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 #define INVALID_GRANT_REF (0)
  25 #define INVALID_EVTCHN    (-1)
  26
  27 struct pci_bus_entry {
  28         struct list_head list;
  29         struct pci_bus *bus;
  30 };
  31
  32 #define _PDEVB_op_active                (0)
  33 #define PDEVB_op_active                 (1 << (_PDEVB_op_active))
  34
  35 struct pcifront_device {
  36         struct xenbus_device *xdev;
  37         struct list_head root_buses;
  38
  39         int evtchn;
  40         int gnt_ref;
  41
  42         int irq;
  43
  44         /* Lock this when doing any operations in sh_info */
  45         spinlock_t sh_info_lock;
  46         struct xen_pci_sharedinfo *sh_info;
  47         struct work_struct op_work;
  48         unsigned long flags;
  49
  50 };
  51
  52 struct pcifront_sd {
  53         struct pci_sysdata sd;
  54         struct pcifront_device *pdev;
  55 };
  56
  57 static inline struct pcifront_device *
  58 pcifront_get_pdev(struct pcifront_sd *sd)
  59 {
  60         return sd->pdev;
  61 }
  62
  63 static inline void pcifront_init_sd(struct pcifront_sd *sd,
  64                                     unsigned int domain, unsigned int bus,
  65                                     struct pcifront_device *pdev)
  66 {
  67         /* Because we do not expose that information via XenBus. */
  68         sd->sd.node = first_online_node;
  69         sd->sd.domain = domain;
  70         sd->pdev = pdev;
  71 }
  72
  73 static DEFINE_SPINLOCK(pcifront_dev_lock);
  74 static struct pcifront_device *pcifront_dev;
  75
  76 static int verbose_request;
  77 module_param(verbose_request, int, 0644);
  78
  79 static int errno_to_pcibios_err(int errno)
  80 {
  81         switch (errno) {
  82         case XEN_PCI_ERR_success:
  83                 return PCIBIOS_SUCCESSFUL;
  84
  85         case XEN_PCI_ERR_dev_not_found:
  86                 return PCIBIOS_DEVICE_NOT_FOUND;
  87
  88         case XEN_PCI_ERR_invalid_offset:
  89         case XEN_PCI_ERR_op_failed:
  90                 return PCIBIOS_BAD_REGISTER_NUMBER;
  91
  92         case XEN_PCI_ERR_not_implemented:
  93                 return PCIBIOS_FUNC_NOT_SUPPORTED;
  94
  95         case XEN_PCI_ERR_access_denied:
  96                 return PCIBIOS_SET_FAILED;
  97         }
  98         return errno;
  99 }
 100
 101 static inline void schedule_pcifront_aer_op(struct pcifront_device *pdev)
 102 {
 103         if (test_bit(_XEN_PCIB_active, (unsigned long *)&pdev->sh_info->flags)
 104                 && !test_and_set_bit(_PDEVB_op_active, &pdev->flags)) {
 105                 dev_dbg(&pdev->xdev->dev, "schedule aer frontend job\n");
 106                 schedule_work(&pdev->op_work);
 107         }
 108 }
 109
 110 static int do_pci_op(struct pcifront_device *pdev, struct xen_pci_op *op)
 111 {
 112         int err = 0;
 113         struct xen_pci_op *active_op = &pdev->sh_info->op;
 114         unsigned long irq_flags;
 115         evtchn_port_t port = pdev->evtchn;
 116         unsigned irq = pdev->irq;
 117         s64 ns, ns_timeout;
 118         struct timeval tv;
 119
 120         spin_lock_irqsave(&pdev->sh_info_lock, irq_flags);
 121
 122         memcpy(active_op, op, sizeof(struct xen_pci_op));
 123
 124         /* Go */
 125         wmb();
 126         set_bit(_XEN_PCIF_active, (unsigned long *)&pdev->sh_info->flags);
 127         notify_remote_via_evtchn(port);
 128
 129         /*
 130          * We set a poll timeout of 3 seconds but give up on return after
 131          * 2 seconds. It is better to time out too late rather than too early
 132          * (in the latter case we end up continually re-executing poll() with a
 133          * timeout in the past). 1s difference gives plenty of slack for error.
