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         pci_lock_rescan_remove();
 475
 476         b = pci_scan_bus_parented(&pdev->xdev->dev, bus,
 477                                   &pcifront_bus_ops, sd);
 478         if (!b) {
 479                 dev_err(&pdev->xdev->dev,
 480                         "Error creating PCI Frontend Bus!\n");
 481                 err = -ENOMEM;
 482                 pci_unlock_rescan_remove();
 483                 goto err_out;
 484         }
 485
 486         bus_entry->bus = b;
 487
 488         list_add(&bus_entry->list, &pdev->root_buses);
 489
 490         /* pci_scan_bus_parented skips devices which do not have a have
 491         * devfn==0. The pcifront_scan_bus enumerates all devfn. */
 492         err = pcifront_scan_bus(pdev, domain, bus, b);
 493
 494         /* Claim resources before going "live" with our devices */
 495         pci_walk_bus(b, pcifront_claim_resource, pdev);
 496
 497         /* Create SysFS and notify udev of the devices. Aka: "going live" */
 498         pci_bus_add_devices(b);
 499
 500         pci_unlock_rescan_remove();
 501         return err;
 502
 503 err_out:
 504         kfree(bus_entry);
 505         kfree(sd);
 506
 507         return err;
 508 }
 509
 510 static int pcifront_rescan_root(struct pcifront_device *pdev,
 511                                    unsigned int domain, unsigned int bus)
 512 {
 513         int err;
 514         struct pci_bus *b;
 515
 516 #ifndef CONFIG_PCI_DOMAINS
 517         if (domain != 0) {
 518                 dev_err(&pdev->xdev->dev,
 519                         "PCI Root in non-zero PCI Domain! domain=%d\n", domain);
 520                 dev_err(&pdev->xdev->dev,
 521                         "Please compile with CONFIG_PCI_DOMAINS\n");
 522                 return -EINVAL;
 523         }
 524 #endif
 525
 526         dev_info(&pdev->xdev->dev, "Rescanning PCI Frontend Bus %04x:%02x\n",
 527                  domain, bus);
 528
 529         b = pci_find_bus(domain, bus);
 530         if (!b)
 531                 /* If the bus is unknown, create it. */
 532                 return pcifront_scan_root(pdev, domain, bus);
 533
 534         err = pcifront_scan_bus(pdev, domain, bus, b);
 535
 536         /* Claim resources before going "live" with our devices */
 537         pci_walk_bus(b, pcifront_claim_resource, pdev);
 538
 539         /* Create SysFS and notify udev of the devices. Aka: "going live" */
 540         pci_bus_add_devices(b);
 541
 542         return err;
 543 }
 544
 545 static void free_root_bus_devs(struct pci_bus *bus)
 546 {
 547         struct pci_dev *dev;
 548
 549         while (!list_empty(&bus->devices)) {
 550                 dev = container_of(bus->devices.next, struct pci_dev,
 551                                    bus_list);
 552                 dev_dbg(&dev->dev, "removing device\n");
 553                 pci_stop_and_remove_bus_device(dev);
 554         }
 555 }
 556
 557 static void pcifront_free_roots(struct pcifront_device *pdev)
 558 {
 559         struct pci_bus_entry *bus_entry, *t;
 560
 561         dev_dbg(&pdev->xdev->dev, "cleaning up root buses\n");
 562
 563         pci_lock_rescan_remove();
 564         list_for_each_entry_safe(bus_entry, t, &pdev->root_buses, list) {
 565                 list_del(&bus_entry->list);
 566
 567                 free_root_bus_devs(bus_entry->bus);
 568
 569                 kfree(bus_entry->bus->sysdata);
 570
 571                 device_unregister(bus_entry->bus->bridge);
 572                 pci_remove_bus(bus_entry->bus);
 573
 574                 kfree(bus_entry);
 575         }
 576         pci_unlock_rescan_remove();
 577 }
 578
 579 static pci_ers_result_t pcifront_common_process(int cmd,
 580                                                 struct pcifront_device *pdev,
 581                                                 pci_channel_state_t state)
 582 {
 583         pci_ers_result_t result;
 584         struct pci_driver *pdrv;
 585         int bus = pdev->sh_info->aer_op.bus;
 586         int devfn = pdev->sh_info->aer_op.devfn;
 587         struct pci_dev *pcidev;
 588         int flag = 0;
 589
 590         dev_dbg(&pdev->xdev->dev,
 591                 "pcifront AER process: cmd %x (bus:%x, devfn%x)",
