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_INT_MASK)
850 irqreturn_t amd_iommu_int_thread(int irq, void *data)
852 struct amd_iommu *iommu = (struct amd_iommu *) data;
853 u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
855 while (status & AMD_IOMMU_INT_MASK) {
856 /* Enable interrupt sources again */
857 writel(AMD_IOMMU_INT_MASK,
858 iommu->mmio_base + MMIO_STATUS_OFFSET);
860 if (status & MMIO_STATUS_EVT_INT_MASK) {
861 pr_devel("Processing IOMMU Event Log\n");
862 iommu_poll_events(iommu);
865 if (status & MMIO_STATUS_PPR_INT_MASK) {
866 pr_devel("Processing IOMMU PPR Log\n");
867 iommu_poll_ppr_log(iommu);
870 #ifdef CONFIG_IRQ_REMAP
871 if (status & MMIO_STATUS_GALOG_INT_MASK) {
872 pr_devel("Processing IOMMU GA Log\n");
873 iommu_poll_ga_log(iommu);
877 if (status & MMIO_STATUS_EVT_OVERFLOW_INT_MASK) {
878 pr_info_ratelimited("IOMMU event log overflow\n");
879 amd_iommu_restart_event_logging(iommu);
883 * Hardware bug: ERBT1312
884 * When re-enabling interrupt (by writing 1
885 * to clear the bit), the hardware might also try to set
886 * the interrupt bit in the event status register.
887 * In this scenario, the bit will be set, and disable
888 * subsequent interrupts.
890 * Workaround: The IOMMU driver should read back the
891 * status register and check if the interrupt bits are cleared.
892 * If not, driver will need to go through the interrupt handler
893 * again and re-clear the bits
895 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
900 irqreturn_t amd_iommu_int_handler(int irq, void *data)
902 return IRQ_WAKE_THREAD;
905 /****************************************************************************
907 * IOMMU command queuing functions
909 ****************************************************************************/
911 static int wait_on_sem(struct amd_iommu *iommu, u64 data)
915 while (*iommu->cmd_sem != data && i < LOOP_TIMEOUT) {
920 if (i == LOOP_TIMEOUT) {
921 pr_alert("Completion-Wait loop timed out\n");
928 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
929 struct iommu_cmd *cmd)
934 /* Copy command to buffer */
935 tail = iommu->cmd_buf_tail;
936 target = iommu->cmd_buf + tail;
937 memcpy(target, cmd, sizeof(*cmd));
939 tail = (tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
940 iommu->cmd_buf_tail = tail;
942 /* Tell the IOMMU about it */
943 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
946 static void build_completion_wait(struct iommu_cmd *cmd,
947 struct amd_iommu *iommu,
950 u64 paddr = iommu_virt_to_phys((void *)iommu->cmd_sem);
952 memset(cmd, 0, sizeof(*cmd));
953 cmd->data[0] = lower_32_bits(paddr) | CMD_COMPL_WAIT_STORE_MASK;
954 cmd->data[1] = upper_32_bits(paddr);
955 cmd->data[2] = lower_32_bits(data);
956 cmd->data[3] = upper_32_bits(data);
957 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
960 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
962 memset(cmd, 0, sizeof(*cmd));
963 cmd->data[0] = devid;
964 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
968 * Builds an invalidation address which is suitable for one page or multiple
969 * pages. Sets the size bit (S) as needed is more than one page is flushed.
971 static inline u64 build_inv_address(u64 address, size_t size)
973 u64 pages, end, msb_diff;
975 pages = iommu_num_pages(address, size, PAGE_SIZE);
978 return address & PAGE_MASK;
980 end = address + size - 1;
983 * msb_diff would hold the index of the most significant bit that
984 * flipped between the start and end.
986 msb_diff = fls64(end ^ address) - 1;
989 * Bits 63:52 are sign extended. If for some reason bit 51 is different
990 * between the start and the end, invalidate everything.
992 if (unlikely(msb_diff > 51)) {
993 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
996 * The msb-bit must be clear on the address. Just set all the
999 address |= (1ull << msb_diff) - 1;
1002 /* Clear bits 11:0 */
1003 address &= PAGE_MASK;
1005 /* Set the size bit - we flush more than one 4kb page */
1006 return address | CMD_INV_IOMMU_PAGES_SIZE_MASK;
1009 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
1010 size_t size, u16 domid, int pde)
1012 u64 inv_address = build_inv_address(address, size);
1014 memset(cmd, 0, sizeof(*cmd));
1015 cmd->data[1] |= domid;
1016 cmd->data[2] = lower_32_bits(inv_address);
1017 cmd->data[3] = upper_32_bits(inv_address);
1018 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
1019 if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
1020 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
1023 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
1024 u64 address, size_t size)
1026 u64 inv_address = build_inv_address(address, size);
1028 memset(cmd, 0, sizeof(*cmd));
1029 cmd->data[0] = devid;
1030 cmd->data[0] |= (qdep & 0xff) << 24;
1031 cmd->data[1] = devid;
1032 cmd->data[2] = lower_32_bits(inv_address);
1033 cmd->data[3] = upper_32_bits(inv_address);
1034 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
1037 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, u32 pasid,
1038 u64 address, bool size)
1040 memset(cmd, 0, sizeof(*cmd));
1042 address &= ~(0xfffULL);
1044 cmd->data[0] = pasid;
1045 cmd->data[1] = domid;
1046 cmd->data[2] = lower_32_bits(address);
1047 cmd->data[3] = upper_32_bits(address);
1048 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
1049 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
1051 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
1052 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
1055 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, u32 pasid,
1056 int qdep, u64 address, bool size)
1058 memset(cmd, 0, sizeof(*cmd));
1060 address &= ~(0xfffULL);
1062 cmd->data[0] = devid;
1063 cmd->data[0] |= ((pasid >> 8) & 0xff) << 16;
1064 cmd->data[0] |= (qdep & 0xff) << 24;
1065 cmd->data[1] = devid;
1066 cmd->data[1] |= (pasid & 0xff) << 16;
1067 cmd->data[2] = lower_32_bits(address);
1068 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
1069 cmd->data[3] = upper_32_bits(address);
1071 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
1072 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
1075 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, u32 pasid,
1076 int status, int tag, bool gn)
1078 memset(cmd, 0, sizeof(*cmd));
1080 cmd->data[0] = devid;
1082 cmd->data[1] = pasid;
1083 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
1085 cmd->data[3] = tag & 0x1ff;
1086 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
1088 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
1091 static void build_inv_all(struct iommu_cmd *cmd)
1093 memset(cmd, 0, sizeof(*cmd));
1094 CMD_SET_TYPE(cmd, CMD_INV_ALL);
1097 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
1099 memset(cmd, 0, sizeof(*cmd));
1100 cmd->data[0] = devid;
1101 CMD_SET_TYPE(cmd, CMD_INV_IRT);
1105 * Writes the command to the IOMMUs command buffer and informs the
1106 * hardware about the new command.
1108 static int __iommu_queue_command_sync(struct amd_iommu *iommu,
1109 struct iommu_cmd *cmd,
1112 unsigned int count = 0;
1113 u32 left, next_tail;
1115 next_tail = (iommu->cmd_buf_tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
1117 left = (iommu->cmd_buf_head - next_tail) % CMD_BUFFER_SIZE;
1120 /* Skip udelay() the first time around */
1122 if (count == LOOP_TIMEOUT) {
1123 pr_err("Command buffer timeout\n");
1130 /* Update head and recheck remaining space */
1131 iommu->cmd_buf_head = readl(iommu->mmio_base +
1132 MMIO_CMD_HEAD_OFFSET);
1137 copy_cmd_to_buffer(iommu, cmd);
1139 /* Do we need to make sure all commands are processed? */
1140 iommu->need_sync = sync;
1145 static int iommu_queue_command_sync(struct amd_iommu *iommu,
1146 struct iommu_cmd *cmd,
1149 unsigned long flags;
1152 raw_spin_lock_irqsave(&iommu->lock, flags);
1153 ret = __iommu_queue_command_sync(iommu, cmd, sync);
1154 raw_spin_unlock_irqrestore(&iommu->lock, flags);
1159 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
1161 return iommu_queue_command_sync(iommu, cmd, true);
1165 * This function queues a completion wait command into the command
1166 * buffer of an IOMMU
1168 static int iommu_completion_wait(struct amd_iommu *iommu)
1170 struct iommu_cmd cmd;
1171 unsigned long flags;
1175 if (!iommu->need_sync)
1178 raw_spin_lock_irqsave(&iommu->lock, flags);
1180 data = ++iommu->cmd_sem_val;
1181 build_completion_wait(&cmd, iommu, data);
1183 ret = __iommu_queue_command_sync(iommu, &cmd, false);
1187 ret = wait_on_sem(iommu, data);
1190 raw_spin_unlock_irqrestore(&iommu->lock, flags);
1195 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
1197 struct iommu_cmd cmd;
1199 build_inv_dte(&cmd, devid);
1201 return iommu_queue_command(iommu, &cmd);
1204 static void amd_iommu_flush_dte_all(struct amd_iommu *iommu)
1207 u16 last_bdf = iommu->pci_seg->last_bdf;
1209 for (devid = 0; devid <= last_bdf; ++devid)
1210 iommu_flush_dte(iommu, devid);
1212 iommu_completion_wait(iommu);
1216 * This function uses heavy locking and may disable irqs for some time. But
1217 * this is no issue because it is only called during resume.