 134          */
 135         do_gettimeofday(&tv);
 136         ns_timeout = timeval_to_ns(&tv) + 2 * (s64)NSEC_PER_SEC;
 137
 138         xen_clear_irq_pending(irq);
 139
 140         while (test_bit(_XEN_PCIF_active,
 141                         (unsigned long *)&pdev->sh_info->flags)) {
 142                 xen_poll_irq_timeout(irq, jiffies + 3*HZ);
 143                 xen_clear_irq_pending(irq);
 144                 do_gettimeofday(&tv);
 145                 ns = timeval_to_ns(&tv);
 146                 if (ns > ns_timeout) {
 147                         dev_err(&pdev->xdev->dev,
 148                                 "pciback not responding!!!\n");
 149                         clear_bit(_XEN_PCIF_active,
 150                                   (unsigned long *)&pdev->sh_info->flags);
 151                         err = XEN_PCI_ERR_dev_not_found;
 152                         goto out;
 153                 }
 154         }
 155
 156         /*
 157         * We might lose backend service request since we
 158         * reuse same evtchn with pci_conf backend response. So re-schedule
 159         * aer pcifront service.
 160         */
 161         if (test_bit(_XEN_PCIB_active,
 162                         (unsigned long *)&pdev->sh_info->flags)) {
 163                 dev_err(&pdev->xdev->dev,
 164                         "schedule aer pcifront service\n");
 165                 schedule_pcifront_aer_op(pdev);
 166         }
 167
 168         memcpy(op, active_op, sizeof(struct xen_pci_op));
 169
 170         err = op->err;
 171 out:
 172         spin_unlock_irqrestore(&pdev->sh_info_lock, irq_flags);
 173         return err;
 174 }
 175
 176 /* Access to this function is spinlocked in drivers/pci/access.c */
 177 static int pcifront_bus_read(struct pci_bus *bus, unsigned int devfn,
 178                              int where, int size, u32 *val)
 179 {
 180         int err = 0;
 181         struct xen_pci_op op = {
 182                 .cmd    = XEN_PCI_OP_conf_read,
 183                 .domain = pci_domain_nr(bus),
 184                 .bus    = bus->number,
 185                 .devfn  = devfn,
 186                 .offset = where,
 187                 .size   = size,
 188         };
 189         struct pcifront_sd *sd = bus->sysdata;
 190         struct pcifront_device *pdev = pcifront_get_pdev(sd);
 191
 192         if (verbose_request)
 193                 dev_info(&pdev->xdev->dev,
 194                          "read dev=%04x:%02x:%02x.%01x - offset %x size %d\n",
 195                          pci_domain_nr(bus), bus->number, PCI_SLOT(devfn),
 196                          PCI_FUNC(devfn), where, size);
 197
 198         err = do_pci_op(pdev, &op);
 199
 200         if (likely(!err)) {
 201                 if (verbose_request)
 202                         dev_info(&pdev->xdev->dev, "read got back value %x\n",
 203                                  op.value);
 204
 205                 *val = op.value;
 206         } else if (err == -ENODEV) {
 207                 /* No device here, pretend that it just returned 0 */
 208                 err = 0;
 209                 *val = 0;
 210         }
 211
 212         return errno_to_pcibios_err(err);
 213 }
 214
 215 /* Access to this function is spinlocked in drivers/pci/access.c */
 216 static int pcifront_bus_write(struct pci_bus *bus, unsigned int devfn,
 217                               int where, int size, u32 val)
 218 {
 219         struct xen_pci_op op = {
 220                 .cmd    = XEN_PCI_OP_conf_write,
 221                 .domain = pci_domain_nr(bus),
 222                 .bus    = bus->number,
 223                 .devfn  = devfn,
 224                 .offset = where,
 225                 .size   = size,
 226                 .value  = val,
 227         };
 228         struct pcifront_sd *sd = bus->sysdata;
 229         struct pcifront_device *pdev = pcifront_get_pdev(sd);
 230
 231         if (verbose_request)
 232                 dev_info(&pdev->xdev->dev,
 233                          "write dev=%04x:%02x:%02x.%01x - "
 234                          "offset %x size %d val %x\n",
 235                          pci_domain_nr(bus), bus->number,
 236                          PCI_SLOT(devfn), PCI_FUNC(devfn), where, size, val);
 237
 238         return errno_to_pcibios_err(do_pci_op(pdev, &op));
 239 }
 240
 241 struct pci_ops pcifront_bus_ops = {
 242         .read = pcifront_bus_read,
 243         .write = pcifront_bus_write,
 244 };
 245
 246 #ifdef CONFIG_PCI_MSI
 247 static int pci_frontend_enable_msix(struct pci_dev *dev,
 248                                     int vector[], int nvec)
 249 {
 250         int err;
 251         int i;
 252         struct xen_pci_op op = {
 253                 .cmd    = XEN_PCI_OP_enable_msix,
 254                 .domain = pci_domain_nr(dev->bus),
 255                 .bus = dev->bus->number,
 256                 .devfn = dev->devfn,
 257                 .value = nvec,
 258         };
 259         struct pcifront_sd *sd = dev->bus->sysdata;
 260         struct pcifront_device *pdev = pcifront_get_pdev(sd);
 261         struct msi_desc *entry;
 262
 263         if (nvec > SH_INFO_MAX_VEC) {
 264                 dev_err(&dev->dev, "too much vector for pci frontend: %x."
