drivers/vfio/pci/vfio_pci.c

   1 /*
   2  * Copyright (C) 2012 Red Hat, Inc.  All rights reserved.
   3  *     Author: Alex Williamson <alex.williamson@redhat.com>
   4  *
   5  * This program is free software; you can redistribute it and/or modify
   6  * it under the terms of the GNU General Public License version 2 as
   7  * published by the Free Software Foundation.
   8  *
   9  * Derived from original vfio:
  10  * Copyright 2010 Cisco Systems, Inc.  All rights reserved.
  11  * Author: Tom Lyon, pugs@cisco.com
  12  */
  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
  29 #include "vfio_pci_private.h"
  30
  31 #define DRIVER_VERSION  "0.2"
  32 #define DRIVER_AUTHOR   "Alex Williamson <alex.williamson@redhat.com>"
  33 #define DRIVER_DESC     "VFIO PCI - User Level meta-driver"
  34
  35 static bool nointxmask;
  36 module_param_named(nointxmask, nointxmask, bool, S_IRUGO | S_IWUSR);
  37 MODULE_PARM_DESC(nointxmask,
  38                   "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.");
  39
  40 static DEFINE_MUTEX(driver_lock);
  41
  42 static void vfio_pci_try_bus_reset(struct vfio_pci_device *vdev);
  43
  44 static int vfio_pci_enable(struct vfio_pci_device *vdev)
  45 {
  46         struct pci_dev *pdev = vdev->pdev;
  47         int ret;
  48         u16 cmd;
  49         u8 msix_pos;
  50
  51         /* Don't allow our initial saved state to include busmaster */
  52         pci_clear_master(pdev);
  53
  54         ret = pci_enable_device(pdev);
  55         if (ret)
  56                 return ret;
  57
  58         vdev->reset_works = (pci_reset_function(pdev) == 0);
  59         pci_save_state(pdev);
  60         vdev->pci_saved_state = pci_store_saved_state(pdev);
  61         if (!vdev->pci_saved_state)
  62                 pr_debug("%s: Couldn't store %s saved state\n",
  63                          __func__, dev_name(&pdev->dev));
  64
  65         ret = vfio_config_init(vdev);
  66         if (ret) {
  67                 kfree(vdev->pci_saved_state);
  68                 vdev->pci_saved_state = NULL;
  69                 pci_disable_device(pdev);
  70                 return ret;
  71         }
  72
  73         if (likely(!nointxmask))
  74                 vdev->pci_2_3 = pci_intx_mask_supported(pdev);
  75
  76         pci_read_config_word(pdev, PCI_COMMAND, &cmd);
  77         if (vdev->pci_2_3 && (cmd & PCI_COMMAND_INTX_DISABLE)) {
  78                 cmd &= ~PCI_COMMAND_INTX_DISABLE;
  79                 pci_write_config_word(pdev, PCI_COMMAND, cmd);
  80         }
  81
  82         msix_pos = pdev->msix_cap;
  83         if (msix_pos) {
  84                 u16 flags;
  85                 u32 table;
  86
  87                 pci_read_config_word(pdev, msix_pos + PCI_MSIX_FLAGS, &flags);
  88                 pci_read_config_dword(pdev, msix_pos + PCI_MSIX_TABLE, &table);
  89
  90                 vdev->msix_bar = table & PCI_MSIX_TABLE_BIR;
  91                 vdev->msix_offset = table & PCI_MSIX_TABLE_OFFSET;
  92                 vdev->msix_size = ((flags & PCI_MSIX_FLAGS_QSIZE) + 1) * 16;
  93         } else
  94                 vdev->msix_bar = 0xFF;
  95
  96 #ifdef CONFIG_VFIO_PCI_VGA
  97         if ((pdev->class >> 8) == PCI_CLASS_DISPLAY_VGA)
  98                 vdev->has_vga = true;
  99 #endif
 100
 101         return 0;
 102 }
 103
 104 static void vfio_pci_disable(struct vfio_pci_device *vdev)
 105 {
 106         struct pci_dev *pdev = vdev->pdev;
 107         int bar;
 108
 109         /* Stop the device from further DMA */
 110         pci_clear_master(pdev);
 111
 112         vfio_pci_set_irqs_ioctl(vdev, VFIO_IRQ_SET_DATA_NONE |
 113                                 VFIO_IRQ_SET_ACTION_TRIGGER,
 114                                 vdev->irq_type, 0, 0, NULL);
 115
 116         vdev->virq_disabled = false;
 117
 118         vfio_config_free(vdev);
 119
 120         for (bar = PCI_STD_RESOURCES; bar <= PCI_STD_RESOURCE_END; bar++) {
 121                 if (!vdev->barmap[bar])
 122                         continue;
 123                 pci_iounmap(pdev, vdev->barmap[bar]);
 124                 pci_release_selected_regions(pdev, 1 << bar);
 125                 vdev->barmap[bar] = NULL;
 126         }
 127
 128         vdev->needs_reset = true;
 129
 130         /*
 131          * If we have saved state, restore it.  If we can reset the device,
 132          * even better.  Resetting with current state seems better than
 133          * nothing, but saving and restoring current state without reset
 134          * is just busy work.
