2 * Copyright (C) 2007-2010 Advanced Micro Devices, Inc.
3 * Author: Joerg Roedel <jroedel@suse.de>
4 * Leo Duran <leo.duran@amd.com>
6 * This program is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 as published
8 * by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, write to the Free Software
17 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
20 #include <linux/ratelimit.h>
21 #include <linux/pci.h>
22 #include <linux/acpi.h>
23 #include <linux/amba/bus.h>
24 #include <linux/platform_device.h>
25 #include <linux/pci-ats.h>
26 #include <linux/bitmap.h>
27 #include <linux/slab.h>
28 #include <linux/debugfs.h>
29 #include <linux/scatterlist.h>
30 #include <linux/dma-mapping.h>
31 #include <linux/dma-direct.h>
32 #include <linux/iommu-helper.h>
33 #include <linux/iommu.h>
34 #include <linux/delay.h>
35 #include <linux/amd-iommu.h>
36 #include <linux/notifier.h>
37 #include <linux/export.h>
38 #include <linux/irq.h>
39 #include <linux/msi.h>
40 #include <linux/dma-contiguous.h>
41 #include <linux/irqdomain.h>
42 #include <linux/percpu.h>
43 #include <linux/iova.h>
44 #include <asm/irq_remapping.h>
45 #include <asm/io_apic.h>
47 #include <asm/hw_irq.h>
48 #include <asm/msidef.h>
49 #include <asm/proto.h>
50 #include <asm/iommu.h>
54 #include "amd_iommu_proto.h"
55 #include "amd_iommu_types.h"
56 #include "irq_remapping.h"
58 #define AMD_IOMMU_MAPPING_ERROR 0
60 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
62 #define LOOP_TIMEOUT 100000
64 /* IO virtual address start page frame number */
65 #define IOVA_START_PFN (1)
66 #define IOVA_PFN(addr) ((addr) >> PAGE_SHIFT)
68 /* Reserved IOVA ranges */
69 #define MSI_RANGE_START (0xfee00000)
70 #define MSI_RANGE_END (0xfeefffff)
71 #define HT_RANGE_START (0xfd00000000ULL)
72 #define HT_RANGE_END (0xffffffffffULL)
75 * This bitmap is used to advertise the page sizes our hardware support
76 * to the IOMMU core, which will then use this information to split
77 * physically contiguous memory regions it is mapping into page sizes
80 * 512GB Pages are not supported due to a hardware bug
82 #define AMD_IOMMU_PGSIZES ((~0xFFFUL) & ~(2ULL << 38))
84 static DEFINE_SPINLOCK(amd_iommu_devtable_lock);
85 static DEFINE_SPINLOCK(pd_bitmap_lock);
87 /* List of all available dev_data structures */
88 static LLIST_HEAD(dev_data_list);
90 LIST_HEAD(ioapic_map);
92 LIST_HEAD(acpihid_map);
95 * Domain for untranslated devices - only allocated
96 * if iommu=pt passed on kernel cmd line.
98 const struct iommu_ops amd_iommu_ops;
100 static ATOMIC_NOTIFIER_HEAD(ppr_notifier);
101 int amd_iommu_max_glx_val = -1;
103 static const struct dma_map_ops amd_iommu_dma_ops;
106 * general struct to manage commands send to an IOMMU
112 struct kmem_cache *amd_iommu_irq_cache;
114 static void update_domain(struct protection_domain *domain);
115 static int protection_domain_init(struct protection_domain *domain);
116 static void detach_device(struct device *dev);
117 static void iova_domain_flush_tlb(struct iova_domain *iovad);
120 * Data container for a dma_ops specific protection domain
122 struct dma_ops_domain {
123 /* generic protection domain information */
124 struct protection_domain domain;
127 struct iova_domain iovad;
130 static struct iova_domain reserved_iova_ranges;
131 static struct lock_class_key reserved_rbtree_key;
133 /****************************************************************************
137 ****************************************************************************/
139 static inline int match_hid_uid(struct device *dev,
140 struct acpihid_map_entry *entry)
142 const char *hid, *uid;
144 hid = acpi_device_hid(ACPI_COMPANION(dev));
145 uid = acpi_device_uid(ACPI_COMPANION(dev));
151 return strcmp(hid, entry->hid);
154 return strcmp(hid, entry->hid);
156 return (strcmp(hid, entry->hid) || strcmp(uid, entry->uid));
159 static inline u16 get_pci_device_id(struct device *dev)
161 struct pci_dev *pdev = to_pci_dev(dev);
163 return PCI_DEVID(pdev->bus->number, pdev->devfn);
166 static inline int get_acpihid_device_id(struct device *dev,
167 struct acpihid_map_entry **entry)
169 struct acpihid_map_entry *p;
171 list_for_each_entry(p, &acpihid_map, list) {
172 if (!match_hid_uid(dev, p)) {
181 static inline int get_device_id(struct device *dev)
186 devid = get_pci_device_id(dev);
188 devid = get_acpihid_device_id(dev, NULL);
193 static struct protection_domain *to_pdomain(struct iommu_domain *dom)
195 return container_of(dom, struct protection_domain, domain);
198 static struct dma_ops_domain* to_dma_ops_domain(struct protection_domain *domain)
200 BUG_ON(domain->flags != PD_DMA_OPS_MASK);
201 return container_of(domain, struct dma_ops_domain, domain);
204 static struct iommu_dev_data *alloc_dev_data(u16 devid)
206 struct iommu_dev_data *dev_data;
208 dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
212 dev_data->devid = devid;
213 ratelimit_default_init(&dev_data->rs);
215 llist_add(&dev_data->dev_data_list, &dev_data_list);
219 static struct iommu_dev_data *search_dev_data(u16 devid)
221 struct iommu_dev_data *dev_data;
222 struct llist_node *node;
224 if (llist_empty(&dev_data_list))
227 node = dev_data_list.first;
228 llist_for_each_entry(dev_data, node, dev_data_list) {
229 if (dev_data->devid == devid)
236 static int __last_alias(struct pci_dev *pdev, u16 alias, void *data)
238 *(u16 *)data = alias;
242 static u16 get_alias(struct device *dev)
244 struct pci_dev *pdev = to_pci_dev(dev);
245 u16 devid, ivrs_alias, pci_alias;
247 /* The callers make sure that get_device_id() does not fail here */
248 devid = get_device_id(dev);
250 /* For ACPI HID devices, we simply return the devid as such */
251 if (!dev_is_pci(dev))
254 ivrs_alias = amd_iommu_alias_table[devid];
256 pci_for_each_dma_alias(pdev, __last_alias, &pci_alias);
258 if (ivrs_alias == pci_alias)
264 * The IVRS is fairly reliable in telling us about aliases, but it
265 * can't know about every screwy device. If we don't have an IVRS
266 * reported alias, use the PCI reported alias. In that case we may
267 * still need to initialize the rlookup and dev_table entries if the
268 * alias is to a non-existent device.
270 if (ivrs_alias == devid) {
271 if (!amd_iommu_rlookup_table[pci_alias]) {
272 amd_iommu_rlookup_table[pci_alias] =
273 amd_iommu_rlookup_table[devid];
274 memcpy(amd_iommu_dev_table[pci_alias].data,
275 amd_iommu_dev_table[devid].data,
276 sizeof(amd_iommu_dev_table[pci_alias].data));
282 pr_info("AMD-Vi: Using IVRS reported alias %02x:%02x.%d "
283 "for device %s[%04x:%04x], kernel reported alias "
284 "%02x:%02x.%d\n", PCI_BUS_NUM(ivrs_alias), PCI_SLOT(ivrs_alias),
285 PCI_FUNC(ivrs_alias), dev_name(dev), pdev->vendor, pdev->device,
286 PCI_BUS_NUM(pci_alias), PCI_SLOT(pci_alias),
287 PCI_FUNC(pci_alias));
290 * If we don't have a PCI DMA alias and the IVRS alias is on the same
291 * bus, then the IVRS table may know about a quirk that we don't.
293 if (pci_alias == devid &&
294 PCI_BUS_NUM(ivrs_alias) == pdev->bus->number) {
295 pci_add_dma_alias(pdev, ivrs_alias & 0xff);
296 pr_info("AMD-Vi: Added PCI DMA alias %02x.%d for %s\n",
297 PCI_SLOT(ivrs_alias), PCI_FUNC(ivrs_alias),
304 static struct iommu_dev_data *find_dev_data(u16 devid)
306 struct iommu_dev_data *dev_data;
307 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
309 dev_data = search_dev_data(devid);
311 if (dev_data == NULL) {
312 dev_data = alloc_dev_data(devid);
316 if (translation_pre_enabled(iommu))
317 dev_data->defer_attach = true;
323 struct iommu_dev_data *get_dev_data(struct device *dev)
325 return dev->archdata.iommu;
327 EXPORT_SYMBOL(get_dev_data);
330 * Find or create an IOMMU group for a acpihid device.
332 static struct iommu_group *acpihid_device_group(struct device *dev)
334 struct acpihid_map_entry *p, *entry = NULL;
337 devid = get_acpihid_device_id(dev, &entry);
339 return ERR_PTR(devid);
341 list_for_each_entry(p, &acpihid_map, list) {
342 if ((devid == p->devid) && p->group)
343 entry->group = p->group;
347 entry->group = generic_device_group(dev);
349 iommu_group_ref_get(entry->group);
354 static bool pci_iommuv2_capable(struct pci_dev *pdev)
356 static const int caps[] = {
359 PCI_EXT_CAP_ID_PASID,
363 if (pci_ats_disabled())
366 for (i = 0; i < 3; ++i) {
367 pos = pci_find_ext_capability(pdev, caps[i]);
375 static bool pdev_pri_erratum(struct pci_dev *pdev, u32 erratum)
377 struct iommu_dev_data *dev_data;
379 dev_data = get_dev_data(&pdev->dev);
381 return dev_data->errata & (1 << erratum) ? true : false;
385 * This function checks if the driver got a valid device from the caller to
386 * avoid dereferencing invalid pointers.
388 static bool check_device(struct device *dev)
392 if (!dev || !dev->dma_mask)
395 devid = get_device_id(dev);
399 /* Out of our scope? */
400 if (devid > amd_iommu_last_bdf)
403 if (amd_iommu_rlookup_table[devid] == NULL)
409 static void init_iommu_group(struct device *dev)
411 struct iommu_group *group;
413 group = iommu_group_get_for_dev(dev);
417 iommu_group_put(group);
420 static int iommu_init_device(struct device *dev)
422 struct iommu_dev_data *dev_data;
423 struct amd_iommu *iommu;
426 if (dev->archdata.iommu)
429 devid = get_device_id(dev);
433 iommu = amd_iommu_rlookup_table[devid];
435 dev_data = find_dev_data(devid);
439 dev_data->alias = get_alias(dev);
441 if (dev_is_pci(dev) && pci_iommuv2_capable(to_pci_dev(dev))) {
442 struct amd_iommu *iommu;
444 iommu = amd_iommu_rlookup_table[dev_data->devid];
445 dev_data->iommu_v2 = iommu->is_iommu_v2;
448 dev->archdata.iommu = dev_data;
450 iommu_device_link(&iommu->iommu, dev);
455 static void iommu_ignore_device(struct device *dev)
460 devid = get_device_id(dev);
464 alias = get_alias(dev);
466 memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
467 memset(&amd_iommu_dev_table[alias], 0, sizeof(struct dev_table_entry));
469 amd_iommu_rlookup_table[devid] = NULL;
470 amd_iommu_rlookup_table[alias] = NULL;
473 static void iommu_uninit_device(struct device *dev)
475 struct iommu_dev_data *dev_data;
476 struct amd_iommu *iommu;
479 devid = get_device_id(dev);
483 iommu = amd_iommu_rlookup_table[devid];
485 dev_data = search_dev_data(devid);
489 if (dev_data->domain)
492 iommu_device_unlink(&iommu->iommu, dev);
494 iommu_group_remove_device(dev);
500 * We keep dev_data around for unplugged devices and reuse it when the
501 * device is re-plugged - not doing so would introduce a ton of races.