 592                 cmd, bus, devfn);
 593         result = PCI_ERS_RESULT_NONE;
 594
 595         pcidev = pci_get_bus_and_slot(bus, devfn);
 596         if (!pcidev || !pcidev->driver) {
 597                 dev_err(&pdev->xdev->dev, "device or AER driver is NULL\n");
 598                 if (pcidev)
 599                         pci_dev_put(pcidev);
 600                 return result;
 601         }
 602         pdrv = pcidev->driver;
 603
 604         if (pdrv) {
 605                 if (pdrv->err_handler && pdrv->err_handler->error_detected) {
 606                         dev_dbg(&pcidev->dev,
 607                                 "trying to call AER service\n");
 608                         if (pcidev) {
 609                                 flag = 1;
 610                                 switch (cmd) {
 611                                 case XEN_PCI_OP_aer_detected:
 612                                         result = pdrv->err_handler->
 613                                                  error_detected(pcidev, state);
 614                                         break;
 615                                 case XEN_PCI_OP_aer_mmio:
 616                                         result = pdrv->err_handler->
 617                                                  mmio_enabled(pcidev);
 618                                         break;
 619                                 case XEN_PCI_OP_aer_slotreset:
 620                                         result = pdrv->err_handler->
 621                                                  slot_reset(pcidev);
 622                                         break;
 623                                 case XEN_PCI_OP_aer_resume:
 624                                         pdrv->err_handler->resume(pcidev);
 625                                         break;
 626                                 default:
 627                                         dev_err(&pdev->xdev->dev,
 628                                                 "bad request in aer recovery "
 629                                                 "operation!\n");
 630
 631                                 }
 632                         }
 633                 }
 634         }
 635         if (!flag)
 636                 result = PCI_ERS_RESULT_NONE;
 637
 638         return result;
 639 }
 640
 641
 642 static void pcifront_do_aer(struct work_struct *data)
 643 {
 644         struct pcifront_device *pdev =
 645                 container_of(data, struct pcifront_device, op_work);
 646         int cmd = pdev->sh_info->aer_op.cmd;
 647         pci_channel_state_t state =
 648                 (pci_channel_state_t)pdev->sh_info->aer_op.err;
 649
 650         /*If a pci_conf op is in progress,
 651                 we have to wait until it is done before service aer op*/
 652         dev_dbg(&pdev->xdev->dev,
 653                 "pcifront service aer bus %x devfn %x\n",
 654                 pdev->sh_info->aer_op.bus, pdev->sh_info->aer_op.devfn);
 655
 656         pdev->sh_info->aer_op.err = pcifront_common_process(cmd, pdev, state);
 657
 658         /* Post the operation to the guest. */
 659         wmb();
 660         clear_bit(_XEN_PCIB_active, (unsigned long *)&pdev->sh_info->flags);
 661         notify_remote_via_evtchn(pdev->evtchn);
 662
 663         /*in case of we lost an aer request in four lines time_window*/
 664         smp_mb__before_clear_bit();
 665         clear_bit(_PDEVB_op_active, &pdev->flags);
 666         smp_mb__after_clear_bit();
 667
 668         schedule_pcifront_aer_op(pdev);
 669
 670 }
 671
 672 static irqreturn_t pcifront_handler_aer(int irq, void *dev)
 673 {
 674         struct pcifront_device *pdev = dev;
 675         schedule_pcifront_aer_op(pdev);
 676         return IRQ_HANDLED;
 677 }
 678 static int pcifront_connect_and_init_dma(struct pcifront_device *pdev)
 679 {
 680         int err = 0;
 681
 682         spin_lock(&pcifront_dev_lock);
 683
 684         if (!pcifront_dev) {
 685                 dev_info(&pdev->xdev->dev, "Installing PCI frontend\n");
 686                 pcifront_dev = pdev;
 687         } else
 688                 err = -EEXIST;
 689
 690         spin_unlock(&pcifront_dev_lock);
 691
 692         if (!err && !swiotlb_nr_tbl()) {
 693                 err = pci_xen_swiotlb_init_late();
