1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (C) 2007-2010 Advanced Micro Devices, Inc.
8 #define pr_fmt(fmt) "AMD-Vi: " fmt
9 #define dev_fmt(fmt) pr_fmt(fmt)
11 #include <linux/ratelimit.h>
12 #include <linux/pci.h>
13 #include <linux/acpi.h>
14 #include <linux/pci-ats.h>
15 #include <linux/bitmap.h>
16 #include <linux/slab.h>
17 #include <linux/debugfs.h>
18 #include <linux/scatterlist.h>
19 #include <linux/dma-map-ops.h>
20 #include <linux/dma-direct.h>
21 #include <linux/iommu-helper.h>
22 #include <linux/delay.h>
23 #include <linux/amd-iommu.h>
24 #include <linux/notifier.h>
25 #include <linux/export.h>
26 #include <linux/irq.h>
27 #include <linux/msi.h>
28 #include <linux/irqdomain.h>
29 #include <linux/percpu.h>
30 #include <linux/io-pgtable.h>
31 #include <linux/cc_platform.h>
32 #include <asm/irq_remapping.h>
33 #include <asm/io_apic.h>
35 #include <asm/hw_irq.h>
36 #include <asm/proto.h>
37 #include <asm/iommu.h>
41 #include "amd_iommu.h"
42 #include "../dma-iommu.h"
43 #include "../irq_remapping.h"
45 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
47 #define LOOP_TIMEOUT 100000
49 /* IO virtual address start page frame number */
50 #define IOVA_START_PFN (1)
51 #define IOVA_PFN(addr) ((addr) >> PAGE_SHIFT)
53 /* Reserved IOVA ranges */
54 #define MSI_RANGE_START (0xfee00000)
55 #define MSI_RANGE_END (0xfeefffff)
56 #define HT_RANGE_START (0xfd00000000ULL)
57 #define HT_RANGE_END (0xffffffffffULL)
59 #define DEFAULT_PGTABLE_LEVEL PAGE_MODE_3_LEVEL
61 static DEFINE_SPINLOCK(pd_bitmap_lock);
63 LIST_HEAD(ioapic_map);
65 LIST_HEAD(acpihid_map);
67 const struct iommu_ops amd_iommu_ops;
69 static ATOMIC_NOTIFIER_HEAD(ppr_notifier);
70 int amd_iommu_max_glx_val = -1;
73 * general struct to manage commands send to an IOMMU
79 struct kmem_cache *amd_iommu_irq_cache;
81 static void detach_device(struct device *dev);
82 static int domain_enable_v2(struct protection_domain *domain, int pasids);
84 /****************************************************************************
88 ****************************************************************************/
90 static inline int get_acpihid_device_id(struct device *dev,
91 struct acpihid_map_entry **entry)
93 struct acpi_device *adev = ACPI_COMPANION(dev);
94 struct acpihid_map_entry *p;
99 list_for_each_entry(p, &acpihid_map, list) {
100 if (acpi_dev_hid_uid_match(adev, p->hid,
101 p->uid[0] ? p->uid : NULL)) {
110 static inline int get_device_sbdf_id(struct device *dev)
115 sbdf = get_pci_sbdf_id(to_pci_dev(dev));
117 sbdf = get_acpihid_device_id(dev, NULL);
122 struct dev_table_entry *get_dev_table(struct amd_iommu *iommu)
124 struct dev_table_entry *dev_table;
125 struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
127 BUG_ON(pci_seg == NULL);
128 dev_table = pci_seg->dev_table;
129 BUG_ON(dev_table == NULL);
134 static inline u16 get_device_segment(struct device *dev)
138 if (dev_is_pci(dev)) {
139 struct pci_dev *pdev = to_pci_dev(dev);
141 seg = pci_domain_nr(pdev->bus);
143 u32 devid = get_acpihid_device_id(dev, NULL);
145 seg = PCI_SBDF_TO_SEGID(devid);
151 /* Writes the specific IOMMU for a device into the PCI segment rlookup table */
152 void amd_iommu_set_rlookup_table(struct amd_iommu *iommu, u16 devid)
154 struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
156 pci_seg->rlookup_table[devid] = iommu;
159 static struct amd_iommu *__rlookup_amd_iommu(u16 seg, u16 devid)
161 struct amd_iommu_pci_seg *pci_seg;
163 for_each_pci_segment(pci_seg) {
164 if (pci_seg->id == seg)
165 return pci_seg->rlookup_table[devid];
170 static struct amd_iommu *rlookup_amd_iommu(struct device *dev)
172 u16 seg = get_device_segment(dev);
173 int devid = get_device_sbdf_id(dev);
177 return __rlookup_amd_iommu(seg, PCI_SBDF_TO_DEVID(devid));
180 static struct protection_domain *to_pdomain(struct iommu_domain *dom)
182 return container_of(dom, struct protection_domain, domain);
185 static struct iommu_dev_data *alloc_dev_data(struct amd_iommu *iommu, u16 devid)
187 struct iommu_dev_data *dev_data;
188 struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
190 dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
194 spin_lock_init(&dev_data->lock);
195 dev_data->devid = devid;
196 ratelimit_default_init(&dev_data->rs);
198 llist_add(&dev_data->dev_data_list, &pci_seg->dev_data_list);
202 static struct iommu_dev_data *search_dev_data(struct amd_iommu *iommu, u16 devid)
204 struct iommu_dev_data *dev_data;
205 struct llist_node *node;
206 struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
208 if (llist_empty(&pci_seg->dev_data_list))
211 node = pci_seg->dev_data_list.first;
212 llist_for_each_entry(dev_data, node, dev_data_list) {
213 if (dev_data->devid == devid)
220 static int clone_alias(struct pci_dev *pdev, u16 alias, void *data)
222 struct amd_iommu *iommu;
223 struct dev_table_entry *dev_table;
224 u16 devid = pci_dev_id(pdev);
229 iommu = rlookup_amd_iommu(&pdev->dev);
233 amd_iommu_set_rlookup_table(iommu, alias);
234 dev_table = get_dev_table(iommu);
235 memcpy(dev_table[alias].data,
236 dev_table[devid].data,
237 sizeof(dev_table[alias].data));
242 static void clone_aliases(struct amd_iommu *iommu, struct device *dev)
244 struct pci_dev *pdev;
246 if (!dev_is_pci(dev))
248 pdev = to_pci_dev(dev);
251 * The IVRS alias stored in the alias table may not be
252 * part of the PCI DMA aliases if it's bus differs
253 * from the original device.
255 clone_alias(pdev, iommu->pci_seg->alias_table[pci_dev_id(pdev)], NULL);
257 pci_for_each_dma_alias(pdev, clone_alias, NULL);
260 static void setup_aliases(struct amd_iommu *iommu, struct device *dev)
262 struct pci_dev *pdev = to_pci_dev(dev);
263 struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
266 /* For ACPI HID devices, there are no aliases */
267 if (!dev_is_pci(dev))
271 * Add the IVRS alias to the pci aliases if it is on the same
272 * bus. The IVRS table may know about a quirk that we don't.
274 ivrs_alias = pci_seg->alias_table[pci_dev_id(pdev)];
275 if (ivrs_alias != pci_dev_id(pdev) &&
276 PCI_BUS_NUM(ivrs_alias) == pdev->bus->number)
277 pci_add_dma_alias(pdev, ivrs_alias & 0xff, 1);
279 clone_aliases(iommu, dev);
282 static struct iommu_dev_data *find_dev_data(struct amd_iommu *iommu, u16 devid)
284 struct iommu_dev_data *dev_data;
286 dev_data = search_dev_data(iommu, devid);
288 if (dev_data == NULL) {
289 dev_data = alloc_dev_data(iommu, devid);
293 if (translation_pre_enabled(iommu))
294 dev_data->defer_attach = true;
301 * Find or create an IOMMU group for a acpihid device.
303 static struct iommu_group *acpihid_device_group(struct device *dev)
305 struct acpihid_map_entry *p, *entry = NULL;
308 devid = get_acpihid_device_id(dev, &entry);
310 return ERR_PTR(devid);
312 list_for_each_entry(p, &acpihid_map, list) {
313 if ((devid == p->devid) && p->group)
314 entry->group = p->group;
318 entry->group = generic_device_group(dev);
320 iommu_group_ref_get(entry->group);
325 static bool pci_iommuv2_capable(struct pci_dev *pdev)
327 static const int caps[] = {
329 PCI_EXT_CAP_ID_PASID,
333 if (!pci_ats_supported(pdev))
336 for (i = 0; i < 2; ++i) {
337 pos = pci_find_ext_capability(pdev, caps[i]);
346 * This function checks if the driver got a valid device from the caller to
347 * avoid dereferencing invalid pointers.
349 static bool check_device(struct device *dev)
351 struct amd_iommu_pci_seg *pci_seg;
352 struct amd_iommu *iommu;
358 sbdf = get_device_sbdf_id(dev);
361 devid = PCI_SBDF_TO_DEVID(sbdf);
363 iommu = rlookup_amd_iommu(dev);
367 /* Out of our scope? */
368 pci_seg = iommu->pci_seg;
369 if (devid > pci_seg->last_bdf)
375 static int iommu_init_device(struct amd_iommu *iommu, struct device *dev)
377 struct iommu_dev_data *dev_data;
380 if (dev_iommu_priv_get(dev))
383 sbdf = get_device_sbdf_id(dev);
387 devid = PCI_SBDF_TO_DEVID(sbdf);
388 dev_data = find_dev_data(iommu, devid);
393 setup_aliases(iommu, dev);
396 * By default we use passthrough mode for IOMMUv2 capable device.
397 * But if amd_iommu=force_isolation is set (e.g. to debug DMA to
398 * invalid address), we ignore the capability for the device so
399 * it'll be forced to go into translation mode.
401 if ((iommu_default_passthrough() || !amd_iommu_force_isolation) &&
402 dev_is_pci(dev) && pci_iommuv2_capable(to_pci_dev(dev))) {
403 dev_data->iommu_v2 = iommu->is_iommu_v2;
406 dev_iommu_priv_set(dev, dev_data);
411 static void iommu_ignore_device(struct amd_iommu *iommu, struct device *dev)
413 struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
414 struct dev_table_entry *dev_table = get_dev_table(iommu);
417 sbdf = get_device_sbdf_id(dev);
421 devid = PCI_SBDF_TO_DEVID(sbdf);
422 pci_seg->rlookup_table[devid] = NULL;
423 memset(&dev_table[devid], 0, sizeof(struct dev_table_entry));
425 setup_aliases(iommu, dev);
428 static void amd_iommu_uninit_device(struct device *dev)
430 struct iommu_dev_data *dev_data;
432 dev_data = dev_iommu_priv_get(dev);
436 if (dev_data->domain)
439 dev_iommu_priv_set(dev, NULL);
442 * We keep dev_data around for unplugged devices and reuse it when the
443 * device is re-plugged - not doing so would introduce a ton of races.
447 /****************************************************************************
449 * Interrupt handling functions
451 ****************************************************************************/
453 static void dump_dte_entry(struct amd_iommu *iommu, u16 devid)
456 struct dev_table_entry *dev_table = get_dev_table(iommu);
458 for (i = 0; i < 4; ++i)
459 pr_err("DTE[%d]: %016llx\n", i, dev_table[devid].data[i]);
462 static void dump_command(unsigned long phys_addr)
464 struct iommu_cmd *cmd = iommu_phys_to_virt(phys_addr);
467 for (i = 0; i < 4; ++i)
468 pr_err("CMD[%d]: %08x\n", i, cmd->data[i]);
471 static void amd_iommu_report_rmp_hw_error(struct amd_iommu *iommu, volatile u32 *event)
473 struct iommu_dev_data *dev_data = NULL;
474 int devid, vmg_tag, flags;
475 struct pci_dev *pdev;
478 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
479 vmg_tag = (event[1]) & 0xFFFF;
480 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
481 spa = ((u64)event[3] << 32) | (event[2] & 0xFFFFFFF8);
483 pdev = pci_get_domain_bus_and_slot(iommu->pci_seg->id, PCI_BUS_NUM(devid),
486 dev_data = dev_iommu_priv_get(&pdev->dev);
489 if (__ratelimit(&dev_data->rs)) {
490 pci_err(pdev, "Event logged [RMP_HW_ERROR vmg_tag=0x%04x, spa=0x%llx, flags=0x%04x]\n",
491 vmg_tag, spa, flags);
494 pr_err_ratelimited("Event logged [RMP_HW_ERROR device=%04x:%02x:%02x.%x, vmg_tag=0x%04x, spa=0x%llx, flags=0x%04x]\n",
495 iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
496 vmg_tag, spa, flags);
503 static void amd_iommu_report_rmp_fault(struct amd_iommu *iommu, volatile u32 *event)
505 struct iommu_dev_data *dev_data = NULL;
506 int devid, flags_rmp, vmg_tag, flags;
507 struct pci_dev *pdev;
510 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
511 flags_rmp = (event[0] >> EVENT_FLAGS_SHIFT) & 0xFF;
512 vmg_tag = (event[1]) & 0xFFFF;
513 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
514 gpa = ((u64)event[3] << 32) | event[2];
516 pdev = pci_get_domain_bus_and_slot(iommu->pci_seg->id, PCI_BUS_NUM(devid),
519 dev_data = dev_iommu_priv_get(&pdev->dev);
522 if (__ratelimit(&dev_data->rs)) {
523 pci_err(pdev, "Event logged [RMP_PAGE_FAULT vmg_tag=0x%04x, gpa=0x%llx, flags_rmp=0x%04x, flags=0x%04x]\n",
524 vmg_tag, gpa, flags_rmp, flags);
527 pr_err_ratelimited("Event logged [RMP_PAGE_FAULT device=%04x:%02x:%02x.%x, vmg_tag=0x%04x, gpa=0x%llx, flags_rmp=0x%04x, flags=0x%04x]\n",
528 iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
529 vmg_tag, gpa, flags_rmp, flags);
536 #define IS_IOMMU_MEM_TRANSACTION(flags) \
537 (((flags) & EVENT_FLAG_I) == 0)
539 #define IS_WRITE_REQUEST(flags) \
540 ((flags) & EVENT_FLAG_RW)
542 static void amd_iommu_report_page_fault(struct amd_iommu *iommu,
543 u16 devid, u16 domain_id,
544 u64 address, int flags)
546 struct iommu_dev_data *dev_data = NULL;
547 struct pci_dev *pdev;
549 pdev = pci_get_domain_bus_and_slot(iommu->pci_seg->id, PCI_BUS_NUM(devid),
552 dev_data = dev_iommu_priv_get(&pdev->dev);
556 * If this is a DMA fault (for which the I(nterrupt)
557 * bit will be unset), allow report_iommu_fault() to
558 * prevent logging it.
