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>
40 #include <uapi/linux/iommufd.h>
42 #include "amd_iommu.h"
43 #include "../dma-iommu.h"
44 #include "../irq_remapping.h"
46 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
48 #define LOOP_TIMEOUT 100000
50 /* IO virtual address start page frame number */
51 #define IOVA_START_PFN (1)
52 #define IOVA_PFN(addr) ((addr) >> PAGE_SHIFT)
54 /* Reserved IOVA ranges */
55 #define MSI_RANGE_START (0xfee00000)
56 #define MSI_RANGE_END (0xfeefffff)
57 #define HT_RANGE_START (0xfd00000000ULL)
58 #define HT_RANGE_END (0xffffffffffULL)
60 #define DEFAULT_PGTABLE_LEVEL PAGE_MODE_3_LEVEL
62 static DEFINE_SPINLOCK(pd_bitmap_lock);
64 LIST_HEAD(ioapic_map);
66 LIST_HEAD(acpihid_map);
68 const struct iommu_ops amd_iommu_ops;
69 const struct iommu_dirty_ops amd_dirty_ops;
71 static ATOMIC_NOTIFIER_HEAD(ppr_notifier);
72 int amd_iommu_max_glx_val = -1;
75 * general struct to manage commands send to an IOMMU
81 struct kmem_cache *amd_iommu_irq_cache;
83 static void detach_device(struct device *dev);
84 static int domain_enable_v2(struct protection_domain *domain, int pasids);
86 /****************************************************************************
90 ****************************************************************************/
92 static inline int get_acpihid_device_id(struct device *dev,
93 struct acpihid_map_entry **entry)
95 struct acpi_device *adev = ACPI_COMPANION(dev);
96 struct acpihid_map_entry *p;
101 list_for_each_entry(p, &acpihid_map, list) {
102 if (acpi_dev_hid_uid_match(adev, p->hid,
103 p->uid[0] ? p->uid : NULL)) {
112 static inline int get_device_sbdf_id(struct device *dev)
117 sbdf = get_pci_sbdf_id(to_pci_dev(dev));
119 sbdf = get_acpihid_device_id(dev, NULL);
124 struct dev_table_entry *get_dev_table(struct amd_iommu *iommu)
126 struct dev_table_entry *dev_table;
127 struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
129 BUG_ON(pci_seg == NULL);
130 dev_table = pci_seg->dev_table;
131 BUG_ON(dev_table == NULL);
136 static inline u16 get_device_segment(struct device *dev)
140 if (dev_is_pci(dev)) {
141 struct pci_dev *pdev = to_pci_dev(dev);
143 seg = pci_domain_nr(pdev->bus);
145 u32 devid = get_acpihid_device_id(dev, NULL);
147 seg = PCI_SBDF_TO_SEGID(devid);
153 /* Writes the specific IOMMU for a device into the PCI segment rlookup table */
154 void amd_iommu_set_rlookup_table(struct amd_iommu *iommu, u16 devid)
156 struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
158 pci_seg->rlookup_table[devid] = iommu;
161 static struct amd_iommu *__rlookup_amd_iommu(u16 seg, u16 devid)
163 struct amd_iommu_pci_seg *pci_seg;
165 for_each_pci_segment(pci_seg) {
166 if (pci_seg->id == seg)
167 return pci_seg->rlookup_table[devid];
172 static struct amd_iommu *rlookup_amd_iommu(struct device *dev)
174 u16 seg = get_device_segment(dev);
175 int devid = get_device_sbdf_id(dev);
179 return __rlookup_amd_iommu(seg, PCI_SBDF_TO_DEVID(devid));
182 static struct protection_domain *to_pdomain(struct iommu_domain *dom)
184 return container_of(dom, struct protection_domain, domain);
187 static struct iommu_dev_data *alloc_dev_data(struct amd_iommu *iommu, u16 devid)
189 struct iommu_dev_data *dev_data;
190 struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
192 dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
196 spin_lock_init(&dev_data->lock);
197 dev_data->devid = devid;
198 ratelimit_default_init(&dev_data->rs);
200 llist_add(&dev_data->dev_data_list, &pci_seg->dev_data_list);
204 static struct iommu_dev_data *search_dev_data(struct amd_iommu *iommu, u16 devid)
206 struct iommu_dev_data *dev_data;
207 struct llist_node *node;
208 struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
210 if (llist_empty(&pci_seg->dev_data_list))
213 node = pci_seg->dev_data_list.first;
214 llist_for_each_entry(dev_data, node, dev_data_list) {
215 if (dev_data->devid == devid)
222 static int clone_alias(struct pci_dev *pdev, u16 alias, void *data)
224 struct amd_iommu *iommu;
225 struct dev_table_entry *dev_table;
226 u16 devid = pci_dev_id(pdev);
231 iommu = rlookup_amd_iommu(&pdev->dev);
235 amd_iommu_set_rlookup_table(iommu, alias);
236 dev_table = get_dev_table(iommu);
237 memcpy(dev_table[alias].data,
238 dev_table[devid].data,
239 sizeof(dev_table[alias].data));
244 static void clone_aliases(struct amd_iommu *iommu, struct device *dev)
246 struct pci_dev *pdev;
248 if (!dev_is_pci(dev))
250 pdev = to_pci_dev(dev);
253 * The IVRS alias stored in the alias table may not be
254 * part of the PCI DMA aliases if it's bus differs
255 * from the original device.
257 clone_alias(pdev, iommu->pci_seg->alias_table[pci_dev_id(pdev)], NULL);
259 pci_for_each_dma_alias(pdev, clone_alias, NULL);
262 static void setup_aliases(struct amd_iommu *iommu, struct device *dev)
264 struct pci_dev *pdev = to_pci_dev(dev);
265 struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
268 /* For ACPI HID devices, there are no aliases */
269 if (!dev_is_pci(dev))
273 * Add the IVRS alias to the pci aliases if it is on the same
274 * bus. The IVRS table may know about a quirk that we don't.
276 ivrs_alias = pci_seg->alias_table[pci_dev_id(pdev)];
277 if (ivrs_alias != pci_dev_id(pdev) &&
278 PCI_BUS_NUM(ivrs_alias) == pdev->bus->number)
279 pci_add_dma_alias(pdev, ivrs_alias & 0xff, 1);
281 clone_aliases(iommu, dev);
284 static struct iommu_dev_data *find_dev_data(struct amd_iommu *iommu, u16 devid)
286 struct iommu_dev_data *dev_data;
288 dev_data = search_dev_data(iommu, devid);
290 if (dev_data == NULL) {
291 dev_data = alloc_dev_data(iommu, devid);
295 if (translation_pre_enabled(iommu))
296 dev_data->defer_attach = true;
303 * Find or create an IOMMU group for a acpihid device.
305 static struct iommu_group *acpihid_device_group(struct device *dev)
307 struct acpihid_map_entry *p, *entry = NULL;
310 devid = get_acpihid_device_id(dev, &entry);
312 return ERR_PTR(devid);
314 list_for_each_entry(p, &acpihid_map, list) {
315 if ((devid == p->devid) && p->group)
316 entry->group = p->group;
320 entry->group = generic_device_group(dev);
322 iommu_group_ref_get(entry->group);
327 static bool pci_iommuv2_capable(struct pci_dev *pdev)
329 static const int caps[] = {
331 PCI_EXT_CAP_ID_PASID,
335 if (!pci_ats_supported(pdev))
338 for (i = 0; i < 2; ++i) {
339 pos = pci_find_ext_capability(pdev, caps[i]);
348 * This function checks if the driver got a valid device from the caller to
349 * avoid dereferencing invalid pointers.
351 static bool check_device(struct device *dev)
353 struct amd_iommu_pci_seg *pci_seg;
354 struct amd_iommu *iommu;
360 sbdf = get_device_sbdf_id(dev);
363 devid = PCI_SBDF_TO_DEVID(sbdf);
365 iommu = rlookup_amd_iommu(dev);
369 /* Out of our scope? */
370 pci_seg = iommu->pci_seg;
371 if (devid > pci_seg->last_bdf)
377 static int iommu_init_device(struct amd_iommu *iommu, struct device *dev)
379 struct iommu_dev_data *dev_data;
382 if (dev_iommu_priv_get(dev))
385 sbdf = get_device_sbdf_id(dev);
389 devid = PCI_SBDF_TO_DEVID(sbdf);
390 dev_data = find_dev_data(iommu, devid);
395 setup_aliases(iommu, dev);
398 * By default we use passthrough mode for IOMMUv2 capable device.
399 * But if amd_iommu=force_isolation is set (e.g. to debug DMA to
400 * invalid address), we ignore the capability for the device so
401 * it'll be forced to go into translation mode.
403 if ((iommu_default_passthrough() || !amd_iommu_force_isolation) &&
404 dev_is_pci(dev) && pci_iommuv2_capable(to_pci_dev(dev))) {
405 dev_data->iommu_v2 = iommu->is_iommu_v2;
408 dev_iommu_priv_set(dev, dev_data);
413 static void iommu_ignore_device(struct amd_iommu *iommu, struct device *dev)
415 struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
416 struct dev_table_entry *dev_table = get_dev_table(iommu);
419 sbdf = get_device_sbdf_id(dev);
423 devid = PCI_SBDF_TO_DEVID(sbdf);
424 pci_seg->rlookup_table[devid] = NULL;
425 memset(&dev_table[devid], 0, sizeof(struct dev_table_entry));
427 setup_aliases(iommu, dev);
430 static void amd_iommu_uninit_device(struct device *dev)
432 struct iommu_dev_data *dev_data;
434 dev_data = dev_iommu_priv_get(dev);
438 if (dev_data->domain)
441 dev_iommu_priv_set(dev, NULL);
444 * We keep dev_data around for unplugged devices and reuse it when the
445 * device is re-plugged - not doing so would introduce a ton of races.
449 /****************************************************************************
451 * Interrupt handling functions
453 ****************************************************************************/
455 static void dump_dte_entry(struct amd_iommu *iommu, u16 devid)
458 struct dev_table_entry *dev_table = get_dev_table(iommu);
460 for (i = 0; i < 4; ++i)
461 pr_err("DTE[%d]: %016llx\n", i, dev_table[devid].data[i]);
464 static void dump_command(unsigned long phys_addr)
466 struct iommu_cmd *cmd = iommu_phys_to_virt(phys_addr);
469 for (i = 0; i < 4; ++i)
470 pr_err("CMD[%d]: %08x\n", i, cmd->data[i]);
473 static void amd_iommu_report_rmp_hw_error(struct amd_iommu *iommu, volatile u32 *event)
475 struct iommu_dev_data *dev_data = NULL;
476 int devid, vmg_tag, flags;
477 struct pci_dev *pdev;
480 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
481 vmg_tag = (event[1]) & 0xFFFF;
482 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
483 spa = ((u64)event[3] << 32) | (event[2] & 0xFFFFFFF8);
485 pdev = pci_get_domain_bus_and_slot(iommu->pci_seg->id, PCI_BUS_NUM(devid),
488 dev_data = dev_iommu_priv_get(&pdev->dev);
491 if (__ratelimit(&dev_data->rs)) {
492 pci_err(pdev, "Event logged [RMP_HW_ERROR vmg_tag=0x%04x, spa=0x%llx, flags=0x%04x]\n",
493 vmg_tag, spa, flags);
496 pr_err_ratelimited("Event logged [RMP_HW_ERROR device=%04x:%02x:%02x.%x, vmg_tag=0x%04x, spa=0x%llx, flags=0x%04x]\n",
497 iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
498 vmg_tag, spa, flags);
505 static void amd_iommu_report_rmp_fault(struct amd_iommu *iommu, volatile u32 *event)
507 struct iommu_dev_data *dev_data = NULL;
508 int devid, flags_rmp, vmg_tag, flags;
509 struct pci_dev *pdev;
512 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
513 flags_rmp = (event[0] >> EVENT_FLAGS_SHIFT) & 0xFF;
514 vmg_tag = (event[1]) & 0xFFFF;
515 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
516 gpa = ((u64)event[3] << 32) | event[2];
518 pdev = pci_get_domain_bus_and_slot(iommu->pci_seg->id, PCI_BUS_NUM(devid),
521 dev_data = dev_iommu_priv_get(&pdev->dev);
524 if (__ratelimit(&dev_data->rs)) {
525 pci_err(pdev, "Event logged [RMP_PAGE_FAULT vmg_tag=0x%04x, gpa=0x%llx, flags_rmp=0x%04x, flags=0x%04x]\n",
526 vmg_tag, gpa, flags_rmp, flags);
529 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",
530 iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
531 vmg_tag, gpa, flags_rmp, flags);
538 #define IS_IOMMU_MEM_TRANSACTION(flags) \
539 (((flags) & EVENT_FLAG_I) == 0)
541 #define IS_WRITE_REQUEST(flags) \
542 ((flags) & EVENT_FLAG_RW)
544 static void amd_iommu_report_page_fault(struct amd_iommu *iommu,
545 u16 devid, u16 domain_id,
546 u64 address, int flags)
548 struct iommu_dev_data *dev_data = NULL;
549 struct pci_dev *pdev;
551 pdev = pci_get_domain_bus_and_slot(iommu->pci_seg->id, PCI_BUS_NUM(devid),
554 dev_data = dev_iommu_priv_get(&pdev->dev);
558 * If this is a DMA fault (for which the I(nterrupt)
559 * bit will be unset), allow report_iommu_fault() to
560 * prevent logging it.
