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 #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 (amd_iommu_gpt_level == PAGE_MODE_5_LEVEL) {
1615 dev_table[devid].data[2] |=
1616 ((u64)GUEST_PGTABLE_5_LEVEL << DTE_GPT_LEVEL_SHIFT);
1619 if (domain->flags & PD_GIOV_MASK)
1620 pte_root |= DTE_FLAG_GIOV;
1623 flags &= ~DEV_DOMID_MASK;
1624 flags |= domain->id;
1626 old_domid = dev_table[devid].data[1] & DEV_DOMID_MASK;
1627 dev_table[devid].data[1] = flags;
1628 dev_table[devid].data[0] = pte_root;
1631 * A kdump kernel might be replacing a domain ID that was copied from
1632 * the previous kernel--if so, it needs to flush the translation cache
1633 * entries for the old domain ID that is being overwritten
1636 amd_iommu_flush_tlb_domid(iommu, old_domid);
1640 static void clear_dte_entry(struct amd_iommu *iommu, u16 devid)
1642 struct dev_table_entry *dev_table = get_dev_table(iommu);
1644 /* remove entry from the device table seen by the hardware */
1645 dev_table[devid].data[0] = DTE_FLAG_V;
1647 if (!amd_iommu_snp_en)
1648 dev_table[devid].data[0] |= DTE_FLAG_TV;
1650 dev_table[devid].data[1] &= DTE_FLAG_MASK;
1652 amd_iommu_apply_erratum_63(iommu, devid);
1655 static void do_attach(struct iommu_dev_data *dev_data,
1656 struct protection_domain *domain)
1658 struct amd_iommu *iommu;
1661 iommu = rlookup_amd_iommu(dev_data->dev);
1664 ats = dev_data->ats.enabled;
1666 /* Update data structures */
1667 dev_data->domain = domain;
1668 list_add(&dev_data->list, &domain->dev_list);
1670 /* Update NUMA Node ID */
1671 if (domain->nid == NUMA_NO_NODE)
1672 domain->nid = dev_to_node(dev_data->dev);
1674 /* Do reference counting */
1675 domain->dev_iommu[iommu->index] += 1;
1676 domain->dev_cnt += 1;
1678 /* Update device table */
1679 set_dte_entry(iommu, dev_data->devid, domain,
1680 ats, dev_data->iommu_v2);
1681 clone_aliases(iommu, dev_data->dev);
1683 device_flush_dte(dev_data);
1686 static void do_detach(struct iommu_dev_data *dev_data)
1688 struct protection_domain *domain = dev_data->domain;
1689 struct amd_iommu *iommu;
1691 iommu = rlookup_amd_iommu(dev_data->dev);
1695 /* Update data structures */
1696 dev_data->domain = NULL;
1697 list_del(&dev_data->list);
1698 clear_dte_entry(iommu, dev_data->devid);
1699 clone_aliases(iommu, dev_data->dev);
1701 /* Flush the DTE entry */
1702 device_flush_dte(dev_data);
1705 amd_iommu_domain_flush_tlb_pde(domain);
1707 /* Wait for the flushes to finish */
1708 amd_iommu_domain_flush_complete(domain);
1710 /* decrease reference counters - needs to happen after the flushes */
1711 domain->dev_iommu[iommu->index] -= 1;
1712 domain->dev_cnt -= 1;
1715 static void pdev_iommuv2_disable(struct pci_dev *pdev)
1717 pci_disable_ats(pdev);
1718 pci_disable_pri(pdev);
1719 pci_disable_pasid(pdev);
1722 static int pdev_pri_ats_enable(struct pci_dev *pdev)
1726 /* Only allow access to user-accessible pages */
1727 ret = pci_enable_pasid(pdev, 0);
1731 /* First reset the PRI state of the device */
1732 ret = pci_reset_pri(pdev);
1737 /* FIXME: Hardcode number of outstanding requests for now */
1738 ret = pci_enable_pri(pdev, 32);
1742 ret = pci_enable_ats(pdev, PAGE_SHIFT);
1749 pci_disable_pri(pdev);
1752 pci_disable_pasid(pdev);
1758 * If a device is not yet associated with a domain, this function makes the
1759 * device visible in the domain
1761 static int attach_device(struct device *dev,
1762 struct protection_domain *domain)
1764 struct iommu_dev_data *dev_data;
1765 struct pci_dev *pdev;
1766 unsigned long flags;
1769 spin_lock_irqsave(&domain->lock, flags);
1771 dev_data = dev_iommu_priv_get(dev);
1773 spin_lock(&dev_data->lock);
1776 if (dev_data->domain != NULL)
1779 if (!dev_is_pci(dev))
1780 goto skip_ats_check;
1782 pdev = to_pci_dev(dev);
1783 if (domain->flags & PD_IOMMUV2_MASK) {
1784 struct iommu_domain *def_domain = iommu_get_dma_domain(dev);
1789 * In case of using AMD_IOMMU_V1 page table mode and the device
1790 * is enabling for PPR/ATS support (using v2 table),
1791 * we need to make sure that the domain type is identity map.
1793 if ((amd_iommu_pgtable == AMD_IOMMU_V1) &&
1794 def_domain->type != IOMMU_DOMAIN_IDENTITY) {
1798 if (dev_data->iommu_v2) {
1799 if (pdev_pri_ats_enable(pdev) != 0)
1802 dev_data->ats.enabled = true;
1803 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
1804 dev_data->pri_tlp = pci_prg_resp_pasid_required(pdev);
1806 } else if (amd_iommu_iotlb_sup &&
1807 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
1808 dev_data->ats.enabled = true;
1809 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
1815 do_attach(dev_data, domain);
1818 * We might boot into a crash-kernel here. The crashed kernel
1819 * left the caches in the IOMMU dirty. So we have to flush
1820 * here to evict all dirty stuff.
1822 amd_iommu_domain_flush_tlb_pde(domain);
1824 amd_iommu_domain_flush_complete(domain);
1827 spin_unlock(&dev_data->lock);
1829 spin_unlock_irqrestore(&domain->lock, flags);
1835 * Removes a device from a protection domain (with devtable_lock held)
1837 static void detach_device(struct device *dev)
1839 struct protection_domain *domain;
1840 struct iommu_dev_data *dev_data;
1841 unsigned long flags;
1843 dev_data = dev_iommu_priv_get(dev);
1844 domain = dev_data->domain;
1846 spin_lock_irqsave(&domain->lock, flags);
1848 spin_lock(&dev_data->lock);
1851 * First check if the device is still attached. It might already
1852 * be detached from its domain because the generic
1853 * iommu_detach_group code detached it and we try again here in
1854 * our alias handling.