505 /****************************************************************************
507 * Interrupt handling functions
509 ****************************************************************************/
511 static void dump_dte_entry(u16 devid)
515 for (i = 0; i < 4; ++i)
516 pr_err("AMD-Vi: DTE[%d]: %016llx\n", i,
517 amd_iommu_dev_table[devid].data[i]);
520 static void dump_command(unsigned long phys_addr)
522 struct iommu_cmd *cmd = iommu_phys_to_virt(phys_addr);
525 for (i = 0; i < 4; ++i)
526 pr_err("AMD-Vi: CMD[%d]: %08x\n", i, cmd->data[i]);
529 static void amd_iommu_report_page_fault(u16 devid, u16 domain_id,
530 u64 address, int flags)
532 struct iommu_dev_data *dev_data = NULL;
533 struct pci_dev *pdev;
535 pdev = pci_get_domain_bus_and_slot(0, PCI_BUS_NUM(devid),
538 dev_data = get_dev_data(&pdev->dev);
540 if (dev_data && __ratelimit(&dev_data->rs)) {
541 dev_err(&pdev->dev, "AMD-Vi: Event logged [IO_PAGE_FAULT domain=0x%04x address=0x%016llx flags=0x%04x]\n",
542 domain_id, address, flags);
543 } else if (printk_ratelimit()) {
544 pr_err("AMD-Vi: Event logged [IO_PAGE_FAULT device=%02x:%02x.%x domain=0x%04x address=0x%016llx flags=0x%04x]\n",
545 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
546 domain_id, address, flags);
553 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
555 struct device *dev = iommu->iommu.dev;
556 int type, devid, pasid, flags, tag;
557 volatile u32 *event = __evt;
562 type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
563 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
564 pasid = PPR_PASID(*(u64 *)&event[0]);
565 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
566 address = (u64)(((u64)event[3]) << 32) | event[2];
569 /* Did we hit the erratum? */
570 if (++count == LOOP_TIMEOUT) {
571 pr_err("AMD-Vi: No event written to event log\n");
578 if (type == EVENT_TYPE_IO_FAULT) {
579 amd_iommu_report_page_fault(devid, pasid, address, flags);
582 dev_err(dev, "AMD-Vi: Event logged [");
586 case EVENT_TYPE_ILL_DEV:
587 dev_err(dev, "ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x pasid=0x%05x address=0x%016llx flags=0x%04x]\n",
588 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
589 pasid, address, flags);
590 dump_dte_entry(devid);
592 case EVENT_TYPE_DEV_TAB_ERR:
593 dev_err(dev, "DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
594 "address=0x%016llx flags=0x%04x]\n",
595 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
598 case EVENT_TYPE_PAGE_TAB_ERR:
599 dev_err(dev, "PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x domain=0x%04x address=0x%016llx flags=0x%04x]\n",
600 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
601 pasid, address, flags);
603 case EVENT_TYPE_ILL_CMD:
604 dev_err(dev, "ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
605 dump_command(address);
607 case EVENT_TYPE_CMD_HARD_ERR:
608 dev_err(dev, "COMMAND_HARDWARE_ERROR address=0x%016llx flags=0x%04x]\n",
611 case EVENT_TYPE_IOTLB_INV_TO:
612 dev_err(dev, "IOTLB_INV_TIMEOUT device=%02x:%02x.%x address=0x%016llx]\n",
613 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
616 case EVENT_TYPE_INV_DEV_REQ:
617 dev_err(dev, "INVALID_DEVICE_REQUEST device=%02x:%02x.%x pasid=0x%05x address=0x%016llx flags=0x%04x]\n",
618 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
619 pasid, address, flags);
621 case EVENT_TYPE_INV_PPR_REQ:
622 pasid = ((event[0] >> 16) & 0xFFFF)
623 | ((event[1] << 6) & 0xF0000);
624 tag = event[1] & 0x03FF;
625 dev_err(dev, "INVALID_PPR_REQUEST device=%02x:%02x.%x pasid=0x%05x address=0x%016llx flags=0x%04x]\n",
626 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
627 pasid, address, flags);
630 dev_err(dev, "UNKNOWN event[0]=0x%08x event[1]=0x%08x event[2]=0x%08x event[3]=0x%08x\n",
631 event[0], event[1], event[2], event[3]);
634 memset(__evt, 0, 4 * sizeof(u32));
637 static void iommu_poll_events(struct amd_iommu *iommu)
641 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
642 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
644 while (head != tail) {
645 iommu_print_event(iommu, iommu->evt_buf + head);
646 head = (head + EVENT_ENTRY_SIZE) % EVT_BUFFER_SIZE;
649 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
652 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
654 struct amd_iommu_fault fault;
656 if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
657 pr_err_ratelimited("AMD-Vi: Unknown PPR request received\n");
661 fault.address = raw[1];
662 fault.pasid = PPR_PASID(raw[0]);
663 fault.device_id = PPR_DEVID(raw[0]);
664 fault.tag = PPR_TAG(raw[0]);
665 fault.flags = PPR_FLAGS(raw[0]);
667 atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
670 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
674 if (iommu->ppr_log == NULL)
677 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
678 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
680 while (head != tail) {
685 raw = (u64 *)(iommu->ppr_log + head);
688 * Hardware bug: Interrupt may arrive before the entry is
689 * written to memory. If this happens we need to wait for the
692 for (i = 0; i < LOOP_TIMEOUT; ++i) {
693 if (PPR_REQ_TYPE(raw[0]) != 0)
698 /* Avoid memcpy function-call overhead */
703 * To detect the hardware bug we need to clear the entry
706 raw[0] = raw[1] = 0UL;
708 /* Update head pointer of hardware ring-buffer */
709 head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
710 writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
712 /* Handle PPR entry */
713 iommu_handle_ppr_entry(iommu, entry);
715 /* Refresh ring-buffer information */
716 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
717 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
721 #ifdef CONFIG_IRQ_REMAP
722 static int (*iommu_ga_log_notifier)(u32);
724 int amd_iommu_register_ga_log_notifier(int (*notifier)(u32))
726 iommu_ga_log_notifier = notifier;
730 EXPORT_SYMBOL(amd_iommu_register_ga_log_notifier);
732 static void iommu_poll_ga_log(struct amd_iommu *iommu)
734 u32 head, tail, cnt = 0;
736 if (iommu->ga_log == NULL)
739 head = readl(iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
740 tail = readl(iommu->mmio_base + MMIO_GA_TAIL_OFFSET);
742 while (head != tail) {
746 raw = (u64 *)(iommu->ga_log + head);
749 /* Avoid memcpy function-call overhead */
752 /* Update head pointer of hardware ring-buffer */
753 head = (head + GA_ENTRY_SIZE) % GA_LOG_SIZE;
754 writel(head, iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
756 /* Handle GA entry */
757 switch (GA_REQ_TYPE(log_entry)) {
759 if (!iommu_ga_log_notifier)
762 pr_debug("AMD-Vi: %s: devid=%#x, ga_tag=%#x\n",
763 __func__, GA_DEVID(log_entry),
766 if (iommu_ga_log_notifier(GA_TAG(log_entry)) != 0)
767 pr_err("AMD-Vi: GA log notifier failed.\n");
774 #endif /* CONFIG_IRQ_REMAP */
776 #define AMD_IOMMU_INT_MASK \
777 (MMIO_STATUS_EVT_INT_MASK | \
778 MMIO_STATUS_PPR_INT_MASK | \
779 MMIO_STATUS_GALOG_INT_MASK)
781 irqreturn_t amd_iommu_int_thread(int irq, void *data)
783 struct amd_iommu *iommu = (struct amd_iommu *) data;
784 u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
786 while (status & AMD_IOMMU_INT_MASK) {
787 /* Enable EVT and PPR and GA interrupts again */
788 writel(AMD_IOMMU_INT_MASK,
789 iommu->mmio_base + MMIO_STATUS_OFFSET);
791 if (status & MMIO_STATUS_EVT_INT_MASK) {
792 pr_devel("AMD-Vi: Processing IOMMU Event Log\n");
793 iommu_poll_events(iommu);
796 if (status & MMIO_STATUS_PPR_INT_MASK) {
797 pr_devel("AMD-Vi: Processing IOMMU PPR Log\n");
798 iommu_poll_ppr_log(iommu);
801 #ifdef CONFIG_IRQ_REMAP
802 if (status & MMIO_STATUS_GALOG_INT_MASK) {
803 pr_devel("AMD-Vi: Processing IOMMU GA Log\n");
804 iommu_poll_ga_log(iommu);
809 * Hardware bug: ERBT1312
810 * When re-enabling interrupt (by writing 1
811 * to clear the bit), the hardware might also try to set
812 * the interrupt bit in the event status register.
813 * In this scenario, the bit will be set, and disable
814 * subsequent interrupts.
816 * Workaround: The IOMMU driver should read back the
817 * status register and check if the interrupt bits are cleared.
818 * If not, driver will need to go through the interrupt handler
819 * again and re-clear the bits
821 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
826 irqreturn_t amd_iommu_int_handler(int irq, void *data)
828 return IRQ_WAKE_THREAD;
831 /****************************************************************************
833 * IOMMU command queuing functions
835 ****************************************************************************/
837 static int wait_on_sem(volatile u64 *sem)
841 while (*sem == 0 && i < LOOP_TIMEOUT) {
846 if (i == LOOP_TIMEOUT) {
847 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
854 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
855 struct iommu_cmd *cmd)
859 target = iommu->cmd_buf + iommu->cmd_buf_tail;
861 iommu->cmd_buf_tail += sizeof(*cmd);
862 iommu->cmd_buf_tail %= CMD_BUFFER_SIZE;
864 /* Copy command to buffer */
865 memcpy(target, cmd, sizeof(*cmd));
867 /* Tell the IOMMU about it */
868 writel(iommu->cmd_buf_tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
871 static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
873 u64 paddr = iommu_virt_to_phys((void *)address);
875 WARN_ON(address & 0x7ULL);
877 memset(cmd, 0, sizeof(*cmd));
878 cmd->data[0] = lower_32_bits(paddr) | CMD_COMPL_WAIT_STORE_MASK;
879 cmd->data[1] = upper_32_bits(paddr);
881 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
884 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
886 memset(cmd, 0, sizeof(*cmd));
887 cmd->data[0] = devid;
888 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
891 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
892 size_t size, u16 domid, int pde)
897 pages = iommu_num_pages(address, size, PAGE_SIZE);
902 * If we have to flush more than one page, flush all
903 * TLB entries for this domain
905 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
909 address &= PAGE_MASK;
911 memset(cmd, 0, sizeof(*cmd));
912 cmd->data[1] |= domid;
913 cmd->data[2] = lower_32_bits(address);
914 cmd->data[3] = upper_32_bits(address);
915 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
916 if (s) /* size bit - we flush more than one 4kb page */
917 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
918 if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
919 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
922 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
923 u64 address, size_t size)
928 pages = iommu_num_pages(address, size, PAGE_SIZE);
933 * If we have to flush more than one page, flush all
934 * TLB entries for this domain
936 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
940 address &= PAGE_MASK;
942 memset(cmd, 0, sizeof(*cmd));
943 cmd->data[0] = devid;
944 cmd->data[0] |= (qdep & 0xff) << 24;
945 cmd->data[1] = devid;
946 cmd->data[2] = lower_32_bits(address);
947 cmd->data[3] = upper_32_bits(address);
948 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
950 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
953 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, int pasid,
954 u64 address, bool size)
956 memset(cmd, 0, sizeof(*cmd));
958 address &= ~(0xfffULL);
960 cmd->data[0] = pasid;
961 cmd->data[1] = domid;
962 cmd->data[2] = lower_32_bits(address);
963 cmd->data[3] = upper_32_bits(address);
964 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
965 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
967 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
968 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
971 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, int pasid,
972 int qdep, u64 address, bool size)
974 memset(cmd, 0, sizeof(*cmd));
976 address &= ~(0xfffULL);
978 cmd->data[0] = devid;
979 cmd->data[0] |= ((pasid >> 8) & 0xff) << 16;
980 cmd->data[0] |= (qdep & 0xff) << 24;
981 cmd->data[1] = devid;
982 cmd->data[1] |= (pasid & 0xff) << 16;
983 cmd->data[2] = lower_32_bits(address);
984 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
985 cmd->data[3] = upper_32_bits(address);
987 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
988 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
991 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, int pasid,
992 int status, int tag, bool gn)
994 memset(cmd, 0, sizeof(*cmd));
996 cmd->data[0] = devid;
998 cmd->data[1] = pasid;
999 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
1001 cmd->data[3] = tag & 0x1ff;
1002 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
1004 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
1007 static void build_inv_all(struct iommu_cmd *cmd)
1009 memset(cmd, 0, sizeof(*cmd));
1010 CMD_SET_TYPE(cmd, CMD_INV_ALL);
1013 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
1015 memset(cmd, 0, sizeof(*cmd));
1016 cmd->data[0] = devid;
1017 CMD_SET_TYPE(cmd, CMD_INV_IRT);
1021 * Writes the command to the IOMMUs command buffer and informs the
1022 * hardware about the new command.
1024 static int __iommu_queue_command_sync(struct amd_iommu *iommu,
1025 struct iommu_cmd *cmd,
1028 unsigned int count = 0;
1029 u32 left, next_tail;
1031 next_tail = (iommu->cmd_buf_tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
1033 left = (iommu->cmd_buf_head - next_tail) % CMD_BUFFER_SIZE;
1036 /* Skip udelay() the first time around */
1038 if (count == LOOP_TIMEOUT) {
1039 pr_err("AMD-Vi: Command buffer timeout\n");
1046 /* Update head and recheck remaining space */
1047 iommu->cmd_buf_head = readl(iommu->mmio_base +
1048 MMIO_CMD_HEAD_OFFSET);
1053 copy_cmd_to_buffer(iommu, cmd);
1055 /* Do we need to make sure all commands are processed? */
1056 iommu->need_sync = sync;
1061 static int iommu_queue_command_sync(struct amd_iommu *iommu,
1062 struct iommu_cmd *cmd,
1065 unsigned long flags;
1068 raw_spin_lock_irqsave(&iommu->lock, flags);
1069 ret = __iommu_queue_command_sync(iommu, cmd, sync);
1070 raw_spin_unlock_irqrestore(&iommu->lock, flags);
1075 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
1077 return iommu_queue_command_sync(iommu, cmd, true);
1081 * This function queues a completion wait command into the command
1082 * buffer of an IOMMU
1084 static int iommu_completion_wait(struct amd_iommu *iommu)
1086 struct iommu_cmd cmd;
1087 unsigned long flags;
1090 if (!iommu->need_sync)
1094 build_completion_wait(&cmd, (u64)&iommu->cmd_sem);
1096 raw_spin_lock_irqsave(&iommu->lock, flags);
1100 ret = __iommu_queue_command_sync(iommu, &cmd, false);
1104 ret = wait_on_sem(&iommu->cmd_sem);
1107 raw_spin_unlock_irqrestore(&iommu->lock, flags);
1112 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
1114 struct iommu_cmd cmd;
1116 build_inv_dte(&cmd, devid);
1118 return iommu_queue_command(iommu, &cmd);
1121 static void amd_iommu_flush_dte_all(struct amd_iommu *iommu)
1125 for (devid = 0; devid <= 0xffff; ++devid)
1126 iommu_flush_dte(iommu, devid);
1128 iommu_completion_wait(iommu);
1132 * This function uses heavy locking and may disable irqs for some time. But
1133 * this is no issue because it is only called during resume.