 265                                    " Increase SH_INFO_MAX_VEC.\n", nvec);
 266                 return -EINVAL;
 267         }
 268
 269         i = 0;
 270         list_for_each_entry(entry, &dev->msi_list, list) {
 271                 op.msix_entries[i].entry = entry->msi_attrib.entry_nr;
 272                 /* Vector is useless at this point. */
 273                 op.msix_entries[i].vector = -1;
 274                 i++;
 275         }
 276
 277         err = do_pci_op(pdev, &op);
 278
 279         if (likely(!err)) {
 280                 if (likely(!op.value)) {
 281                         /* we get the result */
 282                         for (i = 0; i < nvec; i++) {
 283                                 if (op.msix_entries[i].vector <= 0) {
 284                                         dev_warn(&dev->dev, "MSI-X entry %d is invalid: %d!\n",
 285                                                 i, op.msix_entries[i].vector);
 286                                         err = -EINVAL;
 287                                         vector[i] = -1;
 288                                         continue;
 289                                 }
 290                                 vector[i] = op.msix_entries[i].vector;
 291                         }
 292                 } else {
 293                         printk(KERN_DEBUG "enable msix get value %x\n",
 294                                 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 __devinit 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.%02x found.\n", domain, bus,
 438                                  PCI_SLOT(devfn), PCI_FUNC(devfn));
 439         }
 440
 441         return 0;
 442 }
 443
 444 static int __devinit 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 = kzalloc(sizeof(*bus_entry), GFP_KERNEL);
 467         sd = kzalloc(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 __devinit 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_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 (get_driver(&pdrv->driver)) {
 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                 put_driver(&pdrv->driver);
 629         }
 630         if (!flag)
 631                 result = PCI_ERS_RESULT_NONE;
 632
 633         return result;
 634 }
 635
 636
 637 static void pcifront_do_aer(struct work_struct *data)
 638 {
 639         struct pcifront_device *pdev =
 640                 container_of(data, struct pcifront_device, op_work);
 641         int cmd = pdev->sh_info->aer_op.cmd;
 642         pci_channel_state_t state =
 643                 (pci_channel_state_t)pdev->sh_info->aer_op.err;
 644
 645         /*If a pci_conf op is in progress,
 646                 we have to wait until it is done before service aer op*/
 647         dev_dbg(&pdev->xdev->dev,
 648                 "pcifront service aer bus %x devfn %x\n",
 649                 pdev->sh_info->aer_op.bus, pdev->sh_info->aer_op.devfn);
 650
 651         pdev->sh_info->aer_op.err = pcifront_common_process(cmd, pdev, state);
 652
 653         /* Post the operation to the guest. */
 654         wmb();
 655         clear_bit(_XEN_PCIB_active, (unsigned long *)&pdev->sh_info->flags);
 656         notify_remote_via_evtchn(pdev->evtchn);
 657
 658         /*in case of we lost an aer request in four lines time_window*/
 659         smp_mb__before_clear_bit();
 660         clear_bit(_PDEVB_op_active, &pdev->flags);
 661         smp_mb__after_clear_bit();
 662
 663         schedule_pcifront_aer_op(pdev);
 664
 665 }
 666
 667 static irqreturn_t pcifront_handler_aer(int irq, void *dev)
 668 {
 669         struct pcifront_device *pdev = dev;
 670         schedule_pcifront_aer_op(pdev);
 671         return IRQ_HANDLED;
 672 }
 673 static int pcifront_connect(struct pcifront_device *pdev)
 674 {
 675         int err = 0;
 676
 677         spin_lock(&pcifront_dev_lock);
 678
 679         if (!pcifront_dev) {
 680                 dev_info(&pdev->xdev->dev, "Installing PCI frontend\n");
