drivers/xen/xen-pciback/pci_stub.c

   1 /*
   2  * PCI Stub Driver - Grabs devices in backend to be exported later
   3  *
   4  * Ryan Wilson <hap9@epoch.ncsc.mil>
   5  * Chris Bookholt <hap10@epoch.ncsc.mil>
   6  */
   7
   8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   9
  10 #include <linux/module.h>
  11 #include <linux/init.h>
  12 #include <linux/rwsem.h>
  13 #include <linux/list.h>
  14 #include <linux/spinlock.h>
  15 #include <linux/kref.h>
  16 #include <linux/pci.h>
  17 #include <linux/wait.h>
  18 #include <linux/sched.h>
  19 #include <linux/atomic.h>
  20 #include <xen/events.h>
  21 #include <asm/xen/pci.h>
  22 #include <asm/xen/hypervisor.h>
  23 #include <xen/interface/physdev.h>
  24 #include "pciback.h"
  25 #include "conf_space.h"
  26 #include "conf_space_quirks.h"
  27
  28 #define PCISTUB_DRIVER_NAME "pciback"
  29
  30 static char *pci_devs_to_hide;
  31 wait_queue_head_t xen_pcibk_aer_wait_queue;
  32 /*Add sem for sync AER handling and xen_pcibk remove/reconfigue ops,
  33 * We want to avoid in middle of AER ops, xen_pcibk devices is being removed
  34 */
  35 static DECLARE_RWSEM(pcistub_sem);
  36 module_param_named(hide, pci_devs_to_hide, charp, 0444);
  37
  38 struct pcistub_device_id {
  39         struct list_head slot_list;
  40         int domain;
  41         unsigned char bus;
  42         unsigned int devfn;
  43 };
  44 static LIST_HEAD(pcistub_device_ids);
  45 static DEFINE_SPINLOCK(device_ids_lock);
  46
  47 struct pcistub_device {
  48         struct kref kref;
  49         struct list_head dev_list;
  50         spinlock_t lock;
  51
  52         struct pci_dev *dev;
  53         struct xen_pcibk_device *pdev;/* non-NULL if struct pci_dev is in use */
  54 };
  55
  56 /* Access to pcistub_devices & seized_devices lists and the initialize_devices
  57  * flag must be locked with pcistub_devices_lock
  58  */
  59 static DEFINE_SPINLOCK(pcistub_devices_lock);
  60 static LIST_HEAD(pcistub_devices);
  61
  62 /* wait for device_initcall before initializing our devices
  63  * (see pcistub_init_devices_late)
  64  */
  65 static int initialize_devices;
  66 static LIST_HEAD(seized_devices);
  67
  68 static struct pcistub_device *pcistub_device_alloc(struct pci_dev *dev)
  69 {
  70         struct pcistub_device *psdev;
  71
  72         dev_dbg(&dev->dev, "pcistub_device_alloc\n");
  73
  74         psdev = kzalloc(sizeof(*psdev), GFP_KERNEL);
  75         if (!psdev)
  76                 return NULL;
  77
  78         psdev->dev = pci_dev_get(dev);
  79         if (!psdev->dev) {
  80                 kfree(psdev);
  81                 return NULL;
  82         }
  83
  84         kref_init(&psdev->kref);
  85         spin_lock_init(&psdev->lock);
  86
  87         return psdev;
  88 }
  89
  90 /* Don't call this directly as it's called by pcistub_device_put */
  91 static void pcistub_device_release(struct kref *kref)
  92 {
  93         struct pcistub_device *psdev;
  94         struct pci_dev *dev;
  95         struct xen_pcibk_dev_data *dev_data;
  96
  97         psdev = container_of(kref, struct pcistub_device, kref);
  98         dev = psdev->dev;
  99         dev_data = pci_get_drvdata(dev);
 100
 101         dev_dbg(&dev->dev, "pcistub_device_release\n");
 102
 103         xen_unregister_device_domain_owner(dev);
 104
 105         /* Call the reset function which does not take lock as this
 106          * is called from "unbind" which takes a device_lock mutex.
 107          */
 108         __pci_reset_function_locked(dev);
 109         if (dev_data &&
 110             pci_load_and_free_saved_state(dev, &dev_data->pci_saved_state))
 111                 dev_info(&dev->dev, "Could not reload PCI state\n");
 112         else
 113                 pci_restore_state(dev);
 114
 115         if (dev->msix_cap) {
 116                 struct physdev_pci_device ppdev = {
 117                         .seg = pci_domain_nr(dev->bus),
 118                         .bus = dev->bus->number,
 119                         .devfn = dev->devfn
 120                 };
 121                 int err = HYPERVISOR_physdev_op(PHYSDEVOP_release_msix,
 122                                                 &ppdev);
 123
 124                 if (err && err != -ENOSYS)
 125                         dev_warn(&dev->dev, "MSI-X release failed (%d)\n",
 126                                  err);
 127         }
 128
 129         /* Disable the device */
 130         xen_pcibk_reset_device(dev);
 131
 132         kfree(dev_data);
 133         pci_set_drvdata(dev, NULL);
 134
 135         /* Clean-up the device */
 136         xen_pcibk_config_free_dyn_fields(dev);
 137         xen_pcibk_config_free_dev(dev);
 138
 139         pci_clear_dev_assigned(dev);
 140         pci_dev_put(dev);
 141
 142         kfree(psdev);
 143 }
 144
 145 static inline void pcistub_device_get(struct pcistub_device *psdev)
 146 {
 147         kref_get(&psdev->kref);
 148 }
 149
 150 static inline void pcistub_device_put(struct pcistub_device *psdev)
 151 {
 152         kref_put(&psdev->kref, pcistub_device_release);
 153 }
 154
 155 static struct pcistub_device *pcistub_device_find_locked(int domain, int bus,
 156                                                          int slot, int func)
 157 {
 158         struct pcistub_device *psdev;
 159
 160         list_for_each_entry(psdev, &pcistub_devices, dev_list) {
 161                 if (psdev->dev != NULL
 162                     && domain == pci_domain_nr(psdev->dev->bus)
 163                     && bus == psdev->dev->bus->number
 164                     && slot == PCI_SLOT(psdev->dev->devfn)
 165                     && func == PCI_FUNC(psdev->dev->devfn)) {
 166                         return psdev;
 167                 }
 168         }
 169
 170         return NULL;
 171 }
 172
 173 static struct pcistub_device *pcistub_device_find(int domain, int bus,
 174                                                   int slot, int func)
 175 {
 176         struct pcistub_device *psdev;
 177         unsigned long flags;
 178
 179         spin_lock_irqsave(&pcistub_devices_lock, flags);
 180
 181         psdev = pcistub_device_find_locked(domain, bus, slot, func);
 182         if (psdev)
 183                 pcistub_device_get(psdev);
 184
 185         spin_unlock_irqrestore(&pcistub_devices_lock, flags);
 186         return psdev;
 187 }
 188
 189 static struct pci_dev *pcistub_device_get_pci_dev(struct xen_pcibk_device *pdev,
 190                                                   struct pcistub_device *psdev)
 191 {
 192         struct pci_dev *pci_dev = NULL;
 193         unsigned long flags;
 194
 195         pcistub_device_get(psdev);
 196
 197         spin_lock_irqsave(&psdev->lock, flags);
 198         if (!psdev->pdev) {
 199                 psdev->pdev = pdev;
 200                 pci_dev = psdev->dev;
 201         }
 202         spin_unlock_irqrestore(&psdev->lock, flags);
 203
 204         if (!pci_dev)
 205                 pcistub_device_put(psdev);
 206
 207         return pci_dev;
 208 }
 209
 210 struct pci_dev *pcistub_get_pci_dev_by_slot(struct xen_pcibk_device *pdev,
 211                                             int domain, int bus,
 212                                             int slot, int func)
 213 {
 214         struct pcistub_device *psdev;
 215         struct pci_dev *found_dev = NULL;
 216         unsigned long flags;
 217
 218         spin_lock_irqsave(&pcistub_devices_lock, flags);
 219
 220         psdev = pcistub_device_find_locked(domain, bus, slot, func);
 221         if (psdev)
 222                 found_dev = pcistub_device_get_pci_dev(pdev, psdev);
 223
 224         spin_unlock_irqrestore(&pcistub_devices_lock, flags);
 225         return found_dev;
 226 }
 227
 228 struct pci_dev *pcistub_get_pci_dev(struct xen_pcibk_device *pdev,
 229                                     struct pci_dev *dev)
 230 {
 231         struct pcistub_device *psdev;
 232         struct pci_dev *found_dev = NULL;
 233         unsigned long flags;
 234
 235         spin_lock_irqsave(&pcistub_devices_lock, flags);
 236
 237         list_for_each_entry(psdev, &pcistub_devices, dev_list) {
 238                 if (psdev->dev == dev) {
 239                         found_dev = pcistub_device_get_pci_dev(pdev, psdev);
 240                         break;
 241                 }
 242         }
 243
 244         spin_unlock_irqrestore(&pcistub_devices_lock, flags);
 245         return found_dev;
 246 }
 247
 248 /*
 249  * Called when:
 250  *  - XenBus state has been reconfigure (pci unplug). See xen_pcibk_remove_device
 251  *  - XenBus state has been disconnected (guest shutdown). See xen_pcibk_xenbus_remove
 252  *  - 'echo BDF > unbind' on pciback module with no guest attached. See pcistub_remove
 253  *  - 'echo BDF > unbind' with a guest still using it. See pcistub_remove
 254  *
 255  *  As such we have to be careful.
