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 #define dev_fmt pr_fmt
  10
  11 #include <linux/module.h>
  12 #include <linux/init.h>
  13 #include <linux/rwsem.h>
  14 #include <linux/list.h>
  15 #include <linux/spinlock.h>
  16 #include <linux/kref.h>
  17 #include <linux/pci.h>
  18 #include <linux/wait.h>
  19 #include <linux/sched.h>
  20 #include <linux/atomic.h>
  21 #include <xen/events.h>
  22 #include <xen/pci.h>
  23 #include <xen/xen.h>
  24 #include <asm/xen/hypervisor.h>
  25 #include <xen/interface/physdev.h>
  26 #include "pciback.h"
  27 #include "conf_space.h"
  28 #include "conf_space_quirks.h"
  29
  30 #define PCISTUB_DRIVER_NAME "pciback"
  31
  32 static char *pci_devs_to_hide;
  33 wait_queue_head_t xen_pcibk_aer_wait_queue;
  34 /*Add sem for sync AER handling and xen_pcibk remove/reconfigue ops,
  35 * We want to avoid in middle of AER ops, xen_pcibk devices is being removed
  36 */
  37 static DECLARE_RWSEM(pcistub_sem);
  38 module_param_named(hide, pci_devs_to_hide, charp, 0444);
  39
  40 struct pcistub_device_id {
  41         struct list_head slot_list;
  42         int domain;
  43         unsigned char bus;
  44         unsigned int devfn;
  45 };
  46 static LIST_HEAD(pcistub_device_ids);
  47 static DEFINE_SPINLOCK(device_ids_lock);
  48
  49 struct pcistub_device {
  50         struct kref kref;
  51         struct list_head dev_list;
  52         spinlock_t lock;
  53
  54         struct pci_dev *dev;
  55         struct xen_pcibk_device *pdev;/* non-NULL if struct pci_dev is in use */
  56 };
  57
  58 /* Access to pcistub_devices & seized_devices lists and the initialize_devices
  59  * flag must be locked with pcistub_devices_lock
  60  */
  61 static DEFINE_SPINLOCK(pcistub_devices_lock);
  62 static LIST_HEAD(pcistub_devices);
  63
  64 /* wait for device_initcall before initializing our devices
  65  * (see pcistub_init_devices_late)
  66  */
  67 static int initialize_devices;
  68 static LIST_HEAD(seized_devices);
  69
  70 static struct pcistub_device *pcistub_device_alloc(struct pci_dev *dev)
  71 {
  72         struct pcistub_device *psdev;
  73
  74         dev_dbg(&dev->dev, "pcistub_device_alloc\n");
  75
  76         psdev = kzalloc(sizeof(*psdev), GFP_KERNEL);
  77         if (!psdev)
  78                 return NULL;
  79
  80         psdev->dev = pci_dev_get(dev);
  81         if (!psdev->dev) {
  82                 kfree(psdev);
  83                 return NULL;
  84         }
  85
  86         kref_init(&psdev->kref);
  87         spin_lock_init(&psdev->lock);
  88
  89         return psdev;
  90 }
  91
  92 /* Don't call this directly as it's called by pcistub_device_put */
  93 static void pcistub_device_release(struct kref *kref)
  94 {
  95         struct pcistub_device *psdev;
  96         struct pci_dev *dev;
  97         struct xen_pcibk_dev_data *dev_data;
  98
  99         psdev = container_of(kref, struct pcistub_device, kref);
 100         dev = psdev->dev;
 101         dev_data = pci_get_drvdata(dev);
 102
 103         dev_dbg(&dev->dev, "pcistub_device_release\n");
 104
 105         xen_unregister_device_domain_owner(dev);
 106
 107         /* Call the reset function which does not take lock as this
 108          * is called from "unbind" which takes a device_lock mutex.
 109          */
 110         __pci_reset_function_locked(dev);
 111         if (dev_data &&
 112             pci_load_and_free_saved_state(dev, &dev_data->pci_saved_state))
 113                 dev_info(&dev->dev, "Could not reload PCI state\n");
 114         else
 115                 pci_restore_state(dev);
 116
 117         if (dev->msix_cap) {
 118                 struct physdev_pci_device ppdev = {
 119                         .seg = pci_domain_nr(dev->bus),
 120                         .bus = dev->bus->number,
 121                         .devfn = dev->devfn
 122                 };
 123                 int err = HYPERVISOR_physdev_op(PHYSDEVOP_release_msix,
 124                                                 &ppdev);
 125
 126                 if (err && err != -ENOSYS)
 127                         dev_warn(&dev->dev, "MSI-X release failed (%d)\n",
 128                                  err);
 129         }
 130
 131         /* Disable the device */
 132         xen_pcibk_reset_device(dev);
 133
 134         kfree(dev_data);
 135         pci_set_drvdata(dev, NULL);
 136
 137         /* Clean-up the device */
 138         xen_pcibk_config_free_dyn_fields(dev);
 139         xen_pcibk_config_free_dev(dev);
 140
 141         pci_clear_dev_assigned(dev);
 142         pci_dev_put(dev);
 143
 144         kfree(psdev);
 145 }
 146
 147 static inline void pcistub_device_get(struct pcistub_device *psdev)
 148 {
 149         kref_get(&psdev->kref);
 150 }
 151
 152 static inline void pcistub_device_put(struct pcistub_device *psdev)
 153 {
 154         kref_put(&psdev->kref, pcistub_device_release);
 155 }
 156
 157 static struct pcistub_device *pcistub_device_find_locked(int domain, int bus,
 158                                                          int slot, int func)
 159 {
 160         struct pcistub_device *psdev;
 161
 162         list_for_each_entry(psdev, &pcistub_devices, dev_list) {
 163                 if (psdev->dev != NULL
 164                     && domain == pci_domain_nr(psdev->dev->bus)
 165                     && bus == psdev->dev->bus->number
 166                     && slot == PCI_SLOT(psdev->dev->devfn)
 167                     && func == PCI_FUNC(psdev->dev->devfn)) {
 168                         return psdev;
 169                 }
 170         }
 171
 172         return NULL;
 173 }
 174
 175 static struct pcistub_device *pcistub_device_find(int domain, int bus,
 176                                                   int slot, int func)
 177 {
 178         struct pcistub_device *psdev;
 179         unsigned long flags;
 180
 181         spin_lock_irqsave(&pcistub_devices_lock, flags);
 182
 183         psdev = pcistub_device_find_locked(domain, bus, slot, func);
 184         if (psdev)
 185                 pcistub_device_get(psdev);
 186
 187         spin_unlock_irqrestore(&pcistub_devices_lock, flags);
 188         return psdev;
 189 }
 190
 191 static struct pci_dev *pcistub_device_get_pci_dev(struct xen_pcibk_device *pdev,
 192                                                   struct pcistub_device *psdev)
 193 {
 194         struct pci_dev *pci_dev = NULL;
 195         unsigned long flags;
 196
 197         pcistub_device_get(psdev);
 198
 199         spin_lock_irqsave(&psdev->lock, flags);
 200         if (!psdev->pdev) {
 201                 psdev->pdev = pdev;
 202                 pci_dev = psdev->dev;
 203         }
 204         spin_unlock_irqrestore(&psdev->lock, flags);
 205
 206         if (!pci_dev)
 207                 pcistub_device_put(psdev);
 208
 209         return pci_dev;
 210 }
 211
 212 struct pci_dev *pcistub_get_pci_dev_by_slot(struct xen_pcibk_device *pdev,
 213                                             int domain, int bus,
 214                                             int slot, int func)
 215 {
 216         struct pcistub_device *psdev;
 217         struct pci_dev *found_dev = NULL;
 218         unsigned long flags;
 219
 220         spin_lock_irqsave(&pcistub_devices_lock, flags);
 221
 222         psdev = pcistub_device_find_locked(domain, bus, slot, func);
 223         if (psdev)
 224                 found_dev = pcistub_device_get_pci_dev(pdev, psdev);
 225
 226         spin_unlock_irqrestore(&pcistub_devices_lock, flags);
 227         return found_dev;
 228 }
 229
 230 struct pci_dev *pcistub_get_pci_dev(struct xen_pcibk_device *pdev,
 231                                     struct pci_dev *dev)
 232 {
 233         struct pcistub_device *psdev;
 234         struct pci_dev *found_dev = NULL;
 235         unsigned long flags;
 236
 237         spin_lock_irqsave(&pcistub_devices_lock, flags);
 238
 239         list_for_each_entry(psdev, &pcistub_devices, dev_list) {
 240                 if (psdev->dev == dev) {
 241                         found_dev = pcistub_device_get_pci_dev(pdev, psdev);
 242                         break;
 243                 }
 244         }
 245
 246         spin_unlock_irqrestore(&pcistub_devices_lock, flags);
 247         return found_dev;
 248 }
 249
 250 /*
 251  * Called when:
 252  *  - XenBus state has been reconfigure (pci unplug). See xen_pcibk_remove_device
 253  *  - XenBus state has been disconnected (guest shutdown). See xen_pcibk_xenbus_remove
 254  *  - 'echo BDF > unbind' on pciback module with no guest attached. See pcistub_remove
 255  *  - 'echo BDF > unbind' with a guest still using it. See pcistub_remove
 256  *
 257  *  As such we have to be careful.
