arch/x86/pci/xen.c

   1 /*
   2  * Xen PCI - handle PCI (INTx) and MSI infrastructure calls for PV, HVM and
   3  * initial domain support. We also handle the DSDT _PRT callbacks for GSI's
   4  * used in HVM and initial domain mode (PV does not parse ACPI, so it has no
   5  * concept of GSIs). Under PV we hook under the pnbbios API for IRQs and
   6  * 0xcf8 PCI configuration read/write.
   7  *
   8  *   Author: Ryan Wilson <hap9@epoch.ncsc.mil>
   9  *           Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
  10  *           Stefano Stabellini <stefano.stabellini@eu.citrix.com>
  11  */
  12 #include <linux/export.h>
  13 #include <linux/init.h>
  14 #include <linux/pci.h>
  15 #include <linux/acpi.h>
  16
  17 #include <linux/io.h>
  18 #include <asm/io_apic.h>
  19 #include <asm/pci_x86.h>
  20
  21 #include <asm/xen/hypervisor.h>
  22
  23 #include <xen/features.h>
  24 #include <xen/events.h>
  25 #include <asm/xen/pci.h>
  26 #include <asm/xen/cpuid.h>
  27 #include <asm/apic.h>
  28 #include <asm/i8259.h>
  29
  30 static int xen_pcifront_enable_irq(struct pci_dev *dev)
  31 {
  32         int rc;
  33         int share = 1;
  34         int pirq;
  35         u8 gsi;
  36
  37         rc = pci_read_config_byte(dev, PCI_INTERRUPT_LINE, &gsi);
  38         if (rc < 0) {
  39                 dev_warn(&dev->dev, "Xen PCI: failed to read interrupt line: %d\n",
  40                          rc);
  41                 return rc;
  42         }
  43         /* In PV DomU the Xen PCI backend puts the PIRQ in the interrupt line.*/
  44         pirq = gsi;
  45
  46         if (gsi < nr_legacy_irqs())
  47                 share = 0;
  48
  49         rc = xen_bind_pirq_gsi_to_irq(gsi, pirq, share, "pcifront");
  50         if (rc < 0) {
  51                 dev_warn(&dev->dev, "Xen PCI: failed to bind GSI%d (PIRQ%d) to IRQ: %d\n",
  52                          gsi, pirq, rc);
  53                 return rc;
  54         }
  55
  56         dev->irq = rc;
  57         dev_info(&dev->dev, "Xen PCI mapped GSI%d to IRQ%d\n", gsi, dev->irq);
  58         return 0;
  59 }
  60
  61 #ifdef CONFIG_ACPI
  62 static int xen_register_pirq(u32 gsi, int gsi_override, int triggering,
  63                              bool set_pirq)
  64 {
  65         int rc, pirq = -1, irq = -1;
  66         struct physdev_map_pirq map_irq;
  67         int shareable = 0;
  68         char *name;
  69
  70         irq = xen_irq_from_gsi(gsi);
  71         if (irq > 0)
  72                 return irq;
  73
  74         if (set_pirq)
  75                 pirq = gsi;
  76
  77         map_irq.domid = DOMID_SELF;
  78         map_irq.type = MAP_PIRQ_TYPE_GSI;
  79         map_irq.index = gsi;
  80         map_irq.pirq = pirq;
  81
  82         rc = HYPERVISOR_physdev_op(PHYSDEVOP_map_pirq, &map_irq);
  83         if (rc) {
  84                 printk(KERN_WARNING "xen map irq failed %d\n", rc);
  85                 return -1;
  86         }
  87
  88         if (triggering == ACPI_EDGE_SENSITIVE) {
  89                 shareable = 0;
  90                 name = "ioapic-edge";
  91         } else {
  92                 shareable = 1;
  93                 name = "ioapic-level";
  94         }
  95
