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/dma-iommu.h>
24 #include <linux/iommu-helper.h>
25 #include <linux/iommu.h>
26 #include <linux/delay.h>
27 #include <linux/amd-iommu.h>
28 #include <linux/notifier.h>
29 #include <linux/export.h>
30 #include <linux/irq.h>
31 #include <linux/msi.h>
32 #include <linux/dma-contiguous.h>
33 #include <linux/irqdomain.h>
34 #include <linux/percpu.h>
35 #include <linux/iova.h>
36 #include <asm/irq_remapping.h>
37 #include <asm/io_apic.h>
39 #include <asm/hw_irq.h>
40 #include <asm/msidef.h>
41 #include <asm/proto.h>
42 #include <asm/iommu.h>
46 #include "amd_iommu_proto.h"
47 #include "amd_iommu_types.h"
48 #include "irq_remapping.h"
50 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
52 #define LOOP_TIMEOUT 100000
54 /* IO virtual address start page frame number */
55 #define IOVA_START_PFN (1)
56 #define IOVA_PFN(addr) ((addr) >> PAGE_SHIFT)
58 /* Reserved IOVA ranges */
59 #define MSI_RANGE_START (0xfee00000)
60 #define MSI_RANGE_END (0xfeefffff)
61 #define HT_RANGE_START (0xfd00000000ULL)
62 #define HT_RANGE_END (0xffffffffffULL)
65 * This bitmap is used to advertise the page sizes our hardware support
66 * to the IOMMU core, which will then use this information to split
67 * physically contiguous memory regions it is mapping into page sizes
70 * 512GB Pages are not supported due to a hardware bug
72 #define AMD_IOMMU_PGSIZES ((~0xFFFUL) & ~(2ULL << 38))
74 static DEFINE_SPINLOCK(pd_bitmap_lock);
76 /* List of all available dev_data structures */
77 static LLIST_HEAD(dev_data_list);
79 LIST_HEAD(ioapic_map);
81 LIST_HEAD(acpihid_map);
84 * Domain for untranslated devices - only allocated
85 * if iommu=pt passed on kernel cmd line.
87 const struct iommu_ops amd_iommu_ops;
89 static ATOMIC_NOTIFIER_HEAD(ppr_notifier);
90 int amd_iommu_max_glx_val = -1;
93 * general struct to manage commands send to an IOMMU
99 struct kmem_cache *amd_iommu_irq_cache;
101 static void update_domain(struct protection_domain *domain);
102 static int protection_domain_init(struct protection_domain *domain);
103 static void detach_device(struct device *dev);
104 static void update_and_flush_device_table(struct protection_domain *domain,
105 struct domain_pgtable *pgtable);
107 /****************************************************************************
111 ****************************************************************************/
113 static inline u16 get_pci_device_id(struct device *dev)
115 struct pci_dev *pdev = to_pci_dev(dev);
117 return pci_dev_id(pdev);
120 static inline int get_acpihid_device_id(struct device *dev,
121 struct acpihid_map_entry **entry)
123 struct acpi_device *adev = ACPI_COMPANION(dev);
124 struct acpihid_map_entry *p;
129 list_for_each_entry(p, &acpihid_map, list) {
130 if (acpi_dev_hid_uid_match(adev, p->hid,
131 p->uid[0] ? p->uid : NULL)) {
140 static inline int get_device_id(struct device *dev)
145 devid = get_pci_device_id(dev);
147 devid = get_acpihid_device_id(dev, NULL);
152 static struct protection_domain *to_pdomain(struct iommu_domain *dom)
154 return container_of(dom, struct protection_domain, domain);
157 static void amd_iommu_domain_get_pgtable(struct protection_domain *domain,
158 struct domain_pgtable *pgtable)
160 u64 pt_root = atomic64_read(&domain->pt_root);
162 pgtable->root = (u64 *)(pt_root & PAGE_MASK);
163 pgtable->mode = pt_root & 7; /* lowest 3 bits encode pgtable mode */
166 static u64 amd_iommu_domain_encode_pgtable(u64 *root, int mode)
170 /* lowest 3 bits encode pgtable mode */
172 pt_root |= (u64)root;
177 static struct iommu_dev_data *alloc_dev_data(u16 devid)
179 struct iommu_dev_data *dev_data;
181 dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
185 spin_lock_init(&dev_data->lock);
186 dev_data->devid = devid;
187 ratelimit_default_init(&dev_data->rs);
189 llist_add(&dev_data->dev_data_list, &dev_data_list);
193 static struct iommu_dev_data *search_dev_data(u16 devid)
195 struct iommu_dev_data *dev_data;
196 struct llist_node *node;
198 if (llist_empty(&dev_data_list))
201 node = dev_data_list.first;
202 llist_for_each_entry(dev_data, node, dev_data_list) {
203 if (dev_data->devid == devid)
210 static int clone_alias(struct pci_dev *pdev, u16 alias, void *data)
212 u16 devid = pci_dev_id(pdev);
217 amd_iommu_rlookup_table[alias] =
218 amd_iommu_rlookup_table[devid];
219 memcpy(amd_iommu_dev_table[alias].data,
220 amd_iommu_dev_table[devid].data,
221 sizeof(amd_iommu_dev_table[alias].data));
226 static void clone_aliases(struct pci_dev *pdev)
232 * The IVRS alias stored in the alias table may not be
233 * part of the PCI DMA aliases if it's bus differs
234 * from the original device.
236 clone_alias(pdev, amd_iommu_alias_table[pci_dev_id(pdev)], NULL);
238 pci_for_each_dma_alias(pdev, clone_alias, NULL);
241 static struct pci_dev *setup_aliases(struct device *dev)
243 struct pci_dev *pdev = to_pci_dev(dev);
246 /* For ACPI HID devices, there are no aliases */
247 if (!dev_is_pci(dev))
251 * Add the IVRS alias to the pci aliases if it is on the same
252 * bus. The IVRS table may know about a quirk that we don't.
254 ivrs_alias = amd_iommu_alias_table[pci_dev_id(pdev)];
255 if (ivrs_alias != pci_dev_id(pdev) &&
256 PCI_BUS_NUM(ivrs_alias) == pdev->bus->number)
257 pci_add_dma_alias(pdev, ivrs_alias & 0xff, 1);
264 static struct iommu_dev_data *find_dev_data(u16 devid)
266 struct iommu_dev_data *dev_data;
267 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
269 dev_data = search_dev_data(devid);
271 if (dev_data == NULL) {
272 dev_data = alloc_dev_data(devid);
276 if (translation_pre_enabled(iommu))
277 dev_data->defer_attach = true;
283 struct iommu_dev_data *get_dev_data(struct device *dev)
285 return dev->archdata.iommu;
287 EXPORT_SYMBOL(get_dev_data);
290 * Find or create an IOMMU group for a acpihid device.
292 static struct iommu_group *acpihid_device_group(struct device *dev)
294 struct acpihid_map_entry *p, *entry = NULL;
297 devid = get_acpihid_device_id(dev, &entry);
299 return ERR_PTR(devid);
301 list_for_each_entry(p, &acpihid_map, list) {
302 if ((devid == p->devid) && p->group)
303 entry->group = p->group;
307 entry->group = generic_device_group(dev);
309 iommu_group_ref_get(entry->group);
314 static bool pci_iommuv2_capable(struct pci_dev *pdev)
316 static const int caps[] = {
319 PCI_EXT_CAP_ID_PASID,
323 if (pci_ats_disabled())
326 for (i = 0; i < 3; ++i) {
327 pos = pci_find_ext_capability(pdev, caps[i]);
335 static bool pdev_pri_erratum(struct pci_dev *pdev, u32 erratum)
337 struct iommu_dev_data *dev_data;
339 dev_data = get_dev_data(&pdev->dev);
341 return dev_data->errata & (1 << erratum) ? true : false;
345 * This function checks if the driver got a valid device from the caller to
346 * avoid dereferencing invalid pointers.
348 static bool check_device(struct device *dev)
355 devid = get_device_id(dev);
359 /* Out of our scope? */
360 if (devid > amd_iommu_last_bdf)
363 if (amd_iommu_rlookup_table[devid] == NULL)
369 static int iommu_init_device(struct device *dev)
371 struct iommu_dev_data *dev_data;
374 if (dev->archdata.iommu)
377 devid = get_device_id(dev);
381 dev_data = find_dev_data(devid);
385 dev_data->pdev = setup_aliases(dev);
388 * By default we use passthrough mode for IOMMUv2 capable device.
389 * But if amd_iommu=force_isolation is set (e.g. to debug DMA to
390 * invalid address), we ignore the capability for the device so
391 * it'll be forced to go into translation mode.
393 if ((iommu_default_passthrough() || !amd_iommu_force_isolation) &&
394 dev_is_pci(dev) && pci_iommuv2_capable(to_pci_dev(dev))) {
395 struct amd_iommu *iommu;
397 iommu = amd_iommu_rlookup_table[dev_data->devid];
398 dev_data->iommu_v2 = iommu->is_iommu_v2;
401 dev->archdata.iommu = dev_data;
406 static void iommu_ignore_device(struct device *dev)
410 devid = get_device_id(dev);
414 amd_iommu_rlookup_table[devid] = NULL;
415 memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
420 static void amd_iommu_uninit_device(struct device *dev)
422 struct iommu_dev_data *dev_data;
425 devid = get_device_id(dev);
429 dev_data = search_dev_data(devid);
433 if (dev_data->domain)
437 * We keep dev_data around for unplugged devices and reuse it when the
438 * device is re-plugged - not doing so would introduce a ton of races.
