu_group_put(group);
 715
 716	return (i == info->count) ? -EINVAL : 0;
 717}
 718
 719static bool vfio_pci_dev_below_slot(struct pci_dev *pdev, struct pci_slot *slot)
 720{
 721	for (; pdev; pdev = pdev->bus->self)
 722		if (pdev->bus == slot->bus)
 723			return (pdev->slot == slot);
 724	return false;
 725}
 726
 727struct vfio_pci_walk_info {
 728	int (*fn)(struct pci_dev *, void *data);
 729	void *data;
 730	struct pci_dev *pdev;
 731	bool slot;
 732	int ret;
 733};
 734
 735static int vfio_pci_walk_wrapper(struct pci_dev *pdev, void *data)
 736{
 737	struct vfio_pci_walk_info *walk = data;
 738
 739	if (!walk->slot || vfio_pci_dev_below_slot(pdev, walk->pdev->slot))
 740		walk->ret = walk->fn(pdev, walk->data);
 741
 742	return walk->ret;
 743}
 744
 745static int vfio_pci_for_each_slot_or_bus(struct pci_dev *pdev,
 746					 int (*fn)(struct pci_dev *,
 747						   void *data), void *data,
 748					 bool slot)
 749{
 750	struct vfio_pci_walk_info walk = {
 751		.fn = fn, .data = data, .pdev = pdev, .slot = slot, .ret = 0,
 752	};
 753
 754	pci_walk_bus(pdev->bus, vfio_pci_walk_wrapper, &walk);
 755
 756	return walk.ret;
 757}
 758
 759static int msix_mmappable_cap(struct vfio_pci_device *vdev,
 760			      struct vfio_info_cap *caps)
 761{
 762	struct vfio_info_cap_header header = {
 763		.id = VFIO_REGION_INFO_CAP_MSIX_MAPPABLE,
 764		.version = 1
 765	};
 766
 767	return vfio_info_add_capability(caps, &header, sizeof(header));
 768}
 769
 770int vfio_pci_register_dev_region(struct vfio_pci_device *vdev,
 771				 unsigned int type, unsigned int subtype,
 772				 const struct vfio_pci_regops *ops,
 773				 size_t size, u32 flags, void *data)
 774{
 775	struct vfio_pci_region *region;
 776
 777	region = krealloc(vdev->region,
 778			  (vdev->num_regions + 1) * sizeof(*region),
 779			  GFP_KERNEL);
 780	if (!region)
 781		return -ENOMEM;
 782
 783	vdev->region = region;
 784	vdev->region[vdev->num_regions].type = type;
 785	vdev->region[vdev->num_regions].subtype = subtype;
 786	vdev->region[vdev->num_regions].ops = ops;
 787	vdev->region[vdev->num_regions].size = size;
 788	vdev->region[vdev->num_regions].flags = flags;
 789	vdev->region[vdev->num_regions].data = data;
 790
 791	vdev->num_regions++;
 792
 793	return 0;
 794}
 795
 796struct vfio_devices {
 797	struct vfio_device **devices;
 798	int cur_index;
 799	int max_index;
 800};
 801
 802static long vfio_pci_ioctl(void *device_data,
 803			   unsigned int cmd, unsigned long arg)
 804{
 805	struct vfio_pci_device *vdev = device_data;
 806	unsigned long minsz;
 807
 808	if (cmd == VFIO_DEVICE_GET_INFO) {
 809		struct vfio_device_info info;
 810
 811		minsz = offsetofend(struct vfio_device_info, num_irqs);
 812
 813		if (copy_from_user(&info, (void __user *)arg, minsz))
 814			return -EFAULT;
 815
 816		if (info.argsz < minsz)
 817			return -EINVAL;
 818
 819		info.flags = VFIO_DEVICE_FLAGS_PCI;
 820
 821		if (vdev->reset_works)
 822			info.flags |= VFIO_DEVICE_FLAGS_RESET;
 823
 824		info.num_regions = VFIO_PCI_NUM_REGIONS + vdev->num_regions;
 825		info.num_irqs = VFIO_PCI_NUM_IRQS;
 826
 827		return copy_to_user((void __user *)arg, &info, minsz) ?
 828			-EFAULT : 0;
 829
 830	} else if (cmd == VFIO_DEVICE_GET_REGION_INFO) {
 831		struct pci_dev *pdev = vdev->pdev;
 832		struct vfio_region_info info;
 833		struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
 834		int i, ret;
 835
 836		minsz = offsetofend(struct vfio_region_info, offset);
 837
 838		if (copy_from_user(&info, (void __user *)arg, minsz))
 839			return -EFAULT;
 840
 841		if (info.argsz < minsz)
 842			return -EINVAL;
 843
 844		switch (info.index) {
 845		case VFIO_PCI_CONFIG_REGION_INDEX:
 846			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
 847			info.size = pdev->cfg_size;
 848			info.flags = VFIO_REGION_INFO_FLAG_READ |
 849				     VFIO_REGION_INFO_FLAG_WRITE;
 850			break;
 851		case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
 852			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
 853			info.size = pci_resource_len(pdev, info.index);
 854			if (!info.size) {
 855				info.flags = 0;
 856				break;
 857			}
 858
 859			info.flags = VFIO_REGION_INFO_FLAG_READ |
 860				     VFIO_REGION_INFO_FLAG_WRITE;
 861			if (vdev->bar_mmap_supported[info.index]) {
 
 862				info.flags |= VFIO_REGION_INFO_FLAG_MMAP;
 863				if (info.index == vdev->msix_bar) {
 864					ret = msix_mmappable_cap(vdev, &caps);
 865					if (ret)
 866						return ret;
 867				}
 868			}
 869
 870			break;
 871		case VFIO_PCI_ROM_REGION_INDEX:
 872		{
 873			void __iomem *io;
 874			size_t size;
 875			u16 cmd;
 876
 877			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
 878			info.flags = 0;
 879
 880			/* Report the BAR size, not the ROM size */
 881			info.size = pci_resource_len(pdev, info.index);
 882			if (!info.size) {
 883				/* Shadow ROMs appear as PCI option ROMs */
 884				if (pdev->resource[PCI_ROM_RESOURCE].flags &
 885							IORESOURCE_ROM_SHADOW)
 886					info.size = 0x20000;
 887				else
 888					break;
 889			}
 890
 891			/*
 892			 * Is it really there?  Enable memory decode for
 893			 * implicit access in pci_map_rom().
 894			 */
 895			cmd = vfio_pci_memory_lock_and_enable(vdev);
 896			io = pci_map_rom(pdev, &size);
 897			if (io) {
 898				info.flags = VFIO_REGION_INFO_FLAG_READ;
 899				pci_unmap_rom(pdev, io);
 900			} else {
 901				info.size = 0;
 