1219 static void amd_iommu_flush_tlb_all(struct amd_iommu *iommu)
1222 u16 last_bdf = iommu->pci_seg->last_bdf;
1224 for (dom_id = 0; dom_id <= last_bdf; ++dom_id) {
1225 struct iommu_cmd cmd;
1226 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1228 iommu_queue_command(iommu, &cmd);
1231 iommu_completion_wait(iommu);
1234 static void amd_iommu_flush_tlb_domid(struct amd_iommu *iommu, u32 dom_id)
1236 struct iommu_cmd cmd;
1238 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1240 iommu_queue_command(iommu, &cmd);
1242 iommu_completion_wait(iommu);
1245 static void amd_iommu_flush_all(struct amd_iommu *iommu)
1247 struct iommu_cmd cmd;
1249 build_inv_all(&cmd);
1251 iommu_queue_command(iommu, &cmd);
1252 iommu_completion_wait(iommu);
1255 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1257 struct iommu_cmd cmd;
1259 build_inv_irt(&cmd, devid);
1261 iommu_queue_command(iommu, &cmd);
1264 static void amd_iommu_flush_irt_all(struct amd_iommu *iommu)
1267 u16 last_bdf = iommu->pci_seg->last_bdf;
1269 for (devid = 0; devid <= last_bdf; devid++)
1270 iommu_flush_irt(iommu, devid);
1272 iommu_completion_wait(iommu);
1275 void iommu_flush_all_caches(struct amd_iommu *iommu)
1277 if (iommu_feature(iommu, FEATURE_IA)) {
1278 amd_iommu_flush_all(iommu);
1280 amd_iommu_flush_dte_all(iommu);
1281 amd_iommu_flush_irt_all(iommu);
1282 amd_iommu_flush_tlb_all(iommu);
1287 * Command send function for flushing on-device TLB
1289 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1290 u64 address, size_t size)
1292 struct amd_iommu *iommu;
1293 struct iommu_cmd cmd;
1296 qdep = dev_data->ats.qdep;
1297 iommu = rlookup_amd_iommu(dev_data->dev);
1301 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1303 return iommu_queue_command(iommu, &cmd);
1306 static int device_flush_dte_alias(struct pci_dev *pdev, u16 alias, void *data)
1308 struct amd_iommu *iommu = data;
1310 return iommu_flush_dte(iommu, alias);
1314 * Command send function for invalidating a device table entry
1316 static int device_flush_dte(struct iommu_dev_data *dev_data)
1318 struct amd_iommu *iommu;
1319 struct pci_dev *pdev = NULL;
1320 struct amd_iommu_pci_seg *pci_seg;
1324 iommu = rlookup_amd_iommu(dev_data->dev);
1328 if (dev_is_pci(dev_data->dev))
1329 pdev = to_pci_dev(dev_data->dev);
1332 ret = pci_for_each_dma_alias(pdev,
1333 device_flush_dte_alias, iommu);
1335 ret = iommu_flush_dte(iommu, dev_data->devid);
1339 pci_seg = iommu->pci_seg;
1340 alias = pci_seg->alias_table[dev_data->devid];
1341 if (alias != dev_data->devid) {
1342 ret = iommu_flush_dte(iommu, alias);
1347 if (dev_data->ats.enabled)
1348 ret = device_flush_iotlb(dev_data, 0, ~0UL);
1354 * TLB invalidation function which is called from the mapping functions.
1355 * It invalidates a single PTE if the range to flush is within a single
1356 * page. Otherwise it flushes the whole TLB of the IOMMU.
1358 static void __domain_flush_pages(struct protection_domain *domain,
1359 u64 address, size_t size, int pde)
1361 struct iommu_dev_data *dev_data;
1362 struct iommu_cmd cmd;
1365 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1367 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
1368 if (!domain->dev_iommu[i])
1372 * Devices of this domain are behind this IOMMU
1373 * We need a TLB flush
1375 ret |= iommu_queue_command(amd_iommus[i], &cmd);
1378 list_for_each_entry(dev_data, &domain->dev_list, list) {
1380 if (!dev_data->ats.enabled)
1383 ret |= device_flush_iotlb(dev_data, address, size);
1389 static void domain_flush_pages(struct protection_domain *domain,
1390 u64 address, size_t size, int pde)
1392 if (likely(!amd_iommu_np_cache)) {
1393 __domain_flush_pages(domain, address, size, pde);
1398 * When NpCache is on, we infer that we run in a VM and use a vIOMMU.
1399 * In such setups it is best to avoid flushes of ranges which are not
1400 * naturally aligned, since it would lead to flushes of unmodified
1401 * PTEs. Such flushes would require the hypervisor to do more work than
1402 * necessary. Therefore, perform repeated flushes of aligned ranges
1403 * until you cover the range. Each iteration flushes the smaller
1404 * between the natural alignment of the address that we flush and the
1405 * greatest naturally aligned region that fits in the range.
1408 int addr_alignment = __ffs(address);
1409 int size_alignment = __fls(size);
1414 * size is always non-zero, but address might be zero, causing
1415 * addr_alignment to be negative. As the casting of the
1416 * argument in __ffs(address) to long might trim the high bits
1417 * of the address on x86-32, cast to long when doing the check.
1419 if (likely((unsigned long)address != 0))
1420 min_alignment = min(addr_alignment, size_alignment);
1422 min_alignment = size_alignment;
1424 flush_size = 1ul << min_alignment;
1426 __domain_flush_pages(domain, address, flush_size, pde);
1427 address += flush_size;
1432 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1433 void amd_iommu_domain_flush_tlb_pde(struct protection_domain *domain)
1435 domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1438 void amd_iommu_domain_flush_complete(struct protection_domain *domain)
1442 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
1443 if (domain && !domain->dev_iommu[i])
1447 * Devices of this domain are behind this IOMMU
1448 * We need to wait for completion of all commands.
1450 iommu_completion_wait(amd_iommus[i]);
1454 /* Flush the not present cache if it exists */
1455 static void domain_flush_np_cache(struct protection_domain *domain,
1456 dma_addr_t iova, size_t size)
1458 if (unlikely(amd_iommu_np_cache)) {
1459 unsigned long flags;
1461 spin_lock_irqsave(&domain->lock, flags);
1462 domain_flush_pages(domain, iova, size, 1);
1463 amd_iommu_domain_flush_complete(domain);
1464 spin_unlock_irqrestore(&domain->lock, flags);
1470 * This function flushes the DTEs for all devices in domain
1472 static void domain_flush_devices(struct protection_domain *domain)
1474 struct iommu_dev_data *dev_data;
1476 list_for_each_entry(dev_data, &domain->dev_list, list)
1477 device_flush_dte(dev_data);
1480 /****************************************************************************
1482 * The next functions belong to the domain allocation. A domain is
1483 * allocated for every IOMMU as the default domain. If device isolation
1484 * is enabled, every device get its own domain. The most important thing
1485 * about domains is the page table mapping the DMA address space they
1488 ****************************************************************************/
1490 static u16 domain_id_alloc(void)
1494 spin_lock(&pd_bitmap_lock);
1495 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1497 if (id > 0 && id < MAX_DOMAIN_ID)
1498 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1501 spin_unlock(&pd_bitmap_lock);
1506 static void domain_id_free(int id)
1508 spin_lock(&pd_bitmap_lock);
1509 if (id > 0 && id < MAX_DOMAIN_ID)
1510 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1511 spin_unlock(&pd_bitmap_lock);
1514 static void free_gcr3_tbl_level1(u64 *tbl)
1519 for (i = 0; i < 512; ++i) {
1520 if (!(tbl[i] & GCR3_VALID))
1523 ptr = iommu_phys_to_virt(tbl[i] & PAGE_MASK);
1525 free_page((unsigned long)ptr);
1529 static void free_gcr3_tbl_level2(u64 *tbl)
1534 for (i = 0; i < 512; ++i) {
1535 if (!(tbl[i] & GCR3_VALID))
1538 ptr = iommu_phys_to_virt(tbl[i] & PAGE_MASK);
1540 free_gcr3_tbl_level1(ptr);
1544 static void free_gcr3_table(struct protection_domain *domain)
1546 if (domain->glx == 2)
1547 free_gcr3_tbl_level2(domain->gcr3_tbl);
1548 else if (domain->glx == 1)
1549 free_gcr3_tbl_level1(domain->gcr3_tbl);
1551 BUG_ON(domain->glx != 0);
1553 free_page((unsigned long)domain->gcr3_tbl);
1556 static void set_dte_entry(struct amd_iommu *iommu, u16 devid,
1557 struct protection_domain *domain, bool ats, bool ppr)
1562 struct dev_table_entry *dev_table = get_dev_table(iommu);
1564 if (domain->iop.mode != PAGE_MODE_NONE)
1565 pte_root = iommu_virt_to_phys(domain->iop.root);
1567 pte_root |= (domain->iop.mode & DEV_ENTRY_MODE_MASK)
1568 << DEV_ENTRY_MODE_SHIFT;
1570 pte_root |= DTE_FLAG_IR | DTE_FLAG_IW | DTE_FLAG_V;
1573 * When SNP is enabled, Only set TV bit when IOMMU
1574 * page translation is in use.