 256  *
 257  *  To make this easier, the caller has to hold the device lock.
 258  */
 259 void pcistub_put_pci_dev(struct pci_dev *dev)
 260 {
 261         struct pcistub_device *psdev, *found_psdev = NULL;
 262         unsigned long flags;
 263         struct xen_pcibk_dev_data *dev_data;
 264         int ret;
 265
 266         spin_lock_irqsave(&pcistub_devices_lock, flags);
 267
 268         list_for_each_entry(psdev, &pcistub_devices, dev_list) {
 269                 if (psdev->dev == dev) {
 270                         found_psdev = psdev;
 271                         break;
 272                 }
 273         }
 274
 275         spin_unlock_irqrestore(&pcistub_devices_lock, flags);
 276         if (WARN_ON(!found_psdev))
 277                 return;
 278
 279         /*hold this lock for avoiding breaking link between
 280         * pcistub and xen_pcibk when AER is in processing
 281         */
 282         down_write(&pcistub_sem);
 283         /* Cleanup our device
 284          * (so it's ready for the next domain)
 285          */
 286         device_lock_assert(&dev->dev);
 287         __pci_reset_function_locked(dev);
 288
 289         dev_data = pci_get_drvdata(dev);
 290         ret = pci_load_saved_state(dev, dev_data->pci_saved_state);
 291         if (!ret) {
 292                 /*
 293                  * The usual sequence is pci_save_state & pci_restore_state
 294                  * but the guest might have messed the configuration space up.
 295                  * Use the initial version (when device was bound to us).
 296                  */
 297                 pci_restore_state(dev);
 298         } else
 299                 dev_info(&dev->dev, "Could not reload PCI state\n");
 300         /* This disables the device. */
 301         xen_pcibk_reset_device(dev);
 302
 303         /* And cleanup up our emulated fields. */
 304         xen_pcibk_config_reset_dev(dev);
 305         xen_pcibk_config_free_dyn_fields(dev);
 306
 307         xen_unregister_device_domain_owner(dev);
 308
 309         spin_lock_irqsave(&found_psdev->lock, flags);
 310         found_psdev->pdev = NULL;
 311         spin_unlock_irqrestore(&found_psdev->lock, flags);
 312
 313         pcistub_device_put(found_psdev);
 314         up_write(&pcistub_sem);
 315 }
 316
 317 static int pcistub_match_one(struct pci_dev *dev,
 318                              struct pcistub_device_id *pdev_id)
 319 {
 320         /* Match the specified device by domain, bus, slot, func and also if
 321          * any of the device's parent bridges match.
 322          */
 323         for (; dev != NULL; dev = dev->bus->self) {
 324                 if (pci_domain_nr(dev->bus) == pdev_id->domain
 325                     && dev->bus->number == pdev_id->bus
 326                     && dev->devfn == pdev_id->devfn)
 327                         return 1;
 328
 329                 /* Sometimes topmost bridge links to itself. */
 330                 if (dev == dev->bus->self)
 331                         break;
 332         }
 333
 334         return 0;
 335 }
 336
 337 static int pcistub_match(struct pci_dev *dev)
 338 {
 339         struct pcistub_device_id *pdev_id;
 340         unsigned long flags;
 341         int found = 0;
 342
 343         spin_lock_irqsave(&device_ids_lock, flags);
 344         list_for_each_entry(pdev_id, &pcistub_device_ids, slot_list) {
 345                 if (pcistub_match_one(dev, pdev_id)) {
 346                         found = 1;
 347                         break;
 348                 }
 349         }
 350         spin_unlock_irqrestore(&device_ids_lock, flags);
 351
 352         return found;
 353 }
 354
 355 static int pcistub_init_device(struct pci_dev *dev)
 356 {
 357         struct xen_pcibk_dev_data *dev_data;
 358         int err = 0;
 359
 360         dev_dbg(&dev->dev, "initializing...\n");
 361
 362         /* The PCI backend is not intended to be a module (or to work with
 363          * removable PCI devices (yet). If it were, xen_pcibk_config_free()
 364          * would need to be called somewhere to free the memory allocated
 365          * here and then to call kfree(pci_get_drvdata(psdev->dev)).
 366          */
 367         dev_data = kzalloc(sizeof(*dev_data) +  strlen(DRV_NAME "[]")
 368                                 + strlen(pci_name(dev)) + 1, GFP_KERNEL);
 369         if (!dev_data) {
 370                 err = -ENOMEM;
 371                 goto out;
 372         }
 373         pci_set_drvdata(dev, dev_data);
 374
 375         /*
 376          * Setup name for fake IRQ handler. It will only be enabled
 377          * once the device is turned on by the guest.
 378          */
 379         sprintf(dev_data->irq_name, DRV_NAME "[%s]", pci_name(dev));
 380
 381         dev_dbg(&dev->dev, "initializing config\n");
 382
 383         init_waitqueue_head(&xen_pcibk_aer_wait_queue);
 384         err = xen_pcibk_config_init_dev(dev);
 385         if (err)
 386                 goto out;
 387
 388         /* HACK: Force device (& ACPI) to determine what IRQ it's on - we
 389          * must do this here because pcibios_enable_device may specify
 390          * the pci device's true irq (and possibly its other resources)
 391          * if they differ from what's in the configuration space.
 392          * This makes the assumption that the device's resources won't
 393          * change after this point (otherwise this code may break!)