443 * Helper function to get the first pte of a large mapping
445 static u64 *first_pte_l7(u64 *pte, unsigned long *page_size,
446 unsigned long *count)
448 unsigned long pte_mask, pg_size, cnt;
451 pg_size = PTE_PAGE_SIZE(*pte);
452 cnt = PAGE_SIZE_PTE_COUNT(pg_size);
453 pte_mask = ~((cnt << 3) - 1);
454 fpte = (u64 *)(((unsigned long)pte) & pte_mask);
457 *page_size = pg_size;
465 /****************************************************************************
467 * Interrupt handling functions
469 ****************************************************************************/
471 static void dump_dte_entry(u16 devid)
475 for (i = 0; i < 4; ++i)
476 pr_err("DTE[%d]: %016llx\n", i,
477 amd_iommu_dev_table[devid].data[i]);
480 static void dump_command(unsigned long phys_addr)
482 struct iommu_cmd *cmd = iommu_phys_to_virt(phys_addr);
485 for (i = 0; i < 4; ++i)
486 pr_err("CMD[%d]: %08x\n", i, cmd->data[i]);
489 static void amd_iommu_report_page_fault(u16 devid, u16 domain_id,
490 u64 address, int flags)
492 struct iommu_dev_data *dev_data = NULL;
493 struct pci_dev *pdev;
495 pdev = pci_get_domain_bus_and_slot(0, PCI_BUS_NUM(devid),
498 dev_data = get_dev_data(&pdev->dev);
500 if (dev_data && __ratelimit(&dev_data->rs)) {
501 pci_err(pdev, "Event logged [IO_PAGE_FAULT domain=0x%04x address=0x%llx flags=0x%04x]\n",
502 domain_id, address, flags);
503 } else if (printk_ratelimit()) {
504 pr_err("Event logged [IO_PAGE_FAULT device=%02x:%02x.%x domain=0x%04x address=0x%llx flags=0x%04x]\n",
505 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
506 domain_id, address, flags);
513 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
515 struct device *dev = iommu->iommu.dev;
516 int type, devid, pasid, flags, tag;
517 volatile u32 *event = __evt;
522 type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
523 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
524 pasid = (event[0] & EVENT_DOMID_MASK_HI) |
525 (event[1] & EVENT_DOMID_MASK_LO);
526 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
527 address = (u64)(((u64)event[3]) << 32) | event[2];
530 /* Did we hit the erratum? */
531 if (++count == LOOP_TIMEOUT) {
532 pr_err("No event written to event log\n");
539 if (type == EVENT_TYPE_IO_FAULT) {
540 amd_iommu_report_page_fault(devid, pasid, address, flags);
545 case EVENT_TYPE_ILL_DEV:
546 dev_err(dev, "Event logged [ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x pasid=0x%05x address=0x%llx flags=0x%04x]\n",
547 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
548 pasid, address, flags);
549 dump_dte_entry(devid);
551 case EVENT_TYPE_DEV_TAB_ERR:
552 dev_err(dev, "Event logged [DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
553 "address=0x%llx flags=0x%04x]\n",
554 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
557 case EVENT_TYPE_PAGE_TAB_ERR:
558 dev_err(dev, "Event logged [PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x pasid=0x%04x address=0x%llx flags=0x%04x]\n",
559 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
560 pasid, address, flags);
562 case EVENT_TYPE_ILL_CMD:
563 dev_err(dev, "Event logged [ILLEGAL_COMMAND_ERROR address=0x%llx]\n", address);
564 dump_command(address);
566 case EVENT_TYPE_CMD_HARD_ERR:
567 dev_err(dev, "Event logged [COMMAND_HARDWARE_ERROR address=0x%llx flags=0x%04x]\n",
570 case EVENT_TYPE_IOTLB_INV_TO:
571 dev_err(dev, "Event logged [IOTLB_INV_TIMEOUT device=%02x:%02x.%x address=0x%llx]\n",
572 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
575 case EVENT_TYPE_INV_DEV_REQ:
576 dev_err(dev, "Event logged [INVALID_DEVICE_REQUEST device=%02x:%02x.%x pasid=0x%05x address=0x%llx flags=0x%04x]\n",
577 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
578 pasid, address, flags);
580 case EVENT_TYPE_INV_PPR_REQ:
581 pasid = PPR_PASID(*((u64 *)__evt));
582 tag = event[1] & 0x03FF;
583 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",
584 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
585 pasid, address, flags, tag);
588 dev_err(dev, "Event logged [UNKNOWN event[0]=0x%08x event[1]=0x%08x event[2]=0x%08x event[3]=0x%08x\n",
589 event[0], event[1], event[2], event[3]);
592 memset(__evt, 0, 4 * sizeof(u32));
595 static void iommu_poll_events(struct amd_iommu *iommu)
599 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
600 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
602 while (head != tail) {
603 iommu_print_event(iommu, iommu->evt_buf + head);
604 head = (head + EVENT_ENTRY_SIZE) % EVT_BUFFER_SIZE;
607 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
610 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
612 struct amd_iommu_fault fault;
614 if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
615 pr_err_ratelimited("Unknown PPR request received\n");
619 fault.address = raw[1];
620 fault.pasid = PPR_PASID(raw[0]);
621 fault.device_id = PPR_DEVID(raw[0]);
622 fault.tag = PPR_TAG(raw[0]);
623 fault.flags = PPR_FLAGS(raw[0]);
625 atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
628 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
632 if (iommu->ppr_log == NULL)
635 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
636 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
638 while (head != tail) {
643 raw = (u64 *)(iommu->ppr_log + head);
646 * Hardware bug: Interrupt may arrive before the entry is
647 * written to memory. If this happens we need to wait for the
650 for (i = 0; i < LOOP_TIMEOUT; ++i) {
651 if (PPR_REQ_TYPE(raw[0]) != 0)
656 /* Avoid memcpy function-call overhead */
661 * To detect the hardware bug we need to clear the entry
664 raw[0] = raw[1] = 0UL;
666 /* Update head pointer of hardware ring-buffer */
667 head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
668 writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
670 /* Handle PPR entry */
671 iommu_handle_ppr_entry(iommu, entry);
673 /* Refresh ring-buffer information */
674 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
675 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
679 #ifdef CONFIG_IRQ_REMAP
680 static int (*iommu_ga_log_notifier)(u32);
682 int amd_iommu_register_ga_log_notifier(int (*notifier)(u32))
684 iommu_ga_log_notifier = notifier;
688 EXPORT_SYMBOL(amd_iommu_register_ga_log_notifier);
690 static void iommu_poll_ga_log(struct amd_iommu *iommu)
692 u32 head, tail, cnt = 0;
694 if (iommu->ga_log == NULL)
697 head = readl(iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
698 tail = readl(iommu->mmio_base + MMIO_GA_TAIL_OFFSET);
700 while (head != tail) {
704 raw = (u64 *)(iommu->ga_log + head);
707 /* Avoid memcpy function-call overhead */
710 /* Update head pointer of hardware ring-buffer */
711 head = (head + GA_ENTRY_SIZE) % GA_LOG_SIZE;
712 writel(head, iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
714 /* Handle GA entry */
715 switch (GA_REQ_TYPE(log_entry)) {
717 if (!iommu_ga_log_notifier)
720 pr_debug("%s: devid=%#x, ga_tag=%#x\n",
721 __func__, GA_DEVID(log_entry),
724 if (iommu_ga_log_notifier(GA_TAG(log_entry)) != 0)
725 pr_err("GA log notifier failed.\n");
732 #endif /* CONFIG_IRQ_REMAP */
734 #define AMD_IOMMU_INT_MASK \
735 (MMIO_STATUS_EVT_INT_MASK | \
736 MMIO_STATUS_PPR_INT_MASK | \
737 MMIO_STATUS_GALOG_INT_MASK)
739 irqreturn_t amd_iommu_int_thread(int irq, void *data)
741 struct amd_iommu *iommu = (struct amd_iommu *) data;
742 u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
744 while (status & AMD_IOMMU_INT_MASK) {
745 /* Enable EVT and PPR and GA interrupts again */
746 writel(AMD_IOMMU_INT_MASK,
747 iommu->mmio_base + MMIO_STATUS_OFFSET);
749 if (status & MMIO_STATUS_EVT_INT_MASK) {
750 pr_devel("Processing IOMMU Event Log\n");
751 iommu_poll_events(iommu);
754 if (status & MMIO_STATUS_PPR_INT_MASK) {
755 pr_devel("Processing IOMMU PPR Log\n");
756 iommu_poll_ppr_log(iommu);
759 #ifdef CONFIG_IRQ_REMAP
760 if (status & MMIO_STATUS_GALOG_INT_MASK) {
761 pr_devel("Processing IOMMU GA Log\n");
762 iommu_poll_ga_log(iommu);
767 * Hardware bug: ERBT1312
768 * When re-enabling interrupt (by writing 1
769 * to clear the bit), the hardware might also try to set
770 * the interrupt bit in the event status register.
771 * In this scenario, the bit will be set, and disable
772 * subsequent interrupts.
774 * Workaround: The IOMMU driver should read back the
775 * status register and check if the interrupt bits are cleared.
776 * If not, driver will need to go through the interrupt handler
777 * again and re-clear the bits
779 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
784 irqreturn_t amd_iommu_int_handler(int irq, void *data)
786 return IRQ_WAKE_THREAD;
789 /****************************************************************************
791 * IOMMU command queuing functions
793 ****************************************************************************/
795 static int wait_on_sem(volatile u64 *sem)
799 while (*sem == 0 && i < LOOP_TIMEOUT) {
804 if (i == LOOP_TIMEOUT) {
805 pr_alert("Completion-Wait loop timed out\n");
812 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
813 struct iommu_cmd *cmd)
818 /* Copy command to buffer */
819 tail = iommu->cmd_buf_tail;
820 target = iommu->cmd_buf + tail;
821 memcpy(target, cmd, sizeof(*cmd));
823 tail = (tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
824 iommu->cmd_buf_tail = tail;
826 /* Tell the IOMMU about it */
827 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
830 static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
832 u64 paddr = iommu_virt_to_phys((void *)address);
834 WARN_ON(address & 0x7ULL);
836 memset(cmd, 0, sizeof(*cmd));
837 cmd->data[0] = lower_32_bits(paddr) | CMD_COMPL_WAIT_STORE_MASK;
838 cmd->data[1] = upper_32_bits(paddr);
840 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
843 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
845 memset(cmd, 0, sizeof(*cmd));
846 cmd->data[0] = devid;
847 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
850 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
851 size_t size, u16 domid, int pde)
856 pages = iommu_num_pages(address, size, PAGE_SIZE);
861 * If we have to flush more than one page, flush all
862 * TLB entries for this domain
864 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
868 address &= PAGE_MASK;
870 memset(cmd, 0, sizeof(*cmd));
871 cmd->data[1] |= domid;
872 cmd->data[2] = lower_32_bits(address);
873 cmd->data[3] = upper_32_bits(address);
874 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
875 if (s) /* size bit - we flush more than one 4kb page */
876 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
877 if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
878 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
881 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
882 u64 address, size_t size)
887 pages = iommu_num_pages(address, size, PAGE_SIZE);
892 * If we have to flush more than one page, flush all
893 * TLB entries for this domain
895 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
899 address &= PAGE_MASK;
901 memset(cmd, 0, sizeof(*cmd));
902 cmd->data[0] = devid;
903 cmd->data[0] |= (qdep & 0xff) << 24;
904 cmd->data[1] = devid;
905 cmd->data[2] = lower_32_bits(address);
906 cmd->data[3] = upper_32_bits(address);
907 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
909 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
912 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, int pasid,
913 u64 address, bool size)
915 memset(cmd, 0, sizeof(*cmd));
917 address &= ~(0xfffULL);
919 cmd->data[0] = pasid;
920 cmd->data[1] = domid;
921 cmd->data[2] = lower_32_bits(address);
922 cmd->data[3] = upper_32_bits(address);
923 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
924 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
926 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
927 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
930 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, int pasid,
931 int qdep, u64 address, bool size)
933 memset(cmd, 0, sizeof(*cmd));
935 address &= ~(0xfffULL);
937 cmd->data[0] = devid;
938 cmd->data[0] |= ((pasid >> 8) & 0xff) << 16;
939 cmd->data[0] |= (qdep & 0xff) << 24;
940 cmd->data[1] = devid;
941 cmd->data[1] |= (pasid & 0xff) << 16;
942 cmd->data[2] = lower_32_bits(address);
943 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
944 cmd->data[3] = upper_32_bits(address);
946 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
947 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
950 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, int pasid,
951 int status, int tag, bool gn)
953 memset(cmd, 0, sizeof(*cmd));
955 cmd->data[0] = devid;
957 cmd->data[1] = pasid;
958 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
960 cmd->data[3] = tag & 0x1ff;
961 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
963 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
966 static void build_inv_all(struct iommu_cmd *cmd)
968 memset(cmd, 0, sizeof(*cmd));
969 CMD_SET_TYPE(cmd, CMD_INV_ALL);
972 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
974 memset(cmd, 0, sizeof(*cmd));
975 cmd->data[0] = devid;
976 CMD_SET_TYPE(cmd, CMD_INV_IRT);
980 * Writes the command to the IOMMUs command buffer and informs the
981 * hardware about the new command.
983 static int __iommu_queue_command_sync(struct amd_iommu *iommu,
984 struct iommu_cmd *cmd,
987 unsigned int count = 0;
990 next_tail = (iommu->cmd_buf_tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
992 left = (iommu->cmd_buf_head - next_tail) % CMD_BUFFER_SIZE;
995 /* Skip udelay() the first time around */
997 if (count == LOOP_TIMEOUT) {
998 pr_err("Command buffer timeout\n");
1005 /* Update head and recheck remaining space */
1006 iommu->cmd_buf_head = readl(iommu->mmio_base +
1007 MMIO_CMD_HEAD_OFFSET);
1012 copy_cmd_to_buffer(iommu, cmd);
1014 /* Do we need to make sure all commands are processed? */
1015 iommu->need_sync = sync;
1020 static int iommu_queue_command_sync(struct amd_iommu *iommu,
1021 struct iommu_cmd *cmd,
1024 unsigned long flags;
1027 raw_spin_lock_irqsave(&iommu->lock, flags);
1028 ret = __iommu_queue_command_sync(iommu, cmd, sync);
1029 raw_spin_unlock_irqrestore(&iommu->lock, flags);
1034 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
1036 return iommu_queue_command_sync(iommu, cmd, true);
1040 * This function queues a completion wait command into the command
1041 * buffer of an IOMMU
1043 static int iommu_completion_wait(struct amd_iommu *iommu)
1045 struct iommu_cmd cmd;
1046 unsigned long flags;
1049 if (!iommu->need_sync)
1053 build_completion_wait(&cmd, (u64)&iommu->cmd_sem);
1055 raw_spin_lock_irqsave(&iommu->lock, flags);
1059 ret = __iommu_queue_command_sync(iommu, &cmd, false);
1063 ret = wait_on_sem(&iommu->cmd_sem);
1066 raw_spin_unlock_irqrestore(&iommu->lock, flags);
1071 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
1073 struct iommu_cmd cmd;
1075 build_inv_dte(&cmd, devid);
1077 return iommu_queue_command(iommu, &cmd);
1080 static void amd_iommu_flush_dte_all(struct amd_iommu *iommu)
1084 for (devid = 0; devid <= 0xffff; ++devid)
1085 iommu_flush_dte(iommu, devid);
1087 iommu_completion_wait(iommu);
1091 * This function uses heavy locking and may disable irqs for some time. But
1092 * this is no issue because it is only called during resume.
