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 int 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         /*
 418          * Scan the bus for functions and add.
 419          * We omit handling of PCI bridge attachment because pciback prevents
 420          * bridges from being exported.
 421          */
 422         for (devfn = 0; devfn < 0x100; devfn++) {
 423                 d = pci_get_slot(b, devfn);
 424                 if (d) {
 425                         /* Device is already known. */
 426                         pci_dev_put(d);
 427                         continue;
 428                 }
 429
 430                 d = pci_scan_single_device(b, devfn);
 431                 if (d)
 432                         dev_info(&pdev->xdev->dev, "New device on "
 433                                  "%04x:%02x:%02x.%d found.\n", domain, bus,
 434                                  PCI_SLOT(devfn), PCI_FUNC(devfn));
 435         }
 436
 437         return 0;
 438 }
 439
 440 static int pcifront_scan_root(struct pcifront_device *pdev,
 441                                  unsigned int domain, unsigned int bus)
 442 {
 443         struct pci_bus *b;
 444         LIST_HEAD(resources);
 445         struct pcifront_sd *sd = NULL;
 446         struct pci_bus_entry *bus_entry = NULL;
 447         int err = 0;
 448         static struct resource busn_res = {
 449                 .start = 0,
 450                 .end = 255,
 451                 .flags = IORESOURCE_BUS,
 452         };
 453
 454 #ifndef CONFIG_PCI_DOMAINS
 455         if (domain != 0) {
 456                 dev_err(&pdev->xdev->dev,
 457                         "PCI Root in non-zero PCI Domain! domain=%d\n", domain);
 458                 dev_err(&pdev->xdev->dev,
 459                         "Please compile with CONFIG_PCI_DOMAINS\n");
 460                 err = -EINVAL;
 461                 goto err_out;
 462         }
 463 #endif
 464
 465         dev_info(&pdev->xdev->dev, "Creating PCI Frontend Bus %04x:%02x\n",
 466                  domain, bus);
 467
 468         bus_entry = kzalloc(sizeof(*bus_entry), GFP_KERNEL);
 469         sd = kzalloc(sizeof(*sd), GFP_KERNEL);
 470         if (!bus_entry || !sd) {
 471                 err = -ENOMEM;
 472                 goto err_out;
 473         }
 474         pci_add_resource(&resources, &ioport_resource);
 475         pci_add_resource(&resources, &iomem_resource);
 476         pci_add_resource(&resources, &busn_res);
 477         pcifront_init_sd(sd, domain, bus, pdev);
 478
 479         pci_lock_rescan_remove();
 480
 481         b = pci_scan_root_bus(&pdev->xdev->dev, bus,
 482                                   &pcifront_bus_ops, sd, &resources);
 483         if (!b) {
 484                 dev_err(&pdev->xdev->dev,
 485                         "Error creating PCI Frontend Bus!\n");
 486                 err = -ENOMEM;
 487                 pci_unlock_rescan_remove();
 488                 pci_free_resource_list(&resources);
 489                 goto err_out;
 490         }
 491
 492         bus_entry->bus = b;
 493
 494         list_add(&bus_entry->list, &pdev->root_buses);
 495
 496         /*
 497          * pci_scan_root_bus skips devices which do not have a
 498          * devfn==0. The pcifront_scan_bus enumerates all devfn.
