drivers/pci/xen-pcifront.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Xen PCI Frontend
   4  *
   5  * Author: Ryan Wilson <hap9@epoch.ncsc.mil>
   6  */
   7 #include <linux/module.h>
   8 #include <linux/init.h>
   9 #include <linux/mm.h>
  10 #include <xen/xenbus.h>
  11 #include <xen/events.h>
  12 #include <xen/grant_table.h>
  13 #include <xen/page.h>
  14 #include <linux/spinlock.h>
  15 #include <linux/pci.h>
  16 #include <linux/msi.h>
  17 #include <xen/interface/io/pciif.h>
  18 #include <asm/xen/pci.h>
  19 #include <linux/interrupt.h>
  20 #include <linux/atomic.h>
  21 #include <linux/workqueue.h>
  22 #include <linux/bitops.h>
  23 #include <linux/time.h>
  24 #include <linux/ktime.h>
  25 #include <linux/swiotlb.h>
  26 #include <xen/platform_pci.h>
  27
  28 #include <asm/xen/swiotlb-xen.h>
  29
  30 #define INVALID_EVTCHN    (-1)
  31
  32 struct pci_bus_entry {
  33         struct list_head list;
  34         struct pci_bus *bus;
  35 };
  36
  37 #define _PDEVB_op_active                (0)
  38 #define PDEVB_op_active                 (1 << (_PDEVB_op_active))
  39
  40 struct pcifront_device {
  41         struct xenbus_device *xdev;
  42         struct list_head root_buses;
  43
  44         int evtchn;
  45         grant_ref_t gnt_ref;
  46
  47         int irq;
  48
  49         /* Lock this when doing any operations in sh_info */
  50         spinlock_t sh_info_lock;
  51         struct xen_pci_sharedinfo *sh_info;
  52         struct work_struct op_work;
  53         unsigned long flags;
  54
  55 };
  56
  57 struct pcifront_sd {
  58         struct pci_sysdata sd;
  59         struct pcifront_device *pdev;
  60 };
  61
  62 static inline struct pcifront_device *
  63 pcifront_get_pdev(struct pcifront_sd *sd)
  64 {
  65         return sd->pdev;
  66 }
  67
  68 static inline void pcifront_init_sd(struct pcifront_sd *sd,
  69                                     unsigned int domain, unsigned int bus,
  70                                     struct pcifront_device *pdev)
  71 {
  72         /* Because we do not expose that information via XenBus. */
  73         sd->sd.node = first_online_node;
  74         sd->sd.domain = domain;
  75         sd->pdev = pdev;
  76 }
  77
  78 static DEFINE_SPINLOCK(pcifront_dev_lock);
  79 static struct pcifront_device *pcifront_dev;
  80
  81 static int errno_to_pcibios_err(int errno)
  82 {
  83         switch (errno) {
  84         case XEN_PCI_ERR_success:
  85                 return PCIBIOS_SUCCESSFUL;
  86
  87         case XEN_PCI_ERR_dev_not_found:
  88                 return PCIBIOS_DEVICE_NOT_FOUND;
  89
  90         case XEN_PCI_ERR_invalid_offset:
  91         case XEN_PCI_ERR_op_failed:
  92                 return PCIBIOS_BAD_REGISTER_NUMBER;
  93
  94         case XEN_PCI_ERR_not_implemented:
  95                 return PCIBIOS_FUNC_NOT_SUPPORTED;
  96
  97         case XEN_PCI_ERR_access_denied:
  98                 return PCIBIOS_SET_FAILED;
  99         }
 100         return errno;
 101 }
 102
 103 static inline void schedule_pcifront_aer_op(struct pcifront_device *pdev)
 104 {
 105         if (test_bit(_XEN_PCIB_active, (unsigned long *)&pdev->sh_info->flags)
 106                 && !test_and_set_bit(_PDEVB_op_active, &pdev->flags)) {
 107                 dev_dbg(&pdev->xdev->dev, "schedule aer frontend job\n");
 108                 schedule_work(&pdev->op_work);
 109         }
 110 }
 111
 112 static int do_pci_op(struct pcifront_device *pdev, struct xen_pci_op *op)
 113 {
 114         int err = 0;
 115         struct xen_pci_op *active_op = &pdev->sh_info->op;
 116         unsigned long irq_flags;
 117         evtchn_port_t port = pdev->evtchn;
 118         unsigned irq = pdev->irq;
 119         s64 ns, ns_timeout;
 120
 121         spin_lock_irqsave(&pdev->sh_info_lock, irq_flags);
 122
 123         memcpy(active_op, op, sizeof(struct xen_pci_op));
 124
 125         /* Go */
 126         wmb();
 127         set_bit(_XEN_PCIF_active, (unsigned long *)&pdev->sh_info->flags);
 128         notify_remote_via_evtchn(port);
 129
 130         /*
 131          * We set a poll timeout of 3 seconds but give up on return after
 132          * 2 seconds. It is better to time out too late rather than too early
 133          * (in the latter case we end up continually re-executing poll() with a
 134          * timeout in the past). 1s difference gives plenty of slack for error.