 681                 pcifront_dev = pdev;
 682         } else {
 683                 dev_err(&pdev->xdev->dev, "PCI frontend already installed!\n");
 684                 err = -EEXIST;
 685         }
 686
 687         spin_unlock(&pcifront_dev_lock);
 688
 689         return err;
 690 }
 691
 692 static void pcifront_disconnect(struct pcifront_device *pdev)
 693 {
 694         spin_lock(&pcifront_dev_lock);
 695
 696         if (pdev == pcifront_dev) {
 697                 dev_info(&pdev->xdev->dev,
 698                          "Disconnecting PCI Frontend Buses\n");
 699                 pcifront_dev = NULL;
 700         }
 701
 702         spin_unlock(&pcifront_dev_lock);
 703 }
 704 static struct pcifront_device *alloc_pdev(struct xenbus_device *xdev)
 705 {
 706         struct pcifront_device *pdev;
 707
 708         pdev = kzalloc(sizeof(struct pcifront_device), GFP_KERNEL);
 709         if (pdev == NULL)
 710                 goto out;
 711
 712         pdev->sh_info =
 713             (struct xen_pci_sharedinfo *)__get_free_page(GFP_KERNEL);
 714         if (pdev->sh_info == NULL) {
 715                 kfree(pdev);
 716                 pdev = NULL;
 717                 goto out;
 718         }
 719         pdev->sh_info->flags = 0;
 720
 721         /*Flag for registering PV AER handler*/
 722         set_bit(_XEN_PCIB_AERHANDLER, (void *)&pdev->sh_info->flags);
 723
 724         dev_set_drvdata(&xdev->dev, pdev);
 725         pdev->xdev = xdev;
 726
 727         INIT_LIST_HEAD(&pdev->root_buses);
 728
 729         spin_lock_init(&pdev->sh_info_lock);
 730
 731         pdev->evtchn = INVALID_EVTCHN;
 732         pdev->gnt_ref = INVALID_GRANT_REF;
 733         pdev->irq = -1;
 734
 735         INIT_WORK(&pdev->op_work, pcifront_do_aer);
 736
 737         dev_dbg(&xdev->dev, "Allocated pdev @ 0x%p pdev->sh_info @ 0x%p\n",
 738                 pdev, pdev->sh_info);
 739 out:
 740         return pdev;
 741 }
 742
 743 static void free_pdev(struct pcifront_device *pdev)
 744 {
 745         dev_dbg(&pdev->xdev->dev, "freeing pdev @ 0x%p\n", pdev);
 746
 747         pcifront_free_roots(pdev);
 748
 749         cancel_work_sync(&pdev->op_work);
 750
 751         if (pdev->irq >= 0)
 752                 unbind_from_irqhandler(pdev->irq, pdev);
 753
 754         if (pdev->evtchn != INVALID_EVTCHN)
 755                 xenbus_free_evtchn(pdev->xdev, pdev->evtchn);
 756
 757         if (pdev->gnt_ref != INVALID_GRANT_REF)
 758                 gnttab_end_foreign_access(pdev->gnt_ref, 0 /* r/w page */,
 759                                           (unsigned long)pdev->sh_info);
 760         else
 761                 free_page((unsigned long)pdev->sh_info);
 762
 763         dev_set_drvdata(&pdev->xdev->dev, NULL);
 764
 765         kfree(pdev);
 766 }
 767
 768 static int pcifront_publish_info(struct pcifront_device *pdev)
 769 {
 770         int err = 0;
 771         struct xenbus_transaction trans;
 772
 773         err = xenbus_grant_ring(pdev->xdev, virt_to_mfn(pdev->sh_info));
 774         if (err < 0)
 775                 goto out;
 776
 777         pdev->gnt_ref = err;
 778
 779         err = xenbus_alloc_evtchn(pdev->xdev, &pdev->evtchn);
 780         if (err)
 781                 goto out;
 782
 783         err = bind_evtchn_to_irqhandler(pdev->evtchn, pcifront_handler_aer,
 784                 0, "pcifront", pdev);
 785
 786         if (err < 0)
 787                 return err;
 788
 789         pdev->irq = err;
 790
 791 do_publish:
 792         err = xenbus_transaction_start(&trans);
 793         if (err) {
 794                 xenbus_dev_fatal(pdev->xdev, err,
 795                                  "Error writing configuration for backend "
 796                                  "(start transaction)");
 797                 goto out;
 798         }
 799
 800         err = xenbus_printf(trans, pdev->xdev->nodename,
 801                             "pci-op-ref", "%u", pdev->gnt_ref);
 802         if (!err)