 258  *
 259  *  To make this easier, the caller has to hold the device lock.
 260  */
 261 void pcistub_put_pci_dev(struct pci_dev *dev)
 262 {
 263         struct pcistub_device *psdev, *found_psdev = NULL;
 264         unsigned long flags;
 265         struct xen_pcibk_dev_data *dev_data;
 266         int ret;
 267
 268         spin_lock_irqsave(&pcistub_devices_lock, flags);
 269
 270         list_for_each_entry(psdev, &pcistub_devices, dev_list) {
 271                 if (psdev->dev == dev) {
 272                         found_psdev = psdev;
 273                         break;
 274                 }
 275         }
 276
 277         spin_unlock_irqrestore(&pcistub_devices_lock, flags);
 278         if (WARN_ON(!found_psdev))
 279                 return;
 280
 281         /*hold this lock for avoiding breaking link between
 282         * pcistub and xen_pcibk when AER is in processing
 283         */
 284         down_write(&pcistub_sem);
 285         /* Cleanup our device
 286          * (so it's ready for the next domain)
 287          */
 288         device_lock_assert(&dev->dev);
 289         __pci_reset_function_locked(dev);
 290
 291         dev_data = pci_get_drvdata(dev);
 292         ret = pci_load_saved_state(dev, dev_data->pci_saved_state);
 293         if (!ret) {
 294                 /*
 295                  * The usual sequence is pci_save_state & pci_restore_state
 296                  * but the guest might have messed the configuration space up.
 297                  * Use the initial version (when device was bound to us).
 298                  */
 299                 pci_restore_state(dev);
 300         } else
 301                 dev_info(&dev->dev, "Could not reload PCI state\n");
 302         /* This disables the device. */
 303         xen_pcibk_reset_device(dev);
 304
 305         /* And cleanup up our emulated fields. */
 306         xen_pcibk_config_reset_dev(dev);
 307         xen_pcibk_config_free_dyn_fields(dev);
 308
 309         dev_data->allow_interrupt_control = 0;
 310
 311         xen_unregister_device_domain_owner(dev);
 312
 313         spin_lock_irqsave(&found_psdev->lock, flags);
 314         found_psdev->pdev = NULL;
 315         spin_unlock_irqrestore(&found_psdev->lock, flags);
 316
 317         pcistub_device_put(found_psdev);
 318         up_write(&pcistub_sem);
 319 }
 320
 321 static int pcistub_match_one(struct pci_dev *dev,
 322                              struct pcistub_device_id *pdev_id)
 323 {
 324         /* Match the specified device by domain, bus, slot, func and also if
 325          * any of the device's parent bridges match.
 326          */
 327         for (; dev != NULL; dev = dev->bus->self) {
 328                 if (pci_domain_nr(dev->bus) == pdev_id->domain
 329                     && dev->bus->number == pdev_id->bus
 330                     && dev->devfn == pdev_id->devfn)
 331                         return 1;
 332
 333                 /* Sometimes topmost bridge links to itself. */
 334                 if (dev == dev->bus->self)
 335                         break;
 336         }
 337
 338         return 0;
 339 }
 340
 341 static int pcistub_match(struct pci_dev *dev)
 342 {
 343         struct pcistub_device_id *pdev_id;
 344         unsigned long flags;
 345         int found = 0;
 346
 347         spin_lock_irqsave(&device_ids_lock, flags);
 348         list_for_each_entry(pdev_id, &pcistub_device_ids, slot_list) {
 349                 if (pcistub_match_one(dev, pdev_id)) {
 350                         found = 1;
 351                         break;
 352                 }
 353         }
 354         spin_unlock_irqrestore(&device_ids_lock, flags);
 355
 356         return found;
 357 }
 358
 359 static int pcistub_init_device(struct pci_dev *dev)
 360 {
 361         struct xen_pcibk_dev_data *dev_data;
 362         int err = 0;
 363
 364         dev_dbg(&dev->dev, "initializing...\n");
 365
 366         /* The PCI backend is not intended to be a module (or to work with
 367          * removable PCI devices (yet). If it were, xen_pcibk_config_free()
 368          * would need to be called somewhere to free the memory allocated
 369          * here and then to call kfree(pci_get_drvdata(psdev->dev)).
 370          */
 371         dev_data = kzalloc(sizeof(*dev_data) +  strlen(DRV_NAME "[]")
 372                                 + strlen(pci_name(dev)) + 1, GFP_KERNEL);
 373         if (!dev_data) {
 374                 err = -ENOMEM;
 375                 goto out;
 376         }
 377         pci_set_drvdata(dev, dev_data);
 378
 379         /*
 380          * Setup name for fake IRQ handler. It will only be enabled
 381          * once the device is turned on by the guest.
 382          */
 383         sprintf(dev_data->irq_name, DRV_NAME "[%s]", pci_name(dev));
 384
 385         dev_dbg(&dev->dev, "initializing config\n");
 386
 387         init_waitqueue_head(&xen_pcibk_aer_wait_queue);
 388         err = xen_pcibk_config_init_dev(dev);
 389         if (err)
 390                 goto out;
 391
 392         /* HACK: Force device (& ACPI) to determine what IRQ it's on - we
 393          * must do this here because pcibios_enable_device may specify
 394          * the pci device's true irq (and possibly its other resources)
 395          * if they differ from what's in the configuration space.
 396          * This makes the assumption that the device's resources won't
 397          * change after this point (otherwise this code may break!)
 398          */
 399         dev_dbg(&dev->dev, "enabling device\n");
 400         err = pci_enable_device(dev);
 401         if (err)
 402                 goto config_release;
 403
 404         if (dev->msix_cap) {
 405                 struct physdev_pci_device ppdev = {
 406                         .seg = pci_domain_nr(dev->bus),
 407                         .bus = dev->bus->number,
 408                         .devfn = dev->devfn
 409                 };
 410
 411                 err = HYPERVISOR_physdev_op(PHYSDEVOP_prepare_msix, &ppdev);
 412                 if (err && err != -ENOSYS)
 413                         dev_err(&dev->dev, "MSI-X preparation failed (%d)\n",
 414                                 err);
 415         }
 416
 417         /* We need the device active to save the state. */
 418         dev_dbg(&dev->dev, "save state of device\n");
 419         pci_save_state(dev);
 420         dev_data->pci_saved_state = pci_store_saved_state(dev);
 421         if (!dev_data->pci_saved_state)
 422                 dev_err(&dev->dev, "Could not store PCI conf saved state!\n");
 423         else {
 424                 dev_dbg(&dev->dev, "resetting (FLR, D3, etc) the device\n");
 425                 __pci_reset_function_locked(dev);
 426                 pci_restore_state(dev);
 427         }
 428         /* Now disable the device (this also ensures some private device
 429          * data is setup before we export)
 430          */
 431         dev_dbg(&dev->dev, "reset device\n");
 432         xen_pcibk_reset_device(dev);
 433
 434         pci_set_dev_assigned(dev);
 435         return 0;
 436
 437 config_release:
 438         xen_pcibk_config_free_dev(dev);
 439
 440 out:
 441         pci_set_drvdata(dev, NULL);
 442         kfree(dev_data);
 443         return err;
 444 }
 445
 446 /*
 447  * Because some initialization still happens on
 448  * devices during fs_initcall, we need to defer
 449  * full initialization of our devices until
 450  * device_initcall.