 135          */
 136         ns_timeout = ktime_get_ns() + 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                 ns = ktime_get_ns();
 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         dev_dbg(&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                 dev_dbg(&pdev->xdev->dev, "read got back value %x\n",
 200                         op.value);
 201
 202                 *val = op.value;
 203         } else if (err == -ENODEV) {
 204                 /* No device here, pretend that it just returned 0 */
 205                 err = 0;
 206                 *val = 0;
 207         }
 208
 209         return errno_to_pcibios_err(err);
 210 }
 211
 212 /* Access to this function is spinlocked in drivers/pci/access.c */
 213 static int pcifront_bus_write(struct pci_bus *bus, unsigned int devfn,
 214                               int where, int size, u32 val)
 215 {
 216         struct xen_pci_op op = {
 217                 .cmd    = XEN_PCI_OP_conf_write,
 218                 .domain = pci_domain_nr(bus),
 219                 .bus    = bus->number,
 220                 .devfn  = devfn,
 221                 .offset = where,
 222                 .size   = size,
 223                 .value  = val,
 224         };
 225         struct pcifront_sd *sd = bus->sysdata;
 226         struct pcifront_device *pdev = pcifront_get_pdev(sd);
 227
 228         dev_dbg(&pdev->xdev->dev,
 229                 "write dev=%04x:%02x:%02x.%d - offset %x size %d val %x\n",
 230                 pci_domain_nr(bus), bus->number,
 231                 PCI_SLOT(devfn), PCI_FUNC(devfn), where, size, val);
 232
 233         return errno_to_pcibios_err(do_pci_op(pdev, &op));
 234 }
 235
 236 static struct pci_ops pcifront_bus_ops = {
 237         .read = pcifront_bus_read,
 238         .write = pcifront_bus_write,
 239 };
 240
 241 #ifdef CONFIG_PCI_MSI
 242 static int pci_frontend_enable_msix(struct pci_dev *dev,
 243                                     int vector[], int nvec)
 244 {
 245         int err;
 246         int i;
 247         struct xen_pci_op op = {
 248                 .cmd    = XEN_PCI_OP_enable_msix,
 249                 .domain = pci_domain_nr(dev->bus),
 250                 .bus = dev->bus->number,
 251                 .devfn = dev->devfn,
 252                 .value = nvec,
 253         };
 254         struct pcifront_sd *sd = dev->bus->sysdata;
 255         struct pcifront_device *pdev = pcifront_get_pdev(sd);
 256         struct msi_desc *entry;
 257
 258         if (nvec > SH_INFO_MAX_VEC) {
 259                 pci_err(dev, "too many vectors (0x%x) for PCI frontend:"
 260                                    " Increase SH_INFO_MAX_VEC\n", nvec);
 261                 return -EINVAL;
 262         }
 263
 264         i = 0;
 265         for_each_pci_msi_entry(entry, dev) {
 266                 op.msix_entries[i].entry = entry->msi_attrib.entry_nr;
 267                 /* Vector is useless at this point. */
 268                 op.msix_entries[i].vector = -1;
 269                 i++;
 270         }
 271
 272         err = do_pci_op(pdev, &op);
 273
 274         if (likely(!err)) {
 275                 if (likely(!op.value)) {
 276                         /* we get the result */
 277                         for (i = 0; i < nvec; i++) {
 278                                 if (op.msix_entries[i].vector <= 0) {
 279                                         pci_warn(dev, "MSI-X entry %d is invalid: %d!\n",
 280                                                 i, op.msix_entries[i].vector);
 281                                         err = -EINVAL;
 282                                         vector[i] = -1;
 283                                         continue;
 284                                 }
 285                                 vector[i] = op.msix_entries[i].vector;
 286                         }
 287                 } else {
 288                         pr_info("enable msix get value %x\n", op.value);
 289                         err = op.value;
 290                 }
 291         } else {
 292                 pci_err(dev, "enable msix get err %x\n", err);
 293         }
 294         return err;
 295 }
 296
 297 static void pci_frontend_disable_msix(struct pci_dev *dev)
 298 {
 299         int err;
 300         struct xen_pci_op op = {
 301                 .cmd    = XEN_PCI_OP_disable_msix,
 302                 .domain = pci_domain_nr(dev->bus),
 303                 .bus = dev->bus->number,
 304                 .devfn = dev->devfn,
 305         };
 306         struct pcifront_sd *sd = dev->bus->sysdata;
 307         struct pcifront_device *pdev = pcifront_get_pdev(sd);
 308
 309         err = do_pci_op(pdev, &op);
 310
 311         /* What should do for error ? */
 312         if (err)
 313                 pci_err(dev, "pci_disable_msix get err %x\n", err);
 314 }
 315
 316 static int pci_frontend_enable_msi(struct pci_dev *dev, int vector[])
 317 {
 318         int err;
 319         struct xen_pci_op op = {
 320                 .cmd    = XEN_PCI_OP_enable_msi,
 321                 .domain = pci_domain_nr(dev->bus),
 322                 .bus = dev->bus->number,
 323                 .devfn = dev->devfn,
 324         };
 325         struct pcifront_sd *sd = dev->bus->sysdata;
 326         struct pcifront_device *pdev = pcifront_get_pdev(sd);
 327