 394          */
 395         dev_dbg(&dev->dev, "enabling device\n");
 396         err = pci_enable_device(dev);
 397         if (err)
 398                 goto config_release;
 399
 400         if (dev->msix_cap) {
 401                 struct physdev_pci_device ppdev = {
 402                         .seg = pci_domain_nr(dev->bus),
 403                         .bus = dev->bus->number,
 404                         .devfn = dev->devfn
 405                 };
 406
 407                 err = HYPERVISOR_physdev_op(PHYSDEVOP_prepare_msix, &ppdev);
 408                 if (err && err != -ENOSYS)
 409                         dev_err(&dev->dev, "MSI-X preparation failed (%d)\n",
 410                                 err);
 411         }
 412
 413         /* We need the device active to save the state. */
 414         dev_dbg(&dev->dev, "save state of device\n");
 415         pci_save_state(dev);
 416         dev_data->pci_saved_state = pci_store_saved_state(dev);
 417         if (!dev_data->pci_saved_state)
 418                 dev_err(&dev->dev, "Could not store PCI conf saved state!\n");
 419         else {
 420                 dev_dbg(&dev->dev, "resetting (FLR, D3, etc) the device\n");
 421                 __pci_reset_function_locked(dev);
 422                 pci_restore_state(dev);
 423         }
 424         /* Now disable the device (this also ensures some private device
 425          * data is setup before we export)
 426          */
 427         dev_dbg(&dev->dev, "reset device\n");
 428         xen_pcibk_reset_device(dev);
 429
 430         pci_set_dev_assigned(dev);
 431         return 0;
 432
 433 config_release:
 434         xen_pcibk_config_free_dev(dev);
 435
 436 out:
 437         pci_set_drvdata(dev, NULL);
 438         kfree(dev_data);
 439         return err;
 440 }
 441
 442 /*
 443  * Because some initialization still happens on
 444  * devices during fs_initcall, we need to defer
 445  * full initialization of our devices until
 446  * device_initcall.
 447  */
 448 static int __init pcistub_init_devices_late(void)
 449 {
 450         struct pcistub_device *psdev;
 451         unsigned long flags;
 452         int err = 0;
 453
 454         spin_lock_irqsave(&pcistub_devices_lock, flags);
 455
 456         while (!list_empty(&seized_devices)) {
 457                 psdev = container_of(seized_devices.next,
 458                                      struct pcistub_device, dev_list);
 459                 list_del(&psdev->dev_list);
 460
 461                 spin_unlock_irqrestore(&pcistub_devices_lock, flags);
 462
 463                 err = pcistub_init_device(psdev->dev);
 464                 if (err) {
 465                         dev_err(&psdev->dev->dev,
 466                                 "error %d initializing device\n", err);
 467                         kfree(psdev);
 468                         psdev = NULL;
 469                 }
 470
 471                 spin_lock_irqsave(&pcistub_devices_lock, flags);
 472
 473                 if (psdev)
 474                         list_add_tail(&psdev->dev_list, &pcistub_devices);
 475         }
 476
 477         initialize_devices = 1;
 478
 479         spin_unlock_irqrestore(&pcistub_devices_lock, flags);
 480
 481         return 0;
 482 }
 483
 484 static void pcistub_device_id_add_list(struct pcistub_device_id *new,
 485                                        int domain, int bus, unsigned int devfn)
 486 {
 487         struct pcistub_device_id *pci_dev_id;
 488         unsigned long flags;
 489         int found = 0;
 490
 491         spin_lock_irqsave(&device_ids_lock, flags);
 492
 493         list_for_each_entry(pci_dev_id, &pcistub_device_ids, slot_list) {
 494                 if (pci_dev_id->domain == domain && pci_dev_id->bus == bus &&
 495                     pci_dev_id->devfn == devfn) {
 496                         found = 1;
 497                         break;
 498                 }
 499         }
 500
 501         if (!found) {
 502                 new->domain = domain;
 503                 new->bus = bus;
 504                 new->devfn = devfn;
 505                 list_add_tail(&new->slot_list, &pcistub_device_ids);
 506         }
 507
 508         spin_unlock_irqrestore(&device_ids_lock, flags);
 509
 510         if (found)
 511                 kfree(new);
 512 }
 513
 514 static int pcistub_seize(struct pci_dev *dev,
 515                          struct pcistub_device_id *pci_dev_id)
 516 {
 517         struct pcistub_device *psdev;
 518         unsigned long flags;
 519         int err = 0;
 520
 521         psdev = pcistub_device_alloc(dev);
 522         if (!psdev) {
 523                 kfree(pci_dev_id);
 524                 return -ENOMEM;
 525         }
 526
 527         spin_lock_irqsave(&pcistub_devices_lock, flags);
 528
 529         if (initialize_devices) {
 530                 spin_unlock_irqrestore(&pcistub_devices_lock, flags);
 531
 532                 /* don't want irqs disabled when calling pcistub_init_device */
 533                 err = pcistub_init_device(psdev->dev);
 534
 535                 spin_lock_irqsave(&pcistub_devices_lock, flags);
 536
 537                 if (!err)
 538                         list_add(&psdev->dev_list, &pcistub_devices);
 539         } else {
 540                 dev_dbg(&dev->dev, "deferring initialization\n");
 541                 list_add(&psdev->dev_list, &seized_devices);
 542         }
 543
 544         spin_unlock_irqrestore(&pcistub_devices_lock, flags);
 545
 546         if (err) {
 547                 kfree(pci_dev_id);
 548                 pcistub_device_put(psdev);
 549         } else if (pci_dev_id)
 550                 pcistub_device_id_add_list(pci_dev_id, pci_domain_nr(dev->bus),
 551                                            dev->bus->number, dev->devfn);
 552
 553         return err;
 554 }
 555
 556 /* Called when 'bind'. This means we must _NOT_ call pci_reset_function or
 557  * other functions that take the sysfs lock. */
 558 static int pcistub_probe(struct pci_dev *dev, const struct pci_device_id *id)
 559 {
 560         int err = 0, match;
 561         struct pcistub_device_id *pci_dev_id = NULL;
 562
 563         dev_dbg(&dev->dev, "probing...\n");
 564
 565         match = pcistub_match(dev);
 566
 567         if ((dev->driver_override &&
 568              !strcmp(dev->driver_override, PCISTUB_DRIVER_NAME)) ||
 569             match) {
 570
 571                 if (dev->hdr_type != PCI_HEADER_TYPE_NORMAL
 572                     && dev->hdr_type != PCI_HEADER_TYPE_BRIDGE) {
 573                         dev_err(&dev->dev, "can't export pci devices that "
 574                                 "don't have a normal (0) or bridge (1) "
 575                                 "header type!\n");
 576                         err = -ENODEV;
 577                         goto out;
 578                 }
 579
 580                 if (!match) {
 581                         pci_dev_id = kmalloc(sizeof(*pci_dev_id), GFP_KERNEL);
 582                         if (!pci_dev_id) {
 583                                 err = -ENOMEM;
 584                                 goto out;
 585                         }
 586                 }
 587
 588                 dev_info(&dev->dev, "seizing device\n");
 589                 err = pcistub_seize(dev, pci_dev_id);
 590         } else
 591                 /* Didn't find the device */
