2 * Copyright (C) 2007-2010 Advanced Micro Devices, Inc.
3 * Author: Joerg Roedel <jroedel@suse.de>
4 * Leo Duran <leo.duran@amd.com>
6 * This program is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 as published
8 * by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, write to the Free Software
17 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
20 #include <linux/ratelimit.h>
21 #include <linux/pci.h>
22 #include <linux/acpi.h>
23 #include <linux/amba/bus.h>
24 #include <linux/platform_device.h>
25 #include <linux/pci-ats.h>
26 #include <linux/bitmap.h>
27 #include <linux/slab.h>
28 #include <linux/debugfs.h>
29 #include <linux/scatterlist.h>
30 #include <linux/dma-mapping.h>
31 #include <linux/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 * general struct to manage commands send to an IOMMU
112 struct kmem_cache *amd_iommu_irq_cache;
114 static void update_domain(struct protection_domain *domain);
115 static int protection_domain_init(struct protection_domain *domain);
116 static void detach_device(struct device *dev);
117 static void iova_domain_flush_tlb(struct iova_domain *iovad);
120 * Data container for a dma_ops specific protection domain
122 struct dma_ops_domain {
123 /* generic protection domain information */
124 struct protection_domain domain;
127 struct iova_domain iovad;
130 static struct iova_domain reserved_iova_ranges;
131 static struct lock_class_key reserved_rbtree_key;
133 /****************************************************************************
137 ****************************************************************************/
139 static inline int match_hid_uid(struct device *dev,
140 struct acpihid_map_entry *entry)
142 struct acpi_device *adev = ACPI_COMPANION(dev);
143 const char *hid, *uid;
148 hid = acpi_device_hid(adev);
149 uid = acpi_device_uid(adev);
155 return strcmp(hid, entry->hid);
158 return strcmp(hid, entry->hid);
160 return (strcmp(hid, entry->hid) || strcmp(uid, entry->uid));
163 static inline u16 get_pci_device_id(struct device *dev)
165 struct pci_dev *pdev = to_pci_dev(dev);
167 return PCI_DEVID(pdev->bus->number, pdev->devfn);
170 static inline int get_acpihid_device_id(struct device *dev,
171 struct acpihid_map_entry **entry)
173 struct acpihid_map_entry *p;
175 list_for_each_entry(p, &acpihid_map, list) {
176 if (!match_hid_uid(dev, p)) {
185 static inline int get_device_id(struct device *dev)
190 devid = get_pci_device_id(dev);
192 devid = get_acpihid_device_id(dev, NULL);
197 static struct protection_domain *to_pdomain(struct iommu_domain *dom)
199 return container_of(dom, struct protection_domain, domain);
202 static struct dma_ops_domain* to_dma_ops_domain(struct protection_domain *domain)
204 BUG_ON(domain->flags != PD_DMA_OPS_MASK);
205 return container_of(domain, struct dma_ops_domain, domain);
208 static struct iommu_dev_data *alloc_dev_data(u16 devid)
210 struct iommu_dev_data *dev_data;
213 dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
217 dev_data->devid = devid;
219 spin_lock_irqsave(&dev_data_list_lock, flags);
220 list_add_tail(&dev_data->dev_data_list, &dev_data_list);
221 spin_unlock_irqrestore(&dev_data_list_lock, flags);
223 ratelimit_default_init(&dev_data->rs);
228 static struct iommu_dev_data *search_dev_data(u16 devid)
230 struct iommu_dev_data *dev_data;
233 spin_lock_irqsave(&dev_data_list_lock, flags);
234 list_for_each_entry(dev_data, &dev_data_list, dev_data_list) {
235 if (dev_data->devid == devid)
242 spin_unlock_irqrestore(&dev_data_list_lock, flags);
247 static int __last_alias(struct pci_dev *pdev, u16 alias, void *data)
249 *(u16 *)data = alias;
253 static u16 get_alias(struct device *dev)
255 struct pci_dev *pdev = to_pci_dev(dev);
256 u16 devid, ivrs_alias, pci_alias;
258 /* The callers make sure that get_device_id() does not fail here */
259 devid = get_device_id(dev);
261 /* For ACPI HID devices, we simply return the devid as such */
262 if (!dev_is_pci(dev))
265 ivrs_alias = amd_iommu_alias_table[devid];
267 pci_for_each_dma_alias(pdev, __last_alias, &pci_alias);
269 if (ivrs_alias == pci_alias)
275 * The IVRS is fairly reliable in telling us about aliases, but it
276 * can't know about every screwy device. If we don't have an IVRS
277 * reported alias, use the PCI reported alias. In that case we may
278 * still need to initialize the rlookup and dev_table entries if the
279 * alias is to a non-existent device.
281 if (ivrs_alias == devid) {
282 if (!amd_iommu_rlookup_table[pci_alias]) {
283 amd_iommu_rlookup_table[pci_alias] =
284 amd_iommu_rlookup_table[devid];
285 memcpy(amd_iommu_dev_table[pci_alias].data,
286 amd_iommu_dev_table[devid].data,
287 sizeof(amd_iommu_dev_table[pci_alias].data));
293 pr_info("AMD-Vi: Using IVRS reported alias %02x:%02x.%d "
294 "for device %s[%04x:%04x], kernel reported alias "
295 "%02x:%02x.%d\n", PCI_BUS_NUM(ivrs_alias), PCI_SLOT(ivrs_alias),
296 PCI_FUNC(ivrs_alias), dev_name(dev), pdev->vendor, pdev->device,
297 PCI_BUS_NUM(pci_alias), PCI_SLOT(pci_alias),
298 PCI_FUNC(pci_alias));
301 * If we don't have a PCI DMA alias and the IVRS alias is on the same
302 * bus, then the IVRS table may know about a quirk that we don't.
304 if (pci_alias == devid &&
305 PCI_BUS_NUM(ivrs_alias) == pdev->bus->number) {
306 pci_add_dma_alias(pdev, ivrs_alias & 0xff);
307 pr_info("AMD-Vi: Added PCI DMA alias %02x.%d for %s\n",
308 PCI_SLOT(ivrs_alias), PCI_FUNC(ivrs_alias),
315 static struct iommu_dev_data *find_dev_data(u16 devid)
317 struct iommu_dev_data *dev_data;
318 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
320 dev_data = search_dev_data(devid);
322 if (dev_data == NULL) {
323 dev_data = alloc_dev_data(devid);
327 if (translation_pre_enabled(iommu))
328 dev_data->defer_attach = true;
334 struct iommu_dev_data *get_dev_data(struct device *dev)
336 return dev->archdata.iommu;
338 EXPORT_SYMBOL(get_dev_data);
341 * Find or create an IOMMU group for a acpihid device.
343 static struct iommu_group *acpihid_device_group(struct device *dev)
345 struct acpihid_map_entry *p, *entry = NULL;
348 devid = get_acpihid_device_id(dev, &entry);
350 return ERR_PTR(devid);
352 list_for_each_entry(p, &acpihid_map, list) {
353 if ((devid == p->devid) && p->group)
354 entry->group = p->group;
358 entry->group = generic_device_group(dev);
360 iommu_group_ref_get(entry->group);
365 static bool pci_iommuv2_capable(struct pci_dev *pdev)
367 static const int caps[] = {
370 PCI_EXT_CAP_ID_PASID,
374 for (i = 0; i < 3; ++i) {
375 pos = pci_find_ext_capability(pdev, caps[i]);
383 static bool pdev_pri_erratum(struct pci_dev *pdev, u32 erratum)
385 struct iommu_dev_data *dev_data;
387 dev_data = get_dev_data(&pdev->dev);
389 return dev_data->errata & (1 << erratum) ? true : false;
393 * This function checks if the driver got a valid device from the caller to
394 * avoid dereferencing invalid pointers.
396 static bool check_device(struct device *dev)
400 if (!dev || !dev->dma_mask)
403 devid = get_device_id(dev);
407 /* Out of our scope? */
408 if (devid > amd_iommu_last_bdf)
411 if (amd_iommu_rlookup_table[devid] == NULL)
417 static void init_iommu_group(struct device *dev)
419 struct iommu_group *group;
421 group = iommu_group_get_for_dev(dev);
425 iommu_group_put(group);
428 static int iommu_init_device(struct device *dev)
430 struct iommu_dev_data *dev_data;
431 struct amd_iommu *iommu;
434 if (dev->archdata.iommu)
437 devid = get_device_id(dev);
441 iommu = amd_iommu_rlookup_table[devid];
443 dev_data = find_dev_data(devid);
447 dev_data->alias = get_alias(dev);
450 * By default we use passthrough mode for IOMMUv2 capable device.
451 * But if amd_iommu=force_isolation is set (e.g. to debug DMA to
452 * invalid address), we ignore the capability for the device so
453 * it'll be forced to go into translation mode.
455 if ((iommu_pass_through || !amd_iommu_force_isolation) &&
456 dev_is_pci(dev) && pci_iommuv2_capable(to_pci_dev(dev))) {
457 struct amd_iommu *iommu;
459 iommu = amd_iommu_rlookup_table[dev_data->devid];
460 dev_data->iommu_v2 = iommu->is_iommu_v2;
463 dev->archdata.iommu = dev_data;
465 iommu_device_link(&iommu->iommu, dev);
470 static void iommu_ignore_device(struct device *dev)
475 devid = get_device_id(dev);
479 alias = get_alias(dev);
481 memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
482 memset(&amd_iommu_dev_table[alias], 0, sizeof(struct dev_table_entry));
484 amd_iommu_rlookup_table[devid] = NULL;
485 amd_iommu_rlookup_table[alias] = NULL;
488 static void iommu_uninit_device(struct device *dev)
490 struct iommu_dev_data *dev_data;
491 struct amd_iommu *iommu;
494 devid = get_device_id(dev);
498 iommu = amd_iommu_rlookup_table[devid];
500 dev_data = search_dev_data(devid);
504 if (dev_data->domain)
507 iommu_device_unlink(&iommu->iommu, dev);
509 iommu_group_remove_device(dev);
515 * We keep dev_data around for unplugged devices and reuse it when the
516 * device is re-plugged - not doing so would introduce a ton of races.