 328         err = do_pci_op(pdev, &op);
 329         if (likely(!err)) {
 330                 vector[0] = op.value;
 331                 if (op.value <= 0) {
 332                         pci_warn(dev, "MSI entry is invalid: %d!\n",
 333                                 op.value);
 334                         err = -EINVAL;
 335                         vector[0] = -1;
 336                 }
 337         } else {
 338                 pci_err(dev, "pci frontend enable msi failed for dev "
 339                                     "%x:%x\n", op.bus, op.devfn);
 340                 err = -EINVAL;
 341         }
 342         return err;
 343 }
 344
 345 static void pci_frontend_disable_msi(struct pci_dev *dev)
 346 {
 347         int err;
 348         struct xen_pci_op op = {
 349                 .cmd    = XEN_PCI_OP_disable_msi,
 350                 .domain = pci_domain_nr(dev->bus),
 351                 .bus = dev->bus->number,
 352                 .devfn = dev->devfn,
 353         };
 354         struct pcifront_sd *sd = dev->bus->sysdata;
 355         struct pcifront_device *pdev = pcifront_get_pdev(sd);
 356
 357         err = do_pci_op(pdev, &op);
 358         if (err == XEN_PCI_ERR_dev_not_found) {
 359                 /* XXX No response from backend, what shall we do? */
 360                 pr_info("get no response from backend for disable MSI\n");
 361                 return;
 362         }
 363         if (err)
 364                 /* how can pciback notify us fail? */
 365                 pr_info("get fake response from backend\n");
 366 }
 367
 368 static struct xen_pci_frontend_ops pci_frontend_ops = {
 369         .enable_msi = pci_frontend_enable_msi,
 370         .disable_msi = pci_frontend_disable_msi,
 371         .enable_msix = pci_frontend_enable_msix,
 372         .disable_msix = pci_frontend_disable_msix,
 373 };
 374
 375 static void pci_frontend_registrar(int enable)
 376 {
 377         if (enable)
 378                 xen_pci_frontend = &pci_frontend_ops;
 379         else
 380                 xen_pci_frontend = NULL;
 381 };
 382 #else
 383 static inline void pci_frontend_registrar(int enable) { };
 384 #endif /* CONFIG_PCI_MSI */
 385
 386 /* Claim resources for the PCI frontend as-is, backend won't allow changes */
 387 static int pcifront_claim_resource(struct pci_dev *dev, void *data)
 388 {
 389         struct pcifront_device *pdev = data;
 390         int i;
 391         struct resource *r;
 392
 393         for (i = 0; i < PCI_NUM_RESOURCES; i++) {
 394                 r = &dev->resource[i];
 395
 396                 if (!r->parent && r->start && r->flags) {
 397                         dev_info(&pdev->xdev->dev, "claiming resource %s/%d\n",
 398                                 pci_name(dev), i);
 399                         if (pci_claim_resource(dev, i)) {
 400                                 dev_err(&pdev->xdev->dev, "Could not claim resource %s/%d! "
 401                                         "Device offline. Try using e820_host=1 in the guest config.\n",
 402                                         pci_name(dev), i);
 403                         }
 404                 }
 405         }
 406
 407         return 0;
 408 }
 409
 410 static int pcifront_scan_bus(struct pcifront_device *pdev,
 411                                 unsigned int domain, unsigned int bus,
 412                                 struct pci_bus *b)
 413 {
 414         struct pci_dev *d;
 415         unsigned int devfn;
 416
 417         /* Scan the bus for functions and add.
 418          * We omit handling of PCI bridge attachment because pciback prevents
 419          * bridges from being exported.
 420          */
 421         for (devfn = 0; devfn < 0x100; devfn++) {
 422                 d = pci_get_slot(b, devfn);
 423                 if (d) {
 424                         /* Device is already known. */
 425                         pci_dev_put(d);
 426                         continue;
 427                 }
 428
 429                 d = pci_scan_single_device(b, devfn);
 430                 if (d)
 431                         dev_info(&pdev->xdev->dev, "New device on "
 432                                  "%04x:%02x:%02x.%d found.\n", domain, bus,
 433                                  PCI_SLOT(devfn), PCI_FUNC(devfn));
 434         }
 435
 436         return 0;
 437 }
 438
 439 static int pcifront_scan_root(struct pcifront_device *pdev,
 440                                  unsigned int domain, unsigned int bus)
 441 {
 442         struct pci_bus *b;
 443         LIST_HEAD(resources);
 444         struct pcifront_sd *sd = NULL;
 445         struct pci_bus_entry *bus_entry = NULL;
 446         int err = 0;
 447         static struct resource busn_res = {
 448                 .start = 0,
 449                 .end = 255,
 450                 .flags = IORESOURCE_BUS,
 451         };
 452
 453 #ifndef CONFIG_PCI_DOMAINS
 454         if (domain != 0) {
 455                 dev_err(&pdev->xdev->dev,
 456                         "PCI Root in non-zero PCI Domain! domain=%d\n", domain);
 457                 dev_err(&pdev->xdev->dev,
 458                         "Please compile with CONFIG_PCI_DOMAINS\n");
 459                 err = -EINVAL;
 460                 goto err_out;
 461         }
 462 #endif
 463
 464         dev_info(&pdev->xdev->dev, "Creating PCI Frontend Bus %04x:%02x\n",
