1 // SPDX-License-Identifier: GPL-2.0-only
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.
9 * Author: Ryan Wilson <hap9@epoch.ncsc.mil>
10 * Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
11 * Stefano Stabellini <stefano.stabellini@eu.citrix.com>
13 #include <linux/export.h>
14 #include <linux/init.h>
15 #include <linux/pci.h>
16 #include <linux/acpi.h>
19 #include <asm/io_apic.h>
20 #include <asm/pci_x86.h>
22 #include <asm/xen/hypervisor.h>
24 #include <xen/features.h>
25 #include <xen/events.h>
26 #include <asm/xen/pci.h>
27 #include <asm/xen/cpuid.h>
30 #include <asm/i8259.h>
32 static int xen_pcifront_enable_irq(struct pci_dev *dev)
39 rc = pci_read_config_byte(dev, PCI_INTERRUPT_LINE, &gsi);
41 dev_warn(&dev->dev, "Xen PCI: failed to read interrupt line: %d\n",
45 /* In PV DomU the Xen PCI backend puts the PIRQ in the interrupt line.*/
48 if (gsi < nr_legacy_irqs())
51 rc = xen_bind_pirq_gsi_to_irq(gsi, pirq, share, "pcifront");
53 dev_warn(&dev->dev, "Xen PCI: failed to bind GSI%d (PIRQ%d) to IRQ: %d\n",
59 dev_info(&dev->dev, "Xen PCI mapped GSI%d to IRQ%d\n", gsi, dev->irq);
64 static int xen_register_pirq(u32 gsi, int triggering, bool set_pirq)
66 int rc, pirq = -1, irq;
67 struct physdev_map_pirq map_irq;
71 irq = xen_irq_from_gsi(gsi);
78 map_irq.domid = DOMID_SELF;
79 map_irq.type = MAP_PIRQ_TYPE_GSI;
83 rc = HYPERVISOR_physdev_op(PHYSDEVOP_map_pirq, &map_irq);
85 printk(KERN_WARNING "xen map irq failed %d\n", rc);
89 if (triggering == ACPI_EDGE_SENSITIVE) {
94 name = "ioapic-level";
97 irq = xen_bind_pirq_gsi_to_irq(gsi, map_irq.pirq, shareable, name);
101 printk(KERN_DEBUG "xen: --> pirq=%d -> irq=%d (gsi=%d)\n", map_irq.pirq, irq, gsi);
106 static int acpi_register_gsi_xen_hvm(struct device *dev, u32 gsi,
107 int trigger, int polarity)
109 if (!xen_hvm_domain())
112 return xen_register_pirq(gsi, trigger,
113 false /* no mapping of GSI to PIRQ */);
116 #ifdef CONFIG_XEN_PV_DOM0
117 static int xen_register_gsi(u32 gsi, int triggering, int polarity)
120 struct physdev_setup_gsi setup_gsi;
122 if (!xen_pv_domain())
125 printk(KERN_DEBUG "xen: registering gsi %u triggering %d polarity %d\n",
126 gsi, triggering, polarity);
128 irq = xen_register_pirq(gsi, triggering, true);
131 setup_gsi.triggering = (triggering == ACPI_EDGE_SENSITIVE ? 0 : 1);
132 setup_gsi.polarity = (polarity == ACPI_ACTIVE_HIGH ? 0 : 1);
134 rc = HYPERVISOR_physdev_op(PHYSDEVOP_setup_gsi, &setup_gsi);
136 printk(KERN_INFO "Already setup the GSI :%d\n", gsi);
138 printk(KERN_ERR "Failed to setup GSI :%d, err_code:%d\n",
145 static int acpi_register_gsi_xen(struct device *dev, u32 gsi,
146 int trigger, int polarity)
148 return xen_register_gsi(gsi, trigger, polarity);
153 #if defined(CONFIG_PCI_MSI)
154 #include <linux/msi.h>
156 struct xen_pci_frontend_ops *xen_pci_frontend;
157 EXPORT_SYMBOL_GPL(xen_pci_frontend);
160 int (*setup_msi_irqs)(struct pci_dev *dev, int nvec, int type);
161 void (*teardown_msi_irqs)(struct pci_dev *dev);
164 static struct xen_msi_ops xen_msi_ops __ro_after_init;
166 static int xen_setup_msi_irqs(struct pci_dev *dev, int nvec, int type)
169 struct msi_desc *msidesc;
172 if (type == PCI_CAP_ID_MSI && nvec > 1)
175 v = kcalloc(max(1, nvec), sizeof(int), GFP_KERNEL);
179 if (type == PCI_CAP_ID_MSIX)
180 ret = xen_pci_frontend_enable_msix(dev, v, nvec);
182 ret = xen_pci_frontend_enable_msi(dev, v);
186 for_each_pci_msi_entry(msidesc, dev) {
187 irq = xen_bind_pirq_msi_to_irq(dev, msidesc, v[i],
188 (type == PCI_CAP_ID_MSI) ? nvec : 1,
189 (type == PCI_CAP_ID_MSIX) ?
