2 * Copyright (C) 2007-2010 Advanced Micro Devices, Inc.
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/iommu-helper.h>
32 #include <linux/iommu.h>
33 #include <linux/delay.h>
34 #include <linux/amd-iommu.h>
35 #include <linux/notifier.h>
36 #include <linux/export.h>
37 #include <linux/irq.h>
38 #include <linux/msi.h>
39 #include <linux/dma-contiguous.h>
40 #include <linux/irqdomain.h>
41 #include <linux/percpu.h>
42 #include <linux/iova.h>
43 #include <asm/irq_remapping.h>
44 #include <asm/io_apic.h>
46 #include <asm/hw_irq.h>
47 #include <asm/msidef.h>
48 #include <asm/proto.h>
49 #include <asm/iommu.h>
53 #include "amd_iommu_proto.h"
54 #include "amd_iommu_types.h"
55 #include "irq_remapping.h"
57 #define AMD_IOMMU_MAPPING_ERROR 0
59 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
61 #define LOOP_TIMEOUT 100000
63 /* IO virtual address start page frame number */
64 #define IOVA_START_PFN (1)
65 #define IOVA_PFN(addr) ((addr) >> PAGE_SHIFT)
66 #define DMA_32BIT_PFN IOVA_PFN(DMA_BIT_MASK(32))
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_RWLOCK(amd_iommu_devtable_lock);
86 /* List of all available dev_data structures */
87 static LIST_HEAD(dev_data_list);
88 static DEFINE_SPINLOCK(dev_data_list_lock);
90 LIST_HEAD(ioapic_map);
92 LIST_HEAD(acpihid_map);
95 * Domain for untranslated devices - only allocated
96 * if iommu=pt passed on kernel cmd line.
98 const struct iommu_ops amd_iommu_ops;
100 static ATOMIC_NOTIFIER_HEAD(ppr_notifier);
101 int amd_iommu_max_glx_val = -1;
103 static const struct dma_map_ops amd_iommu_dma_ops;
106 * This struct contains device specific data for the IOMMU
108 struct iommu_dev_data {
109 struct list_head list; /* For domain->dev_list */
110 struct list_head dev_data_list; /* For global dev_data_list */
111 struct protection_domain *domain; /* Domain the device is bound to */
112 u16 devid; /* PCI Device ID */
113 u16 alias; /* Alias Device ID */
114 bool iommu_v2; /* Device can make use of IOMMUv2 */
115 bool passthrough; /* Device is identity mapped */
119 } ats; /* ATS state */
120 bool pri_tlp; /* PASID TLB required for
122 u32 errata; /* Bitmap for errata to apply */
123 bool use_vapic; /* Enable device to use vapic mode */
125 struct ratelimit_state rs; /* Ratelimit IOPF messages */
129 * general struct to manage commands send to an IOMMU
135 struct kmem_cache *amd_iommu_irq_cache;
137 static void update_domain(struct protection_domain *domain);
138 static int protection_domain_init(struct protection_domain *domain);
139 static void detach_device(struct device *dev);
140 static void iova_domain_flush_tlb(struct iova_domain *iovad);
143 * Data container for a dma_ops specific protection domain
145 struct dma_ops_domain {
146 /* generic protection domain information */
147 struct protection_domain domain;
150 struct iova_domain iovad;
153 static struct iova_domain reserved_iova_ranges;
154 static struct lock_class_key reserved_rbtree_key;
156 /****************************************************************************
160 ****************************************************************************/
162 static inline int match_hid_uid(struct device *dev,
163 struct acpihid_map_entry *entry)
165 const char *hid, *uid;
167 hid = acpi_device_hid(ACPI_COMPANION(dev));
168 uid = acpi_device_uid(ACPI_COMPANION(dev));
174 return strcmp(hid, entry->hid);
177 return strcmp(hid, entry->hid);
179 return (strcmp(hid, entry->hid) || strcmp(uid, entry->uid));
182 static inline u16 get_pci_device_id(struct device *dev)
184 struct pci_dev *pdev = to_pci_dev(dev);
186 return PCI_DEVID(pdev->bus->number, pdev->devfn);
189 static inline int get_acpihid_device_id(struct device *dev,
190 struct acpihid_map_entry **entry)
192 struct acpihid_map_entry *p;
194 list_for_each_entry(p, &acpihid_map, list) {
195 if (!match_hid_uid(dev, p)) {
204 static inline int get_device_id(struct device *dev)
209 devid = get_pci_device_id(dev);
211 devid = get_acpihid_device_id(dev, NULL);
216 static struct protection_domain *to_pdomain(struct iommu_domain *dom)
218 return container_of(dom, struct protection_domain, domain);
221 static struct dma_ops_domain* to_dma_ops_domain(struct protection_domain *domain)
223 BUG_ON(domain->flags != PD_DMA_OPS_MASK);
224 return container_of(domain, struct dma_ops_domain, domain);
227 static struct iommu_dev_data *alloc_dev_data(u16 devid)
229 struct iommu_dev_data *dev_data;
232 dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
236 dev_data->devid = devid;
238 spin_lock_irqsave(&dev_data_list_lock, flags);
239 list_add_tail(&dev_data->dev_data_list, &dev_data_list);
240 spin_unlock_irqrestore(&dev_data_list_lock, flags);
242 ratelimit_default_init(&dev_data->rs);
247 static struct iommu_dev_data *search_dev_data(u16 devid)
249 struct iommu_dev_data *dev_data;
252 spin_lock_irqsave(&dev_data_list_lock, flags);
253 list_for_each_entry(dev_data, &dev_data_list, dev_data_list) {
254 if (dev_data->devid == devid)
261 spin_unlock_irqrestore(&dev_data_list_lock, flags);
266 static int __last_alias(struct pci_dev *pdev, u16 alias, void *data)
268 *(u16 *)data = alias;
272 static u16 get_alias(struct device *dev)
274 struct pci_dev *pdev = to_pci_dev(dev);
275 u16 devid, ivrs_alias, pci_alias;
277 /* The callers make sure that get_device_id() does not fail here */
278 devid = get_device_id(dev);
279 ivrs_alias = amd_iommu_alias_table[devid];
280 pci_for_each_dma_alias(pdev, __last_alias, &pci_alias);
282 if (ivrs_alias == pci_alias)
288 * The IVRS is fairly reliable in telling us about aliases, but it
289 * can't know about every screwy device. If we don't have an IVRS
290 * reported alias, use the PCI reported alias. In that case we may
291 * still need to initialize the rlookup and dev_table entries if the
292 * alias is to a non-existent device.
294 if (ivrs_alias == devid) {
295 if (!amd_iommu_rlookup_table[pci_alias]) {
296 amd_iommu_rlookup_table[pci_alias] =
297 amd_iommu_rlookup_table[devid];
298 memcpy(amd_iommu_dev_table[pci_alias].data,
299 amd_iommu_dev_table[devid].data,
300 sizeof(amd_iommu_dev_table[pci_alias].data));
306 pr_info("AMD-Vi: Using IVRS reported alias %02x:%02x.%d "
307 "for device %s[%04x:%04x], kernel reported alias "
308 "%02x:%02x.%d\n", PCI_BUS_NUM(ivrs_alias), PCI_SLOT(ivrs_alias),
309 PCI_FUNC(ivrs_alias), dev_name(dev), pdev->vendor, pdev->device,
310 PCI_BUS_NUM(pci_alias), PCI_SLOT(pci_alias),
311 PCI_FUNC(pci_alias));
314 * If we don't have a PCI DMA alias and the IVRS alias is on the same
315 * bus, then the IVRS table may know about a quirk that we don't.
317 if (pci_alias == devid &&
318 PCI_BUS_NUM(ivrs_alias) == pdev->bus->number) {
319 pci_add_dma_alias(pdev, ivrs_alias & 0xff);
320 pr_info("AMD-Vi: Added PCI DMA alias %02x.%d for %s\n",
321 PCI_SLOT(ivrs_alias), PCI_FUNC(ivrs_alias),
328 static struct iommu_dev_data *find_dev_data(u16 devid)
330 struct iommu_dev_data *dev_data;
332 dev_data = search_dev_data(devid);
334 if (dev_data == NULL)
335 dev_data = alloc_dev_data(devid);
340 static struct iommu_dev_data *get_dev_data(struct device *dev)
342 return dev->archdata.iommu;
346 * Find or create an IOMMU group for a acpihid device.
348 static struct iommu_group *acpihid_device_group(struct device *dev)
350 struct acpihid_map_entry *p, *entry = NULL;
353 devid = get_acpihid_device_id(dev, &entry);
355 return ERR_PTR(devid);
357 list_for_each_entry(p, &acpihid_map, list) {
358 if ((devid == p->devid) && p->group)
359 entry->group = p->group;
363 entry->group = generic_device_group(dev);
365 iommu_group_ref_get(entry->group);
370 static bool pci_iommuv2_capable(struct pci_dev *pdev)
372 static const int caps[] = {
375 PCI_EXT_CAP_ID_PASID,
379 for (i = 0; i < 3; ++i) {
380 pos = pci_find_ext_capability(pdev, caps[i]);
388 static bool pdev_pri_erratum(struct pci_dev *pdev, u32 erratum)
390 struct iommu_dev_data *dev_data;
392 dev_data = get_dev_data(&pdev->dev);
394 return dev_data->errata & (1 << erratum) ? true : false;
398 * This function checks if the driver got a valid device from the caller to
399 * avoid dereferencing invalid pointers.
401 static bool check_device(struct device *dev)
405 if (!dev || !dev->dma_mask)
408 devid = get_device_id(dev);
412 /* Out of our scope? */
413 if (devid > amd_iommu_last_bdf)
416 if (amd_iommu_rlookup_table[devid] == NULL)
422 static void init_iommu_group(struct device *dev)
424 struct iommu_group *group;
426 group = iommu_group_get_for_dev(dev);
430 iommu_group_put(group);
433 static int iommu_init_device(struct device *dev)
435 struct iommu_dev_data *dev_data;
436 struct amd_iommu *iommu;
439 if (dev->archdata.iommu)
442 devid = get_device_id(dev);
446 iommu = amd_iommu_rlookup_table[devid];
448 dev_data = find_dev_data(devid);
452 dev_data->alias = get_alias(dev);
454 if (dev_is_pci(dev) && pci_iommuv2_capable(to_pci_dev(dev))) {
455 struct amd_iommu *iommu;
457 iommu = amd_iommu_rlookup_table[dev_data->devid];
458 dev_data->iommu_v2 = iommu->is_iommu_v2;
461 dev->archdata.iommu = dev_data;
463 iommu_device_link(&iommu->iommu, dev);
468 static void iommu_ignore_device(struct device *dev)
473 devid = get_device_id(dev);
477 alias = get_alias(dev);
479 memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
480 memset(&amd_iommu_dev_table[alias], 0, sizeof(struct dev_table_entry));
482 amd_iommu_rlookup_table[devid] = NULL;
483 amd_iommu_rlookup_table[alias] = NULL;
486 static void iommu_uninit_device(struct device *dev)
488 struct iommu_dev_data *dev_data;
489 struct amd_iommu *iommu;
492 devid = get_device_id(dev);
496 iommu = amd_iommu_rlookup_table[devid];
498 dev_data = search_dev_data(devid);
502 if (dev_data->domain)
505 iommu_device_unlink(&iommu->iommu, dev);
507 iommu_group_remove_device(dev);
513 * We keep dev_data around for unplugged devices and reuse it when the
514 * device is re-plugged - not doing so would introduce a ton of races.