 592                 err = -ENODEV;
 593
 594 out:
 595         return err;
 596 }
 597
 598 /* Called when 'unbind'. This means we must _NOT_ call pci_reset_function or
 599  * other functions that take the sysfs lock. */
 600 static void pcistub_remove(struct pci_dev *dev)
 601 {
 602         struct pcistub_device *psdev, *found_psdev = NULL;
 603         unsigned long flags;
 604
 605         dev_dbg(&dev->dev, "removing\n");
 606
 607         spin_lock_irqsave(&pcistub_devices_lock, flags);
 608
 609         xen_pcibk_config_quirk_release(dev);
 610
 611         list_for_each_entry(psdev, &pcistub_devices, dev_list) {
 612                 if (psdev->dev == dev) {
 613                         found_psdev = psdev;
 614                         break;
 615                 }
 616         }
 617
 618         spin_unlock_irqrestore(&pcistub_devices_lock, flags);
 619
 620         if (found_psdev) {
 621                 dev_dbg(&dev->dev, "found device to remove %s\n",
 622                         found_psdev->pdev ? "- in-use" : "");
 623
 624                 if (found_psdev->pdev) {
 625                         int domid = xen_find_device_domain_owner(dev);
 626
 627                         pr_warn("****** removing device %s while still in-use by domain %d! ******\n",
 628                                pci_name(found_psdev->dev), domid);
 629                         pr_warn("****** driver domain may still access this device's i/o resources!\n");
 630                         pr_warn("****** shutdown driver domain before binding device\n");
 631                         pr_warn("****** to other drivers or domains\n");
 632
 633                         /* N.B. This ends up calling pcistub_put_pci_dev which ends up
 634                          * doing the FLR. */
 635                         xen_pcibk_release_pci_dev(found_psdev->pdev,
 636                                                 found_psdev->dev,
 637                                                 false /* caller holds the lock. */);
 638                 }
 639
 640                 spin_lock_irqsave(&pcistub_devices_lock, flags);
 641                 list_del(&found_psdev->dev_list);
 642                 spin_unlock_irqrestore(&pcistub_devices_lock, flags);
 643
 644                 /* the final put for releasing from the list */
 645                 pcistub_device_put(found_psdev);
 646         }
 647 }
 648
 649 static const struct pci_device_id pcistub_ids[] = {
 650         {
 651          .vendor = PCI_ANY_ID,
 652          .device = PCI_ANY_ID,
 653          .subvendor = PCI_ANY_ID,
 654          .subdevice = PCI_ANY_ID,
 655          },
 656         {0,},
 657 };
 658
 659 #define PCI_NODENAME_MAX 40
 660 static void kill_domain_by_device(struct pcistub_device *psdev)
 661 {
 662         struct xenbus_transaction xbt;
 663         int err;
 664         char nodename[PCI_NODENAME_MAX];
 665
 666         BUG_ON(!psdev);
 667         snprintf(nodename, PCI_NODENAME_MAX, "/local/domain/0/backend/pci/%d/0",
 668                 psdev->pdev->xdev->otherend_id);
 669
 670 again:
 671         err = xenbus_transaction_start(&xbt);
 672         if (err) {
 673                 dev_err(&psdev->dev->dev,
 674                         "error %d when start xenbus transaction\n", err);
 675                 return;
 676         }
 677         /*PV AER handlers will set this flag*/
 678         xenbus_printf(xbt, nodename, "aerState" , "aerfail");
 679         err = xenbus_transaction_end(xbt, 0);
 680         if (err) {
 681                 if (err == -EAGAIN)
 682                         goto again;
 683                 dev_err(&psdev->dev->dev,
 684                         "error %d when end xenbus transaction\n", err);
 685                 return;
 686         }
 687 }
 688
 689 /* For each aer recovery step error_detected, mmio_enabled, etc, front_end and
 690  * backend need to have cooperation. In xen_pcibk, those steps will do similar
 691  * jobs: send service request and waiting for front_end response.
 692 */
 693 static pci_ers_result_t common_process(struct pcistub_device *psdev,
 694                                        pci_channel_state_t state, int aer_cmd,
 695                                        pci_ers_result_t result)
 696 {
 697         pci_ers_result_t res = result;
 698         struct xen_pcie_aer_op *aer_op;
 699         struct xen_pcibk_device *pdev = psdev->pdev;
 700         struct xen_pci_sharedinfo *sh_info = pdev->sh_info;
 701         int ret;
 702
 703         /*with PV AER drivers*/
 704         aer_op = &(sh_info->aer_op);
 705         aer_op->cmd = aer_cmd ;
 706         /*useful for error_detected callback*/
 707         aer_op->err = state;
 708         /*pcifront_end BDF*/
 709         ret = xen_pcibk_get_pcifront_dev(psdev->dev, psdev->pdev,
 710                 &aer_op->domain, &aer_op->bus, &aer_op->devfn);
 711         if (!ret) {
 712                 dev_err(&psdev->dev->dev,
 713                         DRV_NAME ": failed to get pcifront device\n");
 714                 return PCI_ERS_RESULT_NONE;
 715         }
 716         wmb();
 717
 718         dev_dbg(&psdev->dev->dev,
 719                         DRV_NAME ": aer_op %x dom %x bus %x devfn %x\n",
 720                         aer_cmd, aer_op->domain, aer_op->bus, aer_op->devfn);
 721         /*local flag to mark there's aer request, xen_pcibk callback will use
 722         * this flag to judge whether we need to check pci-front give aer
 723         * service ack signal
 724         */
 725         set_bit(_PCIB_op_pending, (unsigned long *)&pdev->flags);
 726
 727         /*It is possible that a pcifront conf_read_write ops request invokes
 728         * the callback which cause the spurious execution of wake_up.
 729         * Yet it is harmless and better than a spinlock here
 730         */
 731         set_bit(_XEN_PCIB_active,
 732                 (unsigned long *)&sh_info->flags);
 733         wmb();
 734         notify_remote_via_irq(pdev->evtchn_irq);
 735
 736         /* Enable IRQ to signal "request done". */
 737         xen_pcibk_lateeoi(pdev, 0);
 738
 739         ret = wait_event_timeout(xen_pcibk_aer_wait_queue,
 740                                  !(test_bit(_XEN_PCIB_active, (unsigned long *)
 741                                  &sh_info->flags)), 300*HZ);
 742
 743         /* Enable IRQ for pcifront request if not already active. */
 744         if (!test_bit(_PDEVF_op_active, &pdev->flags))
 745                 xen_pcibk_lateeoi(pdev, 0);
 746
 747         if (!ret) {
 748                 if (test_bit(_XEN_PCIB_active,
 749                         (unsigned long *)&sh_info->flags)) {
 750                         dev_err(&psdev->dev->dev,
 751                                 "pcifront aer process not responding!\n");
 752                         clear_bit(_XEN_PCIB_active,
 753                           (unsigned long *)&sh_info->flags);
 754                         aer_op->err = PCI_ERS_RESULT_NONE;
 755                         return res;
 756                 }
 757         }
 758         clear_bit(_PCIB_op_pending, (unsigned long *)&pdev->flags);
 759
 760         res = (pci_ers_result_t)aer_op->err;
 761         return res;
 762 }
 763
 764 /*
 765 * xen_pcibk_slot_reset: it will send the slot_reset request to  pcifront in case
 766 * of the device driver could provide this service, and then wait for pcifront
 767 * ack.
