1 // SPDX-License-Identifier: GPL-2.0
3 #define pr_fmt(fmt) "DMAR-IR: " fmt
5 #include <linux/interrupt.h>
6 #include <linux/dmar.h>
7 #include <linux/spinlock.h>
8 #include <linux/slab.h>
9 #include <linux/jiffies.h>
10 #include <linux/hpet.h>
11 #include <linux/pci.h>
12 #include <linux/irq.h>
13 #include <linux/intel-iommu.h>
14 #include <linux/acpi.h>
15 #include <linux/irqdomain.h>
16 #include <linux/crash_dump.h>
17 #include <asm/io_apic.h>
21 #include <asm/irq_remapping.h>
22 #include <asm/pci-direct.h>
23 #include <asm/msidef.h>
25 #include "../irq_remapping.h"
33 struct intel_iommu *iommu;
35 unsigned int bus; /* PCI bus number */
36 unsigned int devfn; /* PCI devfn number */
40 struct intel_iommu *iommu;
47 struct intel_iommu *iommu;
54 struct intel_ir_data {
55 struct irq_2_iommu irq_2_iommu;
56 struct irte irte_entry;
58 struct msi_msg msi_entry;
62 #define IR_X2APIC_MODE(mode) (mode ? (1 << 11) : 0)
63 #define IRTE_DEST(dest) ((eim_mode) ? dest : dest << 8)
65 static int __read_mostly eim_mode;
66 static struct ioapic_scope ir_ioapic[MAX_IO_APICS];
67 static struct hpet_scope ir_hpet[MAX_HPET_TBS];
74 * ->iommu->register_lock
76 * intel_irq_remap_ops.{supported,prepare,enable,disable,reenable} are called
77 * in single-threaded environment with interrupt disabled, so no need to tabke
78 * the dmar_global_lock.
80 DEFINE_RAW_SPINLOCK(irq_2_ir_lock);
81 static const struct irq_domain_ops intel_ir_domain_ops;
83 static void iommu_disable_irq_remapping(struct intel_iommu *iommu);
84 static int __init parse_ioapics_under_ir(void);
86 static bool ir_pre_enabled(struct intel_iommu *iommu)
88 return (iommu->flags & VTD_FLAG_IRQ_REMAP_PRE_ENABLED);
91 static void clear_ir_pre_enabled(struct intel_iommu *iommu)
93 iommu->flags &= ~VTD_FLAG_IRQ_REMAP_PRE_ENABLED;
96 static void init_ir_status(struct intel_iommu *iommu)
100 gsts = readl(iommu->reg + DMAR_GSTS_REG);
101 if (gsts & DMA_GSTS_IRES)
102 iommu->flags |= VTD_FLAG_IRQ_REMAP_PRE_ENABLED;
105 static int alloc_irte(struct intel_iommu *iommu,
106 struct irq_2_iommu *irq_iommu, u16 count)
108 struct ir_table *table = iommu->ir_table;
109 unsigned int mask = 0;
113 if (!count || !irq_iommu)
117 count = __roundup_pow_of_two(count);
121 if (mask > ecap_max_handle_mask(iommu->ecap)) {
122 pr_err("Requested mask %x exceeds the max invalidation handle"
123 " mask value %Lx\n", mask,
124 ecap_max_handle_mask(iommu->ecap));
128 raw_spin_lock_irqsave(&irq_2_ir_lock, flags);
129 index = bitmap_find_free_region(table->bitmap,
130 INTR_REMAP_TABLE_ENTRIES, mask);
132 pr_warn("IR%d: can't allocate an IRTE\n", iommu->seq_id);
134 irq_iommu->iommu = iommu;
135 irq_iommu->irte_index = index;
136 irq_iommu->sub_handle = 0;
137 irq_iommu->irte_mask = mask;
138 irq_iommu->mode = IRQ_REMAPPING;
140 raw_spin_unlock_irqrestore(&irq_2_ir_lock, flags);
145 static int qi_flush_iec(struct intel_iommu *iommu, int index, int mask)
149 desc.qw0 = QI_IEC_IIDEX(index) | QI_IEC_TYPE | QI_IEC_IM(mask)
155 return qi_submit_sync(iommu, &desc, 1, 0);
158 static int modify_irte(struct irq_2_iommu *irq_iommu,
159 struct irte *irte_modified)
161 struct intel_iommu *iommu;
169 raw_spin_lock_irqsave(&irq_2_ir_lock, flags);
171 iommu = irq_iommu->iommu;
173 index = irq_iommu->irte_index + irq_iommu->sub_handle;
174 irte = &iommu->ir_table->base[index];
176 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE)
177 if ((irte->pst == 1) || (irte_modified->pst == 1)) {
180 ret = cmpxchg_double(&irte->low, &irte->high,
181 irte->low, irte->high,
182 irte_modified->low, irte_modified->high);
184 * We use cmpxchg16 to atomically update the 128-bit IRTE,
185 * and it cannot be updated by the hardware or other processors
186 * behind us, so the return value of cmpxchg16 should be the
187 * same as the old value.