560 if (IS_IOMMU_MEM_TRANSACTION(flags)) {
561 /* Device not attached to domain properly */
562 if (dev_data->domain == NULL) {
563 pr_err_ratelimited("Event logged [Device not attached to domain properly]\n");
564 pr_err_ratelimited(" device=%04x:%02x:%02x.%x domain=0x%04x\n",
565 iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid),
566 PCI_FUNC(devid), domain_id);
570 if (!report_iommu_fault(&dev_data->domain->domain,
572 IS_WRITE_REQUEST(flags) ?
578 if (__ratelimit(&dev_data->rs)) {
579 pci_err(pdev, "Event logged [IO_PAGE_FAULT domain=0x%04x address=0x%llx flags=0x%04x]\n",
580 domain_id, address, flags);
583 pr_err_ratelimited("Event logged [IO_PAGE_FAULT device=%04x:%02x:%02x.%x domain=0x%04x address=0x%llx flags=0x%04x]\n",
584 iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
585 domain_id, address, flags);
593 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
595 struct device *dev = iommu->iommu.dev;
596 int type, devid, flags, tag;
597 volatile u32 *event = __evt;
603 type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
604 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
605 pasid = (event[0] & EVENT_DOMID_MASK_HI) |
606 (event[1] & EVENT_DOMID_MASK_LO);
607 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
608 address = (u64)(((u64)event[3]) << 32) | event[2];
611 /* Did we hit the erratum? */
612 if (++count == LOOP_TIMEOUT) {
613 pr_err("No event written to event log\n");
620 if (type == EVENT_TYPE_IO_FAULT) {
621 amd_iommu_report_page_fault(iommu, devid, pasid, address, flags);
626 case EVENT_TYPE_ILL_DEV:
627 dev_err(dev, "Event logged [ILLEGAL_DEV_TABLE_ENTRY device=%04x:%02x:%02x.%x pasid=0x%05x address=0x%llx flags=0x%04x]\n",
628 iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
629 pasid, address, flags);
630 dump_dte_entry(iommu, devid);
632 case EVENT_TYPE_DEV_TAB_ERR:
633 dev_err(dev, "Event logged [DEV_TAB_HARDWARE_ERROR device=%04x:%02x:%02x.%x "
634 "address=0x%llx flags=0x%04x]\n",
635 iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
638 case EVENT_TYPE_PAGE_TAB_ERR:
639 dev_err(dev, "Event logged [PAGE_TAB_HARDWARE_ERROR device=%04x:%02x:%02x.%x pasid=0x%04x address=0x%llx flags=0x%04x]\n",
640 iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
641 pasid, address, flags);
643 case EVENT_TYPE_ILL_CMD:
644 dev_err(dev, "Event logged [ILLEGAL_COMMAND_ERROR address=0x%llx]\n", address);
645 dump_command(address);
647 case EVENT_TYPE_CMD_HARD_ERR:
648 dev_err(dev, "Event logged [COMMAND_HARDWARE_ERROR address=0x%llx flags=0x%04x]\n",
651 case EVENT_TYPE_IOTLB_INV_TO:
652 dev_err(dev, "Event logged [IOTLB_INV_TIMEOUT device=%04x:%02x:%02x.%x address=0x%llx]\n",
653 iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
656 case EVENT_TYPE_INV_DEV_REQ:
657 dev_err(dev, "Event logged [INVALID_DEVICE_REQUEST device=%04x:%02x:%02x.%x pasid=0x%05x address=0x%llx flags=0x%04x]\n",
658 iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
659 pasid, address, flags);
661 case EVENT_TYPE_RMP_FAULT:
662 amd_iommu_report_rmp_fault(iommu, event);
664 case EVENT_TYPE_RMP_HW_ERR:
665 amd_iommu_report_rmp_hw_error(iommu, event);
667 case EVENT_TYPE_INV_PPR_REQ:
668 pasid = PPR_PASID(*((u64 *)__evt));
669 tag = event[1] & 0x03FF;
670 dev_err(dev, "Event logged [INVALID_PPR_REQUEST device=%04x:%02x:%02x.%x pasid=0x%05x address=0x%llx flags=0x%04x tag=0x%03x]\n",
671 iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
672 pasid, address, flags, tag);
675 dev_err(dev, "Event logged [UNKNOWN event[0]=0x%08x event[1]=0x%08x event[2]=0x%08x event[3]=0x%08x\n",
676 event[0], event[1], event[2], event[3]);
680 * To detect the hardware errata 732 we need to clear the
681 * entry back to zero. This issue does not exist on SNP
682 * enabled system. Also this buffer is not writeable on
683 * SNP enabled system.
685 if (!amd_iommu_snp_en)
686 memset(__evt, 0, 4 * sizeof(u32));
689 static void iommu_poll_events(struct amd_iommu *iommu)
693 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
694 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
696 while (head != tail) {
697 iommu_print_event(iommu, iommu->evt_buf + head);
698 head = (head + EVENT_ENTRY_SIZE) % EVT_BUFFER_SIZE;
701 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
704 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
706 struct amd_iommu_fault fault;
708 if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
709 pr_err_ratelimited("Unknown PPR request received\n");
713 fault.address = raw[1];
714 fault.pasid = PPR_PASID(raw[0]);
715 fault.sbdf = PCI_SEG_DEVID_TO_SBDF(iommu->pci_seg->id, PPR_DEVID(raw[0]));
716 fault.tag = PPR_TAG(raw[0]);
717 fault.flags = PPR_FLAGS(raw[0]);
719 atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
722 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
726 if (iommu->ppr_log == NULL)
729 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
730 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
732 while (head != tail) {
737 raw = (u64 *)(iommu->ppr_log + head);
740 * Hardware bug: Interrupt may arrive before the entry is
741 * written to memory. If this happens we need to wait for the
744 for (i = 0; i < LOOP_TIMEOUT; ++i) {
745 if (PPR_REQ_TYPE(raw[0]) != 0)
750 /* Avoid memcpy function-call overhead */
755 * To detect the hardware errata 733 we need to clear the
756 * entry back to zero. This issue does not exist on SNP
757 * enabled system. Also this buffer is not writeable on
758 * SNP enabled system.
760 if (!amd_iommu_snp_en)
761 raw[0] = raw[1] = 0UL;
763 /* Update head pointer of hardware ring-buffer */
764 head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
765 writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
767 /* Handle PPR entry */
768 iommu_handle_ppr_entry(iommu, entry);
770 /* Refresh ring-buffer information */
771 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
772 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
776 #ifdef CONFIG_IRQ_REMAP
777 static int (*iommu_ga_log_notifier)(u32);
779 int amd_iommu_register_ga_log_notifier(int (*notifier)(u32))
781 iommu_ga_log_notifier = notifier;
785 EXPORT_SYMBOL(amd_iommu_register_ga_log_notifier);
787 static void iommu_poll_ga_log(struct amd_iommu *iommu)
791 if (iommu->ga_log == NULL)
794 head = readl(iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
795 tail = readl(iommu->mmio_base + MMIO_GA_TAIL_OFFSET);
797 while (head != tail) {
801 raw = (u64 *)(iommu->ga_log + head);
803 /* Avoid memcpy function-call overhead */
806 /* Update head pointer of hardware ring-buffer */
807 head = (head + GA_ENTRY_SIZE) % GA_LOG_SIZE;
808 writel(head, iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
810 /* Handle GA entry */
811 switch (GA_REQ_TYPE(log_entry)) {
813 if (!iommu_ga_log_notifier)
816 pr_debug("%s: devid=%#x, ga_tag=%#x\n",
817 __func__, GA_DEVID(log_entry),
820 if (iommu_ga_log_notifier(GA_TAG(log_entry)) != 0)
821 pr_err("GA log notifier failed.\n");
830 amd_iommu_set_pci_msi_domain(struct device *dev, struct amd_iommu *iommu)
832 if (!irq_remapping_enabled || !dev_is_pci(dev) ||
833 !pci_dev_has_default_msi_parent_domain(to_pci_dev(dev)))
836 dev_set_msi_domain(dev, iommu->ir_domain);
839 #else /* CONFIG_IRQ_REMAP */
841 amd_iommu_set_pci_msi_domain(struct device *dev, struct amd_iommu *iommu) { }
842 #endif /* !CONFIG_IRQ_REMAP */
844 #define AMD_IOMMU_INT_MASK \
845 (MMIO_STATUS_EVT_OVERFLOW_INT_MASK | \
846 MMIO_STATUS_EVT_INT_MASK | \
847 MMIO_STATUS_PPR_INT_MASK | \
848 MMIO_STATUS_GALOG_OVERFLOW_MASK | \
849 MMIO_STATUS_GALOG_INT_MASK)
851 irqreturn_t amd_iommu_int_thread(int irq, void *data)
853 struct amd_iommu *iommu = (struct amd_iommu *) data;
854 u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
856 while (status & AMD_IOMMU_INT_MASK) {
857 /* Enable interrupt sources again */
858 writel(AMD_IOMMU_INT_MASK,
859 iommu->mmio_base + MMIO_STATUS_OFFSET);
861 if (status & MMIO_STATUS_EVT_INT_MASK) {
862 pr_devel("Processing IOMMU Event Log\n");
863 iommu_poll_events(iommu);
866 if (status & MMIO_STATUS_PPR_INT_MASK) {
867 pr_devel("Processing IOMMU PPR Log\n");
868 iommu_poll_ppr_log(iommu);
871 #ifdef CONFIG_IRQ_REMAP
872 if (status & (MMIO_STATUS_GALOG_INT_MASK |
873 MMIO_STATUS_GALOG_OVERFLOW_MASK)) {
874 pr_devel("Processing IOMMU GA Log\n");
875 iommu_poll_ga_log(iommu);
878 if (status & MMIO_STATUS_GALOG_OVERFLOW_MASK) {
879 pr_info_ratelimited("IOMMU GA Log overflow\n");
880 amd_iommu_restart_ga_log(iommu);
884 if (status & MMIO_STATUS_EVT_OVERFLOW_INT_MASK) {
885 pr_info_ratelimited("IOMMU event log overflow\n");
886 amd_iommu_restart_event_logging(iommu);
890 * Hardware bug: ERBT1312
891 * When re-enabling interrupt (by writing 1
892 * to clear the bit), the hardware might also try to set
893 * the interrupt bit in the event status register.
894 * In this scenario, the bit will be set, and disable
895 * subsequent interrupts.
897 * Workaround: The IOMMU driver should read back the
898 * status register and check if the interrupt bits are cleared.