 135          */
 136         if (pci_load_and_free_saved_state(pdev, &vdev->pci_saved_state)) {
 137                 pr_info("%s: Couldn't reload %s saved state\n",
 138                         __func__, dev_name(&pdev->dev));
 139
 140                 if (!vdev->reset_works)
 141                         goto out;
 142
 143                 pci_save_state(pdev);
 144         }
 145
 146         /*
 147          * Disable INTx and MSI, presumably to avoid spurious interrupts
 148          * during reset.  Stolen from pci_reset_function()
 149          */
 150         pci_write_config_word(pdev, PCI_COMMAND, PCI_COMMAND_INTX_DISABLE);
 151
 152         /*
 153          * Try to reset the device.  The success of this is dependent on
 154          * being able to lock the device, which is not always possible.
 155          */
 156         if (vdev->reset_works) {
 157                 int ret = pci_try_reset_function(pdev);
 158                 if (ret)
 159                         pr_warn("%s: Failed to reset device %s (%d)\n",
 160                                 __func__, dev_name(&pdev->dev), ret);
 161                 else
 162                         vdev->needs_reset = false;
 163         }
 164
 165         pci_restore_state(pdev);
 166 out:
 167         pci_disable_device(pdev);
 168
 169         vfio_pci_try_bus_reset(vdev);
 170 }
 171
 172 static void vfio_pci_release(void *device_data)
 173 {
 174         struct vfio_pci_device *vdev = device_data;
 175
 176         mutex_lock(&driver_lock);
 177
 178         if (!(--vdev->refcnt)) {
 179                 vfio_spapr_pci_eeh_release(vdev->pdev);
 180                 vfio_pci_disable(vdev);
 181         }
 182
 183         mutex_unlock(&driver_lock);
 184
 185         module_put(THIS_MODULE);
 186 }
 187
 188 static int vfio_pci_open(void *device_data)
 189 {
 190         struct vfio_pci_device *vdev = device_data;
 191         int ret = 0;
 192
 193         if (!try_module_get(THIS_MODULE))
 194                 return -ENODEV;
 195
 196         mutex_lock(&driver_lock);
 197
 198         if (!vdev->refcnt) {
 199                 ret = vfio_pci_enable(vdev);
 200                 if (ret)
 201                         goto error;
 202
 203                 vfio_spapr_pci_eeh_open(vdev->pdev);
 204         }
 205         vdev->refcnt++;
 206 error:
 207         mutex_unlock(&driver_lock);
 208         if (ret)
 209                 module_put(THIS_MODULE);
 210         return ret;
 211 }
 212
 213 static int vfio_pci_get_irq_count(struct vfio_pci_device *vdev, int irq_type)
 214 {
 215         if (irq_type == VFIO_PCI_INTX_IRQ_INDEX) {
 216                 u8 pin;
 217                 pci_read_config_byte(vdev->pdev, PCI_INTERRUPT_PIN, &pin);
 218                 if (IS_ENABLED(CONFIG_VFIO_PCI_INTX) && pin)
 219                         return 1;
 220
 221         } else if (irq_type == VFIO_PCI_MSI_IRQ_INDEX) {
 222                 u8 pos;
 223                 u16 flags;
 224
 225                 pos = vdev->pdev->msi_cap;
 226                 if (pos) {
 227                         pci_read_config_word(vdev->pdev,
 228                                              pos + PCI_MSI_FLAGS, &flags);
 229                         return 1 << ((flags & PCI_MSI_FLAGS_QMASK) >> 1);
 230                 }
 231         } else if (irq_type == VFIO_PCI_MSIX_IRQ_INDEX) {
 232                 u8 pos;
 233                 u16 flags;
 234
 235                 pos = vdev->pdev->msix_cap;
 236                 if (pos) {
 237                         pci_read_config_word(vdev->pdev,
 238                                              pos + PCI_MSIX_FLAGS, &flags);
 239
 240                         return (flags & PCI_MSIX_FLAGS_QSIZE) + 1;
 241                 }
 242         } else if (irq_type == VFIO_PCI_ERR_IRQ_INDEX)
 243                 if (pci_is_pcie(vdev->pdev))
 244                         return 1;
 245
 246         return 0;
 247 }
 248
 249 static int vfio_pci_count_devs(struct pci_dev *pdev, void *data)
 250 {
 251         (*(int *)data)++;
 252         return 0;
 253 }
 254
 255 struct vfio_pci_fill_info {
 256         int max;
 257         int cur;
 258         struct vfio_pci_dependent_device *devices;
 259 };
 260
 261 static int vfio_pci_fill_devs(struct pci_dev *pdev, void *data)
 262 {
 263         struct vfio_pci_fill_info *fill = data;
 264         struct iommu_group *iommu_group;
 265
 266         if (fill->cur == fill->max)
 267                 return -EAGAIN; /* Something changed, try again */
 268
 269         iommu_group = iommu_group_get(&pdev->dev);
 270         if (!iommu_group)
 271                 return -EPERM; /* Cannot reset non-isolated devices */
 272
 273         fill->devices[fill->cur].group_id = iommu_group_id(iommu_group);
 274         fill->devices[fill->cur].segment = pci_domain_nr(pdev->bus);
 275         fill->devices[fill->cur].bus = pdev->bus->number;
 276         fill->devices[fill->cur].devfn = pdev->devfn;
 277         fill->cur++;
 278         iommu_group_put(iommu_group);
 279         return 0;
 280 }
 281
 282 struct vfio_pci_group_entry {
