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 spin_lock_init(&dev_data->lock);
205 dev_data->devid = devid;
206 ratelimit_default_init(&dev_data->rs);
208 llist_add(&dev_data->dev_data_list, &dev_data_list);
212 static struct iommu_dev_data *search_dev_data(u16 devid)
214 struct iommu_dev_data *dev_data;
215 struct llist_node *node;
217 if (llist_empty(&dev_data_list))
220 node = dev_data_list.first;
221 llist_for_each_entry(dev_data, node, dev_data_list) {
222 if (dev_data->devid == devid)
229 static int clone_alias(struct pci_dev *pdev, u16 alias, void *data)
231 u16 devid = pci_dev_id(pdev);
236 amd_iommu_rlookup_table[alias] =
237 amd_iommu_rlookup_table[devid];
238 memcpy(amd_iommu_dev_table[alias].data,
239 amd_iommu_dev_table[devid].data,
240 sizeof(amd_iommu_dev_table[alias].data));
245 static void clone_aliases(struct pci_dev *pdev)
251 * The IVRS alias stored in the alias table may not be
252 * part of the PCI DMA aliases if it's bus differs
253 * from the original device.
255 clone_alias(pdev, amd_iommu_alias_table[pci_dev_id(pdev)], NULL);
257 pci_for_each_dma_alias(pdev, clone_alias, NULL);
260 static struct pci_dev *setup_aliases(struct device *dev)
262 struct pci_dev *pdev = to_pci_dev(dev);
265 /* For ACPI HID devices, there are no aliases */
266 if (!dev_is_pci(dev))
270 * Add the IVRS alias to the pci aliases if it is on the same
271 * bus. The IVRS table may know about a quirk that we don't.
273 ivrs_alias = amd_iommu_alias_table[pci_dev_id(pdev)];
274 if (ivrs_alias != pci_dev_id(pdev) &&
275 PCI_BUS_NUM(ivrs_alias) == pdev->bus->number)
276 pci_add_dma_alias(pdev, ivrs_alias & 0xff, 1);
283 static struct iommu_dev_data *find_dev_data(u16 devid)
285 struct iommu_dev_data *dev_data;
286 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
288 dev_data = search_dev_data(devid);
290 if (dev_data == NULL) {
291 dev_data = alloc_dev_data(devid);
295 if (translation_pre_enabled(iommu))
296 dev_data->defer_attach = true;
302 struct iommu_dev_data *get_dev_data(struct device *dev)
304 return dev->archdata.iommu;
306 EXPORT_SYMBOL(get_dev_data);
309 * Find or create an IOMMU group for a acpihid device.
311 static struct iommu_group *acpihid_device_group(struct device *dev)
313 struct acpihid_map_entry *p, *entry = NULL;
316 devid = get_acpihid_device_id(dev, &entry);
318 return ERR_PTR(devid);
320 list_for_each_entry(p, &acpihid_map, list) {
321 if ((devid == p->devid) && p->group)
322 entry->group = p->group;
326 entry->group = generic_device_group(dev);
328 iommu_group_ref_get(entry->group);
333 static bool pci_iommuv2_capable(struct pci_dev *pdev)
335 static const int caps[] = {
338 PCI_EXT_CAP_ID_PASID,
342 if (pci_ats_disabled())
345 for (i = 0; i < 3; ++i) {
346 pos = pci_find_ext_capability(pdev, caps[i]);
354 static bool pdev_pri_erratum(struct pci_dev *pdev, u32 erratum)
356 struct iommu_dev_data *dev_data;
358 dev_data = get_dev_data(&pdev->dev);
360 return dev_data->errata & (1 << erratum) ? true : false;
364 * This function checks if the driver got a valid device from the caller to
365 * avoid dereferencing invalid pointers.
367 static bool check_device(struct device *dev)
371 if (!dev || !dev->dma_mask)
374 devid = get_device_id(dev);
378 /* Out of our scope? */
379 if (devid > amd_iommu_last_bdf)
382 if (amd_iommu_rlookup_table[devid] == NULL)
388 static void init_iommu_group(struct device *dev)
390 struct iommu_group *group;
392 group = iommu_group_get_for_dev(dev);
396 iommu_group_put(group);
399 static int iommu_init_device(struct device *dev)
401 struct iommu_dev_data *dev_data;
402 struct amd_iommu *iommu;
405 if (dev->archdata.iommu)
408 devid = get_device_id(dev);
412 iommu = amd_iommu_rlookup_table[devid];
414 dev_data = find_dev_data(devid);
418 dev_data->pdev = setup_aliases(dev);
421 * By default we use passthrough mode for IOMMUv2 capable device.
422 * But if amd_iommu=force_isolation is set (e.g. to debug DMA to
423 * invalid address), we ignore the capability for the device so
424 * it'll be forced to go into translation mode.
426 if ((iommu_default_passthrough() || !amd_iommu_force_isolation) &&
427 dev_is_pci(dev) && pci_iommuv2_capable(to_pci_dev(dev))) {
428 struct amd_iommu *iommu;
430 iommu = amd_iommu_rlookup_table[dev_data->devid];
431 dev_data->iommu_v2 = iommu->is_iommu_v2;
434 dev->archdata.iommu = dev_data;
436 iommu_device_link(&iommu->iommu, dev);
441 static void iommu_ignore_device(struct device *dev)
445 devid = get_device_id(dev);
449 amd_iommu_rlookup_table[devid] = NULL;
450 memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
455 static void iommu_uninit_device(struct device *dev)
457 struct iommu_dev_data *dev_data;
458 struct amd_iommu *iommu;
461 devid = get_device_id(dev);
465 iommu = amd_iommu_rlookup_table[devid];
467 dev_data = search_dev_data(devid);
471 if (dev_data->domain)
474 iommu_device_unlink(&iommu->iommu, dev);
476 iommu_group_remove_device(dev);
482 * We keep dev_data around for unplugged devices and reuse it when the
483 * device is re-plugged - not doing so would introduce a ton of races.
488 * Helper function to get the first pte of a large mapping
490 static u64 *first_pte_l7(u64 *pte, unsigned long *page_size,
491 unsigned long *count)
493 unsigned long pte_mask, pg_size, cnt;
496 pg_size = PTE_PAGE_SIZE(*pte);
497 cnt = PAGE_SIZE_PTE_COUNT(pg_size);
498 pte_mask = ~((cnt << 3) - 1);
499 fpte = (u64 *)(((unsigned long)pte) & pte_mask);
502 *page_size = pg_size;
510 /****************************************************************************
512 * Interrupt handling functions
514 ****************************************************************************/
516 static void dump_dte_entry(u16 devid)
520 for (i = 0; i < 4; ++i)
521 pr_err("DTE[%d]: %016llx\n", i,
522 amd_iommu_dev_table[devid].data[i]);
525 static void dump_command(unsigned long phys_addr)
527 struct iommu_cmd *cmd = iommu_phys_to_virt(phys_addr);
530 for (i = 0; i < 4; ++i)
531 pr_err("CMD[%d]: %08x\n", i, cmd->data[i]);
534 static void amd_iommu_report_page_fault(u16 devid, u16 domain_id,
535 u64 address, int flags)
537 struct iommu_dev_data *dev_data = NULL;
538 struct pci_dev *pdev;
540 pdev = pci_get_domain_bus_and_slot(0, PCI_BUS_NUM(devid),
543 dev_data = get_dev_data(&pdev->dev);
545 if (dev_data && __ratelimit(&dev_data->rs)) {
546 pci_err(pdev, "Event logged [IO_PAGE_FAULT domain=0x%04x address=0x%llx flags=0x%04x]\n",
547 domain_id, address, flags);
548 } else if (printk_ratelimit()) {
549 pr_err("Event logged [IO_PAGE_FAULT device=%02x:%02x.%x domain=0x%04x address=0x%llx flags=0x%04x]\n",
550 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
551 domain_id, address, flags);
558 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
560 struct device *dev = iommu->iommu.dev;
561 int type, devid, pasid, flags, tag;
562 volatile u32 *event = __evt;
567 type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
568 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
569 pasid = (event[0] & EVENT_DOMID_MASK_HI) |
570 (event[1] & EVENT_DOMID_MASK_LO);
571 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
572 address = (u64)(((u64)event[3]) << 32) | event[2];
575 /* Did we hit the erratum? */
576 if (++count == LOOP_TIMEOUT) {
577 pr_err("No event written to event log\n");
584 if (type == EVENT_TYPE_IO_FAULT) {
585 amd_iommu_report_page_fault(devid, pasid, address, flags);
590 case EVENT_TYPE_ILL_DEV:
591 dev_err(dev, "Event logged [ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x pasid=0x%05x address=0x%llx flags=0x%04x]\n",
592 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
593 pasid, address, flags);
594 dump_dte_entry(devid);
596 case EVENT_TYPE_DEV_TAB_ERR:
597 dev_err(dev, "Event logged [DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
598 "address=0x%llx flags=0x%04x]\n",
599 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
602 case EVENT_TYPE_PAGE_TAB_ERR:
603 dev_err(dev, "Event logged [PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x pasid=0x%04x address=0x%llx flags=0x%04x]\n",
604 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
605 pasid, address, flags);
607 case EVENT_TYPE_ILL_CMD:
608 dev_err(dev, "Event logged [ILLEGAL_COMMAND_ERROR address=0x%llx]\n", address);
609 dump_command(address);
611 case EVENT_TYPE_CMD_HARD_ERR:
612 dev_err(dev, "Event logged [COMMAND_HARDWARE_ERROR address=0x%llx flags=0x%04x]\n",
615 case EVENT_TYPE_IOTLB_INV_TO:
616 dev_err(dev, "Event logged [IOTLB_INV_TIMEOUT device=%02x:%02x.%x address=0x%llx]\n",
617 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
620 case EVENT_TYPE_INV_DEV_REQ:
621 dev_err(dev, "Event logged [INVALID_DEVICE_REQUEST device=%02x:%02x.%x pasid=0x%05x address=0x%llx flags=0x%04x]\n",
622 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
623 pasid, address, flags);
625 case EVENT_TYPE_INV_PPR_REQ:
626 pasid = ((event[0] >> 16) & 0xFFFF)
627 | ((event[1] << 6) & 0xF0000);
628 tag = event[1] & 0x03FF;
629 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",
630 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
631 pasid, address, flags, tag);
634 dev_err(dev, "Event logged [UNKNOWN event[0]=0x%08x event[1]=0x%08x event[2]=0x%08x event[3]=0x%08x\n",
635 event[0], event[1], event[2], event[3]);
638 memset(__evt, 0, 4 * sizeof(u32));
641 static void iommu_poll_events(struct amd_iommu *iommu)
645 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
646 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
648 while (head != tail) {
649 iommu_print_event(iommu, iommu->evt_buf + head);
650 head = (head + EVENT_ENTRY_SIZE) % EVT_BUFFER_SIZE;
653 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
656 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
658 struct amd_iommu_fault fault;
660 if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
661 pr_err_ratelimited("Unknown PPR request received\n");
665 fault.address = raw[1];
666 fault.pasid = PPR_PASID(raw[0]);
667 fault.device_id = PPR_DEVID(raw[0]);
668 fault.tag = PPR_TAG(raw[0]);
669 fault.flags = PPR_FLAGS(raw[0]);
671 atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
674 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
678 if (iommu->ppr_log == NULL)
681 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
682 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
684 while (head != tail) {
689 raw = (u64 *)(iommu->ppr_log + head);
692 * Hardware bug: Interrupt may arrive before the entry is
693 * written to memory. If this happens we need to wait for the
696 for (i = 0; i < LOOP_TIMEOUT; ++i) {
697 if (PPR_REQ_TYPE(raw[0]) != 0)
702 /* Avoid memcpy function-call overhead */
707 * To detect the hardware bug we need to clear the entry
710 raw[0] = raw[1] = 0UL;
712 /* Update head pointer of hardware ring-buffer */
713 head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
714 writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
716 /* Handle PPR entry */
717 iommu_handle_ppr_entry(iommu, entry);
719 /* Refresh ring-buffer information */
720 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
721 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
725 #ifdef CONFIG_IRQ_REMAP
726 static int (*iommu_ga_log_notifier)(u32);
728 int amd_iommu_register_ga_log_notifier(int (*notifier)(u32))
730 iommu_ga_log_notifier = notifier;
734 EXPORT_SYMBOL(amd_iommu_register_ga_log_notifier);
736 static void iommu_poll_ga_log(struct amd_iommu *iommu)
738 u32 head, tail, cnt = 0;
740 if (iommu->ga_log == NULL)
743 head = readl(iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
744 tail = readl(iommu->mmio_base + MMIO_GA_TAIL_OFFSET);
746 while (head != tail) {
750 raw = (u64 *)(iommu->ga_log + head);
753 /* Avoid memcpy function-call overhead */
756 /* Update head pointer of hardware ring-buffer */
757 head = (head + GA_ENTRY_SIZE) % GA_LOG_SIZE;
758 writel(head, iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
760 /* Handle GA entry */
761 switch (GA_REQ_TYPE(log_entry)) {
763 if (!iommu_ga_log_notifier)
766 pr_debug("%s: devid=%#x, ga_tag=%#x\n",
767 __func__, GA_DEVID(log_entry),
770 if (iommu_ga_log_notifier(GA_TAG(log_entry)) != 0)
771 pr_err("GA log notifier failed.\n");
778 #endif /* CONFIG_IRQ_REMAP */
780 #define AMD_IOMMU_INT_MASK \
781 (MMIO_STATUS_EVT_INT_MASK | \
782 MMIO_STATUS_PPR_INT_MASK | \
783 MMIO_STATUS_GALOG_INT_MASK)
785 irqreturn_t amd_iommu_int_thread(int irq, void *data)
787 struct amd_iommu *iommu = (struct amd_iommu *) data;
788 u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
790 while (status & AMD_IOMMU_INT_MASK) {
791 /* Enable EVT and PPR and GA interrupts again */
792 writel(AMD_IOMMU_INT_MASK,
793 iommu->mmio_base + MMIO_STATUS_OFFSET);
795 if (status & MMIO_STATUS_EVT_INT_MASK) {
796 pr_devel("Processing IOMMU Event Log\n");
797 iommu_poll_events(iommu);
800 if (status & MMIO_STATUS_PPR_INT_MASK) {
801 pr_devel("Processing IOMMU PPR Log\n");
802 iommu_poll_ppr_log(iommu);
805 #ifdef CONFIG_IRQ_REMAP
806 if (status & MMIO_STATUS_GALOG_INT_MASK) {
807 pr_devel("Processing IOMMU GA Log\n");
808 iommu_poll_ga_log(iommu);
813 * Hardware bug: ERBT1312
814 * When re-enabling interrupt (by writing 1
815 * to clear the bit), the hardware might also try to set
816 * the interrupt bit in the event status register.