193 set_64bit(&irte->low, irte_modified->low);
194 set_64bit(&irte->high, irte_modified->high);
196 __iommu_flush_cache(iommu, irte, sizeof(*irte));
198 rc = qi_flush_iec(iommu, index, 0);
200 /* Update iommu mode according to the IRTE mode */
201 irq_iommu->mode = irte->pst ? IRQ_POSTING : IRQ_REMAPPING;
202 raw_spin_unlock_irqrestore(&irq_2_ir_lock, flags);
207 static struct intel_iommu *map_hpet_to_ir(u8 hpet_id)
211 for (i = 0; i < MAX_HPET_TBS; i++)
212 if (ir_hpet[i].id == hpet_id && ir_hpet[i].iommu)
213 return ir_hpet[i].iommu;
217 static struct intel_iommu *map_ioapic_to_ir(int apic)
221 for (i = 0; i < MAX_IO_APICS; i++)
222 if (ir_ioapic[i].id == apic && ir_ioapic[i].iommu)
223 return ir_ioapic[i].iommu;
227 static struct intel_iommu *map_dev_to_ir(struct pci_dev *dev)
229 struct dmar_drhd_unit *drhd;
231 drhd = dmar_find_matched_drhd_unit(dev);
238 static int clear_entries(struct irq_2_iommu *irq_iommu)
240 struct irte *start, *entry, *end;
241 struct intel_iommu *iommu;
244 if (irq_iommu->sub_handle)
247 iommu = irq_iommu->iommu;
248 index = irq_iommu->irte_index;
250 start = iommu->ir_table->base + index;
251 end = start + (1 << irq_iommu->irte_mask);
253 for (entry = start; entry < end; entry++) {
254 set_64bit(&entry->low, 0);
255 set_64bit(&entry->high, 0);
257 bitmap_release_region(iommu->ir_table->bitmap, index,
258 irq_iommu->irte_mask);
260 return qi_flush_iec(iommu, index, irq_iommu->irte_mask);
264 * source validation type
266 #define SVT_NO_VERIFY 0x0 /* no verification is required */
267 #define SVT_VERIFY_SID_SQ 0x1 /* verify using SID and SQ fields */
268 #define SVT_VERIFY_BUS 0x2 /* verify bus of request-id */
271 * source-id qualifier
273 #define SQ_ALL_16 0x0 /* verify all 16 bits of request-id */
274 #define SQ_13_IGNORE_1 0x1 /* verify most significant 13 bits, ignore
275 * the third least significant bit
277 #define SQ_13_IGNORE_2 0x2 /* verify most significant 13 bits, ignore
278 * the second and third least significant bits
280 #define SQ_13_IGNORE_3 0x3 /* verify most significant 13 bits, ignore
281 * the least three significant bits
285 * set SVT, SQ and SID fields of irte to verify
286 * source ids of interrupt requests
288 static void set_irte_sid(struct irte *irte, unsigned int svt,
289 unsigned int sq, unsigned int sid)
291 if (disable_sourceid_checking)
299 * Set an IRTE to match only the bus number. Interrupt requests that reference
300 * this IRTE must have a requester-id whose bus number is between or equal
301 * to the start_bus and end_bus arguments.
303 static void set_irte_verify_bus(struct irte *irte, unsigned int start_bus,
304 unsigned int end_bus)
306 set_irte_sid(irte, SVT_VERIFY_BUS, SQ_ALL_16,
307 (start_bus << 8) | end_bus);
310 static int set_ioapic_sid(struct irte *irte, int apic)
318 down_read(&dmar_global_lock);
319 for (i = 0; i < MAX_IO_APICS; i++) {
320 if (ir_ioapic[i].iommu && ir_ioapic[i].id == apic) {
321 sid = (ir_ioapic[i].bus << 8) | ir_ioapic[i].devfn;
325 up_read(&dmar_global_lock);
328 pr_warn("Failed to set source-id of IOAPIC (%d)\n", apic);
332 set_irte_sid(irte, SVT_VERIFY_SID_SQ, SQ_ALL_16, sid);
337 static int set_hpet_sid(struct irte *irte, u8 id)
345 down_read(&dmar_global_lock);
346 for (i = 0; i < MAX_HPET_TBS; i++) {
347 if (ir_hpet[i].iommu && ir_hpet[i].id == id) {
348 sid = (ir_hpet[i].bus << 8) | ir_hpet[i].devfn;
352 up_read(&dmar_global_lock);
355 pr_warn("Failed to set source-id of HPET block (%d)\n", id);
360 * Should really use SQ_ALL_16. Some platforms are broken.
361 * While we figure out the right quirks for these broken platforms, use
362 * SQ_13_IGNORE_3 for now.
364 set_irte_sid(irte, SVT_VERIFY_SID_SQ, SQ_13_IGNORE_3, sid);
369 struct set_msi_sid_data {
370 struct pci_dev *pdev;
376 static int set_msi_sid_cb(struct pci_dev *pdev, u16 alias, void *opaque)
378 struct set_msi_sid_data *data = opaque;
380 if (data->count == 0 || PCI_BUS_NUM(alias) == PCI_BUS_NUM(data->alias))
381 data->busmatch_count++;
390 static int set_msi_sid(struct irte *irte, struct pci_dev *dev)
392 struct set_msi_sid_data data;
398 data.busmatch_count = 0;
399 pci_for_each_dma_alias(dev, set_msi_sid_cb, &data);
402 * DMA alias provides us with a PCI device and alias. The only case
403 * where the it will return an alias on a different bus than the
404 * device is the case of a PCIe-to-PCI bridge, where the alias is for
405 * the subordinate bus. In this case we can only verify the bus.
407 * If there are multiple aliases, all with the same bus number,
408 * then all we can do is verify the bus. This is typical in NTB
409 * hardware which use proxy IDs where the device will generate traffic
410 * from multiple devfn numbers on the same bus.
412 * If the alias device is on a different bus than our source device
413 * then we have a topology based alias, use it.
