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(pd_bitmap_lock);
75 /* List of all available dev_data structures */
76 static LLIST_HEAD(dev_data_list);
78 LIST_HEAD(ioapic_map);
80 LIST_HEAD(acpihid_map);
83 * Domain for untranslated devices - only allocated
84 * if iommu=pt passed on kernel cmd line.
86 const struct iommu_ops amd_iommu_ops;
88 static ATOMIC_NOTIFIER_HEAD(ppr_notifier);
89 int amd_iommu_max_glx_val = -1;
91 static const struct dma_map_ops amd_iommu_dma_ops;
94 * general struct to manage commands send to an IOMMU
100 struct kmem_cache *amd_iommu_irq_cache;
102 static void update_domain(struct protection_domain *domain);
103 static int protection_domain_init(struct protection_domain *domain);
104 static void detach_device(struct device *dev);
105 static void iova_domain_flush_tlb(struct iova_domain *iovad);
108 * Data container for a dma_ops specific protection domain
110 struct dma_ops_domain {
111 /* generic protection domain information */
112 struct protection_domain domain;
115 struct iova_domain iovad;
118 static struct iova_domain reserved_iova_ranges;
119 static struct lock_class_key reserved_rbtree_key;
121 /****************************************************************************
125 ****************************************************************************/
127 static inline int match_hid_uid(struct device *dev,
128 struct acpihid_map_entry *entry)
130 struct acpi_device *adev = ACPI_COMPANION(dev);
131 const char *hid, *uid;
136 hid = acpi_device_hid(adev);
137 uid = acpi_device_uid(adev);
143 return strcmp(hid, entry->hid);
146 return strcmp(hid, entry->hid);
148 return (strcmp(hid, entry->hid) || strcmp(uid, entry->uid));
151 static inline u16 get_pci_device_id(struct device *dev)
153 struct pci_dev *pdev = to_pci_dev(dev);
155 return pci_dev_id(pdev);
158 static inline int get_acpihid_device_id(struct device *dev,
159 struct acpihid_map_entry **entry)
161 struct acpihid_map_entry *p;
163 list_for_each_entry(p, &acpihid_map, list) {
164 if (!match_hid_uid(dev, p)) {
173 static inline int get_device_id(struct device *dev)
178 devid = get_pci_device_id(dev);
180 devid = get_acpihid_device_id(dev, NULL);
185 static struct protection_domain *to_pdomain(struct iommu_domain *dom)
187 return container_of(dom, struct protection_domain, domain);
190 static struct dma_ops_domain* to_dma_ops_domain(struct protection_domain *domain)
192 BUG_ON(domain->flags != PD_DMA_OPS_MASK);
193 return container_of(domain, struct dma_ops_domain, domain);
196 static struct iommu_dev_data *alloc_dev_data(u16 devid)
198 struct iommu_dev_data *dev_data;
200 dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
204 dev_data->devid = devid;
205 ratelimit_default_init(&dev_data->rs);
207 llist_add(&dev_data->dev_data_list, &dev_data_list);
211 static struct iommu_dev_data *search_dev_data(u16 devid)
213 struct iommu_dev_data *dev_data;
214 struct llist_node *node;
216 if (llist_empty(&dev_data_list))
219 node = dev_data_list.first;
220 llist_for_each_entry(dev_data, node, dev_data_list) {
221 if (dev_data->devid == devid)
228 static int __last_alias(struct pci_dev *pdev, u16 alias, void *data)
230 *(u16 *)data = alias;
234 static u16 get_alias(struct device *dev)
236 struct pci_dev *pdev = to_pci_dev(dev);
237 u16 devid, ivrs_alias, pci_alias;
239 /* The callers make sure that get_device_id() does not fail here */
240 devid = get_device_id(dev);
242 /* For ACPI HID devices, we simply return the devid as such */
243 if (!dev_is_pci(dev))
246 ivrs_alias = amd_iommu_alias_table[devid];
248 pci_for_each_dma_alias(pdev, __last_alias, &pci_alias);
250 if (ivrs_alias == pci_alias)
256 * The IVRS is fairly reliable in telling us about aliases, but it
257 * can't know about every screwy device. If we don't have an IVRS
258 * reported alias, use the PCI reported alias. In that case we may
259 * still need to initialize the rlookup and dev_table entries if the
260 * alias is to a non-existent device.
262 if (ivrs_alias == devid) {
263 if (!amd_iommu_rlookup_table[pci_alias]) {
264 amd_iommu_rlookup_table[pci_alias] =
265 amd_iommu_rlookup_table[devid];
266 memcpy(amd_iommu_dev_table[pci_alias].data,
267 amd_iommu_dev_table[devid].data,
268 sizeof(amd_iommu_dev_table[pci_alias].data));
274 pci_info(pdev, "Using IVRS reported alias %02x:%02x.%d "
275 "for device [%04x:%04x], kernel reported alias "
276 "%02x:%02x.%d\n", PCI_BUS_NUM(ivrs_alias), PCI_SLOT(ivrs_alias),
277 PCI_FUNC(ivrs_alias), pdev->vendor, pdev->device,
278 PCI_BUS_NUM(pci_alias), PCI_SLOT(pci_alias),
279 PCI_FUNC(pci_alias));
282 * If we don't have a PCI DMA alias and the IVRS alias is on the same
283 * bus, then the IVRS table may know about a quirk that we don't.
285 if (pci_alias == devid &&
286 PCI_BUS_NUM(ivrs_alias) == pdev->bus->number) {
287 pci_add_dma_alias(pdev, ivrs_alias & 0xff);
288 pci_info(pdev, "Added PCI DMA alias %02x.%d\n",
289 PCI_SLOT(ivrs_alias), PCI_FUNC(ivrs_alias));
295 static struct iommu_dev_data *find_dev_data(u16 devid)
297 struct iommu_dev_data *dev_data;
298 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
300 dev_data = search_dev_data(devid);
302 if (dev_data == NULL) {
303 dev_data = alloc_dev_data(devid);
307 if (translation_pre_enabled(iommu))
308 dev_data->defer_attach = true;
314 struct iommu_dev_data *get_dev_data(struct device *dev)
316 return dev->archdata.iommu;
318 EXPORT_SYMBOL(get_dev_data);
321 * Find or create an IOMMU group for a acpihid device.
323 static struct iommu_group *acpihid_device_group(struct device *dev)
325 struct acpihid_map_entry *p, *entry = NULL;
328 devid = get_acpihid_device_id(dev, &entry);
330 return ERR_PTR(devid);
332 list_for_each_entry(p, &acpihid_map, list) {
333 if ((devid == p->devid) && p->group)
334 entry->group = p->group;
338 entry->group = generic_device_group(dev);
340 iommu_group_ref_get(entry->group);
345 static bool pci_iommuv2_capable(struct pci_dev *pdev)
347 static const int caps[] = {
350 PCI_EXT_CAP_ID_PASID,
354 if (pci_ats_disabled())
357 for (i = 0; i < 3; ++i) {
358 pos = pci_find_ext_capability(pdev, caps[i]);
366 static bool pdev_pri_erratum(struct pci_dev *pdev, u32 erratum)
368 struct iommu_dev_data *dev_data;
370 dev_data = get_dev_data(&pdev->dev);
372 return dev_data->errata & (1 << erratum) ? true : false;
376 * This function checks if the driver got a valid device from the caller to
377 * avoid dereferencing invalid pointers.
379 static bool check_device(struct device *dev)
383 if (!dev || !dev->dma_mask)
386 devid = get_device_id(dev);
390 /* Out of our scope? */
391 if (devid > amd_iommu_last_bdf)
394 if (amd_iommu_rlookup_table[devid] == NULL)
400 static void init_iommu_group(struct device *dev)
402 struct iommu_group *group;
404 group = iommu_group_get_for_dev(dev);
408 iommu_group_put(group);
411 static int iommu_init_device(struct device *dev)
413 struct iommu_dev_data *dev_data;
414 struct amd_iommu *iommu;
417 if (dev->archdata.iommu)
420 devid = get_device_id(dev);
424 iommu = amd_iommu_rlookup_table[devid];
426 dev_data = find_dev_data(devid);
430 dev_data->alias = get_alias(dev);
433 * By default we use passthrough mode for IOMMUv2 capable device.
434 * But if amd_iommu=force_isolation is set (e.g. to debug DMA to
435 * invalid address), we ignore the capability for the device so
436 * it'll be forced to go into translation mode.
438 if ((iommu_default_passthrough() || !amd_iommu_force_isolation) &&
439 dev_is_pci(dev) && pci_iommuv2_capable(to_pci_dev(dev))) {
440 struct amd_iommu *iommu;
442 iommu = amd_iommu_rlookup_table[dev_data->devid];
443 dev_data->iommu_v2 = iommu->is_iommu_v2;
446 dev->archdata.iommu = dev_data;
448 iommu_device_link(&iommu->iommu, dev);
453 static void iommu_ignore_device(struct device *dev)
458 devid = get_device_id(dev);
462 alias = get_alias(dev);
464 memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
465 memset(&amd_iommu_dev_table[alias], 0, sizeof(struct dev_table_entry));
467 amd_iommu_rlookup_table[devid] = NULL;
468 amd_iommu_rlookup_table[alias] = NULL;
471 static void iommu_uninit_device(struct device *dev)
473 struct iommu_dev_data *dev_data;
474 struct amd_iommu *iommu;
477 devid = get_device_id(dev);
481 iommu = amd_iommu_rlookup_table[devid];
483 dev_data = search_dev_data(devid);
487 if (dev_data->domain)
490 iommu_device_unlink(&iommu->iommu, dev);
492 iommu_group_remove_device(dev);
498 * We keep dev_data around for unplugged devices and reuse it when the
499 * device is re-plugged - not doing so would introduce a ton of races.
504 * Helper function to get the first pte of a large mapping
506 static u64 *first_pte_l7(u64 *pte, unsigned long *page_size,
507 unsigned long *count)
509 unsigned long pte_mask, pg_size, cnt;
512 pg_size = PTE_PAGE_SIZE(*pte);
513 cnt = PAGE_SIZE_PTE_COUNT(pg_size);
514 pte_mask = ~((cnt << 3) - 1);
515 fpte = (u64 *)(((unsigned long)pte) & pte_mask);
518 *page_size = pg_size;
526 /****************************************************************************
528 * Interrupt handling functions
530 ****************************************************************************/
532 static void dump_dte_entry(u16 devid)
536 for (i = 0; i < 4; ++i)
537 pr_err("DTE[%d]: %016llx\n", i,
538 amd_iommu_dev_table[devid].data[i]);
541 static void dump_command(unsigned long phys_addr)
543 struct iommu_cmd *cmd = iommu_phys_to_virt(phys_addr);
546 for (i = 0; i < 4; ++i)
547 pr_err("CMD[%d]: %08x\n", i, cmd->data[i]);
550 static void amd_iommu_report_page_fault(u16 devid, u16 domain_id,
551 u64 address, int flags)
553 struct iommu_dev_data *dev_data = NULL;
554 struct pci_dev *pdev;
556 pdev = pci_get_domain_bus_and_slot(0, PCI_BUS_NUM(devid),
559 dev_data = get_dev_data(&pdev->dev);
561 if (dev_data && __ratelimit(&dev_data->rs)) {
562 pci_err(pdev, "Event logged [IO_PAGE_FAULT domain=0x%04x address=0x%llx flags=0x%04x]\n",
563 domain_id, address, flags);
564 } else if (printk_ratelimit()) {
565 pr_err("Event logged [IO_PAGE_FAULT device=%02x:%02x.%x domain=0x%04x address=0x%llx flags=0x%04x]\n",
566 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
567 domain_id, address, flags);
574 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
576 struct device *dev = iommu->iommu.dev;
577 int type, devid, pasid, flags, tag;
578 volatile u32 *event = __evt;
583 type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
584 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
585 pasid = PPR_PASID(*(u64 *)&event[0]);
586 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
587 address = (u64)(((u64)event[3]) << 32) | event[2];
590 /* Did we hit the erratum? */
591 if (++count == LOOP_TIMEOUT) {
592 pr_err("No event written to event log\n");
599 if (type == EVENT_TYPE_IO_FAULT) {
600 amd_iommu_report_page_fault(devid, pasid, address, flags);
605 case EVENT_TYPE_ILL_DEV:
606 dev_err(dev, "Event logged [ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x pasid=0x%05x address=0x%llx flags=0x%04x]\n",
607 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
608 pasid, address, flags);
609 dump_dte_entry(devid);
611 case EVENT_TYPE_DEV_TAB_ERR:
612 dev_err(dev, "Event logged [DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
613 "address=0x%llx flags=0x%04x]\n",
614 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
617 case EVENT_TYPE_PAGE_TAB_ERR:
618 dev_err(dev, "Event logged [PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x domain=0x%04x address=0x%llx flags=0x%04x]\n",
619 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
620 pasid, address, flags);
622 case EVENT_TYPE_ILL_CMD:
623 dev_err(dev, "Event logged [ILLEGAL_COMMAND_ERROR address=0x%llx]\n", address);
624 dump_command(address);
626 case EVENT_TYPE_CMD_HARD_ERR:
627 dev_err(dev, "Event logged [COMMAND_HARDWARE_ERROR address=0x%llx flags=0x%04x]\n",
630 case EVENT_TYPE_IOTLB_INV_TO:
631 dev_err(dev, "Event logged [IOTLB_INV_TIMEOUT device=%02x:%02x.%x address=0x%llx]\n",
632 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
635 case EVENT_TYPE_INV_DEV_REQ:
636 dev_err(dev, "Event logged [INVALID_DEVICE_REQUEST device=%02x:%02x.%x pasid=0x%05x address=0x%llx flags=0x%04x]\n",
637 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
638 pasid, address, flags);
640 case EVENT_TYPE_INV_PPR_REQ:
641 pasid = ((event[0] >> 16) & 0xFFFF)
642 | ((event[1] << 6) & 0xF0000);
643 tag = event[1] & 0x03FF;
644 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",
645 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
646 pasid, address, flags, tag);
649 dev_err(dev, "Event logged [UNKNOWN event[0]=0x%08x event[1]=0x%08x event[2]=0x%08x event[3]=0x%08x\n",
650 event[0], event[1], event[2], event[3]);
653 memset(__evt, 0, 4 * sizeof(u32));
656 static void iommu_poll_events(struct amd_iommu *iommu)
660 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
661 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
663 while (head != tail) {
664 iommu_print_event(iommu, iommu->evt_buf + head);
665 head = (head + EVENT_ENTRY_SIZE) % EVT_BUFFER_SIZE;
668 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
