1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (C) 2007-2010 Advanced Micro Devices, Inc.
4 * Author: Joerg Roedel <jroedel@suse.de>
5 * Leo Duran <leo.duran@amd.com>
8 #define pr_fmt(fmt) "AMD-Vi: " fmt
9 #define dev_fmt(fmt) pr_fmt(fmt)
11 #include <linux/ratelimit.h>
12 #include <linux/pci.h>
13 #include <linux/acpi.h>
14 #include <linux/amba/bus.h>
15 #include <linux/platform_device.h>
16 #include <linux/pci-ats.h>
17 #include <linux/bitmap.h>
18 #include <linux/slab.h>
19 #include <linux/debugfs.h>
20 #include <linux/scatterlist.h>
21 #include <linux/dma-mapping.h>
22 #include <linux/dma-direct.h>
23 #include <linux/iommu-helper.h>
24 #include <linux/iommu.h>
25 #include <linux/delay.h>
26 #include <linux/amd-iommu.h>
27 #include <linux/notifier.h>
28 #include <linux/export.h>
29 #include <linux/irq.h>
30 #include <linux/msi.h>
31 #include <linux/dma-contiguous.h>
32 #include <linux/irqdomain.h>
33 #include <linux/percpu.h>
34 #include <linux/iova.h>
35 #include <asm/irq_remapping.h>
36 #include <asm/io_apic.h>
38 #include <asm/hw_irq.h>
39 #include <asm/msidef.h>
40 #include <asm/proto.h>
41 #include <asm/iommu.h>
45 #include "amd_iommu_proto.h"
46 #include "amd_iommu_types.h"
47 #include "irq_remapping.h"
49 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
51 #define LOOP_TIMEOUT 100000
53 /* IO virtual address start page frame number */
54 #define IOVA_START_PFN (1)
55 #define IOVA_PFN(addr) ((addr) >> PAGE_SHIFT)
57 /* Reserved IOVA ranges */
58 #define MSI_RANGE_START (0xfee00000)
59 #define MSI_RANGE_END (0xfeefffff)
60 #define HT_RANGE_START (0xfd00000000ULL)
61 #define HT_RANGE_END (0xffffffffffULL)
64 * This bitmap is used to advertise the page sizes our hardware support
65 * to the IOMMU core, which will then use this information to split
66 * physically contiguous memory regions it is mapping into page sizes
69 * 512GB Pages are not supported due to a hardware bug
71 #define AMD_IOMMU_PGSIZES ((~0xFFFUL) & ~(2ULL << 38))
73 static DEFINE_SPINLOCK(amd_iommu_devtable_lock);
74 static DEFINE_SPINLOCK(pd_bitmap_lock);
76 /* List of all available dev_data structures */
77 static LLIST_HEAD(dev_data_list);
79 LIST_HEAD(ioapic_map);
81 LIST_HEAD(acpihid_map);
84 * Domain for untranslated devices - only allocated
85 * if iommu=pt passed on kernel cmd line.
87 const struct iommu_ops amd_iommu_ops;
89 static ATOMIC_NOTIFIER_HEAD(ppr_notifier);
90 int amd_iommu_max_glx_val = -1;
92 static const struct dma_map_ops amd_iommu_dma_ops;
95 * general struct to manage commands send to an IOMMU
101 struct kmem_cache *amd_iommu_irq_cache;
103 static void update_domain(struct protection_domain *domain);
104 static int protection_domain_init(struct protection_domain *domain);
105 static void detach_device(struct device *dev);
106 static void iova_domain_flush_tlb(struct iova_domain *iovad);
109 * Data container for a dma_ops specific protection domain
111 struct dma_ops_domain {
112 /* generic protection domain information */
113 struct protection_domain domain;
116 struct iova_domain iovad;
119 static struct iova_domain reserved_iova_ranges;
120 static struct lock_class_key reserved_rbtree_key;
122 /****************************************************************************
126 ****************************************************************************/
128 static inline int match_hid_uid(struct device *dev,
129 struct acpihid_map_entry *entry)
131 struct acpi_device *adev = ACPI_COMPANION(dev);
132 const char *hid, *uid;
137 hid = acpi_device_hid(adev);
138 uid = acpi_device_uid(adev);
144 return strcmp(hid, entry->hid);
147 return strcmp(hid, entry->hid);
149 return (strcmp(hid, entry->hid) || strcmp(uid, entry->uid));
152 static inline u16 get_pci_device_id(struct device *dev)
154 struct pci_dev *pdev = to_pci_dev(dev);
156 return pci_dev_id(pdev);
159 static inline int get_acpihid_device_id(struct device *dev,
160 struct acpihid_map_entry **entry)
162 struct acpihid_map_entry *p;
164 list_for_each_entry(p, &acpihid_map, list) {
165 if (!match_hid_uid(dev, p)) {
174 static inline int get_device_id(struct device *dev)
179 devid = get_pci_device_id(dev);
181 devid = get_acpihid_device_id(dev, NULL);
186 static struct protection_domain *to_pdomain(struct iommu_domain *dom)
188 return container_of(dom, struct protection_domain, domain);
191 static struct dma_ops_domain* to_dma_ops_domain(struct protection_domain *domain)
193 BUG_ON(domain->flags != PD_DMA_OPS_MASK);
194 return container_of(domain, struct dma_ops_domain, domain);
197 static struct iommu_dev_data *alloc_dev_data(u16 devid)
199 struct iommu_dev_data *dev_data;
201 dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
205 dev_data->devid = devid;
206 ratelimit_default_init(&dev_data->rs);
208 llist_add(&dev_data->dev_data_list, &dev_data_list);
212 static struct iommu_dev_data *search_dev_data(u16 devid)
214 struct iommu_dev_data *dev_data;
215 struct llist_node *node;
217 if (llist_empty(&dev_data_list))
220 node = dev_data_list.first;
221 llist_for_each_entry(dev_data, node, dev_data_list) {
222 if (dev_data->devid == devid)
229 static int __last_alias(struct pci_dev *pdev, u16 alias, void *data)
231 *(u16 *)data = alias;
235 static u16 get_alias(struct device *dev)
237 struct pci_dev *pdev = to_pci_dev(dev);
238 u16 devid, ivrs_alias, pci_alias;
240 /* The callers make sure that get_device_id() does not fail here */
241 devid = get_device_id(dev);
243 /* For ACPI HID devices, we simply return the devid as such */
244 if (!dev_is_pci(dev))
247 ivrs_alias = amd_iommu_alias_table[devid];
249 pci_for_each_dma_alias(pdev, __last_alias, &pci_alias);
251 if (ivrs_alias == pci_alias)
257 * The IVRS is fairly reliable in telling us about aliases, but it
258 * can't know about every screwy device. If we don't have an IVRS
259 * reported alias, use the PCI reported alias. In that case we may
260 * still need to initialize the rlookup and dev_table entries if the
261 * alias is to a non-existent device.
263 if (ivrs_alias == devid) {
264 if (!amd_iommu_rlookup_table[pci_alias]) {
265 amd_iommu_rlookup_table[pci_alias] =
266 amd_iommu_rlookup_table[devid];
267 memcpy(amd_iommu_dev_table[pci_alias].data,
268 amd_iommu_dev_table[devid].data,
269 sizeof(amd_iommu_dev_table[pci_alias].data));
275 pci_info(pdev, "Using IVRS reported alias %02x:%02x.%d "
276 "for device [%04x:%04x], kernel reported alias "
277 "%02x:%02x.%d\n", PCI_BUS_NUM(ivrs_alias), PCI_SLOT(ivrs_alias),
278 PCI_FUNC(ivrs_alias), pdev->vendor, pdev->device,
279 PCI_BUS_NUM(pci_alias), PCI_SLOT(pci_alias),
280 PCI_FUNC(pci_alias));
283 * If we don't have a PCI DMA alias and the IVRS alias is on the same
284 * bus, then the IVRS table may know about a quirk that we don't.
286 if (pci_alias == devid &&
287 PCI_BUS_NUM(ivrs_alias) == pdev->bus->number) {
288 pci_add_dma_alias(pdev, ivrs_alias & 0xff);
289 pci_info(pdev, "Added PCI DMA alias %02x.%d\n",
290 PCI_SLOT(ivrs_alias), PCI_FUNC(ivrs_alias));
296 static struct iommu_dev_data *find_dev_data(u16 devid)
298 struct iommu_dev_data *dev_data;
299 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
301 dev_data = search_dev_data(devid);
303 if (dev_data == NULL) {
304 dev_data = alloc_dev_data(devid);
308 if (translation_pre_enabled(iommu))
309 dev_data->defer_attach = true;
315 struct iommu_dev_data *get_dev_data(struct device *dev)
317 return dev->archdata.iommu;
319 EXPORT_SYMBOL(get_dev_data);
322 * Find or create an IOMMU group for a acpihid device.
324 static struct iommu_group *acpihid_device_group(struct device *dev)
326 struct acpihid_map_entry *p, *entry = NULL;
329 devid = get_acpihid_device_id(dev, &entry);
331 return ERR_PTR(devid);
333 list_for_each_entry(p, &acpihid_map, list) {
334 if ((devid == p->devid) && p->group)
335 entry->group = p->group;
339 entry->group = generic_device_group(dev);
341 iommu_group_ref_get(entry->group);
346 static bool pci_iommuv2_capable(struct pci_dev *pdev)
348 static const int caps[] = {
351 PCI_EXT_CAP_ID_PASID,
355 if (pci_ats_disabled())
358 for (i = 0; i < 3; ++i) {
359 pos = pci_find_ext_capability(pdev, caps[i]);
367 static bool pdev_pri_erratum(struct pci_dev *pdev, u32 erratum)
369 struct iommu_dev_data *dev_data;
371 dev_data = get_dev_data(&pdev->dev);
373 return dev_data->errata & (1 << erratum) ? true : false;
377 * This function checks if the driver got a valid device from the caller to
378 * avoid dereferencing invalid pointers.
380 static bool check_device(struct device *dev)
384 if (!dev || !dev->dma_mask)
387 devid = get_device_id(dev);
391 /* Out of our scope? */
392 if (devid > amd_iommu_last_bdf)
395 if (amd_iommu_rlookup_table[devid] == NULL)
401 static void init_iommu_group(struct device *dev)
403 struct iommu_group *group;
405 group = iommu_group_get_for_dev(dev);
409 iommu_group_put(group);
412 static int iommu_init_device(struct device *dev)
414 struct iommu_dev_data *dev_data;
415 struct amd_iommu *iommu;
418 if (dev->archdata.iommu)
421 devid = get_device_id(dev);
425 iommu = amd_iommu_rlookup_table[devid];
427 dev_data = find_dev_data(devid);
431 dev_data->alias = get_alias(dev);
434 * By default we use passthrough mode for IOMMUv2 capable device.
435 * But if amd_iommu=force_isolation is set (e.g. to debug DMA to
436 * invalid address), we ignore the capability for the device so
437 * it'll be forced to go into translation mode.
439 if ((iommu_pass_through || !amd_iommu_force_isolation) &&
440 dev_is_pci(dev) && pci_iommuv2_capable(to_pci_dev(dev))) {
441 struct amd_iommu *iommu;
443 iommu = amd_iommu_rlookup_table[dev_data->devid];
444 dev_data->iommu_v2 = iommu->is_iommu_v2;
447 dev->archdata.iommu = dev_data;
449 iommu_device_link(&iommu->iommu, dev);
454 static void iommu_ignore_device(struct device *dev)
459 devid = get_device_id(dev);
463 alias = get_alias(dev);
465 memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
466 memset(&amd_iommu_dev_table[alias], 0, sizeof(struct dev_table_entry));
468 amd_iommu_rlookup_table[devid] = NULL;
469 amd_iommu_rlookup_table[alias] = NULL;
472 static void iommu_uninit_device(struct device *dev)
474 struct iommu_dev_data *dev_data;
475 struct amd_iommu *iommu;
478 devid = get_device_id(dev);
482 iommu = amd_iommu_rlookup_table[devid];
484 dev_data = search_dev_data(devid);
488 if (dev_data->domain)
491 iommu_device_unlink(&iommu->iommu, dev);
493 iommu_group_remove_device(dev);
499 * We keep dev_data around for unplugged devices and reuse it when the
500 * device is re-plugged - not doing so would introduce a ton of races.