1094 static void amd_iommu_flush_tlb_all(struct amd_iommu *iommu)
1098 for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
1099 struct iommu_cmd cmd;
1100 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1102 iommu_queue_command(iommu, &cmd);
1105 iommu_completion_wait(iommu);
1108 static void amd_iommu_flush_tlb_domid(struct amd_iommu *iommu, u32 dom_id)
1110 struct iommu_cmd cmd;
1112 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1114 iommu_queue_command(iommu, &cmd);
1116 iommu_completion_wait(iommu);
1119 static void amd_iommu_flush_all(struct amd_iommu *iommu)
1121 struct iommu_cmd cmd;
1123 build_inv_all(&cmd);
1125 iommu_queue_command(iommu, &cmd);
1126 iommu_completion_wait(iommu);
1129 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1131 struct iommu_cmd cmd;
1133 build_inv_irt(&cmd, devid);
1135 iommu_queue_command(iommu, &cmd);
1138 static void amd_iommu_flush_irt_all(struct amd_iommu *iommu)
1142 for (devid = 0; devid <= MAX_DEV_TABLE_ENTRIES; devid++)
1143 iommu_flush_irt(iommu, devid);
1145 iommu_completion_wait(iommu);
1148 void iommu_flush_all_caches(struct amd_iommu *iommu)
1150 if (iommu_feature(iommu, FEATURE_IA)) {
1151 amd_iommu_flush_all(iommu);
1153 amd_iommu_flush_dte_all(iommu);
1154 amd_iommu_flush_irt_all(iommu);
1155 amd_iommu_flush_tlb_all(iommu);
1160 * Command send function for flushing on-device TLB
1162 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1163 u64 address, size_t size)
1165 struct amd_iommu *iommu;
1166 struct iommu_cmd cmd;
1169 qdep = dev_data->ats.qdep;
1170 iommu = amd_iommu_rlookup_table[dev_data->devid];
1172 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1174 return iommu_queue_command(iommu, &cmd);
1177 static int device_flush_dte_alias(struct pci_dev *pdev, u16 alias, void *data)
1179 struct amd_iommu *iommu = data;
1181 return iommu_flush_dte(iommu, alias);
1185 * Command send function for invalidating a device table entry
1187 static int device_flush_dte(struct iommu_dev_data *dev_data)
1189 struct amd_iommu *iommu;
1193 iommu = amd_iommu_rlookup_table[dev_data->devid];
1196 ret = pci_for_each_dma_alias(dev_data->pdev,
1197 device_flush_dte_alias, iommu);
1199 ret = iommu_flush_dte(iommu, dev_data->devid);
1203 alias = amd_iommu_alias_table[dev_data->devid];
1204 if (alias != dev_data->devid) {
1205 ret = iommu_flush_dte(iommu, alias);
1210 if (dev_data->ats.enabled)
1211 ret = device_flush_iotlb(dev_data, 0, ~0UL);
1217 * TLB invalidation function which is called from the mapping functions.
1218 * It invalidates a single PTE if the range to flush is within a single
1219 * page. Otherwise it flushes the whole TLB of the IOMMU.
1221 static void __domain_flush_pages(struct protection_domain *domain,
1222 u64 address, size_t size, int pde)
1224 struct iommu_dev_data *dev_data;
1225 struct iommu_cmd cmd;
1228 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1230 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
1231 if (!domain->dev_iommu[i])
1235 * Devices of this domain are behind this IOMMU
1236 * We need a TLB flush
1238 ret |= iommu_queue_command(amd_iommus[i], &cmd);
1241 list_for_each_entry(dev_data, &domain->dev_list, list) {
1243 if (!dev_data->ats.enabled)
1246 ret |= device_flush_iotlb(dev_data, address, size);
1252 static void domain_flush_pages(struct protection_domain *domain,
1253 u64 address, size_t size)
1255 __domain_flush_pages(domain, address, size, 0);
1258 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1259 static void domain_flush_tlb_pde(struct protection_domain *domain)
1261 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1264 static void domain_flush_complete(struct protection_domain *domain)
1268 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
1269 if (domain && !domain->dev_iommu[i])
1273 * Devices of this domain are behind this IOMMU
1274 * We need to wait for completion of all commands.
1276 iommu_completion_wait(amd_iommus[i]);
1280 /* Flush the not present cache if it exists */
1281 static void domain_flush_np_cache(struct protection_domain *domain,
1282 dma_addr_t iova, size_t size)
1284 if (unlikely(amd_iommu_np_cache)) {
1285 unsigned long flags;
1287 spin_lock_irqsave(&domain->lock, flags);
1288 domain_flush_pages(domain, iova, size);
1289 domain_flush_complete(domain);
1290 spin_unlock_irqrestore(&domain->lock, flags);
1296 * This function flushes the DTEs for all devices in domain
1298 static void domain_flush_devices(struct protection_domain *domain)
1300 struct iommu_dev_data *dev_data;
1302 list_for_each_entry(dev_data, &domain->dev_list, list)
1303 device_flush_dte(dev_data);
1306 /****************************************************************************
1308 * The functions below are used the create the page table mappings for
1309 * unity mapped regions.
1311 ****************************************************************************/
1313 static void free_page_list(struct page *freelist)
1315 while (freelist != NULL) {
1316 unsigned long p = (unsigned long)page_address(freelist);
1317 freelist = freelist->freelist;
1322 static struct page *free_pt_page(unsigned long pt, struct page *freelist)
1324 struct page *p = virt_to_page((void *)pt);
1326 p->freelist = freelist;
1331 #define DEFINE_FREE_PT_FN(LVL, FN) \
1332 static struct page *free_pt_##LVL (unsigned long __pt, struct page *freelist) \
1340 for (i = 0; i < 512; ++i) { \
1341 /* PTE present? */ \
1342 if (!IOMMU_PTE_PRESENT(pt[i])) \
1346 if (PM_PTE_LEVEL(pt[i]) == 0 || \
1347 PM_PTE_LEVEL(pt[i]) == 7) \
1350 p = (unsigned long)IOMMU_PTE_PAGE(pt[i]); \
1351 freelist = FN(p, freelist); \
1354 return free_pt_page((unsigned long)pt, freelist); \
1357 DEFINE_FREE_PT_FN(l2, free_pt_page)
1358 DEFINE_FREE_PT_FN(l3, free_pt_l2)
1359 DEFINE_FREE_PT_FN(l4, free_pt_l3)
1360 DEFINE_FREE_PT_FN(l5, free_pt_l4)
1361 DEFINE_FREE_PT_FN(l6, free_pt_l5)
1363 static struct page *free_sub_pt(unsigned long root, int mode,
1364 struct page *freelist)
1367 case PAGE_MODE_NONE:
1368 case PAGE_MODE_7_LEVEL:
1370 case PAGE_MODE_1_LEVEL:
1371 freelist = free_pt_page(root, freelist);
1373 case PAGE_MODE_2_LEVEL:
1374 freelist = free_pt_l2(root, freelist);
1376 case PAGE_MODE_3_LEVEL:
1377 freelist = free_pt_l3(root, freelist);
1379 case PAGE_MODE_4_LEVEL:
1380 freelist = free_pt_l4(root, freelist);
1382 case PAGE_MODE_5_LEVEL:
1383 freelist = free_pt_l5(root, freelist);
1385 case PAGE_MODE_6_LEVEL:
1386 freelist = free_pt_l6(root, freelist);
1395 static void free_pagetable(struct protection_domain *domain)
1397 struct domain_pgtable pgtable;
1398 struct page *freelist = NULL;
1401 amd_iommu_domain_get_pgtable(domain, &pgtable);
1402 atomic64_set(&domain->pt_root, 0);
1404 BUG_ON(pgtable.mode < PAGE_MODE_NONE ||
1405 pgtable.mode > PAGE_MODE_6_LEVEL);
1407 root = (unsigned long)pgtable.root;
1408 freelist = free_sub_pt(root, pgtable.mode, freelist);
1410 free_page_list(freelist);
1414 * This function is used to add another level to an IO page table. Adding
1415 * another level increases the size of the address space by 9 bits to a size up
1418 static bool increase_address_space(struct protection_domain *domain,
1419 unsigned long address,
1422 struct domain_pgtable pgtable;
1423 unsigned long flags;
1427 spin_lock_irqsave(&domain->lock, flags);
1429 amd_iommu_domain_get_pgtable(domain, &pgtable);
1431 if (address <= PM_LEVEL_SIZE(pgtable.mode))
1435 if (WARN_ON_ONCE(pgtable.mode == PAGE_MODE_6_LEVEL))
1438 pte = (void *)get_zeroed_page(gfp);
1442 *pte = PM_LEVEL_PDE(pgtable.mode, iommu_virt_to_phys(pgtable.root));
1446 update_and_flush_device_table(domain, &pgtable);
1447 domain_flush_complete(domain);
1450 * Device Table needs to be updated and flushed before the new root can
1453 root = amd_iommu_domain_encode_pgtable(pte, pgtable.mode);
1454 atomic64_set(&domain->pt_root, root);
1459 spin_unlock_irqrestore(&domain->lock, flags);
1464 static u64 *alloc_pte(struct protection_domain *domain,
1465 unsigned long address,
1466 unsigned long page_size,
1471 struct domain_pgtable pgtable;
1475 BUG_ON(!is_power_of_2(page_size));
1477 amd_iommu_domain_get_pgtable(domain, &pgtable);
1479 while (address > PM_LEVEL_SIZE(pgtable.mode)) {
1481 * Return an error if there is no memory to update the
1484 if (!increase_address_space(domain, address, gfp))
1487 /* Read new values to check if update was successful */
1488 amd_iommu_domain_get_pgtable(domain, &pgtable);
1492 level = pgtable.mode - 1;
1493 pte = &pgtable.root[PM_LEVEL_INDEX(level, address)];
1494 address = PAGE_SIZE_ALIGN(address, page_size);
1495 end_lvl = PAGE_SIZE_LEVEL(page_size);
1497 while (level > end_lvl) {
1502 pte_level = PM_PTE_LEVEL(__pte);
1505 * If we replace a series of large PTEs, we need
1506 * to tear down all of them.
1508 if (IOMMU_PTE_PRESENT(__pte) &&
1509 pte_level == PAGE_MODE_7_LEVEL) {
1510 unsigned long count, i;
1513 lpte = first_pte_l7(pte, NULL, &count);
1516 * Unmap the replicated PTEs that still match the
1517 * original large mapping
1519 for (i = 0; i < count; ++i)
1520 cmpxchg64(&lpte[i], __pte, 0ULL);
1526 if (!IOMMU_PTE_PRESENT(__pte) ||
1527 pte_level == PAGE_MODE_NONE) {
1528 page = (u64 *)get_zeroed_page(gfp);
1533 __npte = PM_LEVEL_PDE(level, iommu_virt_to_phys(page));
1535 /* pte could have been changed somewhere. */
1536 if (cmpxchg64(pte, __pte, __npte) != __pte)
1537 free_page((unsigned long)page);
1538 else if (IOMMU_PTE_PRESENT(__pte))
1544 /* No level skipping support yet */
1545 if (pte_level != level)
1550 pte = IOMMU_PTE_PAGE(__pte);
1552 if (pte_page && level == end_lvl)
1555 pte = &pte[PM_LEVEL_INDEX(level, address)];
1562 * This function checks if there is a PTE for a given dma address. If
1563 * there is one, it returns the pointer to it.
1565 static u64 *fetch_pte(struct protection_domain *domain,
1566 unsigned long address,
1567 unsigned long *page_size)
1569 struct domain_pgtable pgtable;
1575 amd_iommu_domain_get_pgtable(domain, &pgtable);
1577 if (address > PM_LEVEL_SIZE(pgtable.mode))
1580 level = pgtable.mode - 1;
1581 pte = &pgtable.root[PM_LEVEL_INDEX(level, address)];
1582 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1587 if (!IOMMU_PTE_PRESENT(*pte))
1591 if (PM_PTE_LEVEL(*pte) == 7 ||
1592 PM_PTE_LEVEL(*pte) == 0)
1595 /* No level skipping support yet */
1596 if (PM_PTE_LEVEL(*pte) != level)
1601 /* Walk to the next level */
1602 pte = IOMMU_PTE_PAGE(*pte);
1603 pte = &pte[PM_LEVEL_INDEX(level, address)];
1604 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1608 * If we have a series of large PTEs, make
1609 * sure to return a pointer to the first one.
1611 if (PM_PTE_LEVEL(*pte) == PAGE_MODE_7_LEVEL)
1612 pte = first_pte_l7(pte, page_size, NULL);
1617 static struct page *free_clear_pte(u64 *pte, u64 pteval, struct page *freelist)
1622 while (cmpxchg64(pte, pteval, 0) != pteval) {
1623 pr_warn("AMD-Vi: IOMMU pte changed since we read it\n");
1627 if (!IOMMU_PTE_PRESENT(pteval))
1630 pt = (unsigned long)IOMMU_PTE_PAGE(pteval);
1631 mode = IOMMU_PTE_MODE(pteval);
1633 return free_sub_pt(pt, mode, freelist);
1637 * Generic mapping functions. It maps a physical address into a DMA
1638 * address space. It allocates the page table pages if necessary.