 451  */
 452 static int __init pcistub_init_devices_late(void)
 453 {
 454         struct pcistub_device *psdev;
 455         unsigned long flags;
 456         int err = 0;
 457
 458         spin_lock_irqsave(&pcistub_devices_lock, flags);
 459
 460         while (!list_empty(&seized_devices)) {
 461                 psdev = container_of(seized_devices.next,
 462                                      struct pcistub_device, dev_list);
 463                 list_del(&psdev->dev_list);
 464
 465                 spin_unlock_irqrestore(&pcistub_devices_lock, flags);
 466
 467                 err = pcistub_init_device(psdev->dev);
 468                 if (err) {
 469                         dev_err(&psdev->dev->dev,
 470                                 "error %d initializing device\n", err);
 471                         kfree(psdev);
 472                         psdev = NULL;
 473                 }
 474
 475                 spin_lock_irqsave(&pcistub_devices_lock, flags);
 476
 477                 if (psdev)
 478                         list_add_tail(&psdev->dev_list, &pcistub_devices);
 479         }
 480
 481         initialize_devices = 1;
 482
 483         spin_unlock_irqrestore(&pcistub_devices_lock, flags);
 484
 485         return 0;
 486 }
 487
 488 static void pcistub_device_id_add_list(struct pcistub_device_id *new,
 489                                        int domain, int bus, unsigned int devfn)
 490 {
 491         struct pcistub_device_id *pci_dev_id;
 492         unsigned long flags;
 493         int found = 0;
 494
 495         spin_lock_irqsave(&device_ids_lock, flags);
 496
 497         list_for_each_entry(pci_dev_id, &pcistub_device_ids, slot_list) {
 498                 if (pci_dev_id->domain == domain && pci_dev_id->bus == bus &&
 499                     pci_dev_id->devfn == devfn) {
 500                         found = 1;
 501                         break;
 502                 }
 503         }
 504
 505         if (!found) {
 506                 new->domain = domain;
 507                 new->bus = bus;
 508                 new->devfn = devfn;
 509                 list_add_tail(&new->slot_list, &pcistub_device_ids);
 510         }
 511
 512         spin_unlock_irqrestore(&device_ids_lock, flags);
 513
 514         if (found)
 515                 kfree(new);
 516 }
 517
 518 static int pcistub_seize(struct pci_dev *dev,
 519                          struct pcistub_device_id *pci_dev_id)
 520 {
 521         struct pcistub_device *psdev;
 522         unsigned long flags;
 523         int err = 0;
 524
 525         psdev = pcistub_device_alloc(dev);
 526         if (!psdev) {
 527                 kfree(pci_dev_id);
 528                 return -ENOMEM;
 529         }
 530
 531         spin_lock_irqsave(&pcistub_devices_lock, flags);
 532
 533         if (initialize_devices) {
 534                 spin_unlock_irqrestore(&pcistub_devices_lock, flags);
 535
 536                 /* don't want irqs disabled when calling pcistub_init_device */
 537                 err = pcistub_init_device(psdev->dev);
 538
 539                 spin_lock_irqsave(&pcistub_devices_lock, flags);
 540
 541                 if (!err)
 542                         list_add(&psdev->dev_list, &pcistub_devices);
 543         } else {
 544                 dev_dbg(&dev->dev, "deferring initialization\n");
 545                 list_add(&psdev->dev_list, &seized_devices);
 546         }
 547
 548         spin_unlock_irqrestore(&pcistub_devices_lock, flags);
 549
 550         if (err) {
 551                 kfree(pci_dev_id);
 552                 pcistub_device_put(psdev);
 553         } else if (pci_dev_id)
 554                 pcistub_device_id_add_list(pci_dev_id, pci_domain_nr(dev->bus),
 555                                            dev->bus->number, dev->devfn);
 556
 557         return err;
 558 }
 559
 560 /* Called when 'bind'. This means we must _NOT_ call pci_reset_function or
 561  * other functions that take the sysfs lock. */
 562 static int pcistub_probe(struct pci_dev *dev, const struct pci_device_id *id)
 563 {
 564         int err = 0, match;
 565         struct pcistub_device_id *pci_dev_id = NULL;
 566
 567         dev_dbg(&dev->dev, "probing...\n");
 568
 569         match = pcistub_match(dev);
 570
 571         if ((dev->driver_override &&
 572              !strcmp(dev->driver_override, PCISTUB_DRIVER_NAME)) ||
 573             match) {
 574
 575                 if (dev->hdr_type != PCI_HEADER_TYPE_NORMAL
 576                     && dev->hdr_type != PCI_HEADER_TYPE_BRIDGE) {
 577                         dev_err(&dev->dev, "can't export pci devices that "
 578                                 "don't have a normal (0) or bridge (1) "
 579                                 "header type!\n");
 580                         err = -ENODEV;
 581                         goto out;
 582                 }
 583
 584                 if (!match) {
 585                         pci_dev_id = kmalloc(sizeof(*pci_dev_id), GFP_KERNEL);
 586                         if (!pci_dev_id) {
 587                                 err = -ENOMEM;
 588                                 goto out;
 589                         }
 590                 }
 591
 592                 dev_info(&dev->dev, "seizing device\n");
 593                 err = pcistub_seize(dev, pci_dev_id);
 594         } else
 595                 /* Didn't find the device */
 596                 err = -ENODEV;
 597
 598 out:
 599         return err;
 600 }
 601
 602 /* Called when 'unbind'. This means we must _NOT_ call pci_reset_function or
 603  * other functions that take the sysfs lock. */
 604 static void pcistub_remove(struct pci_dev *dev)
 605 {
 606         struct pcistub_device *psdev, *found_psdev = NULL;
 607         unsigned long flags;
 608
 609         dev_dbg(&dev->dev, "removing\n");
 610
 611         spin_lock_irqsave(&pcistub_devices_lock, flags);
 612
 613         xen_pcibk_config_quirk_release(dev);
 614
 615         list_for_each_entry(psdev, &pcistub_devices, dev_list) {
 616                 if (psdev->dev == dev) {
 617                         found_psdev = psdev;
 618                         break;
 619                 }
 620         }
 621
 622         spin_unlock_irqrestore(&pcistub_devices_lock, flags);
 623
 624         if (found_psdev) {
 625                 dev_dbg(&dev->dev, "found device to remove %s\n",
 626                         found_psdev->pdev ? "- in-use" : "");
 627
 628                 if (found_psdev->pdev) {
 629                         int domid = xen_find_device_domain_owner(dev);
 630
 631                         dev_warn(&dev->dev, "****** removing device %s while still in-use by domain %d! ******\n",
 632                                pci_name(found_psdev->dev), domid);
 633                         dev_warn(&dev->dev, "****** driver domain may still access this device's i/o resources!\n");
 634                         dev_warn(&dev->dev, "****** shutdown driver domain before binding device\n");
 635                         dev_warn(&dev->dev, "****** to other drivers or domains\n");
 636
 637                         /* N.B. This ends up calling pcistub_put_pci_dev which ends up
 638                          * doing the FLR. */
 639                         xen_pcibk_release_pci_dev(found_psdev->pdev,
 640                                                 found_psdev->dev,
 641                                                 false /* caller holds the lock. */);
 642                 }
 643
 644                 spin_lock_irqsave(&pcistub_devices_lock, flags);
 645                 list_del(&found_psdev->dev_list);
 646                 spin_unlock_irqrestore(&pcistub_devices_lock, flags);
 647
 648                 /* the final put for releasing from the list */
 649                 pcistub_device_put(found_psdev);
 650         }
 651 }
 652
 653 static const struct pci_device_id pcistub_ids[] = {
 654         {
 655          .vendor = PCI_ANY_ID,
 656          .device = PCI_ANY_ID,
 657          .subvendor = PCI_ANY_ID,
 658          .subdevice = PCI_ANY_ID,
 659          },
 660         {0,},
 661 };
 662
 663 #define PCI_NODENAME_MAX 40
 664 static void kill_domain_by_device(struct pcistub_device *psdev)
 665 {
 666         struct xenbus_transaction xbt;
 667         int err;
 668         char nodename[PCI_NODENAME_MAX];
 669
 670         BUG_ON(!psdev);
 671         snprintf(nodename, PCI_NODENAME_MAX, "/local/domain/0/backend/pci/%d/0",
 672                 psdev->pdev->xdev->otherend_id);
 673
 674 again:
 675         err = xenbus_transaction_start(&xbt);
 676         if (err) {
 677                 dev_err(&psdev->dev->dev,
 678                         "error %d when start xenbus transaction\n", err);
 679                 return;
 680         }
 681         /*PV AER handlers will set this flag*/
 682         xenbus_printf(xbt, nodename, "aerState" , "aerfail");
 683         err = xenbus_transaction_end(xbt, 0);
 684         if (err) {
 685                 if (err == -EAGAIN)
 686                         goto again;
 687                 dev_err(&psdev->dev->dev,
 688                         "error %d when end xenbus transaction\n", err);
 689                 return;
 690         }
 691 }
 692
 693 /* For each aer recovery step error_detected, mmio_enabled, etc, front_end and
 694  * backend need to have cooperation. In xen_pcibk, those steps will do similar
 695  * jobs: send service request and waiting for front_end response.