671 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
673 struct amd_iommu_fault fault;
675 if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
676 pr_err_ratelimited("Unknown PPR request received\n");
680 fault.address = raw[1];
681 fault.pasid = PPR_PASID(raw[0]);
682 fault.device_id = PPR_DEVID(raw[0]);
683 fault.tag = PPR_TAG(raw[0]);
684 fault.flags = PPR_FLAGS(raw[0]);
686 atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
689 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
693 if (iommu->ppr_log == NULL)
696 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
697 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
699 while (head != tail) {
704 raw = (u64 *)(iommu->ppr_log + head);
707 * Hardware bug: Interrupt may arrive before the entry is
708 * written to memory. If this happens we need to wait for the
711 for (i = 0; i < LOOP_TIMEOUT; ++i) {
712 if (PPR_REQ_TYPE(raw[0]) != 0)
717 /* Avoid memcpy function-call overhead */
722 * To detect the hardware bug we need to clear the entry
725 raw[0] = raw[1] = 0UL;
727 /* Update head pointer of hardware ring-buffer */
728 head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
729 writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
731 /* Handle PPR entry */
732 iommu_handle_ppr_entry(iommu, entry);
734 /* Refresh ring-buffer information */
735 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
736 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
740 #ifdef CONFIG_IRQ_REMAP
741 static int (*iommu_ga_log_notifier)(u32);
743 int amd_iommu_register_ga_log_notifier(int (*notifier)(u32))
745 iommu_ga_log_notifier = notifier;
749 EXPORT_SYMBOL(amd_iommu_register_ga_log_notifier);
751 static void iommu_poll_ga_log(struct amd_iommu *iommu)
753 u32 head, tail, cnt = 0;
755 if (iommu->ga_log == NULL)
758 head = readl(iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
759 tail = readl(iommu->mmio_base + MMIO_GA_TAIL_OFFSET);
761 while (head != tail) {
765 raw = (u64 *)(iommu->ga_log + head);
768 /* Avoid memcpy function-call overhead */
771 /* Update head pointer of hardware ring-buffer */
772 head = (head + GA_ENTRY_SIZE) % GA_LOG_SIZE;
773 writel(head, iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
775 /* Handle GA entry */
776 switch (GA_REQ_TYPE(log_entry)) {
778 if (!iommu_ga_log_notifier)
781 pr_debug("%s: devid=%#x, ga_tag=%#x\n",
782 __func__, GA_DEVID(log_entry),
785 if (iommu_ga_log_notifier(GA_TAG(log_entry)) != 0)
786 pr_err("GA log notifier failed.\n");
793 #endif /* CONFIG_IRQ_REMAP */
795 #define AMD_IOMMU_INT_MASK \
796 (MMIO_STATUS_EVT_INT_MASK | \
797 MMIO_STATUS_PPR_INT_MASK | \
798 MMIO_STATUS_GALOG_INT_MASK)
800 irqreturn_t amd_iommu_int_thread(int irq, void *data)
802 struct amd_iommu *iommu = (struct amd_iommu *) data;
803 u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
805 while (status & AMD_IOMMU_INT_MASK) {
806 /* Enable EVT and PPR and GA interrupts again */
807 writel(AMD_IOMMU_INT_MASK,
808 iommu->mmio_base + MMIO_STATUS_OFFSET);
810 if (status & MMIO_STATUS_EVT_INT_MASK) {
811 pr_devel("Processing IOMMU Event Log\n");
812 iommu_poll_events(iommu);
815 if (status & MMIO_STATUS_PPR_INT_MASK) {
816 pr_devel("Processing IOMMU PPR Log\n");
817 iommu_poll_ppr_log(iommu);
820 #ifdef CONFIG_IRQ_REMAP
821 if (status & MMIO_STATUS_GALOG_INT_MASK) {
822 pr_devel("Processing IOMMU GA Log\n");
823 iommu_poll_ga_log(iommu);
828 * Hardware bug: ERBT1312
829 * When re-enabling interrupt (by writing 1
830 * to clear the bit), the hardware might also try to set
831 * the interrupt bit in the event status register.
832 * In this scenario, the bit will be set, and disable
833 * subsequent interrupts.
835 * Workaround: The IOMMU driver should read back the
836 * status register and check if the interrupt bits are cleared.
837 * If not, driver will need to go through the interrupt handler
838 * again and re-clear the bits
840 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
845 irqreturn_t amd_iommu_int_handler(int irq, void *data)
847 return IRQ_WAKE_THREAD;
850 /****************************************************************************
852 * IOMMU command queuing functions
854 ****************************************************************************/
856 static int wait_on_sem(volatile u64 *sem)
860 while (*sem == 0 && i < LOOP_TIMEOUT) {
865 if (i == LOOP_TIMEOUT) {
866 pr_alert("Completion-Wait loop timed out\n");
873 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
874 struct iommu_cmd *cmd)
878 target = iommu->cmd_buf + iommu->cmd_buf_tail;
880 iommu->cmd_buf_tail += sizeof(*cmd);
881 iommu->cmd_buf_tail %= CMD_BUFFER_SIZE;
883 /* Copy command to buffer */
884 memcpy(target, cmd, sizeof(*cmd));
886 /* Tell the IOMMU about it */
887 writel(iommu->cmd_buf_tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
890 static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
892 u64 paddr = iommu_virt_to_phys((void *)address);
894 WARN_ON(address & 0x7ULL);
896 memset(cmd, 0, sizeof(*cmd));
897 cmd->data[0] = lower_32_bits(paddr) | CMD_COMPL_WAIT_STORE_MASK;
898 cmd->data[1] = upper_32_bits(paddr);
900 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
903 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
905 memset(cmd, 0, sizeof(*cmd));
906 cmd->data[0] = devid;
907 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
910 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
911 size_t size, u16 domid, int pde)
916 pages = iommu_num_pages(address, size, PAGE_SIZE);
921 * If we have to flush more than one page, flush all
922 * TLB entries for this domain
924 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
928 address &= PAGE_MASK;
930 memset(cmd, 0, sizeof(*cmd));
931 cmd->data[1] |= domid;
932 cmd->data[2] = lower_32_bits(address);
933 cmd->data[3] = upper_32_bits(address);
934 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
935 if (s) /* size bit - we flush more than one 4kb page */
936 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
937 if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
938 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
941 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
942 u64 address, size_t size)
947 pages = iommu_num_pages(address, size, PAGE_SIZE);
952 * If we have to flush more than one page, flush all
953 * TLB entries for this domain
955 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
959 address &= PAGE_MASK;
961 memset(cmd, 0, sizeof(*cmd));
962 cmd->data[0] = devid;
963 cmd->data[0] |= (qdep & 0xff) << 24;
964 cmd->data[1] = devid;
965 cmd->data[2] = lower_32_bits(address);
966 cmd->data[3] = upper_32_bits(address);
967 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
969 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
972 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, int pasid,
973 u64 address, bool size)
975 memset(cmd, 0, sizeof(*cmd));
977 address &= ~(0xfffULL);
979 cmd->data[0] = pasid;
980 cmd->data[1] = domid;
981 cmd->data[2] = lower_32_bits(address);
982 cmd->data[3] = upper_32_bits(address);
983 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
984 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
986 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
987 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
990 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, int pasid,
991 int qdep, u64 address, bool size)
993 memset(cmd, 0, sizeof(*cmd));
995 address &= ~(0xfffULL);
997 cmd->data[0] = devid;
998 cmd->data[0] |= ((pasid >> 8) & 0xff) << 16;
999 cmd->data[0] |= (qdep & 0xff) << 24;
1000 cmd->data[1] = devid;
1001 cmd->data[1] |= (pasid & 0xff) << 16;
1002 cmd->data[2] = lower_32_bits(address);
1003 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
1004 cmd->data[3] = upper_32_bits(address);
1006 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
1007 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
1010 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, int pasid,
1011 int status, int tag, bool gn)
1013 memset(cmd, 0, sizeof(*cmd));
1015 cmd->data[0] = devid;
1017 cmd->data[1] = pasid;
1018 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
1020 cmd->data[3] = tag & 0x1ff;
1021 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
1023 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
1026 static void build_inv_all(struct iommu_cmd *cmd)
1028 memset(cmd, 0, sizeof(*cmd));
1029 CMD_SET_TYPE(cmd, CMD_INV_ALL);
1032 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
1034 memset(cmd, 0, sizeof(*cmd));
1035 cmd->data[0] = devid;
1036 CMD_SET_TYPE(cmd, CMD_INV_IRT);
1040 * Writes the command to the IOMMUs command buffer and informs the
1041 * hardware about the new command.
1043 static int __iommu_queue_command_sync(struct amd_iommu *iommu,
1044 struct iommu_cmd *cmd,
1047 unsigned int count = 0;
1048 u32 left, next_tail;
1050 next_tail = (iommu->cmd_buf_tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
1052 left = (iommu->cmd_buf_head - next_tail) % CMD_BUFFER_SIZE;
1055 /* Skip udelay() the first time around */
1057 if (count == LOOP_TIMEOUT) {
1058 pr_err("Command buffer timeout\n");
1065 /* Update head and recheck remaining space */
1066 iommu->cmd_buf_head = readl(iommu->mmio_base +
1067 MMIO_CMD_HEAD_OFFSET);
1072 copy_cmd_to_buffer(iommu, cmd);
1074 /* Do we need to make sure all commands are processed? */
1075 iommu->need_sync = sync;
1080 static int iommu_queue_command_sync(struct amd_iommu *iommu,
1081 struct iommu_cmd *cmd,
1084 unsigned long flags;
1087 raw_spin_lock_irqsave(&iommu->lock, flags);
1088 ret = __iommu_queue_command_sync(iommu, cmd, sync);
1089 raw_spin_unlock_irqrestore(&iommu->lock, flags);
1094 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
1096 return iommu_queue_command_sync(iommu, cmd, true);
1100 * This function queues a completion wait command into the command
1101 * buffer of an IOMMU
1103 static int iommu_completion_wait(struct amd_iommu *iommu)
1105 struct iommu_cmd cmd;
1106 unsigned long flags;
1109 if (!iommu->need_sync)
1113 build_completion_wait(&cmd, (u64)&iommu->cmd_sem);
1115 raw_spin_lock_irqsave(&iommu->lock, flags);
1119 ret = __iommu_queue_command_sync(iommu, &cmd, false);
1123 ret = wait_on_sem(&iommu->cmd_sem);
1126 raw_spin_unlock_irqrestore(&iommu->lock, flags);
1131 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
1133 struct iommu_cmd cmd;
1135 build_inv_dte(&cmd, devid);
1137 return iommu_queue_command(iommu, &cmd);
1140 static void amd_iommu_flush_dte_all(struct amd_iommu *iommu)
1144 for (devid = 0; devid <= 0xffff; ++devid)
1145 iommu_flush_dte(iommu, devid);
1147 iommu_completion_wait(iommu);
1151 * This function uses heavy locking and may disable irqs for some time. But
1152 * this is no issue because it is only called during resume.
1154 static void amd_iommu_flush_tlb_all(struct amd_iommu *iommu)
1158 for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
1159 struct iommu_cmd cmd;
1160 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1162 iommu_queue_command(iommu, &cmd);
1165 iommu_completion_wait(iommu);
1168 static void amd_iommu_flush_tlb_domid(struct amd_iommu *iommu, u32 dom_id)
1170 struct iommu_cmd cmd;
1172 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1174 iommu_queue_command(iommu, &cmd);
1176 iommu_completion_wait(iommu);
1179 static void amd_iommu_flush_all(struct amd_iommu *iommu)
1181 struct iommu_cmd cmd;
1183 build_inv_all(&cmd);
1185 iommu_queue_command(iommu, &cmd);
1186 iommu_completion_wait(iommu);
1189 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1191 struct iommu_cmd cmd;
1193 build_inv_irt(&cmd, devid);
1195 iommu_queue_command(iommu, &cmd);
1198 static void amd_iommu_flush_irt_all(struct amd_iommu *iommu)
1202 for (devid = 0; devid <= MAX_DEV_TABLE_ENTRIES; devid++)
1203 iommu_flush_irt(iommu, devid);
1205 iommu_completion_wait(iommu);
1208 void iommu_flush_all_caches(struct amd_iommu *iommu)
1210 if (iommu_feature(iommu, FEATURE_IA)) {
1211 amd_iommu_flush_all(iommu);
1213 amd_iommu_flush_dte_all(iommu);
1214 amd_iommu_flush_irt_all(iommu);
1215 amd_iommu_flush_tlb_all(iommu);
1220 * Command send function for flushing on-device TLB
1222 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1223 u64 address, size_t size)
1225 struct amd_iommu *iommu;
1226 struct iommu_cmd cmd;
1229 qdep = dev_data->ats.qdep;
1230 iommu = amd_iommu_rlookup_table[dev_data->devid];
1232 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1234 return iommu_queue_command(iommu, &cmd);
1238 * Command send function for invalidating a device table entry
1240 static int device_flush_dte(struct iommu_dev_data *dev_data)
1242 struct amd_iommu *iommu;
1246 iommu = amd_iommu_rlookup_table[dev_data->devid];
1247 alias = dev_data->alias;
1249 ret = iommu_flush_dte(iommu, dev_data->devid);
1250 if (!ret && alias != dev_data->devid)
1251 ret = iommu_flush_dte(iommu, alias);
1255 if (dev_data->ats.enabled)
1256 ret = device_flush_iotlb(dev_data, 0, ~0UL);
1262 * TLB invalidation function which is called from the mapping functions.
