1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (C) 2007-2010 Advanced Micro Devices, Inc.
4 * Author: Joerg Roedel <jroedel@suse.de>
5 * Leo Duran <leo.duran@amd.com>
8 #define pr_fmt(fmt) "AMD-Vi: " fmt
9 #define dev_fmt(fmt) pr_fmt(fmt)
11 #include <linux/ratelimit.h>
12 #include <linux/pci.h>
13 #include <linux/acpi.h>
14 #include <linux/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 static void amd_iommu_handle_irq(void *data, const char *evt_type,
845 u32 int_mask, u32 overflow_mask,
846 void (*int_handler)(struct amd_iommu *),
847 void (*overflow_handler)(struct amd_iommu *))
849 struct amd_iommu *iommu = (struct amd_iommu *) data;
850 u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
851 u32 mask = int_mask | overflow_mask;
853 while (status & mask) {
854 /* Enable interrupt sources again */
855 writel(mask, iommu->mmio_base + MMIO_STATUS_OFFSET);
858 pr_devel("Processing IOMMU (ivhd%d) %s Log\n",
859 iommu->index, evt_type);
863 if ((status & overflow_mask) && overflow_handler)
864 overflow_handler(iommu);
867 * Hardware bug: ERBT1312
868 * When re-enabling interrupt (by writing 1
869 * to clear the bit), the hardware might also try to set
870 * the interrupt bit in the event status register.
871 * In this scenario, the bit will be set, and disable
872 * subsequent interrupts.
874 * Workaround: The IOMMU driver should read back the
875 * status register and check if the interrupt bits are cleared.
876 * If not, driver will need to go through the interrupt handler
877 * again and re-clear the bits
879 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
883 irqreturn_t amd_iommu_int_thread_evtlog(int irq, void *data)
885 amd_iommu_handle_irq(data, "Evt", MMIO_STATUS_EVT_INT_MASK,
886 MMIO_STATUS_EVT_OVERFLOW_MASK,
887 iommu_poll_events, amd_iommu_restart_event_logging);
892 irqreturn_t amd_iommu_int_thread_pprlog(int irq, void *data)
894 amd_iommu_handle_irq(data, "PPR", MMIO_STATUS_PPR_INT_MASK,
895 MMIO_STATUS_PPR_OVERFLOW_MASK,
896 iommu_poll_ppr_log, amd_iommu_restart_ppr_log);
901 irqreturn_t amd_iommu_int_thread_galog(int irq, void *data)
903 #ifdef CONFIG_IRQ_REMAP
904 amd_iommu_handle_irq(data, "GA", MMIO_STATUS_GALOG_INT_MASK,
905 MMIO_STATUS_GALOG_OVERFLOW_MASK,
906 iommu_poll_ga_log, amd_iommu_restart_ga_log);
912 irqreturn_t amd_iommu_int_thread(int irq, void *data)
914 amd_iommu_int_thread_evtlog(irq, data);
915 amd_iommu_int_thread_pprlog(irq, data);
916 amd_iommu_int_thread_galog(irq, data);
921 irqreturn_t amd_iommu_int_handler(int irq, void *data)
923 return IRQ_WAKE_THREAD;
926 /****************************************************************************
928 * IOMMU command queuing functions
930 ****************************************************************************/
932 static int wait_on_sem(struct amd_iommu *iommu, u64 data)
936 while (*iommu->cmd_sem != data && i < LOOP_TIMEOUT) {
941 if (i == LOOP_TIMEOUT) {
942 pr_alert("Completion-Wait loop timed out\n");
949 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
950 struct iommu_cmd *cmd)
955 /* Copy command to buffer */
956 tail = iommu->cmd_buf_tail;
957 target = iommu->cmd_buf + tail;
958 memcpy(target, cmd, sizeof(*cmd));
960 tail = (tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
961 iommu->cmd_buf_tail = tail;
963 /* Tell the IOMMU about it */
964 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
967 static void build_completion_wait(struct iommu_cmd *cmd,
968 struct amd_iommu *iommu,
971 u64 paddr = iommu_virt_to_phys((void *)iommu->cmd_sem);
973 memset(cmd, 0, sizeof(*cmd));
974 cmd->data[0] = lower_32_bits(paddr) | CMD_COMPL_WAIT_STORE_MASK;
975 cmd->data[1] = upper_32_bits(paddr);
976 cmd->data[2] = lower_32_bits(data);
977 cmd->data[3] = upper_32_bits(data);
978 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
981 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
983 memset(cmd, 0, sizeof(*cmd));
984 cmd->data[0] = devid;
985 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
989 * Builds an invalidation address which is suitable for one page or multiple
990 * pages. Sets the size bit (S) as needed is more than one page is flushed.
992 static inline u64 build_inv_address(u64 address, size_t size)
994 u64 pages, end, msb_diff;
996 pages = iommu_num_pages(address, size, PAGE_SIZE);
999 return address & PAGE_MASK;
1001 end = address + size - 1;
1004 * msb_diff would hold the index of the most significant bit that
1005 * flipped between the start and end.
1007 msb_diff = fls64(end ^ address) - 1;
1010 * Bits 63:52 are sign extended. If for some reason bit 51 is different
1011 * between the start and the end, invalidate everything.
1013 if (unlikely(msb_diff > 51)) {
1014 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
1017 * The msb-bit must be clear on the address. Just set all the
1020 address |= (1ull << msb_diff) - 1;
1023 /* Clear bits 11:0 */
1024 address &= PAGE_MASK;
1026 /* Set the size bit - we flush more than one 4kb page */
1027 return address | CMD_INV_IOMMU_PAGES_SIZE_MASK;
1030 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
1031 size_t size, u16 domid, int pde)
1033 u64 inv_address = build_inv_address(address, size);
1035 memset(cmd, 0, sizeof(*cmd));
1036 cmd->data[1] |= domid;
1037 cmd->data[2] = lower_32_bits(inv_address);
1038 cmd->data[3] = upper_32_bits(inv_address);
1039 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
1040 if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
1041 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
1044 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
1045 u64 address, size_t size)
1047 u64 inv_address = build_inv_address(address, size);
1049 memset(cmd, 0, sizeof(*cmd));
1050 cmd->data[0] = devid;
1051 cmd->data[0] |= (qdep & 0xff) << 24;
1052 cmd->data[1] = devid;
1053 cmd->data[2] = lower_32_bits(inv_address);
1054 cmd->data[3] = upper_32_bits(inv_address);
1055 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
1058 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, u32 pasid,
1059 u64 address, bool size)
1061 memset(cmd, 0, sizeof(*cmd));
1063 address &= ~(0xfffULL);
1065 cmd->data[0] = pasid;
1066 cmd->data[1] = domid;
1067 cmd->data[2] = lower_32_bits(address);
1068 cmd->data[3] = upper_32_bits(address);
1069 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
1070 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
1072 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
1073 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
1076 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, u32 pasid,
1077 int qdep, u64 address, bool size)
1079 memset(cmd, 0, sizeof(*cmd));
1081 address &= ~(0xfffULL);
1083 cmd->data[0] = devid;
1084 cmd->data[0] |= ((pasid >> 8) & 0xff) << 16;
1085 cmd->data[0] |= (qdep & 0xff) << 24;
1086 cmd->data[1] = devid;
1087 cmd->data[1] |= (pasid & 0xff) << 16;
1088 cmd->data[2] = lower_32_bits(address);
1089 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
1090 cmd->data[3] = upper_32_bits(address);
1092 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
1093 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
1096 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, u32 pasid,
1097 int status, int tag, bool gn)
1099 memset(cmd, 0, sizeof(*cmd));
1101 cmd->data[0] = devid;
1103 cmd->data[1] = pasid;
1104 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
1106 cmd->data[3] = tag & 0x1ff;
1107 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
1109 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
1112 static void build_inv_all(struct iommu_cmd *cmd)
1114 memset(cmd, 0, sizeof(*cmd));
1115 CMD_SET_TYPE(cmd, CMD_INV_ALL);
1118 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
1120 memset(cmd, 0, sizeof(*cmd));
1121 cmd->data[0] = devid;
1122 CMD_SET_TYPE(cmd, CMD_INV_IRT);
1126 * Writes the command to the IOMMUs command buffer and informs the
1127 * hardware about the new command.
1129 static int __iommu_queue_command_sync(struct amd_iommu *iommu,
1130 struct iommu_cmd *cmd,
1133 unsigned int count = 0;
1134 u32 left, next_tail;
1136 next_tail = (iommu->cmd_buf_tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
1138 left = (iommu->cmd_buf_head - next_tail) % CMD_BUFFER_SIZE;
1141 /* Skip udelay() the first time around */
1143 if (count == LOOP_TIMEOUT) {
1144 pr_err("Command buffer timeout\n");
1151 /* Update head and recheck remaining space */
1152 iommu->cmd_buf_head = readl(iommu->mmio_base +
1153 MMIO_CMD_HEAD_OFFSET);
1158 copy_cmd_to_buffer(iommu, cmd);
1160 /* Do we need to make sure all commands are processed? */
1161 iommu->need_sync = sync;
1166 static int iommu_queue_command_sync(struct amd_iommu *iommu,
1167 struct iommu_cmd *cmd,
1170 unsigned long flags;
1173 raw_spin_lock_irqsave(&iommu->lock, flags);
1174 ret = __iommu_queue_command_sync(iommu, cmd, sync);
1175 raw_spin_unlock_irqrestore(&iommu->lock, flags);
1180 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
1182 return iommu_queue_command_sync(iommu, cmd, true);
1186 * This function queues a completion wait command into the command
1187 * buffer of an IOMMU
1189 static int iommu_completion_wait(struct amd_iommu *iommu)
1191 struct iommu_cmd cmd;
1192 unsigned long flags;
1196 if (!iommu->need_sync)
1199 data = atomic64_add_return(1, &iommu->cmd_sem_val);
1200 build_completion_wait(&cmd, iommu, data);
1202 raw_spin_lock_irqsave(&iommu->lock, flags);
1204 ret = __iommu_queue_command_sync(iommu, &cmd, false);
1208 ret = wait_on_sem(iommu, data);
1211 raw_spin_unlock_irqrestore(&iommu->lock, flags);
1216 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
1218 struct iommu_cmd cmd;
1220 build_inv_dte(&cmd, devid);
1222 return iommu_queue_command(iommu, &cmd);
1225 static void amd_iommu_flush_dte_all(struct amd_iommu *iommu)
1228 u16 last_bdf = iommu->pci_seg->last_bdf;
1230 for (devid = 0; devid <= last_bdf; ++devid)
1231 iommu_flush_dte(iommu, devid);
1233 iommu_completion_wait(iommu);
1237 * This function uses heavy locking and may disable irqs for some time. But
1238 * this is no issue because it is only called during resume.