 902			}
 903			vfio_pci_memory_unlock_and_restore(vdev, cmd);
 904
 
 905			break;
 906		}
 907		case VFIO_PCI_VGA_REGION_INDEX:
 908			if (!vdev->has_vga)
 909				return -EINVAL;
 910
 911			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
 912			info.size = 0xc0000;
 913			info.flags = VFIO_REGION_INFO_FLAG_READ |
 914				     VFIO_REGION_INFO_FLAG_WRITE;
 915
 916			break;
 917		default:
 918		{
 919			struct vfio_region_info_cap_type cap_type = {
 920					.header.id = VFIO_REGION_INFO_CAP_TYPE,
 921					.header.version = 1 };
 922
 923			if (info.index >=
 924			    VFIO_PCI_NUM_REGIONS + vdev->num_regions)
 925				return -EINVAL;
 926			info.index = array_index_nospec(info.index,
 927							VFIO_PCI_NUM_REGIONS +
 928							vdev->num_regions);
 929
 930			i = info.index - VFIO_PCI_NUM_REGIONS;
 931
 932			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
 933			info.size = vdev->region[i].size;
 934			info.flags = vdev->region[i].flags;
 935
 936			cap_type.type = vdev->region[i].type;
 937			cap_type.subtype = vdev->region[i].subtype;
 938
 939			ret = vfio_info_add_capability(&caps, &cap_type.header,
 940						       sizeof(cap_type));
 941			if (ret)
 942				return ret;
 943
 944			if (vdev->region[i].ops->add_capability) {
 945				ret = vdev->region[i].ops->add_capability(vdev,
 946						&vdev->region[i], &caps);
 947				if (ret)
 948					return ret;
 949			}
 950		}
 951		}
 952
 953		if (caps.size) {
 954			info.flags |= VFIO_REGION_INFO_FLAG_CAPS;
 955			if (info.argsz < sizeof(info) + caps.size) {
 956				info.argsz = sizeof(info) + caps.size;
 957				info.cap_offset = 0;
 958			} else {
 959				vfio_info_cap_shift(&caps, sizeof(info));
 960				if (copy_to_user((void __user *)arg +
 961						  sizeof(info), caps.buf,
 962						  caps.size)) {
 963					kfree(caps.buf);
 964					return -EFAULT;
 965				}
 966				info.cap_offset = sizeof(info);
 967			}
 968
 969			kfree(caps.buf);
 970		}
 971
 972		return copy_to_user((void __user *)arg, &info, minsz) ?
 973			-EFAULT : 0;
 974
 975	} else if (cmd == VFIO_DEVICE_GET_IRQ_INFO) {
 976		struct vfio_irq_info info;
 977
 978		minsz = offsetofend(struct vfio_irq_info, count);
 979
 980		if (copy_from_user(&info, (void __user *)arg, minsz))
 981			return -EFAULT;
 982
 983		if (info.argsz < minsz || info.index >= VFIO_PCI_NUM_IRQS)
 984			return -EINVAL;
 985
 986		switch (info.index) {
 987		case VFIO_PCI_INTX_IRQ_INDEX ... VFIO_PCI_MSIX_IRQ_INDEX:
 988		case VFIO_PCI_REQ_IRQ_INDEX:
 989			break;
 990		case VFIO_PCI_ERR_IRQ_INDEX:
 991			if (pci_is_pcie(vdev->pdev))
 992				break;
 993			fallthrough;
 994		default:
 995			return -EINVAL;
 996		}
 997
 998		info.flags = VFIO_IRQ_INFO_EVENTFD;
 999
1000		info.count = vfio_pci_get_irq_count(vdev, info.index);
1001
1002		if (info.index == VFIO_PCI_INTX_IRQ_INDEX)
1003			info.flags |= (VFIO_IRQ_INFO_MASKABLE |
1004				       VFIO_IRQ_INFO_AUTOMASKED);
1005		else
1006			info.flags |= VFIO_IRQ_INFO_NORESIZE;
1007
1008		return copy_to_user((void __user *)arg, &info, minsz) ?
1009			-EFAULT : 0;
1010
1011	} else if (cmd == VFIO_DEVICE_SET_IRQS) {
1012		struct vfio_irq_set hdr;
1013		u8 *data = NULL;
1014		int max, ret = 0;
1015		size_t data_size = 0;
1016
1017		minsz = offsetofend(struct vfio_irq_set, count);
1018
1019		if (copy_from_user(&hdr, (void __user *)arg, minsz))
1020			return -EFAULT;
1021
1022		max = vfio_pci_get_irq_count(vdev, hdr.index);
 
 
 
1023
1024		ret = vfio_set_irqs_validate_and_prepare(&hdr, max,
1025						 VFIO_PCI_NUM_IRQS, &data_size);
1026		if (ret)
1027			return ret;
 
 
 
 
 
 
 
 
 