 465                  domain, bus);
 466
 467         bus_entry = kzalloc(sizeof(*bus_entry), GFP_KERNEL);
 468         sd = kzalloc(sizeof(*sd), GFP_KERNEL);
 469         if (!bus_entry || !sd) {
 470                 err = -ENOMEM;
 471                 goto err_out;
 472         }
 473         pci_add_resource(&resources, &ioport_resource);
 474         pci_add_resource(&resources, &iomem_resource);
 475         pci_add_resource(&resources, &busn_res);
 476         pcifront_init_sd(sd, domain, bus, pdev);
 477
 478         pci_lock_rescan_remove();
 479
 480         b = pci_scan_root_bus(&pdev->xdev->dev, bus,
 481                                   &pcifront_bus_ops, sd, &resources);
 482         if (!b) {
 483                 dev_err(&pdev->xdev->dev,
 484                         "Error creating PCI Frontend Bus!\n");
 485                 err = -ENOMEM;
 486                 pci_unlock_rescan_remove();
 487                 pci_free_resource_list(&resources);
 488                 goto err_out;
 489         }
 490
 491         bus_entry->bus = b;
 492
 493         list_add(&bus_entry->list, &pdev->root_buses);
 494
 495         /* pci_scan_root_bus skips devices which do not have a
 496         * devfn==0. The pcifront_scan_bus enumerates all devfn. */
 497         err = pcifront_scan_bus(pdev, domain, bus, b);
 498
 499         /* Claim resources before going "live" with our devices */
 500         pci_walk_bus(b, pcifront_claim_resource, pdev);
 501
 502         /* Create SysFS and notify udev of the devices. Aka: "going live" */
 503         pci_bus_add_devices(b);
 504
 505         pci_unlock_rescan_remove();
 506         return err;
 507
 508 err_out:
 509         kfree(bus_entry);
 510         kfree(sd);
 511
 512         return err;
 513 }
 514
 515 static int pcifront_rescan_root(struct pcifront_device *pdev,
 516                                    unsigned int domain, unsigned int bus)
 517 {
 518         int err;
 519         struct pci_bus *b;
 520
 521 #ifndef CONFIG_PCI_DOMAINS
 522         if (domain != 0) {
 523                 dev_err(&pdev->xdev->dev,
 524                         "PCI Root in non-zero PCI Domain! domain=%d\n", domain);
 525                 dev_err(&pdev->xdev->dev,
 526                         "Please compile with CONFIG_PCI_DOMAINS\n");
 527                 return -EINVAL;
 528         }
 529 #endif
 530
 531         dev_info(&pdev->xdev->dev, "Rescanning PCI Frontend Bus %04x:%02x\n",
 532                  domain, bus);
 533
 534         b = pci_find_bus(domain, bus);
 535         if (!b)
 536                 /* If the bus is unknown, create it. */
 537                 return pcifront_scan_root(pdev, domain, bus);
 538
 539         err = pcifront_scan_bus(pdev, domain, bus, b);
 540
 541         /* Claim resources before going "live" with our devices */
 542         pci_walk_bus(b, pcifront_claim_resource, pdev);
 543
 544         /* Create SysFS and notify udev of the devices. Aka: "going live" */
 545         pci_bus_add_devices(b);
 546
 547         return err;
 548 }
 549
 550 static void free_root_bus_devs(struct pci_bus *bus)
 551 {
 552         struct pci_dev *dev;
 553
 554         while (!list_empty(&bus->devices)) {
 555                 dev = container_of(bus->devices.next, struct pci_dev,
 556                                    bus_list);
 557                 pci_dbg(dev, "removing device\n");
 558                 pci_stop_and_remove_bus_device(dev);
 559         }
 560 }
 561
 562 static void pcifront_free_roots(struct pcifront_device *pdev)
 563 {
 564         struct pci_bus_entry *bus_entry, *t;
 565
 566         dev_dbg(&pdev->xdev->dev, "cleaning up root buses\n");
 567
 568         pci_lock_rescan_remove();
 569         list_for_each_entry_safe(bus_entry, t, &pdev->root_buses, list) {
 570                 list_del(&bus_entry->list);
 571
 572                 free_root_bus_devs(bus_entry->bus);
 573
 574                 kfree(bus_entry->bus->sysdata);
 575
 576                 device_unregister(bus_entry->bus->bridge);
 577                 pci_remove_bus(bus_entry->bus);
 578
 579                 kfree(bus_entry);
 580         }
 581         pci_unlock_rescan_remove();
 582 }
 583
 584 static pci_ers_result_t pcifront_common_process(int cmd,
 585                                                 struct pcifront_device *pdev,
 586                                                 pci_channel_state_t state)
 587 {
 588         pci_ers_result_t result;
 589         struct pci_driver *pdrv;
 590         int bus = pdev->sh_info->aer_op.bus;
 591         int devfn = pdev->sh_info->aer_op.devfn;
 592         int domain = pdev->sh_info->aer_op.domain;
 593         struct pci_dev *pcidev;
 594         int flag = 0;
 595
 596         dev_dbg(&pdev->xdev->dev,
 597                 "pcifront AER process: cmd %x (bus:%x, devfn%x)",
 598                 cmd, bus, devfn);
 599         result = PCI_ERS_RESULT_NONE;
 600
 601         pcidev = pci_get_domain_bus_and_slot(domain, bus, devfn);
 602         if (!pcidev || !pcidev->driver) {
 603                 dev_err(&pdev->xdev->dev, "device or AER driver is NULL\n");
 604                 pci_dev_put(pcidev);
 605                 return result;
 606         }
 607         pdrv = pcidev->driver;
 608
 609         if (pdrv) {
 610                 if (pdrv->err_handler && pdrv->err_handler->error_detected) {