518 /****************************************************************************
520 * Interrupt handling functions
522 ****************************************************************************/
524 static void dump_dte_entry(u16 devid)
528 for (i = 0; i < 4; ++i)
529 pr_err("AMD-Vi: DTE[%d]: %016llx\n", i,
530 amd_iommu_dev_table[devid].data[i]);
533 static void dump_command(unsigned long phys_addr)
535 struct iommu_cmd *cmd = phys_to_virt(phys_addr);
538 for (i = 0; i < 4; ++i)
539 pr_err("AMD-Vi: CMD[%d]: %08x\n", i, cmd->data[i]);
542 static void amd_iommu_report_page_fault(u16 devid, u16 domain_id,
543 u64 address, int flags)
545 struct iommu_dev_data *dev_data = NULL;
546 struct pci_dev *pdev;
548 pdev = pci_get_bus_and_slot(PCI_BUS_NUM(devid), devid & 0xff);
550 dev_data = get_dev_data(&pdev->dev);
552 if (dev_data && __ratelimit(&dev_data->rs)) {
553 dev_err(&pdev->dev, "AMD-Vi: Event logged [IO_PAGE_FAULT domain=0x%04x address=0x%016llx flags=0x%04x]\n",
554 domain_id, address, flags);
555 } else if (printk_ratelimit()) {
556 pr_err("AMD-Vi: Event logged [IO_PAGE_FAULT device=%02x:%02x.%x domain=0x%04x address=0x%016llx flags=0x%04x]\n",
557 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
558 domain_id, address, flags);
565 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
567 int type, devid, domid, flags;
568 volatile u32 *event = __evt;
573 type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
574 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
575 domid = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
576 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
577 address = (u64)(((u64)event[3]) << 32) | event[2];
580 /* Did we hit the erratum? */
581 if (++count == LOOP_TIMEOUT) {
582 pr_err("AMD-Vi: No event written to event log\n");
589 if (type == EVENT_TYPE_IO_FAULT) {
590 amd_iommu_report_page_fault(devid, domid, address, flags);
593 printk(KERN_ERR "AMD-Vi: Event logged [");
597 case EVENT_TYPE_ILL_DEV:
598 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
599 "address=0x%016llx flags=0x%04x]\n",
600 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
602 dump_dte_entry(devid);
604 case EVENT_TYPE_DEV_TAB_ERR:
605 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
606 "address=0x%016llx flags=0x%04x]\n",
607 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
610 case EVENT_TYPE_PAGE_TAB_ERR:
611 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
612 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
613 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
614 domid, address, flags);
616 case EVENT_TYPE_ILL_CMD:
617 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
618 dump_command(address);
620 case EVENT_TYPE_CMD_HARD_ERR:
621 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
622 "flags=0x%04x]\n", address, flags);
624 case EVENT_TYPE_IOTLB_INV_TO:
625 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
626 "address=0x%016llx]\n",
627 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
630 case EVENT_TYPE_INV_DEV_REQ:
631 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
632 "address=0x%016llx flags=0x%04x]\n",
633 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
637 printk(KERN_ERR "UNKNOWN type=0x%02x]\n", type);
640 memset(__evt, 0, 4 * sizeof(u32));
643 static void iommu_poll_events(struct amd_iommu *iommu)
647 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
648 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
650 while (head != tail) {
651 iommu_print_event(iommu, iommu->evt_buf + head);
652 head = (head + EVENT_ENTRY_SIZE) % EVT_BUFFER_SIZE;
655 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
658 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
660 struct amd_iommu_fault fault;
662 if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
663 pr_err_ratelimited("AMD-Vi: Unknown PPR request received\n");
667 fault.address = raw[1];
668 fault.pasid = PPR_PASID(raw[0]);
669 fault.device_id = PPR_DEVID(raw[0]);
670 fault.tag = PPR_TAG(raw[0]);
671 fault.flags = PPR_FLAGS(raw[0]);
673 atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
676 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
680 if (iommu->ppr_log == NULL)
683 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
684 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
686 while (head != tail) {
691 raw = (u64 *)(iommu->ppr_log + head);
694 * Hardware bug: Interrupt may arrive before the entry is
695 * written to memory. If this happens we need to wait for the
698 for (i = 0; i < LOOP_TIMEOUT; ++i) {
699 if (PPR_REQ_TYPE(raw[0]) != 0)
704 /* Avoid memcpy function-call overhead */
709 * To detect the hardware bug we need to clear the entry
712 raw[0] = raw[1] = 0UL;
714 /* Update head pointer of hardware ring-buffer */
715 head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
716 writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
718 /* Handle PPR entry */
719 iommu_handle_ppr_entry(iommu, entry);
721 /* Refresh ring-buffer information */
722 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
723 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
727 #ifdef CONFIG_IRQ_REMAP
728 static int (*iommu_ga_log_notifier)(u32);
730 int amd_iommu_register_ga_log_notifier(int (*notifier)(u32))
732 iommu_ga_log_notifier = notifier;
736 EXPORT_SYMBOL(amd_iommu_register_ga_log_notifier);
738 static void iommu_poll_ga_log(struct amd_iommu *iommu)
740 u32 head, tail, cnt = 0;
742 if (iommu->ga_log == NULL)
745 head = readl(iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
746 tail = readl(iommu->mmio_base + MMIO_GA_TAIL_OFFSET);
748 while (head != tail) {
752 raw = (u64 *)(iommu->ga_log + head);
755 /* Avoid memcpy function-call overhead */
758 /* Update head pointer of hardware ring-buffer */
759 head = (head + GA_ENTRY_SIZE) % GA_LOG_SIZE;
760 writel(head, iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
762 /* Handle GA entry */
763 switch (GA_REQ_TYPE(log_entry)) {
765 if (!iommu_ga_log_notifier)
768 pr_debug("AMD-Vi: %s: devid=%#x, ga_tag=%#x\n",
769 __func__, GA_DEVID(log_entry),
772 if (iommu_ga_log_notifier(GA_TAG(log_entry)) != 0)
773 pr_err("AMD-Vi: GA log notifier failed.\n");
780 #endif /* CONFIG_IRQ_REMAP */
782 #define AMD_IOMMU_INT_MASK \
783 (MMIO_STATUS_EVT_INT_MASK | \
784 MMIO_STATUS_PPR_INT_MASK | \
785 MMIO_STATUS_GALOG_INT_MASK)
787 irqreturn_t amd_iommu_int_thread(int irq, void *data)
789 struct amd_iommu *iommu = (struct amd_iommu *) data;
790 u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
792 while (status & AMD_IOMMU_INT_MASK) {
793 /* Enable EVT and PPR and GA interrupts again */
794 writel(AMD_IOMMU_INT_MASK,
795 iommu->mmio_base + MMIO_STATUS_OFFSET);
797 if (status & MMIO_STATUS_EVT_INT_MASK) {
798 pr_devel("AMD-Vi: Processing IOMMU Event Log\n");
799 iommu_poll_events(iommu);
802 if (status & MMIO_STATUS_PPR_INT_MASK) {
803 pr_devel("AMD-Vi: Processing IOMMU PPR Log\n");
804 iommu_poll_ppr_log(iommu);
807 #ifdef CONFIG_IRQ_REMAP
808 if (status & MMIO_STATUS_GALOG_INT_MASK) {
809 pr_devel("AMD-Vi: Processing IOMMU GA Log\n");
810 iommu_poll_ga_log(iommu);
815 * Hardware bug: ERBT1312
816 * When re-enabling interrupt (by writing 1
817 * to clear the bit), the hardware might also try to set
818 * the interrupt bit in the event status register.
819 * In this scenario, the bit will be set, and disable
820 * subsequent interrupts.
822 * Workaround: The IOMMU driver should read back the
823 * status register and check if the interrupt bits are cleared.
824 * If not, driver will need to go through the interrupt handler
825 * again and re-clear the bits
827 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
832 irqreturn_t amd_iommu_int_handler(int irq, void *data)
834 return IRQ_WAKE_THREAD;
837 /****************************************************************************
839 * IOMMU command queuing functions
841 ****************************************************************************/
843 static int wait_on_sem(volatile u64 *sem)
847 while (*sem == 0 && i < LOOP_TIMEOUT) {
852 if (i == LOOP_TIMEOUT) {
853 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
860 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
861 struct iommu_cmd *cmd)
865 target = iommu->cmd_buf + iommu->cmd_buf_tail;
867 iommu->cmd_buf_tail += sizeof(*cmd);
868 iommu->cmd_buf_tail %= CMD_BUFFER_SIZE;
870 /* Copy command to buffer */
871 memcpy(target, cmd, sizeof(*cmd));
873 /* Tell the IOMMU about it */
874 writel(iommu->cmd_buf_tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
877 static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
879 WARN_ON(address & 0x7ULL);
881 memset(cmd, 0, sizeof(*cmd));
882 cmd->data[0] = lower_32_bits(__pa(address)) | CMD_COMPL_WAIT_STORE_MASK;
883 cmd->data[1] = upper_32_bits(__pa(address));
885 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
888 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
890 memset(cmd, 0, sizeof(*cmd));
891 cmd->data[0] = devid;
892 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
895 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
896 size_t size, u16 domid, int pde)
901 pages = iommu_num_pages(address, size, PAGE_SIZE);
906 * If we have to flush more than one page, flush all
907 * TLB entries for this domain
909 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
913 address &= PAGE_MASK;
915 memset(cmd, 0, sizeof(*cmd));
916 cmd->data[1] |= domid;
917 cmd->data[2] = lower_32_bits(address);
918 cmd->data[3] = upper_32_bits(address);
919 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
920 if (s) /* size bit - we flush more than one 4kb page */
921 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
922 if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
923 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
926 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
927 u64 address, size_t size)
932 pages = iommu_num_pages(address, size, PAGE_SIZE);
937 * If we have to flush more than one page, flush all
938 * TLB entries for this domain
940 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
944 address &= PAGE_MASK;
946 memset(cmd, 0, sizeof(*cmd));
947 cmd->data[0] = devid;
948 cmd->data[0] |= (qdep & 0xff) << 24;
949 cmd->data[1] = devid;
950 cmd->data[2] = lower_32_bits(address);
951 cmd->data[3] = upper_32_bits(address);
952 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
954 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
957 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, int pasid,
958 u64 address, bool size)
960 memset(cmd, 0, sizeof(*cmd));
962 address &= ~(0xfffULL);
964 cmd->data[0] = pasid;
965 cmd->data[1] = domid;
966 cmd->data[2] = lower_32_bits(address);
967 cmd->data[3] = upper_32_bits(address);
968 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
969 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
971 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
972 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
975 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, int pasid,
976 int qdep, u64 address, bool size)
978 memset(cmd, 0, sizeof(*cmd));
980 address &= ~(0xfffULL);
982 cmd->data[0] = devid;
983 cmd->data[0] |= ((pasid >> 8) & 0xff) << 16;
984 cmd->data[0] |= (qdep & 0xff) << 24;
985 cmd->data[1] = devid;
986 cmd->data[1] |= (pasid & 0xff) << 16;
987 cmd->data[2] = lower_32_bits(address);
988 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
989 cmd->data[3] = upper_32_bits(address);
991 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
992 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
995 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, int pasid,
996 int status, int tag, bool gn)
998 memset(cmd, 0, sizeof(*cmd));
1000 cmd->data[0] = devid;
1002 cmd->data[1] = pasid;
1003 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
1005 cmd->data[3] = tag & 0x1ff;
1006 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
1008 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
1011 static void build_inv_all(struct iommu_cmd *cmd)
1013 memset(cmd, 0, sizeof(*cmd));
1014 CMD_SET_TYPE(cmd, CMD_INV_ALL);
1017 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
1019 memset(cmd, 0, sizeof(*cmd));
1020 cmd->data[0] = devid;
1021 CMD_SET_TYPE(cmd, CMD_INV_IRT);
1025 * Writes the command to the IOMMUs command buffer and informs the
1026 * hardware about the new command.
1028 static int __iommu_queue_command_sync(struct amd_iommu *iommu,
1029 struct iommu_cmd *cmd,
1032 unsigned int count = 0;
1033 u32 left, next_tail;
1035 next_tail = (iommu->cmd_buf_tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
1037 left = (iommu->cmd_buf_head - next_tail) % CMD_BUFFER_SIZE;
1040 /* Skip udelay() the first time around */
1042 if (count == LOOP_TIMEOUT) {
1043 pr_err("AMD-Vi: Command buffer timeout\n");
1050 /* Update head and recheck remaining space */
1051 iommu->cmd_buf_head = readl(iommu->mmio_base +
1052 MMIO_CMD_HEAD_OFFSET);
1057 copy_cmd_to_buffer(iommu, cmd);
1059 /* Do we need to make sure all commands are processed? */
1060 iommu->need_sync = sync;
1065 static int iommu_queue_command_sync(struct amd_iommu *iommu,
1066 struct iommu_cmd *cmd,
1069 unsigned long flags;
1072 spin_lock_irqsave(&iommu->lock, flags);
1073 ret = __iommu_queue_command_sync(iommu, cmd, sync);
1074 spin_unlock_irqrestore(&iommu->lock, flags);
1079 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
1081 return iommu_queue_command_sync(iommu, cmd, true);
1085 * This function queues a completion wait command into the command
1086 * buffer of an IOMMU
1088 static int iommu_completion_wait(struct amd_iommu *iommu)
1090 struct iommu_cmd cmd;
1091 unsigned long flags;
1094 if (!iommu->need_sync)
1098 build_completion_wait(&cmd, (u64)&iommu->cmd_sem);
1100 spin_lock_irqsave(&iommu->lock, flags);
1104 ret = __iommu_queue_command_sync(iommu, &cmd, false);
1108 ret = wait_on_sem(&iommu->cmd_sem);
1111 spin_unlock_irqrestore(&iommu->lock, flags);
1116 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
1118 struct iommu_cmd cmd;
1120 build_inv_dte(&cmd, devid);
1122 return iommu_queue_command(iommu, &cmd);
1125 static void amd_iommu_flush_dte_all(struct amd_iommu *iommu)
1129 for (devid = 0; devid <= 0xffff; ++devid)
1130 iommu_flush_dte(iommu, devid);
1132 iommu_completion_wait(iommu);
1136 * This function uses heavy locking and may disable irqs for some time. But
1137 * this is no issue because it is only called during resume.