1576 if (!amd_iommu_snp_en || (domain->id != 0))
1577 pte_root |= DTE_FLAG_TV;
1579 flags = dev_table[devid].data[1];
1582 flags |= DTE_FLAG_IOTLB;
1585 if (iommu_feature(iommu, FEATURE_EPHSUP))
1586 pte_root |= 1ULL << DEV_ENTRY_PPR;
1589 if (domain->flags & PD_IOMMUV2_MASK) {
1590 u64 gcr3 = iommu_virt_to_phys(domain->gcr3_tbl);
1591 u64 glx = domain->glx;
1594 pte_root |= DTE_FLAG_GV;
1595 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
1597 /* First mask out possible old values for GCR3 table */
1598 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
1601 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
1604 /* Encode GCR3 table into DTE */
1605 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
1608 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
1611 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
1614 if (domain->flags & PD_GIOV_MASK)
1615 pte_root |= DTE_FLAG_GIOV;
1618 flags &= ~DEV_DOMID_MASK;
1619 flags |= domain->id;
1621 old_domid = dev_table[devid].data[1] & DEV_DOMID_MASK;
1622 dev_table[devid].data[1] = flags;
1623 dev_table[devid].data[0] = pte_root;
1626 * A kdump kernel might be replacing a domain ID that was copied from
1627 * the previous kernel--if so, it needs to flush the translation cache
1628 * entries for the old domain ID that is being overwritten
1631 amd_iommu_flush_tlb_domid(iommu, old_domid);
1635 static void clear_dte_entry(struct amd_iommu *iommu, u16 devid)
1637 struct dev_table_entry *dev_table = get_dev_table(iommu);
1639 /* remove entry from the device table seen by the hardware */
1640 dev_table[devid].data[0] = DTE_FLAG_V;
1642 if (!amd_iommu_snp_en)
1643 dev_table[devid].data[0] |= DTE_FLAG_TV;
1645 dev_table[devid].data[1] &= DTE_FLAG_MASK;
1647 amd_iommu_apply_erratum_63(iommu, devid);
1650 static void do_attach(struct iommu_dev_data *dev_data,
1651 struct protection_domain *domain)
1653 struct amd_iommu *iommu;
1656 iommu = rlookup_amd_iommu(dev_data->dev);
1659 ats = dev_data->ats.enabled;
1661 /* Update data structures */
1662 dev_data->domain = domain;
1663 list_add(&dev_data->list, &domain->dev_list);
1665 /* Do reference counting */
1666 domain->dev_iommu[iommu->index] += 1;
1667 domain->dev_cnt += 1;
1669 /* Override supported page sizes */
1670 if (domain->flags & PD_GIOV_MASK)
1671 domain->domain.pgsize_bitmap = AMD_IOMMU_PGSIZES_V2;
1673 /* Update device table */
1674 set_dte_entry(iommu, dev_data->devid, domain,
1675 ats, dev_data->iommu_v2);
1676 clone_aliases(iommu, dev_data->dev);
1678 device_flush_dte(dev_data);
1681 static void do_detach(struct iommu_dev_data *dev_data)
1683 struct protection_domain *domain = dev_data->domain;
1684 struct amd_iommu *iommu;
1686 iommu = rlookup_amd_iommu(dev_data->dev);
1690 /* Update data structures */
1691 dev_data->domain = NULL;
1692 list_del(&dev_data->list);
1693 clear_dte_entry(iommu, dev_data->devid);
1694 clone_aliases(iommu, dev_data->dev);
1696 /* Flush the DTE entry */
1697 device_flush_dte(dev_data);
1700 amd_iommu_domain_flush_tlb_pde(domain);
1702 /* Wait for the flushes to finish */
1703 amd_iommu_domain_flush_complete(domain);
1705 /* decrease reference counters - needs to happen after the flushes */
1706 domain->dev_iommu[iommu->index] -= 1;
1707 domain->dev_cnt -= 1;
1710 static void pdev_iommuv2_disable(struct pci_dev *pdev)
1712 pci_disable_ats(pdev);
1713 pci_disable_pri(pdev);
1714 pci_disable_pasid(pdev);
1717 static int pdev_pri_ats_enable(struct pci_dev *pdev)
1721 /* Only allow access to user-accessible pages */
1722 ret = pci_enable_pasid(pdev, 0);
1726 /* First reset the PRI state of the device */
1727 ret = pci_reset_pri(pdev);
1732 /* FIXME: Hardcode number of outstanding requests for now */
1733 ret = pci_enable_pri(pdev, 32);
1737 ret = pci_enable_ats(pdev, PAGE_SHIFT);
1744 pci_disable_pri(pdev);
1747 pci_disable_pasid(pdev);
1753 * If a device is not yet associated with a domain, this function makes the
1754 * device visible in the domain
1756 static int attach_device(struct device *dev,
1757 struct protection_domain *domain)
1759 struct iommu_dev_data *dev_data;
1760 struct pci_dev *pdev;
1761 unsigned long flags;
1764 spin_lock_irqsave(&domain->lock, flags);
1766 dev_data = dev_iommu_priv_get(dev);
1768 spin_lock(&dev_data->lock);
1771 if (dev_data->domain != NULL)
1774 if (!dev_is_pci(dev))
1775 goto skip_ats_check;
1777 pdev = to_pci_dev(dev);
1778 if (domain->flags & PD_IOMMUV2_MASK) {
1779 struct iommu_domain *def_domain = iommu_get_dma_domain(dev);
1784 * In case of using AMD_IOMMU_V1 page table mode and the device
1785 * is enabling for PPR/ATS support (using v2 table),
1786 * we need to make sure that the domain type is identity map.
1788 if ((amd_iommu_pgtable == AMD_IOMMU_V1) &&
1789 def_domain->type != IOMMU_DOMAIN_IDENTITY) {
1793 if (dev_data->iommu_v2) {
1794 if (pdev_pri_ats_enable(pdev) != 0)
1797 dev_data->ats.enabled = true;
1798 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
1799 dev_data->pri_tlp = pci_prg_resp_pasid_required(pdev);
1801 } else if (amd_iommu_iotlb_sup &&
1802 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
1803 dev_data->ats.enabled = true;
1804 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
1810 do_attach(dev_data, domain);
1813 * We might boot into a crash-kernel here. The crashed kernel
1814 * left the caches in the IOMMU dirty. So we have to flush
1815 * here to evict all dirty stuff.
1817 amd_iommu_domain_flush_tlb_pde(domain);
1819 amd_iommu_domain_flush_complete(domain);
1822 spin_unlock(&dev_data->lock);
1824 spin_unlock_irqrestore(&domain->lock, flags);
1830 * Removes a device from a protection domain (with devtable_lock held)
1832 static void detach_device(struct device *dev)
1834 struct protection_domain *domain;
1835 struct iommu_dev_data *dev_data;
1836 unsigned long flags;
1838 dev_data = dev_iommu_priv_get(dev);
1839 domain = dev_data->domain;
1841 spin_lock_irqsave(&domain->lock, flags);
1843 spin_lock(&dev_data->lock);
1846 * First check if the device is still attached. It might already
1847 * be detached from its domain because the generic
1848 * iommu_detach_group code detached it and we try again here in
1849 * our alias handling.
1851 if (WARN_ON(!dev_data->domain))
1854 do_detach(dev_data);
1856 if (!dev_is_pci(dev))
1859 if (domain->flags & PD_IOMMUV2_MASK && dev_data->iommu_v2)
1860 pdev_iommuv2_disable(to_pci_dev(dev));
1861 else if (dev_data->ats.enabled)
1862 pci_disable_ats(to_pci_dev(dev));
1864 dev_data->ats.enabled = false;
1867 spin_unlock(&dev_data->lock);
1869 spin_unlock_irqrestore(&domain->lock, flags);
1872 static struct iommu_device *amd_iommu_probe_device(struct device *dev)
1874 struct iommu_device *iommu_dev;
1875 struct amd_iommu *iommu;
1878 if (!check_device(dev))
1879 return ERR_PTR(-ENODEV);
1881 iommu = rlookup_amd_iommu(dev);
1883 return ERR_PTR(-ENODEV);
1885 /* Not registered yet? */
1886 if (!iommu->iommu.ops)
1887 return ERR_PTR(-ENODEV);
1889 if (dev_iommu_priv_get(dev))
1890 return &iommu->iommu;
1892 ret = iommu_init_device(iommu, dev);
1894 if (ret != -ENOTSUPP)
1895 dev_err(dev, "Failed to initialize - trying to proceed anyway\n");
1896 iommu_dev = ERR_PTR(ret);
1897 iommu_ignore_device(iommu, dev);
1899 amd_iommu_set_pci_msi_domain(dev, iommu);
1900 iommu_dev = &iommu->iommu;
1903 iommu_completion_wait(iommu);
1908 static void amd_iommu_probe_finalize(struct device *dev)
1910 /* Domains are initialized for this device - have a look what we ended up with */
1911 set_dma_ops(dev, NULL);
1912 iommu_setup_dma_ops(dev, 0, U64_MAX);
1915 static void amd_iommu_release_device(struct device *dev)
1917 struct amd_iommu *iommu;
1919 if (!check_device(dev))
1922 iommu = rlookup_amd_iommu(dev);
1926 amd_iommu_uninit_device(dev);
1927 iommu_completion_wait(iommu);
1930 static struct iommu_group *amd_iommu_device_group(struct device *dev)
1932 if (dev_is_pci(dev))
1933 return pci_device_group(dev);
1935 return acpihid_device_group(dev);
1938 /*****************************************************************************
1940 * The next functions belong to the dma_ops mapping/unmapping code.
1942 *****************************************************************************/
1944 static void update_device_table(struct protection_domain *domain)
1946 struct iommu_dev_data *dev_data;
1948 list_for_each_entry(dev_data, &domain->dev_list, list) {
1949 struct amd_iommu *iommu = rlookup_amd_iommu(dev_data->dev);
1953 set_dte_entry(iommu, dev_data->devid, domain,
1954 dev_data->ats.enabled, dev_data->iommu_v2);
1955 clone_aliases(iommu, dev_data->dev);
1959 void amd_iommu_update_and_flush_device_table(struct protection_domain *domain)
1961 update_device_table(domain);
1962 domain_flush_devices(domain);
1965 void amd_iommu_domain_update(struct protection_domain *domain)
1967 /* Update device table */
1968 amd_iommu_update_and_flush_device_table(domain);
1970 /* Flush domain TLB(s) and wait for completion */
1971 amd_iommu_domain_flush_tlb_pde(domain);
1972 amd_iommu_domain_flush_complete(domain);
1975 /*****************************************************************************
1977 * The following functions belong to the exported interface of AMD IOMMU
1979 * This interface allows access to lower level functions of the IOMMU
1980 * like protection domain handling and assignement of devices to domains
1981 * which is not possible with the dma_ops interface.
