drivers/pci/xen-pcifront.c

   1 /*
   2  * Xen PCI Frontend.
   3  *
   4  *   Author: Ryan Wilson <hap9@epoch.ncsc.mil>
   5  */
   6 #include <linux/module.h>
   7 #include <linux/init.h>
   8 #include <linux/mm.h>
   9 #include <xen/xenbus.h>
  10 #include <xen/events.h>
  11 #include <xen/grant_table.h>
  12 #include <xen/page.h>
  13 #include <linux/spinlock.h>
  14 #include <linux/pci.h>
  15 #include <linux/msi.h>
  16 #include <xen/interface/io/pciif.h>
  17 #include <asm/xen/pci.h>
  18 #include <linux/interrupt.h>
  19 #include <linux/atomic.h>
  20 #include <linux/workqueue.h>
  21 #include <linux/bitops.h>
  22 #include <linux/time.h>
  23
  24 #include <asm/xen/swiotlb-xen.h>
  25 #define INVALID_GRANT_REF (0)
  26 #define INVALID_EVTCHN    (-1)
  27
  28 struct pci_bus_entry {
  29         struct list_head list;
  30         struct pci_bus *bus;
  31 };
  32
  33 #define _PDEVB_op_active                (0)
  34 #define PDEVB_op_active                 (1 << (_PDEVB_op_active))
  35
  36 struct pcifront_device {
  37         struct xenbus_device *xdev;
  38         struct list_head root_buses;
  39
  40         int evtchn;
  41         int gnt_ref;
  42
  43         int irq;
  44
  45         /* Lock this when doing any operations in sh_info */
  46         spinlock_t sh_info_lock;
  47         struct xen_pci_sharedinfo *sh_info;
  48         struct work_struct op_work;
  49         unsigned long flags;
  50
  51 };
  52
  53 struct pcifront_sd {
  54         int domain;
  55         struct pcifront_device *pdev;
  56 };
  57
  58 static inline struct pcifront_device *
  59 pcifront_get_pdev(struct pcifront_sd *sd)
  60 {
  61         return sd->pdev;
  62 }
  63
  64 static inline void pcifront_init_sd(struct pcifront_sd *sd,
  65                                     unsigned int domain, unsigned int bus,
  66                                     struct pcifront_device *pdev)
  67 {
  68         sd->domain = domain;
  69         sd->pdev = pdev;
  70 }
  71
  72 static DEFINE_SPINLOCK(pcifront_dev_lock);
  73 static struct pcifront_device *pcifront_dev;
  74
  75 static int verbose_request;
  76 module_param(verbose_request, int, 0644);
  77
  78 static int errno_to_pcibios_err(int errno)
  79 {
  80         switch (errno) {
  81         case XEN_PCI_ERR_success:
  82                 return PCIBIOS_SUCCESSFUL;
  83
  84         case XEN_PCI_ERR_dev_not_found:
  85                 return PCIBIOS_DEVICE_NOT_FOUND;
  86
  87         case XEN_PCI_ERR_invalid_offset:
  88         case XEN_PCI_ERR_op_failed:
  89                 return PCIBIOS_BAD_REGISTER_NUMBER;
  90
  91         case XEN_PCI_ERR_not_implemented:
  92                 return PCIBIOS_FUNC_NOT_SUPPORTED;
  93
  94         case XEN_PCI_ERR_access_denied:
  95                 return PCIBIOS_SET_FAILED;
  96         }
  97         return errno;
  98 }
  99
 100 static inline void schedule_pcifront_aer_op(struct pcifront_device *pdev)
 101 {
 102         if (test_bit(_XEN_PCIB_active, (unsigned long *)&pdev->sh_info->flags)
 103                 && !test_and_set_bit(_PDEVB_op_active, &pdev->flags)) {
 104                 dev_dbg(&pdev->xdev->dev, "schedule aer frontend job\n");
 105                 schedule_work(&pdev->op_work);
 106         }
 107 }
 108
 109 static int do_pci_op(struct pcifront_device *pdev, struct xen_pci_op *op)
 110 {
 111         int err = 0;
 112         struct xen_pci_op *active_op = &pdev->sh_info->op;
 113         unsigned long irq_flags;
 114         evtchn_port_t port = pdev->evtchn;
 115         unsigned irq = pdev->irq;
 116         s64 ns, ns_timeout;
 117         struct timeval tv;
 118
 119         spin_lock_irqsave(&pdev->sh_info_lock, irq_flags);
 120
 121         memcpy(active_op, op, sizeof(struct xen_pci_op));
 122
 123         /* Go */
 124         wmb();
 125         set_bit(_XEN_PCIF_active, (unsigned long *)&pdev->sh_info->flags);
 126         notify_remote_via_evtchn(port);
 127
 128         /*
 129          * We set a poll timeout of 3 seconds but give up on return after
 130          * 2 seconds. It is better to time out too late rather than too early
 131          * (in the latter case we end up continually re-executing poll() with a
 132          * timeout in the past). 1s difference gives plenty of slack for error.
