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#define dev_fmt pr_fmt
  13
  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/vfio.h>
  28#include <linux/vgaarb.h>
  29#include <linux/nospec.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 inline bool vfio_vga_disabled(void)
  58{
  59#ifdef CONFIG_VFIO_PCI_VGA
  60	return disable_vga;
  61#else
  62	return true;
  63#endif
  64}
  65
  66/*
  67 * Our VGA arbiter participation is limited since we don't know anything
  68 * about the device itself.  However, if the device is the only VGA device
  69 * downstream of a bridge and VFIO VGA support is disabled, then we can
  70 * safely return legacy VGA IO and memory as not decoded since the user
  71 * has no way to get to it and routing can be disabled externally at the
  72 * bridge.
  73 */
  74static unsigned int vfio_pci_set_vga_decode(void *opaque, bool single_vga)
  75{
  76	struct vfio_pci_device *vdev = opaque;
  77	struct pci_dev *tmp = NULL, *pdev = vdev->pdev;
  78	unsigned char max_busnr;
  79	unsigned int decodes;
  80
  81	if (single_vga || !vfio_vga_disabled() || pci_is_root_bus(pdev->bus))
  82		return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM |
  83		       VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM;
  84
  85	max_busnr = pci_bus_max_busnr(pdev->bus);
  86	decodes = VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
  87
  88	while ((tmp = pci_get_class(PCI_CLASS_DISPLAY_VGA << 8, tmp)) != NULL) {
  89		if (tmp == pdev ||
  90		    pci_domain_nr(tmp->bus) != pci_domain_nr(pdev->bus) ||
  91		    pci_is_root_bus(tmp->bus))
  92			continue;
  93
  94		if (tmp->bus->number >= pdev->bus->number &&
  95		    tmp->bus->number <= max_busnr) {
  96			pci_dev_put(tmp);
  97			decodes |= VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM;
  98			break;
  99		}
 100	}
 101
 102	return decodes;
 103}
 104
 105static inline bool vfio_pci_is_vga(struct pci_dev *pdev)
 106{
 107	return (pdev->class >> 8) == PCI_CLASS_DISPLAY_VGA;
 108}
 109
 110static void vfio_pci_probe_mmaps(struct vfio_pci_device *vdev)
 111{
 112	struct resource *res;
 113	int bar;
 114	struct vfio_pci_dummy_resource *dummy_res;
 115
 116	INIT_LIST_HEAD(&vdev->dummy_resources_list);
 117
 118	for (bar = PCI_STD_RESOURCES; bar <= PCI_STD_RESOURCE_END; bar++) {
 119		res = vdev->pdev->resource + bar;
 120
 121		if (!IS_ENABLED(CONFIG_VFIO_PCI_MMAP))
 122			goto no_mmap;
 123
 124		if (!(res->flags & IORESOURCE_MEM))
 125			goto no_mmap;
 126
 127		/*
 128		 * The PCI core shouldn't set up a resource with a
 129		 * type but zero size. But there may be bugs that
 130		 * cause us to do that.
 131		 */
 132		if (!resource_size(res))
 133			goto no_mmap;
 134
 135		if (resource_size(res) >= PAGE_SIZE) {
 136			vdev->bar_mmap_supported[bar] = true;
 137			continue;
 138		}
 139
 140		if (!(res->start & ~PAGE_MASK)) {
 141			/*
 142			 * Add a dummy resource to reserve the remainder
 143			 * of the exclusive page in case that hot-add
 144			 * device's bar is assigned into it.
 145			 */
 146			dummy_res = kzalloc(sizeof(*dummy_res), GFP_KERNEL);
 147			if (dummy_res == NULL)
 148				goto no_mmap;
 149
 150			dummy_res->resource.name = "vfio sub-page reserved";
 151			dummy_res->resource.start = res->end + 1;
 152			dummy_res->resource.end = res->start + PAGE_SIZE - 1;
 153			dummy_res->resource.flags = res->flags;
 154			if (request_resource(res->parent,
 155						&dummy_res->resource)) {
 156				kfree(dummy_res);
 157				goto no_mmap;
 158			}
 159			dummy_res->index = bar;
 160			list_add(&dummy_res->res_next,
 161					&vdev->dummy_resources_list);
 162			vdev->bar_mmap_supported[bar] = true;
 163			continue;
 164		}
 165		/*
 166		 * Here we don't handle the case when the BAR is not page
 167		 * aligned because we can't expect the BAR will be
 168		 * assigned into the same location in a page in guest
 169		 * when we passthrough the BAR. And it's hard to access
 170		 * this BAR in userspace because we have no way to get
 171		 * the BAR's location in a page.
 172		 */
 173no_mmap:
 174		vdev->bar_mmap_supported[bar] = false;
 175	}
 176}
 177
 178static void vfio_pci_try_bus_reset(struct vfio_pci_device *vdev);
 179static void vfio_pci_disable(struct vfio_pci_device *vdev);
 180
 181/*
 182 * INTx masking requires the ability to disable INTx signaling via PCI_COMMAND
 183 * _and_ the ability detect when the device is asserting INTx via PCI_STATUS.
 184 * If a device implements the former but not the latter we would typically
 185 * expect broken_intx_masking be set and require an exclusive interrupt.
 186 * However since we do have control of the device's ability to assert INTx,
 187 * we can instead pretend that the device does not implement INTx, virtualizing
 188 * the pin register to report zero and maintaining DisINTx set on the host.
 189 */
 190static bool vfio_pci_nointx(struct pci_dev *pdev)
 191{
 192	switch (pdev->vendor) {
 193	case PCI_VENDOR_ID_INTEL:
 194		switch (pdev->device) {
 195		/* All i40e (XL710/X710/XXV710) 10/20/25/40GbE NICs */
 196		case 0x1572:
 197		case 0x1574:
 198		case 0x1580 ... 0x1581:
 199		case 0x1583 ... 0x158b:
 200		case 0x37d0 ... 0x37d2:
 201			return true;
 202		default:
 203			return false;
 204		}
 205	}
 206
 207	return false;
 208}
 209
 210static void vfio_pci_probe_power_state(struct vfio_pci_device *vdev)
 211{
 212	struct pci_dev *pdev = vdev->pdev;
 213	u16 pmcsr;
 214
 215	if (!pdev->pm_cap)
 216		return;
 217
 218	pci_read_config_word(pdev, pdev->pm_cap + PCI_PM_CTRL, &pmcsr);
 219
 220	vdev->needs_pm_restore = !(pmcsr & PCI_PM_CTRL_NO_SOFT_RESET);
 221}
 222
 223/*
 224 * pci_set_power_state() wrapper handling devices which perform a soft reset on
 225 * D3->D0 transition.  Save state prior to D0/1/2->D3, stash it on the vdev,
 226 * restore when returned to D0.  Saved separately from pci_saved_state for use
 227 * by PM capability emulation and separately from pci_dev internal saved state
 228 * to avoid it being overwritten and consumed around other resets.
 229 */
 230int vfio_pci_set_power_state(struct vfio_pci_device *vdev, pci_power_t state)
 231{
 232	struct pci_dev *pdev = vdev->pdev;
 233	bool needs_restore = false, needs_save = false;
 234	int ret;
 235
 236	if (vdev->needs_pm_restore) {
 237		if (pdev->current_state < PCI_D3hot && state >= PCI_D3hot) {
 238			pci_save_state(pdev);
 239			needs_save = true;
 240		}
 241
 242		if (pdev->current_state >= PCI_D3hot && state <= PCI_D0)
 243			needs_restore = true;
 244	}
 245
 246	ret = pci_set_power_state(pdev, state);
 247
 248	if (!ret) {
 249		/* D3 might be unsupported via quirk, skip unless in D3 */
 250		if (needs_save && pdev->current_state >= PCI_D3hot) {
 251			vdev->pm_save = pci_store_saved_state(pdev);
 252		} else if (needs_restore) {
 253			pci_load_and_free_saved_state(pdev, &vdev->pm_save);
 254			pci_restore_state(pdev);
 255		}
 256	}
 257
 258	return ret;
 259}
 260
 261static int vfio_pci_enable(struct vfio_pci_device *vdev)
 262{
 263	struct pci_dev *pdev = vdev->pdev;
 264	int ret;
 265	u16 cmd;
 266	u8 msix_pos;
 267
 268	vfio_pci_set_power_state(vdev, PCI_D0);
 269
 270	/* Don't allow our initial saved state to include busmaster */
 271	pci_clear_master(pdev);
 272
 273	ret = pci_enable_device(pdev);
 274	if (ret)
 275		return ret;
 276
 277	/* If reset fails because of the device lock, fail this path entirely */
 278	ret = pci_try_reset_function(pdev);
 279	if (ret == -EAGAIN) {
 280		pci_disable_device(pdev);
 281		return ret;
 282	}
 283
 284	vdev->reset_works = !ret;
 285	pci_save_state(pdev);
 286	vdev->pci_saved_state = pci_store_saved_state(pdev);
 287	if (!vdev->pci_saved_state)
 288		pci_dbg(pdev, "%s: Couldn't store saved state\n", __func__);
 