 803                 err = xenbus_printf(trans, pdev->xdev->nodename,
 804                                     "event-channel", "%u", pdev->evtchn);
 805         if (!err)
 806                 err = xenbus_printf(trans, pdev->xdev->nodename,
 807                                     "magic", XEN_PCI_MAGIC);
 808
 809         if (err) {
 810                 xenbus_transaction_end(trans, 1);
 811                 xenbus_dev_fatal(pdev->xdev, err,
 812                                  "Error writing configuration for backend");
 813                 goto out;
 814         } else {
 815                 err = xenbus_transaction_end(trans, 0);
 816                 if (err == -EAGAIN)
 817                         goto do_publish;
 818                 else if (err) {
 819                         xenbus_dev_fatal(pdev->xdev, err,
 820                                          "Error completing transaction "
 821                                          "for backend");
 822                         goto out;
 823                 }
 824         }
 825
 826         xenbus_switch_state(pdev->xdev, XenbusStateInitialised);
 827
 828         dev_dbg(&pdev->xdev->dev, "publishing successful!\n");
 829
 830 out:
 831         return err;
 832 }
 833
 834 static int __devinit pcifront_try_connect(struct pcifront_device *pdev)
 835 {
 836         int err = -EFAULT;
 837         int i, num_roots, len;
 838         char str[64];
 839         unsigned int domain, bus;
 840
 841
 842         /* Only connect once */
 843         if (xenbus_read_driver_state(pdev->xdev->nodename) !=
 844             XenbusStateInitialised)
 845                 goto out;
 846
 847         err = pcifront_connect(pdev);
 848         if (err) {
 849                 xenbus_dev_fatal(pdev->xdev, err,
 850                                  "Error connecting PCI Frontend");
 851                 goto out;
 852         }
 853
 854         err = xenbus_scanf(XBT_NIL, pdev->xdev->otherend,
 855                            "root_num", "%d", &num_roots);
 856         if (err == -ENOENT) {
 857                 xenbus_dev_error(pdev->xdev, err,
 858                                  "No PCI Roots found, trying 0000:00");
 859                 err = pcifront_scan_root(pdev, 0, 0);
 860                 num_roots = 0;
 861         } else if (err != 1) {
 862                 if (err == 0)
 863                         err = -EINVAL;
 864                 xenbus_dev_fatal(pdev->xdev, err,
 865                                  "Error reading number of PCI roots");
 866                 goto out;
 867         }
 868
 869         for (i = 0; i < num_roots; i++) {
 870                 len = snprintf(str, sizeof(str), "root-%d", i);
 871                 if (unlikely(len >= (sizeof(str) - 1))) {
 872                         err = -ENOMEM;
 873                         goto out;
 874                 }
 875
 876                 err = xenbus_scanf(XBT_NIL, pdev->xdev->otherend, str,
 877                                    "%x:%x", &domain, &bus);
 878                 if (err != 2) {
 879                         if (err >= 0)
 880                                 err = -EINVAL;
 881                         xenbus_dev_fatal(pdev->xdev, err,
 882                                          "Error reading PCI root %d", i);
 883                         goto out;
 884                 }
 885
 886                 err = pcifront_scan_root(pdev, domain, bus);
 887                 if (err) {
 888                         xenbus_dev_fatal(pdev->xdev, err,
 889                                          "Error scanning PCI root %04x:%02x",
 890                                          domain, bus);
 891                         goto out;
 892                 }
 893         }
 894
 895         err = xenbus_switch_state(pdev->xdev, XenbusStateConnected);
 896
 897 out:
 898         return err;
 899 }
 900
 901 static int pcifront_try_disconnect(struct pcifront_device *pdev)
 902 {
 903         int err = 0;
 904         enum xenbus_state prev_state;
 905
 906