 133          */
 134         do_gettimeofday(&tv);
 135         ns_timeout = timeval_to_ns(&tv) + 2 * (s64)NSEC_PER_SEC;
 136
 137         xen_clear_irq_pending(irq);
 138
 139         while (test_bit(_XEN_PCIF_active,
 140                         (unsigned long *)&pdev->sh_info->flags)) {
 141                 xen_poll_irq_timeout(irq, jiffies + 3*HZ);
 142                 xen_clear_irq_pending(irq);
 143                 do_gettimeofday(&tv);
 144                 ns = timeval_to_ns(&tv);
 145                 if (ns > ns_timeout) {
 146                         dev_err(&pdev->xdev->dev,
 147                                 "pciback not responding!!!\n");
 148                         clear_bit(_XEN_PCIF_active,
 149                                   (unsigned long *)&pdev->sh_info->flags);
 150                         err = XEN_PCI_ERR_dev_not_found;
 151                         goto out;
 152                 }
 153         }
 154
 155         /*
 156         * We might lose backend service request since we
 157         * reuse same evtchn with pci_conf backend response. So re-schedule
 158         * aer pcifront service.
 159         */
 160         if (test_bit(_XEN_PCIB_active,
 161                         (unsigned long *)&pdev->sh_info->flags)) {
 162                 dev_err(&pdev->xdev->dev,
 163                         "schedule aer pcifront service\n");
 164                 schedule_pcifront_aer_op(pdev);
 165         }
 166
 167         memcpy(op, active_op, sizeof(struct xen_pci_op));
 168
 169         err = op->err;
 170 out:
 171         spin_unlock_irqrestore(&pdev->sh_info_lock, irq_flags);
 172         return err;
 173 }
 174
 175 /* Access to this function is spinlocked in drivers/pci/access.c */
 176 static int pcifront_bus_read(struct pci_bus *bus, unsigned int devfn,
 177                              int where, int size, u32 *val)
 178 {
 179         int err = 0;
 180         struct xen_pci_op op = {
 181                 .cmd    = XEN_PCI_OP_conf_read,
 182                 .domain = pci_domain_nr(bus),
 183                 .bus    = bus->number,
 184                 .devfn  = devfn,
 185                 .offset = where,
 186                 .size   = size,
 187         };
 188         struct pcifront_sd *sd = bus->sysdata;
 189         struct pcifront_device *pdev = pcifront_get_pdev(sd);
 190
 191         if (verbose_request)
 192                 dev_info(&pdev->xdev->dev,
 193                          "read dev=%04x:%02x:%02x.%d - offset %x size %d\n",
 194                          pci_domain_nr(bus), bus->number, PCI_SLOT(devfn),
 195                          PCI_FUNC(devfn), where, size);
 196
 197         err = do_pci_op(pdev, &op);
 198
 199         if (likely(!err)) {
 200                 if (verbose_request)
 201                         dev_info(&pdev->xdev->dev, "read got back value %x\n",
 202                                  op.value);
 203
 204                 *val = op.value;
 205         } else if (err == -ENODEV) {
 206                 /* No device here, pretend that it just returned 0 */
 207                 err = 0;
 208                 *val = 0;
 209         }
 210
 211         return errno_to_pcibios_err(err);
 212 }
 213
 214 /* Access to this function is spinlocked in drivers/pci/access.c */
 215 static int pcifront_bus_write(struct pci_bus *bus, unsigned int devfn,
 216                               int where, int size, u32 val)
 217 {
 218         struct xen_pci_op op = {
 219                 .cmd    = XEN_PCI_OP_conf_write,
 220                 .domain = pci_domain_nr(bus),
 221                 .bus    = bus->number,
 222                 .devfn  = devfn,
 223                 .offset = where,
 224                 .size   = size,
 225                 .value  = val,
 226         };
 227         struct pcifront_sd *sd = bus->sysdata;
 228         struct pcifront_device *pdev = pcifront_get_pdev(sd);
 229
 230         if (verbose_request)
 231                 dev_info(&pdev->xdev->dev,
 232                          "write dev=%04x:%02x:%02x.%d - "
 233                          "offset %x size %d val %x\n",
 234                          pci_domain_nr(bus), bus->number,
 235                          PCI_SLOT(devfn), PCI_FUNC(devfn), where, size, val);
 236
 237         return errno_to_pcibios_err(do_pci_op(pdev, &op));
 238 }
 239
 240 static struct pci_ops pcifront_bus_ops = {
 241         .read = pcifront_bus_read,
 242         .write = pcifront_bus_write,
 243 };
 244
 245 #ifdef CONFIG_PCI_MSI
 246 static int pci_frontend_enable_msix(struct pci_dev *dev,
 247                                     int vector[], int nvec)
 248 {
 249         int err;
 250         int i;
 251         struct xen_pci_op op = {
 252                 .cmd    = XEN_PCI_OP_enable_msix,
 253                 .domain = pci_domain_nr(dev->bus),
 254                 .bus = dev->bus->number,
 255                 .devfn = dev->devfn,
 256                 .value = nvec,
 257         };
 258         struct pcifront_sd *sd = dev->bus->sysdata;
 259         struct pcifront_device *pdev = pcifront_get_pdev(sd);
 260         struct msi_desc *entry;
 261
 262         if (nvec > SH_INFO_MAX_VEC) {
 263                 dev_err(&dev->dev, "too much vector for pci frontend: %x."