817 * In this scenario, the bit will be set, and disable
818 * subsequent interrupts.
820 * Workaround: The IOMMU driver should read back the
821 * status register and check if the interrupt bits are cleared.
822 * If not, driver will need to go through the interrupt handler
823 * again and re-clear the bits
825 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
830 irqreturn_t amd_iommu_int_handler(int irq, void *data)
832 return IRQ_WAKE_THREAD;
835 /****************************************************************************
837 * IOMMU command queuing functions
839 ****************************************************************************/
841 static int wait_on_sem(volatile u64 *sem)
845 while (*sem == 0 && i < LOOP_TIMEOUT) {
850 if (i == LOOP_TIMEOUT) {
851 pr_alert("Completion-Wait loop timed out\n");
858 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
859 struct iommu_cmd *cmd)
863 target = iommu->cmd_buf + iommu->cmd_buf_tail;
865 iommu->cmd_buf_tail += sizeof(*cmd);
866 iommu->cmd_buf_tail %= CMD_BUFFER_SIZE;
868 /* Copy command to buffer */
869 memcpy(target, cmd, sizeof(*cmd));
871 /* Tell the IOMMU about it */
872 writel(iommu->cmd_buf_tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
875 static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
877 u64 paddr = iommu_virt_to_phys((void *)address);
879 WARN_ON(address & 0x7ULL);
881 memset(cmd, 0, sizeof(*cmd));
882 cmd->data[0] = lower_32_bits(paddr) | CMD_COMPL_WAIT_STORE_MASK;
883 cmd->data[1] = upper_32_bits(paddr);
885 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
888 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
890 memset(cmd, 0, sizeof(*cmd));
891 cmd->data[0] = devid;
892 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
895 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
896 size_t size, u16 domid, int pde)
901 pages = iommu_num_pages(address, size, PAGE_SIZE);
906 * If we have to flush more than one page, flush all
907 * TLB entries for this domain
909 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
913 address &= PAGE_MASK;
915 memset(cmd, 0, sizeof(*cmd));
916 cmd->data[1] |= domid;
917 cmd->data[2] = lower_32_bits(address);
918 cmd->data[3] = upper_32_bits(address);
919 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
920 if (s) /* size bit - we flush more than one 4kb page */
921 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
922 if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
923 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
926 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
927 u64 address, size_t size)
932 pages = iommu_num_pages(address, size, PAGE_SIZE);
937 * If we have to flush more than one page, flush all
938 * TLB entries for this domain
940 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
944 address &= PAGE_MASK;
946 memset(cmd, 0, sizeof(*cmd));
947 cmd->data[0] = devid;
948 cmd->data[0] |= (qdep & 0xff) << 24;
949 cmd->data[1] = devid;
950 cmd->data[2] = lower_32_bits(address);
951 cmd->data[3] = upper_32_bits(address);
952 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
954 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
957 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, int pasid,
958 u64 address, bool size)
960 memset(cmd, 0, sizeof(*cmd));
962 address &= ~(0xfffULL);
964 cmd->data[0] = pasid;
965 cmd->data[1] = domid;
966 cmd->data[2] = lower_32_bits(address);
967 cmd->data[3] = upper_32_bits(address);
968 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
969 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
971 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
972 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
975 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, int pasid,
976 int qdep, u64 address, bool size)
978 memset(cmd, 0, sizeof(*cmd));
980 address &= ~(0xfffULL);
982 cmd->data[0] = devid;
983 cmd->data[0] |= ((pasid >> 8) & 0xff) << 16;
984 cmd->data[0] |= (qdep & 0xff) << 24;
985 cmd->data[1] = devid;
986 cmd->data[1] |= (pasid & 0xff) << 16;
987 cmd->data[2] = lower_32_bits(address);
988 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
989 cmd->data[3] = upper_32_bits(address);
991 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
992 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
995 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, int pasid,
996 int status, int tag, bool gn)
998 memset(cmd, 0, sizeof(*cmd));
1000 cmd->data[0] = devid;
1002 cmd->data[1] = pasid;
1003 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
1005 cmd->data[3] = tag & 0x1ff;
1006 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
1008 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
1011 static void build_inv_all(struct iommu_cmd *cmd)
1013 memset(cmd, 0, sizeof(*cmd));
1014 CMD_SET_TYPE(cmd, CMD_INV_ALL);
1017 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
1019 memset(cmd, 0, sizeof(*cmd));
1020 cmd->data[0] = devid;
1021 CMD_SET_TYPE(cmd, CMD_INV_IRT);
1025 * Writes the command to the IOMMUs command buffer and informs the
1026 * hardware about the new command.
1028 static int __iommu_queue_command_sync(struct amd_iommu *iommu,
1029 struct iommu_cmd *cmd,
1032 unsigned int count = 0;
1033 u32 left, next_tail;
1035 next_tail = (iommu->cmd_buf_tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
1037 left = (iommu->cmd_buf_head - next_tail) % CMD_BUFFER_SIZE;
1040 /* Skip udelay() the first time around */
1042 if (count == LOOP_TIMEOUT) {
1043 pr_err("Command buffer timeout\n");
1050 /* Update head and recheck remaining space */
1051 iommu->cmd_buf_head = readl(iommu->mmio_base +
1052 MMIO_CMD_HEAD_OFFSET);
1057 copy_cmd_to_buffer(iommu, cmd);
1059 /* Do we need to make sure all commands are processed? */
1060 iommu->need_sync = sync;
1065 static int iommu_queue_command_sync(struct amd_iommu *iommu,
1066 struct iommu_cmd *cmd,
1069 unsigned long flags;
1072 raw_spin_lock_irqsave(&iommu->lock, flags);
1073 ret = __iommu_queue_command_sync(iommu, cmd, sync);
1074 raw_spin_unlock_irqrestore(&iommu->lock, flags);
1079 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
1081 return iommu_queue_command_sync(iommu, cmd, true);
1085 * This function queues a completion wait command into the command
1086 * buffer of an IOMMU
1088 static int iommu_completion_wait(struct amd_iommu *iommu)
1090 struct iommu_cmd cmd;
1091 unsigned long flags;
1094 if (!iommu->need_sync)
1098 build_completion_wait(&cmd, (u64)&iommu->cmd_sem);
1100 raw_spin_lock_irqsave(&iommu->lock, flags);
1104 ret = __iommu_queue_command_sync(iommu, &cmd, false);
1108 ret = wait_on_sem(&iommu->cmd_sem);
1111 raw_spin_unlock_irqrestore(&iommu->lock, flags);
1116 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
1118 struct iommu_cmd cmd;
1120 build_inv_dte(&cmd, devid);
1122 return iommu_queue_command(iommu, &cmd);
1125 static void amd_iommu_flush_dte_all(struct amd_iommu *iommu)
1129 for (devid = 0; devid <= 0xffff; ++devid)
1130 iommu_flush_dte(iommu, devid);
1132 iommu_completion_wait(iommu);
1136 * This function uses heavy locking and may disable irqs for some time. But
1137 * this is no issue because it is only called during resume.
1139 static void amd_iommu_flush_tlb_all(struct amd_iommu *iommu)
1143 for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
1144 struct iommu_cmd cmd;
1145 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1147 iommu_queue_command(iommu, &cmd);
1150 iommu_completion_wait(iommu);
1153 static void amd_iommu_flush_tlb_domid(struct amd_iommu *iommu, u32 dom_id)
1155 struct iommu_cmd cmd;
1157 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1159 iommu_queue_command(iommu, &cmd);
1161 iommu_completion_wait(iommu);
1164 static void amd_iommu_flush_all(struct amd_iommu *iommu)
1166 struct iommu_cmd cmd;
1168 build_inv_all(&cmd);
1170 iommu_queue_command(iommu, &cmd);
1171 iommu_completion_wait(iommu);
1174 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1176 struct iommu_cmd cmd;
1178 build_inv_irt(&cmd, devid);
1180 iommu_queue_command(iommu, &cmd);
1183 static void amd_iommu_flush_irt_all(struct amd_iommu *iommu)
1187 for (devid = 0; devid <= MAX_DEV_TABLE_ENTRIES; devid++)
1188 iommu_flush_irt(iommu, devid);
1190 iommu_completion_wait(iommu);
1193 void iommu_flush_all_caches(struct amd_iommu *iommu)
1195 if (iommu_feature(iommu, FEATURE_IA)) {
1196 amd_iommu_flush_all(iommu);
1198 amd_iommu_flush_dte_all(iommu);
1199 amd_iommu_flush_irt_all(iommu);
1200 amd_iommu_flush_tlb_all(iommu);
1205 * Command send function for flushing on-device TLB
1207 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1208 u64 address, size_t size)
1210 struct amd_iommu *iommu;
1211 struct iommu_cmd cmd;
1214 qdep = dev_data->ats.qdep;
1215 iommu = amd_iommu_rlookup_table[dev_data->devid];
1217 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1219 return iommu_queue_command(iommu, &cmd);
1222 static int device_flush_dte_alias(struct pci_dev *pdev, u16 alias, void *data)
1224 struct amd_iommu *iommu = data;
1226 return iommu_flush_dte(iommu, alias);
1230 * Command send function for invalidating a device table entry
1232 static int device_flush_dte(struct iommu_dev_data *dev_data)
1234 struct amd_iommu *iommu;
1238 iommu = amd_iommu_rlookup_table[dev_data->devid];
1241 ret = pci_for_each_dma_alias(dev_data->pdev,
1242 device_flush_dte_alias, iommu);
1244 ret = iommu_flush_dte(iommu, dev_data->devid);
1248 alias = amd_iommu_alias_table[dev_data->devid];
1249 if (alias != dev_data->devid) {
1250 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 unsigned long flags;
1338 spin_lock_irqsave(&domain->lock, flags);
1339 domain_flush_pages(domain, iova, size);
1340 domain_flush_complete(domain);
1341 spin_unlock_irqrestore(&domain->lock, flags);
1347 * This function flushes the DTEs for all devices in domain
1349 static void domain_flush_devices(struct protection_domain *domain)
1351 struct iommu_dev_data *dev_data;
1353 list_for_each_entry(dev_data, &domain->dev_list, list)
1354 device_flush_dte(dev_data);
1357 /****************************************************************************
1359 * The functions below are used the create the page table mappings for
1360 * unity mapped regions.