1639 * In the future it can be extended to a generic mapping function
1640 * supporting all features of AMD IOMMU page tables like level skipping
1641 * and full 64 bit address spaces.
1643 static int iommu_map_page(struct protection_domain *dom,
1644 unsigned long bus_addr,
1645 unsigned long phys_addr,
1646 unsigned long page_size,
1650 struct page *freelist = NULL;
1651 bool updated = false;
1655 BUG_ON(!IS_ALIGNED(bus_addr, page_size));
1656 BUG_ON(!IS_ALIGNED(phys_addr, page_size));
1659 if (!(prot & IOMMU_PROT_MASK))
1662 count = PAGE_SIZE_PTE_COUNT(page_size);
1663 pte = alloc_pte(dom, bus_addr, page_size, NULL, gfp, &updated);
1669 for (i = 0; i < count; ++i)
1670 freelist = free_clear_pte(&pte[i], pte[i], freelist);
1672 if (freelist != NULL)
1676 __pte = PAGE_SIZE_PTE(__sme_set(phys_addr), page_size);
1677 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_PR | IOMMU_PTE_FC;
1679 __pte = __sme_set(phys_addr) | IOMMU_PTE_PR | IOMMU_PTE_FC;
1681 if (prot & IOMMU_PROT_IR)
1682 __pte |= IOMMU_PTE_IR;
1683 if (prot & IOMMU_PROT_IW)
1684 __pte |= IOMMU_PTE_IW;
1686 for (i = 0; i < count; ++i)
1693 unsigned long flags;
1695 spin_lock_irqsave(&dom->lock, flags);
1697 * Flush domain TLB(s) and wait for completion. Any Device-Table
1698 * Updates and flushing already happened in
1699 * increase_address_space().
1701 domain_flush_tlb_pde(dom);
1702 domain_flush_complete(dom);
1703 spin_unlock_irqrestore(&dom->lock, flags);
1706 /* Everything flushed out, free pages now */
1707 free_page_list(freelist);
1712 static unsigned long iommu_unmap_page(struct protection_domain *dom,
1713 unsigned long bus_addr,
1714 unsigned long page_size)
1716 unsigned long long unmapped;
1717 unsigned long unmap_size;
1720 BUG_ON(!is_power_of_2(page_size));
1724 while (unmapped < page_size) {
1726 pte = fetch_pte(dom, bus_addr, &unmap_size);
1731 count = PAGE_SIZE_PTE_COUNT(unmap_size);
1732 for (i = 0; i < count; i++)
1736 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1737 unmapped += unmap_size;
1740 BUG_ON(unmapped && !is_power_of_2(unmapped));
1745 /****************************************************************************
1747 * The next functions belong to the domain allocation. A domain is
1748 * allocated for every IOMMU as the default domain. If device isolation
1749 * is enabled, every device get its own domain. The most important thing
1750 * about domains is the page table mapping the DMA address space they
1753 ****************************************************************************/
1755 static u16 domain_id_alloc(void)
1759 spin_lock(&pd_bitmap_lock);
1760 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1762 if (id > 0 && id < MAX_DOMAIN_ID)
1763 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1766 spin_unlock(&pd_bitmap_lock);
1771 static void domain_id_free(int id)
1773 spin_lock(&pd_bitmap_lock);
1774 if (id > 0 && id < MAX_DOMAIN_ID)
1775 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1776 spin_unlock(&pd_bitmap_lock);
1779 static void free_gcr3_tbl_level1(u64 *tbl)
1784 for (i = 0; i < 512; ++i) {
1785 if (!(tbl[i] & GCR3_VALID))
1788 ptr = iommu_phys_to_virt(tbl[i] & PAGE_MASK);
1790 free_page((unsigned long)ptr);
1794 static void free_gcr3_tbl_level2(u64 *tbl)
1799 for (i = 0; i < 512; ++i) {
1800 if (!(tbl[i] & GCR3_VALID))
1803 ptr = iommu_phys_to_virt(tbl[i] & PAGE_MASK);
1805 free_gcr3_tbl_level1(ptr);
1809 static void free_gcr3_table(struct protection_domain *domain)
1811 if (domain->glx == 2)
1812 free_gcr3_tbl_level2(domain->gcr3_tbl);
1813 else if (domain->glx == 1)
1814 free_gcr3_tbl_level1(domain->gcr3_tbl);
1816 BUG_ON(domain->glx != 0);
1818 free_page((unsigned long)domain->gcr3_tbl);
1822 * Free a domain, only used if something went wrong in the
1823 * allocation path and we need to free an already allocated page table
1825 static void dma_ops_domain_free(struct protection_domain *domain)
1830 iommu_put_dma_cookie(&domain->domain);
1832 free_pagetable(domain);
1835 domain_id_free(domain->id);
1841 * Allocates a new protection domain usable for the dma_ops functions.
1842 * It also initializes the page table and the address allocator data
1843 * structures required for the dma_ops interface
1845 static struct protection_domain *dma_ops_domain_alloc(void)
1847 struct protection_domain *domain;
1850 domain = kzalloc(sizeof(struct protection_domain), GFP_KERNEL);
1854 if (protection_domain_init(domain))
1857 pt_root = (void *)get_zeroed_page(GFP_KERNEL);
1861 root = amd_iommu_domain_encode_pgtable(pt_root, PAGE_MODE_3_LEVEL);
1862 atomic64_set(&domain->pt_root, root);
1863 domain->flags = PD_DMA_OPS_MASK;
1865 if (iommu_get_dma_cookie(&domain->domain) == -ENOMEM)
1871 dma_ops_domain_free(domain);
1877 * little helper function to check whether a given protection domain is a
1880 static bool dma_ops_domain(struct protection_domain *domain)
1882 return domain->flags & PD_DMA_OPS_MASK;
1885 static void set_dte_entry(u16 devid, struct protection_domain *domain,
1886 struct domain_pgtable *pgtable,
1893 if (pgtable->mode != PAGE_MODE_NONE)
1894 pte_root = iommu_virt_to_phys(pgtable->root);
1896 pte_root |= (pgtable->mode & DEV_ENTRY_MODE_MASK)
1897 << DEV_ENTRY_MODE_SHIFT;
1898 pte_root |= DTE_FLAG_IR | DTE_FLAG_IW | DTE_FLAG_V | DTE_FLAG_TV;
1900 flags = amd_iommu_dev_table[devid].data[1];
1903 flags |= DTE_FLAG_IOTLB;
1906 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
1908 if (iommu_feature(iommu, FEATURE_EPHSUP))
1909 pte_root |= 1ULL << DEV_ENTRY_PPR;
1912 if (domain->flags & PD_IOMMUV2_MASK) {
1913 u64 gcr3 = iommu_virt_to_phys(domain->gcr3_tbl);
1914 u64 glx = domain->glx;
1917 pte_root |= DTE_FLAG_GV;
1918 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
1920 /* First mask out possible old values for GCR3 table */
1921 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
1924 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
1927 /* Encode GCR3 table into DTE */
1928 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
1931 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
1934 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
1938 flags &= ~DEV_DOMID_MASK;
1939 flags |= domain->id;
1941 old_domid = amd_iommu_dev_table[devid].data[1] & DEV_DOMID_MASK;
1942 amd_iommu_dev_table[devid].data[1] = flags;
1943 amd_iommu_dev_table[devid].data[0] = pte_root;
1946 * A kdump kernel might be replacing a domain ID that was copied from
1947 * the previous kernel--if so, it needs to flush the translation cache
1948 * entries for the old domain ID that is being overwritten
1951 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
1953 amd_iommu_flush_tlb_domid(iommu, old_domid);
1957 static void clear_dte_entry(u16 devid)
1959 /* remove entry from the device table seen by the hardware */
1960 amd_iommu_dev_table[devid].data[0] = DTE_FLAG_V | DTE_FLAG_TV;
1961 amd_iommu_dev_table[devid].data[1] &= DTE_FLAG_MASK;
1963 amd_iommu_apply_erratum_63(devid);
1966 static void do_attach(struct iommu_dev_data *dev_data,
1967 struct protection_domain *domain)
1969 struct domain_pgtable pgtable;
1970 struct amd_iommu *iommu;
1973 iommu = amd_iommu_rlookup_table[dev_data->devid];
1974 ats = dev_data->ats.enabled;
1976 /* Update data structures */
1977 dev_data->domain = domain;
1978 list_add(&dev_data->list, &domain->dev_list);
1980 /* Do reference counting */
1981 domain->dev_iommu[iommu->index] += 1;
1982 domain->dev_cnt += 1;
1984 /* Update device table */
1985 amd_iommu_domain_get_pgtable(domain, &pgtable);
1986 set_dte_entry(dev_data->devid, domain, &pgtable,
1987 ats, dev_data->iommu_v2);
1988 clone_aliases(dev_data->pdev);
1990 device_flush_dte(dev_data);
1993 static void do_detach(struct iommu_dev_data *dev_data)
1995 struct protection_domain *domain = dev_data->domain;
1996 struct amd_iommu *iommu;
1998 iommu = amd_iommu_rlookup_table[dev_data->devid];
2000 /* Update data structures */
2001 dev_data->domain = NULL;
2002 list_del(&dev_data->list);
2003 clear_dte_entry(dev_data->devid);
2004 clone_aliases(dev_data->pdev);
2006 /* Flush the DTE entry */
2007 device_flush_dte(dev_data);
2010 domain_flush_tlb_pde(domain);
2012 /* Wait for the flushes to finish */
2013 domain_flush_complete(domain);
2015 /* decrease reference counters - needs to happen after the flushes */
2016 domain->dev_iommu[iommu->index] -= 1;
2017 domain->dev_cnt -= 1;
2020 static void pdev_iommuv2_disable(struct pci_dev *pdev)
2022 pci_disable_ats(pdev);
2023 pci_disable_pri(pdev);
2024 pci_disable_pasid(pdev);
2027 /* FIXME: Change generic reset-function to do the same */
2028 static int pri_reset_while_enabled(struct pci_dev *pdev)
2033 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2037 pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
2038 control |= PCI_PRI_CTRL_RESET;
2039 pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
2044 static int pdev_iommuv2_enable(struct pci_dev *pdev)
2049 /* FIXME: Hardcode number of outstanding requests for now */
2051 if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
2053 reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
2055 /* Only allow access to user-accessible pages */
2056 ret = pci_enable_pasid(pdev, 0);
2060 /* First reset the PRI state of the device */
2061 ret = pci_reset_pri(pdev);
2066 ret = pci_enable_pri(pdev, reqs);
2071 ret = pri_reset_while_enabled(pdev);
2076 ret = pci_enable_ats(pdev, PAGE_SHIFT);
2083 pci_disable_pri(pdev);
2084 pci_disable_pasid(pdev);
2090 * If a device is not yet associated with a domain, this function makes the
2091 * device visible in the domain
2093 static int attach_device(struct device *dev,
2094 struct protection_domain *domain)
2096 struct iommu_dev_data *dev_data;
2097 struct pci_dev *pdev;
2098 unsigned long flags;
2101 spin_lock_irqsave(&domain->lock, flags);
2103 dev_data = get_dev_data(dev);
2105 spin_lock(&dev_data->lock);
2108 if (dev_data->domain != NULL)
2111 if (!dev_is_pci(dev))
2112 goto skip_ats_check;
2114 pdev = to_pci_dev(dev);
2115 if (domain->flags & PD_IOMMUV2_MASK) {
2116 struct iommu_domain *def_domain = iommu_get_dma_domain(dev);
2119 if (def_domain->type != IOMMU_DOMAIN_IDENTITY)
2122 if (dev_data->iommu_v2) {
2123 if (pdev_iommuv2_enable(pdev) != 0)
2126 dev_data->ats.enabled = true;
2127 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2128 dev_data->pri_tlp = pci_prg_resp_pasid_required(pdev);
2130 } else if (amd_iommu_iotlb_sup &&
2131 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
2132 dev_data->ats.enabled = true;
2133 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2139 do_attach(dev_data, domain);
2142 * We might boot into a crash-kernel here. The crashed kernel
2143 * left the caches in the IOMMU dirty. So we have to flush
2144 * here to evict all dirty stuff.
2146 domain_flush_tlb_pde(domain);
2148 domain_flush_complete(domain);
2151 spin_unlock(&dev_data->lock);
2153 spin_unlock_irqrestore(&domain->lock, flags);
2159 * Removes a device from a protection domain (with devtable_lock held)
2161 static void detach_device(struct device *dev)
2163 struct protection_domain *domain;
2164 struct iommu_dev_data *dev_data;
2165 unsigned long flags;
2167 dev_data = get_dev_data(dev);
2168 domain = dev_data->domain;
2170 spin_lock_irqsave(&domain->lock, flags);
2172 spin_lock(&dev_data->lock);
2175 * First check if the device is still attached. It might already
2176 * be detached from its domain because the generic
2177 * iommu_detach_group code detached it and we try again here in
2178 * our alias handling.
