1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Xen PCI - handle PCI (INTx) and MSI infrastructure calls for PV, HVM and 4 * initial domain support. We also handle the DSDT _PRT callbacks for GSI's 5 * used in HVM and initial domain mode (PV does not parse ACPI, so it has no 6 * concept of GSIs). Under PV we hook under the pnbbios API for IRQs and 7 * 0xcf8 PCI configuration read/write. 8 * 9 * Author: Ryan Wilson <hap9@epoch.ncsc.mil> 10 * Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> 11 * Stefano Stabellini <stefano.stabellini@eu.citrix.com> 12 */ 13 #include <linux/export.h> 14 #include <linux/init.h> 15 #include <linux/pci.h> 16 #include <linux/acpi.h> 17 18 #include <linux/io.h> 19 #include <asm/io_apic.h> 20 #include <asm/pci_x86.h> 21 22 #include <asm/xen/hypervisor.h> 23 24 #include <xen/features.h> 25 #include <xen/events.h> 26 #include <asm/xen/pci.h> 27 #include <asm/xen/cpuid.h> 28 #include <asm/apic.h> 29 #include <asm/i8259.h> 30 31 static int xen_pcifront_enable_irq(struct pci_dev *dev) 32 { 33 int rc; 34 int share = 1; 35 int pirq; 36 u8 gsi; 37 38 rc = pci_read_config_byte(dev, PCI_INTERRUPT_LINE, &gsi); 39 if (rc < 0) { 40 dev_warn(&dev->dev, "Xen PCI: failed to read interrupt line: %d\n", 41 rc); 42 return rc; 43 } 44 /* In PV DomU the Xen PCI backend puts the PIRQ in the interrupt line.*/ 45 pirq = gsi; 46 47 if (gsi < nr_legacy_irqs()) 48 share = 0; 49 50 rc = xen_bind_pirq_gsi_to_irq(gsi, pirq, share, "pcifront"); 51 if (rc < 0) { 52 dev_warn(&dev->dev, "Xen PCI: failed to bind GSI%d (PIRQ%d) to IRQ: %d\n", 53 gsi, pirq, rc); 54 return rc; 55 } 56 57 dev->irq = rc; 58 dev_info(&dev->dev, "Xen PCI mapped GSI%d to IRQ%d\n", gsi, dev->irq); 59 return 0; 60 } 61 62 #ifdef CONFIG_ACPI 63 static int xen_register_pirq(u32 gsi, int triggering, bool set_pirq) 64 { 65 int rc, pirq = -1, irq; 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 irq = xen_bind_pirq_gsi_to_irq(gsi, map_irq.pirq, shareable, name); 97 if (irq < 0) 98 goto out; 99 100 printk(KERN_DEBUG "xen: --> pirq=%d -> irq=%d (gsi=%d)\n", map_irq.pirq, irq, gsi); 101 out: 102 return irq; 103 } 104 105 static int acpi_register_gsi_xen_hvm(struct device *dev, u32 gsi, 106 int trigger, int polarity) 107 { 108 if (!xen_hvm_domain()) 109 return -1; 110 111 return xen_register_pirq(gsi, trigger, 112 false /* no mapping of GSI to PIRQ */); 113 } 114 115 #ifdef CONFIG_XEN_DOM0 116 static int xen_register_gsi(u32 gsi, int triggering, int polarity) 117 { 118 int rc, irq; 119 struct physdev_setup_gsi setup_gsi; 120 121 if (!xen_pv_domain()) 122 return -1; 123 124 printk(KERN_DEBUG "xen: registering gsi %u triggering %d polarity %d\n", 125 gsi, triggering, polarity); 126 127 irq = xen_register_pirq(gsi, triggering, true); 128 129 setup_gsi.gsi = gsi; 130 setup_gsi.triggering = (triggering == ACPI_EDGE_SENSITIVE ? 0 : 1); 131 setup_gsi.polarity = (polarity == ACPI_ACTIVE_HIGH ? 