 289
 290	if (likely(!nointxmask)) {
 291		if (vfio_pci_nointx(pdev)) {
 292			pci_info(pdev, "Masking broken INTx support\n");
 293			vdev->nointx = true;
 294			pci_intx(pdev, 0);
 295		} else
 296			vdev->pci_2_3 = pci_intx_mask_supported(pdev);
 297	}
 298
 
 
 
 299	pci_read_config_word(pdev, PCI_COMMAND, &cmd);
 300	if (vdev->pci_2_3 && (cmd & PCI_COMMAND_INTX_DISABLE)) {
 301		cmd &= ~PCI_COMMAND_INTX_DISABLE;
 302		pci_write_config_word(pdev, PCI_COMMAND, cmd);
 303	}
 304
 305	ret = vfio_config_init(vdev);
 306	if (ret) {
 307		kfree(vdev->pci_saved_state);
 308		vdev->pci_saved_state = NULL;
 309		pci_disable_device(pdev);
 310		return ret;
 311	}
 312
 313	msix_pos = pdev->msix_cap;
 314	if (msix_pos) {
 315		u16 flags;
 316		u32 table;
 317
 318		pci_read_config_word(pdev, msix_pos + PCI_MSIX_FLAGS, &flags);
 319		pci_read_config_dword(pdev, msix_pos + PCI_MSIX_TABLE, &table);
 320
 321		vdev->msix_bar = table & PCI_MSIX_TABLE_BIR;
 322		vdev->msix_offset = table & PCI_MSIX_TABLE_OFFSET;
 323		vdev->msix_size = ((flags & PCI_MSIX_FLAGS_QSIZE) + 1) * 16;
 324	} else
 325		vdev->msix_bar = 0xFF;
 326
 327	if (!vfio_vga_disabled() && vfio_pci_is_vga(pdev))
 
 328		vdev->has_vga = true;
 329
 330
 331	if (vfio_pci_is_vga(pdev) &&
 332	    pdev->vendor == PCI_VENDOR_ID_INTEL &&
 333	    IS_ENABLED(CONFIG_VFIO_PCI_IGD)) {
 334		ret = vfio_pci_igd_init(vdev);
 335		if (ret) {
 336			pci_warn(pdev, "Failed to setup Intel IGD regions\n");
 337			goto disable_exit;
 338		}
 339	}
 340
 341	if (pdev->vendor == PCI_VENDOR_ID_NVIDIA &&
 342	    IS_ENABLED(CONFIG_VFIO_PCI_NVLINK2)) {
 343		ret = vfio_pci_nvdia_v100_nvlink2_init(vdev);
 344		if (ret && ret != -ENODEV) {
 345			pci_warn(pdev, "Failed to setup NVIDIA NV2 RAM region\n");
 346			goto disable_exit;
 347		}
 348	}
 349
 350	if (pdev->vendor == PCI_VENDOR_ID_IBM &&
 351	    IS_ENABLED(CONFIG_VFIO_PCI_NVLINK2)) {
 352		ret = vfio_pci_ibm_npu2_init(vdev);
 353		if (ret && ret != -ENODEV) {
 354			pci_warn(pdev, "Failed to setup NVIDIA NV2 ATSD region\n");
 355			goto disable_exit;
 356		}
 357	}
 358
 359	vfio_pci_probe_mmaps(vdev);
 360
 361	return 0;
 362
 363disable_exit:
 364	vfio_pci_disable(vdev);
 365	return ret;
 366}
 367
 368static void vfio_pci_disable(struct vfio_pci_device *vdev)
 369{
 370	struct pci_dev *pdev = vdev->pdev;
 371	struct vfio_pci_dummy_resource *dummy_res, *tmp;
 372	struct vfio_pci_ioeventfd *ioeventfd, *ioeventfd_tmp;
 373	int i, bar;
 374
 375	/* Stop the device from further DMA */
 376	pci_clear_master(pdev);
 377
 378	vfio_pci_set_irqs_ioctl(vdev, VFIO_IRQ_SET_DATA_NONE |
 379				VFIO_IRQ_SET_ACTION_TRIGGER,
 380				vdev->irq_type, 0, 0, NULL);
 381
 382	/* Device closed, don't need mutex here */
 383	list_for_each_entry_safe(ioeventfd, ioeventfd_tmp,
 384				 &vdev->ioeventfds_list, next) {
 385		vfio_virqfd_disable(&ioeventfd->virqfd);
 386		list_del(&ioeventfd->next);
 387		kfree(ioeventfd);
 388	}
 389	vdev->ioeventfds_nr = 0;
 390
 391	vdev->virq_disabled = false;
 392
 393	for (i = 0; i < vdev->num_regions; i++)
 394		vdev->region[i].ops->release(vdev, &vdev->region[i]);
 395
 396	vdev->num_regions = 0;
 397	kfree(vdev->region);
 398	vdev->region = NULL; /* don't krealloc a freed pointer */
 399
 400	vfio_config_free(vdev);
 401
 402	for (bar = PCI_STD_RESOURCES; bar <= PCI_STD_RESOURCE_END; bar++) {
 403		if (!vdev->barmap[bar])
 404			continue;
 405		pci_iounmap(pdev, vdev->barmap[bar]);
 406		pci_release_selected_regions(pdev, 1 << bar);
 407		vdev->barmap[bar] = NULL;
 408	}
 409
 410	list_for_each_entry_safe(dummy_res, tmp,
 411				 &vdev->dummy_resources_list, res_next) {
 412		list_del(&dummy_res->res_next);
 413		release_resource(&dummy_res->resource);
 414		kfree(dummy_res);
 415	}
 416
 417	vdev->needs_reset = true;
 418
 419	/*
 420	 * If we have saved state, restore it.  If we can reset the device,
 421	 * even better.  Resetting with current state seems better than
 422	 * nothing, but saving and restoring current state without reset
 423	 * is just busy work.
 424	 */
 425	if (pci_load_and_free_saved_state(pdev, &vdev->pci_saved_state)) {
 426		pci_info(pdev, "%s: Couldn't reload saved state\n", __func__);
 