899 * If not, driver will need to go through the interrupt handler
900 * again and re-clear the bits
902 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
907 irqreturn_t amd_iommu_int_handler(int irq, void *data)
909 return IRQ_WAKE_THREAD;
912 /****************************************************************************
914 * IOMMU command queuing functions
916 ****************************************************************************/
918 static int wait_on_sem(struct amd_iommu *iommu, u64 data)
922 while (*iommu->cmd_sem != data && i < LOOP_TIMEOUT) {
927 if (i == LOOP_TIMEOUT) {
928 pr_alert("Completion-Wait loop timed out\n");
935 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
936 struct iommu_cmd *cmd)
941 /* Copy command to buffer */
942 tail = iommu->cmd_buf_tail;
943 target = iommu->cmd_buf + tail;
944 memcpy(target, cmd, sizeof(*cmd));
946 tail = (tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
947 iommu->cmd_buf_tail = tail;
949 /* Tell the IOMMU about it */
950 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
953 static void build_completion_wait(struct iommu_cmd *cmd,
954 struct amd_iommu *iommu,
957 u64 paddr = iommu_virt_to_phys((void *)iommu->cmd_sem);
959 memset(cmd, 0, sizeof(*cmd));
960 cmd->data[0] = lower_32_bits(paddr) | CMD_COMPL_WAIT_STORE_MASK;
961 cmd->data[1] = upper_32_bits(paddr);
962 cmd->data[2] = lower_32_bits(data);
963 cmd->data[3] = upper_32_bits(data);
964 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
967 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
969 memset(cmd, 0, sizeof(*cmd));
970 cmd->data[0] = devid;
971 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
975 * Builds an invalidation address which is suitable for one page or multiple
976 * pages. Sets the size bit (S) as needed is more than one page is flushed.
978 static inline u64 build_inv_address(u64 address, size_t size)
980 u64 pages, end, msb_diff;
982 pages = iommu_num_pages(address, size, PAGE_SIZE);
985 return address & PAGE_MASK;
987 end = address + size - 1;
990 * msb_diff would hold the index of the most significant bit that
991 * flipped between the start and end.
993 msb_diff = fls64(end ^ address) - 1;
996 * Bits 63:52 are sign extended. If for some reason bit 51 is different
997 * between the start and the end, invalidate everything.
999 if (unlikely(msb_diff > 51)) {
1000 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
1003 * The msb-bit must be clear on the address. Just set all the
1006 address |= (1ull << msb_diff) - 1;
1009 /* Clear bits 11:0 */
1010 address &= PAGE_MASK;
1012 /* Set the size bit - we flush more than one 4kb page */
1013 return address | CMD_INV_IOMMU_PAGES_SIZE_MASK;
1016 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
1017 size_t size, u16 domid, int pde)
1019 u64 inv_address = build_inv_address(address, size);
1021 memset(cmd, 0, sizeof(*cmd));
1022 cmd->data[1] |= domid;
1023 cmd->data[2] = lower_32_bits(inv_address);
1024 cmd->data[3] = upper_32_bits(inv_address);
1025 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
1026 if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
1027 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
1030 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
1031 u64 address, size_t size)
1033 u64 inv_address = build_inv_address(address, size);
1035 memset(cmd, 0, sizeof(*cmd));
1036 cmd->data[0] = devid;
1037 cmd->data[0] |= (qdep & 0xff) << 24;
1038 cmd->data[1] = devid;
1039 cmd->data[2] = lower_32_bits(inv_address);
1040 cmd->data[3] = upper_32_bits(inv_address);
1041 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
1044 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, u32 pasid,
1045 u64 address, bool size)
1047 memset(cmd, 0, sizeof(*cmd));
1049 address &= ~(0xfffULL);
1051 cmd->data[0] = pasid;
1052 cmd->data[1] = domid;
1053 cmd->data[2] = lower_32_bits(address);
1054 cmd->data[3] = upper_32_bits(address);
1055 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
1056 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
1058 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
1059 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
1062 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, u32 pasid,
1063 int qdep, u64 address, bool size)
1065 memset(cmd, 0, sizeof(*cmd));
1067 address &= ~(0xfffULL);
1069 cmd->data[0] = devid;
1070 cmd->data[0] |= ((pasid >> 8) & 0xff) << 16;
1071 cmd->data[0] |= (qdep & 0xff) << 24;
1072 cmd->data[1] = devid;
1073 cmd->data[1] |= (pasid & 0xff) << 16;
1074 cmd->data[2] = lower_32_bits(address);
1075 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
1076 cmd->data[3] = upper_32_bits(address);
1078 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
1079 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
1082 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, u32 pasid,
1083 int status, int tag, bool gn)
1085 memset(cmd, 0, sizeof(*cmd));
1087 cmd->data[0] = devid;
1089 cmd->data[1] = pasid;
1090 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
1092 cmd->data[3] = tag & 0x1ff;
1093 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
1095 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
1098 static void build_inv_all(struct iommu_cmd *cmd)
1100 memset(cmd, 0, sizeof(*cmd));
1101 CMD_SET_TYPE(cmd, CMD_INV_ALL);
1104 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
1106 memset(cmd, 0, sizeof(*cmd));
1107 cmd->data[0] = devid;
1108 CMD_SET_TYPE(cmd, CMD_INV_IRT);
1112 * Writes the command to the IOMMUs command buffer and informs the
1113 * hardware about the new command.
1115 static int __iommu_queue_command_sync(struct amd_iommu *iommu,
1116 struct iommu_cmd *cmd,
1119 unsigned int count = 0;
1120 u32 left, next_tail;
1122 next_tail = (iommu->cmd_buf_tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
1124 left = (iommu->cmd_buf_head - next_tail) % CMD_BUFFER_SIZE;
1127 /* Skip udelay() the first time around */
1129 if (count == LOOP_TIMEOUT) {
1130 pr_err("Command buffer timeout\n");
1137 /* Update head and recheck remaining space */
1138 iommu->cmd_buf_head = readl(iommu->mmio_base +
1139 MMIO_CMD_HEAD_OFFSET);
1144 copy_cmd_to_buffer(iommu, cmd);
1146 /* Do we need to make sure all commands are processed? */
1147 iommu->need_sync = sync;
1152 static int iommu_queue_command_sync(struct amd_iommu *iommu,
1153 struct iommu_cmd *cmd,
1156 unsigned long flags;
1159 raw_spin_lock_irqsave(&iommu->lock, flags);
1160 ret = __iommu_queue_command_sync(iommu, cmd, sync);
1161 raw_spin_unlock_irqrestore(&iommu->lock, flags);
1166 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
1168 return iommu_queue_command_sync(iommu, cmd, true);
1172 * This function queues a completion wait command into the command
1173 * buffer of an IOMMU
1175 static int iommu_completion_wait(struct amd_iommu *iommu)
1177 struct iommu_cmd cmd;
1178 unsigned long flags;
1182 if (!iommu->need_sync)
1185 data = atomic64_add_return(1, &iommu->cmd_sem_val);
1186 build_completion_wait(&cmd, iommu, data);
1188 raw_spin_lock_irqsave(&iommu->lock, flags);
1190 ret = __iommu_queue_command_sync(iommu, &cmd, false);
1194 ret = wait_on_sem(iommu, data);
1197 raw_spin_unlock_irqrestore(&iommu->lock, flags);
1202 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
1204 struct iommu_cmd cmd;
1206 build_inv_dte(&cmd, devid);
1208 return iommu_queue_command(iommu, &cmd);
1211 static void amd_iommu_flush_dte_all(struct amd_iommu *iommu)
1214 u16 last_bdf = iommu->pci_seg->last_bdf;
1216 for (devid = 0; devid <= last_bdf; ++devid)
1217 iommu_flush_dte(iommu, devid);
1219 iommu_completion_wait(iommu);
1223 * This function uses heavy locking and may disable irqs for some time. But
1224 * this is no issue because it is only called during resume.
1226 static void amd_iommu_flush_tlb_all(struct amd_iommu *iommu)
1229 u16 last_bdf = iommu->pci_seg->last_bdf;
1231 for (dom_id = 0; dom_id <= last_bdf; ++dom_id) {
1232 struct iommu_cmd cmd;
1233 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1235 iommu_queue_command(iommu, &cmd);
1238 iommu_completion_wait(iommu);
1241 static void amd_iommu_flush_tlb_domid(struct amd_iommu *iommu, u32 dom_id)
1243 struct iommu_cmd cmd;
1245 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1247 iommu_queue_command(iommu, &cmd);
1249 iommu_completion_wait(iommu);
1252 static void amd_iommu_flush_all(struct amd_iommu *iommu)
1254 struct iommu_cmd cmd;
1256 build_inv_all(&cmd);
1258 iommu_queue_command(iommu, &cmd);
1259 iommu_completion_wait(iommu);
1262 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1264 struct iommu_cmd cmd;
1266 build_inv_irt(&cmd, devid);
1268 iommu_queue_command(iommu, &cmd);
1271 static void amd_iommu_flush_irt_all(struct amd_iommu *iommu)
1274 u16 last_bdf = iommu->pci_seg->last_bdf;
1276 if (iommu->irtcachedis_enabled)
1279 for (devid = 0; devid <= last_bdf; devid++)
1280 iommu_flush_irt(iommu, devid);
1282 iommu_completion_wait(iommu);
1285 void iommu_flush_all_caches(struct amd_iommu *iommu)
1287 if (iommu_feature(iommu, FEATURE_IA)) {
1288 amd_iommu_flush_all(iommu);
1290 amd_iommu_flush_dte_all(iommu);
1291 amd_iommu_flush_irt_all(iommu);
1292 amd_iommu_flush_tlb_all(iommu);
1297 * Command send function for flushing on-device TLB
1299 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1300 u64 address, size_t size)
1302 struct amd_iommu *iommu;
1303 struct iommu_cmd cmd;
1306 qdep = dev_data->ats.qdep;
1307 iommu = rlookup_amd_iommu(dev_data->dev);
1311 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1313 return iommu_queue_command(iommu, &cmd);
1316 static int device_flush_dte_alias(struct pci_dev *pdev, u16 alias, void *data)
1318 struct amd_iommu *iommu = data;
1320 return iommu_flush_dte(iommu, alias);
1324 * Command send function for invalidating a device table entry
1326 static int device_flush_dte(struct iommu_dev_data *dev_data)
1328 struct amd_iommu *iommu;
1329 struct pci_dev *pdev = NULL;
1330 struct amd_iommu_pci_seg *pci_seg;
1334 iommu = rlookup_amd_iommu(dev_data->dev);
1338 if (dev_is_pci(dev_data->dev))
1339 pdev = to_pci_dev(dev_data->dev);
1342 ret = pci_for_each_dma_alias(pdev,
1343 device_flush_dte_alias, iommu);
1345 ret = iommu_flush_dte(iommu, dev_data->devid);
1349 pci_seg = iommu->pci_seg;
1350 alias = pci_seg->alias_table[dev_data->devid];
1351 if (alias != dev_data->devid) {
1352 ret = iommu_flush_dte(iommu, alias);
1357 if (dev_data->ats.enabled)
1358 ret = device_flush_iotlb(dev_data, 0, ~0UL);
1364 * TLB invalidation function which is called from the mapping functions.
1365 * It invalidates a single PTE if the range to flush is within a single
1366 * page. Otherwise it flushes the whole TLB of the IOMMU.
1368 static void __domain_flush_pages(struct protection_domain *domain,
1369 u64 address, size_t size, int pde)
1371 struct iommu_dev_data *dev_data;
1372 struct iommu_cmd cmd;
1375 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1377 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
1378 if (!domain->dev_iommu[i])
1382 * Devices of this domain are behind this IOMMU
1383 * We need a TLB flush
1385 ret |= iommu_queue_command(amd_iommus[i], &cmd);
1388 list_for_each_entry(dev_data, &domain->dev_list, list) {
1390 if (!dev_data->ats.enabled)
1393 ret |= device_flush_iotlb(dev_data, address, size);
1399 static void domain_flush_pages(struct protection_domain *domain,
1400 u64 address, size_t size, int pde)
1402 if (likely(!amd_iommu_np_cache)) {
1403 __domain_flush_pages(domain, address, size, pde);
1408 * When NpCache is on, we infer that we run in a VM and use a vIOMMU.
1409 * In such setups it is best to avoid flushes of ranges which are not
1410 * naturally aligned, since it would lead to flushes of unmodified
1411 * PTEs. Such flushes would require the hypervisor to do more work than
1412 * necessary. Therefore, perform repeated flushes of aligned ranges
1413 * until you cover the range. Each iteration flushes the smaller
1414 * between the natural alignment of the address that we flush and the
1415 * greatest naturally aligned region that fits in the range.
1418 int addr_alignment = __ffs(address);
1419 int size_alignment = __fls(size);
1424 * size is always non-zero, but address might be zero, causing
1425 * addr_alignment to be negative. As the casting of the
1426 * argument in __ffs(address) to long might trim the high bits
1427 * of the address on x86-32, cast to long when doing the check.
1429 if (likely((unsigned long)address != 0))
1430 min_alignment = min(addr_alignment, size_alignment);
1432 min_alignment = size_alignment;
1434 flush_size = 1ul << min_alignment;
1436 __domain_flush_pages(domain, address, flush_size, pde);
1437 address += flush_size;
1442 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1443 void amd_iommu_domain_flush_tlb_pde(struct protection_domain *domain)
1445 domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1448 void amd_iommu_domain_flush_complete(struct protection_domain *domain)
1452 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
1453 if (domain && !domain->dev_iommu[i])
1457 * Devices of this domain are behind this IOMMU
1458 * We need to wait for completion of all commands.