520 /****************************************************************************
522 * Interrupt handling functions
524 ****************************************************************************/
526 static void dump_dte_entry(u16 devid)
530 for (i = 0; i < 4; ++i)
531 pr_err("AMD-Vi: DTE[%d]: %016llx\n", i,
532 amd_iommu_dev_table[devid].data[i]);
535 static void dump_command(unsigned long phys_addr)
537 struct iommu_cmd *cmd = iommu_phys_to_virt(phys_addr);
540 for (i = 0; i < 4; ++i)
541 pr_err("AMD-Vi: CMD[%d]: %08x\n", i, cmd->data[i]);
544 static void amd_iommu_report_page_fault(u16 devid, u16 domain_id,
545 u64 address, int flags)
547 struct iommu_dev_data *dev_data = NULL;
548 struct pci_dev *pdev;
550 pdev = pci_get_bus_and_slot(PCI_BUS_NUM(devid), devid & 0xff);
552 dev_data = get_dev_data(&pdev->dev);
554 if (dev_data && __ratelimit(&dev_data->rs)) {
555 dev_err(&pdev->dev, "AMD-Vi: Event logged [IO_PAGE_FAULT domain=0x%04x address=0x%016llx flags=0x%04x]\n",
556 domain_id, address, flags);
557 } else if (printk_ratelimit()) {
558 pr_err("AMD-Vi: Event logged [IO_PAGE_FAULT device=%02x:%02x.%x domain=0x%04x address=0x%016llx flags=0x%04x]\n",
559 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
560 domain_id, address, flags);
567 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
569 int type, devid, domid, flags;
570 volatile u32 *event = __evt;
575 type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
576 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
577 domid = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
578 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
579 address = (u64)(((u64)event[3]) << 32) | event[2];
582 /* Did we hit the erratum? */
583 if (++count == LOOP_TIMEOUT) {
584 pr_err("AMD-Vi: No event written to event log\n");
591 if (type == EVENT_TYPE_IO_FAULT) {
592 amd_iommu_report_page_fault(devid, domid, address, flags);
595 printk(KERN_ERR "AMD-Vi: Event logged [");
599 case EVENT_TYPE_ILL_DEV:
600 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
601 "address=0x%016llx flags=0x%04x]\n",
602 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
604 dump_dte_entry(devid);
606 case EVENT_TYPE_DEV_TAB_ERR:
607 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
608 "address=0x%016llx flags=0x%04x]\n",
609 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
612 case EVENT_TYPE_PAGE_TAB_ERR:
613 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
614 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
615 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
616 domid, address, flags);
618 case EVENT_TYPE_ILL_CMD:
619 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
620 dump_command(address);
622 case EVENT_TYPE_CMD_HARD_ERR:
623 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
624 "flags=0x%04x]\n", address, flags);
626 case EVENT_TYPE_IOTLB_INV_TO:
627 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
628 "address=0x%016llx]\n",
629 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
632 case EVENT_TYPE_INV_DEV_REQ:
633 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
634 "address=0x%016llx flags=0x%04x]\n",
635 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
639 printk(KERN_ERR "UNKNOWN type=0x%02x]\n", type);
642 memset(__evt, 0, 4 * sizeof(u32));
645 static void iommu_poll_events(struct amd_iommu *iommu)
649 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
650 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
652 while (head != tail) {
653 iommu_print_event(iommu, iommu->evt_buf + head);
654 head = (head + EVENT_ENTRY_SIZE) % EVT_BUFFER_SIZE;
657 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
660 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
662 struct amd_iommu_fault fault;
664 if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
665 pr_err_ratelimited("AMD-Vi: Unknown PPR request received\n");
669 fault.address = raw[1];
670 fault.pasid = PPR_PASID(raw[0]);
671 fault.device_id = PPR_DEVID(raw[0]);
672 fault.tag = PPR_TAG(raw[0]);
673 fault.flags = PPR_FLAGS(raw[0]);
675 atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
678 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
682 if (iommu->ppr_log == NULL)
685 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
686 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
688 while (head != tail) {
693 raw = (u64 *)(iommu->ppr_log + head);
696 * Hardware bug: Interrupt may arrive before the entry is
697 * written to memory. If this happens we need to wait for the
700 for (i = 0; i < LOOP_TIMEOUT; ++i) {
701 if (PPR_REQ_TYPE(raw[0]) != 0)
706 /* Avoid memcpy function-call overhead */
711 * To detect the hardware bug we need to clear the entry
714 raw[0] = raw[1] = 0UL;
716 /* Update head pointer of hardware ring-buffer */
717 head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
718 writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
720 /* Handle PPR entry */
721 iommu_handle_ppr_entry(iommu, entry);
723 /* Refresh ring-buffer information */
724 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
725 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
729 #ifdef CONFIG_IRQ_REMAP
730 static int (*iommu_ga_log_notifier)(u32);
732 int amd_iommu_register_ga_log_notifier(int (*notifier)(u32))
734 iommu_ga_log_notifier = notifier;
738 EXPORT_SYMBOL(amd_iommu_register_ga_log_notifier);
740 static void iommu_poll_ga_log(struct amd_iommu *iommu)
742 u32 head, tail, cnt = 0;
744 if (iommu->ga_log == NULL)
747 head = readl(iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
748 tail = readl(iommu->mmio_base + MMIO_GA_TAIL_OFFSET);
750 while (head != tail) {
754 raw = (u64 *)(iommu->ga_log + head);
757 /* Avoid memcpy function-call overhead */
760 /* Update head pointer of hardware ring-buffer */
761 head = (head + GA_ENTRY_SIZE) % GA_LOG_SIZE;
762 writel(head, iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
764 /* Handle GA entry */
765 switch (GA_REQ_TYPE(log_entry)) {
767 if (!iommu_ga_log_notifier)
770 pr_debug("AMD-Vi: %s: devid=%#x, ga_tag=%#x\n",
771 __func__, GA_DEVID(log_entry),
774 if (iommu_ga_log_notifier(GA_TAG(log_entry)) != 0)
775 pr_err("AMD-Vi: GA log notifier failed.\n");
782 #endif /* CONFIG_IRQ_REMAP */
784 #define AMD_IOMMU_INT_MASK \
785 (MMIO_STATUS_EVT_INT_MASK | \
786 MMIO_STATUS_PPR_INT_MASK | \
787 MMIO_STATUS_GALOG_INT_MASK)
789 irqreturn_t amd_iommu_int_thread(int irq, void *data)
791 struct amd_iommu *iommu = (struct amd_iommu *) data;
792 u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
794 while (status & AMD_IOMMU_INT_MASK) {
795 /* Enable EVT and PPR and GA interrupts again */
796 writel(AMD_IOMMU_INT_MASK,
797 iommu->mmio_base + MMIO_STATUS_OFFSET);
799 if (status & MMIO_STATUS_EVT_INT_MASK) {
800 pr_devel("AMD-Vi: Processing IOMMU Event Log\n");
801 iommu_poll_events(iommu);
804 if (status & MMIO_STATUS_PPR_INT_MASK) {
805 pr_devel("AMD-Vi: Processing IOMMU PPR Log\n");
806 iommu_poll_ppr_log(iommu);
809 #ifdef CONFIG_IRQ_REMAP
810 if (status & MMIO_STATUS_GALOG_INT_MASK) {
811 pr_devel("AMD-Vi: Processing IOMMU GA Log\n");
812 iommu_poll_ga_log(iommu);
817 * Hardware bug: ERBT1312
818 * When re-enabling interrupt (by writing 1
819 * to clear the bit), the hardware might also try to set
820 * the interrupt bit in the event status register.
821 * In this scenario, the bit will be set, and disable
822 * subsequent interrupts.
824 * Workaround: The IOMMU driver should read back the
825 * status register and check if the interrupt bits are cleared.
826 * If not, driver will need to go through the interrupt handler
827 * again and re-clear the bits
829 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
834 irqreturn_t amd_iommu_int_handler(int irq, void *data)
836 return IRQ_WAKE_THREAD;
839 /****************************************************************************
841 * IOMMU command queuing functions
843 ****************************************************************************/
845 static int wait_on_sem(volatile u64 *sem)
849 while (*sem == 0 && i < LOOP_TIMEOUT) {
854 if (i == LOOP_TIMEOUT) {
855 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
862 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
863 struct iommu_cmd *cmd)
867 target = iommu->cmd_buf + iommu->cmd_buf_tail;
869 iommu->cmd_buf_tail += sizeof(*cmd);
870 iommu->cmd_buf_tail %= CMD_BUFFER_SIZE;
872 /* Copy command to buffer */
873 memcpy(target, cmd, sizeof(*cmd));
875 /* Tell the IOMMU about it */
876 writel(iommu->cmd_buf_tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
879 static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
881 u64 paddr = iommu_virt_to_phys((void *)address);
883 WARN_ON(address & 0x7ULL);
885 memset(cmd, 0, sizeof(*cmd));
886 cmd->data[0] = lower_32_bits(paddr) | CMD_COMPL_WAIT_STORE_MASK;
887 cmd->data[1] = upper_32_bits(paddr);
889 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
892 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
894 memset(cmd, 0, sizeof(*cmd));
895 cmd->data[0] = devid;
896 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
899 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
900 size_t size, u16 domid, int pde)
905 pages = iommu_num_pages(address, size, PAGE_SIZE);
910 * If we have to flush more than one page, flush all
911 * TLB entries for this domain
913 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
917 address &= PAGE_MASK;
919 memset(cmd, 0, sizeof(*cmd));
920 cmd->data[1] |= domid;
921 cmd->data[2] = lower_32_bits(address);
922 cmd->data[3] = upper_32_bits(address);
923 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
924 if (s) /* size bit - we flush more than one 4kb page */
925 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
926 if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
927 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
930 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
931 u64 address, size_t size)
936 pages = iommu_num_pages(address, size, PAGE_SIZE);
941 * If we have to flush more than one page, flush all
942 * TLB entries for this domain
944 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
948 address &= PAGE_MASK;
950 memset(cmd, 0, sizeof(*cmd));
951 cmd->data[0] = devid;
952 cmd->data[0] |= (qdep & 0xff) << 24;
953 cmd->data[1] = devid;
954 cmd->data[2] = lower_32_bits(address);
955 cmd->data[3] = upper_32_bits(address);
956 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
958 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
961 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, int pasid,
962 u64 address, bool size)
964 memset(cmd, 0, sizeof(*cmd));
966 address &= ~(0xfffULL);
968 cmd->data[0] = pasid;
969 cmd->data[1] = domid;
970 cmd->data[2] = lower_32_bits(address);
971 cmd->data[3] = upper_32_bits(address);
972 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
973 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
975 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
976 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
979 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, int pasid,
980 int qdep, u64 address, bool size)
982 memset(cmd, 0, sizeof(*cmd));
984 address &= ~(0xfffULL);
986 cmd->data[0] = devid;
987 cmd->data[0] |= ((pasid >> 8) & 0xff) << 16;
988 cmd->data[0] |= (qdep & 0xff) << 24;
989 cmd->data[1] = devid;
990 cmd->data[1] |= (pasid & 0xff) << 16;
991 cmd->data[2] = lower_32_bits(address);
992 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
993 cmd->data[3] = upper_32_bits(address);
995 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
996 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
999 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, int pasid,
1000 int status, int tag, bool gn)
1002 memset(cmd, 0, sizeof(*cmd));
1004 cmd->data[0] = devid;
1006 cmd->data[1] = pasid;
1007 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
1009 cmd->data[3] = tag & 0x1ff;
1010 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
1012 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
1015 static void build_inv_all(struct iommu_cmd *cmd)
1017 memset(cmd, 0, sizeof(*cmd));
1018 CMD_SET_TYPE(cmd, CMD_INV_ALL);
1021 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
1023 memset(cmd, 0, sizeof(*cmd));
1024 cmd->data[0] = devid;
1025 CMD_SET_TYPE(cmd, CMD_INV_IRT);
1029 * Writes the command to the IOMMUs command buffer and informs the
1030 * hardware about the new command.