 768 * @dev: pointer to PCI devices
 769 * return value is used by aer_core do_recovery policy
 770 */
 771 static pci_ers_result_t xen_pcibk_slot_reset(struct pci_dev *dev)
 772 {
 773         struct pcistub_device *psdev;
 774         pci_ers_result_t result;
 775
 776         result = PCI_ERS_RESULT_RECOVERED;
 777         dev_dbg(&dev->dev, "xen_pcibk_slot_reset(bus:%x,devfn:%x)\n",
 778                 dev->bus->number, dev->devfn);
 779
 780         down_write(&pcistub_sem);
 781         psdev = pcistub_device_find(pci_domain_nr(dev->bus),
 782                                 dev->bus->number,
 783                                 PCI_SLOT(dev->devfn),
 784                                 PCI_FUNC(dev->devfn));
 785
 786         if (!psdev || !psdev->pdev) {
 787                 dev_err(&dev->dev,
 788                         DRV_NAME " device is not found/assigned\n");
 789                 goto end;
 790         }
 791
 792         if (!psdev->pdev->sh_info) {
 793                 dev_err(&dev->dev, DRV_NAME " device is not connected or owned"
 794                         " by HVM, kill it\n");
 795                 kill_domain_by_device(psdev);
 796                 goto end;
 797         }
 798
 799         if (!test_bit(_XEN_PCIB_AERHANDLER,
 800                 (unsigned long *)&psdev->pdev->sh_info->flags)) {
 801                 dev_err(&dev->dev,
 802                         "guest with no AER driver should have been killed\n");
 803                 goto end;
 804         }
 805         result = common_process(psdev, 1, XEN_PCI_OP_aer_slotreset, result);
 806
 807         if (result == PCI_ERS_RESULT_NONE ||
 808                 result == PCI_ERS_RESULT_DISCONNECT) {
 809                 dev_dbg(&dev->dev,
 810                         "No AER slot_reset service or disconnected!\n");
 811                 kill_domain_by_device(psdev);
 812         }
 813 end:
 814         if (psdev)
 815                 pcistub_device_put(psdev);
 816         up_write(&pcistub_sem);
 817         return result;
 818
 819 }
 820
 821
 822 /*xen_pcibk_mmio_enabled: it will send the mmio_enabled request to  pcifront
 823 * in case of the device driver could provide this service, and then wait
 824 * for pcifront ack
 825 * @dev: pointer to PCI devices
 826 * return value is used by aer_core do_recovery policy
 827 */
 828
 829 static pci_ers_result_t xen_pcibk_mmio_enabled(struct pci_dev *dev)
 830 {
 831         struct pcistub_device *psdev;
 832         pci_ers_result_t result;
 833
 834         result = PCI_ERS_RESULT_RECOVERED;
 835         dev_dbg(&dev->dev, "xen_pcibk_mmio_enabled(bus:%x,devfn:%x)\n",
 836                 dev->bus->number, dev->devfn);
 837
 838         down_write(&pcistub_sem);
 839         psdev = pcistub_device_find(pci_domain_nr(dev->bus),
 840                                 dev->bus->number,
 841                                 PCI_SLOT(dev->devfn),
 842                                 PCI_FUNC(dev->devfn));
 843
 844         if (!psdev || !psdev->pdev) {
 845                 dev_err(&dev->dev,
 846                         DRV_NAME " device is not found/assigned\n");
 847                 goto end;
 848         }
 849
 850         if (!psdev->pdev->sh_info) {
 851                 dev_err(&dev->dev, DRV_NAME " device is not connected or owned"
 852                         " by HVM, kill it\n");
 853                 kill_domain_by_device(psdev);
 854                 goto end;
 855         }
 856
 857         if (!test_bit(_XEN_PCIB_AERHANDLER,
 858                 (unsigned long *)&psdev->pdev->sh_info->flags)) {
 859                 dev_err(&dev->dev,
 860                         "guest with no AER driver should have been killed\n");
 861                 goto end;
 862         }
 863         result = common_process(psdev, 1, XEN_PCI_OP_aer_mmio, result);
 864
 865         if (result == PCI_ERS_RESULT_NONE ||
 866                 result == PCI_ERS_RESULT_DISCONNECT) {
 867                 dev_dbg(&dev->dev,
 868                         "No AER mmio_enabled service or disconnected!\n");
 869                 kill_domain_by_device(psdev);
 870         }
 871 end:
 872         if (psdev)
 873                 pcistub_device_put(psdev);
 874         up_write(&pcistub_sem);
 875         return result;
 876 }
 877
 878 /*xen_pcibk_error_detected: it will send the error_detected request to  pcifront
 879 * in case of the device driver could provide this service, and then wait
 880 * for pcifront ack.
 881 * @dev: pointer to PCI devices
 882 * @error: the current PCI connection state
 883 * return value is used by aer_core do_recovery policy
 884 */
 885
 886 static pci_ers_result_t xen_pcibk_error_detected(struct pci_dev *dev,
 887         pci_channel_state_t error)
 888 {
 889         struct pcistub_device *psdev;
 890         pci_ers_result_t result;
 891
 892         result = PCI_ERS_RESULT_CAN_RECOVER;
 893         dev_dbg(&dev->dev, "xen_pcibk_error_detected(bus:%x,devfn:%x)\n",
 894                 dev->bus->number, dev->devfn);
 895
 896         down_write(&pcistub_sem);
 897         psdev = pcistub_device_find(pci_domain_nr(dev->bus),
 898                                 dev->bus->number,
 899                                 PCI_SLOT(dev->devfn),
 900                                 PCI_FUNC(dev->devfn));
 901
 902         if (!psdev || !psdev->pdev) {
 903                 dev_err(&dev->dev,
 904                         DRV_NAME " device is not found/assigned\n");
 905                 goto end;
 906         }
 907
 908         if (!psdev->pdev->sh_info) {
 909                 dev_err(&dev->dev, DRV_NAME " device is not connected or owned"
 910                         " by HVM, kill it\n");
 911                 kill_domain_by_device(psdev);
 912                 goto end;
 913         }
 914
 915         /*Guest owns the device yet no aer handler regiested, kill guest*/
 916         if (!test_bit(_XEN_PCIB_AERHANDLER,
 917                 (unsigned long *)&psdev->pdev->sh_info->flags)) {
 918                 dev_dbg(&dev->dev, "guest may have no aer driver, kill it\n");
 919                 kill_domain_by_device(psdev);
 920                 goto end;
 921         }
 922         result = common_process(psdev, error, XEN_PCI_OP_aer_detected, result);
 923
 924         if (result == PCI_ERS_RESULT_NONE ||
 925                 result == PCI_ERS_RESULT_DISCONNECT) {
 926                 dev_dbg(&dev->dev,
 927                         "No AER error_detected service or disconnected!\n");
 928                 kill_domain_by_device(psdev);
 929         }
 930 end:
 931         if (psdev)
 932                 pcistub_device_put(psdev);
 933         up_write(&pcistub_sem);
 934         return result;
 935 }
 936
 937 /*xen_pcibk_error_resume: it will send the error_resume request to  pcifront
 938 * in case of the device driver could provide this service, and then wait
 939 * for pcifront ack.
 940 * @dev: pointer to PCI devices
 941 */
 942
 943 static void xen_pcibk_error_resume(struct pci_dev *dev)
 944 {
 945         struct pcistub_device *psdev;
 946
 947         dev_dbg(&dev->dev, "xen_pcibk_error_resume(bus:%x,devfn:%x)\n",
 948                 dev->bus->number, dev->devfn);
 949
 950         down_write(&pcistub_sem);
 951         psdev = pcistub_device_find(pci_domain_nr(dev->bus),
 952                                 dev->bus->number,
 953                                 PCI_SLOT(dev->devfn),
 954                                 PCI_FUNC(dev->devfn));
 955
 956         if (!psdev || !psdev->pdev) {
 957                 dev_err(&dev->dev,
 958                         DRV_NAME " device is not found/assigned\n");
 959                 goto end;
 960         }
 961
 962         if (!psdev->pdev->sh_info) {
 963                 dev_err(&dev->dev, DRV_NAME " device is not connected or owned"
 964                         " by HVM, kill it\n");
 965                 kill_domain_by_device(psdev);
 966                 goto end;
 967         }
 968
 969         if (!test_bit(_XEN_PCIB_AERHANDLER,
 970                 (unsigned long *)&psdev->pdev->sh_info->flags)) {
 971                 dev_err(&dev->dev,
 972                         "guest with no AER driver should have been killed\n");
 973                 kill_domain_by_device(psdev);
 974                 goto end;
 975         }
 976         common_process(psdev, 1, XEN_PCI_OP_aer_resume,
 977                        PCI_ERS_RESULT_RECOVERED);
 978 end:
 979         if (psdev)
 980                 pcistub_device_put(psdev);
 981         up_write(&pcistub_sem);
 982         return;
 983 }
 984
 985 /*add xen_pcibk AER handling*/
 986 static const struct pci_error_handlers xen_pcibk_error_handler = {
 987         .error_detected = xen_pcibk_error_detected,
 988         .mmio_enabled = xen_pcibk_mmio_enabled,
 989         .slot_reset = xen_pcibk_slot_reset,
 990         .resume = xen_pcibk_error_resume,
 991 };
 992
 993 /*
 994  * Note: There is no MODULE_DEVICE_TABLE entry here because this isn't
 995  * for a normal device. I don't want it to be loaded automatically.