1983 *****************************************************************************/
1985 static void cleanup_domain(struct protection_domain *domain)
1987 struct iommu_dev_data *entry;
1988 unsigned long flags;
1990 spin_lock_irqsave(&domain->lock, flags);
1992 while (!list_empty(&domain->dev_list)) {
1993 entry = list_first_entry(&domain->dev_list,
1994 struct iommu_dev_data, list);
1995 BUG_ON(!entry->domain);
1999 spin_unlock_irqrestore(&domain->lock, flags);
2002 static void protection_domain_free(struct protection_domain *domain)
2007 if (domain->iop.pgtbl_cfg.tlb)
2008 free_io_pgtable_ops(&domain->iop.iop.ops);
2011 domain_id_free(domain->id);
2016 static int protection_domain_init_v1(struct protection_domain *domain, int mode)
2018 u64 *pt_root = NULL;
2020 BUG_ON(mode < PAGE_MODE_NONE || mode > PAGE_MODE_6_LEVEL);
2022 spin_lock_init(&domain->lock);
2023 domain->id = domain_id_alloc();
2026 INIT_LIST_HEAD(&domain->dev_list);
2028 if (mode != PAGE_MODE_NONE) {
2029 pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2031 domain_id_free(domain->id);
2036 amd_iommu_domain_set_pgtable(domain, pt_root, mode);
2041 static int protection_domain_init_v2(struct protection_domain *domain)
2043 spin_lock_init(&domain->lock);
2044 domain->id = domain_id_alloc();
2047 INIT_LIST_HEAD(&domain->dev_list);
2049 domain->flags |= PD_GIOV_MASK;
2051 if (domain_enable_v2(domain, 1)) {
2052 domain_id_free(domain->id);
2059 static struct protection_domain *protection_domain_alloc(unsigned int type)
2061 struct io_pgtable_ops *pgtbl_ops;
2062 struct protection_domain *domain;
2063 int pgtable = amd_iommu_pgtable;
2064 int mode = DEFAULT_PGTABLE_LEVEL;
2067 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2072 * Force IOMMU v1 page table when iommu=pt and
2073 * when allocating domain for pass-through devices.
2075 if (type == IOMMU_DOMAIN_IDENTITY) {
2076 pgtable = AMD_IOMMU_V1;
2077 mode = PAGE_MODE_NONE;
2078 } else if (type == IOMMU_DOMAIN_UNMANAGED) {
2079 pgtable = AMD_IOMMU_V1;
2084 ret = protection_domain_init_v1(domain, mode);
2087 ret = protection_domain_init_v2(domain);
2096 /* No need to allocate io pgtable ops in passthrough mode */
2097 if (type == IOMMU_DOMAIN_IDENTITY)
2100 pgtbl_ops = alloc_io_pgtable_ops(pgtable, &domain->iop.pgtbl_cfg, domain);
2102 domain_id_free(domain->id);
2112 static struct iommu_domain *amd_iommu_domain_alloc(unsigned type)
2114 struct protection_domain *domain;
2117 * Since DTE[Mode]=0 is prohibited on SNP-enabled system,
2118 * default to use IOMMU_DOMAIN_DMA[_FQ].
2120 if (amd_iommu_snp_en && (type == IOMMU_DOMAIN_IDENTITY))
2123 domain = protection_domain_alloc(type);
2127 domain->domain.geometry.aperture_start = 0;
2128 domain->domain.geometry.aperture_end = ~0ULL;
2129 domain->domain.geometry.force_aperture = true;
2131 return &domain->domain;
2134 static void amd_iommu_domain_free(struct iommu_domain *dom)
2136 struct protection_domain *domain;
2138 domain = to_pdomain(dom);
2140 if (domain->dev_cnt > 0)
2141 cleanup_domain(domain);
2143 BUG_ON(domain->dev_cnt != 0);
2148 if (domain->flags & PD_IOMMUV2_MASK)
2149 free_gcr3_table(domain);
2151 protection_domain_free(domain);
2154 static int amd_iommu_attach_device(struct iommu_domain *dom,
2157 struct iommu_dev_data *dev_data = dev_iommu_priv_get(dev);
2158 struct protection_domain *domain = to_pdomain(dom);
2159 struct amd_iommu *iommu = rlookup_amd_iommu(dev);
2163 * Skip attach device to domain if new domain is same as
2164 * devices current domain
2166 if (dev_data->domain == domain)
2169 dev_data->defer_attach = false;
2171 if (dev_data->domain)
2174 ret = attach_device(dev, domain);
2176 #ifdef CONFIG_IRQ_REMAP
2177 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
2178 if (dom->type == IOMMU_DOMAIN_UNMANAGED)
2179 dev_data->use_vapic = 1;
2181 dev_data->use_vapic = 0;
2185 iommu_completion_wait(iommu);
2190 static void amd_iommu_iotlb_sync_map(struct iommu_domain *dom,
2191 unsigned long iova, size_t size)
2193 struct protection_domain *domain = to_pdomain(dom);
2194 struct io_pgtable_ops *ops = &domain->iop.iop.ops;
2197 domain_flush_np_cache(domain, iova, size);
2200 static int amd_iommu_map_pages(struct iommu_domain *dom, unsigned long iova,
2201 phys_addr_t paddr, size_t pgsize, size_t pgcount,
2202 int iommu_prot, gfp_t gfp, size_t *mapped)
2204 struct protection_domain *domain = to_pdomain(dom);
2205 struct io_pgtable_ops *ops = &domain->iop.iop.ops;
2209 if ((amd_iommu_pgtable == AMD_IOMMU_V1) &&
2210 (domain->iop.mode == PAGE_MODE_NONE))
2213 if (iommu_prot & IOMMU_READ)
2214 prot |= IOMMU_PROT_IR;
2215 if (iommu_prot & IOMMU_WRITE)
2216 prot |= IOMMU_PROT_IW;
2218 if (ops->map_pages) {
2219 ret = ops->map_pages(ops, iova, paddr, pgsize,
2220 pgcount, prot, gfp, mapped);
2226 static void amd_iommu_iotlb_gather_add_page(struct iommu_domain *domain,
2227 struct iommu_iotlb_gather *gather,
2228 unsigned long iova, size_t size)
2231 * AMD's IOMMU can flush as many pages as necessary in a single flush.
2232 * Unless we run in a virtual machine, which can be inferred according
2233 * to whether "non-present cache" is on, it is probably best to prefer
2234 * (potentially) too extensive TLB flushing (i.e., more misses) over
2235 * mutliple TLB flushes (i.e., more flushes). For virtual machines the
2236 * hypervisor needs to synchronize the host IOMMU PTEs with those of
2237 * the guest, and the trade-off is different: unnecessary TLB flushes
2238 * should be avoided.
2240 if (amd_iommu_np_cache &&
2241 iommu_iotlb_gather_is_disjoint(gather, iova, size))
2242 iommu_iotlb_sync(domain, gather);
2244 iommu_iotlb_gather_add_range(gather, iova, size);
2247 static size_t amd_iommu_unmap_pages(struct iommu_domain *dom, unsigned long iova,
2248 size_t pgsize, size_t pgcount,
2249 struct iommu_iotlb_gather *gather)
2251 struct protection_domain *domain = to_pdomain(dom);
2252 struct io_pgtable_ops *ops = &domain->iop.iop.ops;
2255 if ((amd_iommu_pgtable == AMD_IOMMU_V1) &&
2256 (domain->iop.mode == PAGE_MODE_NONE))
2259 r = (ops->unmap_pages) ? ops->unmap_pages(ops, iova, pgsize, pgcount, NULL) : 0;
2262 amd_iommu_iotlb_gather_add_page(dom, gather, iova, r);
2267 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
2270 struct protection_domain *domain = to_pdomain(dom);
2271 struct io_pgtable_ops *ops = &domain->iop.iop.ops;
2273 return ops->iova_to_phys(ops, iova);
2276 static bool amd_iommu_capable(struct device *dev, enum iommu_cap cap)
2279 case IOMMU_CAP_CACHE_COHERENCY:
2281 case IOMMU_CAP_INTR_REMAP:
2282 return (irq_remapping_enabled == 1);
2283 case IOMMU_CAP_NOEXEC:
2285 case IOMMU_CAP_PRE_BOOT_PROTECTION:
2286 return amdr_ivrs_remap_support;
2287 case IOMMU_CAP_ENFORCE_CACHE_COHERENCY:
2296 static void amd_iommu_get_resv_regions(struct device *dev,
2297 struct list_head *head)
2299 struct iommu_resv_region *region;
2300 struct unity_map_entry *entry;
2301 struct amd_iommu *iommu;
2302 struct amd_iommu_pci_seg *pci_seg;
2305 sbdf = get_device_sbdf_id(dev);
2309 devid = PCI_SBDF_TO_DEVID(sbdf);
2310 iommu = rlookup_amd_iommu(dev);
2313 pci_seg = iommu->pci_seg;
2315 list_for_each_entry(entry, &pci_seg->unity_map, list) {
2319 if (devid < entry->devid_start || devid > entry->devid_end)
2322 type = IOMMU_RESV_DIRECT;
2323 length = entry->address_end - entry->address_start;
2324 if (entry->prot & IOMMU_PROT_IR)
2326 if (entry->prot & IOMMU_PROT_IW)
2327 prot |= IOMMU_WRITE;
2328 if (entry->prot & IOMMU_UNITY_MAP_FLAG_EXCL_RANGE)
2329 /* Exclusion range */
2330 type = IOMMU_RESV_RESERVED;
2332 region = iommu_alloc_resv_region(entry->address_start,
2336 dev_err(dev, "Out of memory allocating dm-regions\n");
2339 list_add_tail(®ion->list, head);
2342 region = iommu_alloc_resv_region(MSI_RANGE_START,
2343 MSI_RANGE_END - MSI_RANGE_START + 1,
2344 0, IOMMU_RESV_MSI, GFP_KERNEL);
2347 list_add_tail(®ion->list, head);
2349 region = iommu_alloc_resv_region(HT_RANGE_START,
2350 HT_RANGE_END - HT_RANGE_START + 1,
2351 0, IOMMU_RESV_RESERVED, GFP_KERNEL);
2354 list_add_tail(®ion->list, head);
2357 bool amd_iommu_is_attach_deferred(struct device *dev)
2359 struct iommu_dev_data *dev_data = dev_iommu_priv_get(dev);
2361 return dev_data->defer_attach;
2363 EXPORT_SYMBOL_GPL(amd_iommu_is_attach_deferred);
2365 static void amd_iommu_flush_iotlb_all(struct iommu_domain *domain)
2367 struct protection_domain *dom = to_pdomain(domain);
2368 unsigned long flags;
2370 spin_lock_irqsave(&dom->lock, flags);
2371 amd_iommu_domain_flush_tlb_pde(dom);
2372 amd_iommu_domain_flush_complete(dom);
2373 spin_unlock_irqrestore(&dom->lock, flags);
2376 static void amd_iommu_iotlb_sync(struct iommu_domain *domain,
2377 struct iommu_iotlb_gather *gather)
2379 struct protection_domain *dom = to_pdomain(domain);
2380 unsigned long flags;
2382 spin_lock_irqsave(&dom->lock, flags);
2383 domain_flush_pages(dom, gather->start, gather->end - gather->start, 1);
2384 amd_iommu_domain_flush_complete(dom);
2385 spin_unlock_irqrestore(&dom->lock, flags);
2388 static int amd_iommu_def_domain_type(struct device *dev)
2390 struct iommu_dev_data *dev_data;
2392 dev_data = dev_iommu_priv_get(dev);
2397 * Do not identity map IOMMUv2 capable devices when:
2398 * - memory encryption is active, because some of those devices
2399 * (AMD GPUs) don't have the encryption bit in their DMA-mask
2400 * and require remapping.
