2 * Copyright (C) 2007-2010 Advanced Micro Devices, Inc.
3 * Author: Joerg Roedel <jroedel@suse.de>
4 * Leo Duran <leo.duran@amd.com>
6 * This program is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 as published
8 * by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, write to the Free Software
17 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
20 #include <linux/ratelimit.h>
21 #include <linux/pci.h>
22 #include <linux/acpi.h>
23 #include <linux/amba/bus.h>
24 #include <linux/platform_device.h>
25 #include <linux/pci-ats.h>
26 #include <linux/bitmap.h>
27 #include <linux/slab.h>
28 #include <linux/debugfs.h>
29 #include <linux/scatterlist.h>
30 #include <linux/dma-mapping.h>
31 #include <linux/dma-direct.h>
32 #include <linux/iommu-helper.h>
33 #include <linux/iommu.h>
34 #include <linux/delay.h>
35 #include <linux/amd-iommu.h>
36 #include <linux/notifier.h>
37 #include <linux/export.h>
38 #include <linux/irq.h>
39 #include <linux/msi.h>
40 #include <linux/dma-contiguous.h>
41 #include <linux/irqdomain.h>
42 #include <linux/percpu.h>
43 #include <linux/iova.h>
44 #include <asm/irq_remapping.h>
45 #include <asm/io_apic.h>
47 #include <asm/hw_irq.h>
48 #include <asm/msidef.h>
49 #include <asm/proto.h>
50 #include <asm/iommu.h>
54 #include "amd_iommu_proto.h"
55 #include "amd_iommu_types.h"
56 #include "irq_remapping.h"
58 #define AMD_IOMMU_MAPPING_ERROR 0
60 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
62 #define LOOP_TIMEOUT 100000
64 /* IO virtual address start page frame number */
65 #define IOVA_START_PFN (1)
66 #define IOVA_PFN(addr) ((addr) >> PAGE_SHIFT)
68 /* Reserved IOVA ranges */
69 #define MSI_RANGE_START (0xfee00000)
70 #define MSI_RANGE_END (0xfeefffff)
71 #define HT_RANGE_START (0xfd00000000ULL)
72 #define HT_RANGE_END (0xffffffffffULL)
75 * This bitmap is used to advertise the page sizes our hardware support
76 * to the IOMMU core, which will then use this information to split
77 * physically contiguous memory regions it is mapping into page sizes
80 * 512GB Pages are not supported due to a hardware bug
82 #define AMD_IOMMU_PGSIZES ((~0xFFFUL) & ~(2ULL << 38))
84 static DEFINE_SPINLOCK(amd_iommu_devtable_lock);
85 static DEFINE_SPINLOCK(pd_bitmap_lock);
86 static DEFINE_SPINLOCK(iommu_table_lock);
88 /* List of all available dev_data structures */
89 static LLIST_HEAD(dev_data_list);
91 LIST_HEAD(ioapic_map);
93 LIST_HEAD(acpihid_map);
96 * Domain for untranslated devices - only allocated
97 * if iommu=pt passed on kernel cmd line.
99 const struct iommu_ops amd_iommu_ops;
101 static ATOMIC_NOTIFIER_HEAD(ppr_notifier);
102 int amd_iommu_max_glx_val = -1;
104 static const struct dma_map_ops amd_iommu_dma_ops;
107 * general struct to manage commands send to an IOMMU
113 struct kmem_cache *amd_iommu_irq_cache;
115 static void update_domain(struct protection_domain *domain);
116 static int protection_domain_init(struct protection_domain *domain);
117 static void detach_device(struct device *dev);
118 static void iova_domain_flush_tlb(struct iova_domain *iovad);
121 * Data container for a dma_ops specific protection domain
123 struct dma_ops_domain {
124 /* generic protection domain information */
125 struct protection_domain domain;
128 struct iova_domain iovad;
131 static struct iova_domain reserved_iova_ranges;
132 static struct lock_class_key reserved_rbtree_key;
134 /****************************************************************************
138 ****************************************************************************/
140 static inline int match_hid_uid(struct device *dev,
141 struct acpihid_map_entry *entry)
143 const char *hid, *uid;
145 hid = acpi_device_hid(ACPI_COMPANION(dev));
146 uid = acpi_device_uid(ACPI_COMPANION(dev));
152 return strcmp(hid, entry->hid);
155 return strcmp(hid, entry->hid);
157 return (strcmp(hid, entry->hid) || strcmp(uid, entry->uid));
160 static inline u16 get_pci_device_id(struct device *dev)
162 struct pci_dev *pdev = to_pci_dev(dev);
164 return PCI_DEVID(pdev->bus->number, pdev->devfn);
167 static inline int get_acpihid_device_id(struct device *dev,
168 struct acpihid_map_entry **entry)
170 struct acpihid_map_entry *p;
172 list_for_each_entry(p, &acpihid_map, list) {
173 if (!match_hid_uid(dev, p)) {
182 static inline int get_device_id(struct device *dev)
187 devid = get_pci_device_id(dev);
189 devid = get_acpihid_device_id(dev, NULL);
194 static struct protection_domain *to_pdomain(struct iommu_domain *dom)
196 return container_of(dom, struct protection_domain, domain);
199 static struct dma_ops_domain* to_dma_ops_domain(struct protection_domain *domain)
201 BUG_ON(domain->flags != PD_DMA_OPS_MASK);
202 return container_of(domain, struct dma_ops_domain, domain);
205 static struct iommu_dev_data *alloc_dev_data(u16 devid)
207 struct iommu_dev_data *dev_data;
209 dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
213 dev_data->devid = devid;
214 ratelimit_default_init(&dev_data->rs);
216 llist_add(&dev_data->dev_data_list, &dev_data_list);
220 static struct iommu_dev_data *search_dev_data(u16 devid)
222 struct iommu_dev_data *dev_data;
223 struct llist_node *node;
225 if (llist_empty(&dev_data_list))
228 node = dev_data_list.first;
229 llist_for_each_entry(dev_data, node, dev_data_list) {
230 if (dev_data->devid == devid)
237 static int __last_alias(struct pci_dev *pdev, u16 alias, void *data)
239 *(u16 *)data = alias;
243 static u16 get_alias(struct device *dev)
245 struct pci_dev *pdev = to_pci_dev(dev);
246 u16 devid, ivrs_alias, pci_alias;
248 /* The callers make sure that get_device_id() does not fail here */
249 devid = get_device_id(dev);
250 ivrs_alias = amd_iommu_alias_table[devid];
251 pci_for_each_dma_alias(pdev, __last_alias, &pci_alias);
253 if (ivrs_alias == pci_alias)
259 * The IVRS is fairly reliable in telling us about aliases, but it
260 * can't know about every screwy device. If we don't have an IVRS
261 * reported alias, use the PCI reported alias. In that case we may
262 * still need to initialize the rlookup and dev_table entries if the
263 * alias is to a non-existent device.
265 if (ivrs_alias == devid) {
266 if (!amd_iommu_rlookup_table[pci_alias]) {
267 amd_iommu_rlookup_table[pci_alias] =
268 amd_iommu_rlookup_table[devid];
269 memcpy(amd_iommu_dev_table[pci_alias].data,
270 amd_iommu_dev_table[devid].data,
271 sizeof(amd_iommu_dev_table[pci_alias].data));
277 pr_info("AMD-Vi: Using IVRS reported alias %02x:%02x.%d "
278 "for device %s[%04x:%04x], kernel reported alias "
279 "%02x:%02x.%d\n", PCI_BUS_NUM(ivrs_alias), PCI_SLOT(ivrs_alias),
280 PCI_FUNC(ivrs_alias), dev_name(dev), pdev->vendor, pdev->device,
281 PCI_BUS_NUM(pci_alias), PCI_SLOT(pci_alias),
282 PCI_FUNC(pci_alias));
285 * If we don't have a PCI DMA alias and the IVRS alias is on the same
286 * bus, then the IVRS table may know about a quirk that we don't.
288 if (pci_alias == devid &&
289 PCI_BUS_NUM(ivrs_alias) == pdev->bus->number) {
290 pci_add_dma_alias(pdev, ivrs_alias & 0xff);
291 pr_info("AMD-Vi: Added PCI DMA alias %02x.%d for %s\n",
292 PCI_SLOT(ivrs_alias), PCI_FUNC(ivrs_alias),
299 static struct iommu_dev_data *find_dev_data(u16 devid)
301 struct iommu_dev_data *dev_data;
302 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
304 dev_data = search_dev_data(devid);
306 if (dev_data == NULL) {
307 dev_data = alloc_dev_data(devid);
311 if (translation_pre_enabled(iommu))
312 dev_data->defer_attach = true;
318 struct iommu_dev_data *get_dev_data(struct device *dev)
320 return dev->archdata.iommu;
322 EXPORT_SYMBOL(get_dev_data);
325 * Find or create an IOMMU group for a acpihid device.
327 static struct iommu_group *acpihid_device_group(struct device *dev)
329 struct acpihid_map_entry *p, *entry = NULL;
332 devid = get_acpihid_device_id(dev, &entry);
334 return ERR_PTR(devid);
336 list_for_each_entry(p, &acpihid_map, list) {
337 if ((devid == p->devid) && p->group)
338 entry->group = p->group;
342 entry->group = generic_device_group(dev);
344 iommu_group_ref_get(entry->group);
349 static bool pci_iommuv2_capable(struct pci_dev *pdev)
351 static const int caps[] = {
354 PCI_EXT_CAP_ID_PASID,
358 for (i = 0; i < 3; ++i) {
359 pos = pci_find_ext_capability(pdev, caps[i]);
367 static bool pdev_pri_erratum(struct pci_dev *pdev, u32 erratum)
369 struct iommu_dev_data *dev_data;
371 dev_data = get_dev_data(&pdev->dev);
373 return dev_data->errata & (1 << erratum) ? true : false;
377 * This function checks if the driver got a valid device from the caller to
378 * avoid dereferencing invalid pointers.
380 static bool check_device(struct device *dev)
384 if (!dev || !dev->dma_mask)
387 devid = get_device_id(dev);
391 /* Out of our scope? */
392 if (devid > amd_iommu_last_bdf)
395 if (amd_iommu_rlookup_table[devid] == NULL)
401 static void init_iommu_group(struct device *dev)
403 struct iommu_group *group;
405 group = iommu_group_get_for_dev(dev);
409 iommu_group_put(group);
412 static int iommu_init_device(struct device *dev)
414 struct iommu_dev_data *dev_data;
415 struct amd_iommu *iommu;
418 if (dev->archdata.iommu)
421 devid = get_device_id(dev);
425 iommu = amd_iommu_rlookup_table[devid];
427 dev_data = find_dev_data(devid);
431 dev_data->alias = get_alias(dev);
433 if (dev_is_pci(dev) && pci_iommuv2_capable(to_pci_dev(dev))) {
434 struct amd_iommu *iommu;
436 iommu = amd_iommu_rlookup_table[dev_data->devid];
437 dev_data->iommu_v2 = iommu->is_iommu_v2;
440 dev->archdata.iommu = dev_data;
442 iommu_device_link(&iommu->iommu, dev);
447 static void iommu_ignore_device(struct device *dev)
452 devid = get_device_id(dev);
456 alias = get_alias(dev);
458 memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
459 memset(&amd_iommu_dev_table[alias], 0, sizeof(struct dev_table_entry));
461 amd_iommu_rlookup_table[devid] = NULL;
462 amd_iommu_rlookup_table[alias] = NULL;
465 static void iommu_uninit_device(struct device *dev)
467 struct iommu_dev_data *dev_data;
468 struct amd_iommu *iommu;
471 devid = get_device_id(dev);
475 iommu = amd_iommu_rlookup_table[devid];
477 dev_data = search_dev_data(devid);
481 if (dev_data->domain)
484 iommu_device_unlink(&iommu->iommu, dev);
486 iommu_group_remove_device(dev);
492 * We keep dev_data around for unplugged devices and reuse it when the
493 * device is re-plugged - not doing so would introduce a ton of races.