562 if (IS_IOMMU_MEM_TRANSACTION(flags)) {
563 /* Device not attached to domain properly */
564 if (dev_data->domain == NULL) {
565 pr_err_ratelimited("Event logged [Device not attached to domain properly]\n");
566 pr_err_ratelimited(" device=%04x:%02x:%02x.%x domain=0x%04x\n",
567 iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid),
568 PCI_FUNC(devid), domain_id);
572 if (!report_iommu_fault(&dev_data->domain->domain,
574 IS_WRITE_REQUEST(flags) ?
580 if (__ratelimit(&dev_data->rs)) {
581 pci_err(pdev, "Event logged [IO_PAGE_FAULT domain=0x%04x address=0x%llx flags=0x%04x]\n",
582 domain_id, address, flags);
585 pr_err_ratelimited("Event logged [IO_PAGE_FAULT device=%04x:%02x:%02x.%x domain=0x%04x address=0x%llx flags=0x%04x]\n",
586 iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
587 domain_id, address, flags);
595 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
597 struct device *dev = iommu->iommu.dev;
598 int type, devid, flags, tag;
599 volatile u32 *event = __evt;
605 type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
606 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
607 pasid = (event[0] & EVENT_DOMID_MASK_HI) |
608 (event[1] & EVENT_DOMID_MASK_LO);
609 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
610 address = (u64)(((u64)event[3]) << 32) | event[2];
613 /* Did we hit the erratum? */
614 if (++count == LOOP_TIMEOUT) {
615 pr_err("No event written to event log\n");
622 if (type == EVENT_TYPE_IO_FAULT) {
623 amd_iommu_report_page_fault(iommu, devid, pasid, address, flags);
628 case EVENT_TYPE_ILL_DEV:
629 dev_err(dev, "Event logged [ILLEGAL_DEV_TABLE_ENTRY device=%04x:%02x:%02x.%x pasid=0x%05x address=0x%llx flags=0x%04x]\n",
630 iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
631 pasid, address, flags);
632 dump_dte_entry(iommu, devid);
634 case EVENT_TYPE_DEV_TAB_ERR:
635 dev_err(dev, "Event logged [DEV_TAB_HARDWARE_ERROR device=%04x:%02x:%02x.%x "
636 "address=0x%llx flags=0x%04x]\n",
637 iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
640 case EVENT_TYPE_PAGE_TAB_ERR:
641 dev_err(dev, "Event logged [PAGE_TAB_HARDWARE_ERROR device=%04x:%02x:%02x.%x pasid=0x%04x address=0x%llx flags=0x%04x]\n",
642 iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
643 pasid, address, flags);
645 case EVENT_TYPE_ILL_CMD:
646 dev_err(dev, "Event logged [ILLEGAL_COMMAND_ERROR address=0x%llx]\n", address);
647 dump_command(address);
649 case EVENT_TYPE_CMD_HARD_ERR:
650 dev_err(dev, "Event logged [COMMAND_HARDWARE_ERROR address=0x%llx flags=0x%04x]\n",
653 case EVENT_TYPE_IOTLB_INV_TO:
654 dev_err(dev, "Event logged [IOTLB_INV_TIMEOUT device=%04x:%02x:%02x.%x address=0x%llx]\n",
655 iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
658 case EVENT_TYPE_INV_DEV_REQ:
659 dev_err(dev, "Event logged [INVALID_DEVICE_REQUEST device=%04x:%02x:%02x.%x pasid=0x%05x address=0x%llx flags=0x%04x]\n",
660 iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
661 pasid, address, flags);
663 case EVENT_TYPE_RMP_FAULT:
664 amd_iommu_report_rmp_fault(iommu, event);
666 case EVENT_TYPE_RMP_HW_ERR:
667 amd_iommu_report_rmp_hw_error(iommu, event);
669 case EVENT_TYPE_INV_PPR_REQ:
670 pasid = PPR_PASID(*((u64 *)__evt));
671 tag = event[1] & 0x03FF;
672 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",
673 iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
674 pasid, address, flags, tag);
677 dev_err(dev, "Event logged [UNKNOWN event[0]=0x%08x event[1]=0x%08x event[2]=0x%08x event[3]=0x%08x\n",
678 event[0], event[1], event[2], event[3]);
682 * To detect the hardware errata 732 we need to clear the
683 * entry back to zero. This issue does not exist on SNP
684 * enabled system. Also this buffer is not writeable on
685 * SNP enabled system.
687 if (!amd_iommu_snp_en)
688 memset(__evt, 0, 4 * sizeof(u32));
691 static void iommu_poll_events(struct amd_iommu *iommu)
695 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
696 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
698 while (head != tail) {
699 iommu_print_event(iommu, iommu->evt_buf + head);
700 head = (head + EVENT_ENTRY_SIZE) % EVT_BUFFER_SIZE;
703 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
706 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
708 struct amd_iommu_fault fault;
710 if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
711 pr_err_ratelimited("Unknown PPR request received\n");
715 fault.address = raw[1];
716 fault.pasid = PPR_PASID(raw[0]);
717 fault.sbdf = PCI_SEG_DEVID_TO_SBDF(iommu->pci_seg->id, PPR_DEVID(raw[0]));
718 fault.tag = PPR_TAG(raw[0]);
719 fault.flags = PPR_FLAGS(raw[0]);
721 atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
724 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
728 if (iommu->ppr_log == NULL)
731 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
732 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
734 while (head != tail) {
739 raw = (u64 *)(iommu->ppr_log + head);
742 * Hardware bug: Interrupt may arrive before the entry is
743 * written to memory. If this happens we need to wait for the
746 for (i = 0; i < LOOP_TIMEOUT; ++i) {
747 if (PPR_REQ_TYPE(raw[0]) != 0)
752 /* Avoid memcpy function-call overhead */
757 * To detect the hardware errata 733 we need to clear the
758 * entry back to zero. This issue does not exist on SNP
759 * enabled system. Also this buffer is not writeable on
760 * SNP enabled system.
762 if (!amd_iommu_snp_en)
763 raw[0] = raw[1] = 0UL;
765 /* Update head pointer of hardware ring-buffer */
766 head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
767 writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
769 /* Handle PPR entry */
770 iommu_handle_ppr_entry(iommu, entry);
772 /* Refresh ring-buffer information */
773 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
774 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
778 #ifdef CONFIG_IRQ_REMAP
779 static int (*iommu_ga_log_notifier)(u32);
781 int amd_iommu_register_ga_log_notifier(int (*notifier)(u32))
783 iommu_ga_log_notifier = notifier;
787 EXPORT_SYMBOL(amd_iommu_register_ga_log_notifier);
789 static void iommu_poll_ga_log(struct amd_iommu *iommu)
793 if (iommu->ga_log == NULL)
796 head = readl(iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
797 tail = readl(iommu->mmio_base + MMIO_GA_TAIL_OFFSET);
799 while (head != tail) {
803 raw = (u64 *)(iommu->ga_log + head);
805 /* Avoid memcpy function-call overhead */
808 /* Update head pointer of hardware ring-buffer */
809 head = (head + GA_ENTRY_SIZE) % GA_LOG_SIZE;
810 writel(head, iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
812 /* Handle GA entry */
813 switch (GA_REQ_TYPE(log_entry)) {
815 if (!iommu_ga_log_notifier)
818 pr_debug("%s: devid=%#x, ga_tag=%#x\n",
819 __func__, GA_DEVID(log_entry),
822 if (iommu_ga_log_notifier(GA_TAG(log_entry)) != 0)
823 pr_err("GA log notifier failed.\n");
832 amd_iommu_set_pci_msi_domain(struct device *dev, struct amd_iommu *iommu)
834 if (!irq_remapping_enabled || !dev_is_pci(dev) ||
835 !pci_dev_has_default_msi_parent_domain(to_pci_dev(dev)))
838 dev_set_msi_domain(dev, iommu->ir_domain);
841 #else /* CONFIG_IRQ_REMAP */
843 amd_iommu_set_pci_msi_domain(struct device *dev, struct amd_iommu *iommu) { }
844 #endif /* !CONFIG_IRQ_REMAP */
846 static void amd_iommu_handle_irq(void *data, const char *evt_type,
847 u32 int_mask, u32 overflow_mask,
848 void (*int_handler)(struct amd_iommu *),
849 void (*overflow_handler)(struct amd_iommu *))
851 struct amd_iommu *iommu = (struct amd_iommu *) data;
852 u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
853 u32 mask = int_mask | overflow_mask;
855 while (status & mask) {
856 /* Enable interrupt sources again */
857 writel(mask, iommu->mmio_base + MMIO_STATUS_OFFSET);
860 pr_devel("Processing IOMMU (ivhd%d) %s Log\n",
861 iommu->index, evt_type);
865 if ((status & overflow_mask) && overflow_handler)
866 overflow_handler(iommu);
869 * Hardware bug: ERBT1312
870 * When re-enabling interrupt (by writing 1
871 * to clear the bit), the hardware might also try to set
872 * the interrupt bit in the event status register.
873 * In this scenario, the bit will be set, and disable
874 * subsequent interrupts.
876 * Workaround: The IOMMU driver should read back the
877 * status register and check if the interrupt bits are cleared.
878 * If not, driver will need to go through the interrupt handler
879 * again and re-clear the bits
881 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
885 irqreturn_t amd_iommu_int_thread_evtlog(int irq, void *data)
887 amd_iommu_handle_irq(data, "Evt", MMIO_STATUS_EVT_INT_MASK,
888 MMIO_STATUS_EVT_OVERFLOW_MASK,
889 iommu_poll_events, amd_iommu_restart_event_logging);
894 irqreturn_t amd_iommu_int_thread_pprlog(int irq, void *data)
896 amd_iommu_handle_irq(data, "PPR", MMIO_STATUS_PPR_INT_MASK,
897 MMIO_STATUS_PPR_OVERFLOW_MASK,
898 iommu_poll_ppr_log, amd_iommu_restart_ppr_log);
903 irqreturn_t amd_iommu_int_thread_galog(int irq, void *data)
905 #ifdef CONFIG_IRQ_REMAP
906 amd_iommu_handle_irq(data, "GA", MMIO_STATUS_GALOG_INT_MASK,
907 MMIO_STATUS_GALOG_OVERFLOW_MASK,
908 iommu_poll_ga_log, amd_iommu_restart_ga_log);
914 irqreturn_t amd_iommu_int_thread(int irq, void *data)
916 amd_iommu_int_thread_evtlog(irq, data);
917 amd_iommu_int_thread_pprlog(irq, data);
918 amd_iommu_int_thread_galog(irq, data);
923 irqreturn_t amd_iommu_int_handler(int irq, void *data)
925 return IRQ_WAKE_THREAD;
928 /****************************************************************************
930 * IOMMU command queuing functions
932 ****************************************************************************/
934 static int wait_on_sem(struct amd_iommu *iommu, u64 data)
938 while (*iommu->cmd_sem != data && i < LOOP_TIMEOUT) {
943 if (i == LOOP_TIMEOUT) {
944 pr_alert("Completion-Wait loop timed out\n");
951 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
952 struct iommu_cmd *cmd)
957 /* Copy command to buffer */
958 tail = iommu->cmd_buf_tail;
959 target = iommu->cmd_buf + tail;
960 memcpy(target, cmd, sizeof(*cmd));
962 tail = (tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
963 iommu->cmd_buf_tail = tail;
965 /* Tell the IOMMU about it */
966 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
969 static void build_completion_wait(struct iommu_cmd *cmd,
970 struct amd_iommu *iommu,
973 u64 paddr = iommu_virt_to_phys((void *)iommu->cmd_sem);
975 memset(cmd, 0, sizeof(*cmd));
976 cmd->data[0] = lower_32_bits(paddr) | CMD_COMPL_WAIT_STORE_MASK;
977 cmd->data[1] = upper_32_bits(paddr);
978 cmd->data[2] = lower_32_bits(data);
979 cmd->data[3] = upper_32_bits(data);
980 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
983 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
985 memset(cmd, 0, sizeof(*cmd));
986 cmd->data[0] = devid;
987 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
991 * Builds an invalidation address which is suitable for one page or multiple
992 * pages. Sets the size bit (S) as needed is more than one page is flushed.
994 static inline u64 build_inv_address(u64 address, size_t size)
996 u64 pages, end, msb_diff;
998 pages = iommu_num_pages(address, size, PAGE_SIZE);
1001 return address & PAGE_MASK;
1003 end = address + size - 1;
1006 * msb_diff would hold the index of the most significant bit that
1007 * flipped between the start and end.
1009 msb_diff = fls64(end ^ address) - 1;
1012 * Bits 63:52 are sign extended. If for some reason bit 51 is different
1013 * between the start and the end, invalidate everything.