415 * Otherwise, the alias is for a device DMA quirk and we cannot
416 * assume that MSI uses the same requester ID. Therefore use the
419 if (PCI_BUS_NUM(data.alias) != data.pdev->bus->number)
420 set_irte_verify_bus(irte, PCI_BUS_NUM(data.alias),
422 else if (data.count >= 2 && data.busmatch_count == data.count)
423 set_irte_verify_bus(irte, dev->bus->number, dev->bus->number);
424 else if (data.pdev->bus->number != dev->bus->number)
425 set_irte_sid(irte, SVT_VERIFY_SID_SQ, SQ_ALL_16, data.alias);
427 set_irte_sid(irte, SVT_VERIFY_SID_SQ, SQ_ALL_16,
433 static int iommu_load_old_irte(struct intel_iommu *iommu)
435 struct irte *old_ir_table;
436 phys_addr_t irt_phys;
441 /* Check whether the old ir-table has the same size as ours */
442 irta = dmar_readq(iommu->reg + DMAR_IRTA_REG);
443 if ((irta & INTR_REMAP_TABLE_REG_SIZE_MASK)
444 != INTR_REMAP_TABLE_REG_SIZE)
447 irt_phys = irta & VTD_PAGE_MASK;
448 size = INTR_REMAP_TABLE_ENTRIES*sizeof(struct irte);
450 /* Map the old IR table */
451 old_ir_table = memremap(irt_phys, size, MEMREMAP_WB);
456 memcpy(iommu->ir_table->base, old_ir_table, size);
458 __iommu_flush_cache(iommu, iommu->ir_table->base, size);
461 * Now check the table for used entries and mark those as
462 * allocated in the bitmap
464 for (i = 0; i < INTR_REMAP_TABLE_ENTRIES; i++) {
465 if (iommu->ir_table->base[i].present)
466 bitmap_set(iommu->ir_table->bitmap, i, 1);
469 memunmap(old_ir_table);
475 static void iommu_set_irq_remapping(struct intel_iommu *iommu, int mode)
481 addr = virt_to_phys((void *)iommu->ir_table->base);
483 raw_spin_lock_irqsave(&iommu->register_lock, flags);
485 dmar_writeq(iommu->reg + DMAR_IRTA_REG,
486 (addr) | IR_X2APIC_MODE(mode) | INTR_REMAP_TABLE_REG_SIZE);
488 /* Set interrupt-remapping table pointer */
489 writel(iommu->gcmd | DMA_GCMD_SIRTP, iommu->reg + DMAR_GCMD_REG);
491 IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
492 readl, (sts & DMA_GSTS_IRTPS), sts);
493 raw_spin_unlock_irqrestore(&iommu->register_lock, flags);
496 * Global invalidation of interrupt entry cache to make sure the
497 * hardware uses the new irq remapping table.
499 qi_global_iec(iommu);
502 static void iommu_enable_irq_remapping(struct intel_iommu *iommu)
507 raw_spin_lock_irqsave(&iommu->register_lock, flags);
509 /* Enable interrupt-remapping */
510 iommu->gcmd |= DMA_GCMD_IRE;
511 writel(iommu->gcmd, iommu->reg + DMAR_GCMD_REG);
512 IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
513 readl, (sts & DMA_GSTS_IRES), sts);
515 /* Block compatibility-format MSIs */
516 if (sts & DMA_GSTS_CFIS) {
517 iommu->gcmd &= ~DMA_GCMD_CFI;
518 writel(iommu->gcmd, iommu->reg + DMAR_GCMD_REG);
519 IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
520 readl, !(sts & DMA_GSTS_CFIS), sts);
524 * With CFI clear in the Global Command register, we should be
525 * protected from dangerous (i.e. compatibility) interrupts
526 * regardless of x2apic status. Check just to be sure.
528 if (sts & DMA_GSTS_CFIS)
530 "Compatibility-format IRQs enabled despite intr remapping;\n"
531 "you are vulnerable to IRQ injection.\n");
533 raw_spin_unlock_irqrestore(&iommu->register_lock, flags);
536 static int intel_setup_irq_remapping(struct intel_iommu *iommu)
538 struct ir_table *ir_table;
539 struct fwnode_handle *fn;
540 unsigned long *bitmap;
546 ir_table = kzalloc(sizeof(struct ir_table), GFP_KERNEL);
550 pages = alloc_pages_node(iommu->node, GFP_KERNEL | __GFP_ZERO,
551 INTR_REMAP_PAGE_ORDER);
553 pr_err("IR%d: failed to allocate pages of order %d\n",
554 iommu->seq_id, INTR_REMAP_PAGE_ORDER);
558 bitmap = bitmap_zalloc(INTR_REMAP_TABLE_ENTRIES, GFP_ATOMIC);
559 if (bitmap == NULL) {
560 pr_err("IR%d: failed to allocate bitmap\n", iommu->seq_id);
564 fn = irq_domain_alloc_named_id_fwnode("INTEL-IR", iommu->seq_id);
566 goto out_free_bitmap;
569 irq_domain_create_hierarchy(arch_get_ir_parent_domain(),
570 0, INTR_REMAP_TABLE_ENTRIES,
571 fn, &intel_ir_domain_ops,
573 if (!iommu->ir_domain) {
574 irq_domain_free_fwnode(fn);
575 pr_err("IR%d: failed to allocate irqdomain\n", iommu->seq_id);
576 goto out_free_bitmap;
578 iommu->ir_msi_domain =
579 arch_create_remap_msi_irq_domain(iommu->ir_domain,
583 ir_table->base = page_address(pages);
584 ir_table->bitmap = bitmap;
585 iommu->ir_table = ir_table;
588 * If the queued invalidation is already initialized,
589 * shouldn't disable it.
593 * Clear previous faults.