1856 if (WARN_ON(!dev_data->domain))
1859 do_detach(dev_data);
1861 if (!dev_is_pci(dev))
1864 if (domain->flags & PD_IOMMUV2_MASK && dev_data->iommu_v2)
1865 pdev_iommuv2_disable(to_pci_dev(dev));
1866 else if (dev_data->ats.enabled)
1867 pci_disable_ats(to_pci_dev(dev));
1869 dev_data->ats.enabled = false;
1872 spin_unlock(&dev_data->lock);
1874 spin_unlock_irqrestore(&domain->lock, flags);
1877 static struct iommu_device *amd_iommu_probe_device(struct device *dev)
1879 struct iommu_device *iommu_dev;
1880 struct amd_iommu *iommu;
1883 if (!check_device(dev))
1884 return ERR_PTR(-ENODEV);
1886 iommu = rlookup_amd_iommu(dev);
1888 return ERR_PTR(-ENODEV);
1890 /* Not registered yet? */
1891 if (!iommu->iommu.ops)
1892 return ERR_PTR(-ENODEV);
1894 if (dev_iommu_priv_get(dev))
1895 return &iommu->iommu;
1897 ret = iommu_init_device(iommu, dev);
1899 if (ret != -ENOTSUPP)
1900 dev_err(dev, "Failed to initialize - trying to proceed anyway\n");
1901 iommu_dev = ERR_PTR(ret);
1902 iommu_ignore_device(iommu, dev);
1904 amd_iommu_set_pci_msi_domain(dev, iommu);
1905 iommu_dev = &iommu->iommu;
1908 iommu_completion_wait(iommu);
1913 static void amd_iommu_probe_finalize(struct device *dev)
1915 /* Domains are initialized for this device - have a look what we ended up with */
1916 set_dma_ops(dev, NULL);
1917 iommu_setup_dma_ops(dev, 0, U64_MAX);
1920 static void amd_iommu_release_device(struct device *dev)
1922 struct amd_iommu *iommu;
1924 if (!check_device(dev))
1927 iommu = rlookup_amd_iommu(dev);
1931 amd_iommu_uninit_device(dev);
1932 iommu_completion_wait(iommu);
1935 static struct iommu_group *amd_iommu_device_group(struct device *dev)
1937 if (dev_is_pci(dev))
1938 return pci_device_group(dev);
1940 return acpihid_device_group(dev);
1943 /*****************************************************************************
1945 * The next functions belong to the dma_ops mapping/unmapping code.
1947 *****************************************************************************/
1949 static void update_device_table(struct protection_domain *domain)
1951 struct iommu_dev_data *dev_data;
1953 list_for_each_entry(dev_data, &domain->dev_list, list) {
1954 struct amd_iommu *iommu = rlookup_amd_iommu(dev_data->dev);
1958 set_dte_entry(iommu, dev_data->devid, domain,
1959 dev_data->ats.enabled, dev_data->iommu_v2);
1960 clone_aliases(iommu, dev_data->dev);
1964 void amd_iommu_update_and_flush_device_table(struct protection_domain *domain)
1966 update_device_table(domain);
1967 domain_flush_devices(domain);
1970 void amd_iommu_domain_update(struct protection_domain *domain)
1972 /* Update device table */
1973 amd_iommu_update_and_flush_device_table(domain);
1975 /* Flush domain TLB(s) and wait for completion */
1976 amd_iommu_domain_flush_tlb_pde(domain);
1977 amd_iommu_domain_flush_complete(domain);
1980 /*****************************************************************************
1982 * The following functions belong to the exported interface of AMD IOMMU
1984 * This interface allows access to lower level functions of the IOMMU
1985 * like protection domain handling and assignement of devices to domains
1986 * which is not possible with the dma_ops interface.
1988 *****************************************************************************/
1990 static void cleanup_domain(struct protection_domain *domain)
1992 struct iommu_dev_data *entry;
1993 unsigned long flags;
1995 spin_lock_irqsave(&domain->lock, flags);
1997 while (!list_empty(&domain->dev_list)) {
1998 entry = list_first_entry(&domain->dev_list,
1999 struct iommu_dev_data, list);
2000 BUG_ON(!entry->domain);
2004 spin_unlock_irqrestore(&domain->lock, flags);
2007 static void protection_domain_free(struct protection_domain *domain)
2012 if (domain->iop.pgtbl_cfg.tlb)
2013 free_io_pgtable_ops(&domain->iop.iop.ops);
2016 domain_id_free(domain->id);
2021 static int protection_domain_init_v1(struct protection_domain *domain, int mode)
2023 u64 *pt_root = NULL;
2025 BUG_ON(mode < PAGE_MODE_NONE || mode > PAGE_MODE_6_LEVEL);
2027 spin_lock_init(&domain->lock);
2028 domain->id = domain_id_alloc();
2031 INIT_LIST_HEAD(&domain->dev_list);
2033 if (mode != PAGE_MODE_NONE) {
2034 pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2036 domain_id_free(domain->id);
2041 amd_iommu_domain_set_pgtable(domain, pt_root, mode);
2046 static int protection_domain_init_v2(struct protection_domain *domain)
2048 spin_lock_init(&domain->lock);
2049 domain->id = domain_id_alloc();
2052 INIT_LIST_HEAD(&domain->dev_list);
2054 domain->flags |= PD_GIOV_MASK;
2056 domain->domain.pgsize_bitmap = AMD_IOMMU_PGSIZES_V2;
2058 if (domain_enable_v2(domain, 1)) {
2059 domain_id_free(domain->id);
2066 static struct protection_domain *protection_domain_alloc(unsigned int type)
2068 struct io_pgtable_ops *pgtbl_ops;
2069 struct protection_domain *domain;
2070 int pgtable = amd_iommu_pgtable;
2071 int mode = DEFAULT_PGTABLE_LEVEL;
2074 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2079 * Force IOMMU v1 page table when iommu=pt and
2080 * when allocating domain for pass-through devices.
2082 if (type == IOMMU_DOMAIN_IDENTITY) {
2083 pgtable = AMD_IOMMU_V1;
2084 mode = PAGE_MODE_NONE;
2085 } else if (type == IOMMU_DOMAIN_UNMANAGED) {
2086 pgtable = AMD_IOMMU_V1;
2091 ret = protection_domain_init_v1(domain, mode);
2094 ret = protection_domain_init_v2(domain);
2103 /* No need to allocate io pgtable ops in passthrough mode */
2104 if (type == IOMMU_DOMAIN_IDENTITY)
2107 domain->nid = NUMA_NO_NODE;
2109 pgtbl_ops = alloc_io_pgtable_ops(pgtable, &domain->iop.pgtbl_cfg, domain);
2111 domain_id_free(domain->id);
2121 static struct iommu_domain *amd_iommu_domain_alloc(unsigned type)
2123 struct protection_domain *domain;
2126 * Since DTE[Mode]=0 is prohibited on SNP-enabled system,
2127 * default to use IOMMU_DOMAIN_DMA[_FQ].