1015 if (unlikely(msb_diff > 51)) {
1016 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
1019 * The msb-bit must be clear on the address. Just set all the
1022 address |= (1ull << msb_diff) - 1;
1025 /* Clear bits 11:0 */
1026 address &= PAGE_MASK;
1028 /* Set the size bit - we flush more than one 4kb page */
1029 return address | CMD_INV_IOMMU_PAGES_SIZE_MASK;
1032 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
1033 size_t size, u16 domid, int pde)
1035 u64 inv_address = build_inv_address(address, size);
1037 memset(cmd, 0, sizeof(*cmd));
1038 cmd->data[1] |= domid;
1039 cmd->data[2] = lower_32_bits(inv_address);
1040 cmd->data[3] = upper_32_bits(inv_address);
1041 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
1042 if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
1043 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
1046 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
1047 u64 address, size_t size)
1049 u64 inv_address = build_inv_address(address, size);
1051 memset(cmd, 0, sizeof(*cmd));
1052 cmd->data[0] = devid;
1053 cmd->data[0] |= (qdep & 0xff) << 24;
1054 cmd->data[1] = devid;
1055 cmd->data[2] = lower_32_bits(inv_address);
1056 cmd->data[3] = upper_32_bits(inv_address);
1057 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
1060 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, u32 pasid,
1061 u64 address, bool size)
1063 memset(cmd, 0, sizeof(*cmd));
1065 address &= ~(0xfffULL);
1067 cmd->data[0] = pasid;
1068 cmd->data[1] = domid;
1069 cmd->data[2] = lower_32_bits(address);
1070 cmd->data[3] = upper_32_bits(address);
1071 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
1072 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
1074 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
1075 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
1078 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, u32 pasid,
1079 int qdep, u64 address, bool size)
1081 memset(cmd, 0, sizeof(*cmd));
1083 address &= ~(0xfffULL);
1085 cmd->data[0] = devid;
1086 cmd->data[0] |= ((pasid >> 8) & 0xff) << 16;
1087 cmd->data[0] |= (qdep & 0xff) << 24;
1088 cmd->data[1] = devid;
1089 cmd->data[1] |= (pasid & 0xff) << 16;
1090 cmd->data[2] = lower_32_bits(address);
1091 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
1092 cmd->data[3] = upper_32_bits(address);
1094 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
1095 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
1098 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, u32 pasid,
1099 int status, int tag, bool gn)
1101 memset(cmd, 0, sizeof(*cmd));
1103 cmd->data[0] = devid;
1105 cmd->data[1] = pasid;
1106 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
1108 cmd->data[3] = tag & 0x1ff;
1109 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
1111 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
1114 static void build_inv_all(struct iommu_cmd *cmd)
1116 memset(cmd, 0, sizeof(*cmd));
1117 CMD_SET_TYPE(cmd, CMD_INV_ALL);
1120 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
1122 memset(cmd, 0, sizeof(*cmd));
1123 cmd->data[0] = devid;
1124 CMD_SET_TYPE(cmd, CMD_INV_IRT);
1128 * Writes the command to the IOMMUs command buffer and informs the
1129 * hardware about the new command.
1131 static int __iommu_queue_command_sync(struct amd_iommu *iommu,
1132 struct iommu_cmd *cmd,
1135 unsigned int count = 0;
1136 u32 left, next_tail;
1138 next_tail = (iommu->cmd_buf_tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
1140 left = (iommu->cmd_buf_head - next_tail) % CMD_BUFFER_SIZE;
1143 /* Skip udelay() the first time around */
1145 if (count == LOOP_TIMEOUT) {
1146 pr_err("Command buffer timeout\n");
1153 /* Update head and recheck remaining space */
1154 iommu->cmd_buf_head = readl(iommu->mmio_base +
1155 MMIO_CMD_HEAD_OFFSET);
1160 copy_cmd_to_buffer(iommu, cmd);
1162 /* Do we need to make sure all commands are processed? */
1163 iommu->need_sync = sync;
1168 static int iommu_queue_command_sync(struct amd_iommu *iommu,
1169 struct iommu_cmd *cmd,
1172 unsigned long flags;
1175 raw_spin_lock_irqsave(&iommu->lock, flags);
1176 ret = __iommu_queue_command_sync(iommu, cmd, sync);
1177 raw_spin_unlock_irqrestore(&iommu->lock, flags);
1182 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
1184 return iommu_queue_command_sync(iommu, cmd, true);
1188 * This function queues a completion wait command into the command
1189 * buffer of an IOMMU
1191 static int iommu_completion_wait(struct amd_iommu *iommu)
1193 struct iommu_cmd cmd;
1194 unsigned long flags;
1198 if (!iommu->need_sync)
1201 data = atomic64_add_return(1, &iommu->cmd_sem_val);
1202 build_completion_wait(&cmd, iommu, data);
1204 raw_spin_lock_irqsave(&iommu->lock, flags);
1206 ret = __iommu_queue_command_sync(iommu, &cmd, false);
1210 ret = wait_on_sem(iommu, data);
1213 raw_spin_unlock_irqrestore(&iommu->lock, flags);
1218 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
1220 struct iommu_cmd cmd;
1222 build_inv_dte(&cmd, devid);
1224 return iommu_queue_command(iommu, &cmd);
1227 static void amd_iommu_flush_dte_all(struct amd_iommu *iommu)
1230 u16 last_bdf = iommu->pci_seg->last_bdf;
1232 for (devid = 0; devid <= last_bdf; ++devid)
1233 iommu_flush_dte(iommu, devid);
1235 iommu_completion_wait(iommu);
1239 * This function uses heavy locking and may disable irqs for some time. But
1240 * this is no issue because it is only called during resume.
1242 static void amd_iommu_flush_tlb_all(struct amd_iommu *iommu)
1245 u16 last_bdf = iommu->pci_seg->last_bdf;
1247 for (dom_id = 0; dom_id <= last_bdf; ++dom_id) {
1248 struct iommu_cmd cmd;
1249 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1251 iommu_queue_command(iommu, &cmd);
1254 iommu_completion_wait(iommu);
1257 static void amd_iommu_flush_tlb_domid(struct amd_iommu *iommu, u32 dom_id)
1259 struct iommu_cmd cmd;
1261 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1263 iommu_queue_command(iommu, &cmd);
1265 iommu_completion_wait(iommu);
1268 static void amd_iommu_flush_all(struct amd_iommu *iommu)
1270 struct iommu_cmd cmd;
1272 build_inv_all(&cmd);
1274 iommu_queue_command(iommu, &cmd);
1275 iommu_completion_wait(iommu);
1278 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1280 struct iommu_cmd cmd;
1282 build_inv_irt(&cmd, devid);
1284 iommu_queue_command(iommu, &cmd);
1287 static void amd_iommu_flush_irt_all(struct amd_iommu *iommu)
1290 u16 last_bdf = iommu->pci_seg->last_bdf;
1292 if (iommu->irtcachedis_enabled)
1295 for (devid = 0; devid <= last_bdf; devid++)
1296 iommu_flush_irt(iommu, devid);
1298 iommu_completion_wait(iommu);
1301 void iommu_flush_all_caches(struct amd_iommu *iommu)
1303 if (iommu_feature(iommu, FEATURE_IA)) {
1304 amd_iommu_flush_all(iommu);
1306 amd_iommu_flush_dte_all(iommu);
1307 amd_iommu_flush_irt_all(iommu);
1308 amd_iommu_flush_tlb_all(iommu);
1313 * Command send function for flushing on-device TLB
1315 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1316 u64 address, size_t size)
1318 struct amd_iommu *iommu;
1319 struct iommu_cmd cmd;
1322 qdep = dev_data->ats.qdep;
1323 iommu = rlookup_amd_iommu(dev_data->dev);
1327 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1329 return iommu_queue_command(iommu, &cmd);
1332 static int device_flush_dte_alias(struct pci_dev *pdev, u16 alias, void *data)
1334 struct amd_iommu *iommu = data;
1336 return iommu_flush_dte(iommu, alias);
1340 * Command send function for invalidating a device table entry
1342 static int device_flush_dte(struct iommu_dev_data *dev_data)
1344 struct amd_iommu *iommu;
1345 struct pci_dev *pdev = NULL;
1346 struct amd_iommu_pci_seg *pci_seg;
1350 iommu = rlookup_amd_iommu(dev_data->dev);
1354 if (dev_is_pci(dev_data->dev))
1355 pdev = to_pci_dev(dev_data->dev);
1358 ret = pci_for_each_dma_alias(pdev,
1359 device_flush_dte_alias, iommu);
1361 ret = iommu_flush_dte(iommu, dev_data->devid);
1365 pci_seg = iommu->pci_seg;
1366 alias = pci_seg->alias_table[dev_data->devid];
1367 if (alias != dev_data->devid) {
1368 ret = iommu_flush_dte(iommu, alias);
1373 if (dev_data->ats.enabled)
1374 ret = device_flush_iotlb(dev_data, 0, ~0UL);
1380 * TLB invalidation function which is called from the mapping functions.
1381 * It invalidates a single PTE if the range to flush is within a single
1382 * page. Otherwise it flushes the whole TLB of the IOMMU.
1384 static void __domain_flush_pages(struct protection_domain *domain,
1385 u64 address, size_t size, int pde)
1387 struct iommu_dev_data *dev_data;
1388 struct iommu_cmd cmd;
1391 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1393 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
1394 if (!domain->dev_iommu[i])
1398 * Devices of this domain are behind this IOMMU
1399 * We need a TLB flush
1401 ret |= iommu_queue_command(amd_iommus[i], &cmd);
1404 list_for_each_entry(dev_data, &domain->dev_list, list) {
1406 if (!dev_data->ats.enabled)
1409 ret |= device_flush_iotlb(dev_data, address, size);
1415 static void domain_flush_pages(struct protection_domain *domain,
1416 u64 address, size_t size, int pde)
1418 if (likely(!amd_iommu_np_cache)) {
1419 __domain_flush_pages(domain, address, size, pde);
1424 * When NpCache is on, we infer that we run in a VM and use a vIOMMU.
1425 * In such setups it is best to avoid flushes of ranges which are not
1426 * naturally aligned, since it would lead to flushes of unmodified
1427 * PTEs. Such flushes would require the hypervisor to do more work than
1428 * necessary. Therefore, perform repeated flushes of aligned ranges
1429 * until you cover the range. Each iteration flushes the smaller
1430 * between the natural alignment of the address that we flush and the
1431 * greatest naturally aligned region that fits in the range.
1434 int addr_alignment = __ffs(address);
1435 int size_alignment = __fls(size);
1440 * size is always non-zero, but address might be zero, causing
1441 * addr_alignment to be negative. As the casting of the
1442 * argument in __ffs(address) to long might trim the high bits
1443 * of the address on x86-32, cast to long when doing the check.
1445 if (likely((unsigned long)address != 0))
1446 min_alignment = min(addr_alignment, size_alignment);
1448 min_alignment = size_alignment;
1450 flush_size = 1ul << min_alignment;
1452 __domain_flush_pages(domain, address, flush_size, pde);
1453 address += flush_size;
1458 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1459 void amd_iommu_domain_flush_tlb_pde(struct protection_domain *domain)
1461 domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1464 void amd_iommu_domain_flush_complete(struct protection_domain *domain)
1468 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
1469 if (domain && !domain->dev_iommu[i])
1473 * Devices of this domain are behind this IOMMU
1474 * We need to wait for completion of all commands.
1476 iommu_completion_wait(amd_iommus[i]);
1480 /* Flush the not present cache if it exists */
1481 static void domain_flush_np_cache(struct protection_domain *domain,
1482 dma_addr_t iova, size_t size)
1484 if (unlikely(amd_iommu_np_cache)) {
1485 unsigned long flags;
1487 spin_lock_irqsave(&domain->lock, flags);
1488 domain_flush_pages(domain, iova, size, 1);
1489 amd_iommu_domain_flush_complete(domain);
1490 spin_unlock_irqrestore(&domain->lock, flags);
1496 * This function flushes the DTEs for all devices in domain
1498 static void domain_flush_devices(struct protection_domain *domain)
1500 struct iommu_dev_data *dev_data;
1502 list_for_each_entry(dev_data, &domain->dev_list, list)
1503 device_flush_dte(dev_data);
1506 /****************************************************************************
1508 * The next functions belong to the domain allocation. A domain is
1509 * allocated for every IOMMU as the default domain. If device isolation
1510 * is enabled, every device get its own domain. The most important thing
1511 * about domains is the page table mapping the DMA address space they
1514 ****************************************************************************/
1516 static u16 domain_id_alloc(void)
1520 spin_lock(&pd_bitmap_lock);
1521 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1523 if (id > 0 && id < MAX_DOMAIN_ID)
1524 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1527 spin_unlock(&pd_bitmap_lock);
1532 static void domain_id_free(int id)
1534 spin_lock(&pd_bitmap_lock);
1535 if (id > 0 && id < MAX_DOMAIN_ID)
1536 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1537 spin_unlock(&pd_bitmap_lock);
1540 static void free_gcr3_tbl_level1(u64 *tbl)
1545 for (i = 0; i < 512; ++i) {
1546 if (!(tbl[i] & GCR3_VALID))
1549 ptr = iommu_phys_to_virt(tbl[i] & PAGE_MASK);
1551 free_page((unsigned long)ptr);
1555 static void free_gcr3_tbl_level2(u64 *tbl)
1560 for (i = 0; i < 512; ++i) {
1561 if (!(tbl[i] & GCR3_VALID))
1564 ptr = iommu_phys_to_virt(tbl[i] & PAGE_MASK);
1566 free_gcr3_tbl_level1(ptr);
1570 static void free_gcr3_table(struct protection_domain *domain)
1572 if (domain->glx == 2)
1573 free_gcr3_tbl_level2(domain->gcr3_tbl);
1574 else if (domain->glx == 1)
1575 free_gcr3_tbl_level1(domain->gcr3_tbl);
1577 BUG_ON(domain->glx != 0);
1579 free_page((unsigned long)domain->gcr3_tbl);
1582 static void set_dte_entry(struct amd_iommu *iommu, u16 devid,
1583 struct protection_domain *domain, bool ats, bool ppr)
1588 struct dev_table_entry *dev_table = get_dev_table(iommu);
1590 if (domain->iop.mode != PAGE_MODE_NONE)
1591 pte_root = iommu_virt_to_phys(domain->iop.root);
1593 pte_root |= (domain->iop.mode & DEV_ENTRY_MODE_MASK)
1594 << DEV_ENTRY_MODE_SHIFT;
1596 pte_root |= DTE_FLAG_IR | DTE_FLAG_IW | DTE_FLAG_V;
1599 * When SNP is enabled, Only set TV bit when IOMMU
1600 * page translation is in use.