1240 static void amd_iommu_flush_tlb_all(struct amd_iommu *iommu)
1243 u16 last_bdf = iommu->pci_seg->last_bdf;
1245 for (dom_id = 0; dom_id <= last_bdf; ++dom_id) {
1246 struct iommu_cmd cmd;
1247 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1249 iommu_queue_command(iommu, &cmd);
1252 iommu_completion_wait(iommu);
1255 static void amd_iommu_flush_tlb_domid(struct amd_iommu *iommu, u32 dom_id)
1257 struct iommu_cmd cmd;
1259 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1261 iommu_queue_command(iommu, &cmd);
1263 iommu_completion_wait(iommu);
1266 static void amd_iommu_flush_all(struct amd_iommu *iommu)
1268 struct iommu_cmd cmd;
1270 build_inv_all(&cmd);
1272 iommu_queue_command(iommu, &cmd);
1273 iommu_completion_wait(iommu);
1276 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1278 struct iommu_cmd cmd;
1280 build_inv_irt(&cmd, devid);
1282 iommu_queue_command(iommu, &cmd);
1285 static void amd_iommu_flush_irt_all(struct amd_iommu *iommu)
1288 u16 last_bdf = iommu->pci_seg->last_bdf;
1290 if (iommu->irtcachedis_enabled)
1293 for (devid = 0; devid <= last_bdf; devid++)
1294 iommu_flush_irt(iommu, devid);
1296 iommu_completion_wait(iommu);
1299 void iommu_flush_all_caches(struct amd_iommu *iommu)
1301 if (iommu_feature(iommu, FEATURE_IA)) {
1302 amd_iommu_flush_all(iommu);
1304 amd_iommu_flush_dte_all(iommu);
1305 amd_iommu_flush_irt_all(iommu);
1306 amd_iommu_flush_tlb_all(iommu);
1311 * Command send function for flushing on-device TLB
1313 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1314 u64 address, size_t size)
1316 struct amd_iommu *iommu;
1317 struct iommu_cmd cmd;
1320 qdep = dev_data->ats.qdep;
1321 iommu = rlookup_amd_iommu(dev_data->dev);
1325 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1327 return iommu_queue_command(iommu, &cmd);
1330 static int device_flush_dte_alias(struct pci_dev *pdev, u16 alias, void *data)
1332 struct amd_iommu *iommu = data;
1334 return iommu_flush_dte(iommu, alias);
1338 * Command send function for invalidating a device table entry
1340 static int device_flush_dte(struct iommu_dev_data *dev_data)
1342 struct amd_iommu *iommu;
1343 struct pci_dev *pdev = NULL;
1344 struct amd_iommu_pci_seg *pci_seg;
1348 iommu = rlookup_amd_iommu(dev_data->dev);
1352 if (dev_is_pci(dev_data->dev))
1353 pdev = to_pci_dev(dev_data->dev);
1356 ret = pci_for_each_dma_alias(pdev,
1357 device_flush_dte_alias, iommu);
1359 ret = iommu_flush_dte(iommu, dev_data->devid);
1363 pci_seg = iommu->pci_seg;
1364 alias = pci_seg->alias_table[dev_data->devid];
1365 if (alias != dev_data->devid) {
1366 ret = iommu_flush_dte(iommu, alias);
1371 if (dev_data->ats.enabled)
1372 ret = device_flush_iotlb(dev_data, 0, ~0UL);
1378 * TLB invalidation function which is called from the mapping functions.
1379 * It invalidates a single PTE if the range to flush is within a single
1380 * page. Otherwise it flushes the whole TLB of the IOMMU.
1382 static void __domain_flush_pages(struct protection_domain *domain,
1383 u64 address, size_t size, int pde)
1385 struct iommu_dev_data *dev_data;
1386 struct iommu_cmd cmd;
1389 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1391 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
1392 if (!domain->dev_iommu[i])
1396 * Devices of this domain are behind this IOMMU
1397 * We need a TLB flush
1399 ret |= iommu_queue_command(amd_iommus[i], &cmd);
1402 list_for_each_entry(dev_data, &domain->dev_list, list) {
1404 if (!dev_data->ats.enabled)
1407 ret |= device_flush_iotlb(dev_data, address, size);
1413 static void domain_flush_pages(struct protection_domain *domain,
1414 u64 address, size_t size, int pde)
1416 if (likely(!amd_iommu_np_cache)) {
1417 __domain_flush_pages(domain, address, size, pde);
1422 * When NpCache is on, we infer that we run in a VM and use a vIOMMU.
1423 * In such setups it is best to avoid flushes of ranges which are not
1424 * naturally aligned, since it would lead to flushes of unmodified
1425 * PTEs. Such flushes would require the hypervisor to do more work than
1426 * necessary. Therefore, perform repeated flushes of aligned ranges
1427 * until you cover the range. Each iteration flushes the smaller
1428 * between the natural alignment of the address that we flush and the
1429 * greatest naturally aligned region that fits in the range.
1432 int addr_alignment = __ffs(address);
1433 int size_alignment = __fls(size);
1438 * size is always non-zero, but address might be zero, causing
1439 * addr_alignment to be negative. As the casting of the
1440 * argument in __ffs(address) to long might trim the high bits
1441 * of the address on x86-32, cast to long when doing the check.
1443 if (likely((unsigned long)address != 0))
1444 min_alignment = min(addr_alignment, size_alignment);
1446 min_alignment = size_alignment;
1448 flush_size = 1ul << min_alignment;
1450 __domain_flush_pages(domain, address, flush_size, pde);
1451 address += flush_size;
1456 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1457 void amd_iommu_domain_flush_tlb_pde(struct protection_domain *domain)
1459 domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1462 void amd_iommu_domain_flush_complete(struct protection_domain *domain)
1466 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
1467 if (domain && !domain->dev_iommu[i])
1471 * Devices of this domain are behind this IOMMU
1472 * We need to wait for completion of all commands.
1474 iommu_completion_wait(amd_iommus[i]);
1478 /* Flush the not present cache if it exists */
1479 static void domain_flush_np_cache(struct protection_domain *domain,
1480 dma_addr_t iova, size_t size)
1482 if (unlikely(amd_iommu_np_cache)) {
1483 unsigned long flags;
1485 spin_lock_irqsave(&domain->lock, flags);
1486 domain_flush_pages(domain, iova, size, 1);
1487 amd_iommu_domain_flush_complete(domain);
1488 spin_unlock_irqrestore(&domain->lock, flags);
1494 * This function flushes the DTEs for all devices in domain
1496 static void domain_flush_devices(struct protection_domain *domain)
1498 struct iommu_dev_data *dev_data;
1500 list_for_each_entry(dev_data, &domain->dev_list, list)
1501 device_flush_dte(dev_data);
1504 /****************************************************************************
1506 * The next functions belong to the domain allocation. A domain is
1507 * allocated for every IOMMU as the default domain. If device isolation
1508 * is enabled, every device get its own domain. The most important thing
1509 * about domains is the page table mapping the DMA address space they
1512 ****************************************************************************/
1514 static u16 domain_id_alloc(void)
1518 spin_lock(&pd_bitmap_lock);
1519 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1521 if (id > 0 && id < MAX_DOMAIN_ID)
1522 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1525 spin_unlock(&pd_bitmap_lock);
1530 static void domain_id_free(int id)
1532 spin_lock(&pd_bitmap_lock);
1533 if (id > 0 && id < MAX_DOMAIN_ID)
1534 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1535 spin_unlock(&pd_bitmap_lock);
1538 static void free_gcr3_tbl_level1(u64 *tbl)
1543 for (i = 0; i < 512; ++i) {
1544 if (!(tbl[i] & GCR3_VALID))
1547 ptr = iommu_phys_to_virt(tbl[i] & PAGE_MASK);
1549 free_page((unsigned long)ptr);
1553 static void free_gcr3_tbl_level2(u64 *tbl)
1558 for (i = 0; i < 512; ++i) {
1559 if (!(tbl[i] & GCR3_VALID))
1562 ptr = iommu_phys_to_virt(tbl[i] & PAGE_MASK);
1564 free_gcr3_tbl_level1(ptr);
1568 static void free_gcr3_table(struct protection_domain *domain)
1570 if (domain->glx == 2)
1571 free_gcr3_tbl_level2(domain->gcr3_tbl);
1572 else if (domain->glx == 1)
1573 free_gcr3_tbl_level1(domain->gcr3_tbl);
1575 BUG_ON(domain->glx != 0);
1577 free_page((unsigned long)domain->gcr3_tbl);
1580 static void set_dte_entry(struct amd_iommu *iommu, u16 devid,
1581 struct protection_domain *domain, bool ats, bool ppr)
1586 struct dev_table_entry *dev_table = get_dev_table(iommu);
1588 if (domain->iop.mode != PAGE_MODE_NONE)
1589 pte_root = iommu_virt_to_phys(domain->iop.root);
1591 pte_root |= (domain->iop.mode & DEV_ENTRY_MODE_MASK)
1592 << DEV_ENTRY_MODE_SHIFT;
1594 pte_root |= DTE_FLAG_IR | DTE_FLAG_IW | DTE_FLAG_V;
1597 * When SNP is enabled, Only set TV bit when IOMMU
1598 * page translation is in use.