2129 if (amd_iommu_snp_en && (type == IOMMU_DOMAIN_IDENTITY))
2132 domain = protection_domain_alloc(type);
2136 domain->domain.geometry.aperture_start = 0;
2137 domain->domain.geometry.aperture_end = ~0ULL;
2138 domain->domain.geometry.force_aperture = true;
2140 return &domain->domain;
2143 static void amd_iommu_domain_free(struct iommu_domain *dom)
2145 struct protection_domain *domain;
2147 domain = to_pdomain(dom);
2149 if (domain->dev_cnt > 0)
2150 cleanup_domain(domain);
2152 BUG_ON(domain->dev_cnt != 0);
2157 if (domain->flags & PD_IOMMUV2_MASK)
2158 free_gcr3_table(domain);
2160 protection_domain_free(domain);
2163 static int amd_iommu_attach_device(struct iommu_domain *dom,
2166 struct iommu_dev_data *dev_data = dev_iommu_priv_get(dev);
2167 struct protection_domain *domain = to_pdomain(dom);
2168 struct amd_iommu *iommu = rlookup_amd_iommu(dev);
2172 * Skip attach device to domain if new domain is same as
2173 * devices current domain
2175 if (dev_data->domain == domain)
2178 dev_data->defer_attach = false;
2180 if (dev_data->domain)
2183 ret = attach_device(dev, domain);
2185 #ifdef CONFIG_IRQ_REMAP
2186 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
2187 if (dom->type == IOMMU_DOMAIN_UNMANAGED)
2188 dev_data->use_vapic = 1;
2190 dev_data->use_vapic = 0;
2194 iommu_completion_wait(iommu);
2199 static void amd_iommu_iotlb_sync_map(struct iommu_domain *dom,
2200 unsigned long iova, size_t size)
2202 struct protection_domain *domain = to_pdomain(dom);
2203 struct io_pgtable_ops *ops = &domain->iop.iop.ops;
2206 domain_flush_np_cache(domain, iova, size);
2209 static int amd_iommu_map_pages(struct iommu_domain *dom, unsigned long iova,
2210 phys_addr_t paddr, size_t pgsize, size_t pgcount,
2211 int iommu_prot, gfp_t gfp, size_t *mapped)
2213 struct protection_domain *domain = to_pdomain(dom);
2214 struct io_pgtable_ops *ops = &domain->iop.iop.ops;
2218 if ((amd_iommu_pgtable == AMD_IOMMU_V1) &&
2219 (domain->iop.mode == PAGE_MODE_NONE))
2222 if (iommu_prot & IOMMU_READ)
2223 prot |= IOMMU_PROT_IR;
2224 if (iommu_prot & IOMMU_WRITE)
2225 prot |= IOMMU_PROT_IW;
2227 if (ops->map_pages) {
2228 ret = ops->map_pages(ops, iova, paddr, pgsize,
2229 pgcount, prot, gfp, mapped);
2235 static void amd_iommu_iotlb_gather_add_page(struct iommu_domain *domain,
2236 struct iommu_iotlb_gather *gather,
2237 unsigned long iova, size_t size)
2240 * AMD's IOMMU can flush as many pages as necessary in a single flush.
2241 * Unless we run in a virtual machine, which can be inferred according
2242 * to whether "non-present cache" is on, it is probably best to prefer
2243 * (potentially) too extensive TLB flushing (i.e., more misses) over
2244 * mutliple TLB flushes (i.e., more flushes). For virtual machines the
2245 * hypervisor needs to synchronize the host IOMMU PTEs with those of
2246 * the guest, and the trade-off is different: unnecessary TLB flushes
2247 * should be avoided.
2249 if (amd_iommu_np_cache &&
2250 iommu_iotlb_gather_is_disjoint(gather, iova, size))
2251 iommu_iotlb_sync(domain, gather);
2253 iommu_iotlb_gather_add_range(gather, iova, size);
2256 static size_t amd_iommu_unmap_pages(struct iommu_domain *dom, unsigned long iova,
2257 size_t pgsize, size_t pgcount,
2258 struct iommu_iotlb_gather *gather)
2260 struct protection_domain *domain = to_pdomain(dom);
2261 struct io_pgtable_ops *ops = &domain->iop.iop.ops;
2264 if ((amd_iommu_pgtable == AMD_IOMMU_V1) &&
2265 (domain->iop.mode == PAGE_MODE_NONE))
2268 r = (ops->unmap_pages) ? ops->unmap_pages(ops, iova, pgsize, pgcount, NULL) : 0;
2271 amd_iommu_iotlb_gather_add_page(dom, gather, iova, r);
2276 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
2279 struct protection_domain *domain = to_pdomain(dom);
2280 struct io_pgtable_ops *ops = &domain->iop.iop.ops;
2282 return ops->iova_to_phys(ops, iova);
2285 static bool amd_iommu_capable(struct device *dev, enum iommu_cap cap)
2288 case IOMMU_CAP_CACHE_COHERENCY:
2290 case IOMMU_CAP_NOEXEC:
2292 case IOMMU_CAP_PRE_BOOT_PROTECTION:
2293 return amdr_ivrs_remap_support;
2294 case IOMMU_CAP_ENFORCE_CACHE_COHERENCY:
2303 static void amd_iommu_get_resv_regions(struct device *dev,
2304 struct list_head *head)
2306 struct iommu_resv_region *region;
2307 struct unity_map_entry *entry;
2308 struct amd_iommu *iommu;
2309 struct amd_iommu_pci_seg *pci_seg;
2312 sbdf = get_device_sbdf_id(dev);
2316 devid = PCI_SBDF_TO_DEVID(sbdf);
2317 iommu = rlookup_amd_iommu(dev);
2320 pci_seg = iommu->pci_seg;
2322 list_for_each_entry(entry, &pci_seg->unity_map, list) {
2326 if (devid < entry->devid_start || devid > entry->devid_end)
2329 type = IOMMU_RESV_DIRECT;
2330 length = entry->address_end - entry->address_start;
2331 if (entry->prot & IOMMU_PROT_IR)
2333 if (entry->prot & IOMMU_PROT_IW)
2334 prot |= IOMMU_WRITE;
2335 if (entry->prot & IOMMU_UNITY_MAP_FLAG_EXCL_RANGE)
2336 /* Exclusion range */
2337 type = IOMMU_RESV_RESERVED;
2339 region = iommu_alloc_resv_region(entry->address_start,
2343 dev_err(dev, "Out of memory allocating dm-regions\n");
2346 list_add_tail(®ion->list, head);
2349 region = iommu_alloc_resv_region(MSI_RANGE_START,
2350 MSI_RANGE_END - MSI_RANGE_START + 1,
2351 0, IOMMU_RESV_MSI, GFP_KERNEL);
2354 list_add_tail(®ion->list, head);
2356 region = iommu_alloc_resv_region(HT_RANGE_START,
2357 HT_RANGE_END - HT_RANGE_START + 1,
2358 0, IOMMU_RESV_RESERVED, GFP_KERNEL);
2361 list_add_tail(®ion->list, head);
2364 bool amd_iommu_is_attach_deferred(struct device *dev)
2366 struct iommu_dev_data *dev_data = dev_iommu_priv_get(dev);
2368 return dev_data->defer_attach;
2370 EXPORT_SYMBOL_GPL(amd_iommu_is_attach_deferred);
2372 static void amd_iommu_flush_iotlb_all(struct iommu_domain *domain)
2374 struct protection_domain *dom = to_pdomain(domain);
2375 unsigned long flags;
2377 spin_lock_irqsave(&dom->lock, flags);
2378 amd_iommu_domain_flush_tlb_pde(dom);
2379 amd_iommu_domain_flush_complete(dom);
2380 spin_unlock_irqrestore(&dom->lock, flags);
2383 static void amd_iommu_iotlb_sync(struct iommu_domain *domain,
2384 struct iommu_iotlb_gather *gather)
2386 struct protection_domain *dom = to_pdomain(domain);
2387 unsigned long flags;
2389 spin_lock_irqsave(&dom->lock, flags);
2390 domain_flush_pages(dom, gather->start, gather->end - gather->start, 1);
2391 amd_iommu_domain_flush_complete(dom);
2392 spin_unlock_irqrestore(&dom->lock, flags);
2395 static int amd_iommu_def_domain_type(struct device *dev)
2397 struct iommu_dev_data *dev_data;
2399 dev_data = dev_iommu_priv_get(dev);
2404 * Do not identity map IOMMUv2 capable devices when:
2405 * - memory encryption is active, because some of those devices
2406 * (AMD GPUs) don't have the encryption bit in their DMA-mask
2407 * and require remapping.
2408 * - SNP is enabled, because it prohibits DTE[Mode]=0.