1460 iommu_completion_wait(amd_iommus[i]);
1464 /* Flush the not present cache if it exists */
1465 static void domain_flush_np_cache(struct protection_domain *domain,
1466 dma_addr_t iova, size_t size)
1468 if (unlikely(amd_iommu_np_cache)) {
1469 unsigned long flags;
1471 spin_lock_irqsave(&domain->lock, flags);
1472 domain_flush_pages(domain, iova, size, 1);
1473 amd_iommu_domain_flush_complete(domain);
1474 spin_unlock_irqrestore(&domain->lock, flags);
1480 * This function flushes the DTEs for all devices in domain
1482 static void domain_flush_devices(struct protection_domain *domain)
1484 struct iommu_dev_data *dev_data;
1486 list_for_each_entry(dev_data, &domain->dev_list, list)
1487 device_flush_dte(dev_data);
1490 /****************************************************************************
1492 * The next functions belong to the domain allocation. A domain is
1493 * allocated for every IOMMU as the default domain. If device isolation
1494 * is enabled, every device get its own domain. The most important thing
1495 * about domains is the page table mapping the DMA address space they
1498 ****************************************************************************/
1500 static u16 domain_id_alloc(void)
1504 spin_lock(&pd_bitmap_lock);
1505 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1507 if (id > 0 && id < MAX_DOMAIN_ID)
1508 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1511 spin_unlock(&pd_bitmap_lock);
1516 static void domain_id_free(int id)
1518 spin_lock(&pd_bitmap_lock);
1519 if (id > 0 && id < MAX_DOMAIN_ID)
1520 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1521 spin_unlock(&pd_bitmap_lock);
1524 static void free_gcr3_tbl_level1(u64 *tbl)
1529 for (i = 0; i < 512; ++i) {
1530 if (!(tbl[i] & GCR3_VALID))
1533 ptr = iommu_phys_to_virt(tbl[i] & PAGE_MASK);
1535 free_page((unsigned long)ptr);
1539 static void free_gcr3_tbl_level2(u64 *tbl)
1544 for (i = 0; i < 512; ++i) {
1545 if (!(tbl[i] & GCR3_VALID))
1548 ptr = iommu_phys_to_virt(tbl[i] & PAGE_MASK);
1550 free_gcr3_tbl_level1(ptr);
1554 static void free_gcr3_table(struct protection_domain *domain)
1556 if (domain->glx == 2)
1557 free_gcr3_tbl_level2(domain->gcr3_tbl);
1558 else if (domain->glx == 1)
1559 free_gcr3_tbl_level1(domain->gcr3_tbl);
1561 BUG_ON(domain->glx != 0);
1563 free_page((unsigned long)domain->gcr3_tbl);
1566 static void set_dte_entry(struct amd_iommu *iommu, u16 devid,
1567 struct protection_domain *domain, bool ats, bool ppr)
1572 struct dev_table_entry *dev_table = get_dev_table(iommu);
1574 if (domain->iop.mode != PAGE_MODE_NONE)
1575 pte_root = iommu_virt_to_phys(domain->iop.root);
1577 pte_root |= (domain->iop.mode & DEV_ENTRY_MODE_MASK)
1578 << DEV_ENTRY_MODE_SHIFT;
1580 pte_root |= DTE_FLAG_IR | DTE_FLAG_IW | DTE_FLAG_V;
1583 * When SNP is enabled, Only set TV bit when IOMMU
1584 * page translation is in use.
1586 if (!amd_iommu_snp_en || (domain->id != 0))
1587 pte_root |= DTE_FLAG_TV;
1589 flags = dev_table[devid].data[1];
1592 flags |= DTE_FLAG_IOTLB;
1595 if (iommu_feature(iommu, FEATURE_EPHSUP))
1596 pte_root |= 1ULL << DEV_ENTRY_PPR;
1599 if (domain->flags & PD_IOMMUV2_MASK) {
1600 u64 gcr3 = iommu_virt_to_phys(domain->gcr3_tbl);
1601 u64 glx = domain->glx;
1604 pte_root |= DTE_FLAG_GV;
1605 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
1607 /* First mask out possible old values for GCR3 table */
1608 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
1611 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
1614 /* Encode GCR3 table into DTE */
1615 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
1618 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
1621 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
1624 if (amd_iommu_gpt_level == PAGE_MODE_5_LEVEL) {
1625 dev_table[devid].data[2] |=
1626 ((u64)GUEST_PGTABLE_5_LEVEL << DTE_GPT_LEVEL_SHIFT);
1629 if (domain->flags & PD_GIOV_MASK)
1630 pte_root |= DTE_FLAG_GIOV;
1633 flags &= ~DEV_DOMID_MASK;
1634 flags |= domain->id;
1636 old_domid = dev_table[devid].data[1] & DEV_DOMID_MASK;
1637 dev_table[devid].data[1] = flags;
1638 dev_table[devid].data[0] = pte_root;
1641 * A kdump kernel might be replacing a domain ID that was copied from
1642 * the previous kernel--if so, it needs to flush the translation cache
1643 * entries for the old domain ID that is being overwritten
1646 amd_iommu_flush_tlb_domid(iommu, old_domid);
1650 static void clear_dte_entry(struct amd_iommu *iommu, u16 devid)
1652 struct dev_table_entry *dev_table = get_dev_table(iommu);
1654 /* remove entry from the device table seen by the hardware */
1655 dev_table[devid].data[0] = DTE_FLAG_V;
1657 if (!amd_iommu_snp_en)
1658 dev_table[devid].data[0] |= DTE_FLAG_TV;
1660 dev_table[devid].data[1] &= DTE_FLAG_MASK;
1662 amd_iommu_apply_erratum_63(iommu, devid);
1665 static void do_attach(struct iommu_dev_data *dev_data,
1666 struct protection_domain *domain)
1668 struct amd_iommu *iommu;
1671 iommu = rlookup_amd_iommu(dev_data->dev);
1674 ats = dev_data->ats.enabled;
1676 /* Update data structures */
1677 dev_data->domain = domain;
1678 list_add(&dev_data->list, &domain->dev_list);
1680 /* Update NUMA Node ID */
1681 if (domain->nid == NUMA_NO_NODE)
1682 domain->nid = dev_to_node(dev_data->dev);
1684 /* Do reference counting */
1685 domain->dev_iommu[iommu->index] += 1;
1686 domain->dev_cnt += 1;
1688 /* Update device table */
1689 set_dte_entry(iommu, dev_data->devid, domain,
1690 ats, dev_data->iommu_v2);
1691 clone_aliases(iommu, dev_data->dev);
1693 device_flush_dte(dev_data);
1696 static void do_detach(struct iommu_dev_data *dev_data)
1698 struct protection_domain *domain = dev_data->domain;
1699 struct amd_iommu *iommu;
1701 iommu = rlookup_amd_iommu(dev_data->dev);
1705 /* Update data structures */
1706 dev_data->domain = NULL;
1707 list_del(&dev_data->list);
1708 clear_dte_entry(iommu, dev_data->devid);
1709 clone_aliases(iommu, dev_data->dev);
1711 /* Flush the DTE entry */
1712 device_flush_dte(dev_data);
1715 amd_iommu_domain_flush_tlb_pde(domain);
1717 /* Wait for the flushes to finish */
1718 amd_iommu_domain_flush_complete(domain);
1720 /* decrease reference counters - needs to happen after the flushes */
1721 domain->dev_iommu[iommu->index] -= 1;
1722 domain->dev_cnt -= 1;
1725 static void pdev_iommuv2_disable(struct pci_dev *pdev)
1727 pci_disable_ats(pdev);
1728 pci_disable_pri(pdev);
1729 pci_disable_pasid(pdev);
1732 static int pdev_pri_ats_enable(struct pci_dev *pdev)
1736 /* Only allow access to user-accessible pages */
1737 ret = pci_enable_pasid(pdev, 0);
1741 /* First reset the PRI state of the device */
1742 ret = pci_reset_pri(pdev);
1747 /* FIXME: Hardcode number of outstanding requests for now */
1748 ret = pci_enable_pri(pdev, 32);
1752 ret = pci_enable_ats(pdev, PAGE_SHIFT);
1759 pci_disable_pri(pdev);
1762 pci_disable_pasid(pdev);
1768 * If a device is not yet associated with a domain, this function makes the
1769 * device visible in the domain
1771 static int attach_device(struct device *dev,
1772 struct protection_domain *domain)
1774 struct iommu_dev_data *dev_data;
1775 struct pci_dev *pdev;
1776 unsigned long flags;
1779 spin_lock_irqsave(&domain->lock, flags);
1781 dev_data = dev_iommu_priv_get(dev);
1783 spin_lock(&dev_data->lock);
1786 if (dev_data->domain != NULL)
1789 if (!dev_is_pci(dev))
1790 goto skip_ats_check;
1792 pdev = to_pci_dev(dev);
1793 if (domain->flags & PD_IOMMUV2_MASK) {
1794 struct iommu_domain *def_domain = iommu_get_dma_domain(dev);
1799 * In case of using AMD_IOMMU_V1 page table mode and the device
1800 * is enabling for PPR/ATS support (using v2 table),
1801 * we need to make sure that the domain type is identity map.
1803 if ((amd_iommu_pgtable == AMD_IOMMU_V1) &&
1804 def_domain->type != IOMMU_DOMAIN_IDENTITY) {
1808 if (dev_data->iommu_v2) {
1809 if (pdev_pri_ats_enable(pdev) != 0)
1812 dev_data->ats.enabled = true;
1813 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
1814 dev_data->pri_tlp = pci_prg_resp_pasid_required(pdev);
1816 } else if (amd_iommu_iotlb_sup &&
1817 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
1818 dev_data->ats.enabled = true;
1819 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
1825 do_attach(dev_data, domain);
1828 * We might boot into a crash-kernel here. The crashed kernel
1829 * left the caches in the IOMMU dirty. So we have to flush
1830 * here to evict all dirty stuff.
1832 amd_iommu_domain_flush_tlb_pde(domain);
1834 amd_iommu_domain_flush_complete(domain);
1837 spin_unlock(&dev_data->lock);
1839 spin_unlock_irqrestore(&domain->lock, flags);
1845 * Removes a device from a protection domain (with devtable_lock held)
1847 static void detach_device(struct device *dev)
1849 struct protection_domain *domain;
1850 struct iommu_dev_data *dev_data;
1851 unsigned long flags;
1853 dev_data = dev_iommu_priv_get(dev);
1854 domain = dev_data->domain;
1856 spin_lock_irqsave(&domain->lock, flags);
1858 spin_lock(&dev_data->lock);
1861 * First check if the device is still attached. It might already
1862 * be detached from its domain because the generic
1863 * iommu_detach_group code detached it and we try again here in
1864 * our alias handling.
1866 if (WARN_ON(!dev_data->domain))
1869 do_detach(dev_data);
1871 if (!dev_is_pci(dev))
1874 if (domain->flags & PD_IOMMUV2_MASK && dev_data->iommu_v2)
1875 pdev_iommuv2_disable(to_pci_dev(dev));
1876 else if (dev_data->ats.enabled)
1877 pci_disable_ats(to_pci_dev(dev));
1879 dev_data->ats.enabled = false;
1882 spin_unlock(&dev_data->lock);
1884 spin_unlock_irqrestore(&domain->lock, flags);
1887 static struct iommu_device *amd_iommu_probe_device(struct device *dev)
1889 struct iommu_device *iommu_dev;
1890 struct amd_iommu *iommu;
1893 if (!check_device(dev))
1894 return ERR_PTR(-ENODEV);
1896 iommu = rlookup_amd_iommu(dev);
1898 return ERR_PTR(-ENODEV);
1900 /* Not registered yet? */
1901 if (!iommu->iommu.ops)
1902 return ERR_PTR(-ENODEV);
1904 if (dev_iommu_priv_get(dev))
1905 return &iommu->iommu;
1907 ret = iommu_init_device(iommu, dev);
1909 if (ret != -ENOTSUPP)
1910 dev_err(dev, "Failed to initialize - trying to proceed anyway\n");
1911 iommu_dev = ERR_PTR(ret);
1912 iommu_ignore_device(iommu, dev);
1914 amd_iommu_set_pci_msi_domain(dev, iommu);
1915 iommu_dev = &iommu->iommu;
1918 iommu_completion_wait(iommu);
1923 static void amd_iommu_probe_finalize(struct device *dev)
1925 /* Domains are initialized for this device - have a look what we ended up with */
1926 set_dma_ops(dev, NULL);
1927 iommu_setup_dma_ops(dev, 0, U64_MAX);
1930 static void amd_iommu_release_device(struct device *dev)
1932 struct amd_iommu *iommu;
1934 if (!check_device(dev))
1937 iommu = rlookup_amd_iommu(dev);
1941 amd_iommu_uninit_device(dev);
1942 iommu_completion_wait(iommu);
1945 static struct iommu_group *amd_iommu_device_group(struct device *dev)
1947 if (dev_is_pci(dev))
1948 return pci_device_group(dev);
1950 return acpihid_device_group(dev);
1953 /*****************************************************************************
1955 * The next functions belong to the dma_ops mapping/unmapping code.