1032 static int __iommu_queue_command_sync(struct amd_iommu *iommu,
1033 struct iommu_cmd *cmd,
1036 unsigned int count = 0;
1037 u32 left, next_tail;
1039 next_tail = (iommu->cmd_buf_tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
1041 left = (iommu->cmd_buf_head - next_tail) % CMD_BUFFER_SIZE;
1044 /* Skip udelay() the first time around */
1046 if (count == LOOP_TIMEOUT) {
1047 pr_err("AMD-Vi: Command buffer timeout\n");
1054 /* Update head and recheck remaining space */
1055 iommu->cmd_buf_head = readl(iommu->mmio_base +
1056 MMIO_CMD_HEAD_OFFSET);
1061 copy_cmd_to_buffer(iommu, cmd);
1063 /* Do we need to make sure all commands are processed? */
1064 iommu->need_sync = sync;
1069 static int iommu_queue_command_sync(struct amd_iommu *iommu,
1070 struct iommu_cmd *cmd,
1073 unsigned long flags;
1076 spin_lock_irqsave(&iommu->lock, flags);
1077 ret = __iommu_queue_command_sync(iommu, cmd, sync);
1078 spin_unlock_irqrestore(&iommu->lock, flags);
1083 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
1085 return iommu_queue_command_sync(iommu, cmd, true);
1089 * This function queues a completion wait command into the command
1090 * buffer of an IOMMU
1092 static int iommu_completion_wait(struct amd_iommu *iommu)
1094 struct iommu_cmd cmd;
1095 unsigned long flags;
1098 if (!iommu->need_sync)
1102 build_completion_wait(&cmd, (u64)&iommu->cmd_sem);
1104 spin_lock_irqsave(&iommu->lock, flags);
1108 ret = __iommu_queue_command_sync(iommu, &cmd, false);
1112 ret = wait_on_sem(&iommu->cmd_sem);
1115 spin_unlock_irqrestore(&iommu->lock, flags);
1120 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
1122 struct iommu_cmd cmd;
1124 build_inv_dte(&cmd, devid);
1126 return iommu_queue_command(iommu, &cmd);
1129 static void amd_iommu_flush_dte_all(struct amd_iommu *iommu)
1133 for (devid = 0; devid <= 0xffff; ++devid)
1134 iommu_flush_dte(iommu, devid);
1136 iommu_completion_wait(iommu);
1140 * This function uses heavy locking and may disable irqs for some time. But
1141 * this is no issue because it is only called during resume.
1143 static void amd_iommu_flush_tlb_all(struct amd_iommu *iommu)
1147 for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
1148 struct iommu_cmd cmd;
1149 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1151 iommu_queue_command(iommu, &cmd);
1154 iommu_completion_wait(iommu);
1157 static void amd_iommu_flush_tlb_domid(struct amd_iommu *iommu, u32 dom_id)
1159 struct iommu_cmd cmd;
1161 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1163 iommu_queue_command(iommu, &cmd);
1165 iommu_completion_wait(iommu);
1168 static void amd_iommu_flush_all(struct amd_iommu *iommu)
1170 struct iommu_cmd cmd;
1172 build_inv_all(&cmd);
1174 iommu_queue_command(iommu, &cmd);
1175 iommu_completion_wait(iommu);
1178 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1180 struct iommu_cmd cmd;
1182 build_inv_irt(&cmd, devid);
1184 iommu_queue_command(iommu, &cmd);
1187 static void amd_iommu_flush_irt_all(struct amd_iommu *iommu)
1191 for (devid = 0; devid <= MAX_DEV_TABLE_ENTRIES; devid++)
1192 iommu_flush_irt(iommu, devid);
1194 iommu_completion_wait(iommu);
1197 void iommu_flush_all_caches(struct amd_iommu *iommu)
1199 if (iommu_feature(iommu, FEATURE_IA)) {
1200 amd_iommu_flush_all(iommu);
1202 amd_iommu_flush_dte_all(iommu);
1203 amd_iommu_flush_irt_all(iommu);
1204 amd_iommu_flush_tlb_all(iommu);
1209 * Command send function for flushing on-device TLB
1211 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1212 u64 address, size_t size)
1214 struct amd_iommu *iommu;
1215 struct iommu_cmd cmd;
1218 qdep = dev_data->ats.qdep;
1219 iommu = amd_iommu_rlookup_table[dev_data->devid];
1221 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1223 return iommu_queue_command(iommu, &cmd);
1227 * Command send function for invalidating a device table entry
1229 static int device_flush_dte(struct iommu_dev_data *dev_data)
1231 struct amd_iommu *iommu;
1235 iommu = amd_iommu_rlookup_table[dev_data->devid];
1236 alias = dev_data->alias;
1238 ret = iommu_flush_dte(iommu, dev_data->devid);
1239 if (!ret && alias != dev_data->devid)
1240 ret = iommu_flush_dte(iommu, alias);
1244 if (dev_data->ats.enabled)
1245 ret = device_flush_iotlb(dev_data, 0, ~0UL);
1251 * TLB invalidation function which is called from the mapping functions.
1252 * It invalidates a single PTE if the range to flush is within a single
1253 * page. Otherwise it flushes the whole TLB of the IOMMU.
1255 static void __domain_flush_pages(struct protection_domain *domain,
1256 u64 address, size_t size, int pde)
1258 struct iommu_dev_data *dev_data;
1259 struct iommu_cmd cmd;
1262 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1264 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
1265 if (!domain->dev_iommu[i])
1269 * Devices of this domain are behind this IOMMU
1270 * We need a TLB flush
1272 ret |= iommu_queue_command(amd_iommus[i], &cmd);
1275 list_for_each_entry(dev_data, &domain->dev_list, list) {
1277 if (!dev_data->ats.enabled)
1280 ret |= device_flush_iotlb(dev_data, address, size);
1286 static void domain_flush_pages(struct protection_domain *domain,
1287 u64 address, size_t size)
1289 __domain_flush_pages(domain, address, size, 0);
1292 /* Flush the whole IO/TLB for a given protection domain */
1293 static void domain_flush_tlb(struct protection_domain *domain)
1295 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
1298 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1299 static void domain_flush_tlb_pde(struct protection_domain *domain)
1301 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1304 static void domain_flush_complete(struct protection_domain *domain)
1308 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
1309 if (domain && !domain->dev_iommu[i])
1313 * Devices of this domain are behind this IOMMU
1314 * We need to wait for completion of all commands.
1316 iommu_completion_wait(amd_iommus[i]);
1322 * This function flushes the DTEs for all devices in domain
1324 static void domain_flush_devices(struct protection_domain *domain)
1326 struct iommu_dev_data *dev_data;
1328 list_for_each_entry(dev_data, &domain->dev_list, list)
1329 device_flush_dte(dev_data);
1332 /****************************************************************************
1334 * The functions below are used the create the page table mappings for
1335 * unity mapped regions.
1337 ****************************************************************************/
1340 * This function is used to add another level to an IO page table. Adding
1341 * another level increases the size of the address space by 9 bits to a size up
1344 static void increase_address_space(struct protection_domain *domain,
1347 unsigned long flags;
1350 pte = (void *)get_zeroed_page(gfp);
1354 spin_lock_irqsave(&domain->lock, flags);
1356 if (WARN_ON_ONCE(domain->mode == PAGE_MODE_6_LEVEL))
1357 /* address space already 64 bit large */
1360 *pte = PM_LEVEL_PDE(domain->mode,
1361 iommu_virt_to_phys(domain->pt_root));
1362 domain->pt_root = pte;
1364 domain->updated = true;
1368 spin_unlock_irqrestore(&domain->lock, flags);
1369 free_page((unsigned long)pte);
1374 static u64 *alloc_pte(struct protection_domain *domain,
1375 unsigned long address,
1376 unsigned long page_size,
1383 BUG_ON(!is_power_of_2(page_size));
1385 while (address > PM_LEVEL_SIZE(domain->mode))
1386 increase_address_space(domain, gfp);
1388 level = domain->mode - 1;
1389 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1390 address = PAGE_SIZE_ALIGN(address, page_size);
1391 end_lvl = PAGE_SIZE_LEVEL(page_size);
1393 while (level > end_lvl) {
1398 if (!IOMMU_PTE_PRESENT(__pte)) {
1399 page = (u64 *)get_zeroed_page(gfp);
1403 __npte = PM_LEVEL_PDE(level, iommu_virt_to_phys(page));
1405 /* pte could have been changed somewhere. */
1406 if (cmpxchg64(pte, __pte, __npte) != __pte) {
1407 free_page((unsigned long)page);
1412 /* No level skipping support yet */
1413 if (PM_PTE_LEVEL(*pte) != level)
1418 pte = IOMMU_PTE_PAGE(*pte);
1420 if (pte_page && level == end_lvl)
1423 pte = &pte[PM_LEVEL_INDEX(level, address)];
1430 * This function checks if there is a PTE for a given dma address. If
1431 * there is one, it returns the pointer to it.
1433 static u64 *fetch_pte(struct protection_domain *domain,
1434 unsigned long address,
1435 unsigned long *page_size)
1440 if (address > PM_LEVEL_SIZE(domain->mode))
1443 level = domain->mode - 1;
1444 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1445 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1450 if (!IOMMU_PTE_PRESENT(*pte))
1454 if (PM_PTE_LEVEL(*pte) == 7 ||
1455 PM_PTE_LEVEL(*pte) == 0)
1458 /* No level skipping support yet */
1459 if (PM_PTE_LEVEL(*pte) != level)
1464 /* Walk to the next level */
1465 pte = IOMMU_PTE_PAGE(*pte);
1466 pte = &pte[PM_LEVEL_INDEX(level, address)];
1467 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1470 if (PM_PTE_LEVEL(*pte) == 0x07) {
1471 unsigned long pte_mask;
1474 * If we have a series of large PTEs, make
1475 * sure to return a pointer to the first one.
1477 *page_size = pte_mask = PTE_PAGE_SIZE(*pte);
1478 pte_mask = ~((PAGE_SIZE_PTE_COUNT(pte_mask) << 3) - 1);
1479 pte = (u64 *)(((unsigned long)pte) & pte_mask);
1486 * Generic mapping functions. It maps a physical address into a DMA
1487 * address space. It allocates the page table pages if necessary.