1139 static void amd_iommu_flush_tlb_all(struct amd_iommu *iommu)
1143 for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
1144 struct iommu_cmd cmd;
1145 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1147 iommu_queue_command(iommu, &cmd);
1150 iommu_completion_wait(iommu);
1153 static void amd_iommu_flush_all(struct amd_iommu *iommu)
1155 struct iommu_cmd cmd;
1157 build_inv_all(&cmd);
1159 iommu_queue_command(iommu, &cmd);
1160 iommu_completion_wait(iommu);
1163 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1165 struct iommu_cmd cmd;
1167 build_inv_irt(&cmd, devid);
1169 iommu_queue_command(iommu, &cmd);
1172 static void amd_iommu_flush_irt_all(struct amd_iommu *iommu)
1176 for (devid = 0; devid <= MAX_DEV_TABLE_ENTRIES; devid++)
1177 iommu_flush_irt(iommu, devid);
1179 iommu_completion_wait(iommu);
1182 void iommu_flush_all_caches(struct amd_iommu *iommu)
1184 if (iommu_feature(iommu, FEATURE_IA)) {
1185 amd_iommu_flush_all(iommu);
1187 amd_iommu_flush_dte_all(iommu);
1188 amd_iommu_flush_irt_all(iommu);
1189 amd_iommu_flush_tlb_all(iommu);
1194 * Command send function for flushing on-device TLB
1196 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1197 u64 address, size_t size)
1199 struct amd_iommu *iommu;
1200 struct iommu_cmd cmd;
1203 qdep = dev_data->ats.qdep;
1204 iommu = amd_iommu_rlookup_table[dev_data->devid];
1206 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1208 return iommu_queue_command(iommu, &cmd);
1212 * Command send function for invalidating a device table entry
1214 static int device_flush_dte(struct iommu_dev_data *dev_data)
1216 struct amd_iommu *iommu;
1220 iommu = amd_iommu_rlookup_table[dev_data->devid];
1221 alias = dev_data->alias;
1223 ret = iommu_flush_dte(iommu, dev_data->devid);
1224 if (!ret && alias != dev_data->devid)
1225 ret = iommu_flush_dte(iommu, alias);
1229 if (dev_data->ats.enabled)
1230 ret = device_flush_iotlb(dev_data, 0, ~0UL);
1236 * TLB invalidation function which is called from the mapping functions.
1237 * It invalidates a single PTE if the range to flush is within a single
1238 * page. Otherwise it flushes the whole TLB of the IOMMU.
1240 static void __domain_flush_pages(struct protection_domain *domain,
1241 u64 address, size_t size, int pde)
1243 struct iommu_dev_data *dev_data;
1244 struct iommu_cmd cmd;
1247 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1249 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
1250 if (!domain->dev_iommu[i])
1254 * Devices of this domain are behind this IOMMU
1255 * We need a TLB flush
1257 ret |= iommu_queue_command(amd_iommus[i], &cmd);
1260 list_for_each_entry(dev_data, &domain->dev_list, list) {
1262 if (!dev_data->ats.enabled)
1265 ret |= device_flush_iotlb(dev_data, address, size);
1271 static void domain_flush_pages(struct protection_domain *domain,
1272 u64 address, size_t size)
1274 __domain_flush_pages(domain, address, size, 0);
1277 /* Flush the whole IO/TLB for a given protection domain */
1278 static void domain_flush_tlb(struct protection_domain *domain)
1280 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
1283 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1284 static void domain_flush_tlb_pde(struct protection_domain *domain)
1286 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1289 static void domain_flush_complete(struct protection_domain *domain)
1293 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
1294 if (domain && !domain->dev_iommu[i])
1298 * Devices of this domain are behind this IOMMU
1299 * We need to wait for completion of all commands.
1301 iommu_completion_wait(amd_iommus[i]);
1307 * This function flushes the DTEs for all devices in domain
1309 static void domain_flush_devices(struct protection_domain *domain)
1311 struct iommu_dev_data *dev_data;
1313 list_for_each_entry(dev_data, &domain->dev_list, list)
1314 device_flush_dte(dev_data);
1317 /****************************************************************************
1319 * The functions below are used the create the page table mappings for
1320 * unity mapped regions.
1322 ****************************************************************************/
1325 * This function is used to add another level to an IO page table. Adding
1326 * another level increases the size of the address space by 9 bits to a size up
1329 static bool increase_address_space(struct protection_domain *domain,
1334 if (domain->mode == PAGE_MODE_6_LEVEL)
1335 /* address space already 64 bit large */
1338 pte = (void *)get_zeroed_page(gfp);
1342 *pte = PM_LEVEL_PDE(domain->mode,
1343 virt_to_phys(domain->pt_root));
1344 domain->pt_root = pte;
1346 domain->updated = true;
1351 static u64 *alloc_pte(struct protection_domain *domain,
1352 unsigned long address,
1353 unsigned long page_size,
1360 BUG_ON(!is_power_of_2(page_size));
1362 while (address > PM_LEVEL_SIZE(domain->mode))
1363 increase_address_space(domain, gfp);
1365 level = domain->mode - 1;
1366 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1367 address = PAGE_SIZE_ALIGN(address, page_size);
1368 end_lvl = PAGE_SIZE_LEVEL(page_size);
1370 while (level > end_lvl) {
1375 if (!IOMMU_PTE_PRESENT(__pte)) {
1376 page = (u64 *)get_zeroed_page(gfp);
1380 __npte = PM_LEVEL_PDE(level, virt_to_phys(page));
1382 /* pte could have been changed somewhere. */
1383 if (cmpxchg64(pte, __pte, __npte) != __pte) {
1384 free_page((unsigned long)page);
1389 /* No level skipping support yet */
1390 if (PM_PTE_LEVEL(*pte) != level)
1395 pte = IOMMU_PTE_PAGE(*pte);
1397 if (pte_page && level == end_lvl)
1400 pte = &pte[PM_LEVEL_INDEX(level, address)];
1407 * This function checks if there is a PTE for a given dma address. If
1408 * there is one, it returns the pointer to it.
1410 static u64 *fetch_pte(struct protection_domain *domain,
1411 unsigned long address,
1412 unsigned long *page_size)
1417 if (address > PM_LEVEL_SIZE(domain->mode))
1420 level = domain->mode - 1;
1421 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1422 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1427 if (!IOMMU_PTE_PRESENT(*pte))
1431 if (PM_PTE_LEVEL(*pte) == 7 ||
1432 PM_PTE_LEVEL(*pte) == 0)
1435 /* No level skipping support yet */
1436 if (PM_PTE_LEVEL(*pte) != level)
1441 /* Walk to the next level */
1442 pte = IOMMU_PTE_PAGE(*pte);
1443 pte = &pte[PM_LEVEL_INDEX(level, address)];
1444 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1447 if (PM_PTE_LEVEL(*pte) == 0x07) {
1448 unsigned long pte_mask;
1451 * If we have a series of large PTEs, make
1452 * sure to return a pointer to the first one.
1454 *page_size = pte_mask = PTE_PAGE_SIZE(*pte);
1455 pte_mask = ~((PAGE_SIZE_PTE_COUNT(pte_mask) << 3) - 1);
1456 pte = (u64 *)(((unsigned long)pte) & pte_mask);
1463 * Generic mapping functions. It maps a physical address into a DMA
1464 * address space. It allocates the page table pages if necessary.
1465 * In the future it can be extended to a generic mapping function
1466 * supporting all features of AMD IOMMU page tables like level skipping
1467 * and full 64 bit address spaces.
1469 static int iommu_map_page(struct protection_domain *dom,
1470 unsigned long bus_addr,
1471 unsigned long phys_addr,
1472 unsigned long page_size,
1479 BUG_ON(!IS_ALIGNED(bus_addr, page_size));
1480 BUG_ON(!IS_ALIGNED(phys_addr, page_size));
1482 if (!(prot & IOMMU_PROT_MASK))
1485 count = PAGE_SIZE_PTE_COUNT(page_size);
1486 pte = alloc_pte(dom, bus_addr, page_size, NULL, gfp);
1491 for (i = 0; i < count; ++i)
1492 if (IOMMU_PTE_PRESENT(pte[i]))
1496 __pte = PAGE_SIZE_PTE(phys_addr, page_size);
1497 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_P | IOMMU_PTE_FC;
1499 __pte = phys_addr | IOMMU_PTE_P | IOMMU_PTE_FC;
1501 if (prot & IOMMU_PROT_IR)
1502 __pte |= IOMMU_PTE_IR;
1503 if (prot & IOMMU_PROT_IW)
1504 __pte |= IOMMU_PTE_IW;
1506 for (i = 0; i < count; ++i)
1514 static unsigned long iommu_unmap_page(struct protection_domain *dom,
1515 unsigned long bus_addr,
1516 unsigned long page_size)
1518 unsigned long long unmapped;
1519 unsigned long unmap_size;
1522 BUG_ON(!is_power_of_2(page_size));
1526 while (unmapped < page_size) {
1528 pte = fetch_pte(dom, bus_addr, &unmap_size);
1533 count = PAGE_SIZE_PTE_COUNT(unmap_size);
1534 for (i = 0; i < count; i++)
1538 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1539 unmapped += unmap_size;
1542 BUG_ON(unmapped && !is_power_of_2(unmapped));
1547 /****************************************************************************
1549 * The next functions belong to the address allocator for the dma_ops
1550 * interface functions.