 283         struct vfio_group *group;
 284         int id;
 285 };
 286
 287 struct vfio_pci_group_info {
 288         int count;
 289         struct vfio_pci_group_entry *groups;
 290 };
 291
 292 static int vfio_pci_validate_devs(struct pci_dev *pdev, void *data)
 293 {
 294         struct vfio_pci_group_info *info = data;
 295         struct iommu_group *group;
 296         int id, i;
 297
 298         group = iommu_group_get(&pdev->dev);
 299         if (!group)
 300                 return -EPERM;
 301
 302         id = iommu_group_id(group);
 303
 304         for (i = 0; i < info->count; i++)
 305                 if (info->groups[i].id == id)
 306                         break;
 307
 308         iommu_group_put(group);
 309
 310         return (i == info->count) ? -EINVAL : 0;
 311 }
 312
 313 static bool vfio_pci_dev_below_slot(struct pci_dev *pdev, struct pci_slot *slot)
 314 {
 315         for (; pdev; pdev = pdev->bus->self)
 316                 if (pdev->bus == slot->bus)
 317                         return (pdev->slot == slot);
 318         return false;
 319 }
 320
 321 struct vfio_pci_walk_info {
 322         int (*fn)(struct pci_dev *, void *data);
 323         void *data;
 324         struct pci_dev *pdev;
 325         bool slot;
 326         int ret;
 327 };
 328
 329 static int vfio_pci_walk_wrapper(struct pci_dev *pdev, void *data)
 330 {
 331         struct vfio_pci_walk_info *walk = data;
 332
 333         if (!walk->slot || vfio_pci_dev_below_slot(pdev, walk->pdev->slot))
 334                 walk->ret = walk->fn(pdev, walk->data);
 335
 336         return walk->ret;
 337 }
 338
 339 static int vfio_pci_for_each_slot_or_bus(struct pci_dev *pdev,
 340                                          int (*fn)(struct pci_dev *,
 341                                                    void *data), void *data,
 342                                          bool slot)
 343 {
 344         struct vfio_pci_walk_info walk = {
 345                 .fn = fn, .data = data, .pdev = pdev, .slot = slot, .ret = 0,
 346         };
 347
 348         pci_walk_bus(pdev->bus, vfio_pci_walk_wrapper, &walk);
 349
 350         return walk.ret;
 351 }
 352
 353 static long vfio_pci_ioctl(void *device_data,
 354                            unsigned int cmd, unsigned long arg)
 355 {
 356         struct vfio_pci_device *vdev = device_data;
 357         unsigned long minsz;
 358
 359         if (cmd == VFIO_DEVICE_GET_INFO) {
 360                 struct vfio_device_info info;
 361
 362                 minsz = offsetofend(struct vfio_device_info, num_irqs);
 363
 364                 if (copy_from_user(&info, (void __user *)arg, minsz))
 365                         return -EFAULT;
 366
 367                 if (info.argsz < minsz)
 368                         return -EINVAL;
 369
 370                 info.flags = VFIO_DEVICE_FLAGS_PCI;
 371
 372                 if (vdev->reset_works)
 373                         info.flags |= VFIO_DEVICE_FLAGS_RESET;
 374
 375                 info.num_regions = VFIO_PCI_NUM_REGIONS;
 376                 info.num_irqs = VFIO_PCI_NUM_IRQS;
 377
 378                 return copy_to_user((void __user *)arg, &info, minsz);
 379
 380         } else if (cmd == VFIO_DEVICE_GET_REGION_INFO) {
 381                 struct pci_dev *pdev = vdev->pdev;
 382                 struct vfio_region_info info;
 383
 384                 minsz = offsetofend(struct vfio_region_info, offset);
 385
 386                 if (copy_from_user(&info, (void __user *)arg, minsz))
 387                         return -EFAULT;
 388
 389                 if (info.argsz < minsz)
 390                         return -EINVAL;
 391
 392                 switch (info.index) {
 393                 case VFIO_PCI_CONFIG_REGION_INDEX:
 394                         info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
 395                         info.size = pdev->cfg_size;
 396                         info.flags = VFIO_REGION_INFO_FLAG_READ |
 397                                      VFIO_REGION_INFO_FLAG_WRITE;
 398                         break;
 399                 case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
 400                         info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
 401                         info.size = pci_resource_len(pdev, info.index);
 402                         if (!info.size) {
 403                                 info.flags = 0;
 404                                 break;
 405                         }
 406
 407                         info.flags = VFIO_REGION_INFO_FLAG_READ |
 408                                      VFIO_REGION_INFO_FLAG_WRITE;
 409                         if (IS_ENABLED(CONFIG_VFIO_PCI_MMAP) &&
 410                             pci_resource_flags(pdev, info.index) &
 411                             IORESOURCE_MEM && info.size >= PAGE_SIZE)
 412                                 info.flags |= VFIO_REGION_INFO_FLAG_MMAP;
 413                         break;
 414                 case VFIO_PCI_ROM_REGION_INDEX:
 415                 {
 416                         void __iomem *io;