1488 * In the future it can be extended to a generic mapping function
1489 * supporting all features of AMD IOMMU page tables like level skipping
1490 * and full 64 bit address spaces.
1492 static int iommu_map_page(struct protection_domain *dom,
1493 unsigned long bus_addr,
1494 unsigned long phys_addr,
1495 unsigned long page_size,
1502 BUG_ON(!IS_ALIGNED(bus_addr, page_size));
1503 BUG_ON(!IS_ALIGNED(phys_addr, page_size));
1505 if (!(prot & IOMMU_PROT_MASK))
1508 count = PAGE_SIZE_PTE_COUNT(page_size);
1509 pte = alloc_pte(dom, bus_addr, page_size, NULL, gfp);
1514 for (i = 0; i < count; ++i)
1515 if (IOMMU_PTE_PRESENT(pte[i]))
1519 __pte = PAGE_SIZE_PTE(__sme_set(phys_addr), page_size);
1520 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_PR | IOMMU_PTE_FC;
1522 __pte = __sme_set(phys_addr) | IOMMU_PTE_PR | IOMMU_PTE_FC;
1524 if (prot & IOMMU_PROT_IR)
1525 __pte |= IOMMU_PTE_IR;
1526 if (prot & IOMMU_PROT_IW)
1527 __pte |= IOMMU_PTE_IW;
1529 for (i = 0; i < count; ++i)
1537 static unsigned long iommu_unmap_page(struct protection_domain *dom,
1538 unsigned long bus_addr,
1539 unsigned long page_size)
1541 unsigned long long unmapped;
1542 unsigned long unmap_size;
1545 BUG_ON(!is_power_of_2(page_size));
1549 while (unmapped < page_size) {
1551 pte = fetch_pte(dom, bus_addr, &unmap_size);
1556 count = PAGE_SIZE_PTE_COUNT(unmap_size);
1557 for (i = 0; i < count; i++)
1561 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1562 unmapped += unmap_size;
1565 BUG_ON(unmapped && !is_power_of_2(unmapped));
1570 /****************************************************************************
1572 * The next functions belong to the address allocator for the dma_ops
1573 * interface functions.
1575 ****************************************************************************/
1578 static unsigned long dma_ops_alloc_iova(struct device *dev,
1579 struct dma_ops_domain *dma_dom,
1580 unsigned int pages, u64 dma_mask)
1582 unsigned long pfn = 0;
1584 pages = __roundup_pow_of_two(pages);
1586 if (dma_mask > DMA_BIT_MASK(32))
1587 pfn = alloc_iova_fast(&dma_dom->iovad, pages,
1588 IOVA_PFN(DMA_BIT_MASK(32)));
1591 pfn = alloc_iova_fast(&dma_dom->iovad, pages, IOVA_PFN(dma_mask));
1593 return (pfn << PAGE_SHIFT);
1596 static void dma_ops_free_iova(struct dma_ops_domain *dma_dom,
1597 unsigned long address,
1600 pages = __roundup_pow_of_two(pages);
1601 address >>= PAGE_SHIFT;
1603 free_iova_fast(&dma_dom->iovad, address, pages);
1606 /****************************************************************************
1608 * The next functions belong to the domain allocation. A domain is
1609 * allocated for every IOMMU as the default domain. If device isolation
1610 * is enabled, every device get its own domain. The most important thing
1611 * about domains is the page table mapping the DMA address space they
1614 ****************************************************************************/
1617 * This function adds a protection domain to the global protection domain list
1619 static void add_domain_to_list(struct protection_domain *domain)
1621 unsigned long flags;
1623 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1624 list_add(&domain->list, &amd_iommu_pd_list);
1625 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1629 * This function removes a protection domain to the global
1630 * protection domain list
1632 static void del_domain_from_list(struct protection_domain *domain)
1634 unsigned long flags;
1636 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1637 list_del(&domain->list);
1638 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1641 static u16 domain_id_alloc(void)
1643 unsigned long flags;
1646 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1647 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1649 if (id > 0 && id < MAX_DOMAIN_ID)
1650 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1653 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1658 static void domain_id_free(int id)
1660 unsigned long flags;
1662 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1663 if (id > 0 && id < MAX_DOMAIN_ID)
1664 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1665 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1668 #define DEFINE_FREE_PT_FN(LVL, FN) \
1669 static void free_pt_##LVL (unsigned long __pt) \
1677 for (i = 0; i < 512; ++i) { \
1678 /* PTE present? */ \
1679 if (!IOMMU_PTE_PRESENT(pt[i])) \
1683 if (PM_PTE_LEVEL(pt[i]) == 0 || \
1684 PM_PTE_LEVEL(pt[i]) == 7) \
1687 p = (unsigned long)IOMMU_PTE_PAGE(pt[i]); \
1690 free_page((unsigned long)pt); \
1693 DEFINE_FREE_PT_FN(l2, free_page)
1694 DEFINE_FREE_PT_FN(l3, free_pt_l2)
1695 DEFINE_FREE_PT_FN(l4, free_pt_l3)
1696 DEFINE_FREE_PT_FN(l5, free_pt_l4)
1697 DEFINE_FREE_PT_FN(l6, free_pt_l5)
1699 static void free_pagetable(struct protection_domain *domain)
1701 unsigned long root = (unsigned long)domain->pt_root;
1703 switch (domain->mode) {
1704 case PAGE_MODE_NONE:
1706 case PAGE_MODE_1_LEVEL:
1709 case PAGE_MODE_2_LEVEL:
1712 case PAGE_MODE_3_LEVEL:
1715 case PAGE_MODE_4_LEVEL:
1718 case PAGE_MODE_5_LEVEL:
1721 case PAGE_MODE_6_LEVEL:
1729 static void free_gcr3_tbl_level1(u64 *tbl)
1734 for (i = 0; i < 512; ++i) {
1735 if (!(tbl[i] & GCR3_VALID))
1738 ptr = iommu_phys_to_virt(tbl[i] & PAGE_MASK);
1740 free_page((unsigned long)ptr);
1744 static void free_gcr3_tbl_level2(u64 *tbl)
1749 for (i = 0; i < 512; ++i) {
1750 if (!(tbl[i] & GCR3_VALID))
1753 ptr = iommu_phys_to_virt(tbl[i] & PAGE_MASK);
1755 free_gcr3_tbl_level1(ptr);
1759 static void free_gcr3_table(struct protection_domain *domain)
1761 if (domain->glx == 2)
1762 free_gcr3_tbl_level2(domain->gcr3_tbl);
1763 else if (domain->glx == 1)
1764 free_gcr3_tbl_level1(domain->gcr3_tbl);
1766 BUG_ON(domain->glx != 0);
1768 free_page((unsigned long)domain->gcr3_tbl);
1771 static void dma_ops_domain_flush_tlb(struct dma_ops_domain *dom)
1773 domain_flush_tlb(&dom->domain);
1774 domain_flush_complete(&dom->domain);
1777 static void iova_domain_flush_tlb(struct iova_domain *iovad)
1779 struct dma_ops_domain *dom;
1781 dom = container_of(iovad, struct dma_ops_domain, iovad);
1783 dma_ops_domain_flush_tlb(dom);
1787 * Free a domain, only used if something went wrong in the
1788 * allocation path and we need to free an already allocated page table
1790 static void dma_ops_domain_free(struct dma_ops_domain *dom)
1795 del_domain_from_list(&dom->domain);
1797 put_iova_domain(&dom->iovad);
1799 free_pagetable(&dom->domain);
1802 domain_id_free(dom->domain.id);
1808 * Allocates a new protection domain usable for the dma_ops functions.
1809 * It also initializes the page table and the address allocator data
1810 * structures required for the dma_ops interface
1812 static struct dma_ops_domain *dma_ops_domain_alloc(void)
1814 struct dma_ops_domain *dma_dom;
1816 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
1820 if (protection_domain_init(&dma_dom->domain))
1823 dma_dom->domain.mode = PAGE_MODE_3_LEVEL;
1824 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
1825 dma_dom->domain.flags = PD_DMA_OPS_MASK;
1826 if (!dma_dom->domain.pt_root)
1829 init_iova_domain(&dma_dom->iovad, PAGE_SIZE,
1830 IOVA_START_PFN, DMA_32BIT_PFN);
1832 if (init_iova_flush_queue(&dma_dom->iovad, iova_domain_flush_tlb, NULL))
1835 /* Initialize reserved ranges */
1836 copy_reserved_iova(&reserved_iova_ranges, &dma_dom->iovad);
1838 add_domain_to_list(&dma_dom->domain);
1843 dma_ops_domain_free(dma_dom);
1849 * little helper function to check whether a given protection domain is a
1852 static bool dma_ops_domain(struct protection_domain *domain)
1854 return domain->flags & PD_DMA_OPS_MASK;
1857 static void set_dte_entry(u16 devid, struct protection_domain *domain, bool ats)
1863 if (domain->mode != PAGE_MODE_NONE)
1864 pte_root = iommu_virt_to_phys(domain->pt_root);
1866 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
1867 << DEV_ENTRY_MODE_SHIFT;
1868 pte_root |= DTE_FLAG_IR | DTE_FLAG_IW | DTE_FLAG_V | DTE_FLAG_TV;
1870 flags = amd_iommu_dev_table[devid].data[1];
1873 flags |= DTE_FLAG_IOTLB;
1875 if (domain->flags & PD_IOMMUV2_MASK) {
1876 u64 gcr3 = iommu_virt_to_phys(domain->gcr3_tbl);
1877 u64 glx = domain->glx;
1880 pte_root |= DTE_FLAG_GV;
1881 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
1883 /* First mask out possible old values for GCR3 table */
1884 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
1887 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
1890 /* Encode GCR3 table into DTE */
1891 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
1894 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
1897 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
1901 flags &= ~DEV_DOMID_MASK;
1902 flags |= domain->id;
1904 old_domid = amd_iommu_dev_table[devid].data[1] & DEV_DOMID_MASK;
1905 amd_iommu_dev_table[devid].data[1] = flags;
1906 amd_iommu_dev_table[devid].data[0] = pte_root;
1909 * A kdump kernel might be replacing a domain ID that was copied from
1910 * the previous kernel--if so, it needs to flush the translation cache
1911 * entries for the old domain ID that is being overwritten
1914 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
1916 amd_iommu_flush_tlb_domid(iommu, old_domid);
1920 static void clear_dte_entry(u16 devid)
1922 /* remove entry from the device table seen by the hardware */
1923 amd_iommu_dev_table[devid].data[0] = DTE_FLAG_V | DTE_FLAG_TV;
1924 amd_iommu_dev_table[devid].data[1] &= DTE_FLAG_MASK;
1926 amd_iommu_apply_erratum_63(devid);
1929 static void do_attach(struct iommu_dev_data *dev_data,
1930 struct protection_domain *domain)
1932 struct amd_iommu *iommu;
1936 iommu = amd_iommu_rlookup_table[dev_data->devid];
1937 alias = dev_data->alias;
1938 ats = dev_data->ats.enabled;
1940 /* Update data structures */
1941 dev_data->domain = domain;
1942 list_add(&dev_data->list, &domain->dev_list);
1944 /* Do reference counting */
1945 domain->dev_iommu[iommu->index] += 1;
1946 domain->dev_cnt += 1;
1948 /* Update device table */
1949 set_dte_entry(dev_data->devid, domain, ats);
1950 if (alias != dev_data->devid)
1951 set_dte_entry(alias, domain, ats);
1953 device_flush_dte(dev_data);
1956 static void do_detach(struct iommu_dev_data *dev_data)
1958 struct protection_domain *domain = dev_data->domain;
1959 struct amd_iommu *iommu;
1963 * First check if the device is still attached. It might already
1964 * be detached from its domain because the generic
1965 * iommu_detach_group code detached it and we try again here in
1966 * our alias handling.