504 /****************************************************************************
506 * Interrupt handling functions
508 ****************************************************************************/
510 static void dump_dte_entry(u16 devid)
514 for (i = 0; i < 4; ++i)
515 pr_err("DTE[%d]: %016llx\n", i,
516 amd_iommu_dev_table[devid].data[i]);
519 static void dump_command(unsigned long phys_addr)
521 struct iommu_cmd *cmd = iommu_phys_to_virt(phys_addr);
524 for (i = 0; i < 4; ++i)
525 pr_err("CMD[%d]: %08x\n", i, cmd->data[i]);
528 static void amd_iommu_report_page_fault(u16 devid, u16 domain_id,
529 u64 address, int flags)
531 struct iommu_dev_data *dev_data = NULL;
532 struct pci_dev *pdev;
534 pdev = pci_get_domain_bus_and_slot(0, PCI_BUS_NUM(devid),
537 dev_data = get_dev_data(&pdev->dev);
539 if (dev_data && __ratelimit(&dev_data->rs)) {
540 pci_err(pdev, "Event logged [IO_PAGE_FAULT domain=0x%04x address=0x%llx flags=0x%04x]\n",
541 domain_id, address, flags);
542 } else if (printk_ratelimit()) {
543 pr_err("Event logged [IO_PAGE_FAULT device=%02x:%02x.%x domain=0x%04x address=0x%llx flags=0x%04x]\n",
544 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
545 domain_id, address, flags);
552 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
554 struct device *dev = iommu->iommu.dev;
555 int type, devid, pasid, flags, tag;
556 volatile u32 *event = __evt;
561 type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
562 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
563 pasid = PPR_PASID(*(u64 *)&event[0]);
564 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
565 address = (u64)(((u64)event[3]) << 32) | event[2];
568 /* Did we hit the erratum? */
569 if (++count == LOOP_TIMEOUT) {
570 pr_err("No event written to event log\n");
577 if (type == EVENT_TYPE_IO_FAULT) {
578 amd_iommu_report_page_fault(devid, pasid, address, flags);
583 case EVENT_TYPE_ILL_DEV:
584 dev_err(dev, "Event logged [ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x pasid=0x%05x address=0x%llx flags=0x%04x]\n",
585 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
586 pasid, address, flags);
587 dump_dte_entry(devid);
589 case EVENT_TYPE_DEV_TAB_ERR:
590 dev_err(dev, "Event logged [DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
591 "address=0x%llx flags=0x%04x]\n",
592 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
595 case EVENT_TYPE_PAGE_TAB_ERR:
596 dev_err(dev, "Event logged [PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x domain=0x%04x address=0x%llx flags=0x%04x]\n",
597 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
598 pasid, address, flags);
600 case EVENT_TYPE_ILL_CMD:
601 dev_err(dev, "Event logged [ILLEGAL_COMMAND_ERROR address=0x%llx]\n", address);
602 dump_command(address);
604 case EVENT_TYPE_CMD_HARD_ERR:
605 dev_err(dev, "Event logged [COMMAND_HARDWARE_ERROR address=0x%llx flags=0x%04x]\n",
608 case EVENT_TYPE_IOTLB_INV_TO:
609 dev_err(dev, "Event logged [IOTLB_INV_TIMEOUT device=%02x:%02x.%x address=0x%llx]\n",
610 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
613 case EVENT_TYPE_INV_DEV_REQ:
614 dev_err(dev, "Event logged [INVALID_DEVICE_REQUEST device=%02x:%02x.%x pasid=0x%05x address=0x%llx flags=0x%04x]\n",
615 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
616 pasid, address, flags);
618 case EVENT_TYPE_INV_PPR_REQ:
619 pasid = ((event[0] >> 16) & 0xFFFF)
620 | ((event[1] << 6) & 0xF0000);
621 tag = event[1] & 0x03FF;
622 dev_err(dev, "Event logged [INVALID_PPR_REQUEST device=%02x:%02x.%x pasid=0x%05x address=0x%llx flags=0x%04x tag=0x%03x]\n",
623 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
624 pasid, address, flags, tag);
627 dev_err(dev, "Event logged [UNKNOWN event[0]=0x%08x event[1]=0x%08x event[2]=0x%08x event[3]=0x%08x\n",
628 event[0], event[1], event[2], event[3]);
631 memset(__evt, 0, 4 * sizeof(u32));
634 static void iommu_poll_events(struct amd_iommu *iommu)
638 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
639 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
641 while (head != tail) {
642 iommu_print_event(iommu, iommu->evt_buf + head);
643 head = (head + EVENT_ENTRY_SIZE) % EVT_BUFFER_SIZE;
646 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
649 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
651 struct amd_iommu_fault fault;
653 if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
654 pr_err_ratelimited("Unknown PPR request received\n");
658 fault.address = raw[1];
659 fault.pasid = PPR_PASID(raw[0]);
660 fault.device_id = PPR_DEVID(raw[0]);
661 fault.tag = PPR_TAG(raw[0]);
662 fault.flags = PPR_FLAGS(raw[0]);
664 atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
667 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
671 if (iommu->ppr_log == NULL)
674 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
675 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
677 while (head != tail) {
682 raw = (u64 *)(iommu->ppr_log + head);
685 * Hardware bug: Interrupt may arrive before the entry is
686 * written to memory. If this happens we need to wait for the
689 for (i = 0; i < LOOP_TIMEOUT; ++i) {
690 if (PPR_REQ_TYPE(raw[0]) != 0)
695 /* Avoid memcpy function-call overhead */
700 * To detect the hardware bug we need to clear the entry
703 raw[0] = raw[1] = 0UL;
705 /* Update head pointer of hardware ring-buffer */
706 head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
707 writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
709 /* Handle PPR entry */
710 iommu_handle_ppr_entry(iommu, entry);
712 /* Refresh ring-buffer information */
713 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
714 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
718 #ifdef CONFIG_IRQ_REMAP
719 static int (*iommu_ga_log_notifier)(u32);
721 int amd_iommu_register_ga_log_notifier(int (*notifier)(u32))
723 iommu_ga_log_notifier = notifier;
727 EXPORT_SYMBOL(amd_iommu_register_ga_log_notifier);
729 static void iommu_poll_ga_log(struct amd_iommu *iommu)
731 u32 head, tail, cnt = 0;
733 if (iommu->ga_log == NULL)
736 head = readl(iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
737 tail = readl(iommu->mmio_base + MMIO_GA_TAIL_OFFSET);
739 while (head != tail) {
743 raw = (u64 *)(iommu->ga_log + head);
746 /* Avoid memcpy function-call overhead */
749 /* Update head pointer of hardware ring-buffer */
750 head = (head + GA_ENTRY_SIZE) % GA_LOG_SIZE;
751 writel(head, iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
753 /* Handle GA entry */
754 switch (GA_REQ_TYPE(log_entry)) {
756 if (!iommu_ga_log_notifier)
759 pr_debug("%s: devid=%#x, ga_tag=%#x\n",
760 __func__, GA_DEVID(log_entry),
763 if (iommu_ga_log_notifier(GA_TAG(log_entry)) != 0)
764 pr_err("GA log notifier failed.\n");
771 #endif /* CONFIG_IRQ_REMAP */
773 #define AMD_IOMMU_INT_MASK \
774 (MMIO_STATUS_EVT_INT_MASK | \
775 MMIO_STATUS_PPR_INT_MASK | \
776 MMIO_STATUS_GALOG_INT_MASK)
778 irqreturn_t amd_iommu_int_thread(int irq, void *data)
780 struct amd_iommu *iommu = (struct amd_iommu *) data;
781 u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
783 while (status & AMD_IOMMU_INT_MASK) {
784 /* Enable EVT and PPR and GA interrupts again */
785 writel(AMD_IOMMU_INT_MASK,
786 iommu->mmio_base + MMIO_STATUS_OFFSET);
788 if (status & MMIO_STATUS_EVT_INT_MASK) {
789 pr_devel("Processing IOMMU Event Log\n");
790 iommu_poll_events(iommu);
793 if (status & MMIO_STATUS_PPR_INT_MASK) {
794 pr_devel("Processing IOMMU PPR Log\n");
795 iommu_poll_ppr_log(iommu);
798 #ifdef CONFIG_IRQ_REMAP
799 if (status & MMIO_STATUS_GALOG_INT_MASK) {
800 pr_devel("Processing IOMMU GA Log\n");
801 iommu_poll_ga_log(iommu);
806 * Hardware bug: ERBT1312
807 * When re-enabling interrupt (by writing 1
808 * to clear the bit), the hardware might also try to set
809 * the interrupt bit in the event status register.
810 * In this scenario, the bit will be set, and disable
811 * subsequent interrupts.
813 * Workaround: The IOMMU driver should read back the
814 * status register and check if the interrupt bits are cleared.
815 * If not, driver will need to go through the interrupt handler
816 * again and re-clear the bits
818 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
823 irqreturn_t amd_iommu_int_handler(int irq, void *data)
825 return IRQ_WAKE_THREAD;
828 /****************************************************************************
830 * IOMMU command queuing functions
832 ****************************************************************************/
834 static int wait_on_sem(volatile u64 *sem)
838 while (*sem == 0 && i < LOOP_TIMEOUT) {
843 if (i == LOOP_TIMEOUT) {
844 pr_alert("Completion-Wait loop timed out\n");
851 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
852 struct iommu_cmd *cmd)
856 target = iommu->cmd_buf + iommu->cmd_buf_tail;
858 iommu->cmd_buf_tail += sizeof(*cmd);
859 iommu->cmd_buf_tail %= CMD_BUFFER_SIZE;
861 /* Copy command to buffer */
862 memcpy(target, cmd, sizeof(*cmd));
864 /* Tell the IOMMU about it */
865 writel(iommu->cmd_buf_tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
868 static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
870 u64 paddr = iommu_virt_to_phys((void *)address);
872 WARN_ON(address & 0x7ULL);
874 memset(cmd, 0, sizeof(*cmd));
875 cmd->data[0] = lower_32_bits(paddr) | CMD_COMPL_WAIT_STORE_MASK;
876 cmd->data[1] = upper_32_bits(paddr);
878 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
881 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
883 memset(cmd, 0, sizeof(*cmd));
884 cmd->data[0] = devid;
885 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
888 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
889 size_t size, u16 domid, int pde)
894 pages = iommu_num_pages(address, size, PAGE_SIZE);
899 * If we have to flush more than one page, flush all
900 * TLB entries for this domain
902 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
906 address &= PAGE_MASK;
908 memset(cmd, 0, sizeof(*cmd));
909 cmd->data[1] |= domid;
910 cmd->data[2] = lower_32_bits(address);
911 cmd->data[3] = upper_32_bits(address);
912 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
913 if (s) /* size bit - we flush more than one 4kb page */
914 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
915 if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
916 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
919 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
920 u64 address, size_t size)
925 pages = iommu_num_pages(address, size, PAGE_SIZE);
930 * If we have to flush more than one page, flush all
931 * TLB entries for this domain
933 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
937 address &= PAGE_MASK;
939 memset(cmd, 0, sizeof(*cmd));
940 cmd->data[0] = devid;
941 cmd->data[0] |= (qdep & 0xff) << 24;
942 cmd->data[1] = devid;
943 cmd->data[2] = lower_32_bits(address);
944 cmd->data[3] = upper_32_bits(address);
945 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
947 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
950 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, int pasid,
951 u64 address, bool size)
953 memset(cmd, 0, sizeof(*cmd));
955 address &= ~(0xfffULL);
957 cmd->data[0] = pasid;
958 cmd->data[1] = domid;
959 cmd->data[2] = lower_32_bits(address);
960 cmd->data[3] = upper_32_bits(address);
961 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
962 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
964 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
965 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
968 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, int pasid,
969 int qdep, u64 address, bool size)
971 memset(cmd, 0, sizeof(*cmd));
973 address &= ~(0xfffULL);
975 cmd->data[0] = devid;
976 cmd->data[0] |= ((pasid >> 8) & 0xff) << 16;
977 cmd->data[0] |= (qdep & 0xff) << 24;
978 cmd->data[1] = devid;
979 cmd->data[1] |= (pasid & 0xff) << 16;
980 cmd->data[2] = lower_32_bits(address);
981 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
982 cmd->data[3] = upper_32_bits(address);
984 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
985 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
988 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, int pasid,
989 int status, int tag, bool gn)
991 memset(cmd, 0, sizeof(*cmd));
993 cmd->data[0] = devid;
995 cmd->data[1] = pasid;
996 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
998 cmd->data[3] = tag & 0x1ff;
999 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
1001 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
1004 static void build_inv_all(struct iommu_cmd *cmd)
1006 memset(cmd, 0, sizeof(*cmd));
1007 CMD_SET_TYPE(cmd, CMD_INV_ALL);
1010 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
1012 memset(cmd, 0, sizeof(*cmd));
1013 cmd->data[0] = devid;
1014 CMD_SET_TYPE(cmd, CMD_INV_IRT);
1018 * Writes the command to the IOMMUs command buffer and informs the
1019 * hardware about the new command.
1021 static int __iommu_queue_command_sync(struct amd_iommu *iommu,
1022 struct iommu_cmd *cmd,
1025 unsigned int count = 0;
1026 u32 left, next_tail;
1028 next_tail = (iommu->cmd_buf_tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
1030 left = (iommu->cmd_buf_head - next_tail) % CMD_BUFFER_SIZE;
1033 /* Skip udelay() the first time around */
1035 if (count == LOOP_TIMEOUT) {
1036 pr_err("Command buffer timeout\n");
1043 /* Update head and recheck remaining space */
1044 iommu->cmd_buf_head = readl(iommu->mmio_base +
1045 MMIO_CMD_HEAD_OFFSET);
1050 copy_cmd_to_buffer(iommu, cmd);
1052 /* Do we need to make sure all commands are processed? */
1053 iommu->need_sync = sync;
1058 static int iommu_queue_command_sync(struct amd_iommu *iommu,
1059 struct iommu_cmd *cmd,
1062 unsigned long flags;
1065 raw_spin_lock_irqsave(&iommu->lock, flags);
1066 ret = __iommu_queue_command_sync(iommu, cmd, sync);
1067 raw_spin_unlock_irqrestore(&iommu->lock, flags);
1072 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
1074 return iommu_queue_command_sync(iommu, cmd, true);
1078 * This function queues a completion wait command into the command
1079 * buffer of an IOMMU
1081 static int iommu_completion_wait(struct amd_iommu *iommu)
1083 struct iommu_cmd cmd;
1084 unsigned long flags;
1087 if (!iommu->need_sync)
1091 build_completion_wait(&cmd, (u64)&iommu->cmd_sem);
1093 raw_spin_lock_irqsave(&iommu->lock, flags);
1097 ret = __iommu_queue_command_sync(iommu, &cmd, false);
1101 ret = wait_on_sem(&iommu->cmd_sem);
1104 raw_spin_unlock_irqrestore(&iommu->lock, flags);
1109 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
1111 struct iommu_cmd cmd;
1113 build_inv_dte(&cmd, devid);
1115 return iommu_queue_command(iommu, &cmd);
1118 static void amd_iommu_flush_dte_all(struct amd_iommu *iommu)
1122 for (devid = 0; devid <= 0xffff; ++devid)
1123 iommu_flush_dte(iommu, devid);
1125 iommu_completion_wait(iommu);
1129 * This function uses heavy locking and may disable irqs for some time. But
1130 * this is no issue because it is only called during resume.