1362 ****************************************************************************/
1364 static void free_page_list(struct page *freelist)
1366 while (freelist != NULL) {
1367 unsigned long p = (unsigned long)page_address(freelist);
1368 freelist = freelist->freelist;
1373 static struct page *free_pt_page(unsigned long pt, struct page *freelist)
1375 struct page *p = virt_to_page((void *)pt);
1377 p->freelist = freelist;
1382 #define DEFINE_FREE_PT_FN(LVL, FN) \
1383 static struct page *free_pt_##LVL (unsigned long __pt, struct page *freelist) \
1391 for (i = 0; i < 512; ++i) { \
1392 /* PTE present? */ \
1393 if (!IOMMU_PTE_PRESENT(pt[i])) \
1397 if (PM_PTE_LEVEL(pt[i]) == 0 || \
1398 PM_PTE_LEVEL(pt[i]) == 7) \
1401 p = (unsigned long)IOMMU_PTE_PAGE(pt[i]); \
1402 freelist = FN(p, freelist); \
1405 return free_pt_page((unsigned long)pt, freelist); \
1408 DEFINE_FREE_PT_FN(l2, free_pt_page)
1409 DEFINE_FREE_PT_FN(l3, free_pt_l2)
1410 DEFINE_FREE_PT_FN(l4, free_pt_l3)
1411 DEFINE_FREE_PT_FN(l5, free_pt_l4)
1412 DEFINE_FREE_PT_FN(l6, free_pt_l5)
1414 static struct page *free_sub_pt(unsigned long root, int mode,
1415 struct page *freelist)
1418 case PAGE_MODE_NONE:
1419 case PAGE_MODE_7_LEVEL:
1421 case PAGE_MODE_1_LEVEL:
1422 freelist = free_pt_page(root, freelist);
1424 case PAGE_MODE_2_LEVEL:
1425 freelist = free_pt_l2(root, freelist);
1427 case PAGE_MODE_3_LEVEL:
1428 freelist = free_pt_l3(root, freelist);
1430 case PAGE_MODE_4_LEVEL:
1431 freelist = free_pt_l4(root, freelist);
1433 case PAGE_MODE_5_LEVEL:
1434 freelist = free_pt_l5(root, freelist);
1436 case PAGE_MODE_6_LEVEL:
1437 freelist = free_pt_l6(root, freelist);
1446 static void free_pagetable(struct protection_domain *domain)
1448 unsigned long root = (unsigned long)domain->pt_root;
1449 struct page *freelist = NULL;
1451 BUG_ON(domain->mode < PAGE_MODE_NONE ||
1452 domain->mode > PAGE_MODE_6_LEVEL);
1454 freelist = free_sub_pt(root, domain->mode, freelist);
1456 free_page_list(freelist);
1460 * This function is used to add another level to an IO page table. Adding
1461 * another level increases the size of the address space by 9 bits to a size up
1464 static bool increase_address_space(struct protection_domain *domain,
1465 unsigned long address,
1468 unsigned long flags;
1472 spin_lock_irqsave(&domain->lock, flags);
1474 if (address <= PM_LEVEL_SIZE(domain->mode) ||
1475 WARN_ON_ONCE(domain->mode == PAGE_MODE_6_LEVEL))
1478 pte = (void *)get_zeroed_page(gfp);
1482 *pte = PM_LEVEL_PDE(domain->mode,
1483 iommu_virt_to_phys(domain->pt_root));
1484 domain->pt_root = pte;
1490 spin_unlock_irqrestore(&domain->lock, flags);
1495 static u64 *alloc_pte(struct protection_domain *domain,
1496 unsigned long address,
1497 unsigned long page_size,
1505 BUG_ON(!is_power_of_2(page_size));
1507 while (address > PM_LEVEL_SIZE(domain->mode))
1508 *updated = increase_address_space(domain, address, gfp) || *updated;
1510 level = domain->mode - 1;
1511 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1512 address = PAGE_SIZE_ALIGN(address, page_size);
1513 end_lvl = PAGE_SIZE_LEVEL(page_size);
1515 while (level > end_lvl) {
1520 pte_level = PM_PTE_LEVEL(__pte);
1523 * If we replace a series of large PTEs, we need
1524 * to tear down all of them.
1526 if (IOMMU_PTE_PRESENT(__pte) &&
1527 pte_level == PAGE_MODE_7_LEVEL) {
1528 unsigned long count, i;
1531 lpte = first_pte_l7(pte, NULL, &count);
1534 * Unmap the replicated PTEs that still match the
1535 * original large mapping
1537 for (i = 0; i < count; ++i)
1538 cmpxchg64(&lpte[i], __pte, 0ULL);
1544 if (!IOMMU_PTE_PRESENT(__pte) ||
1545 pte_level == PAGE_MODE_NONE) {
1546 page = (u64 *)get_zeroed_page(gfp);
1551 __npte = PM_LEVEL_PDE(level, iommu_virt_to_phys(page));
1553 /* pte could have been changed somewhere. */
1554 if (cmpxchg64(pte, __pte, __npte) != __pte)
1555 free_page((unsigned long)page);
1556 else if (IOMMU_PTE_PRESENT(__pte))
1562 /* No level skipping support yet */
1563 if (pte_level != level)
1568 pte = IOMMU_PTE_PAGE(__pte);
1570 if (pte_page && level == end_lvl)
1573 pte = &pte[PM_LEVEL_INDEX(level, address)];
1580 * This function checks if there is a PTE for a given dma address. If
1581 * there is one, it returns the pointer to it.
1583 static u64 *fetch_pte(struct protection_domain *domain,
1584 unsigned long address,
1585 unsigned long *page_size)
1592 if (address > PM_LEVEL_SIZE(domain->mode))
1595 level = domain->mode - 1;
1596 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1597 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1602 if (!IOMMU_PTE_PRESENT(*pte))
1606 if (PM_PTE_LEVEL(*pte) == 7 ||
1607 PM_PTE_LEVEL(*pte) == 0)
1610 /* No level skipping support yet */
1611 if (PM_PTE_LEVEL(*pte) != level)
1616 /* Walk to the next level */
1617 pte = IOMMU_PTE_PAGE(*pte);
1618 pte = &pte[PM_LEVEL_INDEX(level, address)];
1619 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1623 * If we have a series of large PTEs, make
1624 * sure to return a pointer to the first one.
1626 if (PM_PTE_LEVEL(*pte) == PAGE_MODE_7_LEVEL)
1627 pte = first_pte_l7(pte, page_size, NULL);
1632 static struct page *free_clear_pte(u64 *pte, u64 pteval, struct page *freelist)
1637 while (cmpxchg64(pte, pteval, 0) != pteval) {
1638 pr_warn("AMD-Vi: IOMMU pte changed since we read it\n");
1642 if (!IOMMU_PTE_PRESENT(pteval))
1645 pt = (unsigned long)IOMMU_PTE_PAGE(pteval);
1646 mode = IOMMU_PTE_MODE(pteval);
1648 return free_sub_pt(pt, mode, freelist);
1652 * Generic mapping functions. It maps a physical address into a DMA
1653 * address space. It allocates the page table pages if necessary.
1654 * In the future it can be extended to a generic mapping function
1655 * supporting all features of AMD IOMMU page tables like level skipping
1656 * and full 64 bit address spaces.
1658 static int iommu_map_page(struct protection_domain *dom,
1659 unsigned long bus_addr,
1660 unsigned long phys_addr,
1661 unsigned long page_size,
1665 struct page *freelist = NULL;
1666 bool updated = false;
1670 BUG_ON(!IS_ALIGNED(bus_addr, page_size));
1671 BUG_ON(!IS_ALIGNED(phys_addr, page_size));
1674 if (!(prot & IOMMU_PROT_MASK))
1677 count = PAGE_SIZE_PTE_COUNT(page_size);
1678 pte = alloc_pte(dom, bus_addr, page_size, NULL, gfp, &updated);
1684 for (i = 0; i < count; ++i)
1685 freelist = free_clear_pte(&pte[i], pte[i], freelist);
1687 if (freelist != NULL)
1691 __pte = PAGE_SIZE_PTE(__sme_set(phys_addr), page_size);
1692 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_PR | IOMMU_PTE_FC;
1694 __pte = __sme_set(phys_addr) | IOMMU_PTE_PR | IOMMU_PTE_FC;
1696 if (prot & IOMMU_PROT_IR)
1697 __pte |= IOMMU_PTE_IR;
1698 if (prot & IOMMU_PROT_IW)
1699 __pte |= IOMMU_PTE_IW;
1701 for (i = 0; i < count; ++i)
1708 unsigned long flags;
1710 spin_lock_irqsave(&dom->lock, flags);
1712 spin_unlock_irqrestore(&dom->lock, flags);
1715 /* Everything flushed out, free pages now */
1716 free_page_list(freelist);
1721 static unsigned long iommu_unmap_page(struct protection_domain *dom,
1722 unsigned long bus_addr,
1723 unsigned long page_size)
1725 unsigned long long unmapped;
1726 unsigned long unmap_size;
1729 BUG_ON(!is_power_of_2(page_size));
1733 while (unmapped < page_size) {
1735 pte = fetch_pte(dom, bus_addr, &unmap_size);
1740 count = PAGE_SIZE_PTE_COUNT(unmap_size);
1741 for (i = 0; i < count; i++)
1745 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1746 unmapped += unmap_size;
1749 BUG_ON(unmapped && !is_power_of_2(unmapped));
1754 /****************************************************************************
1756 * The next functions belong to the address allocator for the dma_ops
1757 * interface functions.
1759 ****************************************************************************/
1762 static unsigned long dma_ops_alloc_iova(struct device *dev,
1763 struct dma_ops_domain *dma_dom,
1764 unsigned int pages, u64 dma_mask)
1766 unsigned long pfn = 0;
1768 pages = __roundup_pow_of_two(pages);
1770 if (dma_mask > DMA_BIT_MASK(32))
1771 pfn = alloc_iova_fast(&dma_dom->iovad, pages,
1772 IOVA_PFN(DMA_BIT_MASK(32)), false);
1775 pfn = alloc_iova_fast(&dma_dom->iovad, pages,
1776 IOVA_PFN(dma_mask), true);
1778 return (pfn << PAGE_SHIFT);
1781 static void dma_ops_free_iova(struct dma_ops_domain *dma_dom,
1782 unsigned long address,
1785 pages = __roundup_pow_of_two(pages);
1786 address >>= PAGE_SHIFT;
1788 free_iova_fast(&dma_dom->iovad, address, pages);
1791 /****************************************************************************
1793 * The next functions belong to the domain allocation. A domain is
1794 * allocated for every IOMMU as the default domain. If device isolation
1795 * is enabled, every device get its own domain. The most important thing
1796 * about domains is the page table mapping the DMA address space they
1799 ****************************************************************************/
1801 static u16 domain_id_alloc(void)
1805 spin_lock(&pd_bitmap_lock);
1806 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1808 if (id > 0 && id < MAX_DOMAIN_ID)
1809 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1812 spin_unlock(&pd_bitmap_lock);
1817 static void domain_id_free(int id)
1819 spin_lock(&pd_bitmap_lock);
1820 if (id > 0 && id < MAX_DOMAIN_ID)
1821 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1822 spin_unlock(&pd_bitmap_lock);
1825 static void free_gcr3_tbl_level1(u64 *tbl)
1830 for (i = 0; i < 512; ++i) {
1831 if (!(tbl[i] & GCR3_VALID))
1834 ptr = iommu_phys_to_virt(tbl[i] & PAGE_MASK);
1836 free_page((unsigned long)ptr);
1840 static void free_gcr3_tbl_level2(u64 *tbl)
1845 for (i = 0; i < 512; ++i) {
1846 if (!(tbl[i] & GCR3_VALID))
1849 ptr = iommu_phys_to_virt(tbl[i] & PAGE_MASK);
1851 free_gcr3_tbl_level1(ptr);
1855 static void free_gcr3_table(struct protection_domain *domain)
1857 if (domain->glx == 2)
1858 free_gcr3_tbl_level2(domain->gcr3_tbl);
1859 else if (domain->glx == 1)
1860 free_gcr3_tbl_level1(domain->gcr3_tbl);
1862 BUG_ON(domain->glx != 0);
1864 free_page((unsigned long)domain->gcr3_tbl);
1867 static void dma_ops_domain_flush_tlb(struct dma_ops_domain *dom)
1869 unsigned long flags;
1871 spin_lock_irqsave(&dom->domain.lock, flags);
1872 domain_flush_tlb(&dom->domain);
1873 domain_flush_complete(&dom->domain);
1874 spin_unlock_irqrestore(&dom->domain.lock, flags);
1877 static void iova_domain_flush_tlb(struct iova_domain *iovad)
1879 struct dma_ops_domain *dom;
1881 dom = container_of(iovad, struct dma_ops_domain, iovad);
1883 dma_ops_domain_flush_tlb(dom);
1887 * Free a domain, only used if something went wrong in the
1888 * allocation path and we need to free an already allocated page table
1890 static void dma_ops_domain_free(struct dma_ops_domain *dom)
1895 put_iova_domain(&dom->iovad);
1897 free_pagetable(&dom->domain);
1900 domain_id_free(dom->domain.id);
1906 * Allocates a new protection domain usable for the dma_ops functions.