1957 *****************************************************************************/
1959 static void update_device_table(struct protection_domain *domain)
1961 struct iommu_dev_data *dev_data;
1963 list_for_each_entry(dev_data, &domain->dev_list, list) {
1964 struct amd_iommu *iommu = rlookup_amd_iommu(dev_data->dev);
1968 set_dte_entry(iommu, dev_data->devid, domain,
1969 dev_data->ats.enabled, dev_data->iommu_v2);
1970 clone_aliases(iommu, dev_data->dev);
1974 void amd_iommu_update_and_flush_device_table(struct protection_domain *domain)
1976 update_device_table(domain);
1977 domain_flush_devices(domain);
1980 void amd_iommu_domain_update(struct protection_domain *domain)
1982 /* Update device table */
1983 amd_iommu_update_and_flush_device_table(domain);
1985 /* Flush domain TLB(s) and wait for completion */
1986 amd_iommu_domain_flush_tlb_pde(domain);
1987 amd_iommu_domain_flush_complete(domain);
1990 /*****************************************************************************
1992 * The following functions belong to the exported interface of AMD IOMMU
1994 * This interface allows access to lower level functions of the IOMMU
1995 * like protection domain handling and assignement of devices to domains
1996 * which is not possible with the dma_ops interface.
1998 *****************************************************************************/
2000 static void cleanup_domain(struct protection_domain *domain)
2002 struct iommu_dev_data *entry;
2003 unsigned long flags;
2005 spin_lock_irqsave(&domain->lock, flags);
2007 while (!list_empty(&domain->dev_list)) {
2008 entry = list_first_entry(&domain->dev_list,
2009 struct iommu_dev_data, list);
2010 BUG_ON(!entry->domain);
2014 spin_unlock_irqrestore(&domain->lock, flags);
2017 static void protection_domain_free(struct protection_domain *domain)
2022 if (domain->iop.pgtbl_cfg.tlb)
2023 free_io_pgtable_ops(&domain->iop.iop.ops);
2026 domain_id_free(domain->id);
2031 static int protection_domain_init_v1(struct protection_domain *domain, int mode)
2033 u64 *pt_root = NULL;
2035 BUG_ON(mode < PAGE_MODE_NONE || mode > PAGE_MODE_6_LEVEL);
2037 spin_lock_init(&domain->lock);
2038 domain->id = domain_id_alloc();
2041 INIT_LIST_HEAD(&domain->dev_list);
2043 if (mode != PAGE_MODE_NONE) {
2044 pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2046 domain_id_free(domain->id);
2051 amd_iommu_domain_set_pgtable(domain, pt_root, mode);
2056 static int protection_domain_init_v2(struct protection_domain *domain)
2058 spin_lock_init(&domain->lock);
2059 domain->id = domain_id_alloc();
2062 INIT_LIST_HEAD(&domain->dev_list);
2064 domain->flags |= PD_GIOV_MASK;
2066 domain->domain.pgsize_bitmap = AMD_IOMMU_PGSIZES_V2;
2068 if (domain_enable_v2(domain, 1)) {
2069 domain_id_free(domain->id);
2076 static struct protection_domain *protection_domain_alloc(unsigned int type)
2078 struct io_pgtable_ops *pgtbl_ops;
2079 struct protection_domain *domain;
2081 int mode = DEFAULT_PGTABLE_LEVEL;
2085 * Force IOMMU v1 page table when iommu=pt and
2086 * when allocating domain for pass-through devices.
2088 if (type == IOMMU_DOMAIN_IDENTITY) {
2089 pgtable = AMD_IOMMU_V1;
2090 mode = PAGE_MODE_NONE;
2091 } else if (type == IOMMU_DOMAIN_UNMANAGED) {
2092 pgtable = AMD_IOMMU_V1;
2093 } else if (type == IOMMU_DOMAIN_DMA || type == IOMMU_DOMAIN_DMA_FQ) {
2094 pgtable = amd_iommu_pgtable;
2099 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2105 ret = protection_domain_init_v1(domain, mode);
2108 ret = protection_domain_init_v2(domain);
2117 /* No need to allocate io pgtable ops in passthrough mode */
2118 if (type == IOMMU_DOMAIN_IDENTITY)
2121 domain->nid = NUMA_NO_NODE;
2123 pgtbl_ops = alloc_io_pgtable_ops(pgtable, &domain->iop.pgtbl_cfg, domain);
2125 domain_id_free(domain->id);
2135 static inline u64 dma_max_address(void)
2137 if (amd_iommu_pgtable == AMD_IOMMU_V1)
2140 /* V2 with 4/5 level page table */
2141 return ((1ULL << PM_LEVEL_SHIFT(amd_iommu_gpt_level)) - 1);
2144 static struct iommu_domain *amd_iommu_domain_alloc(unsigned type)
2146 struct protection_domain *domain;
2149 * Since DTE[Mode]=0 is prohibited on SNP-enabled system,
2150 * default to use IOMMU_DOMAIN_DMA[_FQ].
2152 if (amd_iommu_snp_en && (type == IOMMU_DOMAIN_IDENTITY))
2155 domain = protection_domain_alloc(type);
2159 domain->domain.geometry.aperture_start = 0;
2160 domain->domain.geometry.aperture_end = dma_max_address();
2161 domain->domain.geometry.force_aperture = true;
2163 return &domain->domain;
2166 static void amd_iommu_domain_free(struct iommu_domain *dom)
2168 struct protection_domain *domain;
2170 domain = to_pdomain(dom);
2172 if (domain->dev_cnt > 0)
2173 cleanup_domain(domain);
2175 BUG_ON(domain->dev_cnt != 0);
2180 if (domain->flags & PD_IOMMUV2_MASK)
2181 free_gcr3_table(domain);
2183 protection_domain_free(domain);
2186 static int amd_iommu_attach_device(struct iommu_domain *dom,
2189 struct iommu_dev_data *dev_data = dev_iommu_priv_get(dev);
2190 struct protection_domain *domain = to_pdomain(dom);
2191 struct amd_iommu *iommu = rlookup_amd_iommu(dev);
2195 * Skip attach device to domain if new domain is same as
2196 * devices current domain
2198 if (dev_data->domain == domain)
2201 dev_data->defer_attach = false;
2203 if (dev_data->domain)
2206 ret = attach_device(dev, domain);
2208 #ifdef CONFIG_IRQ_REMAP
2209 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
2210 if (dom->type == IOMMU_DOMAIN_UNMANAGED)
2211 dev_data->use_vapic = 1;
2213 dev_data->use_vapic = 0;
2217 iommu_completion_wait(iommu);
2222 static void amd_iommu_iotlb_sync_map(struct iommu_domain *dom,
2223 unsigned long iova, size_t size)
2225 struct protection_domain *domain = to_pdomain(dom);
2226 struct io_pgtable_ops *ops = &domain->iop.iop.ops;
2229 domain_flush_np_cache(domain, iova, size);
2232 static int amd_iommu_map_pages(struct iommu_domain *dom, unsigned long iova,
2233 phys_addr_t paddr, size_t pgsize, size_t pgcount,
2234 int iommu_prot, gfp_t gfp, size_t *mapped)
2236 struct protection_domain *domain = to_pdomain(dom);
2237 struct io_pgtable_ops *ops = &domain->iop.iop.ops;
2241 if ((amd_iommu_pgtable == AMD_IOMMU_V1) &&
2242 (domain->iop.mode == PAGE_MODE_NONE))
2245 if (iommu_prot & IOMMU_READ)
2246 prot |= IOMMU_PROT_IR;
2247 if (iommu_prot & IOMMU_WRITE)
2248 prot |= IOMMU_PROT_IW;
2250 if (ops->map_pages) {
2251 ret = ops->map_pages(ops, iova, paddr, pgsize,
2252 pgcount, prot, gfp, mapped);
2258 static void amd_iommu_iotlb_gather_add_page(struct iommu_domain *domain,
2259 struct iommu_iotlb_gather *gather,
2260 unsigned long iova, size_t size)
2263 * AMD's IOMMU can flush as many pages as necessary in a single flush.
2264 * Unless we run in a virtual machine, which can be inferred according
2265 * to whether "non-present cache" is on, it is probably best to prefer
2266 * (potentially) too extensive TLB flushing (i.e., more misses) over
2267 * mutliple TLB flushes (i.e., more flushes). For virtual machines the
2268 * hypervisor needs to synchronize the host IOMMU PTEs with those of
2269 * the guest, and the trade-off is different: unnecessary TLB flushes
2270 * should be avoided.
2272 if (amd_iommu_np_cache &&
2273 iommu_iotlb_gather_is_disjoint(gather, iova, size))
2274 iommu_iotlb_sync(domain, gather);
2276 iommu_iotlb_gather_add_range(gather, iova, size);
2279 static size_t amd_iommu_unmap_pages(struct iommu_domain *dom, unsigned long iova,
2280 size_t pgsize, size_t pgcount,
2281 struct iommu_iotlb_gather *gather)
2283 struct protection_domain *domain = to_pdomain(dom);
2284 struct io_pgtable_ops *ops = &domain->iop.iop.ops;
2287 if ((amd_iommu_pgtable == AMD_IOMMU_V1) &&
2288 (domain->iop.mode == PAGE_MODE_NONE))
2291 r = (ops->unmap_pages) ? ops->unmap_pages(ops, iova, pgsize, pgcount, NULL) : 0;
2294 amd_iommu_iotlb_gather_add_page(dom, gather, iova, r);
2299 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
2302 struct protection_domain *domain = to_pdomain(dom);
2303 struct io_pgtable_ops *ops = &domain->iop.iop.ops;
2305 return ops->iova_to_phys(ops, iova);
2308 static bool amd_iommu_capable(struct device *dev, enum iommu_cap cap)
2311 case IOMMU_CAP_CACHE_COHERENCY:
2313 case IOMMU_CAP_NOEXEC:
2315 case IOMMU_CAP_PRE_BOOT_PROTECTION:
2316 return amdr_ivrs_remap_support;
2317 case IOMMU_CAP_ENFORCE_CACHE_COHERENCY:
2319 case IOMMU_CAP_DEFERRED_FLUSH:
2328 static void amd_iommu_get_resv_regions(struct device *dev,
2329 struct list_head *head)
2331 struct iommu_resv_region *region;
2332 struct unity_map_entry *entry;
2333 struct amd_iommu *iommu;
2334 struct amd_iommu_pci_seg *pci_seg;
2337 sbdf = get_device_sbdf_id(dev);
2341 devid = PCI_SBDF_TO_DEVID(sbdf);
2342 iommu = rlookup_amd_iommu(dev);
2345 pci_seg = iommu->pci_seg;
2347 list_for_each_entry(entry, &pci_seg->unity_map, list) {
2351 if (devid < entry->devid_start || devid > entry->devid_end)
2354 type = IOMMU_RESV_DIRECT;
2355 length = entry->address_end - entry->address_start;
2356 if (entry->prot & IOMMU_PROT_IR)
2358 if (entry->prot & IOMMU_PROT_IW)
2359 prot |= IOMMU_WRITE;
2360 if (entry->prot & IOMMU_UNITY_MAP_FLAG_EXCL_RANGE)
2361 /* Exclusion range */
2362 type = IOMMU_RESV_RESERVED;
2364 region = iommu_alloc_resv_region(entry->address_start,
2368 dev_err(dev, "Out of memory allocating dm-regions\n");
2371 list_add_tail(®ion->list, head);
2374 region = iommu_alloc_resv_region(MSI_RANGE_START,
2375 MSI_RANGE_END - MSI_RANGE_START + 1,
2376 0, IOMMU_RESV_MSI, GFP_KERNEL);
2379 list_add_tail(®ion->list, head);
2381 region = iommu_alloc_resv_region(HT_RANGE_START,
2382 HT_RANGE_END - HT_RANGE_START + 1,
2383 0, IOMMU_RESV_RESERVED, GFP_KERNEL);
2386 list_add_tail(®ion->list, head);
2389 bool amd_iommu_is_attach_deferred(struct device *dev)
2391 struct iommu_dev_data *dev_data = dev_iommu_priv_get(dev);
2393 return dev_data->defer_attach;
2395 EXPORT_SYMBOL_GPL(amd_iommu_is_attach_deferred);
2397 static void amd_iommu_flush_iotlb_all(struct iommu_domain *domain)
2399 struct protection_domain *dom = to_pdomain(domain);
2400 unsigned long flags;
2402 spin_lock_irqsave(&dom->lock, flags);
2403 amd_iommu_domain_flush_tlb_pde(dom);
2404 amd_iommu_domain_flush_complete(dom);
2405 spin_unlock_irqrestore(&dom->lock, flags);
2408 static void amd_iommu_iotlb_sync(struct iommu_domain *domain,
2409 struct iommu_iotlb_gather *gather)
2411 struct protection_domain *dom = to_pdomain(domain);
2412 unsigned long flags;
2414 spin_lock_irqsave(&dom->lock, flags);
2415 domain_flush_pages(dom, gather->start, gather->end - gather->start + 1, 1);
2416 amd_iommu_domain_flush_complete(dom);
2417 spin_unlock_irqrestore(&dom->lock, flags);
2420 static int amd_iommu_def_domain_type(struct device *dev)
2422 struct iommu_dev_data *dev_data;
2424 dev_data = dev_iommu_priv_get(dev);
2429 * Do not identity map IOMMUv2 capable devices when:
2430 * - memory encryption is active, because some of those devices
2431 * (AMD GPUs) don't have the encryption bit in their DMA-mask
2432 * and require remapping.