 499          */
 500         err = pcifront_scan_bus(pdev, domain, bus, b);
 501
 502         /* Claim resources before going "live" with our devices */
 503         pci_walk_bus(b, pcifront_claim_resource, pdev);
 504
 505         /* Create SysFS and notify udev of the devices. Aka: "going live" */
 506         pci_bus_add_devices(b);
 507
 508         pci_unlock_rescan_remove();
 509         return err;
 510
 511 err_out:
 512         kfree(bus_entry);
 513         kfree(sd);
 514
 515         return err;
 516 }
 517
 518 static int pcifront_rescan_root(struct pcifront_device *pdev,
 519                                    unsigned int domain, unsigned int bus)
 520 {
 521         int err;
 522         struct pci_bus *b;
 523
 524 #ifndef CONFIG_PCI_DOMAINS
 525         if (domain != 0) {
 526                 dev_err(&pdev->xdev->dev,
 527                         "PCI Root in non-zero PCI Domain! domain=%d\n", domain);
 528                 dev_err(&pdev->xdev->dev,
 529                         "Please compile with CONFIG_PCI_DOMAINS\n");
 530                 return -EINVAL;
 531         }
 532 #endif
 533
 534         dev_info(&pdev->xdev->dev, "Rescanning PCI Frontend Bus %04x:%02x\n",
 535                  domain, bus);
 536
 537         b = pci_find_bus(domain, bus);
 538         if (!b)
 539                 /* If the bus is unknown, create it. */
 540                 return pcifront_scan_root(pdev, domain, bus);
 541
 542         err = pcifront_scan_bus(pdev, domain, bus, b);
 543
 544         /* Claim resources before going "live" with our devices */
 545         pci_walk_bus(b, pcifront_claim_resource, pdev);
 546
 547         /* Create SysFS and notify udev of the devices. Aka: "going live" */
 548         pci_bus_add_devices(b);
 549
 550         return err;
 551 }
 552
 553 static void free_root_bus_devs(struct pci_bus *bus)
 554 {
 555         struct pci_dev *dev;
 556
 557         while (!list_empty(&bus->devices)) {
 558                 dev = container_of(bus->devices.next, struct pci_dev,
 559                                    bus_list);
 560                 pci_dbg(dev, "removing device\n");
 561                 pci_stop_and_remove_bus_device(dev);
 562         }
 563 }
 564
 565 static void pcifront_free_roots(struct pcifront_device *pdev)
 566 {
 567         struct pci_bus_entry *bus_entry, *t;
 568
 569         dev_dbg(&pdev->xdev->dev, "cleaning up root buses\n");
 570
 571         pci_lock_rescan_remove();
 572         list_for_each_entry_safe(bus_entry, t, &pdev->root_buses, list) {
 573                 list_del(&bus_entry->list);
 574
 575                 free_root_bus_devs(bus_entry->bus);
 576
 577                 kfree(bus_entry->bus->sysdata);
 578
 579                 device_unregister(bus_entry->bus->bridge);
 580                 pci_remove_bus(bus_entry->bus);
 581
 582                 kfree(bus_entry);
 583         }
 584         pci_unlock_rescan_remove();
 585 }
 586
 587 static pci_ers_result_t pcifront_common_process(int cmd,
 588                                                 struct pcifront_device *pdev,
 589                                                 pci_channel_state_t state)
 590 {
 591         pci_ers_result_t result;
 592         struct pci_driver *pdrv;
 593         int bus = pdev->sh_info->aer_op.bus;
 594         int devfn = pdev->sh_info->aer_op.devfn;
 595         int domain = pdev->sh_info->aer_op.domain;
 596         struct pci_dev *pcidev;
 597         int flag = 0;
 598
 599         dev_dbg(&pdev->xdev->dev,
 600                 "pcifront AER process: cmd %x (bus:%x, devfn%x)",
 601                 cmd, bus, devfn);
 602         result = PCI_ERS_RESULT_NONE;
 603
 604         pcidev = pci_get_domain_bus_and_slot(domain, bus, devfn);
 605         if (!pcidev || !pcidev->driver) {
 606                 dev_err(&pdev->xdev->dev, "device or AER driver is NULL\n");
 607                 pci_dev_put(pcidev);
 608                 return result;
 609         }
 610         pdrv = pcidev->driver;
 611
 612         if (pdrv) {
 613                 if (pdrv->err_handler && pdrv->err_handler->error_detected) {
 614                         pci_dbg(pcidev, "trying to call AER service\n");
 615                         if (pcidev) {
 616                                 flag = 1;