1907 * It also initializes the page table and the address allocator data
1908 * structures required for the dma_ops interface
1910 static struct dma_ops_domain *dma_ops_domain_alloc(void)
1912 struct dma_ops_domain *dma_dom;
1914 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
1918 if (protection_domain_init(&dma_dom->domain))
1921 dma_dom->domain.mode = PAGE_MODE_3_LEVEL;
1922 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
1923 dma_dom->domain.flags = PD_DMA_OPS_MASK;
1924 if (!dma_dom->domain.pt_root)
1927 init_iova_domain(&dma_dom->iovad, PAGE_SIZE, IOVA_START_PFN);
1929 if (init_iova_flush_queue(&dma_dom->iovad, iova_domain_flush_tlb, NULL))
1932 /* Initialize reserved ranges */
1933 copy_reserved_iova(&reserved_iova_ranges, &dma_dom->iovad);
1938 dma_ops_domain_free(dma_dom);
1944 * little helper function to check whether a given protection domain is a
1947 static bool dma_ops_domain(struct protection_domain *domain)
1949 return domain->flags & PD_DMA_OPS_MASK;
1952 static void set_dte_entry(u16 devid, struct protection_domain *domain,
1959 if (domain->mode != PAGE_MODE_NONE)
1960 pte_root = iommu_virt_to_phys(domain->pt_root);
1962 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
1963 << DEV_ENTRY_MODE_SHIFT;
1964 pte_root |= DTE_FLAG_IR | DTE_FLAG_IW | DTE_FLAG_V | DTE_FLAG_TV;
1966 flags = amd_iommu_dev_table[devid].data[1];
1969 flags |= DTE_FLAG_IOTLB;
1972 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
1974 if (iommu_feature(iommu, FEATURE_EPHSUP))
1975 pte_root |= 1ULL << DEV_ENTRY_PPR;
1978 if (domain->flags & PD_IOMMUV2_MASK) {
1979 u64 gcr3 = iommu_virt_to_phys(domain->gcr3_tbl);
1980 u64 glx = domain->glx;
1983 pte_root |= DTE_FLAG_GV;
1984 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
1986 /* First mask out possible old values for GCR3 table */
1987 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
1990 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
1993 /* Encode GCR3 table into DTE */
1994 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
1997 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
2000 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
2004 flags &= ~DEV_DOMID_MASK;
2005 flags |= domain->id;
2007 old_domid = amd_iommu_dev_table[devid].data[1] & DEV_DOMID_MASK;
2008 amd_iommu_dev_table[devid].data[1] = flags;
2009 amd_iommu_dev_table[devid].data[0] = pte_root;
2012 * A kdump kernel might be replacing a domain ID that was copied from
2013 * the previous kernel--if so, it needs to flush the translation cache
2014 * entries for the old domain ID that is being overwritten
2017 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
2019 amd_iommu_flush_tlb_domid(iommu, old_domid);
2023 static void clear_dte_entry(u16 devid)
2025 /* remove entry from the device table seen by the hardware */
2026 amd_iommu_dev_table[devid].data[0] = DTE_FLAG_V | DTE_FLAG_TV;
2027 amd_iommu_dev_table[devid].data[1] &= DTE_FLAG_MASK;
2029 amd_iommu_apply_erratum_63(devid);
2032 static void do_attach(struct iommu_dev_data *dev_data,
2033 struct protection_domain *domain)
2035 struct amd_iommu *iommu;
2038 iommu = amd_iommu_rlookup_table[dev_data->devid];
2039 ats = dev_data->ats.enabled;
2041 /* Update data structures */
2042 dev_data->domain = domain;
2043 list_add(&dev_data->list, &domain->dev_list);
2045 /* Do reference counting */
2046 domain->dev_iommu[iommu->index] += 1;
2047 domain->dev_cnt += 1;
2049 /* Update device table */
2050 set_dte_entry(dev_data->devid, domain, ats, dev_data->iommu_v2);
2051 clone_aliases(dev_data->pdev);
2053 device_flush_dte(dev_data);
2056 static void do_detach(struct iommu_dev_data *dev_data)
2058 struct protection_domain *domain = dev_data->domain;
2059 struct amd_iommu *iommu;
2061 iommu = amd_iommu_rlookup_table[dev_data->devid];
2063 /* Update data structures */
2064 dev_data->domain = NULL;
2065 list_del(&dev_data->list);
2066 clear_dte_entry(dev_data->devid);
2067 clone_aliases(dev_data->pdev);
2069 /* Flush the DTE entry */
2070 device_flush_dte(dev_data);
2073 domain_flush_tlb_pde(domain);
2075 /* Wait for the flushes to finish */
2076 domain_flush_complete(domain);
2078 /* decrease reference counters - needs to happen after the flushes */
2079 domain->dev_iommu[iommu->index] -= 1;
2080 domain->dev_cnt -= 1;
2083 static void pdev_iommuv2_disable(struct pci_dev *pdev)
2085 pci_disable_ats(pdev);
2086 pci_disable_pri(pdev);
2087 pci_disable_pasid(pdev);
2090 /* FIXME: Change generic reset-function to do the same */
2091 static int pri_reset_while_enabled(struct pci_dev *pdev)
2096 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2100 pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
2101 control |= PCI_PRI_CTRL_RESET;
2102 pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
2107 static int pdev_iommuv2_enable(struct pci_dev *pdev)
2112 /* FIXME: Hardcode number of outstanding requests for now */
2114 if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
2116 reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
2118 /* Only allow access to user-accessible pages */
2119 ret = pci_enable_pasid(pdev, 0);
2123 /* First reset the PRI state of the device */
2124 ret = pci_reset_pri(pdev);
2129 ret = pci_enable_pri(pdev, reqs);
2134 ret = pri_reset_while_enabled(pdev);
2139 ret = pci_enable_ats(pdev, PAGE_SHIFT);
2146 pci_disable_pri(pdev);
2147 pci_disable_pasid(pdev);
2153 * If a device is not yet associated with a domain, this function makes the
2154 * device visible in the domain
2156 static int attach_device(struct device *dev,
2157 struct protection_domain *domain)
2159 struct pci_dev *pdev;
2160 struct iommu_dev_data *dev_data;
2161 unsigned long flags;
2164 spin_lock_irqsave(&domain->lock, flags);
2166 dev_data = get_dev_data(dev);
2168 spin_lock(&dev_data->lock);
2171 if (dev_data->domain != NULL)
2174 if (!dev_is_pci(dev))
2175 goto skip_ats_check;
2177 pdev = to_pci_dev(dev);
2178 if (domain->flags & PD_IOMMUV2_MASK) {
2180 if (!dev_data->passthrough)
2183 if (dev_data->iommu_v2) {
2184 if (pdev_iommuv2_enable(pdev) != 0)
2187 dev_data->ats.enabled = true;
2188 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2189 dev_data->pri_tlp = pci_prg_resp_pasid_required(pdev);
2191 } else if (amd_iommu_iotlb_sup &&
2192 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
2193 dev_data->ats.enabled = true;
2194 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2200 do_attach(dev_data, domain);
2203 * We might boot into a crash-kernel here. The crashed kernel
2204 * left the caches in the IOMMU dirty. So we have to flush
2205 * here to evict all dirty stuff.
2207 domain_flush_tlb_pde(domain);
2209 domain_flush_complete(domain);
2212 spin_unlock(&dev_data->lock);
2214 spin_unlock_irqrestore(&domain->lock, flags);
2220 * Removes a device from a protection domain (with devtable_lock held)
2222 static void detach_device(struct device *dev)
2224 struct protection_domain *domain;
2225 struct iommu_dev_data *dev_data;
2226 unsigned long flags;
2228 dev_data = get_dev_data(dev);
2229 domain = dev_data->domain;
2231 spin_lock_irqsave(&domain->lock, flags);
2233 spin_lock(&dev_data->lock);
2236 * First check if the device is still attached. It might already
2237 * be detached from its domain because the generic
2238 * iommu_detach_group code detached it and we try again here in
2239 * our alias handling.
2241 if (WARN_ON(!dev_data->domain))
2244 do_detach(dev_data);
2246 if (!dev_is_pci(dev))
2249 if (domain->flags & PD_IOMMUV2_MASK && dev_data->iommu_v2)
2250 pdev_iommuv2_disable(to_pci_dev(dev));
2251 else if (dev_data->ats.enabled)
2252 pci_disable_ats(to_pci_dev(dev));
2254 dev_data->ats.enabled = false;
2257 spin_unlock(&dev_data->lock);
2259 spin_unlock_irqrestore(&domain->lock, flags);
2262 static int amd_iommu_add_device(struct device *dev)
2264 struct iommu_dev_data *dev_data;
2265 struct iommu_domain *domain;
2266 struct amd_iommu *iommu;
2269 if (!check_device(dev) || get_dev_data(dev))
2272 devid = get_device_id(dev);
2276 iommu = amd_iommu_rlookup_table[devid];
2278 ret = iommu_init_device(dev);
2280 if (ret != -ENOTSUPP)
2281 dev_err(dev, "Failed to initialize - trying to proceed anyway\n");
2283 iommu_ignore_device(dev);
2284 dev->dma_ops = NULL;
2287 init_iommu_group(dev);
2289 dev_data = get_dev_data(dev);
2293 if (dev_data->iommu_v2)
2294 iommu_request_dm_for_dev(dev);
2296 /* Domains are initialized for this device - have a look what we ended up with */
2297 domain = iommu_get_domain_for_dev(dev);
2298 if (domain->type == IOMMU_DOMAIN_IDENTITY)
2299 dev_data->passthrough = true;
2301 dev->dma_ops = &amd_iommu_dma_ops;
2304 iommu_completion_wait(iommu);
2309 static void amd_iommu_remove_device(struct device *dev)
2311 struct amd_iommu *iommu;
2314 if (!check_device(dev))
2317 devid = get_device_id(dev);
2321 iommu = amd_iommu_rlookup_table[devid];
2323 iommu_uninit_device(dev);
2324 iommu_completion_wait(iommu);
2327 static struct iommu_group *amd_iommu_device_group(struct device *dev)
2329 if (dev_is_pci(dev))
2330 return pci_device_group(dev);
2332 return acpihid_device_group(dev);
2335 /*****************************************************************************
2337 * The next functions belong to the dma_ops mapping/unmapping code.
2339 *****************************************************************************/
2342 * In the dma_ops path we only have the struct device. This function
2343 * finds the corresponding IOMMU, the protection domain and the
2344 * requestor id for a given device.
2345 * If the device is not yet associated with a domain this is also done
2348 static struct protection_domain *get_domain(struct device *dev)
2350 struct protection_domain *domain;
2351 struct iommu_domain *io_domain;
2353 if (!check_device(dev))
2354 return ERR_PTR(-EINVAL);
2356 domain = get_dev_data(dev)->domain;
2357 if (domain == NULL && get_dev_data(dev)->defer_attach) {
2358 get_dev_data(dev)->defer_attach = false;
2359 io_domain = iommu_get_domain_for_dev(dev);
2360 domain = to_pdomain(io_domain);
2361 attach_device(dev, domain);
2364 return ERR_PTR(-EBUSY);
2366 if (!dma_ops_domain(domain))
2367 return ERR_PTR(-EBUSY);
2372 static void update_device_table(struct protection_domain *domain)
2374 struct iommu_dev_data *dev_data;
2376 list_for_each_entry(dev_data, &domain->dev_list, list) {
2377 set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled,
2378 dev_data->iommu_v2);
2379 clone_aliases(dev_data->pdev);
2383 static void update_domain(struct protection_domain *domain)
2385 update_device_table(domain);
2387 domain_flush_devices(domain);
2388 domain_flush_tlb_pde(domain);
2391 static int dir2prot(enum dma_data_direction direction)
2393 if (direction == DMA_TO_DEVICE)
2394 return IOMMU_PROT_IR;
2395 else if (direction == DMA_FROM_DEVICE)
2396 return IOMMU_PROT_IW;
2397 else if (direction == DMA_BIDIRECTIONAL)
2398 return IOMMU_PROT_IW | IOMMU_PROT_IR;
2404 * This function contains common code for mapping of a physically
2405 * contiguous memory region into DMA address space. It is used by all
2406 * mapping functions provided with this IOMMU driver.
2407 * Must be called with the domain lock held.
2409 static dma_addr_t __map_single(struct device *dev,
2410 struct dma_ops_domain *dma_dom,
2413 enum dma_data_direction direction,
2416 dma_addr_t offset = paddr & ~PAGE_MASK;
2417 dma_addr_t address, start, ret;
2418 unsigned long flags;
2423 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
2426 address = dma_ops_alloc_iova(dev, dma_dom, pages, dma_mask);
2430 prot = dir2prot(direction);
2433 for (i = 0; i < pages; ++i) {
2434 ret = iommu_map_page(&dma_dom->domain, start, paddr,
2435 PAGE_SIZE, prot, GFP_ATOMIC);
2444 domain_flush_np_cache(&dma_dom->domain, address, size);
2451 for (--i; i >= 0; --i) {
2453 iommu_unmap_page(&dma_dom->domain, start, PAGE_SIZE);
2456 spin_lock_irqsave(&dma_dom->domain.lock, flags);
2457 domain_flush_tlb(&dma_dom->domain);
2458 domain_flush_complete(&dma_dom->domain);
2459 spin_unlock_irqrestore(&dma_dom->domain.lock, flags);
2461 dma_ops_free_iova(dma_dom, address, pages);
2463 return DMA_MAPPING_ERROR;
2467 * Does the reverse of the __map_single function. Must be called with
2468 * the domain lock held too
2470 static void __unmap_single(struct dma_ops_domain *dma_dom,
2471 dma_addr_t dma_addr,
2475 dma_addr_t i, start;
2478 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
2479 dma_addr &= PAGE_MASK;
2482 for (i = 0; i < pages; ++i) {
2483 iommu_unmap_page(&dma_dom->domain, start, PAGE_SIZE);
2487 if (amd_iommu_unmap_flush) {
2488 unsigned long flags;
2490 spin_lock_irqsave(&dma_dom->domain.lock, flags);
2491 domain_flush_tlb(&dma_dom->domain);
2492 domain_flush_complete(&dma_dom->domain);
2493 spin_unlock_irqrestore(&dma_dom->domain.lock, flags);
2494 dma_ops_free_iova(dma_dom, dma_addr, pages);
2496 pages = __roundup_pow_of_two(pages);
2497 queue_iova(&dma_dom->iovad, dma_addr >> PAGE_SHIFT, pages, 0);
2502 * The exported map_single function for dma_ops.
2504 static dma_addr_t map_page(struct device *dev, struct page *page,
2505 unsigned long offset, size_t size,
2506 enum dma_data_direction dir,
2507 unsigned long attrs)
2509 phys_addr_t paddr = page_to_phys(page) + offset;
2510 struct protection_domain *domain;
2511 struct dma_ops_domain *dma_dom;
2514 domain = get_domain(dev);
2515 if (PTR_ERR(domain) == -EINVAL)
2516 return (dma_addr_t)paddr;
2517 else if (IS_ERR(domain))
2518 return DMA_MAPPING_ERROR;
2520 dma_mask = *dev->dma_mask;
2521 dma_dom = to_dma_ops_domain(domain);
2523 return __map_single(dev, dma_dom, paddr, size, dir, dma_mask);
2527 * The exported unmap_single function for dma_ops.