2433 * - SNP is enabled, because it prohibits DTE[Mode]=0.
2435 if (dev_data->iommu_v2 &&
2436 !cc_platform_has(CC_ATTR_MEM_ENCRYPT) &&
2437 !amd_iommu_snp_en) {
2438 return IOMMU_DOMAIN_IDENTITY;
2444 static bool amd_iommu_enforce_cache_coherency(struct iommu_domain *domain)
2446 /* IOMMU_PTE_FC is always set */
2450 const struct iommu_ops amd_iommu_ops = {
2451 .capable = amd_iommu_capable,
2452 .domain_alloc = amd_iommu_domain_alloc,
2453 .probe_device = amd_iommu_probe_device,
2454 .release_device = amd_iommu_release_device,
2455 .probe_finalize = amd_iommu_probe_finalize,
2456 .device_group = amd_iommu_device_group,
2457 .get_resv_regions = amd_iommu_get_resv_regions,
2458 .is_attach_deferred = amd_iommu_is_attach_deferred,
2459 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
2460 .def_domain_type = amd_iommu_def_domain_type,
2461 .default_domain_ops = &(const struct iommu_domain_ops) {
2462 .attach_dev = amd_iommu_attach_device,
2463 .map_pages = amd_iommu_map_pages,
2464 .unmap_pages = amd_iommu_unmap_pages,
2465 .iotlb_sync_map = amd_iommu_iotlb_sync_map,
2466 .iova_to_phys = amd_iommu_iova_to_phys,
2467 .flush_iotlb_all = amd_iommu_flush_iotlb_all,
2468 .iotlb_sync = amd_iommu_iotlb_sync,
2469 .free = amd_iommu_domain_free,
2470 .enforce_cache_coherency = amd_iommu_enforce_cache_coherency,
2474 /*****************************************************************************
2476 * The next functions do a basic initialization of IOMMU for pass through
2479 * In passthrough mode the IOMMU is initialized and enabled but not used for
2480 * DMA-API translation.
2482 *****************************************************************************/
2484 /* IOMMUv2 specific functions */
2485 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
2487 return atomic_notifier_chain_register(&ppr_notifier, nb);
2489 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
2491 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
2493 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
2495 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
2497 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
2499 struct protection_domain *domain = to_pdomain(dom);
2500 unsigned long flags;
2502 spin_lock_irqsave(&domain->lock, flags);
2504 if (domain->iop.pgtbl_cfg.tlb)
2505 free_io_pgtable_ops(&domain->iop.iop.ops);
2507 spin_unlock_irqrestore(&domain->lock, flags);
2509 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
2511 /* Note: This function expects iommu_domain->lock to be held prior calling the function. */
2512 static int domain_enable_v2(struct protection_domain *domain, int pasids)
2516 /* Number of GCR3 table levels required */
2517 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
2520 if (levels > amd_iommu_max_glx_val)
2523 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
2524 if (domain->gcr3_tbl == NULL)
2527 domain->glx = levels;
2528 domain->flags |= PD_IOMMUV2_MASK;
2530 amd_iommu_domain_update(domain);
2535 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
2537 struct protection_domain *pdom = to_pdomain(dom);
2538 unsigned long flags;
2541 spin_lock_irqsave(&pdom->lock, flags);
2544 * Save us all sanity checks whether devices already in the
2545 * domain support IOMMUv2. Just force that the domain has no
2546 * devices attached when it is switched into IOMMUv2 mode.
2549 if (pdom->dev_cnt > 0 || pdom->flags & PD_IOMMUV2_MASK)
2552 if (!pdom->gcr3_tbl)
2553 ret = domain_enable_v2(pdom, pasids);
2556 spin_unlock_irqrestore(&pdom->lock, flags);
2559 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
2561 static int __flush_pasid(struct protection_domain *domain, u32 pasid,
2562 u64 address, bool size)
2564 struct iommu_dev_data *dev_data;
2565 struct iommu_cmd cmd;
2568 if (!(domain->flags & PD_IOMMUV2_MASK))
2571 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
2574 * IOMMU TLB needs to be flushed before Device TLB to
2575 * prevent device TLB refill from IOMMU TLB
2577 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
2578 if (domain->dev_iommu[i] == 0)
2581 ret = iommu_queue_command(amd_iommus[i], &cmd);
2586 /* Wait until IOMMU TLB flushes are complete */
2587 amd_iommu_domain_flush_complete(domain);
2589 /* Now flush device TLBs */
2590 list_for_each_entry(dev_data, &domain->dev_list, list) {
2591 struct amd_iommu *iommu;
2595 There might be non-IOMMUv2 capable devices in an IOMMUv2
2598 if (!dev_data->ats.enabled)
2601 qdep = dev_data->ats.qdep;
2602 iommu = rlookup_amd_iommu(dev_data->dev);
2605 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
2606 qdep, address, size);
2608 ret = iommu_queue_command(iommu, &cmd);
2613 /* Wait until all device TLBs are flushed */
2614 amd_iommu_domain_flush_complete(domain);
2623 static int __amd_iommu_flush_page(struct protection_domain *domain, u32 pasid,
2626 return __flush_pasid(domain, pasid, address, false);
2629 int amd_iommu_flush_page(struct iommu_domain *dom, u32 pasid,
2632 struct protection_domain *domain = to_pdomain(dom);
2633 unsigned long flags;
2636 spin_lock_irqsave(&domain->lock, flags);
2637 ret = __amd_iommu_flush_page(domain, pasid, address);
2638 spin_unlock_irqrestore(&domain->lock, flags);
2642 EXPORT_SYMBOL(amd_iommu_flush_page);
2644 static int __amd_iommu_flush_tlb(struct protection_domain *domain, u32 pasid)
2646 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
2650 int amd_iommu_flush_tlb(struct iommu_domain *dom, u32 pasid)
2652 struct protection_domain *domain = to_pdomain(dom);
2653 unsigned long flags;
2656 spin_lock_irqsave(&domain->lock, flags);
2657 ret = __amd_iommu_flush_tlb(domain, pasid);
2658 spin_unlock_irqrestore(&domain->lock, flags);
2662 EXPORT_SYMBOL(amd_iommu_flush_tlb);
2664 static u64 *__get_gcr3_pte(u64 *root, int level, u32 pasid, bool alloc)
2671 index = (pasid >> (9 * level)) & 0x1ff;
2677 if (!(*pte & GCR3_VALID)) {
2681 root = (void *)get_zeroed_page(GFP_ATOMIC);
2685 *pte = iommu_virt_to_phys(root) | GCR3_VALID;
2688 root = iommu_phys_to_virt(*pte & PAGE_MASK);
2696 static int __set_gcr3(struct protection_domain *domain, u32 pasid,
2701 if (domain->iop.mode != PAGE_MODE_NONE)
2704 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
2708 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
2710 return __amd_iommu_flush_tlb(domain, pasid);
2713 static int __clear_gcr3(struct protection_domain *domain, u32 pasid)
2717 if (domain->iop.mode != PAGE_MODE_NONE)
2720 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
2726 return __amd_iommu_flush_tlb(domain, pasid);
2729 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, u32 pasid,
2732 struct protection_domain *domain = to_pdomain(dom);
2733 unsigned long flags;
2736 spin_lock_irqsave(&domain->lock, flags);
2737 ret = __set_gcr3(domain, pasid, cr3);
2738 spin_unlock_irqrestore(&domain->lock, flags);
2742 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
2744 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, u32 pasid)
2746 struct protection_domain *domain = to_pdomain(dom);
2747 unsigned long flags;
2750 spin_lock_irqsave(&domain->lock, flags);
2751 ret = __clear_gcr3(domain, pasid);
2752 spin_unlock_irqrestore(&domain->lock, flags);
2756 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
2758 int amd_iommu_complete_ppr(struct pci_dev *pdev, u32 pasid,
2759 int status, int tag)
2761 struct iommu_dev_data *dev_data;
2762 struct amd_iommu *iommu;
2763 struct iommu_cmd cmd;
2765 dev_data = dev_iommu_priv_get(&pdev->dev);
2766 iommu = rlookup_amd_iommu(&pdev->dev);
2770 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
2771 tag, dev_data->pri_tlp);
2773 return iommu_queue_command(iommu, &cmd);
2775 EXPORT_SYMBOL(amd_iommu_complete_ppr);
2777 int amd_iommu_device_info(struct pci_dev *pdev,
2778 struct amd_iommu_device_info *info)
2783 if (pdev == NULL || info == NULL)
2786 if (!amd_iommu_v2_supported())
2789 memset(info, 0, sizeof(*info));
2791 if (pci_ats_supported(pdev))
2792 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
2794 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2796 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
2798 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
2802 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
2803 max_pasids = min(max_pasids, (1 << 20));
2805 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
2806 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
2808 features = pci_pasid_features(pdev);
2809 if (features & PCI_PASID_CAP_EXEC)
2810 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
2811 if (features & PCI_PASID_CAP_PRIV)
2812 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
2817 EXPORT_SYMBOL(amd_iommu_device_info);
2819 #ifdef CONFIG_IRQ_REMAP
2821 /*****************************************************************************
2823 * Interrupt Remapping Implementation
2825 *****************************************************************************/
2827 static struct irq_chip amd_ir_chip;
2828 static DEFINE_SPINLOCK(iommu_table_lock);
2830 static void iommu_flush_irt_and_complete(struct amd_iommu *iommu, u16 devid)
2834 unsigned long flags;
2835 struct iommu_cmd cmd, cmd2;
2837 if (iommu->irtcachedis_enabled)
2840 build_inv_irt(&cmd, devid);
2841 data = atomic64_add_return(1, &iommu->cmd_sem_val);
2842 build_completion_wait(&cmd2, iommu, data);
2844 raw_spin_lock_irqsave(&iommu->lock, flags);
2845 ret = __iommu_queue_command_sync(iommu, &cmd, true);
2848 ret = __iommu_queue_command_sync(iommu, &cmd2, false);
2851 wait_on_sem(iommu, data);
2853 raw_spin_unlock_irqrestore(&iommu->lock, flags);
2856 static void set_dte_irq_entry(struct amd_iommu *iommu, u16 devid,
2857 struct irq_remap_table *table)
2860 struct dev_table_entry *dev_table = get_dev_table(iommu);
2862 dte = dev_table[devid].data[2];
2863 dte &= ~DTE_IRQ_PHYS_ADDR_MASK;
2864 dte |= iommu_virt_to_phys(table->table);
2865 dte |= DTE_IRQ_REMAP_INTCTL;
2866 dte |= DTE_INTTABLEN;
2867 dte |= DTE_IRQ_REMAP_ENABLE;
2869 dev_table[devid].data[2] = dte;
2872 static struct irq_remap_table *get_irq_table(struct amd_iommu *iommu, u16 devid)
2874 struct irq_remap_table *table;
2875 struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
2877 if (WARN_ONCE(!pci_seg->rlookup_table[devid],
2878 "%s: no iommu for devid %x:%x\n",
2879 __func__, pci_seg->id, devid))
2882 table = pci_seg->irq_lookup_table[devid];
2883 if (WARN_ONCE(!table, "%s: no table for devid %x:%x\n",
2884 __func__, pci_seg->id, devid))
2890 static struct irq_remap_table *__alloc_irq_table(void)
2892 struct irq_remap_table *table;
2894 table = kzalloc(sizeof(*table), GFP_KERNEL);
2898 table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_KERNEL);
2899 if (!table->table) {
2903 raw_spin_lock_init(&table->lock);
2905 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
2906 memset(table->table, 0,
2907 MAX_IRQS_PER_TABLE * sizeof(u32));
2909 memset(table->table, 0,
2910 (MAX_IRQS_PER_TABLE * (sizeof(u64) * 2)));
2914 static void set_remap_table_entry(struct amd_iommu *iommu, u16 devid,
2915 struct irq_remap_table *table)
2917 struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
2919 pci_seg->irq_lookup_table[devid] = table;
2920 set_dte_irq_entry(iommu, devid, table);
2921 iommu_flush_dte(iommu, devid);
2924 static int set_remap_table_entry_alias(struct pci_dev *pdev, u16 alias,
2927 struct irq_remap_table *table = data;
2928 struct amd_iommu_pci_seg *pci_seg;
2929 struct amd_iommu *iommu = rlookup_amd_iommu(&pdev->dev);
2934 pci_seg = iommu->pci_seg;
2935 pci_seg->irq_lookup_table[alias] = table;
2936 set_dte_irq_entry(iommu, alias, table);
2937 iommu_flush_dte(pci_seg->rlookup_table[alias], alias);
2942 static struct irq_remap_table *alloc_irq_table(struct amd_iommu *iommu,
2943 u16 devid, struct pci_dev *pdev)
2945 struct irq_remap_table *table = NULL;
2946 struct irq_remap_table *new_table = NULL;
2947 struct amd_iommu_pci_seg *pci_seg;
2948 unsigned long flags;
2951 spin_lock_irqsave(&iommu_table_lock, flags);
2953 pci_seg = iommu->pci_seg;
2954 table = pci_seg->irq_lookup_table[devid];
2958 alias = pci_seg->alias_table[devid];
2959 table = pci_seg->irq_lookup_table[alias];
2961 set_remap_table_entry(iommu, devid, table);
2964 spin_unlock_irqrestore(&iommu_table_lock, flags);
2966 /* Nothing there yet, allocate new irq remapping table */
2967 new_table = __alloc_irq_table();
2971 spin_lock_irqsave(&iommu_table_lock, flags);
2973 table = pci_seg->irq_lookup_table[devid];
2977 table = pci_seg->irq_lookup_table[alias];
2979 set_remap_table_entry(iommu, devid, table);
2987 pci_for_each_dma_alias(pdev, set_remap_table_entry_alias,
2990 set_remap_table_entry(iommu, devid, table);
2993 set_remap_table_entry(iommu, alias, table);
2996 iommu_completion_wait(iommu);
2999 spin_unlock_irqrestore(&iommu_table_lock, flags);
3002 kmem_cache_free(amd_iommu_irq_cache, new_table->table);
3008 static int alloc_irq_index(struct amd_iommu *iommu, u16 devid, int count,
3009 bool align, struct pci_dev *pdev)
3011 struct irq_remap_table *table;
3012 int index, c, alignment = 1;
3013 unsigned long flags;
3015 table = alloc_irq_table(iommu, devid, pdev);
3020 alignment = roundup_pow_of_two(count);
3022 raw_spin_lock_irqsave(&table->lock, flags);
3024 /* Scan table for free entries */
3025 for (index = ALIGN(table->min_index, alignment), c = 0;
3026 index < MAX_IRQS_PER_TABLE;) {
3027 if (!iommu->irte_ops->is_allocated(table, index)) {
3031 index = ALIGN(index + 1, alignment);
3037 iommu->irte_ops->set_allocated(table, index - c + 1);
3049 raw_spin_unlock_irqrestore(&table->lock, flags);
3054 static int modify_irte_ga(struct amd_iommu *iommu, u16 devid, int index,
3055 struct irte_ga *irte)
3057 struct irq_remap_table *table;
3058 struct irte_ga *entry;
3059 unsigned long flags;
3062 table = get_irq_table(iommu, devid);
3066 raw_spin_lock_irqsave(&table->lock, flags);
3068 entry = (struct irte_ga *)table->table;
3069 entry = &entry[index];
3072 * We use cmpxchg16 to atomically update the 128-bit IRTE,
3073 * and it cannot be updated by the hardware or other processors
3074 * behind us, so the return value of cmpxchg16 should be the
3075 * same as the old value.