1263 * It invalidates a single PTE if the range to flush is within a single
1264 * page. Otherwise it flushes the whole TLB of the IOMMU.
1266 static void __domain_flush_pages(struct protection_domain *domain,
1267 u64 address, size_t size, int pde)
1269 struct iommu_dev_data *dev_data;
1270 struct iommu_cmd cmd;
1273 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1275 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
1276 if (!domain->dev_iommu[i])
1280 * Devices of this domain are behind this IOMMU
1281 * We need a TLB flush
1283 ret |= iommu_queue_command(amd_iommus[i], &cmd);
1286 list_for_each_entry(dev_data, &domain->dev_list, list) {
1288 if (!dev_data->ats.enabled)
1291 ret |= device_flush_iotlb(dev_data, address, size);
1297 static void domain_flush_pages(struct protection_domain *domain,
1298 u64 address, size_t size)
1300 __domain_flush_pages(domain, address, size, 0);
1303 /* Flush the whole IO/TLB for a given protection domain */
1304 static void domain_flush_tlb(struct protection_domain *domain)
1306 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
1309 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1310 static void domain_flush_tlb_pde(struct protection_domain *domain)
1312 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1315 static void domain_flush_complete(struct protection_domain *domain)
1319 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
1320 if (domain && !domain->dev_iommu[i])
1324 * Devices of this domain are behind this IOMMU
1325 * We need to wait for completion of all commands.
1327 iommu_completion_wait(amd_iommus[i]);
1331 /* Flush the not present cache if it exists */
1332 static void domain_flush_np_cache(struct protection_domain *domain,
1333 dma_addr_t iova, size_t size)
1335 if (unlikely(amd_iommu_np_cache)) {
1336 domain_flush_pages(domain, iova, size);
1337 domain_flush_complete(domain);
1343 * This function flushes the DTEs for all devices in domain
1345 static void domain_flush_devices(struct protection_domain *domain)
1347 struct iommu_dev_data *dev_data;
1349 list_for_each_entry(dev_data, &domain->dev_list, list)
1350 device_flush_dte(dev_data);
1353 /****************************************************************************
1355 * The functions below are used the create the page table mappings for
1356 * unity mapped regions.
1358 ****************************************************************************/
1360 static void free_page_list(struct page *freelist)
1362 while (freelist != NULL) {
1363 unsigned long p = (unsigned long)page_address(freelist);
1364 freelist = freelist->freelist;
1369 static struct page *free_pt_page(unsigned long pt, struct page *freelist)
1371 struct page *p = virt_to_page((void *)pt);
1373 p->freelist = freelist;
1378 #define DEFINE_FREE_PT_FN(LVL, FN) \
1379 static struct page *free_pt_##LVL (unsigned long __pt, struct page *freelist) \
1387 for (i = 0; i < 512; ++i) { \
1388 /* PTE present? */ \
1389 if (!IOMMU_PTE_PRESENT(pt[i])) \
1393 if (PM_PTE_LEVEL(pt[i]) == 0 || \
1394 PM_PTE_LEVEL(pt[i]) == 7) \
1397 p = (unsigned long)IOMMU_PTE_PAGE(pt[i]); \
1398 freelist = FN(p, freelist); \
1401 return free_pt_page((unsigned long)pt, freelist); \
1404 DEFINE_FREE_PT_FN(l2, free_pt_page)
1405 DEFINE_FREE_PT_FN(l3, free_pt_l2)
1406 DEFINE_FREE_PT_FN(l4, free_pt_l3)
1407 DEFINE_FREE_PT_FN(l5, free_pt_l4)
1408 DEFINE_FREE_PT_FN(l6, free_pt_l5)
1410 static struct page *free_sub_pt(unsigned long root, int mode,
1411 struct page *freelist)
1414 case PAGE_MODE_NONE:
1415 case PAGE_MODE_7_LEVEL:
1417 case PAGE_MODE_1_LEVEL:
1418 freelist = free_pt_page(root, freelist);
1420 case PAGE_MODE_2_LEVEL:
1421 freelist = free_pt_l2(root, freelist);
1423 case PAGE_MODE_3_LEVEL:
1424 freelist = free_pt_l3(root, freelist);
1426 case PAGE_MODE_4_LEVEL:
1427 freelist = free_pt_l4(root, freelist);
1429 case PAGE_MODE_5_LEVEL:
1430 freelist = free_pt_l5(root, freelist);
1432 case PAGE_MODE_6_LEVEL:
1433 freelist = free_pt_l6(root, freelist);
1442 static void free_pagetable(struct protection_domain *domain)
1444 unsigned long root = (unsigned long)domain->pt_root;
1445 struct page *freelist = NULL;
1447 BUG_ON(domain->mode < PAGE_MODE_NONE ||
1448 domain->mode > PAGE_MODE_6_LEVEL);
1450 freelist = free_sub_pt(root, domain->mode, freelist);
1452 free_page_list(freelist);
1456 * This function is used to add another level to an IO page table. Adding
1457 * another level increases the size of the address space by 9 bits to a size up
1460 static bool increase_address_space(struct protection_domain *domain,
1463 unsigned long flags;
1467 spin_lock_irqsave(&domain->lock, flags);
1469 if (WARN_ON_ONCE(domain->mode == PAGE_MODE_6_LEVEL))
1470 /* address space already 64 bit large */
1473 pte = (void *)get_zeroed_page(gfp);
1477 *pte = PM_LEVEL_PDE(domain->mode,
1478 iommu_virt_to_phys(domain->pt_root));
1479 domain->pt_root = pte;
1485 spin_unlock_irqrestore(&domain->lock, flags);
1490 static u64 *alloc_pte(struct protection_domain *domain,
1491 unsigned long address,
1492 unsigned long page_size,
1500 BUG_ON(!is_power_of_2(page_size));
1502 while (address > PM_LEVEL_SIZE(domain->mode))
1503 *updated = increase_address_space(domain, gfp) || *updated;
1505 level = domain->mode - 1;
1506 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1507 address = PAGE_SIZE_ALIGN(address, page_size);
1508 end_lvl = PAGE_SIZE_LEVEL(page_size);
1510 while (level > end_lvl) {
1515 pte_level = PM_PTE_LEVEL(__pte);
1518 * If we replace a series of large PTEs, we need
1519 * to tear down all of them.
1521 if (IOMMU_PTE_PRESENT(__pte) &&
1522 pte_level == PAGE_MODE_7_LEVEL) {
1523 unsigned long count, i;
1526 lpte = first_pte_l7(pte, NULL, &count);
1529 * Unmap the replicated PTEs that still match the
1530 * original large mapping
1532 for (i = 0; i < count; ++i)
1533 cmpxchg64(&lpte[i], __pte, 0ULL);
1539 if (!IOMMU_PTE_PRESENT(__pte) ||
1540 pte_level == PAGE_MODE_NONE) {
1541 page = (u64 *)get_zeroed_page(gfp);
1546 __npte = PM_LEVEL_PDE(level, iommu_virt_to_phys(page));
1548 /* pte could have been changed somewhere. */
1549 if (cmpxchg64(pte, __pte, __npte) != __pte)
1550 free_page((unsigned long)page);
1551 else if (IOMMU_PTE_PRESENT(__pte))
1557 /* No level skipping support yet */
1558 if (pte_level != level)
1563 pte = IOMMU_PTE_PAGE(__pte);
1565 if (pte_page && level == end_lvl)
1568 pte = &pte[PM_LEVEL_INDEX(level, address)];
1575 * This function checks if there is a PTE for a given dma address. If
1576 * there is one, it returns the pointer to it.
1578 static u64 *fetch_pte(struct protection_domain *domain,
1579 unsigned long address,
1580 unsigned long *page_size)
1587 if (address > PM_LEVEL_SIZE(domain->mode))
1590 level = domain->mode - 1;
1591 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1592 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1597 if (!IOMMU_PTE_PRESENT(*pte))
1601 if (PM_PTE_LEVEL(*pte) == 7 ||
1602 PM_PTE_LEVEL(*pte) == 0)
1605 /* No level skipping support yet */
1606 if (PM_PTE_LEVEL(*pte) != level)
1611 /* Walk to the next level */
1612 pte = IOMMU_PTE_PAGE(*pte);
1613 pte = &pte[PM_LEVEL_INDEX(level, address)];
1614 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1618 * If we have a series of large PTEs, make
1619 * sure to return a pointer to the first one.
1621 if (PM_PTE_LEVEL(*pte) == PAGE_MODE_7_LEVEL)
1622 pte = first_pte_l7(pte, page_size, NULL);
1627 static struct page *free_clear_pte(u64 *pte, u64 pteval, struct page *freelist)
1632 while (cmpxchg64(pte, pteval, 0) != pteval) {
1633 pr_warn("AMD-Vi: IOMMU pte changed since we read it\n");
1637 if (!IOMMU_PTE_PRESENT(pteval))
1640 pt = (unsigned long)IOMMU_PTE_PAGE(pteval);
1641 mode = IOMMU_PTE_MODE(pteval);
1643 return free_sub_pt(pt, mode, freelist);
1647 * Generic mapping functions. It maps a physical address into a DMA
1648 * address space. It allocates the page table pages if necessary.
1649 * In the future it can be extended to a generic mapping function
1650 * supporting all features of AMD IOMMU page tables like level skipping
1651 * and full 64 bit address spaces.
1653 static int iommu_map_page(struct protection_domain *dom,
1654 unsigned long bus_addr,
1655 unsigned long phys_addr,
1656 unsigned long page_size,
1660 struct page *freelist = NULL;
1661 bool updated = false;
1665 BUG_ON(!IS_ALIGNED(bus_addr, page_size));
1666 BUG_ON(!IS_ALIGNED(phys_addr, page_size));
1669 if (!(prot & IOMMU_PROT_MASK))
1672 count = PAGE_SIZE_PTE_COUNT(page_size);
1673 pte = alloc_pte(dom, bus_addr, page_size, NULL, gfp, &updated);
1679 for (i = 0; i < count; ++i)
1680 freelist = free_clear_pte(&pte[i], pte[i], freelist);
1682 if (freelist != NULL)
1686 __pte = PAGE_SIZE_PTE(__sme_set(phys_addr), page_size);
1687 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_PR | IOMMU_PTE_FC;
1689 __pte = __sme_set(phys_addr) | IOMMU_PTE_PR | IOMMU_PTE_FC;
1691 if (prot & IOMMU_PROT_IR)
1692 __pte |= IOMMU_PTE_IR;
1693 if (prot & IOMMU_PROT_IW)
1694 __pte |= IOMMU_PTE_IW;
1696 for (i = 0; i < count; ++i)
1705 /* Everything flushed out, free pages now */
1706 free_page_list(freelist);
1711 static unsigned long iommu_unmap_page(struct protection_domain *dom,
1712 unsigned long bus_addr,
1713 unsigned long page_size)
1715 unsigned long long unmapped;
1716 unsigned long unmap_size;
1719 BUG_ON(!is_power_of_2(page_size));
1723 while (unmapped < page_size) {
1725 pte = fetch_pte(dom, bus_addr, &unmap_size);
1730 count = PAGE_SIZE_PTE_COUNT(unmap_size);
1731 for (i = 0; i < count; i++)
1735 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1736 unmapped += unmap_size;
1739 BUG_ON(unmapped && !is_power_of_2(unmapped));
1744 /****************************************************************************
1746 * The next functions belong to the address allocator for the dma_ops
1747 * interface functions.
1749 ****************************************************************************/
1752 static unsigned long dma_ops_alloc_iova(struct device *dev,
1753 struct dma_ops_domain *dma_dom,
1754 unsigned int pages, u64 dma_mask)
1756 unsigned long pfn = 0;
1758 pages = __roundup_pow_of_two(pages);
1760 if (dma_mask > DMA_BIT_MASK(32))
1761 pfn = alloc_iova_fast(&dma_dom->iovad, pages,
1762 IOVA_PFN(DMA_BIT_MASK(32)), false);
1765 pfn = alloc_iova_fast(&dma_dom->iovad, pages,
1766 IOVA_PFN(dma_mask), true);
1768 return (pfn << PAGE_SHIFT);
1771 static void dma_ops_free_iova(struct dma_ops_domain *dma_dom,
1772 unsigned long address,
1775 pages = __roundup_pow_of_two(pages);
1776 address >>= PAGE_SHIFT;
1778 free_iova_fast(&dma_dom->iovad, address, pages);
1781 /****************************************************************************
1783 * The next functions belong to the domain allocation. A domain is
1784 * allocated for every IOMMU as the default domain. If device isolation
1785 * is enabled, every device get its own domain. The most important thing
1786 * about domains is the page table mapping the DMA address space they
1789 ****************************************************************************/
1791 static u16 domain_id_alloc(void)
1795 spin_lock(&pd_bitmap_lock);
1796 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1798 if (id > 0 && id < MAX_DOMAIN_ID)
1799 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1802 spin_unlock(&pd_bitmap_lock);
1807 static void domain_id_free(int id)
1809 spin_lock(&pd_bitmap_lock);
1810 if (id > 0 && id < MAX_DOMAIN_ID)
1811 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1812 spin_unlock(&pd_bitmap_lock);
1815 static void free_gcr3_tbl_level1(u64 *tbl)
1820 for (i = 0; i < 512; ++i) {
1821 if (!(tbl[i] & GCR3_VALID))
1824 ptr = iommu_phys_to_virt(tbl[i] & PAGE_MASK);
1826 free_page((unsigned long)ptr);
1830 static void free_gcr3_tbl_level2(u64 *tbl)
1835 for (i = 0; i < 512; ++i) {
1836 if (!(tbl[i] & GCR3_VALID))
1839 ptr = iommu_phys_to_virt(tbl[i] & PAGE_MASK);
1841 free_gcr3_tbl_level1(ptr);
1845 static void free_gcr3_table(struct protection_domain *domain)
1847 if (domain->glx == 2)
1848 free_gcr3_tbl_level2(domain->gcr3_tbl);
1849 else if (domain->glx == 1)
1850 free_gcr3_tbl_level1(domain->gcr3_tbl);
1852 BUG_ON(domain->glx != 0);
1854 free_page((unsigned long)domain->gcr3_tbl);
1857 static void dma_ops_domain_flush_tlb(struct dma_ops_domain *dom)
1859 domain_flush_tlb(&dom->domain);
1860 domain_flush_complete(&dom->domain);
1863 static void iova_domain_flush_tlb(struct iova_domain *iovad)
1865 struct dma_ops_domain *dom;
1867 dom = container_of(iovad, struct dma_ops_domain, iovad);
1869 dma_ops_domain_flush_tlb(dom);
1873 * Free a domain, only used if something went wrong in the
1874 * allocation path and we need to free an already allocated page table
1876 static void dma_ops_domain_free(struct dma_ops_domain *dom)
1881 put_iova_domain(&dom->iovad);
1883 free_pagetable(&dom->domain);
1886 domain_id_free(dom->domain.id);
1892 * Allocates a new protection domain usable for the dma_ops functions.