2529 static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2530 enum dma_data_direction dir, unsigned long attrs)
2532 struct protection_domain *domain;
2533 struct dma_ops_domain *dma_dom;
2535 domain = get_domain(dev);
2539 dma_dom = to_dma_ops_domain(domain);
2541 __unmap_single(dma_dom, dma_addr, size, dir);
2544 static int sg_num_pages(struct device *dev,
2545 struct scatterlist *sglist,
2548 unsigned long mask, boundary_size;
2549 struct scatterlist *s;
2552 mask = dma_get_seg_boundary(dev);
2553 boundary_size = mask + 1 ? ALIGN(mask + 1, PAGE_SIZE) >> PAGE_SHIFT :
2554 1UL << (BITS_PER_LONG - PAGE_SHIFT);
2556 for_each_sg(sglist, s, nelems, i) {
2559 s->dma_address = npages << PAGE_SHIFT;
2560 p = npages % boundary_size;
2561 n = iommu_num_pages(sg_phys(s), s->length, PAGE_SIZE);
2562 if (p + n > boundary_size)
2563 npages += boundary_size - p;
2571 * The exported map_sg function for dma_ops (handles scatter-gather
2574 static int map_sg(struct device *dev, struct scatterlist *sglist,
2575 int nelems, enum dma_data_direction direction,
2576 unsigned long attrs)
2578 int mapped_pages = 0, npages = 0, prot = 0, i;
2579 struct protection_domain *domain;
2580 struct dma_ops_domain *dma_dom;
2581 struct scatterlist *s;
2582 unsigned long address;
2586 domain = get_domain(dev);
2590 dma_dom = to_dma_ops_domain(domain);
2591 dma_mask = *dev->dma_mask;
2593 npages = sg_num_pages(dev, sglist, nelems);
2595 address = dma_ops_alloc_iova(dev, dma_dom, npages, dma_mask);
2599 prot = dir2prot(direction);
2601 /* Map all sg entries */
2602 for_each_sg(sglist, s, nelems, i) {
2603 int j, pages = iommu_num_pages(sg_phys(s), s->length, PAGE_SIZE);
2605 for (j = 0; j < pages; ++j) {
2606 unsigned long bus_addr, phys_addr;
2608 bus_addr = address + s->dma_address + (j << PAGE_SHIFT);
2609 phys_addr = (sg_phys(s) & PAGE_MASK) + (j << PAGE_SHIFT);
2610 ret = iommu_map_page(domain, bus_addr, phys_addr,
2612 GFP_ATOMIC | __GFP_NOWARN);
2620 /* Everything is mapped - write the right values into s->dma_address */
2621 for_each_sg(sglist, s, nelems, i) {
2623 * Add in the remaining piece of the scatter-gather offset that
2624 * was masked out when we were determining the physical address
2625 * via (sg_phys(s) & PAGE_MASK) earlier.
2627 s->dma_address += address + (s->offset & ~PAGE_MASK);
2628 s->dma_length = s->length;
2632 domain_flush_np_cache(domain, s->dma_address, s->dma_length);
2637 dev_err(dev, "IOMMU mapping error in map_sg (io-pages: %d reason: %d)\n",
2640 for_each_sg(sglist, s, nelems, i) {
2641 int j, pages = iommu_num_pages(sg_phys(s), s->length, PAGE_SIZE);
2643 for (j = 0; j < pages; ++j) {
2644 unsigned long bus_addr;
2646 bus_addr = address + s->dma_address + (j << PAGE_SHIFT);
2647 iommu_unmap_page(domain, bus_addr, PAGE_SIZE);
2649 if (--mapped_pages == 0)
2655 free_iova_fast(&dma_dom->iovad, address >> PAGE_SHIFT, npages);
2662 * The exported map_sg function for dma_ops (handles scatter-gather
2665 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2666 int nelems, enum dma_data_direction dir,
2667 unsigned long attrs)
2669 struct protection_domain *domain;
2670 struct dma_ops_domain *dma_dom;
2671 unsigned long startaddr;
2674 domain = get_domain(dev);
2678 startaddr = sg_dma_address(sglist) & PAGE_MASK;
2679 dma_dom = to_dma_ops_domain(domain);
2680 npages = sg_num_pages(dev, sglist, nelems);
2682 __unmap_single(dma_dom, startaddr, npages << PAGE_SHIFT, dir);
2686 * The exported alloc_coherent function for dma_ops.
2688 static void *alloc_coherent(struct device *dev, size_t size,
2689 dma_addr_t *dma_addr, gfp_t flag,
2690 unsigned long attrs)
2692 u64 dma_mask = dev->coherent_dma_mask;
2693 struct protection_domain *domain;
2694 struct dma_ops_domain *dma_dom;
2697 domain = get_domain(dev);
2698 if (PTR_ERR(domain) == -EINVAL) {
2699 page = alloc_pages(flag, get_order(size));
2700 *dma_addr = page_to_phys(page);
2701 return page_address(page);
2702 } else if (IS_ERR(domain))
2705 dma_dom = to_dma_ops_domain(domain);
2706 size = PAGE_ALIGN(size);
2707 dma_mask = dev->coherent_dma_mask;
2708 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
2711 page = alloc_pages(flag | __GFP_NOWARN, get_order(size));
2713 if (!gfpflags_allow_blocking(flag))
2716 page = dma_alloc_from_contiguous(dev, size >> PAGE_SHIFT,
2717 get_order(size), flag & __GFP_NOWARN);
2723 dma_mask = *dev->dma_mask;
2725 *dma_addr = __map_single(dev, dma_dom, page_to_phys(page),
2726 size, DMA_BIDIRECTIONAL, dma_mask);
2728 if (*dma_addr == DMA_MAPPING_ERROR)
2731 return page_address(page);
2735 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2736 __free_pages(page, get_order(size));
2742 * The exported free_coherent function for dma_ops.
2744 static void free_coherent(struct device *dev, size_t size,
2745 void *virt_addr, dma_addr_t dma_addr,
2746 unsigned long attrs)
2748 struct protection_domain *domain;
2749 struct dma_ops_domain *dma_dom;
2752 page = virt_to_page(virt_addr);
2753 size = PAGE_ALIGN(size);
2755 domain = get_domain(dev);
2759 dma_dom = to_dma_ops_domain(domain);
2761 __unmap_single(dma_dom, dma_addr, size, DMA_BIDIRECTIONAL);
2764 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2765 __free_pages(page, get_order(size));
2769 * This function is called by the DMA layer to find out if we can handle a
2770 * particular device. It is part of the dma_ops.
2772 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
2774 if (!dma_direct_supported(dev, mask))
2776 return check_device(dev);
2779 static const struct dma_map_ops amd_iommu_dma_ops = {
2780 .alloc = alloc_coherent,
2781 .free = free_coherent,
2782 .map_page = map_page,
2783 .unmap_page = unmap_page,
2785 .unmap_sg = unmap_sg,
2786 .dma_supported = amd_iommu_dma_supported,
2787 .mmap = dma_common_mmap,
2788 .get_sgtable = dma_common_get_sgtable,
2791 static int init_reserved_iova_ranges(void)
2793 struct pci_dev *pdev = NULL;
2796 init_iova_domain(&reserved_iova_ranges, PAGE_SIZE, IOVA_START_PFN);
2798 lockdep_set_class(&reserved_iova_ranges.iova_rbtree_lock,
2799 &reserved_rbtree_key);
2801 /* MSI memory range */
2802 val = reserve_iova(&reserved_iova_ranges,
2803 IOVA_PFN(MSI_RANGE_START), IOVA_PFN(MSI_RANGE_END));
2805 pr_err("Reserving MSI range failed\n");
2809 /* HT memory range */
2810 val = reserve_iova(&reserved_iova_ranges,
2811 IOVA_PFN(HT_RANGE_START), IOVA_PFN(HT_RANGE_END));
2813 pr_err("Reserving HT range failed\n");
2818 * Memory used for PCI resources
2819 * FIXME: Check whether we can reserve the PCI-hole completly
2821 for_each_pci_dev(pdev) {
2824 for (i = 0; i < PCI_NUM_RESOURCES; ++i) {
2825 struct resource *r = &pdev->resource[i];
2827 if (!(r->flags & IORESOURCE_MEM))
2830 val = reserve_iova(&reserved_iova_ranges,
2834 pci_err(pdev, "Reserve pci-resource range %pR failed\n", r);
2843 int __init amd_iommu_init_api(void)
2847 ret = iova_cache_get();
2851 ret = init_reserved_iova_ranges();
2855 err = bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
2858 #ifdef CONFIG_ARM_AMBA
2859 err = bus_set_iommu(&amba_bustype, &amd_iommu_ops);
2863 err = bus_set_iommu(&platform_bus_type, &amd_iommu_ops);
2870 int __init amd_iommu_init_dma_ops(void)
2872 swiotlb = (iommu_default_passthrough() || sme_me_mask) ? 1 : 0;
2875 if (amd_iommu_unmap_flush)
2876 pr_info("IO/TLB flush on unmap enabled\n");
2878 pr_info("Lazy IO/TLB flushing enabled\n");
2884 /*****************************************************************************
2886 * The following functions belong to the exported interface of AMD IOMMU
2888 * This interface allows access to lower level functions of the IOMMU
2889 * like protection domain handling and assignement of devices to domains
2890 * which is not possible with the dma_ops interface.
2892 *****************************************************************************/
2894 static void cleanup_domain(struct protection_domain *domain)
2896 struct iommu_dev_data *entry;
2897 unsigned long flags;
2899 spin_lock_irqsave(&domain->lock, flags);
2901 while (!list_empty(&domain->dev_list)) {
2902 entry = list_first_entry(&domain->dev_list,
2903 struct iommu_dev_data, list);
2904 BUG_ON(!entry->domain);
2908 spin_unlock_irqrestore(&domain->lock, flags);
2911 static void protection_domain_free(struct protection_domain *domain)
2917 domain_id_free(domain->id);
2922 static int protection_domain_init(struct protection_domain *domain)
2924 spin_lock_init(&domain->lock);
2925 mutex_init(&domain->api_lock);
2926 domain->id = domain_id_alloc();
2929 INIT_LIST_HEAD(&domain->dev_list);
2934 static struct protection_domain *protection_domain_alloc(void)
2936 struct protection_domain *domain;
2938 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2942 if (protection_domain_init(domain))
2953 static struct iommu_domain *amd_iommu_domain_alloc(unsigned type)
2955 struct protection_domain *pdomain;
2956 struct dma_ops_domain *dma_domain;
2959 case IOMMU_DOMAIN_UNMANAGED:
2960 pdomain = protection_domain_alloc();
2964 pdomain->mode = PAGE_MODE_3_LEVEL;
2965 pdomain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2966 if (!pdomain->pt_root) {
2967 protection_domain_free(pdomain);
2971 pdomain->domain.geometry.aperture_start = 0;
2972 pdomain->domain.geometry.aperture_end = ~0ULL;
2973 pdomain->domain.geometry.force_aperture = true;
2976 case IOMMU_DOMAIN_DMA:
2977 dma_domain = dma_ops_domain_alloc();
2979 pr_err("Failed to allocate\n");
2982 pdomain = &dma_domain->domain;
2984 case IOMMU_DOMAIN_IDENTITY:
2985 pdomain = protection_domain_alloc();
2989 pdomain->mode = PAGE_MODE_NONE;
2995 return &pdomain->domain;
2998 static void amd_iommu_domain_free(struct iommu_domain *dom)
3000 struct protection_domain *domain;
3001 struct dma_ops_domain *dma_dom;
3003 domain = to_pdomain(dom);
3005 if (domain->dev_cnt > 0)
3006 cleanup_domain(domain);
3008 BUG_ON(domain->dev_cnt != 0);
3013 switch (dom->type) {
3014 case IOMMU_DOMAIN_DMA:
3015 /* Now release the domain */
3016 dma_dom = to_dma_ops_domain(domain);
3017 dma_ops_domain_free(dma_dom);
3020 if (domain->mode != PAGE_MODE_NONE)
3021 free_pagetable(domain);
3023 if (domain->flags & PD_IOMMUV2_MASK)
3024 free_gcr3_table(domain);
3026 protection_domain_free(domain);
3031 static void amd_iommu_detach_device(struct iommu_domain *dom,
3034 struct iommu_dev_data *dev_data = dev->archdata.iommu;
3035 struct amd_iommu *iommu;
3038 if (!check_device(dev))
3041 devid = get_device_id(dev);
3045 if (dev_data->domain != NULL)
3048 iommu = amd_iommu_rlookup_table[devid];
3052 #ifdef CONFIG_IRQ_REMAP
3053 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) &&
3054 (dom->type == IOMMU_DOMAIN_UNMANAGED))
3055 dev_data->use_vapic = 0;
3058 iommu_completion_wait(iommu);
3061 static int amd_iommu_attach_device(struct iommu_domain *dom,
3064 struct protection_domain *domain = to_pdomain(dom);
3065 struct iommu_dev_data *dev_data;
3066 struct amd_iommu *iommu;
3069 if (!check_device(dev))
3072 dev_data = dev->archdata.iommu;
3074 iommu = amd_iommu_rlookup_table[dev_data->devid];
3078 if (dev_data->domain)
3081 ret = attach_device(dev, domain);
3083 #ifdef CONFIG_IRQ_REMAP
3084 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
3085 if (dom->type == IOMMU_DOMAIN_UNMANAGED)
3086 dev_data->use_vapic = 1;
3088 dev_data->use_vapic = 0;
3092 iommu_completion_wait(iommu);
3097 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
3098 phys_addr_t paddr, size_t page_size, int iommu_prot)
3100 struct protection_domain *domain = to_pdomain(dom);
3104 if (domain->mode == PAGE_MODE_NONE)
3107 if (iommu_prot & IOMMU_READ)
3108 prot |= IOMMU_PROT_IR;
3109 if (iommu_prot & IOMMU_WRITE)
3110 prot |= IOMMU_PROT_IW;
3112 mutex_lock(&domain->api_lock);
3113 ret = iommu_map_page(domain, iova, paddr, page_size, prot, GFP_KERNEL);
3114 mutex_unlock(&domain->api_lock);
3116 domain_flush_np_cache(domain, iova, page_size);
3121 static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
3123 struct iommu_iotlb_gather *gather)
3125 struct protection_domain *domain = to_pdomain(dom);
3128 if (domain->mode == PAGE_MODE_NONE)
3131 mutex_lock(&domain->api_lock);
3132 unmap_size = iommu_unmap_page(domain, iova, page_size);
3133 mutex_unlock(&domain->api_lock);