 996  */
 997
 998 static struct pci_driver xen_pcibk_pci_driver = {
 999         /* The name should be xen_pciback, but until the tools are updated
1000          * we will keep it as pciback. */
1001         .name = PCISTUB_DRIVER_NAME,
1002         .id_table = pcistub_ids,
1003         .probe = pcistub_probe,
1004         .remove = pcistub_remove,
1005         .err_handler = &xen_pcibk_error_handler,
1006 };
1007
1008 static inline int str_to_slot(const char *buf, int *domain, int *bus,
1009                               int *slot, int *func)
1010 {
1011         int parsed = 0;
1012
1013         switch (sscanf(buf, " %x:%x:%x.%x %n", domain, bus, slot, func,
1014                        &parsed)) {
1015         case 3:
1016                 *func = -1;
1017                 sscanf(buf, " %x:%x:%x.* %n", domain, bus, slot, &parsed);
1018                 break;
1019         case 2:
1020                 *slot = *func = -1;
1021                 sscanf(buf, " %x:%x:*.* %n", domain, bus, &parsed);
1022                 break;
1023         }
1024         if (parsed && !buf[parsed])
1025                 return 0;
1026
1027         /* try again without domain */
1028         *domain = 0;
1029         switch (sscanf(buf, " %x:%x.%x %n", bus, slot, func, &parsed)) {
1030         case 2:
1031                 *func = -1;
1032                 sscanf(buf, " %x:%x.* %n", bus, slot, &parsed);
1033                 break;
1034         case 1:
1035                 *slot = *func = -1;
1036                 sscanf(buf, " %x:*.* %n", bus, &parsed);
1037                 break;
1038         }
1039         if (parsed && !buf[parsed])
1040                 return 0;
1041
1042         return -EINVAL;
1043 }
1044
1045 static inline int str_to_quirk(const char *buf, int *domain, int *bus, int
1046                                *slot, int *func, int *reg, int *size, int *mask)
1047 {
1048         int parsed = 0;
1049
1050         sscanf(buf, " %x:%x:%x.%x-%x:%x:%x %n", domain, bus, slot, func,
1051                reg, size, mask, &parsed);
1052         if (parsed && !buf[parsed])
1053                 return 0;
1054
1055         /* try again without domain */
1056         *domain = 0;
1057         sscanf(buf, " %x:%x.%x-%x:%x:%x %n", bus, slot, func, reg, size,
1058                mask, &parsed);
1059         if (parsed && !buf[parsed])
1060                 return 0;
1061
1062         return -EINVAL;
1063 }
1064
1065 static int pcistub_device_id_add(int domain, int bus, int slot, int func)
1066 {
1067         struct pcistub_device_id *pci_dev_id;
1068         int rc = 0, devfn = PCI_DEVFN(slot, func);
1069
1070         if (slot < 0) {
1071                 for (slot = 0; !rc && slot < 32; ++slot)
1072                         rc = pcistub_device_id_add(domain, bus, slot, func);
1073                 return rc;
1074         }
1075
1076         if (func < 0) {
1077                 for (func = 0; !rc && func < 8; ++func)
1078                         rc = pcistub_device_id_add(domain, bus, slot, func);
1079                 return rc;
1080         }
1081
1082         if ((
1083 #if !defined(MODULE) /* pci_domains_supported is not being exported */ \
1084     || !defined(CONFIG_PCI_DOMAINS)
1085              !pci_domains_supported ? domain :
1086 #endif
1087              domain < 0 || domain > 0xffff)
1088             || bus < 0 || bus > 0xff
1089             || PCI_SLOT(devfn) != slot
1090             || PCI_FUNC(devfn) != func)
1091                 return -EINVAL;
1092
1093         pci_dev_id = kmalloc(sizeof(*pci_dev_id), GFP_KERNEL);
1094         if (!pci_dev_id)
1095                 return -ENOMEM;
1096
1097         pr_debug("wants to seize %04x:%02x:%02x.%d\n",
1098                  domain, bus, slot, func);
1099
1100         pcistub_device_id_add_list(pci_dev_id, domain, bus, devfn);
1101
1102         return 0;
1103 }
1104
1105 static int pcistub_device_id_remove(int domain, int bus, int slot, int func)
1106 {
1107         struct pcistub_device_id *pci_dev_id, *t;
1108         int err = -ENOENT;
1109         unsigned long flags;
1110
1111         spin_lock_irqsave(&device_ids_lock, flags);
1112         list_for_each_entry_safe(pci_dev_id, t, &pcistub_device_ids,
1113                                  slot_list) {
1114                 if (pci_dev_id->domain == domain && pci_dev_id->bus == bus
1115                     && (slot < 0 || PCI_SLOT(pci_dev_id->devfn) == slot)
1116                     && (func < 0 || PCI_FUNC(pci_dev_id->devfn) == func)) {
1117                         /* Don't break; here because it's possible the same
1118                          * slot could be in the list more than once
1119                          */
1120                         list_del(&pci_dev_id->slot_list);
1121                         kfree(pci_dev_id);
1122
1123                         err = 0;
1124
1125                         pr_debug("removed %04x:%02x:%02x.%d from seize list\n",
1126                                  domain, bus, slot, func);
1127                 }
1128         }
1129         spin_unlock_irqrestore(&device_ids_lock, flags);
1130
1131         return err;
1132 }
1133
1134 static int pcistub_reg_add(int domain, int bus, int slot, int func,
1135                            unsigned int reg, unsigned int size,
1136                            unsigned int mask)
1137 {
1138         int err = 0;
1139         struct pcistub_device *psdev;
1140         struct pci_dev *dev;
1141         struct config_field *field;
1142
1143         if (reg > 0xfff || (size < 4 && (mask >> (size * 8))))
1144                 return -EINVAL;
1145
1146         psdev = pcistub_device_find(domain, bus, slot, func);
1147         if (!psdev) {
1148                 err = -ENODEV;
1149                 goto out;
1150         }
1151         dev = psdev->dev;
1152
1153         field = kzalloc(sizeof(*field), GFP_KERNEL);
1154         if (!field) {
1155                 err = -ENOMEM;
1156                 goto out;
1157         }
1158
1159         field->offset = reg;
1160         field->size = size;
1161         field->mask = mask;
1162         field->init = NULL;
1163         field->reset = NULL;
1164         field->release = NULL;
1165         field->clean = xen_pcibk_config_field_free;
1166
1167         err = xen_pcibk_config_quirks_add_field(dev, field);
1168         if (err)
1169                 kfree(field);
1170 out:
1171         if (psdev)
1172                 pcistub_device_put(psdev);
1173         return err;
1174 }
1175
1176 static ssize_t new_slot_store(struct device_driver *drv, const char *buf,
1177                               size_t count)
1178 {
1179         int domain, bus, slot, func;
1180         int err;
1181
1182         err = str_to_slot(buf, &domain, &bus, &slot, &func);
1183         if (err)
1184                 goto out;
1185
1186         err = pcistub_device_id_add(domain, bus, slot, func);
1187
1188 out:
1189         if (!err)
1190                 err = count;
1191         return err;
1192 }
1193 static DRIVER_ATTR_WO(new_slot);
1194
1195 static ssize_t remove_slot_store(struct device_driver *drv, const char *buf,
1196                                  size_t count)
1197 {
1198         int domain, bus, slot, func;
1199         int err;
1200
1201         err = str_to_slot(buf, &domain, &bus, &slot, &func);
1202         if (err)
1203                 goto out;
1204
1205         err = pcistub_device_id_remove(domain, bus, slot, func);
1206
1207 out:
1208         if (!err)
1209                 err = count;
1210         return err;
1211 }
1212 static DRIVER_ATTR_WO(remove_slot);
1213
1214 static ssize_t slots_show(struct device_driver *drv, char *buf)
1215 {
1216         struct pcistub_device_id *pci_dev_id;
1217         size_t count = 0;
1218         unsigned long flags;
1219
1220         spin_lock_irqsave(&device_ids_lock, flags);
1221         list_for_each_entry(pci_dev_id, &pcistub_device_ids, slot_list) {
1222                 if (count >= PAGE_SIZE)
1223                         break;
1224
1225                 count += scnprintf(buf + count, PAGE_SIZE - count,
1226                                    "%04x:%02x:%02x.%d\n",
1227                                    pci_dev_id->domain, pci_dev_id->bus,