 
1028
1029		if (data_size) {
1030			data = memdup_user((void __user *)(arg + minsz),
1031					    data_size);
1032			if (IS_ERR(data))
1033				return PTR_ERR(data);
1034		}
1035
1036		mutex_lock(&vdev->igate);
1037
1038		ret = vfio_pci_set_irqs_ioctl(vdev, hdr.flags, hdr.index,
1039					      hdr.start, hdr.count, data);
1040
1041		mutex_unlock(&vdev->igate);
1042		kfree(data);
1043
1044		return ret;
1045
1046	} else if (cmd == VFIO_DEVICE_RESET) {
1047		int ret;
1048
1049		if (!vdev->reset_works)
1050			return -EINVAL;
1051
1052		vfio_pci_zap_and_down_write_memory_lock(vdev);
1053		ret = pci_try_reset_function(vdev->pdev);
1054		up_write(&vdev->memory_lock);
1055
1056		return ret;
1057
1058	} else if (cmd == VFIO_DEVICE_GET_PCI_HOT_RESET_INFO) {
1059		struct vfio_pci_hot_reset_info hdr;
1060		struct vfio_pci_fill_info fill = { 0 };
1061		struct vfio_pci_dependent_device *devices = NULL;
1062		bool slot = false;
1063		int ret = 0;
1064
1065		minsz = offsetofend(struct vfio_pci_hot_reset_info, count);
1066
1067		if (copy_from_user(&hdr, (void __user *)arg, minsz))
1068			return -EFAULT;
1069
1070		if (hdr.argsz < minsz)
1071			return -EINVAL;
1072
1073		hdr.flags = 0;
1074
1075		/* Can we do a slot or bus reset or neither? */
1076		if (!pci_probe_reset_slot(vdev->pdev->slot))
1077			slot = true;
1078		else if (pci_probe_reset_bus(vdev->pdev->bus))
1079			return -ENODEV;
1080
1081		/* How many devices are affected? */
1082		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1083						    vfio_pci_count_devs,
1084						    &fill.max, slot);
1085		if (ret)
1086			return ret;
1087
1088		WARN_ON(!fill.max); /* Should always be at least one */
1089
1090		/*
1091		 * If there's enough space, fill it now, otherwise return
1092		 * -ENOSPC and the number of devices affected.
1093		 */
1094		if (hdr.argsz < sizeof(hdr) + (fill.max * sizeof(*devices))) {
1095			ret = -ENOSPC;
1096			hdr.count = fill.max;
1097			goto reset_info_exit;
1098		}
1099
1100		devices = kcalloc(fill.max, sizeof(*devices), GFP_KERNEL);
1101		if (!devices)
1102			return -ENOMEM;
1103
1104		fill.devices = devices;
1105
1106		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1107						    vfio_pci_fill_devs,
1108						    &fill, slot);
1109
1110		/*
1111		 * If a device was removed between counting and filling,
1112		 * we may come up short of fill.max.  If a device was
1113		 * added, we'll have a return of -EAGAIN above.
1114		 */
1115		if (!ret)
1116			hdr.count = fill.cur;
1117
1118reset_info_exit:
1119		if (copy_to_user((void __user *)arg, &hdr, minsz))
1120			ret = -EFAULT;
1121
1122		if (!ret) {
1123			if (copy_to_user((void __user *)(arg + minsz), devices,
1124					 hdr.count * sizeof(*devices)))
1125				ret = -EFAULT;
1126		}
1127
1128		kfree(devices);
1129		return ret;
1130
1131	} else if (cmd == VFIO_DEVICE_PCI_HOT_RESET) {
1132		struct vfio_pci_hot_reset hdr;
1133		int32_t *group_fds;
1134		struct vfio_pci_group_entry *groups;
1135		struct vfio_pci_group_info info;
1136		struct vfio_devices devs = { .cur_index = 0 };
1137		bool slot = false;
1138		int i, group_idx, mem_idx = 0, count = 0, ret = 0;
1139
1140		minsz = offsetofend(struct vfio_pci_hot_reset, count);
1141
1142		if (copy_from_user(&hdr, (void __user *)arg, minsz))
1143			return -EFAULT;
1144
1145		if (hdr.argsz < minsz || hdr.flags)
1146			return -EINVAL;
1147
1148		/* Can we do a slot or bus reset or neither? */
1149		if (!pci_probe_reset_slot(vdev->pdev->slot))
1150			slot = true;
1151		else if (pci_probe_reset_bus(vdev->pdev->bus))
1152			return -ENODEV;
1153
1154		/*
1155		 * We can't let userspace give us an arbitrarily large
1156		 * buffer to copy, so verify how many we think there
1157		 * could be.  Note groups can have multiple devices so
1158		 * one group per device is the max.
1159		 */
1160		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1161						    vfio_pci_count_devs,
1162						    &count, slot);
1163		if (ret)
1164			return ret;
1165
1166		/* Somewhere between 1 and count is OK */
1167		if (!hdr.count || hdr.count > count)
1168			return -EINVAL;
1169
1170		group_fds = kcalloc(hdr.count, sizeof(*group_fds), GFP_KERNEL);
1171		groups = kcalloc(hdr.count, sizeof(*groups), GFP_KERNEL);
1172		if (!group_fds || !groups) {
1173			kfree(group_fds);
1174			kfree(groups);
1175			return -ENOMEM;
1176		}
1177
1178		if (copy_from_user(group_fds, (void __user *)(arg + minsz),
1179				   hdr.count * sizeof(*group_fds))) {
1180			kfree(group_fds);
1181			kfree(groups);
1182			return -EFAULT;
1183		}
1184
1185		/*
1186		 * For each group_fd, get the group through the vfio external
1187		 * user interface and store the group and iommu ID.  This
1188		 * ensures the group is held across the reset.
1189		 */
1190		for (group_idx = 0; group_idx < hdr.count; group_idx++) {
1191			struct vfio_group *group;
1192			struct fd f = fdget(group_fds[group_idx]);
1193			if (!f.file) {
1194				ret = -EBADF;
1195				break;
1196			}
1197
1198			group = vfio_group_get_external_user(f.file);
1199			fdput(f);
1200			if (IS_ERR(group)) {
1201				ret = PTR_ERR(group);
1202				break;
1203			}
1204
1205			groups[group_idx].group = group;
1206			groups[group_idx].id =
1207					vfio_external_user_iommu_id(group);
1208		}
1209
1210		kfree(group_fds);
1211
1212		/* release reference to groups on error */
1213		if (ret)
1214			goto hot_reset_release;
1215
1216		info.count = hdr.count;
1217		info.groups = groups;
1218
1219		/*
1220		 * Test whether all the affected devices are contained
1221		 * by the set of groups provided by the user.
1222		 */
1223		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1224						    vfio_pci_validate_devs,
1225						    &info, slot);
1226		if (ret)
1227			goto hot_reset_release;
1228
1229		devs.max_index = count;
1230		devs.devices = kcalloc(count, sizeof(struct vfio_device *),
1231				       GFP_KERNEL);
1232		if (!devs.devices) {
1233			ret = -ENOMEM;
1234			goto hot_reset_release;
1235		}
1236
1237		/*
1238		 * We need to get memory_lock for each device, but devices
1239		 * can share mmap_lock, therefore we need to zap and hold
1240		 * the vma_lock for each device, and only then get each
1241		 * memory_lock.
1242		 */
1243		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1244					    vfio_pci_try_zap_and_vma_lock_cb,
1245					    &devs, slot);
1246		if (ret)
1247			goto hot_reset_release;
1248
1249		for (; mem_idx < devs.cur_index; mem_idx++) {
1250			struct vfio_pci_device *tmp;
1251
1252			tmp = vfio_device_data(devs.devices[mem_idx]);
1253
1254			ret = down_write_trylock(&tmp->memory_lock);