2410 if (dev_data->iommu_v2 &&
2411 !cc_platform_has(CC_ATTR_MEM_ENCRYPT) &&
2412 !amd_iommu_snp_en) {
2413 return IOMMU_DOMAIN_IDENTITY;
2419 static bool amd_iommu_enforce_cache_coherency(struct iommu_domain *domain)
2421 /* IOMMU_PTE_FC is always set */
2425 const struct iommu_ops amd_iommu_ops = {
2426 .capable = amd_iommu_capable,
2427 .domain_alloc = amd_iommu_domain_alloc,
2428 .probe_device = amd_iommu_probe_device,
2429 .release_device = amd_iommu_release_device,
2430 .probe_finalize = amd_iommu_probe_finalize,
2431 .device_group = amd_iommu_device_group,
2432 .get_resv_regions = amd_iommu_get_resv_regions,
2433 .is_attach_deferred = amd_iommu_is_attach_deferred,
2434 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
2435 .def_domain_type = amd_iommu_def_domain_type,
2436 .default_domain_ops = &(const struct iommu_domain_ops) {
2437 .attach_dev = amd_iommu_attach_device,
2438 .map_pages = amd_iommu_map_pages,
2439 .unmap_pages = amd_iommu_unmap_pages,
2440 .iotlb_sync_map = amd_iommu_iotlb_sync_map,
2441 .iova_to_phys = amd_iommu_iova_to_phys,
2442 .flush_iotlb_all = amd_iommu_flush_iotlb_all,
2443 .iotlb_sync = amd_iommu_iotlb_sync,
2444 .free = amd_iommu_domain_free,
2445 .enforce_cache_coherency = amd_iommu_enforce_cache_coherency,
2449 /*****************************************************************************
2451 * The next functions do a basic initialization of IOMMU for pass through
2454 * In passthrough mode the IOMMU is initialized and enabled but not used for
2455 * DMA-API translation.
2457 *****************************************************************************/
2459 /* IOMMUv2 specific functions */
2460 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
2462 return atomic_notifier_chain_register(&ppr_notifier, nb);
2464 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
2466 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
2468 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
2470 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
2472 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
2474 struct protection_domain *domain = to_pdomain(dom);
2475 unsigned long flags;
2477 spin_lock_irqsave(&domain->lock, flags);
2479 if (domain->iop.pgtbl_cfg.tlb)
2480 free_io_pgtable_ops(&domain->iop.iop.ops);
2482 spin_unlock_irqrestore(&domain->lock, flags);
2484 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
2486 /* Note: This function expects iommu_domain->lock to be held prior calling the function. */
2487 static int domain_enable_v2(struct protection_domain *domain, int pasids)
2491 /* Number of GCR3 table levels required */
2492 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
2495 if (levels > amd_iommu_max_glx_val)
2498 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
2499 if (domain->gcr3_tbl == NULL)
2502 domain->glx = levels;
2503 domain->flags |= PD_IOMMUV2_MASK;
2505 amd_iommu_domain_update(domain);
2510 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
2512 struct protection_domain *pdom = to_pdomain(dom);
2513 unsigned long flags;
2516 spin_lock_irqsave(&pdom->lock, flags);
2519 * Save us all sanity checks whether devices already in the
2520 * domain support IOMMUv2. Just force that the domain has no
2521 * devices attached when it is switched into IOMMUv2 mode.
2524 if (pdom->dev_cnt > 0 || pdom->flags & PD_IOMMUV2_MASK)
2527 if (!pdom->gcr3_tbl)
2528 ret = domain_enable_v2(pdom, pasids);
2531 spin_unlock_irqrestore(&pdom->lock, flags);
2534 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
2536 static int __flush_pasid(struct protection_domain *domain, u32 pasid,
2537 u64 address, bool size)
2539 struct iommu_dev_data *dev_data;
2540 struct iommu_cmd cmd;
2543 if (!(domain->flags & PD_IOMMUV2_MASK))
2546 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
2549 * IOMMU TLB needs to be flushed before Device TLB to
2550 * prevent device TLB refill from IOMMU TLB
2552 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
2553 if (domain->dev_iommu[i] == 0)
2556 ret = iommu_queue_command(amd_iommus[i], &cmd);
2561 /* Wait until IOMMU TLB flushes are complete */
2562 amd_iommu_domain_flush_complete(domain);
2564 /* Now flush device TLBs */
2565 list_for_each_entry(dev_data, &domain->dev_list, list) {
2566 struct amd_iommu *iommu;
2570 There might be non-IOMMUv2 capable devices in an IOMMUv2
2573 if (!dev_data->ats.enabled)
2576 qdep = dev_data->ats.qdep;
2577 iommu = rlookup_amd_iommu(dev_data->dev);
2580 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
2581 qdep, address, size);
2583 ret = iommu_queue_command(iommu, &cmd);
2588 /* Wait until all device TLBs are flushed */
2589 amd_iommu_domain_flush_complete(domain);
2598 static int __amd_iommu_flush_page(struct protection_domain *domain, u32 pasid,
2601 return __flush_pasid(domain, pasid, address, false);
2604 int amd_iommu_flush_page(struct iommu_domain *dom, u32 pasid,
2607 struct protection_domain *domain = to_pdomain(dom);
2608 unsigned long flags;
2611 spin_lock_irqsave(&domain->lock, flags);
2612 ret = __amd_iommu_flush_page(domain, pasid, address);
2613 spin_unlock_irqrestore(&domain->lock, flags);
2617 EXPORT_SYMBOL(amd_iommu_flush_page);
2619 static int __amd_iommu_flush_tlb(struct protection_domain *domain, u32 pasid)
2621 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
2625 int amd_iommu_flush_tlb(struct iommu_domain *dom, u32 pasid)
2627 struct protection_domain *domain = to_pdomain(dom);
2628 unsigned long flags;
2631 spin_lock_irqsave(&domain->lock, flags);
2632 ret = __amd_iommu_flush_tlb(domain, pasid);
2633 spin_unlock_irqrestore(&domain->lock, flags);
2637 EXPORT_SYMBOL(amd_iommu_flush_tlb);
2639 static u64 *__get_gcr3_pte(u64 *root, int level, u32 pasid, bool alloc)
2646 index = (pasid >> (9 * level)) & 0x1ff;
2652 if (!(*pte & GCR3_VALID)) {
2656 root = (void *)get_zeroed_page(GFP_ATOMIC);
2660 *pte = iommu_virt_to_phys(root) | GCR3_VALID;
2663 root = iommu_phys_to_virt(*pte & PAGE_MASK);
2671 static int __set_gcr3(struct protection_domain *domain, u32 pasid,
2676 if (domain->iop.mode != PAGE_MODE_NONE)
2679 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
2683 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
2685 return __amd_iommu_flush_tlb(domain, pasid);
2688 static int __clear_gcr3(struct protection_domain *domain, u32 pasid)
2692 if (domain->iop.mode != PAGE_MODE_NONE)
2695 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
2701 return __amd_iommu_flush_tlb(domain, pasid);
2704 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, u32 pasid,
2707 struct protection_domain *domain = to_pdomain(dom);
2708 unsigned long flags;
2711 spin_lock_irqsave(&domain->lock, flags);
2712 ret = __set_gcr3(domain, pasid, cr3);
2713 spin_unlock_irqrestore(&domain->lock, flags);
2717 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