3078 WARN_ON(!try_cmpxchg128(&entry->irte, &old, irte->irte));
3080 raw_spin_unlock_irqrestore(&table->lock, flags);
3082 iommu_flush_irt_and_complete(iommu, devid);
3087 static int modify_irte(struct amd_iommu *iommu,
3088 u16 devid, int index, union irte *irte)
3090 struct irq_remap_table *table;
3091 unsigned long flags;
3093 table = get_irq_table(iommu, devid);
3097 raw_spin_lock_irqsave(&table->lock, flags);
3098 table->table[index] = irte->val;
3099 raw_spin_unlock_irqrestore(&table->lock, flags);
3101 iommu_flush_irt_and_complete(iommu, devid);
3106 static void free_irte(struct amd_iommu *iommu, u16 devid, int index)
3108 struct irq_remap_table *table;
3109 unsigned long flags;
3111 table = get_irq_table(iommu, devid);
3115 raw_spin_lock_irqsave(&table->lock, flags);
3116 iommu->irte_ops->clear_allocated(table, index);
3117 raw_spin_unlock_irqrestore(&table->lock, flags);
3119 iommu_flush_irt_and_complete(iommu, devid);
3122 static void irte_prepare(void *entry,
3123 u32 delivery_mode, bool dest_mode,
3124 u8 vector, u32 dest_apicid, int devid)
3126 union irte *irte = (union irte *) entry;
3129 irte->fields.vector = vector;
3130 irte->fields.int_type = delivery_mode;
3131 irte->fields.destination = dest_apicid;
3132 irte->fields.dm = dest_mode;
3133 irte->fields.valid = 1;
3136 static void irte_ga_prepare(void *entry,
3137 u32 delivery_mode, bool dest_mode,
3138 u8 vector, u32 dest_apicid, int devid)
3140 struct irte_ga *irte = (struct irte_ga *) entry;
3144 irte->lo.fields_remap.int_type = delivery_mode;
3145 irte->lo.fields_remap.dm = dest_mode;
3146 irte->hi.fields.vector = vector;
3147 irte->lo.fields_remap.destination = APICID_TO_IRTE_DEST_LO(dest_apicid);
3148 irte->hi.fields.destination = APICID_TO_IRTE_DEST_HI(dest_apicid);
3149 irte->lo.fields_remap.valid = 1;
3152 static void irte_activate(struct amd_iommu *iommu, void *entry, u16 devid, u16 index)
3154 union irte *irte = (union irte *) entry;
3156 irte->fields.valid = 1;
3157 modify_irte(iommu, devid, index, irte);
3160 static void irte_ga_activate(struct amd_iommu *iommu, void *entry, u16 devid, u16 index)
3162 struct irte_ga *irte = (struct irte_ga *) entry;
3164 irte->lo.fields_remap.valid = 1;
3165 modify_irte_ga(iommu, devid, index, irte);
3168 static void irte_deactivate(struct amd_iommu *iommu, void *entry, u16 devid, u16 index)
3170 union irte *irte = (union irte *) entry;
3172 irte->fields.valid = 0;
3173 modify_irte(iommu, devid, index, irte);
3176 static void irte_ga_deactivate(struct amd_iommu *iommu, void *entry, u16 devid, u16 index)
3178 struct irte_ga *irte = (struct irte_ga *) entry;
3180 irte->lo.fields_remap.valid = 0;
3181 modify_irte_ga(iommu, devid, index, irte);
3184 static void irte_set_affinity(struct amd_iommu *iommu, void *entry, u16 devid, u16 index,
3185 u8 vector, u32 dest_apicid)
3187 union irte *irte = (union irte *) entry;
3189 irte->fields.vector = vector;
3190 irte->fields.destination = dest_apicid;
3191 modify_irte(iommu, devid, index, irte);
3194 static void irte_ga_set_affinity(struct amd_iommu *iommu, void *entry, u16 devid, u16 index,
3195 u8 vector, u32 dest_apicid)
3197 struct irte_ga *irte = (struct irte_ga *) entry;
3199 if (!irte->lo.fields_remap.guest_mode) {
3200 irte->hi.fields.vector = vector;
3201 irte->lo.fields_remap.destination =
3202 APICID_TO_IRTE_DEST_LO(dest_apicid);
3203 irte->hi.fields.destination =
3204 APICID_TO_IRTE_DEST_HI(dest_apicid);
3205 modify_irte_ga(iommu, devid, index, irte);
3209 #define IRTE_ALLOCATED (~1U)
3210 static void irte_set_allocated(struct irq_remap_table *table, int index)
3212 table->table[index] = IRTE_ALLOCATED;
3215 static void irte_ga_set_allocated(struct irq_remap_table *table, int index)
3217 struct irte_ga *ptr = (struct irte_ga *)table->table;
3218 struct irte_ga *irte = &ptr[index];
3220 memset(&irte->lo.val, 0, sizeof(u64));
3221 memset(&irte->hi.val, 0, sizeof(u64));
3222 irte->hi.fields.vector = 0xff;
3225 static bool irte_is_allocated(struct irq_remap_table *table, int index)
3227 union irte *ptr = (union irte *)table->table;
3228 union irte *irte = &ptr[index];
3230 return irte->val != 0;
3233 static bool irte_ga_is_allocated(struct irq_remap_table *table, int index)
3235 struct irte_ga *ptr = (struct irte_ga *)table->table;
3236 struct irte_ga *irte = &ptr[index];
3238 return irte->hi.fields.vector != 0;
3241 static void irte_clear_allocated(struct irq_remap_table *table, int index)
3243 table->table[index] = 0;
3246 static void irte_ga_clear_allocated(struct irq_remap_table *table, int index)
3248 struct irte_ga *ptr = (struct irte_ga *)table->table;
3249 struct irte_ga *irte = &ptr[index];
3251 memset(&irte->lo.val, 0, sizeof(u64));
3252 memset(&irte->hi.val, 0, sizeof(u64));
3255 static int get_devid(struct irq_alloc_info *info)
3257 switch (info->type) {
3258 case X86_IRQ_ALLOC_TYPE_IOAPIC:
3259 return get_ioapic_devid(info->devid);
3260 case X86_IRQ_ALLOC_TYPE_HPET:
3261 return get_hpet_devid(info->devid);
3262 case X86_IRQ_ALLOC_TYPE_PCI_MSI:
3263 case X86_IRQ_ALLOC_TYPE_PCI_MSIX:
3264 return get_device_sbdf_id(msi_desc_to_dev(info->desc));
3271 struct irq_remap_ops amd_iommu_irq_ops = {
3272 .prepare = amd_iommu_prepare,
3273 .enable = amd_iommu_enable,
3274 .disable = amd_iommu_disable,
3275 .reenable = amd_iommu_reenable,
3276 .enable_faulting = amd_iommu_enable_faulting,
3279 static void fill_msi_msg(struct msi_msg *msg, u32 index)
3282 msg->address_lo = 0;
3283 msg->arch_addr_lo.base_address = X86_MSI_BASE_ADDRESS_LOW;
3284 msg->address_hi = X86_MSI_BASE_ADDRESS_HIGH;
3287 static void irq_remapping_prepare_irte(struct amd_ir_data *data,
3288 struct irq_cfg *irq_cfg,
3289 struct irq_alloc_info *info,
3290 int devid, int index, int sub_handle)
3292 struct irq_2_irte *irte_info = &data->irq_2_irte;
3293 struct amd_iommu *iommu = data->iommu;
3298 data->irq_2_irte.devid = devid;
3299 data->irq_2_irte.index = index + sub_handle;
3300 iommu->irte_ops->prepare(data->entry, apic->delivery_mode,
3301 apic->dest_mode_logical, irq_cfg->vector,
3302 irq_cfg->dest_apicid, devid);
3304 switch (info->type) {
3305 case X86_IRQ_ALLOC_TYPE_IOAPIC:
3306 case X86_IRQ_ALLOC_TYPE_HPET:
3307 case X86_IRQ_ALLOC_TYPE_PCI_MSI:
3308 case X86_IRQ_ALLOC_TYPE_PCI_MSIX:
3309 fill_msi_msg(&data->msi_entry, irte_info->index);
3318 struct amd_irte_ops irte_32_ops = {
3319 .prepare = irte_prepare,
3320 .activate = irte_activate,
3321 .deactivate = irte_deactivate,
3322 .set_affinity = irte_set_affinity,
3323 .set_allocated = irte_set_allocated,
3324 .is_allocated = irte_is_allocated,
3325 .clear_allocated = irte_clear_allocated,
3328 struct amd_irte_ops irte_128_ops = {
3329 .prepare = irte_ga_prepare,
3330 .activate = irte_ga_activate,
3331 .deactivate = irte_ga_deactivate,
3332 .set_affinity = irte_ga_set_affinity,
3333 .set_allocated = irte_ga_set_allocated,
3334 .is_allocated = irte_ga_is_allocated,
3335 .clear_allocated = irte_ga_clear_allocated,
3338 static int irq_remapping_alloc(struct irq_domain *domain, unsigned int virq,
3339 unsigned int nr_irqs, void *arg)
3341 struct irq_alloc_info *info = arg;
3342 struct irq_data *irq_data;
3343 struct amd_ir_data *data = NULL;
3344 struct amd_iommu *iommu;
3345 struct irq_cfg *cfg;
3346 int i, ret, devid, seg, sbdf;
3351 if (nr_irqs > 1 && info->type != X86_IRQ_ALLOC_TYPE_PCI_MSI)
3354 sbdf = get_devid(info);
3358 seg = PCI_SBDF_TO_SEGID(sbdf);
3359 devid = PCI_SBDF_TO_DEVID(sbdf);
3360 iommu = __rlookup_amd_iommu(seg, devid);
3364 ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
3368 if (info->type == X86_IRQ_ALLOC_TYPE_IOAPIC) {
3369 struct irq_remap_table *table;
3371 table = alloc_irq_table(iommu, devid, NULL);
3373 if (!table->min_index) {
3375 * Keep the first 32 indexes free for IOAPIC
3378 table->min_index = 32;
3379 for (i = 0; i < 32; ++i)
3380 iommu->irte_ops->set_allocated(table, i);
3382 WARN_ON(table->min_index != 32);
3383 index = info->ioapic.pin;
3387 } else if (info->type == X86_IRQ_ALLOC_TYPE_PCI_MSI ||
3388 info->type == X86_IRQ_ALLOC_TYPE_PCI_MSIX) {
3389 bool align = (info->type == X86_IRQ_ALLOC_TYPE_PCI_MSI);
3391 index = alloc_irq_index(iommu, devid, nr_irqs, align,
3392 msi_desc_to_pci_dev(info->desc));
3394 index = alloc_irq_index(iommu, devid, nr_irqs, false, NULL);
3398 pr_warn("Failed to allocate IRTE\n");
3400 goto out_free_parent;
3403 for (i = 0; i < nr_irqs; i++) {
3404 irq_data = irq_domain_get_irq_data(domain, virq + i);
3405 cfg = irq_data ? irqd_cfg(irq_data) : NULL;
3412 data = kzalloc(sizeof(*data), GFP_KERNEL);
3416 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
3417 data->entry = kzalloc(sizeof(union irte), GFP_KERNEL);
3419 data->entry = kzalloc(sizeof(struct irte_ga),
3426 data->iommu = iommu;
3427 irq_data->hwirq = (devid << 16) + i;
3428 irq_data->chip_data = data;
3429 irq_data->chip = &amd_ir_chip;
3430 irq_remapping_prepare_irte(data, cfg, info, devid, index, i);