204 dev_err(&dev->dev, "Xen PCI frontend has not registered MSI/MSI-X support!\n");
206 dev_err(&dev->dev, "Xen PCI frontend error: %d!\n", ret);
212 static void xen_msi_compose_msg(struct pci_dev *pdev, unsigned int pirq,
216 * We set vector == 0 to tell the hypervisor we don't care about
217 * it, but we want a pirq setup instead. We use the dest_id fields
218 * to pass the pirq that we want.
220 memset(msg, 0, sizeof(*msg));
221 msg->address_hi = X86_MSI_BASE_ADDRESS_HIGH;
222 msg->arch_addr_hi.destid_8_31 = pirq >> 8;
223 msg->arch_addr_lo.destid_0_7 = pirq & 0xFF;
224 msg->arch_addr_lo.base_address = X86_MSI_BASE_ADDRESS_LOW;
225 msg->arch_data.delivery_mode = APIC_DELIVERY_MODE_EXTINT;
228 static int xen_hvm_setup_msi_irqs(struct pci_dev *dev, int nvec, int type)
231 struct msi_desc *msidesc;
234 if (type == PCI_CAP_ID_MSI && nvec > 1)
237 for_each_pci_msi_entry(msidesc, dev) {
238 pirq = xen_allocate_pirq_msi(dev, msidesc);
243 xen_msi_compose_msg(dev, pirq, &msg);
244 __pci_write_msi_msg(msidesc, &msg);
245 dev_dbg(&dev->dev, "xen: msi bound to pirq=%d\n", pirq);
246 irq = xen_bind_pirq_msi_to_irq(dev, msidesc, pirq,
247 (type == PCI_CAP_ID_MSI) ? nvec : 1,
248 (type == PCI_CAP_ID_MSIX) ?