1602 if (!amd_iommu_snp_en || (domain->id != 0))
1603 pte_root |= DTE_FLAG_TV;
1605 flags = dev_table[devid].data[1];
1608 flags |= DTE_FLAG_IOTLB;
1611 if (iommu_feature(iommu, FEATURE_EPHSUP))
1612 pte_root |= 1ULL << DEV_ENTRY_PPR;
1615 if (domain->dirty_tracking)
1616 pte_root |= DTE_FLAG_HAD;
1618 if (domain->flags & PD_IOMMUV2_MASK) {
1619 u64 gcr3 = iommu_virt_to_phys(domain->gcr3_tbl);
1620 u64 glx = domain->glx;
1623 pte_root |= DTE_FLAG_GV;
1624 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
1626 /* First mask out possible old values for GCR3 table */
1627 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
1630 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
1633 /* Encode GCR3 table into DTE */
1634 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
1637 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
1640 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
1643 if (amd_iommu_gpt_level == PAGE_MODE_5_LEVEL) {
1644 dev_table[devid].data[2] |=
1645 ((u64)GUEST_PGTABLE_5_LEVEL << DTE_GPT_LEVEL_SHIFT);
1648 if (domain->flags & PD_GIOV_MASK)
1649 pte_root |= DTE_FLAG_GIOV;
1652 flags &= ~DEV_DOMID_MASK;
1653 flags |= domain->id;
1655 old_domid = dev_table[devid].data[1] & DEV_DOMID_MASK;
1656 dev_table[devid].data[1] = flags;
1657 dev_table[devid].data[0] = pte_root;
1660 * A kdump kernel might be replacing a domain ID that was copied from
1661 * the previous kernel--if so, it needs to flush the translation cache
1662 * entries for the old domain ID that is being overwritten
1665 amd_iommu_flush_tlb_domid(iommu, old_domid);
1669 static void clear_dte_entry(struct amd_iommu *iommu, u16 devid)
1671 struct dev_table_entry *dev_table = get_dev_table(iommu);
1673 /* remove entry from the device table seen by the hardware */
1674 dev_table[devid].data[0] = DTE_FLAG_V;
1676 if (!amd_iommu_snp_en)
1677 dev_table[devid].data[0] |= DTE_FLAG_TV;
1679 dev_table[devid].data[1] &= DTE_FLAG_MASK;
1681 amd_iommu_apply_erratum_63(iommu, devid);
1684 static void do_attach(struct iommu_dev_data *dev_data,
1685 struct protection_domain *domain)
1687 struct amd_iommu *iommu;
1690 iommu = rlookup_amd_iommu(dev_data->dev);
1693 ats = dev_data->ats.enabled;
1695 /* Update data structures */
1696 dev_data->domain = domain;
1697 list_add(&dev_data->list, &domain->dev_list);
1699 /* Update NUMA Node ID */
1700 if (domain->nid == NUMA_NO_NODE)
1701 domain->nid = dev_to_node(dev_data->dev);
1703 /* Do reference counting */
1704 domain->dev_iommu[iommu->index] += 1;
1705 domain->dev_cnt += 1;
1707 /* Update device table */
1708 set_dte_entry(iommu, dev_data->devid, domain,
1709 ats, dev_data->iommu_v2);
1710 clone_aliases(iommu, dev_data->dev);
1712 device_flush_dte(dev_data);
1715 static void do_detach(struct iommu_dev_data *dev_data)
1717 struct protection_domain *domain = dev_data->domain;
1718 struct amd_iommu *iommu;
1720 iommu = rlookup_amd_iommu(dev_data->dev);
1724 /* Update data structures */
1725 dev_data->domain = NULL;
1726 list_del(&dev_data->list);
1727 clear_dte_entry(iommu, dev_data->devid);
1728 clone_aliases(iommu, dev_data->dev);
1730 /* Flush the DTE entry */
1731 device_flush_dte(dev_data);
1734 amd_iommu_domain_flush_tlb_pde(domain);
1736 /* Wait for the flushes to finish */
1737 amd_iommu_domain_flush_complete(domain);
1739 /* decrease reference counters - needs to happen after the flushes */
1740 domain->dev_iommu[iommu->index] -= 1;
1741 domain->dev_cnt -= 1;
1744 static void pdev_iommuv2_disable(struct pci_dev *pdev)
1746 pci_disable_ats(pdev);
1747 pci_disable_pri(pdev);
1748 pci_disable_pasid(pdev);
1751 static int pdev_pri_ats_enable(struct pci_dev *pdev)
1755 /* Only allow access to user-accessible pages */
1756 ret = pci_enable_pasid(pdev, 0);
1760 /* First reset the PRI state of the device */
1761 ret = pci_reset_pri(pdev);
1766 /* FIXME: Hardcode number of outstanding requests for now */
1767 ret = pci_enable_pri(pdev, 32);
1771 ret = pci_enable_ats(pdev, PAGE_SHIFT);
1778 pci_disable_pri(pdev);
1781 pci_disable_pasid(pdev);
1787 * If a device is not yet associated with a domain, this function makes the
1788 * device visible in the domain
1790 static int attach_device(struct device *dev,
1791 struct protection_domain *domain)
1793 struct iommu_dev_data *dev_data;
1794 struct pci_dev *pdev;
1795 unsigned long flags;
1798 spin_lock_irqsave(&domain->lock, flags);
1800 dev_data = dev_iommu_priv_get(dev);
1802 spin_lock(&dev_data->lock);
1805 if (dev_data->domain != NULL)
1808 if (!dev_is_pci(dev))
1809 goto skip_ats_check;
1811 pdev = to_pci_dev(dev);
1812 if (domain->flags & PD_IOMMUV2_MASK) {
1813 struct iommu_domain *def_domain = iommu_get_dma_domain(dev);
1818 * In case of using AMD_IOMMU_V1 page table mode and the device
1819 * is enabling for PPR/ATS support (using v2 table),
1820 * we need to make sure that the domain type is identity map.
1822 if ((amd_iommu_pgtable == AMD_IOMMU_V1) &&
1823 def_domain->type != IOMMU_DOMAIN_IDENTITY) {
1827 if (dev_data->iommu_v2) {
1828 if (pdev_pri_ats_enable(pdev) != 0)
1831 dev_data->ats.enabled = true;
1832 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
1833 dev_data->pri_tlp = pci_prg_resp_pasid_required(pdev);
1835 } else if (amd_iommu_iotlb_sup &&
1836 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
1837 dev_data->ats.enabled = true;
1838 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
1844 do_attach(dev_data, domain);
1847 * We might boot into a crash-kernel here. The crashed kernel
1848 * left the caches in the IOMMU dirty. So we have to flush
1849 * here to evict all dirty stuff.
1851 amd_iommu_domain_flush_tlb_pde(domain);
1853 amd_iommu_domain_flush_complete(domain);
1856 spin_unlock(&dev_data->lock);
1858 spin_unlock_irqrestore(&domain->lock, flags);
1864 * Removes a device from a protection domain (with devtable_lock held)
1866 static void detach_device(struct device *dev)
1868 struct protection_domain *domain;
1869 struct iommu_dev_data *dev_data;
1870 unsigned long flags;
1872 dev_data = dev_iommu_priv_get(dev);
1873 domain = dev_data->domain;
1875 spin_lock_irqsave(&domain->lock, flags);
1877 spin_lock(&dev_data->lock);
1880 * First check if the device is still attached. It might already
1881 * be detached from its domain because the generic
1882 * iommu_detach_group code detached it and we try again here in
1883 * our alias handling.
1885 if (WARN_ON(!dev_data->domain))
1888 do_detach(dev_data);
1890 if (!dev_is_pci(dev))
1893 if (domain->flags & PD_IOMMUV2_MASK && dev_data->iommu_v2)
1894 pdev_iommuv2_disable(to_pci_dev(dev));
1895 else if (dev_data->ats.enabled)
1896 pci_disable_ats(to_pci_dev(dev));
1898 dev_data->ats.enabled = false;
1901 spin_unlock(&dev_data->lock);
1903 spin_unlock_irqrestore(&domain->lock, flags);
1906 static struct iommu_device *amd_iommu_probe_device(struct device *dev)
1908 struct iommu_device *iommu_dev;
1909 struct amd_iommu *iommu;
1912 if (!check_device(dev))
1913 return ERR_PTR(-ENODEV);
1915 iommu = rlookup_amd_iommu(dev);
1917 return ERR_PTR(-ENODEV);
1919 /* Not registered yet? */
1920 if (!iommu->iommu.ops)
1921 return ERR_PTR(-ENODEV);
1923 if (dev_iommu_priv_get(dev))
1924 return &iommu->iommu;
1926 ret = iommu_init_device(iommu, dev);
1928 if (ret != -ENOTSUPP)
1929 dev_err(dev, "Failed to initialize - trying to proceed anyway\n");
1930 iommu_dev = ERR_PTR(ret);
1931 iommu_ignore_device(iommu, dev);
1933 amd_iommu_set_pci_msi_domain(dev, iommu);
1934 iommu_dev = &iommu->iommu;
1937 iommu_completion_wait(iommu);
1942 static void amd_iommu_probe_finalize(struct device *dev)
1944 /* Domains are initialized for this device - have a look what we ended up with */
1945 set_dma_ops(dev, NULL);
1946 iommu_setup_dma_ops(dev, 0, U64_MAX);
1949 static void amd_iommu_release_device(struct device *dev)
1951 struct amd_iommu *iommu;
1953 if (!check_device(dev))
1956 iommu = rlookup_amd_iommu(dev);
1960 amd_iommu_uninit_device(dev);
1961 iommu_completion_wait(iommu);
1964 static struct iommu_group *amd_iommu_device_group(struct device *dev)
1966 if (dev_is_pci(dev))
1967 return pci_device_group(dev);
1969 return acpihid_device_group(dev);
1972 /*****************************************************************************
1974 * The next functions belong to the dma_ops mapping/unmapping code.
1976 *****************************************************************************/
1978 static void update_device_table(struct protection_domain *domain)
1980 struct iommu_dev_data *dev_data;
1982 list_for_each_entry(dev_data, &domain->dev_list, list) {
1983 struct amd_iommu *iommu = rlookup_amd_iommu(dev_data->dev);
1987 set_dte_entry(iommu, dev_data->devid, domain,
1988 dev_data->ats.enabled, dev_data->iommu_v2);
1989 clone_aliases(iommu, dev_data->dev);
1993 void amd_iommu_update_and_flush_device_table(struct protection_domain *domain)
1995 update_device_table(domain);
1996 domain_flush_devices(domain);
1999 void amd_iommu_domain_update(struct protection_domain *domain)
2001 /* Update device table */
2002 amd_iommu_update_and_flush_device_table(domain);
2004 /* Flush domain TLB(s) and wait for completion */
2005 amd_iommu_domain_flush_tlb_pde(domain);
2006 amd_iommu_domain_flush_complete(domain);
2009 /*****************************************************************************
2011 * The following functions belong to the exported interface of AMD IOMMU
2013 * This interface allows access to lower level functions of the IOMMU
2014 * like protection domain handling and assignement of devices to domains
2015 * which is not possible with the dma_ops interface.
2017 *****************************************************************************/
2019 static void cleanup_domain(struct protection_domain *domain)
2021 struct iommu_dev_data *entry;
2022 unsigned long flags;
2024 spin_lock_irqsave(&domain->lock, flags);
2026 while (!list_empty(&domain->dev_list)) {
2027 entry = list_first_entry(&domain->dev_list,
2028 struct iommu_dev_data, list);
2029 BUG_ON(!entry->domain);
2033 spin_unlock_irqrestore(&domain->lock, flags);
2036 static void protection_domain_free(struct protection_domain *domain)
2041 if (domain->iop.pgtbl_cfg.tlb)
2042 free_io_pgtable_ops(&domain->iop.iop.ops);
2045 domain_id_free(domain->id);
2050 static int protection_domain_init_v1(struct protection_domain *domain, int mode)
2052 u64 *pt_root = NULL;
2054 BUG_ON(mode < PAGE_MODE_NONE || mode > PAGE_MODE_6_LEVEL);
2056 spin_lock_init(&domain->lock);
2057 domain->id = domain_id_alloc();
2060 INIT_LIST_HEAD(&domain->dev_list);
2062 if (mode != PAGE_MODE_NONE) {
2063 pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2065 domain_id_free(domain->id);
2070 amd_iommu_domain_set_pgtable(domain, pt_root, mode);
2075 static int protection_domain_init_v2(struct protection_domain *domain)
2077 spin_lock_init(&domain->lock);
2078 domain->id = domain_id_alloc();
2081 INIT_LIST_HEAD(&domain->dev_list);
2083 domain->flags |= PD_GIOV_MASK;
2085 domain->domain.pgsize_bitmap = AMD_IOMMU_PGSIZES_V2;
2087 if (domain_enable_v2(domain, 1)) {
2088 domain_id_free(domain->id);
2095 static struct protection_domain *protection_domain_alloc(unsigned int type)
2097 struct io_pgtable_ops *pgtbl_ops;
2098 struct protection_domain *domain;
2100 int mode = DEFAULT_PGTABLE_LEVEL;
2104 * Force IOMMU v1 page table when iommu=pt and
2105 * when allocating domain for pass-through devices.