595 dmar_fault(-1, iommu);
596 dmar_disable_qi(iommu);
598 if (dmar_enable_qi(iommu)) {
599 pr_err("Failed to enable queued invalidation\n");
600 goto out_free_bitmap;
604 init_ir_status(iommu);
606 if (ir_pre_enabled(iommu)) {
607 if (!is_kdump_kernel()) {
608 pr_warn("IRQ remapping was enabled on %s but we are not in kdump mode\n",
610 clear_ir_pre_enabled(iommu);
611 iommu_disable_irq_remapping(iommu);
612 } else if (iommu_load_old_irte(iommu))
613 pr_err("Failed to copy IR table for %s from previous kernel\n",
616 pr_info("Copied IR table for %s from previous kernel\n",
620 iommu_set_irq_remapping(iommu, eim_mode);
627 __free_pages(pages, INTR_REMAP_PAGE_ORDER);
631 iommu->ir_table = NULL;
636 static void intel_teardown_irq_remapping(struct intel_iommu *iommu)
638 struct fwnode_handle *fn;
640 if (iommu && iommu->ir_table) {
641 if (iommu->ir_msi_domain) {
642 fn = iommu->ir_msi_domain->fwnode;
644 irq_domain_remove(iommu->ir_msi_domain);
645 irq_domain_free_fwnode(fn);
646 iommu->ir_msi_domain = NULL;
648 if (iommu->ir_domain) {
649 fn = iommu->ir_domain->fwnode;
651 irq_domain_remove(iommu->ir_domain);
652 irq_domain_free_fwnode(fn);
653 iommu->ir_domain = NULL;
655 free_pages((unsigned long)iommu->ir_table->base,
656 INTR_REMAP_PAGE_ORDER);
657 bitmap_free(iommu->ir_table->bitmap);
658 kfree(iommu->ir_table);
659 iommu->ir_table = NULL;
664 * Disable Interrupt Remapping.
666 static void iommu_disable_irq_remapping(struct intel_iommu *iommu)
671 if (!ecap_ir_support(iommu->ecap))
675 * global invalidation of interrupt entry cache before disabling
676 * interrupt-remapping.
678 qi_global_iec(iommu);
680 raw_spin_lock_irqsave(&iommu->register_lock, flags);
682 sts = readl(iommu->reg + DMAR_GSTS_REG);
683 if (!(sts & DMA_GSTS_IRES))
686 iommu->gcmd &= ~DMA_GCMD_IRE;
687 writel(iommu->gcmd, iommu->reg + DMAR_GCMD_REG);
689 IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
690 readl, !(sts & DMA_GSTS_IRES), sts);
693 raw_spin_unlock_irqrestore(&iommu->register_lock, flags);
696 static int __init dmar_x2apic_optout(void)
698 struct acpi_table_dmar *dmar;
699 dmar = (struct acpi_table_dmar *)dmar_tbl;
700 if (!dmar || no_x2apic_optout)
702 return dmar->flags & DMAR_X2APIC_OPT_OUT;
705 static void __init intel_cleanup_irq_remapping(void)
707 struct dmar_drhd_unit *drhd;
708 struct intel_iommu *iommu;
710 for_each_iommu(iommu, drhd) {
711 if (ecap_ir_support(iommu->ecap)) {
712 iommu_disable_irq_remapping(iommu);
713 intel_teardown_irq_remapping(iommu);
717 if (x2apic_supported())
718 pr_warn("Failed to enable irq remapping. You are vulnerable to irq-injection attacks.\n");
721 static int __init intel_prepare_irq_remapping(void)
723 struct dmar_drhd_unit *drhd;
724 struct intel_iommu *iommu;
727 if (irq_remap_broken) {
728 pr_warn("This system BIOS has enabled interrupt remapping\n"
729 "on a chipset that contains an erratum making that\n"
730 "feature unstable. To maintain system stability\n"
731 "interrupt remapping is being disabled. Please\n"
732 "contact your BIOS vendor for an update\n");
733 add_taint(TAINT_FIRMWARE_WORKAROUND, LOCKDEP_STILL_OK);
737 if (dmar_table_init() < 0)
740 if (!dmar_ir_support())
743 if (parse_ioapics_under_ir()) {
744 pr_info("Not enabling interrupt remapping\n");
748 /* First make sure all IOMMUs support IRQ remapping */
749 for_each_iommu(iommu, drhd)
750 if (!ecap_ir_support(iommu->ecap))
753 /* Detect remapping mode: lapic or x2apic */
754 if (x2apic_supported()) {
755 eim = !dmar_x2apic_optout();
757 pr_info("x2apic is disabled because BIOS sets x2apic opt out bit.");
758 pr_info("Use 'intremap=no_x2apic_optout' to override the BIOS setting.\n");
762 for_each_iommu(iommu, drhd) {
763 if (eim && !ecap_eim_support(iommu->ecap)) {
764 pr_info("%s does not support EIM\n", iommu->name);
771 pr_info("Queued invalidation will be enabled to support x2apic and Intr-remapping.\n");
773 /* Do the initializations early */
774 for_each_iommu(iommu, drhd) {
775 if (intel_setup_irq_remapping(iommu)) {
776 pr_err("Failed to setup irq remapping for %s\n",
785 intel_cleanup_irq_remapping();
790 * Set Posted-Interrupts capability.
792 static inline void set_irq_posting_cap(void)
794 struct dmar_drhd_unit *drhd;
795 struct intel_iommu *iommu;
797 if (!disable_irq_post) {
799 * If IRTE is in posted format, the 'pda' field goes across the
800 * 64-bit boundary, we need use cmpxchg16b to atomically update
801 * it. We only expose posted-interrupt when X86_FEATURE_CX16
802 * is supported. Actually, hardware platforms supporting PI
803 * should have X86_FEATURE_CX16 support, this has been confirmed
804 * with Intel hardware guys.