1132 static void amd_iommu_flush_tlb_all(struct amd_iommu *iommu)
1136 for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
1137 struct iommu_cmd cmd;
1138 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1140 iommu_queue_command(iommu, &cmd);
1143 iommu_completion_wait(iommu);
1146 static void amd_iommu_flush_tlb_domid(struct amd_iommu *iommu, u32 dom_id)
1148 struct iommu_cmd cmd;
1150 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1152 iommu_queue_command(iommu, &cmd);
1154 iommu_completion_wait(iommu);
1157 static void amd_iommu_flush_all(struct amd_iommu *iommu)
1159 struct iommu_cmd cmd;
1161 build_inv_all(&cmd);
1163 iommu_queue_command(iommu, &cmd);
1164 iommu_completion_wait(iommu);
1167 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1169 struct iommu_cmd cmd;
1171 build_inv_irt(&cmd, devid);
1173 iommu_queue_command(iommu, &cmd);
1176 static void amd_iommu_flush_irt_all(struct amd_iommu *iommu)
1180 for (devid = 0; devid <= MAX_DEV_TABLE_ENTRIES; devid++)
1181 iommu_flush_irt(iommu, devid);
1183 iommu_completion_wait(iommu);
1186 void iommu_flush_all_caches(struct amd_iommu *iommu)
1188 if (iommu_feature(iommu, FEATURE_IA)) {
1189 amd_iommu_flush_all(iommu);
1191 amd_iommu_flush_dte_all(iommu);
1192 amd_iommu_flush_irt_all(iommu);
1193 amd_iommu_flush_tlb_all(iommu);
1198 * Command send function for flushing on-device TLB
1200 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1201 u64 address, size_t size)
1203 struct amd_iommu *iommu;
1204 struct iommu_cmd cmd;
1207 qdep = dev_data->ats.qdep;
1208 iommu = amd_iommu_rlookup_table[dev_data->devid];
1210 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1212 return iommu_queue_command(iommu, &cmd);
1216 * Command send function for invalidating a device table entry
1218 static int device_flush_dte(struct iommu_dev_data *dev_data)
1220 struct amd_iommu *iommu;
1224 iommu = amd_iommu_rlookup_table[dev_data->devid];
1225 alias = dev_data->alias;
1227 ret = iommu_flush_dte(iommu, dev_data->devid);
1228 if (!ret && alias != dev_data->devid)
1229 ret = iommu_flush_dte(iommu, alias);
1233 if (dev_data->ats.enabled)
1234 ret = device_flush_iotlb(dev_data, 0, ~0UL);
1240 * TLB invalidation function which is called from the mapping functions.
1241 * It invalidates a single PTE if the range to flush is within a single
1242 * page. Otherwise it flushes the whole TLB of the IOMMU.
1244 static void __domain_flush_pages(struct protection_domain *domain,
1245 u64 address, size_t size, int pde)
1247 struct iommu_dev_data *dev_data;
1248 struct iommu_cmd cmd;
1251 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1253 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
1254 if (!domain->dev_iommu[i])
1258 * Devices of this domain are behind this IOMMU
1259 * We need a TLB flush
1261 ret |= iommu_queue_command(amd_iommus[i], &cmd);
1264 list_for_each_entry(dev_data, &domain->dev_list, list) {
1266 if (!dev_data->ats.enabled)
1269 ret |= device_flush_iotlb(dev_data, address, size);
1275 static void domain_flush_pages(struct protection_domain *domain,
1276 u64 address, size_t size)
1278 __domain_flush_pages(domain, address, size, 0);
1281 /* Flush the whole IO/TLB for a given protection domain */
1282 static void domain_flush_tlb(struct protection_domain *domain)
1284 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
1287 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1288 static void domain_flush_tlb_pde(struct protection_domain *domain)
1290 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1293 static void domain_flush_complete(struct protection_domain *domain)
1297 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
1298 if (domain && !domain->dev_iommu[i])
1302 * Devices of this domain are behind this IOMMU
1303 * We need to wait for completion of all commands.
1305 iommu_completion_wait(amd_iommus[i]);
1309 /* Flush the not present cache if it exists */
1310 static void domain_flush_np_cache(struct protection_domain *domain,
1311 dma_addr_t iova, size_t size)
1313 if (unlikely(amd_iommu_np_cache)) {
1314 domain_flush_pages(domain, iova, size);
1315 domain_flush_complete(domain);
1321 * This function flushes the DTEs for all devices in domain
1323 static void domain_flush_devices(struct protection_domain *domain)
1325 struct iommu_dev_data *dev_data;
1327 list_for_each_entry(dev_data, &domain->dev_list, list)
1328 device_flush_dte(dev_data);
1331 /****************************************************************************
1333 * The functions below are used the create the page table mappings for
1334 * unity mapped regions.
1336 ****************************************************************************/
1338 static void free_page_list(struct page *freelist)
1340 while (freelist != NULL) {
1341 unsigned long p = (unsigned long)page_address(freelist);
1342 freelist = freelist->freelist;
1347 static struct page *free_pt_page(unsigned long pt, struct page *freelist)
1349 struct page *p = virt_to_page((void *)pt);
1351 p->freelist = freelist;
1356 #define DEFINE_FREE_PT_FN(LVL, FN) \
1357 static struct page *free_pt_##LVL (unsigned long __pt, struct page *freelist) \
1365 for (i = 0; i < 512; ++i) { \
1366 /* PTE present? */ \
1367 if (!IOMMU_PTE_PRESENT(pt[i])) \
1371 if (PM_PTE_LEVEL(pt[i]) == 0 || \
1372 PM_PTE_LEVEL(pt[i]) == 7) \
1375 p = (unsigned long)IOMMU_PTE_PAGE(pt[i]); \
1376 freelist = FN(p, freelist); \
1379 return free_pt_page((unsigned long)pt, freelist); \
1382 DEFINE_FREE_PT_FN(l2, free_pt_page)
1383 DEFINE_FREE_PT_FN(l3, free_pt_l2)
1384 DEFINE_FREE_PT_FN(l4, free_pt_l3)
1385 DEFINE_FREE_PT_FN(l5, free_pt_l4)
1386 DEFINE_FREE_PT_FN(l6, free_pt_l5)
1388 static struct page *free_sub_pt(unsigned long root, int mode,
1389 struct page *freelist)
1392 case PAGE_MODE_NONE:
1393 case PAGE_MODE_7_LEVEL:
1395 case PAGE_MODE_1_LEVEL:
1396 freelist = free_pt_page(root, freelist);
1398 case PAGE_MODE_2_LEVEL:
1399 freelist = free_pt_l2(root, freelist);
1401 case PAGE_MODE_3_LEVEL:
1402 freelist = free_pt_l3(root, freelist);
1404 case PAGE_MODE_4_LEVEL:
1405 freelist = free_pt_l4(root, freelist);
1407 case PAGE_MODE_5_LEVEL:
1408 freelist = free_pt_l5(root, freelist);
1410 case PAGE_MODE_6_LEVEL:
1411 freelist = free_pt_l6(root, freelist);
1420 static void free_pagetable(struct protection_domain *domain)
1422 unsigned long root = (unsigned long)domain->pt_root;
1423 struct page *freelist = NULL;
1425 BUG_ON(domain->mode < PAGE_MODE_NONE ||
1426 domain->mode > PAGE_MODE_6_LEVEL);
1428 free_sub_pt(root, domain->mode, freelist);
1430 free_page_list(freelist);
1434 * This function is used to add another level to an IO page table. Adding
1435 * another level increases the size of the address space by 9 bits to a size up
1438 static bool increase_address_space(struct protection_domain *domain,
1443 if (domain->mode == PAGE_MODE_6_LEVEL)
1444 /* address space already 64 bit large */
1447 pte = (void *)get_zeroed_page(gfp);
1451 *pte = PM_LEVEL_PDE(domain->mode,
1452 iommu_virt_to_phys(domain->pt_root));
1453 domain->pt_root = pte;
1455 domain->updated = true;
1460 static u64 *alloc_pte(struct protection_domain *domain,
1461 unsigned long address,
1462 unsigned long page_size,
1469 BUG_ON(!is_power_of_2(page_size));
1471 while (address > PM_LEVEL_SIZE(domain->mode))
1472 increase_address_space(domain, gfp);
1474 level = domain->mode - 1;
1475 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1476 address = PAGE_SIZE_ALIGN(address, page_size);
1477 end_lvl = PAGE_SIZE_LEVEL(page_size);
1479 while (level > end_lvl) {
1484 pte_level = PM_PTE_LEVEL(__pte);
1486 if (!IOMMU_PTE_PRESENT(__pte) ||
1487 pte_level == PAGE_MODE_7_LEVEL) {
1488 page = (u64 *)get_zeroed_page(gfp);
1492 __npte = PM_LEVEL_PDE(level, iommu_virt_to_phys(page));
1494 /* pte could have been changed somewhere. */
1495 if (cmpxchg64(pte, __pte, __npte) != __pte)
1496 free_page((unsigned long)page);
1497 else if (pte_level == PAGE_MODE_7_LEVEL)
1498 domain->updated = true;
1503 /* No level skipping support yet */
1504 if (pte_level != level)
1509 pte = IOMMU_PTE_PAGE(__pte);
1511 if (pte_page && level == end_lvl)
1514 pte = &pte[PM_LEVEL_INDEX(level, address)];
1521 * This function checks if there is a PTE for a given dma address. If
1522 * there is one, it returns the pointer to it.
1524 static u64 *fetch_pte(struct protection_domain *domain,
1525 unsigned long address,
1526 unsigned long *page_size)
1533 if (address > PM_LEVEL_SIZE(domain->mode))
1536 level = domain->mode - 1;
1537 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1538 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1543 if (!IOMMU_PTE_PRESENT(*pte))
1547 if (PM_PTE_LEVEL(*pte) == 7 ||
1548 PM_PTE_LEVEL(*pte) == 0)
1551 /* No level skipping support yet */
1552 if (PM_PTE_LEVEL(*pte) != level)
1557 /* Walk to the next level */
1558 pte = IOMMU_PTE_PAGE(*pte);
1559 pte = &pte[PM_LEVEL_INDEX(level, address)];
1560 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1563 if (PM_PTE_LEVEL(*pte) == 0x07) {
1564 unsigned long pte_mask;
1567 * If we have a series of large PTEs, make
1568 * sure to return a pointer to the first one.
1570 *page_size = pte_mask = PTE_PAGE_SIZE(*pte);
1571 pte_mask = ~((PAGE_SIZE_PTE_COUNT(pte_mask) << 3) - 1);
1572 pte = (u64 *)(((unsigned long)pte) & pte_mask);
1578 static struct page *free_clear_pte(u64 *pte, u64 pteval, struct page *freelist)
1583 while (cmpxchg64(pte, pteval, 0) != pteval) {
1584 pr_warn("AMD-Vi: IOMMU pte changed since we read it\n");
1588 if (!IOMMU_PTE_PRESENT(pteval))
1591 pt = (unsigned long)IOMMU_PTE_PAGE(pteval);
1592 mode = IOMMU_PTE_MODE(pteval);
1594 return free_sub_pt(pt, mode, freelist);
1598 * Generic mapping functions. It maps a physical address into a DMA
1599 * address space. It allocates the page table pages if necessary.
1600 * In the future it can be extended to a generic mapping function
1601 * supporting all features of AMD IOMMU page tables like level skipping
1602 * and full 64 bit address spaces.
1604 static int iommu_map_page(struct protection_domain *dom,
1605 unsigned long bus_addr,
1606 unsigned long phys_addr,
1607 unsigned long page_size,
1611 struct page *freelist = NULL;
1615 BUG_ON(!IS_ALIGNED(bus_addr, page_size));
1616 BUG_ON(!IS_ALIGNED(phys_addr, page_size));
1618 if (!(prot & IOMMU_PROT_MASK))
1621 count = PAGE_SIZE_PTE_COUNT(page_size);
1622 pte = alloc_pte(dom, bus_addr, page_size, NULL, gfp);
1627 for (i = 0; i < count; ++i)
1628 freelist = free_clear_pte(&pte[i], pte[i], freelist);
1630 if (freelist != NULL)
1631 dom->updated = true;
1634 __pte = PAGE_SIZE_PTE(__sme_set(phys_addr), page_size);
1635 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_PR | IOMMU_PTE_FC;
1637 __pte = __sme_set(phys_addr) | IOMMU_PTE_PR | IOMMU_PTE_FC;
1639 if (prot & IOMMU_PROT_IR)
1640 __pte |= IOMMU_PTE_IR;
1641 if (prot & IOMMU_PROT_IW)
1642 __pte |= IOMMU_PTE_IW;
1644 for (i = 0; i < count; ++i)
1649 /* Everything flushed out, free pages now */
1650 free_page_list(freelist);
1655 static unsigned long iommu_unmap_page(struct protection_domain *dom,
1656 unsigned long bus_addr,
1657 unsigned long page_size)
1659 unsigned long long unmapped;
1660 unsigned long unmap_size;
1663 BUG_ON(!is_power_of_2(page_size));
1667 while (unmapped < page_size) {
1669 pte = fetch_pte(dom, bus_addr, &unmap_size);
1674 count = PAGE_SIZE_PTE_COUNT(unmap_size);
1675 for (i = 0; i < count; i++)
1679 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1680 unmapped += unmap_size;
1683 BUG_ON(unmapped && !is_power_of_2(unmapped));
1688 /****************************************************************************
1690 * The next functions belong to the address allocator for the dma_ops
1691 * interface functions.