1552 ****************************************************************************/
1555 static unsigned long dma_ops_alloc_iova(struct device *dev,
1556 struct dma_ops_domain *dma_dom,
1557 unsigned int pages, u64 dma_mask)
1559 unsigned long pfn = 0;
1561 pages = __roundup_pow_of_two(pages);
1563 if (dma_mask > DMA_BIT_MASK(32))
1564 pfn = alloc_iova_fast(&dma_dom->iovad, pages,
1565 IOVA_PFN(DMA_BIT_MASK(32)));
1568 pfn = alloc_iova_fast(&dma_dom->iovad, pages, IOVA_PFN(dma_mask));
1570 return (pfn << PAGE_SHIFT);
1573 static void dma_ops_free_iova(struct dma_ops_domain *dma_dom,
1574 unsigned long address,
1577 pages = __roundup_pow_of_two(pages);
1578 address >>= PAGE_SHIFT;
1580 free_iova_fast(&dma_dom->iovad, address, pages);
1583 /****************************************************************************
1585 * The next functions belong to the domain allocation. A domain is
1586 * allocated for every IOMMU as the default domain. If device isolation
1587 * is enabled, every device get its own domain. The most important thing
1588 * about domains is the page table mapping the DMA address space they
1591 ****************************************************************************/
1594 * This function adds a protection domain to the global protection domain list
1596 static void add_domain_to_list(struct protection_domain *domain)
1598 unsigned long flags;
1600 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1601 list_add(&domain->list, &amd_iommu_pd_list);
1602 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1606 * This function removes a protection domain to the global
1607 * protection domain list
1609 static void del_domain_from_list(struct protection_domain *domain)
1611 unsigned long flags;
1613 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1614 list_del(&domain->list);
1615 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1618 static u16 domain_id_alloc(void)
1620 unsigned long flags;
1623 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1624 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1626 if (id > 0 && id < MAX_DOMAIN_ID)
1627 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1630 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1635 static void domain_id_free(int id)
1637 unsigned long flags;
1639 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1640 if (id > 0 && id < MAX_DOMAIN_ID)
1641 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1642 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1645 #define DEFINE_FREE_PT_FN(LVL, FN) \
1646 static void free_pt_##LVL (unsigned long __pt) \
1654 for (i = 0; i < 512; ++i) { \
1655 /* PTE present? */ \
1656 if (!IOMMU_PTE_PRESENT(pt[i])) \
1660 if (PM_PTE_LEVEL(pt[i]) == 0 || \
1661 PM_PTE_LEVEL(pt[i]) == 7) \
1664 p = (unsigned long)IOMMU_PTE_PAGE(pt[i]); \
1667 free_page((unsigned long)pt); \
1670 DEFINE_FREE_PT_FN(l2, free_page)
1671 DEFINE_FREE_PT_FN(l3, free_pt_l2)
1672 DEFINE_FREE_PT_FN(l4, free_pt_l3)
1673 DEFINE_FREE_PT_FN(l5, free_pt_l4)
1674 DEFINE_FREE_PT_FN(l6, free_pt_l5)
1676 static void free_pagetable(struct protection_domain *domain)
1678 unsigned long root = (unsigned long)domain->pt_root;
1680 switch (domain->mode) {
1681 case PAGE_MODE_NONE:
1683 case PAGE_MODE_1_LEVEL:
1686 case PAGE_MODE_2_LEVEL:
1689 case PAGE_MODE_3_LEVEL:
1692 case PAGE_MODE_4_LEVEL:
1695 case PAGE_MODE_5_LEVEL:
1698 case PAGE_MODE_6_LEVEL:
1706 static void free_gcr3_tbl_level1(u64 *tbl)
1711 for (i = 0; i < 512; ++i) {
1712 if (!(tbl[i] & GCR3_VALID))
1715 ptr = __va(tbl[i] & PAGE_MASK);
1717 free_page((unsigned long)ptr);
1721 static void free_gcr3_tbl_level2(u64 *tbl)
1726 for (i = 0; i < 512; ++i) {
1727 if (!(tbl[i] & GCR3_VALID))
1730 ptr = __va(tbl[i] & PAGE_MASK);
1732 free_gcr3_tbl_level1(ptr);
1736 static void free_gcr3_table(struct protection_domain *domain)
1738 if (domain->glx == 2)
1739 free_gcr3_tbl_level2(domain->gcr3_tbl);
1740 else if (domain->glx == 1)
1741 free_gcr3_tbl_level1(domain->gcr3_tbl);
1743 BUG_ON(domain->glx != 0);
1745 free_page((unsigned long)domain->gcr3_tbl);
1748 static void dma_ops_domain_flush_tlb(struct dma_ops_domain *dom)
1750 domain_flush_tlb(&dom->domain);
1751 domain_flush_complete(&dom->domain);
1754 static void iova_domain_flush_tlb(struct iova_domain *iovad)
1756 struct dma_ops_domain *dom;
1758 dom = container_of(iovad, struct dma_ops_domain, iovad);
1760 dma_ops_domain_flush_tlb(dom);
1764 * Free a domain, only used if something went wrong in the
1765 * allocation path and we need to free an already allocated page table
1767 static void dma_ops_domain_free(struct dma_ops_domain *dom)
1772 del_domain_from_list(&dom->domain);
1774 put_iova_domain(&dom->iovad);
1776 free_pagetable(&dom->domain);
1779 domain_id_free(dom->domain.id);
1785 * Allocates a new protection domain usable for the dma_ops functions.
1786 * It also initializes the page table and the address allocator data
1787 * structures required for the dma_ops interface
1789 static struct dma_ops_domain *dma_ops_domain_alloc(void)
1791 struct dma_ops_domain *dma_dom;
1793 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
1797 if (protection_domain_init(&dma_dom->domain))
1800 dma_dom->domain.mode = PAGE_MODE_3_LEVEL;
1801 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
1802 dma_dom->domain.flags = PD_DMA_OPS_MASK;
1803 if (!dma_dom->domain.pt_root)
1806 init_iova_domain(&dma_dom->iovad, PAGE_SIZE,
1807 IOVA_START_PFN, DMA_32BIT_PFN);
1809 if (init_iova_flush_queue(&dma_dom->iovad, iova_domain_flush_tlb, NULL))
1812 /* Initialize reserved ranges */
1813 copy_reserved_iova(&reserved_iova_ranges, &dma_dom->iovad);
1815 add_domain_to_list(&dma_dom->domain);
1820 dma_ops_domain_free(dma_dom);
1826 * little helper function to check whether a given protection domain is a
1829 static bool dma_ops_domain(struct protection_domain *domain)
1831 return domain->flags & PD_DMA_OPS_MASK;
1834 static void set_dte_entry(u16 devid, struct protection_domain *domain, bool ats)
1839 if (domain->mode != PAGE_MODE_NONE)
1840 pte_root = virt_to_phys(domain->pt_root);
1842 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
1843 << DEV_ENTRY_MODE_SHIFT;
1844 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;
1846 flags = amd_iommu_dev_table[devid].data[1];
1849 flags |= DTE_FLAG_IOTLB;
1851 if (domain->flags & PD_IOMMUV2_MASK) {
1852 u64 gcr3 = __pa(domain->gcr3_tbl);
1853 u64 glx = domain->glx;
1856 pte_root |= DTE_FLAG_GV;
1857 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
1859 /* First mask out possible old values for GCR3 table */
1860 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
1863 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
1866 /* Encode GCR3 table into DTE */
1867 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
1870 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
1873 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
1878 flags &= ~(DTE_FLAG_SA | 0xffffULL);
1879 flags |= domain->id;
1881 amd_iommu_dev_table[devid].data[1] = flags;
1882 amd_iommu_dev_table[devid].data[0] = pte_root;
1885 static void clear_dte_entry(u16 devid)
1887 /* remove entry from the device table seen by the hardware */
1888 amd_iommu_dev_table[devid].data[0] = IOMMU_PTE_P | IOMMU_PTE_TV;
1889 amd_iommu_dev_table[devid].data[1] &= DTE_FLAG_MASK;
1891 amd_iommu_apply_erratum_63(devid);
1894 static void do_attach(struct iommu_dev_data *dev_data,
1895 struct protection_domain *domain)
1897 struct amd_iommu *iommu;
1901 iommu = amd_iommu_rlookup_table[dev_data->devid];
1902 alias = dev_data->alias;
1903 ats = dev_data->ats.enabled;
1905 /* Update data structures */
1906 dev_data->domain = domain;
1907 list_add(&dev_data->list, &domain->dev_list);
1909 /* Do reference counting */
1910 domain->dev_iommu[iommu->index] += 1;
1911 domain->dev_cnt += 1;
1913 /* Update device table */
1914 set_dte_entry(dev_data->devid, domain, ats);
1915 if (alias != dev_data->devid)
1916 set_dte_entry(alias, domain, ats);
1918 device_flush_dte(dev_data);
1921 static void do_detach(struct iommu_dev_data *dev_data)
1923 struct amd_iommu *iommu;
1927 * First check if the device is still attached. It might already
1928 * be detached from its domain because the generic
1929 * iommu_detach_group code detached it and we try again here in
1930 * our alias handling.
1932 if (!dev_data->domain)
1935 iommu = amd_iommu_rlookup_table[dev_data->devid];
1936 alias = dev_data->alias;
1938 /* decrease reference counters */
1939 dev_data->domain->dev_iommu[iommu->index] -= 1;
1940 dev_data->domain->dev_cnt -= 1;
1942 /* Update data structures */
1943 dev_data->domain = NULL;
1944 list_del(&dev_data->list);
1945 clear_dte_entry(dev_data->devid);
1946 if (alias != dev_data->devid)
1947 clear_dte_entry(alias);
1949 /* Flush the DTE entry */
1950 device_flush_dte(dev_data);
1954 * If a device is not yet associated with a domain, this function does
1955 * assigns it visible for the hardware
1957 static int __attach_device(struct iommu_dev_data *dev_data,
1958 struct protection_domain *domain)
1963 * Must be called with IRQs disabled. Warn here to detect early
1966 WARN_ON(!irqs_disabled());
1969 spin_lock(&domain->lock);
1972 if (dev_data->domain != NULL)
1975 /* Attach alias group root */
1976 do_attach(dev_data, domain);
1983 spin_unlock(&domain->lock);
1989 static void pdev_iommuv2_disable(struct pci_dev *pdev)
1991 pci_disable_ats(pdev);
1992 pci_disable_pri(pdev);
1993 pci_disable_pasid(pdev);
1996 /* FIXME: Change generic reset-function to do the same */
1997 static int pri_reset_while_enabled(struct pci_dev *pdev)
2002 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2006 pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
2007 control |= PCI_PRI_CTRL_RESET;
2008 pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
2013 static int pdev_iommuv2_enable(struct pci_dev *pdev)
2018 /* FIXME: Hardcode number of outstanding requests for now */
2020 if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
2022 reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
2024 /* Only allow access to user-accessible pages */
2025 ret = pci_enable_pasid(pdev, 0);
2029 /* First reset the PRI state of the device */
2030 ret = pci_reset_pri(pdev);
2035 ret = pci_enable_pri(pdev, reqs);
2040 ret = pri_reset_while_enabled(pdev);
2045 ret = pci_enable_ats(pdev, PAGE_SHIFT);
2052 pci_disable_pri(pdev);
2053 pci_disable_pasid(pdev);
2058 /* FIXME: Move this to PCI code */
2059 #define PCI_PRI_TLP_OFF (1 << 15)
2061 static bool pci_pri_tlp_required(struct pci_dev *pdev)
2066 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2070 pci_read_config_word(pdev, pos + PCI_PRI_STATUS, &status);
2072 return (status & PCI_PRI_TLP_OFF) ? true : false;
2076 * If a device is not yet associated with a domain, this function
2077 * assigns it visible for the hardware
2079 static int attach_device(struct device *dev,
2080 struct protection_domain *domain)
2082 struct pci_dev *pdev;
2083 struct iommu_dev_data *dev_data;
2084 unsigned long flags;
2087 dev_data = get_dev_data(dev);
2089 if (!dev_is_pci(dev))
2090 goto skip_ats_check;
2092 pdev = to_pci_dev(dev);
2093 if (domain->flags & PD_IOMMUV2_MASK) {
2094 if (!dev_data->passthrough)
2097 if (dev_data->iommu_v2) {
2098 if (pdev_iommuv2_enable(pdev) != 0)
2101 dev_data->ats.enabled = true;
2102 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2103 dev_data->pri_tlp = pci_pri_tlp_required(pdev);
2105 } else if (amd_iommu_iotlb_sup &&
2106 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
2107 dev_data->ats.enabled = true;
2108 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2112 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2113 ret = __attach_device(dev_data, domain);
2114 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2117 * We might boot into a crash-kernel here. The crashed kernel
2118 * left the caches in the IOMMU dirty. So we have to flush
2119 * here to evict all dirty stuff.