 427
 428		if (!vdev->reset_works)
 429			goto out;
 430
 431		pci_save_state(pdev);
 432	}
 433
 434	/*
 435	 * Disable INTx and MSI, presumably to avoid spurious interrupts
 436	 * during reset.  Stolen from pci_reset_function()
 437	 */
 438	pci_write_config_word(pdev, PCI_COMMAND, PCI_COMMAND_INTX_DISABLE);
 439
 440	/*
 441	 * Try to get the locks ourselves to prevent a deadlock. The
 442	 * success of this is dependent on being able to lock the device,
 443	 * which is not always possible.
 444	 * We can not use the "try" reset interface here, which will
 445	 * overwrite the previously restored configuration information.
 446	 */
 447	if (vdev->reset_works && pci_cfg_access_trylock(pdev)) {
 448		if (device_trylock(&pdev->dev)) {
 449			if (!__pci_reset_function_locked(pdev))
 450				vdev->needs_reset = false;
 451			device_unlock(&pdev->dev);
 452		}
 453		pci_cfg_access_unlock(pdev);
 454	}
 455
 456	pci_restore_state(pdev);
 457out:
 458	pci_disable_device(pdev);
 459
 460	vfio_pci_try_bus_reset(vdev);
 461
 462	if (!disable_idle_d3)
 463		vfio_pci_set_power_state(vdev, PCI_D3hot);
 464}
 465
 466static void vfio_pci_release(void *device_data)
 467{
 468	struct vfio_pci_device *vdev = device_data;
 469
 470	mutex_lock(&vdev->reflck->lock);
 471
 472	if (!(--vdev->refcnt)) {
 473		vfio_spapr_pci_eeh_release(vdev->pdev);
 474		vfio_pci_disable(vdev);
 475	}
 476
 477	mutex_unlock(&vdev->reflck->lock);
 478
 479	module_put(THIS_MODULE);
 480}
 481
 482static int vfio_pci_open(void *device_data)
 483{
 484	struct vfio_pci_device *vdev = device_data;
 485	int ret = 0;
 486
 487	if (!try_module_get(THIS_MODULE))
 488		return -ENODEV;
 489
 490	mutex_lock(&vdev->reflck->lock);
 491
 492	if (!vdev->refcnt) {
 493		ret = vfio_pci_enable(vdev);
 494		if (ret)
 495			goto error;
 496
 497		vfio_spapr_pci_eeh_open(vdev->pdev);
 498	}
 499	vdev->refcnt++;
 500error:
 501	mutex_unlock(&vdev->reflck->lock);
 502	if (ret)
 503		module_put(THIS_MODULE);
 504	return ret;
 505}
 506
 507static int vfio_pci_get_irq_count(struct vfio_pci_device *vdev, int irq_type)
 508{
 509	if (irq_type == VFIO_PCI_INTX_IRQ_INDEX) {
 510		u8 pin;
 511
 512		if (!IS_ENABLED(CONFIG_VFIO_PCI_INTX) ||
 513		    vdev->nointx || vdev->pdev->is_virtfn)
 514			return 0;
 515
 516		pci_read_config_byte(vdev->pdev, PCI_INTERRUPT_PIN, &pin);
 
 
 517
 518		return pin ? 1 : 0;
 519	} else if (irq_type == VFIO_PCI_MSI_IRQ_INDEX) {
 520		u8 pos;
 521		u16 flags;
 522
 523		pos = vdev->pdev->msi_cap;
 524		if (pos) {
 525			pci_read_config_word(vdev->pdev,
 526					     pos + PCI_MSI_FLAGS, &flags);
 527			return 1 << ((flags & PCI_MSI_FLAGS_QMASK) >> 1);
 
 528		}
 529	} else if (irq_type == VFIO_PCI_MSIX_IRQ_INDEX) {
 530		u8 pos;
 531		u16 flags;
 532
 533		pos = vdev->pdev->msix_cap;
 534		if (pos) {
 535			pci_read_config_word(vdev->pdev,
 536					     pos + PCI_MSIX_FLAGS, &flags);
 537
 538			return (flags & PCI_MSIX_FLAGS_QSIZE) + 1;
 539		}
 540	} else if (irq_type == VFIO_PCI_ERR_IRQ_INDEX) {
 541		if (pci_is_pcie(vdev->pdev))
 542			return 1;
 543	} else if (irq_type == VFIO_PCI_REQ_IRQ_INDEX) {
 544		return 1;
 545	}
 546
 547	return 0;
 548}
 549
 550static int vfio_pci_count_devs(struct pci_dev *pdev, void *data)
 551{
 552	(*(int *)data)++;
 553	return 0;
 554}
 555
 556struct vfio_pci_fill_info {
 557	int max;
 558	int cur;
 559	struct vfio_pci_dependent_device *devices;
 560};
 561
 562static int vfio_pci_fill_devs(struct pci_dev *pdev, void *data)
 563{
 564	struct vfio_pci_fill_info *fill = data;
 565	struct iommu_group *iommu_group;
 566
 567	if (fill->cur == fill->max)
 568		return -EAGAIN; /* Something changed, try again */
 569
 570	iommu_group = iommu_group_get(&pdev->dev);
 571	if (!iommu_group)
 572		return -EPERM; /* Cannot reset non-isolated devices */
 573
 574	fill->devices[fill->cur].group_id = iommu_group_id(iommu_group);
 575	fill->devices[fill->cur].segment = pci_domain_nr(pdev->bus);
 576	fill->devices[fill->cur].bus = pdev->bus->number;
 577	fill->devices[fill->cur].devfn = pdev->devfn;
 578	fill->cur++;
 579	iommu_group_put(iommu_group);
 580	return 0;
 581}
 582
 583struct vfio_pci_group_entry {
 584	struct vfio_group *group;
 585	int id;
 586};
 587
 588struct vfio_pci_group_info {
 589	int count;
 590	struct vfio_pci_group_entry *groups;
 591};
 592
 593static int vfio_pci_validate_devs(struct pci_dev *pdev, void *data)
 594{
 595	struct vfio_pci_group_info *info = data;
 596	struct iommu_group *group;
 597	int id, i;
 598
 599	group = iommu_group_get(&pdev->dev);
 600	if (!group)
 601		return -EPERM;
 602
 603	id = iommu_group_id(group);
 604
 605	for (i = 0; i < info->count; i++)
 606		if (info->groups[i].id == id)
 607			break;
 608
 609	iommu_group_put(group);
 610
 611	return (i == info->count) ? -EINVAL : 0;
 612}
 613
 614static bool vfio_pci_dev_below_slot(struct pci_dev *pdev, struct pci_slot *slot)
 615{
 616	for (; pdev; pdev = pdev->bus->self)
 617		if (pdev->bus == slot->bus)
 618			return (pdev->slot == slot);
 619	return false;
 620}
 621
 622struct vfio_pci_walk_info {
 623	int (*fn)(struct pci_dev *, void *data);
 624	void *data;
 625	struct pci_dev *pdev;
 626	bool slot;
 627	int ret;
 628};
 629
 630static int vfio_pci_walk_wrapper(struct pci_dev *pdev, void *data)
 631{
 632	struct vfio_pci_walk_info *walk = data;
 633
 634	if (!walk->slot || vfio_pci_dev_below_slot(pdev, walk->pdev->slot))
 635		walk->ret = walk->fn(pdev, walk->data);
 636
 637	return walk->ret;
 638}
 639
 640static int vfio_pci_for_each_slot_or_bus(struct pci_dev *pdev,
 641					 int (*fn)(struct pci_dev *,
 642						   void *data), void *data,
 643					 bool slot)
 644{
 645	struct vfio_pci_walk_info walk = {
 646		.fn = fn, .data = data, .pdev = pdev, .slot = slot, .ret = 0,
 647	};
 648
 649	pci_walk_bus(pdev->bus, vfio_pci_walk_wrapper, &walk);
 650
 651	return walk.ret;
 652}
 653
 654static int msix_mmappable_cap(struct vfio_pci_device *vdev,
 655			      struct vfio_info_cap *caps)
 656{
 657	struct vfio_info_cap_header header = {
 658		.id = VFIO_REGION_INFO_CAP_MSIX_MAPPABLE,
 659		.version = 1
 660	};
 661
 662	return vfio_info_add_capability(caps, &header, sizeof(header));
 663}
 664
 665int vfio_pci_register_dev_region(struct vfio_pci_device *vdev,
 666				 unsigned int type, unsigned int subtype,
 667				 const struct vfio_pci_regops *ops,
 668				 size_t size, u32 flags, void *data)
 669{
 670	struct vfio_pci_region *region;
 671
 672	region = krealloc(vdev->region,
 673			  (vdev->num_regions + 1) * sizeof(*region),
 674			  GFP_KERNEL);
 675	if (!region)
 676		return -ENOMEM;
 677
 678	vdev->region = region;
 679	vdev->region[vdev->num_regions].type = type;
 680	vdev->region[vdev->num_regions].subtype = subtype;
 681	vdev->region[vdev->num_regions].ops = ops;
 682	vdev->region[vdev->num_regions].size = size;
 683	vdev->region[vdev->num_regions].flags = flags;
 684	vdev->region[vdev->num_regions].data = data;
 685
 686	vdev->num_regions++;
 687
 688	return 0;
 689}
 690
 691static long vfio_pci_ioctl(void *device_data,
 692			   unsigned int cmd, unsigned long arg)
 693{
 694	struct vfio_pci_device *vdev = device_data;
 695	unsigned long minsz;
 696
 697	if (cmd == VFIO_DEVICE_GET_INFO) {
 698		struct vfio_device_info info;
 699
 700		minsz = offsetofend(struct vfio_device_info, num_irqs);
 701
 702		if (copy_from_user(&info, (void __user *)arg, minsz))
 703			return -EFAULT;
 704
 705		if (info.argsz < minsz)
 706			return -EINVAL;
 707
 708		info.flags = VFIO_DEVICE_FLAGS_PCI;
 709
 710		if (vdev->reset_works)
 711			info.flags |= VFIO_DEVICE_FLAGS_RESET;
 712
 713		info.num_regions = VFIO_PCI_NUM_REGIONS + vdev->num_regions;
 714		info.num_irqs = VFIO_PCI_NUM_IRQS;
 715
 716		return copy_to_user((void __user *)arg, &info, minsz) ?
 717			-EFAULT : 0;
 718
 719	} else if (cmd == VFIO_DEVICE_GET_REGION_INFO) {
 720		struct pci_dev *pdev = vdev->pdev;
 721		struct vfio_region_info info;
 722		struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
 723		int i, ret;
 724
 725		minsz = offsetofend(struct vfio_region_info, offset);
 726
 727		if (copy_from_user(&info, (void __user *)arg, minsz))
 728			return -EFAULT;
 729
 730		if (info.argsz < minsz)
 731			return -EINVAL;
 732
 733		switch (info.index) {
 734		case VFIO_PCI_CONFIG_REGION_INDEX:
 735			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
 736			info.size = pdev->cfg_size;
 737			info.flags = VFIO_REGION_INFO_FLAG_READ |
 738				     VFIO_REGION_INFO_FLAG_WRITE;
 739			break;
 740		case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
 741			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
 742			info.size = pci_resource_len(pdev, info.index);
 743			if (!info.size) {
 744				info.flags = 0;
 745				break;
 746			}
 747
 748			info.flags = VFIO_REGION_INFO_FLAG_READ |
 749				     VFIO_REGION_INFO_FLAG_WRITE;
 750			if (vdev->bar_mmap_supported[info.index]) {
 