1968 if (!dev_data->domain)
1971 iommu = amd_iommu_rlookup_table[dev_data->devid];
1972 alias = dev_data->alias;
1974 /* Update data structures */
1975 dev_data->domain = NULL;
1976 list_del(&dev_data->list);
1977 clear_dte_entry(dev_data->devid);
1978 if (alias != dev_data->devid)
1979 clear_dte_entry(alias);
1981 /* Flush the DTE entry */
1982 device_flush_dte(dev_data);
1985 domain_flush_tlb_pde(domain);
1987 /* Wait for the flushes to finish */
1988 domain_flush_complete(domain);
1990 /* decrease reference counters - needs to happen after the flushes */
1991 domain->dev_iommu[iommu->index] -= 1;
1992 domain->dev_cnt -= 1;
1996 * If a device is not yet associated with a domain, this function does
1997 * assigns it visible for the hardware
1999 static int __attach_device(struct iommu_dev_data *dev_data,
2000 struct protection_domain *domain)
2005 * Must be called with IRQs disabled. Warn here to detect early
2008 WARN_ON(!irqs_disabled());
2011 spin_lock(&domain->lock);
2014 if (dev_data->domain != NULL)
2017 /* Attach alias group root */
2018 do_attach(dev_data, domain);
2025 spin_unlock(&domain->lock);
2031 static void pdev_iommuv2_disable(struct pci_dev *pdev)
2033 pci_disable_ats(pdev);
2034 pci_disable_pri(pdev);
2035 pci_disable_pasid(pdev);
2038 /* FIXME: Change generic reset-function to do the same */
2039 static int pri_reset_while_enabled(struct pci_dev *pdev)
2044 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2048 pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
2049 control |= PCI_PRI_CTRL_RESET;
2050 pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
2055 static int pdev_iommuv2_enable(struct pci_dev *pdev)
2060 /* FIXME: Hardcode number of outstanding requests for now */
2062 if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
2064 reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
2066 /* Only allow access to user-accessible pages */
2067 ret = pci_enable_pasid(pdev, 0);
2071 /* First reset the PRI state of the device */
2072 ret = pci_reset_pri(pdev);
2077 ret = pci_enable_pri(pdev, reqs);
2082 ret = pri_reset_while_enabled(pdev);
2087 ret = pci_enable_ats(pdev, PAGE_SHIFT);
2094 pci_disable_pri(pdev);
2095 pci_disable_pasid(pdev);
2100 /* FIXME: Move this to PCI code */
2101 #define PCI_PRI_TLP_OFF (1 << 15)
2103 static bool pci_pri_tlp_required(struct pci_dev *pdev)
2108 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2112 pci_read_config_word(pdev, pos + PCI_PRI_STATUS, &status);
2114 return (status & PCI_PRI_TLP_OFF) ? true : false;
2118 * If a device is not yet associated with a domain, this function
2119 * assigns it visible for the hardware
2121 static int attach_device(struct device *dev,
2122 struct protection_domain *domain)
2124 struct pci_dev *pdev;
2125 struct iommu_dev_data *dev_data;
2126 unsigned long flags;
2129 dev_data = get_dev_data(dev);
2131 if (!dev_is_pci(dev))
2132 goto skip_ats_check;
2134 pdev = to_pci_dev(dev);
2135 if (domain->flags & PD_IOMMUV2_MASK) {
2136 if (!dev_data->passthrough)
2139 if (dev_data->iommu_v2) {
2140 if (pdev_iommuv2_enable(pdev) != 0)
2143 dev_data->ats.enabled = true;
2144 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2145 dev_data->pri_tlp = pci_pri_tlp_required(pdev);
2147 } else if (amd_iommu_iotlb_sup &&
2148 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
2149 dev_data->ats.enabled = true;
2150 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2154 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2155 ret = __attach_device(dev_data, domain);
2156 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2159 * We might boot into a crash-kernel here. The crashed kernel
2160 * left the caches in the IOMMU dirty. So we have to flush
2161 * here to evict all dirty stuff.
2163 domain_flush_tlb_pde(domain);
2165 domain_flush_complete(domain);
2171 * Removes a device from a protection domain (unlocked)
2173 static void __detach_device(struct iommu_dev_data *dev_data)
2175 struct protection_domain *domain;
2178 * Must be called with IRQs disabled. Warn here to detect early
2181 WARN_ON(!irqs_disabled());
2183 if (WARN_ON(!dev_data->domain))
2186 domain = dev_data->domain;
2188 spin_lock(&domain->lock);
2190 do_detach(dev_data);
2192 spin_unlock(&domain->lock);
2196 * Removes a device from a protection domain (with devtable_lock held)
2198 static void detach_device(struct device *dev)
2200 struct protection_domain *domain;
2201 struct iommu_dev_data *dev_data;
2202 unsigned long flags;
2204 dev_data = get_dev_data(dev);
2205 domain = dev_data->domain;
2207 /* lock device table */
2208 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2209 __detach_device(dev_data);
2210 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2212 if (!dev_is_pci(dev))
2215 if (domain->flags & PD_IOMMUV2_MASK && dev_data->iommu_v2)
2216 pdev_iommuv2_disable(to_pci_dev(dev));
2217 else if (dev_data->ats.enabled)
2218 pci_disable_ats(to_pci_dev(dev));
2220 dev_data->ats.enabled = false;
2223 static int amd_iommu_add_device(struct device *dev)
2225 struct iommu_dev_data *dev_data;
2226 struct iommu_domain *domain;
2227 struct amd_iommu *iommu;
2230 if (!check_device(dev) || get_dev_data(dev))
2233 devid = get_device_id(dev);
2237 iommu = amd_iommu_rlookup_table[devid];
2239 ret = iommu_init_device(dev);
2241 if (ret != -ENOTSUPP)
2242 pr_err("Failed to initialize device %s - trying to proceed anyway\n",
2245 iommu_ignore_device(dev);
2246 dev->dma_ops = &nommu_dma_ops;
2249 init_iommu_group(dev);
2251 dev_data = get_dev_data(dev);
2255 if (iommu_pass_through || dev_data->iommu_v2)
2256 iommu_request_dm_for_dev(dev);
2258 /* Domains are initialized for this device - have a look what we ended up with */
2259 domain = iommu_get_domain_for_dev(dev);
2260 if (domain->type == IOMMU_DOMAIN_IDENTITY)
2261 dev_data->passthrough = true;
2263 dev->dma_ops = &amd_iommu_dma_ops;
2266 iommu_completion_wait(iommu);
2271 static void amd_iommu_remove_device(struct device *dev)
2273 struct amd_iommu *iommu;
2276 if (!check_device(dev))
2279 devid = get_device_id(dev);
2283 iommu = amd_iommu_rlookup_table[devid];
2285 iommu_uninit_device(dev);
2286 iommu_completion_wait(iommu);
2289 static struct iommu_group *amd_iommu_device_group(struct device *dev)
2291 if (dev_is_pci(dev))
2292 return pci_device_group(dev);
2294 return acpihid_device_group(dev);
2297 /*****************************************************************************
2299 * The next functions belong to the dma_ops mapping/unmapping code.
2301 *****************************************************************************/
2304 * In the dma_ops path we only have the struct device. This function
2305 * finds the corresponding IOMMU, the protection domain and the
2306 * requestor id for a given device.
2307 * If the device is not yet associated with a domain this is also done
2310 static struct protection_domain *get_domain(struct device *dev)
2312 struct protection_domain *domain;
2313 struct iommu_domain *io_domain;
2315 if (!check_device(dev))
2316 return ERR_PTR(-EINVAL);
2318 domain = get_dev_data(dev)->domain;
2319 if (domain == NULL && get_dev_data(dev)->defer_attach) {
2320 get_dev_data(dev)->defer_attach = false;
2321 io_domain = iommu_get_domain_for_dev(dev);
2322 domain = to_pdomain(io_domain);
2323 attach_device(dev, domain);
2326 return ERR_PTR(-EBUSY);
2328 if (!dma_ops_domain(domain))
2329 return ERR_PTR(-EBUSY);
2334 static void update_device_table(struct protection_domain *domain)
2336 struct iommu_dev_data *dev_data;
2338 list_for_each_entry(dev_data, &domain->dev_list, list) {
2339 set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled);
2341 if (dev_data->devid == dev_data->alias)
2344 /* There is an alias, update device table entry for it */
2345 set_dte_entry(dev_data->alias, domain, dev_data->ats.enabled);
2349 static void update_domain(struct protection_domain *domain)
2351 if (!domain->updated)
2354 update_device_table(domain);
2356 domain_flush_devices(domain);
2357 domain_flush_tlb_pde(domain);
2359 domain->updated = false;
2362 static int dir2prot(enum dma_data_direction direction)
2364 if (direction == DMA_TO_DEVICE)
2365 return IOMMU_PROT_IR;
2366 else if (direction == DMA_FROM_DEVICE)
2367 return IOMMU_PROT_IW;
2368 else if (direction == DMA_BIDIRECTIONAL)
2369 return IOMMU_PROT_IW | IOMMU_PROT_IR;
2375 * This function contains common code for mapping of a physically
2376 * contiguous memory region into DMA address space. It is used by all
2377 * mapping functions provided with this IOMMU driver.
2378 * Must be called with the domain lock held.