1693 ****************************************************************************/
1696 static unsigned long dma_ops_alloc_iova(struct device *dev,
1697 struct dma_ops_domain *dma_dom,
1698 unsigned int pages, u64 dma_mask)
1700 unsigned long pfn = 0;
1702 pages = __roundup_pow_of_two(pages);
1704 if (dma_mask > DMA_BIT_MASK(32))
1705 pfn = alloc_iova_fast(&dma_dom->iovad, pages,
1706 IOVA_PFN(DMA_BIT_MASK(32)), false);
1709 pfn = alloc_iova_fast(&dma_dom->iovad, pages,
1710 IOVA_PFN(dma_mask), true);
1712 return (pfn << PAGE_SHIFT);
1715 static void dma_ops_free_iova(struct dma_ops_domain *dma_dom,
1716 unsigned long address,
1719 pages = __roundup_pow_of_two(pages);
1720 address >>= PAGE_SHIFT;
1722 free_iova_fast(&dma_dom->iovad, address, pages);
1725 /****************************************************************************
1727 * The next functions belong to the domain allocation. A domain is
1728 * allocated for every IOMMU as the default domain. If device isolation
1729 * is enabled, every device get its own domain. The most important thing
1730 * about domains is the page table mapping the DMA address space they
1733 ****************************************************************************/
1735 static u16 domain_id_alloc(void)
1739 spin_lock(&pd_bitmap_lock);
1740 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1742 if (id > 0 && id < MAX_DOMAIN_ID)
1743 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1746 spin_unlock(&pd_bitmap_lock);
1751 static void domain_id_free(int id)
1753 spin_lock(&pd_bitmap_lock);
1754 if (id > 0 && id < MAX_DOMAIN_ID)
1755 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1756 spin_unlock(&pd_bitmap_lock);
1759 static void free_gcr3_tbl_level1(u64 *tbl)
1764 for (i = 0; i < 512; ++i) {
1765 if (!(tbl[i] & GCR3_VALID))
1768 ptr = iommu_phys_to_virt(tbl[i] & PAGE_MASK);
1770 free_page((unsigned long)ptr);
1774 static void free_gcr3_tbl_level2(u64 *tbl)
1779 for (i = 0; i < 512; ++i) {
1780 if (!(tbl[i] & GCR3_VALID))
1783 ptr = iommu_phys_to_virt(tbl[i] & PAGE_MASK);
1785 free_gcr3_tbl_level1(ptr);
1789 static void free_gcr3_table(struct protection_domain *domain)
1791 if (domain->glx == 2)
1792 free_gcr3_tbl_level2(domain->gcr3_tbl);
1793 else if (domain->glx == 1)
1794 free_gcr3_tbl_level1(domain->gcr3_tbl);
1796 BUG_ON(domain->glx != 0);
1798 free_page((unsigned long)domain->gcr3_tbl);
1801 static void dma_ops_domain_flush_tlb(struct dma_ops_domain *dom)
1803 domain_flush_tlb(&dom->domain);
1804 domain_flush_complete(&dom->domain);
1807 static void iova_domain_flush_tlb(struct iova_domain *iovad)
1809 struct dma_ops_domain *dom;
1811 dom = container_of(iovad, struct dma_ops_domain, iovad);
1813 dma_ops_domain_flush_tlb(dom);
1817 * Free a domain, only used if something went wrong in the
1818 * allocation path and we need to free an already allocated page table
1820 static void dma_ops_domain_free(struct dma_ops_domain *dom)
1825 put_iova_domain(&dom->iovad);
1827 free_pagetable(&dom->domain);
1830 domain_id_free(dom->domain.id);
1836 * Allocates a new protection domain usable for the dma_ops functions.
1837 * It also initializes the page table and the address allocator data
1838 * structures required for the dma_ops interface
1840 static struct dma_ops_domain *dma_ops_domain_alloc(void)
1842 struct dma_ops_domain *dma_dom;
1844 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
1848 if (protection_domain_init(&dma_dom->domain))
1851 dma_dom->domain.mode = PAGE_MODE_3_LEVEL;
1852 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
1853 dma_dom->domain.flags = PD_DMA_OPS_MASK;
1854 if (!dma_dom->domain.pt_root)
1857 init_iova_domain(&dma_dom->iovad, PAGE_SIZE, IOVA_START_PFN);
1859 if (init_iova_flush_queue(&dma_dom->iovad, iova_domain_flush_tlb, NULL))
1862 /* Initialize reserved ranges */
1863 copy_reserved_iova(&reserved_iova_ranges, &dma_dom->iovad);
1868 dma_ops_domain_free(dma_dom);
1874 * little helper function to check whether a given protection domain is a
1877 static bool dma_ops_domain(struct protection_domain *domain)
1879 return domain->flags & PD_DMA_OPS_MASK;
1882 static void set_dte_entry(u16 devid, struct protection_domain *domain,
1889 if (domain->mode != PAGE_MODE_NONE)
1890 pte_root = iommu_virt_to_phys(domain->pt_root);
1892 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
1893 << DEV_ENTRY_MODE_SHIFT;
1894 pte_root |= DTE_FLAG_IR | DTE_FLAG_IW | DTE_FLAG_V | DTE_FLAG_TV;
1896 flags = amd_iommu_dev_table[devid].data[1];
1899 flags |= DTE_FLAG_IOTLB;
1902 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
1904 if (iommu_feature(iommu, FEATURE_EPHSUP))
1905 pte_root |= 1ULL << DEV_ENTRY_PPR;
1908 if (domain->flags & PD_IOMMUV2_MASK) {
1909 u64 gcr3 = iommu_virt_to_phys(domain->gcr3_tbl);
1910 u64 glx = domain->glx;
1913 pte_root |= DTE_FLAG_GV;
1914 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
1916 /* First mask out possible old values for GCR3 table */
1917 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
1920 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
1923 /* Encode GCR3 table into DTE */
1924 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
1927 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
1930 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
1934 flags &= ~DEV_DOMID_MASK;
1935 flags |= domain->id;
1937 old_domid = amd_iommu_dev_table[devid].data[1] & DEV_DOMID_MASK;
1938 amd_iommu_dev_table[devid].data[1] = flags;
1939 amd_iommu_dev_table[devid].data[0] = pte_root;
1942 * A kdump kernel might be replacing a domain ID that was copied from
1943 * the previous kernel--if so, it needs to flush the translation cache
1944 * entries for the old domain ID that is being overwritten
1947 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
1949 amd_iommu_flush_tlb_domid(iommu, old_domid);
1953 static void clear_dte_entry(u16 devid)
1955 /* remove entry from the device table seen by the hardware */
1956 amd_iommu_dev_table[devid].data[0] = DTE_FLAG_V | DTE_FLAG_TV;
1957 amd_iommu_dev_table[devid].data[1] &= DTE_FLAG_MASK;
1959 amd_iommu_apply_erratum_63(devid);
1962 static void do_attach(struct iommu_dev_data *dev_data,
1963 struct protection_domain *domain)
1965 struct amd_iommu *iommu;
1969 iommu = amd_iommu_rlookup_table[dev_data->devid];
1970 alias = dev_data->alias;
1971 ats = dev_data->ats.enabled;
1973 /* Update data structures */
1974 dev_data->domain = domain;
1975 list_add(&dev_data->list, &domain->dev_list);
1977 /* Do reference counting */
1978 domain->dev_iommu[iommu->index] += 1;
1979 domain->dev_cnt += 1;
1981 /* Update device table */
1982 set_dte_entry(dev_data->devid, domain, ats, dev_data->iommu_v2);
1983 if (alias != dev_data->devid)
1984 set_dte_entry(alias, domain, ats, dev_data->iommu_v2);
1986 device_flush_dte(dev_data);
1989 static void do_detach(struct iommu_dev_data *dev_data)
1991 struct protection_domain *domain = dev_data->domain;
1992 struct amd_iommu *iommu;
1995 iommu = amd_iommu_rlookup_table[dev_data->devid];
1996 alias = dev_data->alias;
1998 /* Update data structures */
1999 dev_data->domain = NULL;
2000 list_del(&dev_data->list);
2001 clear_dte_entry(dev_data->devid);
2002 if (alias != dev_data->devid)
2003 clear_dte_entry(alias);
2005 /* Flush the DTE entry */
2006 device_flush_dte(dev_data);
2009 domain_flush_tlb_pde(domain);
2011 /* Wait for the flushes to finish */
2012 domain_flush_complete(domain);
2014 /* decrease reference counters - needs to happen after the flushes */
2015 domain->dev_iommu[iommu->index] -= 1;
2016 domain->dev_cnt -= 1;
2020 * If a device is not yet associated with a domain, this function makes the
2021 * device visible in the domain
2023 static int __attach_device(struct iommu_dev_data *dev_data,
2024 struct protection_domain *domain)
2029 spin_lock(&domain->lock);
2032 if (dev_data->domain != NULL)
2035 /* Attach alias group root */
2036 do_attach(dev_data, domain);
2043 spin_unlock(&domain->lock);
2049 static void pdev_iommuv2_disable(struct pci_dev *pdev)
2051 pci_disable_ats(pdev);
2052 pci_disable_pri(pdev);
2053 pci_disable_pasid(pdev);
2056 /* FIXME: Change generic reset-function to do the same */
2057 static int pri_reset_while_enabled(struct pci_dev *pdev)
2062 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2066 pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
2067 control |= PCI_PRI_CTRL_RESET;
2068 pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
2073 static int pdev_iommuv2_enable(struct pci_dev *pdev)
2078 /* FIXME: Hardcode number of outstanding requests for now */
2080 if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
2082 reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
2084 /* Only allow access to user-accessible pages */
2085 ret = pci_enable_pasid(pdev, 0);
2089 /* First reset the PRI state of the device */
2090 ret = pci_reset_pri(pdev);
2095 ret = pci_enable_pri(pdev, reqs);
2100 ret = pri_reset_while_enabled(pdev);
2105 ret = pci_enable_ats(pdev, PAGE_SHIFT);
2112 pci_disable_pri(pdev);
2113 pci_disable_pasid(pdev);
2119 * If a device is not yet associated with a domain, this function makes the
2120 * device visible in the domain
2122 static int attach_device(struct device *dev,
2123 struct protection_domain *domain)
2125 struct pci_dev *pdev;
2126 struct iommu_dev_data *dev_data;
2127 unsigned long flags;
2130 dev_data = get_dev_data(dev);
2132 if (!dev_is_pci(dev))
2133 goto skip_ats_check;
2135 pdev = to_pci_dev(dev);
2136 if (domain->flags & PD_IOMMUV2_MASK) {
2137 if (!dev_data->passthrough)
2140 if (dev_data->iommu_v2) {
2141 if (pdev_iommuv2_enable(pdev) != 0)
2144 dev_data->ats.enabled = true;
2145 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2146 dev_data->pri_tlp = pci_prg_resp_pasid_required(pdev);
2148 } else if (amd_iommu_iotlb_sup &&
2149 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
2150 dev_data->ats.enabled = true;
2151 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2155 spin_lock_irqsave(&amd_iommu_devtable_lock, flags);
2156 ret = __attach_device(dev_data, domain);
2157 spin_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2160 * We might boot into a crash-kernel here. The crashed kernel
2161 * left the caches in the IOMMU dirty. So we have to flush
2162 * here to evict all dirty stuff.
2164 domain_flush_tlb_pde(domain);
2170 * Removes a device from a protection domain (unlocked)
2172 static void __detach_device(struct iommu_dev_data *dev_data)
2174 struct protection_domain *domain;
2176 domain = dev_data->domain;
2178 spin_lock(&domain->lock);
2180 do_detach(dev_data);
2182 spin_unlock(&domain->lock);
2186 * Removes a device from a protection domain (with devtable_lock held)
2188 static void detach_device(struct device *dev)
2190 struct protection_domain *domain;
2191 struct iommu_dev_data *dev_data;
2192 unsigned long flags;
2194 dev_data = get_dev_data(dev);
2195 domain = dev_data->domain;
2198 * First check if the device is still attached. It might already
2199 * be detached from its domain because the generic
2200 * iommu_detach_group code detached it and we try again here in
2201 * our alias handling.
2203 if (WARN_ON(!dev_data->domain))
2206 /* lock device table */
2207 spin_lock_irqsave(&amd_iommu_devtable_lock, flags);
2208 __detach_device(dev_data);
2209 spin_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2211 if (!dev_is_pci(dev))
2214 if (domain->flags & PD_IOMMUV2_MASK && dev_data->iommu_v2)
2215 pdev_iommuv2_disable(to_pci_dev(dev));
2216 else if (dev_data->ats.enabled)
2217 pci_disable_ats(to_pci_dev(dev));
2219 dev_data->ats.enabled = false;
2222 static int amd_iommu_add_device(struct device *dev)
2224 struct iommu_dev_data *dev_data;
2225 struct iommu_domain *domain;
2226 struct amd_iommu *iommu;
2229 if (!check_device(dev) || get_dev_data(dev))
2232 devid = get_device_id(dev);
2236 iommu = amd_iommu_rlookup_table[devid];
2238 ret = iommu_init_device(dev);
2240 if (ret != -ENOTSUPP)
2241 dev_err(dev, "Failed to initialize - trying to proceed anyway\n");
2243 iommu_ignore_device(dev);
2244 dev->dma_ops = NULL;
2247 init_iommu_group(dev);
2249 dev_data = get_dev_data(dev);
2253 if (iommu_pass_through || dev_data->iommu_v2)
2254 iommu_request_dm_for_dev(dev);
2256 /* Domains are initialized for this device - have a look what we ended up with */
2257 domain = iommu_get_domain_for_dev(dev);
2258 if (domain->type == IOMMU_DOMAIN_IDENTITY)
2259 dev_data->passthrough = true;
2261 dev->dma_ops = &amd_iommu_dma_ops;
2264 iommu_completion_wait(iommu);
2269 static void amd_iommu_remove_device(struct device *dev)
2271 struct amd_iommu *iommu;
2274 if (!check_device(dev))
2277 devid = get_device_id(dev);
2281 iommu = amd_iommu_rlookup_table[devid];
2283 iommu_uninit_device(dev);
2284 iommu_completion_wait(iommu);
2287 static struct iommu_group *amd_iommu_device_group(struct device *dev)
2289 if (dev_is_pci(dev))
2290 return pci_device_group(dev);
2292 return acpihid_device_group(dev);
2295 /*****************************************************************************
2297 * The next functions belong to the dma_ops mapping/unmapping code.
2299 *****************************************************************************/
2302 * In the dma_ops path we only have the struct device. This function
2303 * finds the corresponding IOMMU, the protection domain and the
2304 * requestor id for a given device.