806 if (boot_cpu_has(X86_FEATURE_CX16))
807 intel_irq_remap_ops.capability |= 1 << IRQ_POSTING_CAP;
809 for_each_iommu(iommu, drhd)
810 if (!cap_pi_support(iommu->cap)) {
811 intel_irq_remap_ops.capability &=
812 ~(1 << IRQ_POSTING_CAP);
818 static int __init intel_enable_irq_remapping(void)
820 struct dmar_drhd_unit *drhd;
821 struct intel_iommu *iommu;
825 * Setup Interrupt-remapping for all the DRHD's now.
827 for_each_iommu(iommu, drhd) {
828 if (!ir_pre_enabled(iommu))
829 iommu_enable_irq_remapping(iommu);
836 irq_remapping_enabled = 1;
838 set_irq_posting_cap();
840 pr_info("Enabled IRQ remapping in %s mode\n", eim_mode ? "x2apic" : "xapic");
842 return eim_mode ? IRQ_REMAP_X2APIC_MODE : IRQ_REMAP_XAPIC_MODE;
845 intel_cleanup_irq_remapping();
849 static int ir_parse_one_hpet_scope(struct acpi_dmar_device_scope *scope,
850 struct intel_iommu *iommu,
851 struct acpi_dmar_hardware_unit *drhd)
853 struct acpi_dmar_pci_path *path;
855 int count, free = -1;
858 path = (struct acpi_dmar_pci_path *)(scope + 1);
859 count = (scope->length - sizeof(struct acpi_dmar_device_scope))
860 / sizeof(struct acpi_dmar_pci_path);
862 while (--count > 0) {
864 * Access PCI directly due to the PCI
865 * subsystem isn't initialized yet.
867 bus = read_pci_config_byte(bus, path->device, path->function,
872 for (count = 0; count < MAX_HPET_TBS; count++) {
873 if (ir_hpet[count].iommu == iommu &&
874 ir_hpet[count].id == scope->enumeration_id)
876 else if (ir_hpet[count].iommu == NULL && free == -1)
880 pr_warn("Exceeded Max HPET blocks\n");
884 ir_hpet[free].iommu = iommu;
885 ir_hpet[free].id = scope->enumeration_id;
886 ir_hpet[free].bus = bus;
887 ir_hpet[free].devfn = PCI_DEVFN(path->device, path->function);
888 pr_info("HPET id %d under DRHD base 0x%Lx\n",
889 scope->enumeration_id, drhd->address);
894 static int ir_parse_one_ioapic_scope(struct acpi_dmar_device_scope *scope,
895 struct intel_iommu *iommu,
896 struct acpi_dmar_hardware_unit *drhd)
898 struct acpi_dmar_pci_path *path;
900 int count, free = -1;
903 path = (struct acpi_dmar_pci_path *)(scope + 1);
904 count = (scope->length - sizeof(struct acpi_dmar_device_scope))
905 / sizeof(struct acpi_dmar_pci_path);
907 while (--count > 0) {
909 * Access PCI directly due to the PCI
910 * subsystem isn't initialized yet.
912 bus = read_pci_config_byte(bus, path->device, path->function,
917 for (count = 0; count < MAX_IO_APICS; count++) {
918 if (ir_ioapic[count].iommu == iommu &&
919 ir_ioapic[count].id == scope->enumeration_id)
921 else if (ir_ioapic[count].iommu == NULL && free == -1)
925 pr_warn("Exceeded Max IO APICS\n");
929 ir_ioapic[free].bus = bus;
930 ir_ioapic[free].devfn = PCI_DEVFN(path->device, path->function);
931 ir_ioapic[free].iommu = iommu;
932 ir_ioapic[free].id = scope->enumeration_id;
933 pr_info("IOAPIC id %d under DRHD base 0x%Lx IOMMU %d\n",
934 scope->enumeration_id, drhd->address, iommu->seq_id);
939 static int ir_parse_ioapic_hpet_scope(struct acpi_dmar_header *header,
940 struct intel_iommu *iommu)
943 struct acpi_dmar_hardware_unit *drhd;
944 struct acpi_dmar_device_scope *scope;
947 drhd = (struct acpi_dmar_hardware_unit *)header;
948 start = (void *)(drhd + 1);
949 end = ((void *)drhd) + header->length;
951 while (start < end && ret == 0) {
953 if (scope->entry_type == ACPI_DMAR_SCOPE_TYPE_IOAPIC)
954 ret = ir_parse_one_ioapic_scope(scope, iommu, drhd);
955 else if (scope->entry_type == ACPI_DMAR_SCOPE_TYPE_HPET)
956 ret = ir_parse_one_hpet_scope(scope, iommu, drhd);
957 start += scope->length;
963 static void ir_remove_ioapic_hpet_scope(struct intel_iommu *iommu)
967 for (i = 0; i < MAX_HPET_TBS; i++)
968 if (ir_hpet[i].iommu == iommu)
969 ir_hpet[i].iommu = NULL;
971 for (i = 0; i < MAX_IO_APICS; i++)
972 if (ir_ioapic[i].iommu == iommu)
973 ir_ioapic[i].iommu = NULL;
977 * Finds the assocaition between IOAPIC's and its Interrupt-remapping
980 static int __init parse_ioapics_under_ir(void)
982 struct dmar_drhd_unit *drhd;
983 struct intel_iommu *iommu;
984 bool ir_supported = false;
987 for_each_iommu(iommu, drhd) {
990 if (!ecap_ir_support(iommu->ecap))
993 ret = ir_parse_ioapic_hpet_scope(drhd->hdr, iommu);
1003 for (ioapic_idx = 0; ioapic_idx < nr_ioapics; ioapic_idx++) {
1004 int ioapic_id = mpc_ioapic_id(ioapic_idx);
1005 if (!map_ioapic_to_ir(ioapic_id)) {
1006 pr_err(FW_BUG "ioapic %d has no mapping iommu, "
1007 "interrupt remapping will be disabled\n",
1016 static int __init ir_dev_scope_init(void)
1020 if (!irq_remapping_enabled)
1023 down_write(&dmar_global_lock);
1024 ret = dmar_dev_scope_init();
1025 up_write(&dmar_global_lock);
1029 rootfs_initcall(ir_dev_scope_init);
1031 static void disable_irq_remapping(void)
1033 struct dmar_drhd_unit *drhd;
1034 struct intel_iommu *iommu = NULL;
1037 * Disable Interrupt-remapping for all the DRHD's now.