2121 domain_flush_tlb_pde(domain);
2127 * Removes a device from a protection domain (unlocked)
2129 static void __detach_device(struct iommu_dev_data *dev_data)
2131 struct protection_domain *domain;
2134 * Must be called with IRQs disabled. Warn here to detect early
2137 WARN_ON(!irqs_disabled());
2139 if (WARN_ON(!dev_data->domain))
2142 domain = dev_data->domain;
2144 spin_lock(&domain->lock);
2146 do_detach(dev_data);
2148 spin_unlock(&domain->lock);
2152 * Removes a device from a protection domain (with devtable_lock held)
2154 static void detach_device(struct device *dev)
2156 struct protection_domain *domain;
2157 struct iommu_dev_data *dev_data;
2158 unsigned long flags;
2160 dev_data = get_dev_data(dev);
2161 domain = dev_data->domain;
2163 /* lock device table */
2164 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2165 __detach_device(dev_data);
2166 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2168 if (!dev_is_pci(dev))
2171 if (domain->flags & PD_IOMMUV2_MASK && dev_data->iommu_v2)
2172 pdev_iommuv2_disable(to_pci_dev(dev));
2173 else if (dev_data->ats.enabled)
2174 pci_disable_ats(to_pci_dev(dev));
2176 dev_data->ats.enabled = false;
2179 static int amd_iommu_add_device(struct device *dev)
2181 struct iommu_dev_data *dev_data;
2182 struct iommu_domain *domain;
2183 struct amd_iommu *iommu;
2186 if (!check_device(dev) || get_dev_data(dev))
2189 devid = get_device_id(dev);
2193 iommu = amd_iommu_rlookup_table[devid];
2195 ret = iommu_init_device(dev);
2197 if (ret != -ENOTSUPP)
2198 pr_err("Failed to initialize device %s - trying to proceed anyway\n",
2201 iommu_ignore_device(dev);
2202 dev->dma_ops = &nommu_dma_ops;
2205 init_iommu_group(dev);
2207 dev_data = get_dev_data(dev);
2211 if (iommu_pass_through || dev_data->iommu_v2)
2212 iommu_request_dm_for_dev(dev);
2214 /* Domains are initialized for this device - have a look what we ended up with */
2215 domain = iommu_get_domain_for_dev(dev);
2216 if (domain->type == IOMMU_DOMAIN_IDENTITY)
2217 dev_data->passthrough = true;
2219 dev->dma_ops = &amd_iommu_dma_ops;
2222 iommu_completion_wait(iommu);
2227 static void amd_iommu_remove_device(struct device *dev)
2229 struct amd_iommu *iommu;
2232 if (!check_device(dev))
2235 devid = get_device_id(dev);
2239 iommu = amd_iommu_rlookup_table[devid];
2241 iommu_uninit_device(dev);
2242 iommu_completion_wait(iommu);
2245 static struct iommu_group *amd_iommu_device_group(struct device *dev)
2247 if (dev_is_pci(dev))
2248 return pci_device_group(dev);
2250 return acpihid_device_group(dev);
2253 /*****************************************************************************
2255 * The next functions belong to the dma_ops mapping/unmapping code.
2257 *****************************************************************************/
2260 * In the dma_ops path we only have the struct device. This function
2261 * finds the corresponding IOMMU, the protection domain and the
2262 * requestor id for a given device.
2263 * If the device is not yet associated with a domain this is also done
2266 static struct protection_domain *get_domain(struct device *dev)
2268 struct protection_domain *domain;
2270 if (!check_device(dev))
2271 return ERR_PTR(-EINVAL);
2273 domain = get_dev_data(dev)->domain;
2274 if (!dma_ops_domain(domain))
2275 return ERR_PTR(-EBUSY);
2280 static void update_device_table(struct protection_domain *domain)
2282 struct iommu_dev_data *dev_data;
2284 list_for_each_entry(dev_data, &domain->dev_list, list) {
2285 set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled);
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);
2295 static void update_domain(struct protection_domain *domain)
2297 if (!domain->updated)
2300 update_device_table(domain);
2302 domain_flush_devices(domain);
2303 domain_flush_tlb_pde(domain);
2305 domain->updated = false;
2308 static int dir2prot(enum dma_data_direction direction)
2310 if (direction == DMA_TO_DEVICE)
2311 return IOMMU_PROT_IR;
2312 else if (direction == DMA_FROM_DEVICE)
2313 return IOMMU_PROT_IW;
2314 else if (direction == DMA_BIDIRECTIONAL)
2315 return IOMMU_PROT_IW | IOMMU_PROT_IR;
2320 * This function contains common code for mapping of a physically
2321 * contiguous memory region into DMA address space. It is used by all
2322 * mapping functions provided with this IOMMU driver.
2323 * Must be called with the domain lock held.
2325 static dma_addr_t __map_single(struct device *dev,
2326 struct dma_ops_domain *dma_dom,
2329 enum dma_data_direction direction,
2332 dma_addr_t offset = paddr & ~PAGE_MASK;
2333 dma_addr_t address, start, ret;
2338 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
2341 address = dma_ops_alloc_iova(dev, dma_dom, pages, dma_mask);
2342 if (address == AMD_IOMMU_MAPPING_ERROR)
2345 prot = dir2prot(direction);
2348 for (i = 0; i < pages; ++i) {
2349 ret = iommu_map_page(&dma_dom->domain, start, paddr,
2350 PAGE_SIZE, prot, GFP_ATOMIC);
2359 if (unlikely(amd_iommu_np_cache)) {
2360 domain_flush_pages(&dma_dom->domain, address, size);
2361 domain_flush_complete(&dma_dom->domain);
2369 for (--i; i >= 0; --i) {
2371 iommu_unmap_page(&dma_dom->domain, start, PAGE_SIZE);
2374 domain_flush_tlb(&dma_dom->domain);
2375 domain_flush_complete(&dma_dom->domain);
2377 dma_ops_free_iova(dma_dom, address, pages);
2379 return AMD_IOMMU_MAPPING_ERROR;
2383 * Does the reverse of the __map_single function. Must be called with
2384 * the domain lock held too
2386 static void __unmap_single(struct dma_ops_domain *dma_dom,
2387 dma_addr_t dma_addr,
2391 dma_addr_t flush_addr;
2392 dma_addr_t i, start;
2395 flush_addr = dma_addr;
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;
2598 struct dma_ops_domain *dma_dom;
2601 domain = get_domain(dev);
2602 if (PTR_ERR(domain) == -EINVAL) {
2603 page = alloc_pages(flag, get_order(size));
2604 *dma_addr = page_to_phys(page);
2605 return page_address(page);
2606 } else if (IS_ERR(domain))
2609 dma_dom = to_dma_ops_domain(domain);
2610 size = PAGE_ALIGN(size);
2611 dma_mask = dev->coherent_dma_mask;
2612 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
2615 page = alloc_pages(flag | __GFP_NOWARN, get_order(size));
2617 if (!gfpflags_allow_blocking(flag))
2620 page = dma_alloc_from_contiguous(dev, size >> PAGE_SHIFT,
2621 get_order(size), flag);
2627 dma_mask = *dev->dma_mask;
2629 *dma_addr = __map_single(dev, dma_dom, page_to_phys(page),
2630 size, DMA_BIDIRECTIONAL, dma_mask);
2632 if (*dma_addr == AMD_IOMMU_MAPPING_ERROR)
2635 return page_address(page);
2639 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2640 __free_pages(page, get_order(size));
2646 * The exported free_coherent function for dma_ops.
2648 static void free_coherent(struct device *dev, size_t size,
2649 void *virt_addr, dma_addr_t dma_addr,
2650 unsigned long attrs)
2652 struct protection_domain *domain;
2653 struct dma_ops_domain *dma_dom;
2656 page = virt_to_page(virt_addr);
2657 size = PAGE_ALIGN(size);
2659 domain = get_domain(dev);
2663 dma_dom = to_dma_ops_domain(domain);
2665 __unmap_single(dma_dom, dma_addr, size, DMA_BIDIRECTIONAL);
2668 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2669 __free_pages(page, get_order(size));
2673 * This function is called by the DMA layer to find out if we can handle a
2674 * particular device. It is part of the dma_ops.
2676 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
2678 if (!x86_dma_supported(dev, mask))
2680 return check_device(dev);
2683 static int amd_iommu_mapping_error(struct device *dev, dma_addr_t dma_addr)
2685 return dma_addr == AMD_IOMMU_MAPPING_ERROR;
2688 static const struct dma_map_ops amd_iommu_dma_ops = {
2689 .alloc = alloc_coherent,
2690 .free = free_coherent,
2691 .map_page = map_page,
2692 .unmap_page = unmap_page,
2694 .unmap_sg = unmap_sg,
2695 .dma_supported = amd_iommu_dma_supported,
2696 .mapping_error = amd_iommu_mapping_error,
2699 static int init_reserved_iova_ranges(void)
2701 struct pci_dev *pdev = NULL;
2704 init_iova_domain(&reserved_iova_ranges, PAGE_SIZE,
2705 IOVA_START_PFN, DMA_32BIT_PFN);
2707 lockdep_set_class(&reserved_iova_ranges.iova_rbtree_lock,
2708 &reserved_rbtree_key);
2710 /* MSI memory range */
2711 val = reserve_iova(&reserved_iova_ranges,
2712 IOVA_PFN(MSI_RANGE_START), IOVA_PFN(MSI_RANGE_END));
2714 pr_err("Reserving MSI range failed\n");
2718 /* HT memory range */
2719 val = reserve_iova(&reserved_iova_ranges,
2720 IOVA_PFN(HT_RANGE_START), IOVA_PFN(HT_RANGE_END));
2722 pr_err("Reserving HT range failed\n");
2727 * Memory used for PCI resources
2728 * FIXME: Check whether we can reserve the PCI-hole completly
2730 for_each_pci_dev(pdev) {
2733 for (i = 0; i < PCI_NUM_RESOURCES; ++i) {
2734 struct resource *r = &pdev->resource[i];
2736 if (!(r->flags & IORESOURCE_MEM))
2739 val = reserve_iova(&reserved_iova_ranges,
2743 pr_err("Reserve pci-resource range failed\n");
2752 int __init amd_iommu_init_api(void)
2756 ret = iova_cache_get();
2760 ret = init_reserved_iova_ranges();
2764 err = bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
2767 #ifdef CONFIG_ARM_AMBA
2768 err = bus_set_iommu(&amba_bustype, &amd_iommu_ops);
2772 err = bus_set_iommu(&platform_bus_type, &amd_iommu_ops);
2779 int __init amd_iommu_init_dma_ops(void)
2781 swiotlb = iommu_pass_through ? 1 : 0;
2785 * In case we don't initialize SWIOTLB (actually the common case
2786 * when AMD IOMMU is enabled), make sure there are global
2787 * dma_ops set as a fall-back for devices not handled by this
2788 * driver (for example non-PCI devices).
2791 dma_ops = &nommu_dma_ops;
2793 if (amd_iommu_unmap_flush)
2794 pr_info("AMD-Vi: IO/TLB flush on unmap enabled\n");
2796 pr_info("AMD-Vi: Lazy IO/TLB flushing enabled\n");
2802 /*****************************************************************************
2804 * The following functions belong to the exported interface of AMD IOMMU
2806 * This interface allows access to lower level functions of the IOMMU
2807 * like protection domain handling and assignement of devices to domains
2808 * which is not possible with the dma_ops interface.