2180 if (WARN_ON(!dev_data->domain))
2183 do_detach(dev_data);
2185 if (!dev_is_pci(dev))
2188 if (domain->flags & PD_IOMMUV2_MASK && dev_data->iommu_v2)
2189 pdev_iommuv2_disable(to_pci_dev(dev));
2190 else if (dev_data->ats.enabled)
2191 pci_disable_ats(to_pci_dev(dev));
2193 dev_data->ats.enabled = false;
2196 spin_unlock(&dev_data->lock);
2198 spin_unlock_irqrestore(&domain->lock, flags);
2201 static struct iommu_device *amd_iommu_probe_device(struct device *dev)
2203 struct iommu_device *iommu_dev;
2204 struct amd_iommu *iommu;
2207 if (!check_device(dev))
2208 return ERR_PTR(-ENODEV);
2210 devid = get_device_id(dev);
2212 return ERR_PTR(devid);
2214 iommu = amd_iommu_rlookup_table[devid];
2216 if (get_dev_data(dev))
2217 return &iommu->iommu;
2219 ret = iommu_init_device(dev);
2221 if (ret != -ENOTSUPP)
2222 dev_err(dev, "Failed to initialize - trying to proceed anyway\n");
2223 iommu_dev = ERR_PTR(ret);
2224 iommu_ignore_device(dev);
2226 iommu_dev = &iommu->iommu;
2229 iommu_completion_wait(iommu);
2234 static void amd_iommu_probe_finalize(struct device *dev)
2236 struct iommu_domain *domain;
2238 /* Domains are initialized for this device - have a look what we ended up with */
2239 domain = iommu_get_domain_for_dev(dev);
2240 if (domain->type == IOMMU_DOMAIN_DMA)
2241 iommu_setup_dma_ops(dev, IOVA_START_PFN << PAGE_SHIFT, 0);
2244 static void amd_iommu_release_device(struct device *dev)
2246 struct amd_iommu *iommu;
2249 if (!check_device(dev))
2252 devid = get_device_id(dev);
2256 iommu = amd_iommu_rlookup_table[devid];
2258 amd_iommu_uninit_device(dev);
2259 iommu_completion_wait(iommu);
2262 static struct iommu_group *amd_iommu_device_group(struct device *dev)
2264 if (dev_is_pci(dev))
2265 return pci_device_group(dev);
2267 return acpihid_device_group(dev);
2270 static int amd_iommu_domain_get_attr(struct iommu_domain *domain,
2271 enum iommu_attr attr, void *data)
2273 switch (domain->type) {
2274 case IOMMU_DOMAIN_UNMANAGED:
2276 case IOMMU_DOMAIN_DMA:
2278 case DOMAIN_ATTR_DMA_USE_FLUSH_QUEUE:
2279 *(int *)data = !amd_iommu_unmap_flush;
2290 /*****************************************************************************
2292 * The next functions belong to the dma_ops mapping/unmapping code.
2294 *****************************************************************************/
2296 static void update_device_table(struct protection_domain *domain,
2297 struct domain_pgtable *pgtable)
2299 struct iommu_dev_data *dev_data;
2301 list_for_each_entry(dev_data, &domain->dev_list, list) {
2302 set_dte_entry(dev_data->devid, domain, pgtable,
2303 dev_data->ats.enabled, dev_data->iommu_v2);
2304 clone_aliases(dev_data->pdev);
2308 static void update_and_flush_device_table(struct protection_domain *domain,
2309 struct domain_pgtable *pgtable)
2311 update_device_table(domain, pgtable);
2312 domain_flush_devices(domain);
2315 static void update_domain(struct protection_domain *domain)
2317 struct domain_pgtable pgtable;
2319 /* Update device table */
2320 amd_iommu_domain_get_pgtable(domain, &pgtable);
2321 update_and_flush_device_table(domain, &pgtable);
2323 /* Flush domain TLB(s) and wait for completion */
2324 domain_flush_tlb_pde(domain);
2325 domain_flush_complete(domain);
2328 int __init amd_iommu_init_api(void)
2332 ret = iova_cache_get();
2336 err = bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
2339 #ifdef CONFIG_ARM_AMBA
2340 err = bus_set_iommu(&amba_bustype, &amd_iommu_ops);
2344 err = bus_set_iommu(&platform_bus_type, &amd_iommu_ops);
2351 int __init amd_iommu_init_dma_ops(void)
2353 swiotlb = (iommu_default_passthrough() || sme_me_mask) ? 1 : 0;
2355 if (amd_iommu_unmap_flush)
2356 pr_info("IO/TLB flush on unmap enabled\n");
2358 pr_info("Lazy IO/TLB flushing enabled\n");
2364 /*****************************************************************************
2366 * The following functions belong to the exported interface of AMD IOMMU
2368 * This interface allows access to lower level functions of the IOMMU
2369 * like protection domain handling and assignement of devices to domains
2370 * which is not possible with the dma_ops interface.
2372 *****************************************************************************/
2374 static void cleanup_domain(struct protection_domain *domain)
2376 struct iommu_dev_data *entry;
2377 unsigned long flags;
2379 spin_lock_irqsave(&domain->lock, flags);
2381 while (!list_empty(&domain->dev_list)) {
2382 entry = list_first_entry(&domain->dev_list,
2383 struct iommu_dev_data, list);
2384 BUG_ON(!entry->domain);
2388 spin_unlock_irqrestore(&domain->lock, flags);
2391 static void protection_domain_free(struct protection_domain *domain)
2397 domain_id_free(domain->id);
2402 static int protection_domain_init(struct protection_domain *domain)
2404 spin_lock_init(&domain->lock);
2405 domain->id = domain_id_alloc();
2408 INIT_LIST_HEAD(&domain->dev_list);
2413 static struct protection_domain *protection_domain_alloc(void)
2415 struct protection_domain *domain;
2417 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2421 if (protection_domain_init(domain))
2432 static struct iommu_domain *amd_iommu_domain_alloc(unsigned type)
2434 struct protection_domain *pdomain;
2438 case IOMMU_DOMAIN_UNMANAGED:
2439 pdomain = protection_domain_alloc();
2443 pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2445 protection_domain_free(pdomain);
2449 root = amd_iommu_domain_encode_pgtable(pt_root, PAGE_MODE_3_LEVEL);
2450 atomic64_set(&pdomain->pt_root, root);
2452 pdomain->domain.geometry.aperture_start = 0;
2453 pdomain->domain.geometry.aperture_end = ~0ULL;
2454 pdomain->domain.geometry.force_aperture = true;
2457 case IOMMU_DOMAIN_DMA:
2458 pdomain = dma_ops_domain_alloc();
2460 pr_err("Failed to allocate\n");
2464 case IOMMU_DOMAIN_IDENTITY:
2465 pdomain = protection_domain_alloc();
2469 atomic64_set(&pdomain->pt_root, PAGE_MODE_NONE);
2475 return &pdomain->domain;
2478 static void amd_iommu_domain_free(struct iommu_domain *dom)
2480 struct protection_domain *domain;
2481 struct domain_pgtable pgtable;
2483 domain = to_pdomain(dom);
2485 if (domain->dev_cnt > 0)
2486 cleanup_domain(domain);
2488 BUG_ON(domain->dev_cnt != 0);
2493 switch (dom->type) {
2494 case IOMMU_DOMAIN_DMA:
2495 /* Now release the domain */
2496 dma_ops_domain_free(domain);
2499 amd_iommu_domain_get_pgtable(domain, &pgtable);
2501 if (pgtable.mode != PAGE_MODE_NONE)
2502 free_pagetable(domain);
2504 if (domain->flags & PD_IOMMUV2_MASK)
2505 free_gcr3_table(domain);
2507 protection_domain_free(domain);
2512 static void amd_iommu_detach_device(struct iommu_domain *dom,
2515 struct iommu_dev_data *dev_data = dev->archdata.iommu;
2516 struct amd_iommu *iommu;
2519 if (!check_device(dev))
2522 devid = get_device_id(dev);
2526 if (dev_data->domain != NULL)
2529 iommu = amd_iommu_rlookup_table[devid];
2533 #ifdef CONFIG_IRQ_REMAP
2534 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) &&
2535 (dom->type == IOMMU_DOMAIN_UNMANAGED))
2536 dev_data->use_vapic = 0;
2539 iommu_completion_wait(iommu);
2542 static int amd_iommu_attach_device(struct iommu_domain *dom,
2545 struct protection_domain *domain = to_pdomain(dom);
2546 struct iommu_dev_data *dev_data;
2547 struct amd_iommu *iommu;
2550 if (!check_device(dev))
2553 dev_data = dev->archdata.iommu;
2554 dev_data->defer_attach = false;
2556 iommu = amd_iommu_rlookup_table[dev_data->devid];
2560 if (dev_data->domain)
2563 ret = attach_device(dev, domain);
2565 #ifdef CONFIG_IRQ_REMAP
2566 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
2567 if (dom->type == IOMMU_DOMAIN_UNMANAGED)
2568 dev_data->use_vapic = 1;
2570 dev_data->use_vapic = 0;
2574 iommu_completion_wait(iommu);
2579 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
2580 phys_addr_t paddr, size_t page_size, int iommu_prot,
2583 struct protection_domain *domain = to_pdomain(dom);
2584 struct domain_pgtable pgtable;
2588 amd_iommu_domain_get_pgtable(domain, &pgtable);
2589 if (pgtable.mode == PAGE_MODE_NONE)
2592 if (iommu_prot & IOMMU_READ)
2593 prot |= IOMMU_PROT_IR;
2594 if (iommu_prot & IOMMU_WRITE)
2595 prot |= IOMMU_PROT_IW;
2597 ret = iommu_map_page(domain, iova, paddr, page_size, prot, gfp);
2599 domain_flush_np_cache(domain, iova, page_size);
2604 static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
2606 struct iommu_iotlb_gather *gather)
2608 struct protection_domain *domain = to_pdomain(dom);
2609 struct domain_pgtable pgtable;
2611 amd_iommu_domain_get_pgtable(domain, &pgtable);
2612 if (pgtable.mode == PAGE_MODE_NONE)
2615 return iommu_unmap_page(domain, iova, page_size);
2618 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
2621 struct protection_domain *domain = to_pdomain(dom);
2622 unsigned long offset_mask, pte_pgsize;
2623 struct domain_pgtable pgtable;
2626 amd_iommu_domain_get_pgtable(domain, &pgtable);
2627 if (pgtable.mode == PAGE_MODE_NONE)
2630 pte = fetch_pte(domain, iova, &pte_pgsize);
2632 if (!pte || !IOMMU_PTE_PRESENT(*pte))
2635 offset_mask = pte_pgsize - 1;
2636 __pte = __sme_clr(*pte & PM_ADDR_MASK);
2638 return (__pte & ~offset_mask) | (iova & offset_mask);
2641 static bool amd_iommu_capable(enum iommu_cap cap)
2644 case IOMMU_CAP_CACHE_COHERENCY:
2646 case IOMMU_CAP_INTR_REMAP:
2647 return (irq_remapping_enabled == 1);
2648 case IOMMU_CAP_NOEXEC:
2657 static void amd_iommu_get_resv_regions(struct device *dev,
2658 struct list_head *head)
2660 struct iommu_resv_region *region;
2661 struct unity_map_entry *entry;
2664 devid = get_device_id(dev);
2668 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
2672 if (devid < entry->devid_start || devid > entry->devid_end)
2675 type = IOMMU_RESV_DIRECT;
2676 length = entry->address_end - entry->address_start;
2677 if (entry->prot & IOMMU_PROT_IR)
2679 if (entry->prot & IOMMU_PROT_IW)
2680 prot |= IOMMU_WRITE;
2681 if (entry->prot & IOMMU_UNITY_MAP_FLAG_EXCL_RANGE)
2682 /* Exclusion range */
2683 type = IOMMU_RESV_RESERVED;
2685 region = iommu_alloc_resv_region(entry->address_start,
2686 length, prot, type);
2688 dev_err(dev, "Out of memory allocating dm-regions\n");
2691 list_add_tail(®ion->list, head);
2694 region = iommu_alloc_resv_region(MSI_RANGE_START,
2695 MSI_RANGE_END - MSI_RANGE_START + 1,
2699 list_add_tail(®ion->list, head);
2701 region = iommu_alloc_resv_region(HT_RANGE_START,
2702 HT_RANGE_END - HT_RANGE_START + 1,
2703 0, IOMMU_RESV_RESERVED);
2706 list_add_tail(®ion->list, head);
2709 static bool amd_iommu_is_attach_deferred(struct iommu_domain *domain,
2712 struct iommu_dev_data *dev_data = dev->archdata.iommu;
2713 return dev_data->defer_attach;
2716 static void amd_iommu_flush_iotlb_all(struct iommu_domain *domain)
2718 struct protection_domain *dom = to_pdomain(domain);
2719 unsigned long flags;
2721 spin_lock_irqsave(&dom->lock, flags);
2722 domain_flush_tlb_pde(dom);
2723 domain_flush_complete(dom);
2724 spin_unlock_irqrestore(&dom->lock, flags);
2727 static void amd_iommu_iotlb_sync(struct iommu_domain *domain,
2728 struct iommu_iotlb_gather *gather)
2730 amd_iommu_flush_iotlb_all(domain);
2733 static int amd_iommu_def_domain_type(struct device *dev)
2735 struct iommu_dev_data *dev_data;
2737 dev_data = get_dev_data(dev);
2741 if (dev_data->iommu_v2)
2742 return IOMMU_DOMAIN_IDENTITY;
2747 const struct iommu_ops amd_iommu_ops = {
2748 .capable = amd_iommu_capable,
2749 .domain_alloc = amd_iommu_domain_alloc,
2750 .domain_free = amd_iommu_domain_free,
2751 .attach_dev = amd_iommu_attach_device,
2752 .detach_dev = amd_iommu_detach_device,
2753 .map = amd_iommu_map,
2754 .unmap = amd_iommu_unmap,
2755 .iova_to_phys = amd_iommu_iova_to_phys,
2756 .probe_device = amd_iommu_probe_device,
2757 .release_device = amd_iommu_release_device,
2758 .probe_finalize = amd_iommu_probe_finalize,
2759 .device_group = amd_iommu_device_group,
2760 .domain_get_attr = amd_iommu_domain_get_attr,
2761 .get_resv_regions = amd_iommu_get_resv_regions,
2762 .put_resv_regions = generic_iommu_put_resv_regions,
2763 .is_attach_deferred = amd_iommu_is_attach_deferred,
2764 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
2765 .flush_iotlb_all = amd_iommu_flush_iotlb_all,
2766 .iotlb_sync = amd_iommu_iotlb_sync,
2767 .def_domain_type = amd_iommu_def_domain_type,
2770 /*****************************************************************************
2772 * The next functions do a basic initialization of IOMMU for pass through
2775 * In passthrough mode the IOMMU is initialized and enabled but not used for
2776 * DMA-API translation.