497 /****************************************************************************
499 * Interrupt handling functions
501 ****************************************************************************/
503 static void dump_dte_entry(u16 devid)
507 for (i = 0; i < 4; ++i)
508 pr_err("AMD-Vi: DTE[%d]: %016llx\n", i,
509 amd_iommu_dev_table[devid].data[i]);
512 static void dump_command(unsigned long phys_addr)
514 struct iommu_cmd *cmd = iommu_phys_to_virt(phys_addr);
517 for (i = 0; i < 4; ++i)
518 pr_err("AMD-Vi: CMD[%d]: %08x\n", i, cmd->data[i]);
521 static void amd_iommu_report_page_fault(u16 devid, u16 domain_id,
522 u64 address, int flags)
524 struct iommu_dev_data *dev_data = NULL;
525 struct pci_dev *pdev;
527 pdev = pci_get_domain_bus_and_slot(0, PCI_BUS_NUM(devid),
530 dev_data = get_dev_data(&pdev->dev);
532 if (dev_data && __ratelimit(&dev_data->rs)) {
533 dev_err(&pdev->dev, "AMD-Vi: Event logged [IO_PAGE_FAULT domain=0x%04x address=0x%016llx flags=0x%04x]\n",
534 domain_id, address, flags);
535 } else if (printk_ratelimit()) {
536 pr_err("AMD-Vi: Event logged [IO_PAGE_FAULT device=%02x:%02x.%x domain=0x%04x address=0x%016llx flags=0x%04x]\n",
537 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
538 domain_id, address, flags);
545 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
547 struct device *dev = iommu->iommu.dev;
548 int type, devid, domid, flags;
549 volatile u32 *event = __evt;
554 type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
555 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
556 domid = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
557 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
558 address = (u64)(((u64)event[3]) << 32) | event[2];
561 /* Did we hit the erratum? */
562 if (++count == LOOP_TIMEOUT) {
563 pr_err("AMD-Vi: No event written to event log\n");
570 if (type == EVENT_TYPE_IO_FAULT) {
571 amd_iommu_report_page_fault(devid, domid, address, flags);
574 dev_err(dev, "AMD-Vi: Event logged [");
578 case EVENT_TYPE_ILL_DEV:
579 dev_err(dev, "ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
580 "address=0x%016llx flags=0x%04x]\n",
581 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
583 dump_dte_entry(devid);
585 case EVENT_TYPE_DEV_TAB_ERR:
586 dev_err(dev, "DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
587 "address=0x%016llx flags=0x%04x]\n",
588 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
591 case EVENT_TYPE_PAGE_TAB_ERR:
592 dev_err(dev, "PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
593 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
594 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
595 domid, address, flags);
597 case EVENT_TYPE_ILL_CMD:
598 dev_err(dev, "ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
599 dump_command(address);
601 case EVENT_TYPE_CMD_HARD_ERR:
602 dev_err(dev, "COMMAND_HARDWARE_ERROR address=0x%016llx "
603 "flags=0x%04x]\n", address, flags);
605 case EVENT_TYPE_IOTLB_INV_TO:
606 dev_err(dev, "IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
607 "address=0x%016llx]\n",
608 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
611 case EVENT_TYPE_INV_DEV_REQ:
612 dev_err(dev, "INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
613 "address=0x%016llx flags=0x%04x]\n",
614 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
618 dev_err(dev, KERN_ERR "UNKNOWN event[0]=0x%08x event[1]=0x%08x "
619 "event[2]=0x%08x event[3]=0x%08x\n",
620 event[0], event[1], event[2], event[3]);
623 memset(__evt, 0, 4 * sizeof(u32));
626 static void iommu_poll_events(struct amd_iommu *iommu)
630 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
631 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
633 while (head != tail) {
634 iommu_print_event(iommu, iommu->evt_buf + head);
635 head = (head + EVENT_ENTRY_SIZE) % EVT_BUFFER_SIZE;
638 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
641 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
643 struct amd_iommu_fault fault;
645 if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
646 pr_err_ratelimited("AMD-Vi: Unknown PPR request received\n");
650 fault.address = raw[1];
651 fault.pasid = PPR_PASID(raw[0]);
652 fault.device_id = PPR_DEVID(raw[0]);
653 fault.tag = PPR_TAG(raw[0]);
654 fault.flags = PPR_FLAGS(raw[0]);
656 atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
659 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
663 if (iommu->ppr_log == NULL)
666 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
667 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
669 while (head != tail) {
674 raw = (u64 *)(iommu->ppr_log + head);
677 * Hardware bug: Interrupt may arrive before the entry is
678 * written to memory. If this happens we need to wait for the
681 for (i = 0; i < LOOP_TIMEOUT; ++i) {
682 if (PPR_REQ_TYPE(raw[0]) != 0)
687 /* Avoid memcpy function-call overhead */
692 * To detect the hardware bug we need to clear the entry
695 raw[0] = raw[1] = 0UL;
697 /* Update head pointer of hardware ring-buffer */
698 head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
699 writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
701 /* Handle PPR entry */
702 iommu_handle_ppr_entry(iommu, entry);
704 /* Refresh ring-buffer information */
705 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
706 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
710 #ifdef CONFIG_IRQ_REMAP
711 static int (*iommu_ga_log_notifier)(u32);
713 int amd_iommu_register_ga_log_notifier(int (*notifier)(u32))
715 iommu_ga_log_notifier = notifier;
719 EXPORT_SYMBOL(amd_iommu_register_ga_log_notifier);
721 static void iommu_poll_ga_log(struct amd_iommu *iommu)
723 u32 head, tail, cnt = 0;
725 if (iommu->ga_log == NULL)
728 head = readl(iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
729 tail = readl(iommu->mmio_base + MMIO_GA_TAIL_OFFSET);
731 while (head != tail) {
735 raw = (u64 *)(iommu->ga_log + head);
738 /* Avoid memcpy function-call overhead */
741 /* Update head pointer of hardware ring-buffer */
742 head = (head + GA_ENTRY_SIZE) % GA_LOG_SIZE;
743 writel(head, iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
745 /* Handle GA entry */
746 switch (GA_REQ_TYPE(log_entry)) {
748 if (!iommu_ga_log_notifier)
751 pr_debug("AMD-Vi: %s: devid=%#x, ga_tag=%#x\n",
752 __func__, GA_DEVID(log_entry),
755 if (iommu_ga_log_notifier(GA_TAG(log_entry)) != 0)
756 pr_err("AMD-Vi: GA log notifier failed.\n");
763 #endif /* CONFIG_IRQ_REMAP */
765 #define AMD_IOMMU_INT_MASK \
766 (MMIO_STATUS_EVT_INT_MASK | \
767 MMIO_STATUS_PPR_INT_MASK | \
768 MMIO_STATUS_GALOG_INT_MASK)
770 irqreturn_t amd_iommu_int_thread(int irq, void *data)
772 struct amd_iommu *iommu = (struct amd_iommu *) data;
773 u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
775 while (status & AMD_IOMMU_INT_MASK) {
776 /* Enable EVT and PPR and GA interrupts again */
777 writel(AMD_IOMMU_INT_MASK,
778 iommu->mmio_base + MMIO_STATUS_OFFSET);
780 if (status & MMIO_STATUS_EVT_INT_MASK) {
781 pr_devel("AMD-Vi: Processing IOMMU Event Log\n");
782 iommu_poll_events(iommu);
785 if (status & MMIO_STATUS_PPR_INT_MASK) {
786 pr_devel("AMD-Vi: Processing IOMMU PPR Log\n");
787 iommu_poll_ppr_log(iommu);
790 #ifdef CONFIG_IRQ_REMAP
791 if (status & MMIO_STATUS_GALOG_INT_MASK) {
792 pr_devel("AMD-Vi: Processing IOMMU GA Log\n");
793 iommu_poll_ga_log(iommu);
798 * Hardware bug: ERBT1312
799 * When re-enabling interrupt (by writing 1
800 * to clear the bit), the hardware might also try to set
801 * the interrupt bit in the event status register.
802 * In this scenario, the bit will be set, and disable
803 * subsequent interrupts.
805 * Workaround: The IOMMU driver should read back the
806 * status register and check if the interrupt bits are cleared.
807 * If not, driver will need to go through the interrupt handler
808 * again and re-clear the bits
810 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
815 irqreturn_t amd_iommu_int_handler(int irq, void *data)
817 return IRQ_WAKE_THREAD;
820 /****************************************************************************
822 * IOMMU command queuing functions
824 ****************************************************************************/
826 static int wait_on_sem(volatile u64 *sem)
830 while (*sem == 0 && i < LOOP_TIMEOUT) {
835 if (i == LOOP_TIMEOUT) {
836 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
843 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
844 struct iommu_cmd *cmd)
848 target = iommu->cmd_buf + iommu->cmd_buf_tail;
850 iommu->cmd_buf_tail += sizeof(*cmd);
851 iommu->cmd_buf_tail %= CMD_BUFFER_SIZE;
853 /* Copy command to buffer */
854 memcpy(target, cmd, sizeof(*cmd));
856 /* Tell the IOMMU about it */
857 writel(iommu->cmd_buf_tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
860 static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
862 u64 paddr = iommu_virt_to_phys((void *)address);
864 WARN_ON(address & 0x7ULL);
866 memset(cmd, 0, sizeof(*cmd));
867 cmd->data[0] = lower_32_bits(paddr) | CMD_COMPL_WAIT_STORE_MASK;
868 cmd->data[1] = upper_32_bits(paddr);
870 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
873 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
875 memset(cmd, 0, sizeof(*cmd));
876 cmd->data[0] = devid;
877 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
880 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
881 size_t size, u16 domid, int pde)
886 pages = iommu_num_pages(address, size, PAGE_SIZE);
891 * If we have to flush more than one page, flush all
892 * TLB entries for this domain
894 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
898 address &= PAGE_MASK;
900 memset(cmd, 0, sizeof(*cmd));
901 cmd->data[1] |= domid;
902 cmd->data[2] = lower_32_bits(address);
903 cmd->data[3] = upper_32_bits(address);
904 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
905 if (s) /* size bit - we flush more than one 4kb page */
906 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
907 if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
908 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
911 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
912 u64 address, size_t size)
917 pages = iommu_num_pages(address, size, PAGE_SIZE);
922 * If we have to flush more than one page, flush all
923 * TLB entries for this domain
925 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
929 address &= PAGE_MASK;
931 memset(cmd, 0, sizeof(*cmd));
932 cmd->data[0] = devid;
933 cmd->data[0] |= (qdep & 0xff) << 24;
934 cmd->data[1] = devid;
935 cmd->data[2] = lower_32_bits(address);
936 cmd->data[3] = upper_32_bits(address);
937 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
939 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
942 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, int pasid,
943 u64 address, bool size)
945 memset(cmd, 0, sizeof(*cmd));
947 address &= ~(0xfffULL);
949 cmd->data[0] = pasid;
950 cmd->data[1] = domid;
951 cmd->data[2] = lower_32_bits(address);
952 cmd->data[3] = upper_32_bits(address);
953 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
954 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
956 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
957 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
960 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, int pasid,
961 int qdep, u64 address, bool size)
963 memset(cmd, 0, sizeof(*cmd));
965 address &= ~(0xfffULL);
967 cmd->data[0] = devid;
968 cmd->data[0] |= ((pasid >> 8) & 0xff) << 16;
969 cmd->data[0] |= (qdep & 0xff) << 24;
970 cmd->data[1] = devid;
971 cmd->data[1] |= (pasid & 0xff) << 16;
972 cmd->data[2] = lower_32_bits(address);
973 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
974 cmd->data[3] = upper_32_bits(address);
976 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
977 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
980 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, int pasid,
981 int status, int tag, bool gn)
983 memset(cmd, 0, sizeof(*cmd));
985 cmd->data[0] = devid;
987 cmd->data[1] = pasid;
988 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
990 cmd->data[3] = tag & 0x1ff;
991 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
993 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
996 static void build_inv_all(struct iommu_cmd *cmd)
998 memset(cmd, 0, sizeof(*cmd));
999 CMD_SET_TYPE(cmd, CMD_INV_ALL);
1002 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
1004 memset(cmd, 0, sizeof(*cmd));
1005 cmd->data[0] = devid;
1006 CMD_SET_TYPE(cmd, CMD_INV_IRT);
1010 * Writes the command to the IOMMUs command buffer and informs the
1011 * hardware about the new command.
1013 static int __iommu_queue_command_sync(struct amd_iommu *iommu,
1014 struct iommu_cmd *cmd,
1017 unsigned int count = 0;
1018 u32 left, next_tail;
1020 next_tail = (iommu->cmd_buf_tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
1022 left = (iommu->cmd_buf_head - next_tail) % CMD_BUFFER_SIZE;
1025 /* Skip udelay() the first time around */
1027 if (count == LOOP_TIMEOUT) {
1028 pr_err("AMD-Vi: Command buffer timeout\n");
1035 /* Update head and recheck remaining space */
1036 iommu->cmd_buf_head = readl(iommu->mmio_base +
1037 MMIO_CMD_HEAD_OFFSET);
1042 copy_cmd_to_buffer(iommu, cmd);
1044 /* Do we need to make sure all commands are processed? */
1045 iommu->need_sync = sync;
1050 static int iommu_queue_command_sync(struct amd_iommu *iommu,
1051 struct iommu_cmd *cmd,
1054 unsigned long flags;
1057 raw_spin_lock_irqsave(&iommu->lock, flags);
1058 ret = __iommu_queue_command_sync(iommu, cmd, sync);
1059 raw_spin_unlock_irqrestore(&iommu->lock, flags);
1064 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
1066 return iommu_queue_command_sync(iommu, cmd, true);
1070 * This function queues a completion wait command into the command
1071 * buffer of an IOMMU
1073 static int iommu_completion_wait(struct amd_iommu *iommu)
1075 struct iommu_cmd cmd;
1076 unsigned long flags;
1079 if (!iommu->need_sync)
1083 build_completion_wait(&cmd, (u64)&iommu->cmd_sem);
1085 raw_spin_lock_irqsave(&iommu->lock, flags);
1089 ret = __iommu_queue_command_sync(iommu, &cmd, false);
1093 ret = wait_on_sem(&iommu->cmd_sem);
1096 raw_spin_unlock_irqrestore(&iommu->lock, flags);
1101 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
1103 struct iommu_cmd cmd;
1105 build_inv_dte(&cmd, devid);
1107 return iommu_queue_command(iommu, &cmd);
1110 static void amd_iommu_flush_dte_all(struct amd_iommu *iommu)
1114 for (devid = 0; devid <= 0xffff; ++devid)
1115 iommu_flush_dte(iommu, devid);
1117 iommu_completion_wait(iommu);
1121 * This function uses heavy locking and may disable irqs for some time. But
1122 * this is no issue because it is only called during resume.