2107 if (type == IOMMU_DOMAIN_IDENTITY) {
2108 pgtable = AMD_IOMMU_V1;
2109 mode = PAGE_MODE_NONE;
2110 } else if (type == IOMMU_DOMAIN_UNMANAGED) {
2111 pgtable = AMD_IOMMU_V1;
2112 } else if (type == IOMMU_DOMAIN_DMA || type == IOMMU_DOMAIN_DMA_FQ) {
2113 pgtable = amd_iommu_pgtable;
2118 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2124 ret = protection_domain_init_v1(domain, mode);
2127 ret = protection_domain_init_v2(domain);
2136 /* No need to allocate io pgtable ops in passthrough mode */
2137 if (type == IOMMU_DOMAIN_IDENTITY)
2140 domain->nid = NUMA_NO_NODE;
2142 pgtbl_ops = alloc_io_pgtable_ops(pgtable, &domain->iop.pgtbl_cfg, domain);
2144 domain_id_free(domain->id);
2154 static inline u64 dma_max_address(void)
2156 if (amd_iommu_pgtable == AMD_IOMMU_V1)
2159 /* V2 with 4/5 level page table */
2160 return ((1ULL << PM_LEVEL_SHIFT(amd_iommu_gpt_level)) - 1);
2163 static bool amd_iommu_hd_support(struct amd_iommu *iommu)
2165 return iommu && (iommu->features & FEATURE_HDSUP);
2168 static struct iommu_domain *do_iommu_domain_alloc(unsigned int type,
2169 struct device *dev, u32 flags)
2171 bool dirty_tracking = flags & IOMMU_HWPT_ALLOC_DIRTY_TRACKING;
2172 struct protection_domain *domain;
2173 struct amd_iommu *iommu = NULL;
2176 iommu = rlookup_amd_iommu(dev);
2178 return ERR_PTR(-ENODEV);
2182 * Since DTE[Mode]=0 is prohibited on SNP-enabled system,
2183 * default to use IOMMU_DOMAIN_DMA[_FQ].
2185 if (amd_iommu_snp_en && (type == IOMMU_DOMAIN_IDENTITY))
2186 return ERR_PTR(-EINVAL);
2188 if (dirty_tracking && !amd_iommu_hd_support(iommu))
2189 return ERR_PTR(-EOPNOTSUPP);
2191 domain = protection_domain_alloc(type);
2193 return ERR_PTR(-ENOMEM);
2195 domain->domain.geometry.aperture_start = 0;
2196 domain->domain.geometry.aperture_end = dma_max_address();
2197 domain->domain.geometry.force_aperture = true;
2200 domain->domain.type = type;
2201 domain->domain.pgsize_bitmap = iommu->iommu.ops->pgsize_bitmap;
2202 domain->domain.ops = iommu->iommu.ops->default_domain_ops;
2205 domain->domain.dirty_ops = &amd_dirty_ops;
2208 return &domain->domain;
2211 static struct iommu_domain *amd_iommu_domain_alloc(unsigned int type)
2213 struct iommu_domain *domain;
2215 domain = do_iommu_domain_alloc(type, NULL, 0);
2222 static struct iommu_domain *
2223 amd_iommu_domain_alloc_user(struct device *dev, u32 flags,
2224 struct iommu_domain *parent,
2225 const struct iommu_user_data *user_data)
2228 unsigned int type = IOMMU_DOMAIN_UNMANAGED;
2230 if ((flags & ~IOMMU_HWPT_ALLOC_DIRTY_TRACKING) || parent || user_data)
2231 return ERR_PTR(-EOPNOTSUPP);
2233 return do_iommu_domain_alloc(type, dev, flags);
2236 static void amd_iommu_domain_free(struct iommu_domain *dom)
2238 struct protection_domain *domain;
2240 domain = to_pdomain(dom);
2242 if (domain->dev_cnt > 0)
2243 cleanup_domain(domain);
2245 BUG_ON(domain->dev_cnt != 0);
2250 if (domain->flags & PD_IOMMUV2_MASK)
2251 free_gcr3_table(domain);
2253 protection_domain_free(domain);
2256 static int amd_iommu_attach_device(struct iommu_domain *dom,
2259 struct iommu_dev_data *dev_data = dev_iommu_priv_get(dev);
2260 struct protection_domain *domain = to_pdomain(dom);
2261 struct amd_iommu *iommu = rlookup_amd_iommu(dev);
2265 * Skip attach device to domain if new domain is same as
2266 * devices current domain
2268 if (dev_data->domain == domain)
2271 dev_data->defer_attach = false;
2274 * Restrict to devices with compatible IOMMU hardware support
2275 * when enforcement of dirty tracking is enabled.
2277 if (dom->dirty_ops && !amd_iommu_hd_support(iommu))
2280 if (dev_data->domain)
2283 ret = attach_device(dev, domain);
2285 #ifdef CONFIG_IRQ_REMAP
2286 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
2287 if (dom->type == IOMMU_DOMAIN_UNMANAGED)
2288 dev_data->use_vapic = 1;
2290 dev_data->use_vapic = 0;
2294 iommu_completion_wait(iommu);
2299 static void amd_iommu_iotlb_sync_map(struct iommu_domain *dom,
2300 unsigned long iova, size_t size)
2302 struct protection_domain *domain = to_pdomain(dom);
2303 struct io_pgtable_ops *ops = &domain->iop.iop.ops;
2306 domain_flush_np_cache(domain, iova, size);
2309 static int amd_iommu_map_pages(struct iommu_domain *dom, unsigned long iova,
2310 phys_addr_t paddr, size_t pgsize, size_t pgcount,
2311 int iommu_prot, gfp_t gfp, size_t *mapped)
2313 struct protection_domain *domain = to_pdomain(dom);
2314 struct io_pgtable_ops *ops = &domain->iop.iop.ops;
2318 if ((amd_iommu_pgtable == AMD_IOMMU_V1) &&
2319 (domain->iop.mode == PAGE_MODE_NONE))
2322 if (iommu_prot & IOMMU_READ)
2323 prot |= IOMMU_PROT_IR;
2324 if (iommu_prot & IOMMU_WRITE)
2325 prot |= IOMMU_PROT_IW;
2327 if (ops->map_pages) {
2328 ret = ops->map_pages(ops, iova, paddr, pgsize,
2329 pgcount, prot, gfp, mapped);
2335 static void amd_iommu_iotlb_gather_add_page(struct iommu_domain *domain,
2336 struct iommu_iotlb_gather *gather,
2337 unsigned long iova, size_t size)
2340 * AMD's IOMMU can flush as many pages as necessary in a single flush.
2341 * Unless we run in a virtual machine, which can be inferred according
2342 * to whether "non-present cache" is on, it is probably best to prefer
2343 * (potentially) too extensive TLB flushing (i.e., more misses) over
2344 * mutliple TLB flushes (i.e., more flushes). For virtual machines the
2345 * hypervisor needs to synchronize the host IOMMU PTEs with those of
2346 * the guest, and the trade-off is different: unnecessary TLB flushes
2347 * should be avoided.
2349 if (amd_iommu_np_cache &&
2350 iommu_iotlb_gather_is_disjoint(gather, iova, size))
2351 iommu_iotlb_sync(domain, gather);
2353 iommu_iotlb_gather_add_range(gather, iova, size);
2356 static size_t amd_iommu_unmap_pages(struct iommu_domain *dom, unsigned long iova,
2357 size_t pgsize, size_t pgcount,
2358 struct iommu_iotlb_gather *gather)
2360 struct protection_domain *domain = to_pdomain(dom);
2361 struct io_pgtable_ops *ops = &domain->iop.iop.ops;
2364 if ((amd_iommu_pgtable == AMD_IOMMU_V1) &&
2365 (domain->iop.mode == PAGE_MODE_NONE))
2368 r = (ops->unmap_pages) ? ops->unmap_pages(ops, iova, pgsize, pgcount, NULL) : 0;
2371 amd_iommu_iotlb_gather_add_page(dom, gather, iova, r);
2376 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
2379 struct protection_domain *domain = to_pdomain(dom);
2380 struct io_pgtable_ops *ops = &domain->iop.iop.ops;
2382 return ops->iova_to_phys(ops, iova);
2385 static bool amd_iommu_capable(struct device *dev, enum iommu_cap cap)
2388 case IOMMU_CAP_CACHE_COHERENCY:
2390 case IOMMU_CAP_NOEXEC:
2392 case IOMMU_CAP_PRE_BOOT_PROTECTION:
2393 return amdr_ivrs_remap_support;
2394 case IOMMU_CAP_ENFORCE_CACHE_COHERENCY:
2396 case IOMMU_CAP_DEFERRED_FLUSH:
2398 case IOMMU_CAP_DIRTY_TRACKING: {
2399 struct amd_iommu *iommu = rlookup_amd_iommu(dev);
2401 return amd_iommu_hd_support(iommu);
2410 static int amd_iommu_set_dirty_tracking(struct iommu_domain *domain,
2413 struct protection_domain *pdomain = to_pdomain(domain);
2414 struct dev_table_entry *dev_table;
2415 struct iommu_dev_data *dev_data;
2416 bool domain_flush = false;
2417 struct amd_iommu *iommu;
2418 unsigned long flags;
2421 spin_lock_irqsave(&pdomain->lock, flags);
2422 if (!(pdomain->dirty_tracking ^ enable)) {
2423 spin_unlock_irqrestore(&pdomain->lock, flags);
2427 list_for_each_entry(dev_data, &pdomain->dev_list, list) {
2428 iommu = rlookup_amd_iommu(dev_data->dev);
2432 dev_table = get_dev_table(iommu);
2433 pte_root = dev_table[dev_data->devid].data[0];
2435 pte_root = (enable ? pte_root | DTE_FLAG_HAD :
2436 pte_root & ~DTE_FLAG_HAD);
2438 /* Flush device DTE */
2439 dev_table[dev_data->devid].data[0] = pte_root;
2440 device_flush_dte(dev_data);
2441 domain_flush = true;
2444 /* Flush IOTLB to mark IOPTE dirty on the next translation(s) */
2446 amd_iommu_domain_flush_tlb_pde(pdomain);
2447 amd_iommu_domain_flush_complete(pdomain);
2449 pdomain->dirty_tracking = enable;
2450 spin_unlock_irqrestore(&pdomain->lock, flags);
2455 static int amd_iommu_read_and_clear_dirty(struct iommu_domain *domain,
2456 unsigned long iova, size_t size,
2457 unsigned long flags,
2458 struct iommu_dirty_bitmap *dirty)
2460 struct protection_domain *pdomain = to_pdomain(domain);
2461 struct io_pgtable_ops *ops = &pdomain->iop.iop.ops;
2462 unsigned long lflags;
2464 if (!ops || !ops->read_and_clear_dirty)
2467 spin_lock_irqsave(&pdomain->lock, lflags);
2468 if (!pdomain->dirty_tracking && dirty->bitmap) {
2469 spin_unlock_irqrestore(&pdomain->lock, lflags);
2472 spin_unlock_irqrestore(&pdomain->lock, lflags);
2474 return ops->read_and_clear_dirty(ops, iova, size, flags, dirty);
2477 static void amd_iommu_get_resv_regions(struct device *dev,
2478 struct list_head *head)
2480 struct iommu_resv_region *region;
2481 struct unity_map_entry *entry;
2482 struct amd_iommu *iommu;
2483 struct amd_iommu_pci_seg *pci_seg;
2486 sbdf = get_device_sbdf_id(dev);
2490 devid = PCI_SBDF_TO_DEVID(sbdf);
2491 iommu = rlookup_amd_iommu(dev);
2494 pci_seg = iommu->pci_seg;
2496 list_for_each_entry(entry, &pci_seg->unity_map, list) {
2500 if (devid < entry->devid_start || devid > entry->devid_end)
2503 type = IOMMU_RESV_DIRECT;
2504 length = entry->address_end - entry->address_start;
2505 if (entry->prot & IOMMU_PROT_IR)
2507 if (entry->prot & IOMMU_PROT_IW)
2508 prot |= IOMMU_WRITE;
2509 if (entry->prot & IOMMU_UNITY_MAP_FLAG_EXCL_RANGE)
2510 /* Exclusion range */
2511 type = IOMMU_RESV_RESERVED;
2513 region = iommu_alloc_resv_region(entry->address_start,
2517 dev_err(dev, "Out of memory allocating dm-regions\n");
2520 list_add_tail(®ion->list, head);
2523 region = iommu_alloc_resv_region(MSI_RANGE_START,
2524 MSI_RANGE_END - MSI_RANGE_START + 1,
2525 0, IOMMU_RESV_MSI, GFP_KERNEL);
2528 list_add_tail(®ion->list, head);
2530 region = iommu_alloc_resv_region(HT_RANGE_START,
2531 HT_RANGE_END - HT_RANGE_START + 1,
2532 0, IOMMU_RESV_RESERVED, GFP_KERNEL);
2535 list_add_tail(®ion->list, head);
2538 bool amd_iommu_is_attach_deferred(struct device *dev)
2540 struct iommu_dev_data *dev_data = dev_iommu_priv_get(dev);
2542 return dev_data->defer_attach;
2544 EXPORT_SYMBOL_GPL(amd_iommu_is_attach_deferred);
2546 static void amd_iommu_flush_iotlb_all(struct iommu_domain *domain)
2548 struct protection_domain *dom = to_pdomain(domain);
2549 unsigned long flags;
2551 spin_lock_irqsave(&dom->lock, flags);
2552 amd_iommu_domain_flush_tlb_pde(dom);
2553 amd_iommu_domain_flush_complete(dom);
2554 spin_unlock_irqrestore(&dom->lock, flags);
2557 static void amd_iommu_iotlb_sync(struct iommu_domain *domain,
2558 struct iommu_iotlb_gather *gather)
2560 struct protection_domain *dom = to_pdomain(domain);
2561 unsigned long flags;
2563 spin_lock_irqsave(&dom->lock, flags);
2564 domain_flush_pages(dom, gather->start, gather->end - gather->start + 1, 1);
2565 amd_iommu_domain_flush_complete(dom);
2566 spin_unlock_irqrestore(&dom->lock, flags);
2569 static int amd_iommu_def_domain_type(struct device *dev)
2571 struct iommu_dev_data *dev_data;
2573 dev_data = dev_iommu_priv_get(dev);
2578 * Do not identity map IOMMUv2 capable devices when:
2579 * - memory encryption is active, because some of those devices
2580 * (AMD GPUs) don't have the encryption bit in their DMA-mask
2581 * and require remapping.