1600 if (!amd_iommu_snp_en || (domain->id != 0))
1601 pte_root |= DTE_FLAG_TV;
1603 flags = dev_table[devid].data[1];
1606 flags |= DTE_FLAG_IOTLB;
1609 if (iommu_feature(iommu, FEATURE_EPHSUP))
1610 pte_root |= 1ULL << DEV_ENTRY_PPR;
1613 if (domain->flags & PD_IOMMUV2_MASK) {
1614 u64 gcr3 = iommu_virt_to_phys(domain->gcr3_tbl);
1615 u64 glx = domain->glx;
1618 pte_root |= DTE_FLAG_GV;
1619 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
1621 /* First mask out possible old values for GCR3 table */
1622 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
1625 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
1628 /* Encode GCR3 table into DTE */
1629 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
1632 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
1635 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
1638 if (amd_iommu_gpt_level == PAGE_MODE_5_LEVEL) {
1639 dev_table[devid].data[2] |=
1640 ((u64)GUEST_PGTABLE_5_LEVEL << DTE_GPT_LEVEL_SHIFT);
1643 if (domain->flags & PD_GIOV_MASK)
1644 pte_root |= DTE_FLAG_GIOV;
1647 flags &= ~DEV_DOMID_MASK;
1648 flags |= domain->id;
1650 old_domid = dev_table[devid].data[1] & DEV_DOMID_MASK;
1651 dev_table[devid].data[1] = flags;
1652 dev_table[devid].data[0] = pte_root;
1655 * A kdump kernel might be replacing a domain ID that was copied from
1656 * the previous kernel--if so, it needs to flush the translation cache
1657 * entries for the old domain ID that is being overwritten
1660 amd_iommu_flush_tlb_domid(iommu, old_domid);
1664 static void clear_dte_entry(struct amd_iommu *iommu, u16 devid)
1666 struct dev_table_entry *dev_table = get_dev_table(iommu);
1668 /* remove entry from the device table seen by the hardware */
1669 dev_table[devid].data[0] = DTE_FLAG_V;
1671 if (!amd_iommu_snp_en)
1672 dev_table[devid].data[0] |= DTE_FLAG_TV;
1674 dev_table[devid].data[1] &= DTE_FLAG_MASK;
1676 amd_iommu_apply_erratum_63(iommu, devid);
1679 static void do_attach(struct iommu_dev_data *dev_data,
1680 struct protection_domain *domain)
1682 struct amd_iommu *iommu;
1685 iommu = rlookup_amd_iommu(dev_data->dev);
1688 ats = dev_data->ats.enabled;
1690 /* Update data structures */
1691 dev_data->domain = domain;
1692 list_add(&dev_data->list, &domain->dev_list);
1694 /* Update NUMA Node ID */
1695 if (domain->nid == NUMA_NO_NODE)
1696 domain->nid = dev_to_node(dev_data->dev);
1698 /* Do reference counting */
1699 domain->dev_iommu[iommu->index] += 1;
1700 domain->dev_cnt += 1;
1702 /* Update device table */
1703 set_dte_entry(iommu, dev_data->devid, domain,
1704 ats, dev_data->iommu_v2);
1705 clone_aliases(iommu, dev_data->dev);
1707 device_flush_dte(dev_data);
1710 static void do_detach(struct iommu_dev_data *dev_data)
1712 struct protection_domain *domain = dev_data->domain;
1713 struct amd_iommu *iommu;
1715 iommu = rlookup_amd_iommu(dev_data->dev);
1719 /* Update data structures */
1720 dev_data->domain = NULL;
1721 list_del(&dev_data->list);
1722 clear_dte_entry(iommu, dev_data->devid);
1723 clone_aliases(iommu, dev_data->dev);
1725 /* Flush the DTE entry */
1726 device_flush_dte(dev_data);
1729 amd_iommu_domain_flush_tlb_pde(domain);
1731 /* Wait for the flushes to finish */
1732 amd_iommu_domain_flush_complete(domain);
1734 /* decrease reference counters - needs to happen after the flushes */
1735 domain->dev_iommu[iommu->index] -= 1;
1736 domain->dev_cnt -= 1;
1739 static void pdev_iommuv2_disable(struct pci_dev *pdev)
1741 pci_disable_ats(pdev);
1742 pci_disable_pri(pdev);
1743 pci_disable_pasid(pdev);
1746 static int pdev_pri_ats_enable(struct pci_dev *pdev)
1750 /* Only allow access to user-accessible pages */
1751 ret = pci_enable_pasid(pdev, 0);
1755 /* First reset the PRI state of the device */
1756 ret = pci_reset_pri(pdev);
1761 /* FIXME: Hardcode number of outstanding requests for now */
1762 ret = pci_enable_pri(pdev, 32);
1766 ret = pci_enable_ats(pdev, PAGE_SHIFT);
1773 pci_disable_pri(pdev);
1776 pci_disable_pasid(pdev);
1782 * If a device is not yet associated with a domain, this function makes the
1783 * device visible in the domain
1785 static int attach_device(struct device *dev,
1786 struct protection_domain *domain)
1788 struct iommu_dev_data *dev_data;
1789 struct pci_dev *pdev;
1790 unsigned long flags;
1793 spin_lock_irqsave(&domain->lock, flags);
1795 dev_data = dev_iommu_priv_get(dev);
1797 spin_lock(&dev_data->lock);
1800 if (dev_data->domain != NULL)
1803 if (!dev_is_pci(dev))
1804 goto skip_ats_check;
1806 pdev = to_pci_dev(dev);
1807 if (domain->flags & PD_IOMMUV2_MASK) {
1808 struct iommu_domain *def_domain = iommu_get_dma_domain(dev);
1813 * In case of using AMD_IOMMU_V1 page table mode and the device
1814 * is enabling for PPR/ATS support (using v2 table),
1815 * we need to make sure that the domain type is identity map.
1817 if ((amd_iommu_pgtable == AMD_IOMMU_V1) &&
1818 def_domain->type != IOMMU_DOMAIN_IDENTITY) {
1822 if (dev_data->iommu_v2) {
1823 if (pdev_pri_ats_enable(pdev) != 0)
1826 dev_data->ats.enabled = true;
1827 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
1828 dev_data->pri_tlp = pci_prg_resp_pasid_required(pdev);
1830 } else if (amd_iommu_iotlb_sup &&
1831 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
1832 dev_data->ats.enabled = true;
1833 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
1839 do_attach(dev_data, domain);
1842 * We might boot into a crash-kernel here. The crashed kernel
1843 * left the caches in the IOMMU dirty. So we have to flush
1844 * here to evict all dirty stuff.
1846 amd_iommu_domain_flush_tlb_pde(domain);
1848 amd_iommu_domain_flush_complete(domain);
1851 spin_unlock(&dev_data->lock);
1853 spin_unlock_irqrestore(&domain->lock, flags);
1859 * Removes a device from a protection domain (with devtable_lock held)
1861 static void detach_device(struct device *dev)
1863 struct protection_domain *domain;
1864 struct iommu_dev_data *dev_data;
1865 unsigned long flags;
1867 dev_data = dev_iommu_priv_get(dev);
1868 domain = dev_data->domain;
1870 spin_lock_irqsave(&domain->lock, flags);
1872 spin_lock(&dev_data->lock);
1875 * First check if the device is still attached. It might already
1876 * be detached from its domain because the generic
1877 * iommu_detach_group code detached it and we try again here in
1878 * our alias handling.
1880 if (WARN_ON(!dev_data->domain))
1883 do_detach(dev_data);
1885 if (!dev_is_pci(dev))
1888 if (domain->flags & PD_IOMMUV2_MASK && dev_data->iommu_v2)
1889 pdev_iommuv2_disable(to_pci_dev(dev));
1890 else if (dev_data->ats.enabled)
1891 pci_disable_ats(to_pci_dev(dev));
1893 dev_data->ats.enabled = false;
1896 spin_unlock(&dev_data->lock);
1898 spin_unlock_irqrestore(&domain->lock, flags);
1901 static struct iommu_device *amd_iommu_probe_device(struct device *dev)
1903 struct iommu_device *iommu_dev;
1904 struct amd_iommu *iommu;
1907 if (!check_device(dev))
1908 return ERR_PTR(-ENODEV);
1910 iommu = rlookup_amd_iommu(dev);
1912 return ERR_PTR(-ENODEV);
1914 /* Not registered yet? */
1915 if (!iommu->iommu.ops)
1916 return ERR_PTR(-ENODEV);
1918 if (dev_iommu_priv_get(dev))
1919 return &iommu->iommu;
1921 ret = iommu_init_device(iommu, dev);
1923 if (ret != -ENOTSUPP)
1924 dev_err(dev, "Failed to initialize - trying to proceed anyway\n");
1925 iommu_dev = ERR_PTR(ret);
1926 iommu_ignore_device(iommu, dev);
1928 amd_iommu_set_pci_msi_domain(dev, iommu);
1929 iommu_dev = &iommu->iommu;
1932 iommu_completion_wait(iommu);
1937 static void amd_iommu_probe_finalize(struct device *dev)
1939 /* Domains are initialized for this device - have a look what we ended up with */
1940 set_dma_ops(dev, NULL);
1941 iommu_setup_dma_ops(dev, 0, U64_MAX);
1944 static void amd_iommu_release_device(struct device *dev)
1946 struct amd_iommu *iommu;
1948 if (!check_device(dev))
1951 iommu = rlookup_amd_iommu(dev);
1955 amd_iommu_uninit_device(dev);
1956 iommu_completion_wait(iommu);
1959 static struct iommu_group *amd_iommu_device_group(struct device *dev)
1961 if (dev_is_pci(dev))
1962 return pci_device_group(dev);
1964 return acpihid_device_group(dev);
1967 /*****************************************************************************
1969 * The next functions belong to the dma_ops mapping/unmapping code.