1255			if (!ret) {
1256				ret = -EBUSY;
1257				goto hot_reset_release;
1258			}
1259			mutex_unlock(&tmp->vma_lock);
1260		}
1261
1262		/* User has access, do the reset */
1263		ret = pci_reset_bus(vdev->pdev);
1264
1265hot_reset_release:
1266		for (i = 0; i < devs.cur_index; i++) {
1267			struct vfio_device *device;
1268			struct vfio_pci_device *tmp;
1269
1270			device = devs.devices[i];
1271			tmp = vfio_device_data(device);
1272
1273			if (i < mem_idx)
1274				up_write(&tmp->memory_lock);
1275			else
1276				mutex_unlock(&tmp->vma_lock);
1277			vfio_device_put(device);
1278		}
1279		kfree(devs.devices);
1280
1281		for (group_idx--; group_idx >= 0; group_idx--)
1282			vfio_group_put_external_user(groups[group_idx].group);
1283
1284		kfree(groups);
1285		return ret;
1286	} else if (cmd == VFIO_DEVICE_IOEVENTFD) {
1287		struct vfio_device_ioeventfd ioeventfd;
1288		int count;
1289
1290		minsz = offsetofend(struct vfio_device_ioeventfd, fd);
1291
1292		if (copy_from_user(&ioeventfd, (void __user *)arg, minsz))
1293			return -EFAULT;
1294
1295		if (ioeventfd.argsz < minsz)
1296			return -EINVAL;
1297
1298		if (ioeventfd.flags & ~VFIO_DEVICE_IOEVENTFD_SIZE_MASK)
1299			return -EINVAL;
1300
1301		count = ioeventfd.flags & VFIO_DEVICE_IOEVENTFD_SIZE_MASK;
1302
1303		if (hweight8(count) != 1 || ioeventfd.fd < -1)
1304			return -EINVAL;
1305
1306		return vfio_pci_ioeventfd(vdev, ioeventfd.offset,
1307					  ioeventfd.data, count, ioeventfd.fd);
1308	} else if (cmd == VFIO_DEVICE_FEATURE) {
1309		struct vfio_device_feature feature;
1310		uuid_t uuid;
1311
1312		minsz = offsetofend(struct vfio_device_feature, flags);
1313
1314		if (copy_from_user(&feature, (void __user *)arg, minsz))
1315			return -EFAULT;
1316
1317		if (feature.argsz < minsz)
1318			return -EINVAL;
1319
1320		/* Check unknown flags */
1321		if (feature.flags & ~(VFIO_DEVICE_FEATURE_MASK |
1322				      VFIO_DEVICE_FEATURE_SET |
1323				      VFIO_DEVICE_FEATURE_GET |
1324				      VFIO_DEVICE_FEATURE_PROBE))
1325			return -EINVAL;
1326
1327		/* GET & SET are mutually exclusive except with PROBE */
1328		if (!(feature.flags & VFIO_DEVICE_FEATURE_PROBE) &&
1329		    (feature.flags & VFIO_DEVICE_FEATURE_SET) &&
1330		    (feature.flags & VFIO_DEVICE_FEATURE_GET))
1331			return -EINVAL;
1332
1333		switch (feature.flags & VFIO_DEVICE_FEATURE_MASK) {
1334		case VFIO_DEVICE_FEATURE_PCI_VF_TOKEN:
1335			if (!vdev->vf_token)
1336				return -ENOTTY;
1337
1338			/*
1339			 * We do not support GET of the VF Token UUID as this
1340			 * could expose the token of the previous device user.
1341			 */
1342			if (feature.flags & VFIO_DEVICE_FEATURE_GET)
1343				return -EINVAL;
1344
1345			if (feature.flags & VFIO_DEVICE_FEATURE_PROBE)
1346				return 0;
1347
1348			/* Don't SET unless told to do so */
1349			if (!(feature.flags & VFIO_DEVICE_FEATURE_SET))
1350				return -EINVAL;
1351
1352			if (feature.argsz < minsz + sizeof(uuid))
1353				return -EINVAL;
1354
1355			if (copy_from_user(&uuid, (void __user *)(arg + minsz),
1356					   sizeof(uuid)))
1357				return -EFAULT;
1358
1359			mutex_lock(&vdev->vf_token->lock);
1360			uuid_copy(&vdev->vf_token->uuid, &uuid);
1361			mutex_unlock(&vdev->vf_token->lock);
1362
1363			return 0;
1364		default:
1365			return -ENOTTY;
1366		}
1367	}
1368
1369	return -ENOTTY;
1370}
1371
1372static ssize_t vfio_pci_rw(void *device_data, char __user *buf,
1373			   size_t count, loff_t *ppos, bool iswrite)
1374{
1375	unsigned int index = VFIO_PCI_OFFSET_TO_INDEX(*ppos);
1376	struct vfio_pci_device *vdev = device_data;
1377
1378	if (index >= VFIO_PCI_NUM_REGIONS + vdev->num_regions)
1379		return -EINVAL;
1380
1381	switch (index) {
1382	case VFIO_PCI_CONFIG_REGION_INDEX:
1383		return vfio_pci_config_rw(vdev, buf, count, ppos, iswrite);
1384
1385	case VFIO_PCI_ROM_REGION_INDEX:
1386		if (iswrite)
1387			return -EINVAL;
1388		return vfio_pci_bar_rw(vdev, buf, count, ppos, false);
1389
1390	case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
1391		return vfio_pci_bar_rw(vdev, buf, count, ppos, iswrite);
1392
1393	case VFIO_PCI_VGA_REGION_INDEX:
1394		return vfio_pci_vga_rw(vdev, buf, count, ppos, iswrite);
1395	default:
1396		index -= VFIO_PCI_NUM_REGIONS;
1397		return vdev->region[index].ops->rw(vdev, buf,
1398						   count, ppos, iswrite);
1399	}
1400
1401	return -EINVAL;
1402}
1403
1404static ssize_t vfio_pci_read(void *device_data, char __user *buf,
1405			     size_t count, loff_t *ppos)
1406{
1407	if (!count)
1408		return 0;
1409
1410	return vfio_pci_rw(device_data, buf, count, ppos, false);
1411}
1412
1413static ssize_t vfio_pci_write(void *device_data, const char __user *buf,
1414			      size_t count, loff_t *ppos)
1415{
1416	if (!count)
1417		return 0;
1418
1419	return vfio_pci_rw(device_data, (char __user *)buf, count, ppos, true);
1420}
1421
1422/* Return 1 on zap and vma_lock acquired, 0 on contention (only with @try) */
1423static int vfio_pci_zap_and_vma_lock(struct vfio_pci_device *vdev, bool try)
1424{
1425	struct vfio_pci_mmap_vma *mmap_vma, *tmp;
1426
1427	/*
1428	 * Lock ordering:
1429	 * vma_lock is nested under mmap_lock for vm_ops callback paths.
1430	 * The memory_lock semaphore is used by both code paths calling
1431	 * into this function to zap vmas and the vm_ops.fault callback
1432	 * to protect the memory enable state of the device.
1433	 *
1434	 * When zapping vmas we need to maintain the mmap_lock => vma_lock
1435	 * ordering, which requires using vma_lock to walk vma_list to
1436	 * acquire an mm, then dropping vma_lock to get the mmap_lock and
1437	 * reacquiring vma_lock.  This logic is derived from similar
1438	 * requirements in uverbs_user_mmap_disassociate().
1439	 *
1440	 * mmap_lock must always be the top-level lock when it is taken.
1441	 * Therefore we can only hold the memory_lock write lock when
1442	 * vma_list is empty, as we'd need to take mmap_lock to clear
1443	 * entries.  vma_list can only be guaranteed empty when holding
1444	 * vma_lock, thus memory_lock is nested under vma_lock.
1445	 *
1446	 * This enables the vm_ops.fault callback to acquire vma_lock,
1447	 * followed by memory_lock read lock, while already holding
1448	 * mmap_lock without risk of deadlock.
1449	 */
1450	while (1) {
1451		struct mm_struct *mm = NULL;
1452
1453		if (try) {
1454			if (!mutex_trylock(&vdev->vma_lock))
1455				return 0;
1456		} else {
1457			mutex_lock(&vdev->vma_lock);
1458		}
1459		while (!list_empty(&vdev->vma_list)) {
1460			mmap_vma = list_first_entry(&vdev->vma_list,
1461						    struct vfio_pci_mmap_vma,