2582 * - SNP is enabled, because it prohibits DTE[Mode]=0.
2584 if (dev_data->iommu_v2 &&
2585 !cc_platform_has(CC_ATTR_MEM_ENCRYPT) &&
2586 !amd_iommu_snp_en) {
2587 return IOMMU_DOMAIN_IDENTITY;
2593 static bool amd_iommu_enforce_cache_coherency(struct iommu_domain *domain)
2595 /* IOMMU_PTE_FC is always set */
2599 const struct iommu_dirty_ops amd_dirty_ops = {
2600 .set_dirty_tracking = amd_iommu_set_dirty_tracking,
2601 .read_and_clear_dirty = amd_iommu_read_and_clear_dirty,
2604 const struct iommu_ops amd_iommu_ops = {
2605 .capable = amd_iommu_capable,
2606 .domain_alloc = amd_iommu_domain_alloc,
2607 .domain_alloc_user = amd_iommu_domain_alloc_user,
2608 .probe_device = amd_iommu_probe_device,
2609 .release_device = amd_iommu_release_device,
2610 .probe_finalize = amd_iommu_probe_finalize,
2611 .device_group = amd_iommu_device_group,
2612 .get_resv_regions = amd_iommu_get_resv_regions,
2613 .is_attach_deferred = amd_iommu_is_attach_deferred,
2614 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
2615 .def_domain_type = amd_iommu_def_domain_type,
2616 .default_domain_ops = &(const struct iommu_domain_ops) {
2617 .attach_dev = amd_iommu_attach_device,
2618 .map_pages = amd_iommu_map_pages,
2619 .unmap_pages = amd_iommu_unmap_pages,
2620 .iotlb_sync_map = amd_iommu_iotlb_sync_map,
2621 .iova_to_phys = amd_iommu_iova_to_phys,
2622 .flush_iotlb_all = amd_iommu_flush_iotlb_all,
2623 .iotlb_sync = amd_iommu_iotlb_sync,
2624 .free = amd_iommu_domain_free,
2625 .enforce_cache_coherency = amd_iommu_enforce_cache_coherency,
2629 /*****************************************************************************
2631 * The next functions do a basic initialization of IOMMU for pass through
2634 * In passthrough mode the IOMMU is initialized and enabled but not used for
2635 * DMA-API translation.
2637 *****************************************************************************/
2639 /* IOMMUv2 specific functions */
2640 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
2642 return atomic_notifier_chain_register(&ppr_notifier, nb);
2644 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
2646 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
2648 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
2650 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
2652 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
2654 struct protection_domain *domain = to_pdomain(dom);
2655 unsigned long flags;
2657 spin_lock_irqsave(&domain->lock, flags);
2659 if (domain->iop.pgtbl_cfg.tlb)
2660 free_io_pgtable_ops(&domain->iop.iop.ops);
2662 spin_unlock_irqrestore(&domain->lock, flags);
2664 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
2666 /* Note: This function expects iommu_domain->lock to be held prior calling the function. */
2667 static int domain_enable_v2(struct protection_domain *domain, int pasids)
2671 /* Number of GCR3 table levels required */
2672 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
2675 if (levels > amd_iommu_max_glx_val)
2678 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
2679 if (domain->gcr3_tbl == NULL)
2682 domain->glx = levels;
2683 domain->flags |= PD_IOMMUV2_MASK;
2685 amd_iommu_domain_update(domain);
2690 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
2692 struct protection_domain *pdom = to_pdomain(dom);
2693 unsigned long flags;
2696 spin_lock_irqsave(&pdom->lock, flags);
2699 * Save us all sanity checks whether devices already in the
2700 * domain support IOMMUv2. Just force that the domain has no
2701 * devices attached when it is switched into IOMMUv2 mode.
2704 if (pdom->dev_cnt > 0 || pdom->flags & PD_IOMMUV2_MASK)
2707 if (!pdom->gcr3_tbl)
2708 ret = domain_enable_v2(pdom, pasids);
2711 spin_unlock_irqrestore(&pdom->lock, flags);
2714 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
2716 static int __flush_pasid(struct protection_domain *domain, u32 pasid,
2717 u64 address, bool size)
2719 struct iommu_dev_data *dev_data;
2720 struct iommu_cmd cmd;
2723 if (!(domain->flags & PD_IOMMUV2_MASK))
2726 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
2729 * IOMMU TLB needs to be flushed before Device TLB to
2730 * prevent device TLB refill from IOMMU TLB
2732 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
2733 if (domain->dev_iommu[i] == 0)
2736 ret = iommu_queue_command(amd_iommus[i], &cmd);
2741 /* Wait until IOMMU TLB flushes are complete */
2742 amd_iommu_domain_flush_complete(domain);
2744 /* Now flush device TLBs */
2745 list_for_each_entry(dev_data, &domain->dev_list, list) {
2746 struct amd_iommu *iommu;
2750 There might be non-IOMMUv2 capable devices in an IOMMUv2
2753 if (!dev_data->ats.enabled)
2756 qdep = dev_data->ats.qdep;
2757 iommu = rlookup_amd_iommu(dev_data->dev);
2760 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
2761 qdep, address, size);
2763 ret = iommu_queue_command(iommu, &cmd);
2768 /* Wait until all device TLBs are flushed */
2769 amd_iommu_domain_flush_complete(domain);
2778 static int __amd_iommu_flush_page(struct protection_domain *domain, u32 pasid,
2781 return __flush_pasid(domain, pasid, address, false);
2784 int amd_iommu_flush_page(struct iommu_domain *dom, u32 pasid,
2787 struct protection_domain *domain = to_pdomain(dom);
2788 unsigned long flags;
2791 spin_lock_irqsave(&domain->lock, flags);
2792 ret = __amd_iommu_flush_page(domain, pasid, address);
2793 spin_unlock_irqrestore(&domain->lock, flags);
2797 EXPORT_SYMBOL(amd_iommu_flush_page);
2799 static int __amd_iommu_flush_tlb(struct protection_domain *domain, u32 pasid)
2801 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
2805 int amd_iommu_flush_tlb(struct iommu_domain *dom, u32 pasid)
2807 struct protection_domain *domain = to_pdomain(dom);
2808 unsigned long flags;
2811 spin_lock_irqsave(&domain->lock, flags);
2812 ret = __amd_iommu_flush_tlb(domain, pasid);
2813 spin_unlock_irqrestore(&domain->lock, flags);
2817 EXPORT_SYMBOL(amd_iommu_flush_tlb);
2819 static u64 *__get_gcr3_pte(u64 *root, int level, u32 pasid, bool alloc)
2826 index = (pasid >> (9 * level)) & 0x1ff;
2832 if (!(*pte & GCR3_VALID)) {
2836 root = (void *)get_zeroed_page(GFP_ATOMIC);
2840 *pte = iommu_virt_to_phys(root) | GCR3_VALID;
2843 root = iommu_phys_to_virt(*pte & PAGE_MASK);
2851 static int __set_gcr3(struct protection_domain *domain, u32 pasid,
2856 if (domain->iop.mode != PAGE_MODE_NONE)
2859 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
2863 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
2865 return __amd_iommu_flush_tlb(domain, pasid);
2868 static int __clear_gcr3(struct protection_domain *domain, u32 pasid)
2872 if (domain->iop.mode != PAGE_MODE_NONE)
2875 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
2881 return __amd_iommu_flush_tlb(domain, pasid);
2884 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, u32 pasid,
2887 struct protection_domain *domain = to_pdomain(dom);
2888 unsigned long flags;
2891 spin_lock_irqsave(&domain->lock, flags);
2892 ret = __set_gcr3(domain, pasid, cr3);
2893 spin_unlock_irqrestore(&domain->lock, flags);
2897 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
2899 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, u32 pasid)
2901 struct protection_domain *domain = to_pdomain(dom);
2902 unsigned long flags;
2905 spin_lock_irqsave(&domain->lock, flags);
2906 ret = __clear_gcr3(domain, pasid);
2907 spin_unlock_irqrestore(&domain->lock, flags);
2911 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
2913 int amd_iommu_complete_ppr(struct pci_dev *pdev, u32 pasid,
2914 int status, int tag)
2916 struct iommu_dev_data *dev_data;
2917 struct amd_iommu *iommu;
2918 struct iommu_cmd cmd;
2920 dev_data = dev_iommu_priv_get(&pdev->dev);
2921 iommu = rlookup_amd_iommu(&pdev->dev);
2925 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
2926 tag, dev_data->pri_tlp);
2928 return iommu_queue_command(iommu, &cmd);
2930 EXPORT_SYMBOL(amd_iommu_complete_ppr);
2932 int amd_iommu_device_info(struct pci_dev *pdev,
2933 struct amd_iommu_device_info *info)
2938 if (pdev == NULL || info == NULL)
2941 if (!amd_iommu_v2_supported())
2944 memset(info, 0, sizeof(*info));
2946 if (pci_ats_supported(pdev))
2947 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
2949 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2951 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
2953 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
2957 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
2958 max_pasids = min(max_pasids, (1 << 20));
2960 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
2961 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
2963 features = pci_pasid_features(pdev);
2964 if (features & PCI_PASID_CAP_EXEC)
2965 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
2966 if (features & PCI_PASID_CAP_PRIV)
2967 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
2972 EXPORT_SYMBOL(amd_iommu_device_info);
2974 #ifdef CONFIG_IRQ_REMAP
2976 /*****************************************************************************
2978 * Interrupt Remapping Implementation
2980 *****************************************************************************/
2982 static struct irq_chip amd_ir_chip;
2983 static DEFINE_SPINLOCK(iommu_table_lock);
2985 static void iommu_flush_irt_and_complete(struct amd_iommu *iommu, u16 devid)
2989 unsigned long flags;
2990 struct iommu_cmd cmd, cmd2;
2992 if (iommu->irtcachedis_enabled)
2995 build_inv_irt(&cmd, devid);
2996 data = atomic64_add_return(1, &iommu->cmd_sem_val);
2997 build_completion_wait(&cmd2, iommu, data);
2999 raw_spin_lock_irqsave(&iommu->lock, flags);
3000 ret = __iommu_queue_command_sync(iommu, &cmd, true);
3003 ret = __iommu_queue_command_sync(iommu, &cmd2, false);
3006 wait_on_sem(iommu, data);
3008 raw_spin_unlock_irqrestore(&iommu->lock, flags);
3011 static void set_dte_irq_entry(struct amd_iommu *iommu, u16 devid,
3012 struct irq_remap_table *table)
3015 struct dev_table_entry *dev_table = get_dev_table(iommu);
3017 dte = dev_table[devid].data[2];
3018 dte &= ~DTE_IRQ_PHYS_ADDR_MASK;
3019 dte |= iommu_virt_to_phys(table->table);
3020 dte |= DTE_IRQ_REMAP_INTCTL;
3021 dte |= DTE_INTTABLEN;
3022 dte |= DTE_IRQ_REMAP_ENABLE;
3024 dev_table[devid].data[2] = dte;
3027 static struct irq_remap_table *get_irq_table(struct amd_iommu *iommu, u16 devid)
3029 struct irq_remap_table *table;
3030 struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
3032 if (WARN_ONCE(!pci_seg->rlookup_table[devid],
3033 "%s: no iommu for devid %x:%x\n",
3034 __func__, pci_seg->id, devid))
3037 table = pci_seg->irq_lookup_table[devid];
3038 if (WARN_ONCE(!table, "%s: no table for devid %x:%x\n",
3039 __func__, pci_seg->id, devid))
3045 static struct irq_remap_table *__alloc_irq_table(void)