1971 *****************************************************************************/
1973 static void update_device_table(struct protection_domain *domain)
1975 struct iommu_dev_data *dev_data;
1977 list_for_each_entry(dev_data, &domain->dev_list, list) {
1978 struct amd_iommu *iommu = rlookup_amd_iommu(dev_data->dev);
1982 set_dte_entry(iommu, dev_data->devid, domain,
1983 dev_data->ats.enabled, dev_data->iommu_v2);
1984 clone_aliases(iommu, dev_data->dev);
1988 void amd_iommu_update_and_flush_device_table(struct protection_domain *domain)
1990 update_device_table(domain);
1991 domain_flush_devices(domain);
1994 void amd_iommu_domain_update(struct protection_domain *domain)
1996 /* Update device table */
1997 amd_iommu_update_and_flush_device_table(domain);
1999 /* Flush domain TLB(s) and wait for completion */
2000 amd_iommu_domain_flush_tlb_pde(domain);
2001 amd_iommu_domain_flush_complete(domain);
2004 /*****************************************************************************
2006 * The following functions belong to the exported interface of AMD IOMMU
2008 * This interface allows access to lower level functions of the IOMMU
2009 * like protection domain handling and assignement of devices to domains
2010 * which is not possible with the dma_ops interface.
2012 *****************************************************************************/
2014 static void cleanup_domain(struct protection_domain *domain)
2016 struct iommu_dev_data *entry;
2017 unsigned long flags;
2019 spin_lock_irqsave(&domain->lock, flags);
2021 while (!list_empty(&domain->dev_list)) {
2022 entry = list_first_entry(&domain->dev_list,
2023 struct iommu_dev_data, list);
2024 BUG_ON(!entry->domain);
2028 spin_unlock_irqrestore(&domain->lock, flags);
2031 static void protection_domain_free(struct protection_domain *domain)
2036 if (domain->iop.pgtbl_cfg.tlb)
2037 free_io_pgtable_ops(&domain->iop.iop.ops);
2040 domain_id_free(domain->id);
2045 static int protection_domain_init_v1(struct protection_domain *domain, int mode)
2047 u64 *pt_root = NULL;
2049 BUG_ON(mode < PAGE_MODE_NONE || mode > PAGE_MODE_6_LEVEL);
2051 spin_lock_init(&domain->lock);
2052 domain->id = domain_id_alloc();
2055 INIT_LIST_HEAD(&domain->dev_list);
2057 if (mode != PAGE_MODE_NONE) {
2058 pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2060 domain_id_free(domain->id);
2065 amd_iommu_domain_set_pgtable(domain, pt_root, mode);
2070 static int protection_domain_init_v2(struct protection_domain *domain)
2072 spin_lock_init(&domain->lock);
2073 domain->id = domain_id_alloc();
2076 INIT_LIST_HEAD(&domain->dev_list);
2078 domain->flags |= PD_GIOV_MASK;
2080 domain->domain.pgsize_bitmap = AMD_IOMMU_PGSIZES_V2;
2082 if (domain_enable_v2(domain, 1)) {
2083 domain_id_free(domain->id);
2090 static struct protection_domain *protection_domain_alloc(unsigned int type)
2092 struct io_pgtable_ops *pgtbl_ops;
2093 struct protection_domain *domain;
2095 int mode = DEFAULT_PGTABLE_LEVEL;
2099 * Force IOMMU v1 page table when iommu=pt and
2100 * when allocating domain for pass-through devices.
2102 if (type == IOMMU_DOMAIN_IDENTITY) {
2103 pgtable = AMD_IOMMU_V1;
2104 mode = PAGE_MODE_NONE;
2105 } else if (type == IOMMU_DOMAIN_UNMANAGED) {
2106 pgtable = AMD_IOMMU_V1;
2107 } else if (type == IOMMU_DOMAIN_DMA || type == IOMMU_DOMAIN_DMA_FQ) {
2108 pgtable = amd_iommu_pgtable;
2113 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2119 ret = protection_domain_init_v1(domain, mode);
2122 ret = protection_domain_init_v2(domain);
2131 /* No need to allocate io pgtable ops in passthrough mode */
2132 if (type == IOMMU_DOMAIN_IDENTITY)
2135 domain->nid = NUMA_NO_NODE;
2137 pgtbl_ops = alloc_io_pgtable_ops(pgtable, &domain->iop.pgtbl_cfg, domain);
2139 domain_id_free(domain->id);
2149 static inline u64 dma_max_address(void)
2151 if (amd_iommu_pgtable == AMD_IOMMU_V1)
2154 /* V2 with 4/5 level page table */
2155 return ((1ULL << PM_LEVEL_SHIFT(amd_iommu_gpt_level)) - 1);
2158 static struct iommu_domain *amd_iommu_domain_alloc(unsigned type)
2160 struct protection_domain *domain;
2163 * Since DTE[Mode]=0 is prohibited on SNP-enabled system,
2164 * default to use IOMMU_DOMAIN_DMA[_FQ].
2166 if (amd_iommu_snp_en && (type == IOMMU_DOMAIN_IDENTITY))
2169 domain = protection_domain_alloc(type);
2173 domain->domain.geometry.aperture_start = 0;
2174 domain->domain.geometry.aperture_end = dma_max_address();
2175 domain->domain.geometry.force_aperture = true;
2177 return &domain->domain;
2180 static void amd_iommu_domain_free(struct iommu_domain *dom)
2182 struct protection_domain *domain;
2184 domain = to_pdomain(dom);
2186 if (domain->dev_cnt > 0)
2187 cleanup_domain(domain);
2189 BUG_ON(domain->dev_cnt != 0);
2194 if (domain->flags & PD_IOMMUV2_MASK)
2195 free_gcr3_table(domain);
2197 protection_domain_free(domain);
2200 static int amd_iommu_attach_device(struct iommu_domain *dom,
2203 struct iommu_dev_data *dev_data = dev_iommu_priv_get(dev);
2204 struct protection_domain *domain = to_pdomain(dom);
2205 struct amd_iommu *iommu = rlookup_amd_iommu(dev);
2209 * Skip attach device to domain if new domain is same as
2210 * devices current domain
2212 if (dev_data->domain == domain)
2215 dev_data->defer_attach = false;
2217 if (dev_data->domain)
2220 ret = attach_device(dev, domain);
2222 #ifdef CONFIG_IRQ_REMAP
2223 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
2224 if (dom->type == IOMMU_DOMAIN_UNMANAGED)
2225 dev_data->use_vapic = 1;
2227 dev_data->use_vapic = 0;
2231 iommu_completion_wait(iommu);
2236 static void amd_iommu_iotlb_sync_map(struct iommu_domain *dom,
2237 unsigned long iova, size_t size)
2239 struct protection_domain *domain = to_pdomain(dom);
2240 struct io_pgtable_ops *ops = &domain->iop.iop.ops;
2243 domain_flush_np_cache(domain, iova, size);
2246 static int amd_iommu_map_pages(struct iommu_domain *dom, unsigned long iova,
2247 phys_addr_t paddr, size_t pgsize, size_t pgcount,
2248 int iommu_prot, gfp_t gfp, size_t *mapped)
2250 struct protection_domain *domain = to_pdomain(dom);
2251 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 if (iommu_prot & IOMMU_READ)
2260 prot |= IOMMU_PROT_IR;
2261 if (iommu_prot & IOMMU_WRITE)
2262 prot |= IOMMU_PROT_IW;
2264 if (ops->map_pages) {
2265 ret = ops->map_pages(ops, iova, paddr, pgsize,
2266 pgcount, prot, gfp, mapped);
2272 static void amd_iommu_iotlb_gather_add_page(struct iommu_domain *domain,
2273 struct iommu_iotlb_gather *gather,
2274 unsigned long iova, size_t size)
2277 * AMD's IOMMU can flush as many pages as necessary in a single flush.
2278 * Unless we run in a virtual machine, which can be inferred according
2279 * to whether "non-present cache" is on, it is probably best to prefer
2280 * (potentially) too extensive TLB flushing (i.e., more misses) over
2281 * mutliple TLB flushes (i.e., more flushes). For virtual machines the
2282 * hypervisor needs to synchronize the host IOMMU PTEs with those of
2283 * the guest, and the trade-off is different: unnecessary TLB flushes
2284 * should be avoided.