2305 * If the device is not yet associated with a domain this is also done
2308 static struct protection_domain *get_domain(struct device *dev)
2310 struct protection_domain *domain;
2311 struct iommu_domain *io_domain;
2313 if (!check_device(dev))
2314 return ERR_PTR(-EINVAL);
2316 domain = get_dev_data(dev)->domain;
2317 if (domain == NULL && get_dev_data(dev)->defer_attach) {
2318 get_dev_data(dev)->defer_attach = false;
2319 io_domain = iommu_get_domain_for_dev(dev);
2320 domain = to_pdomain(io_domain);
2321 attach_device(dev, domain);
2324 return ERR_PTR(-EBUSY);
2326 if (!dma_ops_domain(domain))
2327 return ERR_PTR(-EBUSY);
2332 static void update_device_table(struct protection_domain *domain)
2334 struct iommu_dev_data *dev_data;
2336 list_for_each_entry(dev_data, &domain->dev_list, list) {
2337 set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled,
2338 dev_data->iommu_v2);
2340 if (dev_data->devid == dev_data->alias)
2343 /* There is an alias, update device table entry for it */
2344 set_dte_entry(dev_data->alias, domain, dev_data->ats.enabled,
2345 dev_data->iommu_v2);
2349 static void update_domain(struct protection_domain *domain)
2351 if (!domain->updated)
2354 update_device_table(domain);
2356 domain_flush_devices(domain);
2357 domain_flush_tlb_pde(domain);
2359 domain->updated = false;
2362 static int dir2prot(enum dma_data_direction direction)
2364 if (direction == DMA_TO_DEVICE)
2365 return IOMMU_PROT_IR;
2366 else if (direction == DMA_FROM_DEVICE)
2367 return IOMMU_PROT_IW;
2368 else if (direction == DMA_BIDIRECTIONAL)
2369 return IOMMU_PROT_IW | IOMMU_PROT_IR;
2375 * This function contains common code for mapping of a physically
2376 * contiguous memory region into DMA address space. It is used by all
2377 * mapping functions provided with this IOMMU driver.
2378 * Must be called with the domain lock held.
2380 static dma_addr_t __map_single(struct device *dev,
2381 struct dma_ops_domain *dma_dom,
2384 enum dma_data_direction direction,
2387 dma_addr_t offset = paddr & ~PAGE_MASK;
2388 dma_addr_t address, start, ret;
2393 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
2396 address = dma_ops_alloc_iova(dev, dma_dom, pages, dma_mask);
2400 prot = dir2prot(direction);
2403 for (i = 0; i < pages; ++i) {
2404 ret = iommu_map_page(&dma_dom->domain, start, paddr,
2405 PAGE_SIZE, prot, GFP_ATOMIC);
2414 domain_flush_np_cache(&dma_dom->domain, address, size);
2421 for (--i; i >= 0; --i) {
2423 iommu_unmap_page(&dma_dom->domain, start, PAGE_SIZE);
2426 domain_flush_tlb(&dma_dom->domain);
2427 domain_flush_complete(&dma_dom->domain);
2429 dma_ops_free_iova(dma_dom, address, pages);
2431 return DMA_MAPPING_ERROR;
2435 * Does the reverse of the __map_single function. Must be called with
2436 * the domain lock held too
2438 static void __unmap_single(struct dma_ops_domain *dma_dom,
2439 dma_addr_t dma_addr,
2443 dma_addr_t i, start;
2446 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
2447 dma_addr &= PAGE_MASK;
2450 for (i = 0; i < pages; ++i) {
2451 iommu_unmap_page(&dma_dom->domain, start, PAGE_SIZE);
2455 if (amd_iommu_unmap_flush) {
2456 domain_flush_tlb(&dma_dom->domain);
2457 domain_flush_complete(&dma_dom->domain);
2458 dma_ops_free_iova(dma_dom, dma_addr, pages);
2460 pages = __roundup_pow_of_two(pages);
2461 queue_iova(&dma_dom->iovad, dma_addr >> PAGE_SHIFT, pages, 0);
2466 * The exported map_single function for dma_ops.
2468 static dma_addr_t map_page(struct device *dev, struct page *page,
2469 unsigned long offset, size_t size,
2470 enum dma_data_direction dir,
2471 unsigned long attrs)
2473 phys_addr_t paddr = page_to_phys(page) + offset;
2474 struct protection_domain *domain;
2475 struct dma_ops_domain *dma_dom;
2478 domain = get_domain(dev);
2479 if (PTR_ERR(domain) == -EINVAL)
2480 return (dma_addr_t)paddr;
2481 else if (IS_ERR(domain))
2482 return DMA_MAPPING_ERROR;
2484 dma_mask = *dev->dma_mask;
2485 dma_dom = to_dma_ops_domain(domain);
2487 return __map_single(dev, dma_dom, paddr, size, dir, dma_mask);
2491 * The exported unmap_single function for dma_ops.
2493 static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2494 enum dma_data_direction dir, unsigned long attrs)
2496 struct protection_domain *domain;
2497 struct dma_ops_domain *dma_dom;
2499 domain = get_domain(dev);
2503 dma_dom = to_dma_ops_domain(domain);
2505 __unmap_single(dma_dom, dma_addr, size, dir);
2508 static int sg_num_pages(struct device *dev,
2509 struct scatterlist *sglist,
2512 unsigned long mask, boundary_size;
2513 struct scatterlist *s;
2516 mask = dma_get_seg_boundary(dev);
2517 boundary_size = mask + 1 ? ALIGN(mask + 1, PAGE_SIZE) >> PAGE_SHIFT :
2518 1UL << (BITS_PER_LONG - PAGE_SHIFT);
2520 for_each_sg(sglist, s, nelems, i) {
2523 s->dma_address = npages << PAGE_SHIFT;
2524 p = npages % boundary_size;
2525 n = iommu_num_pages(sg_phys(s), s->length, PAGE_SIZE);
2526 if (p + n > boundary_size)
2527 npages += boundary_size - p;
2535 * The exported map_sg function for dma_ops (handles scatter-gather
2538 static int map_sg(struct device *dev, struct scatterlist *sglist,
2539 int nelems, enum dma_data_direction direction,
2540 unsigned long attrs)
2542 int mapped_pages = 0, npages = 0, prot = 0, i;
2543 struct protection_domain *domain;
2544 struct dma_ops_domain *dma_dom;
2545 struct scatterlist *s;
2546 unsigned long address;
2550 domain = get_domain(dev);
2554 dma_dom = to_dma_ops_domain(domain);
2555 dma_mask = *dev->dma_mask;
2557 npages = sg_num_pages(dev, sglist, nelems);
2559 address = dma_ops_alloc_iova(dev, dma_dom, npages, dma_mask);
2563 prot = dir2prot(direction);
2565 /* Map all sg entries */
2566 for_each_sg(sglist, s, nelems, i) {
2567 int j, pages = iommu_num_pages(sg_phys(s), s->length, PAGE_SIZE);
2569 for (j = 0; j < pages; ++j) {
2570 unsigned long bus_addr, phys_addr;
2572 bus_addr = address + s->dma_address + (j << PAGE_SHIFT);
2573 phys_addr = (sg_phys(s) & PAGE_MASK) + (j << PAGE_SHIFT);
2574 ret = iommu_map_page(domain, bus_addr, phys_addr, PAGE_SIZE, prot, GFP_ATOMIC);
2582 /* Everything is mapped - write the right values into s->dma_address */
2583 for_each_sg(sglist, s, nelems, i) {
2585 * Add in the remaining piece of the scatter-gather offset that
2586 * was masked out when we were determining the physical address
2587 * via (sg_phys(s) & PAGE_MASK) earlier.
2589 s->dma_address += address + (s->offset & ~PAGE_MASK);
2590 s->dma_length = s->length;
2594 domain_flush_np_cache(domain, s->dma_address, s->dma_length);
2599 dev_err(dev, "IOMMU mapping error in map_sg (io-pages: %d reason: %d)\n",
2602 for_each_sg(sglist, s, nelems, i) {
2603 int j, pages = iommu_num_pages(sg_phys(s), s->length, PAGE_SIZE);
2605 for (j = 0; j < pages; ++j) {
2606 unsigned long bus_addr;
2608 bus_addr = address + s->dma_address + (j << PAGE_SHIFT);
2609 iommu_unmap_page(domain, bus_addr, PAGE_SIZE);
2611 if (--mapped_pages == 0)
2617 free_iova_fast(&dma_dom->iovad, address >> PAGE_SHIFT, npages);
2624 * The exported map_sg function for dma_ops (handles scatter-gather
2627 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2628 int nelems, enum dma_data_direction dir,
2629 unsigned long attrs)
2631 struct protection_domain *domain;
2632 struct dma_ops_domain *dma_dom;
2633 unsigned long startaddr;
2636 domain = get_domain(dev);
2640 startaddr = sg_dma_address(sglist) & PAGE_MASK;
2641 dma_dom = to_dma_ops_domain(domain);
2642 npages = sg_num_pages(dev, sglist, nelems);
2644 __unmap_single(dma_dom, startaddr, npages << PAGE_SHIFT, dir);
2648 * The exported alloc_coherent function for dma_ops.
2650 static void *alloc_coherent(struct device *dev, size_t size,
2651 dma_addr_t *dma_addr, gfp_t flag,
2652 unsigned long attrs)
2654 u64 dma_mask = dev->coherent_dma_mask;
2655 struct protection_domain *domain;
2656 struct dma_ops_domain *dma_dom;
2659 domain = get_domain(dev);
2660 if (PTR_ERR(domain) == -EINVAL) {
2661 page = alloc_pages(flag, get_order(size));
2662 *dma_addr = page_to_phys(page);
2663 return page_address(page);
2664 } else if (IS_ERR(domain))
2667 dma_dom = to_dma_ops_domain(domain);
2668 size = PAGE_ALIGN(size);
2669 dma_mask = dev->coherent_dma_mask;
2670 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
2673 page = alloc_pages(flag | __GFP_NOWARN, get_order(size));
2675 if (!gfpflags_allow_blocking(flag))
2678 page = dma_alloc_from_contiguous(dev, size >> PAGE_SHIFT,
2679 get_order(size), flag & __GFP_NOWARN);
2685 dma_mask = *dev->dma_mask;
2687 *dma_addr = __map_single(dev, dma_dom, page_to_phys(page),
2688 size, DMA_BIDIRECTIONAL, dma_mask);
2690 if (*dma_addr == DMA_MAPPING_ERROR)
2693 return page_address(page);
2697 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2698 __free_pages(page, get_order(size));
2704 * The exported free_coherent function for dma_ops.
2706 static void free_coherent(struct device *dev, size_t size,
2707 void *virt_addr, dma_addr_t dma_addr,
2708 unsigned long attrs)
2710 struct protection_domain *domain;
2711 struct dma_ops_domain *dma_dom;
2714 page = virt_to_page(virt_addr);
2715 size = PAGE_ALIGN(size);
2717 domain = get_domain(dev);
2721 dma_dom = to_dma_ops_domain(domain);
2723 __unmap_single(dma_dom, dma_addr, size, DMA_BIDIRECTIONAL);
2726 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2727 __free_pages(page, get_order(size));
2731 * This function is called by the DMA layer to find out if we can handle a
2732 * particular device. It is part of the dma_ops.
2734 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
2736 if (!dma_direct_supported(dev, mask))
2738 return check_device(dev);
2741 static const struct dma_map_ops amd_iommu_dma_ops = {
2742 .alloc = alloc_coherent,
2743 .free = free_coherent,
2744 .map_page = map_page,
2745 .unmap_page = unmap_page,
2747 .unmap_sg = unmap_sg,
2748 .dma_supported = amd_iommu_dma_supported,
2751 static int init_reserved_iova_ranges(void)
2753 struct pci_dev *pdev = NULL;
2756 init_iova_domain(&reserved_iova_ranges, PAGE_SIZE, IOVA_START_PFN);
2758 lockdep_set_class(&reserved_iova_ranges.iova_rbtree_lock,
2759 &reserved_rbtree_key);
2761 /* MSI memory range */
2762 val = reserve_iova(&reserved_iova_ranges,
2763 IOVA_PFN(MSI_RANGE_START), IOVA_PFN(MSI_RANGE_END));
2765 pr_err("Reserving MSI range failed\n");
2769 /* HT memory range */
2770 val = reserve_iova(&reserved_iova_ranges,
2771 IOVA_PFN(HT_RANGE_START), IOVA_PFN(HT_RANGE_END));
2773 pr_err("Reserving HT range failed\n");
2778 * Memory used for PCI resources
2779 * FIXME: Check whether we can reserve the PCI-hole completly
2781 for_each_pci_dev(pdev) {
2784 for (i = 0; i < PCI_NUM_RESOURCES; ++i) {
2785 struct resource *r = &pdev->resource[i];
2787 if (!(r->flags & IORESOURCE_MEM))
2790 val = reserve_iova(&reserved_iova_ranges,
2794 pci_err(pdev, "Reserve pci-resource range %pR failed\n", r);
2803 int __init amd_iommu_init_api(void)
2807 ret = iova_cache_get();
2811 ret = init_reserved_iova_ranges();
2815 err = bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
2818 #ifdef CONFIG_ARM_AMBA
2819 err = bus_set_iommu(&amba_bustype, &amd_iommu_ops);
2823 err = bus_set_iommu(&platform_bus_type, &amd_iommu_ops);
2830 int __init amd_iommu_init_dma_ops(void)
2832 swiotlb = (iommu_pass_through || sme_me_mask) ? 1 : 0;
2835 if (amd_iommu_unmap_flush)
2836 pr_info("IO/TLB flush on unmap enabled\n");
2838 pr_info("Lazy IO/TLB flushing enabled\n");
2844 /*****************************************************************************
2846 * The following functions belong to the exported interface of AMD IOMMU
2848 * This interface allows access to lower level functions of the IOMMU
2849 * like protection domain handling and assignement of devices to domains
2850 * which is not possible with the dma_ops interface.