2719 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, u32 pasid)
2721 struct protection_domain *domain = to_pdomain(dom);
2722 unsigned long flags;
2725 spin_lock_irqsave(&domain->lock, flags);
2726 ret = __clear_gcr3(domain, pasid);
2727 spin_unlock_irqrestore(&domain->lock, flags);
2731 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
2733 int amd_iommu_complete_ppr(struct pci_dev *pdev, u32 pasid,
2734 int status, int tag)
2736 struct iommu_dev_data *dev_data;
2737 struct amd_iommu *iommu;
2738 struct iommu_cmd cmd;
2740 dev_data = dev_iommu_priv_get(&pdev->dev);
2741 iommu = rlookup_amd_iommu(&pdev->dev);
2745 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
2746 tag, dev_data->pri_tlp);
2748 return iommu_queue_command(iommu, &cmd);
2750 EXPORT_SYMBOL(amd_iommu_complete_ppr);
2752 int amd_iommu_device_info(struct pci_dev *pdev,
2753 struct amd_iommu_device_info *info)
2758 if (pdev == NULL || info == NULL)
2761 if (!amd_iommu_v2_supported())
2764 memset(info, 0, sizeof(*info));
2766 if (pci_ats_supported(pdev))
2767 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
2769 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2771 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
2773 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
2777 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
2778 max_pasids = min(max_pasids, (1 << 20));
2780 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
2781 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
2783 features = pci_pasid_features(pdev);
2784 if (features & PCI_PASID_CAP_EXEC)
2785 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
2786 if (features & PCI_PASID_CAP_PRIV)
2787 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
2792 EXPORT_SYMBOL(amd_iommu_device_info);
2794 #ifdef CONFIG_IRQ_REMAP
2796 /*****************************************************************************
2798 * Interrupt Remapping Implementation
2800 *****************************************************************************/
2802 static struct irq_chip amd_ir_chip;
2803 static DEFINE_SPINLOCK(iommu_table_lock);
2805 static void set_dte_irq_entry(struct amd_iommu *iommu, u16 devid,
2806 struct irq_remap_table *table)
2809 struct dev_table_entry *dev_table = get_dev_table(iommu);
2811 dte = dev_table[devid].data[2];
2812 dte &= ~DTE_IRQ_PHYS_ADDR_MASK;
2813 dte |= iommu_virt_to_phys(table->table);
2814 dte |= DTE_IRQ_REMAP_INTCTL;
2815 dte |= DTE_INTTABLEN;
2816 dte |= DTE_IRQ_REMAP_ENABLE;
2818 dev_table[devid].data[2] = dte;
2821 static struct irq_remap_table *get_irq_table(struct amd_iommu *iommu, u16 devid)
2823 struct irq_remap_table *table;
2824 struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
2826 if (WARN_ONCE(!pci_seg->rlookup_table[devid],
2827 "%s: no iommu for devid %x:%x\n",
2828 __func__, pci_seg->id, devid))
2831 table = pci_seg->irq_lookup_table[devid];
2832 if (WARN_ONCE(!table, "%s: no table for devid %x:%x\n",
2833 __func__, pci_seg->id, devid))
2839 static struct irq_remap_table *__alloc_irq_table(void)
2841 struct irq_remap_table *table;
2843 table = kzalloc(sizeof(*table), GFP_KERNEL);
2847 table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_KERNEL);
2848 if (!table->table) {
2852 raw_spin_lock_init(&table->lock);
2854 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
2855 memset(table->table, 0,
2856 MAX_IRQS_PER_TABLE * sizeof(u32));
2858 memset(table->table, 0,
2859 (MAX_IRQS_PER_TABLE * (sizeof(u64) * 2)));
2863 static void set_remap_table_entry(struct amd_iommu *iommu, u16 devid,
2864 struct irq_remap_table *table)
2866 struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
2868 pci_seg->irq_lookup_table[devid] = table;
2869 set_dte_irq_entry(iommu, devid, table);
2870 iommu_flush_dte(iommu, devid);
2873 static int set_remap_table_entry_alias(struct pci_dev *pdev, u16 alias,
2876 struct irq_remap_table *table = data;
2877 struct amd_iommu_pci_seg *pci_seg;
2878 struct amd_iommu *iommu = rlookup_amd_iommu(&pdev->dev);
2883 pci_seg = iommu->pci_seg;
2884 pci_seg->irq_lookup_table[alias] = table;
2885 set_dte_irq_entry(iommu, alias, table);
2886 iommu_flush_dte(pci_seg->rlookup_table[alias], alias);
2891 static struct irq_remap_table *alloc_irq_table(struct amd_iommu *iommu,
2892 u16 devid, struct pci_dev *pdev)
2894 struct irq_remap_table *table = NULL;
2895 struct irq_remap_table *new_table = NULL;
2896 struct amd_iommu_pci_seg *pci_seg;
2897 unsigned long flags;
2900 spin_lock_irqsave(&iommu_table_lock, flags);
2902 pci_seg = iommu->pci_seg;
2903 table = pci_seg->irq_lookup_table[devid];
2907 alias = pci_seg->alias_table[devid];
2908 table = pci_seg->irq_lookup_table[alias];
2910 set_remap_table_entry(iommu, devid, table);
2913 spin_unlock_irqrestore(&iommu_table_lock, flags);
2915 /* Nothing there yet, allocate new irq remapping table */
2916 new_table = __alloc_irq_table();
2920 spin_lock_irqsave(&iommu_table_lock, flags);
2922 table = pci_seg->irq_lookup_table[devid];
2926 table = pci_seg->irq_lookup_table[alias];
2928 set_remap_table_entry(iommu, devid, table);
2936 pci_for_each_dma_alias(pdev, set_remap_table_entry_alias,
2939 set_remap_table_entry(iommu, devid, table);
2942 set_remap_table_entry(iommu, alias, table);
2945 iommu_completion_wait(iommu);
2948 spin_unlock_irqrestore(&iommu_table_lock, flags);
2951 kmem_cache_free(amd_iommu_irq_cache, new_table->table);
2957 static int alloc_irq_index(struct amd_iommu *iommu, u16 devid, int count,
2958 bool align, struct pci_dev *pdev)
2960 struct irq_remap_table *table;
2961 int index, c, alignment = 1;
2962 unsigned long flags;
2964 table = alloc_irq_table(iommu, devid, pdev);
2969 alignment = roundup_pow_of_two(count);
2971 raw_spin_lock_irqsave(&table->lock, flags);
2973 /* Scan table for free entries */
2974 for (index = ALIGN(table->min_index, alignment), c = 0;
2975 index < MAX_IRQS_PER_TABLE;) {
2976 if (!iommu->irte_ops->is_allocated(table, index)) {
2980 index = ALIGN(index + 1, alignment);
2986 iommu->irte_ops->set_allocated(table, index - c + 1);
2998 raw_spin_unlock_irqrestore(&table->lock, flags);
3003 static int modify_irte_ga(struct amd_iommu *iommu, u16 devid, int index,
3004 struct irte_ga *irte, struct amd_ir_data *data)
3007 struct irq_remap_table *table;
3008 unsigned long flags;
3009 struct irte_ga *entry;
3011 table = get_irq_table(iommu, devid);
3015 raw_spin_lock_irqsave(&table->lock, flags);
3017 entry = (struct irte_ga *)table->table;
3018 entry = &entry[index];
3020 ret = cmpxchg_double(&entry->lo.val, &entry->hi.val,
3021 entry->lo.val, entry->hi.val,
3022 irte->lo.val, irte->hi.val);
3024 * We use cmpxchg16 to atomically update the 128-bit IRTE,
3025 * and it cannot be updated by the hardware or other processors
3026 * behind us, so the return value of cmpxchg16 should be the
3027 * same as the old value.