 611                         pci_dbg(pcidev, "trying to call AER service\n");
 612                         if (pcidev) {
 613                                 flag = 1;
 614                                 switch (cmd) {
 615                                 case XEN_PCI_OP_aer_detected:
 616                                         result = pdrv->err_handler->
 617                                                  error_detected(pcidev, state);
 618                                         break;
 619                                 case XEN_PCI_OP_aer_mmio:
 620                                         result = pdrv->err_handler->
 621                                                  mmio_enabled(pcidev);
 622                                         break;
 623                                 case XEN_PCI_OP_aer_slotreset:
 624                                         result = pdrv->err_handler->
 625                                                  slot_reset(pcidev);
 626                                         break;
 627                                 case XEN_PCI_OP_aer_resume:
 628                                         pdrv->err_handler->resume(pcidev);
 629                                         break;
 630                                 default:
 631                                         dev_err(&pdev->xdev->dev,
 632                                                 "bad request in aer recovery "
 633                                                 "operation!\n");
 634
 635                                 }
 636                         }
 637                 }
 638         }
 639         if (!flag)
 640                 result = PCI_ERS_RESULT_NONE;
 641
 642         return result;
 643 }
 644
 645
 646 static void pcifront_do_aer(struct work_struct *data)
 647 {
 648         struct pcifront_device *pdev =
 649                 container_of(data, struct pcifront_device, op_work);
 650         int cmd = pdev->sh_info->aer_op.cmd;
 651         pci_channel_state_t state =
 652                 (pci_channel_state_t)pdev->sh_info->aer_op.err;
 653
 654         /*If a pci_conf op is in progress,
 655                 we have to wait until it is done before service aer op*/
 656         dev_dbg(&pdev->xdev->dev,
 657                 "pcifront service aer bus %x devfn %x\n",
 658                 pdev->sh_info->aer_op.bus, pdev->sh_info->aer_op.devfn);
 659
 660         pdev->sh_info->aer_op.err = pcifront_common_process(cmd, pdev, state);
 661
 662         /* Post the operation to the guest. */
 663         wmb();
 664         clear_bit(_XEN_PCIB_active, (unsigned long *)&pdev->sh_info->flags);
 665         notify_remote_via_evtchn(pdev->evtchn);
 666
 667         /*in case of we lost an aer request in four lines time_window*/
 668         smp_mb__before_atomic();
 669         clear_bit(_PDEVB_op_active, &pdev->flags);
 670         smp_mb__after_atomic();
 671
 672         schedule_pcifront_aer_op(pdev);
 673
 674 }
 675
 676 static irqreturn_t pcifront_handler_aer(int irq, void *dev)
 677 {
 678         struct pcifront_device *pdev = dev;
 679         schedule_pcifront_aer_op(pdev);
 680         return IRQ_HANDLED;
 681 }
 682 static int pcifront_connect_and_init_dma(struct pcifront_device *pdev)
 683 {
 684         int err = 0;
 685
 686         spin_lock(&pcifront_dev_lock);
 687
 688         if (!pcifront_dev) {
 689                 dev_info(&pdev->xdev->dev, "Installing PCI frontend\n");
 690                 pcifront_dev = pdev;
 691         } else
 692                 err = -EEXIST;
 693
 694         spin_unlock(&pcifront_dev_lock);
 695
 696         if (!err && !is_swiotlb_active()) {
 697                 err = pci_xen_swiotlb_init_late();
 698                 if (err)
 699                         dev_err(&pdev->xdev->dev, "Could not setup SWIOTLB!\n");
 700         }
 701         return err;
 702 }
 703
 704 static void pcifront_disconnect(struct pcifront_device *pdev)
 705 {
 706         spin_lock(&pcifront_dev_lock);
 707
 708         if (pdev == pcifront_dev) {
 709                 dev_info(&pdev->xdev->dev,
 710                          "Disconnecting PCI Frontend Buses\n");
 711                 pcifront_dev = NULL;
 712         }
 713
 714         spin_unlock(&pcifront_dev_lock);
 715 }
 716 static struct pcifront_device *alloc_pdev(struct xenbus_device *xdev)
 717 {
 718         struct pcifront_device *pdev;
 719
 720         pdev = kzalloc(sizeof(struct pcifront_device), GFP_KERNEL);
 721         if (pdev == NULL)
 722                 goto out;
 723
 724         pdev->sh_info =
 725             (struct xen_pci_sharedinfo *)__get_free_page(GFP_KERNEL);
 726         if (pdev->sh_info == NULL) {
 727                 kfree(pdev);
 728                 pdev = NULL;
 729                 goto out;
 730         }
 731         pdev->sh_info->flags = 0;
 732
 733         /*Flag for registering PV AER handler*/
 734         set_bit(_XEN_PCIB_AERHANDLER, (void *)&pdev->sh_info->flags);
 735
 736         dev_set_drvdata(&xdev->dev, pdev);
 737         pdev->xdev = xdev;
 738
 739         INIT_LIST_HEAD(&pdev->root_buses);
 740
 741         spin_lock_init(&pdev->sh_info_lock);
 742
 743         pdev->evtchn = INVALID_EVTCHN;
 744         pdev->gnt_ref = INVALID_GRANT_REF;
 745         pdev->irq = -1;
 746
 747         INIT_WORK(&pdev->op_work, pcifront_do_aer);