2852 *****************************************************************************/
2854 static void cleanup_domain(struct protection_domain *domain)
2856 struct iommu_dev_data *entry;
2857 unsigned long flags;
2859 spin_lock_irqsave(&amd_iommu_devtable_lock, flags);
2861 while (!list_empty(&domain->dev_list)) {
2862 entry = list_first_entry(&domain->dev_list,
2863 struct iommu_dev_data, list);
2864 BUG_ON(!entry->domain);
2865 __detach_device(entry);
2868 spin_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2871 static void protection_domain_free(struct protection_domain *domain)
2877 domain_id_free(domain->id);
2882 static int protection_domain_init(struct protection_domain *domain)
2884 spin_lock_init(&domain->lock);
2885 mutex_init(&domain->api_lock);
2886 domain->id = domain_id_alloc();
2889 INIT_LIST_HEAD(&domain->dev_list);
2894 static struct protection_domain *protection_domain_alloc(void)
2896 struct protection_domain *domain;
2898 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2902 if (protection_domain_init(domain))
2913 static struct iommu_domain *amd_iommu_domain_alloc(unsigned type)
2915 struct protection_domain *pdomain;
2916 struct dma_ops_domain *dma_domain;
2919 case IOMMU_DOMAIN_UNMANAGED:
2920 pdomain = protection_domain_alloc();
2924 pdomain->mode = PAGE_MODE_3_LEVEL;
2925 pdomain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2926 if (!pdomain->pt_root) {
2927 protection_domain_free(pdomain);
2931 pdomain->domain.geometry.aperture_start = 0;
2932 pdomain->domain.geometry.aperture_end = ~0ULL;
2933 pdomain->domain.geometry.force_aperture = true;
2936 case IOMMU_DOMAIN_DMA:
2937 dma_domain = dma_ops_domain_alloc();
2939 pr_err("Failed to allocate\n");
2942 pdomain = &dma_domain->domain;
2944 case IOMMU_DOMAIN_IDENTITY:
2945 pdomain = protection_domain_alloc();
2949 pdomain->mode = PAGE_MODE_NONE;
2955 return &pdomain->domain;
2958 static void amd_iommu_domain_free(struct iommu_domain *dom)
2960 struct protection_domain *domain;
2961 struct dma_ops_domain *dma_dom;
2963 domain = to_pdomain(dom);
2965 if (domain->dev_cnt > 0)
2966 cleanup_domain(domain);
2968 BUG_ON(domain->dev_cnt != 0);
2973 switch (dom->type) {
2974 case IOMMU_DOMAIN_DMA:
2975 /* Now release the domain */
2976 dma_dom = to_dma_ops_domain(domain);
2977 dma_ops_domain_free(dma_dom);
2980 if (domain->mode != PAGE_MODE_NONE)
2981 free_pagetable(domain);
2983 if (domain->flags & PD_IOMMUV2_MASK)
2984 free_gcr3_table(domain);
2986 protection_domain_free(domain);
2991 static void amd_iommu_detach_device(struct iommu_domain *dom,
2994 struct iommu_dev_data *dev_data = dev->archdata.iommu;
2995 struct amd_iommu *iommu;
2998 if (!check_device(dev))
3001 devid = get_device_id(dev);
3005 if (dev_data->domain != NULL)
3008 iommu = amd_iommu_rlookup_table[devid];
3012 #ifdef CONFIG_IRQ_REMAP
3013 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) &&
3014 (dom->type == IOMMU_DOMAIN_UNMANAGED))
3015 dev_data->use_vapic = 0;
3018 iommu_completion_wait(iommu);
3021 static int amd_iommu_attach_device(struct iommu_domain *dom,
3024 struct protection_domain *domain = to_pdomain(dom);
3025 struct iommu_dev_data *dev_data;
3026 struct amd_iommu *iommu;
3029 if (!check_device(dev))
3032 dev_data = dev->archdata.iommu;
3034 iommu = amd_iommu_rlookup_table[dev_data->devid];
3038 if (dev_data->domain)
3041 ret = attach_device(dev, domain);
3043 #ifdef CONFIG_IRQ_REMAP
3044 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
3045 if (dom->type == IOMMU_DOMAIN_UNMANAGED)
3046 dev_data->use_vapic = 1;
3048 dev_data->use_vapic = 0;
3052 iommu_completion_wait(iommu);
3057 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
3058 phys_addr_t paddr, size_t page_size, int iommu_prot)
3060 struct protection_domain *domain = to_pdomain(dom);
3064 if (domain->mode == PAGE_MODE_NONE)
3067 if (iommu_prot & IOMMU_READ)
3068 prot |= IOMMU_PROT_IR;
3069 if (iommu_prot & IOMMU_WRITE)
3070 prot |= IOMMU_PROT_IW;
3072 mutex_lock(&domain->api_lock);
3073 ret = iommu_map_page(domain, iova, paddr, page_size, prot, GFP_KERNEL);
3074 mutex_unlock(&domain->api_lock);
3076 domain_flush_np_cache(domain, iova, page_size);
3081 static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
3084 struct protection_domain *domain = to_pdomain(dom);
3087 if (domain->mode == PAGE_MODE_NONE)
3090 mutex_lock(&domain->api_lock);
3091 unmap_size = iommu_unmap_page(domain, iova, page_size);
3092 mutex_unlock(&domain->api_lock);
3097 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
3100 struct protection_domain *domain = to_pdomain(dom);
3101 unsigned long offset_mask, pte_pgsize;
3104 if (domain->mode == PAGE_MODE_NONE)
3107 pte = fetch_pte(domain, iova, &pte_pgsize);
3109 if (!pte || !IOMMU_PTE_PRESENT(*pte))
3112 offset_mask = pte_pgsize - 1;
3113 __pte = __sme_clr(*pte & PM_ADDR_MASK);
3115 return (__pte & ~offset_mask) | (iova & offset_mask);
3118 static bool amd_iommu_capable(enum iommu_cap cap)
3121 case IOMMU_CAP_CACHE_COHERENCY:
3123 case IOMMU_CAP_INTR_REMAP:
3124 return (irq_remapping_enabled == 1);
3125 case IOMMU_CAP_NOEXEC:
3134 static void amd_iommu_get_resv_regions(struct device *dev,
3135 struct list_head *head)
3137 struct iommu_resv_region *region;
3138 struct unity_map_entry *entry;
3141 devid = get_device_id(dev);
3145 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
3149 if (devid < entry->devid_start || devid > entry->devid_end)
3152 type = IOMMU_RESV_DIRECT;
3153 length = entry->address_end - entry->address_start;
3154 if (entry->prot & IOMMU_PROT_IR)
3156 if (entry->prot & IOMMU_PROT_IW)
3157 prot |= IOMMU_WRITE;
3158 if (entry->prot & IOMMU_UNITY_MAP_FLAG_EXCL_RANGE)
3159 /* Exclusion range */
3160 type = IOMMU_RESV_RESERVED;
3162 region = iommu_alloc_resv_region(entry->address_start,
3163 length, prot, type);
3165 dev_err(dev, "Out of memory allocating dm-regions\n");
3168 list_add_tail(®ion->list, head);
3171 region = iommu_alloc_resv_region(MSI_RANGE_START,
3172 MSI_RANGE_END - MSI_RANGE_START + 1,
3176 list_add_tail(®ion->list, head);
3178 region = iommu_alloc_resv_region(HT_RANGE_START,
3179 HT_RANGE_END - HT_RANGE_START + 1,
3180 0, IOMMU_RESV_RESERVED);
3183 list_add_tail(®ion->list, head);
3186 static void amd_iommu_put_resv_regions(struct device *dev,
3187 struct list_head *head)
3189 struct iommu_resv_region *entry, *next;
3191 list_for_each_entry_safe(entry, next, head, list)
3195 static void amd_iommu_apply_resv_region(struct device *dev,
3196 struct iommu_domain *domain,
3197 struct iommu_resv_region *region)
3199 struct dma_ops_domain *dma_dom = to_dma_ops_domain(to_pdomain(domain));
3200 unsigned long start, end;
3202 start = IOVA_PFN(region->start);
3203 end = IOVA_PFN(region->start + region->length - 1);
3205 WARN_ON_ONCE(reserve_iova(&dma_dom->iovad, start, end) == NULL);
3208 static bool amd_iommu_is_attach_deferred(struct iommu_domain *domain,
3211 struct iommu_dev_data *dev_data = dev->archdata.iommu;
3212 return dev_data->defer_attach;
3215 static void amd_iommu_flush_iotlb_all(struct iommu_domain *domain)
3217 struct protection_domain *dom = to_pdomain(domain);
3219 domain_flush_tlb_pde(dom);
3220 domain_flush_complete(dom);
3223 static void amd_iommu_iotlb_range_add(struct iommu_domain *domain,
3224 unsigned long iova, size_t size)
3228 const struct iommu_ops amd_iommu_ops = {
3229 .capable = amd_iommu_capable,
3230 .domain_alloc = amd_iommu_domain_alloc,
3231 .domain_free = amd_iommu_domain_free,
3232 .attach_dev = amd_iommu_attach_device,
3233 .detach_dev = amd_iommu_detach_device,
3234 .map = amd_iommu_map,
3235 .unmap = amd_iommu_unmap,
3236 .iova_to_phys = amd_iommu_iova_to_phys,
3237 .add_device = amd_iommu_add_device,
3238 .remove_device = amd_iommu_remove_device,
3239 .device_group = amd_iommu_device_group,
3240 .get_resv_regions = amd_iommu_get_resv_regions,
3241 .put_resv_regions = amd_iommu_put_resv_regions,
3242 .apply_resv_region = amd_iommu_apply_resv_region,
3243 .is_attach_deferred = amd_iommu_is_attach_deferred,
3244 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
3245 .flush_iotlb_all = amd_iommu_flush_iotlb_all,
3246 .iotlb_range_add = amd_iommu_iotlb_range_add,
3247 .iotlb_sync = amd_iommu_flush_iotlb_all,
3250 /*****************************************************************************
3252 * The next functions do a basic initialization of IOMMU for pass through
3255 * In passthrough mode the IOMMU is initialized and enabled but not used for
3256 * DMA-API translation.
3258 *****************************************************************************/
3260 /* IOMMUv2 specific functions */
3261 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
3263 return atomic_notifier_chain_register(&ppr_notifier, nb);
3265 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
3267 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
3269 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
3271 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
3273 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
3275 struct protection_domain *domain = to_pdomain(dom);
3276 unsigned long flags;
3278 spin_lock_irqsave(&domain->lock, flags);
3280 /* Update data structure */
3281 domain->mode = PAGE_MODE_NONE;
3282 domain->updated = true;
3284 /* Make changes visible to IOMMUs */
3285 update_domain(domain);
3287 /* Page-table is not visible to IOMMU anymore, so free it */
3288 free_pagetable(domain);
3290 spin_unlock_irqrestore(&domain->lock, flags);
3292 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
3294 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
3296 struct protection_domain *domain = to_pdomain(dom);
3297 unsigned long flags;
3300 if (pasids <= 0 || pasids > (PASID_MASK + 1))
3303 /* Number of GCR3 table levels required */
3304 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
3307 if (levels > amd_iommu_max_glx_val)
3310 spin_lock_irqsave(&domain->lock, flags);
3313 * Save us all sanity checks whether devices already in the
3314 * domain support IOMMUv2. Just force that the domain has no
3315 * devices attached when it is switched into IOMMUv2 mode.