 264                                    " Increase SH_INFO_MAX_VEC.\n", nvec);
 265                 return -EINVAL;
 266         }
 267
 268         i = 0;
 269         list_for_each_entry(entry, &dev->msi_list, list) {
 270                 op.msix_entries[i].entry = entry->msi_attrib.entry_nr;
 271                 /* Vector is useless at this point. */
 272                 op.msix_entries[i].vector = -1;
 273                 i++;
 274         }
 275
 276         err = do_pci_op(pdev, &op);
 277
 278         if (likely(!err)) {
 279                 if (likely(!op.value)) {
 280                         /* we get the result */
 281                         for (i = 0; i < nvec; i++) {
 282                                 if (op.msix_entries[i].vector <= 0) {
 283                                         dev_warn(&dev->dev, "MSI-X entry %d is invalid: %d!\n",
 284                                                 i, op.msix_entries[i].vector);
 285                                         err = -EINVAL;
 286                                         vector[i] = -1;
 287                                         continue;
 288                                 }
 289                                 vector[i] = op.msix_entries[i].vector;
 290                         }
 291                 } else {
 292                         printk(KERN_DEBUG "enable msix get value %x\n",
 293                                 op.value);
 294                         err = op.value;
 295                 }
 296         } else {
 297                 dev_err(&dev->dev, "enable msix get err %x\n", err);
 298         }
 299         return err;
 300 }
 301
 302 static void pci_frontend_disable_msix(struct pci_dev *dev)
 303 {
 304         int err;
 305         struct xen_pci_op op = {
 306                 .cmd    = XEN_PCI_OP_disable_msix,
 307                 .domain = pci_domain_nr(dev->bus),
 308                 .bus = dev->bus->number,
 309                 .devfn = dev->devfn,
 310         };
 311         struct pcifront_sd *sd = dev->bus->sysdata;
 312         struct pcifront_device *pdev = pcifront_get_pdev(sd);
 313
 314         err = do_pci_op(pdev, &op);
 315
 316         /* What should do for error ? */
 317         if (err)
 318                 dev_err(&dev->dev, "pci_disable_msix get err %x\n", err);
 319 }
 320
 321 static int pci_frontend_enable_msi(struct pci_dev *dev, int vector[])
 322 {
 323         int err;
 324         struct xen_pci_op op = {
 325                 .cmd    = XEN_PCI_OP_enable_msi,
 326                 .domain = pci_domain_nr(dev->bus),
 327                 .bus = dev->bus->number,
 328                 .devfn = dev->devfn,
 329         };
 330         struct pcifront_sd *sd = dev->bus->sysdata;
 331         struct pcifront_device *pdev = pcifront_get_pdev(sd);
 332
 333         err = do_pci_op(pdev, &op);
 334         if (likely(!err)) {
 335                 vector[0] = op.value;
 336                 if (op.value <= 0) {
 337                         dev_warn(&dev->dev, "MSI entry is invalid: %d!\n",
 338                                 op.value);
 339                         err = -EINVAL;
 340                         vector[0] = -1;
 341                 }
 342         } else {
 343                 dev_err(&dev->dev, "pci frontend enable msi failed for dev "
 344                                     "%x:%x\n", op.bus, op.devfn);
 345                 err = -EINVAL;
 346         }
 347         return err;
 348 }
 349
 350 static void pci_frontend_disable_msi(struct pci_dev *dev)
 351 {
 352         int err;
 353         struct xen_pci_op op = {
 354                 .cmd    = XEN_PCI_OP_disable_msi,
 355                 .domain = pci_domain_nr(dev->bus),
 356                 .bus = dev->bus->number,
 357                 .devfn = dev->devfn,
 358         };
 359         struct pcifront_sd *sd = dev->bus->sysdata;
 360         struct pcifront_device *pdev = pcifront_get_pdev(sd);
 361
 362         err = do_pci_op(pdev, &op);
 363         if (err == XEN_PCI_ERR_dev_not_found) {
 364                 /* XXX No response from backend, what shall we do? */
 365                 printk(KERN_DEBUG "get no response from backend for disable MSI\n");
 366                 return;
 367         }
 368         if (err)
 369                 /* how can pciback notify us fail? */
 370                 printk(KERN_DEBUG "get fake response frombackend\n");
 371 }
 372
 373 static struct xen_pci_frontend_ops pci_frontend_ops = {
 374         .enable_msi = pci_frontend_enable_msi,
 375         .disable_msi = pci_frontend_disable_msi,
 376         .enable_msix = pci_frontend_enable_msix,
 377         .disable_msix = pci_frontend_disable_msix,
 378 };
 379
 380 static void pci_frontend_registrar(int enable)
 381 {
 382         if (enable)
 383                 xen_pci_frontend = &pci_frontend_ops;
 384         else
 385                 xen_pci_frontend = NULL;
 386 };
 387 #else
 388 static inline void pci_frontend_registrar(int enable) { };
 389 #endif /* CONFIG_PCI_MSI */
 390
 391 /* Claim resources for the PCI frontend as-is, backend won't allow changes */
 392 static int pcifront_claim_resource(struct pci_dev *dev, void *data)
 393 {
 394         struct pcifront_device *pdev = data;
 395         int i;
 396         struct resource *r;
 397
 398         for (i = 0; i < PCI_NUM_RESOURCES; i++) {
 399                 r = &dev->resource[i];
 400
 401                 if (!r->parent && r->start && r->flags) {
 402                         dev_info(&pdev->xdev->dev, "claiming resource %s/%d\n",
 403                                 pci_name(dev), i);
 404                         if (pci_claim_resource(dev, i)) {
 405                                 dev_err(&pdev->xdev->dev, "Could not claim resource %s/%d! "
 406                                         "Device offline. Try using e820_host=1 in the guest config.\n",
 407                                         pci_name(dev), i);
 408                         }
 409                 }
 410         }
 411
 412         return 0;
 413 }
 414
 415 static int pcifront_scan_bus(struct pcifront_device *pdev,
 416                                 unsigned int domain, unsigned int bus,
 417                                 struct pci_bus *b)
 418 {
 419         struct pci_dev *d;
 420         unsigned int devfn;
 421
 422         /* Scan the bus for functions and add.