 617                                 switch (cmd) {
 618                                 case XEN_PCI_OP_aer_detected:
 619                                         result = pdrv->err_handler->
 620                                                  error_detected(pcidev, state);
 621                                         break;
 622                                 case XEN_PCI_OP_aer_mmio:
 623                                         result = pdrv->err_handler->
 624                                                  mmio_enabled(pcidev);
 625                                         break;
 626                                 case XEN_PCI_OP_aer_slotreset:
 627                                         result = pdrv->err_handler->
 628                                                  slot_reset(pcidev);
 629                                         break;
 630                                 case XEN_PCI_OP_aer_resume:
 631                                         pdrv->err_handler->resume(pcidev);
 632                                         break;
 633                                 default:
 634                                         dev_err(&pdev->xdev->dev,
 635                                                 "bad request in aer recovery "
 636                                                 "operation!\n");
 637
 638                                 }
 639                         }
 640                 }
 641         }
 642         if (!flag)
 643                 result = PCI_ERS_RESULT_NONE;
 644
 645         return result;
 646 }
 647
 648
 649 static void pcifront_do_aer(struct work_struct *data)
 650 {
 651         struct pcifront_device *pdev =
 652                 container_of(data, struct pcifront_device, op_work);
 653         int cmd = pdev->sh_info->aer_op.cmd;
 654         pci_channel_state_t state =
 655                 (pci_channel_state_t)pdev->sh_info->aer_op.err;
 656
 657         /*
 658          * If a pci_conf op is in progress, we have to wait until it is done
 659          * before service aer op
 660          */
 661         dev_dbg(&pdev->xdev->dev,
 662                 "pcifront service aer bus %x devfn %x\n",
 663                 pdev->sh_info->aer_op.bus, pdev->sh_info->aer_op.devfn);
 664
 665         pdev->sh_info->aer_op.err = pcifront_common_process(cmd, pdev, state);
 666
 667         /* Post the operation to the guest. */
 668         wmb();
 669         clear_bit(_XEN_PCIB_active, (unsigned long *)&pdev->sh_info->flags);
 670         notify_remote_via_evtchn(pdev->evtchn);
 671
 672         /*in case of we lost an aer request in four lines time_window*/
 673         smp_mb__before_atomic();
 674         clear_bit(_PDEVB_op_active, &pdev->flags);
 675         smp_mb__after_atomic();
 676
 677         schedule_pcifront_aer_op(pdev);
 678
 679 }
 680
 681 static irqreturn_t pcifront_handler_aer(int irq, void *dev)
 682 {
 683         struct pcifront_device *pdev = dev;
 684
 685         schedule_pcifront_aer_op(pdev);
 686         return IRQ_HANDLED;
 687 }
 688 static int pcifront_connect_and_init_dma(struct pcifront_device *pdev)
 689 {
 690         int err = 0;
 691
 692         spin_lock(&pcifront_dev_lock);
 693
 694         if (!pcifront_dev) {
 695                 dev_info(&pdev->xdev->dev, "Installing PCI frontend\n");
 696                 pcifront_dev = pdev;
 697         } else
 698                 err = -EEXIST;
 699
 700         spin_unlock(&pcifront_dev_lock);
 701
 702         if (!err && !is_swiotlb_active(&pdev->xdev->dev)) {
 703                 err = pci_xen_swiotlb_init_late();
 704                 if (err)
 705                         dev_err(&pdev->xdev->dev, "Could not setup SWIOTLB!\n");
 706         }
 707         return err;
 708 }
 709
 710 static void pcifront_disconnect(struct pcifront_device *pdev)
 711 {
 712         spin_lock(&pcifront_dev_lock);
 713
 714         if (pdev == pcifront_dev) {
 715                 dev_info(&pdev->xdev->dev,
 716                          "Disconnecting PCI Frontend Buses\n");
 717                 pcifront_dev = NULL;
 718         }
 719
 720         spin_unlock(&pcifront_dev_lock);
 721 }
 722 static struct pcifront_device *alloc_pdev(struct xenbus_device *xdev)
 723 {