 696 */
 697 static pci_ers_result_t common_process(struct pcistub_device *psdev,
 698                                        pci_channel_state_t state, int aer_cmd,
 699                                        pci_ers_result_t result)
 700 {
 701         pci_ers_result_t res = result;
 702         struct xen_pcie_aer_op *aer_op;
 703         struct xen_pcibk_device *pdev = psdev->pdev;
 704         struct xen_pci_sharedinfo *sh_info = pdev->sh_info;
 705         int ret;
 706
 707         /*with PV AER drivers*/
 708         aer_op = &(sh_info->aer_op);
 709         aer_op->cmd = aer_cmd ;
 710         /*useful for error_detected callback*/
 711         aer_op->err = state;
 712         /*pcifront_end BDF*/
 713         ret = xen_pcibk_get_pcifront_dev(psdev->dev, psdev->pdev,
 714                 &aer_op->domain, &aer_op->bus, &aer_op->devfn);
 715         if (!ret) {
 716                 dev_err(&psdev->dev->dev, "failed to get pcifront device\n");
 717                 return PCI_ERS_RESULT_NONE;
 718         }
 719         wmb();
 720
 721         dev_dbg(&psdev->dev->dev, "aer_op %x dom %x bus %x devfn %x\n",
 722                         aer_cmd, aer_op->domain, aer_op->bus, aer_op->devfn);
 723         /*local flag to mark there's aer request, xen_pcibk callback will use
 724         * this flag to judge whether we need to check pci-front give aer
 725         * service ack signal
 726         */
 727         set_bit(_PCIB_op_pending, (unsigned long *)&pdev->flags);
 728
 729         /*It is possible that a pcifront conf_read_write ops request invokes
 730         * the callback which cause the spurious execution of wake_up.
 731         * Yet it is harmless and better than a spinlock here
 732         */
 733         set_bit(_XEN_PCIB_active,
 734                 (unsigned long *)&sh_info->flags);
 735         wmb();
 736         notify_remote_via_irq(pdev->evtchn_irq);
 737
 738         /* Enable IRQ to signal "request done". */
 739         xen_pcibk_lateeoi(pdev, 0);
 740
 741         ret = wait_event_timeout(xen_pcibk_aer_wait_queue,
 742                                  !(test_bit(_XEN_PCIB_active, (unsigned long *)
 743                                  &sh_info->flags)), 300*HZ);
 744
 745         /* Enable IRQ for pcifront request if not already active. */
 746         if (!test_bit(_PDEVF_op_active, &pdev->flags))
 747                 xen_pcibk_lateeoi(pdev, 0);
 748
 749         if (!ret) {
 750                 if (test_bit(_XEN_PCIB_active,
 751                         (unsigned long *)&sh_info->flags)) {
 752                         dev_err(&psdev->dev->dev,
 753                                 "pcifront aer process not responding!\n");
 754                         clear_bit(_XEN_PCIB_active,
 755                           (unsigned long *)&sh_info->flags);
 756                         aer_op->err = PCI_ERS_RESULT_NONE;
 757                         return res;
 758                 }
 759         }
 760         clear_bit(_PCIB_op_pending, (unsigned long *)&pdev->flags);
 761
 762         res = (pci_ers_result_t)aer_op->err;
 763         return res;
 764 }
 765
 766 /*
 767 * xen_pcibk_slot_reset: it will send the slot_reset request to  pcifront in case
 768 * of the device driver could provide this service, and then wait for pcifront
 769 * ack.