1039 for_each_iommu(iommu, drhd) {
1040 if (!ecap_ir_support(iommu->ecap))
1043 iommu_disable_irq_remapping(iommu);
1047 * Clear Posted-Interrupts capability.
1049 if (!disable_irq_post)
1050 intel_irq_remap_ops.capability &= ~(1 << IRQ_POSTING_CAP);
1053 static int reenable_irq_remapping(int eim)
1055 struct dmar_drhd_unit *drhd;
1057 struct intel_iommu *iommu = NULL;
1059 for_each_iommu(iommu, drhd)
1061 dmar_reenable_qi(iommu);
1064 * Setup Interrupt-remapping for all the DRHD's now.
1066 for_each_iommu(iommu, drhd) {
1067 if (!ecap_ir_support(iommu->ecap))
1070 /* Set up interrupt remapping for iommu.*/
1071 iommu_set_irq_remapping(iommu, eim);
1072 iommu_enable_irq_remapping(iommu);
1079 set_irq_posting_cap();
1085 * handle error condition gracefully here!
1090 static void prepare_irte(struct irte *irte, int vector, unsigned int dest)
1092 memset(irte, 0, sizeof(*irte));
1095 irte->dst_mode = apic->irq_dest_mode;
1097 * Trigger mode in the IRTE will always be edge, and for IO-APIC, the
1098 * actual level or edge trigger will be setup in the IO-APIC
1099 * RTE. This will help simplify level triggered irq migration.
1100 * For more details, see the comments (in io_apic.c) explainig IO-APIC
1101 * irq migration in the presence of interrupt-remapping.
1103 irte->trigger_mode = 0;
1104 irte->dlvry_mode = apic->irq_delivery_mode;
1105 irte->vector = vector;
1106 irte->dest_id = IRTE_DEST(dest);
1107 irte->redir_hint = 1;
1110 static struct irq_domain *intel_get_ir_irq_domain(struct irq_alloc_info *info)
1112 struct intel_iommu *iommu = NULL;
1117 switch (info->type) {
1118 case X86_IRQ_ALLOC_TYPE_IOAPIC_GET_PARENT:
1119 iommu = map_ioapic_to_ir(info->ioapic_id);
1121 case X86_IRQ_ALLOC_TYPE_HPET_GET_PARENT:
1122 iommu = map_hpet_to_ir(info->hpet_id);
1129 return iommu ? iommu->ir_domain : NULL;
1132 static struct irq_domain *intel_get_irq_domain(struct irq_alloc_info *info)
1134 struct intel_iommu *iommu;
1139 switch (info->type) {
1140 case X86_IRQ_ALLOC_TYPE_PCI_MSI:
1141 case X86_IRQ_ALLOC_TYPE_PCI_MSIX:
1142 iommu = map_dev_to_ir(info->msi_dev);
1144 return iommu->ir_msi_domain;
1153 struct irq_remap_ops intel_irq_remap_ops = {
1154 .prepare = intel_prepare_irq_remapping,
1155 .enable = intel_enable_irq_remapping,
1156 .disable = disable_irq_remapping,
1157 .reenable = reenable_irq_remapping,
1158 .enable_faulting = enable_drhd_fault_handling,
1159 .get_ir_irq_domain = intel_get_ir_irq_domain,
1160 .get_irq_domain = intel_get_irq_domain,
1163 static void intel_ir_reconfigure_irte(struct irq_data *irqd, bool force)
1165 struct intel_ir_data *ir_data = irqd->chip_data;
1166 struct irte *irte = &ir_data->irte_entry;
1167 struct irq_cfg *cfg = irqd_cfg(irqd);
1170 * Atomically updates the IRTE with the new destination, vector
1171 * and flushes the interrupt entry cache.
1173 irte->vector = cfg->vector;
1174 irte->dest_id = IRTE_DEST(cfg->dest_apicid);
1176 /* Update the hardware only if the interrupt is in remapped mode. */
1177 if (force || ir_data->irq_2_iommu.mode == IRQ_REMAPPING)
1178 modify_irte(&ir_data->irq_2_iommu, irte);
1182 * Migrate the IO-APIC irq in the presence of intr-remapping.
1184 * For both level and edge triggered, irq migration is a simple atomic
1185 * update(of vector and cpu destination) of IRTE and flush the hardware cache.
1187 * For level triggered, we eliminate the io-apic RTE modification (with the
1188 * updated vector information), by using a virtual vector (io-apic pin number).
1189 * Real vector that is used for interrupting cpu will be coming from
1190 * the interrupt-remapping table entry.
1192 * As the migration is a simple atomic update of IRTE, the same mechanism
1193 * is used to migrate MSI irq's in the presence of interrupt-remapping.
1196 intel_ir_set_affinity(struct irq_data *data, const struct cpumask *mask,
1199 struct irq_data *parent = data->parent_data;
1200 struct irq_cfg *cfg = irqd_cfg(data);
1203 ret = parent->chip->irq_set_affinity(parent, mask, force);
1204 if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
1207 intel_ir_reconfigure_irte(data, false);
1209 * After this point, all the interrupts will start arriving
1210 * at the new destination. So, time to cleanup the previous
1211 * vector allocation.