 724         struct pcifront_device *pdev;
 725
 726         pdev = kzalloc(sizeof(struct pcifront_device), GFP_KERNEL);
 727         if (pdev == NULL)
 728                 goto out;
 729
 730         pdev->sh_info =
 731             (struct xen_pci_sharedinfo *)__get_free_page(GFP_KERNEL);
 732         if (pdev->sh_info == NULL) {
 733                 kfree(pdev);
 734                 pdev = NULL;
 735                 goto out;
 736         }
 737         pdev->sh_info->flags = 0;
 738
 739         /*Flag for registering PV AER handler*/
 740         set_bit(_XEN_PCIB_AERHANDLER, (void *)&pdev->sh_info->flags);
 741
 742         dev_set_drvdata(&xdev->dev, pdev);
 743         pdev->xdev = xdev;
 744
 745         INIT_LIST_HEAD(&pdev->root_buses);
 746
 747         spin_lock_init(&pdev->sh_info_lock);
 748
 749         pdev->evtchn = INVALID_EVTCHN;
 750         pdev->gnt_ref = INVALID_GRANT_REF;
 751         pdev->irq = -1;
 752
 753         INIT_WORK(&pdev->op_work, pcifront_do_aer);
 754
 755         dev_dbg(&xdev->dev, "Allocated pdev @ 0x%p pdev->sh_info @ 0x%p\n",
 756                 pdev, pdev->sh_info);
 757 out:
 758         return pdev;
 759 }
 760
 761 static void free_pdev(struct pcifront_device *pdev)
 762 {
 763         dev_dbg(&pdev->xdev->dev, "freeing pdev @ 0x%p\n", pdev);
 764
 765         pcifront_free_roots(pdev);
 766
 767         cancel_work_sync(&pdev->op_work);
 768
 769         if (pdev->irq >= 0)
 770                 unbind_from_irqhandler(pdev->irq, pdev);
 771
 772         if (pdev->evtchn != INVALID_EVTCHN)
 773                 xenbus_free_evtchn(pdev->xdev, pdev->evtchn);
 774
 775         if (pdev->gnt_ref != INVALID_GRANT_REF)
 776                 gnttab_end_foreign_access(pdev->gnt_ref, 0 /* r/w page */,
 777                                           (unsigned long)pdev->sh_info);
 778         else
 779                 free_page((unsigned long)pdev->sh_info);
 780
 781         dev_set_drvdata(&pdev->xdev->dev, NULL);
 782
 783         kfree(pdev);
 784 }
 785
 786 static int pcifront_publish_info(struct pcifront_device *pdev)
 787 {
 788         int err = 0;
 789         struct xenbus_transaction trans;
 790         grant_ref_t gref;
 791
 792         err = xenbus_grant_ring(pdev->xdev, pdev->sh_info, 1, &gref);
 793         if (err < 0)
 794                 goto out;
 795
 796         pdev->gnt_ref = gref;
 797
 798         err = xenbus_alloc_evtchn(pdev->xdev, &pdev->evtchn);
 799         if (err)
 800                 goto out;
 801
 802         err = bind_evtchn_to_irqhandler(pdev->evtchn, pcifront_handler_aer,
 803                 0, "pcifront", pdev);
 804
 805         if (err < 0)
 806                 return err;
 807
 808         pdev->irq = err;
 809
 810 do_publish:
 811         err = xenbus_transaction_start(&trans);
 812         if (err) {
 813                 xenbus_dev_fatal(pdev->xdev, err,
 814                                  "Error writing configuration for backend "
 815                                  "(start transaction)");
 816                 goto out;
 817         }
 818
 819         err = xenbus_printf(trans, pdev->xdev->nodename,
 820                             "pci-op-ref", "%u", pdev->gnt_ref);
 821         if (!err)
 822                 err = xenbus_printf(trans, pdev->xdev->nodename,
 823                                     "event-channel", "%u", pdev->evtchn);
 824         if (!err)
 825                 err = xenbus_printf(trans, pdev->xdev->nodename,
 826                                     "magic", XEN_PCI_MAGIC);
 827
 828         if (err) {
 829                 xenbus_transaction_end(trans, 1);
 830                 xenbus_dev_fatal(pdev->xdev, err,
 831                                  "Error writing configuration for backend");
 832                 goto out;
 833         } else {
 834                 err = xenbus_transaction_end(trans, 0);
 835                 if (err == -EAGAIN)
 836                         goto do_publish;
 837                 else if (err) {
 838                         xenbus_dev_fatal(pdev->xdev, err,
 839                                          "Error completing transaction "
 840                                          "for backend");
 841                         goto out;
 842                 }
 843         }