 770 * @dev: pointer to PCI devices
 771 * return value is used by aer_core do_recovery policy
 772 */
 773 static pci_ers_result_t xen_pcibk_slot_reset(struct pci_dev *dev)
 774 {
 775         struct pcistub_device *psdev;
 776         pci_ers_result_t result;
 777
 778         result = PCI_ERS_RESULT_RECOVERED;
 779         dev_dbg(&dev->dev, "xen_pcibk_slot_reset(bus:%x,devfn:%x)\n",
 780                 dev->bus->number, dev->devfn);
 781
 782         down_write(&pcistub_sem);
 783         psdev = pcistub_device_find(pci_domain_nr(dev->bus),
 784                                 dev->bus->number,
 785                                 PCI_SLOT(dev->devfn),
 786                                 PCI_FUNC(dev->devfn));
 787
 788         if (!psdev || !psdev->pdev) {
 789                 dev_err(&dev->dev, "device is not found/assigned\n");
 790                 goto end;
 791         }
 792
 793         if (!psdev->pdev->sh_info) {
 794                 dev_err(&dev->dev, "device is not connected or owned"
 795                         " by HVM, kill it\n");
 796                 kill_domain_by_device(psdev);
 797                 goto end;
 798         }
 799
 800         if (!test_bit(_XEN_PCIB_AERHANDLER,
 801                 (unsigned long *)&psdev->pdev->sh_info->flags)) {
 802                 dev_err(&dev->dev,
 803                         "guest with no AER driver should have been killed\n");
 804                 goto end;
 805         }
 806         result = common_process(psdev, pci_channel_io_normal, XEN_PCI_OP_aer_slotreset, result);
 807
 808         if (result == PCI_ERS_RESULT_NONE ||
 809                 result == PCI_ERS_RESULT_DISCONNECT) {
 810                 dev_dbg(&dev->dev,
 811                         "No AER slot_reset service or disconnected!\n");
 812                 kill_domain_by_device(psdev);
 813         }
 814 end:
 815         if (psdev)
 816                 pcistub_device_put(psdev);
 817         up_write(&pcistub_sem);
 818         return result;
 819
 820 }
 821
 822
 823 /*xen_pcibk_mmio_enabled: it will send the mmio_enabled request to  pcifront
 824 * in case of the device driver could provide this service, and then wait
 825 * for pcifront ack
 826 * @dev: pointer to PCI devices
 827 * return value is used by aer_core do_recovery policy
 828 */
 829
 830 static pci_ers_result_t xen_pcibk_mmio_enabled(struct pci_dev *dev)
 831 {
 832         struct pcistub_device *psdev;
 833         pci_ers_result_t result;
 834
 835         result = PCI_ERS_RESULT_RECOVERED;
 836         dev_dbg(&dev->dev, "xen_pcibk_mmio_enabled(bus:%x,devfn:%x)\n",
 837                 dev->bus->number, dev->devfn);
 838
 839         down_write(&pcistub_sem);
 840         psdev = pcistub_device_find(pci_domain_nr(dev->bus),
 841                                 dev->bus->number,
 842                                 PCI_SLOT(dev->devfn),
 843                                 PCI_FUNC(dev->devfn));
 844
 845         if (!psdev || !psdev->pdev) {
 846                 dev_err(&dev->dev, "device is not found/assigned\n");
 847                 goto end;
 848         }
 849
 850         if (!psdev->pdev->sh_info) {
 851                 dev_err(&dev->dev, "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, pci_channel_io_normal, 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, "device is not found/assigned\n");
 904                 goto end;
 905         }
 906
 907         if (!psdev->pdev->sh_info) {
 908                 dev_err(&dev->dev, "device is not connected or owned"
 909                         " by HVM, kill it\n");
 910                 kill_domain_by_device(psdev);
 911                 goto end;
 912         }
 913
 914         /*Guest owns the device yet no aer handler regiested, kill guest*/
 915         if (!test_bit(_XEN_PCIB_AERHANDLER,
 916                 (unsigned long *)&psdev->pdev->sh_info->flags)) {
 917                 dev_dbg(&dev->dev, "guest may have no aer driver, kill it\n");
 918                 kill_domain_by_device(psdev);
 919                 goto end;
 920         }
 921         result = common_process(psdev, error, XEN_PCI_OP_aer_detected, result);
 922
 923         if (result == PCI_ERS_RESULT_NONE ||
 924                 result == PCI_ERS_RESULT_DISCONNECT) {
 925                 dev_dbg(&dev->dev,
 926                         "No AER error_detected service or disconnected!\n");
 927                 kill_domain_by_device(psdev);
 928         }
 929 end:
 930         if (psdev)
 931                 pcistub_device_put(psdev);
 932         up_write(&pcistub_sem);
 933         return result;
 934 }
 935
 936 /*xen_pcibk_error_resume: it will send the error_resume request to  pcifront
 937 * in case of the device driver could provide this service, and then wait
 938 * for pcifront ack.
 939 * @dev: pointer to PCI devices
 940 */
 941
 942 static void xen_pcibk_error_resume(struct pci_dev *dev)
 943 {
 944         struct pcistub_device *psdev;
 945
 946         dev_dbg(&dev->dev, "xen_pcibk_error_resume(bus:%x,devfn:%x)\n",
 947                 dev->bus->number, dev->devfn);
 948
 949         down_write(&pcistub_sem);
 950         psdev = pcistub_device_find(pci_domain_nr(dev->bus),
 951                                 dev->bus->number,
 952                                 PCI_SLOT(dev->devfn),
 953                                 PCI_FUNC(dev->devfn));
 954
 955         if (!psdev || !psdev->pdev) {
 956                 dev_err(&dev->dev, "device is not found/assigned\n");
 957                 goto end;
 958         }
 959
 960         if (!psdev->pdev->sh_info) {
 961                 dev_err(&dev->dev, "device is not connected or owned"
 962                         " by HVM, kill it\n");
 963                 kill_domain_by_device(psdev);
 964                 goto end;
 965         }
 966
 967         if (!test_bit(_XEN_PCIB_AERHANDLER,
 968                 (unsigned long *)&psdev->pdev->sh_info->flags)) {
 969                 dev_err(&dev->dev,
 970                         "guest with no AER driver should have been killed\n");
 971                 kill_domain_by_device(psdev);
 972                 goto end;
 973         }
 974         common_process(psdev, pci_channel_io_normal, XEN_PCI_OP_aer_resume,
 975                        PCI_ERS_RESULT_RECOVERED);
 976 end:
 977         if (psdev)
 978                 pcistub_device_put(psdev);
 979         up_write(&pcistub_sem);
 980         return;
 981 }
 982
 983 /*add xen_pcibk AER handling*/
 984 static const struct pci_error_handlers xen_pcibk_error_handler = {
 985         .error_detected = xen_pcibk_error_detected,
 986         .mmio_enabled = xen_pcibk_mmio_enabled,
 987         .slot_reset = xen_pcibk_slot_reset,
 988         .resume = xen_pcibk_error_resume,
 989 };
 990
 991 /*
 992  * Note: There is no MODULE_DEVICE_TABLE entry here because this isn't
 993  * for a normal device. I don't want it to be loaded automatically.
 994  */
 995
 996 static struct pci_driver xen_pcibk_pci_driver = {
 997         /* The name should be xen_pciback, but until the tools are updated
 998          * we will keep it as pciback. */
 999         .name = PCISTUB_DRIVER_NAME,
1000         .id_table = pcistub_ids,
1001         .probe = pcistub_probe,
1002         .remove = pcistub_remove,
1003         .err_handler = &xen_pcibk_error_handler,
1004 };
1005
1006 static inline int str_to_slot(const char *buf, int *domain, int *bus,
1007                               int *slot, int *func)
1008 {
1009         int parsed = 0;
1010
1011         switch (sscanf(buf, " %x:%x:%x.%x %n", domain, bus, slot, func,
1012                        &parsed)) {
1013         case 3:
1014                 *func = -1;
1015                 sscanf(buf, " %x:%x:%x.* %n", domain, bus, slot, &parsed);
1016                 break;
1017         case 2:
1018                 *slot = *func = -1;
1019                 sscanf(buf, " %x:%x:*.* %n", domain, bus, &parsed);
1020                 break;
1021         }
1022         if (parsed && !buf[parsed])
1023                 return 0;
1024
1025         /* try again without domain */
1026         *domain = 0;
1027         switch (sscanf(buf, " %x:%x.%x %n", bus, slot, func, &parsed)) {
1028         case 2:
1029                 *func = -1;
1030                 sscanf(buf, " %x:%x.* %n", bus, slot, &parsed);
1031                 break;
1032         case 1:
1033                 *slot = *func = -1;
1034                 sscanf(buf, " %x:*.* %n", bus, &parsed);
1035                 break;
1036         }