 694                 if (err)
 695                         dev_err(&pdev->xdev->dev, "Could not setup SWIOTLB!\n");
 696         }
 697         return err;
 698 }
 699
 700 static void pcifront_disconnect(struct pcifront_device *pdev)
 701 {
 702         spin_lock(&pcifront_dev_lock);
 703
 704         if (pdev == pcifront_dev) {
 705                 dev_info(&pdev->xdev->dev,
 706                          "Disconnecting PCI Frontend Buses\n");
 707                 pcifront_dev = NULL;
 708         }
 709
 710         spin_unlock(&pcifront_dev_lock);
 711 }
 712 static struct pcifront_device *alloc_pdev(struct xenbus_device *xdev)
 713 {
 714         struct pcifront_device *pdev;
 715
 716         pdev = kzalloc(sizeof(struct pcifront_device), GFP_KERNEL);
 717         if (pdev == NULL)
 718                 goto out;
 719
 720         pdev->sh_info =
 721             (struct xen_pci_sharedinfo *)__get_free_page(GFP_KERNEL);
 722         if (pdev->sh_info == NULL) {
 723                 kfree(pdev);
 724                 pdev = NULL;
 725                 goto out;
 726         }
 727         pdev->sh_info->flags = 0;
 728
 729         /*Flag for registering PV AER handler*/
 730         set_bit(_XEN_PCIB_AERHANDLER, (void *)&pdev->sh_info->flags);
 731
 732         dev_set_drvdata(&xdev->dev, pdev);
 733         pdev->xdev = xdev;
 734
 735         INIT_LIST_HEAD(&pdev->root_buses);
 736
 737         spin_lock_init(&pdev->sh_info_lock);
 738
 739         pdev->evtchn = INVALID_EVTCHN;
 740         pdev->gnt_ref = INVALID_GRANT_REF;
 741         pdev->irq = -1;
 742
 743         INIT_WORK(&pdev->op_work, pcifront_do_aer);
 744
 745         dev_dbg(&xdev->dev, "Allocated pdev @ 0x%p pdev->sh_info @ 0x%p\n",
 746                 pdev, pdev->sh_info);
 747 out:
 748         return pdev;
 749 }
 750
 751 static void free_pdev(struct pcifront_device *pdev)
 752 {
 753         dev_dbg(&pdev->xdev->dev, "freeing pdev @ 0x%p\n", pdev);
 754
 755         pcifront_free_roots(pdev);
 756
 757         cancel_work_sync(&pdev->op_work);
 758
 759         if (pdev->irq >= 0)
 760                 unbind_from_irqhandler(pdev->irq, pdev);
 761
 762         if (pdev->evtchn != INVALID_EVTCHN)
 763                 xenbus_free_evtchn(pdev->xdev, pdev->evtchn);
 764
 765         if (pdev->gnt_ref != INVALID_GRANT_REF)
 766                 gnttab_end_foreign_access(pdev->gnt_ref, 0 /* r/w page */,
 767                                           (unsigned long)pdev->sh_info);
 768         else
 769                 free_page((unsigned long)pdev->sh_info);
 770
 771         dev_set_drvdata(&pdev->xdev->dev, NULL);
 772
 773         kfree(pdev);
 774 }
 775
 776 static int pcifront_publish_info(struct pcifront_device *pdev)
 777 {
 778         int err = 0;
 779         struct xenbus_transaction trans;
 780
 781         err = xenbus_grant_ring(pdev->xdev, virt_to_mfn(pdev->sh_info));
 782         if (err < 0)
 783                 goto out;
 784
 785         pdev->gnt_ref = err;
 786
 787         err = xenbus_alloc_evtchn(pdev->xdev, &pdev->evtchn);
 788         if (err)
 789                 goto out;
 790
 791         err = bind_evtchn_to_irqhandler(pdev->evtchn, pcifront_handler_aer,
 792                 0, "pcifront", pdev);
 793
 794         if (err < 0)
 795                 return err;
 796
 797         pdev->irq = err;
 798
 799 do_publish:
 800         err = xenbus_transaction_start(&trans);
 801         if (err) {
 802                 xenbus_dev_fatal(pdev->xdev, err,
 803                                  "Error writing configuration for backend "
 804                                  "(start transaction)");
 805                 goto out;
 806         }
 807
 808         err = xenbus_printf(trans, pdev->xdev->nodename,
 809                             "pci-op-ref", "%u", pdev->gnt_ref);
 810         if (!err)
 811                 err = xenbus_printf(trans, pdev->xdev->nodename,
 812                                     "event-channel", "%u", pdev->evtchn);
 813         if (!err)
 814                 err = xenbus_printf(trans, pdev->xdev->nodename,
 815                                     "magic", XEN_PCI_MAGIC);
 816
 817         if (err) {
 818                 xenbus_transaction_end(trans, 1);