3047 struct irq_remap_table *table;
3049 table = kzalloc(sizeof(*table), GFP_KERNEL);
3053 table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_KERNEL);
3054 if (!table->table) {
3058 raw_spin_lock_init(&table->lock);
3060 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
3061 memset(table->table, 0,
3062 MAX_IRQS_PER_TABLE * sizeof(u32));
3064 memset(table->table, 0,
3065 (MAX_IRQS_PER_TABLE * (sizeof(u64) * 2)));
3069 static void set_remap_table_entry(struct amd_iommu *iommu, u16 devid,
3070 struct irq_remap_table *table)
3072 struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg;
3074 pci_seg->irq_lookup_table[devid] = table;
3075 set_dte_irq_entry(iommu, devid, table);
3076 iommu_flush_dte(iommu, devid);
3079 static int set_remap_table_entry_alias(struct pci_dev *pdev, u16 alias,
3082 struct irq_remap_table *table = data;
3083 struct amd_iommu_pci_seg *pci_seg;
3084 struct amd_iommu *iommu = rlookup_amd_iommu(&pdev->dev);
3089 pci_seg = iommu->pci_seg;
3090 pci_seg->irq_lookup_table[alias] = table;
3091 set_dte_irq_entry(iommu, alias, table);
3092 iommu_flush_dte(pci_seg->rlookup_table[alias], alias);
3097 static struct irq_remap_table *alloc_irq_table(struct amd_iommu *iommu,
3098 u16 devid, struct pci_dev *pdev)
3100 struct irq_remap_table *table = NULL;
3101 struct irq_remap_table *new_table = NULL;
3102 struct amd_iommu_pci_seg *pci_seg;
3103 unsigned long flags;
3106 spin_lock_irqsave(&iommu_table_lock, flags);
3108 pci_seg = iommu->pci_seg;
3109 table = pci_seg->irq_lookup_table[devid];
3113 alias = pci_seg->alias_table[devid];
3114 table = pci_seg->irq_lookup_table[alias];
3116 set_remap_table_entry(iommu, devid, table);
3119 spin_unlock_irqrestore(&iommu_table_lock, flags);
3121 /* Nothing there yet, allocate new irq remapping table */
3122 new_table = __alloc_irq_table();
3126 spin_lock_irqsave(&iommu_table_lock, flags);
3128 table = pci_seg->irq_lookup_table[devid];
3132 table = pci_seg->irq_lookup_table[alias];
3134 set_remap_table_entry(iommu, devid, table);
3142 pci_for_each_dma_alias(pdev, set_remap_table_entry_alias,
3145 set_remap_table_entry(iommu, devid, table);
3148 set_remap_table_entry(iommu, alias, table);
3151 iommu_completion_wait(iommu);
3154 spin_unlock_irqrestore(&iommu_table_lock, flags);
3157 kmem_cache_free(amd_iommu_irq_cache, new_table->table);
3163 static int alloc_irq_index(struct amd_iommu *iommu, u16 devid, int count,
3164 bool align, struct pci_dev *pdev)
3166 struct irq_remap_table *table;
3167 int index, c, alignment = 1;
3168 unsigned long flags;
3170 table = alloc_irq_table(iommu, devid, pdev);
3175 alignment = roundup_pow_of_two(count);
3177 raw_spin_lock_irqsave(&table->lock, flags);
3179 /* Scan table for free entries */
3180 for (index = ALIGN(table->min_index, alignment), c = 0;
3181 index < MAX_IRQS_PER_TABLE;) {
3182 if (!iommu->irte_ops->is_allocated(table, index)) {
3186 index = ALIGN(index + 1, alignment);
3192 iommu->irte_ops->set_allocated(table, index - c + 1);
3204 raw_spin_unlock_irqrestore(&table->lock, flags);
3209 static int modify_irte_ga(struct amd_iommu *iommu, u16 devid, int index,
3210 struct irte_ga *irte)
3212 struct irq_remap_table *table;
3213 struct irte_ga *entry;
3214 unsigned long flags;
3217 table = get_irq_table(iommu, devid);
3221 raw_spin_lock_irqsave(&table->lock, flags);
3223 entry = (struct irte_ga *)table->table;
3224 entry = &entry[index];
3227 * We use cmpxchg16 to atomically update the 128-bit IRTE,
3228 * and it cannot be updated by the hardware or other processors
3229 * behind us, so the return value of cmpxchg16 should be the
3230 * same as the old value.
3233 WARN_ON(!try_cmpxchg128(&entry->irte, &old, irte->irte));
3235 raw_spin_unlock_irqrestore(&table->lock, flags);
3237 iommu_flush_irt_and_complete(iommu, devid);
3242 static int modify_irte(struct amd_iommu *iommu,
3243 u16 devid, int index, union irte *irte)
3245 struct irq_remap_table *table;
3246 unsigned long flags;
3248 table = get_irq_table(iommu, devid);
3252 raw_spin_lock_irqsave(&table->lock, flags);
3253 table->table[index] = irte->val;
3254 raw_spin_unlock_irqrestore(&table->lock, flags);
3256 iommu_flush_irt_and_complete(iommu, devid);
3261 static void free_irte(struct amd_iommu *iommu, u16 devid, int index)
3263 struct irq_remap_table *table;
3264 unsigned long flags;
3266 table = get_irq_table(iommu, devid);
3270 raw_spin_lock_irqsave(&table->lock, flags);
3271 iommu->irte_ops->clear_allocated(table, index);
3272 raw_spin_unlock_irqrestore(&table->lock, flags);
3274 iommu_flush_irt_and_complete(iommu, devid);
3277 static void irte_prepare(void *entry,
3278 u32 delivery_mode, bool dest_mode,
3279 u8 vector, u32 dest_apicid, int devid)
3281 union irte *irte = (union irte *) entry;
3284 irte->fields.vector = vector;
3285 irte->fields.int_type = delivery_mode;
3286 irte->fields.destination = dest_apicid;
3287 irte->fields.dm = dest_mode;
3288 irte->fields.valid = 1;
3291 static void irte_ga_prepare(void *entry,
3292 u32 delivery_mode, bool dest_mode,
3293 u8 vector, u32 dest_apicid, int devid)
3295 struct irte_ga *irte = (struct irte_ga *) entry;
3299 irte->lo.fields_remap.int_type = delivery_mode;
3300 irte->lo.fields_remap.dm = dest_mode;
3301 irte->hi.fields.vector = vector;
3302 irte->lo.fields_remap.destination = APICID_TO_IRTE_DEST_LO(dest_apicid);
3303 irte->hi.fields.destination = APICID_TO_IRTE_DEST_HI(dest_apicid);
3304 irte->lo.fields_remap.valid = 1;
3307 static void irte_activate(struct amd_iommu *iommu, void *entry, u16 devid, u16 index)
3309 union irte *irte = (union irte *) entry;
3311 irte->fields.valid = 1;
3312 modify_irte(iommu, devid, index, irte);
3315 static void irte_ga_activate(struct amd_iommu *iommu, void *entry, u16 devid, u16 index)
3317 struct irte_ga *irte = (struct irte_ga *) entry;
3319 irte->lo.fields_remap.valid = 1;
3320 modify_irte_ga(iommu, devid, index, irte);
3323 static void irte_deactivate(struct amd_iommu *iommu, void *entry, u16 devid, u16 index)
3325 union irte *irte = (union irte *) entry;
3327 irte->fields.valid = 0;
3328 modify_irte(iommu, devid, index, irte);
3331 static void irte_ga_deactivate(struct amd_iommu *iommu, void *entry, u16 devid, u16 index)
3333 struct irte_ga *irte = (struct irte_ga *) entry;
3335 irte->lo.fields_remap.valid = 0;
3336 modify_irte_ga(iommu, devid, index, irte);
3339 static void irte_set_affinity(struct amd_iommu *iommu, void *entry, u16 devid, u16 index,
3340 u8 vector, u32 dest_apicid)
3342 union irte *irte = (union irte *) entry;
3344 irte->fields.vector = vector;
3345 irte->fields.destination = dest_apicid;
3346 modify_irte(iommu, devid, index, irte);
3349 static void irte_ga_set_affinity(struct amd_iommu *iommu, void *entry, u16 devid, u16 index,
3350 u8 vector, u32 dest_apicid)
3352 struct irte_ga *irte = (struct irte_ga *) entry;
3354 if (!irte->lo.fields_remap.guest_mode) {
3355 irte->hi.fields.vector = vector;
3356 irte->lo.fields_remap.destination =
3357 APICID_TO_IRTE_DEST_LO(dest_apicid);
3358 irte->hi.fields.destination =
3359 APICID_TO_IRTE_DEST_HI(dest_apicid);
3360 modify_irte_ga(iommu, devid, index, irte);
3364 #define IRTE_ALLOCATED (~1U)
3365 static void irte_set_allocated(struct irq_remap_table *table, int index)
3367 table->table[index] = IRTE_ALLOCATED;
3370 static void irte_ga_set_allocated(struct irq_remap_table *table, int index)
3372 struct irte_ga *ptr = (struct irte_ga *)table->table;
3373 struct irte_ga *irte = &ptr[index];
3375 memset(&irte->lo.val, 0, sizeof(u64));
3376 memset(&irte->hi.val, 0, sizeof(u64));
3377 irte->hi.fields.vector = 0xff;
3380 static bool irte_is_allocated(struct irq_remap_table *table, int index)
3382 union irte *ptr = (union irte *)table->table;
3383 union irte *irte = &ptr[index];
3385 return irte->val != 0;
3388 static bool irte_ga_is_allocated(struct irq_remap_table *table, int index)
3390 struct irte_ga *ptr = (struct irte_ga *)table->table;
3391 struct irte_ga *irte = &ptr[index];
3393 return irte->hi.fields.vector != 0;
3396 static void irte_clear_allocated(struct irq_remap_table *table, int index)
3398 table->table[index] = 0;
3401 static void irte_ga_clear_allocated(struct irq_remap_table *table, int index)
3403 struct irte_ga *ptr = (struct irte_ga *)table->table;
3404 struct irte_ga *irte = &ptr[index];
3406 memset(&irte->lo.val, 0, sizeof(u64));
3407 memset(&irte->hi.val, 0, sizeof(u64));
3410 static int get_devid(struct irq_alloc_info *info)
3412 switch (info->type) {
3413 case X86_IRQ_ALLOC_TYPE_IOAPIC:
3414 return get_ioapic_devid(info->devid);
3415 case X86_IRQ_ALLOC_TYPE_HPET:
3416 return get_hpet_devid(info->devid);
3417 case X86_IRQ_ALLOC_TYPE_PCI_MSI:
3418 case X86_IRQ_ALLOC_TYPE_PCI_MSIX:
3419 return get_device_sbdf_id(msi_desc_to_dev(info->desc));
3426 struct irq_remap_ops amd_iommu_irq_ops = {
3427 .prepare = amd_iommu_prepare,
3428 .enable = amd_iommu_enable,
3429 .disable = amd_iommu_disable,
3430 .reenable = amd_iommu_reenable,
3431 .enable_faulting = amd_iommu_enable_faulting,
3434 static void fill_msi_msg(struct msi_msg *msg, u32 index)
3437 msg->address_lo = 0;
3438 msg->arch_addr_lo.base_address = X86_MSI_BASE_ADDRESS_LOW;
3439 msg->address_hi = X86_MSI_BASE_ADDRESS_HIGH;
3442 static void irq_remapping_prepare_irte(struct amd_ir_data *data,
3443 struct irq_cfg *irq_cfg,
3444 struct irq_alloc_info *info,
3445 int devid, int index, int sub_handle)
3447 struct irq_2_irte *irte_info = &data->irq_2_irte;
3448 struct amd_iommu *iommu = data->iommu;
3453 data->irq_2_irte.devid = devid;
3454 data->irq_2_irte.index = index + sub_handle;
3455 iommu->irte_ops->prepare(data->entry, apic->delivery_mode,
3456 apic->dest_mode_logical, irq_cfg->vector,
3457 irq_cfg->dest_apicid, devid);
3459 switch (info->type) {
3460 case X86_IRQ_ALLOC_TYPE_IOAPIC:
3461 case X86_IRQ_ALLOC_TYPE_HPET:
3462 case X86_IRQ_ALLOC_TYPE_PCI_MSI:
3463 case X86_IRQ_ALLOC_TYPE_PCI_MSIX:
3464 fill_msi_msg(&data->msi_entry, irte_info->index);
3473 struct amd_irte_ops irte_32_ops = {
3474 .prepare = irte_prepare,
3475 .activate = irte_activate,
3476 .deactivate = irte_deactivate,
3477 .set_affinity = irte_set_affinity,
3478 .set_allocated = irte_set_allocated,
3479 .is_allocated = irte_is_allocated,
3480 .clear_allocated = irte_clear_allocated,
3483 struct amd_irte_ops irte_128_ops = {
3484 .prepare = irte_ga_prepare,
3485 .activate = irte_ga_activate,
3486 .deactivate = irte_ga_deactivate,
3487 .set_affinity = irte_ga_set_affinity,
3488 .set_allocated = irte_ga_set_allocated,
3489 .is_allocated = irte_ga_is_allocated,
3490 .clear_allocated = irte_ga_clear_allocated,
3493 static int irq_remapping_alloc(struct irq_domain *domain, unsigned int virq,
3494 unsigned int nr_irqs, void *arg)
3496 struct irq_alloc_info *info = arg;