 748
 749         dev_dbg(&xdev->dev, "Allocated pdev @ 0x%p pdev->sh_info @ 0x%p\n",
 750                 pdev, pdev->sh_info);
 751 out:
 752         return pdev;
 753 }
 754
 755 static void free_pdev(struct pcifront_device *pdev)
 756 {
 757         dev_dbg(&pdev->xdev->dev, "freeing pdev @ 0x%p\n", pdev);
 758
 759         pcifront_free_roots(pdev);
 760
 761         cancel_work_sync(&pdev->op_work);
 762
 763         if (pdev->irq >= 0)
 764                 unbind_from_irqhandler(pdev->irq, pdev);
 765
 766         if (pdev->evtchn != INVALID_EVTCHN)
 767                 xenbus_free_evtchn(pdev->xdev, pdev->evtchn);
 768
 769         if (pdev->gnt_ref != INVALID_GRANT_REF)
 770                 gnttab_end_foreign_access(pdev->gnt_ref, 0 /* r/w page */,
 771                                           (unsigned long)pdev->sh_info);
 772         else
 773                 free_page((unsigned long)pdev->sh_info);
 774
 775         dev_set_drvdata(&pdev->xdev->dev, NULL);
 776
 777         kfree(pdev);
 778 }
 779
 780 static int pcifront_publish_info(struct pcifront_device *pdev)
 781 {
 782         int err = 0;
 783         struct xenbus_transaction trans;
 784         grant_ref_t gref;
 785
 786         err = xenbus_grant_ring(pdev->xdev, pdev->sh_info, 1, &gref);
 787         if (err < 0)
 788                 goto out;
 789
 790         pdev->gnt_ref = gref;
 791
 792         err = xenbus_alloc_evtchn(pdev->xdev, &pdev->evtchn);
 793         if (err)
 794                 goto out;
 795
 796         err = bind_evtchn_to_irqhandler(pdev->evtchn, pcifront_handler_aer,
 797                 0, "pcifront", pdev);
 798
 799         if (err < 0)
 800                 return err;
 801
 802         pdev->irq = err;
 803
 804 do_publish:
 805         err = xenbus_transaction_start(&trans);
 806         if (err) {
 807                 xenbus_dev_fatal(pdev->xdev, err,
 808                                  "Error writing configuration for backend "
 809                                  "(start transaction)");
 810                 goto out;
 811         }
 812
 813         err = xenbus_printf(trans, pdev->xdev->nodename,
 814                             "pci-op-ref", "%u", pdev->gnt_ref);
 815         if (!err)
 816                 err = xenbus_printf(trans, pdev->xdev->nodename,
 817                                     "event-channel", "%u", pdev->evtchn);
 818         if (!err)
 819                 err = xenbus_printf(trans, pdev->xdev->nodename,
 820                                     "magic", XEN_PCI_MAGIC);
 821
 822         if (err) {
 823                 xenbus_transaction_end(trans, 1);
 824                 xenbus_dev_fatal(pdev->xdev, err,
 825                                  "Error writing configuration for backend");
 826                 goto out;
 827         } else {
 828                 err = xenbus_transaction_end(trans, 0);
 829                 if (err == -EAGAIN)
 830                         goto do_publish;
 831                 else if (err) {
 832                         xenbus_dev_fatal(pdev->xdev, err,
 833                                          "Error completing transaction "
 834                                          "for backend");
 835                         goto out;
 836                 }
 837         }
 838
 839         xenbus_switch_state(pdev->xdev, XenbusStateInitialised);
 840
 841         dev_dbg(&pdev->xdev->dev, "publishing successful!\n");
 842
 843 out:
 844         return err;
 845 }
 846
 847 static int pcifront_try_connect(struct pcifront_device *pdev)
 848 {
 849         int err = -EFAULT;
 850         int i, num_roots, len;
 851         char str[64];
 852         unsigned int domain, bus;
 853
 854
 855         /* Only connect once */
 856         if (xenbus_read_driver_state(pdev->xdev->nodename) !=
 857             XenbusStateInitialised)
 858                 goto out;
 859
 860         err = pcifront_connect_and_init_dma(pdev);
 861         if (err && err != -EEXIST) {
 862                 xenbus_dev_fatal(pdev->xdev, err,
 863                                  "Error setting up PCI Frontend");
 864                 goto out;
 865         }
 866
 867         err = xenbus_scanf(XBT_NIL, pdev->xdev->otherend,
 868                            "root_num", "%d", &num_roots);
 869         if (err == -ENOENT) {
 870                 xenbus_dev_error(pdev->xdev, err,
 871                                  "No PCI Roots found, trying 0000:00");
 872                 err = pcifront_scan_root(pdev, 0, 0);
 873                 if (err) {
 874                         xenbus_dev_fatal(pdev->xdev, err,
 875                                          "Error scanning PCI root 0000:00");
 876                         goto out;
 877                 }
 878                 num_roots = 0;
 879         } else if (err != 1) {
 880                 if (err == 0)
 881                         err = -EINVAL;
 882                 xenbus_dev_fatal(pdev->xdev, err,
 883                                  "Error reading number of PCI roots");
 884                 goto out;
 885         }
 886
 887         for (i = 0; i < num_roots; i++) {
 888                 len = snprintf(str, sizeof(str), "root-%d", i);
 889                 if (unlikely(len >= (sizeof(str) - 1))) {
 890                         err = -ENOMEM;
 891                         goto out;
 892                 }