 417                         size_t size;
 418
 419                         info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
 420                         info.flags = 0;
 421
 422                         /* Report the BAR size, not the ROM size */
 423                         info.size = pci_resource_len(pdev, info.index);
 424                         if (!info.size)
 425                                 break;
 426
 427                         /* Is it really there? */
 428                         io = pci_map_rom(pdev, &size);
 429                         if (!io || !size) {
 430                                 info.size = 0;
 431                                 break;
 432                         }
 433                         pci_unmap_rom(pdev, io);
 434
 435                         info.flags = VFIO_REGION_INFO_FLAG_READ;
 436                         break;
 437                 }
 438                 case VFIO_PCI_VGA_REGION_INDEX:
 439                         if (!vdev->has_vga)
 440                                 return -EINVAL;
 441
 442                         info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
 443                         info.size = 0xc0000;
 444                         info.flags = VFIO_REGION_INFO_FLAG_READ |
 445                                      VFIO_REGION_INFO_FLAG_WRITE;
 446
 447                         break;
 448                 default:
 449                         return -EINVAL;
 450                 }
 451
 452                 return copy_to_user((void __user *)arg, &info, minsz);
 453
 454         } else if (cmd == VFIO_DEVICE_GET_IRQ_INFO) {
 455                 struct vfio_irq_info info;
 456
 457                 minsz = offsetofend(struct vfio_irq_info, count);
 458
 459                 if (copy_from_user(&info, (void __user *)arg, minsz))
 460                         return -EFAULT;
 461
 462                 if (info.argsz < minsz || info.index >= VFIO_PCI_NUM_IRQS)
 463                         return -EINVAL;
 464
 465                 switch (info.index) {
 466                 case VFIO_PCI_INTX_IRQ_INDEX ... VFIO_PCI_MSIX_IRQ_INDEX:
 467                         break;
 468                 case VFIO_PCI_ERR_IRQ_INDEX:
 469                         if (pci_is_pcie(vdev->pdev))
 470                                 break;
 471                 /* pass thru to return error */
 472                 default:
 473                         return -EINVAL;
 474                 }
 475
 476                 info.flags = VFIO_IRQ_INFO_EVENTFD;
 477
 478                 info.count = vfio_pci_get_irq_count(vdev, info.index);
 479
 480                 if (info.index == VFIO_PCI_INTX_IRQ_INDEX)
 481                         info.flags |= (VFIO_IRQ_INFO_MASKABLE |
 482                                        VFIO_IRQ_INFO_AUTOMASKED);
 483                 else
 484                         info.flags |= VFIO_IRQ_INFO_NORESIZE;
 485
 486                 return copy_to_user((void __user *)arg, &info, minsz);
 487
 488         } else if (cmd == VFIO_DEVICE_SET_IRQS) {
 489                 struct vfio_irq_set hdr;
 490                 u8 *data = NULL;
 491                 int ret = 0;
 492
 493                 minsz = offsetofend(struct vfio_irq_set, count);
 494
 495                 if (copy_from_user(&hdr, (void __user *)arg, minsz))
 496                         return -EFAULT;
 497
 498                 if (hdr.argsz < minsz || hdr.index >= VFIO_PCI_NUM_IRQS ||
 499                     hdr.flags & ~(VFIO_IRQ_SET_DATA_TYPE_MASK |
 500                                   VFIO_IRQ_SET_ACTION_TYPE_MASK))
 501                         return -EINVAL;
 502
 503                 if (!(hdr.flags & VFIO_IRQ_SET_DATA_NONE)) {
 504                         size_t size;
 505                         int max = vfio_pci_get_irq_count(vdev, hdr.index);
 506
 507                         if (hdr.flags & VFIO_IRQ_SET_DATA_BOOL)
 508                                 size = sizeof(uint8_t);
 509                         else if (hdr.flags & VFIO_IRQ_SET_DATA_EVENTFD)
 510                                 size = sizeof(int32_t);
 511                         else
 512                                 return -EINVAL;
 513
 514                         if (hdr.argsz - minsz < hdr.count * size ||
 515                             hdr.start >= max || hdr.start + hdr.count > max)
 516                                 return -EINVAL;
 517
 518                         data = memdup_user((void __user *)(arg + minsz),
 519                                            hdr.count * size);
 520                         if (IS_ERR(data))
 521                                 return PTR_ERR(data);
 522                 }
 523
 524                 mutex_lock(&vdev->igate);
 525
 526                 ret = vfio_pci_set_irqs_ioctl(vdev, hdr.flags, hdr.index,
 527                                               hdr.start, hdr.count, data);
 528
 529                 mutex_unlock(&vdev->igate);
 530                 kfree(data);
 531
 532                 return ret;
 533
 534         } else if (cmd == VFIO_DEVICE_RESET) {
 535                 return vdev->reset_works ?