3318 if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
3322 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
3323 if (domain->gcr3_tbl == NULL)
3326 domain->glx = levels;
3327 domain->flags |= PD_IOMMUV2_MASK;
3328 domain->updated = true;
3330 update_domain(domain);
3335 spin_unlock_irqrestore(&domain->lock, flags);
3339 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
3341 static int __flush_pasid(struct protection_domain *domain, int pasid,
3342 u64 address, bool size)
3344 struct iommu_dev_data *dev_data;
3345 struct iommu_cmd cmd;
3348 if (!(domain->flags & PD_IOMMUV2_MASK))
3351 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
3354 * IOMMU TLB needs to be flushed before Device TLB to
3355 * prevent device TLB refill from IOMMU TLB
3357 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
3358 if (domain->dev_iommu[i] == 0)
3361 ret = iommu_queue_command(amd_iommus[i], &cmd);
3366 /* Wait until IOMMU TLB flushes are complete */
3367 domain_flush_complete(domain);
3369 /* Now flush device TLBs */
3370 list_for_each_entry(dev_data, &domain->dev_list, list) {
3371 struct amd_iommu *iommu;
3375 There might be non-IOMMUv2 capable devices in an IOMMUv2
3378 if (!dev_data->ats.enabled)
3381 qdep = dev_data->ats.qdep;
3382 iommu = amd_iommu_rlookup_table[dev_data->devid];
3384 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
3385 qdep, address, size);
3387 ret = iommu_queue_command(iommu, &cmd);
3392 /* Wait until all device TLBs are flushed */
3393 domain_flush_complete(domain);
3402 static int __amd_iommu_flush_page(struct protection_domain *domain, int pasid,
3405 return __flush_pasid(domain, pasid, address, false);
3408 int amd_iommu_flush_page(struct iommu_domain *dom, int pasid,
3411 struct protection_domain *domain = to_pdomain(dom);
3412 unsigned long flags;
3415 spin_lock_irqsave(&domain->lock, flags);
3416 ret = __amd_iommu_flush_page(domain, pasid, address);
3417 spin_unlock_irqrestore(&domain->lock, flags);
3421 EXPORT_SYMBOL(amd_iommu_flush_page);
3423 static int __amd_iommu_flush_tlb(struct protection_domain *domain, int pasid)
3425 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
3429 int amd_iommu_flush_tlb(struct iommu_domain *dom, int pasid)
3431 struct protection_domain *domain = to_pdomain(dom);
3432 unsigned long flags;
3435 spin_lock_irqsave(&domain->lock, flags);
3436 ret = __amd_iommu_flush_tlb(domain, pasid);
3437 spin_unlock_irqrestore(&domain->lock, flags);
3441 EXPORT_SYMBOL(amd_iommu_flush_tlb);
3443 static u64 *__get_gcr3_pte(u64 *root, int level, int pasid, bool alloc)
3450 index = (pasid >> (9 * level)) & 0x1ff;
3456 if (!(*pte & GCR3_VALID)) {
3460 root = (void *)get_zeroed_page(GFP_ATOMIC);
3464 *pte = iommu_virt_to_phys(root) | GCR3_VALID;
3467 root = iommu_phys_to_virt(*pte & PAGE_MASK);
3475 static int __set_gcr3(struct protection_domain *domain, int pasid,
3480 if (domain->mode != PAGE_MODE_NONE)
3483 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3487 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3489 return __amd_iommu_flush_tlb(domain, pasid);
3492 static int __clear_gcr3(struct protection_domain *domain, int pasid)
3496 if (domain->mode != PAGE_MODE_NONE)
3499 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3505 return __amd_iommu_flush_tlb(domain, pasid);
3508 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, int pasid,
3511 struct protection_domain *domain = to_pdomain(dom);
3512 unsigned long flags;
3515 spin_lock_irqsave(&domain->lock, flags);
3516 ret = __set_gcr3(domain, pasid, cr3);
3517 spin_unlock_irqrestore(&domain->lock, flags);
3521 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3523 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, int pasid)
3525 struct protection_domain *domain = to_pdomain(dom);
3526 unsigned long flags;
3529 spin_lock_irqsave(&domain->lock, flags);
3530 ret = __clear_gcr3(domain, pasid);
3531 spin_unlock_irqrestore(&domain->lock, flags);
3535 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3537 int amd_iommu_complete_ppr(struct pci_dev *pdev, int pasid,
3538 int status, int tag)
3540 struct iommu_dev_data *dev_data;
3541 struct amd_iommu *iommu;
3542 struct iommu_cmd cmd;
3544 dev_data = get_dev_data(&pdev->dev);
3545 iommu = amd_iommu_rlookup_table[dev_data->devid];
3547 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3548 tag, dev_data->pri_tlp);
3550 return iommu_queue_command(iommu, &cmd);
3552 EXPORT_SYMBOL(amd_iommu_complete_ppr);
3554 struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3556 struct protection_domain *pdomain;
3558 pdomain = get_domain(&pdev->dev);
3559 if (IS_ERR(pdomain))
3562 /* Only return IOMMUv2 domains */
3563 if (!(pdomain->flags & PD_IOMMUV2_MASK))
3566 return &pdomain->domain;
3568 EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3570 void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3572 struct iommu_dev_data *dev_data;
3574 if (!amd_iommu_v2_supported())
3577 dev_data = get_dev_data(&pdev->dev);
3578 dev_data->errata |= (1 << erratum);
3580 EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3582 int amd_iommu_device_info(struct pci_dev *pdev,
3583 struct amd_iommu_device_info *info)
3588 if (pdev == NULL || info == NULL)
3591 if (!amd_iommu_v2_supported())
3594 memset(info, 0, sizeof(*info));
3596 if (!pci_ats_disabled()) {
3597 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS);
3599 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3602 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3604 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3606 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3610 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3611 max_pasids = min(max_pasids, (1 << 20));
3613 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3614 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3616 features = pci_pasid_features(pdev);
3617 if (features & PCI_PASID_CAP_EXEC)
3618 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3619 if (features & PCI_PASID_CAP_PRIV)
3620 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3625 EXPORT_SYMBOL(amd_iommu_device_info);
3627 #ifdef CONFIG_IRQ_REMAP
3629 /*****************************************************************************
3631 * Interrupt Remapping Implementation
3633 *****************************************************************************/
3635 static struct irq_chip amd_ir_chip;
3636 static DEFINE_SPINLOCK(iommu_table_lock);
3638 static void set_dte_irq_entry(u16 devid, struct irq_remap_table *table)
3642 dte = amd_iommu_dev_table[devid].data[2];
3643 dte &= ~DTE_IRQ_PHYS_ADDR_MASK;
3644 dte |= iommu_virt_to_phys(table->table);
3645 dte |= DTE_IRQ_REMAP_INTCTL;
3646 dte |= DTE_IRQ_TABLE_LEN;
3647 dte |= DTE_IRQ_REMAP_ENABLE;
3649 amd_iommu_dev_table[devid].data[2] = dte;
3652 static struct irq_remap_table *get_irq_table(u16 devid)
3654 struct irq_remap_table *table;
3656 if (WARN_ONCE(!amd_iommu_rlookup_table[devid],
3657 "%s: no iommu for devid %x\n", __func__, devid))
3660 table = irq_lookup_table[devid];
3661 if (WARN_ONCE(!table, "%s: no table for devid %x\n", __func__, devid))
3667 static struct irq_remap_table *__alloc_irq_table(void)
3669 struct irq_remap_table *table;
3671 table = kzalloc(sizeof(*table), GFP_KERNEL);
3675 table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_KERNEL);
3676 if (!table->table) {
3680 raw_spin_lock_init(&table->lock);
3682 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
3683 memset(table->table, 0,
3684 MAX_IRQS_PER_TABLE * sizeof(u32));
3686 memset(table->table, 0,
3687 (MAX_IRQS_PER_TABLE * (sizeof(u64) * 2)));
3691 static void set_remap_table_entry(struct amd_iommu *iommu, u16 devid,
3692 struct irq_remap_table *table)
3694 irq_lookup_table[devid] = table;
3695 set_dte_irq_entry(devid, table);
3696 iommu_flush_dte(iommu, devid);
3699 static struct irq_remap_table *alloc_irq_table(u16 devid)
3701 struct irq_remap_table *table = NULL;
3702 struct irq_remap_table *new_table = NULL;
3703 struct amd_iommu *iommu;
3704 unsigned long flags;
3707 spin_lock_irqsave(&iommu_table_lock, flags);
3709 iommu = amd_iommu_rlookup_table[devid];
3713 table = irq_lookup_table[devid];
3717 alias = amd_iommu_alias_table[devid];
3718 table = irq_lookup_table[alias];
3720 set_remap_table_entry(iommu, devid, table);
3723 spin_unlock_irqrestore(&iommu_table_lock, flags);
3725 /* Nothing there yet, allocate new irq remapping table */
3726 new_table = __alloc_irq_table();
3730 spin_lock_irqsave(&iommu_table_lock, flags);
3732 table = irq_lookup_table[devid];
3736 table = irq_lookup_table[alias];
3738 set_remap_table_entry(iommu, devid, table);
3745 set_remap_table_entry(iommu, devid, table);
3747 set_remap_table_entry(iommu, alias, table);
3750 iommu_completion_wait(iommu);
3753 spin_unlock_irqrestore(&iommu_table_lock, flags);
3756 kmem_cache_free(amd_iommu_irq_cache, new_table->table);
3762 static int alloc_irq_index(u16 devid, int count, bool align)
3764 struct irq_remap_table *table;
3765 int index, c, alignment = 1;
3766 unsigned long flags;
3767 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
3772 table = alloc_irq_table(devid);
3777 alignment = roundup_pow_of_two(count);
3779 raw_spin_lock_irqsave(&table->lock, flags);
3781 /* Scan table for free entries */
3782 for (index = ALIGN(table->min_index, alignment), c = 0;
3783 index < MAX_IRQS_PER_TABLE;) {
3784 if (!iommu->irte_ops->is_allocated(table, index)) {
3788 index = ALIGN(index + 1, alignment);
3794 iommu->irte_ops->set_allocated(table, index - c + 1);
3806 raw_spin_unlock_irqrestore(&table->lock, flags);
3811 static int modify_irte_ga(u16 devid, int index, struct irte_ga *irte,
3812 struct amd_ir_data *data)
3814 struct irq_remap_table *table;
3815 struct amd_iommu *iommu;
3816 unsigned long flags;
3817 struct irte_ga *entry;
3819 iommu = amd_iommu_rlookup_table[devid];
3823 table = get_irq_table(devid);
3827 raw_spin_lock_irqsave(&table->lock, flags);
3829 entry = (struct irte_ga *)table->table;
3830 entry = &entry[index];
3831 entry->lo.fields_remap.valid = 0;
3832 entry->hi.val = irte->hi.val;
3833 entry->lo.val = irte->lo.val;
3834 entry->lo.fields_remap.valid = 1;
3838 raw_spin_unlock_irqrestore(&table->lock, flags);
3840 iommu_flush_irt(iommu, devid);
3841 iommu_completion_wait(iommu);
3846 static int modify_irte(u16 devid, int index, union irte *irte)
3848 struct irq_remap_table *table;
3849 struct amd_iommu *iommu;
3850 unsigned long flags;
3852 iommu = amd_iommu_rlookup_table[devid];
3856 table = get_irq_table(devid);
3860 raw_spin_lock_irqsave(&table->lock, flags);
3861 table->table[index] = irte->val;
3862 raw_spin_unlock_irqrestore(&table->lock, flags);
3864 iommu_flush_irt(iommu, devid);
3865 iommu_completion_wait(iommu);
3870 static void free_irte(u16 devid, int index)
3872 struct irq_remap_table *table;
3873 struct amd_iommu *iommu;
3874 unsigned long flags;
3876 iommu = amd_iommu_rlookup_table[devid];
3880 table = get_irq_table(devid);
3884 raw_spin_lock_irqsave(&table->lock, flags);
3885 iommu->irte_ops->clear_allocated(table, index);
3886 raw_spin_unlock_irqrestore(&table->lock, flags);
3888 iommu_flush_irt(iommu, devid);
3889 iommu_completion_wait(iommu);
3892 static void irte_prepare(void *entry,
3893 u32 delivery_mode, u32 dest_mode,
3894 u8 vector, u32 dest_apicid, int devid)
3896 union irte *irte = (union irte *) entry;
3899 irte->fields.vector = vector;
3900 irte->fields.int_type = delivery_mode;
3901 irte->fields.destination = dest_apicid;
3902 irte->fields.dm = dest_mode;
3903 irte->fields.valid = 1;
3906 static void irte_ga_prepare(void *entry,
3907 u32 delivery_mode, u32 dest_mode,
3908 u8 vector, u32 dest_apicid, int devid)
3910 struct irte_ga *irte = (struct irte_ga *) entry;
3914 irte->lo.fields_remap.int_type = delivery_mode;
3915 irte->lo.fields_remap.dm = dest_mode;
3916 irte->hi.fields.vector = vector;
3917 irte->lo.fields_remap.destination = APICID_TO_IRTE_DEST_LO(dest_apicid);
3918 irte->hi.fields.destination = APICID_TO_IRTE_DEST_HI(dest_apicid);
3919 irte->lo.fields_remap.valid = 1;
3922 static void irte_activate(void *entry, u16 devid, u16 index)
3924 union irte *irte = (union irte *) entry;
3926 irte->fields.valid = 1;
3927 modify_irte(devid, index, irte);
3930 static void irte_ga_activate(void *entry, u16 devid, u16 index)
3932 struct irte_ga *irte = (struct irte_ga *) entry;
3934 irte->lo.fields_remap.valid = 1;
3935 modify_irte_ga(devid, index, irte, NULL);
3938 static void irte_deactivate(void *entry, u16 devid, u16 index)
3940 union irte *irte = (union irte *) entry;
3942 irte->fields.valid = 0;
3943 modify_irte(devid, index, irte);
3946 static void irte_ga_deactivate(void *entry, u16 devid, u16 index)
3948 struct irte_ga *irte = (struct irte_ga *) entry;
3950 irte->lo.fields_remap.valid = 0;
3951 modify_irte_ga(devid, index, irte, NULL);
3954 static void irte_set_affinity(void *entry, u16 devid, u16 index,
3955 u8 vector, u32 dest_apicid)
3957 union irte *irte = (union irte *) entry;
3959 irte->fields.vector = vector;
3960 irte->fields.destination = dest_apicid;
3961 modify_irte(devid, index, irte);
3964 static void irte_ga_set_affinity(void *entry, u16 devid, u16 index,
3965 u8 vector, u32 dest_apicid)
3967 struct irte_ga *irte = (struct irte_ga *) entry;
3969 if (!irte->lo.fields_remap.guest_mode) {
3970 irte->hi.fields.vector = vector;
3971 irte->lo.fields_remap.destination =
3972 APICID_TO_IRTE_DEST_LO(dest_apicid);
3973 irte->hi.fields.destination =
3974 APICID_TO_IRTE_DEST_HI(dest_apicid);
3975 modify_irte_ga(devid, index, irte, NULL);
3979 #define IRTE_ALLOCATED (~1U)
3980 static void irte_set_allocated(struct irq_remap_table *table, int index)
3982 table->table[index] = IRTE_ALLOCATED;
3985 static void irte_ga_set_allocated(struct irq_remap_table *table, int index)
3987 struct irte_ga *ptr = (struct irte_ga *)table->table;
3988 struct irte_ga *irte = &ptr[index];
3990 memset(&irte->lo.val, 0, sizeof(u64));
3991 memset(&irte->hi.val, 0, sizeof(u64));
3992 irte->hi.fields.vector = 0xff;
3995 static bool irte_is_allocated(struct irq_remap_table *table, int index)
3997 union irte *ptr = (union irte *)table->table;
3998 union irte *irte = &ptr[index];
4000 return irte->val != 0;
4003 static bool irte_ga_is_allocated(struct irq_remap_table *table, int index)
4005 struct irte_ga *ptr = (struct irte_ga *)table->table;
4006 struct irte_ga *irte = &ptr[index];
4008 return irte->hi.fields.vector != 0;
4011 static void irte_clear_allocated(struct irq_remap_table *table, int index)
4013 table->table[index] = 0;
4016 static void irte_ga_clear_allocated(struct irq_remap_table *table, int index)
4018 struct irte_ga *ptr = (struct irte_ga *)table->table;
4019 struct irte_ga *irte = &ptr[index];
4021 memset(&irte->lo.val, 0, sizeof(u64));
4022 memset(&irte->hi.val, 0, sizeof(u64));
4025 static int get_devid(struct irq_alloc_info *info)
4029 switch (info->type) {
4030 case X86_IRQ_ALLOC_TYPE_IOAPIC:
4031 devid = get_ioapic_devid(info->ioapic_id);
4033 case X86_IRQ_ALLOC_TYPE_HPET:
4034 devid = get_hpet_devid(info->hpet_id);
4036 case X86_IRQ_ALLOC_TYPE_MSI:
4037 case X86_IRQ_ALLOC_TYPE_MSIX:
4038 devid = get_device_id(&info->msi_dev->dev);
4048 static struct irq_domain *get_ir_irq_domain(struct irq_alloc_info *info)
4050 struct amd_iommu *iommu;
4056 devid = get_devid(info);
4058 iommu = amd_iommu_rlookup_table[devid];
4060 return iommu->ir_domain;
4066 static struct irq_domain *get_irq_domain(struct irq_alloc_info *info)
4068 struct amd_iommu *iommu;
4074 switch (info->type) {
4075 case X86_IRQ_ALLOC_TYPE_MSI:
4076 case X86_IRQ_ALLOC_TYPE_MSIX:
4077 devid = get_device_id(&info->msi_dev->dev);
4081 iommu = amd_iommu_rlookup_table[devid];
4083 return iommu->msi_domain;
4092 struct irq_remap_ops amd_iommu_irq_ops = {
4093 .prepare = amd_iommu_prepare,
4094 .enable = amd_iommu_enable,
4095 .disable = amd_iommu_disable,
4096 .reenable = amd_iommu_reenable,
4097 .enable_faulting = amd_iommu_enable_faulting,
4098 .get_ir_irq_domain = get_ir_irq_domain,
4099 .get_irq_domain = get_irq_domain,
4102 static void irq_remapping_prepare_irte(struct amd_ir_data *data,
4103 struct irq_cfg *irq_cfg,
4104 struct irq_alloc_info *info,
4105 int devid, int index, int sub_handle)
4107 struct irq_2_irte *irte_info = &data->irq_2_irte;
4108 struct msi_msg *msg = &data->msi_entry;
4109 struct IO_APIC_route_entry *entry;
4110 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
4115 data->irq_2_irte.devid = devid;
4116 data->irq_2_irte.index = index + sub_handle;
4117 iommu->irte_ops->prepare(data->entry, apic->irq_delivery_mode,
4118 apic->irq_dest_mode, irq_cfg->vector,
4119 irq_cfg->dest_apicid, devid);
4121 switch (info->type) {
4122 case X86_IRQ_ALLOC_TYPE_IOAPIC:
4123 /* Setup IOAPIC entry */
4124 entry = info->ioapic_entry;
4125 info->ioapic_entry = NULL;
4126 memset(entry, 0, sizeof(*entry));
4127 entry->vector = index;
4129 entry->trigger = info->ioapic_trigger;
4130 entry->polarity = info->ioapic_polarity;
4131 /* Mask level triggered irqs. */
4132 if (info->ioapic_trigger)
4136 case X86_IRQ_ALLOC_TYPE_HPET:
4137 case X86_IRQ_ALLOC_TYPE_MSI:
4138 case X86_IRQ_ALLOC_TYPE_MSIX:
4139 msg->address_hi = MSI_ADDR_BASE_HI;
4140 msg->address_lo = MSI_ADDR_BASE_LO;
4141 msg->data = irte_info->index;
4150 struct amd_irte_ops irte_32_ops = {
4151 .prepare = irte_prepare,
4152 .activate = irte_activate,
4153 .deactivate = irte_deactivate,
4154 .set_affinity = irte_set_affinity,
4155 .set_allocated = irte_set_allocated,
4156 .is_allocated = irte_is_allocated,
4157 .clear_allocated = irte_clear_allocated,
4160 struct amd_irte_ops irte_128_ops = {
4161 .prepare = irte_ga_prepare,
4162 .activate = irte_ga_activate,
4163 .deactivate = irte_ga_deactivate,
4164 .set_affinity = irte_ga_set_affinity,
4165 .set_allocated = irte_ga_set_allocated,
4166 .is_allocated = irte_ga_is_allocated,
4167 .clear_allocated = irte_ga_clear_allocated,
4170 static int irq_remapping_alloc(struct irq_domain *domain, unsigned int virq,
4171 unsigned int nr_irqs, void *arg)
4173 struct irq_alloc_info *info = arg;
4174 struct irq_data *irq_data;
4175 struct amd_ir_data *data = NULL;
4176 struct irq_cfg *cfg;
4182 if (nr_irqs > 1 && info->type != X86_IRQ_ALLOC_TYPE_MSI &&
4183 info->type != X86_IRQ_ALLOC_TYPE_MSIX)
4187 * With IRQ remapping enabled, don't need contiguous CPU vectors
4188 * to support multiple MSI interrupts.