 907         prev_state = xenbus_read_driver_state(pdev->xdev->nodename);
 908
 909         if (prev_state >= XenbusStateClosing)
 910                 goto out;
 911
 912         if (prev_state == XenbusStateConnected) {
 913                 pcifront_free_roots(pdev);
 914                 pcifront_disconnect(pdev);
 915         }
 916
 917         err = xenbus_switch_state(pdev->xdev, XenbusStateClosed);
 918
 919 out:
 920
 921         return err;
 922 }
 923
 924 static int __devinit pcifront_attach_devices(struct pcifront_device *pdev)
 925 {
 926         int err = -EFAULT;
 927         int i, num_roots, len;
 928         unsigned int domain, bus;
 929         char str[64];
 930
 931         if (xenbus_read_driver_state(pdev->xdev->nodename) !=
 932             XenbusStateReconfiguring)
 933                 goto out;
 934
 935         err = xenbus_scanf(XBT_NIL, pdev->xdev->otherend,
 936                            "root_num", "%d", &num_roots);
 937         if (err == -ENOENT) {
 938                 xenbus_dev_error(pdev->xdev, err,
 939                                  "No PCI Roots found, trying 0000:00");
 940                 err = pcifront_rescan_root(pdev, 0, 0);
 941                 num_roots = 0;
 942         } else if (err != 1) {
 943                 if (err == 0)
 944                         err = -EINVAL;
 945                 xenbus_dev_fatal(pdev->xdev, err,
 946                                  "Error reading number of PCI roots");
 947                 goto out;
 948         }
 949
 950         for (i = 0; i < num_roots; i++) {
 951                 len = snprintf(str, sizeof(str), "root-%d", i);
 952                 if (unlikely(len >= (sizeof(str) - 1))) {
 953                         err = -ENOMEM;
 954                         goto out;
 955                 }
 956
 957                 err = xenbus_scanf(XBT_NIL, pdev->xdev->otherend, str,
 958                                    "%x:%x", &domain, &bus);
 959                 if (err != 2) {
 960                         if (err >= 0)
 961                                 err = -EINVAL;
 962                         xenbus_dev_fatal(pdev->xdev, err,
 963                                          "Error reading PCI root %d", i);
 964                         goto out;
 965                 }
 966
 967                 err = pcifront_rescan_root(pdev, domain, bus);
 968                 if (err) {
 969                         xenbus_dev_fatal(pdev->xdev, err,
 970                                          "Error scanning PCI root %04x:%02x",
 971                                          domain, bus);
 972                         goto out;
 973                 }
 974         }
 975
 976         xenbus_switch_state(pdev->xdev, XenbusStateConnected);
 977
 978 out:
 979         return err;
 980 }
 981
 982 static int pcifront_detach_devices(struct pcifront_device *pdev)
 983 {
 984         int err = 0;
 985         int i, num_devs;
 986         unsigned int domain, bus, slot, func;
 987         struct pci_bus *pci_bus;
 988         struct pci_dev *pci_dev;
 989         char str[64];
 990
 991         if (xenbus_read_driver_state(pdev->xdev->nodename) !=
 992             XenbusStateConnected)
 993                 goto out;
 994
 995         err = xenbus_scanf(XBT_NIL, pdev->xdev->otherend, "num_devs", "%d",
 996                            &num_devs);
 997         if (err != 1) {
 998                 if (err >= 0)
 999                         err = -EINVAL;
1000                 xenbus_dev_fatal(pdev->xdev, err,
1001                                  "Error reading number of PCI devices");
1002                 goto out;
1003         }
1004
1005         /* Find devices being detached and remove them. */
1006         for (i = 0; i < num_devs; i++) {
1007                 int l, state;
1008                 l = snprintf(str, sizeof(str), "state-%d", i);
1009                 if (unlikely(l >= (sizeof(str) - 1))) {
1010                         err = -ENOMEM;
1011                         goto out;