 751				info.flags |= VFIO_REGION_INFO_FLAG_MMAP;
 752				if (info.index == vdev->msix_bar) {
 753					ret = msix_mmappable_cap(vdev, &caps);
 754					if (ret)
 755						return ret;
 756				}
 757			}
 758
 759			break;
 760		case VFIO_PCI_ROM_REGION_INDEX:
 761		{
 762			void __iomem *io;
 763			size_t size;
 764			u16 orig_cmd;
 765
 766			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
 767			info.flags = 0;
 768
 769			/* Report the BAR size, not the ROM size */
 770			info.size = pci_resource_len(pdev, info.index);
 771			if (!info.size) {
 772				/* Shadow ROMs appear as PCI option ROMs */
 773				if (pdev->resource[PCI_ROM_RESOURCE].flags &
 774							IORESOURCE_ROM_SHADOW)
 775					info.size = 0x20000;
 776				else
 777					break;
 778			}
 779
 780			/*
 781			 * Is it really there?  Enable memory decode for
 782			 * implicit access in pci_map_rom().
 783			 */
 784			pci_read_config_word(pdev, PCI_COMMAND, &orig_cmd);
 785			pci_write_config_word(pdev, PCI_COMMAND,
 786					      orig_cmd | PCI_COMMAND_MEMORY);
 787
 
 788			io = pci_map_rom(pdev, &size);
 789			if (io) {
 790				info.flags = VFIO_REGION_INFO_FLAG_READ;
 791				pci_unmap_rom(pdev, io);
 792			} else {
 793				info.size = 0;
 
 794			}
 
 795
 796			pci_write_config_word(pdev, PCI_COMMAND, orig_cmd);
 797			break;
 798		}
 799		case VFIO_PCI_VGA_REGION_INDEX:
 800			if (!vdev->has_vga)
 801				return -EINVAL;
 802
 803			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
 804			info.size = 0xc0000;
 805			info.flags = VFIO_REGION_INFO_FLAG_READ |
 806				     VFIO_REGION_INFO_FLAG_WRITE;
 807
 808			break;
 809		default:
 810		{
 811			struct vfio_region_info_cap_type cap_type = {
 812					.header.id = VFIO_REGION_INFO_CAP_TYPE,
 813					.header.version = 1 };
 814
 815			if (info.index >=
 816			    VFIO_PCI_NUM_REGIONS + vdev->num_regions)
 817				return -EINVAL;
 818			info.index = array_index_nospec(info.index,
 819							VFIO_PCI_NUM_REGIONS +
 820							vdev->num_regions);
 821
 822			i = info.index - VFIO_PCI_NUM_REGIONS;
 823
 824			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
 825			info.size = vdev->region[i].size;
 826			info.flags = vdev->region[i].flags;
 827
 828			cap_type.type = vdev->region[i].type;
 829			cap_type.subtype = vdev->region[i].subtype;
 830
 831			ret = vfio_info_add_capability(&caps, &cap_type.header,
 832						       sizeof(cap_type));
 833			if (ret)
 834				return ret;
 835
 836			if (vdev->region[i].ops->add_capability) {
 837				ret = vdev->region[i].ops->add_capability(vdev,
 838						&vdev->region[i], &caps);
 839				if (ret)
 840					return ret;
 841			}
 842		}
 843		}
 844
 845		if (caps.size) {
 846			info.flags |= VFIO_REGION_INFO_FLAG_CAPS;
 847			if (info.argsz < sizeof(info) + caps.size) {
 848				info.argsz = sizeof(info) + caps.size;
 849				info.cap_offset = 0;
 850			} else {
 851				vfio_info_cap_shift(&caps, sizeof(info));
 852				if (copy_to_user((void __user *)arg +
 853						  sizeof(info), caps.buf,
 854						  caps.size)) {
 855					kfree(caps.buf);
 856					return -EFAULT;
 857				}
 858				info.cap_offset = sizeof(info);
 859			}
 860
 861			kfree(caps.buf);
 862		}
 863
 864		return copy_to_user((void __user *)arg, &info, minsz) ?
 865			-EFAULT : 0;
 866
 867	} else if (cmd == VFIO_DEVICE_GET_IRQ_INFO) {
 868		struct vfio_irq_info info;
 869
 870		minsz = offsetofend(struct vfio_irq_info, count);
 871
 872		if (copy_from_user(&info, (void __user *)arg, minsz))
 873			return -EFAULT;
 874
 875		if (info.argsz < minsz || info.index >= VFIO_PCI_NUM_IRQS)
 876			return -EINVAL;
 877
 878		switch (info.index) {
 879		case VFIO_PCI_INTX_IRQ_INDEX ... VFIO_PCI_MSIX_IRQ_INDEX:
 880		case VFIO_PCI_REQ_IRQ_INDEX:
 881			break;
 882		case VFIO_PCI_ERR_IRQ_INDEX:
 883			if (pci_is_pcie(vdev->pdev))
 884				break;
 885		/* fall through */
 886		default:
 887			return -EINVAL;
 888		}
 889
 890		info.flags = VFIO_IRQ_INFO_EVENTFD;
 891
 892		info.count = vfio_pci_get_irq_count(vdev, info.index);
 893
 894		if (info.index == VFIO_PCI_INTX_IRQ_INDEX)
 895			info.flags |= (VFIO_IRQ_INFO_MASKABLE |
 896				       VFIO_IRQ_INFO_AUTOMASKED);
 897		else
 898			info.flags |= VFIO_IRQ_INFO_NORESIZE;
 899
 900		return copy_to_user((void __user *)arg, &info, minsz) ?
 901			-EFAULT : 0;
 902
 903	} else if (cmd == VFIO_DEVICE_SET_IRQS) {
 904		struct vfio_irq_set hdr;
 905		u8 *data = NULL;
 906		int max, ret = 0;
 907		size_t data_size = 0;
 908
 909		minsz = offsetofend(struct vfio_irq_set, count);
 910
 911		if (copy_from_user(&hdr, (void __user *)arg, minsz))
 912			return -EFAULT;
 913
 914		max = vfio_pci_get_irq_count(vdev, hdr.index);
 