254 "xen: msi --> pirq=%d --> irq=%d\n", pirq, irq);
259 dev_err(&dev->dev, "Failed to create MSI%s! ret=%d!\n",
260 type == PCI_CAP_ID_MSI ? "" : "-X", irq);
264 #ifdef CONFIG_XEN_PV_DOM0
265 static bool __read_mostly pci_seg_supported = true;
267 static int xen_initdom_setup_msi_irqs(struct pci_dev *dev, int nvec, int type)
270 struct msi_desc *msidesc;
272 for_each_pci_msi_entry(msidesc, dev) {
273 struct physdev_map_pirq map_irq;
276 domid = ret = xen_find_device_domain_owner(dev);
277 /* N.B. Casting int's -ENODEV to uint16_t results in 0xFFED,
278 * hence check ret value for < 0. */
282 memset(&map_irq, 0, sizeof(map_irq));
283 map_irq.domid = domid;
284 map_irq.type = MAP_PIRQ_TYPE_MSI_SEG;
287 map_irq.bus = dev->bus->number |
288 (pci_domain_nr(dev->bus) << 16);
289 map_irq.devfn = dev->devfn;
291 if (type == PCI_CAP_ID_MSI && nvec > 1) {
292 map_irq.type = MAP_PIRQ_TYPE_MULTI_MSI;
293 map_irq.entry_nr = nvec;
294 } else if (type == PCI_CAP_ID_MSIX) {
297 u32 table_offset, bir;
300 pci_read_config_dword(dev, pos + PCI_MSIX_TABLE,
302 bir = (u8)(table_offset & PCI_MSIX_TABLE_BIR);
303 flags = pci_resource_flags(dev, bir);
304 if (!flags || (flags & IORESOURCE_UNSET))
307 map_irq.table_base = pci_resource_start(dev, bir);
308 map_irq.entry_nr = msidesc->msi_attrib.entry_nr;
312 if (pci_seg_supported)
313 ret = HYPERVISOR_physdev_op(PHYSDEVOP_map_pirq,
315 if (type == PCI_CAP_ID_MSI && nvec > 1 && ret) {
317 * If MAP_PIRQ_TYPE_MULTI_MSI is not available
318 * there's nothing else we can do in this case.
319 * Just set ret > 0 so driver can retry with
325 if (ret == -EINVAL && !pci_domain_nr(dev->bus)) {
326 map_irq.type = MAP_PIRQ_TYPE_MSI;
329 map_irq.bus = dev->bus->number;
330 ret = HYPERVISOR_physdev_op(PHYSDEVOP_map_pirq,
333 pci_seg_supported = false;
336 dev_warn(&dev->dev, "xen map irq failed %d for %d domain\n",
341 ret = xen_bind_pirq_msi_to_irq(dev, msidesc, map_irq.pirq,
342 (type == PCI_CAP_ID_MSI) ? nvec : 1,
343 (type == PCI_CAP_ID_MSIX) ? "msi-x" : "msi",
353 static void xen_initdom_restore_msi_irqs(struct pci_dev *dev)
357 if (pci_seg_supported) {
358 struct physdev_pci_device restore_ext;
360 restore_ext.seg = pci_domain_nr(dev->bus);
361 restore_ext.bus = dev->bus->number;
362 restore_ext.devfn = dev->devfn;
363 ret = HYPERVISOR_physdev_op(PHYSDEVOP_restore_msi_ext,
366 pci_seg_supported = false;
367 WARN(ret && ret != -ENOSYS, "restore_msi_ext -> %d\n", ret);
369 if (!pci_seg_supported) {
370 struct physdev_restore_msi restore;
372 restore.bus = dev->bus->number;
373 restore.devfn = dev->devfn;
374 ret = HYPERVISOR_physdev_op(PHYSDEVOP_restore_msi, &restore);
375 WARN(ret && ret != -ENOSYS, "restore_msi -> %d\n", ret);
378 #else /* CONFIG_XEN_PV_DOM0 */
379 #define xen_initdom_setup_msi_irqs NULL
380 #define xen_initdom_restore_msi_irqs NULL
381 #endif /* !CONFIG_XEN_PV_DOM0 */
383 static void xen_teardown_msi_irqs(struct pci_dev *dev)
385 struct msi_desc *msidesc;
388 for_each_pci_msi_entry(msidesc, dev) {
390 for (i = 0; i < msidesc->nvec_used; i++)
391 xen_destroy_irq(msidesc->irq + i);
396 static void xen_pv_teardown_msi_irqs(struct pci_dev *dev)
398 struct msi_desc *msidesc = first_pci_msi_entry(dev);
400 if (msidesc->msi_attrib.is_msix)
401 xen_pci_frontend_disable_msix(dev);
403 xen_pci_frontend_disable_msi(dev);
405 xen_teardown_msi_irqs(dev);