1124 static void amd_iommu_flush_tlb_all(struct amd_iommu *iommu)
1128 for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
1129 struct iommu_cmd cmd;
1130 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1132 iommu_queue_command(iommu, &cmd);
1135 iommu_completion_wait(iommu);
1138 static void amd_iommu_flush_all(struct amd_iommu *iommu)
1140 struct iommu_cmd cmd;
1142 build_inv_all(&cmd);
1144 iommu_queue_command(iommu, &cmd);
1145 iommu_completion_wait(iommu);
1148 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1150 struct iommu_cmd cmd;
1152 build_inv_irt(&cmd, devid);
1154 iommu_queue_command(iommu, &cmd);
1157 static void amd_iommu_flush_irt_all(struct amd_iommu *iommu)
1161 for (devid = 0; devid <= MAX_DEV_TABLE_ENTRIES; devid++)
1162 iommu_flush_irt(iommu, devid);
1164 iommu_completion_wait(iommu);
1167 void iommu_flush_all_caches(struct amd_iommu *iommu)
1169 if (iommu_feature(iommu, FEATURE_IA)) {
1170 amd_iommu_flush_all(iommu);
1172 amd_iommu_flush_dte_all(iommu);
1173 amd_iommu_flush_irt_all(iommu);
1174 amd_iommu_flush_tlb_all(iommu);
1179 * Command send function for flushing on-device TLB
1181 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1182 u64 address, size_t size)
1184 struct amd_iommu *iommu;
1185 struct iommu_cmd cmd;
1188 qdep = dev_data->ats.qdep;
1189 iommu = amd_iommu_rlookup_table[dev_data->devid];
1191 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1193 return iommu_queue_command(iommu, &cmd);
1197 * Command send function for invalidating a device table entry
1199 static int device_flush_dte(struct iommu_dev_data *dev_data)
1201 struct amd_iommu *iommu;
1205 iommu = amd_iommu_rlookup_table[dev_data->devid];
1206 alias = dev_data->alias;
1208 ret = iommu_flush_dte(iommu, dev_data->devid);
1209 if (!ret && alias != dev_data->devid)
1210 ret = iommu_flush_dte(iommu, alias);
1214 if (dev_data->ats.enabled)
1215 ret = device_flush_iotlb(dev_data, 0, ~0UL);
1221 * TLB invalidation function which is called from the mapping functions.
1222 * It invalidates a single PTE if the range to flush is within a single
1223 * page. Otherwise it flushes the whole TLB of the IOMMU.
1225 static void __domain_flush_pages(struct protection_domain *domain,
1226 u64 address, size_t size, int pde)
1228 struct iommu_dev_data *dev_data;
1229 struct iommu_cmd cmd;
1232 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1234 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
1235 if (!domain->dev_iommu[i])
1239 * Devices of this domain are behind this IOMMU
1240 * We need a TLB flush
1242 ret |= iommu_queue_command(amd_iommus[i], &cmd);
1245 list_for_each_entry(dev_data, &domain->dev_list, list) {
1247 if (!dev_data->ats.enabled)
1250 ret |= device_flush_iotlb(dev_data, address, size);
1256 static void domain_flush_pages(struct protection_domain *domain,
1257 u64 address, size_t size)
1259 __domain_flush_pages(domain, address, size, 0);
1262 /* Flush the whole IO/TLB for a given protection domain */
1263 static void domain_flush_tlb(struct protection_domain *domain)
1265 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
1268 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1269 static void domain_flush_tlb_pde(struct protection_domain *domain)
1271 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1274 static void domain_flush_complete(struct protection_domain *domain)
1278 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
1279 if (domain && !domain->dev_iommu[i])
1283 * Devices of this domain are behind this IOMMU
1284 * We need to wait for completion of all commands.
1286 iommu_completion_wait(amd_iommus[i]);
1292 * This function flushes the DTEs for all devices in domain
1294 static void domain_flush_devices(struct protection_domain *domain)
1296 struct iommu_dev_data *dev_data;
1298 list_for_each_entry(dev_data, &domain->dev_list, list)
1299 device_flush_dte(dev_data);
1302 /****************************************************************************
1304 * The functions below are used the create the page table mappings for
1305 * unity mapped regions.
1307 ****************************************************************************/
1310 * This function is used to add another level to an IO page table. Adding
1311 * another level increases the size of the address space by 9 bits to a size up
1314 static bool increase_address_space(struct protection_domain *domain,
1319 if (domain->mode == PAGE_MODE_6_LEVEL)
1320 /* address space already 64 bit large */
1323 pte = (void *)get_zeroed_page(gfp);
1327 *pte = PM_LEVEL_PDE(domain->mode,
1328 iommu_virt_to_phys(domain->pt_root));
1329 domain->pt_root = pte;
1331 domain->updated = true;
1336 static u64 *alloc_pte(struct protection_domain *domain,
1337 unsigned long address,
1338 unsigned long page_size,
1345 BUG_ON(!is_power_of_2(page_size));
1347 while (address > PM_LEVEL_SIZE(domain->mode))
1348 increase_address_space(domain, gfp);
1350 level = domain->mode - 1;
1351 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1352 address = PAGE_SIZE_ALIGN(address, page_size);
1353 end_lvl = PAGE_SIZE_LEVEL(page_size);
1355 while (level > end_lvl) {
1360 if (!IOMMU_PTE_PRESENT(__pte)) {
1361 page = (u64 *)get_zeroed_page(gfp);
1365 __npte = PM_LEVEL_PDE(level, iommu_virt_to_phys(page));
1367 /* pte could have been changed somewhere. */
1368 if (cmpxchg64(pte, __pte, __npte) != __pte) {
1369 free_page((unsigned long)page);
1374 /* No level skipping support yet */
1375 if (PM_PTE_LEVEL(*pte) != level)
1380 pte = IOMMU_PTE_PAGE(*pte);
1382 if (pte_page && level == end_lvl)
1385 pte = &pte[PM_LEVEL_INDEX(level, address)];
1392 * This function checks if there is a PTE for a given dma address. If
1393 * there is one, it returns the pointer to it.
1395 static u64 *fetch_pte(struct protection_domain *domain,
1396 unsigned long address,
1397 unsigned long *page_size)
1402 if (address > PM_LEVEL_SIZE(domain->mode))
1405 level = domain->mode - 1;
1406 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1407 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1412 if (!IOMMU_PTE_PRESENT(*pte))
1416 if (PM_PTE_LEVEL(*pte) == 7 ||
1417 PM_PTE_LEVEL(*pte) == 0)
1420 /* No level skipping support yet */
1421 if (PM_PTE_LEVEL(*pte) != level)
1426 /* Walk to the next level */
1427 pte = IOMMU_PTE_PAGE(*pte);
1428 pte = &pte[PM_LEVEL_INDEX(level, address)];
1429 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1432 if (PM_PTE_LEVEL(*pte) == 0x07) {
1433 unsigned long pte_mask;
1436 * If we have a series of large PTEs, make
1437 * sure to return a pointer to the first one.
1439 *page_size = pte_mask = PTE_PAGE_SIZE(*pte);
1440 pte_mask = ~((PAGE_SIZE_PTE_COUNT(pte_mask) << 3) - 1);
1441 pte = (u64 *)(((unsigned long)pte) & pte_mask);
1448 * Generic mapping functions. It maps a physical address into a DMA
1449 * address space. It allocates the page table pages if necessary.
1450 * In the future it can be extended to a generic mapping function
1451 * supporting all features of AMD IOMMU page tables like level skipping
1452 * and full 64 bit address spaces.
1454 static int iommu_map_page(struct protection_domain *dom,
1455 unsigned long bus_addr,
1456 unsigned long phys_addr,
1457 unsigned long page_size,
1464 BUG_ON(!IS_ALIGNED(bus_addr, page_size));
1465 BUG_ON(!IS_ALIGNED(phys_addr, page_size));
1467 if (!(prot & IOMMU_PROT_MASK))
1470 count = PAGE_SIZE_PTE_COUNT(page_size);
1471 pte = alloc_pte(dom, bus_addr, page_size, NULL, gfp);
1476 for (i = 0; i < count; ++i)
1477 if (IOMMU_PTE_PRESENT(pte[i]))
1481 __pte = PAGE_SIZE_PTE(__sme_set(phys_addr), page_size);
1482 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_PR | IOMMU_PTE_FC;
1484 __pte = __sme_set(phys_addr) | IOMMU_PTE_PR | IOMMU_PTE_FC;
1486 if (prot & IOMMU_PROT_IR)
1487 __pte |= IOMMU_PTE_IR;
1488 if (prot & IOMMU_PROT_IW)
1489 __pte |= IOMMU_PTE_IW;
1491 for (i = 0; i < count; ++i)
1499 static unsigned long iommu_unmap_page(struct protection_domain *dom,
1500 unsigned long bus_addr,
1501 unsigned long page_size)
1503 unsigned long long unmapped;
1504 unsigned long unmap_size;
1507 BUG_ON(!is_power_of_2(page_size));
1511 while (unmapped < page_size) {
1513 pte = fetch_pte(dom, bus_addr, &unmap_size);
1518 count = PAGE_SIZE_PTE_COUNT(unmap_size);
1519 for (i = 0; i < count; i++)
1523 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1524 unmapped += unmap_size;
1527 BUG_ON(unmapped && !is_power_of_2(unmapped));
1532 /****************************************************************************
1534 * The next functions belong to the address allocator for the dma_ops
1535 * interface functions.
1537 ****************************************************************************/
1540 static unsigned long dma_ops_alloc_iova(struct device *dev,
1541 struct dma_ops_domain *dma_dom,
1542 unsigned int pages, u64 dma_mask)
1544 unsigned long pfn = 0;
1546 pages = __roundup_pow_of_two(pages);
1548 if (dma_mask > DMA_BIT_MASK(32))
1549 pfn = alloc_iova_fast(&dma_dom->iovad, pages,
1550 IOVA_PFN(DMA_BIT_MASK(32)), false);
1553 pfn = alloc_iova_fast(&dma_dom->iovad, pages,
1554 IOVA_PFN(dma_mask), true);
1556 return (pfn << PAGE_SHIFT);
1559 static void dma_ops_free_iova(struct dma_ops_domain *dma_dom,
1560 unsigned long address,
1563 pages = __roundup_pow_of_two(pages);
1564 address >>= PAGE_SHIFT;
1566 free_iova_fast(&dma_dom->iovad, address, pages);
1569 /****************************************************************************
1571 * The next functions belong to the domain allocation. A domain is
1572 * allocated for every IOMMU as the default domain. If device isolation
1573 * is enabled, every device get its own domain. The most important thing
1574 * about domains is the page table mapping the DMA address space they
1577 ****************************************************************************/
1580 * This function adds a protection domain to the global protection domain list
1582 static void add_domain_to_list(struct protection_domain *domain)
1584 unsigned long flags;
1586 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1587 list_add(&domain->list, &amd_iommu_pd_list);
1588 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1592 * This function removes a protection domain to the global
1593 * protection domain list
1595 static void del_domain_from_list(struct protection_domain *domain)
1597 unsigned long flags;
1599 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1600 list_del(&domain->list);
1601 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1604 static u16 domain_id_alloc(void)
1608 spin_lock(&pd_bitmap_lock);
1609 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1611 if (id > 0 && id < MAX_DOMAIN_ID)
1612 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1615 spin_unlock(&pd_bitmap_lock);
1620 static void domain_id_free(int id)
1622 spin_lock(&pd_bitmap_lock);
1623 if (id > 0 && id < MAX_DOMAIN_ID)
1624 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1625 spin_unlock(&pd_bitmap_lock);
1628 #define DEFINE_FREE_PT_FN(LVL, FN) \
1629 static void free_pt_##LVL (unsigned long __pt) \
1637 for (i = 0; i < 512; ++i) { \
1638 /* PTE present? */ \
1639 if (!IOMMU_PTE_PRESENT(pt[i])) \
1643 if (PM_PTE_LEVEL(pt[i]) == 0 || \
1644 PM_PTE_LEVEL(pt[i]) == 7) \
1647 p = (unsigned long)IOMMU_PTE_PAGE(pt[i]); \
1650 free_page((unsigned long)pt); \
1653 DEFINE_FREE_PT_FN(l2, free_page)
1654 DEFINE_FREE_PT_FN(l3, free_pt_l2)
1655 DEFINE_FREE_PT_FN(l4, free_pt_l3)
1656 DEFINE_FREE_PT_FN(l5, free_pt_l4)
1657 DEFINE_FREE_PT_FN(l6, free_pt_l5)
1659 static void free_pagetable(struct protection_domain *domain)
1661 unsigned long root = (unsigned long)domain->pt_root;
1663 switch (domain->mode) {
1664 case PAGE_MODE_NONE:
1666 case PAGE_MODE_1_LEVEL:
1669 case PAGE_MODE_2_LEVEL:
1672 case PAGE_MODE_3_LEVEL:
1675 case PAGE_MODE_4_LEVEL:
1678 case PAGE_MODE_5_LEVEL:
1681 case PAGE_MODE_6_LEVEL:
1689 static void free_gcr3_tbl_level1(u64 *tbl)
1694 for (i = 0; i < 512; ++i) {
1695 if (!(tbl[i] & GCR3_VALID))
1698 ptr = iommu_phys_to_virt(tbl[i] & PAGE_MASK);
1700 free_page((unsigned long)ptr);
1704 static void free_gcr3_tbl_level2(u64 *tbl)
1709 for (i = 0; i < 512; ++i) {
1710 if (!(tbl[i] & GCR3_VALID))
1713 ptr = iommu_phys_to_virt(tbl[i] & PAGE_MASK);
1715 free_gcr3_tbl_level1(ptr);
1719 static void free_gcr3_table(struct protection_domain *domain)
1721 if (domain->glx == 2)
1722 free_gcr3_tbl_level2(domain->gcr3_tbl);
1723 else if (domain->glx == 1)
1724 free_gcr3_tbl_level1(domain->gcr3_tbl);
1726 BUG_ON(domain->glx != 0);
1728 free_page((unsigned long)domain->gcr3_tbl);
1731 static void dma_ops_domain_flush_tlb(struct dma_ops_domain *dom)
1733 domain_flush_tlb(&dom->domain);
1734 domain_flush_complete(&dom->domain);
1737 static void iova_domain_flush_tlb(struct iova_domain *iovad)
1739 struct dma_ops_domain *dom;
1741 dom = container_of(iovad, struct dma_ops_domain, iovad);
1743 dma_ops_domain_flush_tlb(dom);
1747 * Free a domain, only used if something went wrong in the
1748 * allocation path and we need to free an already allocated page table
1750 static void dma_ops_domain_free(struct dma_ops_domain *dom)
1755 del_domain_from_list(&dom->domain);
1757 put_iova_domain(&dom->iovad);
1759 free_pagetable(&dom->domain);
1762 domain_id_free(dom->domain.id);
1768 * Allocates a new protection domain usable for the dma_ops functions.