0 : 1); 132 133 rc = HYPERVISOR_physdev_op(PHYSDEVOP_setup_gsi, &setup_gsi); 134 if (rc == -EEXIST) 135 printk(KERN_INFO "Already setup the GSI :%d\n", gsi); 136 else if (rc) { 137 printk(KERN_ERR "Failed to setup GSI :%d, err_code:%d\n", 138 gsi, rc); 139 } 140 141 return irq; 142 } 143 144 static int acpi_register_gsi_xen(struct device *dev, u32 gsi, 145 int trigger, int polarity) 146 { 147 return xen_register_gsi(gsi, trigger, polarity); 148 } 149 #endif 150 #endif 151 152 #if defined(CONFIG_PCI_MSI) 153 #include <linux/msi.h> 154 #include <asm/msidef.h> 155 156 struct xen_pci_frontend_ops *xen_pci_frontend; 157 EXPORT_SYMBOL_GPL(xen_pci_frontend); 158 159 static int xen_setup_msi_irqs(struct pci_dev *dev, int nvec, int type) 160 { 161 int irq, ret, i; 162 struct msi_desc *msidesc; 163 int *v; 164 165 if (type == PCI_CAP_ID_MSI && nvec > 1) 166 return 1; 167 168 v = kcalloc(max(1, nvec), sizeof(int), GFP_KERNEL); 169 if (!v) 170 return -ENOMEM; 171 172 if (type == PCI_CAP_ID_MSIX) 173 ret = xen_pci_frontend_enable_msix(dev, v, nvec); 174 else 175 ret = xen_pci_frontend_enable_msi(dev, v); 176 if (ret) 177 goto error; 178 i = 0; 179 for_each_pci_msi_entry(msidesc, dev) { 180 irq = xen_bind_pirq_msi_to_irq(dev, msidesc, v[i], 181 (type == PCI_CAP_ID_MSI) ? nvec : 1, 182 (type == PCI_CAP_ID_MSIX) ? 183 "pcifront-msi-x" : 184 "pcifront-msi", 185 DOMID_SELF); 186 if (irq < 0) { 187 ret = irq; 188 goto free; 189 } 190 i++; 191 } 192 kfree(v); 193 return 0; 194 195 error: 196 if (ret == -ENOSYS) 197 dev_err(&dev->dev, "Xen PCI frontend has not registered MSI/MSI-X support!\n"); 198 else if (ret) 199 dev_err(&dev->dev, "Xen PCI frontend error: %d!\n", ret); 200 free: 201 kfree(v); 202 return ret; 203 } 204 205 #define XEN_PIRQ_MSI_DATA (MSI_DATA_TRIGGER_EDGE | \ 206 MSI_DATA_LEVEL_ASSERT | (3 << 8) | MSI_DATA_VECTOR(0)) 207 208 static void xen_msi_compose_msg(struct pci_dev *pdev, unsigned int pirq, 209 struct msi_msg *msg) 210 { 211 /* We set vector == 0 to tell the hypervisor we don't care about it, 212 * but we want a pirq setup instead. 213 * We use the dest_id field to pass the pirq that we want. */ 214 msg->address_hi = MSI_ADDR_BASE_HI | MSI_ADDR_EXT_DEST_ID(pirq); 215 msg->address_lo = 216 MSI_ADDR_BASE_LO | 217 MSI_ADDR_DEST_MODE_PHYSICAL | 218 MSI_ADDR_REDIRECTION_CPU | 219 MSI_ADDR_DEST_ID(pirq); 220 221 msg->data = XEN_PIRQ_MSI_DATA; 222 } 223 224 static int xen_hvm_setup_msi_irqs(struct pci_dev *dev, int nvec, int type) 225 { 226 int irq, pirq; 227 struct msi_desc *msidesc; 228 struct msi_msg msg; 229 230 if (type == PCI_CAP_ID_MSI && nvec > 1) 231 return 1; 232 233 for_each_pci_msi_entry(msidesc, dev) { 234 pirq = xen_allocate_pirq_msi(dev, msidesc); 235 if (pirq < 0) { 236 irq = -ENODEV; 237 goto error; 238 } 239 xen_msi_compose_msg(dev, pirq, &msg); 240 __pci_write_msi_msg(msidesc, &msg); 241 dev_dbg(&dev->dev, "xen: msi bound to pirq=%d\n", pirq); 242 irq = xen_bind_pirq_msi_to_irq(dev, msidesc, pirq, 243 (type == PCI_CAP_ID_MSI) ? nvec : 1, 244 (type == PCI_CAP_ID_MSIX) ? 