1893 * It also initializes the page table and the address allocator data
1894 * structures required for the dma_ops interface
1896 static struct dma_ops_domain *dma_ops_domain_alloc(void)
1898 struct dma_ops_domain *dma_dom;
1900 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
1904 if (protection_domain_init(&dma_dom->domain))
1907 dma_dom->domain.mode = PAGE_MODE_3_LEVEL;
1908 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
1909 dma_dom->domain.flags = PD_DMA_OPS_MASK;
1910 if (!dma_dom->domain.pt_root)
1913 init_iova_domain(&dma_dom->iovad, PAGE_SIZE, IOVA_START_PFN);
1915 if (init_iova_flush_queue(&dma_dom->iovad, iova_domain_flush_tlb, NULL))
1918 /* Initialize reserved ranges */
1919 copy_reserved_iova(&reserved_iova_ranges, &dma_dom->iovad);
1924 dma_ops_domain_free(dma_dom);
1930 * little helper function to check whether a given protection domain is a
1933 static bool dma_ops_domain(struct protection_domain *domain)
1935 return domain->flags & PD_DMA_OPS_MASK;
1938 static void set_dte_entry(u16 devid, struct protection_domain *domain,
1945 if (domain->mode != PAGE_MODE_NONE)
1946 pte_root = iommu_virt_to_phys(domain->pt_root);
1948 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
1949 << DEV_ENTRY_MODE_SHIFT;
1950 pte_root |= DTE_FLAG_IR | DTE_FLAG_IW | DTE_FLAG_V | DTE_FLAG_TV;
1952 flags = amd_iommu_dev_table[devid].data[1];
1955 flags |= DTE_FLAG_IOTLB;
1958 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
1960 if (iommu_feature(iommu, FEATURE_EPHSUP))
1961 pte_root |= 1ULL << DEV_ENTRY_PPR;
1964 if (domain->flags & PD_IOMMUV2_MASK) {
1965 u64 gcr3 = iommu_virt_to_phys(domain->gcr3_tbl);
1966 u64 glx = domain->glx;
1969 pte_root |= DTE_FLAG_GV;
1970 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
1972 /* First mask out possible old values for GCR3 table */
1973 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
1976 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
1979 /* Encode GCR3 table into DTE */
1980 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
1983 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
1986 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
1990 flags &= ~DEV_DOMID_MASK;
1991 flags |= domain->id;
1993 old_domid = amd_iommu_dev_table[devid].data[1] & DEV_DOMID_MASK;
1994 amd_iommu_dev_table[devid].data[1] = flags;
1995 amd_iommu_dev_table[devid].data[0] = pte_root;
1998 * A kdump kernel might be replacing a domain ID that was copied from
1999 * the previous kernel--if so, it needs to flush the translation cache
2000 * entries for the old domain ID that is being overwritten
2003 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
2005 amd_iommu_flush_tlb_domid(iommu, old_domid);
2009 static void clear_dte_entry(u16 devid)
2011 /* remove entry from the device table seen by the hardware */
2012 amd_iommu_dev_table[devid].data[0] = DTE_FLAG_V | DTE_FLAG_TV;
2013 amd_iommu_dev_table[devid].data[1] &= DTE_FLAG_MASK;
2015 amd_iommu_apply_erratum_63(devid);
2018 static void do_attach(struct iommu_dev_data *dev_data,
2019 struct protection_domain *domain)
2021 struct amd_iommu *iommu;
2025 iommu = amd_iommu_rlookup_table[dev_data->devid];
2026 alias = dev_data->alias;
2027 ats = dev_data->ats.enabled;
2029 /* Update data structures */
2030 dev_data->domain = domain;
2031 list_add(&dev_data->list, &domain->dev_list);
2033 /* Do reference counting */
2034 domain->dev_iommu[iommu->index] += 1;
2035 domain->dev_cnt += 1;
2037 /* Update device table */
2038 set_dte_entry(dev_data->devid, domain, ats, dev_data->iommu_v2);
2039 if (alias != dev_data->devid)
2040 set_dte_entry(alias, domain, ats, dev_data->iommu_v2);
2042 device_flush_dte(dev_data);
2045 static void do_detach(struct iommu_dev_data *dev_data)
2047 struct protection_domain *domain = dev_data->domain;
2048 struct amd_iommu *iommu;
2051 iommu = amd_iommu_rlookup_table[dev_data->devid];
2052 alias = dev_data->alias;
2054 /* Update data structures */
2055 dev_data->domain = NULL;
2056 list_del(&dev_data->list);
2057 clear_dte_entry(dev_data->devid);
2058 if (alias != dev_data->devid)
2059 clear_dte_entry(alias);
2061 /* Flush the DTE entry */
2062 device_flush_dte(dev_data);
2065 domain_flush_tlb_pde(domain);
2067 /* Wait for the flushes to finish */
2068 domain_flush_complete(domain);
2070 /* decrease reference counters - needs to happen after the flushes */
2071 domain->dev_iommu[iommu->index] -= 1;
2072 domain->dev_cnt -= 1;
2076 * If a device is not yet associated with a domain, this function makes the
2077 * device visible in the domain
2079 static int __attach_device(struct iommu_dev_data *dev_data,
2080 struct protection_domain *domain)
2082 unsigned long flags;
2086 spin_lock_irqsave(&domain->lock, flags);
2089 if (dev_data->domain != NULL)
2092 /* Attach alias group root */
2093 do_attach(dev_data, domain);
2100 spin_unlock_irqrestore(&domain->lock, flags);
2106 static void pdev_iommuv2_disable(struct pci_dev *pdev)
2108 pci_disable_ats(pdev);
2109 pci_disable_pri(pdev);
2110 pci_disable_pasid(pdev);
2113 /* FIXME: Change generic reset-function to do the same */
2114 static int pri_reset_while_enabled(struct pci_dev *pdev)
2119 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2123 pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
2124 control |= PCI_PRI_CTRL_RESET;
2125 pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
2130 static int pdev_iommuv2_enable(struct pci_dev *pdev)
2135 /* FIXME: Hardcode number of outstanding requests for now */
2137 if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
2139 reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
2141 /* Only allow access to user-accessible pages */
2142 ret = pci_enable_pasid(pdev, 0);
2146 /* First reset the PRI state of the device */
2147 ret = pci_reset_pri(pdev);
2152 ret = pci_enable_pri(pdev, reqs);
2157 ret = pri_reset_while_enabled(pdev);
2162 ret = pci_enable_ats(pdev, PAGE_SHIFT);
2169 pci_disable_pri(pdev);
2170 pci_disable_pasid(pdev);
2176 * If a device is not yet associated with a domain, this function makes the
2177 * device visible in the domain
2179 static int attach_device(struct device *dev,
2180 struct protection_domain *domain)
2182 struct pci_dev *pdev;
2183 struct iommu_dev_data *dev_data;
2186 dev_data = get_dev_data(dev);
2188 if (!dev_is_pci(dev))
2189 goto skip_ats_check;
2191 pdev = to_pci_dev(dev);
2192 if (domain->flags & PD_IOMMUV2_MASK) {
2193 if (!dev_data->passthrough)
2196 if (dev_data->iommu_v2) {
2197 if (pdev_iommuv2_enable(pdev) != 0)
2200 dev_data->ats.enabled = true;
2201 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2202 dev_data->pri_tlp = pci_prg_resp_pasid_required(pdev);
2204 } else if (amd_iommu_iotlb_sup &&
2205 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
2206 dev_data->ats.enabled = true;
2207 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2211 ret = __attach_device(dev_data, domain);
2214 * We might boot into a crash-kernel here. The crashed kernel
2215 * left the caches in the IOMMU dirty. So we have to flush
2216 * here to evict all dirty stuff.
2218 domain_flush_tlb_pde(domain);
2220 domain_flush_complete(domain);
2226 * Removes a device from a protection domain (unlocked)
2228 static void __detach_device(struct iommu_dev_data *dev_data)
2230 struct protection_domain *domain;
2231 unsigned long flags;
2233 domain = dev_data->domain;
2235 spin_lock_irqsave(&domain->lock, flags);
2237 do_detach(dev_data);
2239 spin_unlock_irqrestore(&domain->lock, flags);
2243 * Removes a device from a protection domain (with devtable_lock held)
2245 static void detach_device(struct device *dev)
2247 struct protection_domain *domain;
2248 struct iommu_dev_data *dev_data;
2250 dev_data = get_dev_data(dev);
2251 domain = dev_data->domain;
2254 * First check if the device is still attached. It might already
2255 * be detached from its domain because the generic
2256 * iommu_detach_group code detached it and we try again here in
2257 * our alias handling.
2259 if (WARN_ON(!dev_data->domain))
2262 __detach_device(dev_data);
2264 if (!dev_is_pci(dev))
2267 if (domain->flags & PD_IOMMUV2_MASK && dev_data->iommu_v2)
2268 pdev_iommuv2_disable(to_pci_dev(dev));
2269 else if (dev_data->ats.enabled)
2270 pci_disable_ats(to_pci_dev(dev));
2272 dev_data->ats.enabled = false;
2275 static int amd_iommu_add_device(struct device *dev)
2277 struct iommu_dev_data *dev_data;
2278 struct iommu_domain *domain;
2279 struct amd_iommu *iommu;
2282 if (!check_device(dev) || get_dev_data(dev))
2285 devid = get_device_id(dev);
2289 iommu = amd_iommu_rlookup_table[devid];
2291 ret = iommu_init_device(dev);
2293 if (ret != -ENOTSUPP)
2294 dev_err(dev, "Failed to initialize - trying to proceed anyway\n");
2296 iommu_ignore_device(dev);
2297 dev->dma_ops = NULL;
2300 init_iommu_group(dev);
2302 dev_data = get_dev_data(dev);
2306 if (dev_data->iommu_v2)
2307 iommu_request_dm_for_dev(dev);
2309 /* Domains are initialized for this device - have a look what we ended up with */
2310 domain = iommu_get_domain_for_dev(dev);
2311 if (domain->type == IOMMU_DOMAIN_IDENTITY)
2312 dev_data->passthrough = true;
2314 dev->dma_ops = &amd_iommu_dma_ops;
2317 iommu_completion_wait(iommu);
2322 static void amd_iommu_remove_device(struct device *dev)
2324 struct amd_iommu *iommu;
2327 if (!check_device(dev))
2330 devid = get_device_id(dev);
2334 iommu = amd_iommu_rlookup_table[devid];
2336 iommu_uninit_device(dev);
2337 iommu_completion_wait(iommu);
2340 static struct iommu_group *amd_iommu_device_group(struct device *dev)
2342 if (dev_is_pci(dev))
2343 return pci_device_group(dev);
2345 return acpihid_device_group(dev);
2348 /*****************************************************************************
2350 * The next functions belong to the dma_ops mapping/unmapping code.
2352 *****************************************************************************/
2355 * In the dma_ops path we only have the struct device. This function
2356 * finds the corresponding IOMMU, the protection domain and the
2357 * requestor id for a given device.
2358 * If the device is not yet associated with a domain this is also done
2361 static struct protection_domain *get_domain(struct device *dev)
2363 struct protection_domain *domain;
2364 struct iommu_domain *io_domain;
2366 if (!check_device(dev))
2367 return ERR_PTR(-EINVAL);
2369 domain = get_dev_data(dev)->domain;
2370 if (domain == NULL && get_dev_data(dev)->defer_attach) {
2371 get_dev_data(dev)->defer_attach = false;
2372 io_domain = iommu_get_domain_for_dev(dev);
2373 domain = to_pdomain(io_domain);
2374 attach_device(dev, domain);
2377 return ERR_PTR(-EBUSY);
2379 if (!dma_ops_domain(domain))
2380 return ERR_PTR(-EBUSY);
2385 static void update_device_table(struct protection_domain *domain)
2387 struct iommu_dev_data *dev_data;
2389 list_for_each_entry(dev_data, &domain->dev_list, list) {
2390 set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled,
2391 dev_data->iommu_v2);
2393 if (dev_data->devid == dev_data->alias)
2396 /* There is an alias, update device table entry for it */
2397 set_dte_entry(dev_data->alias, domain, dev_data->ats.enabled,
2398 dev_data->iommu_v2);
2402 static void update_domain(struct protection_domain *domain)
2404 update_device_table(domain);
2406 domain_flush_devices(domain);
2407 domain_flush_tlb_pde(domain);
2410 static int dir2prot(enum dma_data_direction direction)
2412 if (direction == DMA_TO_DEVICE)
2413 return IOMMU_PROT_IR;
2414 else if (direction == DMA_FROM_DEVICE)
2415 return IOMMU_PROT_IW;
2416 else if (direction == DMA_BIDIRECTIONAL)
2417 return IOMMU_PROT_IW | IOMMU_PROT_IR;
2423 * This function contains common code for mapping of a physically
2424 * contiguous memory region into DMA address space. It is used by all
2425 * mapping functions provided with this IOMMU driver.