3431 irq_set_status_flags(virq + i, IRQ_MOVE_PCNTXT);
3437 for (i--; i >= 0; i--) {
3438 irq_data = irq_domain_get_irq_data(domain, virq + i);
3440 kfree(irq_data->chip_data);
3442 for (i = 0; i < nr_irqs; i++)
3443 free_irte(iommu, devid, index + i);
3445 irq_domain_free_irqs_common(domain, virq, nr_irqs);
3449 static void irq_remapping_free(struct irq_domain *domain, unsigned int virq,
3450 unsigned int nr_irqs)
3452 struct irq_2_irte *irte_info;
3453 struct irq_data *irq_data;
3454 struct amd_ir_data *data;
3457 for (i = 0; i < nr_irqs; i++) {
3458 irq_data = irq_domain_get_irq_data(domain, virq + i);
3459 if (irq_data && irq_data->chip_data) {
3460 data = irq_data->chip_data;
3461 irte_info = &data->irq_2_irte;
3462 free_irte(data->iommu, irte_info->devid, irte_info->index);
3467 irq_domain_free_irqs_common(domain, virq, nr_irqs);
3470 static void amd_ir_update_irte(struct irq_data *irqd, struct amd_iommu *iommu,
3471 struct amd_ir_data *ir_data,
3472 struct irq_2_irte *irte_info,
3473 struct irq_cfg *cfg);
3475 static int irq_remapping_activate(struct irq_domain *domain,
3476 struct irq_data *irq_data, bool reserve)
3478 struct amd_ir_data *data = irq_data->chip_data;
3479 struct irq_2_irte *irte_info = &data->irq_2_irte;
3480 struct amd_iommu *iommu = data->iommu;
3481 struct irq_cfg *cfg = irqd_cfg(irq_data);
3486 iommu->irte_ops->activate(iommu, data->entry, irte_info->devid,
3488 amd_ir_update_irte(irq_data, iommu, data, irte_info, cfg);
3492 static void irq_remapping_deactivate(struct irq_domain *domain,
3493 struct irq_data *irq_data)
3495 struct amd_ir_data *data = irq_data->chip_data;
3496 struct irq_2_irte *irte_info = &data->irq_2_irte;
3497 struct amd_iommu *iommu = data->iommu;
3500 iommu->irte_ops->deactivate(iommu, data->entry, irte_info->devid,
3504 static int irq_remapping_select(struct irq_domain *d, struct irq_fwspec *fwspec,
3505 enum irq_domain_bus_token bus_token)
3507 struct amd_iommu *iommu;
3510 if (!amd_iommu_irq_remap)
3513 if (x86_fwspec_is_ioapic(fwspec))
3514 devid = get_ioapic_devid(fwspec->param[0]);
3515 else if (x86_fwspec_is_hpet(fwspec))
3516 devid = get_hpet_devid(fwspec->param[0]);
3520 iommu = __rlookup_amd_iommu((devid >> 16), (devid & 0xffff));
3522 return iommu && iommu->ir_domain == d;
3525 static const struct irq_domain_ops amd_ir_domain_ops = {
3526 .select = irq_remapping_select,
3527 .alloc = irq_remapping_alloc,
3528 .free = irq_remapping_free,
3529 .activate = irq_remapping_activate,
3530 .deactivate = irq_remapping_deactivate,
3533 int amd_iommu_activate_guest_mode(void *data)
3535 struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
3536 struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
3539 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) || !entry)
3542 valid = entry->lo.fields_vapic.valid;
3547 entry->lo.fields_vapic.valid = valid;
3548 entry->lo.fields_vapic.guest_mode = 1;
3549 entry->lo.fields_vapic.ga_log_intr = 1;
3550 entry->hi.fields.ga_root_ptr = ir_data->ga_root_ptr;
3551 entry->hi.fields.vector = ir_data->ga_vector;
3552 entry->lo.fields_vapic.ga_tag = ir_data->ga_tag;
3554 return modify_irte_ga(ir_data->iommu, ir_data->irq_2_irte.devid,
3555 ir_data->irq_2_irte.index, entry);
3557 EXPORT_SYMBOL(amd_iommu_activate_guest_mode);
3559 int amd_iommu_deactivate_guest_mode(void *data)
3561 struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
3562 struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
3563 struct irq_cfg *cfg = ir_data->cfg;
3566 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
3567 !entry || !entry->lo.fields_vapic.guest_mode)
3570 valid = entry->lo.fields_remap.valid;
3575 entry->lo.fields_remap.valid = valid;
3576 entry->lo.fields_remap.dm = apic->dest_mode_logical;
3577 entry->lo.fields_remap.int_type = apic->delivery_mode;
3578 entry->hi.fields.vector = cfg->vector;
3579 entry->lo.fields_remap.destination =
3580 APICID_TO_IRTE_DEST_LO(cfg->dest_apicid);
3581 entry->hi.fields.destination =
3582 APICID_TO_IRTE_DEST_HI(cfg->dest_apicid);
3584 return modify_irte_ga(ir_data->iommu, ir_data->irq_2_irte.devid,
3585 ir_data->irq_2_irte.index, entry);
3587 EXPORT_SYMBOL(amd_iommu_deactivate_guest_mode);
3589 static int amd_ir_set_vcpu_affinity(struct irq_data *data, void *vcpu_info)
3592 struct amd_iommu_pi_data *pi_data = vcpu_info;
3593 struct vcpu_data *vcpu_pi_info = pi_data->vcpu_data;
3594 struct amd_ir_data *ir_data = data->chip_data;
3595 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
3596 struct iommu_dev_data *dev_data;
3598 if (ir_data->iommu == NULL)
3601 dev_data = search_dev_data(ir_data->iommu, irte_info->devid);
3604 * This device has never been set up for guest mode.
3605 * we should not modify the IRTE
3607 if (!dev_data || !dev_data->use_vapic)
3610 ir_data->cfg = irqd_cfg(data);
3611 pi_data->ir_data = ir_data;
3614 * SVM tries to set up for VAPIC mode, but we are in
3615 * legacy mode. So, we force legacy mode instead.
3617 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
3618 pr_debug("%s: Fall back to using intr legacy remap\n",
3620 pi_data->is_guest_mode = false;
3623 pi_data->prev_ga_tag = ir_data->cached_ga_tag;
3624 if (pi_data->is_guest_mode) {
3625 ir_data->ga_root_ptr = (pi_data->base >> 12);
3626 ir_data->ga_vector = vcpu_pi_info->vector;
3627 ir_data->ga_tag = pi_data->ga_tag;
3628 ret = amd_iommu_activate_guest_mode(ir_data);
3630 ir_data->cached_ga_tag = pi_data->ga_tag;
3632 ret = amd_iommu_deactivate_guest_mode(ir_data);
3635 * This communicates the ga_tag back to the caller
3636 * so that it can do all the necessary clean up.
3639 ir_data->cached_ga_tag = 0;
3646 static void amd_ir_update_irte(struct irq_data *irqd, struct amd_iommu *iommu,
3647 struct amd_ir_data *ir_data,
3648 struct irq_2_irte *irte_info,
3649 struct irq_cfg *cfg)
3653 * Atomically updates the IRTE with the new destination, vector
3654 * and flushes the interrupt entry cache.
3656 iommu->irte_ops->set_affinity(iommu, ir_data->entry, irte_info->devid,
3657 irte_info->index, cfg->vector,
3661 static int amd_ir_set_affinity(struct irq_data *data,
3662 const struct cpumask *mask, bool force)
3664 struct amd_ir_data *ir_data = data->chip_data;
3665 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
3666 struct irq_cfg *cfg = irqd_cfg(data);
3667 struct irq_data *parent = data->parent_data;
3668 struct amd_iommu *iommu = ir_data->iommu;
3674 ret = parent->chip->irq_set_affinity(parent, mask, force);
3675 if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
3678 amd_ir_update_irte(data, iommu, ir_data, irte_info, cfg);
3680 * After this point, all the interrupts will start arriving
3681 * at the new destination. So, time to cleanup the previous
3682 * vector allocation.
3684 send_cleanup_vector(cfg);
3686 return IRQ_SET_MASK_OK_DONE;
3689 static void ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg)
3691 struct amd_ir_data *ir_data = irq_data->chip_data;
3693 *msg = ir_data->msi_entry;
3696 static struct irq_chip amd_ir_chip = {
3698 .irq_ack = apic_ack_irq,
3699 .irq_set_affinity = amd_ir_set_affinity,
3700 .irq_set_vcpu_affinity = amd_ir_set_vcpu_affinity,
3701 .irq_compose_msi_msg = ir_compose_msi_msg,
3704 static const struct msi_parent_ops amdvi_msi_parent_ops = {
3705 .supported_flags = X86_VECTOR_MSI_FLAGS_SUPPORTED |
3706 MSI_FLAG_MULTI_PCI_MSI |
3709 .init_dev_msi_info = msi_parent_init_dev_msi_info,
3712 static const struct msi_parent_ops virt_amdvi_msi_parent_ops = {
3713 .supported_flags = X86_VECTOR_MSI_FLAGS_SUPPORTED |
3714 MSI_FLAG_MULTI_PCI_MSI,
3716 .init_dev_msi_info = msi_parent_init_dev_msi_info,
3719 int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
3721 struct fwnode_handle *fn;
3723 fn = irq_domain_alloc_named_id_fwnode("AMD-IR", iommu->index);
3726 iommu->ir_domain = irq_domain_create_hierarchy(arch_get_ir_parent_domain(), 0, 0,
3727 fn, &amd_ir_domain_ops, iommu);
3728 if (!iommu->ir_domain) {
3729 irq_domain_free_fwnode(fn);
3733 irq_domain_update_bus_token(iommu->ir_domain, DOMAIN_BUS_AMDVI);
3734 iommu->ir_domain->flags |= IRQ_DOMAIN_FLAG_MSI_PARENT |
3735 IRQ_DOMAIN_FLAG_ISOLATED_MSI;
3737 if (amd_iommu_np_cache)
3738 iommu->ir_domain->msi_parent_ops = &virt_amdvi_msi_parent_ops;
3740 iommu->ir_domain->msi_parent_ops = &amdvi_msi_parent_ops;
3745 int amd_iommu_update_ga(int cpu, bool is_run, void *data)
3747 struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
3748 struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
3750 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
3751 !entry || !entry->lo.fields_vapic.guest_mode)
3754 if (!ir_data->iommu)
3758 entry->lo.fields_vapic.destination =
3759 APICID_TO_IRTE_DEST_LO(cpu);
3760 entry->hi.fields.destination =
3761 APICID_TO_IRTE_DEST_HI(cpu);
3763 entry->lo.fields_vapic.is_run = is_run;
3765 return modify_irte_ga(ir_data->iommu, ir_data->irq_2_irte.devid,
3766 ir_data->irq_2_irte.index, entry);
3768 EXPORT_SYMBOL(amd_iommu_update_ga);