2778 *****************************************************************************/
2780 /* IOMMUv2 specific functions */
2781 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
2783 return atomic_notifier_chain_register(&ppr_notifier, nb);
2785 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
2787 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
2789 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
2791 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
2793 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
2795 struct protection_domain *domain = to_pdomain(dom);
2796 struct domain_pgtable pgtable;
2797 unsigned long flags;
2800 spin_lock_irqsave(&domain->lock, flags);
2802 /* First save pgtable configuration*/
2803 amd_iommu_domain_get_pgtable(domain, &pgtable);
2805 /* Update data structure */
2806 pt_root = amd_iommu_domain_encode_pgtable(NULL, PAGE_MODE_NONE);
2807 atomic64_set(&domain->pt_root, pt_root);
2809 /* Make changes visible to IOMMUs */
2810 update_domain(domain);
2812 /* Restore old pgtable in domain->ptroot to free page-table */
2813 pt_root = amd_iommu_domain_encode_pgtable(pgtable.root, pgtable.mode);
2814 atomic64_set(&domain->pt_root, pt_root);
2816 /* Page-table is not visible to IOMMU anymore, so free it */
2817 free_pagetable(domain);
2819 spin_unlock_irqrestore(&domain->lock, flags);
2821 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
2823 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
2825 struct protection_domain *domain = to_pdomain(dom);
2826 unsigned long flags;
2829 if (pasids <= 0 || pasids > (PASID_MASK + 1))
2832 /* Number of GCR3 table levels required */
2833 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
2836 if (levels > amd_iommu_max_glx_val)
2839 spin_lock_irqsave(&domain->lock, flags);
2842 * Save us all sanity checks whether devices already in the
2843 * domain support IOMMUv2. Just force that the domain has no
2844 * devices attached when it is switched into IOMMUv2 mode.
2847 if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
2851 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
2852 if (domain->gcr3_tbl == NULL)
2855 domain->glx = levels;
2856 domain->flags |= PD_IOMMUV2_MASK;
2858 update_domain(domain);
2863 spin_unlock_irqrestore(&domain->lock, flags);
2867 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
2869 static int __flush_pasid(struct protection_domain *domain, int pasid,
2870 u64 address, bool size)
2872 struct iommu_dev_data *dev_data;
2873 struct iommu_cmd cmd;
2876 if (!(domain->flags & PD_IOMMUV2_MASK))
2879 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
2882 * IOMMU TLB needs to be flushed before Device TLB to
2883 * prevent device TLB refill from IOMMU TLB
2885 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
2886 if (domain->dev_iommu[i] == 0)
2889 ret = iommu_queue_command(amd_iommus[i], &cmd);
2894 /* Wait until IOMMU TLB flushes are complete */
2895 domain_flush_complete(domain);
2897 /* Now flush device TLBs */
2898 list_for_each_entry(dev_data, &domain->dev_list, list) {
2899 struct amd_iommu *iommu;
2903 There might be non-IOMMUv2 capable devices in an IOMMUv2
2906 if (!dev_data->ats.enabled)
2909 qdep = dev_data->ats.qdep;
2910 iommu = amd_iommu_rlookup_table[dev_data->devid];
2912 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
2913 qdep, address, size);
2915 ret = iommu_queue_command(iommu, &cmd);
2920 /* Wait until all device TLBs are flushed */
2921 domain_flush_complete(domain);
2930 static int __amd_iommu_flush_page(struct protection_domain *domain, int pasid,
2933 return __flush_pasid(domain, pasid, address, false);
2936 int amd_iommu_flush_page(struct iommu_domain *dom, int pasid,
2939 struct protection_domain *domain = to_pdomain(dom);
2940 unsigned long flags;
2943 spin_lock_irqsave(&domain->lock, flags);
2944 ret = __amd_iommu_flush_page(domain, pasid, address);
2945 spin_unlock_irqrestore(&domain->lock, flags);
2949 EXPORT_SYMBOL(amd_iommu_flush_page);
2951 static int __amd_iommu_flush_tlb(struct protection_domain *domain, int pasid)
2953 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
2957 int amd_iommu_flush_tlb(struct iommu_domain *dom, int pasid)
2959 struct protection_domain *domain = to_pdomain(dom);
2960 unsigned long flags;
2963 spin_lock_irqsave(&domain->lock, flags);
2964 ret = __amd_iommu_flush_tlb(domain, pasid);
2965 spin_unlock_irqrestore(&domain->lock, flags);
2969 EXPORT_SYMBOL(amd_iommu_flush_tlb);
2971 static u64 *__get_gcr3_pte(u64 *root, int level, int pasid, bool alloc)
2978 index = (pasid >> (9 * level)) & 0x1ff;
2984 if (!(*pte & GCR3_VALID)) {
2988 root = (void *)get_zeroed_page(GFP_ATOMIC);
2992 *pte = iommu_virt_to_phys(root) | GCR3_VALID;
2995 root = iommu_phys_to_virt(*pte & PAGE_MASK);
3003 static int __set_gcr3(struct protection_domain *domain, int pasid,
3006 struct domain_pgtable pgtable;
3009 amd_iommu_domain_get_pgtable(domain, &pgtable);
3010 if (pgtable.mode != PAGE_MODE_NONE)
3013 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3017 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3019 return __amd_iommu_flush_tlb(domain, pasid);
3022 static int __clear_gcr3(struct protection_domain *domain, int pasid)
3024 struct domain_pgtable pgtable;
3027 amd_iommu_domain_get_pgtable(domain, &pgtable);
3028 if (pgtable.mode != PAGE_MODE_NONE)
3031 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3037 return __amd_iommu_flush_tlb(domain, pasid);
3040 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, int pasid,
3043 struct protection_domain *domain = to_pdomain(dom);
3044 unsigned long flags;
3047 spin_lock_irqsave(&domain->lock, flags);
3048 ret = __set_gcr3(domain, pasid, cr3);
3049 spin_unlock_irqrestore(&domain->lock, flags);
3053 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3055 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, int pasid)
3057 struct protection_domain *domain = to_pdomain(dom);
3058 unsigned long flags;
3061 spin_lock_irqsave(&domain->lock, flags);
3062 ret = __clear_gcr3(domain, pasid);
3063 spin_unlock_irqrestore(&domain->lock, flags);
3067 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3069 int amd_iommu_complete_ppr(struct pci_dev *pdev, int pasid,
3070 int status, int tag)
3072 struct iommu_dev_data *dev_data;
3073 struct amd_iommu *iommu;
3074 struct iommu_cmd cmd;
3076 dev_data = get_dev_data(&pdev->dev);
3077 iommu = amd_iommu_rlookup_table[dev_data->devid];
3079 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3080 tag, dev_data->pri_tlp);
3082 return iommu_queue_command(iommu, &cmd);
3084 EXPORT_SYMBOL(amd_iommu_complete_ppr);
3086 struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3088 struct protection_domain *pdomain;
3089 struct iommu_domain *io_domain;
3090 struct device *dev = &pdev->dev;
3092 if (!check_device(dev))
3095 pdomain = get_dev_data(dev)->domain;
3096 if (pdomain == NULL && get_dev_data(dev)->defer_attach) {
3097 get_dev_data(dev)->defer_attach = false;
3098 io_domain = iommu_get_domain_for_dev(dev);
3099 pdomain = to_pdomain(io_domain);
3100 attach_device(dev, pdomain);
3102 if (pdomain == NULL)
3105 if (!dma_ops_domain(pdomain))
3108 /* Only return IOMMUv2 domains */
3109 if (!(pdomain->flags & PD_IOMMUV2_MASK))
3112 return &pdomain->domain;
3114 EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3116 void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3118 struct iommu_dev_data *dev_data;
3120 if (!amd_iommu_v2_supported())
3123 dev_data = get_dev_data(&pdev->dev);
3124 dev_data->errata |= (1 << erratum);
3126 EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3128 int amd_iommu_device_info(struct pci_dev *pdev,
3129 struct amd_iommu_device_info *info)
3134 if (pdev == NULL || info == NULL)
3137 if (!amd_iommu_v2_supported())
3140 memset(info, 0, sizeof(*info));
3142 if (!pci_ats_disabled()) {
3143 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS);
3145 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3148 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3150 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3152 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3156 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3157 max_pasids = min(max_pasids, (1 << 20));
3159 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3160 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3162 features = pci_pasid_features(pdev);
3163 if (features & PCI_PASID_CAP_EXEC)
3164 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3165 if (features & PCI_PASID_CAP_PRIV)
3166 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3171 EXPORT_SYMBOL(amd_iommu_device_info);
3173 #ifdef CONFIG_IRQ_REMAP
3175 /*****************************************************************************
3177 * Interrupt Remapping Implementation
3179 *****************************************************************************/
3181 static struct irq_chip amd_ir_chip;
3182 static DEFINE_SPINLOCK(iommu_table_lock);
3184 static void set_dte_irq_entry(u16 devid, struct irq_remap_table *table)
3188 dte = amd_iommu_dev_table[devid].data[2];
3189 dte &= ~DTE_IRQ_PHYS_ADDR_MASK;
3190 dte |= iommu_virt_to_phys(table->table);
3191 dte |= DTE_IRQ_REMAP_INTCTL;
3192 dte |= DTE_IRQ_TABLE_LEN;
3193 dte |= DTE_IRQ_REMAP_ENABLE;
3195 amd_iommu_dev_table[devid].data[2] = dte;
3198 static struct irq_remap_table *get_irq_table(u16 devid)
3200 struct irq_remap_table *table;
3202 if (WARN_ONCE(!amd_iommu_rlookup_table[devid],
3203 "%s: no iommu for devid %x\n", __func__, devid))
3206 table = irq_lookup_table[devid];
3207 if (WARN_ONCE(!table, "%s: no table for devid %x\n", __func__, devid))
3213 static struct irq_remap_table *__alloc_irq_table(void)
3215 struct irq_remap_table *table;
3217 table = kzalloc(sizeof(*table), GFP_KERNEL);
3221 table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_KERNEL);
3222 if (!table->table) {
3226 raw_spin_lock_init(&table->lock);
3228 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
3229 memset(table->table, 0,
3230 MAX_IRQS_PER_TABLE * sizeof(u32));
3232 memset(table->table, 0,
3233 (MAX_IRQS_PER_TABLE * (sizeof(u64) * 2)));
3237 static void set_remap_table_entry(struct amd_iommu *iommu, u16 devid,
3238 struct irq_remap_table *table)
3240 irq_lookup_table[devid] = table;
3241 set_dte_irq_entry(devid, table);
3242 iommu_flush_dte(iommu, devid);
3245 static int set_remap_table_entry_alias(struct pci_dev *pdev, u16 alias,
3248 struct irq_remap_table *table = data;
3250 irq_lookup_table[alias] = table;
3251 set_dte_irq_entry(alias, table);
3253 iommu_flush_dte(amd_iommu_rlookup_table[alias], alias);