2286 if (amd_iommu_np_cache &&
2287 iommu_iotlb_gather_is_disjoint(gather, iova, size))
2288 iommu_iotlb_sync(domain, gather);
2290 iommu_iotlb_gather_add_range(gather, iova, size);
2293 static size_t amd_iommu_unmap_pages(struct iommu_domain *dom, unsigned long iova,
2294 size_t pgsize, size_t pgcount,
2295 struct iommu_iotlb_gather *gather)
2297 struct protection_domain *domain = to_pdomain(dom);
2298 struct io_pgtable_ops *ops = &domain->iop.iop.ops;
2301 if ((amd_iommu_pgtable == AMD_IOMMU_V1) &&
2302 (domain->iop.mode == PAGE_MODE_NONE))
2305 r = (ops->unmap_pages) ? ops->unmap_pages(ops, iova, pgsize, pgcount, NULL) : 0;
2308 amd_iommu_iotlb_gather_add_page(dom, gather, iova, r);
2313 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
2316 struct protection_domain *domain = to_pdomain(dom);
2317 struct io_pgtable_ops *ops = &domain->iop.iop.ops;
2319 return ops->iova_to_phys(ops, iova);
2322 static bool amd_iommu_capable(struct device *dev, enum iommu_cap cap)
2325 case IOMMU_CAP_CACHE_COHERENCY:
2327 case IOMMU_CAP_NOEXEC:
2329 case IOMMU_CAP_PRE_BOOT_PROTECTION:
2330 return amdr_ivrs_remap_support;
2331 case IOMMU_CAP_ENFORCE_CACHE_COHERENCY:
2333 case IOMMU_CAP_DEFERRED_FLUSH:
2342 static void amd_iommu_get_resv_regions(struct device *dev,
2343 struct list_head *head)
2345 struct iommu_resv_region *region;
2346 struct unity_map_entry *entry;
2347 struct amd_iommu *iommu;
2348 struct amd_iommu_pci_seg *pci_seg;
2351 sbdf = get_device_sbdf_id(dev);
2355 devid = PCI_SBDF_TO_DEVID(sbdf);
2356 iommu = rlookup_amd_iommu(dev);
2359 pci_seg = iommu->pci_seg;
2361 list_for_each_entry(entry, &pci_seg->unity_map, list) {
2365 if (devid < entry->devid_start || devid > entry->devid_end)
2368 type = IOMMU_RESV_DIRECT;
2369 length = entry->address_end - entry->address_start;
2370 if (entry->prot & IOMMU_PROT_IR)
2372 if (entry->prot & IOMMU_PROT_IW)
2373 prot |= IOMMU_WRITE;
2374 if (entry->prot & IOMMU_UNITY_MAP_FLAG_EXCL_RANGE)
2375 /* Exclusion range */
2376 type = IOMMU_RESV_RESERVED;
2378 region = iommu_alloc_resv_region(entry->address_start,
2382 dev_err(dev, "Out of memory allocating dm-regions\n");
2385 list_add_tail(®ion->list, head);
2388 region = iommu_alloc_resv_region(MSI_RANGE_START,
2389 MSI_RANGE_END - MSI_RANGE_START + 1,
2390 0, IOMMU_RESV_MSI, GFP_KERNEL);
2393 list_add_tail(®ion->list, head);
2395 region = iommu_alloc_resv_region(HT_RANGE_START,
2396 HT_RANGE_END - HT_RANGE_START + 1,
2397 0, IOMMU_RESV_RESERVED, GFP_KERNEL);
2400 list_add_tail(®ion->list, head);
2403 bool amd_iommu_is_attach_deferred(struct device *dev)
2405 struct iommu_dev_data *dev_data = dev_iommu_priv_get(dev);
2407 return dev_data->defer_attach;
2409 EXPORT_SYMBOL_GPL(amd_iommu_is_attach_deferred);
2411 static void amd_iommu_flush_iotlb_all(struct iommu_domain *domain)
2413 struct protection_domain *dom = to_pdomain(domain);
2414 unsigned long flags;
2416 spin_lock_irqsave(&dom->lock, flags);
2417 amd_iommu_domain_flush_tlb_pde(dom);
2418 amd_iommu_domain_flush_complete(dom);
2419 spin_unlock_irqrestore(&dom->lock, flags);
2422 static void amd_iommu_iotlb_sync(struct iommu_domain *domain,
2423 struct iommu_iotlb_gather *gather)
2425 struct protection_domain *dom = to_pdomain(domain);
2426 unsigned long flags;
2428 spin_lock_irqsave(&dom->lock, flags);
2429 domain_flush_pages(dom, gather->start, gather->end - gather->start + 1, 1);
2430 amd_iommu_domain_flush_complete(dom);
2431 spin_unlock_irqrestore(&dom->lock, flags);
2434 static int amd_iommu_def_domain_type(struct device *dev)
2436 struct iommu_dev_data *dev_data;
2438 dev_data = dev_iommu_priv_get(dev);
2443 * Do not identity map IOMMUv2 capable devices when:
2444 * - memory encryption is active, because some of those devices
2445 * (AMD GPUs) don't have the encryption bit in their DMA-mask
2446 * and require remapping.
2447 * - SNP is enabled, because it prohibits DTE[Mode]=0.
2449 if (dev_data->iommu_v2 &&
2450 !cc_platform_has(CC_ATTR_MEM_ENCRYPT) &&
2451 !amd_iommu_snp_en) {
2452 return IOMMU_DOMAIN_IDENTITY;
2458 static bool amd_iommu_enforce_cache_coherency(struct iommu_domain *domain)
2460 /* IOMMU_PTE_FC is always set */
2464 const struct iommu_ops amd_iommu_ops = {
2465 .capable = amd_iommu_capable,
2466 .domain_alloc = amd_iommu_domain_alloc,
2467 .probe_device = amd_iommu_probe_device,
2468 .release_device = amd_iommu_release_device,
2469 .probe_finalize = amd_iommu_probe_finalize,
2470 .device_group = amd_iommu_device_group,
2471 .get_resv_regions = amd_iommu_get_resv_regions,
2472 .is_attach_deferred = amd_iommu_is_attach_deferred,
2473 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
2474 .def_domain_type = amd_iommu_def_domain_type,
2475 .default_domain_ops = &(const struct iommu_domain_ops) {
2476 .attach_dev = amd_iommu_attach_device,
2477 .map_pages = amd_iommu_map_pages,
2478 .unmap_pages = amd_iommu_unmap_pages,
2479 .iotlb_sync_map = amd_iommu_iotlb_sync_map,
2480 .iova_to_phys = amd_iommu_iova_to_phys,
2481 .flush_iotlb_all = amd_iommu_flush_iotlb_all,
2482 .iotlb_sync = amd_iommu_iotlb_sync,
2483 .free = amd_iommu_domain_free,
2484 .enforce_cache_coherency = amd_iommu_enforce_cache_coherency,
2488 /*****************************************************************************
2490 * The next functions do a basic initialization of IOMMU for pass through
2493 * In passthrough mode the IOMMU is initialized and enabled but not used for
2494 * DMA-API translation.
2496 *****************************************************************************/
2498 /* IOMMUv2 specific functions */
2499 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
2501 return atomic_notifier_chain_register(&ppr_notifier, nb);
2503 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
2505 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
2507 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
2509 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
2511 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
2513 struct protection_domain *domain = to_pdomain(dom);
2514 unsigned long flags;
2516 spin_lock_irqsave(&domain->lock, flags);
2518 if (domain->iop.pgtbl_cfg.tlb)
2519 free_io_pgtable_ops(&domain->iop.iop.ops);
2521 spin_unlock_irqrestore(&domain->lock, flags);
2523 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
2525 /* Note: This function expects iommu_domain->lock to be held prior calling the function. */
2526 static int domain_enable_v2(struct protection_domain *domain, int pasids)
2530 /* Number of GCR3 table levels required */
2531 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
2534 if (levels > amd_iommu_max_glx_val)
2537 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
2538 if (domain->gcr3_tbl == NULL)
2541 domain->glx = levels;
2542 domain->flags |= PD_IOMMUV2_MASK;
2544 amd_iommu_domain_update(domain);
2549 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
2551 struct protection_domain *pdom = to_pdomain(dom);
2552 unsigned long flags;
2555 spin_lock_irqsave(&pdom->lock, flags);
2558 * Save us all sanity checks whether devices already in the
2559 * domain support IOMMUv2. Just force that the domain has no
2560 * devices attached when it is switched into IOMMUv2 mode.
2563 if (pdom->dev_cnt > 0 || pdom->flags & PD_IOMMUV2_MASK)
2566 if (!pdom->gcr3_tbl)
2567 ret = domain_enable_v2(pdom, pasids);
2570 spin_unlock_irqrestore(&pdom->lock, flags);
2573 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
2575 static int __flush_pasid(struct protection_domain *domain, u32 pasid,
2576 u64 address, bool size)
2578 struct iommu_dev_data *dev_data;
2579 struct iommu_cmd cmd;
2582 if (!(domain->flags & PD_IOMMUV2_MASK))
2585 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
2588 * IOMMU TLB needs to be flushed before Device TLB to
2589 * prevent device TLB refill from IOMMU TLB
2591 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
2592 if (domain->dev_iommu[i] == 0)
2595 ret = iommu_queue_command(amd_iommus[i], &cmd);
2600 /* Wait until IOMMU TLB flushes are complete */
2601 amd_iommu_domain_flush_complete(domain);
2603 /* Now flush device TLBs */
2604 list_for_each_entry(dev_data, &domain->dev_list, list) {
2605 struct amd_iommu *iommu;
2609 There might be non-IOMMUv2 capable devices in an IOMMUv2
2612 if (!dev_data->ats.enabled)
2615 qdep = dev_data->ats.qdep;
2616 iommu = rlookup_amd_iommu(dev_data->dev);
2619 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
2620 qdep, address, size);
2622 ret = iommu_queue_command(iommu, &cmd);
2627 /* Wait until all device TLBs are flushed */
2628 amd_iommu_domain_flush_complete(domain);
2637 static int __amd_iommu_flush_page(struct protection_domain *domain, u32 pasid,
2640 return __flush_pasid(domain, pasid, address, false);
2643 int amd_iommu_flush_page(struct iommu_domain *dom, u32 pasid,
2646 struct protection_domain *domain = to_pdomain(dom);
2647 unsigned long flags;
2650 spin_lock_irqsave(&domain->lock, flags);