 844
 845         xenbus_switch_state(pdev->xdev, XenbusStateInitialised);
 846
 847         dev_dbg(&pdev->xdev->dev, "publishing successful!\n");
 848
 849 out:
 850         return err;
 851 }
 852
 853 static int pcifront_try_connect(struct pcifront_device *pdev)
 854 {
 855         int err = -EFAULT;
 856         int i, num_roots, len;
 857         char str[64];
 858         unsigned int domain, bus;
 859
 860
 861         /* Only connect once */
 862         if (xenbus_read_driver_state(pdev->xdev->nodename) !=
 863             XenbusStateInitialised)
 864                 goto out;
 865
 866         err = pcifront_connect_and_init_dma(pdev);
 867         if (err && err != -EEXIST) {
 868                 xenbus_dev_fatal(pdev->xdev, err,
 869                                  "Error setting up PCI Frontend");
 870                 goto out;
 871         }
 872
 873         err = xenbus_scanf(XBT_NIL, pdev->xdev->otherend,
 874                            "root_num", "%d", &num_roots);
 875         if (err == -ENOENT) {
 876                 xenbus_dev_error(pdev->xdev, err,
 877                                  "No PCI Roots found, trying 0000:00");
 878                 err = pcifront_scan_root(pdev, 0, 0);
 879                 if (err) {
 880                         xenbus_dev_fatal(pdev->xdev, err,
 881                                          "Error scanning PCI root 0000:00");
 882                         goto out;
 883                 }
 884                 num_roots = 0;
 885         } else if (err != 1) {
 886                 if (err == 0)
 887                         err = -EINVAL;
 888                 xenbus_dev_fatal(pdev->xdev, err,
 889                                  "Error reading number of PCI roots");
 890                 goto out;
 891         }
 892
 893         for (i = 0; i < num_roots; i++) {
 894                 len = snprintf(str, sizeof(str), "root-%d", i);
 895                 if (unlikely(len >= (sizeof(str) - 1))) {
 896                         err = -ENOMEM;
 897                         goto out;
 898                 }
 899
 900                 err = xenbus_scanf(XBT_NIL, pdev->xdev->otherend, str,
 901                                    "%x:%x", &domain, &bus);
 902                 if (err != 2) {
 903                         if (err >= 0)
 904                                 err = -EINVAL;
 905                         xenbus_dev_fatal(pdev->xdev, err,
 906                                          "Error reading PCI root %d", i);
 907                         goto out;
 908                 }
 909
 910                 err = pcifront_scan_root(pdev, domain, bus);
 911                 if (err) {
 912                         xenbus_dev_fatal(pdev->xdev, err,
 913                                          "Error scanning PCI root %04x:%02x",
 914                                          domain, bus);
 915                         goto out;
 916                 }
 917         }
 918
 919         err = xenbus_switch_state(pdev->xdev, XenbusStateConnected);
 920
 921 out:
 922         return err;
 923 }
 924
 925 static int pcifront_try_disconnect(struct pcifront_device *pdev)
 926 {
 927         int err = 0;
 928         enum xenbus_state prev_state;
 929
 930
 931         prev_state = xenbus_read_driver_state(pdev->xdev->nodename);
 932
 933         if (prev_state >= XenbusStateClosing)
 934                 goto out;
 935
 936         if (prev_state == XenbusStateConnected) {
 937                 pcifront_free_roots(pdev);
 938                 pcifront_disconnect(pdev);
 939         }
 940
 941         err = xenbus_switch_state(pdev->xdev, XenbusStateClosed);
 942
 943 out:
 944
 945         return err;
 946 }
 947
 948 static int pcifront_attach_devices(struct pcifront_device *pdev)
 949 {
 950         int err = -EFAULT;
 951         int i, num_roots, len;
 952         unsigned int domain, bus;
 953         char str[64];
 954
 955         if (xenbus_read_driver_state(pdev->xdev->nodename) !=
 956             XenbusStateReconfiguring)
 957                 goto out;
 958
 959         err = xenbus_scanf(XBT_NIL, pdev->xdev->otherend,
 960                            "root_num", "%d", &num_roots);
 961         if (err == -ENOENT) {
 962                 xenbus_dev_error(pdev->xdev, err,