3258 static struct irq_remap_table *alloc_irq_table(u16 devid, struct pci_dev *pdev)
3260 struct irq_remap_table *table = NULL;
3261 struct irq_remap_table *new_table = NULL;
3262 struct amd_iommu *iommu;
3263 unsigned long flags;
3266 spin_lock_irqsave(&iommu_table_lock, flags);
3268 iommu = amd_iommu_rlookup_table[devid];
3272 table = irq_lookup_table[devid];
3276 alias = amd_iommu_alias_table[devid];
3277 table = irq_lookup_table[alias];
3279 set_remap_table_entry(iommu, devid, table);
3282 spin_unlock_irqrestore(&iommu_table_lock, flags);
3284 /* Nothing there yet, allocate new irq remapping table */
3285 new_table = __alloc_irq_table();
3289 spin_lock_irqsave(&iommu_table_lock, flags);
3291 table = irq_lookup_table[devid];
3295 table = irq_lookup_table[alias];
3297 set_remap_table_entry(iommu, devid, table);
3305 pci_for_each_dma_alias(pdev, set_remap_table_entry_alias,
3308 set_remap_table_entry(iommu, devid, table);
3311 set_remap_table_entry(iommu, alias, table);
3314 iommu_completion_wait(iommu);
3317 spin_unlock_irqrestore(&iommu_table_lock, flags);
3320 kmem_cache_free(amd_iommu_irq_cache, new_table->table);
3326 static int alloc_irq_index(u16 devid, int count, bool align,
3327 struct pci_dev *pdev)
3329 struct irq_remap_table *table;
3330 int index, c, alignment = 1;
3331 unsigned long flags;
3332 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
3337 table = alloc_irq_table(devid, pdev);
3342 alignment = roundup_pow_of_two(count);
3344 raw_spin_lock_irqsave(&table->lock, flags);
3346 /* Scan table for free entries */
3347 for (index = ALIGN(table->min_index, alignment), c = 0;
3348 index < MAX_IRQS_PER_TABLE;) {
3349 if (!iommu->irte_ops->is_allocated(table, index)) {
3353 index = ALIGN(index + 1, alignment);
3359 iommu->irte_ops->set_allocated(table, index - c + 1);
3371 raw_spin_unlock_irqrestore(&table->lock, flags);
3376 static int modify_irte_ga(u16 devid, int index, struct irte_ga *irte,
3377 struct amd_ir_data *data)
3379 struct irq_remap_table *table;
3380 struct amd_iommu *iommu;
3381 unsigned long flags;
3382 struct irte_ga *entry;
3384 iommu = amd_iommu_rlookup_table[devid];
3388 table = get_irq_table(devid);
3392 raw_spin_lock_irqsave(&table->lock, flags);
3394 entry = (struct irte_ga *)table->table;
3395 entry = &entry[index];
3396 entry->lo.fields_remap.valid = 0;
3397 entry->hi.val = irte->hi.val;
3398 entry->lo.val = irte->lo.val;
3399 entry->lo.fields_remap.valid = 1;
3403 raw_spin_unlock_irqrestore(&table->lock, flags);
3405 iommu_flush_irt(iommu, devid);
3406 iommu_completion_wait(iommu);
3411 static int modify_irte(u16 devid, int index, union irte *irte)
3413 struct irq_remap_table *table;
3414 struct amd_iommu *iommu;
3415 unsigned long flags;
3417 iommu = amd_iommu_rlookup_table[devid];
3421 table = get_irq_table(devid);
3425 raw_spin_lock_irqsave(&table->lock, flags);
3426 table->table[index] = irte->val;
3427 raw_spin_unlock_irqrestore(&table->lock, flags);
3429 iommu_flush_irt(iommu, devid);
3430 iommu_completion_wait(iommu);
3435 static void free_irte(u16 devid, int index)
3437 struct irq_remap_table *table;
3438 struct amd_iommu *iommu;
3439 unsigned long flags;
3441 iommu = amd_iommu_rlookup_table[devid];
3445 table = get_irq_table(devid);
3449 raw_spin_lock_irqsave(&table->lock, flags);
3450 iommu->irte_ops->clear_allocated(table, index);
3451 raw_spin_unlock_irqrestore(&table->lock, flags);
3453 iommu_flush_irt(iommu, devid);
3454 iommu_completion_wait(iommu);
3457 static void irte_prepare(void *entry,
3458 u32 delivery_mode, u32 dest_mode,
3459 u8 vector, u32 dest_apicid, int devid)
3461 union irte *irte = (union irte *) entry;
3464 irte->fields.vector = vector;
3465 irte->fields.int_type = delivery_mode;
3466 irte->fields.destination = dest_apicid;
3467 irte->fields.dm = dest_mode;
3468 irte->fields.valid = 1;
3471 static void irte_ga_prepare(void *entry,
3472 u32 delivery_mode, u32 dest_mode,
3473 u8 vector, u32 dest_apicid, int devid)
3475 struct irte_ga *irte = (struct irte_ga *) entry;
3479 irte->lo.fields_remap.int_type = delivery_mode;
3480 irte->lo.fields_remap.dm = dest_mode;
3481 irte->hi.fields.vector = vector;
3482 irte->lo.fields_remap.destination = APICID_TO_IRTE_DEST_LO(dest_apicid);
3483 irte->hi.fields.destination = APICID_TO_IRTE_DEST_HI(dest_apicid);
3484 irte->lo.fields_remap.valid = 1;
3487 static void irte_activate(void *entry, u16 devid, u16 index)
3489 union irte *irte = (union irte *) entry;
3491 irte->fields.valid = 1;
3492 modify_irte(devid, index, irte);
3495 static void irte_ga_activate(void *entry, u16 devid, u16 index)
3497 struct irte_ga *irte = (struct irte_ga *) entry;
3499 irte->lo.fields_remap.valid = 1;
3500 modify_irte_ga(devid, index, irte, NULL);
3503 static void irte_deactivate(void *entry, u16 devid, u16 index)
3505 union irte *irte = (union irte *) entry;
3507 irte->fields.valid = 0;
3508 modify_irte(devid, index, irte);
3511 static void irte_ga_deactivate(void *entry, u16 devid, u16 index)
3513 struct irte_ga *irte = (struct irte_ga *) entry;
3515 irte->lo.fields_remap.valid = 0;
3516 modify_irte_ga(devid, index, irte, NULL);
3519 static void irte_set_affinity(void *entry, u16 devid, u16 index,
3520 u8 vector, u32 dest_apicid)
3522 union irte *irte = (union irte *) entry;
3524 irte->fields.vector = vector;
3525 irte->fields.destination = dest_apicid;
3526 modify_irte(devid, index, irte);
3529 static void irte_ga_set_affinity(void *entry, u16 devid, u16 index,
3530 u8 vector, u32 dest_apicid)
3532 struct irte_ga *irte = (struct irte_ga *) entry;
3534 if (!irte->lo.fields_remap.guest_mode) {
3535 irte->hi.fields.vector = vector;
3536 irte->lo.fields_remap.destination =
3537 APICID_TO_IRTE_DEST_LO(dest_apicid);
3538 irte->hi.fields.destination =
3539 APICID_TO_IRTE_DEST_HI(dest_apicid);
3540 modify_irte_ga(devid, index, irte, NULL);
3544 #define IRTE_ALLOCATED (~1U)
3545 static void irte_set_allocated(struct irq_remap_table *table, int index)
3547 table->table[index] = IRTE_ALLOCATED;
3550 static void irte_ga_set_allocated(struct irq_remap_table *table, int index)
3552 struct irte_ga *ptr = (struct irte_ga *)table->table;
3553 struct irte_ga *irte = &ptr[index];
3555 memset(&irte->lo.val, 0, sizeof(u64));
3556 memset(&irte->hi.val, 0, sizeof(u64));
3557 irte->hi.fields.vector = 0xff;
3560 static bool irte_is_allocated(struct irq_remap_table *table, int index)
3562 union irte *ptr = (union irte *)table->table;
3563 union irte *irte = &ptr[index];
3565 return irte->val != 0;
3568 static bool irte_ga_is_allocated(struct irq_remap_table *table, int index)
3570 struct irte_ga *ptr = (struct irte_ga *)table->table;
3571 struct irte_ga *irte = &ptr[index];
3573 return irte->hi.fields.vector != 0;
3576 static void irte_clear_allocated(struct irq_remap_table *table, int index)
3578 table->table[index] = 0;
3581 static void irte_ga_clear_allocated(struct irq_remap_table *table, int index)
3583 struct irte_ga *ptr = (struct irte_ga *)table->table;
3584 struct irte_ga *irte = &ptr[index];
3586 memset(&irte->lo.val, 0, sizeof(u64));
3587 memset(&irte->hi.val, 0, sizeof(u64));
3590 static int get_devid(struct irq_alloc_info *info)
3594 switch (info->type) {
3595 case X86_IRQ_ALLOC_TYPE_IOAPIC:
3596 devid = get_ioapic_devid(info->ioapic_id);
3598 case X86_IRQ_ALLOC_TYPE_HPET:
3599 devid = get_hpet_devid(info->hpet_id);
3601 case X86_IRQ_ALLOC_TYPE_MSI:
3602 case X86_IRQ_ALLOC_TYPE_MSIX:
3603 devid = get_device_id(&info->msi_dev->dev);
3613 static struct irq_domain *get_ir_irq_domain(struct irq_alloc_info *info)
3615 struct amd_iommu *iommu;
3621 devid = get_devid(info);
3623 iommu = amd_iommu_rlookup_table[devid];
3625 return iommu->ir_domain;
3631 static struct irq_domain *get_irq_domain(struct irq_alloc_info *info)
3633 struct amd_iommu *iommu;
3639 switch (info->type) {
3640 case X86_IRQ_ALLOC_TYPE_MSI:
3641 case X86_IRQ_ALLOC_TYPE_MSIX:
3642 devid = get_device_id(&info->msi_dev->dev);
3646 iommu = amd_iommu_rlookup_table[devid];
3648 return iommu->msi_domain;
3657 struct irq_remap_ops amd_iommu_irq_ops = {
3658 .prepare = amd_iommu_prepare,
3659 .enable = amd_iommu_enable,
3660 .disable = amd_iommu_disable,
3661 .reenable = amd_iommu_reenable,
3662 .enable_faulting = amd_iommu_enable_faulting,
3663 .get_ir_irq_domain = get_ir_irq_domain,
3664 .get_irq_domain = get_irq_domain,
3667 static void irq_remapping_prepare_irte(struct amd_ir_data *data,
3668 struct irq_cfg *irq_cfg,
3669 struct irq_alloc_info *info,
3670 int devid, int index, int sub_handle)
3672 struct irq_2_irte *irte_info = &data->irq_2_irte;
3673 struct msi_msg *msg = &data->msi_entry;
3674 struct IO_APIC_route_entry *entry;
3675 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
3680 data->irq_2_irte.devid = devid;
3681 data->irq_2_irte.index = index + sub_handle;
3682 iommu->irte_ops->prepare(data->entry, apic->irq_delivery_mode,
3683 apic->irq_dest_mode, irq_cfg->vector,
3684 irq_cfg->dest_apicid, devid);
3686 switch (info->type) {
3687 case X86_IRQ_ALLOC_TYPE_IOAPIC:
3688 /* Setup IOAPIC entry */
3689 entry = info->ioapic_entry;
3690 info->ioapic_entry = NULL;
3691 memset(entry, 0, sizeof(*entry));
3692 entry->vector = index;
3694 entry->trigger = info->ioapic_trigger;
3695 entry->polarity = info->ioapic_polarity;
3696 /* Mask level triggered irqs. */
3697 if (info->ioapic_trigger)
3701 case X86_IRQ_ALLOC_TYPE_HPET:
3702 case X86_IRQ_ALLOC_TYPE_MSI:
3703 case X86_IRQ_ALLOC_TYPE_MSIX:
3704 msg->address_hi = MSI_ADDR_BASE_HI;
3705 msg->address_lo = MSI_ADDR_BASE_LO;
3706 msg->data = irte_info->index;
3715 struct amd_irte_ops irte_32_ops = {
3716 .prepare = irte_prepare,
3717 .activate = irte_activate,
3718 .deactivate = irte_deactivate,
3719 .set_affinity = irte_set_affinity,
3720 .set_allocated = irte_set_allocated,
3721 .is_allocated = irte_is_allocated,
3722 .clear_allocated = irte_clear_allocated,
3725 struct amd_irte_ops irte_128_ops = {
3726 .prepare = irte_ga_prepare,
3727 .activate = irte_ga_activate,
3728 .deactivate = irte_ga_deactivate,
3729 .set_affinity = irte_ga_set_affinity,
3730 .set_allocated = irte_ga_set_allocated,
3731 .is_allocated = irte_ga_is_allocated,
3732 .clear_allocated = irte_ga_clear_allocated,
3735 static int irq_remapping_alloc(struct irq_domain *domain, unsigned int virq,
3736 unsigned int nr_irqs, void *arg)
3738 struct irq_alloc_info *info = arg;
3739 struct irq_data *irq_data;
3740 struct amd_ir_data *data = NULL;
3741 struct irq_cfg *cfg;
3747 if (nr_irqs > 1 && info->type != X86_IRQ_ALLOC_TYPE_MSI &&
3748 info->type != X86_IRQ_ALLOC_TYPE_MSIX)
3752 * With IRQ remapping enabled, don't need contiguous CPU vectors
3753 * to support multiple MSI interrupts.