2651 ret = __amd_iommu_flush_page(domain, pasid, address);
2652 spin_unlock_irqrestore(&domain->lock, flags);
2656 EXPORT_SYMBOL(amd_iommu_flush_page);
2658 static int __amd_iommu_flush_tlb(struct protection_domain *domain, u32 pasid)
2660 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
2664 int amd_iommu_flush_tlb(struct iommu_domain *dom, u32 pasid)
2666 struct protection_domain *domain = to_pdomain(dom);
2667 unsigned long flags;
2670 spin_lock_irqsave(&domain->lock, flags);
2671 ret = __amd_iommu_flush_tlb(domain, pasid);
2672 spin_unlock_irqrestore(&domain->lock, flags);
2676 EXPORT_SYMBOL(amd_iommu_flush_tlb);
2678 static u64 *__get_gcr3_pte(u64 *root, int level, u32 pasid, bool alloc)
2685 index = (pasid >> (9 * level)) & 0x1ff;
2691 if (!(*pte & GCR3_VALID)) {
2695 root = (void *)get_zeroed_page(GFP_ATOMIC);
2699 *pte = iommu_virt_to_phys(root) | GCR3_VALID;
2702 root = iommu_phys_to_virt(*pte & PAGE_MASK);
2710 static int __set_gcr3(struct protection_domain *domain, u32 pasid,
2715 if (domain->iop.mode != PAGE_MODE_NONE)
2718 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
2722 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
2724 return __amd_iommu_flush_tlb(domain, pasid);
2727 static int __clear_gcr3(struct protection_domain *domain, u32 pasid)
2731 if (domain->iop.mode != PAGE_MODE_NONE)
2734 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
2740 return __amd_iommu_flush_tlb(domain, pasid);
2743 int amd_iommu_domain_set_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 = __set_gcr3(domain, pasid, cr3);
2752 spin_unlock_irqrestore(&domain->lock, flags);
2756 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
2758 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, u32 pasid)
2760 struct protection_domain *domain = to_pdomain(dom);
2761 unsigned long flags;
2764 spin_lock_irqsave(&domain->lock, flags);
2765 ret = __clear_gcr3(domain, pasid);
2766 spin_unlock_irqrestore(&domain->lock, flags);
2770 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
2772 int amd_iommu_complete_ppr(struct pci_dev *pdev, u32 pasid,
2773 int status, int tag)
2775 struct iommu_dev_data *dev_data;
2776 struct amd_iommu *iommu;
2777 struct iommu_cmd cmd;
2779 dev_data = dev_iommu_priv_get(&pdev->dev);
2780 iommu = rlookup_amd_iommu(&pdev->dev);
2784 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
2785 tag, dev_data->pri_tlp);
2787 return iommu_queue_command(iommu, &cmd);
2789 EXPORT_SYMBOL(amd_iommu_complete_ppr);
2791 int amd_iommu_device_info(struct pci_dev *pdev,
2792 struct amd_iommu_device_info *info)
2797 if (pdev == NULL || info == NULL)
2800 if (!amd_iommu_v2_supported())
2803 memset(info, 0, sizeof(*info));
2805 if (pci_ats_supported(pdev))
2806 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
2808 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2810 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
2812 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
2816 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
2817 max_pasids = min(max_pasids, (1 << 20));
2819 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
2820 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
2822 features = pci_pasid_features(pdev);
2823 if (features & PCI_PASID_CAP_EXEC)
2824 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
2825 if (features & PCI_PASID_CAP_PRIV)
2826 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
2831 EXPORT_SYMBOL(amd_iommu_device_info);
2833 #ifdef CONFIG_IRQ_REMAP
2835 /*****************************************************************************
2837 * Interrupt Remapping Implementation
2839 *****************************************************************************/
2841 static struct irq_chip amd_ir_chip;
2842 static DEFINE_SPINLOCK(iommu_table_lock);
2844 static void iommu_flush_irt_and_complete(struct amd_iommu *iommu, u16 devid)
2848 unsigned long flags;
2849 struct iommu_cmd cmd, cmd2;
2851 if (iommu->irtcachedis_enabled)
2854 build_inv_irt(&cmd, devid);
2855 data = atomic64_add_return(1, &iommu->cmd_sem_val);
2856 build_completion_wait(&cmd2, iommu, data);
2858 raw_spin_lock_irqsave(&iommu->lock, flags);
2859 ret = __iommu_queue_command_sync(iommu, &cmd, true);
2862 ret = __iommu_queue_command_sync(iommu, &cmd2, false);
2865 wait_on_sem(iommu, data);
2867 raw_spin_unlock_irqrestore(&iommu->lock, flags);
2870 static void set_dte_irq_entry(struct amd_iommu *iommu, u16 devid,
2871 struct irq_remap_table *table)
2874 struct dev_table_entry *dev_table = get_dev_table(iommu);
2876 dte = dev_table[devid].data[2];
2877 dte &= ~DTE_IRQ_PHYS_ADDR_MASK;
2878 dte |= iommu_virt_to_phys(table->table);
2879 dte |= DTE_IRQ_REMAP_INTCTL;
2880 dte |= DTE_INTTABLEN;
2881 dte |= DTE_IRQ_REMAP_ENABLE;
2883 dev_table[devid].data[2] = dte;
2886 static struct irq_remap_table *get_irq_table(struct amd_iommu *iommu, u16 devid)
2888 struct irq_remap_table *table;
2889 struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
2891 if (WARN_ONCE(!pci_seg->rlookup_table[devid],
2892 "%s: no iommu for devid %x:%x\n",
2893 __func__, pci_seg->id, devid))
2896 table = pci_seg->irq_lookup_table[devid];
2897 if (WARN_ONCE(!table, "%s: no table for devid %x:%x\n",
2898 __func__, pci_seg->id, devid))
2904 static struct irq_remap_table *__alloc_irq_table(void)
2906 struct irq_remap_table *table;
2908 table = kzalloc(sizeof(*table), GFP_KERNEL);
2912 table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_KERNEL);
2913 if (!table->table) {
2917 raw_spin_lock_init(&table->lock);
2919 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
2920 memset(table->table, 0,
2921 MAX_IRQS_PER_TABLE * sizeof(u32));
2923 memset(table->table, 0,
2924 (MAX_IRQS_PER_TABLE * (sizeof(u64) * 2)));
2928 static void set_remap_table_entry(struct amd_iommu *iommu, u16 devid,
2929 struct irq_remap_table *table)
2931 struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
2933 pci_seg->irq_lookup_table[devid] = table;
2934 set_dte_irq_entry(iommu, devid, table);
2935 iommu_flush_dte(iommu, devid);
2938 static int set_remap_table_entry_alias(struct pci_dev *pdev, u16 alias,
2941 struct irq_remap_table *table = data;
2942 struct amd_iommu_pci_seg *pci_seg;
2943 struct amd_iommu *iommu = rlookup_amd_iommu(&pdev->dev);
2948 pci_seg = iommu->pci_seg;
2949 pci_seg->irq_lookup_table[alias] = table;
2950 set_dte_irq_entry(iommu, alias, table);
2951 iommu_flush_dte(pci_seg->rlookup_table[alias], alias);
2956 static struct irq_remap_table *alloc_irq_table(struct amd_iommu *iommu,
2957 u16 devid, struct pci_dev *pdev)
2959 struct irq_remap_table *table = NULL;
2960 struct irq_remap_table *new_table = NULL;
2961 struct amd_iommu_pci_seg *pci_seg;
2962 unsigned long flags;
2965 spin_lock_irqsave(&iommu_table_lock, flags);
2967 pci_seg = iommu->pci_seg;
2968 table = pci_seg->irq_lookup_table[devid];
2972 alias = pci_seg->alias_table[devid];
2973 table = pci_seg->irq_lookup_table[alias];
2975 set_remap_table_entry(iommu, devid, table);
2978 spin_unlock_irqrestore(&iommu_table_lock, flags);
2980 /* Nothing there yet, allocate new irq remapping table */
2981 new_table = __alloc_irq_table();
2985 spin_lock_irqsave(&iommu_table_lock, flags);
2987 table = pci_seg->irq_lookup_table[devid];
2991 table = pci_seg->irq_lookup_table[alias];
2993 set_remap_table_entry(iommu, devid, table);
3001 pci_for_each_dma_alias(pdev, set_remap_table_entry_alias,
3004 set_remap_table_entry(iommu, devid, table);
3007 set_remap_table_entry(iommu, alias, table);
3010 iommu_completion_wait(iommu);
3013 spin_unlock_irqrestore(&iommu_table_lock, flags);
3016 kmem_cache_free(amd_iommu_irq_cache, new_table->table);
3022 static int alloc_irq_index(struct amd_iommu *iommu, u16 devid, int count,
3023 bool align, struct pci_dev *pdev)
3025 struct irq_remap_table *table;
3026 int index, c, alignment = 1;
3027 unsigned long flags;
3029 table = alloc_irq_table(iommu, devid, pdev);
3034 alignment = roundup_pow_of_two(count);
3036 raw_spin_lock_irqsave(&table->lock, flags);
3038 /* Scan table for free entries */
3039 for (index = ALIGN(table->min_index, alignment), c = 0;
3040 index < MAX_IRQS_PER_TABLE;) {
3041 if (!iommu->irte_ops->is_allocated(table, index)) {
3045 index = ALIGN(index + 1, alignment);
3051 iommu->irte_ops->set_allocated(table, index - c + 1);
3063 raw_spin_unlock_irqrestore(&table->lock, flags);
3068 static int modify_irte_ga(struct amd_iommu *iommu, u16 devid, int index,
3069 struct irte_ga *irte)
3071 struct irq_remap_table *table;
3072 struct irte_ga *entry;
3073 unsigned long flags;
3076 table = get_irq_table(iommu, devid);
3080 raw_spin_lock_irqsave(&table->lock, flags);
3082 entry = (struct irte_ga *)table->table;
3083 entry = &entry[index];
3086 * We use cmpxchg16 to atomically update the 128-bit IRTE,
3087 * and it cannot be updated by the hardware or other processors
3088 * behind us, so the return value of cmpxchg16 should be the
3089 * same as the old value.