2380 static dma_addr_t __map_single(struct device *dev,
2381 struct dma_ops_domain *dma_dom,
2384 enum dma_data_direction direction,
2387 dma_addr_t offset = paddr & ~PAGE_MASK;
2388 dma_addr_t address, start, ret;
2393 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
2396 address = dma_ops_alloc_iova(dev, dma_dom, pages, dma_mask);
2397 if (address == AMD_IOMMU_MAPPING_ERROR)
2400 prot = dir2prot(direction);
2403 for (i = 0; i < pages; ++i) {
2404 ret = iommu_map_page(&dma_dom->domain, start, paddr,
2405 PAGE_SIZE, prot, GFP_ATOMIC);
2414 if (unlikely(amd_iommu_np_cache)) {
2415 domain_flush_pages(&dma_dom->domain, address, size);
2416 domain_flush_complete(&dma_dom->domain);
2424 for (--i; i >= 0; --i) {
2426 iommu_unmap_page(&dma_dom->domain, start, PAGE_SIZE);
2429 domain_flush_tlb(&dma_dom->domain);
2430 domain_flush_complete(&dma_dom->domain);
2432 dma_ops_free_iova(dma_dom, address, pages);
2434 return AMD_IOMMU_MAPPING_ERROR;
2438 * Does the reverse of the __map_single function. Must be called with
2439 * the domain lock held too
2441 static void __unmap_single(struct dma_ops_domain *dma_dom,
2442 dma_addr_t dma_addr,
2446 dma_addr_t flush_addr;
2447 dma_addr_t i, start;
2450 flush_addr = dma_addr;
2451 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
2452 dma_addr &= PAGE_MASK;
2455 for (i = 0; i < pages; ++i) {
2456 iommu_unmap_page(&dma_dom->domain, start, PAGE_SIZE);
2460 if (amd_iommu_unmap_flush) {
2461 domain_flush_tlb(&dma_dom->domain);
2462 domain_flush_complete(&dma_dom->domain);
2463 dma_ops_free_iova(dma_dom, dma_addr, pages);
2465 pages = __roundup_pow_of_two(pages);
2466 queue_iova(&dma_dom->iovad, dma_addr >> PAGE_SHIFT, pages, 0);
2471 * The exported map_single function for dma_ops.
2473 static dma_addr_t map_page(struct device *dev, struct page *page,
2474 unsigned long offset, size_t size,
2475 enum dma_data_direction dir,
2476 unsigned long attrs)
2478 phys_addr_t paddr = page_to_phys(page) + offset;
2479 struct protection_domain *domain;
2480 struct dma_ops_domain *dma_dom;
2483 domain = get_domain(dev);
2484 if (PTR_ERR(domain) == -EINVAL)
2485 return (dma_addr_t)paddr;
2486 else if (IS_ERR(domain))
2487 return AMD_IOMMU_MAPPING_ERROR;
2489 dma_mask = *dev->dma_mask;
2490 dma_dom = to_dma_ops_domain(domain);
2492 return __map_single(dev, dma_dom, paddr, size, dir, dma_mask);
2496 * The exported unmap_single function for dma_ops.
2498 static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2499 enum dma_data_direction dir, unsigned long attrs)
2501 struct protection_domain *domain;
2502 struct dma_ops_domain *dma_dom;
2504 domain = get_domain(dev);
2508 dma_dom = to_dma_ops_domain(domain);
2510 __unmap_single(dma_dom, dma_addr, size, dir);
2513 static int sg_num_pages(struct device *dev,
2514 struct scatterlist *sglist,
2517 unsigned long mask, boundary_size;
2518 struct scatterlist *s;
2521 mask = dma_get_seg_boundary(dev);
2522 boundary_size = mask + 1 ? ALIGN(mask + 1, PAGE_SIZE) >> PAGE_SHIFT :
2523 1UL << (BITS_PER_LONG - PAGE_SHIFT);
2525 for_each_sg(sglist, s, nelems, i) {
2528 s->dma_address = npages << PAGE_SHIFT;
2529 p = npages % boundary_size;
2530 n = iommu_num_pages(sg_phys(s), s->length, PAGE_SIZE);
2531 if (p + n > boundary_size)
2532 npages += boundary_size - p;
2540 * The exported map_sg function for dma_ops (handles scatter-gather
2543 static int map_sg(struct device *dev, struct scatterlist *sglist,
2544 int nelems, enum dma_data_direction direction,
2545 unsigned long attrs)
2547 int mapped_pages = 0, npages = 0, prot = 0, i;
2548 struct protection_domain *domain;
2549 struct dma_ops_domain *dma_dom;
2550 struct scatterlist *s;
2551 unsigned long address;
2554 domain = get_domain(dev);
2558 dma_dom = to_dma_ops_domain(domain);
2559 dma_mask = *dev->dma_mask;
2561 npages = sg_num_pages(dev, sglist, nelems);
2563 address = dma_ops_alloc_iova(dev, dma_dom, npages, dma_mask);
2564 if (address == AMD_IOMMU_MAPPING_ERROR)
2567 prot = dir2prot(direction);
2569 /* Map all sg entries */
2570 for_each_sg(sglist, s, nelems, i) {
2571 int j, pages = iommu_num_pages(sg_phys(s), s->length, PAGE_SIZE);
2573 for (j = 0; j < pages; ++j) {
2574 unsigned long bus_addr, phys_addr;
2577 bus_addr = address + s->dma_address + (j << PAGE_SHIFT);
2578 phys_addr = (sg_phys(s) & PAGE_MASK) + (j << PAGE_SHIFT);
2579 ret = iommu_map_page(domain, bus_addr, phys_addr,
2581 GFP_ATOMIC | __GFP_NOWARN);
2589 /* Everything is mapped - write the right values into s->dma_address */
2590 for_each_sg(sglist, s, nelems, i) {
2592 * Add in the remaining piece of the scatter-gather offset that
2593 * was masked out when we were determining the physical address
2594 * via (sg_phys(s) & PAGE_MASK) earlier.
2596 s->dma_address += address + (s->offset & ~PAGE_MASK);
2597 s->dma_length = s->length;
2603 pr_err("%s: IOMMU mapping error in map_sg (io-pages: %d)\n",
2604 dev_name(dev), npages);
2606 for_each_sg(sglist, s, nelems, i) {
2607 int j, pages = iommu_num_pages(sg_phys(s), s->length, PAGE_SIZE);
2609 for (j = 0; j < pages; ++j) {
2610 unsigned long bus_addr;
2612 bus_addr = address + s->dma_address + (j << PAGE_SHIFT);
2613 iommu_unmap_page(domain, bus_addr, PAGE_SIZE);
2615 if (--mapped_pages == 0)
2621 free_iova_fast(&dma_dom->iovad, address >> PAGE_SHIFT, npages);
2628 * The exported map_sg function for dma_ops (handles scatter-gather
2631 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2632 int nelems, enum dma_data_direction dir,
2633 unsigned long attrs)
2635 struct protection_domain *domain;
2636 struct dma_ops_domain *dma_dom;
2637 unsigned long startaddr;
2640 domain = get_domain(dev);
2644 startaddr = sg_dma_address(sglist) & PAGE_MASK;
2645 dma_dom = to_dma_ops_domain(domain);
2646 npages = sg_num_pages(dev, sglist, nelems);
2648 __unmap_single(dma_dom, startaddr, npages << PAGE_SHIFT, dir);
2652 * The exported alloc_coherent function for dma_ops.
2654 static void *alloc_coherent(struct device *dev, size_t size,
2655 dma_addr_t *dma_addr, gfp_t flag,
2656 unsigned long attrs)
2658 u64 dma_mask = dev->coherent_dma_mask;
2659 struct protection_domain *domain;
2660 struct dma_ops_domain *dma_dom;
2663 domain = get_domain(dev);
2664 if (PTR_ERR(domain) == -EINVAL) {
2665 page = alloc_pages(flag, get_order(size));
2666 *dma_addr = page_to_phys(page);
2667 return page_address(page);
2668 } else if (IS_ERR(domain))
2671 dma_dom = to_dma_ops_domain(domain);
2672 size = PAGE_ALIGN(size);
2673 dma_mask = dev->coherent_dma_mask;
2674 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
2677 page = alloc_pages(flag | __GFP_NOWARN, get_order(size));
2679 if (!gfpflags_allow_blocking(flag))
2682 page = dma_alloc_from_contiguous(dev, size >> PAGE_SHIFT,
2683 get_order(size), flag);
2689 dma_mask = *dev->dma_mask;
2691 *dma_addr = __map_single(dev, dma_dom, page_to_phys(page),
2692 size, DMA_BIDIRECTIONAL, dma_mask);
2694 if (*dma_addr == AMD_IOMMU_MAPPING_ERROR)
2697 return page_address(page);
2701 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2702 __free_pages(page, get_order(size));
2708 * The exported free_coherent function for dma_ops.
2710 static void free_coherent(struct device *dev, size_t size,
2711 void *virt_addr, dma_addr_t dma_addr,
2712 unsigned long attrs)
2714 struct protection_domain *domain;
2715 struct dma_ops_domain *dma_dom;
2718 page = virt_to_page(virt_addr);
2719 size = PAGE_ALIGN(size);
2721 domain = get_domain(dev);
2725 dma_dom = to_dma_ops_domain(domain);
2727 __unmap_single(dma_dom, dma_addr, size, DMA_BIDIRECTIONAL);
2730 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2731 __free_pages(page, get_order(size));
2735 * This function is called by the DMA layer to find out if we can handle a
2736 * particular device. It is part of the dma_ops.
2738 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
2740 if (!x86_dma_supported(dev, mask))
2742 return check_device(dev);
2745 static int amd_iommu_mapping_error(struct device *dev, dma_addr_t dma_addr)
2747 return dma_addr == AMD_IOMMU_MAPPING_ERROR;
2750 static const struct dma_map_ops amd_iommu_dma_ops = {
2751 .alloc = alloc_coherent,
2752 .free = free_coherent,
2753 .map_page = map_page,
2754 .unmap_page = unmap_page,
2756 .unmap_sg = unmap_sg,
2757 .dma_supported = amd_iommu_dma_supported,
2758 .mapping_error = amd_iommu_mapping_error,
2761 static int init_reserved_iova_ranges(void)
2763 struct pci_dev *pdev = NULL;
2766 init_iova_domain(&reserved_iova_ranges, PAGE_SIZE,
2767 IOVA_START_PFN, DMA_32BIT_PFN);
2769 lockdep_set_class(&reserved_iova_ranges.iova_rbtree_lock,
2770 &reserved_rbtree_key);
2772 /* MSI memory range */
2773 val = reserve_iova(&reserved_iova_ranges,
2774 IOVA_PFN(MSI_RANGE_START), IOVA_PFN(MSI_RANGE_END));
2776 pr_err("Reserving MSI range failed\n");
2780 /* HT memory range */
2781 val = reserve_iova(&reserved_iova_ranges,
2782 IOVA_PFN(HT_RANGE_START), IOVA_PFN(HT_RANGE_END));
2784 pr_err("Reserving HT range failed\n");
2789 * Memory used for PCI resources
2790 * FIXME: Check whether we can reserve the PCI-hole completly
2792 for_each_pci_dev(pdev) {
2795 for (i = 0; i < PCI_NUM_RESOURCES; ++i) {
2796 struct resource *r = &pdev->resource[i];
2798 if (!(r->flags & IORESOURCE_MEM))
2801 val = reserve_iova(&reserved_iova_ranges,
2805 pr_err("Reserve pci-resource range failed\n");
2814 int __init amd_iommu_init_api(void)
2818 ret = iova_cache_get();
2822 ret = init_reserved_iova_ranges();
2826 err = bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
2829 #ifdef CONFIG_ARM_AMBA
2830 err = bus_set_iommu(&amba_bustype, &amd_iommu_ops);
2834 err = bus_set_iommu(&platform_bus_type, &amd_iommu_ops);
2841 int __init amd_iommu_init_dma_ops(void)
2843 swiotlb = (iommu_pass_through || sme_me_mask) ? 1 : 0;
2847 * In case we don't initialize SWIOTLB (actually the common case
2848 * when AMD IOMMU is enabled and SME is not active), make sure there
2849 * are global dma_ops set as a fall-back for devices not handled by
2850 * this driver (for example non-PCI devices). When SME is active,
2851 * make sure that swiotlb variable remains set so the global dma_ops
2852 * continue to be SWIOTLB.