1769 * It also initializes the page table and the address allocator data
1770 * structures required for the dma_ops interface
1772 static struct dma_ops_domain *dma_ops_domain_alloc(void)
1774 struct dma_ops_domain *dma_dom;
1776 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
1780 if (protection_domain_init(&dma_dom->domain))
1783 dma_dom->domain.mode = PAGE_MODE_3_LEVEL;
1784 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
1785 dma_dom->domain.flags = PD_DMA_OPS_MASK;
1786 if (!dma_dom->domain.pt_root)
1789 init_iova_domain(&dma_dom->iovad, PAGE_SIZE, IOVA_START_PFN);
1791 if (init_iova_flush_queue(&dma_dom->iovad, iova_domain_flush_tlb, NULL))
1794 /* Initialize reserved ranges */
1795 copy_reserved_iova(&reserved_iova_ranges, &dma_dom->iovad);
1797 add_domain_to_list(&dma_dom->domain);
1802 dma_ops_domain_free(dma_dom);
1808 * little helper function to check whether a given protection domain is a
1811 static bool dma_ops_domain(struct protection_domain *domain)
1813 return domain->flags & PD_DMA_OPS_MASK;
1816 static void set_dte_entry(u16 devid, struct protection_domain *domain,
1822 if (domain->mode != PAGE_MODE_NONE)
1823 pte_root = iommu_virt_to_phys(domain->pt_root);
1825 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
1826 << DEV_ENTRY_MODE_SHIFT;
1827 pte_root |= DTE_FLAG_IR | DTE_FLAG_IW | DTE_FLAG_V | DTE_FLAG_TV;
1829 flags = amd_iommu_dev_table[devid].data[1];
1832 flags |= DTE_FLAG_IOTLB;
1835 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
1837 if (iommu_feature(iommu, FEATURE_EPHSUP))
1838 pte_root |= 1ULL << DEV_ENTRY_PPR;
1841 if (domain->flags & PD_IOMMUV2_MASK) {
1842 u64 gcr3 = iommu_virt_to_phys(domain->gcr3_tbl);
1843 u64 glx = domain->glx;
1846 pte_root |= DTE_FLAG_GV;
1847 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
1849 /* First mask out possible old values for GCR3 table */
1850 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
1853 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
1856 /* Encode GCR3 table into DTE */
1857 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
1860 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
1863 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
1867 flags &= ~DEV_DOMID_MASK;
1868 flags |= domain->id;
1870 amd_iommu_dev_table[devid].data[1] = flags;
1871 amd_iommu_dev_table[devid].data[0] = pte_root;
1874 static void clear_dte_entry(u16 devid)
1876 /* remove entry from the device table seen by the hardware */
1877 amd_iommu_dev_table[devid].data[0] = DTE_FLAG_V | DTE_FLAG_TV;
1878 amd_iommu_dev_table[devid].data[1] &= DTE_FLAG_MASK;
1880 amd_iommu_apply_erratum_63(devid);
1883 static void do_attach(struct iommu_dev_data *dev_data,
1884 struct protection_domain *domain)
1886 struct amd_iommu *iommu;
1890 iommu = amd_iommu_rlookup_table[dev_data->devid];
1891 alias = dev_data->alias;
1892 ats = dev_data->ats.enabled;
1894 /* Update data structures */
1895 dev_data->domain = domain;
1896 list_add(&dev_data->list, &domain->dev_list);
1898 /* Do reference counting */
1899 domain->dev_iommu[iommu->index] += 1;
1900 domain->dev_cnt += 1;
1902 /* Update device table */
1903 set_dte_entry(dev_data->devid, domain, ats, dev_data->iommu_v2);
1904 if (alias != dev_data->devid)
1905 set_dte_entry(alias, domain, ats, dev_data->iommu_v2);
1907 device_flush_dte(dev_data);
1910 static void do_detach(struct iommu_dev_data *dev_data)
1912 struct amd_iommu *iommu;
1916 * First check if the device is still attached. It might already
1917 * be detached from its domain because the generic
1918 * iommu_detach_group code detached it and we try again here in
1919 * our alias handling.
1921 if (!dev_data->domain)
1924 iommu = amd_iommu_rlookup_table[dev_data->devid];
1925 alias = dev_data->alias;
1927 /* decrease reference counters */
1928 dev_data->domain->dev_iommu[iommu->index] -= 1;
1929 dev_data->domain->dev_cnt -= 1;
1931 /* Update data structures */
1932 dev_data->domain = NULL;
1933 list_del(&dev_data->list);
1934 clear_dte_entry(dev_data->devid);
1935 if (alias != dev_data->devid)
1936 clear_dte_entry(alias);
1938 /* Flush the DTE entry */
1939 device_flush_dte(dev_data);
1943 * If a device is not yet associated with a domain, this function does
1944 * assigns it visible for the hardware
1946 static int __attach_device(struct iommu_dev_data *dev_data,
1947 struct protection_domain *domain)
1952 * Must be called with IRQs disabled. Warn here to detect early
1955 WARN_ON(!irqs_disabled());
1958 spin_lock(&domain->lock);
1961 if (dev_data->domain != NULL)
1964 /* Attach alias group root */
1965 do_attach(dev_data, domain);
1972 spin_unlock(&domain->lock);
1978 static void pdev_iommuv2_disable(struct pci_dev *pdev)
1980 pci_disable_ats(pdev);
1981 pci_disable_pri(pdev);
1982 pci_disable_pasid(pdev);
1985 /* FIXME: Change generic reset-function to do the same */
1986 static int pri_reset_while_enabled(struct pci_dev *pdev)
1991 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
1995 pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
1996 control |= PCI_PRI_CTRL_RESET;
1997 pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
2002 static int pdev_iommuv2_enable(struct pci_dev *pdev)
2007 /* FIXME: Hardcode number of outstanding requests for now */
2009 if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
2011 reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
2013 /* Only allow access to user-accessible pages */
2014 ret = pci_enable_pasid(pdev, 0);
2018 /* First reset the PRI state of the device */
2019 ret = pci_reset_pri(pdev);
2024 ret = pci_enable_pri(pdev, reqs);
2029 ret = pri_reset_while_enabled(pdev);
2034 ret = pci_enable_ats(pdev, PAGE_SHIFT);
2041 pci_disable_pri(pdev);
2042 pci_disable_pasid(pdev);
2047 /* FIXME: Move this to PCI code */
2048 #define PCI_PRI_TLP_OFF (1 << 15)
2050 static bool pci_pri_tlp_required(struct pci_dev *pdev)
2055 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2059 pci_read_config_word(pdev, pos + PCI_PRI_STATUS, &status);
2061 return (status & PCI_PRI_TLP_OFF) ? true : false;
2065 * If a device is not yet associated with a domain, this function
2066 * assigns it visible for the hardware
2068 static int attach_device(struct device *dev,
2069 struct protection_domain *domain)
2071 struct pci_dev *pdev;
2072 struct iommu_dev_data *dev_data;
2073 unsigned long flags;
2076 dev_data = get_dev_data(dev);
2078 if (!dev_is_pci(dev))
2079 goto skip_ats_check;
2081 pdev = to_pci_dev(dev);
2082 if (domain->flags & PD_IOMMUV2_MASK) {
2083 if (!dev_data->passthrough)
2086 if (dev_data->iommu_v2) {
2087 if (pdev_iommuv2_enable(pdev) != 0)
2090 dev_data->ats.enabled = true;
2091 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2092 dev_data->pri_tlp = pci_pri_tlp_required(pdev);
2094 } else if (amd_iommu_iotlb_sup &&
2095 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
2096 dev_data->ats.enabled = true;
2097 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2101 spin_lock_irqsave(&amd_iommu_devtable_lock, flags);
2102 ret = __attach_device(dev_data, domain);
2103 spin_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2106 * We might boot into a crash-kernel here. The crashed kernel
2107 * left the caches in the IOMMU dirty. So we have to flush
2108 * here to evict all dirty stuff.
2110 domain_flush_tlb_pde(domain);
2116 * Removes a device from a protection domain (unlocked)
2118 static void __detach_device(struct iommu_dev_data *dev_data)
2120 struct protection_domain *domain;
2123 * Must be called with IRQs disabled. Warn here to detect early
2126 WARN_ON(!irqs_disabled());
2128 if (WARN_ON(!dev_data->domain))
2131 domain = dev_data->domain;
2133 spin_lock(&domain->lock);
2135 do_detach(dev_data);
2137 spin_unlock(&domain->lock);
2141 * Removes a device from a protection domain (with devtable_lock held)
2143 static void detach_device(struct device *dev)
2145 struct protection_domain *domain;
2146 struct iommu_dev_data *dev_data;
2147 unsigned long flags;
2149 dev_data = get_dev_data(dev);
2150 domain = dev_data->domain;
2152 /* lock device table */
2153 spin_lock_irqsave(&amd_iommu_devtable_lock, flags);
2154 __detach_device(dev_data);
2155 spin_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2157 if (!dev_is_pci(dev))
2160 if (domain->flags & PD_IOMMUV2_MASK && dev_data->iommu_v2)
2161 pdev_iommuv2_disable(to_pci_dev(dev));
2162 else if (dev_data->ats.enabled)
2163 pci_disable_ats(to_pci_dev(dev));
2165 dev_data->ats.enabled = false;
2168 static int amd_iommu_add_device(struct device *dev)
2170 struct iommu_dev_data *dev_data;
2171 struct iommu_domain *domain;
2172 struct amd_iommu *iommu;
2175 if (!check_device(dev) || get_dev_data(dev))
2178 devid = get_device_id(dev);
2182 iommu = amd_iommu_rlookup_table[devid];
2184 ret = iommu_init_device(dev);
2186 if (ret != -ENOTSUPP)
2187 pr_err("Failed to initialize device %s - trying to proceed anyway\n",
2190 iommu_ignore_device(dev);
2191 dev->dma_ops = &dma_direct_ops;
2194 init_iommu_group(dev);
2196 dev_data = get_dev_data(dev);
2200 if (iommu_pass_through || dev_data->iommu_v2)
2201 iommu_request_dm_for_dev(dev);
2203 /* Domains are initialized for this device - have a look what we ended up with */
2204 domain = iommu_get_domain_for_dev(dev);
2205 if (domain->type == IOMMU_DOMAIN_IDENTITY)
2206 dev_data->passthrough = true;
2208 dev->dma_ops = &amd_iommu_dma_ops;
2211 iommu_completion_wait(iommu);
2216 static void amd_iommu_remove_device(struct device *dev)
2218 struct amd_iommu *iommu;
2221 if (!check_device(dev))
2224 devid = get_device_id(dev);
2228 iommu = amd_iommu_rlookup_table[devid];
2230 iommu_uninit_device(dev);
2231 iommu_completion_wait(iommu);
2234 static struct iommu_group *amd_iommu_device_group(struct device *dev)
2236 if (dev_is_pci(dev))
2237 return pci_device_group(dev);
2239 return acpihid_device_group(dev);
2242 /*****************************************************************************
2244 * The next functions belong to the dma_ops mapping/unmapping code.
2246 *****************************************************************************/
2249 * In the dma_ops path we only have the struct device. This function
2250 * finds the corresponding IOMMU, the protection domain and the
2251 * requestor id for a given device.
2252 * If the device is not yet associated with a domain this is also done
2255 static struct protection_domain *get_domain(struct device *dev)
2257 struct protection_domain *domain;
2258 struct iommu_domain *io_domain;
2260 if (!check_device(dev))
2261 return ERR_PTR(-EINVAL);
2263 domain = get_dev_data(dev)->domain;
2264 if (domain == NULL && get_dev_data(dev)->defer_attach) {
2265 get_dev_data(dev)->defer_attach = false;
2266 io_domain = iommu_get_domain_for_dev(dev);
2267 domain = to_pdomain(io_domain);
2268 attach_device(dev, domain);
2271 return ERR_PTR(-EBUSY);
2273 if (!dma_ops_domain(domain))
2274 return ERR_PTR(-EBUSY);
2279 static void update_device_table(struct protection_domain *domain)
2281 struct iommu_dev_data *dev_data;
2283 list_for_each_entry(dev_data, &domain->dev_list, list) {
2284 set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled,
2285 dev_data->iommu_v2);
2287 if (dev_data->devid == dev_data->alias)
2290 /* There is an alias, update device table entry for it */
2291 set_dte_entry(dev_data->alias, domain, dev_data->ats.enabled,
2292 dev_data->iommu_v2);
2296 static void update_domain(struct protection_domain *domain)
2298 if (!domain->updated)
2301 update_device_table(domain);
2303 domain_flush_devices(domain);
2304 domain_flush_tlb_pde(domain);
2306 domain->updated = false;
2309 static int dir2prot(enum dma_data_direction direction)
2311 if (direction == DMA_TO_DEVICE)
2312 return IOMMU_PROT_IR;
2313 else if (direction == DMA_FROM_DEVICE)
2314 return IOMMU_PROT_IW;
2315 else if (direction == DMA_BIDIRECTIONAL)
2316 return IOMMU_PROT_IW | IOMMU_PROT_IR;
2322 * This function contains common code for mapping of a physically
2323 * contiguous memory region into DMA address space. It is used by all
2324 * mapping functions provided with this IOMMU driver.