1135 static void amd_iommu_flush_tlb_all(struct amd_iommu *iommu)
1139 for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
1140 struct iommu_cmd cmd;
1141 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1143 iommu_queue_command(iommu, &cmd);
1146 iommu_completion_wait(iommu);
1149 static void amd_iommu_flush_all(struct amd_iommu *iommu)
1151 struct iommu_cmd cmd;
1153 build_inv_all(&cmd);
1155 iommu_queue_command(iommu, &cmd);
1156 iommu_completion_wait(iommu);
1159 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1161 struct iommu_cmd cmd;
1163 build_inv_irt(&cmd, devid);
1165 iommu_queue_command(iommu, &cmd);
1168 static void amd_iommu_flush_irt_all(struct amd_iommu *iommu)
1172 for (devid = 0; devid <= MAX_DEV_TABLE_ENTRIES; devid++)
1173 iommu_flush_irt(iommu, devid);
1175 iommu_completion_wait(iommu);
1178 void iommu_flush_all_caches(struct amd_iommu *iommu)
1180 if (iommu_feature(iommu, FEATURE_IA)) {
1181 amd_iommu_flush_all(iommu);
1183 amd_iommu_flush_dte_all(iommu);
1184 amd_iommu_flush_irt_all(iommu);
1185 amd_iommu_flush_tlb_all(iommu);
1190 * Command send function for flushing on-device TLB
1192 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1193 u64 address, size_t size)
1195 struct amd_iommu *iommu;
1196 struct iommu_cmd cmd;
1199 qdep = dev_data->ats.qdep;
1200 iommu = amd_iommu_rlookup_table[dev_data->devid];
1202 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1204 return iommu_queue_command(iommu, &cmd);
1208 * Command send function for invalidating a device table entry
1210 static int device_flush_dte(struct iommu_dev_data *dev_data)
1212 struct amd_iommu *iommu;
1216 iommu = amd_iommu_rlookup_table[dev_data->devid];
1217 alias = dev_data->alias;
1219 ret = iommu_flush_dte(iommu, dev_data->devid);
1220 if (!ret && alias != dev_data->devid)
1221 ret = iommu_flush_dte(iommu, alias);
1225 if (dev_data->ats.enabled)
1226 ret = device_flush_iotlb(dev_data, 0, ~0UL);
1232 * TLB invalidation function which is called from the mapping functions.
1233 * It invalidates a single PTE if the range to flush is within a single
1234 * page. Otherwise it flushes the whole TLB of the IOMMU.
1236 static void __domain_flush_pages(struct protection_domain *domain,
1237 u64 address, size_t size, int pde)
1239 struct iommu_dev_data *dev_data;
1240 struct iommu_cmd cmd;
1243 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1245 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
1246 if (!domain->dev_iommu[i])
1250 * Devices of this domain are behind this IOMMU
1251 * We need a TLB flush
1253 ret |= iommu_queue_command(amd_iommus[i], &cmd);
1256 list_for_each_entry(dev_data, &domain->dev_list, list) {
1258 if (!dev_data->ats.enabled)
1261 ret |= device_flush_iotlb(dev_data, address, size);
1267 static void domain_flush_pages(struct protection_domain *domain,
1268 u64 address, size_t size)
1270 __domain_flush_pages(domain, address, size, 0);
1273 /* Flush the whole IO/TLB for a given protection domain */
1274 static void domain_flush_tlb(struct protection_domain *domain)
1276 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
1279 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1280 static void domain_flush_tlb_pde(struct protection_domain *domain)
1282 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1285 static void domain_flush_complete(struct protection_domain *domain)
1289 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
1290 if (domain && !domain->dev_iommu[i])
1294 * Devices of this domain are behind this IOMMU
1295 * We need to wait for completion of all commands.
1297 iommu_completion_wait(amd_iommus[i]);
1303 * This function flushes the DTEs for all devices in domain
1305 static void domain_flush_devices(struct protection_domain *domain)
1307 struct iommu_dev_data *dev_data;
1309 list_for_each_entry(dev_data, &domain->dev_list, list)
1310 device_flush_dte(dev_data);
1313 /****************************************************************************
1315 * The functions below are used the create the page table mappings for
1316 * unity mapped regions.
1318 ****************************************************************************/
1321 * This function is used to add another level to an IO page table. Adding
1322 * another level increases the size of the address space by 9 bits to a size up
1325 static bool increase_address_space(struct protection_domain *domain,
1330 if (domain->mode == PAGE_MODE_6_LEVEL)
1331 /* address space already 64 bit large */
1334 pte = (void *)get_zeroed_page(gfp);
1338 *pte = PM_LEVEL_PDE(domain->mode,
1339 iommu_virt_to_phys(domain->pt_root));
1340 domain->pt_root = pte;
1342 domain->updated = true;
1347 static u64 *alloc_pte(struct protection_domain *domain,
1348 unsigned long address,
1349 unsigned long page_size,
1356 BUG_ON(!is_power_of_2(page_size));
1358 while (address > PM_LEVEL_SIZE(domain->mode))
1359 increase_address_space(domain, gfp);
1361 level = domain->mode - 1;
1362 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1363 address = PAGE_SIZE_ALIGN(address, page_size);
1364 end_lvl = PAGE_SIZE_LEVEL(page_size);
1366 while (level > end_lvl) {
1371 if (!IOMMU_PTE_PRESENT(__pte)) {
1372 page = (u64 *)get_zeroed_page(gfp);
1376 __npte = PM_LEVEL_PDE(level, iommu_virt_to_phys(page));
1378 /* pte could have been changed somewhere. */
1379 if (cmpxchg64(pte, __pte, __npte) != __pte) {
1380 free_page((unsigned long)page);
1385 /* No level skipping support yet */
1386 if (PM_PTE_LEVEL(*pte) != level)
1391 pte = IOMMU_PTE_PAGE(*pte);
1393 if (pte_page && level == end_lvl)
1396 pte = &pte[PM_LEVEL_INDEX(level, address)];
1403 * This function checks if there is a PTE for a given dma address. If
1404 * there is one, it returns the pointer to it.
1406 static u64 *fetch_pte(struct protection_domain *domain,
1407 unsigned long address,
1408 unsigned long *page_size)
1415 if (address > PM_LEVEL_SIZE(domain->mode))
1418 level = domain->mode - 1;
1419 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1420 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1425 if (!IOMMU_PTE_PRESENT(*pte))
1429 if (PM_PTE_LEVEL(*pte) == 7 ||
1430 PM_PTE_LEVEL(*pte) == 0)
1433 /* No level skipping support yet */
1434 if (PM_PTE_LEVEL(*pte) != level)
1439 /* Walk to the next level */
1440 pte = IOMMU_PTE_PAGE(*pte);
1441 pte = &pte[PM_LEVEL_INDEX(level, address)];
1442 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1445 if (PM_PTE_LEVEL(*pte) == 0x07) {
1446 unsigned long pte_mask;
1449 * If we have a series of large PTEs, make
1450 * sure to return a pointer to the first one.
1452 *page_size = pte_mask = PTE_PAGE_SIZE(*pte);
1453 pte_mask = ~((PAGE_SIZE_PTE_COUNT(pte_mask) << 3) - 1);
1454 pte = (u64 *)(((unsigned long)pte) & pte_mask);
1461 * Generic mapping functions. It maps a physical address into a DMA
1462 * address space. It allocates the page table pages if necessary.
1463 * In the future it can be extended to a generic mapping function
1464 * supporting all features of AMD IOMMU page tables like level skipping
1465 * and full 64 bit address spaces.
1467 static int iommu_map_page(struct protection_domain *dom,
1468 unsigned long bus_addr,
1469 unsigned long phys_addr,
1470 unsigned long page_size,
1477 BUG_ON(!IS_ALIGNED(bus_addr, page_size));
1478 BUG_ON(!IS_ALIGNED(phys_addr, page_size));
1480 if (!(prot & IOMMU_PROT_MASK))
1483 count = PAGE_SIZE_PTE_COUNT(page_size);
1484 pte = alloc_pte(dom, bus_addr, page_size, NULL, gfp);
1489 for (i = 0; i < count; ++i)
1490 if (IOMMU_PTE_PRESENT(pte[i]))
1494 __pte = PAGE_SIZE_PTE(__sme_set(phys_addr), page_size);
1495 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_PR | IOMMU_PTE_FC;
1497 __pte = __sme_set(phys_addr) | IOMMU_PTE_PR | IOMMU_PTE_FC;
1499 if (prot & IOMMU_PROT_IR)
1500 __pte |= IOMMU_PTE_IR;
1501 if (prot & IOMMU_PROT_IW)
1502 __pte |= IOMMU_PTE_IW;
1504 for (i = 0; i < count; ++i)
1512 static unsigned long iommu_unmap_page(struct protection_domain *dom,
1513 unsigned long bus_addr,
1514 unsigned long page_size)
1516 unsigned long long unmapped;
1517 unsigned long unmap_size;
1520 BUG_ON(!is_power_of_2(page_size));
1524 while (unmapped < page_size) {
1526 pte = fetch_pte(dom, bus_addr, &unmap_size);
1531 count = PAGE_SIZE_PTE_COUNT(unmap_size);
1532 for (i = 0; i < count; i++)
1536 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1537 unmapped += unmap_size;
1540 BUG_ON(unmapped && !is_power_of_2(unmapped));
1545 /****************************************************************************
1547 * The next functions belong to the address allocator for the dma_ops
1548 * interface functions.
1550 ****************************************************************************/
1553 static unsigned long dma_ops_alloc_iova(struct device *dev,
1554 struct dma_ops_domain *dma_dom,
1555 unsigned int pages, u64 dma_mask)
1557 unsigned long pfn = 0;
1559 pages = __roundup_pow_of_two(pages);
1561 if (dma_mask > DMA_BIT_MASK(32))
1562 pfn = alloc_iova_fast(&dma_dom->iovad, pages,
1563 IOVA_PFN(DMA_BIT_MASK(32)), false);
1566 pfn = alloc_iova_fast(&dma_dom->iovad, pages,
1567 IOVA_PFN(dma_mask), true);
1569 return (pfn << PAGE_SHIFT);
1572 static void dma_ops_free_iova(struct dma_ops_domain *dma_dom,
1573 unsigned long address,
1576 pages = __roundup_pow_of_two(pages);
1577 address >>= PAGE_SHIFT;
1579 free_iova_fast(&dma_dom->iovad, address, pages);
1582 /****************************************************************************
1584 * The next functions belong to the domain allocation. A domain is
1585 * allocated for every IOMMU as the default domain. If device isolation
1586 * is enabled, every device get its own domain. The most important thing
1587 * about domains is the page table mapping the DMA address space they
1590 ****************************************************************************/
1593 * This function adds a protection domain to the global protection domain list
1595 static void add_domain_to_list(struct protection_domain *domain)
1597 unsigned long flags;
1599 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1600 list_add(&domain->list, &amd_iommu_pd_list);
1601 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1605 * This function removes a protection domain to the global
1606 * protection domain list
1608 static void del_domain_from_list(struct protection_domain *domain)
1610 unsigned long flags;
1612 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1613 list_del(&domain->list);
1614 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1617 static u16 domain_id_alloc(void)
1621 spin_lock(&pd_bitmap_lock);
1622 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1624 if (id > 0 && id < MAX_DOMAIN_ID)
1625 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1628 spin_unlock(&pd_bitmap_lock);
1633 static void domain_id_free(int id)
1635 spin_lock(&pd_bitmap_lock);
1636 if (id > 0 && id < MAX_DOMAIN_ID)
1637 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1638 spin_unlock(&pd_bitmap_lock);
1641 #define DEFINE_FREE_PT_FN(LVL, FN) \
1642 static void free_pt_##LVL (unsigned long __pt) \
1650 for (i = 0; i < 512; ++i) { \
1651 /* PTE present? */ \
1652 if (!IOMMU_PTE_PRESENT(pt[i])) \
1656 if (PM_PTE_LEVEL(pt[i]) == 0 || \
1657 PM_PTE_LEVEL(pt[i]) == 7) \
1660 p = (unsigned long)IOMMU_PTE_PAGE(pt[i]); \
1663 free_page((unsigned long)pt); \
1666 DEFINE_FREE_PT_FN(l2, free_page)
1667 DEFINE_FREE_PT_FN(l3, free_pt_l2)
1668 DEFINE_FREE_PT_FN(l4, free_pt_l3)
1669 DEFINE_FREE_PT_FN(l5, free_pt_l4)
1670 DEFINE_FREE_PT_FN(l6, free_pt_l5)
1672 static void free_pagetable(struct protection_domain *domain)
1674 unsigned long root = (unsigned long)domain->pt_root;
1676 switch (domain->mode) {
1677 case PAGE_MODE_NONE:
1679 case PAGE_MODE_1_LEVEL:
1682 case PAGE_MODE_2_LEVEL:
1685 case PAGE_MODE_3_LEVEL:
1688 case PAGE_MODE_4_LEVEL:
1691 case PAGE_MODE_5_LEVEL:
1694 case PAGE_MODE_6_LEVEL:
1702 static void free_gcr3_tbl_level1(u64 *tbl)
1707 for (i = 0; i < 512; ++i) {
1708 if (!(tbl[i] & GCR3_VALID))
1711 ptr = iommu_phys_to_virt(tbl[i] & PAGE_MASK);
1713 free_page((unsigned long)ptr);
1717 static void free_gcr3_tbl_level2(u64 *tbl)
1722 for (i = 0; i < 512; ++i) {
1723 if (!(tbl[i] & GCR3_VALID))
1726 ptr = iommu_phys_to_virt(tbl[i] & PAGE_MASK);
1728 free_gcr3_tbl_level1(ptr);
1732 static void free_gcr3_table(struct protection_domain *domain)
1734 if (domain->glx == 2)
1735 free_gcr3_tbl_level2(domain->gcr3_tbl);
1736 else if (domain->glx == 1)
1737 free_gcr3_tbl_level1(domain->gcr3_tbl);
1739 BUG_ON(domain->glx != 0);
1741 free_page((unsigned long)domain->gcr3_tbl);
1744 static void dma_ops_domain_flush_tlb(struct dma_ops_domain *dom)
1746 domain_flush_tlb(&dom->domain);
1747 domain_flush_complete(&dom->domain);
1750 static void iova_domain_flush_tlb(struct iova_domain *iovad)
1752 struct dma_ops_domain *dom;
1754 dom = container_of(iovad, struct dma_ops_domain, iovad);
1756 dma_ops_domain_flush_tlb(dom);
1760 * Free a domain, only used if something went wrong in the
1761 * allocation path and we need to free an already allocated page table
1763 static void dma_ops_domain_free(struct dma_ops_domain *dom)
1768 del_domain_from_list(&dom->domain);
1770 put_iova_domain(&dom->iovad);
1772 free_pagetable(&dom->domain);
1775 domain_id_free(dom->domain.id);
1781 * Allocates a new protection domain usable for the dma_ops functions.