3138 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
3141 struct protection_domain *domain = to_pdomain(dom);
3142 unsigned long offset_mask, pte_pgsize;
3145 if (domain->mode == PAGE_MODE_NONE)
3148 pte = fetch_pte(domain, iova, &pte_pgsize);
3150 if (!pte || !IOMMU_PTE_PRESENT(*pte))
3153 offset_mask = pte_pgsize - 1;
3154 __pte = __sme_clr(*pte & PM_ADDR_MASK);
3156 return (__pte & ~offset_mask) | (iova & offset_mask);
3159 static bool amd_iommu_capable(enum iommu_cap cap)
3162 case IOMMU_CAP_CACHE_COHERENCY:
3164 case IOMMU_CAP_INTR_REMAP:
3165 return (irq_remapping_enabled == 1);
3166 case IOMMU_CAP_NOEXEC:
3175 static void amd_iommu_get_resv_regions(struct device *dev,
3176 struct list_head *head)
3178 struct iommu_resv_region *region;
3179 struct unity_map_entry *entry;
3182 devid = get_device_id(dev);
3186 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
3190 if (devid < entry->devid_start || devid > entry->devid_end)
3193 type = IOMMU_RESV_DIRECT;
3194 length = entry->address_end - entry->address_start;
3195 if (entry->prot & IOMMU_PROT_IR)
3197 if (entry->prot & IOMMU_PROT_IW)
3198 prot |= IOMMU_WRITE;
3199 if (entry->prot & IOMMU_UNITY_MAP_FLAG_EXCL_RANGE)
3200 /* Exclusion range */
3201 type = IOMMU_RESV_RESERVED;
3203 region = iommu_alloc_resv_region(entry->address_start,
3204 length, prot, type);
3206 dev_err(dev, "Out of memory allocating dm-regions\n");
3209 list_add_tail(®ion->list, head);
3212 region = iommu_alloc_resv_region(MSI_RANGE_START,
3213 MSI_RANGE_END - MSI_RANGE_START + 1,
3217 list_add_tail(®ion->list, head);
3219 region = iommu_alloc_resv_region(HT_RANGE_START,
3220 HT_RANGE_END - HT_RANGE_START + 1,
3221 0, IOMMU_RESV_RESERVED);
3224 list_add_tail(®ion->list, head);
3227 static void amd_iommu_put_resv_regions(struct device *dev,
3228 struct list_head *head)
3230 struct iommu_resv_region *entry, *next;
3232 list_for_each_entry_safe(entry, next, head, list)
3236 static void amd_iommu_apply_resv_region(struct device *dev,
3237 struct iommu_domain *domain,
3238 struct iommu_resv_region *region)
3240 struct dma_ops_domain *dma_dom = to_dma_ops_domain(to_pdomain(domain));
3241 unsigned long start, end;
3243 start = IOVA_PFN(region->start);
3244 end = IOVA_PFN(region->start + region->length - 1);
3246 WARN_ON_ONCE(reserve_iova(&dma_dom->iovad, start, end) == NULL);
3249 static bool amd_iommu_is_attach_deferred(struct iommu_domain *domain,
3252 struct iommu_dev_data *dev_data = dev->archdata.iommu;
3253 return dev_data->defer_attach;
3256 static void amd_iommu_flush_iotlb_all(struct iommu_domain *domain)
3258 struct protection_domain *dom = to_pdomain(domain);
3259 unsigned long flags;
3261 spin_lock_irqsave(&dom->lock, flags);
3262 domain_flush_tlb_pde(dom);
3263 domain_flush_complete(dom);
3264 spin_unlock_irqrestore(&dom->lock, flags);
3267 static void amd_iommu_iotlb_sync(struct iommu_domain *domain,
3268 struct iommu_iotlb_gather *gather)
3270 amd_iommu_flush_iotlb_all(domain);
3273 const struct iommu_ops amd_iommu_ops = {
3274 .capable = amd_iommu_capable,
3275 .domain_alloc = amd_iommu_domain_alloc,
3276 .domain_free = amd_iommu_domain_free,
3277 .attach_dev = amd_iommu_attach_device,
3278 .detach_dev = amd_iommu_detach_device,
3279 .map = amd_iommu_map,
3280 .unmap = amd_iommu_unmap,
3281 .iova_to_phys = amd_iommu_iova_to_phys,
3282 .add_device = amd_iommu_add_device,
3283 .remove_device = amd_iommu_remove_device,
3284 .device_group = amd_iommu_device_group,
3285 .get_resv_regions = amd_iommu_get_resv_regions,
3286 .put_resv_regions = amd_iommu_put_resv_regions,
3287 .apply_resv_region = amd_iommu_apply_resv_region,
3288 .is_attach_deferred = amd_iommu_is_attach_deferred,
3289 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
3290 .flush_iotlb_all = amd_iommu_flush_iotlb_all,
3291 .iotlb_sync = amd_iommu_iotlb_sync,
3294 /*****************************************************************************
3296 * The next functions do a basic initialization of IOMMU for pass through
3299 * In passthrough mode the IOMMU is initialized and enabled but not used for
3300 * DMA-API translation.
3302 *****************************************************************************/
3304 /* IOMMUv2 specific functions */
3305 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
3307 return atomic_notifier_chain_register(&ppr_notifier, nb);
3309 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
3311 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
3313 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
3315 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
3317 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
3319 struct protection_domain *domain = to_pdomain(dom);
3320 unsigned long flags;
3322 spin_lock_irqsave(&domain->lock, flags);
3324 /* Update data structure */
3325 domain->mode = PAGE_MODE_NONE;
3327 /* Make changes visible to IOMMUs */
3328 update_domain(domain);
3330 /* Page-table is not visible to IOMMU anymore, so free it */
3331 free_pagetable(domain);
3333 spin_unlock_irqrestore(&domain->lock, flags);
3335 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
3337 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
3339 struct protection_domain *domain = to_pdomain(dom);
3340 unsigned long flags;
3343 if (pasids <= 0 || pasids > (PASID_MASK + 1))
3346 /* Number of GCR3 table levels required */
3347 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
3350 if (levels > amd_iommu_max_glx_val)
3353 spin_lock_irqsave(&domain->lock, flags);
3356 * Save us all sanity checks whether devices already in the
3357 * domain support IOMMUv2. Just force that the domain has no
3358 * devices attached when it is switched into IOMMUv2 mode.
3361 if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
3365 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
3366 if (domain->gcr3_tbl == NULL)
3369 domain->glx = levels;
3370 domain->flags |= PD_IOMMUV2_MASK;
3372 update_domain(domain);
3377 spin_unlock_irqrestore(&domain->lock, flags);
3381 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
3383 static int __flush_pasid(struct protection_domain *domain, int pasid,
3384 u64 address, bool size)
3386 struct iommu_dev_data *dev_data;
3387 struct iommu_cmd cmd;
3390 if (!(domain->flags & PD_IOMMUV2_MASK))
3393 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
3396 * IOMMU TLB needs to be flushed before Device TLB to
3397 * prevent device TLB refill from IOMMU TLB
3399 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
3400 if (domain->dev_iommu[i] == 0)
3403 ret = iommu_queue_command(amd_iommus[i], &cmd);
3408 /* Wait until IOMMU TLB flushes are complete */
3409 domain_flush_complete(domain);
3411 /* Now flush device TLBs */
3412 list_for_each_entry(dev_data, &domain->dev_list, list) {
3413 struct amd_iommu *iommu;
3417 There might be non-IOMMUv2 capable devices in an IOMMUv2
3420 if (!dev_data->ats.enabled)
3423 qdep = dev_data->ats.qdep;
3424 iommu = amd_iommu_rlookup_table[dev_data->devid];
3426 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
3427 qdep, address, size);
3429 ret = iommu_queue_command(iommu, &cmd);
3434 /* Wait until all device TLBs are flushed */
3435 domain_flush_complete(domain);
3444 static int __amd_iommu_flush_page(struct protection_domain *domain, int pasid,
3447 return __flush_pasid(domain, pasid, address, false);
3450 int amd_iommu_flush_page(struct iommu_domain *dom, int pasid,
3453 struct protection_domain *domain = to_pdomain(dom);
3454 unsigned long flags;
3457 spin_lock_irqsave(&domain->lock, flags);
3458 ret = __amd_iommu_flush_page(domain, pasid, address);
3459 spin_unlock_irqrestore(&domain->lock, flags);
3463 EXPORT_SYMBOL(amd_iommu_flush_page);
3465 static int __amd_iommu_flush_tlb(struct protection_domain *domain, int pasid)
3467 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
3471 int amd_iommu_flush_tlb(struct iommu_domain *dom, int pasid)
3473 struct protection_domain *domain = to_pdomain(dom);
3474 unsigned long flags;
3477 spin_lock_irqsave(&domain->lock, flags);
3478 ret = __amd_iommu_flush_tlb(domain, pasid);
3479 spin_unlock_irqrestore(&domain->lock, flags);
3483 EXPORT_SYMBOL(amd_iommu_flush_tlb);
3485 static u64 *__get_gcr3_pte(u64 *root, int level, int pasid, bool alloc)
3492 index = (pasid >> (9 * level)) & 0x1ff;
3498 if (!(*pte & GCR3_VALID)) {
3502 root = (void *)get_zeroed_page(GFP_ATOMIC);
3506 *pte = iommu_virt_to_phys(root) | GCR3_VALID;
3509 root = iommu_phys_to_virt(*pte & PAGE_MASK);
3517 static int __set_gcr3(struct protection_domain *domain, int pasid,
3522 if (domain->mode != PAGE_MODE_NONE)
3525 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3529 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3531 return __amd_iommu_flush_tlb(domain, pasid);
3534 static int __clear_gcr3(struct protection_domain *domain, int pasid)
3538 if (domain->mode != PAGE_MODE_NONE)
3541 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3547 return __amd_iommu_flush_tlb(domain, pasid);
3550 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, int pasid,
3553 struct protection_domain *domain = to_pdomain(dom);
3554 unsigned long flags;
3557 spin_lock_irqsave(&domain->lock, flags);
3558 ret = __set_gcr3(domain, pasid, cr3);
3559 spin_unlock_irqrestore(&domain->lock, flags);
3563 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3565 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, int pasid)
3567 struct protection_domain *domain = to_pdomain(dom);
3568 unsigned long flags;
3571 spin_lock_irqsave(&domain->lock, flags);
3572 ret = __clear_gcr3(domain, pasid);
3573 spin_unlock_irqrestore(&domain->lock, flags);
3577 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3579 int amd_iommu_complete_ppr(struct pci_dev *pdev, int pasid,
3580 int status, int tag)
3582 struct iommu_dev_data *dev_data;
3583 struct amd_iommu *iommu;
3584 struct iommu_cmd cmd;
3586 dev_data = get_dev_data(&pdev->dev);
3587 iommu = amd_iommu_rlookup_table[dev_data->devid];
3589 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3590 tag, dev_data->pri_tlp);
3592 return iommu_queue_command(iommu, &cmd);
3594 EXPORT_SYMBOL(amd_iommu_complete_ppr);
3596 struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3598 struct protection_domain *pdomain;
3600 pdomain = get_domain(&pdev->dev);
3601 if (IS_ERR(pdomain))
3604 /* Only return IOMMUv2 domains */
3605 if (!(pdomain->flags & PD_IOMMUV2_MASK))
3608 return &pdomain->domain;
3610 EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3612 void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3614 struct iommu_dev_data *dev_data;
3616 if (!amd_iommu_v2_supported())
3619 dev_data = get_dev_data(&pdev->dev);
3620 dev_data->errata |= (1 << erratum);
3622 EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3624 int amd_iommu_device_info(struct pci_dev *pdev,
3625 struct amd_iommu_device_info *info)
3630 if (pdev == NULL || info == NULL)
3633 if (!amd_iommu_v2_supported())
3636 memset(info, 0, sizeof(*info));
3638 if (!pci_ats_disabled()) {
3639 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS);
3641 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3644 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3646 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3648 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3652 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3653 max_pasids = min(max_pasids, (1 << 20));
3655 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3656 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3658 features = pci_pasid_features(pdev);
3659 if (features & PCI_PASID_CAP_EXEC)
3660 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3661 if (features & PCI_PASID_CAP_PRIV)
3662 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3667 EXPORT_SYMBOL(amd_iommu_device_info);
3669 #ifdef CONFIG_IRQ_REMAP
3671 /*****************************************************************************
3673 * Interrupt Remapping Implementation
3675 *****************************************************************************/
3677 static struct irq_chip amd_ir_chip;
3678 static DEFINE_SPINLOCK(iommu_table_lock);
3680 static void set_dte_irq_entry(u16 devid, struct irq_remap_table *table)