 536                         pci_try_reset_function(vdev->pdev) : -EINVAL;
 537
 538         } else if (cmd == VFIO_DEVICE_GET_PCI_HOT_RESET_INFO) {
 539                 struct vfio_pci_hot_reset_info hdr;
 540                 struct vfio_pci_fill_info fill = { 0 };
 541                 struct vfio_pci_dependent_device *devices = NULL;
 542                 bool slot = false;
 543                 int ret = 0;
 544
 545                 minsz = offsetofend(struct vfio_pci_hot_reset_info, count);
 546
 547                 if (copy_from_user(&hdr, (void __user *)arg, minsz))
 548                         return -EFAULT;
 549
 550                 if (hdr.argsz < minsz)
 551                         return -EINVAL;
 552
 553                 hdr.flags = 0;
 554
 555                 /* Can we do a slot or bus reset or neither? */
 556                 if (!pci_probe_reset_slot(vdev->pdev->slot))
 557                         slot = true;
 558                 else if (pci_probe_reset_bus(vdev->pdev->bus))
 559                         return -ENODEV;
 560
 561                 /* How many devices are affected? */
 562                 ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
 563                                                     vfio_pci_count_devs,
 564                                                     &fill.max, slot);
 565                 if (ret)
 566                         return ret;
 567
 568                 WARN_ON(!fill.max); /* Should always be at least one */
 569
 570                 /*
 571                  * If there's enough space, fill it now, otherwise return
 572                  * -ENOSPC and the number of devices affected.
 573                  */
 574                 if (hdr.argsz < sizeof(hdr) + (fill.max * sizeof(*devices))) {
 575                         ret = -ENOSPC;
 576                         hdr.count = fill.max;
 577                         goto reset_info_exit;
 578                 }
 579
 580                 devices = kcalloc(fill.max, sizeof(*devices), GFP_KERNEL);
 581                 if (!devices)
 582                         return -ENOMEM;
 583
 584                 fill.devices = devices;
 585
 586                 ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
 587                                                     vfio_pci_fill_devs,
 588                                                     &fill, slot);
 589
 590                 /*
 591                  * If a device was removed between counting and filling,
 592                  * we may come up short of fill.max.  If a device was
 593                  * added, we'll have a return of -EAGAIN above.
 594                  */
 595                 if (!ret)
 596                         hdr.count = fill.cur;
 597
 598 reset_info_exit:
 599                 if (copy_to_user((void __user *)arg, &hdr, minsz))
 600                         ret = -EFAULT;
 601
 602                 if (!ret) {
 603                         if (copy_to_user((void __user *)(arg + minsz), devices,
 604                                          hdr.count * sizeof(*devices)))
 605                                 ret = -EFAULT;
 606                 }
 607
 608                 kfree(devices);
 609                 return ret;
 610
 611         } else if (cmd == VFIO_DEVICE_PCI_HOT_RESET) {
 612                 struct vfio_pci_hot_reset hdr;
 613                 int32_t *group_fds;
 614                 struct vfio_pci_group_entry *groups;
 615                 struct vfio_pci_group_info info;
 616                 bool slot = false;
 617                 int i, count = 0, ret = 0;
 618
 619                 minsz = offsetofend(struct vfio_pci_hot_reset, count);
 620
 621                 if (copy_from_user(&hdr, (void __user *)arg, minsz))
 622                         return -EFAULT;
 623
 624                 if (hdr.argsz < minsz || hdr.flags)
 625                         return -EINVAL;
 626
 627                 /* Can we do a slot or bus reset or neither? */
 628                 if (!pci_probe_reset_slot(vdev->pdev->slot))
 629                         slot = true;
 630                 else if (pci_probe_reset_bus(vdev->pdev->bus))
 631                         return -ENODEV;
 632
 633                 /*
 634                  * We can't let userspace give us an arbitrarily large
 635                  * buffer to copy, so verify how many we think there
 636                  * could be.  Note groups can have multiple devices so
 637                  * one group per device is the max.