2426 * Must be called with the domain lock held.
2428 static dma_addr_t __map_single(struct device *dev,
2429 struct dma_ops_domain *dma_dom,
2432 enum dma_data_direction direction,
2435 dma_addr_t offset = paddr & ~PAGE_MASK;
2436 dma_addr_t address, start, ret;
2441 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
2444 address = dma_ops_alloc_iova(dev, dma_dom, pages, dma_mask);
2448 prot = dir2prot(direction);
2451 for (i = 0; i < pages; ++i) {
2452 ret = iommu_map_page(&dma_dom->domain, start, paddr,
2453 PAGE_SIZE, prot, GFP_ATOMIC);
2462 domain_flush_np_cache(&dma_dom->domain, address, size);
2469 for (--i; i >= 0; --i) {
2471 iommu_unmap_page(&dma_dom->domain, start, PAGE_SIZE);
2474 domain_flush_tlb(&dma_dom->domain);
2475 domain_flush_complete(&dma_dom->domain);
2477 dma_ops_free_iova(dma_dom, address, pages);
2479 return DMA_MAPPING_ERROR;
2483 * Does the reverse of the __map_single function. Must be called with
2484 * the domain lock held too
2486 static void __unmap_single(struct dma_ops_domain *dma_dom,
2487 dma_addr_t dma_addr,
2491 dma_addr_t i, start;
2494 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
2495 dma_addr &= PAGE_MASK;
2498 for (i = 0; i < pages; ++i) {
2499 iommu_unmap_page(&dma_dom->domain, start, PAGE_SIZE);
2503 if (amd_iommu_unmap_flush) {
2504 domain_flush_tlb(&dma_dom->domain);
2505 domain_flush_complete(&dma_dom->domain);
2506 dma_ops_free_iova(dma_dom, dma_addr, pages);
2508 pages = __roundup_pow_of_two(pages);
2509 queue_iova(&dma_dom->iovad, dma_addr >> PAGE_SHIFT, pages, 0);
2514 * The exported map_single function for dma_ops.
2516 static dma_addr_t map_page(struct device *dev, struct page *page,
2517 unsigned long offset, size_t size,
2518 enum dma_data_direction dir,
2519 unsigned long attrs)
2521 phys_addr_t paddr = page_to_phys(page) + offset;
2522 struct protection_domain *domain;
2523 struct dma_ops_domain *dma_dom;
2526 domain = get_domain(dev);
2527 if (PTR_ERR(domain) == -EINVAL)
2528 return (dma_addr_t)paddr;
2529 else if (IS_ERR(domain))
2530 return DMA_MAPPING_ERROR;
2532 dma_mask = *dev->dma_mask;
2533 dma_dom = to_dma_ops_domain(domain);
2535 return __map_single(dev, dma_dom, paddr, size, dir, dma_mask);
2539 * The exported unmap_single function for dma_ops.
2541 static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2542 enum dma_data_direction dir, unsigned long attrs)
2544 struct protection_domain *domain;
2545 struct dma_ops_domain *dma_dom;
2547 domain = get_domain(dev);
2551 dma_dom = to_dma_ops_domain(domain);
2553 __unmap_single(dma_dom, dma_addr, size, dir);
2556 static int sg_num_pages(struct device *dev,
2557 struct scatterlist *sglist,
2560 unsigned long mask, boundary_size;
2561 struct scatterlist *s;
2564 mask = dma_get_seg_boundary(dev);
2565 boundary_size = mask + 1 ? ALIGN(mask + 1, PAGE_SIZE) >> PAGE_SHIFT :
2566 1UL << (BITS_PER_LONG - PAGE_SHIFT);
2568 for_each_sg(sglist, s, nelems, i) {
2571 s->dma_address = npages << PAGE_SHIFT;
2572 p = npages % boundary_size;
2573 n = iommu_num_pages(sg_phys(s), s->length, PAGE_SIZE);
2574 if (p + n > boundary_size)
2575 npages += boundary_size - p;
2583 * The exported map_sg function for dma_ops (handles scatter-gather
2586 static int map_sg(struct device *dev, struct scatterlist *sglist,
2587 int nelems, enum dma_data_direction direction,
2588 unsigned long attrs)
2590 int mapped_pages = 0, npages = 0, prot = 0, i;
2591 struct protection_domain *domain;
2592 struct dma_ops_domain *dma_dom;
2593 struct scatterlist *s;
2594 unsigned long address;
2598 domain = get_domain(dev);
2602 dma_dom = to_dma_ops_domain(domain);
2603 dma_mask = *dev->dma_mask;
2605 npages = sg_num_pages(dev, sglist, nelems);
2607 address = dma_ops_alloc_iova(dev, dma_dom, npages, dma_mask);
2611 prot = dir2prot(direction);
2613 /* Map all sg entries */
2614 for_each_sg(sglist, s, nelems, i) {
2615 int j, pages = iommu_num_pages(sg_phys(s), s->length, PAGE_SIZE);
2617 for (j = 0; j < pages; ++j) {
2618 unsigned long bus_addr, phys_addr;
2620 bus_addr = address + s->dma_address + (j << PAGE_SHIFT);
2621 phys_addr = (sg_phys(s) & PAGE_MASK) + (j << PAGE_SHIFT);
2622 ret = iommu_map_page(domain, bus_addr, phys_addr,
2624 GFP_ATOMIC | __GFP_NOWARN);
2632 /* Everything is mapped - write the right values into s->dma_address */
2633 for_each_sg(sglist, s, nelems, i) {
2635 * Add in the remaining piece of the scatter-gather offset that
2636 * was masked out when we were determining the physical address
2637 * via (sg_phys(s) & PAGE_MASK) earlier.
2639 s->dma_address += address + (s->offset & ~PAGE_MASK);
2640 s->dma_length = s->length;
2644 domain_flush_np_cache(domain, s->dma_address, s->dma_length);
2649 dev_err(dev, "IOMMU mapping error in map_sg (io-pages: %d reason: %d)\n",
2652 for_each_sg(sglist, s, nelems, i) {
2653 int j, pages = iommu_num_pages(sg_phys(s), s->length, PAGE_SIZE);
2655 for (j = 0; j < pages; ++j) {
2656 unsigned long bus_addr;
2658 bus_addr = address + s->dma_address + (j << PAGE_SHIFT);
2659 iommu_unmap_page(domain, bus_addr, PAGE_SIZE);
2661 if (--mapped_pages == 0)
2667 free_iova_fast(&dma_dom->iovad, address >> PAGE_SHIFT, npages);
2674 * The exported map_sg function for dma_ops (handles scatter-gather
2677 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2678 int nelems, enum dma_data_direction dir,
2679 unsigned long attrs)
2681 struct protection_domain *domain;
2682 struct dma_ops_domain *dma_dom;
2683 unsigned long startaddr;
2686 domain = get_domain(dev);
2690 startaddr = sg_dma_address(sglist) & PAGE_MASK;
2691 dma_dom = to_dma_ops_domain(domain);
2692 npages = sg_num_pages(dev, sglist, nelems);
2694 __unmap_single(dma_dom, startaddr, npages << PAGE_SHIFT, dir);
2698 * The exported alloc_coherent function for dma_ops.
2700 static void *alloc_coherent(struct device *dev, size_t size,
2701 dma_addr_t *dma_addr, gfp_t flag,
2702 unsigned long attrs)
2704 u64 dma_mask = dev->coherent_dma_mask;
2705 struct protection_domain *domain;
2706 struct dma_ops_domain *dma_dom;
2709 domain = get_domain(dev);
2710 if (PTR_ERR(domain) == -EINVAL) {
2711 page = alloc_pages(flag, get_order(size));
2712 *dma_addr = page_to_phys(page);
2713 return page_address(page);
2714 } else if (IS_ERR(domain))
2717 dma_dom = to_dma_ops_domain(domain);
2718 size = PAGE_ALIGN(size);
2719 dma_mask = dev->coherent_dma_mask;
2720 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
2723 page = alloc_pages(flag | __GFP_NOWARN, get_order(size));
2725 if (!gfpflags_allow_blocking(flag))
2728 page = dma_alloc_from_contiguous(dev, size >> PAGE_SHIFT,
2729 get_order(size), flag & __GFP_NOWARN);
2735 dma_mask = *dev->dma_mask;
2737 *dma_addr = __map_single(dev, dma_dom, page_to_phys(page),
2738 size, DMA_BIDIRECTIONAL, dma_mask);
2740 if (*dma_addr == DMA_MAPPING_ERROR)
2743 return page_address(page);
2747 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2748 __free_pages(page, get_order(size));
2754 * The exported free_coherent function for dma_ops.
2756 static void free_coherent(struct device *dev, size_t size,
2757 void *virt_addr, dma_addr_t dma_addr,
2758 unsigned long attrs)
2760 struct protection_domain *domain;
2761 struct dma_ops_domain *dma_dom;
2764 page = virt_to_page(virt_addr);
2765 size = PAGE_ALIGN(size);
2767 domain = get_domain(dev);
2771 dma_dom = to_dma_ops_domain(domain);
2773 __unmap_single(dma_dom, dma_addr, size, DMA_BIDIRECTIONAL);
2776 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2777 __free_pages(page, get_order(size));
2781 * This function is called by the DMA layer to find out if we can handle a
2782 * particular device. It is part of the dma_ops.
2784 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
2786 if (!dma_direct_supported(dev, mask))
2788 return check_device(dev);
2791 static const struct dma_map_ops amd_iommu_dma_ops = {
2792 .alloc = alloc_coherent,
2793 .free = free_coherent,
2794 .map_page = map_page,
2795 .unmap_page = unmap_page,
2797 .unmap_sg = unmap_sg,
2798 .dma_supported = amd_iommu_dma_supported,
2801 static int init_reserved_iova_ranges(void)
2803 struct pci_dev *pdev = NULL;
2806 init_iova_domain(&reserved_iova_ranges, PAGE_SIZE, IOVA_START_PFN);
2808 lockdep_set_class(&reserved_iova_ranges.iova_rbtree_lock,
2809 &reserved_rbtree_key);
2811 /* MSI memory range */
2812 val = reserve_iova(&reserved_iova_ranges,
2813 IOVA_PFN(MSI_RANGE_START), IOVA_PFN(MSI_RANGE_END));
2815 pr_err("Reserving MSI range failed\n");
2819 /* HT memory range */
2820 val = reserve_iova(&reserved_iova_ranges,
2821 IOVA_PFN(HT_RANGE_START), IOVA_PFN(HT_RANGE_END));
2823 pr_err("Reserving HT range failed\n");
2828 * Memory used for PCI resources
2829 * FIXME: Check whether we can reserve the PCI-hole completly
2831 for_each_pci_dev(pdev) {
2834 for (i = 0; i < PCI_NUM_RESOURCES; ++i) {
2835 struct resource *r = &pdev->resource[i];
2837 if (!(r->flags & IORESOURCE_MEM))
2840 val = reserve_iova(&reserved_iova_ranges,
2844 pci_err(pdev, "Reserve pci-resource range %pR failed\n", r);
2853 int __init amd_iommu_init_api(void)
2857 ret = iova_cache_get();
2861 ret = init_reserved_iova_ranges();
2865 err = bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
2868 #ifdef CONFIG_ARM_AMBA
2869 err = bus_set_iommu(&amba_bustype, &amd_iommu_ops);
2873 err = bus_set_iommu(&platform_bus_type, &amd_iommu_ops);
2880 int __init amd_iommu_init_dma_ops(void)
2882 swiotlb = (iommu_default_passthrough() || sme_me_mask) ? 1 : 0;
2885 if (amd_iommu_unmap_flush)
2886 pr_info("IO/TLB flush on unmap enabled\n");
2888 pr_info("Lazy IO/TLB flushing enabled\n");
2894 /*****************************************************************************
2896 * The following functions belong to the exported interface of AMD IOMMU
2898 * This interface allows access to lower level functions of the IOMMU
2899 * like protection domain handling and assignement of devices to domains
2900 * which is not possible with the dma_ops interface.