2325 * Must be called with the domain lock held.
2327 static dma_addr_t __map_single(struct device *dev,
2328 struct dma_ops_domain *dma_dom,
2331 enum dma_data_direction direction,
2334 dma_addr_t offset = paddr & ~PAGE_MASK;
2335 dma_addr_t address, start, ret;
2340 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
2343 address = dma_ops_alloc_iova(dev, dma_dom, pages, dma_mask);
2344 if (address == AMD_IOMMU_MAPPING_ERROR)
2347 prot = dir2prot(direction);
2350 for (i = 0; i < pages; ++i) {
2351 ret = iommu_map_page(&dma_dom->domain, start, paddr,
2352 PAGE_SIZE, prot, GFP_ATOMIC);
2361 if (unlikely(amd_iommu_np_cache)) {
2362 domain_flush_pages(&dma_dom->domain, address, size);
2363 domain_flush_complete(&dma_dom->domain);
2371 for (--i; i >= 0; --i) {
2373 iommu_unmap_page(&dma_dom->domain, start, PAGE_SIZE);
2376 domain_flush_tlb(&dma_dom->domain);
2377 domain_flush_complete(&dma_dom->domain);
2379 dma_ops_free_iova(dma_dom, address, pages);
2381 return AMD_IOMMU_MAPPING_ERROR;
2385 * Does the reverse of the __map_single function. Must be called with
2386 * the domain lock held too
2388 static void __unmap_single(struct dma_ops_domain *dma_dom,
2389 dma_addr_t dma_addr,
2393 dma_addr_t i, start;
2396 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
2397 dma_addr &= PAGE_MASK;
2400 for (i = 0; i < pages; ++i) {
2401 iommu_unmap_page(&dma_dom->domain, start, PAGE_SIZE);
2405 if (amd_iommu_unmap_flush) {
2406 dma_ops_free_iova(dma_dom, dma_addr, pages);
2407 domain_flush_tlb(&dma_dom->domain);
2408 domain_flush_complete(&dma_dom->domain);
2410 pages = __roundup_pow_of_two(pages);
2411 queue_iova(&dma_dom->iovad, dma_addr >> PAGE_SHIFT, pages, 0);
2416 * The exported map_single function for dma_ops.
2418 static dma_addr_t map_page(struct device *dev, struct page *page,
2419 unsigned long offset, size_t size,
2420 enum dma_data_direction dir,
2421 unsigned long attrs)
2423 phys_addr_t paddr = page_to_phys(page) + offset;
2424 struct protection_domain *domain;
2425 struct dma_ops_domain *dma_dom;
2428 domain = get_domain(dev);
2429 if (PTR_ERR(domain) == -EINVAL)
2430 return (dma_addr_t)paddr;
2431 else if (IS_ERR(domain))
2432 return AMD_IOMMU_MAPPING_ERROR;
2434 dma_mask = *dev->dma_mask;
2435 dma_dom = to_dma_ops_domain(domain);
2437 return __map_single(dev, dma_dom, paddr, size, dir, dma_mask);
2441 * The exported unmap_single function for dma_ops.
2443 static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2444 enum dma_data_direction dir, unsigned long attrs)
2446 struct protection_domain *domain;
2447 struct dma_ops_domain *dma_dom;
2449 domain = get_domain(dev);
2453 dma_dom = to_dma_ops_domain(domain);
2455 __unmap_single(dma_dom, dma_addr, size, dir);
2458 static int sg_num_pages(struct device *dev,
2459 struct scatterlist *sglist,
2462 unsigned long mask, boundary_size;
2463 struct scatterlist *s;
2466 mask = dma_get_seg_boundary(dev);
2467 boundary_size = mask + 1 ? ALIGN(mask + 1, PAGE_SIZE) >> PAGE_SHIFT :
2468 1UL << (BITS_PER_LONG - PAGE_SHIFT);
2470 for_each_sg(sglist, s, nelems, i) {
2473 s->dma_address = npages << PAGE_SHIFT;
2474 p = npages % boundary_size;
2475 n = iommu_num_pages(sg_phys(s), s->length, PAGE_SIZE);
2476 if (p + n > boundary_size)
2477 npages += boundary_size - p;
2485 * The exported map_sg function for dma_ops (handles scatter-gather
2488 static int map_sg(struct device *dev, struct scatterlist *sglist,
2489 int nelems, enum dma_data_direction direction,
2490 unsigned long attrs)
2492 int mapped_pages = 0, npages = 0, prot = 0, i;
2493 struct protection_domain *domain;
2494 struct dma_ops_domain *dma_dom;
2495 struct scatterlist *s;
2496 unsigned long address;
2499 domain = get_domain(dev);
2503 dma_dom = to_dma_ops_domain(domain);
2504 dma_mask = *dev->dma_mask;
2506 npages = sg_num_pages(dev, sglist, nelems);
2508 address = dma_ops_alloc_iova(dev, dma_dom, npages, dma_mask);
2509 if (address == AMD_IOMMU_MAPPING_ERROR)
2512 prot = dir2prot(direction);
2514 /* Map all sg entries */
2515 for_each_sg(sglist, s, nelems, i) {
2516 int j, pages = iommu_num_pages(sg_phys(s), s->length, PAGE_SIZE);
2518 for (j = 0; j < pages; ++j) {
2519 unsigned long bus_addr, phys_addr;
2522 bus_addr = address + s->dma_address + (j << PAGE_SHIFT);
2523 phys_addr = (sg_phys(s) & PAGE_MASK) + (j << PAGE_SHIFT);
2524 ret = iommu_map_page(domain, bus_addr, phys_addr, PAGE_SIZE, prot, GFP_ATOMIC);
2532 /* Everything is mapped - write the right values into s->dma_address */
2533 for_each_sg(sglist, s, nelems, i) {
2534 s->dma_address += address + s->offset;
2535 s->dma_length = s->length;
2541 pr_err("%s: IOMMU mapping error in map_sg (io-pages: %d)\n",
2542 dev_name(dev), npages);
2544 for_each_sg(sglist, s, nelems, i) {
2545 int j, pages = iommu_num_pages(sg_phys(s), s->length, PAGE_SIZE);
2547 for (j = 0; j < pages; ++j) {
2548 unsigned long bus_addr;
2550 bus_addr = address + s->dma_address + (j << PAGE_SHIFT);
2551 iommu_unmap_page(domain, bus_addr, PAGE_SIZE);
2559 free_iova_fast(&dma_dom->iovad, address, npages);
2566 * The exported map_sg function for dma_ops (handles scatter-gather
2569 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2570 int nelems, enum dma_data_direction dir,
2571 unsigned long attrs)
2573 struct protection_domain *domain;
2574 struct dma_ops_domain *dma_dom;
2575 unsigned long startaddr;
2578 domain = get_domain(dev);
2582 startaddr = sg_dma_address(sglist) & PAGE_MASK;
2583 dma_dom = to_dma_ops_domain(domain);
2584 npages = sg_num_pages(dev, sglist, nelems);
2586 __unmap_single(dma_dom, startaddr, npages << PAGE_SHIFT, dir);
2590 * The exported alloc_coherent function for dma_ops.
2592 static void *alloc_coherent(struct device *dev, size_t size,
2593 dma_addr_t *dma_addr, gfp_t flag,
2594 unsigned long attrs)
2596 u64 dma_mask = dev->coherent_dma_mask;
2597 struct protection_domain *domain = get_domain(dev);
2598 bool is_direct = false;
2601 if (IS_ERR(domain)) {
2602 if (PTR_ERR(domain) != -EINVAL)
2607 virt_addr = dma_direct_alloc(dev, size, dma_addr, flag, attrs);
2608 if (!virt_addr || is_direct)
2612 dma_mask = *dev->dma_mask;
2614 *dma_addr = __map_single(dev, to_dma_ops_domain(domain),
2615 virt_to_phys(virt_addr), PAGE_ALIGN(size),
2616 DMA_BIDIRECTIONAL, dma_mask);
2617 if (*dma_addr == AMD_IOMMU_MAPPING_ERROR)
2622 dma_direct_free(dev, size, virt_addr, *dma_addr, attrs);
2627 * The exported free_coherent function for dma_ops.
2629 static void free_coherent(struct device *dev, size_t size,
2630 void *virt_addr, dma_addr_t dma_addr,
2631 unsigned long attrs)
2633 struct protection_domain *domain = get_domain(dev);
2635 size = PAGE_ALIGN(size);
2637 if (!IS_ERR(domain)) {
2638 struct dma_ops_domain *dma_dom = to_dma_ops_domain(domain);
2640 __unmap_single(dma_dom, dma_addr, size, DMA_BIDIRECTIONAL);
2643 dma_direct_free(dev, size, virt_addr, dma_addr, attrs);
2647 * This function is called by the DMA layer to find out if we can handle a
2648 * particular device. It is part of the dma_ops.
2650 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
2652 if (!dma_direct_supported(dev, mask))
2654 return check_device(dev);
2657 static int amd_iommu_mapping_error(struct device *dev, dma_addr_t dma_addr)
2659 return dma_addr == AMD_IOMMU_MAPPING_ERROR;
2662 static const struct dma_map_ops amd_iommu_dma_ops = {
2663 .alloc = alloc_coherent,
2664 .free = free_coherent,
2665 .map_page = map_page,
2666 .unmap_page = unmap_page,
2668 .unmap_sg = unmap_sg,
2669 .dma_supported = amd_iommu_dma_supported,
2670 .mapping_error = amd_iommu_mapping_error,
2673 static int init_reserved_iova_ranges(void)
2675 struct pci_dev *pdev = NULL;
2678 init_iova_domain(&reserved_iova_ranges, PAGE_SIZE, IOVA_START_PFN);
2680 lockdep_set_class(&reserved_iova_ranges.iova_rbtree_lock,
2681 &reserved_rbtree_key);
2683 /* MSI memory range */
2684 val = reserve_iova(&reserved_iova_ranges,
2685 IOVA_PFN(MSI_RANGE_START), IOVA_PFN(MSI_RANGE_END));
2687 pr_err("Reserving MSI range failed\n");
2691 /* HT memory range */
2692 val = reserve_iova(&reserved_iova_ranges,
2693 IOVA_PFN(HT_RANGE_START), IOVA_PFN(HT_RANGE_END));
2695 pr_err("Reserving HT range failed\n");
2700 * Memory used for PCI resources
2701 * FIXME: Check whether we can reserve the PCI-hole completly
2703 for_each_pci_dev(pdev) {
2706 for (i = 0; i < PCI_NUM_RESOURCES; ++i) {
2707 struct resource *r = &pdev->resource[i];
2709 if (!(r->flags & IORESOURCE_MEM))
2712 val = reserve_iova(&reserved_iova_ranges,
2716 pr_err("Reserve pci-resource range failed\n");
2725 int __init amd_iommu_init_api(void)
2729 ret = iova_cache_get();
2733 ret = init_reserved_iova_ranges();
2737 err = bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
2740 #ifdef CONFIG_ARM_AMBA
2741 err = bus_set_iommu(&amba_bustype, &amd_iommu_ops);
2745 err = bus_set_iommu(&platform_bus_type, &amd_iommu_ops);
2752 int __init amd_iommu_init_dma_ops(void)
2754 swiotlb = (iommu_pass_through || sme_me_mask) ? 1 : 0;
2758 * In case we don't initialize SWIOTLB (actually the common case
2759 * when AMD IOMMU is enabled and SME is not active), make sure there
2760 * are global dma_ops set as a fall-back for devices not handled by
2761 * this driver (for example non-PCI devices). When SME is active,
2762 * make sure that swiotlb variable remains set so the global dma_ops
2763 * continue to be SWIOTLB.
2766 dma_ops = &dma_direct_ops;
2768 if (amd_iommu_unmap_flush)
2769 pr_info("AMD-Vi: IO/TLB flush on unmap enabled\n");
2771 pr_info("AMD-Vi: Lazy IO/TLB flushing enabled\n");
2777 /*****************************************************************************
2779 * The following functions belong to the exported interface of AMD IOMMU
2781 * This interface allows access to lower level functions of the IOMMU
2782 * like protection domain handling and assignement of devices to domains
2783 * which is not possible with the dma_ops interface.