  96         if (gsi_override >= 0)
  97                 gsi = gsi_override;
  98
  99         irq = xen_bind_pirq_gsi_to_irq(gsi, map_irq.pirq, shareable, name);
 100         if (irq < 0)
 101                 goto out;
 102
 103         printk(KERN_DEBUG "xen: --> pirq=%d -> irq=%d (gsi=%d)\n", map_irq.pirq, irq, gsi);
 104 out:
 105         return irq;
 106 }
 107
 108 static int acpi_register_gsi_xen_hvm(struct device *dev, u32 gsi,
 109                                      int trigger, int polarity)
 110 {
 111         if (!xen_hvm_domain())
 112                 return -1;
 113
 114         return xen_register_pirq(gsi, -1 /* no GSI override */, trigger,
 115                                  false /* no mapping of GSI to PIRQ */);
 116 }
 117
 118 #ifdef CONFIG_XEN_DOM0
 119 static int xen_register_gsi(u32 gsi, int gsi_override, int triggering, int polarity)
 120 {
 121         int rc, irq;
 122         struct physdev_setup_gsi setup_gsi;
 123
 124         if (!xen_pv_domain())
 125                 return -1;
 126
 127         printk(KERN_DEBUG "xen: registering gsi %u triggering %d polarity %d\n",
 128                         gsi, triggering, polarity);
 129
 130         irq = xen_register_pirq(gsi, gsi_override, triggering, true);
 131
 132         setup_gsi.gsi = gsi;
 133         setup_gsi.triggering = (triggering == ACPI_EDGE_SENSITIVE ? 0 : 1);
 134         setup_gsi.polarity = (polarity == ACPI_ACTIVE_HIGH ? 0 : 1);
 135
 136         rc = HYPERVISOR_physdev_op(PHYSDEVOP_setup_gsi, &setup_gsi);
 137         if (rc == -EEXIST)
 138                 printk(KERN_INFO "Already setup the GSI :%d\n", gsi);
 139         else if (rc) {
 140                 printk(KERN_ERR "Failed to setup GSI :%d, err_code:%d\n",
 141                                 gsi, rc);
 142         }
 143
 144         return irq;
 145 }
 146
 147 static int acpi_register_gsi_xen(struct device *dev, u32 gsi,
 148                                  int trigger, int polarity)
 149 {
 150         return xen_register_gsi(gsi, -1 /* no GSI override */, trigger, polarity);
 151 }
 152 #endif
 153 #endif
 154
 155 #if defined(CONFIG_PCI_MSI)
 156 #include <linux/msi.h>
 157 #include <asm/msidef.h>
 158
 159 struct xen_pci_frontend_ops *xen_pci_frontend;
 160 EXPORT_SYMBOL_GPL(xen_pci_frontend);
 161
 162 static int xen_setup_msi_irqs(struct pci_dev *dev, int nvec, int type)
 163 {
 164         int irq, ret, i;
 165         struct msi_desc *msidesc;
 166         int *v;
 167
 168         if (type == PCI_CAP_ID_MSI && nvec > 1)
 169                 return 1;
 170
 171         v = kcalloc(max(1, nvec), sizeof(int), GFP_KERNEL);
 172         if (!v)
 173                 return -ENOMEM;
 174
 175         if (type == PCI_CAP_ID_MSIX)
 176                 ret = xen_pci_frontend_enable_msix(dev, v, nvec);
 177         else
 178                 ret = xen_pci_frontend_enable_msi(dev, v);
 179         if (ret)
 180                 goto error;
 181         i = 0;
 182         for_each_pci_msi_entry(msidesc, dev) {
 183                 irq = xen_bind_pirq_msi_to_irq(dev, msidesc, v[i],
 184                                                (type == PCI_CAP_ID_MSI) ? nvec : 1,
 185                                                (type == PCI_CAP_ID_MSIX) ?