2401 * - SNP is enabled, because it prohibits DTE[Mode]=0.
2403 if (dev_data->iommu_v2 &&
2404 !cc_platform_has(CC_ATTR_MEM_ENCRYPT) &&
2405 !amd_iommu_snp_en) {
2406 return IOMMU_DOMAIN_IDENTITY;
2412 static bool amd_iommu_enforce_cache_coherency(struct iommu_domain *domain)
2414 /* IOMMU_PTE_FC is always set */
2418 const struct iommu_ops amd_iommu_ops = {
2419 .capable = amd_iommu_capable,
2420 .domain_alloc = amd_iommu_domain_alloc,
2421 .probe_device = amd_iommu_probe_device,
2422 .release_device = amd_iommu_release_device,
2423 .probe_finalize = amd_iommu_probe_finalize,
2424 .device_group = amd_iommu_device_group,
2425 .get_resv_regions = amd_iommu_get_resv_regions,
2426 .is_attach_deferred = amd_iommu_is_attach_deferred,
2427 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
2428 .def_domain_type = amd_iommu_def_domain_type,
2429 .default_domain_ops = &(const struct iommu_domain_ops) {
2430 .attach_dev = amd_iommu_attach_device,
2431 .map_pages = amd_iommu_map_pages,
2432 .unmap_pages = amd_iommu_unmap_pages,
2433 .iotlb_sync_map = amd_iommu_iotlb_sync_map,
2434 .iova_to_phys = amd_iommu_iova_to_phys,
2435 .flush_iotlb_all = amd_iommu_flush_iotlb_all,
2436 .iotlb_sync = amd_iommu_iotlb_sync,
2437 .free = amd_iommu_domain_free,
2438 .enforce_cache_coherency = amd_iommu_enforce_cache_coherency,
2442 /*****************************************************************************
2444 * The next functions do a basic initialization of IOMMU for pass through
2447 * In passthrough mode the IOMMU is initialized and enabled but not used for
2448 * DMA-API translation.
2450 *****************************************************************************/
2452 /* IOMMUv2 specific functions */
2453 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
2455 return atomic_notifier_chain_register(&ppr_notifier, nb);
2457 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
2459 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
2461 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
2463 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
2465 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
2467 struct protection_domain *domain = to_pdomain(dom);
2468 unsigned long flags;
2470 spin_lock_irqsave(&domain->lock, flags);
2472 if (domain->iop.pgtbl_cfg.tlb)
2473 free_io_pgtable_ops(&domain->iop.iop.ops);
2475 spin_unlock_irqrestore(&domain->lock, flags);
2477 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
2479 /* Note: This function expects iommu_domain->lock to be held prior calling the function. */
2480 static int domain_enable_v2(struct protection_domain *domain, int pasids)
2484 /* Number of GCR3 table levels required */
2485 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
2488 if (levels > amd_iommu_max_glx_val)
2491 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
2492 if (domain->gcr3_tbl == NULL)
2495 domain->glx = levels;
2496 domain->flags |= PD_IOMMUV2_MASK;
2498 amd_iommu_domain_update(domain);
2503 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
2505 struct protection_domain *pdom = to_pdomain(dom);
2506 unsigned long flags;
2509 spin_lock_irqsave(&pdom->lock, flags);
2512 * Save us all sanity checks whether devices already in the
2513 * domain support IOMMUv2. Just force that the domain has no
2514 * devices attached when it is switched into IOMMUv2 mode.
2517 if (pdom->dev_cnt > 0 || pdom->flags & PD_IOMMUV2_MASK)
2520 if (!pdom->gcr3_tbl)
2521 ret = domain_enable_v2(pdom, pasids);
2524 spin_unlock_irqrestore(&pdom->lock, flags);
2527 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
2529 static int __flush_pasid(struct protection_domain *domain, u32 pasid,
2530 u64 address, bool size)
2532 struct iommu_dev_data *dev_data;
2533 struct iommu_cmd cmd;
2536 if (!(domain->flags & PD_IOMMUV2_MASK))
2539 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
2542 * IOMMU TLB needs to be flushed before Device TLB to
2543 * prevent device TLB refill from IOMMU TLB
2545 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
2546 if (domain->dev_iommu[i] == 0)
2549 ret = iommu_queue_command(amd_iommus[i], &cmd);
2554 /* Wait until IOMMU TLB flushes are complete */
2555 amd_iommu_domain_flush_complete(domain);
2557 /* Now flush device TLBs */
2558 list_for_each_entry(dev_data, &domain->dev_list, list) {
2559 struct amd_iommu *iommu;
2563 There might be non-IOMMUv2 capable devices in an IOMMUv2
2566 if (!dev_data->ats.enabled)
2569 qdep = dev_data->ats.qdep;
2570 iommu = rlookup_amd_iommu(dev_data->dev);
2573 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
2574 qdep, address, size);
2576 ret = iommu_queue_command(iommu, &cmd);
2581 /* Wait until all device TLBs are flushed */
2582 amd_iommu_domain_flush_complete(domain);
2591 static int __amd_iommu_flush_page(struct protection_domain *domain, u32 pasid,
2594 return __flush_pasid(domain, pasid, address, false);
2597 int amd_iommu_flush_page(struct iommu_domain *dom, u32 pasid,
2600 struct protection_domain *domain = to_pdomain(dom);
2601 unsigned long flags;
2604 spin_lock_irqsave(&domain->lock, flags);
2605 ret = __amd_iommu_flush_page(domain, pasid, address);
2606 spin_unlock_irqrestore(&domain->lock, flags);
2610 EXPORT_SYMBOL(amd_iommu_flush_page);
2612 static int __amd_iommu_flush_tlb(struct protection_domain *domain, u32 pasid)
2614 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
2618 int amd_iommu_flush_tlb(struct iommu_domain *dom, u32 pasid)
2620 struct protection_domain *domain = to_pdomain(dom);
2621 unsigned long flags;
2624 spin_lock_irqsave(&domain->lock, flags);
2625 ret = __amd_iommu_flush_tlb(domain, pasid);
2626 spin_unlock_irqrestore(&domain->lock, flags);
2630 EXPORT_SYMBOL(amd_iommu_flush_tlb);
2632 static u64 *__get_gcr3_pte(u64 *root, int level, u32 pasid, bool alloc)
2639 index = (pasid >> (9 * level)) & 0x1ff;
2645 if (!(*pte & GCR3_VALID)) {
2649 root = (void *)get_zeroed_page(GFP_ATOMIC);
2653 *pte = iommu_virt_to_phys(root) | GCR3_VALID;
2656 root = iommu_phys_to_virt(*pte & PAGE_MASK);
2664 static int __set_gcr3(struct protection_domain *domain, u32 pasid,
2669 if (domain->iop.mode != PAGE_MODE_NONE)
2672 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
2676 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
2678 return __amd_iommu_flush_tlb(domain, pasid);
2681 static int __clear_gcr3(struct protection_domain *domain, u32 pasid)
2685 if (domain->iop.mode != PAGE_MODE_NONE)
2688 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
2694 return __amd_iommu_flush_tlb(domain, pasid);
2697 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, u32 pasid,
2700 struct protection_domain *domain = to_pdomain(dom);
2701 unsigned long flags;
2704 spin_lock_irqsave(&domain->lock, flags);
2705 ret = __set_gcr3(domain, pasid, cr3);
2706 spin_unlock_irqrestore(&domain->lock, flags);
2710 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
2712 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, u32 pasid)
2714 struct protection_domain *domain = to_pdomain(dom);
2715 unsigned long flags;
2718 spin_lock_irqsave(&domain->lock, flags);
2719 ret = __clear_gcr3(domain, pasid);
2720 spin_unlock_irqrestore(&domain->lock, flags);
2724 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
2726 int amd_iommu_complete_ppr(struct pci_dev *pdev, u32 pasid,
2727 int status, int tag)
2729 struct iommu_dev_data *dev_data;
2730 struct amd_iommu *iommu;
2731 struct iommu_cmd cmd;
2733 dev_data = dev_iommu_priv_get(&pdev->dev);
2734 iommu = rlookup_amd_iommu(&pdev->dev);
2738 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
2739 tag, dev_data->pri_tlp);
2741 return iommu_queue_command(iommu, &cmd);
2743 EXPORT_SYMBOL(amd_iommu_complete_ppr);
2745 int amd_iommu_device_info(struct pci_dev *pdev,
2746 struct amd_iommu_device_info *info)
2751 if (pdev == NULL || info == NULL)
2754 if (!amd_iommu_v2_supported())
2757 memset(info, 0, sizeof(*info));
2759 if (pci_ats_supported(pdev))
2760 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