1213 send_cleanup_vector(cfg);
1215 return IRQ_SET_MASK_OK_DONE;
1218 static void intel_ir_compose_msi_msg(struct irq_data *irq_data,
1219 struct msi_msg *msg)
1221 struct intel_ir_data *ir_data = irq_data->chip_data;
1223 *msg = ir_data->msi_entry;
1226 static int intel_ir_set_vcpu_affinity(struct irq_data *data, void *info)
1228 struct intel_ir_data *ir_data = data->chip_data;
1229 struct vcpu_data *vcpu_pi_info = info;
1231 /* stop posting interrupts, back to remapping mode */
1232 if (!vcpu_pi_info) {
1233 modify_irte(&ir_data->irq_2_iommu, &ir_data->irte_entry);
1235 struct irte irte_pi;
1238 * We are not caching the posted interrupt entry. We
1239 * copy the data from the remapped entry and modify
1240 * the fields which are relevant for posted mode. The
1241 * cached remapped entry is used for switching back to
1244 memset(&irte_pi, 0, sizeof(irte_pi));
1245 dmar_copy_shared_irte(&irte_pi, &ir_data->irte_entry);
1247 /* Update the posted mode fields */
1249 irte_pi.p_urgent = 0;
1250 irte_pi.p_vector = vcpu_pi_info->vector;
1251 irte_pi.pda_l = (vcpu_pi_info->pi_desc_addr >>
1252 (32 - PDA_LOW_BIT)) & ~(-1UL << PDA_LOW_BIT);
1253 irte_pi.pda_h = (vcpu_pi_info->pi_desc_addr >> 32) &
1254 ~(-1UL << PDA_HIGH_BIT);
1256 modify_irte(&ir_data->irq_2_iommu, &irte_pi);
1262 static struct irq_chip intel_ir_chip = {
1264 .irq_ack = apic_ack_irq,
1265 .irq_set_affinity = intel_ir_set_affinity,
1266 .irq_compose_msi_msg = intel_ir_compose_msi_msg,
1267 .irq_set_vcpu_affinity = intel_ir_set_vcpu_affinity,
1270 static void intel_irq_remapping_prepare_irte(struct intel_ir_data *data,
1271 struct irq_cfg *irq_cfg,
1272 struct irq_alloc_info *info,
1273 int index, int sub_handle)
1275 struct IR_IO_APIC_route_entry *entry;
1276 struct irte *irte = &data->irte_entry;
1277 struct msi_msg *msg = &data->msi_entry;
1279 prepare_irte(irte, irq_cfg->vector, irq_cfg->dest_apicid);
1280 switch (info->type) {
1281 case X86_IRQ_ALLOC_TYPE_IOAPIC:
1282 /* Set source-id of interrupt request */
1283 set_ioapic_sid(irte, info->ioapic_id);
1284 apic_printk(APIC_VERBOSE, KERN_DEBUG "IOAPIC[%d]: Set IRTE entry (P:%d FPD:%d Dst_Mode:%d Redir_hint:%d Trig_Mode:%d Dlvry_Mode:%X Avail:%X Vector:%02X Dest:%08X SID:%04X SQ:%X SVT:%X)\n",
1285 info->ioapic_id, irte->present, irte->fpd,
1286 irte->dst_mode, irte->redir_hint,
1287 irte->trigger_mode, irte->dlvry_mode,
1288 irte->avail, irte->vector, irte->dest_id,
1289 irte->sid, irte->sq, irte->svt);
1291 entry = (struct IR_IO_APIC_route_entry *)info->ioapic_entry;
1292 info->ioapic_entry = NULL;
1293 memset(entry, 0, sizeof(*entry));
1294 entry->index2 = (index >> 15) & 0x1;
1297 entry->index = (index & 0x7fff);
1299 * IO-APIC RTE will be configured with virtual vector.
1300 * irq handler will do the explicit EOI to the io-apic.
1302 entry->vector = info->ioapic_pin;
1303 entry->mask = 0; /* enable IRQ */
1304 entry->trigger = info->ioapic_trigger;
1305 entry->polarity = info->ioapic_polarity;
1306 if (info->ioapic_trigger)
1307 entry->mask = 1; /* Mask level triggered irqs. */
1310 case X86_IRQ_ALLOC_TYPE_HPET:
1311 case X86_IRQ_ALLOC_TYPE_PCI_MSI:
1312 case X86_IRQ_ALLOC_TYPE_PCI_MSIX:
1313 if (info->type == X86_IRQ_ALLOC_TYPE_HPET)
1314 set_hpet_sid(irte, info->hpet_id);
1316 set_msi_sid(irte, info->msi_dev);
1318 msg->address_hi = MSI_ADDR_BASE_HI;
1319 msg->data = sub_handle;
1320 msg->address_lo = MSI_ADDR_BASE_LO | MSI_ADDR_IR_EXT_INT |
1322 MSI_ADDR_IR_INDEX1(index) |
1323 MSI_ADDR_IR_INDEX2(index);
1332 static void intel_free_irq_resources(struct irq_domain *domain,
1333 unsigned int virq, unsigned int nr_irqs)
1335 struct irq_data *irq_data;
1336 struct intel_ir_data *data;
1337 struct irq_2_iommu *irq_iommu;
1338 unsigned long flags;
1340 for (i = 0; i < nr_irqs; i++) {
1341 irq_data = irq_domain_get_irq_data(domain, virq + i);
1342 if (irq_data && irq_data->chip_data) {
1343 data = irq_data->chip_data;
1344 irq_iommu = &data->irq_2_iommu;
1345 raw_spin_lock_irqsave(&irq_2_ir_lock, flags);
1346 clear_entries(irq_iommu);
1347 raw_spin_unlock_irqrestore(&irq_2_ir_lock, flags);
1348 irq_domain_reset_irq_data(irq_data);
1354 static int intel_irq_remapping_alloc(struct irq_domain *domain,
1355 unsigned int virq, unsigned int nr_irqs,
1358 struct intel_iommu *iommu = domain->host_data;
1359 struct irq_alloc_info *info = arg;
1360 struct intel_ir_data *data, *ird;
1361 struct irq_data *irq_data;
1362 struct irq_cfg *irq_cfg;
1365 if (!info || !iommu)
1367 if (nr_irqs > 1 && info->type != X86_IRQ_ALLOC_TYPE_PCI_MSI &&
1368 info->type != X86_IRQ_ALLOC_TYPE_PCI_MSIX)
1372 * With IRQ remapping enabled, don't need contiguous CPU vectors
1373 * to support multiple MSI interrupts.