1782 * It also initializes the page table and the address allocator data
1783 * structures required for the dma_ops interface
1785 static struct dma_ops_domain *dma_ops_domain_alloc(void)
1787 struct dma_ops_domain *dma_dom;
1789 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
1793 if (protection_domain_init(&dma_dom->domain))
1796 dma_dom->domain.mode = PAGE_MODE_3_LEVEL;
1797 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
1798 dma_dom->domain.flags = PD_DMA_OPS_MASK;
1799 if (!dma_dom->domain.pt_root)
1802 init_iova_domain(&dma_dom->iovad, PAGE_SIZE, IOVA_START_PFN);
1804 if (init_iova_flush_queue(&dma_dom->iovad, iova_domain_flush_tlb, NULL))
1807 /* Initialize reserved ranges */
1808 copy_reserved_iova(&reserved_iova_ranges, &dma_dom->iovad);
1810 add_domain_to_list(&dma_dom->domain);
1815 dma_ops_domain_free(dma_dom);
1821 * little helper function to check whether a given protection domain is a
1824 static bool dma_ops_domain(struct protection_domain *domain)
1826 return domain->flags & PD_DMA_OPS_MASK;
1829 static void set_dte_entry(u16 devid, struct protection_domain *domain,
1835 if (domain->mode != PAGE_MODE_NONE)
1836 pte_root = iommu_virt_to_phys(domain->pt_root);
1838 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
1839 << DEV_ENTRY_MODE_SHIFT;
1840 pte_root |= DTE_FLAG_IR | DTE_FLAG_IW | DTE_FLAG_V | DTE_FLAG_TV;
1842 flags = amd_iommu_dev_table[devid].data[1];
1845 flags |= DTE_FLAG_IOTLB;
1848 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
1850 if (iommu_feature(iommu, FEATURE_EPHSUP))
1851 pte_root |= 1ULL << DEV_ENTRY_PPR;
1854 if (domain->flags & PD_IOMMUV2_MASK) {
1855 u64 gcr3 = iommu_virt_to_phys(domain->gcr3_tbl);
1856 u64 glx = domain->glx;
1859 pte_root |= DTE_FLAG_GV;
1860 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
1862 /* First mask out possible old values for GCR3 table */
1863 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
1866 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
1869 /* Encode GCR3 table into DTE */
1870 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
1873 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
1876 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
1880 flags &= ~DEV_DOMID_MASK;
1881 flags |= domain->id;
1883 amd_iommu_dev_table[devid].data[1] = flags;
1884 amd_iommu_dev_table[devid].data[0] = pte_root;
1887 static void clear_dte_entry(u16 devid)
1889 /* remove entry from the device table seen by the hardware */
1890 amd_iommu_dev_table[devid].data[0] = DTE_FLAG_V | DTE_FLAG_TV;
1891 amd_iommu_dev_table[devid].data[1] &= DTE_FLAG_MASK;
1893 amd_iommu_apply_erratum_63(devid);
1896 static void do_attach(struct iommu_dev_data *dev_data,
1897 struct protection_domain *domain)
1899 struct amd_iommu *iommu;
1903 iommu = amd_iommu_rlookup_table[dev_data->devid];
1904 alias = dev_data->alias;
1905 ats = dev_data->ats.enabled;
1907 /* Update data structures */
1908 dev_data->domain = domain;
1909 list_add(&dev_data->list, &domain->dev_list);
1911 /* Do reference counting */
1912 domain->dev_iommu[iommu->index] += 1;
1913 domain->dev_cnt += 1;
1915 /* Update device table */
1916 set_dte_entry(dev_data->devid, domain, ats, dev_data->iommu_v2);
1917 if (alias != dev_data->devid)
1918 set_dte_entry(alias, domain, ats, dev_data->iommu_v2);
1920 device_flush_dte(dev_data);
1923 static void do_detach(struct iommu_dev_data *dev_data)
1925 struct amd_iommu *iommu;
1928 iommu = amd_iommu_rlookup_table[dev_data->devid];
1929 alias = dev_data->alias;
1931 /* decrease reference counters */
1932 dev_data->domain->dev_iommu[iommu->index] -= 1;
1933 dev_data->domain->dev_cnt -= 1;
1935 /* Update data structures */
1936 dev_data->domain = NULL;
1937 list_del(&dev_data->list);
1938 clear_dte_entry(dev_data->devid);
1939 if (alias != dev_data->devid)
1940 clear_dte_entry(alias);
1942 /* Flush the DTE entry */
1943 device_flush_dte(dev_data);
1947 * If a device is not yet associated with a domain, this function makes the
1948 * device visible in the domain
1950 static int __attach_device(struct iommu_dev_data *dev_data,
1951 struct protection_domain *domain)
1956 spin_lock(&domain->lock);
1959 if (dev_data->domain != NULL)
1962 /* Attach alias group root */
1963 do_attach(dev_data, domain);
1970 spin_unlock(&domain->lock);
1976 static void pdev_iommuv2_disable(struct pci_dev *pdev)
1978 pci_disable_ats(pdev);
1979 pci_disable_pri(pdev);
1980 pci_disable_pasid(pdev);
1983 /* FIXME: Change generic reset-function to do the same */
1984 static int pri_reset_while_enabled(struct pci_dev *pdev)
1989 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
1993 pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
1994 control |= PCI_PRI_CTRL_RESET;
1995 pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
2000 static int pdev_iommuv2_enable(struct pci_dev *pdev)
2005 /* FIXME: Hardcode number of outstanding requests for now */
2007 if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
2009 reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
2011 /* Only allow access to user-accessible pages */
2012 ret = pci_enable_pasid(pdev, 0);
2016 /* First reset the PRI state of the device */
2017 ret = pci_reset_pri(pdev);
2022 ret = pci_enable_pri(pdev, reqs);
2027 ret = pri_reset_while_enabled(pdev);
2032 ret = pci_enable_ats(pdev, PAGE_SHIFT);
2039 pci_disable_pri(pdev);
2040 pci_disable_pasid(pdev);
2045 /* FIXME: Move this to PCI code */
2046 #define PCI_PRI_TLP_OFF (1 << 15)
2048 static bool pci_pri_tlp_required(struct pci_dev *pdev)
2053 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2057 pci_read_config_word(pdev, pos + PCI_PRI_STATUS, &status);
2059 return (status & PCI_PRI_TLP_OFF) ? true : false;
2063 * If a device is not yet associated with a domain, this function makes the
2064 * device visible in the domain
2066 static int attach_device(struct device *dev,
2067 struct protection_domain *domain)
2069 struct pci_dev *pdev;
2070 struct iommu_dev_data *dev_data;
2071 unsigned long flags;
2074 dev_data = get_dev_data(dev);
2076 if (!dev_is_pci(dev))
2077 goto skip_ats_check;
2079 pdev = to_pci_dev(dev);
2080 if (domain->flags & PD_IOMMUV2_MASK) {
2081 if (!dev_data->passthrough)
2084 if (dev_data->iommu_v2) {
2085 if (pdev_iommuv2_enable(pdev) != 0)
2088 dev_data->ats.enabled = true;
2089 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2090 dev_data->pri_tlp = pci_pri_tlp_required(pdev);
2092 } else if (amd_iommu_iotlb_sup &&
2093 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
2094 dev_data->ats.enabled = true;
2095 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2099 spin_lock_irqsave(&amd_iommu_devtable_lock, flags);
2100 ret = __attach_device(dev_data, domain);
2101 spin_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2104 * We might boot into a crash-kernel here. The crashed kernel
2105 * left the caches in the IOMMU dirty. So we have to flush
2106 * here to evict all dirty stuff.
2108 domain_flush_tlb_pde(domain);
2114 * Removes a device from a protection domain (unlocked)
2116 static void __detach_device(struct iommu_dev_data *dev_data)
2118 struct protection_domain *domain;
2120 domain = dev_data->domain;
2122 spin_lock(&domain->lock);
2124 do_detach(dev_data);
2126 spin_unlock(&domain->lock);
2130 * Removes a device from a protection domain (with devtable_lock held)
2132 static void detach_device(struct device *dev)
2134 struct protection_domain *domain;
2135 struct iommu_dev_data *dev_data;
2136 unsigned long flags;
2138 dev_data = get_dev_data(dev);
2139 domain = dev_data->domain;
2142 * First check if the device is still attached. It might already
2143 * be detached from its domain because the generic
2144 * iommu_detach_group code detached it and we try again here in
2145 * our alias handling.
2147 if (WARN_ON(!dev_data->domain))
2150 /* lock device table */
2151 spin_lock_irqsave(&amd_iommu_devtable_lock, flags);
2152 __detach_device(dev_data);
2153 spin_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2155 if (!dev_is_pci(dev))
2158 if (domain->flags & PD_IOMMUV2_MASK && dev_data->iommu_v2)
2159 pdev_iommuv2_disable(to_pci_dev(dev));
2160 else if (dev_data->ats.enabled)
2161 pci_disable_ats(to_pci_dev(dev));
2163 dev_data->ats.enabled = false;
2166 static int amd_iommu_add_device(struct device *dev)
2168 struct iommu_dev_data *dev_data;
2169 struct iommu_domain *domain;
2170 struct amd_iommu *iommu;
2173 if (!check_device(dev) || get_dev_data(dev))
2176 devid = get_device_id(dev);
2180 iommu = amd_iommu_rlookup_table[devid];
2182 ret = iommu_init_device(dev);
2184 if (ret != -ENOTSUPP)
2185 pr_err("Failed to initialize device %s - trying to proceed anyway\n",
2188 iommu_ignore_device(dev);
2189 dev->dma_ops = &dma_direct_ops;
2192 init_iommu_group(dev);
2194 dev_data = get_dev_data(dev);
2198 if (iommu_pass_through || dev_data->iommu_v2)
2199 iommu_request_dm_for_dev(dev);
2201 /* Domains are initialized for this device - have a look what we ended up with */
2202 domain = iommu_get_domain_for_dev(dev);
2203 if (domain->type == IOMMU_DOMAIN_IDENTITY)
2204 dev_data->passthrough = true;
2206 dev->dma_ops = &amd_iommu_dma_ops;
2209 iommu_completion_wait(iommu);
2214 static void amd_iommu_remove_device(struct device *dev)
2216 struct amd_iommu *iommu;
2219 if (!check_device(dev))
2222 devid = get_device_id(dev);
2226 iommu = amd_iommu_rlookup_table[devid];
2228 iommu_uninit_device(dev);
2229 iommu_completion_wait(iommu);
2232 static struct iommu_group *amd_iommu_device_group(struct device *dev)
2234 if (dev_is_pci(dev))
2235 return pci_device_group(dev);
2237 return acpihid_device_group(dev);
2240 /*****************************************************************************
2242 * The next functions belong to the dma_ops mapping/unmapping code.
2244 *****************************************************************************/
2247 * In the dma_ops path we only have the struct device. This function
2248 * finds the corresponding IOMMU, the protection domain and the
2249 * requestor id for a given device.
2250 * If the device is not yet associated with a domain this is also done
2253 static struct protection_domain *get_domain(struct device *dev)
2255 struct protection_domain *domain;
2256 struct iommu_domain *io_domain;
2258 if (!check_device(dev))
2259 return ERR_PTR(-EINVAL);
2261 domain = get_dev_data(dev)->domain;
2262 if (domain == NULL && get_dev_data(dev)->defer_attach) {
2263 get_dev_data(dev)->defer_attach = false;
2264 io_domain = iommu_get_domain_for_dev(dev);
2265 domain = to_pdomain(io_domain);
2266 attach_device(dev, domain);
2269 return ERR_PTR(-EBUSY);
2271 if (!dma_ops_domain(domain))
2272 return ERR_PTR(-EBUSY);
2277 static void update_device_table(struct protection_domain *domain)
2279 struct iommu_dev_data *dev_data;
2281 list_for_each_entry(dev_data, &domain->dev_list, list) {
2282 set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled,
2283 dev_data->iommu_v2);
2285 if (dev_data->devid == dev_data->alias)
2288 /* There is an alias, update device table entry for it */
2289 set_dte_entry(dev_data->alias, domain, dev_data->ats.enabled,
2290 dev_data->iommu_v2);
2294 static void update_domain(struct protection_domain *domain)
2296 if (!domain->updated)
2299 update_device_table(domain);
2301 domain_flush_devices(domain);
2302 domain_flush_tlb_pde(domain);
2304 domain->updated = false;
2307 static int dir2prot(enum dma_data_direction direction)
2309 if (direction == DMA_TO_DEVICE)
2310 return IOMMU_PROT_IR;
2311 else if (direction == DMA_FROM_DEVICE)
2312 return IOMMU_PROT_IW;
2313 else if (direction == DMA_BIDIRECTIONAL)
2314 return IOMMU_PROT_IW | IOMMU_PROT_IR;
2320 * This function contains common code for mapping of a physically
2321 * contiguous memory region into DMA address space. It is used by all
2322 * mapping functions provided with this IOMMU driver.