 423          * We omit handling of PCI bridge attachment because pciback prevents
 424          * bridges from being exported.
 425          */
 426         for (devfn = 0; devfn < 0x100; devfn++) {
 427                 d = pci_get_slot(b, devfn);
 428                 if (d) {
 429                         /* Device is already known. */
 430                         pci_dev_put(d);
 431                         continue;
 432                 }
 433
 434                 d = pci_scan_single_device(b, devfn);
 435                 if (d)
 436                         dev_info(&pdev->xdev->dev, "New device on "
 437                                  "%04x:%02x:%02x.%d found.\n", domain, bus,
 438                                  PCI_SLOT(devfn), PCI_FUNC(devfn));
 439         }
 440
 441         return 0;
 442 }
 443
 444 static int pcifront_scan_root(struct pcifront_device *pdev,
 445                                  unsigned int domain, unsigned int bus)
 446 {
 447         struct pci_bus *b;
 448         struct pcifront_sd *sd = NULL;
 449         struct pci_bus_entry *bus_entry = NULL;
 450         int err = 0;
 451
 452 #ifndef CONFIG_PCI_DOMAINS
 453         if (domain != 0) {
 454                 dev_err(&pdev->xdev->dev,
 455                         "PCI Root in non-zero PCI Domain! domain=%d\n", domain);
 456                 dev_err(&pdev->xdev->dev,
 457                         "Please compile with CONFIG_PCI_DOMAINS\n");
 458                 err = -EINVAL;
 459                 goto err_out;
 460         }
 461 #endif
 462
 463         dev_info(&pdev->xdev->dev, "Creating PCI Frontend Bus %04x:%02x\n",
 464                  domain, bus);
 465
 466         bus_entry = kmalloc(sizeof(*bus_entry), GFP_KERNEL);
 467         sd = kmalloc(sizeof(*sd), GFP_KERNEL);
 468         if (!bus_entry || !sd) {
 469                 err = -ENOMEM;
 470                 goto err_out;
 471         }
 472         pcifront_init_sd(sd, domain, bus, pdev);
 473
 474         b = pci_scan_bus_parented(&pdev->xdev->dev, bus,
 475                                   &pcifront_bus_ops, sd);
 476         if (!b) {
 477                 dev_err(&pdev->xdev->dev,
 478                         "Error creating PCI Frontend Bus!\n");
 479                 err = -ENOMEM;
 480                 goto err_out;
 481         }
 482
 483         bus_entry->bus = b;
 484
 485         list_add(&bus_entry->list, &pdev->root_buses);
 486
 487         /* pci_scan_bus_parented skips devices which do not have a have
 488         * devfn==0. The pcifront_scan_bus enumerates all devfn. */
 489         err = pcifront_scan_bus(pdev, domain, bus, b);
 490
 491         /* Claim resources before going "live" with our devices */
 492         pci_walk_bus(b, pcifront_claim_resource, pdev);
 493
 494         /* Create SysFS and notify udev of the devices. Aka: "going live" */
 495         pci_bus_add_devices(b);
 496
 497         return err;
 498
 499 err_out:
 500         kfree(bus_entry);
 501         kfree(sd);
 502
 503         return err;
 504 }
 505
 506 static int pcifront_rescan_root(struct pcifront_device *pdev,
 507                                    unsigned int domain, unsigned int bus)
 508 {
 509         int err;
 510         struct pci_bus *b;
 511
 512 #ifndef CONFIG_PCI_DOMAINS
 513         if (domain != 0) {
 514                 dev_err(&pdev->xdev->dev,
 515                         "PCI Root in non-zero PCI Domain! domain=%d\n", domain);
 516                 dev_err(&pdev->xdev->dev,
 517                         "Please compile with CONFIG_PCI_DOMAINS\n");
 518                 return -EINVAL;
 519         }
 520 #endif
 521
 522         dev_info(&pdev->xdev->dev, "Rescanning PCI Frontend Bus %04x:%02x\n",
 523                  domain, bus);
 524
 525         b = pci_find_bus(domain, bus);
 526         if (!b)
 527                 /* If the bus is unknown, create it. */
 528                 return pcifront_scan_root(pdev, domain, bus);
 529
 530         err = pcifront_scan_bus(pdev, domain, bus, b);
 531
 532         /* Claim resources before going "live" with our devices */
 533         pci_walk_bus(b, pcifront_claim_resource, pdev);
 534
 535         /* Create SysFS and notify udev of the devices. Aka: "going live" */
 536         pci_bus_add_devices(b);
 537
 538         return err;
 539 }
 540
 541 static void free_root_bus_devs(struct pci_bus *bus)
 542 {
 543         struct pci_dev *dev;
 544
 545         while (!list_empty(&bus->devices)) {
 546                 dev = container_of(bus->devices.next, struct pci_dev,
 547                                    bus_list);
 548                 dev_dbg(&dev->dev, "removing device\n");
 549                 pci_stop_and_remove_bus_device(dev);
 550         }
 551 }
 552
 553 static void pcifront_free_roots(struct pcifront_device *pdev)
 554 {
 555         struct pci_bus_entry *bus_entry, *t;
 556
 557         dev_dbg(&pdev->xdev->dev, "cleaning up root buses\n");
 558
 559         list_for_each_entry_safe(bus_entry, t, &pdev->root_buses, list) {
 560                 list_del(&bus_entry->list);
 561
 562                 free_root_bus_devs(bus_entry->bus);
 563
 564                 kfree(bus_entry->bus->sysdata);
 565
 566                 device_unregister(bus_entry->bus->bridge);
 567                 pci_remove_bus(bus_entry->bus);
 568
 569                 kfree(bus_entry);
 570         }
 571 }
 572
 573 static pci_ers_result_t pcifront_common_process(int cmd,