 819                 xenbus_dev_fatal(pdev->xdev, err,
 820                                  "Error writing configuration for backend");
 821                 goto out;
 822         } else {
 823                 err = xenbus_transaction_end(trans, 0);
 824                 if (err == -EAGAIN)
 825                         goto do_publish;
 826                 else if (err) {
 827                         xenbus_dev_fatal(pdev->xdev, err,
 828                                          "Error completing transaction "
 829                                          "for backend");
 830                         goto out;
 831                 }
 832         }
 833
 834         xenbus_switch_state(pdev->xdev, XenbusStateInitialised);
 835
 836         dev_dbg(&pdev->xdev->dev, "publishing successful!\n");
 837
 838 out:
 839         return err;
 840 }
 841
 842 static int pcifront_try_connect(struct pcifront_device *pdev)
 843 {
 844         int err = -EFAULT;
 845         int i, num_roots, len;
 846         char str[64];
 847         unsigned int domain, bus;
 848
 849
 850         /* Only connect once */
 851         if (xenbus_read_driver_state(pdev->xdev->nodename) !=
 852             XenbusStateInitialised)
 853                 goto out;
 854
 855         err = pcifront_connect_and_init_dma(pdev);
 856         if (err && err != -EEXIST) {
 857                 xenbus_dev_fatal(pdev->xdev, err,
 858                                  "Error setting up PCI Frontend");
 859                 goto out;
 860         }
 861
 862         err = xenbus_scanf(XBT_NIL, pdev->xdev->otherend,
 863                            "root_num", "%d", &num_roots);
 864         if (err == -ENOENT) {
 865                 xenbus_dev_error(pdev->xdev, err,
 866                                  "No PCI Roots found, trying 0000:00");
 867                 err = pcifront_scan_root(pdev, 0, 0);
 868                 num_roots = 0;
 869         } else if (err != 1) {
 870                 if (err == 0)
 871                         err = -EINVAL;
 872                 xenbus_dev_fatal(pdev->xdev, err,
 873                                  "Error reading number of PCI roots");
 874                 goto out;
 875         }
 876
 877         for (i = 0; i < num_roots; i++) {
 878                 len = snprintf(str, sizeof(str), "root-%d", i);
 879                 if (unlikely(len >= (sizeof(str) - 1))) {
 880                         err = -ENOMEM;
 881                         goto out;
 882                 }
 883
 884                 err = xenbus_scanf(XBT_NIL, pdev->xdev->otherend, str,
 885                                    "%x:%x", &domain, &bus);
 886                 if (err != 2) {
 887                         if (err >= 0)
 888                                 err = -EINVAL;
 889                         xenbus_dev_fatal(pdev->xdev, err,
 890                                          "Error reading PCI root %d", i);
 891                         goto out;
 892                 }
 893
 894                 err = pcifront_scan_root(pdev, domain, bus);
 895                 if (err) {
 896                         xenbus_dev_fatal(pdev->xdev, err,
 897                                          "Error scanning PCI root %04x:%02x",
 898                                          domain, bus);
 899                         goto out;
 900                 }
 901         }
 902
 903         err = xenbus_switch_state(pdev->xdev, XenbusStateConnected);
 904
 905 out:
 906         return err;
 907 }
 908
 909 static int pcifront_try_disconnect(struct pcifront_device *pdev)
 910 {
 911         int err = 0;
 912         enum xenbus_state prev_state;
 913
 914
 915         prev_state = xenbus_read_driver_state(pdev->xdev->nodename);
 916
 917         if (prev_state >= XenbusStateClosing)
 918                 goto out;
 919
 920         if (prev_state == XenbusStateConnected) {
 921                 pcifront_free_roots(pdev);
 922                 pcifront_disconnect(pdev);
 923         }
 924
 925         err = xenbus_switch_state(pdev->xdev, XenbusStateClosed);
 926
 927 out:
 928
 929         return err;
 930 }
 931
 932 static int pcifront_attach_devices(struct pcifront_device *pdev)
 933 {
 934         int err = -EFAULT;
 935         int i, num_roots, len;
 936         unsigned int domain, bus;
 937         char str[64];