2323 * Must be called with the domain lock held.
2325 static dma_addr_t __map_single(struct device *dev,
2326 struct dma_ops_domain *dma_dom,
2329 enum dma_data_direction direction,
2332 dma_addr_t offset = paddr & ~PAGE_MASK;
2333 dma_addr_t address, start, ret;
2338 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
2341 address = dma_ops_alloc_iova(dev, dma_dom, pages, dma_mask);
2342 if (address == AMD_IOMMU_MAPPING_ERROR)
2345 prot = dir2prot(direction);
2348 for (i = 0; i < pages; ++i) {
2349 ret = iommu_map_page(&dma_dom->domain, start, paddr,
2350 PAGE_SIZE, prot, GFP_ATOMIC);
2359 if (unlikely(amd_iommu_np_cache)) {
2360 domain_flush_pages(&dma_dom->domain, address, size);
2361 domain_flush_complete(&dma_dom->domain);
2369 for (--i; i >= 0; --i) {
2371 iommu_unmap_page(&dma_dom->domain, start, PAGE_SIZE);
2374 domain_flush_tlb(&dma_dom->domain);
2375 domain_flush_complete(&dma_dom->domain);
2377 dma_ops_free_iova(dma_dom, address, pages);
2379 return AMD_IOMMU_MAPPING_ERROR;
2383 * Does the reverse of the __map_single function. Must be called with
2384 * the domain lock held too
2386 static void __unmap_single(struct dma_ops_domain *dma_dom,
2387 dma_addr_t dma_addr,
2391 dma_addr_t i, start;
2394 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
2395 dma_addr &= PAGE_MASK;
2398 for (i = 0; i < pages; ++i) {
2399 iommu_unmap_page(&dma_dom->domain, start, PAGE_SIZE);
2403 if (amd_iommu_unmap_flush) {
2404 domain_flush_tlb(&dma_dom->domain);
2405 domain_flush_complete(&dma_dom->domain);
2406 dma_ops_free_iova(dma_dom, dma_addr, pages);
2408 pages = __roundup_pow_of_two(pages);
2409 queue_iova(&dma_dom->iovad, dma_addr >> PAGE_SHIFT, pages, 0);
2414 * The exported map_single function for dma_ops.
2416 static dma_addr_t map_page(struct device *dev, struct page *page,
2417 unsigned long offset, size_t size,
2418 enum dma_data_direction dir,
2419 unsigned long attrs)
2421 phys_addr_t paddr = page_to_phys(page) + offset;
2422 struct protection_domain *domain;
2423 struct dma_ops_domain *dma_dom;
2426 domain = get_domain(dev);
2427 if (PTR_ERR(domain) == -EINVAL)
2428 return (dma_addr_t)paddr;
2429 else if (IS_ERR(domain))
2430 return AMD_IOMMU_MAPPING_ERROR;
2432 dma_mask = *dev->dma_mask;
2433 dma_dom = to_dma_ops_domain(domain);
2435 return __map_single(dev, dma_dom, paddr, size, dir, dma_mask);
2439 * The exported unmap_single function for dma_ops.
2441 static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2442 enum dma_data_direction dir, unsigned long attrs)
2444 struct protection_domain *domain;
2445 struct dma_ops_domain *dma_dom;
2447 domain = get_domain(dev);
2451 dma_dom = to_dma_ops_domain(domain);
2453 __unmap_single(dma_dom, dma_addr, size, dir);
2456 static int sg_num_pages(struct device *dev,
2457 struct scatterlist *sglist,
2460 unsigned long mask, boundary_size;
2461 struct scatterlist *s;
2464 mask = dma_get_seg_boundary(dev);
2465 boundary_size = mask + 1 ? ALIGN(mask + 1, PAGE_SIZE) >> PAGE_SHIFT :
2466 1UL << (BITS_PER_LONG - PAGE_SHIFT);
2468 for_each_sg(sglist, s, nelems, i) {
2471 s->dma_address = npages << PAGE_SHIFT;
2472 p = npages % boundary_size;
2473 n = iommu_num_pages(sg_phys(s), s->length, PAGE_SIZE);
2474 if (p + n > boundary_size)
2475 npages += boundary_size - p;
2483 * The exported map_sg function for dma_ops (handles scatter-gather
2486 static int map_sg(struct device *dev, struct scatterlist *sglist,
2487 int nelems, enum dma_data_direction direction,
2488 unsigned long attrs)
2490 int mapped_pages = 0, npages = 0, prot = 0, i;
2491 struct protection_domain *domain;
2492 struct dma_ops_domain *dma_dom;
2493 struct scatterlist *s;
2494 unsigned long address;
2497 domain = get_domain(dev);
2501 dma_dom = to_dma_ops_domain(domain);
2502 dma_mask = *dev->dma_mask;
2504 npages = sg_num_pages(dev, sglist, nelems);
2506 address = dma_ops_alloc_iova(dev, dma_dom, npages, dma_mask);
2507 if (address == AMD_IOMMU_MAPPING_ERROR)
2510 prot = dir2prot(direction);
2512 /* Map all sg entries */
2513 for_each_sg(sglist, s, nelems, i) {
2514 int j, pages = iommu_num_pages(sg_phys(s), s->length, PAGE_SIZE);
2516 for (j = 0; j < pages; ++j) {
2517 unsigned long bus_addr, phys_addr;
2520 bus_addr = address + s->dma_address + (j << PAGE_SHIFT);
2521 phys_addr = (sg_phys(s) & PAGE_MASK) + (j << PAGE_SHIFT);
2522 ret = iommu_map_page(domain, bus_addr, phys_addr, PAGE_SIZE, prot, GFP_ATOMIC);
2530 /* Everything is mapped - write the right values into s->dma_address */
2531 for_each_sg(sglist, s, nelems, i) {
2532 s->dma_address += address + s->offset;
2533 s->dma_length = s->length;
2539 pr_err("%s: IOMMU mapping error in map_sg (io-pages: %d)\n",
2540 dev_name(dev), npages);
2542 for_each_sg(sglist, s, nelems, i) {
2543 int j, pages = iommu_num_pages(sg_phys(s), s->length, PAGE_SIZE);
2545 for (j = 0; j < pages; ++j) {
2546 unsigned long bus_addr;
2548 bus_addr = address + s->dma_address + (j << PAGE_SHIFT);
2549 iommu_unmap_page(domain, bus_addr, PAGE_SIZE);
2557 free_iova_fast(&dma_dom->iovad, address, npages);
2564 * The exported map_sg function for dma_ops (handles scatter-gather
2567 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2568 int nelems, enum dma_data_direction dir,
2569 unsigned long attrs)
2571 struct protection_domain *domain;
2572 struct dma_ops_domain *dma_dom;
2573 unsigned long startaddr;
2576 domain = get_domain(dev);
2580 startaddr = sg_dma_address(sglist) & PAGE_MASK;
2581 dma_dom = to_dma_ops_domain(domain);
2582 npages = sg_num_pages(dev, sglist, nelems);
2584 __unmap_single(dma_dom, startaddr, npages << PAGE_SHIFT, dir);
2588 * The exported alloc_coherent function for dma_ops.
2590 static void *alloc_coherent(struct device *dev, size_t size,
2591 dma_addr_t *dma_addr, gfp_t flag,
2592 unsigned long attrs)
2594 u64 dma_mask = dev->coherent_dma_mask;
2595 struct protection_domain *domain;
2596 struct dma_ops_domain *dma_dom;
2599 domain = get_domain(dev);
2600 if (PTR_ERR(domain) == -EINVAL) {
2601 page = alloc_pages(flag, get_order(size));
2602 *dma_addr = page_to_phys(page);
2603 return page_address(page);
2604 } else if (IS_ERR(domain))
2607 dma_dom = to_dma_ops_domain(domain);
2608 size = PAGE_ALIGN(size);
2609 dma_mask = dev->coherent_dma_mask;
2610 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
2613 page = alloc_pages(flag | __GFP_NOWARN, get_order(size));
2615 if (!gfpflags_allow_blocking(flag))
2618 page = dma_alloc_from_contiguous(dev, size >> PAGE_SHIFT,
2619 get_order(size), flag & __GFP_NOWARN);
2625 dma_mask = *dev->dma_mask;
2627 *dma_addr = __map_single(dev, dma_dom, page_to_phys(page),
2628 size, DMA_BIDIRECTIONAL, dma_mask);
2630 if (*dma_addr == AMD_IOMMU_MAPPING_ERROR)
2633 return page_address(page);
2637 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2638 __free_pages(page, get_order(size));
2644 * The exported free_coherent function for dma_ops.
2646 static void free_coherent(struct device *dev, size_t size,
2647 void *virt_addr, dma_addr_t dma_addr,
2648 unsigned long attrs)
2650 struct protection_domain *domain;
2651 struct dma_ops_domain *dma_dom;
2654 page = virt_to_page(virt_addr);
2655 size = PAGE_ALIGN(size);
2657 domain = get_domain(dev);
2661 dma_dom = to_dma_ops_domain(domain);
2663 __unmap_single(dma_dom, dma_addr, size, DMA_BIDIRECTIONAL);
2666 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2667 __free_pages(page, get_order(size));
2671 * This function is called by the DMA layer to find out if we can handle a
2672 * particular device. It is part of the dma_ops.
2674 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
2676 if (!dma_direct_supported(dev, mask))
2678 return check_device(dev);
2681 static int amd_iommu_mapping_error(struct device *dev, dma_addr_t dma_addr)
2683 return dma_addr == AMD_IOMMU_MAPPING_ERROR;
2686 static const struct dma_map_ops amd_iommu_dma_ops = {
2687 .alloc = alloc_coherent,
2688 .free = free_coherent,
2689 .map_page = map_page,
2690 .unmap_page = unmap_page,
2692 .unmap_sg = unmap_sg,
2693 .dma_supported = amd_iommu_dma_supported,
2694 .mapping_error = amd_iommu_mapping_error,
2697 static int init_reserved_iova_ranges(void)
2699 struct pci_dev *pdev = NULL;
2702 init_iova_domain(&reserved_iova_ranges, PAGE_SIZE, IOVA_START_PFN);
2704 lockdep_set_class(&reserved_iova_ranges.iova_rbtree_lock,
2705 &reserved_rbtree_key);
2707 /* MSI memory range */
2708 val = reserve_iova(&reserved_iova_ranges,
2709 IOVA_PFN(MSI_RANGE_START), IOVA_PFN(MSI_RANGE_END));
2711 pr_err("Reserving MSI range failed\n");
2715 /* HT memory range */
2716 val = reserve_iova(&reserved_iova_ranges,
2717 IOVA_PFN(HT_RANGE_START), IOVA_PFN(HT_RANGE_END));
2719 pr_err("Reserving HT range failed\n");
2724 * Memory used for PCI resources
2725 * FIXME: Check whether we can reserve the PCI-hole completly
2727 for_each_pci_dev(pdev) {
2730 for (i = 0; i < PCI_NUM_RESOURCES; ++i) {
2731 struct resource *r = &pdev->resource[i];
2733 if (!(r->flags & IORESOURCE_MEM))
2736 val = reserve_iova(&reserved_iova_ranges,
2740 pr_err("Reserve pci-resource range failed\n");
2749 int __init amd_iommu_init_api(void)
2753 ret = iova_cache_get();
2757 ret = init_reserved_iova_ranges();
2761 err = bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
2764 #ifdef CONFIG_ARM_AMBA
2765 err = bus_set_iommu(&amba_bustype, &amd_iommu_ops);
2769 err = bus_set_iommu(&platform_bus_type, &amd_iommu_ops);
2776 int __init amd_iommu_init_dma_ops(void)
2778 swiotlb = (iommu_pass_through || sme_me_mask) ? 1 : 0;
2782 * In case we don't initialize SWIOTLB (actually the common case
2783 * when AMD IOMMU is enabled and SME is not active), make sure there
2784 * are global dma_ops set as a fall-back for devices not handled by
2785 * this driver (for example non-PCI devices). When SME is active,
2786 * make sure that swiotlb variable remains set so the global dma_ops
2787 * continue to be SWIOTLB.
2790 dma_ops = &dma_direct_ops;
2792 if (amd_iommu_unmap_flush)
2793 pr_info("AMD-Vi: IO/TLB flush on unmap enabled\n");
2795 pr_info("AMD-Vi: Lazy IO/TLB flushing enabled\n");
2801 /*****************************************************************************
2803 * The following functions belong to the exported interface of AMD IOMMU
2805 * This interface allows access to lower level functions of the IOMMU
2806 * like protection domain handling and assignement of devices to domains
2807 * which is not possible with the dma_ops interface.