1228                                    PCI_SLOT(pci_dev_id->devfn),
1229                                    PCI_FUNC(pci_dev_id->devfn));
1230         }
1231         spin_unlock_irqrestore(&device_ids_lock, flags);
1232
1233         return count;
1234 }
1235 static DRIVER_ATTR_RO(slots);
1236
1237 static ssize_t irq_handlers_show(struct device_driver *drv, char *buf)
1238 {
1239         struct pcistub_device *psdev;
1240         struct xen_pcibk_dev_data *dev_data;
1241         size_t count = 0;
1242         unsigned long flags;
1243
1244         spin_lock_irqsave(&pcistub_devices_lock, flags);
1245         list_for_each_entry(psdev, &pcistub_devices, dev_list) {
1246                 if (count >= PAGE_SIZE)
1247                         break;
1248                 if (!psdev->dev)
1249                         continue;
1250                 dev_data = pci_get_drvdata(psdev->dev);
1251                 if (!dev_data)
1252                         continue;
1253                 count +=
1254                     scnprintf(buf + count, PAGE_SIZE - count,
1255                               "%s:%s:%sing:%ld\n",
1256                               pci_name(psdev->dev),
1257                               dev_data->isr_on ? "on" : "off",
1258                               dev_data->ack_intr ? "ack" : "not ack",
1259                               dev_data->handled);
1260         }
1261         spin_unlock_irqrestore(&pcistub_devices_lock, flags);
1262         return count;
1263 }
1264 static DRIVER_ATTR_RO(irq_handlers);
1265
1266 static ssize_t irq_handler_state_store(struct device_driver *drv,
1267                                        const char *buf, size_t count)
1268 {
1269         struct pcistub_device *psdev;
1270         struct xen_pcibk_dev_data *dev_data;
1271         int domain, bus, slot, func;
1272         int err;
1273
1274         err = str_to_slot(buf, &domain, &bus, &slot, &func);
1275         if (err)
1276                 return err;
1277
1278         psdev = pcistub_device_find(domain, bus, slot, func);
1279         if (!psdev) {
1280                 err = -ENOENT;
1281                 goto out;
1282         }
1283
1284         dev_data = pci_get_drvdata(psdev->dev);
1285         if (!dev_data) {
1286                 err = -ENOENT;
1287                 goto out;
1288         }
1289
1290         dev_dbg(&psdev->dev->dev, "%s fake irq handler: %d->%d\n",
1291                 dev_data->irq_name, dev_data->isr_on,
1292                 !dev_data->isr_on);
1293
1294         dev_data->isr_on = !(dev_data->isr_on);
1295         if (dev_data->isr_on)
1296                 dev_data->ack_intr = 1;
1297 out:
1298         if (psdev)
1299                 pcistub_device_put(psdev);
1300         if (!err)
1301                 err = count;
1302         return err;
1303 }
1304 static DRIVER_ATTR_WO(irq_handler_state);
1305
1306 static ssize_t quirks_store(struct device_driver *drv, const char *buf,
1307                             size_t count)
1308 {
1309         int domain, bus, slot, func, reg, size, mask;
1310         int err;
1311
1312         err = str_to_quirk(buf, &domain, &bus, &slot, &func, &reg, &size,
1313                            &mask);
1314         if (err)
1315                 goto out;
1316
1317         err = pcistub_reg_add(domain, bus, slot, func, reg, size, mask);
1318
1319 out:
1320         if (!err)
1321                 err = count;
1322         return err;
1323 }
1324
1325 static ssize_t quirks_show(struct device_driver *drv, char *buf)
1326 {
1327         int count = 0;
1328         unsigned long flags;
1329         struct xen_pcibk_config_quirk *quirk;
1330         struct xen_pcibk_dev_data *dev_data;
1331         const struct config_field *field;
1332         const struct config_field_entry *cfg_entry;
1333
1334         spin_lock_irqsave(&device_ids_lock, flags);
1335         list_for_each_entry(quirk, &xen_pcibk_quirks, quirks_list) {
1336                 if (count >= PAGE_SIZE)
1337                         goto out;
1338
1339                 count += scnprintf(buf + count, PAGE_SIZE - count,
1340                                    "%02x:%02x.%01x\n\t%04x:%04x:%04x:%04x\n",
1341                                    quirk->pdev->bus->number,
1342                                    PCI_SLOT(quirk->pdev->devfn),
1343                                    PCI_FUNC(quirk->pdev->devfn),
1344                                    quirk->devid.vendor, quirk->devid.device,
1345                                    quirk->devid.subvendor,
1346                                    quirk->devid.subdevice);
1347
1348                 dev_data = pci_get_drvdata(quirk->pdev);
1349
1350                 list_for_each_entry(cfg_entry, &dev_data->config_fields, list) {
1351                         field = cfg_entry->field;
1352                         if (count >= PAGE_SIZE)
1353                                 goto out;
1354
1355                         count += scnprintf(buf + count, PAGE_SIZE - count,
1356                                            "\t\t%08x:%01x:%08x\n",
1357                                            cfg_entry->base_offset +
1358                                            field->offset, field->size,
1359                                            field->mask);
1360                 }
1361         }
1362
1363 out:
1364         spin_unlock_irqrestore(&device_ids_lock, flags);
1365
1366         return count;
1367 }
1368 static DRIVER_ATTR_RW(quirks);
1369
1370 static ssize_t permissive_store(struct device_driver *drv, const char *buf,
1371                                 size_t count)
1372 {
1373         int domain, bus, slot, func;
1374         int err;
1375         struct pcistub_device *psdev;
1376         struct xen_pcibk_dev_data *dev_data;
1377
1378         err = str_to_slot(buf, &domain, &bus, &slot, &func);
1379         if (err)
1380                 goto out;
1381
1382         psdev = pcistub_device_find(domain, bus, slot, func);
1383         if (!psdev) {
1384                 err = -ENODEV;
1385                 goto out;
1386         }
1387
1388         dev_data = pci_get_drvdata(psdev->dev);
1389         /* the driver data for a device should never be null at this point */
1390         if (!dev_data) {
1391                 err = -ENXIO;
1392                 goto release;
1393         }
1394         if (!dev_data->permissive) {
1395                 dev_data->permissive = 1;
1396                 /* Let user know that what they're doing could be unsafe */
1397                 dev_warn(&psdev->dev->dev, "enabling permissive mode "
1398                          "configuration space accesses!\n");
1399                 dev_warn(&psdev->dev->dev,
1400                          "permissive mode is potentially unsafe!\n");
1401         }
1402 release:
1403         pcistub_device_put(psdev);
1404 out:
1405         if (!err)
1406                 err = count;
1407         return err;
1408 }
1409
1410 static ssize_t permissive_show(struct device_driver *drv, char *buf)
1411 {
1412         struct pcistub_device *psdev;
1413         struct xen_pcibk_dev_data *dev_data;
1414         size_t count = 0;
1415         unsigned long flags;
1416         spin_lock_irqsave(&pcistub_devices_lock, flags);
1417         list_for_each_entry(psdev, &pcistub_devices, dev_list) {
1418                 if (count >= PAGE_SIZE)
1419                         break;
1420                 if (!psdev->dev)
1421                         continue;
1422                 dev_data = pci_get_drvdata(psdev->dev);
1423                 if (!dev_data || !dev_data->permissive)
1424                         continue;
1425                 count +=
1426                     scnprintf(buf + count, PAGE_SIZE - count, "%s\n",
1427                               pci_name(psdev->dev));
1428         }
1429         spin_unlock_irqrestore(&pcistub_devices_lock, flags);
1430         return count;
1431 }
1432 static DRIVER_ATTR_RW(permissive);
1433
1434 static void pcistub_exit(void)
1435 {
1436         driver_remove_file(&xen_pcibk_pci_driver.driver, &driver_attr_new_slot);
1437         driver_remove_file(&xen_pcibk_pci_driver.driver,
1438                            &driver_attr_remove_slot);
1439         driver_remove_file(&xen_pcibk_pci_driver.driver, &driver_attr_slots);