 938
 939         if (xenbus_read_driver_state(pdev->xdev->nodename) !=
 940             XenbusStateReconfiguring)
 941                 goto out;
 942
 943         err = xenbus_scanf(XBT_NIL, pdev->xdev->otherend,
 944                            "root_num", "%d", &num_roots);
 945         if (err == -ENOENT) {
 946                 xenbus_dev_error(pdev->xdev, err,
 947                                  "No PCI Roots found, trying 0000:00");
 948                 err = pcifront_rescan_root(pdev, 0, 0);
 949                 num_roots = 0;
 950         } else if (err != 1) {
 951                 if (err == 0)
 952                         err = -EINVAL;
 953                 xenbus_dev_fatal(pdev->xdev, err,
 954                                  "Error reading number of PCI roots");
 955                 goto out;
 956         }
 957
 958         for (i = 0; i < num_roots; i++) {
 959                 len = snprintf(str, sizeof(str), "root-%d", i);
 960                 if (unlikely(len >= (sizeof(str) - 1))) {
 961                         err = -ENOMEM;
 962                         goto out;
 963                 }
 964
 965                 err = xenbus_scanf(XBT_NIL, pdev->xdev->otherend, str,
 966                                    "%x:%x", &domain, &bus);
 967                 if (err != 2) {
 968                         if (err >= 0)
 969                                 err = -EINVAL;
 970                         xenbus_dev_fatal(pdev->xdev, err,
 971                                          "Error reading PCI root %d", i);
 972                         goto out;
 973                 }
 974
 975                 err = pcifront_rescan_root(pdev, domain, bus);
 976                 if (err) {
 977                         xenbus_dev_fatal(pdev->xdev, err,
 978                                          "Error scanning PCI root %04x:%02x",
 979                                          domain, bus);
 980                         goto out;
 981                 }
 982         }
 983
 984         xenbus_switch_state(pdev->xdev, XenbusStateConnected);
 985
 986 out:
 987         return err;
 988 }
 989
 990 static int pcifront_detach_devices(struct pcifront_device *pdev)
 991 {
 992         int err = 0;
 993         int i, num_devs;
 994         unsigned int domain, bus, slot, func;
 995         struct pci_dev *pci_dev;
 996         char str[64];
 997
 998         if (xenbus_read_driver_state(pdev->xdev->nodename) !=
 999             XenbusStateConnected)
1000                 goto out;
1001
1002         err = xenbus_scanf(XBT_NIL, pdev->xdev->otherend, "num_devs", "%d",
1003                            &num_devs);
1004         if (err != 1) {
1005                 if (err >= 0)
1006                         err = -EINVAL;
1007                 xenbus_dev_fatal(pdev->xdev, err,
1008                                  "Error reading number of PCI devices");
1009                 goto out;
1010         }
1011
1012         /* Find devices being detached and remove them. */
1013         for (i = 0; i < num_devs; i++) {
1014                 int l, state;
1015                 l = snprintf(str, sizeof(str), "state-%d", i);
1016                 if (unlikely(l >= (sizeof(str) - 1))) {
1017                         err = -ENOMEM;
1018                         goto out;
1019                 }
1020                 err = xenbus_scanf(XBT_NIL, pdev->xdev->otherend, str, "%d",
1021                                    &state);
1022                 if (err != 1)
1023                         state = XenbusStateUnknown;
1024
1025                 if (state != XenbusStateClosing)
1026                         continue;
1027
1028                 /* Remove device. */
1029                 l = snprintf(str, sizeof(str), "vdev-%d", i);
1030                 if (unlikely(l >= (sizeof(str) - 1))) {
1031                         err = -ENOMEM;
1032                         goto out;
1033                 }
1034                 err = xenbus_scanf(XBT_NIL, pdev->xdev->otherend, str,
1035                                    "%x:%x:%x.%x", &domain, &bus, &slot, &func);
1036                 if (err != 4) {
1037                         if (err >= 0)
1038                                 err = -EINVAL;
1039                         xenbus_dev_fatal(pdev->xdev, err,