3034 raw_spin_unlock_irqrestore(&table->lock, flags);
3036 iommu_flush_irt(iommu, devid);
3037 iommu_completion_wait(iommu);
3042 static int modify_irte(struct amd_iommu *iommu,
3043 u16 devid, int index, union irte *irte)
3045 struct irq_remap_table *table;
3046 unsigned long flags;
3048 table = get_irq_table(iommu, devid);
3052 raw_spin_lock_irqsave(&table->lock, flags);
3053 table->table[index] = irte->val;
3054 raw_spin_unlock_irqrestore(&table->lock, flags);
3056 iommu_flush_irt(iommu, devid);
3057 iommu_completion_wait(iommu);
3062 static void free_irte(struct amd_iommu *iommu, u16 devid, int index)
3064 struct irq_remap_table *table;
3065 unsigned long flags;
3067 table = get_irq_table(iommu, devid);
3071 raw_spin_lock_irqsave(&table->lock, flags);
3072 iommu->irte_ops->clear_allocated(table, index);
3073 raw_spin_unlock_irqrestore(&table->lock, flags);
3075 iommu_flush_irt(iommu, devid);
3076 iommu_completion_wait(iommu);
3079 static void irte_prepare(void *entry,
3080 u32 delivery_mode, bool dest_mode,
3081 u8 vector, u32 dest_apicid, int devid)
3083 union irte *irte = (union irte *) entry;
3086 irte->fields.vector = vector;
3087 irte->fields.int_type = delivery_mode;
3088 irte->fields.destination = dest_apicid;
3089 irte->fields.dm = dest_mode;
3090 irte->fields.valid = 1;
3093 static void irte_ga_prepare(void *entry,
3094 u32 delivery_mode, bool dest_mode,
3095 u8 vector, u32 dest_apicid, int devid)
3097 struct irte_ga *irte = (struct irte_ga *) entry;
3101 irte->lo.fields_remap.int_type = delivery_mode;
3102 irte->lo.fields_remap.dm = dest_mode;
3103 irte->hi.fields.vector = vector;
3104 irte->lo.fields_remap.destination = APICID_TO_IRTE_DEST_LO(dest_apicid);
3105 irte->hi.fields.destination = APICID_TO_IRTE_DEST_HI(dest_apicid);
3106 irte->lo.fields_remap.valid = 1;
3109 static void irte_activate(struct amd_iommu *iommu, void *entry, u16 devid, u16 index)
3111 union irte *irte = (union irte *) entry;
3113 irte->fields.valid = 1;
3114 modify_irte(iommu, devid, index, irte);
3117 static void irte_ga_activate(struct amd_iommu *iommu, void *entry, u16 devid, u16 index)
3119 struct irte_ga *irte = (struct irte_ga *) entry;
3121 irte->lo.fields_remap.valid = 1;
3122 modify_irte_ga(iommu, devid, index, irte, NULL);
3125 static void irte_deactivate(struct amd_iommu *iommu, void *entry, u16 devid, u16 index)
3127 union irte *irte = (union irte *) entry;
3129 irte->fields.valid = 0;
3130 modify_irte(iommu, devid, index, irte);
3133 static void irte_ga_deactivate(struct amd_iommu *iommu, void *entry, u16 devid, u16 index)
3135 struct irte_ga *irte = (struct irte_ga *) entry;
3137 irte->lo.fields_remap.valid = 0;
3138 modify_irte_ga(iommu, devid, index, irte, NULL);
3141 static void irte_set_affinity(struct amd_iommu *iommu, void *entry, u16 devid, u16 index,
3142 u8 vector, u32 dest_apicid)
3144 union irte *irte = (union irte *) entry;
3146 irte->fields.vector = vector;
3147 irte->fields.destination = dest_apicid;
3148 modify_irte(iommu, devid, index, irte);
3151 static void irte_ga_set_affinity(struct amd_iommu *iommu, void *entry, u16 devid, u16 index,
3152 u8 vector, u32 dest_apicid)
3154 struct irte_ga *irte = (struct irte_ga *) entry;
3156 if (!irte->lo.fields_remap.guest_mode) {
3157 irte->hi.fields.vector = vector;
3158 irte->lo.fields_remap.destination =
3159 APICID_TO_IRTE_DEST_LO(dest_apicid);
3160 irte->hi.fields.destination =
3161 APICID_TO_IRTE_DEST_HI(dest_apicid);
3162 modify_irte_ga(iommu, devid, index, irte, NULL);
3166 #define IRTE_ALLOCATED (~1U)
3167 static void irte_set_allocated(struct irq_remap_table *table, int index)
3169 table->table[index] = IRTE_ALLOCATED;
3172 static void irte_ga_set_allocated(struct irq_remap_table *table, int index)
3174 struct irte_ga *ptr = (struct irte_ga *)table->table;
3175 struct irte_ga *irte = &ptr[index];
3177 memset(&irte->lo.val, 0, sizeof(u64));
3178 memset(&irte->hi.val, 0, sizeof(u64));
3179 irte->hi.fields.vector = 0xff;
3182 static bool irte_is_allocated(struct irq_remap_table *table, int index)
3184 union irte *ptr = (union irte *)table->table;
3185 union irte *irte = &ptr[index];
3187 return irte->val != 0;
3190 static bool irte_ga_is_allocated(struct irq_remap_table *table, int index)
3192 struct irte_ga *ptr = (struct irte_ga *)table->table;
3193 struct irte_ga *irte = &ptr[index];
3195 return irte->hi.fields.vector != 0;
3198 static void irte_clear_allocated(struct irq_remap_table *table, int index)
3200 table->table[index] = 0;
3203 static void irte_ga_clear_allocated(struct irq_remap_table *table, int index)
3205 struct irte_ga *ptr = (struct irte_ga *)table->table;
3206 struct irte_ga *irte = &ptr[index];
3208 memset(&irte->lo.val, 0, sizeof(u64));
3209 memset(&irte->hi.val, 0, sizeof(u64));
3212 static int get_devid(struct irq_alloc_info *info)
3214 switch (info->type) {
3215 case X86_IRQ_ALLOC_TYPE_IOAPIC:
3216 return get_ioapic_devid(info->devid);
3217 case X86_IRQ_ALLOC_TYPE_HPET:
3218 return get_hpet_devid(info->devid);
3219 case X86_IRQ_ALLOC_TYPE_PCI_MSI:
3220 case X86_IRQ_ALLOC_TYPE_PCI_MSIX:
3221 return get_device_sbdf_id(msi_desc_to_dev(info->desc));
3228 struct irq_remap_ops amd_iommu_irq_ops = {
3229 .prepare = amd_iommu_prepare,
3230 .enable = amd_iommu_enable,
3231 .disable = amd_iommu_disable,
3232 .reenable = amd_iommu_reenable,
3233 .enable_faulting = amd_iommu_enable_faulting,
3236 static void fill_msi_msg(struct msi_msg *msg, u32 index)
3239 msg->address_lo = 0;
3240 msg->arch_addr_lo.base_address = X86_MSI_BASE_ADDRESS_LOW;
3241 msg->address_hi = X86_MSI_BASE_ADDRESS_HIGH;
3244 static void irq_remapping_prepare_irte(struct amd_ir_data *data,
3245 struct irq_cfg *irq_cfg,
3246 struct irq_alloc_info *info,
3247 int devid, int index, int sub_handle)
3249 struct irq_2_irte *irte_info = &data->irq_2_irte;
3250 struct amd_iommu *iommu = data->iommu;
3255 data->irq_2_irte.devid = devid;
3256 data->irq_2_irte.index = index + sub_handle;
3257 iommu->irte_ops->prepare(data->entry, apic->delivery_mode,
3258 apic->dest_mode_logical, irq_cfg->vector,
3259 irq_cfg->dest_apicid, devid);
3261 switch (info->type) {
3262 case X86_IRQ_ALLOC_TYPE_IOAPIC:
3263 case X86_IRQ_ALLOC_TYPE_HPET:
3264 case X86_IRQ_ALLOC_TYPE_PCI_MSI:
3265 case X86_IRQ_ALLOC_TYPE_PCI_MSIX:
3266 fill_msi_msg(&data->msi_entry, irte_info->index);
3275 struct amd_irte_ops irte_32_ops = {
3276 .prepare = irte_prepare,
3277 .activate = irte_activate,
3278 .deactivate = irte_deactivate,
3279 .set_affinity = irte_set_affinity,
3280 .set_allocated = irte_set_allocated,
3281 .is_allocated = irte_is_allocated,
3282 .clear_allocated = irte_clear_allocated,
3285 struct amd_irte_ops irte_128_ops = {
3286 .prepare = irte_ga_prepare,
3287 .activate = irte_ga_activate,
3288 .deactivate = irte_ga_deactivate,
3289 .set_affinity = irte_ga_set_affinity,