4190 if (info->type == X86_IRQ_ALLOC_TYPE_MSI)
4191 info->flags &= ~X86_IRQ_ALLOC_CONTIGUOUS_VECTORS;
4193 devid = get_devid(info);
4197 ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
4201 if (info->type == X86_IRQ_ALLOC_TYPE_IOAPIC) {
4202 struct irq_remap_table *table;
4203 struct amd_iommu *iommu;
4205 table = alloc_irq_table(devid);
4207 if (!table->min_index) {
4209 * Keep the first 32 indexes free for IOAPIC
4212 table->min_index = 32;
4213 iommu = amd_iommu_rlookup_table[devid];
4214 for (i = 0; i < 32; ++i)
4215 iommu->irte_ops->set_allocated(table, i);
4217 WARN_ON(table->min_index != 32);
4218 index = info->ioapic_pin;
4223 bool align = (info->type == X86_IRQ_ALLOC_TYPE_MSI);
4225 index = alloc_irq_index(devid, nr_irqs, align);
4228 pr_warn("Failed to allocate IRTE\n");
4230 goto out_free_parent;
4233 for (i = 0; i < nr_irqs; i++) {
4234 irq_data = irq_domain_get_irq_data(domain, virq + i);
4235 cfg = irqd_cfg(irq_data);
4236 if (!irq_data || !cfg) {
4242 data = kzalloc(sizeof(*data), GFP_KERNEL);
4246 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
4247 data->entry = kzalloc(sizeof(union irte), GFP_KERNEL);
4249 data->entry = kzalloc(sizeof(struct irte_ga),
4256 irq_data->hwirq = (devid << 16) + i;
4257 irq_data->chip_data = data;
4258 irq_data->chip = &amd_ir_chip;
4259 irq_remapping_prepare_irte(data, cfg, info, devid, index, i);
4260 irq_set_status_flags(virq + i, IRQ_MOVE_PCNTXT);
4266 for (i--; i >= 0; i--) {
4267 irq_data = irq_domain_get_irq_data(domain, virq + i);
4269 kfree(irq_data->chip_data);
4271 for (i = 0; i < nr_irqs; i++)
4272 free_irte(devid, index + i);
4274 irq_domain_free_irqs_common(domain, virq, nr_irqs);
4278 static void irq_remapping_free(struct irq_domain *domain, unsigned int virq,
4279 unsigned int nr_irqs)
4281 struct irq_2_irte *irte_info;
4282 struct irq_data *irq_data;
4283 struct amd_ir_data *data;
4286 for (i = 0; i < nr_irqs; i++) {
4287 irq_data = irq_domain_get_irq_data(domain, virq + i);
4288 if (irq_data && irq_data->chip_data) {
4289 data = irq_data->chip_data;
4290 irte_info = &data->irq_2_irte;
4291 free_irte(irte_info->devid, irte_info->index);
4296 irq_domain_free_irqs_common(domain, virq, nr_irqs);
4299 static void amd_ir_update_irte(struct irq_data *irqd, struct amd_iommu *iommu,
4300 struct amd_ir_data *ir_data,
4301 struct irq_2_irte *irte_info,
4302 struct irq_cfg *cfg);
4304 static int irq_remapping_activate(struct irq_domain *domain,
4305 struct irq_data *irq_data, bool reserve)
4307 struct amd_ir_data *data = irq_data->chip_data;
4308 struct irq_2_irte *irte_info = &data->irq_2_irte;
4309 struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4310 struct irq_cfg *cfg = irqd_cfg(irq_data);
4315 iommu->irte_ops->activate(data->entry, irte_info->devid,
4317 amd_ir_update_irte(irq_data, iommu, data, irte_info, cfg);
4321 static void irq_remapping_deactivate(struct irq_domain *domain,
4322 struct irq_data *irq_data)
4324 struct amd_ir_data *data = irq_data->chip_data;
4325 struct irq_2_irte *irte_info = &data->irq_2_irte;
4326 struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4329 iommu->irte_ops->deactivate(data->entry, irte_info->devid,
4333 static const struct irq_domain_ops amd_ir_domain_ops = {
4334 .alloc = irq_remapping_alloc,
4335 .free = irq_remapping_free,
4336 .activate = irq_remapping_activate,
4337 .deactivate = irq_remapping_deactivate,
4340 static int amd_ir_set_vcpu_affinity(struct irq_data *data, void *vcpu_info)
4342 struct amd_iommu *iommu;
4343 struct amd_iommu_pi_data *pi_data = vcpu_info;
4344 struct vcpu_data *vcpu_pi_info = pi_data->vcpu_data;
4345 struct amd_ir_data *ir_data = data->chip_data;
4346 struct irte_ga *irte = (struct irte_ga *) ir_data->entry;
4347 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
4348 struct iommu_dev_data *dev_data = search_dev_data(irte_info->devid);
4351 * This device has never been set up for guest mode.
4352 * we should not modify the IRTE
4354 if (!dev_data || !dev_data->use_vapic)
4357 pi_data->ir_data = ir_data;
4360 * SVM tries to set up for VAPIC mode, but we are in
4361 * legacy mode. So, we force legacy mode instead.
4363 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
4364 pr_debug("%s: Fall back to using intr legacy remap\n",
4366 pi_data->is_guest_mode = false;
4369 iommu = amd_iommu_rlookup_table[irte_info->devid];
4373 pi_data->prev_ga_tag = ir_data->cached_ga_tag;
4374 if (pi_data->is_guest_mode) {
4376 irte->hi.fields.ga_root_ptr = (pi_data->base >> 12);
4377 irte->hi.fields.vector = vcpu_pi_info->vector;
4378 irte->lo.fields_vapic.ga_log_intr = 1;
4379 irte->lo.fields_vapic.guest_mode = 1;
4380 irte->lo.fields_vapic.ga_tag = pi_data->ga_tag;
4382 ir_data->cached_ga_tag = pi_data->ga_tag;
4385 struct irq_cfg *cfg = irqd_cfg(data);
4389 irte->hi.fields.vector = cfg->vector;
4390 irte->lo.fields_remap.guest_mode = 0;
4391 irte->lo.fields_remap.destination =
4392 APICID_TO_IRTE_DEST_LO(cfg->dest_apicid);
4393 irte->hi.fields.destination =
4394 APICID_TO_IRTE_DEST_HI(cfg->dest_apicid);
4395 irte->lo.fields_remap.int_type = apic->irq_delivery_mode;
4396 irte->lo.fields_remap.dm = apic->irq_dest_mode;
4399 * This communicates the ga_tag back to the caller
4400 * so that it can do all the necessary clean up.
4402 ir_data->cached_ga_tag = 0;
4405 return modify_irte_ga(irte_info->devid, irte_info->index, irte, ir_data);
4409 static void amd_ir_update_irte(struct irq_data *irqd, struct amd_iommu *iommu,
4410 struct amd_ir_data *ir_data,
4411 struct irq_2_irte *irte_info,
4412 struct irq_cfg *cfg)
4416 * Atomically updates the IRTE with the new destination, vector
4417 * and flushes the interrupt entry cache.
4419 iommu->irte_ops->set_affinity(ir_data->entry, irte_info->devid,
4420 irte_info->index, cfg->vector,
4424 static int amd_ir_set_affinity(struct irq_data *data,
4425 const struct cpumask *mask, bool force)
4427 struct amd_ir_data *ir_data = data->chip_data;
4428 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
4429 struct irq_cfg *cfg = irqd_cfg(data);
4430 struct irq_data *parent = data->parent_data;
4431 struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4437 ret = parent->chip->irq_set_affinity(parent, mask, force);
4438 if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
4441 amd_ir_update_irte(data, iommu, ir_data, irte_info, cfg);
4443 * After this point, all the interrupts will start arriving
4444 * at the new destination. So, time to cleanup the previous
4445 * vector allocation.
4447 send_cleanup_vector(cfg);
4449 return IRQ_SET_MASK_OK_DONE;
4452 static void ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg)
4454 struct amd_ir_data *ir_data = irq_data->chip_data;
4456 *msg = ir_data->msi_entry;
4459 static struct irq_chip amd_ir_chip = {
4461 .irq_ack = apic_ack_irq,
4462 .irq_set_affinity = amd_ir_set_affinity,
4463 .irq_set_vcpu_affinity = amd_ir_set_vcpu_affinity,
4464 .irq_compose_msi_msg = ir_compose_msi_msg,
4467 int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
4469 struct fwnode_handle *fn;
4471 fn = irq_domain_alloc_named_id_fwnode("AMD-IR", iommu->index);
4474 iommu->ir_domain = irq_domain_create_tree(fn, &amd_ir_domain_ops, iommu);
4475 irq_domain_free_fwnode(fn);
4476 if (!iommu->ir_domain)
4479 iommu->ir_domain->parent = arch_get_ir_parent_domain();
4480 iommu->msi_domain = arch_create_remap_msi_irq_domain(iommu->ir_domain,
4486 int amd_iommu_update_ga(int cpu, bool is_run, void *data)
4488 unsigned long flags;
4489 struct amd_iommu *iommu;
4490 struct irq_remap_table *table;
4491 struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
4492 int devid = ir_data->irq_2_irte.devid;
4493 struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
4494 struct irte_ga *ref = (struct irte_ga *) ir_data->ref;
4496 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
4497 !ref || !entry || !entry->lo.fields_vapic.guest_mode)
4500 iommu = amd_iommu_rlookup_table[devid];
4504 table = get_irq_table(devid);
4508 raw_spin_lock_irqsave(&table->lock, flags);
4510 if (ref->lo.fields_vapic.guest_mode) {
4512 ref->lo.fields_vapic.destination =
4513 APICID_TO_IRTE_DEST_LO(cpu);
4514 ref->hi.fields.destination =
4515 APICID_TO_IRTE_DEST_HI(cpu);
4517 ref->lo.fields_vapic.is_run = is_run;
4521 raw_spin_unlock_irqrestore(&table->lock, flags);
4523 iommu_flush_irt(iommu, devid);
4524 iommu_completion_wait(iommu);
4527 EXPORT_SYMBOL(amd_iommu_update_ga);