1037         if (parsed && !buf[parsed])
1038                 return 0;
1039
1040         return -EINVAL;
1041 }
1042
1043 static inline int str_to_quirk(const char *buf, int *domain, int *bus, int
1044                                *slot, int *func, int *reg, int *size, int *mask)
1045 {
1046         int parsed = 0;
1047
1048         sscanf(buf, " %x:%x:%x.%x-%x:%x:%x %n", domain, bus, slot, func,
1049                reg, size, mask, &parsed);
1050         if (parsed && !buf[parsed])
1051                 return 0;
1052
1053         /* try again without domain */
1054         *domain = 0;
1055         sscanf(buf, " %x:%x.%x-%x:%x:%x %n", bus, slot, func, reg, size,
1056                mask, &parsed);
1057         if (parsed && !buf[parsed])
1058                 return 0;
1059
1060         return -EINVAL;
1061 }
1062
1063 static int pcistub_device_id_add(int domain, int bus, int slot, int func)
1064 {
1065         struct pcistub_device_id *pci_dev_id;
1066         int rc = 0, devfn = PCI_DEVFN(slot, func);
1067
1068         if (slot < 0) {
1069                 for (slot = 0; !rc && slot < 32; ++slot)
1070                         rc = pcistub_device_id_add(domain, bus, slot, func);
1071                 return rc;
1072         }
1073
1074         if (func < 0) {
1075                 for (func = 0; !rc && func < 8; ++func)
1076                         rc = pcistub_device_id_add(domain, bus, slot, func);
1077                 return rc;
1078         }
1079
1080         if ((
1081 #if !defined(MODULE) /* pci_domains_supported is not being exported */ \
1082     || !defined(CONFIG_PCI_DOMAINS)
1083              !pci_domains_supported ? domain :
1084 #endif
1085              domain < 0 || domain > 0xffff)
1086             || bus < 0 || bus > 0xff
1087             || PCI_SLOT(devfn) != slot
1088             || PCI_FUNC(devfn) != func)
1089                 return -EINVAL;
1090
1091         pci_dev_id = kmalloc(sizeof(*pci_dev_id), GFP_KERNEL);
1092         if (!pci_dev_id)
1093                 return -ENOMEM;
1094
1095         pr_debug("wants to seize %04x:%02x:%02x.%d\n",
1096                  domain, bus, slot, func);
1097
1098         pcistub_device_id_add_list(pci_dev_id, domain, bus, devfn);
1099
1100         return 0;
1101 }
1102
1103 static int pcistub_device_id_remove(int domain, int bus, int slot, int func)
1104 {
1105         struct pcistub_device_id *pci_dev_id, *t;
1106         int err = -ENOENT;
1107         unsigned long flags;
1108
1109         spin_lock_irqsave(&device_ids_lock, flags);
1110         list_for_each_entry_safe(pci_dev_id, t, &pcistub_device_ids,
1111                                  slot_list) {
1112                 if (pci_dev_id->domain == domain && pci_dev_id->bus == bus
1113                     && (slot < 0 || PCI_SLOT(pci_dev_id->devfn) == slot)
1114                     && (func < 0 || PCI_FUNC(pci_dev_id->devfn) == func)) {
1115                         /* Don't break; here because it's possible the same
1116                          * slot could be in the list more than once
1117                          */
1118                         list_del(&pci_dev_id->slot_list);
1119                         kfree(pci_dev_id);
1120
1121                         err = 0;
1122
1123                         pr_debug("removed %04x:%02x:%02x.%d from seize list\n",
1124                                  domain, bus, slot, func);
1125                 }
1126         }
1127         spin_unlock_irqrestore(&device_ids_lock, flags);
1128
1129         return err;
1130 }
1131
1132 static int pcistub_reg_add(int domain, int bus, int slot, int func,
1133                            unsigned int reg, unsigned int size,
1134                            unsigned int mask)
1135 {
1136         int err = 0;
1137         struct pcistub_device *psdev;
1138         struct pci_dev *dev;
1139         struct config_field *field;
1140
1141         if (reg > 0xfff || (size < 4 && (mask >> (size * 8))))
1142                 return -EINVAL;
1143
1144         psdev = pcistub_device_find(domain, bus, slot, func);
1145         if (!psdev) {
1146                 err = -ENODEV;
1147                 goto out;
1148         }
1149         dev = psdev->dev;
1150
1151         field = kzalloc(sizeof(*field), GFP_KERNEL);
1152         if (!field) {
1153                 err = -ENOMEM;
1154                 goto out;
1155         }
1156
1157         field->offset = reg;
1158         field->size = size;
1159         field->mask = mask;
1160         field->init = NULL;
1161         field->reset = NULL;
1162         field->release = NULL;
1163         field->clean = xen_pcibk_config_field_free;
1164
1165         err = xen_pcibk_config_quirks_add_field(dev, field);
1166         if (err)
1167                 kfree(field);
1168 out:
1169         if (psdev)
1170                 pcistub_device_put(psdev);
1171         return err;
1172 }
1173
1174 static ssize_t new_slot_store(struct device_driver *drv, const char *buf,
1175                               size_t count)
1176 {
1177         int domain, bus, slot, func;
1178         int err;
1179
1180         err = str_to_slot(buf, &domain, &bus, &slot, &func);
1181         if (err)
1182                 goto out;
1183
1184         err = pcistub_device_id_add(domain, bus, slot, func);
1185
1186 out:
1187         if (!err)
1188                 err = count;
1189         return err;
1190 }
1191 static DRIVER_ATTR_WO(new_slot);
1192
1193 static ssize_t remove_slot_store(struct device_driver *drv, const char *buf,
1194                                  size_t count)
1195 {
1196         int domain, bus, slot, func;
1197         int err;
1198
1199         err = str_to_slot(buf, &domain, &bus, &slot, &func);
1200         if (err)
1201                 goto out;
1202
1203         err = pcistub_device_id_remove(domain, bus, slot, func);
1204
1205 out:
1206         if (!err)
1207                 err = count;
1208         return err;
1209 }
1210 static DRIVER_ATTR_WO(remove_slot);
1211
1212 static ssize_t slots_show(struct device_driver *drv, char *buf)
1213 {
1214         struct pcistub_device_id *pci_dev_id;
1215         size_t count = 0;
1216         unsigned long flags;
1217
1218         spin_lock_irqsave(&device_ids_lock, flags);
1219         list_for_each_entry(pci_dev_id, &pcistub_device_ids, slot_list) {
1220                 if (count >= PAGE_SIZE)
1221                         break;
1222
1223                 count += scnprintf(buf + count, PAGE_SIZE - count,
1224                                    "%04x:%02x:%02x.%d\n",
1225                                    pci_dev_id->domain, pci_dev_id->bus,
1226                                    PCI_SLOT(pci_dev_id->devfn),
1227                                    PCI_FUNC(pci_dev_id->devfn));
1228         }
1229         spin_unlock_irqrestore(&device_ids_lock, flags);
1230
1231         return count;
1232 }
1233 static DRIVER_ATTR_RO(slots);
1234
1235 static ssize_t irq_handlers_show(struct device_driver *drv, char *buf)
1236 {
1237         struct pcistub_device *psdev;
1238         struct xen_pcibk_dev_data *dev_data;
1239         size_t count = 0;
1240         unsigned long flags;
1241
1242         spin_lock_irqsave(&pcistub_devices_lock, flags);
1243         list_for_each_entry(psdev, &pcistub_devices, dev_list) {
1244                 if (count >= PAGE_SIZE)
1245                         break;
1246                 if (!psdev->dev)
1247                         continue;
1248                 dev_data = pci_get_drvdata(psdev->dev);
1249                 if (!dev_data)
1250                         continue;
1251                 count +=
1252                     scnprintf(buf + count, PAGE_SIZE - count,
1253                               "%s:%s:%sing:%ld\n",
1254                               pci_name(psdev->dev),
1255                               dev_data->isr_on ? "on" : "off",
1256                               dev_data->ack_intr ? "ack" : "not ack",
1257                               dev_data->handled);
1258         }
1259         spin_unlock_irqrestore(&pcistub_devices_lock, flags);
1260         return count;
1261 }
1262 static DRIVER_ATTR_RO(irq_handlers);
1263
1264 static ssize_t irq_handler_state_store(struct device_driver *drv,
1265                                        const char *buf, size_t count)
1266 {
1267         struct pcistub_device *psdev;
1268         struct xen_pcibk_dev_data *dev_data;
1269         int domain, bus, slot, func;
1270         int err;
1271
1272         err = str_to_slot(buf, &domain, &bus, &slot, &func);