2762 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2764 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
2766 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
2770 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
2771 max_pasids = min(max_pasids, (1 << 20));
2773 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
2774 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
2776 features = pci_pasid_features(pdev);
2777 if (features & PCI_PASID_CAP_EXEC)
2778 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
2779 if (features & PCI_PASID_CAP_PRIV)
2780 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
2785 EXPORT_SYMBOL(amd_iommu_device_info);
2787 #ifdef CONFIG_IRQ_REMAP
2789 /*****************************************************************************
2791 * Interrupt Remapping Implementation
2793 *****************************************************************************/
2795 static struct irq_chip amd_ir_chip;
2796 static DEFINE_SPINLOCK(iommu_table_lock);
2798 static void set_dte_irq_entry(struct amd_iommu *iommu, u16 devid,
2799 struct irq_remap_table *table)
2802 struct dev_table_entry *dev_table = get_dev_table(iommu);
2804 dte = dev_table[devid].data[2];
2805 dte &= ~DTE_IRQ_PHYS_ADDR_MASK;
2806 dte |= iommu_virt_to_phys(table->table);
2807 dte |= DTE_IRQ_REMAP_INTCTL;
2808 dte |= DTE_INTTABLEN;
2809 dte |= DTE_IRQ_REMAP_ENABLE;
2811 dev_table[devid].data[2] = dte;
2814 static struct irq_remap_table *get_irq_table(struct amd_iommu *iommu, u16 devid)
2816 struct irq_remap_table *table;
2817 struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
2819 if (WARN_ONCE(!pci_seg->rlookup_table[devid],
2820 "%s: no iommu for devid %x:%x\n",
2821 __func__, pci_seg->id, devid))
2824 table = pci_seg->irq_lookup_table[devid];
2825 if (WARN_ONCE(!table, "%s: no table for devid %x:%x\n",
2826 __func__, pci_seg->id, devid))
2832 static struct irq_remap_table *__alloc_irq_table(void)
2834 struct irq_remap_table *table;
2836 table = kzalloc(sizeof(*table), GFP_KERNEL);
2840 table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_KERNEL);
2841 if (!table->table) {
2845 raw_spin_lock_init(&table->lock);
2847 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
2848 memset(table->table, 0,
2849 MAX_IRQS_PER_TABLE * sizeof(u32));
2851 memset(table->table, 0,
2852 (MAX_IRQS_PER_TABLE * (sizeof(u64) * 2)));
2856 static void set_remap_table_entry(struct amd_iommu *iommu, u16 devid,
2857 struct irq_remap_table *table)
2859 struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
2861 pci_seg->irq_lookup_table[devid] = table;
2862 set_dte_irq_entry(iommu, devid, table);
2863 iommu_flush_dte(iommu, devid);
2866 static int set_remap_table_entry_alias(struct pci_dev *pdev, u16 alias,
2869 struct irq_remap_table *table = data;
2870 struct amd_iommu_pci_seg *pci_seg;
2871 struct amd_iommu *iommu = rlookup_amd_iommu(&pdev->dev);
2876 pci_seg = iommu->pci_seg;
2877 pci_seg->irq_lookup_table[alias] = table;
2878 set_dte_irq_entry(iommu, alias, table);
2879 iommu_flush_dte(pci_seg->rlookup_table[alias], alias);
2884 static struct irq_remap_table *alloc_irq_table(struct amd_iommu *iommu,
2885 u16 devid, struct pci_dev *pdev)
2887 struct irq_remap_table *table = NULL;
2888 struct irq_remap_table *new_table = NULL;
2889 struct amd_iommu_pci_seg *pci_seg;
2890 unsigned long flags;
2893 spin_lock_irqsave(&iommu_table_lock, flags);
2895 pci_seg = iommu->pci_seg;
2896 table = pci_seg->irq_lookup_table[devid];
2900 alias = pci_seg->alias_table[devid];
2901 table = pci_seg->irq_lookup_table[alias];
2903 set_remap_table_entry(iommu, devid, table);
2906 spin_unlock_irqrestore(&iommu_table_lock, flags);
2908 /* Nothing there yet, allocate new irq remapping table */
2909 new_table = __alloc_irq_table();
2913 spin_lock_irqsave(&iommu_table_lock, flags);
2915 table = pci_seg->irq_lookup_table[devid];
2919 table = pci_seg->irq_lookup_table[alias];
2921 set_remap_table_entry(iommu, devid, table);
2929 pci_for_each_dma_alias(pdev, set_remap_table_entry_alias,
2932 set_remap_table_entry(iommu, devid, table);
2935 set_remap_table_entry(iommu, alias, table);
2938 iommu_completion_wait(iommu);
2941 spin_unlock_irqrestore(&iommu_table_lock, flags);
2944 kmem_cache_free(amd_iommu_irq_cache, new_table->table);
2950 static int alloc_irq_index(struct amd_iommu *iommu, u16 devid, int count,
2951 bool align, struct pci_dev *pdev)
2953 struct irq_remap_table *table;
2954 int index, c, alignment = 1;
2955 unsigned long flags;
2957 table = alloc_irq_table(iommu, devid, pdev);
2962 alignment = roundup_pow_of_two(count);
2964 raw_spin_lock_irqsave(&table->lock, flags);
2966 /* Scan table for free entries */
2967 for (index = ALIGN(table->min_index, alignment), c = 0;
2968 index < MAX_IRQS_PER_TABLE;) {
2969 if (!iommu->irte_ops->is_allocated(table, index)) {
2973 index = ALIGN(index + 1, alignment);
2979 iommu->irte_ops->set_allocated(table, index - c + 1);
2991 raw_spin_unlock_irqrestore(&table->lock, flags);
2996 static int modify_irte_ga(struct amd_iommu *iommu, u16 devid, int index,
2997 struct irte_ga *irte, struct amd_ir_data *data)
3000 struct irq_remap_table *table;
3001 unsigned long flags;
3002 struct irte_ga *entry;
3004 table = get_irq_table(iommu, devid);
3008 raw_spin_lock_irqsave(&table->lock, flags);
3010 entry = (struct irte_ga *)table->table;
3011 entry = &entry[index];
3013 ret = cmpxchg_double(&entry->lo.val, &entry->hi.val,
3014 entry->lo.val, entry->hi.val,
3015 irte->lo.val, irte->hi.val);
3017 * We use cmpxchg16 to atomically update the 128-bit IRTE,
3018 * and it cannot be updated by the hardware or other processors
3019 * behind us, so the return value of cmpxchg16 should be the
3020 * same as the old value.
3027 raw_spin_unlock_irqrestore(&table->lock, flags);
3029 iommu_flush_irt(iommu, devid);
3030 iommu_completion_wait(iommu);
3035 static int modify_irte(struct amd_iommu *iommu,
3036 u16 devid, int index, union irte *irte)
3038 struct irq_remap_table *table;
3039 unsigned long flags;
3041 table = get_irq_table(iommu, devid);
3045 raw_spin_lock_irqsave(&table->lock, flags);
3046 table->table[index] = irte->val;
3047 raw_spin_unlock_irqrestore(&table->lock, flags);
3049 iommu_flush_irt(iommu, devid);
3050 iommu_completion_wait(iommu);
3055 static void free_irte(struct amd_iommu *iommu, u16 devid, int index)
3057 struct irq_remap_table *table;
3058 unsigned long flags;
3060 table = get_irq_table(iommu, devid);
3064 raw_spin_lock_irqsave(&table->lock, flags);
3065 iommu->irte_ops->clear_allocated(table, index);
3066 raw_spin_unlock_irqrestore(&table->lock, flags);
3068 iommu_flush_irt(iommu, devid);
3069 iommu_completion_wait(iommu);
3072 static void irte_prepare(void *entry,
3073 u32 delivery_mode, bool dest_mode,
3074 u8 vector, u32 dest_apicid, int devid)
3076 union irte *irte = (union irte *) entry;
3079 irte->fields.vector = vector;
3080 irte->fields.int_type = delivery_mode;
3081 irte->fields.destination = dest_apicid;
3082 irte->fields.dm = dest_mode;
3083 irte->fields.valid = 1;
3086 static void irte_ga_prepare(void *entry,
3087 u32 delivery_mode, bool dest_mode,
3088 u8 vector, u32 dest_apicid, int devid)
3090 struct irte_ga *irte = (struct irte_ga *) entry;
3094 irte->lo.fields_remap.int_type = delivery_mode;
3095 irte->lo.fields_remap.dm = dest_mode;
3096 irte->hi.fields.vector = vector;
3097 irte->lo.fields_remap.destination = APICID_TO_IRTE_DEST_LO(dest_apicid);
3098 irte->hi.fields.destination = APICID_TO_IRTE_DEST_HI(dest_apicid);
3099 irte->lo.fields_remap.valid = 1;
3102 static void irte_activate(struct amd_iommu *iommu, void *entry, u16 devid, u16 index)
3104 union irte *irte = (union irte *) entry;
3106 irte->fields.valid = 1;
3107 modify_irte(iommu, devid, index, irte);
3110 static void irte_ga_activate(struct amd_iommu *iommu, void *entry, u16 devid, u16 index)
3112 struct irte_ga *irte = (struct irte_ga *) entry;
3114 irte->lo.fields_remap.valid = 1;
3115 modify_irte_ga(iommu, devid, index, irte, NULL);