 638                  */
 639                 ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
 640                                                     vfio_pci_count_devs,
 641                                                     &count, slot);
 642                 if (ret)
 643                         return ret;
 644
 645                 /* Somewhere between 1 and count is OK */
 646                 if (!hdr.count || hdr.count > count)
 647                         return -EINVAL;
 648
 649                 group_fds = kcalloc(hdr.count, sizeof(*group_fds), GFP_KERNEL);
 650                 groups = kcalloc(hdr.count, sizeof(*groups), GFP_KERNEL);
 651                 if (!group_fds || !groups) {
 652                         kfree(group_fds);
 653                         kfree(groups);
 654                         return -ENOMEM;
 655                 }
 656
 657                 if (copy_from_user(group_fds, (void __user *)(arg + minsz),
 658                                    hdr.count * sizeof(*group_fds))) {
 659                         kfree(group_fds);
 660                         kfree(groups);
 661                         return -EFAULT;
 662                 }
 663
 664                 /*
 665                  * For each group_fd, get the group through the vfio external
 666                  * user interface and store the group and iommu ID.  This
 667                  * ensures the group is held across the reset.
 668                  */
 669                 for (i = 0; i < hdr.count; i++) {
 670                         struct vfio_group *group;
 671                         struct fd f = fdget(group_fds[i]);
 672                         if (!f.file) {
 673                                 ret = -EBADF;
 674                                 break;
 675                         }
 676
 677                         group = vfio_group_get_external_user(f.file);
 678                         fdput(f);
 679                         if (IS_ERR(group)) {
 680                                 ret = PTR_ERR(group);
 681                                 break;
 682                         }
 683
 684                         groups[i].group = group;
 685                         groups[i].id = vfio_external_user_iommu_id(group);
 686                 }
 687
 688                 kfree(group_fds);
 689
 690                 /* release reference to groups on error */
 691                 if (ret)
 692                         goto hot_reset_release;
 693
 694                 info.count = hdr.count;
 695                 info.groups = groups;
 696
 697                 /*
 698                  * Test whether all the affected devices are contained
 699                  * by the set of groups provided by the user.
 700                  */
 701                 ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
 702                                                     vfio_pci_validate_devs,
 703                                                     &info, slot);
 704                 if (!ret)
 705                         /* User has access, do the reset */
 706                         ret = slot ? pci_try_reset_slot(vdev->pdev->slot) :
 707                                      pci_try_reset_bus(vdev->pdev->bus);
 708
 709 hot_reset_release:
 710                 for (i--; i >= 0; i--)
 711                         vfio_group_put_external_user(groups[i].group);
 712
 713                 kfree(groups);
 714                 return ret;
 715         }
 716
 717         return -ENOTTY;
 718 }
 719
 720 static ssize_t vfio_pci_rw(void *device_data, char __user *buf,
 721                            size_t count, loff_t *ppos, bool iswrite)
 722 {
 723         unsigned int index = VFIO_PCI_OFFSET_TO_INDEX(*ppos);
 724         struct vfio_pci_device *vdev = device_data;
 725
 726         if (index >= VFIO_PCI_NUM_REGIONS)
 727                 return -EINVAL;
 728
 729         switch (index) {
 730         case VFIO_PCI_CONFIG_REGION_INDEX:
 731                 return vfio_pci_config_rw(vdev, buf, count, ppos, iswrite);
 732
 733         case VFIO_PCI_ROM_REGION_INDEX:
 734                 if (iswrite)
 735                         return -EINVAL;
 736                 return vfio_pci_bar_rw(vdev, buf, count, ppos, false);
 737
 738         case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
 739                 return vfio_pci_bar_rw(vdev, buf, count, ppos, iswrite);
 740
 741         case VFIO_PCI_VGA_REGION_INDEX:
 742                 return vfio_pci_vga_rw(vdev, buf, count, ppos, iswrite);
 743         }
 744
 745         return -EINVAL;
 746 }
 747
 748 static ssize_t vfio_pci_read(void *device_data, char __user *buf,
 749                              size_t count, loff_t *ppos)
 750 {
 751         if (!count)
 752                 return 0;
 753
 754         return vfio_pci_rw(device_data, buf, count, ppos, false);
 755 }
 756
 757 static ssize_t vfio_pci_write(void *device_data, const char __user *buf,
 758                               size_t count, loff_t *ppos)
 759 {
 760         if (!count)
 761                 return 0;
 762
 763         return vfio_pci_rw(device_data, (char __user *)buf, count, ppos, true);
 764 }
 765
 766 static int vfio_pci_mmap(void *device_data, struct vm_area_struct *vma)
 767 {
 768         struct vfio_pci_device *vdev = device_data;
 769         struct pci_dev *pdev = vdev->pdev;
 770         unsigned int index;
 771         u64 phys_len, req_len, pgoff, req_start;
 772         int ret;
 773
 774         index = vma->vm_pgoff >> (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT);
 775
 776         if (vma->vm_end < vma->vm_start)
 777                 return -EINVAL;
 778         if ((vma->vm_flags & VM_SHARED) == 0)
 779                 return -EINVAL;
 780         if (index >= VFIO_PCI_ROM_REGION_INDEX)
 781                 return -EINVAL;
 782         if (!(pci_resource_flags(pdev, index) & IORESOURCE_MEM))
 783                 return -EINVAL;
 784
 785         phys_len = pci_resource_len(pdev, index);
 786         req_len = vma->vm_end - vma->vm_start;
 787         pgoff = vma->vm_pgoff &
 788                 ((1U << (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT)) - 1);
 789         req_start = pgoff << PAGE_SHIFT;
 790
 791         if (phys_len < PAGE_SIZE || req_start + req_len > phys_len)
 792                 return -EINVAL;
 793
 794         if (index == vdev->msix_bar) {
 795                 /*
 796                  * Disallow mmaps overlapping the MSI-X table; users don't
 797                  * get to touch this directly.  We could find somewhere
 798                  * else to map the overlap, but page granularity is only
 799                  * a recommendation, not a requirement, so the user needs
 800                  * to know which bits are real.  Requiring them to mmap
 801                  * around the table makes that clear.