408 static int xen_msi_domain_alloc_irqs(struct irq_domain *domain,
409 struct device *dev, int nvec)
413 if (WARN_ON_ONCE(!dev_is_pci(dev)))
416 if (first_msi_entry(dev)->msi_attrib.is_msix)
417 type = PCI_CAP_ID_MSIX;
419 type = PCI_CAP_ID_MSI;
421 return xen_msi_ops.setup_msi_irqs(to_pci_dev(dev), nvec, type);
424 static void xen_msi_domain_free_irqs(struct irq_domain *domain,
427 if (WARN_ON_ONCE(!dev_is_pci(dev)))
430 xen_msi_ops.teardown_msi_irqs(to_pci_dev(dev));
433 static struct msi_domain_ops xen_pci_msi_domain_ops = {
434 .domain_alloc_irqs = xen_msi_domain_alloc_irqs,
435 .domain_free_irqs = xen_msi_domain_free_irqs,
438 static struct msi_domain_info xen_pci_msi_domain_info = {
439 .ops = &xen_pci_msi_domain_ops,
443 * This irq domain is a blatant violation of the irq domain design, but
444 * distangling XEN into real irq domains is not a job for mere mortals with
445 * limited XENology. But it's the least dangerous way for a mere mortal to
446 * get rid of the arch_*_msi_irqs() hackery in order to store the irq
447 * domain pointer in struct device. This irq domain wrappery allows to do
448 * that without breaking XEN terminally.
450 static __init struct irq_domain *xen_create_pci_msi_domain(void)
452 struct irq_domain *d = NULL;
453 struct fwnode_handle *fn;
455 fn = irq_domain_alloc_named_fwnode("XEN-MSI");
457 d = msi_create_irq_domain(fn, &xen_pci_msi_domain_info, NULL);
459 /* FIXME: No idea how to survive if this fails */
465 static __init void xen_setup_pci_msi(void)
467 if (xen_pv_domain()) {
468 if (xen_initial_domain()) {
469 xen_msi_ops.setup_msi_irqs = xen_initdom_setup_msi_irqs;
470 x86_msi.restore_msi_irqs = xen_initdom_restore_msi_irqs;
472 xen_msi_ops.setup_msi_irqs = xen_setup_msi_irqs;
474 xen_msi_ops.teardown_msi_irqs = xen_pv_teardown_msi_irqs;
475 } else if (xen_hvm_domain()) {
476 xen_msi_ops.setup_msi_irqs = xen_hvm_setup_msi_irqs;
477 xen_msi_ops.teardown_msi_irqs = xen_teardown_msi_irqs;
484 * Override the PCI/MSI irq domain init function. No point
485 * in allocating the native domain and never use it.
487 x86_init.irqs.create_pci_msi_domain = xen_create_pci_msi_domain;
489 * With XEN PIRQ/Eventchannels in use PCI/MSI[-X] masking is solely
490 * controlled by the hypervisor.
492 pci_msi_ignore_mask = 1;
495 #else /* CONFIG_PCI_MSI */
496 static inline void xen_setup_pci_msi(void) { }
497 #endif /* CONFIG_PCI_MSI */
499 int __init pci_xen_init(void)
501 if (!xen_pv_domain() || xen_initial_domain())
504 printk(KERN_INFO "PCI: setting up Xen PCI frontend stub\n");
506 pcibios_set_cache_line_size();
508 pcibios_enable_irq = xen_pcifront_enable_irq;
509 pcibios_disable_irq = NULL;
511 /* Keep ACPI out of the picture */
518 #ifdef CONFIG_PCI_MSI
519 static void __init xen_hvm_msi_init(void)
523 * If hardware supports (x2)APIC virtualization (as indicated
524 * by hypervisor's leaf 4) then we don't need to use pirqs/
525 * event channels for MSI handling and instead use regular
528 uint32_t eax = cpuid_eax(xen_cpuid_base() + 4);
530 if (((eax & XEN_HVM_CPUID_X2APIC_VIRT) && x2apic_mode) ||
531 ((eax & XEN_HVM_CPUID_APIC_ACCESS_VIRT) && boot_cpu_has(X86_FEATURE_APIC)))
538 int __init pci_xen_hvm_init(void)
540 if (!xen_have_vector_callback || !xen_feature(XENFEAT_hvm_pirqs))
545 * We don't want to change the actual ACPI delivery model,
546 * just how GSIs get registered.