2902 *****************************************************************************/
2904 static void cleanup_domain(struct protection_domain *domain)
2906 struct iommu_dev_data *entry;
2908 while (!list_empty(&domain->dev_list)) {
2909 entry = list_first_entry(&domain->dev_list,
2910 struct iommu_dev_data, list);
2911 BUG_ON(!entry->domain);
2912 __detach_device(entry);
2916 static void protection_domain_free(struct protection_domain *domain)
2922 domain_id_free(domain->id);
2927 static int protection_domain_init(struct protection_domain *domain)
2929 spin_lock_init(&domain->lock);
2930 mutex_init(&domain->api_lock);
2931 domain->id = domain_id_alloc();
2934 INIT_LIST_HEAD(&domain->dev_list);
2939 static struct protection_domain *protection_domain_alloc(void)
2941 struct protection_domain *domain;
2943 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2947 if (protection_domain_init(domain))
2958 static struct iommu_domain *amd_iommu_domain_alloc(unsigned type)
2960 struct protection_domain *pdomain;
2961 struct dma_ops_domain *dma_domain;
2964 case IOMMU_DOMAIN_UNMANAGED:
2965 pdomain = protection_domain_alloc();
2969 pdomain->mode = PAGE_MODE_3_LEVEL;
2970 pdomain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2971 if (!pdomain->pt_root) {
2972 protection_domain_free(pdomain);
2976 pdomain->domain.geometry.aperture_start = 0;
2977 pdomain->domain.geometry.aperture_end = ~0ULL;
2978 pdomain->domain.geometry.force_aperture = true;
2981 case IOMMU_DOMAIN_DMA:
2982 dma_domain = dma_ops_domain_alloc();
2984 pr_err("Failed to allocate\n");
2987 pdomain = &dma_domain->domain;
2989 case IOMMU_DOMAIN_IDENTITY:
2990 pdomain = protection_domain_alloc();
2994 pdomain->mode = PAGE_MODE_NONE;
3000 return &pdomain->domain;
3003 static void amd_iommu_domain_free(struct iommu_domain *dom)
3005 struct protection_domain *domain;
3006 struct dma_ops_domain *dma_dom;
3008 domain = to_pdomain(dom);
3010 if (domain->dev_cnt > 0)
3011 cleanup_domain(domain);
3013 BUG_ON(domain->dev_cnt != 0);
3018 switch (dom->type) {
3019 case IOMMU_DOMAIN_DMA:
3020 /* Now release the domain */
3021 dma_dom = to_dma_ops_domain(domain);
3022 dma_ops_domain_free(dma_dom);
3025 if (domain->mode != PAGE_MODE_NONE)
3026 free_pagetable(domain);
3028 if (domain->flags & PD_IOMMUV2_MASK)
3029 free_gcr3_table(domain);
3031 protection_domain_free(domain);
3036 static void amd_iommu_detach_device(struct iommu_domain *dom,
3039 struct iommu_dev_data *dev_data = dev->archdata.iommu;
3040 struct amd_iommu *iommu;
3043 if (!check_device(dev))
3046 devid = get_device_id(dev);
3050 if (dev_data->domain != NULL)
3053 iommu = amd_iommu_rlookup_table[devid];
3057 #ifdef CONFIG_IRQ_REMAP
3058 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) &&
3059 (dom->type == IOMMU_DOMAIN_UNMANAGED))
3060 dev_data->use_vapic = 0;
3063 iommu_completion_wait(iommu);
3066 static int amd_iommu_attach_device(struct iommu_domain *dom,
3069 struct protection_domain *domain = to_pdomain(dom);
3070 struct iommu_dev_data *dev_data;
3071 struct amd_iommu *iommu;
3074 if (!check_device(dev))
3077 dev_data = dev->archdata.iommu;
3079 iommu = amd_iommu_rlookup_table[dev_data->devid];
3083 if (dev_data->domain)
3086 ret = attach_device(dev, domain);
3088 #ifdef CONFIG_IRQ_REMAP
3089 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
3090 if (dom->type == IOMMU_DOMAIN_UNMANAGED)
3091 dev_data->use_vapic = 1;
3093 dev_data->use_vapic = 0;
3097 iommu_completion_wait(iommu);
3102 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
3103 phys_addr_t paddr, size_t page_size, int iommu_prot)
3105 struct protection_domain *domain = to_pdomain(dom);
3109 if (domain->mode == PAGE_MODE_NONE)
3112 if (iommu_prot & IOMMU_READ)
3113 prot |= IOMMU_PROT_IR;
3114 if (iommu_prot & IOMMU_WRITE)
3115 prot |= IOMMU_PROT_IW;
3117 mutex_lock(&domain->api_lock);
3118 ret = iommu_map_page(domain, iova, paddr, page_size, prot, GFP_KERNEL);
3119 mutex_unlock(&domain->api_lock);
3121 domain_flush_np_cache(domain, iova, page_size);
3126 static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
3128 struct iommu_iotlb_gather *gather)
3130 struct protection_domain *domain = to_pdomain(dom);
3133 if (domain->mode == PAGE_MODE_NONE)
3136 mutex_lock(&domain->api_lock);
3137 unmap_size = iommu_unmap_page(domain, iova, page_size);
3138 mutex_unlock(&domain->api_lock);
3143 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
3146 struct protection_domain *domain = to_pdomain(dom);
3147 unsigned long offset_mask, pte_pgsize;
3150 if (domain->mode == PAGE_MODE_NONE)
3153 pte = fetch_pte(domain, iova, &pte_pgsize);
3155 if (!pte || !IOMMU_PTE_PRESENT(*pte))
3158 offset_mask = pte_pgsize - 1;
3159 __pte = __sme_clr(*pte & PM_ADDR_MASK);
3161 return (__pte & ~offset_mask) | (iova & offset_mask);
3164 static bool amd_iommu_capable(enum iommu_cap cap)
3167 case IOMMU_CAP_CACHE_COHERENCY:
3169 case IOMMU_CAP_INTR_REMAP:
3170 return (irq_remapping_enabled == 1);
3171 case IOMMU_CAP_NOEXEC:
3180 static void amd_iommu_get_resv_regions(struct device *dev,
3181 struct list_head *head)
3183 struct iommu_resv_region *region;
3184 struct unity_map_entry *entry;
3187 devid = get_device_id(dev);
3191 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
3195 if (devid < entry->devid_start || devid > entry->devid_end)
3198 type = IOMMU_RESV_DIRECT;
3199 length = entry->address_end - entry->address_start;
3200 if (entry->prot & IOMMU_PROT_IR)
3202 if (entry->prot & IOMMU_PROT_IW)
3203 prot |= IOMMU_WRITE;
3204 if (entry->prot & IOMMU_UNITY_MAP_FLAG_EXCL_RANGE)
3205 /* Exclusion range */
3206 type = IOMMU_RESV_RESERVED;
3208 region = iommu_alloc_resv_region(entry->address_start,
3209 length, prot, type);
3211 dev_err(dev, "Out of memory allocating dm-regions\n");
3214 list_add_tail(®ion->list, head);
3217 region = iommu_alloc_resv_region(MSI_RANGE_START,
3218 MSI_RANGE_END - MSI_RANGE_START + 1,
3222 list_add_tail(®ion->list, head);
3224 region = iommu_alloc_resv_region(HT_RANGE_START,
3225 HT_RANGE_END - HT_RANGE_START + 1,
3226 0, IOMMU_RESV_RESERVED);
3229 list_add_tail(®ion->list, head);
3232 static void amd_iommu_put_resv_regions(struct device *dev,
3233 struct list_head *head)
3235 struct iommu_resv_region *entry, *next;
3237 list_for_each_entry_safe(entry, next, head, list)
3241 static void amd_iommu_apply_resv_region(struct device *dev,
3242 struct iommu_domain *domain,
3243 struct iommu_resv_region *region)
3245 struct dma_ops_domain *dma_dom = to_dma_ops_domain(to_pdomain(domain));
3246 unsigned long start, end;
3248 start = IOVA_PFN(region->start);
3249 end = IOVA_PFN(region->start + region->length - 1);
3251 WARN_ON_ONCE(reserve_iova(&dma_dom->iovad, start, end) == NULL);
3254 static bool amd_iommu_is_attach_deferred(struct iommu_domain *domain,
3257 struct iommu_dev_data *dev_data = dev->archdata.iommu;
3258 return dev_data->defer_attach;
3261 static void amd_iommu_flush_iotlb_all(struct iommu_domain *domain)
3263 struct protection_domain *dom = to_pdomain(domain);
3265 domain_flush_tlb_pde(dom);
3266 domain_flush_complete(dom);
3269 static void amd_iommu_iotlb_sync(struct iommu_domain *domain,
3270 struct iommu_iotlb_gather *gather)
3272 amd_iommu_flush_iotlb_all(domain);
3275 const struct iommu_ops amd_iommu_ops = {
3276 .capable = amd_iommu_capable,
3277 .domain_alloc = amd_iommu_domain_alloc,
3278 .domain_free = amd_iommu_domain_free,
3279 .attach_dev = amd_iommu_attach_device,
3280 .detach_dev = amd_iommu_detach_device,
3281 .map = amd_iommu_map,
3282 .unmap = amd_iommu_unmap,
3283 .iova_to_phys = amd_iommu_iova_to_phys,
3284 .add_device = amd_iommu_add_device,
3285 .remove_device = amd_iommu_remove_device,
3286 .device_group = amd_iommu_device_group,
3287 .get_resv_regions = amd_iommu_get_resv_regions,
3288 .put_resv_regions = amd_iommu_put_resv_regions,
3289 .apply_resv_region = amd_iommu_apply_resv_region,
3290 .is_attach_deferred = amd_iommu_is_attach_deferred,
3291 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
3292 .flush_iotlb_all = amd_iommu_flush_iotlb_all,
3293 .iotlb_sync = amd_iommu_iotlb_sync,
3296 /*****************************************************************************
3298 * The next functions do a basic initialization of IOMMU for pass through
3301 * In passthrough mode the IOMMU is initialized and enabled but not used for
3302 * DMA-API translation.
3304 *****************************************************************************/
3306 /* IOMMUv2 specific functions */
3307 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
3309 return atomic_notifier_chain_register(&ppr_notifier, nb);
3311 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
3313 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
3315 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
3317 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
3319 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
3321 struct protection_domain *domain = to_pdomain(dom);
3322 unsigned long flags;
3324 spin_lock_irqsave(&domain->lock, flags);
3326 /* Update data structure */
3327 domain->mode = PAGE_MODE_NONE;
3329 /* Make changes visible to IOMMUs */
3330 update_domain(domain);
3332 /* Page-table is not visible to IOMMU anymore, so free it */
3333 free_pagetable(domain);
3335 spin_unlock_irqrestore(&domain->lock, flags);
3337 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
3339 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
3341 struct protection_domain *domain = to_pdomain(dom);
3342 unsigned long flags;
3345 if (pasids <= 0 || pasids > (PASID_MASK + 1))
3348 /* Number of GCR3 table levels required */
3349 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
3352 if (levels > amd_iommu_max_glx_val)
3355 spin_lock_irqsave(&domain->lock, flags);
3358 * Save us all sanity checks whether devices already in the
3359 * domain support IOMMUv2. Just force that the domain has no
3360 * devices attached when it is switched into IOMMUv2 mode.
3363 if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
3367 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
3368 if (domain->gcr3_tbl == NULL)
3371 domain->glx = levels;
3372 domain->flags |= PD_IOMMUV2_MASK;
3374 update_domain(domain);
3379 spin_unlock_irqrestore(&domain->lock, flags);
3383 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
3385 static int __flush_pasid(struct protection_domain *domain, int pasid,
3386 u64 address, bool size)
3388 struct iommu_dev_data *dev_data;
3389 struct iommu_cmd cmd;
3392 if (!(domain->flags & PD_IOMMUV2_MASK))
3395 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
3398 * IOMMU TLB needs to be flushed before Device TLB to
3399 * prevent device TLB refill from IOMMU TLB
3401 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
3402 if (domain->dev_iommu[i] == 0)
3405 ret = iommu_queue_command(amd_iommus[i], &cmd);
3410 /* Wait until IOMMU TLB flushes are complete */
3411 domain_flush_complete(domain);
3413 /* Now flush device TLBs */
3414 list_for_each_entry(dev_data, &domain->dev_list, list) {
3415 struct amd_iommu *iommu;
3419 There might be non-IOMMUv2 capable devices in an IOMMUv2
3422 if (!dev_data->ats.enabled)
3425 qdep = dev_data->ats.qdep;
3426 iommu = amd_iommu_rlookup_table[dev_data->devid];
3428 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
3429 qdep, address, size);
3431 ret = iommu_queue_command(iommu, &cmd);
3436 /* Wait until all device TLBs are flushed */
3437 domain_flush_complete(domain);
3446 static int __amd_iommu_flush_page(struct protection_domain *domain, int pasid,
3449 return __flush_pasid(domain, pasid, address, false);
3452 int amd_iommu_flush_page(struct iommu_domain *dom, int pasid,
3455 struct protection_domain *domain = to_pdomain(dom);
3456 unsigned long flags;
3459 spin_lock_irqsave(&domain->lock, flags);
3460 ret = __amd_iommu_flush_page(domain, pasid, address);
3461 spin_unlock_irqrestore(&domain->lock, flags);
3465 EXPORT_SYMBOL(amd_iommu_flush_page);
3467 static int __amd_iommu_flush_tlb(struct protection_domain *domain, int pasid)
3469 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
3473 int amd_iommu_flush_tlb(struct iommu_domain *dom, int pasid)
3475 struct protection_domain *domain = to_pdomain(dom);
3476 unsigned long flags;
3479 spin_lock_irqsave(&domain->lock, flags);
3480 ret = __amd_iommu_flush_tlb(domain, pasid);
3481 spin_unlock_irqrestore(&domain->lock, flags);
3485 EXPORT_SYMBOL(amd_iommu_flush_tlb);
3487 static u64 *__get_gcr3_pte(u64 *root, int level, int pasid, bool alloc)
3494 index = (pasid >> (9 * level)) & 0x1ff;
3500 if (!(*pte & GCR3_VALID)) {
3504 root = (void *)get_zeroed_page(GFP_ATOMIC);
3508 *pte = iommu_virt_to_phys(root) | GCR3_VALID;
3511 root = iommu_phys_to_virt(*pte & PAGE_MASK);
3519 static int __set_gcr3(struct protection_domain *domain, int pasid,
3524 if (domain->mode != PAGE_MODE_NONE)
3527 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3531 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3533 return __amd_iommu_flush_tlb(domain, pasid);
3536 static int __clear_gcr3(struct protection_domain *domain, int pasid)
3540 if (domain->mode != PAGE_MODE_NONE)
3543 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3549 return __amd_iommu_flush_tlb(domain, pasid);
3552 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, int pasid,
3555 struct protection_domain *domain = to_pdomain(dom);
3556 unsigned long flags;
3559 spin_lock_irqsave(&domain->lock, flags);
3560 ret = __set_gcr3(domain, pasid, cr3);
3561 spin_unlock_irqrestore(&domain->lock, flags);
3565 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3567 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, int pasid)
3569 struct protection_domain *domain = to_pdomain(dom);
3570 unsigned long flags;
3573 spin_lock_irqsave(&domain->lock, flags);
3574 ret = __clear_gcr3(domain, pasid);
3575 spin_unlock_irqrestore(&domain->lock, flags);
3579 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3581 int amd_iommu_complete_ppr(struct pci_dev *pdev, int pasid,
3582 int status, int tag)
3584 struct iommu_dev_data *dev_data;
3585 struct amd_iommu *iommu;
3586 struct iommu_cmd cmd;
3588 dev_data = get_dev_data(&pdev->dev);