 
 
 915
 916		ret = vfio_set_irqs_validate_and_prepare(&hdr, max,
 917						 VFIO_PCI_NUM_IRQS, &data_size);
 918		if (ret)
 919			return ret;
 
 
 
 
 
 
 
 
 
 
 920
 921		if (data_size) {
 922			data = memdup_user((void __user *)(arg + minsz),
 923					    data_size);
 924			if (IS_ERR(data))
 925				return PTR_ERR(data);
 926		}
 927
 928		mutex_lock(&vdev->igate);
 929
 930		ret = vfio_pci_set_irqs_ioctl(vdev, hdr.flags, hdr.index,
 931					      hdr.start, hdr.count, data);
 932
 933		mutex_unlock(&vdev->igate);
 934		kfree(data);
 935
 936		return ret;
 937
 938	} else if (cmd == VFIO_DEVICE_RESET) {
 939		return vdev->reset_works ?
 940			pci_try_reset_function(vdev->pdev) : -EINVAL;
 941
 942	} else if (cmd == VFIO_DEVICE_GET_PCI_HOT_RESET_INFO) {
 943		struct vfio_pci_hot_reset_info hdr;
 944		struct vfio_pci_fill_info fill = { 0 };
 945		struct vfio_pci_dependent_device *devices = NULL;
 946		bool slot = false;
 947		int ret = 0;
 948
 949		minsz = offsetofend(struct vfio_pci_hot_reset_info, count);
 950
 951		if (copy_from_user(&hdr, (void __user *)arg, minsz))
 952			return -EFAULT;
 953
 954		if (hdr.argsz < minsz)
 955			return -EINVAL;
 956
 957		hdr.flags = 0;
 958
 959		/* Can we do a slot or bus reset or neither? */
 960		if (!pci_probe_reset_slot(vdev->pdev->slot))
 961			slot = true;
 962		else if (pci_probe_reset_bus(vdev->pdev->bus))
 963			return -ENODEV;
 964
 965		/* How many devices are affected? */
 966		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
 967						    vfio_pci_count_devs,
 968						    &fill.max, slot);
 969		if (ret)
 970			return ret;
 971
 972		WARN_ON(!fill.max); /* Should always be at least one */
 973
 974		/*
 975		 * If there's enough space, fill it now, otherwise return
 976		 * -ENOSPC and the number of devices affected.
 977		 */
 978		if (hdr.argsz < sizeof(hdr) + (fill.max * sizeof(*devices))) {
 979			ret = -ENOSPC;
 980			hdr.count = fill.max;
 981			goto reset_info_exit;
 982		}
 983
 984		devices = kcalloc(fill.max, sizeof(*devices), GFP_KERNEL);
 985		if (!devices)
 986			return -ENOMEM;
 987
 988		fill.devices = devices;
 989
 990		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
 991						    vfio_pci_fill_devs,
 992						    &fill, slot);
 993
 994		/*
 995		 * If a device was removed between counting and filling,
 996		 * we may come up short of fill.max.  If a device was
 997		 * added, we'll have a return of -EAGAIN above.
 998		 */
 999		if (!ret)
1000			hdr.count = fill.cur;
1001
1002reset_info_exit:
1003		if (copy_to_user((void __user *)arg, &hdr, minsz))
1004			ret = -EFAULT;
1005
1006		if (!ret) {
1007			if (copy_to_user((void __user *)(arg + minsz), devices,
1008					 hdr.count * sizeof(*devices)))
1009				ret = -EFAULT;
1010		}
1011
1012		kfree(devices);
1013		return ret;
1014
1015	} else if (cmd == VFIO_DEVICE_PCI_HOT_RESET) {
1016		struct vfio_pci_hot_reset hdr;
1017		int32_t *group_fds;
1018		struct vfio_pci_group_entry *groups;
1019		struct vfio_pci_group_info info;
1020		bool slot = false;
1021		int i, count = 0, ret = 0;
1022
1023		minsz = offsetofend(struct vfio_pci_hot_reset, count);
1024
1025		if (copy_from_user(&hdr, (void __user *)arg, minsz))
1026			return -EFAULT;
1027
1028		if (hdr.argsz < minsz || hdr.flags)
1029			return -EINVAL;
1030
1031		/* Can we do a slot or bus reset or neither? */
1032		if (!pci_probe_reset_slot(vdev->pdev->slot))
1033			slot = true;
1034		else if (pci_probe_reset_bus(vdev->pdev->bus))
1035			return -ENODEV;
1036
1037		/*
1038		 * We can't let userspace give us an arbitrarily large
1039		 * buffer to copy, so verify how many we think there
1040		 * could be.  Note groups can have multiple devices so
1041		 * one group per device is the max.
1042		 */
1043		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1044						    vfio_pci_count_devs,
1045						    &count, slot);
1046		if (ret)
1047			return ret;
1048
1049		/* Somewhere between 1 and count is OK */
1050		if (!hdr.count || hdr.count > count)
1051			return -EINVAL;
1052
1053		group_fds = kcalloc(hdr.count, sizeof(*group_fds), GFP_KERNEL);
1054		groups = kcalloc(hdr.count, sizeof(*groups), GFP_KERNEL);
1055		if (!group_fds || !groups) {
1056			kfree(group_fds);
1057			kfree(groups);
1058			return -ENOMEM;
1059		}
1060
1061		if (copy_from_user(group_fds, (void __user *)(arg + minsz),
1062				   hdr.count * sizeof(*group_fds))) {
1063			kfree(group_fds);
1064			kfree(groups);
1065			return -EFAULT;
1066		}
1067
1068		/*
1069		 * For each group_fd, get the group through the vfio external
1070		 * user interface and store the group and iommu ID.  This
1071		 * ensures the group is held across the reset.
1072		 */
1073		for (i = 0; i < hdr.count; i++) {
1074			struct vfio_group *group;
1075			struct fd f = fdget(group_fds[i]);
1076			if (!f.file) {
1077				ret = -EBADF;
1078				break;
1079			}
1080
1081			group = vfio_group_get_external_user(f.file);
1082			fdput(f);
1083			if (IS_ERR(group)) {
1084				ret = PTR_ERR(group);
1085				break;
1086			}
1087
1088			groups[i].group = group;
1089			groups[i].id = vfio_external_user_iommu_id(group);
1090		}
1091
1092		kfree(group_fds);
1093
1094		/* release reference to groups on error */
1095		if (ret)
1096			goto hot_reset_release;
1097
1098		info.count = hdr.count;
1099		info.groups = groups;
1100
1101		/*
1102		 * Test whether all the affected devices are contained
1103		 * by the set of groups provided by the user.
1104		 */
1105		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1106						    vfio_pci_validate_devs,
1107						    &info, slot);
1108		if (!ret)
1109			/* User has access, do the reset */
1110			ret = pci_reset_bus(vdev->pdev);
 