 574                                                 struct pcifront_device *pdev,
 575                                                 pci_channel_state_t state)
 576 {
 577         pci_ers_result_t result;
 578         struct pci_driver *pdrv;
 579         int bus = pdev->sh_info->aer_op.bus;
 580         int devfn = pdev->sh_info->aer_op.devfn;
 581         struct pci_dev *pcidev;
 582         int flag = 0;
 583
 584         dev_dbg(&pdev->xdev->dev,
 585                 "pcifront AER process: cmd %x (bus:%x, devfn%x)",
 586                 cmd, bus, devfn);
 587         result = PCI_ERS_RESULT_NONE;
 588
 589         pcidev = pci_get_bus_and_slot(bus, devfn);
 590         if (!pcidev || !pcidev->driver) {
 591                 dev_err(&pdev->xdev->dev, "device or AER driver is NULL\n");
 592                 if (pcidev)
 593                         pci_dev_put(pcidev);
 594                 return result;
 595         }
 596         pdrv = pcidev->driver;
 597
 598         if (pdrv) {
 599                 if (pdrv->err_handler && pdrv->err_handler->error_detected) {
 600                         dev_dbg(&pcidev->dev,
 601                                 "trying to call AER service\n");
 602                         if (pcidev) {
 603                                 flag = 1;
 604                                 switch (cmd) {
 605                                 case XEN_PCI_OP_aer_detected:
 606                                         result = pdrv->err_handler->
 607                                                  error_detected(pcidev, state);
 608                                         break;
 609                                 case XEN_PCI_OP_aer_mmio:
 610                                         result = pdrv->err_handler->
 611                                                  mmio_enabled(pcidev);
 612                                         break;
 613                                 case XEN_PCI_OP_aer_slotreset:
 614                                         result = pdrv->err_handler->
 615                                                  slot_reset(pcidev);
 616                                         break;
 617                                 case XEN_PCI_OP_aer_resume:
 618                                         pdrv->err_handler->resume(pcidev);
 619                                         break;
 620                                 default:
 621                                         dev_err(&pdev->xdev->dev,
 622                                                 "bad request in aer recovery "
 623                                                 "operation!\n");
 624
 625                                 }
 626                         }
 627                 }
 628         }
 629         if (!flag)
 630                 result = PCI_ERS_RESULT_NONE;
 631
 632         return result;
 633 }
 634
 635
 636 static void pcifront_do_aer(struct work_struct *data)
 637 {
 638         struct pcifront_device *pdev =
 639                 container_of(data, struct pcifront_device, op_work);
 640         int cmd = pdev->sh_info->aer_op.cmd;
 641         pci_channel_state_t state =
 642                 (pci_channel_state_t)pdev->sh_info->aer_op.err;
 643
 644         /*If a pci_conf op is in progress,
 645                 we have to wait until it is done before service aer op*/
 646         dev_dbg(&pdev->xdev->dev,
 647                 "pcifront service aer bus %x devfn %x\n",
 648                 pdev->sh_info->aer_op.bus, pdev->sh_info->aer_op.devfn);
 649
 650         pdev->sh_info->aer_op.err = pcifront_common_process(cmd, pdev, state);
 651
 652         /* Post the operation to the guest. */
 653         wmb();
 654         clear_bit(_XEN_PCIB_active, (unsigned long *)&pdev->sh_info->flags);
 655         notify_remote_via_evtchn(pdev->evtchn);
 656
 657         /*in case of we lost an aer request in four lines time_window*/
 658         smp_mb__before_clear_bit();
 659         clear_bit(_PDEVB_op_active, &pdev->flags);
 660         smp_mb__after_clear_bit();
 661
 662         schedule_pcifront_aer_op(pdev);
 663
 664 }
 665
 666 static irqreturn_t pcifront_handler_aer(int irq, void *dev)
 667 {
 668         struct pcifront_device *pdev = dev;
 669         schedule_pcifront_aer_op(pdev);
 670         return IRQ_HANDLED;
 671 }
 672 static int pcifront_connect_and_init_dma(struct pcifront_device *pdev)
 673 {
 674         int err = 0;
 675
 676         spin_lock(&pcifront_dev_lock);
 677
 678         if (!pcifront_dev) {
 679                 dev_info(&pdev->xdev->dev, "Installing PCI frontend\n");
 680                 pcifront_dev = pdev;
 681         } else {
 682                 dev_err(&pdev->xdev->dev, "PCI frontend already installed!\n");
 683                 err = -EEXIST;
 684         }
 685         spin_unlock(&pcifront_dev_lock);
 686
 687         if (!err && !swiotlb_nr_tbl()) {
 688                 err = pci_xen_swiotlb_init_late();
 689                 if (err)
 690                         dev_err(&pdev->xdev->dev, "Could not setup SWIOTLB!\n");
 691         }
 692         return err;
 693 }
 694
 695 static void pcifront_disconnect(struct pcifront_device *pdev)
 696 {
 697         spin_lock(&pcifront_dev_lock);
 698
 699         if (pdev == pcifront_dev) {
 700                 dev_info(&pdev->xdev->dev,
 701                          "Disconnecting PCI Frontend Buses\n");
 702                 pcifront_dev = NULL;
 703         }
 704
 705         spin_unlock(&pcifront_dev_lock);
 706 }