 893
 894                 err = xenbus_scanf(XBT_NIL, pdev->xdev->otherend, str,
 895                                    "%x:%x", &domain, &bus);
 896                 if (err != 2) {
 897                         if (err >= 0)
 898                                 err = -EINVAL;
 899                         xenbus_dev_fatal(pdev->xdev, err,
 900                                          "Error reading PCI root %d", i);
 901                         goto out;
 902                 }
 903
 904                 err = pcifront_scan_root(pdev, domain, bus);
 905                 if (err) {
 906                         xenbus_dev_fatal(pdev->xdev, err,
 907                                          "Error scanning PCI root %04x:%02x",
 908                                          domain, bus);
 909                         goto out;
 910                 }
 911         }
 912
 913         err = xenbus_switch_state(pdev->xdev, XenbusStateConnected);
 914
 915 out:
 916         return err;
 917 }
 918
 919 static int pcifront_try_disconnect(struct pcifront_device *pdev)
 920 {
 921         int err = 0;
 922         enum xenbus_state prev_state;
 923
 924
 925         prev_state = xenbus_read_driver_state(pdev->xdev->nodename);
 926
 927         if (prev_state >= XenbusStateClosing)
 928                 goto out;
 929
 930         if (prev_state == XenbusStateConnected) {
 931                 pcifront_free_roots(pdev);
 932                 pcifront_disconnect(pdev);
 933         }
 934
 935         err = xenbus_switch_state(pdev->xdev, XenbusStateClosed);
 936
 937 out:
 938
 939         return err;
 940 }
 941
 942 static int pcifront_attach_devices(struct pcifront_device *pdev)
 943 {
 944         int err = -EFAULT;
 945         int i, num_roots, len;
 946         unsigned int domain, bus;
 947         char str[64];
 948
 949         if (xenbus_read_driver_state(pdev->xdev->nodename) !=
 950             XenbusStateReconfiguring)
 951                 goto out;
 952
 953         err = xenbus_scanf(XBT_NIL, pdev->xdev->otherend,
 954                            "root_num", "%d", &num_roots);
 955         if (err == -ENOENT) {
 956                 xenbus_dev_error(pdev->xdev, err,
 957                                  "No PCI Roots found, trying 0000:00");
 958                 err = pcifront_rescan_root(pdev, 0, 0);
 959                 if (err) {
 960                         xenbus_dev_fatal(pdev->xdev, err,
 961                                          "Error scanning PCI root 0000:00");
 962                         goto out;
 963                 }
 964                 num_roots = 0;
 965         } else if (err != 1) {
 966                 if (err == 0)
 967                         err = -EINVAL;
 968                 xenbus_dev_fatal(pdev->xdev, err,
 969                                  "Error reading number of PCI roots");
 970                 goto out;
 971         }
 972
 973         for (i = 0; i < num_roots; i++) {
 974                 len = snprintf(str, sizeof(str), "root-%d", i);
 975                 if (unlikely(len >= (sizeof(str) - 1))) {
 976                         err = -ENOMEM;
 977                         goto out;
 978                 }
 979
 980                 err = xenbus_scanf(XBT_NIL, pdev->xdev->otherend, str,
 981                                    "%x:%x", &domain, &bus);
 982                 if (err != 2) {
 983                         if (err >= 0)
 984                                 err = -EINVAL;
 985                         xenbus_dev_fatal(pdev->xdev, err,
 986                                          "Error reading PCI root %d", i);
 987                         goto out;
 988                 }
 989
 990                 err = pcifront_rescan_root(pdev, domain, bus);
 991                 if (err) {
 992                         xenbus_dev_fatal(pdev->xdev, err,
 993                                          "Error scanning PCI root %04x:%02x",
 994                                          domain, bus);
 995                         goto out;
 996                 }
 997         }
 998
 999         xenbus_switch_state(pdev->xdev, XenbusStateConnected);
1000
1001 out:
1002         return err;
1003 }
1004
1005 static int pcifront_detach_devices(struct pcifront_device *pdev)
1006 {
1007         int err = 0;
1008         int i, num_devs;
1009         unsigned int domain, bus, slot, func;
1010         struct pci_dev *pci_dev;
1011         char str[64];
1012
1013         if (xenbus_read_driver_state(pdev->xdev->nodename) !=
1014             XenbusStateConnected)
1015                 goto out;
1016
1017         err = xenbus_scanf(XBT_NIL, pdev->xdev->otherend, "num_devs", "%d",
1018                            &num_devs);
1019         if (err != 1) {
1020                 if (err >= 0)
1021                         err = -EINVAL;
1022                 xenbus_dev_fatal(pdev->xdev, err,
1023                                  "Error reading number of PCI devices");
1024                 goto out;
1025         }
1026
1027         /* Find devices being detached and remove them. */
1028         for (i = 0; i < num_devs; i++) {
1029                 int l, state;
1030                 l = snprintf(str, sizeof(str), "state-%d", i);
1031                 if (unlikely(l >= (sizeof(str) - 1))) {
1032                         err = -ENOMEM;
1033                         goto out;