2855 dma_ops = &nommu_dma_ops;
2857 if (amd_iommu_unmap_flush)
2858 pr_info("AMD-Vi: IO/TLB flush on unmap enabled\n");
2860 pr_info("AMD-Vi: Lazy IO/TLB flushing enabled\n");
2866 /*****************************************************************************
2868 * The following functions belong to the exported interface of AMD IOMMU
2870 * This interface allows access to lower level functions of the IOMMU
2871 * like protection domain handling and assignement of devices to domains
2872 * which is not possible with the dma_ops interface.
2874 *****************************************************************************/
2876 static void cleanup_domain(struct protection_domain *domain)
2878 struct iommu_dev_data *entry;
2879 unsigned long flags;
2881 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2883 while (!list_empty(&domain->dev_list)) {
2884 entry = list_first_entry(&domain->dev_list,
2885 struct iommu_dev_data, list);
2886 __detach_device(entry);
2889 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2892 static void protection_domain_free(struct protection_domain *domain)
2897 del_domain_from_list(domain);
2900 domain_id_free(domain->id);
2905 static int protection_domain_init(struct protection_domain *domain)
2907 spin_lock_init(&domain->lock);
2908 mutex_init(&domain->api_lock);
2909 domain->id = domain_id_alloc();
2912 INIT_LIST_HEAD(&domain->dev_list);
2917 static struct protection_domain *protection_domain_alloc(void)
2919 struct protection_domain *domain;
2921 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2925 if (protection_domain_init(domain))
2928 add_domain_to_list(domain);
2938 static struct iommu_domain *amd_iommu_domain_alloc(unsigned type)
2940 struct protection_domain *pdomain;
2941 struct dma_ops_domain *dma_domain;
2944 case IOMMU_DOMAIN_UNMANAGED:
2945 pdomain = protection_domain_alloc();
2949 pdomain->mode = PAGE_MODE_3_LEVEL;
2950 pdomain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2951 if (!pdomain->pt_root) {
2952 protection_domain_free(pdomain);
2956 pdomain->domain.geometry.aperture_start = 0;
2957 pdomain->domain.geometry.aperture_end = ~0ULL;
2958 pdomain->domain.geometry.force_aperture = true;
2961 case IOMMU_DOMAIN_DMA:
2962 dma_domain = dma_ops_domain_alloc();
2964 pr_err("AMD-Vi: Failed to allocate\n");
2967 pdomain = &dma_domain->domain;
2969 case IOMMU_DOMAIN_IDENTITY:
2970 pdomain = protection_domain_alloc();
2974 pdomain->mode = PAGE_MODE_NONE;
2980 return &pdomain->domain;
2983 static void amd_iommu_domain_free(struct iommu_domain *dom)
2985 struct protection_domain *domain;
2986 struct dma_ops_domain *dma_dom;
2988 domain = to_pdomain(dom);
2990 if (domain->dev_cnt > 0)
2991 cleanup_domain(domain);
2993 BUG_ON(domain->dev_cnt != 0);
2998 switch (dom->type) {
2999 case IOMMU_DOMAIN_DMA:
3000 /* Now release the domain */
3001 dma_dom = to_dma_ops_domain(domain);
3002 dma_ops_domain_free(dma_dom);
3005 if (domain->mode != PAGE_MODE_NONE)
3006 free_pagetable(domain);
3008 if (domain->flags & PD_IOMMUV2_MASK)
3009 free_gcr3_table(domain);
3011 protection_domain_free(domain);
3016 static void amd_iommu_detach_device(struct iommu_domain *dom,
3019 struct iommu_dev_data *dev_data = dev->archdata.iommu;
3020 struct amd_iommu *iommu;
3023 if (!check_device(dev))
3026 devid = get_device_id(dev);
3030 if (dev_data->domain != NULL)
3033 iommu = amd_iommu_rlookup_table[devid];
3037 #ifdef CONFIG_IRQ_REMAP
3038 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) &&
3039 (dom->type == IOMMU_DOMAIN_UNMANAGED))
3040 dev_data->use_vapic = 0;
3043 iommu_completion_wait(iommu);
3046 static int amd_iommu_attach_device(struct iommu_domain *dom,
3049 struct protection_domain *domain = to_pdomain(dom);
3050 struct iommu_dev_data *dev_data;
3051 struct amd_iommu *iommu;
3054 if (!check_device(dev))
3057 dev_data = dev->archdata.iommu;
3059 iommu = amd_iommu_rlookup_table[dev_data->devid];
3063 if (dev_data->domain)
3066 ret = attach_device(dev, domain);
3068 #ifdef CONFIG_IRQ_REMAP
3069 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
3070 if (dom->type == IOMMU_DOMAIN_UNMANAGED)
3071 dev_data->use_vapic = 1;
3073 dev_data->use_vapic = 0;
3077 iommu_completion_wait(iommu);
3082 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
3083 phys_addr_t paddr, size_t page_size, int iommu_prot)
3085 struct protection_domain *domain = to_pdomain(dom);
3089 if (domain->mode == PAGE_MODE_NONE)
3092 if (iommu_prot & IOMMU_READ)
3093 prot |= IOMMU_PROT_IR;
3094 if (iommu_prot & IOMMU_WRITE)
3095 prot |= IOMMU_PROT_IW;
3097 mutex_lock(&domain->api_lock);
3098 ret = iommu_map_page(domain, iova, paddr, page_size, prot, GFP_KERNEL);
3099 mutex_unlock(&domain->api_lock);
3104 static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
3107 struct protection_domain *domain = to_pdomain(dom);
3110 if (domain->mode == PAGE_MODE_NONE)
3113 mutex_lock(&domain->api_lock);
3114 unmap_size = iommu_unmap_page(domain, iova, page_size);
3115 mutex_unlock(&domain->api_lock);
3117 domain_flush_tlb_pde(domain);
3118 domain_flush_complete(domain);
3123 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
3126 struct protection_domain *domain = to_pdomain(dom);
3127 unsigned long offset_mask, pte_pgsize;
3130 if (domain->mode == PAGE_MODE_NONE)
3133 pte = fetch_pte(domain, iova, &pte_pgsize);
3135 if (!pte || !IOMMU_PTE_PRESENT(*pte))
3138 offset_mask = pte_pgsize - 1;
3139 __pte = __sme_clr(*pte & PM_ADDR_MASK);
3141 return (__pte & ~offset_mask) | (iova & offset_mask);
3144 static bool amd_iommu_capable(enum iommu_cap cap)
3147 case IOMMU_CAP_CACHE_COHERENCY:
3149 case IOMMU_CAP_INTR_REMAP:
3150 return (irq_remapping_enabled == 1);
3151 case IOMMU_CAP_NOEXEC:
3158 static void amd_iommu_get_resv_regions(struct device *dev,
3159 struct list_head *head)
3161 struct iommu_resv_region *region;
3162 struct unity_map_entry *entry;
3165 devid = get_device_id(dev);
3169 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
3173 if (devid < entry->devid_start || devid > entry->devid_end)
3176 type = IOMMU_RESV_DIRECT;
3177 length = entry->address_end - entry->address_start;
3178 if (entry->prot & IOMMU_PROT_IR)
3180 if (entry->prot & IOMMU_PROT_IW)
3181 prot |= IOMMU_WRITE;
3182 if (entry->prot & IOMMU_UNITY_MAP_FLAG_EXCL_RANGE)
3183 /* Exclusion range */
3184 type = IOMMU_RESV_RESERVED;
3186 region = iommu_alloc_resv_region(entry->address_start,
3187 length, prot, type);
3189 pr_err("Out of memory allocating dm-regions for %s\n",
3193 list_add_tail(®ion->list, head);
3196 region = iommu_alloc_resv_region(MSI_RANGE_START,
3197 MSI_RANGE_END - MSI_RANGE_START + 1,
3201 list_add_tail(®ion->list, head);
3203 region = iommu_alloc_resv_region(HT_RANGE_START,
3204 HT_RANGE_END - HT_RANGE_START + 1,
3205 0, IOMMU_RESV_RESERVED);
3208 list_add_tail(®ion->list, head);
3211 static void amd_iommu_put_resv_regions(struct device *dev,
3212 struct list_head *head)
3214 struct iommu_resv_region *entry, *next;
3216 list_for_each_entry_safe(entry, next, head, list)
3220 static void amd_iommu_apply_resv_region(struct device *dev,
3221 struct iommu_domain *domain,
3222 struct iommu_resv_region *region)
3224 struct dma_ops_domain *dma_dom = to_dma_ops_domain(to_pdomain(domain));
3225 unsigned long start, end;
3227 start = IOVA_PFN(region->start);
3228 end = IOVA_PFN(region->start + region->length - 1);
3230 WARN_ON_ONCE(reserve_iova(&dma_dom->iovad, start, end) == NULL);
3233 static bool amd_iommu_is_attach_deferred(struct iommu_domain *domain,
3236 struct iommu_dev_data *dev_data = dev->archdata.iommu;
3237 return dev_data->defer_attach;
3240 const struct iommu_ops amd_iommu_ops = {
3241 .capable = amd_iommu_capable,
3242 .domain_alloc = amd_iommu_domain_alloc,
3243 .domain_free = amd_iommu_domain_free,
3244 .attach_dev = amd_iommu_attach_device,
3245 .detach_dev = amd_iommu_detach_device,
3246 .map = amd_iommu_map,
3247 .unmap = amd_iommu_unmap,
3248 .map_sg = default_iommu_map_sg,
3249 .iova_to_phys = amd_iommu_iova_to_phys,
3250 .add_device = amd_iommu_add_device,
3251 .remove_device = amd_iommu_remove_device,
3252 .device_group = amd_iommu_device_group,
3253 .get_resv_regions = amd_iommu_get_resv_regions,
3254 .put_resv_regions = amd_iommu_put_resv_regions,
3255 .apply_resv_region = amd_iommu_apply_resv_region,
3256 .is_attach_deferred = amd_iommu_is_attach_deferred,
3257 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
3260 /*****************************************************************************
3262 * The next functions do a basic initialization of IOMMU for pass through
3265 * In passthrough mode the IOMMU is initialized and enabled but not used for
3266 * DMA-API translation.