 707 static struct pcifront_device *alloc_pdev(struct xenbus_device *xdev)
 708 {
 709         struct pcifront_device *pdev;
 710
 711         pdev = kzalloc(sizeof(struct pcifront_device), GFP_KERNEL);
 712         if (pdev == NULL)
 713                 goto out;
 714
 715         pdev->sh_info =
 716             (struct xen_pci_sharedinfo *)__get_free_page(GFP_KERNEL);
 717         if (pdev->sh_info == NULL) {
 718                 kfree(pdev);
 719                 pdev = NULL;
 720                 goto out;
 721         }
 722         pdev->sh_info->flags = 0;
 723
 724         /*Flag for registering PV AER handler*/
 725         set_bit(_XEN_PCIB_AERHANDLER, (void *)&pdev->sh_info->flags);
 726
 727         dev_set_drvdata(&xdev->dev, pdev);
 728         pdev->xdev = xdev;
 729
 730         INIT_LIST_HEAD(&pdev->root_buses);
 731
 732         spin_lock_init(&pdev->sh_info_lock);
 733
 734         pdev->evtchn = INVALID_EVTCHN;
 735         pdev->gnt_ref = INVALID_GRANT_REF;
 736         pdev->irq = -1;
 737
 738         INIT_WORK(&pdev->op_work, pcifront_do_aer);
 739
 740         dev_dbg(&xdev->dev, "Allocated pdev @ 0x%p pdev->sh_info @ 0x%p\n",
 741                 pdev, pdev->sh_info);
 742 out:
 743         return pdev;
 744 }
 745
 746 static void free_pdev(struct pcifront_device *pdev)
 747 {
 748         dev_dbg(&pdev->xdev->dev, "freeing pdev @ 0x%p\n", pdev);
 749
 750         pcifront_free_roots(pdev);
 751
 752         cancel_work_sync(&pdev->op_work);
 753
 754         if (pdev->irq >= 0)
 755                 unbind_from_irqhandler(pdev->irq, pdev);
 756
 757         if (pdev->evtchn != INVALID_EVTCHN)
 758                 xenbus_free_evtchn(pdev->xdev, pdev->evtchn);
 759
 760         if (pdev->gnt_ref != INVALID_GRANT_REF)
 761                 gnttab_end_foreign_access(pdev->gnt_ref, 0 /* r/w page */,
 762                                           (unsigned long)pdev->sh_info);
 763         else
 764                 free_page((unsigned long)pdev->sh_info);
 765
 766         dev_set_drvdata(&pdev->xdev->dev, NULL);
 767
 768         kfree(pdev);
 769 }
 770
 771 static int pcifront_publish_info(struct pcifront_device *pdev)
 772 {
 773         int err = 0;
 774         struct xenbus_transaction trans;
 775
 776         err = xenbus_grant_ring(pdev->xdev, virt_to_mfn(pdev->sh_info));
 777         if (err < 0)
 778                 goto out;
 779
 780         pdev->gnt_ref = err;
 781
 782         err = xenbus_alloc_evtchn(pdev->xdev, &pdev->evtchn);
 783         if (err)
 784                 goto out;
 785
 786         err = bind_evtchn_to_irqhandler(pdev->evtchn, pcifront_handler_aer,
 787                 0, "pcifront", pdev);
 788
 789         if (err < 0)
 790                 return err;
 791
 792         pdev->irq = err;
 793
 794 do_publish:
 795         err = xenbus_transaction_start(&trans);
 796         if (err) {
 797                 xenbus_dev_fatal(pdev->xdev, err,
 798                                  "Error writing configuration for backend "
 799                                  "(start transaction)");
 800                 goto out;
 801         }
 802
 803         err = xenbus_printf(trans, pdev->xdev->nodename,
 804                             "pci-op-ref", "%u", pdev->gnt_ref);
 805         if (!err)
 806                 err = xenbus_printf(trans, pdev->xdev->nodename,
 807                                     "event-channel", "%u", pdev->evtchn);
 808         if (!err)
 809                 err = xenbus_printf(trans, pdev->xdev->nodename,
 810                                     "magic", XEN_PCI_MAGIC);
 811
 812         if (err) {
 813                 xenbus_transaction_end(trans, 1);
 814                 xenbus_dev_fatal(pdev->xdev, err,
 815                                  "Error writing configuration for backend");
 816                 goto out;
 817         } else {
 818                 err = xenbus_transaction_end(trans, 0);
 819                 if (err == -EAGAIN)
 820                         goto do_publish;
 821                 else if (err) {
 822                         xenbus_dev_fatal(pdev->xdev, err,
 823                                          "Error completing transaction "
 824                                          "for backend");
 825                         goto out;
 826                 }
 827         }
 828
 829         xenbus_switch_state(pdev->xdev, XenbusStateInitialised);
 830
 831         dev_dbg(&pdev->xdev->dev, "publishing successful!\n");
 832
 833 out:
 834         return err;
 835 }
 836
 837 static int pcifront_try_connect(struct pcifront_device *pdev)
 838 {
 839         int err = -EFAULT;
 840         int i, num_roots, len;
 841         char str[64];
 842         unsigned int domain, bus;
 843
 844
 845         /* Only connect once */
 846         if (xenbus_read_driver_state(pdev->xdev->nodename) !=
 847             XenbusStateInitialised)
 848                 goto out;
 849
 850         err = pcifront_connect_and_init_dma(pdev);
 851         if (err) {
 852                 xenbus_dev_fatal(pdev->xdev, err,
 853                                  "Error setting up PCI Frontend");
 854                 goto out;
 855         }
 856
 857         err = xenbus_scanf(XBT_NIL, pdev->xdev->otherend,
 858                            "root_num", "%d", &num_roots);
 859         if (err == -ENOENT) {
 860                 xenbus_dev_error(pdev->xdev, err,
 861                                  "No PCI Roots found, trying 0000:00");