1440         driver_remove_file(&xen_pcibk_pci_driver.driver, &driver_attr_quirks);
1441         driver_remove_file(&xen_pcibk_pci_driver.driver,
1442                            &driver_attr_permissive);
1443         driver_remove_file(&xen_pcibk_pci_driver.driver,
1444                            &driver_attr_irq_handlers);
1445         driver_remove_file(&xen_pcibk_pci_driver.driver,
1446                            &driver_attr_irq_handler_state);
1447         pci_unregister_driver(&xen_pcibk_pci_driver);
1448 }
1449
1450 static int __init pcistub_init(void)
1451 {
1452         int pos = 0;
1453         int err = 0;
1454         int domain, bus, slot, func;
1455         int parsed;
1456
1457         if (pci_devs_to_hide && *pci_devs_to_hide) {
1458                 do {
1459                         parsed = 0;
1460
1461                         err = sscanf(pci_devs_to_hide + pos,
1462                                      " (%x:%x:%x.%x) %n",
1463                                      &domain, &bus, &slot, &func, &parsed);
1464                         switch (err) {
1465                         case 3:
1466                                 func = -1;
1467                                 sscanf(pci_devs_to_hide + pos,
1468                                        " (%x:%x:%x.*) %n",
1469                                        &domain, &bus, &slot, &parsed);
1470                                 break;
1471                         case 2:
1472                                 slot = func = -1;
1473                                 sscanf(pci_devs_to_hide + pos,
1474                                        " (%x:%x:*.*) %n",
1475                                        &domain, &bus, &parsed);
1476                                 break;
1477                         }
1478
1479                         if (!parsed) {
1480                                 domain = 0;
1481                                 err = sscanf(pci_devs_to_hide + pos,
1482                                              " (%x:%x.%x) %n",
1483                                              &bus, &slot, &func, &parsed);
1484                                 switch (err) {
1485                                 case 2:
1486                                         func = -1;
1487                                         sscanf(pci_devs_to_hide + pos,
1488                                                " (%x:%x.*) %n",
1489                                                &bus, &slot, &parsed);
1490                                         break;
1491                                 case 1:
1492                                         slot = func = -1;
1493                                         sscanf(pci_devs_to_hide + pos,
1494                                                " (%x:*.*) %n",
1495                                                &bus, &parsed);
1496                                         break;
1497                                 }
1498                         }
1499
1500                         if (parsed <= 0)
1501                                 goto parse_error;
1502
1503                         err = pcistub_device_id_add(domain, bus, slot, func);
1504                         if (err)
1505                                 goto out;
1506
1507                         pos += parsed;
1508                 } while (pci_devs_to_hide[pos]);
1509         }
1510
1511         /* If we're the first PCI Device Driver to register, we're the
1512          * first one to get offered PCI devices as they become
1513          * available (and thus we can be the first to grab them)
1514          */
1515         err = pci_register_driver(&xen_pcibk_pci_driver);
1516         if (err < 0)
1517                 goto out;
1518
1519         err = driver_create_file(&xen_pcibk_pci_driver.driver,
1520                                  &driver_attr_new_slot);
1521         if (!err)
1522                 err = driver_create_file(&xen_pcibk_pci_driver.driver,
1523                                          &driver_attr_remove_slot);
1524         if (!err)
1525                 err = driver_create_file(&xen_pcibk_pci_driver.driver,
1526                                          &driver_attr_slots);
1527         if (!err)
1528                 err = driver_create_file(&xen_pcibk_pci_driver.driver,
1529                                          &driver_attr_quirks);
1530         if (!err)
1531                 err = driver_create_file(&xen_pcibk_pci_driver.driver,
1532                                          &driver_attr_permissive);
1533
1534         if (!err)
1535                 err = driver_create_file(&xen_pcibk_pci_driver.driver,
1536                                          &driver_attr_irq_handlers);
1537         if (!err)
1538                 err = driver_create_file(&xen_pcibk_pci_driver.driver,
1539                                         &driver_attr_irq_handler_state);
1540         if (err)
1541                 pcistub_exit();
1542
1543 out:
1544         return err;
1545
1546 parse_error:
1547         pr_err("Error parsing pci_devs_to_hide at \"%s\"\n",
1548                pci_devs_to_hide + pos);
1549         return -EINVAL;
1550 }
1551
1552 #ifndef MODULE
1553 /*
1554  * fs_initcall happens before device_initcall
1555  * so xen_pcibk *should* get called first (b/c we
1556  * want to suck up any device before other drivers
1557  * get a chance by being the first pci device
1558  * driver to register)
1559  */
1560 fs_initcall(pcistub_init);
1561 #endif
1562
1563 #ifdef CONFIG_PCI_IOV
1564 static struct pcistub_device *find_vfs(const struct pci_dev *pdev)
1565 {
1566         struct pcistub_device *psdev = NULL;
1567         unsigned long flags;
1568         bool found = false;
1569
1570         spin_lock_irqsave(&pcistub_devices_lock, flags);
1571         list_for_each_entry(psdev, &pcistub_devices, dev_list) {
1572                 if (!psdev->pdev && psdev->dev != pdev
1573                     && pci_physfn(psdev->dev) == pdev) {
1574                         found = true;
1575                         break;
1576                 }
1577         }
1578         spin_unlock_irqrestore(&pcistub_devices_lock, flags);
1579         if (found)
1580                 return psdev;
1581         return NULL;
1582 }
1583
1584 static int pci_stub_notifier(struct notifier_block *nb,
1585                              unsigned long action, void *data)
1586 {
1587         struct device *dev = data;
1588         const struct pci_dev *pdev = to_pci_dev(dev);
1589
1590         if (action != BUS_NOTIFY_UNBIND_DRIVER)
1591                 return NOTIFY_DONE;
1592
1593         if (!pdev->is_physfn)
1594                 return NOTIFY_DONE;
1595
1596         for (;;) {
1597                 struct pcistub_device *psdev = find_vfs(pdev);
1598                 if (!psdev)
1599                         break;
1600                 device_release_driver(&psdev->dev->dev);
1601         }
1602         return NOTIFY_DONE;
1603 }
1604
1605 static struct notifier_block pci_stub_nb = {
1606         .notifier_call = pci_stub_notifier,
1607 };
1608 #endif
1609
1610 static int __init xen_pcibk_init(void)
1611 {
1612         int err;
1613
1614         if (!xen_initial_domain())
1615                 return -ENODEV;
1616
1617         err = xen_pcibk_config_init();
1618         if (err)
1619                 return err;
1620
1621 #ifdef MODULE
1622         err = pcistub_init();
1623         if (err < 0)
1624                 return err;
1625 #endif
1626
1627         pcistub_init_devices_late();
1628         err = xen_pcibk_xenbus_register();
1629         if (err)
1630                 pcistub_exit();
1631 #ifdef CONFIG_PCI_IOV
1632         else
1633                 bus_register_notifier(&pci_bus_type, &pci_stub_nb);
1634 #endif
1635
1636         return err;
1637 }
1638
1639 static void __exit xen_pcibk_cleanup(void)
1640 {
1641 #ifdef CONFIG_PCI_IOV
1642         bus_unregister_notifier(&pci_bus_type, &pci_stub_nb);
1643 #endif
1644         xen_pcibk_xenbus_unregister();
1645         pcistub_exit();
1646 }
1647
1648 module_init(xen_pcibk_init);
1649 module_exit(xen_pcibk_cleanup);
1650
1651 MODULE_LICENSE("Dual BSD/GPL");
1652 MODULE_ALIAS("xen-backend:pci");