1111
1112hot_reset_release:
1113		for (i--; i >= 0; i--)
1114			vfio_group_put_external_user(groups[i].group);
1115
1116		kfree(groups);
1117		return ret;
1118	} else if (cmd == VFIO_DEVICE_IOEVENTFD) {
1119		struct vfio_device_ioeventfd ioeventfd;
1120		int count;
1121
1122		minsz = offsetofend(struct vfio_device_ioeventfd, fd);
1123
1124		if (copy_from_user(&ioeventfd, (void __user *)arg, minsz))
1125			return -EFAULT;
1126
1127		if (ioeventfd.argsz < minsz)
1128			return -EINVAL;
1129
1130		if (ioeventfd.flags & ~VFIO_DEVICE_IOEVENTFD_SIZE_MASK)
1131			return -EINVAL;
1132
1133		count = ioeventfd.flags & VFIO_DEVICE_IOEVENTFD_SIZE_MASK;
1134
1135		if (hweight8(count) != 1 || ioeventfd.fd < -1)
1136			return -EINVAL;
1137
1138		return vfio_pci_ioeventfd(vdev, ioeventfd.offset,
1139					  ioeventfd.data, count, ioeventfd.fd);
1140	}
1141
1142	return -ENOTTY;
1143}
1144
1145static ssize_t vfio_pci_rw(void *device_data, char __user *buf,
1146			   size_t count, loff_t *ppos, bool iswrite)
1147{
1148	unsigned int index = VFIO_PCI_OFFSET_TO_INDEX(*ppos);
1149	struct vfio_pci_device *vdev = device_data;
1150
1151	if (index >= VFIO_PCI_NUM_REGIONS + vdev->num_regions)
1152		return -EINVAL;
1153
1154	switch (index) {
1155	case VFIO_PCI_CONFIG_REGION_INDEX:
1156		return vfio_pci_config_rw(vdev, buf, count, ppos, iswrite);
1157
1158	case VFIO_PCI_ROM_REGION_INDEX:
1159		if (iswrite)
1160			return -EINVAL;
1161		return vfio_pci_bar_rw(vdev, buf, count, ppos, false);
1162
1163	case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
1164		return vfio_pci_bar_rw(vdev, buf, count, ppos, iswrite);
1165
1166	case VFIO_PCI_VGA_REGION_INDEX:
1167		return vfio_pci_vga_rw(vdev, buf, count, ppos, iswrite);
1168	default:
1169		index -= VFIO_PCI_NUM_REGIONS;
1170		return vdev->region[index].ops->rw(vdev, buf,
1171						   count, ppos, iswrite);
1172	}
1173
1174	return -EINVAL;
1175}
1176
1177static ssize_t vfio_pci_read(void *device_data, char __user *buf,
1178			     size_t count, loff_t *ppos)
1179{
1180	if (!count)
1181		return 0;
1182
1183	return vfio_pci_rw(device_data, buf, count, ppos, false);
1184}
1185
1186static ssize_t vfio_pci_write(void *device_data, const char __user *buf,
1187			      size_t count, loff_t *ppos)
1188{
1189	if (!count)
1190		return 0;
1191
1192	return vfio_pci_rw(device_data, (char __user *)buf, count, ppos, true);
1193}
1194
1195static int vfio_pci_mmap(void *device_data, struct vm_area_struct *vma)
1196{
1197	struct vfio_pci_device *vdev = device_data;
1198	struct pci_dev *pdev = vdev->pdev;
1199	unsigned int index;
1200	u64 phys_len, req_len, pgoff, req_start;
1201	int ret;
1202
1203	index = vma->vm_pgoff >> (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT);
1204
1205	if (vma->vm_end < vma->vm_start)
1206		return -EINVAL;
1207	if ((vma->vm_flags & VM_SHARED) == 0)
1208		return -EINVAL;
1209	if (index >= VFIO_PCI_NUM_REGIONS) {
1210		int regnum = index - VFIO_PCI_NUM_REGIONS;
1211		struct vfio_pci_region *region = vdev->region + regnum;
1212
1213		if (region && region->ops && region->ops->mmap &&
1214		    (region->flags & VFIO_REGION_INFO_FLAG_MMAP))
1215			return region->ops->mmap(vdev, region, vma);
1216		return -EINVAL;
1217	}
1218	if (index >= VFIO_PCI_ROM_REGION_INDEX)
1219		return -EINVAL;
1220	if (!vdev->bar_mmap_supported[index])
1221		return -EINVAL;
1222
1223	phys_len = PAGE_ALIGN(pci_resource_len(pdev, index));
1224	req_len = vma->vm_end - vma->vm_start;
1225	pgoff = vma->vm_pgoff &
1226		((1U << (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT)) - 1);
1227	req_start = pgoff << PAGE_SHIFT;
1228
1229	if (req_start + req_len > phys_len)
1230		return -EINVAL;
1231
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1232	/*
1233	 * Even though we don't make use of the barmap for the mmap,
1234	 * we need to request the region and the barmap tracks that.
1235	 */
1236	if (!vdev->barmap[index]) {
1237		ret = pci_request_selected_regions(pdev,
1238						   1 << index, "vfio-pci");
1239		if (ret)
1240			return ret;
1241
1242		vdev->barmap[index] = pci_iomap(pdev, index, 0);
1243		if (!vdev->barmap[index]) {
1244			pci_release_selected_regions(pdev, 1 << index);
1245			return -ENOMEM;
1246		}
1247	}
1248
1249	vma->vm_private_data = vdev;
1250	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1251	vma->vm_pgoff = (pci_resource_start(pdev, index) >> PAGE_SHIFT) + pgoff;
1252
1253	return remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
1254			       req_len, vma->vm_page_prot);
1255}
1256
1257static void vfio_pci_request(void *device_data, unsigned int count)
1258{
1259	struct vfio_pci_device *vdev = device_data;
1260	struct pci_dev *pdev = vdev->pdev;
1261
1262	mutex_lock(&vdev->igate);
1263
1264	if (vdev->req_trigger) {
1265		if (!(count % 10))
1266			pci_notice_ratelimited(pdev,
1267				"Relaying device request to user (#%u)\n",
1268				count);
1269		eventfd_signal(vdev->req_trigger, 1);
1270	} else if (count == 0) {
1271		pci_warn(pdev,
1272			"No device request channel registered, blocked until released by user\n");
1273	}
1274
1275	mutex_unlock(&vdev->igate);
1276}
1277
1278static const struct vfio_device_ops vfio_pci_ops = {
1279	.name		= "vfio-pci",
1280	.open		= vfio_pci_open,
1281	.release	= vfio_pci_release,
1282	.ioctl		= vfio_pci_ioctl,
1283	.read		= vfio_pci_read,
1284	.write		= vfio_pci_write,
1285	.mmap		= vfio_pci_mmap,
1286	.request	= vfio_pci_request,
1287};
1288
1289static int vfio_pci_reflck_attach(struct vfio_pci_device *vdev);
1290static void vfio_pci_reflck_put(struct vfio_pci_reflck *reflck);
1291
1292static int vfio_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
1293{
 
