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