1273         if (err)
1274                 return err;
1275
1276         psdev = pcistub_device_find(domain, bus, slot, func);
1277         if (!psdev) {
1278                 err = -ENOENT;
1279                 goto out;
1280         }
1281
1282         dev_data = pci_get_drvdata(psdev->dev);
1283         if (!dev_data) {
1284                 err = -ENOENT;
1285                 goto out;
1286         }
1287
1288         dev_dbg(&psdev->dev->dev, "%s fake irq handler: %d->%d\n",
1289                 dev_data->irq_name, dev_data->isr_on,
1290                 !dev_data->isr_on);
1291
1292         dev_data->isr_on = !(dev_data->isr_on);
1293         if (dev_data->isr_on)
1294                 dev_data->ack_intr = 1;
1295 out:
1296         if (psdev)
1297                 pcistub_device_put(psdev);
1298         if (!err)
1299                 err = count;
1300         return err;
1301 }
1302 static DRIVER_ATTR_WO(irq_handler_state);
1303
1304 static ssize_t quirks_store(struct device_driver *drv, const char *buf,
1305                             size_t count)
1306 {
1307         int domain, bus, slot, func, reg, size, mask;
1308         int err;
1309
1310         err = str_to_quirk(buf, &domain, &bus, &slot, &func, &reg, &size,
1311                            &mask);
1312         if (err)
1313                 goto out;
1314
1315         err = pcistub_reg_add(domain, bus, slot, func, reg, size, mask);
1316
1317 out:
1318         if (!err)
1319                 err = count;
1320         return err;
1321 }
1322
1323 static ssize_t quirks_show(struct device_driver *drv, char *buf)
1324 {
1325         int count = 0;
1326         unsigned long flags;
1327         struct xen_pcibk_config_quirk *quirk;
1328         struct xen_pcibk_dev_data *dev_data;
1329         const struct config_field *field;
1330         const struct config_field_entry *cfg_entry;
1331
1332         spin_lock_irqsave(&device_ids_lock, flags);
1333         list_for_each_entry(quirk, &xen_pcibk_quirks, quirks_list) {
1334                 if (count >= PAGE_SIZE)
1335                         goto out;
1336
1337                 count += scnprintf(buf + count, PAGE_SIZE - count,
1338                                    "%02x:%02x.%01x\n\t%04x:%04x:%04x:%04x\n",
1339                                    quirk->pdev->bus->number,
1340                                    PCI_SLOT(quirk->pdev->devfn),
1341                                    PCI_FUNC(quirk->pdev->devfn),
1342                                    quirk->devid.vendor, quirk->devid.device,
1343                                    quirk->devid.subvendor,
1344                                    quirk->devid.subdevice);
1345
1346                 dev_data = pci_get_drvdata(quirk->pdev);
1347
1348                 list_for_each_entry(cfg_entry, &dev_data->config_fields, list) {
1349                         field = cfg_entry->field;
1350                         if (count >= PAGE_SIZE)
1351                                 goto out;
1352
1353                         count += scnprintf(buf + count, PAGE_SIZE - count,
1354                                            "\t\t%08x:%01x:%08x\n",
1355                                            cfg_entry->base_offset +
1356                                            field->offset, field->size,
1357                                            field->mask);
1358                 }
1359         }
1360
1361 out:
1362         spin_unlock_irqrestore(&device_ids_lock, flags);
1363
1364         return count;
1365 }
1366 static DRIVER_ATTR_RW(quirks);
1367
1368 static ssize_t permissive_store(struct device_driver *drv, const char *buf,
1369                                 size_t count)
1370 {
1371         int domain, bus, slot, func;
1372         int err;
1373         struct pcistub_device *psdev;
1374         struct xen_pcibk_dev_data *dev_data;
1375
1376         err = str_to_slot(buf, &domain, &bus, &slot, &func);
1377         if (err)
1378                 goto out;
1379
1380         psdev = pcistub_device_find(domain, bus, slot, func);
1381         if (!psdev) {
1382                 err = -ENODEV;
1383                 goto out;
1384         }
1385
1386         dev_data = pci_get_drvdata(psdev->dev);
1387         /* the driver data for a device should never be null at this point */
1388         if (!dev_data) {
1389                 err = -ENXIO;
1390                 goto release;
1391         }
1392         if (!dev_data->permissive) {
1393                 dev_data->permissive = 1;
1394                 /* Let user know that what they're doing could be unsafe */
1395                 dev_warn(&psdev->dev->dev, "enabling permissive mode "
1396                          "configuration space accesses!\n");
1397                 dev_warn(&psdev->dev->dev,
1398                          "permissive mode is potentially unsafe!\n");
1399         }
1400 release:
1401         pcistub_device_put(psdev);
1402 out:
1403         if (!err)
1404                 err = count;
1405         return err;
1406 }
1407
1408 static ssize_t permissive_show(struct device_driver *drv, char *buf)
1409 {
1410         struct pcistub_device *psdev;
1411         struct xen_pcibk_dev_data *dev_data;
1412         size_t count = 0;
1413         unsigned long flags;
1414         spin_lock_irqsave(&pcistub_devices_lock, flags);
1415         list_for_each_entry(psdev, &pcistub_devices, dev_list) {
1416                 if (count >= PAGE_SIZE)
1417                         break;
1418                 if (!psdev->dev)
1419                         continue;
1420                 dev_data = pci_get_drvdata(psdev->dev);
1421                 if (!dev_data || !dev_data->permissive)
1422                         continue;
1423                 count +=
1424                     scnprintf(buf + count, PAGE_SIZE - count, "%s\n",
1425                               pci_name(psdev->dev));
1426         }
1427         spin_unlock_irqrestore(&pcistub_devices_lock, flags);
1428         return count;
1429 }
1430 static DRIVER_ATTR_RW(permissive);
1431
1432 static ssize_t allow_interrupt_control_store(struct device_driver *drv,
1433                                              const char *buf, size_t count)
1434 {
1435         int domain, bus, slot, func;
1436         int err;
1437         struct pcistub_device *psdev;
1438         struct xen_pcibk_dev_data *dev_data;
1439
1440         err = str_to_slot(buf, &domain, &bus, &slot, &func);
1441         if (err)
1442                 goto out;
1443
1444         psdev = pcistub_device_find(domain, bus, slot, func);
1445         if (!psdev) {
1446                 err = -ENODEV;
1447                 goto out;
1448         }
1449
1450         dev_data = pci_get_drvdata(psdev->dev);
1451         /* the driver data for a device should never be null at this point */
1452         if (!dev_data) {
1453                 err = -ENXIO;
1454                 goto release;
1455         }
1456         dev_data->allow_interrupt_control = 1;
1457 release:
1458         pcistub_device_put(psdev);
1459 out:
1460         if (!err)
1461                 err = count;
1462         return err;
1463 }
1464
1465 static ssize_t allow_interrupt_control_show(struct device_driver *drv,
1466                                             char *buf)
1467 {
1468         struct pcistub_device *psdev;
1469         struct xen_pcibk_dev_data *dev_data;
1470         size_t count = 0;
1471         unsigned long flags;
1472
1473         spin_lock_irqsave(&pcistub_devices_lock, flags);
1474         list_for_each_entry(psdev, &pcistub_devices, dev_list) {
1475                 if (count >= PAGE_SIZE)
1476                         break;
1477                 if (!psdev->dev)
1478                         continue;
1479                 dev_data = pci_get_drvdata(psdev->dev);
1480                 if (!dev_data || !dev_data->allow_interrupt_control)
1481                         continue;
1482                 count +=
1483                     scnprintf(buf + count, PAGE_SIZE - count, "%s\n",
1484                               pci_name(psdev->dev));
1485         }
1486         spin_unlock_irqrestore(&pcistub_devices_lock, flags);
1487         return count;
1488 }
1489 static DRIVER_ATTR_RW(allow_interrupt_control);
1490
1491 static void pcistub_exit(void)
1492 {
1493         driver_remove_file(&xen_pcibk_pci_driver.driver, &driver_attr_new_slot);
1494         driver_remove_file(&xen_pcibk_pci_driver.driver,
1495                            &driver_attr_remove_slot);
1496         driver_remove_file(&xen_pcibk_pci_driver.driver, &driver_attr_slots);