3684 dte = amd_iommu_dev_table[devid].data[2];
3685 dte &= ~DTE_IRQ_PHYS_ADDR_MASK;
3686 dte |= iommu_virt_to_phys(table->table);
3687 dte |= DTE_IRQ_REMAP_INTCTL;
3688 dte |= DTE_IRQ_TABLE_LEN;
3689 dte |= DTE_IRQ_REMAP_ENABLE;
3691 amd_iommu_dev_table[devid].data[2] = dte;
3694 static struct irq_remap_table *get_irq_table(u16 devid)
3696 struct irq_remap_table *table;
3698 if (WARN_ONCE(!amd_iommu_rlookup_table[devid],
3699 "%s: no iommu for devid %x\n", __func__, devid))
3702 table = irq_lookup_table[devid];
3703 if (WARN_ONCE(!table, "%s: no table for devid %x\n", __func__, devid))
3709 static struct irq_remap_table *__alloc_irq_table(void)
3711 struct irq_remap_table *table;
3713 table = kzalloc(sizeof(*table), GFP_KERNEL);
3717 table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_KERNEL);
3718 if (!table->table) {
3722 raw_spin_lock_init(&table->lock);
3724 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
3725 memset(table->table, 0,
3726 MAX_IRQS_PER_TABLE * sizeof(u32));
3728 memset(table->table, 0,
3729 (MAX_IRQS_PER_TABLE * (sizeof(u64) * 2)));
3733 static void set_remap_table_entry(struct amd_iommu *iommu, u16 devid,
3734 struct irq_remap_table *table)
3736 irq_lookup_table[devid] = table;
3737 set_dte_irq_entry(devid, table);
3738 iommu_flush_dte(iommu, devid);
3741 static int set_remap_table_entry_alias(struct pci_dev *pdev, u16 alias,
3744 struct irq_remap_table *table = data;
3746 irq_lookup_table[alias] = table;
3747 set_dte_irq_entry(alias, table);
3749 iommu_flush_dte(amd_iommu_rlookup_table[alias], alias);
3754 static struct irq_remap_table *alloc_irq_table(u16 devid, struct pci_dev *pdev)
3756 struct irq_remap_table *table = NULL;
3757 struct irq_remap_table *new_table = NULL;
3758 struct amd_iommu *iommu;
3759 unsigned long flags;
3762 spin_lock_irqsave(&iommu_table_lock, flags);
3764 iommu = amd_iommu_rlookup_table[devid];
3768 table = irq_lookup_table[devid];
3772 alias = amd_iommu_alias_table[devid];
3773 table = irq_lookup_table[alias];
3775 set_remap_table_entry(iommu, devid, table);
3778 spin_unlock_irqrestore(&iommu_table_lock, flags);
3780 /* Nothing there yet, allocate new irq remapping table */
3781 new_table = __alloc_irq_table();
3785 spin_lock_irqsave(&iommu_table_lock, flags);
3787 table = irq_lookup_table[devid];
3791 table = irq_lookup_table[alias];
3793 set_remap_table_entry(iommu, devid, table);
3801 pci_for_each_dma_alias(pdev, set_remap_table_entry_alias,
3804 set_remap_table_entry(iommu, devid, table);
3807 set_remap_table_entry(iommu, alias, table);
3810 iommu_completion_wait(iommu);
3813 spin_unlock_irqrestore(&iommu_table_lock, flags);
3816 kmem_cache_free(amd_iommu_irq_cache, new_table->table);
3822 static int alloc_irq_index(u16 devid, int count, bool align,
3823 struct pci_dev *pdev)
3825 struct irq_remap_table *table;
3826 int index, c, alignment = 1;
3827 unsigned long flags;
3828 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
3833 table = alloc_irq_table(devid, pdev);
3838 alignment = roundup_pow_of_two(count);
3840 raw_spin_lock_irqsave(&table->lock, flags);
3842 /* Scan table for free entries */
3843 for (index = ALIGN(table->min_index, alignment), c = 0;
3844 index < MAX_IRQS_PER_TABLE;) {
3845 if (!iommu->irte_ops->is_allocated(table, index)) {
3849 index = ALIGN(index + 1, alignment);
3855 iommu->irte_ops->set_allocated(table, index - c + 1);
3867 raw_spin_unlock_irqrestore(&table->lock, flags);
3872 static int modify_irte_ga(u16 devid, int index, struct irte_ga *irte,
3873 struct amd_ir_data *data)
3875 struct irq_remap_table *table;
3876 struct amd_iommu *iommu;
3877 unsigned long flags;
3878 struct irte_ga *entry;
3880 iommu = amd_iommu_rlookup_table[devid];
3884 table = get_irq_table(devid);
3888 raw_spin_lock_irqsave(&table->lock, flags);
3890 entry = (struct irte_ga *)table->table;
3891 entry = &entry[index];
3892 entry->lo.fields_remap.valid = 0;
3893 entry->hi.val = irte->hi.val;
3894 entry->lo.val = irte->lo.val;
3895 entry->lo.fields_remap.valid = 1;
3899 raw_spin_unlock_irqrestore(&table->lock, flags);
3901 iommu_flush_irt(iommu, devid);
3902 iommu_completion_wait(iommu);
3907 static int modify_irte(u16 devid, int index, union irte *irte)
3909 struct irq_remap_table *table;
3910 struct amd_iommu *iommu;
3911 unsigned long flags;
3913 iommu = amd_iommu_rlookup_table[devid];
3917 table = get_irq_table(devid);
3921 raw_spin_lock_irqsave(&table->lock, flags);
3922 table->table[index] = irte->val;
3923 raw_spin_unlock_irqrestore(&table->lock, flags);
3925 iommu_flush_irt(iommu, devid);
3926 iommu_completion_wait(iommu);
3931 static void free_irte(u16 devid, int index)
3933 struct irq_remap_table *table;
3934 struct amd_iommu *iommu;
3935 unsigned long flags;
3937 iommu = amd_iommu_rlookup_table[devid];
3941 table = get_irq_table(devid);
3945 raw_spin_lock_irqsave(&table->lock, flags);
3946 iommu->irte_ops->clear_allocated(table, index);
3947 raw_spin_unlock_irqrestore(&table->lock, flags);
3949 iommu_flush_irt(iommu, devid);
3950 iommu_completion_wait(iommu);
3953 static void irte_prepare(void *entry,
3954 u32 delivery_mode, u32 dest_mode,
3955 u8 vector, u32 dest_apicid, int devid)
3957 union irte *irte = (union irte *) entry;
3960 irte->fields.vector = vector;
3961 irte->fields.int_type = delivery_mode;
3962 irte->fields.destination = dest_apicid;
3963 irte->fields.dm = dest_mode;
3964 irte->fields.valid = 1;
3967 static void irte_ga_prepare(void *entry,
3968 u32 delivery_mode, u32 dest_mode,
3969 u8 vector, u32 dest_apicid, int devid)
3971 struct irte_ga *irte = (struct irte_ga *) entry;
3975 irte->lo.fields_remap.int_type = delivery_mode;
3976 irte->lo.fields_remap.dm = dest_mode;
3977 irte->hi.fields.vector = vector;
3978 irte->lo.fields_remap.destination = APICID_TO_IRTE_DEST_LO(dest_apicid);
3979 irte->hi.fields.destination = APICID_TO_IRTE_DEST_HI(dest_apicid);
3980 irte->lo.fields_remap.valid = 1;
3983 static void irte_activate(void *entry, u16 devid, u16 index)
3985 union irte *irte = (union irte *) entry;
3987 irte->fields.valid = 1;
3988 modify_irte(devid, index, irte);
3991 static void irte_ga_activate(void *entry, u16 devid, u16 index)
3993 struct irte_ga *irte = (struct irte_ga *) entry;
3995 irte->lo.fields_remap.valid = 1;
3996 modify_irte_ga(devid, index, irte, NULL);
3999 static void irte_deactivate(void *entry, u16 devid, u16 index)
4001 union irte *irte = (union irte *) entry;
4003 irte->fields.valid = 0;
4004 modify_irte(devid, index, irte);
4007 static void irte_ga_deactivate(void *entry, u16 devid, u16 index)
4009 struct irte_ga *irte = (struct irte_ga *) entry;
4011 irte->lo.fields_remap.valid = 0;
4012 modify_irte_ga(devid, index, irte, NULL);
4015 static void irte_set_affinity(void *entry, u16 devid, u16 index,
4016 u8 vector, u32 dest_apicid)
4018 union irte *irte = (union irte *) entry;
4020 irte->fields.vector = vector;
4021 irte->fields.destination = dest_apicid;
4022 modify_irte(devid, index, irte);
4025 static void irte_ga_set_affinity(void *entry, u16 devid, u16 index,
4026 u8 vector, u32 dest_apicid)
4028 struct irte_ga *irte = (struct irte_ga *) entry;
4030 if (!irte->lo.fields_remap.guest_mode) {
4031 irte->hi.fields.vector = vector;
4032 irte->lo.fields_remap.destination =
4033 APICID_TO_IRTE_DEST_LO(dest_apicid);
4034 irte->hi.fields.destination =
4035 APICID_TO_IRTE_DEST_HI(dest_apicid);
4036 modify_irte_ga(devid, index, irte, NULL);
4040 #define IRTE_ALLOCATED (~1U)
4041 static void irte_set_allocated(struct irq_remap_table *table, int index)
4043 table->table[index] = IRTE_ALLOCATED;
4046 static void irte_ga_set_allocated(struct irq_remap_table *table, int index)
4048 struct irte_ga *ptr = (struct irte_ga *)table->table;
4049 struct irte_ga *irte = &ptr[index];
4051 memset(&irte->lo.val, 0, sizeof(u64));
4052 memset(&irte->hi.val, 0, sizeof(u64));
4053 irte->hi.fields.vector = 0xff;
4056 static bool irte_is_allocated(struct irq_remap_table *table, int index)
4058 union irte *ptr = (union irte *)table->table;
4059 union irte *irte = &ptr[index];
4061 return irte->val != 0;
4064 static bool irte_ga_is_allocated(struct irq_remap_table *table, int index)
4066 struct irte_ga *ptr = (struct irte_ga *)table->table;
4067 struct irte_ga *irte = &ptr[index];
4069 return irte->hi.fields.vector != 0;
4072 static void irte_clear_allocated(struct irq_remap_table *table, int index)
4074 table->table[index] = 0;
4077 static void irte_ga_clear_allocated(struct irq_remap_table *table, int index)
4079 struct irte_ga *ptr = (struct irte_ga *)table->table;
4080 struct irte_ga *irte = &ptr[index];
4082 memset(&irte->lo.val, 0, sizeof(u64));
4083 memset(&irte->hi.val, 0, sizeof(u64));
4086 static int get_devid(struct irq_alloc_info *info)
4090 switch (info->type) {
4091 case X86_IRQ_ALLOC_TYPE_IOAPIC:
4092 devid = get_ioapic_devid(info->ioapic_id);
4094 case X86_IRQ_ALLOC_TYPE_HPET:
4095 devid = get_hpet_devid(info->hpet_id);
4097 case X86_IRQ_ALLOC_TYPE_MSI:
4098 case X86_IRQ_ALLOC_TYPE_MSIX:
4099 devid = get_device_id(&info->msi_dev->dev);
4109 static struct irq_domain *get_ir_irq_domain(struct irq_alloc_info *info)
4111 struct amd_iommu *iommu;
4117 devid = get_devid(info);
4119 iommu = amd_iommu_rlookup_table[devid];
4121 return iommu->ir_domain;
4127 static struct irq_domain *get_irq_domain(struct irq_alloc_info *info)
4129 struct amd_iommu *iommu;
4135 switch (info->type) {
4136 case X86_IRQ_ALLOC_TYPE_MSI:
4137 case X86_IRQ_ALLOC_TYPE_MSIX:
4138 devid = get_device_id(&info->msi_dev->dev);
4142 iommu = amd_iommu_rlookup_table[devid];
4144 return iommu->msi_domain;
4153 struct irq_remap_ops amd_iommu_irq_ops = {
4154 .prepare = amd_iommu_prepare,
4155 .enable = amd_iommu_enable,
4156 .disable = amd_iommu_disable,
4157 .reenable = amd_iommu_reenable,
4158 .enable_faulting = amd_iommu_enable_faulting,
4159 .get_ir_irq_domain = get_ir_irq_domain,
4160 .get_irq_domain = get_irq_domain,
4163 static void irq_remapping_prepare_irte(struct amd_ir_data *data,
4164 struct irq_cfg *irq_cfg,
4165 struct irq_alloc_info *info,
4166 int devid, int index, int sub_handle)
4168 struct irq_2_irte *irte_info = &data->irq_2_irte;
4169 struct msi_msg *msg = &data->msi_entry;
4170 struct IO_APIC_route_entry *entry;
4171 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
4176 data->irq_2_irte.devid = devid;
4177 data->irq_2_irte.index = index + sub_handle;
4178 iommu->irte_ops->prepare(data->entry, apic->irq_delivery_mode,
4179 apic->irq_dest_mode, irq_cfg->vector,
4180 irq_cfg->dest_apicid, devid);
4182 switch (info->type) {
4183 case X86_IRQ_ALLOC_TYPE_IOAPIC:
4184 /* Setup IOAPIC entry */
4185 entry = info->ioapic_entry;
4186 info->ioapic_entry = NULL;
4187 memset(entry, 0, sizeof(*entry));
4188 entry->vector = index;
4190 entry->trigger = info->ioapic_trigger;
4191 entry->polarity = info->ioapic_polarity;
4192 /* Mask level triggered irqs. */
4193 if (info->ioapic_trigger)
4197 case X86_IRQ_ALLOC_TYPE_HPET:
4198 case X86_IRQ_ALLOC_TYPE_MSI:
4199 case X86_IRQ_ALLOC_TYPE_MSIX:
4200 msg->address_hi = MSI_ADDR_BASE_HI;
4201 msg->address_lo = MSI_ADDR_BASE_LO;
4202 msg->data = irte_info->index;
4211 struct amd_irte_ops irte_32_ops = {
4212 .prepare = irte_prepare,
4213 .activate = irte_activate,
4214 .deactivate = irte_deactivate,
4215 .set_affinity = irte_set_affinity,
4216 .set_allocated = irte_set_allocated,
4217 .is_allocated = irte_is_allocated,
4218 .clear_allocated = irte_clear_allocated,
4221 struct amd_irte_ops irte_128_ops = {
4222 .prepare = irte_ga_prepare,
4223 .activate = irte_ga_activate,
4224 .deactivate = irte_ga_deactivate,
4225 .set_affinity = irte_ga_set_affinity,
4226 .set_allocated = irte_ga_set_allocated,
4227 .is_allocated = irte_ga_is_allocated,
4228 .clear_allocated = irte_ga_clear_allocated,
4231 static int irq_remapping_alloc(struct irq_domain *domain, unsigned int virq,
4232 unsigned int nr_irqs, void *arg)
4234 struct irq_alloc_info *info = arg;
4235 struct irq_data *irq_data;
4236 struct amd_ir_data *data = NULL;
4237 struct irq_cfg *cfg;
4243 if (nr_irqs > 1 && info->type != X86_IRQ_ALLOC_TYPE_MSI &&
4244 info->type != X86_IRQ_ALLOC_TYPE_MSIX)
4248 * With IRQ remapping enabled, don't need contiguous CPU vectors
4249 * to support multiple MSI interrupts.