2809 *****************************************************************************/
2811 static void cleanup_domain(struct protection_domain *domain)
2813 struct iommu_dev_data *entry;
2814 unsigned long flags;
2816 spin_lock_irqsave(&amd_iommu_devtable_lock, flags);
2818 while (!list_empty(&domain->dev_list)) {
2819 entry = list_first_entry(&domain->dev_list,
2820 struct iommu_dev_data, list);
2821 BUG_ON(!entry->domain);
2822 __detach_device(entry);
2825 spin_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2828 static void protection_domain_free(struct protection_domain *domain)
2833 del_domain_from_list(domain);
2836 domain_id_free(domain->id);
2841 static int protection_domain_init(struct protection_domain *domain)
2843 spin_lock_init(&domain->lock);
2844 mutex_init(&domain->api_lock);
2845 domain->id = domain_id_alloc();
2848 INIT_LIST_HEAD(&domain->dev_list);
2853 static struct protection_domain *protection_domain_alloc(void)
2855 struct protection_domain *domain;
2857 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2861 if (protection_domain_init(domain))
2864 add_domain_to_list(domain);
2874 static struct iommu_domain *amd_iommu_domain_alloc(unsigned type)
2876 struct protection_domain *pdomain;
2877 struct dma_ops_domain *dma_domain;
2880 case IOMMU_DOMAIN_UNMANAGED:
2881 pdomain = protection_domain_alloc();
2885 pdomain->mode = PAGE_MODE_3_LEVEL;
2886 pdomain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2887 if (!pdomain->pt_root) {
2888 protection_domain_free(pdomain);
2892 pdomain->domain.geometry.aperture_start = 0;
2893 pdomain->domain.geometry.aperture_end = ~0ULL;
2894 pdomain->domain.geometry.force_aperture = true;
2897 case IOMMU_DOMAIN_DMA:
2898 dma_domain = dma_ops_domain_alloc();
2900 pr_err("AMD-Vi: Failed to allocate\n");
2903 pdomain = &dma_domain->domain;
2905 case IOMMU_DOMAIN_IDENTITY:
2906 pdomain = protection_domain_alloc();
2910 pdomain->mode = PAGE_MODE_NONE;
2916 return &pdomain->domain;
2919 static void amd_iommu_domain_free(struct iommu_domain *dom)
2921 struct protection_domain *domain;
2922 struct dma_ops_domain *dma_dom;
2924 domain = to_pdomain(dom);
2926 if (domain->dev_cnt > 0)
2927 cleanup_domain(domain);
2929 BUG_ON(domain->dev_cnt != 0);
2934 switch (dom->type) {
2935 case IOMMU_DOMAIN_DMA:
2936 /* Now release the domain */
2937 dma_dom = to_dma_ops_domain(domain);
2938 dma_ops_domain_free(dma_dom);
2941 if (domain->mode != PAGE_MODE_NONE)
2942 free_pagetable(domain);
2944 if (domain->flags & PD_IOMMUV2_MASK)
2945 free_gcr3_table(domain);
2947 protection_domain_free(domain);
2952 static void amd_iommu_detach_device(struct iommu_domain *dom,
2955 struct iommu_dev_data *dev_data = dev->archdata.iommu;
2956 struct amd_iommu *iommu;
2959 if (!check_device(dev))
2962 devid = get_device_id(dev);
2966 if (dev_data->domain != NULL)
2969 iommu = amd_iommu_rlookup_table[devid];
2973 #ifdef CONFIG_IRQ_REMAP
2974 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) &&
2975 (dom->type == IOMMU_DOMAIN_UNMANAGED))
2976 dev_data->use_vapic = 0;
2979 iommu_completion_wait(iommu);
2982 static int amd_iommu_attach_device(struct iommu_domain *dom,
2985 struct protection_domain *domain = to_pdomain(dom);
2986 struct iommu_dev_data *dev_data;
2987 struct amd_iommu *iommu;
2990 if (!check_device(dev))
2993 dev_data = dev->archdata.iommu;
2995 iommu = amd_iommu_rlookup_table[dev_data->devid];
2999 if (dev_data->domain)
3002 ret = attach_device(dev, domain);
3004 #ifdef CONFIG_IRQ_REMAP
3005 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
3006 if (dom->type == IOMMU_DOMAIN_UNMANAGED)
3007 dev_data->use_vapic = 1;
3009 dev_data->use_vapic = 0;
3013 iommu_completion_wait(iommu);
3018 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
3019 phys_addr_t paddr, size_t page_size, int iommu_prot)
3021 struct protection_domain *domain = to_pdomain(dom);
3025 if (domain->mode == PAGE_MODE_NONE)
3028 if (iommu_prot & IOMMU_READ)
3029 prot |= IOMMU_PROT_IR;
3030 if (iommu_prot & IOMMU_WRITE)
3031 prot |= IOMMU_PROT_IW;
3033 mutex_lock(&domain->api_lock);
3034 ret = iommu_map_page(domain, iova, paddr, page_size, prot, GFP_KERNEL);
3035 mutex_unlock(&domain->api_lock);
3040 static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
3043 struct protection_domain *domain = to_pdomain(dom);
3046 if (domain->mode == PAGE_MODE_NONE)
3049 mutex_lock(&domain->api_lock);
3050 unmap_size = iommu_unmap_page(domain, iova, page_size);
3051 mutex_unlock(&domain->api_lock);
3056 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
3059 struct protection_domain *domain = to_pdomain(dom);
3060 unsigned long offset_mask, pte_pgsize;
3063 if (domain->mode == PAGE_MODE_NONE)
3066 pte = fetch_pte(domain, iova, &pte_pgsize);
3068 if (!pte || !IOMMU_PTE_PRESENT(*pte))
3071 offset_mask = pte_pgsize - 1;
3072 __pte = __sme_clr(*pte & PM_ADDR_MASK);
3074 return (__pte & ~offset_mask) | (iova & offset_mask);
3077 static bool amd_iommu_capable(enum iommu_cap cap)
3080 case IOMMU_CAP_CACHE_COHERENCY:
3082 case IOMMU_CAP_INTR_REMAP:
3083 return (irq_remapping_enabled == 1);
3084 case IOMMU_CAP_NOEXEC:
3093 static void amd_iommu_get_resv_regions(struct device *dev,
3094 struct list_head *head)
3096 struct iommu_resv_region *region;
3097 struct unity_map_entry *entry;
3100 devid = get_device_id(dev);
3104 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
3108 if (devid < entry->devid_start || devid > entry->devid_end)
3111 length = entry->address_end - entry->address_start;
3112 if (entry->prot & IOMMU_PROT_IR)
3114 if (entry->prot & IOMMU_PROT_IW)
3115 prot |= IOMMU_WRITE;
3117 region = iommu_alloc_resv_region(entry->address_start,
3121 pr_err("Out of memory allocating dm-regions for %s\n",
3125 list_add_tail(®ion->list, head);
3128 region = iommu_alloc_resv_region(MSI_RANGE_START,
3129 MSI_RANGE_END - MSI_RANGE_START + 1,
3133 list_add_tail(®ion->list, head);
3135 region = iommu_alloc_resv_region(HT_RANGE_START,
3136 HT_RANGE_END - HT_RANGE_START + 1,
3137 0, IOMMU_RESV_RESERVED);
3140 list_add_tail(®ion->list, head);
3143 static void amd_iommu_put_resv_regions(struct device *dev,
3144 struct list_head *head)
3146 struct iommu_resv_region *entry, *next;
3148 list_for_each_entry_safe(entry, next, head, list)
3152 static void amd_iommu_apply_resv_region(struct device *dev,
3153 struct iommu_domain *domain,
3154 struct iommu_resv_region *region)
3156 struct dma_ops_domain *dma_dom = to_dma_ops_domain(to_pdomain(domain));
3157 unsigned long start, end;
3159 start = IOVA_PFN(region->start);
3160 end = IOVA_PFN(region->start + region->length - 1);
3162 WARN_ON_ONCE(reserve_iova(&dma_dom->iovad, start, end) == NULL);
3165 static bool amd_iommu_is_attach_deferred(struct iommu_domain *domain,
3168 struct iommu_dev_data *dev_data = dev->archdata.iommu;
3169 return dev_data->defer_attach;
3172 static void amd_iommu_flush_iotlb_all(struct iommu_domain *domain)
3174 struct protection_domain *dom = to_pdomain(domain);
3176 domain_flush_tlb_pde(dom);
3177 domain_flush_complete(dom);
3180 static void amd_iommu_iotlb_range_add(struct iommu_domain *domain,
3181 unsigned long iova, size_t size)
3185 const struct iommu_ops amd_iommu_ops = {
3186 .capable = amd_iommu_capable,
3187 .domain_alloc = amd_iommu_domain_alloc,
3188 .domain_free = amd_iommu_domain_free,
3189 .attach_dev = amd_iommu_attach_device,
3190 .detach_dev = amd_iommu_detach_device,
3191 .map = amd_iommu_map,
3192 .unmap = amd_iommu_unmap,
3193 .iova_to_phys = amd_iommu_iova_to_phys,
3194 .add_device = amd_iommu_add_device,
3195 .remove_device = amd_iommu_remove_device,
3196 .device_group = amd_iommu_device_group,
3197 .get_resv_regions = amd_iommu_get_resv_regions,
3198 .put_resv_regions = amd_iommu_put_resv_regions,
3199 .apply_resv_region = amd_iommu_apply_resv_region,
3200 .is_attach_deferred = amd_iommu_is_attach_deferred,
3201 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
3202 .flush_iotlb_all = amd_iommu_flush_iotlb_all,
3203 .iotlb_range_add = amd_iommu_iotlb_range_add,
3204 .iotlb_sync = amd_iommu_flush_iotlb_all,
3207 /*****************************************************************************
3209 * The next functions do a basic initialization of IOMMU for pass through
3212 * In passthrough mode the IOMMU is initialized and enabled but not used for
3213 * DMA-API translation.
3215 *****************************************************************************/
3217 /* IOMMUv2 specific functions */
3218 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
3220 return atomic_notifier_chain_register(&ppr_notifier, nb);
3222 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
3224 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
3226 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
3228 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
3230 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
3232 struct protection_domain *domain = to_pdomain(dom);
3233 unsigned long flags;
3235 spin_lock_irqsave(&domain->lock, flags);
3237 /* Update data structure */
3238 domain->mode = PAGE_MODE_NONE;
3239 domain->updated = true;
3241 /* Make changes visible to IOMMUs */
3242 update_domain(domain);
3244 /* Page-table is not visible to IOMMU anymore, so free it */
3245 free_pagetable(domain);
3247 spin_unlock_irqrestore(&domain->lock, flags);
3249 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
3251 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
3253 struct protection_domain *domain = to_pdomain(dom);
3254 unsigned long flags;
3257 if (pasids <= 0 || pasids > (PASID_MASK + 1))
3260 /* Number of GCR3 table levels required */
3261 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
3264 if (levels > amd_iommu_max_glx_val)
3267 spin_lock_irqsave(&domain->lock, flags);
3270 * Save us all sanity checks whether devices already in the
3271 * domain support IOMMUv2. Just force that the domain has no
3272 * devices attached when it is switched into IOMMUv2 mode.