3290 .set_allocated = irte_ga_set_allocated,
3291 .is_allocated = irte_ga_is_allocated,
3292 .clear_allocated = irte_ga_clear_allocated,
3295 static int irq_remapping_alloc(struct irq_domain *domain, unsigned int virq,
3296 unsigned int nr_irqs, void *arg)
3298 struct irq_alloc_info *info = arg;
3299 struct irq_data *irq_data;
3300 struct amd_ir_data *data = NULL;
3301 struct amd_iommu *iommu;
3302 struct irq_cfg *cfg;
3303 int i, ret, devid, seg, sbdf;
3308 if (nr_irqs > 1 && info->type != X86_IRQ_ALLOC_TYPE_PCI_MSI)
3311 sbdf = get_devid(info);
3315 seg = PCI_SBDF_TO_SEGID(sbdf);
3316 devid = PCI_SBDF_TO_DEVID(sbdf);
3317 iommu = __rlookup_amd_iommu(seg, devid);
3321 ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
3325 if (info->type == X86_IRQ_ALLOC_TYPE_IOAPIC) {
3326 struct irq_remap_table *table;
3328 table = alloc_irq_table(iommu, devid, NULL);
3330 if (!table->min_index) {
3332 * Keep the first 32 indexes free for IOAPIC
3335 table->min_index = 32;
3336 for (i = 0; i < 32; ++i)
3337 iommu->irte_ops->set_allocated(table, i);
3339 WARN_ON(table->min_index != 32);
3340 index = info->ioapic.pin;
3344 } else if (info->type == X86_IRQ_ALLOC_TYPE_PCI_MSI ||
3345 info->type == X86_IRQ_ALLOC_TYPE_PCI_MSIX) {
3346 bool align = (info->type == X86_IRQ_ALLOC_TYPE_PCI_MSI);
3348 index = alloc_irq_index(iommu, devid, nr_irqs, align,
3349 msi_desc_to_pci_dev(info->desc));
3351 index = alloc_irq_index(iommu, devid, nr_irqs, false, NULL);
3355 pr_warn("Failed to allocate IRTE\n");
3357 goto out_free_parent;
3360 for (i = 0; i < nr_irqs; i++) {
3361 irq_data = irq_domain_get_irq_data(domain, virq + i);
3362 cfg = irq_data ? irqd_cfg(irq_data) : NULL;
3369 data = kzalloc(sizeof(*data), GFP_KERNEL);
3373 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
3374 data->entry = kzalloc(sizeof(union irte), GFP_KERNEL);
3376 data->entry = kzalloc(sizeof(struct irte_ga),
3383 data->iommu = iommu;
3384 irq_data->hwirq = (devid << 16) + i;
3385 irq_data->chip_data = data;
3386 irq_data->chip = &amd_ir_chip;
3387 irq_remapping_prepare_irte(data, cfg, info, devid, index, i);
3388 irq_set_status_flags(virq + i, IRQ_MOVE_PCNTXT);
3394 for (i--; i >= 0; i--) {
3395 irq_data = irq_domain_get_irq_data(domain, virq + i);
3397 kfree(irq_data->chip_data);
3399 for (i = 0; i < nr_irqs; i++)
3400 free_irte(iommu, devid, index + i);
3402 irq_domain_free_irqs_common(domain, virq, nr_irqs);
3406 static void irq_remapping_free(struct irq_domain *domain, unsigned int virq,
3407 unsigned int nr_irqs)
3409 struct irq_2_irte *irte_info;
3410 struct irq_data *irq_data;
3411 struct amd_ir_data *data;
3414 for (i = 0; i < nr_irqs; i++) {
3415 irq_data = irq_domain_get_irq_data(domain, virq + i);
3416 if (irq_data && irq_data->chip_data) {
3417 data = irq_data->chip_data;
3418 irte_info = &data->irq_2_irte;
3419 free_irte(data->iommu, irte_info->devid, irte_info->index);
3424 irq_domain_free_irqs_common(domain, virq, nr_irqs);
3427 static void amd_ir_update_irte(struct irq_data *irqd, struct amd_iommu *iommu,
3428 struct amd_ir_data *ir_data,
3429 struct irq_2_irte *irte_info,
3430 struct irq_cfg *cfg);
3432 static int irq_remapping_activate(struct irq_domain *domain,
3433 struct irq_data *irq_data, bool reserve)
3435 struct amd_ir_data *data = irq_data->chip_data;
3436 struct irq_2_irte *irte_info = &data->irq_2_irte;
3437 struct amd_iommu *iommu = data->iommu;
3438 struct irq_cfg *cfg = irqd_cfg(irq_data);
3443 iommu->irte_ops->activate(iommu, data->entry, irte_info->devid,
3445 amd_ir_update_irte(irq_data, iommu, data, irte_info, cfg);
3449 static void irq_remapping_deactivate(struct irq_domain *domain,
3450 struct irq_data *irq_data)
3452 struct amd_ir_data *data = irq_data->chip_data;
3453 struct irq_2_irte *irte_info = &data->irq_2_irte;
3454 struct amd_iommu *iommu = data->iommu;
3457 iommu->irte_ops->deactivate(iommu, data->entry, irte_info->devid,
3461 static int irq_remapping_select(struct irq_domain *d, struct irq_fwspec *fwspec,
3462 enum irq_domain_bus_token bus_token)
3464 struct amd_iommu *iommu;
3467 if (!amd_iommu_irq_remap)
3470 if (x86_fwspec_is_ioapic(fwspec))
3471 devid = get_ioapic_devid(fwspec->param[0]);
3472 else if (x86_fwspec_is_hpet(fwspec))
3473 devid = get_hpet_devid(fwspec->param[0]);
3477 iommu = __rlookup_amd_iommu((devid >> 16), (devid & 0xffff));
3479 return iommu && iommu->ir_domain == d;
3482 static const struct irq_domain_ops amd_ir_domain_ops = {
3483 .select = irq_remapping_select,
3484 .alloc = irq_remapping_alloc,
3485 .free = irq_remapping_free,
3486 .activate = irq_remapping_activate,
3487 .deactivate = irq_remapping_deactivate,
3490 int amd_iommu_activate_guest_mode(void *data)
3492 struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
3493 struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
3496 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
3497 !entry || entry->lo.fields_vapic.guest_mode)
3500 valid = entry->lo.fields_vapic.valid;
3505 entry->lo.fields_vapic.valid = valid;
3506 entry->lo.fields_vapic.guest_mode = 1;
3507 entry->lo.fields_vapic.ga_log_intr = 1;
3508 entry->hi.fields.ga_root_ptr = ir_data->ga_root_ptr;
3509 entry->hi.fields.vector = ir_data->ga_vector;
3510 entry->lo.fields_vapic.ga_tag = ir_data->ga_tag;
3512 return modify_irte_ga(ir_data->iommu, ir_data->irq_2_irte.devid,
3513 ir_data->irq_2_irte.index, entry, ir_data);
3515 EXPORT_SYMBOL(amd_iommu_activate_guest_mode);
3517 int amd_iommu_deactivate_guest_mode(void *data)
3519 struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
3520 struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
3521 struct irq_cfg *cfg = ir_data->cfg;
3524 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
3525 !entry || !entry->lo.fields_vapic.guest_mode)
3528 valid = entry->lo.fields_remap.valid;
3533 entry->lo.fields_remap.valid = valid;
3534 entry->lo.fields_remap.dm = apic->dest_mode_logical;
3535 entry->lo.fields_remap.int_type = apic->delivery_mode;
3536 entry->hi.fields.vector = cfg->vector;
3537 entry->lo.fields_remap.destination =
3538 APICID_TO_IRTE_DEST_LO(cfg->dest_apicid);
3539 entry->hi.fields.destination =
3540 APICID_TO_IRTE_DEST_HI(cfg->dest_apicid);
3542 return modify_irte_ga(ir_data->iommu, ir_data->irq_2_irte.devid,
3543 ir_data->irq_2_irte.index, entry, ir_data);
3545 EXPORT_SYMBOL(amd_iommu_deactivate_guest_mode);
3547 static int amd_ir_set_vcpu_affinity(struct irq_data *data, void *vcpu_info)
3550 struct amd_iommu_pi_data *pi_data = vcpu_info;
3551 struct vcpu_data *vcpu_pi_info = pi_data->vcpu_data;
3552 struct amd_ir_data *ir_data = data->chip_data;
3553 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
3554 struct iommu_dev_data *dev_data;
3556 if (ir_data->iommu == NULL)
3559 dev_data = search_dev_data(ir_data->iommu, irte_info->devid);
3562 * This device has never been set up for guest mode.