4251 if (info->type == X86_IRQ_ALLOC_TYPE_MSI)
4252 info->flags &= ~X86_IRQ_ALLOC_CONTIGUOUS_VECTORS;
4254 devid = get_devid(info);
4258 ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
4262 if (info->type == X86_IRQ_ALLOC_TYPE_IOAPIC) {
4263 struct irq_remap_table *table;
4264 struct amd_iommu *iommu;
4266 table = alloc_irq_table(devid, NULL);
4268 if (!table->min_index) {
4270 * Keep the first 32 indexes free for IOAPIC
4273 table->min_index = 32;
4274 iommu = amd_iommu_rlookup_table[devid];
4275 for (i = 0; i < 32; ++i)
4276 iommu->irte_ops->set_allocated(table, i);
4278 WARN_ON(table->min_index != 32);
4279 index = info->ioapic_pin;
4283 } else if (info->type == X86_IRQ_ALLOC_TYPE_MSI ||
4284 info->type == X86_IRQ_ALLOC_TYPE_MSIX) {
4285 bool align = (info->type == X86_IRQ_ALLOC_TYPE_MSI);
4287 index = alloc_irq_index(devid, nr_irqs, align, info->msi_dev);
4289 index = alloc_irq_index(devid, nr_irqs, false, NULL);
4293 pr_warn("Failed to allocate IRTE\n");
4295 goto out_free_parent;
4298 for (i = 0; i < nr_irqs; i++) {
4299 irq_data = irq_domain_get_irq_data(domain, virq + i);
4300 cfg = irqd_cfg(irq_data);
4301 if (!irq_data || !cfg) {
4307 data = kzalloc(sizeof(*data), GFP_KERNEL);
4311 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
4312 data->entry = kzalloc(sizeof(union irte), GFP_KERNEL);
4314 data->entry = kzalloc(sizeof(struct irte_ga),
4321 irq_data->hwirq = (devid << 16) + i;
4322 irq_data->chip_data = data;
4323 irq_data->chip = &amd_ir_chip;
4324 irq_remapping_prepare_irte(data, cfg, info, devid, index, i);
4325 irq_set_status_flags(virq + i, IRQ_MOVE_PCNTXT);
4331 for (i--; i >= 0; i--) {
4332 irq_data = irq_domain_get_irq_data(domain, virq + i);
4334 kfree(irq_data->chip_data);
4336 for (i = 0; i < nr_irqs; i++)
4337 free_irte(devid, index + i);
4339 irq_domain_free_irqs_common(domain, virq, nr_irqs);
4343 static void irq_remapping_free(struct irq_domain *domain, unsigned int virq,
4344 unsigned int nr_irqs)
4346 struct irq_2_irte *irte_info;
4347 struct irq_data *irq_data;
4348 struct amd_ir_data *data;
4351 for (i = 0; i < nr_irqs; i++) {
4352 irq_data = irq_domain_get_irq_data(domain, virq + i);
4353 if (irq_data && irq_data->chip_data) {
4354 data = irq_data->chip_data;
4355 irte_info = &data->irq_2_irte;
4356 free_irte(irte_info->devid, irte_info->index);
4361 irq_domain_free_irqs_common(domain, virq, nr_irqs);
4364 static void amd_ir_update_irte(struct irq_data *irqd, struct amd_iommu *iommu,
4365 struct amd_ir_data *ir_data,
4366 struct irq_2_irte *irte_info,
4367 struct irq_cfg *cfg);
4369 static int irq_remapping_activate(struct irq_domain *domain,
4370 struct irq_data *irq_data, bool reserve)
4372 struct amd_ir_data *data = irq_data->chip_data;
4373 struct irq_2_irte *irte_info = &data->irq_2_irte;
4374 struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4375 struct irq_cfg *cfg = irqd_cfg(irq_data);
4380 iommu->irte_ops->activate(data->entry, irte_info->devid,
4382 amd_ir_update_irte(irq_data, iommu, data, irte_info, cfg);
4386 static void irq_remapping_deactivate(struct irq_domain *domain,
4387 struct irq_data *irq_data)
4389 struct amd_ir_data *data = irq_data->chip_data;
4390 struct irq_2_irte *irte_info = &data->irq_2_irte;
4391 struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4394 iommu->irte_ops->deactivate(data->entry, irte_info->devid,
4398 static const struct irq_domain_ops amd_ir_domain_ops = {
4399 .alloc = irq_remapping_alloc,
4400 .free = irq_remapping_free,
4401 .activate = irq_remapping_activate,
4402 .deactivate = irq_remapping_deactivate,
4405 int amd_iommu_activate_guest_mode(void *data)
4407 struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
4408 struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
4410 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
4411 !entry || entry->lo.fields_vapic.guest_mode)
4417 entry->lo.fields_vapic.guest_mode = 1;
4418 entry->lo.fields_vapic.ga_log_intr = 1;
4419 entry->hi.fields.ga_root_ptr = ir_data->ga_root_ptr;
4420 entry->hi.fields.vector = ir_data->ga_vector;
4421 entry->lo.fields_vapic.ga_tag = ir_data->ga_tag;
4423 return modify_irte_ga(ir_data->irq_2_irte.devid,
4424 ir_data->irq_2_irte.index, entry, NULL);
4426 EXPORT_SYMBOL(amd_iommu_activate_guest_mode);
4428 int amd_iommu_deactivate_guest_mode(void *data)
4430 struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
4431 struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
4432 struct irq_cfg *cfg = ir_data->cfg;
4434 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
4435 !entry || !entry->lo.fields_vapic.guest_mode)
4441 entry->lo.fields_remap.dm = apic->irq_dest_mode;
4442 entry->lo.fields_remap.int_type = apic->irq_delivery_mode;
4443 entry->hi.fields.vector = cfg->vector;
4444 entry->lo.fields_remap.destination =
4445 APICID_TO_IRTE_DEST_LO(cfg->dest_apicid);
4446 entry->hi.fields.destination =
4447 APICID_TO_IRTE_DEST_HI(cfg->dest_apicid);
4449 return modify_irte_ga(ir_data->irq_2_irte.devid,
4450 ir_data->irq_2_irte.index, entry, NULL);
4452 EXPORT_SYMBOL(amd_iommu_deactivate_guest_mode);
4454 static int amd_ir_set_vcpu_affinity(struct irq_data *data, void *vcpu_info)
4457 struct amd_iommu *iommu;
4458 struct amd_iommu_pi_data *pi_data = vcpu_info;
4459 struct vcpu_data *vcpu_pi_info = pi_data->vcpu_data;
4460 struct amd_ir_data *ir_data = data->chip_data;
4461 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
4462 struct iommu_dev_data *dev_data = search_dev_data(irte_info->devid);
4465 * This device has never been set up for guest mode.
4466 * we should not modify the IRTE
4468 if (!dev_data || !dev_data->use_vapic)
4471 ir_data->cfg = irqd_cfg(data);
4472 pi_data->ir_data = ir_data;
4475 * SVM tries to set up for VAPIC mode, but we are in
4476 * legacy mode. So, we force legacy mode instead.
4478 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
4479 pr_debug("%s: Fall back to using intr legacy remap\n",
4481 pi_data->is_guest_mode = false;
4484 iommu = amd_iommu_rlookup_table[irte_info->devid];
4488 pi_data->prev_ga_tag = ir_data->cached_ga_tag;
4489 if (pi_data->is_guest_mode) {
4490 ir_data->ga_root_ptr = (pi_data->base >> 12);
4491 ir_data->ga_vector = vcpu_pi_info->vector;
4492 ir_data->ga_tag = pi_data->ga_tag;
4493 ret = amd_iommu_activate_guest_mode(ir_data);
4495 ir_data->cached_ga_tag = pi_data->ga_tag;
4497 ret = amd_iommu_deactivate_guest_mode(ir_data);
4500 * This communicates the ga_tag back to the caller
4501 * so that it can do all the necessary clean up.
4504 ir_data->cached_ga_tag = 0;
4511 static void amd_ir_update_irte(struct irq_data *irqd, struct amd_iommu *iommu,
4512 struct amd_ir_data *ir_data,
4513 struct irq_2_irte *irte_info,
4514 struct irq_cfg *cfg)
4518 * Atomically updates the IRTE with the new destination, vector
4519 * and flushes the interrupt entry cache.
4521 iommu->irte_ops->set_affinity(ir_data->entry, irte_info->devid,
4522 irte_info->index, cfg->vector,
4526 static int amd_ir_set_affinity(struct irq_data *data,
4527 const struct cpumask *mask, bool force)
4529 struct amd_ir_data *ir_data = data->chip_data;
4530 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
4531 struct irq_cfg *cfg = irqd_cfg(data);
4532 struct irq_data *parent = data->parent_data;
4533 struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4539 ret = parent->chip->irq_set_affinity(parent, mask, force);
4540 if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
4543 amd_ir_update_irte(data, iommu, ir_data, irte_info, cfg);
4545 * After this point, all the interrupts will start arriving
4546 * at the new destination. So, time to cleanup the previous
4547 * vector allocation.
4549 send_cleanup_vector(cfg);
4551 return IRQ_SET_MASK_OK_DONE;
4554 static void ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg)
4556 struct amd_ir_data *ir_data = irq_data->chip_data;
4558 *msg = ir_data->msi_entry;
4561 static struct irq_chip amd_ir_chip = {
4563 .irq_ack = apic_ack_irq,
4564 .irq_set_affinity = amd_ir_set_affinity,
4565 .irq_set_vcpu_affinity = amd_ir_set_vcpu_affinity,
4566 .irq_compose_msi_msg = ir_compose_msi_msg,
4569 int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
4571 struct fwnode_handle *fn;
4573 fn = irq_domain_alloc_named_id_fwnode("AMD-IR", iommu->index);
4576 iommu->ir_domain = irq_domain_create_tree(fn, &amd_ir_domain_ops, iommu);
4577 irq_domain_free_fwnode(fn);
4578 if (!iommu->ir_domain)
4581 iommu->ir_domain->parent = arch_get_ir_parent_domain();
4582 iommu->msi_domain = arch_create_remap_msi_irq_domain(iommu->ir_domain,
4588 int amd_iommu_update_ga(int cpu, bool is_run, void *data)
4590 unsigned long flags;
4591 struct amd_iommu *iommu;
4592 struct irq_remap_table *table;
4593 struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
4594 int devid = ir_data->irq_2_irte.devid;
4595 struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
4596 struct irte_ga *ref = (struct irte_ga *) ir_data->ref;
4598 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
4599 !ref || !entry || !entry->lo.fields_vapic.guest_mode)
4602 iommu = amd_iommu_rlookup_table[devid];
4606 table = get_irq_table(devid);
4610 raw_spin_lock_irqsave(&table->lock, flags);
4612 if (ref->lo.fields_vapic.guest_mode) {
4614 ref->lo.fields_vapic.destination =
4615 APICID_TO_IRTE_DEST_LO(cpu);
4616 ref->hi.fields.destination =
4617 APICID_TO_IRTE_DEST_HI(cpu);
4619 ref->lo.fields_vapic.is_run = is_run;
4623 raw_spin_unlock_irqrestore(&table->lock, flags);
4625 iommu_flush_irt(iommu, devid);
4626 iommu_completion_wait(iommu);
4629 EXPORT_SYMBOL(amd_iommu_update_ga);