2810 *****************************************************************************/
2812 static void cleanup_domain(struct protection_domain *domain)
2814 struct iommu_dev_data *entry;
2815 unsigned long flags;
2817 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2819 while (!list_empty(&domain->dev_list)) {
2820 entry = list_first_entry(&domain->dev_list,
2821 struct iommu_dev_data, list);
2822 __detach_device(entry);
2825 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2828 static void protection_domain_free(struct protection_domain *domain)
2833 del_domain_from_list(domain);
2836 domain_id_free(domain->id);
2841 static int protection_domain_init(struct protection_domain *domain)
2843 spin_lock_init(&domain->lock);
2844 mutex_init(&domain->api_lock);
2845 domain->id = domain_id_alloc();
2848 INIT_LIST_HEAD(&domain->dev_list);
2853 static struct protection_domain *protection_domain_alloc(void)
2855 struct protection_domain *domain;
2857 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2861 if (protection_domain_init(domain))
2864 add_domain_to_list(domain);
2874 static struct iommu_domain *amd_iommu_domain_alloc(unsigned type)
2876 struct protection_domain *pdomain;
2877 struct dma_ops_domain *dma_domain;
2880 case IOMMU_DOMAIN_UNMANAGED:
2881 pdomain = protection_domain_alloc();
2885 pdomain->mode = PAGE_MODE_3_LEVEL;
2886 pdomain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2887 if (!pdomain->pt_root) {
2888 protection_domain_free(pdomain);
2892 pdomain->domain.geometry.aperture_start = 0;
2893 pdomain->domain.geometry.aperture_end = ~0ULL;
2894 pdomain->domain.geometry.force_aperture = true;
2897 case IOMMU_DOMAIN_DMA:
2898 dma_domain = dma_ops_domain_alloc();
2900 pr_err("AMD-Vi: Failed to allocate\n");
2903 pdomain = &dma_domain->domain;
2905 case IOMMU_DOMAIN_IDENTITY:
2906 pdomain = protection_domain_alloc();
2910 pdomain->mode = PAGE_MODE_NONE;
2916 return &pdomain->domain;
2919 static void amd_iommu_domain_free(struct iommu_domain *dom)
2921 struct protection_domain *domain;
2922 struct dma_ops_domain *dma_dom;
2924 domain = to_pdomain(dom);
2926 if (domain->dev_cnt > 0)
2927 cleanup_domain(domain);
2929 BUG_ON(domain->dev_cnt != 0);
2934 switch (dom->type) {
2935 case IOMMU_DOMAIN_DMA:
2936 /* Now release the domain */
2937 dma_dom = to_dma_ops_domain(domain);
2938 dma_ops_domain_free(dma_dom);
2941 if (domain->mode != PAGE_MODE_NONE)
2942 free_pagetable(domain);
2944 if (domain->flags & PD_IOMMUV2_MASK)
2945 free_gcr3_table(domain);
2947 protection_domain_free(domain);
2952 static void amd_iommu_detach_device(struct iommu_domain *dom,
2955 struct iommu_dev_data *dev_data = dev->archdata.iommu;
2956 struct amd_iommu *iommu;
2959 if (!check_device(dev))
2962 devid = get_device_id(dev);
2966 if (dev_data->domain != NULL)
2969 iommu = amd_iommu_rlookup_table[devid];
2973 #ifdef CONFIG_IRQ_REMAP
2974 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) &&
2975 (dom->type == IOMMU_DOMAIN_UNMANAGED))
2976 dev_data->use_vapic = 0;
2979 iommu_completion_wait(iommu);
2982 static int amd_iommu_attach_device(struct iommu_domain *dom,
2985 struct protection_domain *domain = to_pdomain(dom);
2986 struct iommu_dev_data *dev_data;
2987 struct amd_iommu *iommu;
2990 if (!check_device(dev))
2993 dev_data = dev->archdata.iommu;
2995 iommu = amd_iommu_rlookup_table[dev_data->devid];
2999 if (dev_data->domain)
3002 ret = attach_device(dev, domain);
3004 #ifdef CONFIG_IRQ_REMAP
3005 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
3006 if (dom->type == IOMMU_DOMAIN_UNMANAGED)
3007 dev_data->use_vapic = 1;
3009 dev_data->use_vapic = 0;
3013 iommu_completion_wait(iommu);
3018 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
3019 phys_addr_t paddr, size_t page_size, int iommu_prot)
3021 struct protection_domain *domain = to_pdomain(dom);
3025 if (domain->mode == PAGE_MODE_NONE)
3028 if (iommu_prot & IOMMU_READ)
3029 prot |= IOMMU_PROT_IR;
3030 if (iommu_prot & IOMMU_WRITE)
3031 prot |= IOMMU_PROT_IW;
3033 mutex_lock(&domain->api_lock);
3034 ret = iommu_map_page(domain, iova, paddr, page_size, prot, GFP_KERNEL);
3035 mutex_unlock(&domain->api_lock);
3040 static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
3043 struct protection_domain *domain = to_pdomain(dom);
3046 if (domain->mode == PAGE_MODE_NONE)
3049 mutex_lock(&domain->api_lock);
3050 unmap_size = iommu_unmap_page(domain, iova, page_size);
3051 mutex_unlock(&domain->api_lock);
3053 domain_flush_tlb_pde(domain);
3058 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
3061 struct protection_domain *domain = to_pdomain(dom);
3062 unsigned long offset_mask, pte_pgsize;
3065 if (domain->mode == PAGE_MODE_NONE)
3068 pte = fetch_pte(domain, iova, &pte_pgsize);
3070 if (!pte || !IOMMU_PTE_PRESENT(*pte))
3073 offset_mask = pte_pgsize - 1;
3074 __pte = *pte & PM_ADDR_MASK;
3076 return (__pte & ~offset_mask) | (iova & offset_mask);
3079 static bool amd_iommu_capable(enum iommu_cap cap)
3082 case IOMMU_CAP_CACHE_COHERENCY:
3084 case IOMMU_CAP_INTR_REMAP:
3085 return (irq_remapping_enabled == 1);
3086 case IOMMU_CAP_NOEXEC:
3093 static void amd_iommu_get_resv_regions(struct device *dev,
3094 struct list_head *head)
3096 struct iommu_resv_region *region;
3097 struct unity_map_entry *entry;
3100 devid = get_device_id(dev);
3104 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
3108 if (devid < entry->devid_start || devid > entry->devid_end)
3111 length = entry->address_end - entry->address_start;
3112 if (entry->prot & IOMMU_PROT_IR)
3114 if (entry->prot & IOMMU_PROT_IW)
3115 prot |= IOMMU_WRITE;
3117 region = iommu_alloc_resv_region(entry->address_start,
3121 pr_err("Out of memory allocating dm-regions for %s\n",
3125 list_add_tail(®ion->list, head);
3128 region = iommu_alloc_resv_region(MSI_RANGE_START,
3129 MSI_RANGE_END - MSI_RANGE_START + 1,
3133 list_add_tail(®ion->list, head);
3135 region = iommu_alloc_resv_region(HT_RANGE_START,
3136 HT_RANGE_END - HT_RANGE_START + 1,
3137 0, IOMMU_RESV_RESERVED);
3140 list_add_tail(®ion->list, head);
3143 static void amd_iommu_put_resv_regions(struct device *dev,
3144 struct list_head *head)
3146 struct iommu_resv_region *entry, *next;
3148 list_for_each_entry_safe(entry, next, head, list)
3152 static void amd_iommu_apply_resv_region(struct device *dev,
3153 struct iommu_domain *domain,
3154 struct iommu_resv_region *region)
3156 struct dma_ops_domain *dma_dom = to_dma_ops_domain(to_pdomain(domain));
3157 unsigned long start, end;
3159 start = IOVA_PFN(region->start);
3160 end = IOVA_PFN(region->start + region->length);
3162 WARN_ON_ONCE(reserve_iova(&dma_dom->iovad, start, end) == NULL);
3165 const struct iommu_ops amd_iommu_ops = {
3166 .capable = amd_iommu_capable,
3167 .domain_alloc = amd_iommu_domain_alloc,
3168 .domain_free = amd_iommu_domain_free,
3169 .attach_dev = amd_iommu_attach_device,
3170 .detach_dev = amd_iommu_detach_device,
3171 .map = amd_iommu_map,
3172 .unmap = amd_iommu_unmap,
3173 .map_sg = default_iommu_map_sg,
3174 .iova_to_phys = amd_iommu_iova_to_phys,
3175 .add_device = amd_iommu_add_device,
3176 .remove_device = amd_iommu_remove_device,
3177 .device_group = amd_iommu_device_group,
3178 .get_resv_regions = amd_iommu_get_resv_regions,
3179 .put_resv_regions = amd_iommu_put_resv_regions,
3180 .apply_resv_region = amd_iommu_apply_resv_region,
3181 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
3184 /*****************************************************************************
3186 * The next functions do a basic initialization of IOMMU for pass through
3189 * In passthrough mode the IOMMU is initialized and enabled but not used for
3190 * DMA-API translation.
3192 *****************************************************************************/
3194 /* IOMMUv2 specific functions */
3195 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
3197 return atomic_notifier_chain_register(&ppr_notifier, nb);
3199 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
3201 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
3203 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
3205 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
3207 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
3209 struct protection_domain *domain = to_pdomain(dom);
3210 unsigned long flags;
3212 spin_lock_irqsave(&domain->lock, flags);
3214 /* Update data structure */
3215 domain->mode = PAGE_MODE_NONE;
3216 domain->updated = true;
3218 /* Make changes visible to IOMMUs */
3219 update_domain(domain);
3221 /* Page-table is not visible to IOMMU anymore, so free it */
3222 free_pagetable(domain);
3224 spin_unlock_irqrestore(&domain->lock, flags);
3226 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
3228 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
3230 struct protection_domain *domain = to_pdomain(dom);
3231 unsigned long flags;
3234 if (pasids <= 0 || pasids > (PASID_MASK + 1))
3237 /* Number of GCR3 table levels required */
3238 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
3241 if (levels > amd_iommu_max_glx_val)
3244 spin_lock_irqsave(&domain->lock, flags);
3247 * Save us all sanity checks whether devices already in the
3248 * domain support IOMMUv2. Just force that the domain has no
3249 * devices attached when it is switched into IOMMUv2 mode.
3252 if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
3256 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
3257 if (domain->gcr3_tbl == NULL)
3260 domain->glx = levels;
3261 domain->flags |= PD_IOMMUV2_MASK;
3262 domain->updated = true;
3264 update_domain(domain);
3269 spin_unlock_irqrestore(&domain->lock, flags);
3273 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
3275 static int __flush_pasid(struct protection_domain *domain, int pasid,
3276 u64 address, bool size)
3278 struct iommu_dev_data *dev_data;
3279 struct iommu_cmd cmd;
3282 if (!(domain->flags & PD_IOMMUV2_MASK))
3285 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
3288 * IOMMU TLB needs to be flushed before Device TLB to
3289 * prevent device TLB refill from IOMMU TLB
3291 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
3292 if (domain->dev_iommu[i] == 0)
3295 ret = iommu_queue_command(amd_iommus[i], &cmd);
3300 /* Wait until IOMMU TLB flushes are complete */
3301 domain_flush_complete(domain);
3303 /* Now flush device TLBs */
3304 list_for_each_entry(dev_data, &domain->dev_list, list) {
3305 struct amd_iommu *iommu;
3309 There might be non-IOMMUv2 capable devices in an IOMMUv2
3312 if (!dev_data->ats.enabled)
3315 qdep = dev_data->ats.qdep;
3316 iommu = amd_iommu_rlookup_table[dev_data->devid];
3318 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
3319 qdep, address, size);
3321 ret = iommu_queue_command(iommu, &cmd);
3326 /* Wait until all device TLBs are flushed */
3327 domain_flush_complete(domain);
3336 static int __amd_iommu_flush_page(struct protection_domain *domain, int pasid,
3339 return __flush_pasid(domain, pasid, address, false);
3342 int amd_iommu_flush_page(struct iommu_domain *dom, int pasid,
3345 struct protection_domain *domain = to_pdomain(dom);
3346 unsigned long flags;
3349 spin_lock_irqsave(&domain->lock, flags);
3350 ret = __amd_iommu_flush_page(domain, pasid, address);
3351 spin_unlock_irqrestore(&domain->lock, flags);
3355 EXPORT_SYMBOL(amd_iommu_flush_page);
3357 static int __amd_iommu_flush_tlb(struct protection_domain *domain, int pasid)