 862                 err = pcifront_scan_root(pdev, 0, 0);
 863                 num_roots = 0;
 864         } else if (err != 1) {
 865                 if (err == 0)
 866                         err = -EINVAL;
 867                 xenbus_dev_fatal(pdev->xdev, err,
 868                                  "Error reading number of PCI roots");
 869                 goto out;
 870         }
 871
 872         for (i = 0; i < num_roots; i++) {
 873                 len = snprintf(str, sizeof(str), "root-%d", i);
 874                 if (unlikely(len >= (sizeof(str) - 1))) {
 875                         err = -ENOMEM;
 876                         goto out;
 877                 }
 878
 879                 err = xenbus_scanf(XBT_NIL, pdev->xdev->otherend, str,
 880                                    "%x:%x", &domain, &bus);
 881                 if (err != 2) {
 882                         if (err >= 0)
 883                                 err = -EINVAL;
 884                         xenbus_dev_fatal(pdev->xdev, err,
 885                                          "Error reading PCI root %d", i);
 886                         goto out;
 887                 }
 888
 889                 err = pcifront_scan_root(pdev, domain, bus);
 890                 if (err) {
 891                         xenbus_dev_fatal(pdev->xdev, err,
 892                                          "Error scanning PCI root %04x:%02x",
 893                                          domain, bus);
 894                         goto out;
 895                 }
 896         }
 897
 898         err = xenbus_switch_state(pdev->xdev, XenbusStateConnected);
 899
 900 out:
 901         return err;
 902 }
 903
 904 static int pcifront_try_disconnect(struct pcifront_device *pdev)
 905 {
 906         int err = 0;
 907         enum xenbus_state prev_state;
 908
 909
 910         prev_state = xenbus_read_driver_state(pdev->xdev->nodename);
 911
 912         if (prev_state >= XenbusStateClosing)
 913                 goto out;
 914
 915         if (prev_state == XenbusStateConnected) {
 916                 pcifront_free_roots(pdev);
 917                 pcifront_disconnect(pdev);
 918         }
 919
 920         err = xenbus_switch_state(pdev->xdev, XenbusStateClosed);
 921
 922 out:
 923
 924         return err;
 925 }
 926
 927 static int pcifront_attach_devices(struct pcifront_device *pdev)
 928 {
 929         int err = -EFAULT;
 930         int i, num_roots, len;
 931         unsigned int domain, bus;
 932         char str[64];
 933
 934         if (xenbus_read_driver_state(pdev->xdev->nodename) !=
 935             XenbusStateReconfiguring)
 936                 goto out;
 937
 938         err = xenbus_scanf(XBT_NIL, pdev->xdev->otherend,
 939                            "root_num", "%d", &num_roots);
 940         if (err == -ENOENT) {
 941                 xenbus_dev_error(pdev->xdev, err,
 942                                  "No PCI Roots found, trying 0000:00");
 943                 err = pcifront_rescan_root(pdev, 0, 0);
 944                 num_roots = 0;
 945         } else if (err != 1) {
 946                 if (err == 0)
 947                         err = -EINVAL;
 948                 xenbus_dev_fatal(pdev->xdev, err,
 949                                  "Error reading number of PCI roots");
 950                 goto out;
 951         }
 952
 953         for (i = 0; i < num_roots; i++) {
 954                 len = snprintf(str, sizeof(str), "root-%d", i);
 955                 if (unlikely(len >= (sizeof(str) - 1))) {
 956                         err = -ENOMEM;
 957                         goto out;
 958                 }
 959
 960                 err = xenbus_scanf(XBT_NIL, pdev->xdev->otherend, str,
 961                                    "%x:%x", &domain, &bus);
 962                 if (err != 2) {
 963                         if (err >= 0)
 964                                 err = -EINVAL;
 965                         xenbus_dev_fatal(pdev->xdev, err,
 966                                          "Error reading PCI root %d", i);
 967                         goto out;
 968                 }
 969
 970                 err = pcifront_rescan_root(pdev, domain, bus);
 971                 if (err) {
 972                         xenbus_dev_fatal(pdev->xdev, err,
 973                                          "Error scanning PCI root %04x:%02x",
 974                                          domain, bus);
 975                         goto out;
 976                 }
 977         }
 978
 979         xenbus_switch_state(pdev->xdev, XenbusStateConnected);
 980
 981 out:
 982         return err;
 983 }
 984
 985 static int pcifront_detach_devices(struct pcifront_device *pdev)
 986 {
 987         int err = 0;
 988         int i, num_devs;
 989         unsigned int domain, bus, slot, func;
 990         struct pci_dev *pci_dev;
 991         char str[64];
 992
 993         if (xenbus_read_driver_state(pdev->xdev->nodename) !=
 994             XenbusStateConnected)
 995                 goto out;
 996
 997         err = xenbus_scanf(XBT_NIL, pdev->xdev->otherend, "num_devs", "%d",
 998                            &num_devs);
 999         if (err != 1) {
1000                 if (err >= 0)
1001                         err = -EINVAL;
1002                 xenbus_dev_fatal(pdev->xdev, err,
1003                                  "Error reading number of PCI devices");
1004                 goto out;
1005         }
1006
1007         /* Find devices being detached and remove them. */