3497 struct irq_data *irq_data;
3498 struct amd_ir_data *data = NULL;
3499 struct amd_iommu *iommu;
3500 struct irq_cfg *cfg;
3501 int i, ret, devid, seg, sbdf;
3506 if (nr_irqs > 1 && info->type != X86_IRQ_ALLOC_TYPE_PCI_MSI)
3509 sbdf = get_devid(info);
3513 seg = PCI_SBDF_TO_SEGID(sbdf);
3514 devid = PCI_SBDF_TO_DEVID(sbdf);
3515 iommu = __rlookup_amd_iommu(seg, devid);
3519 ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
3523 if (info->type == X86_IRQ_ALLOC_TYPE_IOAPIC) {
3524 struct irq_remap_table *table;
3526 table = alloc_irq_table(iommu, devid, NULL);
3528 if (!table->min_index) {
3530 * Keep the first 32 indexes free for IOAPIC
3533 table->min_index = 32;
3534 for (i = 0; i < 32; ++i)
3535 iommu->irte_ops->set_allocated(table, i);
3537 WARN_ON(table->min_index != 32);
3538 index = info->ioapic.pin;
3542 } else if (info->type == X86_IRQ_ALLOC_TYPE_PCI_MSI ||
3543 info->type == X86_IRQ_ALLOC_TYPE_PCI_MSIX) {
3544 bool align = (info->type == X86_IRQ_ALLOC_TYPE_PCI_MSI);
3546 index = alloc_irq_index(iommu, devid, nr_irqs, align,
3547 msi_desc_to_pci_dev(info->desc));
3549 index = alloc_irq_index(iommu, devid, nr_irqs, false, NULL);
3553 pr_warn("Failed to allocate IRTE\n");
3555 goto out_free_parent;
3558 for (i = 0; i < nr_irqs; i++) {
3559 irq_data = irq_domain_get_irq_data(domain, virq + i);
3560 cfg = irq_data ? irqd_cfg(irq_data) : NULL;
3567 data = kzalloc(sizeof(*data), GFP_KERNEL);
3571 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
3572 data->entry = kzalloc(sizeof(union irte), GFP_KERNEL);
3574 data->entry = kzalloc(sizeof(struct irte_ga),
3581 data->iommu = iommu;
3582 irq_data->hwirq = (devid << 16) + i;
3583 irq_data->chip_data = data;
3584 irq_data->chip = &amd_ir_chip;
3585 irq_remapping_prepare_irte(data, cfg, info, devid, index, i);
3586 irq_set_status_flags(virq + i, IRQ_MOVE_PCNTXT);
3592 for (i--; i >= 0; i--) {
3593 irq_data = irq_domain_get_irq_data(domain, virq + i);
3595 kfree(irq_data->chip_data);
3597 for (i = 0; i < nr_irqs; i++)
3598 free_irte(iommu, devid, index + i);
3600 irq_domain_free_irqs_common(domain, virq, nr_irqs);
3604 static void irq_remapping_free(struct irq_domain *domain, unsigned int virq,
3605 unsigned int nr_irqs)
3607 struct irq_2_irte *irte_info;
3608 struct irq_data *irq_data;
3609 struct amd_ir_data *data;
3612 for (i = 0; i < nr_irqs; i++) {
3613 irq_data = irq_domain_get_irq_data(domain, virq + i);
3614 if (irq_data && irq_data->chip_data) {
3615 data = irq_data->chip_data;
3616 irte_info = &data->irq_2_irte;
3617 free_irte(data->iommu, irte_info->devid, irte_info->index);
3622 irq_domain_free_irqs_common(domain, virq, nr_irqs);
3625 static void amd_ir_update_irte(struct irq_data *irqd, struct amd_iommu *iommu,
3626 struct amd_ir_data *ir_data,
3627 struct irq_2_irte *irte_info,
3628 struct irq_cfg *cfg);
3630 static int irq_remapping_activate(struct irq_domain *domain,
3631 struct irq_data *irq_data, bool reserve)
3633 struct amd_ir_data *data = irq_data->chip_data;
3634 struct irq_2_irte *irte_info = &data->irq_2_irte;
3635 struct amd_iommu *iommu = data->iommu;
3636 struct irq_cfg *cfg = irqd_cfg(irq_data);
3641 iommu->irte_ops->activate(iommu, data->entry, irte_info->devid,
3643 amd_ir_update_irte(irq_data, iommu, data, irte_info, cfg);
3647 static void irq_remapping_deactivate(struct irq_domain *domain,
3648 struct irq_data *irq_data)
3650 struct amd_ir_data *data = irq_data->chip_data;
3651 struct irq_2_irte *irte_info = &data->irq_2_irte;
3652 struct amd_iommu *iommu = data->iommu;
3655 iommu->irte_ops->deactivate(iommu, data->entry, irte_info->devid,
3659 static int irq_remapping_select(struct irq_domain *d, struct irq_fwspec *fwspec,
3660 enum irq_domain_bus_token bus_token)
3662 struct amd_iommu *iommu;
3665 if (!amd_iommu_irq_remap)
3668 if (x86_fwspec_is_ioapic(fwspec))
3669 devid = get_ioapic_devid(fwspec->param[0]);
3670 else if (x86_fwspec_is_hpet(fwspec))
3671 devid = get_hpet_devid(fwspec->param[0]);
3675 iommu = __rlookup_amd_iommu((devid >> 16), (devid & 0xffff));
3677 return iommu && iommu->ir_domain == d;
3680 static const struct irq_domain_ops amd_ir_domain_ops = {
3681 .select = irq_remapping_select,
3682 .alloc = irq_remapping_alloc,
3683 .free = irq_remapping_free,
3684 .activate = irq_remapping_activate,
3685 .deactivate = irq_remapping_deactivate,
3688 int amd_iommu_activate_guest_mode(void *data)
3690 struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
3691 struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
3694 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) || !entry)
3697 valid = entry->lo.fields_vapic.valid;
3702 entry->lo.fields_vapic.valid = valid;
3703 entry->lo.fields_vapic.guest_mode = 1;
3704 entry->lo.fields_vapic.ga_log_intr = 1;
3705 entry->hi.fields.ga_root_ptr = ir_data->ga_root_ptr;
3706 entry->hi.fields.vector = ir_data->ga_vector;
3707 entry->lo.fields_vapic.ga_tag = ir_data->ga_tag;
3709 return modify_irte_ga(ir_data->iommu, ir_data->irq_2_irte.devid,
3710 ir_data->irq_2_irte.index, entry);
3712 EXPORT_SYMBOL(amd_iommu_activate_guest_mode);
3714 int amd_iommu_deactivate_guest_mode(void *data)
3716 struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
3717 struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
3718 struct irq_cfg *cfg = ir_data->cfg;
3721 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
3722 !entry || !entry->lo.fields_vapic.guest_mode)
3725 valid = entry->lo.fields_remap.valid;
3730 entry->lo.fields_remap.valid = valid;
3731 entry->lo.fields_remap.dm = apic->dest_mode_logical;
3732 entry->lo.fields_remap.int_type = apic->delivery_mode;
3733 entry->hi.fields.vector = cfg->vector;
3734 entry->lo.fields_remap.destination =
3735 APICID_TO_IRTE_DEST_LO(cfg->dest_apicid);
3736 entry->hi.fields.destination =
3737 APICID_TO_IRTE_DEST_HI(cfg->dest_apicid);
3739 return modify_irte_ga(ir_data->iommu, ir_data->irq_2_irte.devid,
3740 ir_data->irq_2_irte.index, entry);
3742 EXPORT_SYMBOL(amd_iommu_deactivate_guest_mode);
3744 static int amd_ir_set_vcpu_affinity(struct irq_data *data, void *vcpu_info)
3747 struct amd_iommu_pi_data *pi_data = vcpu_info;
3748 struct vcpu_data *vcpu_pi_info = pi_data->vcpu_data;
3749 struct amd_ir_data *ir_data = data->chip_data;
3750 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
3751 struct iommu_dev_data *dev_data;
3753 if (ir_data->iommu == NULL)
3756 dev_data = search_dev_data(ir_data->iommu, irte_info->devid);
3759 * This device has never been set up for guest mode.
3760 * we should not modify the IRTE
3762 if (!dev_data || !dev_data->use_vapic)
3765 ir_data->cfg = irqd_cfg(data);
3766 pi_data->ir_data = ir_data;
3769 * SVM tries to set up for VAPIC mode, but we are in
3770 * legacy mode. So, we force legacy mode instead.
3772 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
3773 pr_debug("%s: Fall back to using intr legacy remap\n",
3775 pi_data->is_guest_mode = false;
3778 pi_data->prev_ga_tag = ir_data->cached_ga_tag;
3779 if (pi_data->is_guest_mode) {
3780 ir_data->ga_root_ptr = (pi_data->base >> 12);
3781 ir_data->ga_vector = vcpu_pi_info->vector;
3782 ir_data->ga_tag = pi_data->ga_tag;
3783 ret = amd_iommu_activate_guest_mode(ir_data);
3785 ir_data->cached_ga_tag = pi_data->ga_tag;
3787 ret = amd_iommu_deactivate_guest_mode(ir_data);
3790 * This communicates the ga_tag back to the caller
3791 * so that it can do all the necessary clean up.
3794 ir_data->cached_ga_tag = 0;
3801 static void amd_ir_update_irte(struct irq_data *irqd, struct amd_iommu *iommu,
3802 struct amd_ir_data *ir_data,
3803 struct irq_2_irte *irte_info,
3804 struct irq_cfg *cfg)
3808 * Atomically updates the IRTE with the new destination, vector
3809 * and flushes the interrupt entry cache.
3811 iommu->irte_ops->set_affinity(iommu, ir_data->entry, irte_info->devid,
3812 irte_info->index, cfg->vector,
3816 static int amd_ir_set_affinity(struct irq_data *data,
3817 const struct cpumask *mask, bool force)
3819 struct amd_ir_data *ir_data = data->chip_data;
3820 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
3821 struct irq_cfg *cfg = irqd_cfg(data);
3822 struct irq_data *parent = data->parent_data;
3823 struct amd_iommu *iommu = ir_data->iommu;
3829 ret = parent->chip->irq_set_affinity(parent, mask, force);
3830 if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
3833 amd_ir_update_irte(data, iommu, ir_data, irte_info, cfg);
3835 * After this point, all the interrupts will start arriving
3836 * at the new destination. So, time to cleanup the previous
3837 * vector allocation.
3839 vector_schedule_cleanup(cfg);
3841 return IRQ_SET_MASK_OK_DONE;
3844 static void ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg)
3846 struct amd_ir_data *ir_data = irq_data->chip_data;
3848 *msg = ir_data->msi_entry;
3851 static struct irq_chip amd_ir_chip = {
3853 .irq_ack = apic_ack_irq,
3854 .irq_set_affinity = amd_ir_set_affinity,
3855 .irq_set_vcpu_affinity = amd_ir_set_vcpu_affinity,
3856 .irq_compose_msi_msg = ir_compose_msi_msg,
3859 static const struct msi_parent_ops amdvi_msi_parent_ops = {
3860 .supported_flags = X86_VECTOR_MSI_FLAGS_SUPPORTED |
3861 MSI_FLAG_MULTI_PCI_MSI |
3864 .init_dev_msi_info = msi_parent_init_dev_msi_info,
3867 static const struct msi_parent_ops virt_amdvi_msi_parent_ops = {
3868 .supported_flags = X86_VECTOR_MSI_FLAGS_SUPPORTED |
3869 MSI_FLAG_MULTI_PCI_MSI,
3871 .init_dev_msi_info = msi_parent_init_dev_msi_info,
3874 int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
3876 struct fwnode_handle *fn;
3878 fn = irq_domain_alloc_named_id_fwnode("AMD-IR", iommu->index);
3881 iommu->ir_domain = irq_domain_create_hierarchy(arch_get_ir_parent_domain(), 0, 0,
3882 fn, &amd_ir_domain_ops, iommu);
3883 if (!iommu->ir_domain) {
3884 irq_domain_free_fwnode(fn);
3888 irq_domain_update_bus_token(iommu->ir_domain, DOMAIN_BUS_AMDVI);
3889 iommu->ir_domain->flags |= IRQ_DOMAIN_FLAG_MSI_PARENT |
3890 IRQ_DOMAIN_FLAG_ISOLATED_MSI;
3892 if (amd_iommu_np_cache)
3893 iommu->ir_domain->msi_parent_ops = &virt_amdvi_msi_parent_ops;
3895 iommu->ir_domain->msi_parent_ops = &amdvi_msi_parent_ops;
3900 int amd_iommu_update_ga(int cpu, bool is_run, void *data)
3902 struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
3903 struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
3905 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
3906 !entry || !entry->lo.fields_vapic.guest_mode)
3909 if (!ir_data->iommu)
3913 entry->lo.fields_vapic.destination =
3914 APICID_TO_IRTE_DEST_LO(cpu);
3915 entry->hi.fields.destination =
3916 APICID_TO_IRTE_DEST_HI(cpu);
3918 entry->lo.fields_vapic.is_run = is_run;
3920 return modify_irte_ga(ir_data->iommu, ir_data->irq_2_irte.devid,
3921 ir_data->irq_2_irte.index, entry);
3923 EXPORT_SYMBOL(amd_iommu_update_ga);