3359 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
3363 int amd_iommu_flush_tlb(struct iommu_domain *dom, int pasid)
3365 struct protection_domain *domain = to_pdomain(dom);
3366 unsigned long flags;
3369 spin_lock_irqsave(&domain->lock, flags);
3370 ret = __amd_iommu_flush_tlb(domain, pasid);
3371 spin_unlock_irqrestore(&domain->lock, flags);
3375 EXPORT_SYMBOL(amd_iommu_flush_tlb);
3377 static u64 *__get_gcr3_pte(u64 *root, int level, int pasid, bool alloc)
3384 index = (pasid >> (9 * level)) & 0x1ff;
3390 if (!(*pte & GCR3_VALID)) {
3394 root = (void *)get_zeroed_page(GFP_ATOMIC);
3398 *pte = __pa(root) | GCR3_VALID;
3401 root = __va(*pte & PAGE_MASK);
3409 static int __set_gcr3(struct protection_domain *domain, int pasid,
3414 if (domain->mode != PAGE_MODE_NONE)
3417 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3421 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3423 return __amd_iommu_flush_tlb(domain, pasid);
3426 static int __clear_gcr3(struct protection_domain *domain, int pasid)
3430 if (domain->mode != PAGE_MODE_NONE)
3433 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3439 return __amd_iommu_flush_tlb(domain, pasid);
3442 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, int pasid,
3445 struct protection_domain *domain = to_pdomain(dom);
3446 unsigned long flags;
3449 spin_lock_irqsave(&domain->lock, flags);
3450 ret = __set_gcr3(domain, pasid, cr3);
3451 spin_unlock_irqrestore(&domain->lock, flags);
3455 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3457 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, int pasid)
3459 struct protection_domain *domain = to_pdomain(dom);
3460 unsigned long flags;
3463 spin_lock_irqsave(&domain->lock, flags);
3464 ret = __clear_gcr3(domain, pasid);
3465 spin_unlock_irqrestore(&domain->lock, flags);
3469 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3471 int amd_iommu_complete_ppr(struct pci_dev *pdev, int pasid,
3472 int status, int tag)
3474 struct iommu_dev_data *dev_data;
3475 struct amd_iommu *iommu;
3476 struct iommu_cmd cmd;
3478 dev_data = get_dev_data(&pdev->dev);
3479 iommu = amd_iommu_rlookup_table[dev_data->devid];
3481 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3482 tag, dev_data->pri_tlp);
3484 return iommu_queue_command(iommu, &cmd);
3486 EXPORT_SYMBOL(amd_iommu_complete_ppr);
3488 struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3490 struct protection_domain *pdomain;
3492 pdomain = get_domain(&pdev->dev);
3493 if (IS_ERR(pdomain))
3496 /* Only return IOMMUv2 domains */
3497 if (!(pdomain->flags & PD_IOMMUV2_MASK))
3500 return &pdomain->domain;
3502 EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3504 void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3506 struct iommu_dev_data *dev_data;
3508 if (!amd_iommu_v2_supported())
3511 dev_data = get_dev_data(&pdev->dev);
3512 dev_data->errata |= (1 << erratum);
3514 EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3516 int amd_iommu_device_info(struct pci_dev *pdev,
3517 struct amd_iommu_device_info *info)
3522 if (pdev == NULL || info == NULL)
3525 if (!amd_iommu_v2_supported())
3528 memset(info, 0, sizeof(*info));
3530 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS);
3532 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3534 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3536 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3538 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3542 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3543 max_pasids = min(max_pasids, (1 << 20));
3545 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3546 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3548 features = pci_pasid_features(pdev);
3549 if (features & PCI_PASID_CAP_EXEC)
3550 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3551 if (features & PCI_PASID_CAP_PRIV)
3552 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3557 EXPORT_SYMBOL(amd_iommu_device_info);
3559 #ifdef CONFIG_IRQ_REMAP
3561 /*****************************************************************************
3563 * Interrupt Remapping Implementation
3565 *****************************************************************************/
3567 static struct irq_chip amd_ir_chip;
3569 #define DTE_IRQ_PHYS_ADDR_MASK (((1ULL << 45)-1) << 6)
3570 #define DTE_IRQ_REMAP_INTCTL (2ULL << 60)
3571 #define DTE_IRQ_TABLE_LEN (8ULL << 1)
3572 #define DTE_IRQ_REMAP_ENABLE 1ULL
3574 static void set_dte_irq_entry(u16 devid, struct irq_remap_table *table)
3578 dte = amd_iommu_dev_table[devid].data[2];
3579 dte &= ~DTE_IRQ_PHYS_ADDR_MASK;
3580 dte |= virt_to_phys(table->table);
3581 dte |= DTE_IRQ_REMAP_INTCTL;
3582 dte |= DTE_IRQ_TABLE_LEN;
3583 dte |= DTE_IRQ_REMAP_ENABLE;
3585 amd_iommu_dev_table[devid].data[2] = dte;
3588 static struct irq_remap_table *get_irq_table(u16 devid, bool ioapic)
3590 struct irq_remap_table *table = NULL;
3591 struct amd_iommu *iommu;
3592 unsigned long flags;
3595 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3597 iommu = amd_iommu_rlookup_table[devid];
3601 table = irq_lookup_table[devid];
3605 alias = amd_iommu_alias_table[devid];
3606 table = irq_lookup_table[alias];
3608 irq_lookup_table[devid] = table;
3609 set_dte_irq_entry(devid, table);
3610 iommu_flush_dte(iommu, devid);
3614 /* Nothing there yet, allocate new irq remapping table */
3615 table = kzalloc(sizeof(*table), GFP_ATOMIC);
3619 /* Initialize table spin-lock */
3620 spin_lock_init(&table->lock);
3623 /* Keep the first 32 indexes free for IOAPIC interrupts */
3624 table->min_index = 32;
3626 table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_ATOMIC);
3627 if (!table->table) {
3633 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
3634 memset(table->table, 0,
3635 MAX_IRQS_PER_TABLE * sizeof(u32));
3637 memset(table->table, 0,
3638 (MAX_IRQS_PER_TABLE * (sizeof(u64) * 2)));
3643 for (i = 0; i < 32; ++i)
3644 iommu->irte_ops->set_allocated(table, i);
3647 irq_lookup_table[devid] = table;
3648 set_dte_irq_entry(devid, table);
3649 iommu_flush_dte(iommu, devid);
3650 if (devid != alias) {
3651 irq_lookup_table[alias] = table;
3652 set_dte_irq_entry(alias, table);
3653 iommu_flush_dte(iommu, alias);
3657 iommu_completion_wait(iommu);
3660 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3665 static int alloc_irq_index(u16 devid, int count)
3667 struct irq_remap_table *table;
3668 unsigned long flags;
3670 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
3675 table = get_irq_table(devid, false);
3679 spin_lock_irqsave(&table->lock, flags);
3681 /* Scan table for free entries */
3682 for (c = 0, index = table->min_index;
3683 index < MAX_IRQS_PER_TABLE;
3685 if (!iommu->irte_ops->is_allocated(table, index))
3692 iommu->irte_ops->set_allocated(table, index - c + 1);
3702 spin_unlock_irqrestore(&table->lock, flags);
3707 static int modify_irte_ga(u16 devid, int index, struct irte_ga *irte,
3708 struct amd_ir_data *data)
3710 struct irq_remap_table *table;
3711 struct amd_iommu *iommu;
3712 unsigned long flags;
3713 struct irte_ga *entry;
3715 iommu = amd_iommu_rlookup_table[devid];
3719 table = get_irq_table(devid, false);
3723 spin_lock_irqsave(&table->lock, flags);
3725 entry = (struct irte_ga *)table->table;
3726 entry = &entry[index];
3727 entry->lo.fields_remap.valid = 0;
3728 entry->hi.val = irte->hi.val;
3729 entry->lo.val = irte->lo.val;
3730 entry->lo.fields_remap.valid = 1;
3734 spin_unlock_irqrestore(&table->lock, flags);
3736 iommu_flush_irt(iommu, devid);
3737 iommu_completion_wait(iommu);
3742 static int modify_irte(u16 devid, int index, union irte *irte)
3744 struct irq_remap_table *table;
3745 struct amd_iommu *iommu;
3746 unsigned long flags;
3748 iommu = amd_iommu_rlookup_table[devid];
3752 table = get_irq_table(devid, false);
3756 spin_lock_irqsave(&table->lock, flags);
3757 table->table[index] = irte->val;
3758 spin_unlock_irqrestore(&table->lock, flags);
3760 iommu_flush_irt(iommu, devid);
3761 iommu_completion_wait(iommu);
3766 static void free_irte(u16 devid, int index)
3768 struct irq_remap_table *table;
3769 struct amd_iommu *iommu;
3770 unsigned long flags;
3772 iommu = amd_iommu_rlookup_table[devid];
3776 table = get_irq_table(devid, false);
3780 spin_lock_irqsave(&table->lock, flags);
3781 iommu->irte_ops->clear_allocated(table, index);
3782 spin_unlock_irqrestore(&table->lock, flags);
3784 iommu_flush_irt(iommu, devid);
3785 iommu_completion_wait(iommu);
3788 static void irte_prepare(void *entry,
3789 u32 delivery_mode, u32 dest_mode,
3790 u8 vector, u32 dest_apicid, int devid)
3792 union irte *irte = (union irte *) entry;
3795 irte->fields.vector = vector;
3796 irte->fields.int_type = delivery_mode;
3797 irte->fields.destination = dest_apicid;
3798 irte->fields.dm = dest_mode;
3799 irte->fields.valid = 1;
3802 static void irte_ga_prepare(void *entry,
3803 u32 delivery_mode, u32 dest_mode,
3804 u8 vector, u32 dest_apicid, int devid)
3806 struct irte_ga *irte = (struct irte_ga *) entry;
3810 irte->lo.fields_remap.int_type = delivery_mode;
3811 irte->lo.fields_remap.dm = dest_mode;
3812 irte->hi.fields.vector = vector;
3813 irte->lo.fields_remap.destination = dest_apicid;
3814 irte->lo.fields_remap.valid = 1;
3817 static void irte_activate(void *entry, u16 devid, u16 index)
3819 union irte *irte = (union irte *) entry;
3821 irte->fields.valid = 1;
3822 modify_irte(devid, index, irte);
3825 static void irte_ga_activate(void *entry, u16 devid, u16 index)
3827 struct irte_ga *irte = (struct irte_ga *) entry;
3829 irte->lo.fields_remap.valid = 1;
3830 modify_irte_ga(devid, index, irte, NULL);
3833 static void irte_deactivate(void *entry, u16 devid, u16 index)
3835 union irte *irte = (union irte *) entry;
3837 irte->fields.valid = 0;
3838 modify_irte(devid, index, irte);
3841 static void irte_ga_deactivate(void *entry, u16 devid, u16 index)
3843 struct irte_ga *irte = (struct irte_ga *) entry;
3845 irte->lo.fields_remap.valid = 0;
3846 modify_irte_ga(devid, index, irte, NULL);
3849 static void irte_set_affinity(void *entry, u16 devid, u16 index,
3850 u8 vector, u32 dest_apicid)
3852 union irte *irte = (union irte *) entry;
3854 irte->fields.vector = vector;
3855 irte->fields.destination = dest_apicid;
3856 modify_irte(devid, index, irte);
3859 static void irte_ga_set_affinity(void *entry, u16 devid, u16 index,
3860 u8 vector, u32 dest_apicid)
3862 struct irte_ga *irte = (struct irte_ga *) entry;
3863 struct iommu_dev_data *dev_data = search_dev_data(devid);
3865 if (!dev_data || !dev_data->use_vapic ||
3866 !irte->lo.fields_remap.guest_mode) {
3867 irte->hi.fields.vector = vector;
3868 irte->lo.fields_remap.destination = dest_apicid;
3869 modify_irte_ga(devid, index, irte, NULL);
3873 #define IRTE_ALLOCATED (~1U)
3874 static void irte_set_allocated(struct irq_remap_table *table, int index)
3876 table->table[index] = IRTE_ALLOCATED;
3879 static void irte_ga_set_allocated(struct irq_remap_table *table, int index)
3881 struct irte_ga *ptr = (struct irte_ga *)table->table;
3882 struct irte_ga *irte = &ptr[index];
3884 memset(&irte->lo.val, 0, sizeof(u64));
3885 memset(&irte->hi.val, 0, sizeof(u64));
3886 irte->hi.fields.vector = 0xff;
3889 static bool irte_is_allocated(struct irq_remap_table *table, int index)
3891 union irte *ptr = (union irte *)table->table;
3892 union irte *irte = &ptr[index];
3894 return irte->val != 0;