 802                  */
 803
 804                 /* If neither entirely above nor below, then it overlaps */
 805                 if (!(req_start >= vdev->msix_offset + vdev->msix_size ||
 806                       req_start + req_len <= vdev->msix_offset))
 807                         return -EINVAL;
 808         }
 809
 810         /*
 811          * Even though we don't make use of the barmap for the mmap,
 812          * we need to request the region and the barmap tracks that.
 813          */
 814         if (!vdev->barmap[index]) {
 815                 ret = pci_request_selected_regions(pdev,
 816                                                    1 << index, "vfio-pci");
 817                 if (ret)
 818                         return ret;
 819
 820                 vdev->barmap[index] = pci_iomap(pdev, index, 0);
 821         }
 822
 823         vma->vm_private_data = vdev;
 824         vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
 825         vma->vm_pgoff = (pci_resource_start(pdev, index) >> PAGE_SHIFT) + pgoff;
 826
 827         return remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
 828                                req_len, vma->vm_page_prot);
 829 }
 830
 831 static const struct vfio_device_ops vfio_pci_ops = {
 832         .name           = "vfio-pci",
 833         .open           = vfio_pci_open,
 834         .release        = vfio_pci_release,
 835         .ioctl          = vfio_pci_ioctl,
 836         .read           = vfio_pci_read,
 837         .write          = vfio_pci_write,
 838         .mmap           = vfio_pci_mmap,
 839 };
 840
 841 static int vfio_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
 842 {
 843         u8 type;
 844         struct vfio_pci_device *vdev;
 845         struct iommu_group *group;
 846         int ret;
 847
 848         pci_read_config_byte(pdev, PCI_HEADER_TYPE, &type);
 849         if ((type & PCI_HEADER_TYPE) != PCI_HEADER_TYPE_NORMAL)
 850                 return -EINVAL;
 851
 852         group = iommu_group_get(&pdev->dev);
 853         if (!group)
 854                 return -EINVAL;
 855
 856         vdev = kzalloc(sizeof(*vdev), GFP_KERNEL);
 857         if (!vdev) {
 858                 iommu_group_put(group);
 859                 return -ENOMEM;
 860         }
 861
 862         vdev->pdev = pdev;
 863         vdev->irq_type = VFIO_PCI_NUM_IRQS;
 864         mutex_init(&vdev->igate);
 865         spin_lock_init(&vdev->irqlock);
 866
 867         ret = vfio_add_group_dev(&pdev->dev, &vfio_pci_ops, vdev);
 868         if (ret) {
 869                 iommu_group_put(group);
 870                 kfree(vdev);
 871         }
 872
 873         return ret;
 874 }
 875
 876 static void vfio_pci_remove(struct pci_dev *pdev)
 877 {
 878         struct vfio_pci_device *vdev;
 879
 880         vdev = vfio_del_group_dev(&pdev->dev);
 881         if (vdev) {
 882                 iommu_group_put(pdev->dev.iommu_group);
 883                 kfree(vdev);
 884         }
 885 }
 886
 887 static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
 888                                                   pci_channel_state_t state)
 889 {
 890         struct vfio_pci_device *vdev;
 891         struct vfio_device *device;
 892
 893         device = vfio_device_get_from_dev(&pdev->dev);
 894         if (device == NULL)
 895                 return PCI_ERS_RESULT_DISCONNECT;
 896
 897         vdev = vfio_device_data(device);
 898         if (vdev == NULL) {
 899                 vfio_device_put(device);
 900                 return PCI_ERS_RESULT_DISCONNECT;
 901         }
 902
 903         mutex_lock(&vdev->igate);
 904
 905         if (vdev->err_trigger)
 906                 eventfd_signal(vdev->err_trigger, 1);
 907
 908         mutex_unlock(&vdev->igate);
 909
 910         vfio_device_put(device);
 911
 912         return PCI_ERS_RESULT_CAN_RECOVER;
 913 }
 914
 915 static struct pci_error_handlers vfio_err_handlers = {
 916         .error_detected = vfio_pci_aer_err_detected,
 917 };
 918
 919 static struct pci_driver vfio_pci_driver = {
 920         .name           = "vfio-pci",
 921         .id_table       = NULL, /* only dynamic ids */
 922         .probe          = vfio_pci_probe,
 923         .remove         = vfio_pci_remove,
 924         .err_handler    = &vfio_err_handlers,
 925 };
 926
 927 struct vfio_devices {
 928         struct vfio_device **devices;
 929         int cur_index;
 930         int max_index;
 931 };
 932
 933 static int vfio_pci_get_devs(struct pci_dev *pdev, void *data)
 934 {
 935         struct vfio_devices *devs = data;
 936         struct pci_driver *pci_drv = ACCESS_ONCE(pdev->driver);
 937
 938         if (pci_drv != &vfio_pci_driver)
 939                 return -EBUSY;
 940
 941         if (devs->cur_index == devs->max_index)
 942                 return -ENOSPC;
 943
 944         devs->devices[devs->cur_index] = vfio_device_get_from_dev(&pdev->dev);
 945         if (!devs->devices[devs->cur_index])
 946                 return -EINVAL;
 947
 948         devs->cur_index++;
 949         return 0;
 950 }
 951
 952 /*
 953  * Attempt to do a bus/slot reset if there are devices affected by a reset for
 954  * this device that are needs_reset and all of the affected devices are unused
 955  * (!refcnt).  Callers are required to hold driver_lock when calling this to
 956  * prevent device opens and concurrent bus reset attempts.  We prevent device
 957  * unbinds by acquiring and holding a reference to the vfio_device.