3118 static void irte_deactivate(struct amd_iommu *iommu, void *entry, u16 devid, u16 index)
3120 union irte *irte = (union irte *) entry;
3122 irte->fields.valid = 0;
3123 modify_irte(iommu, devid, index, irte);
3126 static void irte_ga_deactivate(struct amd_iommu *iommu, void *entry, u16 devid, u16 index)
3128 struct irte_ga *irte = (struct irte_ga *) entry;
3130 irte->lo.fields_remap.valid = 0;
3131 modify_irte_ga(iommu, devid, index, irte, NULL);
3134 static void irte_set_affinity(struct amd_iommu *iommu, void *entry, u16 devid, u16 index,
3135 u8 vector, u32 dest_apicid)
3137 union irte *irte = (union irte *) entry;
3139 irte->fields.vector = vector;
3140 irte->fields.destination = dest_apicid;
3141 modify_irte(iommu, devid, index, irte);
3144 static void irte_ga_set_affinity(struct amd_iommu *iommu, void *entry, u16 devid, u16 index,
3145 u8 vector, u32 dest_apicid)
3147 struct irte_ga *irte = (struct irte_ga *) entry;
3149 if (!irte->lo.fields_remap.guest_mode) {
3150 irte->hi.fields.vector = vector;
3151 irte->lo.fields_remap.destination =
3152 APICID_TO_IRTE_DEST_LO(dest_apicid);
3153 irte->hi.fields.destination =
3154 APICID_TO_IRTE_DEST_HI(dest_apicid);
3155 modify_irte_ga(iommu, devid, index, irte, NULL);
3159 #define IRTE_ALLOCATED (~1U)
3160 static void irte_set_allocated(struct irq_remap_table *table, int index)
3162 table->table[index] = IRTE_ALLOCATED;
3165 static void irte_ga_set_allocated(struct irq_remap_table *table, int index)
3167 struct irte_ga *ptr = (struct irte_ga *)table->table;
3168 struct irte_ga *irte = &ptr[index];
3170 memset(&irte->lo.val, 0, sizeof(u64));
3171 memset(&irte->hi.val, 0, sizeof(u64));
3172 irte->hi.fields.vector = 0xff;
3175 static bool irte_is_allocated(struct irq_remap_table *table, int index)
3177 union irte *ptr = (union irte *)table->table;
3178 union irte *irte = &ptr[index];
3180 return irte->val != 0;
3183 static bool irte_ga_is_allocated(struct irq_remap_table *table, int index)
3185 struct irte_ga *ptr = (struct irte_ga *)table->table;
3186 struct irte_ga *irte = &ptr[index];
3188 return irte->hi.fields.vector != 0;
3191 static void irte_clear_allocated(struct irq_remap_table *table, int index)
3193 table->table[index] = 0;
3196 static void irte_ga_clear_allocated(struct irq_remap_table *table, int index)
3198 struct irte_ga *ptr = (struct irte_ga *)table->table;
3199 struct irte_ga *irte = &ptr[index];
3201 memset(&irte->lo.val, 0, sizeof(u64));
3202 memset(&irte->hi.val, 0, sizeof(u64));
3205 static int get_devid(struct irq_alloc_info *info)
3207 switch (info->type) {
3208 case X86_IRQ_ALLOC_TYPE_IOAPIC:
3209 return get_ioapic_devid(info->devid);
3210 case X86_IRQ_ALLOC_TYPE_HPET:
3211 return get_hpet_devid(info->devid);
3212 case X86_IRQ_ALLOC_TYPE_PCI_MSI:
3213 case X86_IRQ_ALLOC_TYPE_PCI_MSIX:
3214 return get_device_sbdf_id(msi_desc_to_dev(info->desc));
3221 struct irq_remap_ops amd_iommu_irq_ops = {
3222 .prepare = amd_iommu_prepare,
3223 .enable = amd_iommu_enable,
3224 .disable = amd_iommu_disable,
3225 .reenable = amd_iommu_reenable,
3226 .enable_faulting = amd_iommu_enable_faulting,
3229 static void fill_msi_msg(struct msi_msg *msg, u32 index)
3232 msg->address_lo = 0;
3233 msg->arch_addr_lo.base_address = X86_MSI_BASE_ADDRESS_LOW;
3234 msg->address_hi = X86_MSI_BASE_ADDRESS_HIGH;
3237 static void irq_remapping_prepare_irte(struct amd_ir_data *data,
3238 struct irq_cfg *irq_cfg,
3239 struct irq_alloc_info *info,
3240 int devid, int index, int sub_handle)
3242 struct irq_2_irte *irte_info = &data->irq_2_irte;
3243 struct amd_iommu *iommu = data->iommu;
3248 data->irq_2_irte.devid = devid;
3249 data->irq_2_irte.index = index + sub_handle;
3250 iommu->irte_ops->prepare(data->entry, apic->delivery_mode,
3251 apic->dest_mode_logical, irq_cfg->vector,
3252 irq_cfg->dest_apicid, devid);
3254 switch (info->type) {
3255 case X86_IRQ_ALLOC_TYPE_IOAPIC:
3256 case X86_IRQ_ALLOC_TYPE_HPET:
3257 case X86_IRQ_ALLOC_TYPE_PCI_MSI:
3258 case X86_IRQ_ALLOC_TYPE_PCI_MSIX:
3259 fill_msi_msg(&data->msi_entry, irte_info->index);
3268 struct amd_irte_ops irte_32_ops = {
3269 .prepare = irte_prepare,
3270 .activate = irte_activate,
3271 .deactivate = irte_deactivate,
3272 .set_affinity = irte_set_affinity,
3273 .set_allocated = irte_set_allocated,
3274 .is_allocated = irte_is_allocated,
3275 .clear_allocated = irte_clear_allocated,
3278 struct amd_irte_ops irte_128_ops = {
3279 .prepare = irte_ga_prepare,
3280 .activate = irte_ga_activate,
3281 .deactivate = irte_ga_deactivate,
3282 .set_affinity = irte_ga_set_affinity,
3283 .set_allocated = irte_ga_set_allocated,
3284 .is_allocated = irte_ga_is_allocated,
3285 .clear_allocated = irte_ga_clear_allocated,
3288 static int irq_remapping_alloc(struct irq_domain *domain, unsigned int virq,
3289 unsigned int nr_irqs, void *arg)
3291 struct irq_alloc_info *info = arg;
3292 struct irq_data *irq_data;
3293 struct amd_ir_data *data = NULL;
3294 struct amd_iommu *iommu;
3295 struct irq_cfg *cfg;
3296 int i, ret, devid, seg, sbdf;
3301 if (nr_irqs > 1 && info->type != X86_IRQ_ALLOC_TYPE_PCI_MSI)
3304 sbdf = get_devid(info);
3308 seg = PCI_SBDF_TO_SEGID(sbdf);
3309 devid = PCI_SBDF_TO_DEVID(sbdf);
3310 iommu = __rlookup_amd_iommu(seg, devid);
3314 ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
3318 if (info->type == X86_IRQ_ALLOC_TYPE_IOAPIC) {
3319 struct irq_remap_table *table;
3321 table = alloc_irq_table(iommu, devid, NULL);
3323 if (!table->min_index) {
3325 * Keep the first 32 indexes free for IOAPIC
3328 table->min_index = 32;
3329 for (i = 0; i < 32; ++i)
3330 iommu->irte_ops->set_allocated(table, i);
3332 WARN_ON(table->min_index != 32);
3333 index = info->ioapic.pin;
3337 } else if (info->type == X86_IRQ_ALLOC_TYPE_PCI_MSI ||
3338 info->type == X86_IRQ_ALLOC_TYPE_PCI_MSIX) {
3339 bool align = (info->type == X86_IRQ_ALLOC_TYPE_PCI_MSI);
3341 index = alloc_irq_index(iommu, devid, nr_irqs, align,
3342 msi_desc_to_pci_dev(info->desc));
3344 index = alloc_irq_index(iommu, devid, nr_irqs, false, NULL);
3348 pr_warn("Failed to allocate IRTE\n");
3350 goto out_free_parent;
3353 for (i = 0; i < nr_irqs; i++) {
3354 irq_data = irq_domain_get_irq_data(domain, virq + i);
3355 cfg = irq_data ? irqd_cfg(irq_data) : NULL;
3362 data = kzalloc(sizeof(*data), GFP_KERNEL);
3366 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
3367 data->entry = kzalloc(sizeof(union irte), GFP_KERNEL);
3369 data->entry = kzalloc(sizeof(struct irte_ga),
3376 data->iommu = iommu;
3377 irq_data->hwirq = (devid << 16) + i;
3378 irq_data->chip_data = data;
3379 irq_data->chip = &amd_ir_chip;
3380 irq_remapping_prepare_irte(data, cfg, info, devid, index, i);
3381 irq_set_status_flags(virq + i, IRQ_MOVE_PCNTXT);
3387 for (i--; i >= 0; i--) {
3388 irq_data = irq_domain_get_irq_data(domain, virq + i);
3390 kfree(irq_data->chip_data);
3392 for (i = 0; i < nr_irqs; i++)
3393 free_irte(iommu, devid, index + i);
3395 irq_domain_free_irqs_common(domain, virq, nr_irqs);
3399 static void irq_remapping_free(struct irq_domain *domain, unsigned int virq,
3400 unsigned int nr_irqs)
3402 struct irq_2_irte *irte_info;
3403 struct irq_data *irq_data;
3404 struct amd_ir_data *data;
3407 for (i = 0; i < nr_irqs; i++) {
3408 irq_data = irq_domain_get_irq_data(domain, virq + i);
3409 if (irq_data && irq_data->chip_data) {
3410 data = irq_data->chip_data;
3411 irte_info = &data->irq_2_irte;
3412 free_irte(data->iommu, irte_info->devid, irte_info->index);
3417 irq_domain_free_irqs_common(domain, virq, nr_irqs);
3420 static void amd_ir_update_irte(struct irq_data *irqd, struct amd_iommu *iommu,
3421 struct amd_ir_data *ir_data,
3422 struct irq_2_irte *irte_info,
3423 struct irq_cfg *cfg);
3425 static int irq_remapping_activate(struct irq_domain *domain,
3426 struct irq_data *irq_data, bool reserve)
3428 struct amd_ir_data *data = irq_data->chip_data;
3429 struct irq_2_irte *irte_info = &data->irq_2_irte;
3430 struct amd_iommu *iommu = data->iommu;
3431 struct irq_cfg *cfg = irqd_cfg(irq_data);