1012                 }
1013                 err = xenbus_scanf(XBT_NIL, pdev->xdev->otherend, str, "%d",
1014                                    &state);
1015                 if (err != 1)
1016                         state = XenbusStateUnknown;
1017
1018                 if (state != XenbusStateClosing)
1019                         continue;
1020
1021                 /* Remove device. */
1022                 l = snprintf(str, sizeof(str), "vdev-%d", i);
1023                 if (unlikely(l >= (sizeof(str) - 1))) {
1024                         err = -ENOMEM;
1025                         goto out;
1026                 }
1027                 err = xenbus_scanf(XBT_NIL, pdev->xdev->otherend, str,
1028                                    "%x:%x:%x.%x", &domain, &bus, &slot, &func);
1029                 if (err != 4) {
1030                         if (err >= 0)
1031                                 err = -EINVAL;
1032                         xenbus_dev_fatal(pdev->xdev, err,
1033                                          "Error reading PCI device %d", i);
1034                         goto out;
1035                 }
1036
1037                 pci_bus = pci_find_bus(domain, bus);
1038                 if (!pci_bus) {
1039                         dev_dbg(&pdev->xdev->dev, "Cannot get bus %04x:%02x\n",
1040                                 domain, bus);
1041                         continue;
1042                 }
1043                 pci_dev = pci_get_slot(pci_bus, PCI_DEVFN(slot, func));
1044                 if (!pci_dev) {
1045                         dev_dbg(&pdev->xdev->dev,
1046                                 "Cannot get PCI device %04x:%02x:%02x.%02x\n",
1047                                 domain, bus, slot, func);
1048                         continue;
1049                 }
1050                 pci_remove_bus_device(pci_dev);
1051                 pci_dev_put(pci_dev);
1052
1053                 dev_dbg(&pdev->xdev->dev,
1054                         "PCI device %04x:%02x:%02x.%02x removed.\n",
1055                         domain, bus, slot, func);
1056         }
1057
1058         err = xenbus_switch_state(pdev->xdev, XenbusStateReconfiguring);
1059
1060 out:
1061         return err;
1062 }
1063
1064 static void __init_refok pcifront_backend_changed(struct xenbus_device *xdev,
1065                                                   enum xenbus_state be_state)
1066 {
1067         struct pcifront_device *pdev = dev_get_drvdata(&xdev->dev);
1068
1069         switch (be_state) {
1070         case XenbusStateUnknown:
1071         case XenbusStateInitialising:
1072         case XenbusStateInitWait:
1073         case XenbusStateInitialised:
1074         case XenbusStateClosed:
1075                 break;
1076
1077         case XenbusStateConnected:
1078                 pcifront_try_connect(pdev);
1079                 break;
1080
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 struct xenbus_driver xenbus_pcifront_driver = {
1132         .name                   = "pcifront",
1133         .owner                  = THIS_MODULE,
1134         .ids                    = xenpci_ids,
1135         .probe                  = pcifront_xenbus_probe,
1136         .remove                 = pcifront_xenbus_remove,
1137         .otherend_changed       = pcifront_backend_changed,
1138 };
1139
1140 static int __init pcifront_init(void)
1141 {
1142         if (!xen_pv_domain() || xen_initial_domain())
1143                 return -ENODEV;
1144
1145         pci_frontend_registrar(1 /* enable */);
1146
1147         return xenbus_register_frontend(&xenbus_pcifront_driver);
1148 }
1149
1150 static void __exit pcifront_cleanup(void)
1151 {
1152         xenbus_unregister_driver(&xenbus_pcifront_driver);
1153         pci_frontend_registrar(0 /* disable */);
1154 }
1155 module_init(pcifront_init);
1156 module_exit(pcifront_cleanup);
1157
1158 MODULE_DESCRIPTION("Xen PCI passthrough frontend.");
1159 MODULE_LICENSE("GPL");
1160 MODULE_ALIAS("xen:pci");