1375 if (info->type == X86_IRQ_ALLOC_TYPE_PCI_MSI)
1376 info->flags &= ~X86_IRQ_ALLOC_CONTIGUOUS_VECTORS;
1378 ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
1383 data = kzalloc(sizeof(*data), GFP_KERNEL);
1385 goto out_free_parent;
1387 down_read(&dmar_global_lock);
1388 index = alloc_irte(iommu, &data->irq_2_iommu, nr_irqs);
1389 up_read(&dmar_global_lock);
1391 pr_warn("Failed to allocate IRTE\n");
1393 goto out_free_parent;
1396 for (i = 0; i < nr_irqs; i++) {
1397 irq_data = irq_domain_get_irq_data(domain, virq + i);
1398 irq_cfg = irqd_cfg(irq_data);
1399 if (!irq_data || !irq_cfg) {
1405 ird = kzalloc(sizeof(*ird), GFP_KERNEL);
1408 /* Initialize the common data */
1409 ird->irq_2_iommu = data->irq_2_iommu;
1410 ird->irq_2_iommu.sub_handle = i;
1415 irq_data->hwirq = (index << 16) + i;
1416 irq_data->chip_data = ird;
1417 irq_data->chip = &intel_ir_chip;
1418 intel_irq_remapping_prepare_irte(ird, irq_cfg, info, index, i);
1419 irq_set_status_flags(virq + i, IRQ_MOVE_PCNTXT);
1424 intel_free_irq_resources(domain, virq, i);
1426 irq_domain_free_irqs_common(domain, virq, nr_irqs);
1430 static void intel_irq_remapping_free(struct irq_domain *domain,
1431 unsigned int virq, unsigned int nr_irqs)
1433 intel_free_irq_resources(domain, virq, nr_irqs);
1434 irq_domain_free_irqs_common(domain, virq, nr_irqs);
1437 static int intel_irq_remapping_activate(struct irq_domain *domain,
1438 struct irq_data *irq_data, bool reserve)
1440 intel_ir_reconfigure_irte(irq_data, true);
1444 static void intel_irq_remapping_deactivate(struct irq_domain *domain,
1445 struct irq_data *irq_data)
1447 struct intel_ir_data *data = irq_data->chip_data;
1450 memset(&entry, 0, sizeof(entry));
1451 modify_irte(&data->irq_2_iommu, &entry);
1454 static const struct irq_domain_ops intel_ir_domain_ops = {
1455 .alloc = intel_irq_remapping_alloc,
1456 .free = intel_irq_remapping_free,
1457 .activate = intel_irq_remapping_activate,
1458 .deactivate = intel_irq_remapping_deactivate,
1462 * Support of Interrupt Remapping Unit Hotplug
1464 static int dmar_ir_add(struct dmar_drhd_unit *dmaru, struct intel_iommu *iommu)
1467 int eim = x2apic_enabled();
1469 if (eim && !ecap_eim_support(iommu->ecap)) {
1470 pr_info("DRHD %Lx: EIM not supported by DRHD, ecap %Lx\n",
1471 iommu->reg_phys, iommu->ecap);
1475 if (ir_parse_ioapic_hpet_scope(dmaru->hdr, iommu)) {
1476 pr_warn("DRHD %Lx: failed to parse managed IOAPIC/HPET\n",
1481 /* TODO: check all IOAPICs are covered by IOMMU */
1483 /* Setup Interrupt-remapping now. */
1484 ret = intel_setup_irq_remapping(iommu);
1486 pr_err("Failed to setup irq remapping for %s\n",
1488 intel_teardown_irq_remapping(iommu);
1489 ir_remove_ioapic_hpet_scope(iommu);
1491 iommu_enable_irq_remapping(iommu);
1497 int dmar_ir_hotplug(struct dmar_drhd_unit *dmaru, bool insert)
1500 struct intel_iommu *iommu = dmaru->iommu;
1502 if (!irq_remapping_enabled)
1506 if (!ecap_ir_support(iommu->ecap))
1508 if (irq_remapping_cap(IRQ_POSTING_CAP) &&
1509 !cap_pi_support(iommu->cap))
1513 if (!iommu->ir_table)
1514 ret = dmar_ir_add(dmaru, iommu);
1516 if (iommu->ir_table) {
1517 if (!bitmap_empty(iommu->ir_table->bitmap,
1518 INTR_REMAP_TABLE_ENTRIES)) {
1521 iommu_disable_irq_remapping(iommu);
1522 intel_teardown_irq_remapping(iommu);
1523 ir_remove_ioapic_hpet_scope(iommu);