 963                                  "No PCI Roots found, trying 0000:00");
 964                 err = pcifront_rescan_root(pdev, 0, 0);
 965                 if (err) {
 966                         xenbus_dev_fatal(pdev->xdev, err,
 967                                          "Error scanning PCI root 0000:00");
 968                         goto out;
 969                 }
 970                 num_roots = 0;
 971         } else if (err != 1) {
 972                 if (err == 0)
 973                         err = -EINVAL;
 974                 xenbus_dev_fatal(pdev->xdev, err,
 975                                  "Error reading number of PCI roots");
 976                 goto out;
 977         }
 978
 979         for (i = 0; i < num_roots; i++) {
 980                 len = snprintf(str, sizeof(str), "root-%d", i);
 981                 if (unlikely(len >= (sizeof(str) - 1))) {
 982                         err = -ENOMEM;
 983                         goto out;
 984                 }
 985
 986                 err = xenbus_scanf(XBT_NIL, pdev->xdev->otherend, str,
 987                                    "%x:%x", &domain, &bus);
 988                 if (err != 2) {
 989                         if (err >= 0)
 990                                 err = -EINVAL;
 991                         xenbus_dev_fatal(pdev->xdev, err,
 992                                          "Error reading PCI root %d", i);
 993                         goto out;
 994                 }
 995
 996                 err = pcifront_rescan_root(pdev, domain, bus);
 997                 if (err) {
 998                         xenbus_dev_fatal(pdev->xdev, err,
 999                                          "Error scanning PCI root %04x:%02x",
1000                                          domain, bus);
1001                         goto out;
1002                 }
1003         }
1004
1005         xenbus_switch_state(pdev->xdev, XenbusStateConnected);
1006
1007 out:
1008         return err;
1009 }
1010
1011 static int pcifront_detach_devices(struct pcifront_device *pdev)
1012 {
1013         int err = 0;
1014         int i, num_devs;
1015         unsigned int domain, bus, slot, func;
1016         struct pci_dev *pci_dev;
1017         char str[64];
1018
1019         if (xenbus_read_driver_state(pdev->xdev->nodename) !=
1020             XenbusStateConnected)
1021                 goto out;
1022
1023         err = xenbus_scanf(XBT_NIL, pdev->xdev->otherend, "num_devs", "%d",
1024                            &num_devs);
1025         if (err != 1) {
1026                 if (err >= 0)
1027                         err = -EINVAL;
1028                 xenbus_dev_fatal(pdev->xdev, err,
1029                                  "Error reading number of PCI devices");
1030                 goto out;
1031         }
1032
1033         /* Find devices being detached and remove them. */
1034         for (i = 0; i < num_devs; i++) {
1035                 int l, state;
1036
1037                 l = snprintf(str, sizeof(str), "state-%d", i);
1038                 if (unlikely(l >= (sizeof(str) - 1))) {
1039                         err = -ENOMEM;
1040                         goto out;
1041                 }
1042                 state = xenbus_read_unsigned(pdev->xdev->otherend, str,
1043                                              XenbusStateUnknown);
1044
1045                 if (state != XenbusStateClosing)
1046                         continue;
1047
1048                 /* Remove device. */
1049                 l = snprintf(str, sizeof(str), "vdev-%d", i);
1050                 if (unlikely(l >= (sizeof(str) - 1))) {
1051                         err = -ENOMEM;
1052                         goto out;
1053                 }
1054                 err = xenbus_scanf(XBT_NIL, pdev->xdev->otherend, str,
1055                                    "%x:%x:%x.%x", &domain, &bus, &slot, &func);
1056                 if (err != 4) {
1057                         if (err >= 0)
1058                                 err = -EINVAL;
1059                         xenbus_dev_fatal(pdev->xdev, err,
1060                                          "Error reading PCI device %d", i);
1061                         goto out;
1062                 }
1063
1064                 pci_dev = pci_get_domain_bus_and_slot(domain, bus,