1294	struct vfio_pci_device *vdev;
1295	struct iommu_group *group;
1296	int ret;
1297
1298	if (pdev->hdr_type != PCI_HEADER_TYPE_NORMAL)
 
1299		return -EINVAL;
1300
1301	/*
1302	 * Prevent binding to PFs with VFs enabled, this too easily allows
1303	 * userspace instance with VFs and PFs from the same device, which
1304	 * cannot work.  Disabling SR-IOV here would initiate removing the
1305	 * VFs, which would unbind the driver, which is prone to blocking
1306	 * if that VF is also in use by vfio-pci.  Just reject these PFs
1307	 * and let the user sort it out.
1308	 */
1309	if (pci_num_vf(pdev)) {
1310		pci_warn(pdev, "Cannot bind to PF with SR-IOV enabled\n");
1311		return -EBUSY;
1312	}
1313
1314	group = vfio_iommu_group_get(&pdev->dev);
1315	if (!group)
1316		return -EINVAL;
1317
1318	vdev = kzalloc(sizeof(*vdev), GFP_KERNEL);
1319	if (!vdev) {
1320		vfio_iommu_group_put(group, &pdev->dev);
1321		return -ENOMEM;
1322	}
1323
1324	vdev->pdev = pdev;
1325	vdev->irq_type = VFIO_PCI_NUM_IRQS;
1326	mutex_init(&vdev->igate);
1327	spin_lock_init(&vdev->irqlock);
1328	mutex_init(&vdev->ioeventfds_lock);
1329	INIT_LIST_HEAD(&vdev->ioeventfds_list);
1330
1331	ret = vfio_add_group_dev(&pdev->dev, &vfio_pci_ops, vdev);
1332	if (ret) {
1333		vfio_iommu_group_put(group, &pdev->dev);
1334		kfree(vdev);
1335		return ret;
1336	}
1337
1338	ret = vfio_pci_reflck_attach(vdev);
1339	if (ret) {
1340		vfio_del_group_dev(&pdev->dev);
1341		vfio_iommu_group_put(group, &pdev->dev);
1342		kfree(vdev);
1343		return ret;
1344	}
1345
1346	if (vfio_pci_is_vga(pdev)) {
1347		vga_client_register(pdev, vdev, NULL, vfio_pci_set_vga_decode);
1348		vga_set_legacy_decoding(pdev,
1349					vfio_pci_set_vga_decode(vdev, false));
1350	}
1351
1352	vfio_pci_probe_power_state(vdev);
1353
1354	if (!disable_idle_d3) {
1355		/*
1356		 * pci-core sets the device power state to an unknown value at
1357		 * bootup and after being removed from a driver.  The only
1358		 * transition it allows from this unknown state is to D0, which
1359		 * typically happens when a driver calls pci_enable_device().
1360		 * We're not ready to enable the device yet, but we do want to
1361		 * be able to get to D3.  Therefore first do a D0 transition
1362		 * before going to D3.
1363		 */
1364		vfio_pci_set_power_state(vdev, PCI_D0);
1365		vfio_pci_set_power_state(vdev, PCI_D3hot);
1366	}
1367
1368	return ret;
1369}
1370
1371static void vfio_pci_remove(struct pci_dev *pdev)
1372{
1373	struct vfio_pci_device *vdev;
1374
1375	vdev = vfio_del_group_dev(&pdev->dev);
1376	if (!vdev)
1377		return;
1378
1379	vfio_pci_reflck_put(vdev->reflck);
1380
1381	vfio_iommu_group_put(pdev->dev.iommu_group, &pdev->dev);
1382	kfree(vdev->region);
1383	mutex_destroy(&vdev->ioeventfds_lock);
1384
1385	if (!disable_idle_d3)
1386		vfio_pci_set_power_state(vdev, PCI_D0);
1387
1388	kfree(vdev->pm_save);
1389	kfree(vdev);
1390
1391	if (vfio_pci_is_vga(pdev)) {
1392		vga_client_register(pdev, NULL, NULL, NULL);
1393		vga_set_legacy_decoding(pdev,
1394				VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM |
1395				VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM);
1396	}
1397}
1398
1399static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
1400						  pci_channel_state_t state)
1401{
1402	struct vfio_pci_device *vdev;
1403	struct vfio_device *device;
1404
1405	device = vfio_device_get_from_dev(&pdev->dev);
1406	if (device == NULL)
1407		return PCI_ERS_RESULT_DISCONNECT;
1408
1409	vdev = vfio_device_data(device);
1410	if (vdev == NULL) {
1411		vfio_device_put(device);
1412		return PCI_ERS_RESULT_DISCONNECT;
1413	}
1414
1415	mutex_lock(&vdev->igate);
1416
1417	if (vdev->err_trigger)
1418		eventfd_signal(vdev->err_trigger, 1);
1419
1420	mutex_unlock(&vdev->igate);
1421
1422	vfio_device_put(device);
1423
1424	return PCI_ERS_RESULT_CAN_RECOVER;
1425}
1426
1427static const struct pci_error_handlers vfio_err_handlers = {
1428	.error_detected = vfio_pci_aer_err_detected,
1429};
1430
1431static struct pci_driver vfio_pci_driver = {
1432	.name		= "vfio-pci",
1433	.id_table	= NULL, /* only dynamic ids */
1434	.probe		= vfio_pci_probe,
1435	.remove		= vfio_pci_remove,
1436	.err_handler	= &vfio_err_handlers,
1437};
1438
1439static DEFINE_MUTEX(reflck_lock);
1440
1441static struct vfio_pci_reflck *vfio_pci_reflck_alloc(void)
1442{
1443	struct vfio_pci_reflck *reflck;
1444
1445	reflck = kzalloc(sizeof(*reflck), GFP_KERNEL);
1446	if (!reflck)
1447		return ERR_PTR(-ENOMEM);
1448
1449	kref_init(&reflck->kref);
1450	mutex_init(&reflck->lock);
1451
1452	return reflck;
1453}
1454
1455static void vfio_pci_reflck_get(struct vfio_pci_reflck *reflck)
1456{
1457	kref_get(&reflck->kref);
1458}
1459
1460static int vfio_pci_reflck_find(struct pci_dev *pdev, void *data)
1461{
1462	struct vfio_pci_reflck **preflck = data;
1463	struct vfio_device *device;
1464	struct vfio_pci_device *vdev;
1465
1466	device = vfio_device_get_from_dev(&pdev->dev);
1467	if (!device)
1468		return 0;
1469
1470	if (pci_dev_driver(pdev) != &vfio_pci_driver) {
1471		vfio_device_put(device);
1472		return 0;
1473	}
1474
1475	vdev = vfio_device_data(device);
1476
1477	if (vdev->reflck) {
1478		vfio_pci_reflck_get(vdev->reflck);
1479		*preflck = vdev->reflck;
1480		vfio_device_put(device);
1481		return 1;
1482	}
1483
1484	vfio_device_put(device);
1485	return 0;
1486}
1487
1488static int vfio_pci_reflck_attach(struct vfio_pci_device *vdev)
1489{
1490	bool slot = !pci_probe_reset_slot(vdev->pdev->slot);
1491
1492	mutex_lock(&reflck_lock);
1493
1494	if (pci_is_root_bus(vdev->pdev->bus) ||
1495	    vfio_pci_for_each_slot_or_bus(vdev->pdev, vfio_pci_reflck_find,
1496					  &vdev->reflck, slot) <= 0)
1497		vdev->reflck = vfio_pci_reflck_alloc();
1498
1499	mutex_unlock(&reflck_lock);
1500
1501	return PTR_ERR_OR_ZERO(vdev->reflck);
1502}
1503
1504static void vfio_pci_reflck_release(struct kref *kref)
1505{
1506	struct vfio_pci_reflck *reflck = container_of(kref,
1507						      struct vfio_pci_reflck,
1508						      kref);
1509
1510	kfree(reflck);
1511	mutex_unlock(&reflck_lock);
1512}
1513
1514static void vfio_pci_reflck_put(struct vfio_pci_reflck *reflck)
1515{
1516	kref_put_mutex(&reflck->kref, vfio_pci_reflck_release, &reflck_lock);
1517}
1518
1519struct vfio_devices {
1520	struct vfio_device **devices;
1521	int cur_index;
1522	int max_index;
1523};
1524
1525static int vfio_pci_get_unused_devs(struct pci_dev *pdev, void *data)
1526{
1527	struct vfio_devices *devs = data;
1528	struct vfio_device *device;
1529	struct vfio_pci_device *vdev;
1530
1531	if (devs->cur_index == devs->max_index)
1532		return -ENOSPC;
1533
1534	device = vfio_device_get_from_dev(&pdev->dev);
1535	if (!device)
1536		return -EINVAL;
1537
1538	if (pci_dev_driver(pdev) != &vfio_pci_driver) {
1539		vfio_device_put(device);
1540		return -EBUSY;
1541	}
1542
1543	vdev = vfio_device_data(device);
1544
1545	/* Fault if the device is not unused */
1546	if (vdev->refcnt) {
1547		vfio_device_put(device);
1548		return -EBUSY;
1549	}
1550
1551	devs->devices[devs->cur_index++] = device;
1552	return 0;
1553}
1554
1555/*
1556 * If a bus or slot reset is available for the provided device and:
1557 *  - All of the devices affected by that bus or slot reset are unused
1558 *    (!refcnt)
1559 *  - At least one of the affected devices is marked dirty via
1560 *    needs_reset (such as by lack of FLR support)
1561 * Then attempt to perform that bus or slot reset.  Callers are required
1562 * to hold vdev->reflck->lock, protecting the bus/slot reset group from
1563 * concurrent opens.  A vfio_device reference is acquired for each device
1564 * to prevent unbinds during the reset operation.
1565 *
1566 * NB: vfio-core considers a group to be viable even if some devices are
1567 * bound to drivers like pci-stub or pcieport.  Here we require all devices
1568 * to be bound to vfio_pci since that's the only way we can be sure they
1569 * stay put.
1570 */
1571static void vfio_pci_try_bus_reset(struct vfio_pci_device *vdev)
1572{
1573	struct vfio_devices devs = { .cur_index = 0 };
1574	int i = 0, ret = -EINVAL;
1575	bool slot = false;
1576	struct vfio_pci_device *tmp;
1577
1578	if (!pci_probe_reset_slot(vdev->pdev->slot))
1579		slot = true;
1580	else if (pci_probe_reset_bus(vdev->pdev->bus))
1581		return;
1582
1583	if (vfio_pci_for_each_slot_or_bus(vdev->pdev, vfio_pci_count_devs,
1584					  &i, slot) || !i)
1585		return;
1586
1587	devs.max_index = i;
1588	devs.devices = kcalloc(i, sizeof(struct vfio_device *), GFP_KERNEL);
1589	if (!devs.devices)
1590		return;
1591
1592	if (vfio_pci_for_each_slot_or_bus(vdev->pdev,
1593					  vfio_pci_get_unused_devs,
1594					  &devs, slot))
1595		goto put_devs;
1596
1597	/* Does at least one need a reset? */
1598	for (i = 0; i < devs.cur_index; i++) {
1599		tmp = vfio_device_data(devs.devices[i]);
1600		if (tmp->needs_reset) {
1601			ret = pci_reset_bus(vdev->pdev);
1602			break;
1603		}
1604	}
1605
1606put_devs:
1607	for (i = 0; i < devs.cur_index; i++) {
1608		tmp = vfio_device_data(devs.devices[i]);
1609
1610		/*
1611		 * If reset was successful, affected devices no longer need
1612		 * a reset and we should return all the collateral devices
1613		 * to low power.  If not successful, we either didn't reset
1614		 * the bus or timed out waiting for it, so let's not touch
1615		 * the power state.
1616		 */
1617		if (!ret) {
1618			tmp->needs_reset = false;
1619
1620			if (tmp != vdev && !disable_idle_d3)
1621				vfio_pci_set_power_state(tmp, PCI_D3hot);
1622		}
1623
1624		vfio_device_put(devs.devices[i]);
1625	}
1626
1627	kfree(devs.devices);
1628}
1629
1630static void __exit vfio_pci_cleanup(void)
1631{
1632	pci_unregister_driver(&vfio_pci_driver);
 