 186                                                "pcifront-msi-x" :
 187                                                "pcifront-msi",
 188                                                 DOMID_SELF);
 189                 if (irq < 0) {
 190                         ret = irq;
 191                         goto free;
 192                 }
 193                 i++;
 194         }
 195         kfree(v);
 196         return 0;
 197
 198 error:
 199         if (ret == -ENOSYS)
 200                 dev_err(&dev->dev, "Xen PCI frontend has not registered MSI/MSI-X support!\n");
 201         else if (ret)
 202                 dev_err(&dev->dev, "Xen PCI frontend error: %d!\n", ret);
 203 free:
 204         kfree(v);
 205         return ret;
 206 }
 207
 208 #define XEN_PIRQ_MSI_DATA  (MSI_DATA_TRIGGER_EDGE | \
 209                 MSI_DATA_LEVEL_ASSERT | (3 << 8) | MSI_DATA_VECTOR(0))
 210
 211 static void xen_msi_compose_msg(struct pci_dev *pdev, unsigned int pirq,
 212                 struct msi_msg *msg)
 213 {
 214         /* We set vector == 0 to tell the hypervisor we don't care about it,
 215          * but we want a pirq setup instead.
 216          * We use the dest_id field to pass the pirq that we want. */
 217         msg->address_hi = MSI_ADDR_BASE_HI | MSI_ADDR_EXT_DEST_ID(pirq);
 218         msg->address_lo =
 219                 MSI_ADDR_BASE_LO |
 220                 MSI_ADDR_DEST_MODE_PHYSICAL |
 221                 MSI_ADDR_REDIRECTION_CPU |
 222                 MSI_ADDR_DEST_ID(pirq);
 223
 224         msg->data = XEN_PIRQ_MSI_DATA;
 225 }
 226
 227 static int xen_hvm_setup_msi_irqs(struct pci_dev *dev, int nvec, int type)
 228 {
 229         int irq, pirq;
 230         struct msi_desc *msidesc;
 231         struct msi_msg msg;
 232
 233         if (type == PCI_CAP_ID_MSI && nvec > 1)
 234                 return 1;
 235
 236         for_each_pci_msi_entry(msidesc, dev) {
 237                 pirq = xen_allocate_pirq_msi(dev, msidesc);
 238                 if (pirq < 0) {
 239                         irq = -ENODEV;
 240                         goto error;
 241                 }
 242                 xen_msi_compose_msg(dev, pirq, &msg);
 243                 __pci_write_msi_msg(msidesc, &msg);
 244                 dev_dbg(&dev->dev, "xen: msi bound to pirq=%d\n", pirq);
 245                 irq = xen_bind_pirq_msi_to_irq(dev, msidesc, pirq,
 246                                                (type == PCI_CAP_ID_MSI) ? nvec : 1,
 247                                                (type == PCI_CAP_ID_MSIX) ?
 248                                                "msi-x" : "msi",
 249                                                DOMID_SELF);
 250                 if (irq < 0)
 251                         goto error;
 252                 dev_dbg(&dev->dev,
 253                         "xen: msi --> pirq=%d --> irq=%d\n", pirq, irq);
 254         }
 255         return 0;
 256
 257 error:
 258         dev_err(&dev->dev, "Failed to create MSI%s! ret=%d!\n",
 259                 type == PCI_CAP_ID_MSI ? "" : "-X", irq);
 260         return irq;
 261 }
 262
 263 #ifdef CONFIG_XEN_DOM0
 264 static bool __read_mostly pci_seg_supported = true;
 265
 266 static int xen_initdom_setup_msi_irqs(struct pci_dev *dev, int nvec, int type)
 267 {
 268         int ret = 0;
 269         struct msi_desc *msidesc;
 270
 271         for_each_pci_msi_entry(msidesc, dev) {
 272                 struct physdev_map_pirq map_irq;
 273                 domid_t domid;
 274
 275                 domid = ret = xen_find_device_domain_owner(dev);
 276                 /* N.B. Casting int's -ENODEV to uint16_t results in 0xFFED,
 277                  * hence check ret value for < 0. */
 278                 if (ret < 0)
 279                         domid = DOMID_SELF;
 280
 281                 memset(&map_irq, 0, sizeof(map_irq));
 282                 map_irq.domid = domid;
 283                 map_irq.type = MAP_PIRQ_TYPE_MSI_SEG;
 284                 map_irq.index = -1;