1034                 }
1035                 state = xenbus_read_unsigned(pdev->xdev->otherend, str,
1036                                              XenbusStateUnknown);
1037
1038                 if (state != XenbusStateClosing)
1039                         continue;
1040
1041                 /* Remove device. */
1042                 l = snprintf(str, sizeof(str), "vdev-%d", i);
1043                 if (unlikely(l >= (sizeof(str) - 1))) {
1044                         err = -ENOMEM;
1045                         goto out;
1046                 }
1047                 err = xenbus_scanf(XBT_NIL, pdev->xdev->otherend, str,
1048                                    "%x:%x:%x.%x", &domain, &bus, &slot, &func);
1049                 if (err != 4) {
1050                         if (err >= 0)
1051                                 err = -EINVAL;
1052                         xenbus_dev_fatal(pdev->xdev, err,
1053                                          "Error reading PCI device %d", i);
1054                         goto out;
1055                 }
1056
1057                 pci_dev = pci_get_domain_bus_and_slot(domain, bus,
1058                                 PCI_DEVFN(slot, func));
1059                 if (!pci_dev) {
1060                         dev_dbg(&pdev->xdev->dev,
1061                                 "Cannot get PCI device %04x:%02x:%02x.%d\n",
1062                                 domain, bus, slot, func);
1063                         continue;
1064                 }
1065                 pci_lock_rescan_remove();
1066                 pci_stop_and_remove_bus_device(pci_dev);
1067                 pci_dev_put(pci_dev);
1068                 pci_unlock_rescan_remove();
1069
1070                 dev_dbg(&pdev->xdev->dev,
1071                         "PCI device %04x:%02x:%02x.%d removed.\n",
1072                         domain, bus, slot, func);
1073         }
1074
1075         err = xenbus_switch_state(pdev->xdev, XenbusStateReconfiguring);
1076
1077 out:
1078         return err;
1079 }
1080
1081 static void __ref pcifront_backend_changed(struct xenbus_device *xdev,
1082                                                   enum xenbus_state be_state)
1083 {
1084         struct pcifront_device *pdev = dev_get_drvdata(&xdev->dev);
1085
1086         switch (be_state) {
1087         case XenbusStateUnknown:
1088         case XenbusStateInitialising:
1089         case XenbusStateInitWait:
1090         case XenbusStateInitialised:
1091                 break;
1092
1093         case XenbusStateConnected:
1094                 pcifront_try_connect(pdev);
1095                 break;
1096
1097         case XenbusStateClosed:
1098                 if (xdev->state == XenbusStateClosed)
1099                         break;
1100                 fallthrough;    /* Missed the backend's CLOSING state */
1101         case XenbusStateClosing:
1102                 dev_warn(&xdev->dev, "backend going away!\n");
1103                 pcifront_try_disconnect(pdev);
1104                 break;
1105
1106         case XenbusStateReconfiguring:
1107                 pcifront_detach_devices(pdev);
1108                 break;
1109
1110         case XenbusStateReconfigured:
1111                 pcifront_attach_devices(pdev);
1112                 break;
1113         }
1114 }
1115
1116 static int pcifront_xenbus_probe(struct xenbus_device *xdev,
1117                                  const struct xenbus_device_id *id)
1118 {
1119         int err = 0;
1120         struct pcifront_device *pdev = alloc_pdev(xdev);
1121
1122         if (pdev == NULL) {
1123                 err = -ENOMEM;
1124                 xenbus_dev_fatal(xdev, err,
1125                                  "Error allocating pcifront_device struct");
1126                 goto out;
1127         }
1128
1129         err = pcifront_publish_info(pdev);
1130         if (err)
1131                 free_pdev(pdev);
1132
1133 out:
1134         return err;
1135 }
1136
1137 static int pcifront_xenbus_remove(struct xenbus_device *xdev)
1138 {
1139         struct pcifront_device *pdev = dev_get_drvdata(&xdev->dev);
1140         if (pdev)
1141                 free_pdev(pdev);
1142
1143         return 0;
1144 }
1145
1146 static const struct xenbus_device_id xenpci_ids[] = {
1147         {"pci"},
1148         {""},
1149 };
1150
1151 static struct xenbus_driver xenpci_driver = {
1152         .name                   = "pcifront",
1153         .ids                    = xenpci_ids,
1154         .probe                  = pcifront_xenbus_probe,
1155         .remove                 = pcifront_xenbus_remove,
1156         .otherend_changed       = pcifront_backend_changed,
1157 };
1158
1159 static int __init pcifront_init(void)
1160 {
1161         if (!xen_pv_domain() || xen_initial_domain())
1162                 return -ENODEV;
1163
1164         if (!xen_has_pv_devices())
1165                 return -ENODEV;
1166
1167         pci_frontend_registrar(1 /* enable */);
1168
1169         return xenbus_register_frontend(&xenpci_driver);
1170 }
1171
1172 static void __exit pcifront_cleanup(void)
1173 {
1174         xenbus_unregister_driver(&xenpci_driver);
1175         pci_frontend_registrar(0 /* disable */);
1176 }
1177 module_init(pcifront_init);
1178 module_exit(pcifront_cleanup);
1179
1180 MODULE_DESCRIPTION("Xen PCI passthrough frontend.");
1181 MODULE_LICENSE("GPL");
1182 MODULE_ALIAS("xen:pci");