1497         driver_remove_file(&xen_pcibk_pci_driver.driver, &driver_attr_quirks);
1498         driver_remove_file(&xen_pcibk_pci_driver.driver,
1499                            &driver_attr_permissive);
1500         driver_remove_file(&xen_pcibk_pci_driver.driver,
1501                            &driver_attr_allow_interrupt_control);
1502         driver_remove_file(&xen_pcibk_pci_driver.driver,
1503                            &driver_attr_irq_handlers);
1504         driver_remove_file(&xen_pcibk_pci_driver.driver,
1505                            &driver_attr_irq_handler_state);
1506         pci_unregister_driver(&xen_pcibk_pci_driver);
1507 }
1508
1509 static int __init pcistub_init(void)
1510 {
1511         int pos = 0;
1512         int err = 0;
1513         int domain, bus, slot, func;
1514         int parsed;
1515
1516         if (pci_devs_to_hide && *pci_devs_to_hide) {
1517                 do {
1518                         parsed = 0;
1519
1520                         err = sscanf(pci_devs_to_hide + pos,
1521                                      " (%x:%x:%x.%x) %n",
1522                                      &domain, &bus, &slot, &func, &parsed);
1523                         switch (err) {
1524                         case 3:
1525                                 func = -1;
1526                                 sscanf(pci_devs_to_hide + pos,
1527                                        " (%x:%x:%x.*) %n",
1528                                        &domain, &bus, &slot, &parsed);
1529                                 break;
1530                         case 2:
1531                                 slot = func = -1;
1532                                 sscanf(pci_devs_to_hide + pos,
1533                                        " (%x:%x:*.*) %n",
1534                                        &domain, &bus, &parsed);
1535                                 break;
1536                         }
1537
1538                         if (!parsed) {
1539                                 domain = 0;
1540                                 err = sscanf(pci_devs_to_hide + pos,
1541                                              " (%x:%x.%x) %n",
1542                                              &bus, &slot, &func, &parsed);
1543                                 switch (err) {
1544                                 case 2:
1545                                         func = -1;
1546                                         sscanf(pci_devs_to_hide + pos,
1547                                                " (%x:%x.*) %n",
1548                                                &bus, &slot, &parsed);
1549                                         break;
1550                                 case 1:
1551                                         slot = func = -1;
1552                                         sscanf(pci_devs_to_hide + pos,
1553                                                " (%x:*.*) %n",
1554                                                &bus, &parsed);
1555                                         break;
1556                                 }
1557                         }
1558
1559                         if (parsed <= 0)
1560                                 goto parse_error;
1561
1562                         err = pcistub_device_id_add(domain, bus, slot, func);
1563                         if (err)
1564                                 goto out;
1565
1566                         pos += parsed;
1567                 } while (pci_devs_to_hide[pos]);
1568         }
1569
1570         /* If we're the first PCI Device Driver to register, we're the
1571          * first one to get offered PCI devices as they become
1572          * available (and thus we can be the first to grab them)
1573          */
1574         err = pci_register_driver(&xen_pcibk_pci_driver);
1575         if (err < 0)
1576                 goto out;
1577
1578         err = driver_create_file(&xen_pcibk_pci_driver.driver,
1579                                  &driver_attr_new_slot);
1580         if (!err)
1581                 err = driver_create_file(&xen_pcibk_pci_driver.driver,
1582                                          &driver_attr_remove_slot);
1583         if (!err)
1584                 err = driver_create_file(&xen_pcibk_pci_driver.driver,
1585                                          &driver_attr_slots);
1586         if (!err)
1587                 err = driver_create_file(&xen_pcibk_pci_driver.driver,
1588                                          &driver_attr_quirks);
1589         if (!err)
1590                 err = driver_create_file(&xen_pcibk_pci_driver.driver,
1591                                          &driver_attr_permissive);
1592         if (!err)
1593                 err = driver_create_file(&xen_pcibk_pci_driver.driver,
1594                                          &driver_attr_allow_interrupt_control);
1595
1596         if (!err)
1597                 err = driver_create_file(&xen_pcibk_pci_driver.driver,
1598                                          &driver_attr_irq_handlers);
1599         if (!err)
1600                 err = driver_create_file(&xen_pcibk_pci_driver.driver,
1601                                         &driver_attr_irq_handler_state);
1602         if (err)
1603                 pcistub_exit();
1604
1605 out:
1606         return err;
1607
1608 parse_error:
1609         pr_err("Error parsing pci_devs_to_hide at \"%s\"\n",
1610                pci_devs_to_hide + pos);
1611         return -EINVAL;
1612 }
1613
1614 #ifndef MODULE
1615 /*
1616  * fs_initcall happens before device_initcall
1617  * so xen_pcibk *should* get called first (b/c we
1618  * want to suck up any device before other drivers
1619  * get a chance by being the first pci device
1620  * driver to register)
1621  */
1622 fs_initcall(pcistub_init);
1623 #endif
1624
1625 #ifdef CONFIG_PCI_IOV
1626 static struct pcistub_device *find_vfs(const struct pci_dev *pdev)
1627 {
1628         struct pcistub_device *psdev = NULL;
1629         unsigned long flags;
1630         bool found = false;
1631
1632         spin_lock_irqsave(&pcistub_devices_lock, flags);
1633         list_for_each_entry(psdev, &pcistub_devices, dev_list) {
1634                 if (!psdev->pdev && psdev->dev != pdev
1635                     && pci_physfn(psdev->dev) == pdev) {
1636                         found = true;
1637                         break;
1638                 }
1639         }
1640         spin_unlock_irqrestore(&pcistub_devices_lock, flags);
1641         if (found)
1642                 return psdev;
1643         return NULL;
1644 }
1645
1646 static int pci_stub_notifier(struct notifier_block *nb,
1647                              unsigned long action, void *data)
1648 {
1649         struct device *dev = data;
1650         const struct pci_dev *pdev = to_pci_dev(dev);
1651
1652         if (action != BUS_NOTIFY_UNBIND_DRIVER)
1653                 return NOTIFY_DONE;
1654
1655         if (!pdev->is_physfn)
1656                 return NOTIFY_DONE;
1657
1658         for (;;) {
1659                 struct pcistub_device *psdev = find_vfs(pdev);
1660                 if (!psdev)
1661                         break;
1662                 device_release_driver(&psdev->dev->dev);
1663         }
1664         return NOTIFY_DONE;
1665 }
1666
1667 static struct notifier_block pci_stub_nb = {
1668         .notifier_call = pci_stub_notifier,
1669 };
1670 #endif
1671
1672 static int __init xen_pcibk_init(void)
1673 {
1674         int err;
1675
1676         if (!xen_initial_domain())
1677                 return -ENODEV;
1678
1679         err = xen_pcibk_config_init();
1680         if (err)
1681                 return err;
1682
1683 #ifdef MODULE
1684         err = pcistub_init();
1685         if (err < 0)
1686                 return err;
1687 #endif
1688
1689         pcistub_init_devices_late();
1690         err = xen_pcibk_xenbus_register();
1691         if (err)
1692                 pcistub_exit();
1693 #ifdef CONFIG_PCI_IOV
1694         else
1695                 bus_register_notifier(&pci_bus_type, &pci_stub_nb);
1696 #endif
1697
1698         return err;
1699 }
1700
1701 static void __exit xen_pcibk_cleanup(void)
1702 {
1703 #ifdef CONFIG_PCI_IOV
1704         bus_unregister_notifier(&pci_bus_type, &pci_stub_nb);
1705 #endif
1706         xen_pcibk_xenbus_unregister();
1707         pcistub_exit();
1708 }
1709
1710 module_init(xen_pcibk_init);
1711 module_exit(xen_pcibk_cleanup);
1712
1713 MODULE_LICENSE("Dual BSD/GPL");
1714 MODULE_ALIAS("xen-backend:pci");