 958  *
 959  * NB: vfio-core considers a group to be viable even if some devices are
 960  * bound to drivers like pci-stub or pcieport.  Here we require all devices
 961  * to be bound to vfio_pci since that's the only way we can be sure they
 962  * stay put.
 963  */
 964 static void vfio_pci_try_bus_reset(struct vfio_pci_device *vdev)
 965 {
 966         struct vfio_devices devs = { .cur_index = 0 };
 967         int i = 0, ret = -EINVAL;
 968         bool needs_reset = false, slot = false;
 969         struct vfio_pci_device *tmp;
 970
 971         if (!pci_probe_reset_slot(vdev->pdev->slot))
 972                 slot = true;
 973         else if (pci_probe_reset_bus(vdev->pdev->bus))
 974                 return;
 975
 976         if (vfio_pci_for_each_slot_or_bus(vdev->pdev, vfio_pci_count_devs,
 977                                           &i, slot) || !i)
 978                 return;
 979
 980         devs.max_index = i;
 981         devs.devices = kcalloc(i, sizeof(struct vfio_device *), GFP_KERNEL);
 982         if (!devs.devices)
 983                 return;
 984
 985         if (vfio_pci_for_each_slot_or_bus(vdev->pdev,
 986                                           vfio_pci_get_devs, &devs, slot))
 987                 goto put_devs;
 988
 989         for (i = 0; i < devs.cur_index; i++) {
 990                 tmp = vfio_device_data(devs.devices[i]);
 991                 if (tmp->needs_reset)
 992                         needs_reset = true;
 993                 if (tmp->refcnt)
 994                         goto put_devs;
 995         }
 996
 997         if (needs_reset)
 998                 ret = slot ? pci_try_reset_slot(vdev->pdev->slot) :
 999                              pci_try_reset_bus(vdev->pdev->bus);
1000
1001 put_devs:
1002         for (i = 0; i < devs.cur_index; i++) {
1003                 if (!ret) {
1004                         tmp = vfio_device_data(devs.devices[i]);
1005                         tmp->needs_reset = false;
1006                 }
1007                 vfio_device_put(devs.devices[i]);
1008         }
1009
1010         kfree(devs.devices);
1011 }
1012
1013 static void __exit vfio_pci_cleanup(void)
1014 {
1015         pci_unregister_driver(&vfio_pci_driver);
1016         vfio_pci_virqfd_exit();
1017         vfio_pci_uninit_perm_bits();
1018 }
1019
1020 static int __init vfio_pci_init(void)
1021 {
1022         int ret;
1023
1024         /* Allocate shared config space permision data used by all devices */
1025         ret = vfio_pci_init_perm_bits();
1026         if (ret)
1027                 return ret;
1028
1029         /* Start the virqfd cleanup handler */
1030         ret = vfio_pci_virqfd_init();
1031         if (ret)
1032                 goto out_virqfd;
1033
1034         /* Register and scan for devices */
1035         ret = pci_register_driver(&vfio_pci_driver);
1036         if (ret)
1037                 goto out_driver;
1038
1039         return 0;
1040
1041 out_driver:
1042         vfio_pci_virqfd_exit();
1043 out_virqfd:
1044         vfio_pci_uninit_perm_bits();
1045         return ret;
1046 }
1047
1048 module_init(vfio_pci_init);
1049 module_exit(vfio_pci_cleanup);
1050
1051 MODULE_VERSION(DRIVER_VERSION);
1052 MODULE_LICENSE("GPL v2");
1053 MODULE_AUTHOR(DRIVER_AUTHOR);
1054 MODULE_DESCRIPTION(DRIVER_DESC);