3436 iommu->irte_ops->activate(iommu, data->entry, irte_info->devid,
3438 amd_ir_update_irte(irq_data, iommu, data, irte_info, cfg);
3442 static void irq_remapping_deactivate(struct irq_domain *domain,
3443 struct irq_data *irq_data)
3445 struct amd_ir_data *data = irq_data->chip_data;
3446 struct irq_2_irte *irte_info = &data->irq_2_irte;
3447 struct amd_iommu *iommu = data->iommu;
3450 iommu->irte_ops->deactivate(iommu, data->entry, irte_info->devid,
3454 static int irq_remapping_select(struct irq_domain *d, struct irq_fwspec *fwspec,
3455 enum irq_domain_bus_token bus_token)
3457 struct amd_iommu *iommu;
3460 if (!amd_iommu_irq_remap)
3463 if (x86_fwspec_is_ioapic(fwspec))
3464 devid = get_ioapic_devid(fwspec->param[0]);
3465 else if (x86_fwspec_is_hpet(fwspec))
3466 devid = get_hpet_devid(fwspec->param[0]);
3470 iommu = __rlookup_amd_iommu((devid >> 16), (devid & 0xffff));
3472 return iommu && iommu->ir_domain == d;
3475 static const struct irq_domain_ops amd_ir_domain_ops = {
3476 .select = irq_remapping_select,
3477 .alloc = irq_remapping_alloc,
3478 .free = irq_remapping_free,
3479 .activate = irq_remapping_activate,
3480 .deactivate = irq_remapping_deactivate,
3483 int amd_iommu_activate_guest_mode(void *data)
3485 struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
3486 struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
3489 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
3490 !entry || entry->lo.fields_vapic.guest_mode)
3493 valid = entry->lo.fields_vapic.valid;
3498 entry->lo.fields_vapic.valid = valid;
3499 entry->lo.fields_vapic.guest_mode = 1;
3500 entry->lo.fields_vapic.ga_log_intr = 1;
3501 entry->hi.fields.ga_root_ptr = ir_data->ga_root_ptr;
3502 entry->hi.fields.vector = ir_data->ga_vector;
3503 entry->lo.fields_vapic.ga_tag = ir_data->ga_tag;
3505 return modify_irte_ga(ir_data->iommu, ir_data->irq_2_irte.devid,
3506 ir_data->irq_2_irte.index, entry, ir_data);
3508 EXPORT_SYMBOL(amd_iommu_activate_guest_mode);
3510 int amd_iommu_deactivate_guest_mode(void *data)
3512 struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
3513 struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
3514 struct irq_cfg *cfg = ir_data->cfg;
3517 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
3518 !entry || !entry->lo.fields_vapic.guest_mode)
3521 valid = entry->lo.fields_remap.valid;
3526 entry->lo.fields_remap.valid = valid;
3527 entry->lo.fields_remap.dm = apic->dest_mode_logical;
3528 entry->lo.fields_remap.int_type = apic->delivery_mode;
3529 entry->hi.fields.vector = cfg->vector;
3530 entry->lo.fields_remap.destination =
3531 APICID_TO_IRTE_DEST_LO(cfg->dest_apicid);
3532 entry->hi.fields.destination =
3533 APICID_TO_IRTE_DEST_HI(cfg->dest_apicid);
3535 return modify_irte_ga(ir_data->iommu, ir_data->irq_2_irte.devid,
3536 ir_data->irq_2_irte.index, entry, ir_data);
3538 EXPORT_SYMBOL(amd_iommu_deactivate_guest_mode);
3540 static int amd_ir_set_vcpu_affinity(struct irq_data *data, void *vcpu_info)
3543 struct amd_iommu_pi_data *pi_data = vcpu_info;
3544 struct vcpu_data *vcpu_pi_info = pi_data->vcpu_data;
3545 struct amd_ir_data *ir_data = data->chip_data;
3546 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
3547 struct iommu_dev_data *dev_data;
3549 if (ir_data->iommu == NULL)
3552 dev_data = search_dev_data(ir_data->iommu, irte_info->devid);
3555 * This device has never been set up for guest mode.
3556 * we should not modify the IRTE
3558 if (!dev_data || !dev_data->use_vapic)
3561 ir_data->cfg = irqd_cfg(data);
3562 pi_data->ir_data = ir_data;
3565 * SVM tries to set up for VAPIC mode, but we are in
3566 * legacy mode. So, we force legacy mode instead.
3568 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
3569 pr_debug("%s: Fall back to using intr legacy remap\n",
3571 pi_data->is_guest_mode = false;
3574 pi_data->prev_ga_tag = ir_data->cached_ga_tag;
3575 if (pi_data->is_guest_mode) {
3576 ir_data->ga_root_ptr = (pi_data->base >> 12);
3577 ir_data->ga_vector = vcpu_pi_info->vector;
3578 ir_data->ga_tag = pi_data->ga_tag;
3579 ret = amd_iommu_activate_guest_mode(ir_data);
3581 ir_data->cached_ga_tag = pi_data->ga_tag;
3583 ret = amd_iommu_deactivate_guest_mode(ir_data);
3586 * This communicates the ga_tag back to the caller
3587 * so that it can do all the necessary clean up.
3590 ir_data->cached_ga_tag = 0;
3597 static void amd_ir_update_irte(struct irq_data *irqd, struct amd_iommu *iommu,
3598 struct amd_ir_data *ir_data,
3599 struct irq_2_irte *irte_info,
3600 struct irq_cfg *cfg)
3604 * Atomically updates the IRTE with the new destination, vector
3605 * and flushes the interrupt entry cache.
3607 iommu->irte_ops->set_affinity(iommu, ir_data->entry, irte_info->devid,
3608 irte_info->index, cfg->vector,
3612 static int amd_ir_set_affinity(struct irq_data *data,
3613 const struct cpumask *mask, bool force)
3615 struct amd_ir_data *ir_data = data->chip_data;
3616 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
3617 struct irq_cfg *cfg = irqd_cfg(data);
3618 struct irq_data *parent = data->parent_data;
3619 struct amd_iommu *iommu = ir_data->iommu;
3625 ret = parent->chip->irq_set_affinity(parent, mask, force);
3626 if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
3629 amd_ir_update_irte(data, iommu, ir_data, irte_info, cfg);
3631 * After this point, all the interrupts will start arriving
3632 * at the new destination. So, time to cleanup the previous
3633 * vector allocation.
3635 send_cleanup_vector(cfg);
3637 return IRQ_SET_MASK_OK_DONE;
3640 static void ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg)
3642 struct amd_ir_data *ir_data = irq_data->chip_data;
3644 *msg = ir_data->msi_entry;
3647 static struct irq_chip amd_ir_chip = {
3649 .irq_ack = apic_ack_irq,
3650 .irq_set_affinity = amd_ir_set_affinity,
3651 .irq_set_vcpu_affinity = amd_ir_set_vcpu_affinity,
3652 .irq_compose_msi_msg = ir_compose_msi_msg,
3655 static const struct msi_parent_ops amdvi_msi_parent_ops = {
3656 .supported_flags = X86_VECTOR_MSI_FLAGS_SUPPORTED |
3657 MSI_FLAG_MULTI_PCI_MSI |
3660 .init_dev_msi_info = msi_parent_init_dev_msi_info,
3663 static const struct msi_parent_ops virt_amdvi_msi_parent_ops = {
3664 .supported_flags = X86_VECTOR_MSI_FLAGS_SUPPORTED |
3665 MSI_FLAG_MULTI_PCI_MSI,
3667 .init_dev_msi_info = msi_parent_init_dev_msi_info,
3670 int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
3672 struct fwnode_handle *fn;
3674 fn = irq_domain_alloc_named_id_fwnode("AMD-IR", iommu->index);
3677 iommu->ir_domain = irq_domain_create_hierarchy(arch_get_ir_parent_domain(), 0, 0,
3678 fn, &amd_ir_domain_ops, iommu);
3679 if (!iommu->ir_domain) {
3680 irq_domain_free_fwnode(fn);
3684 irq_domain_update_bus_token(iommu->ir_domain, DOMAIN_BUS_AMDVI);
3685 iommu->ir_domain->flags |= IRQ_DOMAIN_FLAG_MSI_PARENT;
3687 if (amd_iommu_np_cache)
3688 iommu->ir_domain->msi_parent_ops = &virt_amdvi_msi_parent_ops;
3690 iommu->ir_domain->msi_parent_ops = &amdvi_msi_parent_ops;
3695 int amd_iommu_update_ga(int cpu, bool is_run, void *data)
3697 unsigned long flags;
3698 struct amd_iommu *iommu;
3699 struct irq_remap_table *table;
3700 struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
3701 int devid = ir_data->irq_2_irte.devid;
3702 struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
3703 struct irte_ga *ref = (struct irte_ga *) ir_data->ref;
3705 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
3706 !ref || !entry || !entry->lo.fields_vapic.guest_mode)
3709 iommu = ir_data->iommu;
3713 table = get_irq_table(iommu, devid);
3717 raw_spin_lock_irqsave(&table->lock, flags);
3719 if (ref->lo.fields_vapic.guest_mode) {
3721 ref->lo.fields_vapic.destination =
3722 APICID_TO_IRTE_DEST_LO(cpu);
3723 ref->hi.fields.destination =
3724 APICID_TO_IRTE_DEST_HI(cpu);
3726 ref->lo.fields_vapic.is_run = is_run;
3730 raw_spin_unlock_irqrestore(&table->lock, flags);
3732 iommu_flush_irt(iommu, devid);
3733 iommu_completion_wait(iommu);
3736 EXPORT_SYMBOL(amd_iommu_update_ga);