2785 *****************************************************************************/
2787 static void cleanup_domain(struct protection_domain *domain)
2789 struct iommu_dev_data *entry;
2790 unsigned long flags;
2792 spin_lock_irqsave(&amd_iommu_devtable_lock, flags);
2794 while (!list_empty(&domain->dev_list)) {
2795 entry = list_first_entry(&domain->dev_list,
2796 struct iommu_dev_data, list);
2797 __detach_device(entry);
2800 spin_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2803 static void protection_domain_free(struct protection_domain *domain)
2808 del_domain_from_list(domain);
2811 domain_id_free(domain->id);
2816 static int protection_domain_init(struct protection_domain *domain)
2818 spin_lock_init(&domain->lock);
2819 mutex_init(&domain->api_lock);
2820 domain->id = domain_id_alloc();
2823 INIT_LIST_HEAD(&domain->dev_list);
2828 static struct protection_domain *protection_domain_alloc(void)
2830 struct protection_domain *domain;
2832 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2836 if (protection_domain_init(domain))
2839 add_domain_to_list(domain);
2849 static struct iommu_domain *amd_iommu_domain_alloc(unsigned type)
2851 struct protection_domain *pdomain;
2852 struct dma_ops_domain *dma_domain;
2855 case IOMMU_DOMAIN_UNMANAGED:
2856 pdomain = protection_domain_alloc();
2860 pdomain->mode = PAGE_MODE_3_LEVEL;
2861 pdomain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2862 if (!pdomain->pt_root) {
2863 protection_domain_free(pdomain);
2867 pdomain->domain.geometry.aperture_start = 0;
2868 pdomain->domain.geometry.aperture_end = ~0ULL;
2869 pdomain->domain.geometry.force_aperture = true;
2872 case IOMMU_DOMAIN_DMA:
2873 dma_domain = dma_ops_domain_alloc();
2875 pr_err("AMD-Vi: Failed to allocate\n");
2878 pdomain = &dma_domain->domain;
2880 case IOMMU_DOMAIN_IDENTITY:
2881 pdomain = protection_domain_alloc();
2885 pdomain->mode = PAGE_MODE_NONE;
2891 return &pdomain->domain;
2894 static void amd_iommu_domain_free(struct iommu_domain *dom)
2896 struct protection_domain *domain;
2897 struct dma_ops_domain *dma_dom;
2899 domain = to_pdomain(dom);
2901 if (domain->dev_cnt > 0)
2902 cleanup_domain(domain);
2904 BUG_ON(domain->dev_cnt != 0);
2909 switch (dom->type) {
2910 case IOMMU_DOMAIN_DMA:
2911 /* Now release the domain */
2912 dma_dom = to_dma_ops_domain(domain);
2913 dma_ops_domain_free(dma_dom);
2916 if (domain->mode != PAGE_MODE_NONE)
2917 free_pagetable(domain);
2919 if (domain->flags & PD_IOMMUV2_MASK)
2920 free_gcr3_table(domain);
2922 protection_domain_free(domain);
2927 static void amd_iommu_detach_device(struct iommu_domain *dom,
2930 struct iommu_dev_data *dev_data = dev->archdata.iommu;
2931 struct amd_iommu *iommu;
2934 if (!check_device(dev))
2937 devid = get_device_id(dev);
2941 if (dev_data->domain != NULL)
2944 iommu = amd_iommu_rlookup_table[devid];
2948 #ifdef CONFIG_IRQ_REMAP
2949 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) &&
2950 (dom->type == IOMMU_DOMAIN_UNMANAGED))
2951 dev_data->use_vapic = 0;
2954 iommu_completion_wait(iommu);
2957 static int amd_iommu_attach_device(struct iommu_domain *dom,
2960 struct protection_domain *domain = to_pdomain(dom);
2961 struct iommu_dev_data *dev_data;
2962 struct amd_iommu *iommu;
2965 if (!check_device(dev))
2968 dev_data = dev->archdata.iommu;
2970 iommu = amd_iommu_rlookup_table[dev_data->devid];
2974 if (dev_data->domain)
2977 ret = attach_device(dev, domain);
2979 #ifdef CONFIG_IRQ_REMAP
2980 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
2981 if (dom->type == IOMMU_DOMAIN_UNMANAGED)
2982 dev_data->use_vapic = 1;
2984 dev_data->use_vapic = 0;
2988 iommu_completion_wait(iommu);
2993 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
2994 phys_addr_t paddr, size_t page_size, int iommu_prot)
2996 struct protection_domain *domain = to_pdomain(dom);
3000 if (domain->mode == PAGE_MODE_NONE)
3003 if (iommu_prot & IOMMU_READ)
3004 prot |= IOMMU_PROT_IR;
3005 if (iommu_prot & IOMMU_WRITE)
3006 prot |= IOMMU_PROT_IW;
3008 mutex_lock(&domain->api_lock);
3009 ret = iommu_map_page(domain, iova, paddr, page_size, prot, GFP_KERNEL);
3010 mutex_unlock(&domain->api_lock);
3015 static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
3018 struct protection_domain *domain = to_pdomain(dom);
3021 if (domain->mode == PAGE_MODE_NONE)
3024 mutex_lock(&domain->api_lock);
3025 unmap_size = iommu_unmap_page(domain, iova, page_size);
3026 mutex_unlock(&domain->api_lock);
3031 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
3034 struct protection_domain *domain = to_pdomain(dom);
3035 unsigned long offset_mask, pte_pgsize;
3038 if (domain->mode == PAGE_MODE_NONE)
3041 pte = fetch_pte(domain, iova, &pte_pgsize);
3043 if (!pte || !IOMMU_PTE_PRESENT(*pte))
3046 offset_mask = pte_pgsize - 1;
3047 __pte = *pte & PM_ADDR_MASK;
3049 return (__pte & ~offset_mask) | (iova & offset_mask);
3052 static bool amd_iommu_capable(enum iommu_cap cap)
3055 case IOMMU_CAP_CACHE_COHERENCY:
3057 case IOMMU_CAP_INTR_REMAP:
3058 return (irq_remapping_enabled == 1);
3059 case IOMMU_CAP_NOEXEC:
3066 static void amd_iommu_get_resv_regions(struct device *dev,
3067 struct list_head *head)
3069 struct iommu_resv_region *region;
3070 struct unity_map_entry *entry;
3073 devid = get_device_id(dev);
3077 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
3081 if (devid < entry->devid_start || devid > entry->devid_end)
3084 length = entry->address_end - entry->address_start;
3085 if (entry->prot & IOMMU_PROT_IR)
3087 if (entry->prot & IOMMU_PROT_IW)
3088 prot |= IOMMU_WRITE;
3090 region = iommu_alloc_resv_region(entry->address_start,
3094 pr_err("Out of memory allocating dm-regions for %s\n",
3098 list_add_tail(®ion->list, head);
3101 region = iommu_alloc_resv_region(MSI_RANGE_START,
3102 MSI_RANGE_END - MSI_RANGE_START + 1,
3106 list_add_tail(®ion->list, head);
3108 region = iommu_alloc_resv_region(HT_RANGE_START,
3109 HT_RANGE_END - HT_RANGE_START + 1,
3110 0, IOMMU_RESV_RESERVED);
3113 list_add_tail(®ion->list, head);
3116 static void amd_iommu_put_resv_regions(struct device *dev,
3117 struct list_head *head)
3119 struct iommu_resv_region *entry, *next;
3121 list_for_each_entry_safe(entry, next, head, list)
3125 static void amd_iommu_apply_resv_region(struct device *dev,
3126 struct iommu_domain *domain,
3127 struct iommu_resv_region *region)
3129 struct dma_ops_domain *dma_dom = to_dma_ops_domain(to_pdomain(domain));
3130 unsigned long start, end;
3132 start = IOVA_PFN(region->start);
3133 end = IOVA_PFN(region->start + region->length - 1);
3135 WARN_ON_ONCE(reserve_iova(&dma_dom->iovad, start, end) == NULL);
3138 static bool amd_iommu_is_attach_deferred(struct iommu_domain *domain,
3141 struct iommu_dev_data *dev_data = dev->archdata.iommu;
3142 return dev_data->defer_attach;
3145 static void amd_iommu_flush_iotlb_all(struct iommu_domain *domain)
3147 struct protection_domain *dom = to_pdomain(domain);
3149 domain_flush_tlb_pde(dom);
3150 domain_flush_complete(dom);
3153 static void amd_iommu_iotlb_range_add(struct iommu_domain *domain,
3154 unsigned long iova, size_t size)
3158 const struct iommu_ops amd_iommu_ops = {
3159 .capable = amd_iommu_capable,
3160 .domain_alloc = amd_iommu_domain_alloc,
3161 .domain_free = amd_iommu_domain_free,
3162 .attach_dev = amd_iommu_attach_device,
3163 .detach_dev = amd_iommu_detach_device,
3164 .map = amd_iommu_map,
3165 .unmap = amd_iommu_unmap,
3166 .map_sg = default_iommu_map_sg,
3167 .iova_to_phys = amd_iommu_iova_to_phys,
3168 .add_device = amd_iommu_add_device,
3169 .remove_device = amd_iommu_remove_device,
3170 .device_group = amd_iommu_device_group,
3171 .get_resv_regions = amd_iommu_get_resv_regions,
3172 .put_resv_regions = amd_iommu_put_resv_regions,
3173 .apply_resv_region = amd_iommu_apply_resv_region,
3174 .is_attach_deferred = amd_iommu_is_attach_deferred,
3175 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
3176 .flush_iotlb_all = amd_iommu_flush_iotlb_all,
3177 .iotlb_range_add = amd_iommu_iotlb_range_add,
3178 .iotlb_sync = amd_iommu_flush_iotlb_all,
3181 /*****************************************************************************
3183 * The next functions do a basic initialization of IOMMU for pass through
3186 * In passthrough mode the IOMMU is initialized and enabled but not used for
3187 * DMA-API translation.
3189 *****************************************************************************/
3191 /* IOMMUv2 specific functions */
3192 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
3194 return atomic_notifier_chain_register(&ppr_notifier, nb);
3196 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
3198 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
3200 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
3202 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
3204 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
3206 struct protection_domain *domain = to_pdomain(dom);
3207 unsigned long flags;
3209 spin_lock_irqsave(&domain->lock, flags);
3211 /* Update data structure */
3212 domain->mode = PAGE_MODE_NONE;
3213 domain->updated = true;
3215 /* Make changes visible to IOMMUs */
3216 update_domain(domain);
3218 /* Page-table is not visible to IOMMU anymore, so free it */
3219 free_pagetable(domain);
3221 spin_unlock_irqrestore(&domain->lock, flags);
3223 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
3225 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
3227 struct protection_domain *domain = to_pdomain(dom);
3228 unsigned long flags;
3231 if (pasids <= 0 || pasids > (PASID_MASK + 1))
3234 /* Number of GCR3 table levels required */
3235 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
3238 if (levels > amd_iommu_max_glx_val)
3241 spin_lock_irqsave(&domain->lock, flags);
3244 * Save us all sanity checks whether devices already in the
3245 * domain support IOMMUv2. Just force that the domain has no
3246 * devices attached when it is switched into IOMMUv2 mode.