1040                                          "Error reading PCI device %d", i);
1041                         goto out;
1042                 }
1043
1044                 pci_dev = pci_get_domain_bus_and_slot(domain, bus,
1045                                 PCI_DEVFN(slot, func));
1046                 if (!pci_dev) {
1047                         dev_dbg(&pdev->xdev->dev,
1048                                 "Cannot get PCI device %04x:%02x:%02x.%d\n",
1049                                 domain, bus, slot, func);
1050                         continue;
1051                 }
1052                 pci_lock_rescan_remove();
1053                 pci_stop_and_remove_bus_device(pci_dev);
1054                 pci_dev_put(pci_dev);
1055                 pci_unlock_rescan_remove();
1056
1057                 dev_dbg(&pdev->xdev->dev,
1058                         "PCI device %04x:%02x:%02x.%d removed.\n",
1059                         domain, bus, slot, func);
1060         }
1061
1062         err = xenbus_switch_state(pdev->xdev, XenbusStateReconfiguring);
1063
1064 out:
1065         return err;
1066 }
1067
1068 static void __init_refok pcifront_backend_changed(struct xenbus_device *xdev,
1069                                                   enum xenbus_state be_state)
1070 {
1071         struct pcifront_device *pdev = dev_get_drvdata(&xdev->dev);
1072
1073         switch (be_state) {
1074         case XenbusStateUnknown:
1075         case XenbusStateInitialising:
1076         case XenbusStateInitWait:
1077         case XenbusStateInitialised:
1078                 break;
1079
1080         case XenbusStateConnected:
1081                 pcifront_try_connect(pdev);
1082                 break;
1083
1084         case XenbusStateClosed:
1085                 if (xdev->state == XenbusStateClosed)
1086                         break;
1087                 /* Missed the backend's CLOSING state -- fallthrough */
1088         case XenbusStateClosing:
1089                 dev_warn(&xdev->dev, "backend going away!\n");
1090                 pcifront_try_disconnect(pdev);
1091                 break;
1092
1093         case XenbusStateReconfiguring:
1094                 pcifront_detach_devices(pdev);
1095                 break;
1096
1097         case XenbusStateReconfigured:
1098                 pcifront_attach_devices(pdev);
1099                 break;
1100         }
1101 }
1102
1103 static int pcifront_xenbus_probe(struct xenbus_device *xdev,
1104                                  const struct xenbus_device_id *id)
1105 {
1106         int err = 0;
1107         struct pcifront_device *pdev = alloc_pdev(xdev);
1108
1109         if (pdev == NULL) {
1110                 err = -ENOMEM;
1111                 xenbus_dev_fatal(xdev, err,
1112                                  "Error allocating pcifront_device struct");
1113                 goto out;
1114         }
1115
1116         err = pcifront_publish_info(pdev);
1117         if (err)
1118                 free_pdev(pdev);
1119
1120 out:
1121         return err;
1122 }
1123
1124 static int pcifront_xenbus_remove(struct xenbus_device *xdev)
1125 {
1126         struct pcifront_device *pdev = dev_get_drvdata(&xdev->dev);
1127         if (pdev)
1128                 free_pdev(pdev);
1129
1130         return 0;
1131 }
1132
1133 static const struct xenbus_device_id xenpci_ids[] = {
1134         {"pci"},
1135         {""},
1136 };
1137
1138 static DEFINE_XENBUS_DRIVER(xenpci, "pcifront",
1139         .probe                  = pcifront_xenbus_probe,
1140         .remove                 = pcifront_xenbus_remove,
1141         .otherend_changed       = pcifront_backend_changed,
1142 );
1143
1144 static int __init pcifront_init(void)
1145 {
1146         if (!xen_pv_domain() || xen_initial_domain())
1147                 return -ENODEV;
1148
1149         pci_frontend_registrar(1 /* enable */);
1150
1151         return xenbus_register_frontend(&xenpci_driver);
1152 }
1153
1154 static void __exit pcifront_cleanup(void)
1155 {
1156         xenbus_unregister_driver(&xenpci_driver);
1157         pci_frontend_registrar(0 /* disable */);
1158 }
1159 module_init(pcifront_init);
1160 module_exit(pcifront_cleanup);
1161
1162 MODULE_DESCRIPTION("Xen PCI passthrough frontend.");
1163 MODULE_LICENSE("GPL");
1164 MODULE_ALIAS("xen:pci");