1462						    vma_next);
1463			mm = mmap_vma->vma->vm_mm;
1464			if (mmget_not_zero(mm))
1465				break;
1466
1467			list_del(&mmap_vma->vma_next);
1468			kfree(mmap_vma);
1469			mm = NULL;
1470		}
1471		if (!mm)
1472			return 1;
1473		mutex_unlock(&vdev->vma_lock);
1474
1475		if (try) {
1476			if (!mmap_read_trylock(mm)) {
1477				mmput(mm);
1478				return 0;
1479			}
1480		} else {
1481			mmap_read_lock(mm);
1482		}
1483		if (mmget_still_valid(mm)) {
1484			if (try) {
1485				if (!mutex_trylock(&vdev->vma_lock)) {
1486					mmap_read_unlock(mm);
1487					mmput(mm);
1488					return 0;
1489				}
1490			} else {
1491				mutex_lock(&vdev->vma_lock);
1492			}
1493			list_for_each_entry_safe(mmap_vma, tmp,
1494						 &vdev->vma_list, vma_next) {
1495				struct vm_area_struct *vma = mmap_vma->vma;
1496
1497				if (vma->vm_mm != mm)
1498					continue;
1499
1500				list_del(&mmap_vma->vma_next);
1501				kfree(mmap_vma);
1502
1503				zap_vma_ptes(vma, vma->vm_start,
1504					     vma->vm_end - vma->vm_start);
1505			}
1506			mutex_unlock(&vdev->vma_lock);
1507		}
1508		mmap_read_unlock(mm);
1509		mmput(mm);
1510	}
1511}
1512
1513void vfio_pci_zap_and_down_write_memory_lock(struct vfio_pci_device *vdev)
1514{
1515	vfio_pci_zap_and_vma_lock(vdev, false);
1516	down_write(&vdev->memory_lock);
1517	mutex_unlock(&vdev->vma_lock);
1518}
1519
1520u16 vfio_pci_memory_lock_and_enable(struct vfio_pci_device *vdev)
1521{
1522	u16 cmd;
1523
1524	down_write(&vdev->memory_lock);
1525	pci_read_config_word(vdev->pdev, PCI_COMMAND, &cmd);
1526	if (!(cmd & PCI_COMMAND_MEMORY))
1527		pci_write_config_word(vdev->pdev, PCI_COMMAND,
1528				      cmd | PCI_COMMAND_MEMORY);
1529
1530	return cmd;
1531}
1532
1533void vfio_pci_memory_unlock_and_restore(struct vfio_pci_device *vdev, u16 cmd)
1534{
1535	pci_write_config_word(vdev->pdev, PCI_COMMAND, cmd);
1536	up_write(&vdev->memory_lock);
1537}
1538
1539/* Caller holds vma_lock */
1540static int __vfio_pci_add_vma(struct vfio_pci_device *vdev,
1541			      struct vm_area_struct *vma)
1542{
1543	struct vfio_pci_mmap_vma *mmap_vma;
1544
1545	mmap_vma = kmalloc(sizeof(*mmap_vma), GFP_KERNEL);
1546	if (!mmap_vma)
1547		return -ENOMEM;
1548
1549	mmap_vma->vma = vma;
1550	list_add(&mmap_vma->vma_next, &vdev->vma_list);
1551
1552	return 0;
1553}
1554
1555/*
1556 * Zap mmaps on open so that we can fault them in on access and therefore
1557 * our vma_list only tracks mappings accessed since last zap.
1558 */
1559static void vfio_pci_mmap_open(struct vm_area_struct *vma)
1560{
1561	zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
1562}
1563
1564static void vfio_pci_mmap_close(struct vm_area_struct *vma)
1565{
1566	struct vfio_pci_device *vdev = vma->vm_private_data;
1567	struct vfio_pci_mmap_vma *mmap_vma;
1568
1569	mutex_lock(&vdev->vma_lock);
1570	list_for_each_entry(mmap_vma, &vdev->vma_list, vma_next) {
1571		if (mmap_vma->vma == vma) {
1572			list_del(&mmap_vma->vma_next);
1573			kfree(mmap_vma);
1574			break;
1575		}
1576	}
1577	mutex_unlock(&vdev->vma_lock);
1578}
1579
1580static vm_fault_t vfio_pci_mmap_fault(struct vm_fault *vmf)
1581{
1582	struct vm_area_struct *vma = vmf->vma;
1583	struct vfio_pci_device *vdev = vma->vm_private_data;
1584	vm_fault_t ret = VM_FAULT_NOPAGE;
1585
1586	mutex_lock(&vdev->vma_lock);
1587	down_read(&vdev->memory_lock);
1588
1589	if (!__vfio_pci_memory_enabled(vdev)) {
1590		ret = VM_FAULT_SIGBUS;
1591		mutex_unlock(&vdev->vma_lock);
1592		goto up_out;
1593	}
1594
1595	if (__vfio_pci_add_vma(vdev, vma)) {
1596		ret = VM_FAULT_OOM;
1597		mutex_unlock(&vdev->vma_lock);
1598		goto up_out;
1599	}
1600
1601	mutex_unlock(&vdev->vma_lock);
1602
1603	if (remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
1604			    vma->vm_end - vma->vm_start, vma->vm_page_prot))
1605		ret = VM_FAULT_SIGBUS;
1606
1607up_out:
1608	up_read(&vdev->memory_lock);
1609	return ret;
1610}
1611
1612static const struct vm_operations_struct vfio_pci_mmap_ops = {
1613	.open = vfio_pci_mmap_open,
1614	.close = vfio_pci_mmap_close,
1615	.fault = vfio_pci_mmap_fault,
1616};
1617
1618static int vfio_pci_mmap(void *device_data, struct vm_area_struct *vma)
1619{
1620	struct vfio_pci_device *vdev = device_data;
1621	struct pci_dev *pdev = vdev->pdev;
1622	unsigned int index;
1623	u64 phys_len, req_len, pgoff, req_start;
1624	int ret;
1625
1626	index = vma->vm_pgoff >> (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT);
1627
1628	if (vma->vm_end < vma->vm_start)
1629		return -EINVAL;
1630	if ((vma->vm_flags & VM_SHARED) == 0)
1631		return -EINVAL;
1632	if (index >= VFIO_PCI_NUM_REGIONS) {
1633		int regnum = index - VFIO_PCI_NUM_REGIONS;
1634		struct vfio_pci_region *region = vdev->region + regnum;
1635
1636		if (region && region->ops && region->ops->mmap &&
1637		    (region->flags & VFIO_REGION_INFO_FLAG_MMAP))
1638			return region->ops->mmap(vdev, region, vma);
1639		return -EINVAL;
1640	}
1641	if (index >= VFIO_PCI_ROM_REGION_INDEX)
1642		return -EINVAL;
1643	if (!vdev->bar_mmap_supported[index])
1644		return -EINVAL;
1645
1646	phys_len = PAGE_ALIGN(pci_resource_len(pdev, index));
1647	req_len = vma->vm_end - vma->vm_start;
1648	pgoff = vma->vm_pgoff &
1649		((1U << (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT)) - 1);
1650	req_start = pgoff << PAGE_SHIFT;
1651
1652	if (req_start + req_len > phys_len)
1653		return -EINVAL;
1654
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1655	/*
1656	 * Even though we don't make use of the barmap for the mmap,
1657	 * we need to request the region and the barmap tracks that.
1658	 */
1659	if (!vdev->barmap[index]) {
1660		ret = pci_request_selected_regions(pdev,
1661						   1 << index, "vfio-pci");
1662		if (ret)
1663			return ret;
1664
1665		vdev->barmap[index] = pci_iomap(pdev, index, 0);
1666		if (!vdev->barmap[index]) {
1667			pci_release_selected_regions(pdev, 1 << index);
1668			return -ENOMEM;
1669		}
1670	}
1671
1672	vma->vm_private_data = vdev;
1673	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1674	vma->vm_pgoff = (pci_resource_start(pdev, index) >> PAGE_SHIFT) + pgoff;
1675
1676	/*
1677	 * See remap_pfn_range(), called from vfio_pci_fault() but we can't
1678	 * change vm_flags within the fault handler.  Set them now.
1679	 */
1680	vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1681	vma->vm_ops = &vfio_pci_mmap_ops;
1682
1683	return 0;
1684}
1685
1686static void vfio_pci_request(void *device_data, unsigned int count)
1687{
1688	struct vfio_pci_device *vdev = device_data;
1689	struct pci_dev *pdev = vdev->pdev;
1690
1691	mutex_lock(&vdev->igate);
1692