3589 iommu = amd_iommu_rlookup_table[dev_data->devid];
3591 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3592 tag, dev_data->pri_tlp);
3594 return iommu_queue_command(iommu, &cmd);
3596 EXPORT_SYMBOL(amd_iommu_complete_ppr);
3598 struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3600 struct protection_domain *pdomain;
3602 pdomain = get_domain(&pdev->dev);
3603 if (IS_ERR(pdomain))
3606 /* Only return IOMMUv2 domains */
3607 if (!(pdomain->flags & PD_IOMMUV2_MASK))
3610 return &pdomain->domain;
3612 EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3614 void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3616 struct iommu_dev_data *dev_data;
3618 if (!amd_iommu_v2_supported())
3621 dev_data = get_dev_data(&pdev->dev);
3622 dev_data->errata |= (1 << erratum);
3624 EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3626 int amd_iommu_device_info(struct pci_dev *pdev,
3627 struct amd_iommu_device_info *info)
3632 if (pdev == NULL || info == NULL)
3635 if (!amd_iommu_v2_supported())
3638 memset(info, 0, sizeof(*info));
3640 if (!pci_ats_disabled()) {
3641 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS);
3643 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3646 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3648 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3650 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3654 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3655 max_pasids = min(max_pasids, (1 << 20));
3657 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3658 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3660 features = pci_pasid_features(pdev);
3661 if (features & PCI_PASID_CAP_EXEC)
3662 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3663 if (features & PCI_PASID_CAP_PRIV)
3664 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3669 EXPORT_SYMBOL(amd_iommu_device_info);
3671 #ifdef CONFIG_IRQ_REMAP
3673 /*****************************************************************************
3675 * Interrupt Remapping Implementation
3677 *****************************************************************************/
3679 static struct irq_chip amd_ir_chip;
3680 static DEFINE_SPINLOCK(iommu_table_lock);
3682 static void set_dte_irq_entry(u16 devid, struct irq_remap_table *table)
3686 dte = amd_iommu_dev_table[devid].data[2];
3687 dte &= ~DTE_IRQ_PHYS_ADDR_MASK;
3688 dte |= iommu_virt_to_phys(table->table);
3689 dte |= DTE_IRQ_REMAP_INTCTL;
3690 dte |= DTE_IRQ_TABLE_LEN;
3691 dte |= DTE_IRQ_REMAP_ENABLE;
3693 amd_iommu_dev_table[devid].data[2] = dte;
3696 static struct irq_remap_table *get_irq_table(u16 devid)
3698 struct irq_remap_table *table;
3700 if (WARN_ONCE(!amd_iommu_rlookup_table[devid],
3701 "%s: no iommu for devid %x\n", __func__, devid))
3704 table = irq_lookup_table[devid];
3705 if (WARN_ONCE(!table, "%s: no table for devid %x\n", __func__, devid))
3711 static struct irq_remap_table *__alloc_irq_table(void)
3713 struct irq_remap_table *table;
3715 table = kzalloc(sizeof(*table), GFP_KERNEL);
3719 table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_KERNEL);
3720 if (!table->table) {
3724 raw_spin_lock_init(&table->lock);
3726 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
3727 memset(table->table, 0,
3728 MAX_IRQS_PER_TABLE * sizeof(u32));
3730 memset(table->table, 0,
3731 (MAX_IRQS_PER_TABLE * (sizeof(u64) * 2)));
3735 static void set_remap_table_entry(struct amd_iommu *iommu, u16 devid,
3736 struct irq_remap_table *table)
3738 irq_lookup_table[devid] = table;
3739 set_dte_irq_entry(devid, table);
3740 iommu_flush_dte(iommu, devid);
3743 static struct irq_remap_table *alloc_irq_table(u16 devid)
3745 struct irq_remap_table *table = NULL;
3746 struct irq_remap_table *new_table = NULL;
3747 struct amd_iommu *iommu;
3748 unsigned long flags;
3751 spin_lock_irqsave(&iommu_table_lock, flags);
3753 iommu = amd_iommu_rlookup_table[devid];
3757 table = irq_lookup_table[devid];
3761 alias = amd_iommu_alias_table[devid];
3762 table = irq_lookup_table[alias];
3764 set_remap_table_entry(iommu, devid, table);
3767 spin_unlock_irqrestore(&iommu_table_lock, flags);
3769 /* Nothing there yet, allocate new irq remapping table */
3770 new_table = __alloc_irq_table();
3774 spin_lock_irqsave(&iommu_table_lock, flags);
3776 table = irq_lookup_table[devid];
3780 table = irq_lookup_table[alias];
3782 set_remap_table_entry(iommu, devid, table);
3789 set_remap_table_entry(iommu, devid, table);
3791 set_remap_table_entry(iommu, alias, table);
3794 iommu_completion_wait(iommu);
3797 spin_unlock_irqrestore(&iommu_table_lock, flags);
3800 kmem_cache_free(amd_iommu_irq_cache, new_table->table);
3806 static int alloc_irq_index(u16 devid, int count, bool align)
3808 struct irq_remap_table *table;
3809 int index, c, alignment = 1;
3810 unsigned long flags;
3811 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
3816 table = alloc_irq_table(devid);
3821 alignment = roundup_pow_of_two(count);
3823 raw_spin_lock_irqsave(&table->lock, flags);
3825 /* Scan table for free entries */
3826 for (index = ALIGN(table->min_index, alignment), c = 0;
3827 index < MAX_IRQS_PER_TABLE;) {
3828 if (!iommu->irte_ops->is_allocated(table, index)) {
3832 index = ALIGN(index + 1, alignment);
3838 iommu->irte_ops->set_allocated(table, index - c + 1);
3850 raw_spin_unlock_irqrestore(&table->lock, flags);
3855 static int modify_irte_ga(u16 devid, int index, struct irte_ga *irte,
3856 struct amd_ir_data *data)
3858 struct irq_remap_table *table;
3859 struct amd_iommu *iommu;
3860 unsigned long flags;
3861 struct irte_ga *entry;
3863 iommu = amd_iommu_rlookup_table[devid];
3867 table = get_irq_table(devid);
3871 raw_spin_lock_irqsave(&table->lock, flags);
3873 entry = (struct irte_ga *)table->table;
3874 entry = &entry[index];
3875 entry->lo.fields_remap.valid = 0;
3876 entry->hi.val = irte->hi.val;
3877 entry->lo.val = irte->lo.val;
3878 entry->lo.fields_remap.valid = 1;
3882 raw_spin_unlock_irqrestore(&table->lock, flags);
3884 iommu_flush_irt(iommu, devid);
3885 iommu_completion_wait(iommu);
3890 static int modify_irte(u16 devid, int index, union irte *irte)
3892 struct irq_remap_table *table;
3893 struct amd_iommu *iommu;
3894 unsigned long flags;
3896 iommu = amd_iommu_rlookup_table[devid];
3900 table = get_irq_table(devid);
3904 raw_spin_lock_irqsave(&table->lock, flags);
3905 table->table[index] = irte->val;
3906 raw_spin_unlock_irqrestore(&table->lock, flags);
3908 iommu_flush_irt(iommu, devid);
3909 iommu_completion_wait(iommu);
3914 static void free_irte(u16 devid, int index)
3916 struct irq_remap_table *table;
3917 struct amd_iommu *iommu;
3918 unsigned long flags;
3920 iommu = amd_iommu_rlookup_table[devid];
3924 table = get_irq_table(devid);
3928 raw_spin_lock_irqsave(&table->lock, flags);
3929 iommu->irte_ops->clear_allocated(table, index);
3930 raw_spin_unlock_irqrestore(&table->lock, flags);
3932 iommu_flush_irt(iommu, devid);
3933 iommu_completion_wait(iommu);
3936 static void irte_prepare(void *entry,
3937 u32 delivery_mode, u32 dest_mode,
3938 u8 vector, u32 dest_apicid, int devid)
3940 union irte *irte = (union irte *) entry;
3943 irte->fields.vector = vector;
3944 irte->fields.int_type = delivery_mode;
3945 irte->fields.destination = dest_apicid;
3946 irte->fields.dm = dest_mode;
3947 irte->fields.valid = 1;
3950 static void irte_ga_prepare(void *entry,
3951 u32 delivery_mode, u32 dest_mode,
3952 u8 vector, u32 dest_apicid, int devid)
3954 struct irte_ga *irte = (struct irte_ga *) entry;
3958 irte->lo.fields_remap.int_type = delivery_mode;
3959 irte->lo.fields_remap.dm = dest_mode;
3960 irte->hi.fields.vector = vector;
3961 irte->lo.fields_remap.destination = APICID_TO_IRTE_DEST_LO(dest_apicid);
3962 irte->hi.fields.destination = APICID_TO_IRTE_DEST_HI(dest_apicid);
3963 irte->lo.fields_remap.valid = 1;
3966 static void irte_activate(void *entry, u16 devid, u16 index)
3968 union irte *irte = (union irte *) entry;
3970 irte->fields.valid = 1;
3971 modify_irte(devid, index, irte);
3974 static void irte_ga_activate(void *entry, u16 devid, u16 index)
3976 struct irte_ga *irte = (struct irte_ga *) entry;
3978 irte->lo.fields_remap.valid = 1;
3979 modify_irte_ga(devid, index, irte, NULL);
3982 static void irte_deactivate(void *entry, u16 devid, u16 index)
3984 union irte *irte = (union irte *) entry;
3986 irte->fields.valid = 0;
3987 modify_irte(devid, index, irte);
3990 static void irte_ga_deactivate(void *entry, u16 devid, u16 index)
3992 struct irte_ga *irte = (struct irte_ga *) entry;
3994 irte->lo.fields_remap.valid = 0;
3995 modify_irte_ga(devid, index, irte, NULL);
3998 static void irte_set_affinity(void *entry, u16 devid, u16 index,
3999 u8 vector, u32 dest_apicid)
4001 union irte *irte = (union irte *) entry;
4003 irte->fields.vector = vector;
4004 irte->fields.destination = dest_apicid;
4005 modify_irte(devid, index, irte);
4008 static void irte_ga_set_affinity(void *entry, u16 devid, u16 index,
4009 u8 vector, u32 dest_apicid)
4011 struct irte_ga *irte = (struct irte_ga *) entry;
4013 if (!irte->lo.fields_remap.guest_mode) {
4014 irte->hi.fields.vector = vector;
4015 irte->lo.fields_remap.destination =
4016 APICID_TO_IRTE_DEST_LO(dest_apicid);
4017 irte->hi.fields.destination =
4018 APICID_TO_IRTE_DEST_HI(dest_apicid);
4019 modify_irte_ga(devid, index, irte, NULL);
4023 #define IRTE_ALLOCATED (~1U)
4024 static void irte_set_allocated(struct irq_remap_table *table, int index)
4026 table->table[index] = IRTE_ALLOCATED;
4029 static void irte_ga_set_allocated(struct irq_remap_table *table, int index)
4031 struct irte_ga *ptr = (struct irte_ga *)table->table;
4032 struct irte_ga *irte = &ptr[index];
4034 memset(&irte->lo.val, 0, sizeof(u64));
4035 memset(&irte->hi.val, 0, sizeof(u64));
4036 irte->hi.fields.vector = 0xff;
4039 static bool irte_is_allocated(struct irq_remap_table *table, int index)
4041 union irte *ptr = (union irte *)table->table;
4042 union irte *irte = &ptr[index];
4044 return irte->val != 0;
4047 static bool irte_ga_is_allocated(struct irq_remap_table *table, int index)
4049 struct irte_ga *ptr = (struct irte_ga *)table->table;
4050 struct irte_ga *irte = &ptr[index];
4052 return irte->hi.fields.vector != 0;
4055 static void irte_clear_allocated(struct irq_remap_table *table, int index)
4057 table->table[index] = 0;
4060 static void irte_ga_clear_allocated(struct irq_remap_table *table, int index)
4062 struct irte_ga *ptr = (struct irte_ga *)table->table;
4063 struct irte_ga *irte = &ptr[index];
4065 memset(&irte->lo.val, 0, sizeof(u64));
4066 memset(&irte->hi.val, 0, sizeof(u64));
4069 static int get_devid(struct irq_alloc_info *info)
4073 switch (info->type) {
4074 case X86_IRQ_ALLOC_TYPE_IOAPIC:
4075 devid = get_ioapic_devid(info->ioapic_id);
4077 case X86_IRQ_ALLOC_TYPE_HPET:
4078 devid = get_hpet_devid(info->hpet_id);
4080 case X86_IRQ_ALLOC_TYPE_MSI:
4081 case X86_IRQ_ALLOC_TYPE_MSIX:
4082 devid = get_device_id(&info->msi_dev->dev);
4092 static struct irq_domain *get_ir_irq_domain(struct irq_alloc_info *info)
4094 struct amd_iommu *iommu;
4100 devid = get_devid(info);
4102 iommu = amd_iommu_rlookup_table[devid];
4104 return iommu->ir_domain;
4110 static struct irq_domain *get_irq_domain(struct irq_alloc_info *info)
4112 struct amd_iommu *iommu;
4118 switch (info->type) {
4119 case X86_IRQ_ALLOC_TYPE_MSI:
4120 case X86_IRQ_ALLOC_TYPE_MSIX:
4121 devid = get_device_id(&info->msi_dev->dev);
4125 iommu = amd_iommu_rlookup_table[devid];
4127 return iommu->msi_domain;
4136 struct irq_remap_ops amd_iommu_irq_ops = {
4137 .prepare = amd_iommu_prepare,
4138 .enable = amd_iommu_enable,
4139 .disable = amd_iommu_disable,
4140 .reenable = amd_iommu_reenable,
4141 .enable_faulting = amd_iommu_enable_faulting,
4142 .get_ir_irq_domain = get_ir_irq_domain,
4143 .get_irq_domain = get_irq_domain,
4146 static void irq_remapping_prepare_irte(struct amd_ir_data *data,
4147 struct irq_cfg *irq_cfg,
4148 struct irq_alloc_info *info,
4149 int devid, int index, int sub_handle)
4151 struct irq_2_irte *irte_info = &data->irq_2_irte;
4152 struct msi_msg *msg = &data->msi_entry;
4153 struct IO_APIC_route_entry *entry;
4154 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
4159 data->irq_2_irte.devid = devid;
4160 data->irq_2_irte.index = index + sub_handle;
4161 iommu->irte_ops->prepare(data->entry, apic->irq_delivery_mode,
4162 apic->irq_dest_mode, irq_cfg->vector,
4163 irq_cfg->dest_apicid, devid);
4165 switch (info->type) {
4166 case X86_IRQ_ALLOC_TYPE_IOAPIC:
4167 /* Setup IOAPIC entry */
4168 entry = info->ioapic_entry;
4169 info->ioapic_entry = NULL;
4170 memset(entry, 0, sizeof(*entry));
4171 entry->vector = index;
4173 entry->trigger = info->ioapic_trigger;
4174 entry->polarity = info->ioapic_polarity;
4175 /* Mask level triggered irqs. */
4176 if (info->ioapic_trigger)
4180 case X86_IRQ_ALLOC_TYPE_HPET:
4181 case X86_IRQ_ALLOC_TYPE_MSI:
4182 case X86_IRQ_ALLOC_TYPE_MSIX:
4183 msg->address_hi = MSI_ADDR_BASE_HI;
4184 msg->address_lo = MSI_ADDR_BASE_LO;
4185 msg->data = irte_info->index;
4194 struct amd_irte_ops irte_32_ops = {
4195 .prepare = irte_prepare,
4196 .activate = irte_activate,
4197 .deactivate = irte_deactivate,
4198 .set_affinity = irte_set_affinity,
4199 .set_allocated = irte_set_allocated,
4200 .is_allocated = irte_is_allocated,
4201 .clear_allocated = irte_clear_allocated,
4204 struct amd_irte_ops irte_128_ops = {
4205 .prepare = irte_ga_prepare,
4206 .activate = irte_ga_activate,
4207 .deactivate = irte_ga_deactivate,
4208 .set_affinity = irte_ga_set_affinity,
4209 .set_allocated = irte_ga_set_allocated,
4210 .is_allocated = irte_ga_is_allocated,
4211 .clear_allocated = irte_ga_clear_allocated,
4214 static int irq_remapping_alloc(struct irq_domain *domain, unsigned int virq,
4215 unsigned int nr_irqs, void *arg)
4217 struct irq_alloc_info *info = arg;
4218 struct irq_data *irq_data;
4219 struct amd_ir_data *data = NULL;
4220 struct irq_cfg *cfg;
4226 if (nr_irqs > 1 && info->type != X86_IRQ_ALLOC_TYPE_MSI &&
4227 info->type != X86_IRQ_ALLOC_TYPE_MSIX)
4231 * With IRQ remapping enabled, don't need contiguous CPU vectors
4232 * to support multiple MSI interrupts.