245 "msi-x" : "msi", 246 DOMID_SELF); 247 if (irq < 0) 248 goto error; 249 dev_dbg(&dev->dev, 250 "xen: msi --> pirq=%d --> irq=%d\n", pirq, irq); 251 } 252 return 0; 253 254 error: 255 dev_err(&dev->dev, "Failed to create MSI%s! ret=%d!\n", 256 type == PCI_CAP_ID_MSI ? "" : "-X", irq); 257 return irq; 258 } 259 260 #ifdef CONFIG_XEN_DOM0 261 static bool __read_mostly pci_seg_supported = true; 262 263 static int xen_initdom_setup_msi_irqs(struct pci_dev *dev, int nvec, int type) 264 { 265 int ret = 0; 266 struct msi_desc *msidesc; 267 268 for_each_pci_msi_entry(msidesc, dev) { 269 struct physdev_map_pirq map_irq; 270 domid_t domid; 271 272 domid = ret = xen_find_device_domain_owner(dev); 273 /* N.B. Casting int's -ENODEV to uint16_t results in 0xFFED, 274 * hence check ret value for < 0. */ 275 if (ret < 0) 276 domid = DOMID_SELF; 277 278 memset(&map_irq, 0, sizeof(map_irq)); 279 map_irq.domid = domid; 280 map_irq.type = MAP_PIRQ_TYPE_MSI_SEG; 281 map_irq.index = -1; 282 map_irq.pirq = -1; 283 map_irq.bus = dev->bus->number | 284 (pci_domain_nr(dev->bus) << 16); 285 map_irq.devfn = dev->devfn; 286 287 if (type == PCI_CAP_ID_MSI && nvec > 1) { 288 map_irq.type = MAP_PIRQ_TYPE_MULTI_MSI; 289 map_irq.entry_nr = nvec; 290 } else if (type == PCI_CAP_ID_MSIX) { 291 int pos; 292 unsigned long flags; 293 u32 table_offset, bir; 294 295 pos = dev->msix_cap; 296 pci_read_config_dword(dev, pos + PCI_MSIX_TABLE, 297 &table_offset); 298 bir = (u8)(table_offset & PCI_MSIX_TABLE_BIR); 299 flags = pci_resource_flags(dev, bir); 300 if (!flags || (flags & IORESOURCE_UNSET)) 301 return -EINVAL; 302 303 map_irq.table_base = pci_resource_start(dev, bir); 304 map_irq.entry_nr = msidesc->msi_attrib.entry_nr; 305 } 306 307 ret = -EINVAL; 308 if (pci_seg_supported) 309 ret = HYPERVISOR_physdev_op(PHYSDEVOP_map_pirq, 310 &map_irq); 311 if (type == PCI_CAP_ID_MSI && nvec > 1 && ret) { 312 /* 313 * If MAP_PIRQ_TYPE_MULTI_MSI is not available 314 * there's nothing else we can do in this case. 315 * Just set ret > 0 so driver can retry with 316 * single MSI. 317 */ 318 ret = 1; 319 goto out; 320 } 321 if (ret == -EINVAL && !pci_domain_nr(dev->bus)) { 322 map_irq.type = MAP_PIRQ_TYPE_MSI; 323 map_irq.index = -1; 324 map_irq.pirq = -1; 325 map_irq.bus = dev->bus->number; 326 ret = HYPERVISOR_physdev_op(PHYSDEVOP_map_pirq, 327 &map_irq); 328 if (ret != -EINVAL) 329 pci_seg_supported = false; 330 } 331 if (ret) { 332 dev_warn(&dev->dev, "xen map irq failed %d for %d domain\n", 333 ret, domid); 334 goto out; 335 } 336 337 ret = xen_bind_pirq_msi_to_irq(dev, msidesc, map_irq.pirq, 338 (type == PCI_CAP_ID_MSI) ? nvec : 1, 339 (type == PCI_CAP_ID_MSIX) ? "msi-x" : "msi", 340 domid); 341 if (ret < 0) 342 goto out; 343 } 344 ret = 0; 345 out: 346 return ret; 347 } 348 349 static void xen_initdom_restore_msi_irqs(struct pci_dev *dev) 350 { 351 int ret = 0; 352 353 if (pci_seg_supported) { 354 struct physdev_pci_device restore_ext; 355 356 restore_ext.seg = pci_domain_nr(dev->bus); 357 restore_ext.bus = dev->bus->number; 358 restore_ext.devfn = dev->devfn; 359 ret = HYPERVISOR_physdev_op(PHYSDEVOP_restore_msi_ext, 360 &restore_ext); 361 if (ret == -ENOSYS) 362 pci_seg_supported = false; 363 WARN(ret && ret != -ENOSYS, "restore_msi_ext -> %d\n", ret); 364 } 365 if (!pci_seg_supported) { 366 struct physdev_restore_msi restore; 367 368 restore.bus = dev->bus->number; 369 restore.devfn = dev->devfn; 370 ret = HYPERVISOR_physdev_op(PHYSDEVOP_restore_msi, &restore); 371 WARN(ret && ret != -ENOSYS, "restore_msi -> %d\n", ret); 372 } 373 } 374 #endif 375 376 static void xen_teardown_msi_irqs(struct pci_dev *dev) 377 { 378 struct msi_desc *msidesc; 379 380 msidesc = first_pci_msi_entry(dev); 381 if (msidesc->msi_attrib.is_msix) 382 xen_pci_frontend_disable_msix(dev); 383 else 384 xen_pci_frontend_disable_msi(dev); 385 386 /* Free the IRQ's and the msidesc using the generic code. */ 387 default_teardown_msi_irqs(dev); 388 } 389 390 static void xen_teardown_msi_irq(unsigned int irq) 391 { 392 xen_destroy_irq(irq); 393 } 394 395 #endif 396 397 int __init pci_xen_init(void) 398 { 399 if (!xen_pv_domain() || xen_initial_domain()) 400 return -ENODEV; 401 402 printk(KERN_INFO "PCI: setting up Xen PCI frontend stub\n"); 403 404 pcibios_set_cache_line_size(); 405 406 pcibios_enable_irq = xen_pcifront_enable_irq; 407 pcibios_disable_irq = NULL; 408 409 /* Keep ACPI out of the picture */ 410 acpi_noirq_set(); 411 412 #ifdef CONFIG_PCI_MSI 413 x86_msi.setup_msi_irqs = xen_setup_msi_irqs; 414 x86_msi.teardown_msi_irq = xen_teardown_msi_irq; 415 x86_msi.teardown_msi_irqs = xen_teardown_msi_irqs; 416 pci_msi_ignore_mask = 1; 417 #endif 418 return 0; 419 } 420 421 #ifdef CONFIG_PCI_MSI 422 void __init xen_msi_init(void) 423 { 424 if (!disable_apic) { 425 /* 426 * If hardware supports (x2)APIC virtualization (as indicated 427 * by hypervisor's leaf 4) then we don't need to use pirqs/ 428 * event channels for MSI handling and instead use regular 429 * APIC processing 430 */ 431 uint32_t eax = cpuid_eax(xen_cpuid_base() + 4); 432 433 if (((eax & XEN_HVM_CPUID_X2APIC_VIRT) && x2apic_mode) || 434 ((eax & XEN_HVM_CPUID_APIC_ACCESS_VIRT) && boot_cpu_has(X86_FEATURE_APIC))) 435 return; 436 } 437 438 x86_msi.setup_msi_irqs = xen_hvm_setup_msi_irqs; 439 x86_msi.teardown_msi_irq = xen_teardown_msi_irq; 440 } 441 #endif 442 443 int __init pci_xen_hvm_init(void) 444 { 445 if (!xen_have_vector_callback || !xen_feature(XENFEAT_hvm_pirqs)) 446 return 0; 447 448 #ifdef CONFIG_ACPI 449 /* 450 * We don't want to change the actual ACPI delivery model, 451 * just how GSIs get registered. 