3275 if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
3279 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
3280 if (domain->gcr3_tbl == NULL)
3283 domain->glx = levels;
3284 domain->flags |= PD_IOMMUV2_MASK;
3285 domain->updated = true;
3287 update_domain(domain);
3292 spin_unlock_irqrestore(&domain->lock, flags);
3296 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
3298 static int __flush_pasid(struct protection_domain *domain, int pasid,
3299 u64 address, bool size)
3301 struct iommu_dev_data *dev_data;
3302 struct iommu_cmd cmd;
3305 if (!(domain->flags & PD_IOMMUV2_MASK))
3308 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
3311 * IOMMU TLB needs to be flushed before Device TLB to
3312 * prevent device TLB refill from IOMMU TLB
3314 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
3315 if (domain->dev_iommu[i] == 0)
3318 ret = iommu_queue_command(amd_iommus[i], &cmd);
3323 /* Wait until IOMMU TLB flushes are complete */
3324 domain_flush_complete(domain);
3326 /* Now flush device TLBs */
3327 list_for_each_entry(dev_data, &domain->dev_list, list) {
3328 struct amd_iommu *iommu;
3332 There might be non-IOMMUv2 capable devices in an IOMMUv2
3335 if (!dev_data->ats.enabled)
3338 qdep = dev_data->ats.qdep;
3339 iommu = amd_iommu_rlookup_table[dev_data->devid];
3341 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
3342 qdep, address, size);
3344 ret = iommu_queue_command(iommu, &cmd);
3349 /* Wait until all device TLBs are flushed */
3350 domain_flush_complete(domain);
3359 static int __amd_iommu_flush_page(struct protection_domain *domain, int pasid,
3362 return __flush_pasid(domain, pasid, address, false);
3365 int amd_iommu_flush_page(struct iommu_domain *dom, int pasid,
3368 struct protection_domain *domain = to_pdomain(dom);
3369 unsigned long flags;
3372 spin_lock_irqsave(&domain->lock, flags);
3373 ret = __amd_iommu_flush_page(domain, pasid, address);
3374 spin_unlock_irqrestore(&domain->lock, flags);
3378 EXPORT_SYMBOL(amd_iommu_flush_page);
3380 static int __amd_iommu_flush_tlb(struct protection_domain *domain, int pasid)
3382 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
3386 int amd_iommu_flush_tlb(struct iommu_domain *dom, int pasid)
3388 struct protection_domain *domain = to_pdomain(dom);
3389 unsigned long flags;
3392 spin_lock_irqsave(&domain->lock, flags);
3393 ret = __amd_iommu_flush_tlb(domain, pasid);
3394 spin_unlock_irqrestore(&domain->lock, flags);
3398 EXPORT_SYMBOL(amd_iommu_flush_tlb);
3400 static u64 *__get_gcr3_pte(u64 *root, int level, int pasid, bool alloc)
3407 index = (pasid >> (9 * level)) & 0x1ff;
3413 if (!(*pte & GCR3_VALID)) {
3417 root = (void *)get_zeroed_page(GFP_ATOMIC);
3421 *pte = iommu_virt_to_phys(root) | GCR3_VALID;
3424 root = iommu_phys_to_virt(*pte & PAGE_MASK);
3432 static int __set_gcr3(struct protection_domain *domain, int pasid,
3437 if (domain->mode != PAGE_MODE_NONE)
3440 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3444 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3446 return __amd_iommu_flush_tlb(domain, pasid);
3449 static int __clear_gcr3(struct protection_domain *domain, int pasid)
3453 if (domain->mode != PAGE_MODE_NONE)
3456 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3462 return __amd_iommu_flush_tlb(domain, pasid);
3465 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, int pasid,
3468 struct protection_domain *domain = to_pdomain(dom);
3469 unsigned long flags;
3472 spin_lock_irqsave(&domain->lock, flags);
3473 ret = __set_gcr3(domain, pasid, cr3);
3474 spin_unlock_irqrestore(&domain->lock, flags);
3478 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3480 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, int pasid)
3482 struct protection_domain *domain = to_pdomain(dom);
3483 unsigned long flags;
3486 spin_lock_irqsave(&domain->lock, flags);
3487 ret = __clear_gcr3(domain, pasid);
3488 spin_unlock_irqrestore(&domain->lock, flags);
3492 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3494 int amd_iommu_complete_ppr(struct pci_dev *pdev, int pasid,
3495 int status, int tag)
3497 struct iommu_dev_data *dev_data;
3498 struct amd_iommu *iommu;
3499 struct iommu_cmd cmd;
3501 dev_data = get_dev_data(&pdev->dev);
3502 iommu = amd_iommu_rlookup_table[dev_data->devid];
3504 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3505 tag, dev_data->pri_tlp);
3507 return iommu_queue_command(iommu, &cmd);
3509 EXPORT_SYMBOL(amd_iommu_complete_ppr);
3511 struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3513 struct protection_domain *pdomain;
3515 pdomain = get_domain(&pdev->dev);
3516 if (IS_ERR(pdomain))
3519 /* Only return IOMMUv2 domains */
3520 if (!(pdomain->flags & PD_IOMMUV2_MASK))
3523 return &pdomain->domain;
3525 EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3527 void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3529 struct iommu_dev_data *dev_data;
3531 if (!amd_iommu_v2_supported())
3534 dev_data = get_dev_data(&pdev->dev);
3535 dev_data->errata |= (1 << erratum);
3537 EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3539 int amd_iommu_device_info(struct pci_dev *pdev,
3540 struct amd_iommu_device_info *info)
3545 if (pdev == NULL || info == NULL)
3548 if (!amd_iommu_v2_supported())
3551 memset(info, 0, sizeof(*info));
3553 if (!pci_ats_disabled()) {
3554 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS);
3556 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3559 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3561 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3563 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3567 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3568 max_pasids = min(max_pasids, (1 << 20));
3570 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3571 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3573 features = pci_pasid_features(pdev);
3574 if (features & PCI_PASID_CAP_EXEC)
3575 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3576 if (features & PCI_PASID_CAP_PRIV)
3577 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3582 EXPORT_SYMBOL(amd_iommu_device_info);
3584 #ifdef CONFIG_IRQ_REMAP
3586 /*****************************************************************************
3588 * Interrupt Remapping Implementation
3590 *****************************************************************************/
3592 static struct irq_chip amd_ir_chip;
3593 static DEFINE_SPINLOCK(iommu_table_lock);
3595 static void set_dte_irq_entry(u16 devid, struct irq_remap_table *table)
3599 dte = amd_iommu_dev_table[devid].data[2];
3600 dte &= ~DTE_IRQ_PHYS_ADDR_MASK;
3601 dte |= iommu_virt_to_phys(table->table);
3602 dte |= DTE_IRQ_REMAP_INTCTL;
3603 dte |= DTE_IRQ_TABLE_LEN;
3604 dte |= DTE_IRQ_REMAP_ENABLE;
3606 amd_iommu_dev_table[devid].data[2] = dte;
3609 static struct irq_remap_table *get_irq_table(u16 devid)
3611 struct irq_remap_table *table;
3613 if (WARN_ONCE(!amd_iommu_rlookup_table[devid],
3614 "%s: no iommu for devid %x\n", __func__, devid))
3617 table = irq_lookup_table[devid];
3618 if (WARN_ONCE(!table, "%s: no table for devid %x\n", __func__, devid))
3624 static struct irq_remap_table *__alloc_irq_table(void)
3626 struct irq_remap_table *table;
3628 table = kzalloc(sizeof(*table), GFP_KERNEL);
3632 table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_KERNEL);
3633 if (!table->table) {
3637 raw_spin_lock_init(&table->lock);
3639 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
3640 memset(table->table, 0,
3641 MAX_IRQS_PER_TABLE * sizeof(u32));
3643 memset(table->table, 0,
3644 (MAX_IRQS_PER_TABLE * (sizeof(u64) * 2)));
3648 static void set_remap_table_entry(struct amd_iommu *iommu, u16 devid,
3649 struct irq_remap_table *table)
3651 irq_lookup_table[devid] = table;
3652 set_dte_irq_entry(devid, table);
3653 iommu_flush_dte(iommu, devid);
3656 static struct irq_remap_table *alloc_irq_table(u16 devid)
3658 struct irq_remap_table *table = NULL;
3659 struct irq_remap_table *new_table = NULL;
3660 struct amd_iommu *iommu;
3661 unsigned long flags;
3664 spin_lock_irqsave(&iommu_table_lock, flags);
3666 iommu = amd_iommu_rlookup_table[devid];
3670 table = irq_lookup_table[devid];
3674 alias = amd_iommu_alias_table[devid];
3675 table = irq_lookup_table[alias];
3677 set_remap_table_entry(iommu, devid, table);
3680 spin_unlock_irqrestore(&iommu_table_lock, flags);
3682 /* Nothing there yet, allocate new irq remapping table */
3683 new_table = __alloc_irq_table();
3687 spin_lock_irqsave(&iommu_table_lock, flags);
3689 table = irq_lookup_table[devid];
3693 table = irq_lookup_table[alias];
3695 set_remap_table_entry(iommu, devid, table);
3702 set_remap_table_entry(iommu, devid, table);
3704 set_remap_table_entry(iommu, alias, table);
3707 iommu_completion_wait(iommu);
3710 spin_unlock_irqrestore(&iommu_table_lock, flags);
3713 kmem_cache_free(amd_iommu_irq_cache, new_table->table);
3719 static int alloc_irq_index(u16 devid, int count, bool align)
3721 struct irq_remap_table *table;
3722 int index, c, alignment = 1;
3723 unsigned long flags;
3724 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
3729 table = alloc_irq_table(devid);
3734 alignment = roundup_pow_of_two(count);
3736 raw_spin_lock_irqsave(&table->lock, flags);
3738 /* Scan table for free entries */
3739 for (index = ALIGN(table->min_index, alignment), c = 0;
3740 index < MAX_IRQS_PER_TABLE;) {
3741 if (!iommu->irte_ops->is_allocated(table, index)) {
3745 index = ALIGN(index + 1, alignment);
3751 iommu->irte_ops->set_allocated(table, index - c + 1);
3763 raw_spin_unlock_irqrestore(&table->lock, flags);
3768 static int modify_irte_ga(u16 devid, int index, struct irte_ga *irte,
3769 struct amd_ir_data *data)
3771 struct irq_remap_table *table;
3772 struct amd_iommu *iommu;
3773 unsigned long flags;
3774 struct irte_ga *entry;
3776 iommu = amd_iommu_rlookup_table[devid];
3780 table = get_irq_table(devid);
3784 raw_spin_lock_irqsave(&table->lock, flags);
3786 entry = (struct irte_ga *)table->table;
3787 entry = &entry[index];
3788 entry->lo.fields_remap.valid = 0;
3789 entry->hi.val = irte->hi.val;
3790 entry->lo.val = irte->lo.val;
3791 entry->lo.fields_remap.valid = 1;
3795 raw_spin_unlock_irqrestore(&table->lock, flags);
3797 iommu_flush_irt(iommu, devid);
3798 iommu_completion_wait(iommu);
3803 static int modify_irte(u16 devid, int index, union irte *irte)
3805 struct irq_remap_table *table;
3806 struct amd_iommu *iommu;
3807 unsigned long flags;
3809 iommu = amd_iommu_rlookup_table[devid];
3813 table = get_irq_table(devid);
3817 raw_spin_lock_irqsave(&table->lock, flags);
3818 table->table[index] = irte->val;
3819 raw_spin_unlock_irqrestore(&table->lock, flags);
3821 iommu_flush_irt(iommu, devid);
3822 iommu_completion_wait(iommu);
3827 static void free_irte(u16 devid, int index)
3829 struct irq_remap_table *table;
3830 struct amd_iommu *iommu;
3831 unsigned long flags;
3833 iommu = amd_iommu_rlookup_table[devid];
3837 table = get_irq_table(devid);
3841 raw_spin_lock_irqsave(&table->lock, flags);
3842 iommu->irte_ops->clear_allocated(table, index);
3843 raw_spin_unlock_irqrestore(&table->lock, flags);
3845 iommu_flush_irt(iommu, devid);
3846 iommu_completion_wait(iommu);
3849 static void irte_prepare(void *entry,
3850 u32 delivery_mode, u32 dest_mode,
3851 u8 vector, u32 dest_apicid, int devid)
3853 union irte *irte = (union irte *) entry;
3856 irte->fields.vector = vector;
3857 irte->fields.int_type = delivery_mode;
3858 irte->fields.destination = dest_apicid;
3859 irte->fields.dm = dest_mode;
3860 irte->fields.valid = 1;
3863 static void irte_ga_prepare(void *entry,
3864 u32 delivery_mode, u32 dest_mode,
3865 u8 vector, u32 dest_apicid, int devid)
3867 struct irte_ga *irte = (struct irte_ga *) entry;
3871 irte->lo.fields_remap.int_type = delivery_mode;
3872 irte->lo.fields_remap.dm = dest_mode;
3873 irte->hi.fields.vector = vector;
3874 irte->lo.fields_remap.destination = APICID_TO_IRTE_DEST_LO(dest_apicid);
3875 irte->hi.fields.destination = APICID_TO_IRTE_DEST_HI(dest_apicid);
3876 irte->lo.fields_remap.valid = 1;
3879 static void irte_activate(void *entry, u16 devid, u16 index)
3881 union irte *irte = (union irte *) entry;
3883 irte->fields.valid = 1;
3884 modify_irte(devid, index, irte);
3887 static void irte_ga_activate(void *entry, u16 devid, u16 index)
3889 struct irte_ga *irte = (struct irte_ga *) entry;
3891 irte->lo.fields_remap.valid = 1;
3892 modify_irte_ga(devid, index, irte, NULL);
3895 static void irte_deactivate(void *entry, u16 devid, u16 index)
3897 union irte *irte = (union irte *) entry;
3899 irte->fields.valid = 0;
3900 modify_irte(devid, index, irte);
3903 static void irte_ga_deactivate(void *entry, u16 devid, u16 index)
3905 struct irte_ga *irte = (struct irte_ga *) entry;
3907 irte->lo.fields_remap.valid = 0;
3908 modify_irte_ga(devid, index, irte, NULL);
3911 static void irte_set_affinity(void *entry, u16 devid, u16 index,
3912 u8 vector, u32 dest_apicid)
3914 union irte *irte = (union irte *) entry;
3916 irte->fields.vector = vector;
3917 irte->fields.destination = dest_apicid;
3918 modify_irte(devid, index, irte);
3921 static void irte_ga_set_affinity(void *entry, u16 devid, u16 index,
3922 u8 vector, u32 dest_apicid)
3924 struct irte_ga *irte = (struct irte_ga *) entry;
3926 if (!irte->lo.fields_remap.guest_mode) {
3927 irte->hi.fields.vector = vector;
3928 irte->lo.fields_remap.destination =
3929 APICID_TO_IRTE_DEST_LO(dest_apicid);
3930 irte->hi.fields.destination =
3931 APICID_TO_IRTE_DEST_HI(dest_apicid);
3932 modify_irte_ga(devid, index, irte, NULL);
3936 #define IRTE_ALLOCATED (~1U)
3937 static void irte_set_allocated(struct irq_remap_table *table, int index)
3939 table->table[index] = IRTE_ALLOCATED;
3942 static void irte_ga_set_allocated(struct irq_remap_table *table, int index)
3944 struct irte_ga *ptr = (struct irte_ga *)table->table;
3945 struct irte_ga *irte = &ptr[index];
3947 memset(&irte->lo.val, 0, sizeof(u64));
3948 memset(&irte->hi.val, 0, sizeof(u64));
3949 irte->hi.fields.vector = 0xff;
3952 static bool irte_is_allocated(struct irq_remap_table *table, int index)
3954 union irte *ptr = (union irte *)table->table;
3955 union irte *irte = &ptr[index];
3957 return irte->val != 0;
3960 static bool irte_ga_is_allocated(struct irq_remap_table *table, int index)
3962 struct irte_ga *ptr = (struct irte_ga *)table->table;
3963 struct irte_ga *irte = &ptr[index];
3965 return irte->hi.fields.vector != 0;
3968 static void irte_clear_allocated(struct irq_remap_table *table, int index)
3970 table->table[index] = 0;
3973 static void irte_ga_clear_allocated(struct irq_remap_table *table, int index)
3975 struct irte_ga *ptr = (struct irte_ga *)table->table;
3976 struct irte_ga *irte = &ptr[index];
3978 memset(&irte->lo.val, 0, sizeof(u64));
3979 memset(&irte->hi.val, 0, sizeof(u64));
3982 static int get_devid(struct irq_alloc_info *info)
3986 switch (info->type) {
3987 case X86_IRQ_ALLOC_TYPE_IOAPIC:
3988 devid = get_ioapic_devid(info->ioapic_id);
3990 case X86_IRQ_ALLOC_TYPE_HPET:
3991 devid = get_hpet_devid(info->hpet_id);
3993 case X86_IRQ_ALLOC_TYPE_MSI:
3994 case X86_IRQ_ALLOC_TYPE_MSIX:
3995 devid = get_device_id(&info->msi_dev->dev);
4005 static struct irq_domain *get_ir_irq_domain(struct irq_alloc_info *info)
4007 struct amd_iommu *iommu;
4013 devid = get_devid(info);
4015 iommu = amd_iommu_rlookup_table[devid];
4017 return iommu->ir_domain;
4023 static struct irq_domain *get_irq_domain(struct irq_alloc_info *info)
4025 struct amd_iommu *iommu;
4031 switch (info->type) {
4032 case X86_IRQ_ALLOC_TYPE_MSI:
4033 case X86_IRQ_ALLOC_TYPE_MSIX:
4034 devid = get_device_id(&info->msi_dev->dev);
4038 iommu = amd_iommu_rlookup_table[devid];
4040 return iommu->msi_domain;
4049 struct irq_remap_ops amd_iommu_irq_ops = {
4050 .prepare = amd_iommu_prepare,
4051 .enable = amd_iommu_enable,
4052 .disable = amd_iommu_disable,
4053 .reenable = amd_iommu_reenable,
4054 .enable_faulting = amd_iommu_enable_faulting,
4055 .get_ir_irq_domain = get_ir_irq_domain,
4056 .get_irq_domain = get_irq_domain,
4059 static void irq_remapping_prepare_irte(struct amd_ir_data *data,
4060 struct irq_cfg *irq_cfg,
4061 struct irq_alloc_info *info,
4062 int devid, int index, int sub_handle)
4064 struct irq_2_irte *irte_info = &data->irq_2_irte;
4065 struct msi_msg *msg = &data->msi_entry;
4066 struct IO_APIC_route_entry *entry;
4067 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
4072 data->irq_2_irte.devid = devid;
4073 data->irq_2_irte.index = index + sub_handle;
4074 iommu->irte_ops->prepare(data->entry, apic->irq_delivery_mode,
4075 apic->irq_dest_mode, irq_cfg->vector,
4076 irq_cfg->dest_apicid, devid);
4078 switch (info->type) {
4079 case X86_IRQ_ALLOC_TYPE_IOAPIC:
4080 /* Setup IOAPIC entry */
4081 entry = info->ioapic_entry;
4082 info->ioapic_entry = NULL;
4083 memset(entry, 0, sizeof(*entry));
4084 entry->vector = index;
4086 entry->trigger = info->ioapic_trigger;
4087 entry->polarity = info->ioapic_polarity;
4088 /* Mask level triggered irqs. */
4089 if (info->ioapic_trigger)
4093 case X86_IRQ_ALLOC_TYPE_HPET:
4094 case X86_IRQ_ALLOC_TYPE_MSI:
4095 case X86_IRQ_ALLOC_TYPE_MSIX:
4096 msg->address_hi = MSI_ADDR_BASE_HI;
4097 msg->address_lo = MSI_ADDR_BASE_LO;
4098 msg->data = irte_info->index;
4107 struct amd_irte_ops irte_32_ops = {
4108 .prepare = irte_prepare,
4109 .activate = irte_activate,
4110 .deactivate = irte_deactivate,
4111 .set_affinity = irte_set_affinity,
4112 .set_allocated = irte_set_allocated,
4113 .is_allocated = irte_is_allocated,
4114 .clear_allocated = irte_clear_allocated,
4117 struct amd_irte_ops irte_128_ops = {
4118 .prepare = irte_ga_prepare,
4119 .activate = irte_ga_activate,
4120 .deactivate = irte_ga_deactivate,
4121 .set_affinity = irte_ga_set_affinity,
4122 .set_allocated = irte_ga_set_allocated,
4123 .is_allocated = irte_ga_is_allocated,
4124 .clear_allocated = irte_ga_clear_allocated,
4127 static int irq_remapping_alloc(struct irq_domain *domain, unsigned int virq,
4128 unsigned int nr_irqs, void *arg)
4130 struct irq_alloc_info *info = arg;
4131 struct irq_data *irq_data;
4132 struct amd_ir_data *data = NULL;
4133 struct irq_cfg *cfg;
4139 if (nr_irqs > 1 && info->type != X86_IRQ_ALLOC_TYPE_MSI &&
4140 info->type != X86_IRQ_ALLOC_TYPE_MSIX)
4144 * With IRQ remapping enabled, don't need contiguous CPU vectors
4145 * to support multiple MSI interrupts.