1065                                 PCI_DEVFN(slot, func));
1066                 if (!pci_dev) {
1067                         dev_dbg(&pdev->xdev->dev,
1068                                 "Cannot get PCI device %04x:%02x:%02x.%d\n",
1069                                 domain, bus, slot, func);
1070                         continue;
1071                 }
1072                 pci_lock_rescan_remove();
1073                 pci_stop_and_remove_bus_device(pci_dev);
1074                 pci_dev_put(pci_dev);
1075                 pci_unlock_rescan_remove();
1076
1077                 dev_dbg(&pdev->xdev->dev,
1078                         "PCI device %04x:%02x:%02x.%d removed.\n",
1079                         domain, bus, slot, func);
1080         }
1081
1082         err = xenbus_switch_state(pdev->xdev, XenbusStateReconfiguring);
1083
1084 out:
1085         return err;
1086 }
1087
1088 static void pcifront_backend_changed(struct xenbus_device *xdev,
1089                                                   enum xenbus_state be_state)
1090 {
1091         struct pcifront_device *pdev = dev_get_drvdata(&xdev->dev);
1092
1093         switch (be_state) {
1094         case XenbusStateUnknown:
1095         case XenbusStateInitialising:
1096         case XenbusStateInitWait:
1097         case XenbusStateInitialised:
1098                 break;
1099
1100         case XenbusStateConnected:
1101                 pcifront_try_connect(pdev);
1102                 break;
1103
1104         case XenbusStateClosed:
1105                 if (xdev->state == XenbusStateClosed)
1106                         break;
1107                 fallthrough;    /* Missed the backend's CLOSING state */
1108         case XenbusStateClosing:
1109                 dev_warn(&xdev->dev, "backend going away!\n");
1110                 pcifront_try_disconnect(pdev);
1111                 break;
1112
1113         case XenbusStateReconfiguring:
1114                 pcifront_detach_devices(pdev);
1115                 break;
1116
1117         case XenbusStateReconfigured:
1118                 pcifront_attach_devices(pdev);
1119                 break;
1120         }
1121 }
1122
1123 static int pcifront_xenbus_probe(struct xenbus_device *xdev,
1124                                  const struct xenbus_device_id *id)
1125 {
1126         int err = 0;
1127         struct pcifront_device *pdev = alloc_pdev(xdev);
1128
1129         if (pdev == NULL) {
1130                 err = -ENOMEM;
1131                 xenbus_dev_fatal(xdev, err,
1132                                  "Error allocating pcifront_device struct");
1133                 goto out;
1134         }
1135
1136         err = pcifront_publish_info(pdev);
1137         if (err)
1138                 free_pdev(pdev);
1139
1140 out:
1141         return err;
1142 }
1143
1144 static int pcifront_xenbus_remove(struct xenbus_device *xdev)
1145 {
1146         struct pcifront_device *pdev = dev_get_drvdata(&xdev->dev);
1147
1148         if (pdev)
1149                 free_pdev(pdev);
1150
1151         return 0;
1152 }
1153
1154 static const struct xenbus_device_id xenpci_ids[] = {
1155         {"pci"},
1156         {""},
1157 };
1158
1159 static struct xenbus_driver xenpci_driver = {
1160         .name                   = "pcifront",
1161         .ids                    = xenpci_ids,
1162         .probe                  = pcifront_xenbus_probe,
1163         .remove                 = pcifront_xenbus_remove,
1164         .otherend_changed       = pcifront_backend_changed,
1165 };
1166
1167 static int __init pcifront_init(void)
1168 {
1169         if (!xen_pv_domain() || xen_initial_domain())
1170                 return -ENODEV;
1171
1172         if (!xen_has_pv_devices())
1173                 return -ENODEV;
1174
1175         pci_frontend_registrar(1 /* enable */);
1176
1177         return xenbus_register_frontend(&xenpci_driver);
1178 }
1179
1180 static void __exit pcifront_cleanup(void)
1181 {
1182         xenbus_unregister_driver(&xenpci_driver);
1183         pci_frontend_registrar(0 /* disable */);
1184 }
1185 module_init(pcifront_init);
1186 module_exit(pcifront_cleanup);
1187
1188 MODULE_DESCRIPTION("Xen PCI passthrough frontend.");
1189 MODULE_LICENSE("GPL");
1190 MODULE_ALIAS("xen:pci");