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