1008         for (i = 0; i < num_devs; i++) {
1009                 int l, state;
1010                 l = snprintf(str, sizeof(str), "state-%d", i);
1011                 if (unlikely(l >= (sizeof(str) - 1))) {
1012                         err = -ENOMEM;
1013                         goto out;
1014                 }
1015                 err = xenbus_scanf(XBT_NIL, pdev->xdev->otherend, str, "%d",
1016                                    &state);
1017                 if (err != 1)
1018                         state = XenbusStateUnknown;
1019
1020                 if (state != XenbusStateClosing)
1021                         continue;
1022
1023                 /* Remove device. */
1024                 l = snprintf(str, sizeof(str), "vdev-%d", i);
1025                 if (unlikely(l >= (sizeof(str) - 1))) {
1026                         err = -ENOMEM;
1027                         goto out;
1028                 }
1029                 err = xenbus_scanf(XBT_NIL, pdev->xdev->otherend, str,
1030                                    "%x:%x:%x.%x", &domain, &bus, &slot, &func);
1031                 if (err != 4) {
1032                         if (err >= 0)
1033                                 err = -EINVAL;
1034                         xenbus_dev_fatal(pdev->xdev, err,
1035                                          "Error reading PCI device %d", i);
1036                         goto out;
1037                 }
1038
1039                 pci_dev = pci_get_domain_bus_and_slot(domain, bus,
1040                                 PCI_DEVFN(slot, func));
1041                 if (!pci_dev) {
1042                         dev_dbg(&pdev->xdev->dev,
1043                                 "Cannot get PCI device %04x:%02x:%02x.%d\n",
1044                                 domain, bus, slot, func);
1045                         continue;
1046                 }
1047                 pci_stop_and_remove_bus_device(pci_dev);
1048                 pci_dev_put(pci_dev);
1049
1050                 dev_dbg(&pdev->xdev->dev,
1051                         "PCI device %04x:%02x:%02x.%d removed.\n",
1052                         domain, bus, slot, func);
1053         }
1054
1055         err = xenbus_switch_state(pdev->xdev, XenbusStateReconfiguring);
1056
1057 out:
1058         return err;
1059 }
1060
1061 static void __init_refok pcifront_backend_changed(struct xenbus_device *xdev,
1062                                                   enum xenbus_state be_state)
1063 {
1064         struct pcifront_device *pdev = dev_get_drvdata(&xdev->dev);
1065
1066         switch (be_state) {
1067         case XenbusStateUnknown:
1068         case XenbusStateInitialising:
1069         case XenbusStateInitWait:
1070         case XenbusStateInitialised:
1071                 break;
1072
1073         case XenbusStateConnected:
1074                 pcifront_try_connect(pdev);
1075                 break;
1076
1077         case XenbusStateClosed:
1078                 if (xdev->state == XenbusStateClosed)
1079                         break;
1080                 /* Missed the backend's CLOSING state -- fallthrough */
1081         case XenbusStateClosing:
1082                 dev_warn(&xdev->dev, "backend going away!\n");
1083                 pcifront_try_disconnect(pdev);
1084                 break;
1085
1086         case XenbusStateReconfiguring:
1087                 pcifront_detach_devices(pdev);
1088                 break;
1089
1090         case XenbusStateReconfigured:
1091                 pcifront_attach_devices(pdev);
1092                 break;
1093         }
1094 }
1095
1096 static int pcifront_xenbus_probe(struct xenbus_device *xdev,
1097                                  const struct xenbus_device_id *id)
1098 {
1099         int err = 0;
1100         struct pcifront_device *pdev = alloc_pdev(xdev);
1101
1102         if (pdev == NULL) {
1103                 err = -ENOMEM;
1104                 xenbus_dev_fatal(xdev, err,
1105                                  "Error allocating pcifront_device struct");
1106                 goto out;
1107         }
1108
1109         err = pcifront_publish_info(pdev);
1110         if (err)
1111                 free_pdev(pdev);
1112
1113 out:
1114         return err;
1115 }
1116
1117 static int pcifront_xenbus_remove(struct xenbus_device *xdev)
1118 {
1119         struct pcifront_device *pdev = dev_get_drvdata(&xdev->dev);
1120         if (pdev)
1121                 free_pdev(pdev);
1122
1123         return 0;
1124 }
1125
1126 static const struct xenbus_device_id xenpci_ids[] = {
1127         {"pci"},
1128         {""},
1129 };
1130
1131 static DEFINE_XENBUS_DRIVER(xenpci, "pcifront",
1132         .probe                  = pcifront_xenbus_probe,
1133         .remove                 = pcifront_xenbus_remove,
1134         .otherend_changed       = pcifront_backend_changed,
1135 );
1136
1137 static int __init pcifront_init(void)
1138 {
1139         if (!xen_pv_domain() || xen_initial_domain())
1140                 return -ENODEV;
1141
1142         pci_frontend_registrar(1 /* enable */);
1143
1144         return xenbus_register_frontend(&xenpci_driver);
1145 }
1146
1147 static void __exit pcifront_cleanup(void)
1148 {
1149         xenbus_unregister_driver(&xenpci_driver);
1150         pci_frontend_registrar(0 /* disable */);
1151 }
1152 module_init(pcifront_init);
1153 module_exit(pcifront_cleanup);
1154
1155 MODULE_DESCRIPTION("Xen PCI passthrough frontend.");
1156 MODULE_LICENSE("GPL");
1157 MODULE_ALIAS("xen:pci");