1633	vfio_pci_uninit_perm_bits();
1634}
1635
1636static void __init vfio_pci_fill_ids(void)
1637{
1638	char *p, *id;
1639	int rc;
1640
1641	/* no ids passed actually */
1642	if (ids[0] == '\0')
1643		return;
1644
1645	/* add ids specified in the module parameter */
1646	p = ids;
1647	while ((id = strsep(&p, ","))) {
1648		unsigned int vendor, device, subvendor = PCI_ANY_ID,
1649			subdevice = PCI_ANY_ID, class = 0, class_mask = 0;
1650		int fields;
1651
1652		if (!strlen(id))
1653			continue;
1654
1655		fields = sscanf(id, "%x:%x:%x:%x:%x:%x",
1656				&vendor, &device, &subvendor, &subdevice,
1657				&class, &class_mask);
1658
1659		if (fields < 2) {
1660			pr_warn("invalid id string \"%s\"\n", id);
1661			continue;
1662		}
1663
1664		rc = pci_add_dynid(&vfio_pci_driver, vendor, device,
1665				   subvendor, subdevice, class, class_mask, 0);
1666		if (rc)
1667			pr_warn("failed to add dynamic id [%04x:%04x[%04x:%04x]] class %#08x/%08x (%d)\n",
1668				vendor, device, subvendor, subdevice,
1669				class, class_mask, rc);
1670		else
1671			pr_info("add [%04x:%04x[%04x:%04x]] class %#08x/%08x\n",
1672				vendor, device, subvendor, subdevice,
1673				class, class_mask);
1674	}
1675}
1676
1677static int __init vfio_pci_init(void)
1678{
1679	int ret;
1680
1681	/* Allocate shared config space permision data used by all devices */
1682	ret = vfio_pci_init_perm_bits();
1683	if (ret)
1684		return ret;
1685
 
 
 
 
 
1686	/* Register and scan for devices */
1687	ret = pci_register_driver(&vfio_pci_driver);
1688	if (ret)
1689		goto out_driver;
1690
1691	vfio_pci_fill_ids();
1692
1693	return 0;
1694
1695out_driver:
 
 
1696	vfio_pci_uninit_perm_bits();
1697	return ret;
1698}
1699
1700module_init(vfio_pci_init);
1701module_exit(vfio_pci_cleanup);
1702
1703MODULE_VERSION(DRIVER_VERSION);
1704MODULE_LICENSE("GPL v2");
1705MODULE_AUTHOR(DRIVER_AUTHOR);
1706MODULE_DESCRIPTION(DRIVER_DESC);