3563 * we should not modify the IRTE
3565 if (!dev_data || !dev_data->use_vapic)
3568 ir_data->cfg = irqd_cfg(data);
3569 pi_data->ir_data = ir_data;
3572 * SVM tries to set up for VAPIC mode, but we are in
3573 * legacy mode. So, we force legacy mode instead.
3575 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
3576 pr_debug("%s: Fall back to using intr legacy remap\n",
3578 pi_data->is_guest_mode = false;
3581 pi_data->prev_ga_tag = ir_data->cached_ga_tag;
3582 if (pi_data->is_guest_mode) {
3583 ir_data->ga_root_ptr = (pi_data->base >> 12);
3584 ir_data->ga_vector = vcpu_pi_info->vector;
3585 ir_data->ga_tag = pi_data->ga_tag;
3586 ret = amd_iommu_activate_guest_mode(ir_data);
3588 ir_data->cached_ga_tag = pi_data->ga_tag;
3590 ret = amd_iommu_deactivate_guest_mode(ir_data);
3593 * This communicates the ga_tag back to the caller
3594 * so that it can do all the necessary clean up.
3597 ir_data->cached_ga_tag = 0;
3604 static void amd_ir_update_irte(struct irq_data *irqd, struct amd_iommu *iommu,
3605 struct amd_ir_data *ir_data,
3606 struct irq_2_irte *irte_info,
3607 struct irq_cfg *cfg)
3611 * Atomically updates the IRTE with the new destination, vector
3612 * and flushes the interrupt entry cache.
3614 iommu->irte_ops->set_affinity(iommu, ir_data->entry, irte_info->devid,
3615 irte_info->index, cfg->vector,
3619 static int amd_ir_set_affinity(struct irq_data *data,
3620 const struct cpumask *mask, bool force)
3622 struct amd_ir_data *ir_data = data->chip_data;
3623 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
3624 struct irq_cfg *cfg = irqd_cfg(data);
3625 struct irq_data *parent = data->parent_data;
3626 struct amd_iommu *iommu = ir_data->iommu;
3632 ret = parent->chip->irq_set_affinity(parent, mask, force);
3633 if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
3636 amd_ir_update_irte(data, iommu, ir_data, irte_info, cfg);
3638 * After this point, all the interrupts will start arriving
3639 * at the new destination. So, time to cleanup the previous
3640 * vector allocation.
3642 send_cleanup_vector(cfg);
3644 return IRQ_SET_MASK_OK_DONE;
3647 static void ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg)
3649 struct amd_ir_data *ir_data = irq_data->chip_data;
3651 *msg = ir_data->msi_entry;
3654 static struct irq_chip amd_ir_chip = {
3656 .irq_ack = apic_ack_irq,
3657 .irq_set_affinity = amd_ir_set_affinity,
3658 .irq_set_vcpu_affinity = amd_ir_set_vcpu_affinity,
3659 .irq_compose_msi_msg = ir_compose_msi_msg,
3662 static const struct msi_parent_ops amdvi_msi_parent_ops = {
3663 .supported_flags = X86_VECTOR_MSI_FLAGS_SUPPORTED |
3664 MSI_FLAG_MULTI_PCI_MSI |
3667 .init_dev_msi_info = msi_parent_init_dev_msi_info,
3670 static const struct msi_parent_ops virt_amdvi_msi_parent_ops = {
3671 .supported_flags = X86_VECTOR_MSI_FLAGS_SUPPORTED |
3672 MSI_FLAG_MULTI_PCI_MSI,
3674 .init_dev_msi_info = msi_parent_init_dev_msi_info,
3677 int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
3679 struct fwnode_handle *fn;
3681 fn = irq_domain_alloc_named_id_fwnode("AMD-IR", iommu->index);
3684 iommu->ir_domain = irq_domain_create_hierarchy(arch_get_ir_parent_domain(), 0, 0,
3685 fn, &amd_ir_domain_ops, iommu);
3686 if (!iommu->ir_domain) {
3687 irq_domain_free_fwnode(fn);
3691 irq_domain_update_bus_token(iommu->ir_domain, DOMAIN_BUS_AMDVI);
3692 iommu->ir_domain->flags |= IRQ_DOMAIN_FLAG_MSI_PARENT |
3693 IRQ_DOMAIN_FLAG_ISOLATED_MSI;
3695 if (amd_iommu_np_cache)
3696 iommu->ir_domain->msi_parent_ops = &virt_amdvi_msi_parent_ops;
3698 iommu->ir_domain->msi_parent_ops = &amdvi_msi_parent_ops;
3703 int amd_iommu_update_ga(int cpu, bool is_run, void *data)
3705 unsigned long flags;
3706 struct amd_iommu *iommu;
3707 struct irq_remap_table *table;
3708 struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
3709 int devid = ir_data->irq_2_irte.devid;
3710 struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
3711 struct irte_ga *ref = (struct irte_ga *) ir_data->ref;
3713 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
3714 !ref || !entry || !entry->lo.fields_vapic.guest_mode)
3717 iommu = ir_data->iommu;
3721 table = get_irq_table(iommu, devid);
3725 raw_spin_lock_irqsave(&table->lock, flags);
3727 if (ref->lo.fields_vapic.guest_mode) {
3729 ref->lo.fields_vapic.destination =
3730 APICID_TO_IRTE_DEST_LO(cpu);
3731 ref->hi.fields.destination =
3732 APICID_TO_IRTE_DEST_HI(cpu);
3734 ref->lo.fields_vapic.is_run = is_run;
3738 raw_spin_unlock_irqrestore(&table->lock, flags);
3740 iommu_flush_irt(iommu, devid);
3741 iommu_completion_wait(iommu);
3744 EXPORT_SYMBOL(amd_iommu_update_ga);