548 __acpi_register_gsi = acpi_register_gsi_xen_hvm;
549 __acpi_unregister_gsi = NULL;
552 #ifdef CONFIG_PCI_MSI
554 * We need to wait until after x2apic is initialized
555 * before we can set MSI IRQ ops.
557 x86_platform.apic_post_init = xen_hvm_msi_init;
562 #ifdef CONFIG_XEN_PV_DOM0
563 int __init pci_xen_initial_domain(void)
568 __acpi_register_gsi = acpi_register_gsi_xen;
569 __acpi_unregister_gsi = NULL;
571 * Pre-allocate the legacy IRQs. Use NR_LEGACY_IRQS here
572 * because we don't have a PIC and thus nr_legacy_irqs() is zero.
574 for (irq = 0; irq < NR_IRQS_LEGACY; irq++) {
575 int trigger, polarity;
577 if (acpi_get_override_irq(irq, &trigger, &polarity) == -1)
580 xen_register_pirq(irq,
581 trigger ? ACPI_LEVEL_SENSITIVE : ACPI_EDGE_SENSITIVE,
582 true /* Map GSI to PIRQ */);
584 if (0 == nr_ioapics) {
585 for (irq = 0; irq < nr_legacy_irqs(); irq++)
586 xen_bind_pirq_gsi_to_irq(irq, irq, 0, "xt-pic");
592 #ifdef CONFIG_XEN_DOM0
594 struct xen_device_domain_owner {
597 struct list_head list;
600 static DEFINE_SPINLOCK(dev_domain_list_spinlock);
601 static struct list_head dev_domain_list = LIST_HEAD_INIT(dev_domain_list);
603 static struct xen_device_domain_owner *find_device(struct pci_dev *dev)
605 struct xen_device_domain_owner *owner;
607 list_for_each_entry(owner, &dev_domain_list, list) {
608 if (owner->dev == dev)
614 int xen_find_device_domain_owner(struct pci_dev *dev)
616 struct xen_device_domain_owner *owner;
617 int domain = -ENODEV;
619 spin_lock(&dev_domain_list_spinlock);
620 owner = find_device(dev);
622 domain = owner->domain;
623 spin_unlock(&dev_domain_list_spinlock);
626 EXPORT_SYMBOL_GPL(xen_find_device_domain_owner);
628 int xen_register_device_domain_owner(struct pci_dev *dev, uint16_t domain)
630 struct xen_device_domain_owner *owner;
632 owner = kzalloc(sizeof(struct xen_device_domain_owner), GFP_KERNEL);
636 spin_lock(&dev_domain_list_spinlock);
637 if (find_device(dev)) {
638 spin_unlock(&dev_domain_list_spinlock);
642 owner->domain = domain;
644 list_add_tail(&owner->list, &dev_domain_list);
645 spin_unlock(&dev_domain_list_spinlock);
648 EXPORT_SYMBOL_GPL(xen_register_device_domain_owner);
650 int xen_unregister_device_domain_owner(struct pci_dev *dev)
652 struct xen_device_domain_owner *owner;
654 spin_lock(&dev_domain_list_spinlock);
655 owner = find_device(dev);
657 spin_unlock(&dev_domain_list_spinlock);
660 list_del(&owner->list);
661 spin_unlock(&dev_domain_list_spinlock);
665 EXPORT_SYMBOL_GPL(xen_unregister_device_domain_owner);
666 #endif /* CONFIG_XEN_DOM0 */