1693	if (vdev->req_trigger) {
1694		if (!(count % 10))
1695			pci_notice_ratelimited(pdev,
1696				"Relaying device request to user (#%u)\n",
1697				count);
1698		eventfd_signal(vdev->req_trigger, 1);
1699	} else if (count == 0) {
1700		pci_warn(pdev,
1701			"No device request channel registered, blocked until released by user\n");
1702	}
1703
1704	mutex_unlock(&vdev->igate);
1705}
1706
1707static int vfio_pci_validate_vf_token(struct vfio_pci_device *vdev,
1708				      bool vf_token, uuid_t *uuid)
1709{
1710	/*
1711	 * There's always some degree of trust or collaboration between SR-IOV
1712	 * PF and VFs, even if just that the PF hosts the SR-IOV capability and
1713	 * can disrupt VFs with a reset, but often the PF has more explicit
1714	 * access to deny service to the VF or access data passed through the
1715	 * VF.  We therefore require an opt-in via a shared VF token (UUID) to
1716	 * represent this trust.  This both prevents that a VF driver might
1717	 * assume the PF driver is a trusted, in-kernel driver, and also that
1718	 * a PF driver might be replaced with a rogue driver, unknown to in-use
1719	 * VF drivers.
1720	 *
1721	 * Therefore when presented with a VF, if the PF is a vfio device and
1722	 * it is bound to the vfio-pci driver, the user needs to provide a VF
1723	 * token to access the device, in the form of appending a vf_token to
1724	 * the device name, for example:
1725	 *
1726	 * "0000:04:10.0 vf_token=bd8d9d2b-5a5f-4f5a-a211-f591514ba1f3"
1727	 *
1728	 * When presented with a PF which has VFs in use, the user must also
1729	 * provide the current VF token to prove collaboration with existing
1730	 * VF users.  If VFs are not in use, the VF token provided for the PF
1731	 * device will act to set the VF token.
1732	 *
1733	 * If the VF token is provided but unused, an error is generated.
1734	 */
1735	if (!vdev->pdev->is_virtfn && !vdev->vf_token && !vf_token)
1736		return 0; /* No VF token provided or required */
1737
1738	if (vdev->pdev->is_virtfn) {
1739		struct vfio_device *pf_dev;
1740		struct vfio_pci_device *pf_vdev = get_pf_vdev(vdev, &pf_dev);
1741		bool match;
1742
1743		if (!pf_vdev) {
1744			if (!vf_token)
1745				return 0; /* PF is not vfio-pci, no VF token */
1746
1747			pci_info_ratelimited(vdev->pdev,
1748				"VF token incorrectly provided, PF not bound to vfio-pci\n");
1749			return -EINVAL;
1750		}
1751
1752		if (!vf_token) {
1753			vfio_device_put(pf_dev);
1754			pci_info_ratelimited(vdev->pdev,
1755				"VF token required to access device\n");
1756			return -EACCES;
1757		}
1758
1759		mutex_lock(&pf_vdev->vf_token->lock);
1760		match = uuid_equal(uuid, &pf_vdev->vf_token->uuid);
1761		mutex_unlock(&pf_vdev->vf_token->lock);
1762
1763		vfio_device_put(pf_dev);
1764
1765		if (!match) {
1766			pci_info_ratelimited(vdev->pdev,
1767				"Incorrect VF token provided for device\n");
1768			return -EACCES;
1769		}
1770	} else if (vdev->vf_token) {
1771		mutex_lock(&vdev->vf_token->lock);
1772		if (vdev->vf_token->users) {
1773			if (!vf_token) {
1774				mutex_unlock(&vdev->vf_token->lock);
1775				pci_info_ratelimited(vdev->pdev,
1776					"VF token required to access device\n");
1777				return -EACCES;
1778			}
1779
1780			if (!uuid_equal(uuid, &vdev->vf_token->uuid)) {
1781				mutex_unlock(&vdev->vf_token->lock);
1782				pci_info_ratelimited(vdev->pdev,
1783					"Incorrect VF token provided for device\n");
1784				return -EACCES;
1785			}
1786		} else if (vf_token) {
1787			uuid_copy(&vdev->vf_token->uuid, uuid);
1788		}
1789
1790		mutex_unlock(&vdev->vf_token->lock);
1791	} else if (vf_token) {
1792		pci_info_ratelimited(vdev->pdev,
1793			"VF token incorrectly provided, not a PF or VF\n");
1794		return -EINVAL;
1795	}
1796
1797	return 0;
1798}
1799
1800#define VF_TOKEN_ARG "vf_token="
1801
1802static int vfio_pci_match(void *device_data, char *buf)
1803{
1804	struct vfio_pci_device *vdev = device_data;
1805	bool vf_token = false;
1806	uuid_t uuid;
1807	int ret;
1808
1809	if (strncmp(pci_name(vdev->pdev), buf, strlen(pci_name(vdev->pdev))))
1810		return 0; /* No match */
1811
1812	if (strlen(buf) > strlen(pci_name(vdev->pdev))) {
1813		buf += strlen(pci_name(vdev->pdev));
1814
1815		if (*buf != ' ')
1816			return 0; /* No match: non-whitespace after name */
1817
1818		while (*buf) {
1819			if (*buf == ' ') {
1820				buf++;
1821				continue;
1822			}
1823
1824			if (!vf_token && !strncmp(buf, VF_TOKEN_ARG,
1825						  strlen(VF_TOKEN_ARG))) {
1826				buf += strlen(VF_TOKEN_ARG);
1827
1828				if (strlen(buf) < UUID_STRING_LEN)
1829					return -EINVAL;
1830
1831				ret = uuid_parse(buf, &uuid);
1832				if (ret)
1833					return ret;
1834
1835				vf_token = true;
1836				buf += UUID_STRING_LEN;
1837			} else {
1838				/* Unknown/duplicate option */
1839				return -EINVAL;
1840			}
1841		}
1842	}
1843
1844	ret = vfio_pci_validate_vf_token(vdev, vf_token, &uuid);
1845	if (ret)
1846		return ret;
1847
1848	return 1; /* Match */
1849}
1850
1851static const struct vfio_device_ops vfio_pci_ops = {
1852	.name		= "vfio-pci",
1853	.open		= vfio_pci_open,
1854	.release	= vfio_pci_release,
1855	.ioctl		= vfio_pci_ioctl,
1856	.read		= vfio_pci_read,
1857	.write		= vfio_pci_write,
1858	.mmap		= vfio_pci_mmap,
1859	.request	= vfio_pci_request,
1860	.match		= vfio_pci_match,
1861};
1862
1863static int vfio_pci_reflck_attach(struct vfio_pci_device *vdev);
1864static void vfio_pci_reflck_put(struct vfio_pci_reflck *reflck);
1865static struct pci_driver vfio_pci_driver;
1866
1867static int vfio_pci_bus_notifier(struct notifier_block *nb,
1868				 unsigned long action, void *data)
1869{
1870	struct vfio_pci_device *vdev = container_of(nb,
1871						    struct vfio_pci_device, nb);
1872	struct device *dev = data;
1873	struct pci_dev *pdev = to_pci_dev(dev);
1874	struct pci_dev *physfn = pci_physfn(pdev);
1875
1876	if (action == BUS_NOTIFY_ADD_DEVICE &&
1877	    pdev->is_virtfn && physfn == vdev->pdev) {
1878		pci_info(vdev->pdev, "Captured SR-IOV VF %s driver_override\n",
1879			 pci_name(pdev));
1880		pdev->driver_override = kasprintf(GFP_KERNEL, "%s",
1881						  vfio_pci_ops.name);
1882	} else if (action == BUS_NOTIFY_BOUND_DRIVER &&
1883		   pdev->is_virtfn && physfn == vdev->pdev) {
1884		struct pci_driver *drv = pci_dev_driver(pdev);
1885
1886		if (drv && drv != &vfio_pci_driver)
1887			pci_warn(vdev->pdev,
1888				 "VF %s bound to driver %s while PF bound to vfio-pci\n",
1889				 pci_name(pdev), drv->name);
1890	}
1891
1892	return 0;
1893}
1894
1895static int vfio_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
1896{
 