3249 if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
3253 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
3254 if (domain->gcr3_tbl == NULL)
3257 domain->glx = levels;
3258 domain->flags |= PD_IOMMUV2_MASK;
3259 domain->updated = true;
3261 update_domain(domain);
3266 spin_unlock_irqrestore(&domain->lock, flags);
3270 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
3272 static int __flush_pasid(struct protection_domain *domain, int pasid,
3273 u64 address, bool size)
3275 struct iommu_dev_data *dev_data;
3276 struct iommu_cmd cmd;
3279 if (!(domain->flags & PD_IOMMUV2_MASK))
3282 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
3285 * IOMMU TLB needs to be flushed before Device TLB to
3286 * prevent device TLB refill from IOMMU TLB
3288 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
3289 if (domain->dev_iommu[i] == 0)
3292 ret = iommu_queue_command(amd_iommus[i], &cmd);
3297 /* Wait until IOMMU TLB flushes are complete */
3298 domain_flush_complete(domain);
3300 /* Now flush device TLBs */
3301 list_for_each_entry(dev_data, &domain->dev_list, list) {
3302 struct amd_iommu *iommu;
3306 There might be non-IOMMUv2 capable devices in an IOMMUv2
3309 if (!dev_data->ats.enabled)
3312 qdep = dev_data->ats.qdep;
3313 iommu = amd_iommu_rlookup_table[dev_data->devid];
3315 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
3316 qdep, address, size);
3318 ret = iommu_queue_command(iommu, &cmd);
3323 /* Wait until all device TLBs are flushed */
3324 domain_flush_complete(domain);
3333 static int __amd_iommu_flush_page(struct protection_domain *domain, int pasid,
3336 return __flush_pasid(domain, pasid, address, false);
3339 int amd_iommu_flush_page(struct iommu_domain *dom, int pasid,
3342 struct protection_domain *domain = to_pdomain(dom);
3343 unsigned long flags;
3346 spin_lock_irqsave(&domain->lock, flags);
3347 ret = __amd_iommu_flush_page(domain, pasid, address);
3348 spin_unlock_irqrestore(&domain->lock, flags);
3352 EXPORT_SYMBOL(amd_iommu_flush_page);
3354 static int __amd_iommu_flush_tlb(struct protection_domain *domain, int pasid)
3356 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
3360 int amd_iommu_flush_tlb(struct iommu_domain *dom, int pasid)
3362 struct protection_domain *domain = to_pdomain(dom);
3363 unsigned long flags;
3366 spin_lock_irqsave(&domain->lock, flags);
3367 ret = __amd_iommu_flush_tlb(domain, pasid);
3368 spin_unlock_irqrestore(&domain->lock, flags);
3372 EXPORT_SYMBOL(amd_iommu_flush_tlb);
3374 static u64 *__get_gcr3_pte(u64 *root, int level, int pasid, bool alloc)
3381 index = (pasid >> (9 * level)) & 0x1ff;
3387 if (!(*pte & GCR3_VALID)) {
3391 root = (void *)get_zeroed_page(GFP_ATOMIC);
3395 *pte = iommu_virt_to_phys(root) | GCR3_VALID;
3398 root = iommu_phys_to_virt(*pte & PAGE_MASK);
3406 static int __set_gcr3(struct protection_domain *domain, int pasid,
3411 if (domain->mode != PAGE_MODE_NONE)
3414 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3418 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3420 return __amd_iommu_flush_tlb(domain, pasid);
3423 static int __clear_gcr3(struct protection_domain *domain, int pasid)
3427 if (domain->mode != PAGE_MODE_NONE)
3430 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3436 return __amd_iommu_flush_tlb(domain, pasid);
3439 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, int pasid,
3442 struct protection_domain *domain = to_pdomain(dom);
3443 unsigned long flags;
3446 spin_lock_irqsave(&domain->lock, flags);
3447 ret = __set_gcr3(domain, pasid, cr3);
3448 spin_unlock_irqrestore(&domain->lock, flags);
3452 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3454 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, int pasid)
3456 struct protection_domain *domain = to_pdomain(dom);
3457 unsigned long flags;
3460 spin_lock_irqsave(&domain->lock, flags);
3461 ret = __clear_gcr3(domain, pasid);
3462 spin_unlock_irqrestore(&domain->lock, flags);
3466 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3468 int amd_iommu_complete_ppr(struct pci_dev *pdev, int pasid,
3469 int status, int tag)
3471 struct iommu_dev_data *dev_data;
3472 struct amd_iommu *iommu;
3473 struct iommu_cmd cmd;
3475 dev_data = get_dev_data(&pdev->dev);
3476 iommu = amd_iommu_rlookup_table[dev_data->devid];
3478 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3479 tag, dev_data->pri_tlp);
3481 return iommu_queue_command(iommu, &cmd);
3483 EXPORT_SYMBOL(amd_iommu_complete_ppr);
3485 struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3487 struct protection_domain *pdomain;
3489 pdomain = get_domain(&pdev->dev);
3490 if (IS_ERR(pdomain))
3493 /* Only return IOMMUv2 domains */
3494 if (!(pdomain->flags & PD_IOMMUV2_MASK))
3497 return &pdomain->domain;
3499 EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3501 void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3503 struct iommu_dev_data *dev_data;
3505 if (!amd_iommu_v2_supported())
3508 dev_data = get_dev_data(&pdev->dev);
3509 dev_data->errata |= (1 << erratum);
3511 EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3513 int amd_iommu_device_info(struct pci_dev *pdev,
3514 struct amd_iommu_device_info *info)
3519 if (pdev == NULL || info == NULL)
3522 if (!amd_iommu_v2_supported())
3525 memset(info, 0, sizeof(*info));
3527 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS);
3529 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3531 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3533 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3535 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3539 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3540 max_pasids = min(max_pasids, (1 << 20));
3542 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3543 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3545 features = pci_pasid_features(pdev);
3546 if (features & PCI_PASID_CAP_EXEC)
3547 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3548 if (features & PCI_PASID_CAP_PRIV)
3549 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3554 EXPORT_SYMBOL(amd_iommu_device_info);
3556 #ifdef CONFIG_IRQ_REMAP
3558 /*****************************************************************************
3560 * Interrupt Remapping Implementation
3562 *****************************************************************************/
3564 static struct irq_chip amd_ir_chip;
3566 static void set_dte_irq_entry(u16 devid, struct irq_remap_table *table)
3570 dte = amd_iommu_dev_table[devid].data[2];
3571 dte &= ~DTE_IRQ_PHYS_ADDR_MASK;
3572 dte |= iommu_virt_to_phys(table->table);
3573 dte |= DTE_IRQ_REMAP_INTCTL;
3574 dte |= DTE_IRQ_TABLE_LEN;
3575 dte |= DTE_IRQ_REMAP_ENABLE;
3577 amd_iommu_dev_table[devid].data[2] = dte;
3580 static struct irq_remap_table *get_irq_table(u16 devid)
3582 struct irq_remap_table *table;
3584 if (WARN_ONCE(!amd_iommu_rlookup_table[devid],
3585 "%s: no iommu for devid %x\n", __func__, devid))
3588 table = irq_lookup_table[devid];
3589 if (WARN_ONCE(!table, "%s: no table for devid %x\n", __func__, devid))
3595 static struct irq_remap_table *__alloc_irq_table(void)
3597 struct irq_remap_table *table;
3599 table = kzalloc(sizeof(*table), GFP_KERNEL);
3603 table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_KERNEL);
3604 if (!table->table) {
3608 raw_spin_lock_init(&table->lock);
3610 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
3611 memset(table->table, 0,
3612 MAX_IRQS_PER_TABLE * sizeof(u32));
3614 memset(table->table, 0,
3615 (MAX_IRQS_PER_TABLE * (sizeof(u64) * 2)));
3619 static void set_remap_table_entry(struct amd_iommu *iommu, u16 devid,
3620 struct irq_remap_table *table)
3622 irq_lookup_table[devid] = table;
3623 set_dte_irq_entry(devid, table);
3624 iommu_flush_dte(iommu, devid);
3627 static struct irq_remap_table *alloc_irq_table(u16 devid)
3629 struct irq_remap_table *table = NULL;
3630 struct irq_remap_table *new_table = NULL;
3631 struct amd_iommu *iommu;
3632 unsigned long flags;
3635 spin_lock_irqsave(&iommu_table_lock, flags);
3637 iommu = amd_iommu_rlookup_table[devid];
3641 table = irq_lookup_table[devid];
3645 alias = amd_iommu_alias_table[devid];
3646 table = irq_lookup_table[alias];
3648 set_remap_table_entry(iommu, devid, table);
3651 spin_unlock_irqrestore(&iommu_table_lock, flags);
3653 /* Nothing there yet, allocate new irq remapping table */
3654 new_table = __alloc_irq_table();
3658 spin_lock_irqsave(&iommu_table_lock, flags);
3660 table = irq_lookup_table[devid];
3664 table = irq_lookup_table[alias];
3666 set_remap_table_entry(iommu, devid, table);
3673 set_remap_table_entry(iommu, devid, table);
3675 set_remap_table_entry(iommu, alias, table);
3678 iommu_completion_wait(iommu);
3681 spin_unlock_irqrestore(&iommu_table_lock, flags);
3684 kmem_cache_free(amd_iommu_irq_cache, new_table->table);
3690 static int alloc_irq_index(u16 devid, int count, bool align)
3692 struct irq_remap_table *table;
3693 int index, c, alignment = 1;
3694 unsigned long flags;
3695 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
3700 table = alloc_irq_table(devid);
3705 alignment = roundup_pow_of_two(count);
3707 raw_spin_lock_irqsave(&table->lock, flags);
3709 /* Scan table for free entries */
3710 for (index = ALIGN(table->min_index, alignment), c = 0;
3711 index < MAX_IRQS_PER_TABLE;) {
3712 if (!iommu->irte_ops->is_allocated(table, index)) {
3716 index = ALIGN(index + 1, alignment);
3722 iommu->irte_ops->set_allocated(table, index - c + 1);
3734 raw_spin_unlock_irqrestore(&table->lock, flags);
3739 static int modify_irte_ga(u16 devid, int index, struct irte_ga *irte,
3740 struct amd_ir_data *data)
3742 struct irq_remap_table *table;
3743 struct amd_iommu *iommu;
3744 unsigned long flags;
3745 struct irte_ga *entry;
3747 iommu = amd_iommu_rlookup_table[devid];
3751 table = get_irq_table(devid);
3755 raw_spin_lock_irqsave(&table->lock, flags);
3757 entry = (struct irte_ga *)table->table;
3758 entry = &entry[index];
3759 entry->lo.fields_remap.valid = 0;
3760 entry->hi.val = irte->hi.val;
3761 entry->lo.val = irte->lo.val;
3762 entry->lo.fields_remap.valid = 1;
3766 raw_spin_unlock_irqrestore(&table->lock, flags);
3768 iommu_flush_irt(iommu, devid);
3769 iommu_completion_wait(iommu);
3774 static int modify_irte(u16 devid, int index, union irte *irte)
3776 struct irq_remap_table *table;
3777 struct amd_iommu *iommu;
3778 unsigned long flags;
3780 iommu = amd_iommu_rlookup_table[devid];
3784 table = get_irq_table(devid);
3788 raw_spin_lock_irqsave(&table->lock, flags);
3789 table->table[index] = irte->val;
3790 raw_spin_unlock_irqrestore(&table->lock, flags);
3792 iommu_flush_irt(iommu, devid);
3793 iommu_completion_wait(iommu);
3798 static void free_irte(u16 devid, int index)
3800 struct irq_remap_table *table;
3801 struct amd_iommu *iommu;
3802 unsigned long flags;
3804 iommu = amd_iommu_rlookup_table[devid];
3808 table = get_irq_table(devid);
3812 raw_spin_lock_irqsave(&table->lock, flags);
3813 iommu->irte_ops->clear_allocated(table, index);
3814 raw_spin_unlock_irqrestore(&table->lock, flags);
3816 iommu_flush_irt(iommu, devid);
3817 iommu_completion_wait(iommu);
3820 static void irte_prepare(void *entry,
3821 u32 delivery_mode, u32 dest_mode,
3822 u8 vector, u32 dest_apicid, int devid)
3824 union irte *irte = (union irte *) entry;
3827 irte->fields.vector = vector;
3828 irte->fields.int_type = delivery_mode;
3829 irte->fields.destination = dest_apicid;
3830 irte->fields.dm = dest_mode;
3831 irte->fields.valid = 1;
3834 static void irte_ga_prepare(void *entry,
3835 u32 delivery_mode, u32 dest_mode,
3836 u8 vector, u32 dest_apicid, int devid)
3838 struct irte_ga *irte = (struct irte_ga *) entry;
3842 irte->lo.fields_remap.int_type = delivery_mode;
3843 irte->lo.fields_remap.dm = dest_mode;
3844 irte->hi.fields.vector = vector;
3845 irte->lo.fields_remap.destination = dest_apicid;
3846 irte->lo.fields_remap.valid = 1;
3849 static void irte_activate(void *entry, u16 devid, u16 index)
3851 union irte *irte = (union irte *) entry;
3853 irte->fields.valid = 1;
3854 modify_irte(devid, index, irte);
3857 static void irte_ga_activate(void *entry, u16 devid, u16 index)
3859 struct irte_ga *irte = (struct irte_ga *) entry;
3861 irte->lo.fields_remap.valid = 1;
3862 modify_irte_ga(devid, index, irte, NULL);
3865 static void irte_deactivate(void *entry, u16 devid, u16 index)
3867 union irte *irte = (union irte *) entry;
3869 irte->fields.valid = 0;
3870 modify_irte(devid, index, irte);
3873 static void irte_ga_deactivate(void *entry, u16 devid, u16 index)
3875 struct irte_ga *irte = (struct irte_ga *) entry;
3877 irte->lo.fields_remap.valid = 0;
3878 modify_irte_ga(devid, index, irte, NULL);
3881 static void irte_set_affinity(void *entry, u16 devid, u16 index,
3882 u8 vector, u32 dest_apicid)
3884 union irte *irte = (union irte *) entry;
3886 irte->fields.vector = vector;
3887 irte->fields.destination = dest_apicid;
3888 modify_irte(devid, index, irte);
3891 static void irte_ga_set_affinity(void *entry, u16 devid, u16 index,
3892 u8 vector, u32 dest_apicid)
3894 struct irte_ga *irte = (struct irte_ga *) entry;
3896 if (!irte->lo.fields_remap.guest_mode) {
3897 irte->hi.fields.vector = vector;
3898 irte->lo.fields_remap.destination = dest_apicid;
3899 modify_irte_ga(devid, index, irte, NULL);
3903 #define IRTE_ALLOCATED (~1U)
3904 static void irte_set_allocated(struct irq_remap_table *table, int index)
3906 table->table[index] = IRTE_ALLOCATED;
3909 static void irte_ga_set_allocated(struct irq_remap_table *table, int index)
3911 struct irte_ga *ptr = (struct irte_ga *)table->table;
3912 struct irte_ga *irte = &ptr[index];
3914 memset(&irte->lo.val, 0, sizeof(u64));
3915 memset(&irte->hi.val, 0, sizeof(u64));
3916 irte->hi.fields.vector = 0xff;
3919 static bool irte_is_allocated(struct irq_remap_table *table, int index)
3921 union irte *ptr = (union irte *)table->table;
3922 union irte *irte = &ptr[index];
3924 return irte->val != 0;
3927 static bool irte_ga_is_allocated(struct irq_remap_table *table, int index)
3929 struct irte_ga *ptr = (struct irte_ga *)table->table;
3930 struct irte_ga *irte = &ptr[index];
3932 return irte->hi.fields.vector != 0;
3935 static void irte_clear_allocated(struct irq_remap_table *table, int index)
3937 table->table[index] = 0;
3940 static void irte_ga_clear_allocated(struct irq_remap_table *table, int index)
3942 struct irte_ga *ptr = (struct irte_ga *)table->table;
3943 struct irte_ga *irte = &ptr[index];
3945 memset(&irte->lo.val, 0, sizeof(u64));
3946 memset(&irte->hi.val, 0, sizeof(u64));
3949 static int get_devid(struct irq_alloc_info *info)
3953 switch (info->type) {
3954 case X86_IRQ_ALLOC_TYPE_IOAPIC:
3955 devid = get_ioapic_devid(info->ioapic_id);
3957 case X86_IRQ_ALLOC_TYPE_HPET:
3958 devid = get_hpet_devid(info->hpet_id);
3960 case X86_IRQ_ALLOC_TYPE_MSI:
3961 case X86_IRQ_ALLOC_TYPE_MSIX:
3962 devid = get_device_id(&info->msi_dev->dev);
3972 static struct irq_domain *get_ir_irq_domain(struct irq_alloc_info *info)
3974 struct amd_iommu *iommu;
3980 devid = get_devid(info);
3982 iommu = amd_iommu_rlookup_table[devid];
3984 return iommu->ir_domain;
3990 static struct irq_domain *get_irq_domain(struct irq_alloc_info *info)
3992 struct amd_iommu *iommu;
3998 switch (info->type) {
3999 case X86_IRQ_ALLOC_TYPE_MSI:
4000 case X86_IRQ_ALLOC_TYPE_MSIX:
4001 devid = get_device_id(&info->msi_dev->dev);
4005 iommu = amd_iommu_rlookup_table[devid];
4007 return iommu->msi_domain;
4016 struct irq_remap_ops amd_iommu_irq_ops = {
4017 .prepare = amd_iommu_prepare,
4018 .enable = amd_iommu_enable,
4019 .disable = amd_iommu_disable,
4020 .reenable = amd_iommu_reenable,
4021 .enable_faulting = amd_iommu_enable_faulting,
4022 .get_ir_irq_domain = get_ir_irq_domain,
4023 .get_irq_domain = get_irq_domain,
4026 static void irq_remapping_prepare_irte(struct amd_ir_data *data,
4027 struct irq_cfg *irq_cfg,
4028 struct irq_alloc_info *info,
4029 int devid, int index, int sub_handle)
4031 struct irq_2_irte *irte_info = &data->irq_2_irte;
4032 struct msi_msg *msg = &data->msi_entry;
4033 struct IO_APIC_route_entry *entry;
4034 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
4039 data->irq_2_irte.devid = devid;
4040 data->irq_2_irte.index = index + sub_handle;
4041 iommu->irte_ops->prepare(data->entry, apic->irq_delivery_mode,
4042 apic->irq_dest_mode, irq_cfg->vector,
4043 irq_cfg->dest_apicid, devid);
4045 switch (info->type) {
4046 case X86_IRQ_ALLOC_TYPE_IOAPIC:
4047 /* Setup IOAPIC entry */
4048 entry = info->ioapic_entry;
4049 info->ioapic_entry = NULL;
4050 memset(entry, 0, sizeof(*entry));
4051 entry->vector = index;
4053 entry->trigger = info->ioapic_trigger;
4054 entry->polarity = info->ioapic_polarity;
4055 /* Mask level triggered irqs. */
4056 if (info->ioapic_trigger)
4060 case X86_IRQ_ALLOC_TYPE_HPET:
4061 case X86_IRQ_ALLOC_TYPE_MSI:
4062 case X86_IRQ_ALLOC_TYPE_MSIX:
4063 msg->address_hi = MSI_ADDR_BASE_HI;
4064 msg->address_lo = MSI_ADDR_BASE_LO;
4065 msg->data = irte_info->index;
4074 struct amd_irte_ops irte_32_ops = {
4075 .prepare = irte_prepare,
4076 .activate = irte_activate,
4077 .deactivate = irte_deactivate,
4078 .set_affinity = irte_set_affinity,
4079 .set_allocated = irte_set_allocated,
4080 .is_allocated = irte_is_allocated,
4081 .clear_allocated = irte_clear_allocated,
4084 struct amd_irte_ops irte_128_ops = {
4085 .prepare = irte_ga_prepare,
4086 .activate = irte_ga_activate,
4087 .deactivate = irte_ga_deactivate,
4088 .set_affinity = irte_ga_set_affinity,
4089 .set_allocated = irte_ga_set_allocated,
4090 .is_allocated = irte_ga_is_allocated,
4091 .clear_allocated = irte_ga_clear_allocated,
4094 static int irq_remapping_alloc(struct irq_domain *domain, unsigned int virq,
4095 unsigned int nr_irqs, void *arg)
4097 struct irq_alloc_info *info = arg;
4098 struct irq_data *irq_data;
4099 struct amd_ir_data *data = NULL;
4100 struct irq_cfg *cfg;
4106 if (nr_irqs > 1 && info->type != X86_IRQ_ALLOC_TYPE_MSI &&
4107 info->type != X86_IRQ_ALLOC_TYPE_MSIX)
4111 * With IRQ remapping enabled, don't need contiguous CPU vectors
4112 * to support multiple MSI interrupts.