 285                 map_irq.pirq = -1;
 286                 map_irq.bus = dev->bus->number |
 287                               (pci_domain_nr(dev->bus) << 16);
 288                 map_irq.devfn = dev->devfn;
 289
 290                 if (type == PCI_CAP_ID_MSI && nvec > 1) {
 291                         map_irq.type = MAP_PIRQ_TYPE_MULTI_MSI;
 292                         map_irq.entry_nr = nvec;
 293                 } else if (type == PCI_CAP_ID_MSIX) {
 294                         int pos;
 295                         unsigned long flags;
 296                         u32 table_offset, bir;
 297
 298                         pos = dev->msix_cap;
 299                         pci_read_config_dword(dev, pos + PCI_MSIX_TABLE,
 300                                               &table_offset);
 301                         bir = (u8)(table_offset & PCI_MSIX_TABLE_BIR);
 302                         flags = pci_resource_flags(dev, bir);
 303                         if (!flags || (flags & IORESOURCE_UNSET))
 304                                 return -EINVAL;
 305
 306                         map_irq.table_base = pci_resource_start(dev, bir);
 307                         map_irq.entry_nr = msidesc->msi_attrib.entry_nr;
 308                 }
 309
 310                 ret = -EINVAL;
 311                 if (pci_seg_supported)
 312                         ret = HYPERVISOR_physdev_op(PHYSDEVOP_map_pirq,
 313                                                     &map_irq);
 314                 if (type == PCI_CAP_ID_MSI && nvec > 1 && ret) {
 315                         /*
 316                          * If MAP_PIRQ_TYPE_MULTI_MSI is not available
 317                          * there's nothing else we can do in this case.
 318                          * Just set ret > 0 so driver can retry with
 319                          * single MSI.
 320                          */
 321                         ret = 1;
 322                         goto out;
 323                 }
 324                 if (ret == -EINVAL && !pci_domain_nr(dev->bus)) {
 325                         map_irq.type = MAP_PIRQ_TYPE_MSI;
 326                         map_irq.index = -1;
 327                         map_irq.pirq = -1;
 328                         map_irq.bus = dev->bus->number;
 329                         ret = HYPERVISOR_physdev_op(PHYSDEVOP_map_pirq,
 330                                                     &map_irq);
 331                         if (ret != -EINVAL)
 332                                 pci_seg_supported = false;
 333                 }
 334                 if (ret) {
 335                         dev_warn(&dev->dev, "xen map irq failed %d for %d domain\n",
 336                                  ret, domid);
 337                         goto out;
 338                 }
 339
 340                 ret = xen_bind_pirq_msi_to_irq(dev, msidesc, map_irq.pirq,
 341                                                (type == PCI_CAP_ID_MSI) ? nvec : 1,
 342                                                (type == PCI_CAP_ID_MSIX) ? "msi-x" : "msi",
 343                                                domid);
 344                 if (ret < 0)
 345                         goto out;
 346         }
 347         ret = 0;
 348 out:
 349         return ret;
 350 }
 351
 352 static void xen_initdom_restore_msi_irqs(struct pci_dev *dev)
 353 {
 354         int ret = 0;
 355
 356         if (pci_seg_supported) {
 357                 struct physdev_pci_device restore_ext;
 358
 359                 restore_ext.seg = pci_domain_nr(dev->bus);
 360                 restore_ext.bus = dev->bus->number;
 361                 restore_ext.devfn = dev->devfn;
 362                 ret = HYPERVISOR_physdev_op(PHYSDEVOP_restore_msi_ext,
 363                                         &restore_ext);
 364                 if (ret == -ENOSYS)
 365                         pci_seg_supported = false;
 366                 WARN(ret && ret != -ENOSYS, "restore_msi_ext -> %d\n", ret);
 367         }
 368         if (!pci_seg_supported) {