1897	struct vfio_pci_device *vdev;
1898	struct iommu_group *group;
1899	int ret;
1900
1901	if (vfio_pci_is_denylisted(pdev))
1902		return -EINVAL;
1903
1904	if (pdev->hdr_type != PCI_HEADER_TYPE_NORMAL)
1905		return -EINVAL;
1906
1907	/*
1908	 * Prevent binding to PFs with VFs enabled, the VFs might be in use
1909	 * by the host or other users.  We cannot capture the VFs if they
1910	 * already exist, nor can we track VF users.  Disabling SR-IOV here
1911	 * would initiate removing the VFs, which would unbind the driver,
1912	 * which is prone to blocking if that VF is also in use by vfio-pci.
1913	 * Just reject these PFs and let the user sort it out.
1914	 */
1915	if (pci_num_vf(pdev)) {
1916		pci_warn(pdev, "Cannot bind to PF with SR-IOV enabled\n");
1917		return -EBUSY;
1918	}
1919
1920	group = vfio_iommu_group_get(&pdev->dev);
1921	if (!group)
1922		return -EINVAL;
1923
1924	vdev = kzalloc(sizeof(*vdev), GFP_KERNEL);
1925	if (!vdev) {
1926		ret = -ENOMEM;
1927		goto out_group_put;
1928	}
1929
1930	vdev->pdev = pdev;
1931	vdev->irq_type = VFIO_PCI_NUM_IRQS;
1932	mutex_init(&vdev->igate);
1933	spin_lock_init(&vdev->irqlock);
1934	mutex_init(&vdev->ioeventfds_lock);
1935	INIT_LIST_HEAD(&vdev->ioeventfds_list);
1936	mutex_init(&vdev->vma_lock);
1937	INIT_LIST_HEAD(&vdev->vma_list);
1938	init_rwsem(&vdev->memory_lock);
1939
1940	ret = vfio_add_group_dev(&pdev->dev, &vfio_pci_ops, vdev);
1941	if (ret)
1942		goto out_free;
1943
1944	ret = vfio_pci_reflck_attach(vdev);
1945	if (ret)
1946		goto out_del_group_dev;
1947
1948	if (pdev->is_physfn) {
1949		vdev->vf_token = kzalloc(sizeof(*vdev->vf_token), GFP_KERNEL);
1950		if (!vdev->vf_token) {
1951			ret = -ENOMEM;
1952			goto out_reflck;
1953		}
1954
1955		mutex_init(&vdev->vf_token->lock);
1956		uuid_gen(&vdev->vf_token->uuid);
1957
1958		vdev->nb.notifier_call = vfio_pci_bus_notifier;
1959		ret = bus_register_notifier(&pci_bus_type, &vdev->nb);
1960		if (ret)
1961			goto out_vf_token;
1962	}
1963
1964	if (vfio_pci_is_vga(pdev)) {
1965		vga_client_register(pdev, vdev, NULL, vfio_pci_set_vga_decode);
1966		vga_set_legacy_decoding(pdev,
1967					vfio_pci_set_vga_decode(vdev, false));
1968	}
1969
1970	vfio_pci_probe_power_state(vdev);
1971
1972	if (!disable_idle_d3) {
1973		/*
1974		 * pci-core sets the device power state to an unknown value at
1975		 * bootup and after being removed from a driver.  The only
1976		 * transition it allows from this unknown state is to D0, which
1977		 * typically happens when a driver calls pci_enable_device().
1978		 * We're not ready to enable the device yet, but we do want to
1979		 * be able to get to D3.  Therefore first do a D0 transition
1980		 * before going to D3.
1981		 */
1982		vfio_pci_set_power_state(vdev, PCI_D0);
1983		vfio_pci_set_power_state(vdev, PCI_D3hot);
1984	}
1985
1986	return ret;
1987
1988out_vf_token:
1989	kfree(vdev->vf_token);
1990out_reflck:
1991	vfio_pci_reflck_put(vdev->reflck);
1992out_del_group_dev:
1993	vfio_del_group_dev(&pdev->dev);
1994out_free:
1995	kfree(vdev);
1996out_group_put:
1997	vfio_iommu_group_put(group, &pdev->dev);
1998	return ret;
1999}
2000
2001static void vfio_pci_remove(struct pci_dev *pdev)
2002{
2003	struct vfio_pci_device *vdev;
2004
2005	pci_disable_sriov(pdev);
2006
2007	vdev = vfio_del_group_dev(&pdev->dev);
2008	if (!vdev)
2009		return;
2010
2011	if (vdev->vf_token) {
2012		WARN_ON(vdev->vf_token->users);
2013		mutex_destroy(&vdev->vf_token->lock);
2014		kfree(vdev->vf_token);
2015	}
2016
2017	if (vdev->nb.notifier_call)
2018		bus_unregister_notifier(&pci_bus_type, &vdev->nb);
2019
2020	vfio_pci_reflck_put(vdev->reflck);
2021
2022	vfio_iommu_group_put(pdev->dev.iommu_group, &pdev->dev);
2023	kfree(vdev->region);
2024	mutex_destroy(&vdev->ioeventfds_lock);
2025
2026	if (!disable_idle_d3)
2027		vfio_pci_set_power_state(vdev, PCI_D0);
2028
2029	kfree(vdev->pm_save);
2030	kfree(vdev);
2031
2032	if (vfio_pci_is_vga(pdev)) {
2033		vga_client_register(pdev, NULL, NULL, NULL);
2034		vga_set_legacy_decoding(pdev,
2035				VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM |
2036				VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM);
2037	}
2038}
2039
2040static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
2041						  pci_channel_state_t state)
2042{
2043	struct vfio_pci_device *vdev;
2044	struct vfio_device *device;
2045
2046	device = vfio_device_get_from_dev(&pdev->dev);
2047	if (device == NULL)
2048		return PCI_ERS_RESULT_DISCONNECT;
2049
2050	vdev = vfio_device_data(device);
2051	if (vdev == NULL) {
2052		vfio_device_put(device);
2053		return PCI_ERS_RESULT_DISCONNECT;
2054	}
2055
2056	mutex_lock(&vdev->igate);
2057
2058	if (vdev->err_trigger)
2059		eventfd_signal(vdev->err_trigger, 1);
2060
2061	mutex_unlock(&vdev->igate);
2062
2063	vfio_device_put(device);
2064
2065	return PCI_ERS_RESULT_CAN_RECOVER;
2066}
2067
2068static int vfio_pci_sriov_configure(struct pci_dev *pdev, int nr_virtfn)
2069{
2070	struct vfio_pci_device *vdev;
2071	struct vfio_device *device;
2072	int ret = 0;
2073
2074	might_sleep();
2075
2076	if (!enable_sriov)
2077		return -ENOENT;
2078
2079	device = vfio_device_get_from_dev(&pdev->dev);
2080	if (!device)
2081		return -ENODEV;
2082
2083	vdev = vfio_device_data(device);
2084	if (!vdev) {
2085		vfio_device_put(device);
2086		return -ENODEV;
2087	}
2088
2089	if (nr_virtfn == 0)
2090		pci_disable_sriov(pdev);
2091	else
2092		ret = pci_enable_sriov(pdev, nr_virtfn);
2093
2094	vfio_device_put(device);
2095
2096	return ret < 0 ? ret : nr_virtfn;
2097}
2098
2099static const struct pci_error_handlers vfio_err_handlers = {
2100	.error_detected = vfio_pci_aer_err_detected,
2101};
2102
2103static struct pci_driver vfio_pci_driver = {
2104	.name			= "vfio-pci",
2105	.id_table		= NULL, /* only dynamic ids */
2106	.probe			= vfio_pci_probe,
2107	.remove			= vfio_pci_remove,
2108	.sriov_configure	= vfio_pci_sriov_configure,
2109	.err_handler		= &vfio_err_handlers,
2110};
2111
2112static DEFINE_MUTEX(reflck_lock);
2113
2114static struct vfio_pci_reflck *vfio_pci_reflck_alloc(void)
2115{
2116	struct vfio_pci_reflck *reflck;
2117
2118	reflck = kzalloc(sizeof(*reflck), GFP_KERNEL);
2119	if (!reflck)
2120		return ERR_PTR(-ENOMEM);
2121
2122	kref_init(&reflck->kref);
2123	mutex_init(&reflck->lock);
2124
2125	return reflck;
2126}
2127
2128static void vfio_pci_reflck_get(struct vfio_pci_reflck *reflck)
2129{
2130	kref_get(&reflck->kref);
2131}
2132
2133static int vfio_pci_reflck_find(struct pci_dev *pdev, void *data)
2134{
2135	struct vfio_pci_reflck **preflck = data;
2136	struct vfio_device *device;
2137	struct vfio_pci_device *vdev;
2138
2139	device = vfio_device_get_from_dev(&pdev->dev);
2140	if (!device)
2141		return 0;
2142
2143	if (pci_dev_driver(pdev) != &vfio_pci_driver) {
2144		vfio_device_put(device);
2145		return 0;
2146	}
2147
2148	vdev = vfio_device_data(device);
2149
2150	if (vdev->reflck) {
2151		vfio_pci_reflck_get(vdev->reflck);
2152		*preflck = vdev->reflck;
2153		vfio_device_put(device);
2154		return 1;
2155	}
2156
2157	vfio_device_put(device);
2158	return 0;
2159}
2160
2161static int vfio_pci_reflck_attach(struct vfio_pci_device *vdev)
2162{
2163	bool slot = !pci_probe_reset_slot(vdev->pdev->slot);
2164
2165	mutex_lock(&reflck_lock);
2166
2167	if (pci_is_root_bus(vdev->pdev->bus) ||
2168	    vfio_pci_for_each_slot_or_bus(vdev->pdev, vfio_pci_reflck_find,
2169					  &vdev->reflck, slot) <= 0)
2170		vdev->reflck = vfio_pci_reflck_alloc();
2171
2172	mutex_unlock(&reflck_lock);
2173
2174	return PTR_ERR_OR_ZERO(vdev->reflck);
2175}
2176
2177static void vfio_pci_reflck_release(struct kref *kref)
2178{
2179	struct vfio_pci_reflck *reflck = container_of(kref,
2180						      struct vfio_pci_reflck,
2181						      kref);
2182
2183	kfree(reflck);
2184	mutex_unlock(&reflck_lock);
2185}
2186
2187static void vfio_pci_reflck_put(struct vfio_pci_reflck *reflck)
2188{
2189	kref_put_mutex(&reflck->kref, vfio_pci_reflck_release, &reflck_lock);
2190}
2191
2192static int vfio_pci_get_unused_devs(struct pci_dev *pdev, void *data)
2193{
2194	struct vfio_devices *devs = data;
2195	struct vfio_device *device;
2196	struct vfio_pci_device *vdev;
2197
2198	if (devs->cur_index == devs->max_index)
2199		return -ENOSPC;
2200
2201	device = vfio_device_get_from_dev(&pdev->dev);
2202	if (!device)
2203		return -EINVAL;
2204
2205	if (pci_dev_driver(pdev) != &vfio_pci_driver) {
2206		vfio_device_put(device);
2207		return -EBUSY;
2208	}
2209
2210	vdev = vfio_device_data(device);
2211
2212	/* Fault if the device is not unused */
2213	if (vdev->refcnt) {
2214		vfio_device_put(device);
2215		return -EBUSY;
2216	}
2217
2218	devs->devices[devs->cur_index++] = device;
2219	return 0;
2220}
2221
2222static int vfio_pci_try_zap_and_vma_lock_cb(struct pci_dev *pdev, void *data)
2223{
2224	struct vfio_devices *devs = data;
2225	struct vfio_device *device;
2226	struct vfio_pci_device *vdev;
2227
2228	if (devs->cur_index == devs->max_index)
2229		return -ENOSPC;
2230
2231	device = vfio_device_get_from_dev(&pdev->dev);
2232	if (!device)
2233		return -EINVAL;
2234
2235	if (pci_dev_driver(pdev) != &vfio_pci_driver) {
2236		vfio_device_put(device);
2237		return -EBUSY;
2238	}
2239
2240	vdev = vfio_device_data(device);
2241
2242	/*
2243	 * Locking multiple devices is prone to deadlock, runaway and
2244	 * unwind if we hit contention.
2245	 */
2246	if (!vfio_pci_zap_and_vma_lock(vdev, true)) {
2247		vfio_device_put(device);
2248		return -EBUSY;
2249	}
2250
2251	devs->devices[devs->cur_index++] = device;
2252	return 0;
2253}
2254
2255/*
2256 * If a bus or slot reset is available for the provided device and:
2257 *  - All of the devices affected by that bus or slot reset are unused
2258 *    (!refcnt)
2259 *  - At least one of the affected devices is marked dirty via
2260 *    needs_reset (such as by lack of FLR support)
2261 * Then attempt to perform that bus or slot reset.  Callers are required
2262 * to hold vdev->reflck->lock, protecting the bus/slot reset group from
2263 * concurrent opens.  A vfio_device reference is acquired for each device
2264 * to prevent unbinds during the reset operation.
2265 *
2266 * NB: vfio-core considers a group to be viable even if some devices are
2267 * bound to drivers like pci-stub or pcieport.  Here we require all devices
2268 * to be bound to vfio_pci since that's the only way we can be sure they
2269 * stay put.
2270 */
2271static void vfio_pci_try_bus_reset(struct vfio_pci_device *vdev)
2272{
2273	struct vfio_devices devs = { .cur_index = 0 };
2274	int i = 0, ret = -EINVAL;
2275	bool slot = false;
2276	struct vfio_pci_device *tmp;
2277
2278	if (!pci_probe_reset_slot(vdev->pdev->slot))
2279		slot = true;
2280	else if (pci_probe_reset_bus(vdev->pdev->bus))
2281		return;
2282
2283	if (vfio_pci_for_each_slot_or_bus(vdev->pdev, vfio_pci_count_devs,
2284					  &i, slot) || !i)
2285		return;
2286
2287	devs.max_index = i;
2288	devs.devices = kcalloc(i, sizeof(struct vfio_device *), GFP_KERNEL);
2289	if (!devs.devices)
2290		return;
2291
2292	if (vfio_pci_for_each_slot_or_bus(vdev->pdev,
2293					  vfio_pci_get_unused_devs,
2294					  &devs, slot))
2295		goto put_devs;
2296
2297	/* Does at least one need a reset? */
2298	for (i = 0; i < devs.cur_index; i++) {
2299		tmp = vfio_device_data(devs.devices[i]);
2300		if (tmp->needs_reset) {
2301			ret = pci_reset_bus(vdev->pdev);
2302			break;
2303		}
2304	}
2305
2306put_devs:
2307	for (i = 0; i < devs.cur_index; i++) {
2308		tmp = vfio_device_data(devs.devices[i]);
2309
2310		/*
2311		 * If reset was successful, affected devices no longer need
2312		 * a reset and we should return all the collateral devices
2313		 * to low power.  If not successful, we either didn't reset
2314		 * the bus or timed out waiting for it, so let's not touch
2315		 * the power state.
2316		 */
2317		if (!ret) {
2318			tmp->needs_reset = false;
2319
2320			if (tmp != vdev && !disable_idle_d3)
2321				vfio_pci_set_power_state(tmp, PCI_D3hot);
2322		}
2323
2324		vfio_device_put(devs.devices[i]);
2325	}
2326
2327	kfree(devs.devices);
2328}
2329
2330static void __exit vfio_pci_cleanup(void)
2331{
2332	pci_unregister_driver(&vfio_pci_driver);
 