3897 static bool irte_ga_is_allocated(struct irq_remap_table *table, int index)
3899 struct irte_ga *ptr = (struct irte_ga *)table->table;
3900 struct irte_ga *irte = &ptr[index];
3902 return irte->hi.fields.vector != 0;
3905 static void irte_clear_allocated(struct irq_remap_table *table, int index)
3907 table->table[index] = 0;
3910 static void irte_ga_clear_allocated(struct irq_remap_table *table, int index)
3912 struct irte_ga *ptr = (struct irte_ga *)table->table;
3913 struct irte_ga *irte = &ptr[index];
3915 memset(&irte->lo.val, 0, sizeof(u64));
3916 memset(&irte->hi.val, 0, sizeof(u64));
3919 static int get_devid(struct irq_alloc_info *info)
3923 switch (info->type) {
3924 case X86_IRQ_ALLOC_TYPE_IOAPIC:
3925 devid = get_ioapic_devid(info->ioapic_id);
3927 case X86_IRQ_ALLOC_TYPE_HPET:
3928 devid = get_hpet_devid(info->hpet_id);
3930 case X86_IRQ_ALLOC_TYPE_MSI:
3931 case X86_IRQ_ALLOC_TYPE_MSIX:
3932 devid = get_device_id(&info->msi_dev->dev);
3942 static struct irq_domain *get_ir_irq_domain(struct irq_alloc_info *info)
3944 struct amd_iommu *iommu;
3950 devid = get_devid(info);
3952 iommu = amd_iommu_rlookup_table[devid];
3954 return iommu->ir_domain;
3960 static struct irq_domain *get_irq_domain(struct irq_alloc_info *info)
3962 struct amd_iommu *iommu;
3968 switch (info->type) {
3969 case X86_IRQ_ALLOC_TYPE_MSI:
3970 case X86_IRQ_ALLOC_TYPE_MSIX:
3971 devid = get_device_id(&info->msi_dev->dev);
3975 iommu = amd_iommu_rlookup_table[devid];
3977 return iommu->msi_domain;
3986 struct irq_remap_ops amd_iommu_irq_ops = {
3987 .prepare = amd_iommu_prepare,
3988 .enable = amd_iommu_enable,
3989 .disable = amd_iommu_disable,
3990 .reenable = amd_iommu_reenable,
3991 .enable_faulting = amd_iommu_enable_faulting,
3992 .get_ir_irq_domain = get_ir_irq_domain,
3993 .get_irq_domain = get_irq_domain,
3996 static void irq_remapping_prepare_irte(struct amd_ir_data *data,
3997 struct irq_cfg *irq_cfg,
3998 struct irq_alloc_info *info,
3999 int devid, int index, int sub_handle)
4001 struct irq_2_irte *irte_info = &data->irq_2_irte;
4002 struct msi_msg *msg = &data->msi_entry;
4003 struct IO_APIC_route_entry *entry;
4004 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
4009 data->irq_2_irte.devid = devid;
4010 data->irq_2_irte.index = index + sub_handle;
4011 iommu->irte_ops->prepare(data->entry, apic->irq_delivery_mode,
4012 apic->irq_dest_mode, irq_cfg->vector,
4013 irq_cfg->dest_apicid, devid);
4015 switch (info->type) {
4016 case X86_IRQ_ALLOC_TYPE_IOAPIC:
4017 /* Setup IOAPIC entry */
4018 entry = info->ioapic_entry;
4019 info->ioapic_entry = NULL;
4020 memset(entry, 0, sizeof(*entry));
4021 entry->vector = index;
4023 entry->trigger = info->ioapic_trigger;
4024 entry->polarity = info->ioapic_polarity;
4025 /* Mask level triggered irqs. */
4026 if (info->ioapic_trigger)
4030 case X86_IRQ_ALLOC_TYPE_HPET:
4031 case X86_IRQ_ALLOC_TYPE_MSI:
4032 case X86_IRQ_ALLOC_TYPE_MSIX:
4033 msg->address_hi = MSI_ADDR_BASE_HI;
4034 msg->address_lo = MSI_ADDR_BASE_LO;
4035 msg->data = irte_info->index;
4044 struct amd_irte_ops irte_32_ops = {
4045 .prepare = irte_prepare,
4046 .activate = irte_activate,
4047 .deactivate = irte_deactivate,
4048 .set_affinity = irte_set_affinity,
4049 .set_allocated = irte_set_allocated,
4050 .is_allocated = irte_is_allocated,
4051 .clear_allocated = irte_clear_allocated,
4054 struct amd_irte_ops irte_128_ops = {
4055 .prepare = irte_ga_prepare,
4056 .activate = irte_ga_activate,
4057 .deactivate = irte_ga_deactivate,
4058 .set_affinity = irte_ga_set_affinity,
4059 .set_allocated = irte_ga_set_allocated,
4060 .is_allocated = irte_ga_is_allocated,
4061 .clear_allocated = irte_ga_clear_allocated,
4064 static int irq_remapping_alloc(struct irq_domain *domain, unsigned int virq,
4065 unsigned int nr_irqs, void *arg)
4067 struct irq_alloc_info *info = arg;
4068 struct irq_data *irq_data;
4069 struct amd_ir_data *data = NULL;
4070 struct irq_cfg *cfg;
4076 if (nr_irqs > 1 && info->type != X86_IRQ_ALLOC_TYPE_MSI &&
4077 info->type != X86_IRQ_ALLOC_TYPE_MSIX)
4081 * With IRQ remapping enabled, don't need contiguous CPU vectors
4082 * to support multiple MSI interrupts.
4084 if (info->type == X86_IRQ_ALLOC_TYPE_MSI)
4085 info->flags &= ~X86_IRQ_ALLOC_CONTIGUOUS_VECTORS;
4087 devid = get_devid(info);
4091 ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
4095 if (info->type == X86_IRQ_ALLOC_TYPE_IOAPIC) {
4096 if (get_irq_table(devid, true))
4097 index = info->ioapic_pin;
4101 index = alloc_irq_index(devid, nr_irqs);
4104 pr_warn("Failed to allocate IRTE\n");
4106 goto out_free_parent;
4109 for (i = 0; i < nr_irqs; i++) {
4110 irq_data = irq_domain_get_irq_data(domain, virq + i);
4111 cfg = irqd_cfg(irq_data);
4112 if (!irq_data || !cfg) {
4118 data = kzalloc(sizeof(*data), GFP_KERNEL);
4122 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
4123 data->entry = kzalloc(sizeof(union irte), GFP_KERNEL);
4125 data->entry = kzalloc(sizeof(struct irte_ga),
4132 irq_data->hwirq = (devid << 16) + i;
4133 irq_data->chip_data = data;
4134 irq_data->chip = &amd_ir_chip;
4135 irq_remapping_prepare_irte(data, cfg, info, devid, index, i);
4136 irq_set_status_flags(virq + i, IRQ_MOVE_PCNTXT);
4142 for (i--; i >= 0; i--) {
4143 irq_data = irq_domain_get_irq_data(domain, virq + i);
4145 kfree(irq_data->chip_data);
4147 for (i = 0; i < nr_irqs; i++)
4148 free_irte(devid, index + i);
4150 irq_domain_free_irqs_common(domain, virq, nr_irqs);
4154 static void irq_remapping_free(struct irq_domain *domain, unsigned int virq,
4155 unsigned int nr_irqs)
4157 struct irq_2_irte *irte_info;
4158 struct irq_data *irq_data;
4159 struct amd_ir_data *data;
4162 for (i = 0; i < nr_irqs; i++) {
4163 irq_data = irq_domain_get_irq_data(domain, virq + i);
4164 if (irq_data && irq_data->chip_data) {
4165 data = irq_data->chip_data;
4166 irte_info = &data->irq_2_irte;
4167 free_irte(irte_info->devid, irte_info->index);
4172 irq_domain_free_irqs_common(domain, virq, nr_irqs);
4175 static void irq_remapping_activate(struct irq_domain *domain,
4176 struct irq_data *irq_data)
4178 struct amd_ir_data *data = irq_data->chip_data;
4179 struct irq_2_irte *irte_info = &data->irq_2_irte;
4180 struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4183 iommu->irte_ops->activate(data->entry, irte_info->devid,
4187 static void irq_remapping_deactivate(struct irq_domain *domain,
4188 struct irq_data *irq_data)
4190 struct amd_ir_data *data = irq_data->chip_data;
4191 struct irq_2_irte *irte_info = &data->irq_2_irte;
4192 struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4195 iommu->irte_ops->deactivate(data->entry, irte_info->devid,
4199 static const struct irq_domain_ops amd_ir_domain_ops = {
4200 .alloc = irq_remapping_alloc,
4201 .free = irq_remapping_free,
4202 .activate = irq_remapping_activate,
4203 .deactivate = irq_remapping_deactivate,
4206 static int amd_ir_set_vcpu_affinity(struct irq_data *data, void *vcpu_info)
4208 struct amd_iommu *iommu;
4209 struct amd_iommu_pi_data *pi_data = vcpu_info;
4210 struct vcpu_data *vcpu_pi_info = pi_data->vcpu_data;
4211 struct amd_ir_data *ir_data = data->chip_data;
4212 struct irte_ga *irte = (struct irte_ga *) ir_data->entry;
4213 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
4214 struct iommu_dev_data *dev_data = search_dev_data(irte_info->devid);
4217 * This device has never been set up for guest mode.
4218 * we should not modify the IRTE
4220 if (!dev_data || !dev_data->use_vapic)
4223 pi_data->ir_data = ir_data;
4226 * SVM tries to set up for VAPIC mode, but we are in
4227 * legacy mode. So, we force legacy mode instead.
4229 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
4230 pr_debug("AMD-Vi: %s: Fall back to using intr legacy remap\n",
4232 pi_data->is_guest_mode = false;
4235 iommu = amd_iommu_rlookup_table[irte_info->devid];
4239 pi_data->prev_ga_tag = ir_data->cached_ga_tag;
4240 if (pi_data->is_guest_mode) {
4242 irte->hi.fields.ga_root_ptr = (pi_data->base >> 12);
4243 irte->hi.fields.vector = vcpu_pi_info->vector;
4244 irte->lo.fields_vapic.guest_mode = 1;
4245 irte->lo.fields_vapic.ga_tag = pi_data->ga_tag;
4247 ir_data->cached_ga_tag = pi_data->ga_tag;
4250 struct irq_cfg *cfg = irqd_cfg(data);
4254 irte->hi.fields.vector = cfg->vector;
4255 irte->lo.fields_remap.guest_mode = 0;
4256 irte->lo.fields_remap.destination = cfg->dest_apicid;
4257 irte->lo.fields_remap.int_type = apic->irq_delivery_mode;
4258 irte->lo.fields_remap.dm = apic->irq_dest_mode;
4261 * This communicates the ga_tag back to the caller
4262 * so that it can do all the necessary clean up.
4264 ir_data->cached_ga_tag = 0;
4267 return modify_irte_ga(irte_info->devid, irte_info->index, irte, ir_data);
4270 static int amd_ir_set_affinity(struct irq_data *data,
4271 const struct cpumask *mask, bool force)
4273 struct amd_ir_data *ir_data = data->chip_data;
4274 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
4275 struct irq_cfg *cfg = irqd_cfg(data);
4276 struct irq_data *parent = data->parent_data;
4277 struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4283 ret = parent->chip->irq_set_affinity(parent, mask, force);
4284 if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
4288 * Atomically updates the IRTE with the new destination, vector
4289 * and flushes the interrupt entry cache.
4291 iommu->irte_ops->set_affinity(ir_data->entry, irte_info->devid,
4292 irte_info->index, cfg->vector, cfg->dest_apicid);
4295 * After this point, all the interrupts will start arriving
4296 * at the new destination. So, time to cleanup the previous
4297 * vector allocation.
4299 send_cleanup_vector(cfg);
4301 return IRQ_SET_MASK_OK_DONE;
4304 static void ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg)
4306 struct amd_ir_data *ir_data = irq_data->chip_data;
4308 *msg = ir_data->msi_entry;
4311 static struct irq_chip amd_ir_chip = {
4313 .irq_ack = ir_ack_apic_edge,
4314 .irq_set_affinity = amd_ir_set_affinity,
4315 .irq_set_vcpu_affinity = amd_ir_set_vcpu_affinity,
4316 .irq_compose_msi_msg = ir_compose_msi_msg,
4319 int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
4321 struct fwnode_handle *fn;
4323 fn = irq_domain_alloc_named_id_fwnode("AMD-IR", iommu->index);
4326 iommu->ir_domain = irq_domain_create_tree(fn, &amd_ir_domain_ops, iommu);
4327 irq_domain_free_fwnode(fn);
4328 if (!iommu->ir_domain)
4331 iommu->ir_domain->parent = arch_get_ir_parent_domain();
4332 iommu->msi_domain = arch_create_remap_msi_irq_domain(iommu->ir_domain,
4338 int amd_iommu_update_ga(int cpu, bool is_run, void *data)
4340 unsigned long flags;
4341 struct amd_iommu *iommu;
4342 struct irq_remap_table *irt;
4343 struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
4344 int devid = ir_data->irq_2_irte.devid;
4345 struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
4346 struct irte_ga *ref = (struct irte_ga *) ir_data->ref;
4348 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
4349 !ref || !entry || !entry->lo.fields_vapic.guest_mode)
4352 iommu = amd_iommu_rlookup_table[devid];
4356 irt = get_irq_table(devid, false);
4360 spin_lock_irqsave(&irt->lock, flags);
4362 if (ref->lo.fields_vapic.guest_mode) {
4364 ref->lo.fields_vapic.destination = cpu;
4365 ref->lo.fields_vapic.is_run = is_run;
4369 spin_unlock_irqrestore(&irt->lock, flags);
4371 iommu_flush_irt(iommu, devid);
4372 iommu_completion_wait(iommu);
4375 EXPORT_SYMBOL(amd_iommu_update_ga);