 369                 struct physdev_restore_msi restore;
 370
 371                 restore.bus = dev->bus->number;
 372                 restore.devfn = dev->devfn;
 373                 ret = HYPERVISOR_physdev_op(PHYSDEVOP_restore_msi, &restore);
 374                 WARN(ret && ret != -ENOSYS, "restore_msi -> %d\n", ret);
 375         }
 376 }
 377 #endif
 378
 379 static void xen_teardown_msi_irqs(struct pci_dev *dev)
 380 {
 381         struct msi_desc *msidesc;
 382
 383         msidesc = first_pci_msi_entry(dev);
 384         if (msidesc->msi_attrib.is_msix)
 385                 xen_pci_frontend_disable_msix(dev);
 386         else
 387                 xen_pci_frontend_disable_msi(dev);
 388
 389         /* Free the IRQ's and the msidesc using the generic code. */
 390         default_teardown_msi_irqs(dev);
 391 }
 392
 393 static void xen_teardown_msi_irq(unsigned int irq)
 394 {
 395         xen_destroy_irq(irq);
 396 }
 397
 398 #endif
 399
 400 int __init pci_xen_init(void)
 401 {
 402         if (!xen_pv_domain() || xen_initial_domain())
 403                 return -ENODEV;
 404
 405         printk(KERN_INFO "PCI: setting up Xen PCI frontend stub\n");
 406
 407         pcibios_set_cache_line_size();
 408
 409         pcibios_enable_irq = xen_pcifront_enable_irq;
 410         pcibios_disable_irq = NULL;
 411
 412         /* Keep ACPI out of the picture */
 413         acpi_noirq_set();
 414
 415 #ifdef CONFIG_PCI_MSI
 416         x86_msi.setup_msi_irqs = xen_setup_msi_irqs;
 417         x86_msi.teardown_msi_irq = xen_teardown_msi_irq;
 418         x86_msi.teardown_msi_irqs = xen_teardown_msi_irqs;
 419         pci_msi_ignore_mask = 1;
 420 #endif
 421         return 0;
 422 }
 423
 424 #ifdef CONFIG_PCI_MSI
 425 void __init xen_msi_init(void)
 426 {
 427         if (!disable_apic) {
 428                 /*
 429                  * If hardware supports (x2)APIC virtualization (as indicated
 430                  * by hypervisor's leaf 4) then we don't need to use pirqs/
 431                  * event channels for MSI handling and instead use regular
 432                  * APIC processing
 433                  */
 434                 uint32_t eax = cpuid_eax(xen_cpuid_base() + 4);
 435
 436                 if (((eax & XEN_HVM_CPUID_X2APIC_VIRT) && x2apic_mode) ||
 437                     ((eax & XEN_HVM_CPUID_APIC_ACCESS_VIRT) && boot_cpu_has(X86_FEATURE_APIC)))
 438                         return;
 439         }
 440
 441         x86_msi.setup_msi_irqs = xen_hvm_setup_msi_irqs;
 442         x86_msi.teardown_msi_irq = xen_teardown_msi_irq;
 443 }
 444 #endif
 445
 446 int __init pci_xen_hvm_init(void)
 447 {
 448         if (!xen_have_vector_callback || !xen_feature(XENFEAT_hvm_pirqs))
 449                 return 0;
 450
 451 #ifdef CONFIG_ACPI
 452         /*
 453          * We don't want to change the actual ACPI delivery model,
 454          * just how GSIs get registered.
 455          */
 456         __acpi_register_gsi = acpi_register_gsi_xen_hvm;
 457         __acpi_unregister_gsi = NULL;
 458 #endif
 459
 460 #ifdef CONFIG_PCI_MSI
 461         /*
 462          * We need to wait until after x2apic is initialized
 463          * before we can set MSI IRQ ops.
 464          */
 465         x86_platform.apic_post_init = xen_msi_init;
 466 #endif
 467         return 0;
 468 }
 469
 470 #ifdef CONFIG_XEN_DOM0
 471 int __init pci_xen_initial_domain(void)
 472 {
 473         int irq;
 474
 475 #ifdef CONFIG_PCI_MSI
 476         x86_msi.setup_msi_irqs = xen_initdom_setup_msi_irqs;
 477         x86_msi.teardown_msi_irq = xen_teardown_msi_irq;
 478         x86_msi.restore_msi_irqs = xen_initdom_restore_msi_irqs;
 479         pci_msi_ignore_mask = 1;
 480 #endif
 481         __acpi_register_gsi = acpi_register_gsi_xen;
 482         __acpi_unregister_gsi = NULL;
 483         /*
 484          * Pre-allocate the legacy IRQs.  Use NR_LEGACY_IRQS here
 485          * because we don't have a PIC and thus nr_legacy_irqs() is zero.