2333	vfio_pci_uninit_perm_bits();
2334}
2335
2336static void __init vfio_pci_fill_ids(void)
2337{
2338	char *p, *id;
2339	int rc;
2340
2341	/* no ids passed actually */
2342	if (ids[0] == '\0')
2343		return;
2344
2345	/* add ids specified in the module parameter */
2346	p = ids;
2347	while ((id = strsep(&p, ","))) {
2348		unsigned int vendor, device, subvendor = PCI_ANY_ID,
2349			subdevice = PCI_ANY_ID, class = 0, class_mask = 0;
2350		int fields;
2351
2352		if (!strlen(id))
2353			continue;
2354
2355		fields = sscanf(id, "%x:%x:%x:%x:%x:%x",
2356				&vendor, &device, &subvendor, &subdevice,
2357				&class, &class_mask);
2358
2359		if (fields < 2) {
2360			pr_warn("invalid id string \"%s\"\n", id);
2361			continue;
2362		}
2363
2364		rc = pci_add_dynid(&vfio_pci_driver, vendor, device,
2365				   subvendor, subdevice, class, class_mask, 0);
2366		if (rc)
2367			pr_warn("failed to add dynamic id [%04x:%04x[%04x:%04x]] class %#08x/%08x (%d)\n",
2368				vendor, device, subvendor, subdevice,
2369				class, class_mask, rc);
2370		else
2371			pr_info("add [%04x:%04x[%04x:%04x]] class %#08x/%08x\n",
2372				vendor, device, subvendor, subdevice,
2373				class, class_mask);
2374	}
2375}
2376
2377static int __init vfio_pci_init(void)
2378{
2379	int ret;
2380
2381	/* Allocate shared config space permision data used by all devices */
2382	ret = vfio_pci_init_perm_bits();
2383	if (ret)
2384		return ret;
2385
 
 
 
 
 
2386	/* Register and scan for devices */
2387	ret = pci_register_driver(&vfio_pci_driver);
2388	if (ret)
2389		goto out_driver;
2390
2391	vfio_pci_fill_ids();
2392
2393	if (disable_denylist)
2394		pr_warn("device denylist disabled.\n");
2395
2396	return 0;
2397
2398out_driver:
 
 
2399	vfio_pci_uninit_perm_bits();
2400	return ret;
2401}
2402
2403module_init(vfio_pci_init);
2404module_exit(vfio_pci_cleanup);
2405
2406MODULE_VERSION(DRIVER_VERSION);
2407MODULE_LICENSE("GPL v2");
2408MODULE_AUTHOR(DRIVER_AUTHOR);
2409MODULE_DESCRIPTION(DRIVER_DESC);