3092 WARN_ON(!try_cmpxchg128(&entry->irte, &old, irte->irte));
3094 raw_spin_unlock_irqrestore(&table->lock, flags);
3096 iommu_flush_irt_and_complete(iommu, devid);
3101 static int modify_irte(struct amd_iommu *iommu,
3102 u16 devid, int index, union irte *irte)
3104 struct irq_remap_table *table;
3105 unsigned long flags;
3107 table = get_irq_table(iommu, devid);
3111 raw_spin_lock_irqsave(&table->lock, flags);
3112 table->table[index] = irte->val;
3113 raw_spin_unlock_irqrestore(&table->lock, flags);
3115 iommu_flush_irt_and_complete(iommu, devid);
3120 static void free_irte(struct amd_iommu *iommu, u16 devid, int index)
3122 struct irq_remap_table *table;
3123 unsigned long flags;
3125 table = get_irq_table(iommu, devid);
3129 raw_spin_lock_irqsave(&table->lock, flags);
3130 iommu->irte_ops->clear_allocated(table, index);
3131 raw_spin_unlock_irqrestore(&table->lock, flags);
3133 iommu_flush_irt_and_complete(iommu, devid);
3136 static void irte_prepare(void *entry,
3137 u32 delivery_mode, bool dest_mode,
3138 u8 vector, u32 dest_apicid, int devid)
3140 union irte *irte = (union irte *) entry;
3143 irte->fields.vector = vector;
3144 irte->fields.int_type = delivery_mode;
3145 irte->fields.destination = dest_apicid;
3146 irte->fields.dm = dest_mode;
3147 irte->fields.valid = 1;
3150 static void irte_ga_prepare(void *entry,
3151 u32 delivery_mode, bool dest_mode,
3152 u8 vector, u32 dest_apicid, int devid)
3154 struct irte_ga *irte = (struct irte_ga *) entry;
3158 irte->lo.fields_remap.int_type = delivery_mode;
3159 irte->lo.fields_remap.dm = dest_mode;
3160 irte->hi.fields.vector = vector;
3161 irte->lo.fields_remap.destination = APICID_TO_IRTE_DEST_LO(dest_apicid);
3162 irte->hi.fields.destination = APICID_TO_IRTE_DEST_HI(dest_apicid);
3163 irte->lo.fields_remap.valid = 1;
3166 static void irte_activate(struct amd_iommu *iommu, void *entry, u16 devid, u16 index)
3168 union irte *irte = (union irte *) entry;
3170 irte->fields.valid = 1;
3171 modify_irte(iommu, devid, index, irte);
3174 static void irte_ga_activate(struct amd_iommu *iommu, void *entry, u16 devid, u16 index)
3176 struct irte_ga *irte = (struct irte_ga *) entry;
3178 irte->lo.fields_remap.valid = 1;
3179 modify_irte_ga(iommu, devid, index, irte);
3182 static void irte_deactivate(struct amd_iommu *iommu, void *entry, u16 devid, u16 index)
3184 union irte *irte = (union irte *) entry;
3186 irte->fields.valid = 0;
3187 modify_irte(iommu, devid, index, irte);
3190 static void irte_ga_deactivate(struct amd_iommu *iommu, void *entry, u16 devid, u16 index)
3192 struct irte_ga *irte = (struct irte_ga *) entry;
3194 irte->lo.fields_remap.valid = 0;
3195 modify_irte_ga(iommu, devid, index, irte);
3198 static void irte_set_affinity(struct amd_iommu *iommu, void *entry, u16 devid, u16 index,
3199 u8 vector, u32 dest_apicid)
3201 union irte *irte = (union irte *) entry;
3203 irte->fields.vector = vector;
3204 irte->fields.destination = dest_apicid;
3205 modify_irte(iommu, devid, index, irte);
3208 static void irte_ga_set_affinity(struct amd_iommu *iommu, void *entry, u16 devid, u16 index,
3209 u8 vector, u32 dest_apicid)
3211 struct irte_ga *irte = (struct irte_ga *) entry;
3213 if (!irte->lo.fields_remap.guest_mode) {
3214 irte->hi.fields.vector = vector;
3215 irte->lo.fields_remap.destination =
3216 APICID_TO_IRTE_DEST_LO(dest_apicid);
3217 irte->hi.fields.destination =
3218 APICID_TO_IRTE_DEST_HI(dest_apicid);
3219 modify_irte_ga(iommu, devid, index, irte);
3223 #define IRTE_ALLOCATED (~1U)
3224 static void irte_set_allocated(struct irq_remap_table *table, int index)
3226 table->table[index] = IRTE_ALLOCATED;
3229 static void irte_ga_set_allocated(struct irq_remap_table *table, int index)
3231 struct irte_ga *ptr = (struct irte_ga *)table->table;
3232 struct irte_ga *irte = &ptr[index];
3234 memset(&irte->lo.val, 0, sizeof(u64));
3235 memset(&irte->hi.val, 0, sizeof(u64));
3236 irte->hi.fields.vector = 0xff;
3239 static bool irte_is_allocated(struct irq_remap_table *table, int index)
3241 union irte *ptr = (union irte *)table->table;
3242 union irte *irte = &ptr[index];
3244 return irte->val != 0;
3247 static bool irte_ga_is_allocated(struct irq_remap_table *table, int index)
3249 struct irte_ga *ptr = (struct irte_ga *)table->table;
3250 struct irte_ga *irte = &ptr[index];
3252 return irte->hi.fields.vector != 0;
3255 static void irte_clear_allocated(struct irq_remap_table *table, int index)
3257 table->table[index] = 0;
3260 static void irte_ga_clear_allocated(struct irq_remap_table *table, int index)
3262 struct irte_ga *ptr = (struct irte_ga *)table->table;
3263 struct irte_ga *irte = &ptr[index];
3265 memset(&irte->lo.val, 0, sizeof(u64));
3266 memset(&irte->hi.val, 0, sizeof(u64));
3269 static int get_devid(struct irq_alloc_info *info)
3271 switch (info->type) {
3272 case X86_IRQ_ALLOC_TYPE_IOAPIC:
3273 return get_ioapic_devid(info->devid);
3274 case X86_IRQ_ALLOC_TYPE_HPET:
3275 return get_hpet_devid(info->devid);
3276 case X86_IRQ_ALLOC_TYPE_PCI_MSI:
3277 case X86_IRQ_ALLOC_TYPE_PCI_MSIX:
3278 return get_device_sbdf_id(msi_desc_to_dev(info->desc));
3285 struct irq_remap_ops amd_iommu_irq_ops = {
3286 .prepare = amd_iommu_prepare,
3287 .enable = amd_iommu_enable,
3288 .disable = amd_iommu_disable,
3289 .reenable = amd_iommu_reenable,
3290 .enable_faulting = amd_iommu_enable_faulting,
3293 static void fill_msi_msg(struct msi_msg *msg, u32 index)
3296 msg->address_lo = 0;
3297 msg->arch_addr_lo.base_address = X86_MSI_BASE_ADDRESS_LOW;
3298 msg->address_hi = X86_MSI_BASE_ADDRESS_HIGH;
3301 static void irq_remapping_prepare_irte(struct amd_ir_data *data,
3302 struct irq_cfg *irq_cfg,
3303 struct irq_alloc_info *info,
3304 int devid, int index, int sub_handle)
3306 struct irq_2_irte *irte_info = &data->irq_2_irte;
3307 struct amd_iommu *iommu = data->iommu;
3312 data->irq_2_irte.devid = devid;
3313 data->irq_2_irte.index = index + sub_handle;
3314 iommu->irte_ops->prepare(data->entry, apic->delivery_mode,
3315 apic->dest_mode_logical, irq_cfg->vector,
3316 irq_cfg->dest_apicid, devid);
3318 switch (info->type) {
3319 case X86_IRQ_ALLOC_TYPE_IOAPIC:
3320 case X86_IRQ_ALLOC_TYPE_HPET:
3321 case X86_IRQ_ALLOC_TYPE_PCI_MSI:
3322 case X86_IRQ_ALLOC_TYPE_PCI_MSIX:
3323 fill_msi_msg(&data->msi_entry, irte_info->index);
3332 struct amd_irte_ops irte_32_ops = {
3333 .prepare = irte_prepare,
3334 .activate = irte_activate,
3335 .deactivate = irte_deactivate,
3336 .set_affinity = irte_set_affinity,
3337 .set_allocated = irte_set_allocated,
3338 .is_allocated = irte_is_allocated,
3339 .clear_allocated = irte_clear_allocated,
3342 struct amd_irte_ops irte_128_ops = {
3343 .prepare = irte_ga_prepare,
3344 .activate = irte_ga_activate,
3345 .deactivate = irte_ga_deactivate,
3346 .set_affinity = irte_ga_set_affinity,
3347 .set_allocated = irte_ga_set_allocated,
3348 .is_allocated = irte_ga_is_allocated,
3349 .clear_allocated = irte_ga_clear_allocated,
3352 static int irq_remapping_alloc(struct irq_domain *domain, unsigned int virq,
3353 unsigned int nr_irqs, void *arg)
3355 struct irq_alloc_info *info = arg;
3356 struct irq_data *irq_data;
3357 struct amd_ir_data *data = NULL;
3358 struct amd_iommu *iommu;
3359 struct irq_cfg *cfg;
3360 int i, ret, devid, seg, sbdf;
3365 if (nr_irqs > 1 && info->type != X86_IRQ_ALLOC_TYPE_PCI_MSI)
3368 sbdf = get_devid(info);
3372 seg = PCI_SBDF_TO_SEGID(sbdf);
3373 devid = PCI_SBDF_TO_DEVID(sbdf);
3374 iommu = __rlookup_amd_iommu(seg, devid);
3378 ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
3382 if (info->type == X86_IRQ_ALLOC_TYPE_IOAPIC) {
3383 struct irq_remap_table *table;
3385 table = alloc_irq_table(iommu, devid, NULL);
3387 if (!table->min_index) {
3389 * Keep the first 32 indexes free for IOAPIC
3392 table->min_index = 32;
3393 for (i = 0; i < 32; ++i)
3394 iommu->irte_ops->set_allocated(table, i);
3396 WARN_ON(table->min_index != 32);
3397 index = info->ioapic.pin;
3401 } else if (info->type == X86_IRQ_ALLOC_TYPE_PCI_MSI ||
3402 info->type == X86_IRQ_ALLOC_TYPE_PCI_MSIX) {
3403 bool align = (info->type == X86_IRQ_ALLOC_TYPE_PCI_MSI);
3405 index = alloc_irq_index(iommu, devid, nr_irqs, align,
3406 msi_desc_to_pci_dev(info->desc));
3408 index = alloc_irq_index(iommu, devid, nr_irqs, false, NULL);
3412 pr_warn("Failed to allocate IRTE\n");
3414 goto out_free_parent;
3417 for (i = 0; i < nr_irqs; i++) {
3418 irq_data = irq_domain_get_irq_data(domain, virq + i);
3419 cfg = irq_data ? irqd_cfg(irq_data) : NULL;
3426 data = kzalloc(sizeof(*data), GFP_KERNEL);
3430 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
3431 data->entry = kzalloc(sizeof(union irte), GFP_KERNEL);
3433 data->entry = kzalloc(sizeof(struct irte_ga),
3440 data->iommu = iommu;
3441 irq_data->hwirq = (devid << 16) + i;
3442 irq_data->chip_data = data;
3443 irq_data->chip = &amd_ir_chip;
3444 irq_remapping_prepare_irte(data, cfg, info, devid, index, i);