3755 if (info->type == X86_IRQ_ALLOC_TYPE_MSI)
3756 info->flags &= ~X86_IRQ_ALLOC_CONTIGUOUS_VECTORS;
3758 devid = get_devid(info);
3762 ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
3766 if (info->type == X86_IRQ_ALLOC_TYPE_IOAPIC) {
3767 struct irq_remap_table *table;
3768 struct amd_iommu *iommu;
3770 table = alloc_irq_table(devid, NULL);
3772 if (!table->min_index) {
3774 * Keep the first 32 indexes free for IOAPIC
3777 table->min_index = 32;
3778 iommu = amd_iommu_rlookup_table[devid];
3779 for (i = 0; i < 32; ++i)
3780 iommu->irte_ops->set_allocated(table, i);
3782 WARN_ON(table->min_index != 32);
3783 index = info->ioapic_pin;
3787 } else if (info->type == X86_IRQ_ALLOC_TYPE_MSI ||
3788 info->type == X86_IRQ_ALLOC_TYPE_MSIX) {
3789 bool align = (info->type == X86_IRQ_ALLOC_TYPE_MSI);
3791 index = alloc_irq_index(devid, nr_irqs, align, info->msi_dev);
3793 index = alloc_irq_index(devid, nr_irqs, false, NULL);
3797 pr_warn("Failed to allocate IRTE\n");
3799 goto out_free_parent;
3802 for (i = 0; i < nr_irqs; i++) {
3803 irq_data = irq_domain_get_irq_data(domain, virq + i);
3804 cfg = irqd_cfg(irq_data);
3805 if (!irq_data || !cfg) {
3811 data = kzalloc(sizeof(*data), GFP_KERNEL);
3815 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
3816 data->entry = kzalloc(sizeof(union irte), GFP_KERNEL);
3818 data->entry = kzalloc(sizeof(struct irte_ga),
3825 irq_data->hwirq = (devid << 16) + i;
3826 irq_data->chip_data = data;
3827 irq_data->chip = &amd_ir_chip;
3828 irq_remapping_prepare_irte(data, cfg, info, devid, index, i);
3829 irq_set_status_flags(virq + i, IRQ_MOVE_PCNTXT);
3835 for (i--; i >= 0; i--) {
3836 irq_data = irq_domain_get_irq_data(domain, virq + i);
3838 kfree(irq_data->chip_data);
3840 for (i = 0; i < nr_irqs; i++)
3841 free_irte(devid, index + i);
3843 irq_domain_free_irqs_common(domain, virq, nr_irqs);
3847 static void irq_remapping_free(struct irq_domain *domain, unsigned int virq,
3848 unsigned int nr_irqs)
3850 struct irq_2_irte *irte_info;
3851 struct irq_data *irq_data;
3852 struct amd_ir_data *data;
3855 for (i = 0; i < nr_irqs; i++) {
3856 irq_data = irq_domain_get_irq_data(domain, virq + i);
3857 if (irq_data && irq_data->chip_data) {
3858 data = irq_data->chip_data;
3859 irte_info = &data->irq_2_irte;
3860 free_irte(irte_info->devid, irte_info->index);
3865 irq_domain_free_irqs_common(domain, virq, nr_irqs);
3868 static void amd_ir_update_irte(struct irq_data *irqd, struct amd_iommu *iommu,
3869 struct amd_ir_data *ir_data,
3870 struct irq_2_irte *irte_info,
3871 struct irq_cfg *cfg);
3873 static int irq_remapping_activate(struct irq_domain *domain,
3874 struct irq_data *irq_data, bool reserve)
3876 struct amd_ir_data *data = irq_data->chip_data;
3877 struct irq_2_irte *irte_info = &data->irq_2_irte;
3878 struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
3879 struct irq_cfg *cfg = irqd_cfg(irq_data);
3884 iommu->irte_ops->activate(data->entry, irte_info->devid,
3886 amd_ir_update_irte(irq_data, iommu, data, irte_info, cfg);
3890 static void irq_remapping_deactivate(struct irq_domain *domain,
3891 struct irq_data *irq_data)
3893 struct amd_ir_data *data = irq_data->chip_data;
3894 struct irq_2_irte *irte_info = &data->irq_2_irte;
3895 struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
3898 iommu->irte_ops->deactivate(data->entry, irte_info->devid,
3902 static const struct irq_domain_ops amd_ir_domain_ops = {
3903 .alloc = irq_remapping_alloc,
3904 .free = irq_remapping_free,
3905 .activate = irq_remapping_activate,
3906 .deactivate = irq_remapping_deactivate,
3909 int amd_iommu_activate_guest_mode(void *data)
3911 struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
3912 struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
3914 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
3915 !entry || entry->lo.fields_vapic.guest_mode)
3921 entry->lo.fields_vapic.guest_mode = 1;
3922 entry->lo.fields_vapic.ga_log_intr = 1;
3923 entry->hi.fields.ga_root_ptr = ir_data->ga_root_ptr;
3924 entry->hi.fields.vector = ir_data->ga_vector;
3925 entry->lo.fields_vapic.ga_tag = ir_data->ga_tag;
3927 return modify_irte_ga(ir_data->irq_2_irte.devid,
3928 ir_data->irq_2_irte.index, entry, ir_data);
3930 EXPORT_SYMBOL(amd_iommu_activate_guest_mode);
3932 int amd_iommu_deactivate_guest_mode(void *data)
3934 struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
3935 struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
3936 struct irq_cfg *cfg = ir_data->cfg;
3938 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
3939 !entry || !entry->lo.fields_vapic.guest_mode)
3945 entry->lo.fields_remap.dm = apic->irq_dest_mode;
3946 entry->lo.fields_remap.int_type = apic->irq_delivery_mode;
3947 entry->hi.fields.vector = cfg->vector;
3948 entry->lo.fields_remap.destination =
3949 APICID_TO_IRTE_DEST_LO(cfg->dest_apicid);
3950 entry->hi.fields.destination =
3951 APICID_TO_IRTE_DEST_HI(cfg->dest_apicid);
3953 return modify_irte_ga(ir_data->irq_2_irte.devid,
3954 ir_data->irq_2_irte.index, entry, ir_data);
3956 EXPORT_SYMBOL(amd_iommu_deactivate_guest_mode);
3958 static int amd_ir_set_vcpu_affinity(struct irq_data *data, void *vcpu_info)
3961 struct amd_iommu *iommu;
3962 struct amd_iommu_pi_data *pi_data = vcpu_info;
3963 struct vcpu_data *vcpu_pi_info = pi_data->vcpu_data;
3964 struct amd_ir_data *ir_data = data->chip_data;
3965 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
3966 struct iommu_dev_data *dev_data = search_dev_data(irte_info->devid);
3969 * This device has never been set up for guest mode.
3970 * we should not modify the IRTE
3972 if (!dev_data || !dev_data->use_vapic)
3975 ir_data->cfg = irqd_cfg(data);
3976 pi_data->ir_data = ir_data;
3979 * SVM tries to set up for VAPIC mode, but we are in
3980 * legacy mode. So, we force legacy mode instead.
3982 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
3983 pr_debug("%s: Fall back to using intr legacy remap\n",
3985 pi_data->is_guest_mode = false;
3988 iommu = amd_iommu_rlookup_table[irte_info->devid];
3992 pi_data->prev_ga_tag = ir_data->cached_ga_tag;
3993 if (pi_data->is_guest_mode) {
3994 ir_data->ga_root_ptr = (pi_data->base >> 12);
3995 ir_data->ga_vector = vcpu_pi_info->vector;
3996 ir_data->ga_tag = pi_data->ga_tag;
3997 ret = amd_iommu_activate_guest_mode(ir_data);
3999 ir_data->cached_ga_tag = pi_data->ga_tag;
4001 ret = amd_iommu_deactivate_guest_mode(ir_data);
4004 * This communicates the ga_tag back to the caller
4005 * so that it can do all the necessary clean up.
4008 ir_data->cached_ga_tag = 0;
4015 static void amd_ir_update_irte(struct irq_data *irqd, struct amd_iommu *iommu,
4016 struct amd_ir_data *ir_data,
4017 struct irq_2_irte *irte_info,
4018 struct irq_cfg *cfg)
4022 * Atomically updates the IRTE with the new destination, vector
4023 * and flushes the interrupt entry cache.
4025 iommu->irte_ops->set_affinity(ir_data->entry, irte_info->devid,
4026 irte_info->index, cfg->vector,
4030 static int amd_ir_set_affinity(struct irq_data *data,
4031 const struct cpumask *mask, bool force)
4033 struct amd_ir_data *ir_data = data->chip_data;
4034 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
4035 struct irq_cfg *cfg = irqd_cfg(data);
4036 struct irq_data *parent = data->parent_data;
4037 struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4043 ret = parent->chip->irq_set_affinity(parent, mask, force);
4044 if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
4047 amd_ir_update_irte(data, iommu, ir_data, irte_info, cfg);
4049 * After this point, all the interrupts will start arriving
4050 * at the new destination. So, time to cleanup the previous
4051 * vector allocation.
4053 send_cleanup_vector(cfg);
4055 return IRQ_SET_MASK_OK_DONE;
4058 static void ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg)
4060 struct amd_ir_data *ir_data = irq_data->chip_data;
4062 *msg = ir_data->msi_entry;
4065 static struct irq_chip amd_ir_chip = {
4067 .irq_ack = apic_ack_irq,
4068 .irq_set_affinity = amd_ir_set_affinity,
4069 .irq_set_vcpu_affinity = amd_ir_set_vcpu_affinity,
4070 .irq_compose_msi_msg = ir_compose_msi_msg,
4073 int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
4075 struct fwnode_handle *fn;
4077 fn = irq_domain_alloc_named_id_fwnode("AMD-IR", iommu->index);
4080 iommu->ir_domain = irq_domain_create_tree(fn, &amd_ir_domain_ops, iommu);
4081 irq_domain_free_fwnode(fn);
4082 if (!iommu->ir_domain)
4085 iommu->ir_domain->parent = arch_get_ir_parent_domain();
4086 iommu->msi_domain = arch_create_remap_msi_irq_domain(iommu->ir_domain,
4092 int amd_iommu_update_ga(int cpu, bool is_run, void *data)
4094 unsigned long flags;
4095 struct amd_iommu *iommu;
4096 struct irq_remap_table *table;
4097 struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
4098 int devid = ir_data->irq_2_irte.devid;
4099 struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
4100 struct irte_ga *ref = (struct irte_ga *) ir_data->ref;
4102 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
4103 !ref || !entry || !entry->lo.fields_vapic.guest_mode)
4106 iommu = amd_iommu_rlookup_table[devid];
4110 table = get_irq_table(devid);
4114 raw_spin_lock_irqsave(&table->lock, flags);
4116 if (ref->lo.fields_vapic.guest_mode) {
4118 ref->lo.fields_vapic.destination =
4119 APICID_TO_IRTE_DEST_LO(cpu);
4120 ref->hi.fields.destination =
4121 APICID_TO_IRTE_DEST_HI(cpu);
4123 ref->lo.fields_vapic.is_run = is_run;
4127 raw_spin_unlock_irqrestore(&table->lock, flags);
4129 iommu_flush_irt(iommu, devid);
4130 iommu_completion_wait(iommu);
4133 EXPORT_SYMBOL(amd_iommu_update_ga);