452 */ 453 __acpi_register_gsi = acpi_register_gsi_xen_hvm; 454 __acpi_unregister_gsi = NULL; 455 #endif 456 457 #ifdef CONFIG_PCI_MSI 458 /* 459 * We need to wait until after x2apic is initialized 460 * before we can set MSI IRQ ops. 461 */ 462 x86_platform.apic_post_init = xen_msi_init; 463 #endif 464 return 0; 465 } 466 467 #ifdef CONFIG_XEN_DOM0 468 int __init pci_xen_initial_domain(void) 469 { 470 int irq; 471 472 #ifdef CONFIG_PCI_MSI 473 x86_msi.setup_msi_irqs = xen_initdom_setup_msi_irqs; 474 x86_msi.teardown_msi_irq = xen_teardown_msi_irq; 475 x86_msi.restore_msi_irqs = xen_initdom_restore_msi_irqs; 476 pci_msi_ignore_mask = 1; 477 #endif 478 __acpi_register_gsi = acpi_register_gsi_xen; 479 __acpi_unregister_gsi = NULL; 480 /* 481 * Pre-allocate the legacy IRQs. Use NR_LEGACY_IRQS here 482 * because we don't have a PIC and thus nr_legacy_irqs() is zero. 483 */ 484 for (irq = 0; irq < NR_IRQS_LEGACY; irq++) { 485 int trigger, polarity; 486 487 if (acpi_get_override_irq(irq, &trigger, &polarity) == -1) 488 continue; 489 490 xen_register_pirq(irq, 491 trigger ? ACPI_LEVEL_SENSITIVE : ACPI_EDGE_SENSITIVE, 492 true /* Map GSI to PIRQ */); 493 } 494 if (0 == nr_ioapics) { 495 for (irq = 0; irq < nr_legacy_irqs(); irq++) 496 xen_bind_pirq_gsi_to_irq(irq, irq, 0, "xt-pic"); 497 } 498 return 0; 499 } 500 501 struct xen_device_domain_owner { 502 domid_t domain; 503 struct pci_dev *dev; 504 struct list_head list; 505 }; 506 507 static DEFINE_SPINLOCK(dev_domain_list_spinlock); 508 static struct list_head dev_domain_list = LIST_HEAD_INIT(dev_domain_list); 509 510 static struct xen_device_domain_owner *find_device(struct pci_dev *dev) 511 { 512 struct xen_device_domain_owner *owner; 513 514 list_for_each_entry(owner, &dev_domain_list, list) { 515 if (owner->dev == dev) 516 return owner; 517 } 518 return NULL; 519 } 520 521 int xen_find_device_domain_owner(struct pci_dev *dev) 522 { 523 struct xen_device_domain_owner *owner; 524 int domain = -ENODEV; 525 526 spin_lock(&dev_domain_list_spinlock); 527 owner = find_device(dev); 528 if (owner) 529 domain = owner->domain; 530 spin_unlock(&dev_domain_list_spinlock); 531 return domain; 532 } 533 EXPORT_SYMBOL_GPL(xen_find_device_domain_owner); 534 535 int xen_register_device_domain_owner(struct pci_dev *dev, uint16_t domain) 536 { 537 struct xen_device_domain_owner *owner; 538 539 owner = kzalloc(sizeof(struct xen_device_domain_owner), GFP_KERNEL); 540 if (!owner) 541 return -ENODEV; 542 543 spin_lock(&dev_domain_list_spinlock); 544 if (find_device(dev)) { 545 spin_unlock(&dev_domain_list_spinlock); 546 kfree(owner); 547 return -EEXIST; 548 } 549 owner->domain = domain; 550 owner->dev = dev; 551 list_add_tail(&owner->list, &dev_domain_list); 552 spin_unlock(&dev_domain_list_spinlock); 553 return 0; 554 } 555 EXPORT_SYMBOL_GPL(xen_register_device_domain_owner); 556 557 int xen_unregister_device_domain_owner(struct pci_dev *dev) 558 { 559 struct xen_device_domain_owner *owner; 560 561 spin_lock(&dev_domain_list_spinlock); 562 owner = find_device(dev); 563 if (!owner) { 564 spin_unlock(&dev_domain_list_spinlock); 565 return -ENODEV; 566 } 567 list_del(&owner->list); 568 spin_unlock(&dev_domain_list_spinlock); 569 kfree(owner); 570 return 0; 571 } 572 EXPORT_SYMBOL_GPL(xen_unregister_device_domain_owner); 573 #endif 574