4234 if (info->type == X86_IRQ_ALLOC_TYPE_MSI)
4235 info->flags &= ~X86_IRQ_ALLOC_CONTIGUOUS_VECTORS;
4237 devid = get_devid(info);
4241 ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
4245 if (info->type == X86_IRQ_ALLOC_TYPE_IOAPIC) {
4246 struct irq_remap_table *table;
4247 struct amd_iommu *iommu;
4249 table = alloc_irq_table(devid);
4251 if (!table->min_index) {
4253 * Keep the first 32 indexes free for IOAPIC
4256 table->min_index = 32;
4257 iommu = amd_iommu_rlookup_table[devid];
4258 for (i = 0; i < 32; ++i)
4259 iommu->irte_ops->set_allocated(table, i);
4261 WARN_ON(table->min_index != 32);
4262 index = info->ioapic_pin;
4267 bool align = (info->type == X86_IRQ_ALLOC_TYPE_MSI);
4269 index = alloc_irq_index(devid, nr_irqs, align);
4272 pr_warn("Failed to allocate IRTE\n");
4274 goto out_free_parent;
4277 for (i = 0; i < nr_irqs; i++) {
4278 irq_data = irq_domain_get_irq_data(domain, virq + i);
4279 cfg = irqd_cfg(irq_data);
4280 if (!irq_data || !cfg) {
4286 data = kzalloc(sizeof(*data), GFP_KERNEL);
4290 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
4291 data->entry = kzalloc(sizeof(union irte), GFP_KERNEL);
4293 data->entry = kzalloc(sizeof(struct irte_ga),
4300 irq_data->hwirq = (devid << 16) + i;
4301 irq_data->chip_data = data;
4302 irq_data->chip = &amd_ir_chip;
4303 irq_remapping_prepare_irte(data, cfg, info, devid, index, i);
4304 irq_set_status_flags(virq + i, IRQ_MOVE_PCNTXT);
4310 for (i--; i >= 0; i--) {
4311 irq_data = irq_domain_get_irq_data(domain, virq + i);
4313 kfree(irq_data->chip_data);
4315 for (i = 0; i < nr_irqs; i++)
4316 free_irte(devid, index + i);
4318 irq_domain_free_irqs_common(domain, virq, nr_irqs);
4322 static void irq_remapping_free(struct irq_domain *domain, unsigned int virq,
4323 unsigned int nr_irqs)
4325 struct irq_2_irte *irte_info;
4326 struct irq_data *irq_data;
4327 struct amd_ir_data *data;
4330 for (i = 0; i < nr_irqs; i++) {
4331 irq_data = irq_domain_get_irq_data(domain, virq + i);
4332 if (irq_data && irq_data->chip_data) {
4333 data = irq_data->chip_data;
4334 irte_info = &data->irq_2_irte;
4335 free_irte(irte_info->devid, irte_info->index);
4340 irq_domain_free_irqs_common(domain, virq, nr_irqs);
4343 static void amd_ir_update_irte(struct irq_data *irqd, struct amd_iommu *iommu,
4344 struct amd_ir_data *ir_data,
4345 struct irq_2_irte *irte_info,
4346 struct irq_cfg *cfg);
4348 static int irq_remapping_activate(struct irq_domain *domain,
4349 struct irq_data *irq_data, bool reserve)
4351 struct amd_ir_data *data = irq_data->chip_data;
4352 struct irq_2_irte *irte_info = &data->irq_2_irte;
4353 struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4354 struct irq_cfg *cfg = irqd_cfg(irq_data);
4359 iommu->irte_ops->activate(data->entry, irte_info->devid,
4361 amd_ir_update_irte(irq_data, iommu, data, irte_info, cfg);
4365 static void irq_remapping_deactivate(struct irq_domain *domain,
4366 struct irq_data *irq_data)
4368 struct amd_ir_data *data = irq_data->chip_data;
4369 struct irq_2_irte *irte_info = &data->irq_2_irte;
4370 struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4373 iommu->irte_ops->deactivate(data->entry, irte_info->devid,
4377 static const struct irq_domain_ops amd_ir_domain_ops = {
4378 .alloc = irq_remapping_alloc,
4379 .free = irq_remapping_free,
4380 .activate = irq_remapping_activate,
4381 .deactivate = irq_remapping_deactivate,
4384 int amd_iommu_activate_guest_mode(void *data)
4386 struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
4387 struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
4389 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
4390 !entry || entry->lo.fields_vapic.guest_mode)
4396 entry->lo.fields_vapic.guest_mode = 1;
4397 entry->lo.fields_vapic.ga_log_intr = 1;
4398 entry->hi.fields.ga_root_ptr = ir_data->ga_root_ptr;
4399 entry->hi.fields.vector = ir_data->ga_vector;
4400 entry->lo.fields_vapic.ga_tag = ir_data->ga_tag;
4402 return modify_irte_ga(ir_data->irq_2_irte.devid,
4403 ir_data->irq_2_irte.index, entry, NULL);
4405 EXPORT_SYMBOL(amd_iommu_activate_guest_mode);
4407 int amd_iommu_deactivate_guest_mode(void *data)
4409 struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
4410 struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
4411 struct irq_cfg *cfg = ir_data->cfg;
4413 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
4414 !entry || !entry->lo.fields_vapic.guest_mode)
4420 entry->lo.fields_remap.dm = apic->irq_dest_mode;
4421 entry->lo.fields_remap.int_type = apic->irq_delivery_mode;
4422 entry->hi.fields.vector = cfg->vector;
4423 entry->lo.fields_remap.destination =
4424 APICID_TO_IRTE_DEST_LO(cfg->dest_apicid);
4425 entry->hi.fields.destination =
4426 APICID_TO_IRTE_DEST_HI(cfg->dest_apicid);
4428 return modify_irte_ga(ir_data->irq_2_irte.devid,
4429 ir_data->irq_2_irte.index, entry, NULL);
4431 EXPORT_SYMBOL(amd_iommu_deactivate_guest_mode);
4433 static int amd_ir_set_vcpu_affinity(struct irq_data *data, void *vcpu_info)
4436 struct amd_iommu *iommu;
4437 struct amd_iommu_pi_data *pi_data = vcpu_info;
4438 struct vcpu_data *vcpu_pi_info = pi_data->vcpu_data;
4439 struct amd_ir_data *ir_data = data->chip_data;
4440 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
4441 struct iommu_dev_data *dev_data = search_dev_data(irte_info->devid);
4444 * This device has never been set up for guest mode.
4445 * we should not modify the IRTE
4447 if (!dev_data || !dev_data->use_vapic)
4450 ir_data->cfg = irqd_cfg(data);
4451 pi_data->ir_data = ir_data;
4454 * SVM tries to set up for VAPIC mode, but we are in
4455 * legacy mode. So, we force legacy mode instead.
4457 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
4458 pr_debug("%s: Fall back to using intr legacy remap\n",
4460 pi_data->is_guest_mode = false;
4463 iommu = amd_iommu_rlookup_table[irte_info->devid];
4467 pi_data->prev_ga_tag = ir_data->cached_ga_tag;
4468 if (pi_data->is_guest_mode) {
4469 ir_data->ga_root_ptr = (pi_data->base >> 12);
4470 ir_data->ga_vector = vcpu_pi_info->vector;
4471 ir_data->ga_tag = pi_data->ga_tag;
4472 ret = amd_iommu_activate_guest_mode(ir_data);
4474 ir_data->cached_ga_tag = pi_data->ga_tag;
4476 ret = amd_iommu_deactivate_guest_mode(ir_data);
4479 * This communicates the ga_tag back to the caller
4480 * so that it can do all the necessary clean up.
4483 ir_data->cached_ga_tag = 0;
4490 static void amd_ir_update_irte(struct irq_data *irqd, struct amd_iommu *iommu,
4491 struct amd_ir_data *ir_data,
4492 struct irq_2_irte *irte_info,
4493 struct irq_cfg *cfg)
4497 * Atomically updates the IRTE with the new destination, vector
4498 * and flushes the interrupt entry cache.
4500 iommu->irte_ops->set_affinity(ir_data->entry, irte_info->devid,
4501 irte_info->index, cfg->vector,
4505 static int amd_ir_set_affinity(struct irq_data *data,
4506 const struct cpumask *mask, bool force)
4508 struct amd_ir_data *ir_data = data->chip_data;
4509 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
4510 struct irq_cfg *cfg = irqd_cfg(data);
4511 struct irq_data *parent = data->parent_data;
4512 struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4518 ret = parent->chip->irq_set_affinity(parent, mask, force);
4519 if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
4522 amd_ir_update_irte(data, iommu, ir_data, irte_info, cfg);
4524 * After this point, all the interrupts will start arriving
4525 * at the new destination. So, time to cleanup the previous
4526 * vector allocation.
4528 send_cleanup_vector(cfg);
4530 return IRQ_SET_MASK_OK_DONE;
4533 static void ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg)
4535 struct amd_ir_data *ir_data = irq_data->chip_data;
4537 *msg = ir_data->msi_entry;
4540 static struct irq_chip amd_ir_chip = {
4542 .irq_ack = apic_ack_irq,
4543 .irq_set_affinity = amd_ir_set_affinity,
4544 .irq_set_vcpu_affinity = amd_ir_set_vcpu_affinity,
4545 .irq_compose_msi_msg = ir_compose_msi_msg,
4548 int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
4550 struct fwnode_handle *fn;
4552 fn = irq_domain_alloc_named_id_fwnode("AMD-IR", iommu->index);
4555 iommu->ir_domain = irq_domain_create_tree(fn, &amd_ir_domain_ops, iommu);
4556 irq_domain_free_fwnode(fn);
4557 if (!iommu->ir_domain)
4560 iommu->ir_domain->parent = arch_get_ir_parent_domain();
4561 iommu->msi_domain = arch_create_remap_msi_irq_domain(iommu->ir_domain,
4567 int amd_iommu_update_ga(int cpu, bool is_run, void *data)
4569 unsigned long flags;
4570 struct amd_iommu *iommu;
4571 struct irq_remap_table *table;
4572 struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
4573 int devid = ir_data->irq_2_irte.devid;
4574 struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
4575 struct irte_ga *ref = (struct irte_ga *) ir_data->ref;
4577 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
4578 !ref || !entry || !entry->lo.fields_vapic.guest_mode)
4581 iommu = amd_iommu_rlookup_table[devid];
4585 table = get_irq_table(devid);
4589 raw_spin_lock_irqsave(&table->lock, flags);
4591 if (ref->lo.fields_vapic.guest_mode) {
4593 ref->lo.fields_vapic.destination =
4594 APICID_TO_IRTE_DEST_LO(cpu);
4595 ref->hi.fields.destination =
4596 APICID_TO_IRTE_DEST_HI(cpu);
4598 ref->lo.fields_vapic.is_run = is_run;
4602 raw_spin_unlock_irqrestore(&table->lock, flags);
4604 iommu_flush_irt(iommu, devid);
4605 iommu_completion_wait(iommu);
4608 EXPORT_SYMBOL(amd_iommu_update_ga);