4147 if (info->type == X86_IRQ_ALLOC_TYPE_MSI)
4148 info->flags &= ~X86_IRQ_ALLOC_CONTIGUOUS_VECTORS;
4150 devid = get_devid(info);
4154 ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
4158 if (info->type == X86_IRQ_ALLOC_TYPE_IOAPIC) {
4159 struct irq_remap_table *table;
4160 struct amd_iommu *iommu;
4162 table = alloc_irq_table(devid);
4164 if (!table->min_index) {
4166 * Keep the first 32 indexes free for IOAPIC
4169 table->min_index = 32;
4170 iommu = amd_iommu_rlookup_table[devid];
4171 for (i = 0; i < 32; ++i)
4172 iommu->irte_ops->set_allocated(table, i);
4174 WARN_ON(table->min_index != 32);
4175 index = info->ioapic_pin;
4180 bool align = (info->type == X86_IRQ_ALLOC_TYPE_MSI);
4182 index = alloc_irq_index(devid, nr_irqs, align);
4185 pr_warn("Failed to allocate IRTE\n");
4187 goto out_free_parent;
4190 for (i = 0; i < nr_irqs; i++) {
4191 irq_data = irq_domain_get_irq_data(domain, virq + i);
4192 cfg = irqd_cfg(irq_data);
4193 if (!irq_data || !cfg) {
4199 data = kzalloc(sizeof(*data), GFP_KERNEL);
4203 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
4204 data->entry = kzalloc(sizeof(union irte), GFP_KERNEL);
4206 data->entry = kzalloc(sizeof(struct irte_ga),
4213 irq_data->hwirq = (devid << 16) + i;
4214 irq_data->chip_data = data;
4215 irq_data->chip = &amd_ir_chip;
4216 irq_remapping_prepare_irte(data, cfg, info, devid, index, i);
4217 irq_set_status_flags(virq + i, IRQ_MOVE_PCNTXT);
4223 for (i--; i >= 0; i--) {
4224 irq_data = irq_domain_get_irq_data(domain, virq + i);
4226 kfree(irq_data->chip_data);
4228 for (i = 0; i < nr_irqs; i++)
4229 free_irte(devid, index + i);
4231 irq_domain_free_irqs_common(domain, virq, nr_irqs);
4235 static void irq_remapping_free(struct irq_domain *domain, unsigned int virq,
4236 unsigned int nr_irqs)
4238 struct irq_2_irte *irte_info;
4239 struct irq_data *irq_data;
4240 struct amd_ir_data *data;
4243 for (i = 0; i < nr_irqs; i++) {
4244 irq_data = irq_domain_get_irq_data(domain, virq + i);
4245 if (irq_data && irq_data->chip_data) {
4246 data = irq_data->chip_data;
4247 irte_info = &data->irq_2_irte;
4248 free_irte(irte_info->devid, irte_info->index);
4253 irq_domain_free_irqs_common(domain, virq, nr_irqs);
4256 static void amd_ir_update_irte(struct irq_data *irqd, struct amd_iommu *iommu,
4257 struct amd_ir_data *ir_data,
4258 struct irq_2_irte *irte_info,
4259 struct irq_cfg *cfg);
4261 static int irq_remapping_activate(struct irq_domain *domain,
4262 struct irq_data *irq_data, bool reserve)
4264 struct amd_ir_data *data = irq_data->chip_data;
4265 struct irq_2_irte *irte_info = &data->irq_2_irte;
4266 struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4267 struct irq_cfg *cfg = irqd_cfg(irq_data);
4272 iommu->irte_ops->activate(data->entry, irte_info->devid,
4274 amd_ir_update_irte(irq_data, iommu, data, irte_info, cfg);
4278 static void irq_remapping_deactivate(struct irq_domain *domain,
4279 struct irq_data *irq_data)
4281 struct amd_ir_data *data = irq_data->chip_data;
4282 struct irq_2_irte *irte_info = &data->irq_2_irte;
4283 struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4286 iommu->irte_ops->deactivate(data->entry, irte_info->devid,
4290 static const struct irq_domain_ops amd_ir_domain_ops = {
4291 .alloc = irq_remapping_alloc,
4292 .free = irq_remapping_free,
4293 .activate = irq_remapping_activate,
4294 .deactivate = irq_remapping_deactivate,
4297 static int amd_ir_set_vcpu_affinity(struct irq_data *data, void *vcpu_info)
4299 struct amd_iommu *iommu;
4300 struct amd_iommu_pi_data *pi_data = vcpu_info;
4301 struct vcpu_data *vcpu_pi_info = pi_data->vcpu_data;
4302 struct amd_ir_data *ir_data = data->chip_data;
4303 struct irte_ga *irte = (struct irte_ga *) ir_data->entry;
4304 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
4305 struct iommu_dev_data *dev_data = search_dev_data(irte_info->devid);
4308 * This device has never been set up for guest mode.
4309 * we should not modify the IRTE
4311 if (!dev_data || !dev_data->use_vapic)
4314 pi_data->ir_data = ir_data;
4317 * SVM tries to set up for VAPIC mode, but we are in
4318 * legacy mode. So, we force legacy mode instead.
4320 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
4321 pr_debug("AMD-Vi: %s: Fall back to using intr legacy remap\n",
4323 pi_data->is_guest_mode = false;
4326 iommu = amd_iommu_rlookup_table[irte_info->devid];
4330 pi_data->prev_ga_tag = ir_data->cached_ga_tag;
4331 if (pi_data->is_guest_mode) {
4333 irte->hi.fields.ga_root_ptr = (pi_data->base >> 12);
4334 irte->hi.fields.vector = vcpu_pi_info->vector;
4335 irte->lo.fields_vapic.ga_log_intr = 1;
4336 irte->lo.fields_vapic.guest_mode = 1;
4337 irte->lo.fields_vapic.ga_tag = pi_data->ga_tag;
4339 ir_data->cached_ga_tag = pi_data->ga_tag;
4342 struct irq_cfg *cfg = irqd_cfg(data);
4346 irte->hi.fields.vector = cfg->vector;
4347 irte->lo.fields_remap.guest_mode = 0;
4348 irte->lo.fields_remap.destination =
4349 APICID_TO_IRTE_DEST_LO(cfg->dest_apicid);
4350 irte->hi.fields.destination =
4351 APICID_TO_IRTE_DEST_HI(cfg->dest_apicid);
4352 irte->lo.fields_remap.int_type = apic->irq_delivery_mode;
4353 irte->lo.fields_remap.dm = apic->irq_dest_mode;
4356 * This communicates the ga_tag back to the caller
4357 * so that it can do all the necessary clean up.
4359 ir_data->cached_ga_tag = 0;
4362 return modify_irte_ga(irte_info->devid, irte_info->index, irte, ir_data);
4366 static void amd_ir_update_irte(struct irq_data *irqd, struct amd_iommu *iommu,
4367 struct amd_ir_data *ir_data,
4368 struct irq_2_irte *irte_info,
4369 struct irq_cfg *cfg)
4373 * Atomically updates the IRTE with the new destination, vector
4374 * and flushes the interrupt entry cache.
4376 iommu->irte_ops->set_affinity(ir_data->entry, irte_info->devid,
4377 irte_info->index, cfg->vector,
4381 static int amd_ir_set_affinity(struct irq_data *data,
4382 const struct cpumask *mask, bool force)
4384 struct amd_ir_data *ir_data = data->chip_data;
4385 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
4386 struct irq_cfg *cfg = irqd_cfg(data);
4387 struct irq_data *parent = data->parent_data;
4388 struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4394 ret = parent->chip->irq_set_affinity(parent, mask, force);
4395 if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
4398 amd_ir_update_irte(data, iommu, ir_data, irte_info, cfg);
4400 * After this point, all the interrupts will start arriving
4401 * at the new destination. So, time to cleanup the previous
4402 * vector allocation.
4404 send_cleanup_vector(cfg);
4406 return IRQ_SET_MASK_OK_DONE;
4409 static void ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg)
4411 struct amd_ir_data *ir_data = irq_data->chip_data;
4413 *msg = ir_data->msi_entry;
4416 static struct irq_chip amd_ir_chip = {
4418 .irq_ack = apic_ack_irq,
4419 .irq_set_affinity = amd_ir_set_affinity,
4420 .irq_set_vcpu_affinity = amd_ir_set_vcpu_affinity,
4421 .irq_compose_msi_msg = ir_compose_msi_msg,
4424 int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
4426 struct fwnode_handle *fn;
4428 fn = irq_domain_alloc_named_id_fwnode("AMD-IR", iommu->index);
4431 iommu->ir_domain = irq_domain_create_tree(fn, &amd_ir_domain_ops, iommu);
4432 irq_domain_free_fwnode(fn);
4433 if (!iommu->ir_domain)
4436 iommu->ir_domain->parent = arch_get_ir_parent_domain();
4437 iommu->msi_domain = arch_create_remap_msi_irq_domain(iommu->ir_domain,
4443 int amd_iommu_update_ga(int cpu, bool is_run, void *data)
4445 unsigned long flags;
4446 struct amd_iommu *iommu;
4447 struct irq_remap_table *table;
4448 struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
4449 int devid = ir_data->irq_2_irte.devid;
4450 struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
4451 struct irte_ga *ref = (struct irte_ga *) ir_data->ref;
4453 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
4454 !ref || !entry || !entry->lo.fields_vapic.guest_mode)
4457 iommu = amd_iommu_rlookup_table[devid];
4461 table = get_irq_table(devid);
4465 raw_spin_lock_irqsave(&table->lock, flags);
4467 if (ref->lo.fields_vapic.guest_mode) {
4469 ref->lo.fields_vapic.destination =
4470 APICID_TO_IRTE_DEST_LO(cpu);
4471 ref->hi.fields.destination =
4472 APICID_TO_IRTE_DEST_HI(cpu);
4474 ref->lo.fields_vapic.is_run = is_run;
4478 raw_spin_unlock_irqrestore(&table->lock, flags);
4480 iommu_flush_irt(iommu, devid);
4481 iommu_completion_wait(iommu);
4484 EXPORT_SYMBOL(amd_iommu_update_ga);