4114 if (info->type == X86_IRQ_ALLOC_TYPE_MSI)
4115 info->flags &= ~X86_IRQ_ALLOC_CONTIGUOUS_VECTORS;
4117 devid = get_devid(info);
4121 ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
4125 if (info->type == X86_IRQ_ALLOC_TYPE_IOAPIC) {
4126 struct irq_remap_table *table;
4127 struct amd_iommu *iommu;
4129 table = alloc_irq_table(devid);
4131 if (!table->min_index) {
4133 * Keep the first 32 indexes free for IOAPIC
4136 table->min_index = 32;
4137 iommu = amd_iommu_rlookup_table[devid];
4138 for (i = 0; i < 32; ++i)
4139 iommu->irte_ops->set_allocated(table, i);
4141 WARN_ON(table->min_index != 32);
4142 index = info->ioapic_pin;
4147 bool align = (info->type == X86_IRQ_ALLOC_TYPE_MSI);
4149 index = alloc_irq_index(devid, nr_irqs, align);
4152 pr_warn("Failed to allocate IRTE\n");
4154 goto out_free_parent;
4157 for (i = 0; i < nr_irqs; i++) {
4158 irq_data = irq_domain_get_irq_data(domain, virq + i);
4159 cfg = irqd_cfg(irq_data);
4160 if (!irq_data || !cfg) {
4166 data = kzalloc(sizeof(*data), GFP_KERNEL);
4170 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
4171 data->entry = kzalloc(sizeof(union irte), GFP_KERNEL);
4173 data->entry = kzalloc(sizeof(struct irte_ga),
4180 irq_data->hwirq = (devid << 16) + i;
4181 irq_data->chip_data = data;
4182 irq_data->chip = &amd_ir_chip;
4183 irq_remapping_prepare_irte(data, cfg, info, devid, index, i);
4184 irq_set_status_flags(virq + i, IRQ_MOVE_PCNTXT);
4190 for (i--; i >= 0; i--) {
4191 irq_data = irq_domain_get_irq_data(domain, virq + i);
4193 kfree(irq_data->chip_data);
4195 for (i = 0; i < nr_irqs; i++)
4196 free_irte(devid, index + i);
4198 irq_domain_free_irqs_common(domain, virq, nr_irqs);
4202 static void irq_remapping_free(struct irq_domain *domain, unsigned int virq,
4203 unsigned int nr_irqs)
4205 struct irq_2_irte *irte_info;
4206 struct irq_data *irq_data;
4207 struct amd_ir_data *data;
4210 for (i = 0; i < nr_irqs; i++) {
4211 irq_data = irq_domain_get_irq_data(domain, virq + i);
4212 if (irq_data && irq_data->chip_data) {
4213 data = irq_data->chip_data;
4214 irte_info = &data->irq_2_irte;
4215 free_irte(irte_info->devid, irte_info->index);
4220 irq_domain_free_irqs_common(domain, virq, nr_irqs);
4223 static void amd_ir_update_irte(struct irq_data *irqd, struct amd_iommu *iommu,
4224 struct amd_ir_data *ir_data,
4225 struct irq_2_irte *irte_info,
4226 struct irq_cfg *cfg);
4228 static int irq_remapping_activate(struct irq_domain *domain,
4229 struct irq_data *irq_data, bool reserve)
4231 struct amd_ir_data *data = irq_data->chip_data;
4232 struct irq_2_irte *irte_info = &data->irq_2_irte;
4233 struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4234 struct irq_cfg *cfg = irqd_cfg(irq_data);
4239 iommu->irte_ops->activate(data->entry, irte_info->devid,
4241 amd_ir_update_irte(irq_data, iommu, data, irte_info, cfg);
4245 static void irq_remapping_deactivate(struct irq_domain *domain,
4246 struct irq_data *irq_data)
4248 struct amd_ir_data *data = irq_data->chip_data;
4249 struct irq_2_irte *irte_info = &data->irq_2_irte;
4250 struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4253 iommu->irte_ops->deactivate(data->entry, irte_info->devid,
4257 static const struct irq_domain_ops amd_ir_domain_ops = {
4258 .alloc = irq_remapping_alloc,
4259 .free = irq_remapping_free,
4260 .activate = irq_remapping_activate,
4261 .deactivate = irq_remapping_deactivate,
4264 static int amd_ir_set_vcpu_affinity(struct irq_data *data, void *vcpu_info)
4266 struct amd_iommu *iommu;
4267 struct amd_iommu_pi_data *pi_data = vcpu_info;
4268 struct vcpu_data *vcpu_pi_info = pi_data->vcpu_data;
4269 struct amd_ir_data *ir_data = data->chip_data;
4270 struct irte_ga *irte = (struct irte_ga *) ir_data->entry;
4271 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
4272 struct iommu_dev_data *dev_data = search_dev_data(irte_info->devid);
4275 * This device has never been set up for guest mode.
4276 * we should not modify the IRTE
4278 if (!dev_data || !dev_data->use_vapic)
4281 pi_data->ir_data = ir_data;
4284 * SVM tries to set up for VAPIC mode, but we are in
4285 * legacy mode. So, we force legacy mode instead.
4287 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
4288 pr_debug("AMD-Vi: %s: Fall back to using intr legacy remap\n",
4290 pi_data->is_guest_mode = false;
4293 iommu = amd_iommu_rlookup_table[irte_info->devid];
4297 pi_data->prev_ga_tag = ir_data->cached_ga_tag;
4298 if (pi_data->is_guest_mode) {
4300 irte->hi.fields.ga_root_ptr = (pi_data->base >> 12);
4301 irte->hi.fields.vector = vcpu_pi_info->vector;
4302 irte->lo.fields_vapic.ga_log_intr = 1;
4303 irte->lo.fields_vapic.guest_mode = 1;
4304 irte->lo.fields_vapic.ga_tag = pi_data->ga_tag;
4306 ir_data->cached_ga_tag = pi_data->ga_tag;
4309 struct irq_cfg *cfg = irqd_cfg(data);
4313 irte->hi.fields.vector = cfg->vector;
4314 irte->lo.fields_remap.guest_mode = 0;
4315 irte->lo.fields_remap.destination = cfg->dest_apicid;
4316 irte->lo.fields_remap.int_type = apic->irq_delivery_mode;
4317 irte->lo.fields_remap.dm = apic->irq_dest_mode;
4320 * This communicates the ga_tag back to the caller
4321 * so that it can do all the necessary clean up.
4323 ir_data->cached_ga_tag = 0;
4326 return modify_irte_ga(irte_info->devid, irte_info->index, irte, ir_data);
4330 static void amd_ir_update_irte(struct irq_data *irqd, struct amd_iommu *iommu,
4331 struct amd_ir_data *ir_data,
4332 struct irq_2_irte *irte_info,
4333 struct irq_cfg *cfg)
4337 * Atomically updates the IRTE with the new destination, vector
4338 * and flushes the interrupt entry cache.
4340 iommu->irte_ops->set_affinity(ir_data->entry, irte_info->devid,
4341 irte_info->index, cfg->vector,
4345 static int amd_ir_set_affinity(struct irq_data *data,
4346 const struct cpumask *mask, bool force)
4348 struct amd_ir_data *ir_data = data->chip_data;
4349 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
4350 struct irq_cfg *cfg = irqd_cfg(data);
4351 struct irq_data *parent = data->parent_data;
4352 struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4358 ret = parent->chip->irq_set_affinity(parent, mask, force);
4359 if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
4362 amd_ir_update_irte(data, iommu, ir_data, irte_info, cfg);
4364 * After this point, all the interrupts will start arriving
4365 * at the new destination. So, time to cleanup the previous
4366 * vector allocation.
4368 send_cleanup_vector(cfg);
4370 return IRQ_SET_MASK_OK_DONE;
4373 static void ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg)
4375 struct amd_ir_data *ir_data = irq_data->chip_data;
4377 *msg = ir_data->msi_entry;
4380 static struct irq_chip amd_ir_chip = {
4382 .irq_ack = ir_ack_apic_edge,
4383 .irq_set_affinity = amd_ir_set_affinity,
4384 .irq_set_vcpu_affinity = amd_ir_set_vcpu_affinity,
4385 .irq_compose_msi_msg = ir_compose_msi_msg,
4388 int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
4390 struct fwnode_handle *fn;
4392 fn = irq_domain_alloc_named_id_fwnode("AMD-IR", iommu->index);
4395 iommu->ir_domain = irq_domain_create_tree(fn, &amd_ir_domain_ops, iommu);
4396 irq_domain_free_fwnode(fn);
4397 if (!iommu->ir_domain)
4400 iommu->ir_domain->parent = arch_get_ir_parent_domain();
4401 iommu->msi_domain = arch_create_remap_msi_irq_domain(iommu->ir_domain,
4407 int amd_iommu_update_ga(int cpu, bool is_run, void *data)
4409 unsigned long flags;
4410 struct amd_iommu *iommu;
4411 struct irq_remap_table *table;
4412 struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
4413 int devid = ir_data->irq_2_irte.devid;
4414 struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
4415 struct irte_ga *ref = (struct irte_ga *) ir_data->ref;
4417 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
4418 !ref || !entry || !entry->lo.fields_vapic.guest_mode)
4421 iommu = amd_iommu_rlookup_table[devid];
4425 table = get_irq_table(devid);
4429 raw_spin_lock_irqsave(&table->lock, flags);
4431 if (ref->lo.fields_vapic.guest_mode) {
4433 ref->lo.fields_vapic.destination = cpu;
4434 ref->lo.fields_vapic.is_run = is_run;
4438 raw_spin_unlock_irqrestore(&table->lock, flags);
4440 iommu_flush_irt(iommu, devid);
4441 iommu_completion_wait(iommu);
4444 EXPORT_SYMBOL(amd_iommu_update_ga);