3445 irq_set_status_flags(virq + i, IRQ_MOVE_PCNTXT);
3451 for (i--; i >= 0; i--) {
3452 irq_data = irq_domain_get_irq_data(domain, virq + i);
3454 kfree(irq_data->chip_data);
3456 for (i = 0; i < nr_irqs; i++)
3457 free_irte(iommu, devid, index + i);
3459 irq_domain_free_irqs_common(domain, virq, nr_irqs);
3463 static void irq_remapping_free(struct irq_domain *domain, unsigned int virq,
3464 unsigned int nr_irqs)
3466 struct irq_2_irte *irte_info;
3467 struct irq_data *irq_data;
3468 struct amd_ir_data *data;
3471 for (i = 0; i < nr_irqs; i++) {
3472 irq_data = irq_domain_get_irq_data(domain, virq + i);
3473 if (irq_data && irq_data->chip_data) {
3474 data = irq_data->chip_data;
3475 irte_info = &data->irq_2_irte;
3476 free_irte(data->iommu, irte_info->devid, irte_info->index);
3481 irq_domain_free_irqs_common(domain, virq, nr_irqs);
3484 static void amd_ir_update_irte(struct irq_data *irqd, struct amd_iommu *iommu,
3485 struct amd_ir_data *ir_data,
3486 struct irq_2_irte *irte_info,
3487 struct irq_cfg *cfg);
3489 static int irq_remapping_activate(struct irq_domain *domain,
3490 struct irq_data *irq_data, bool reserve)
3492 struct amd_ir_data *data = irq_data->chip_data;
3493 struct irq_2_irte *irte_info = &data->irq_2_irte;
3494 struct amd_iommu *iommu = data->iommu;
3495 struct irq_cfg *cfg = irqd_cfg(irq_data);
3500 iommu->irte_ops->activate(iommu, data->entry, irte_info->devid,
3502 amd_ir_update_irte(irq_data, iommu, data, irte_info, cfg);
3506 static void irq_remapping_deactivate(struct irq_domain *domain,
3507 struct irq_data *irq_data)
3509 struct amd_ir_data *data = irq_data->chip_data;
3510 struct irq_2_irte *irte_info = &data->irq_2_irte;
3511 struct amd_iommu *iommu = data->iommu;
3514 iommu->irte_ops->deactivate(iommu, data->entry, irte_info->devid,
3518 static int irq_remapping_select(struct irq_domain *d, struct irq_fwspec *fwspec,
3519 enum irq_domain_bus_token bus_token)
3521 struct amd_iommu *iommu;
3524 if (!amd_iommu_irq_remap)
3527 if (x86_fwspec_is_ioapic(fwspec))
3528 devid = get_ioapic_devid(fwspec->param[0]);
3529 else if (x86_fwspec_is_hpet(fwspec))
3530 devid = get_hpet_devid(fwspec->param[0]);
3534 iommu = __rlookup_amd_iommu((devid >> 16), (devid & 0xffff));
3536 return iommu && iommu->ir_domain == d;
3539 static const struct irq_domain_ops amd_ir_domain_ops = {
3540 .select = irq_remapping_select,
3541 .alloc = irq_remapping_alloc,
3542 .free = irq_remapping_free,
3543 .activate = irq_remapping_activate,
3544 .deactivate = irq_remapping_deactivate,
3547 int amd_iommu_activate_guest_mode(void *data)
3549 struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
3550 struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
3553 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) || !entry)
3556 valid = entry->lo.fields_vapic.valid;
3561 entry->lo.fields_vapic.valid = valid;
3562 entry->lo.fields_vapic.guest_mode = 1;
3563 entry->lo.fields_vapic.ga_log_intr = 1;
3564 entry->hi.fields.ga_root_ptr = ir_data->ga_root_ptr;
3565 entry->hi.fields.vector = ir_data->ga_vector;
3566 entry->lo.fields_vapic.ga_tag = ir_data->ga_tag;
3568 return modify_irte_ga(ir_data->iommu, ir_data->irq_2_irte.devid,
3569 ir_data->irq_2_irte.index, entry);
3571 EXPORT_SYMBOL(amd_iommu_activate_guest_mode);
3573 int amd_iommu_deactivate_guest_mode(void *data)
3575 struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
3576 struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
3577 struct irq_cfg *cfg = ir_data->cfg;
3580 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
3581 !entry || !entry->lo.fields_vapic.guest_mode)
3584 valid = entry->lo.fields_remap.valid;
3589 entry->lo.fields_remap.valid = valid;
3590 entry->lo.fields_remap.dm = apic->dest_mode_logical;
3591 entry->lo.fields_remap.int_type = apic->delivery_mode;
3592 entry->hi.fields.vector = cfg->vector;
3593 entry->lo.fields_remap.destination =
3594 APICID_TO_IRTE_DEST_LO(cfg->dest_apicid);
3595 entry->hi.fields.destination =
3596 APICID_TO_IRTE_DEST_HI(cfg->dest_apicid);
3598 return modify_irte_ga(ir_data->iommu, ir_data->irq_2_irte.devid,
3599 ir_data->irq_2_irte.index, entry);
3601 EXPORT_SYMBOL(amd_iommu_deactivate_guest_mode);
3603 static int amd_ir_set_vcpu_affinity(struct irq_data *data, void *vcpu_info)
3606 struct amd_iommu_pi_data *pi_data = vcpu_info;
3607 struct vcpu_data *vcpu_pi_info = pi_data->vcpu_data;
3608 struct amd_ir_data *ir_data = data->chip_data;
3609 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
3610 struct iommu_dev_data *dev_data;
3612 if (ir_data->iommu == NULL)
3615 dev_data = search_dev_data(ir_data->iommu, irte_info->devid);
3618 * This device has never been set up for guest mode.
3619 * we should not modify the IRTE
3621 if (!dev_data || !dev_data->use_vapic)
3624 ir_data->cfg = irqd_cfg(data);
3625 pi_data->ir_data = ir_data;
3628 * SVM tries to set up for VAPIC mode, but we are in
3629 * legacy mode. So, we force legacy mode instead.
3631 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
3632 pr_debug("%s: Fall back to using intr legacy remap\n",
3634 pi_data->is_guest_mode = false;
3637 pi_data->prev_ga_tag = ir_data->cached_ga_tag;
3638 if (pi_data->is_guest_mode) {
3639 ir_data->ga_root_ptr = (pi_data->base >> 12);
3640 ir_data->ga_vector = vcpu_pi_info->vector;
3641 ir_data->ga_tag = pi_data->ga_tag;
3642 ret = amd_iommu_activate_guest_mode(ir_data);
3644 ir_data->cached_ga_tag = pi_data->ga_tag;
3646 ret = amd_iommu_deactivate_guest_mode(ir_data);
3649 * This communicates the ga_tag back to the caller
3650 * so that it can do all the necessary clean up.
3653 ir_data->cached_ga_tag = 0;
3660 static void amd_ir_update_irte(struct irq_data *irqd, struct amd_iommu *iommu,
3661 struct amd_ir_data *ir_data,
3662 struct irq_2_irte *irte_info,
3663 struct irq_cfg *cfg)
3667 * Atomically updates the IRTE with the new destination, vector
3668 * and flushes the interrupt entry cache.
3670 iommu->irte_ops->set_affinity(iommu, ir_data->entry, irte_info->devid,
3671 irte_info->index, cfg->vector,
3675 static int amd_ir_set_affinity(struct irq_data *data,
3676 const struct cpumask *mask, bool force)
3678 struct amd_ir_data *ir_data = data->chip_data;
3679 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
3680 struct irq_cfg *cfg = irqd_cfg(data);
3681 struct irq_data *parent = data->parent_data;
3682 struct amd_iommu *iommu = ir_data->iommu;
3688 ret = parent->chip->irq_set_affinity(parent, mask, force);
3689 if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
3692 amd_ir_update_irte(data, iommu, ir_data, irte_info, cfg);
3694 * After this point, all the interrupts will start arriving
3695 * at the new destination. So, time to cleanup the previous
3696 * vector allocation.
3698 vector_schedule_cleanup(cfg);
3700 return IRQ_SET_MASK_OK_DONE;
3703 static void ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg)
3705 struct amd_ir_data *ir_data = irq_data->chip_data;
3707 *msg = ir_data->msi_entry;
3710 static struct irq_chip amd_ir_chip = {
3712 .irq_ack = apic_ack_irq,
3713 .irq_set_affinity = amd_ir_set_affinity,
3714 .irq_set_vcpu_affinity = amd_ir_set_vcpu_affinity,
3715 .irq_compose_msi_msg = ir_compose_msi_msg,
3718 static const struct msi_parent_ops amdvi_msi_parent_ops = {
3719 .supported_flags = X86_VECTOR_MSI_FLAGS_SUPPORTED |
3720 MSI_FLAG_MULTI_PCI_MSI |
3723 .init_dev_msi_info = msi_parent_init_dev_msi_info,
3726 static const struct msi_parent_ops virt_amdvi_msi_parent_ops = {
3727 .supported_flags = X86_VECTOR_MSI_FLAGS_SUPPORTED |
3728 MSI_FLAG_MULTI_PCI_MSI,
3730 .init_dev_msi_info = msi_parent_init_dev_msi_info,
3733 int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
3735 struct fwnode_handle *fn;
3737 fn = irq_domain_alloc_named_id_fwnode("AMD-IR", iommu->index);
3740 iommu->ir_domain = irq_domain_create_hierarchy(arch_get_ir_parent_domain(), 0, 0,
3741 fn, &amd_ir_domain_ops, iommu);
3742 if (!iommu->ir_domain) {
3743 irq_domain_free_fwnode(fn);
3747 irq_domain_update_bus_token(iommu->ir_domain, DOMAIN_BUS_AMDVI);
3748 iommu->ir_domain->flags |= IRQ_DOMAIN_FLAG_MSI_PARENT |
3749 IRQ_DOMAIN_FLAG_ISOLATED_MSI;
3751 if (amd_iommu_np_cache)
3752 iommu->ir_domain->msi_parent_ops = &virt_amdvi_msi_parent_ops;
3754 iommu->ir_domain->msi_parent_ops = &amdvi_msi_parent_ops;
3759 int amd_iommu_update_ga(int cpu, bool is_run, void *data)
3761 struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
3762 struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
3764 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
3765 !entry || !entry->lo.fields_vapic.guest_mode)
3768 if (!ir_data->iommu)
3772 entry->lo.fields_vapic.destination =
3773 APICID_TO_IRTE_DEST_LO(cpu);
3774 entry->hi.fields.destination =
3775 APICID_TO_IRTE_DEST_HI(cpu);
3777 entry->lo.fields_vapic.is_run = is_run;
3779 return modify_irte_ga(ir_data->iommu, ir_data->irq_2_irte.devid,
3780 ir_data->irq_2_irte.index, entry);
3782 EXPORT_SYMBOL(amd_iommu_update_ga);