 486          */
 487         for (irq = 0; irq < NR_IRQS_LEGACY; irq++) {
 488                 int trigger, polarity;
 489
 490                 if (acpi_get_override_irq(irq, &trigger, &polarity) == -1)
 491                         continue;
 492
 493                 xen_register_pirq(irq, -1 /* no GSI override */,
 494                         trigger ? ACPI_LEVEL_SENSITIVE : ACPI_EDGE_SENSITIVE,
 495                         true /* Map GSI to PIRQ */);
 496         }
 497         if (0 == nr_ioapics) {
 498                 for (irq = 0; irq < nr_legacy_irqs(); irq++)
 499                         xen_bind_pirq_gsi_to_irq(irq, irq, 0, "xt-pic");
 500         }
 501         return 0;
 502 }
 503
 504 struct xen_device_domain_owner {
 505         domid_t domain;
 506         struct pci_dev *dev;
 507         struct list_head list;
 508 };
 509
 510 static DEFINE_SPINLOCK(dev_domain_list_spinlock);
 511 static struct list_head dev_domain_list = LIST_HEAD_INIT(dev_domain_list);
 512
 513 static struct xen_device_domain_owner *find_device(struct pci_dev *dev)
 514 {
 515         struct xen_device_domain_owner *owner;
 516
 517         list_for_each_entry(owner, &dev_domain_list, list) {
 518                 if (owner->dev == dev)
 519                         return owner;
 520         }
 521         return NULL;
 522 }
 523
 524 int xen_find_device_domain_owner(struct pci_dev *dev)
 525 {
 526         struct xen_device_domain_owner *owner;
 527         int domain = -ENODEV;
 528
 529         spin_lock(&dev_domain_list_spinlock);
 530         owner = find_device(dev);
 531         if (owner)
 532                 domain = owner->domain;
 533         spin_unlock(&dev_domain_list_spinlock);
 534         return domain;
 535 }
 536 EXPORT_SYMBOL_GPL(xen_find_device_domain_owner);
 537
 538 int xen_register_device_domain_owner(struct pci_dev *dev, uint16_t domain)
 539 {
 540         struct xen_device_domain_owner *owner;
 541
 542         owner = kzalloc(sizeof(struct xen_device_domain_owner), GFP_KERNEL);
 543         if (!owner)
 544                 return -ENODEV;
 545
 546         spin_lock(&dev_domain_list_spinlock);
 547         if (find_device(dev)) {
 548                 spin_unlock(&dev_domain_list_spinlock);
 549                 kfree(owner);
 550                 return -EEXIST;
 551         }
 552         owner->domain = domain;
 553         owner->dev = dev;
 554         list_add_tail(&owner->list, &dev_domain_list);
 555         spin_unlock(&dev_domain_list_spinlock);
 556         return 0;
 557 }
 558 EXPORT_SYMBOL_GPL(xen_register_device_domain_owner);
 559
 560 int xen_unregister_device_domain_owner(struct pci_dev *dev)
 561 {
 562         struct xen_device_domain_owner *owner;
 563
 564         spin_lock(&dev_domain_list_spinlock);
 565         owner = find_device(dev);
 566         if (!owner) {
 567                 spin_unlock(&dev_domain_list_spinlock);
 568                 return -ENODEV;
 569         }
 570         list_del(&owner->list);
 571         spin_unlock(&dev_domain_list_spinlock);
 572         kfree(owner);
 573         return 0;
 574 }
 575 EXPORT_SYMBOL_GPL(xen_unregister_device_domain_owner);
 576 #endif