3268 *****************************************************************************/
3270 /* IOMMUv2 specific functions */
3271 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
3273 return atomic_notifier_chain_register(&ppr_notifier, nb);
3275 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
3277 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
3279 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
3281 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
3283 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
3285 struct protection_domain *domain = to_pdomain(dom);
3286 unsigned long flags;
3288 spin_lock_irqsave(&domain->lock, flags);
3290 /* Update data structure */
3291 domain->mode = PAGE_MODE_NONE;
3292 domain->updated = true;
3294 /* Make changes visible to IOMMUs */
3295 update_domain(domain);
3297 /* Page-table is not visible to IOMMU anymore, so free it */
3298 free_pagetable(domain);
3300 spin_unlock_irqrestore(&domain->lock, flags);
3302 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
3304 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
3306 struct protection_domain *domain = to_pdomain(dom);
3307 unsigned long flags;
3310 if (pasids <= 0 || pasids > (PASID_MASK + 1))
3313 /* Number of GCR3 table levels required */
3314 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
3317 if (levels > amd_iommu_max_glx_val)
3320 spin_lock_irqsave(&domain->lock, flags);
3323 * Save us all sanity checks whether devices already in the
3324 * domain support IOMMUv2. Just force that the domain has no
3325 * devices attached when it is switched into IOMMUv2 mode.
3328 if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
3332 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
3333 if (domain->gcr3_tbl == NULL)
3336 domain->glx = levels;
3337 domain->flags |= PD_IOMMUV2_MASK;
3338 domain->updated = true;
3340 update_domain(domain);
3345 spin_unlock_irqrestore(&domain->lock, flags);
3349 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
3351 static int __flush_pasid(struct protection_domain *domain, int pasid,
3352 u64 address, bool size)
3354 struct iommu_dev_data *dev_data;
3355 struct iommu_cmd cmd;
3358 if (!(domain->flags & PD_IOMMUV2_MASK))
3361 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
3364 * IOMMU TLB needs to be flushed before Device TLB to
3365 * prevent device TLB refill from IOMMU TLB
3367 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
3368 if (domain->dev_iommu[i] == 0)
3371 ret = iommu_queue_command(amd_iommus[i], &cmd);
3376 /* Wait until IOMMU TLB flushes are complete */
3377 domain_flush_complete(domain);
3379 /* Now flush device TLBs */
3380 list_for_each_entry(dev_data, &domain->dev_list, list) {
3381 struct amd_iommu *iommu;
3385 There might be non-IOMMUv2 capable devices in an IOMMUv2
3388 if (!dev_data->ats.enabled)
3391 qdep = dev_data->ats.qdep;
3392 iommu = amd_iommu_rlookup_table[dev_data->devid];
3394 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
3395 qdep, address, size);
3397 ret = iommu_queue_command(iommu, &cmd);
3402 /* Wait until all device TLBs are flushed */
3403 domain_flush_complete(domain);
3412 static int __amd_iommu_flush_page(struct protection_domain *domain, int pasid,
3415 return __flush_pasid(domain, pasid, address, false);
3418 int amd_iommu_flush_page(struct iommu_domain *dom, int pasid,
3421 struct protection_domain *domain = to_pdomain(dom);
3422 unsigned long flags;
3425 spin_lock_irqsave(&domain->lock, flags);
3426 ret = __amd_iommu_flush_page(domain, pasid, address);
3427 spin_unlock_irqrestore(&domain->lock, flags);
3431 EXPORT_SYMBOL(amd_iommu_flush_page);
3433 static int __amd_iommu_flush_tlb(struct protection_domain *domain, int pasid)
3435 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
3439 int amd_iommu_flush_tlb(struct iommu_domain *dom, int pasid)
3441 struct protection_domain *domain = to_pdomain(dom);
3442 unsigned long flags;
3445 spin_lock_irqsave(&domain->lock, flags);
3446 ret = __amd_iommu_flush_tlb(domain, pasid);
3447 spin_unlock_irqrestore(&domain->lock, flags);
3451 EXPORT_SYMBOL(amd_iommu_flush_tlb);
3453 static u64 *__get_gcr3_pte(u64 *root, int level, int pasid, bool alloc)
3460 index = (pasid >> (9 * level)) & 0x1ff;
3466 if (!(*pte & GCR3_VALID)) {
3470 root = (void *)get_zeroed_page(GFP_ATOMIC);
3474 *pte = iommu_virt_to_phys(root) | GCR3_VALID;
3477 root = iommu_phys_to_virt(*pte & PAGE_MASK);
3485 static int __set_gcr3(struct protection_domain *domain, int pasid,
3490 if (domain->mode != PAGE_MODE_NONE)
3493 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3497 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3499 return __amd_iommu_flush_tlb(domain, pasid);
3502 static int __clear_gcr3(struct protection_domain *domain, int pasid)
3506 if (domain->mode != PAGE_MODE_NONE)
3509 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3515 return __amd_iommu_flush_tlb(domain, pasid);
3518 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, int pasid,
3521 struct protection_domain *domain = to_pdomain(dom);
3522 unsigned long flags;
3525 spin_lock_irqsave(&domain->lock, flags);
3526 ret = __set_gcr3(domain, pasid, cr3);
3527 spin_unlock_irqrestore(&domain->lock, flags);
3531 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3533 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, int pasid)
3535 struct protection_domain *domain = to_pdomain(dom);
3536 unsigned long flags;
3539 spin_lock_irqsave(&domain->lock, flags);
3540 ret = __clear_gcr3(domain, pasid);
3541 spin_unlock_irqrestore(&domain->lock, flags);
3545 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3547 int amd_iommu_complete_ppr(struct pci_dev *pdev, int pasid,
3548 int status, int tag)
3550 struct iommu_dev_data *dev_data;
3551 struct amd_iommu *iommu;
3552 struct iommu_cmd cmd;
3554 dev_data = get_dev_data(&pdev->dev);
3555 iommu = amd_iommu_rlookup_table[dev_data->devid];
3557 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3558 tag, dev_data->pri_tlp);
3560 return iommu_queue_command(iommu, &cmd);
3562 EXPORT_SYMBOL(amd_iommu_complete_ppr);
3564 struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3566 struct protection_domain *pdomain;
3568 pdomain = get_domain(&pdev->dev);
3569 if (IS_ERR(pdomain))
3572 /* Only return IOMMUv2 domains */
3573 if (!(pdomain->flags & PD_IOMMUV2_MASK))
3576 return &pdomain->domain;
3578 EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3580 void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3582 struct iommu_dev_data *dev_data;
3584 if (!amd_iommu_v2_supported())
3587 dev_data = get_dev_data(&pdev->dev);
3588 dev_data->errata |= (1 << erratum);
3590 EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3592 int amd_iommu_device_info(struct pci_dev *pdev,
3593 struct amd_iommu_device_info *info)
3598 if (pdev == NULL || info == NULL)
3601 if (!amd_iommu_v2_supported())
3604 memset(info, 0, sizeof(*info));
3606 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS);
3608 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3610 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3612 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3614 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3618 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3619 max_pasids = min(max_pasids, (1 << 20));
3621 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3622 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3624 features = pci_pasid_features(pdev);
3625 if (features & PCI_PASID_CAP_EXEC)
3626 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3627 if (features & PCI_PASID_CAP_PRIV)
3628 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3633 EXPORT_SYMBOL(amd_iommu_device_info);
3635 #ifdef CONFIG_IRQ_REMAP
3637 /*****************************************************************************
3639 * Interrupt Remapping Implementation
3641 *****************************************************************************/
3643 static struct irq_chip amd_ir_chip;
3645 static void set_dte_irq_entry(u16 devid, struct irq_remap_table *table)
3649 dte = amd_iommu_dev_table[devid].data[2];
3650 dte &= ~DTE_IRQ_PHYS_ADDR_MASK;
3651 dte |= iommu_virt_to_phys(table->table);
3652 dte |= DTE_IRQ_REMAP_INTCTL;
3653 dte |= DTE_IRQ_TABLE_LEN;
3654 dte |= DTE_IRQ_REMAP_ENABLE;
3656 amd_iommu_dev_table[devid].data[2] = dte;
3659 static struct irq_remap_table *get_irq_table(u16 devid, bool ioapic)
3661 struct irq_remap_table *table = NULL;
3662 struct amd_iommu *iommu;
3663 unsigned long flags;
3666 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3668 iommu = amd_iommu_rlookup_table[devid];
3672 table = irq_lookup_table[devid];
3676 alias = amd_iommu_alias_table[devid];
3677 table = irq_lookup_table[alias];
3679 irq_lookup_table[devid] = table;
3680 set_dte_irq_entry(devid, table);
3681 iommu_flush_dte(iommu, devid);
3685 /* Nothing there yet, allocate new irq remapping table */
3686 table = kzalloc(sizeof(*table), GFP_ATOMIC);
3690 /* Initialize table spin-lock */
3691 spin_lock_init(&table->lock);
3694 /* Keep the first 32 indexes free for IOAPIC interrupts */
3695 table->min_index = 32;
3697 table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_ATOMIC);
3698 if (!table->table) {
3704 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
3705 memset(table->table, 0,
3706 MAX_IRQS_PER_TABLE * sizeof(u32));
3708 memset(table->table, 0,
3709 (MAX_IRQS_PER_TABLE * (sizeof(u64) * 2)));
3714 for (i = 0; i < 32; ++i)
3715 iommu->irte_ops->set_allocated(table, i);
3718 irq_lookup_table[devid] = table;
3719 set_dte_irq_entry(devid, table);
3720 iommu_flush_dte(iommu, devid);
3721 if (devid != alias) {
3722 irq_lookup_table[alias] = table;
3723 set_dte_irq_entry(alias, table);
3724 iommu_flush_dte(iommu, alias);
3728 iommu_completion_wait(iommu);
3731 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3736 static int alloc_irq_index(u16 devid, int count)
3738 struct irq_remap_table *table;
3739 unsigned long flags;
3741 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
3746 table = get_irq_table(devid, false);
3750 spin_lock_irqsave(&table->lock, flags);
3752 /* Scan table for free entries */
3753 for (c = 0, index = table->min_index;
3754 index < MAX_IRQS_PER_TABLE;
3756 if (!iommu->irte_ops->is_allocated(table, index))
3763 iommu->irte_ops->set_allocated(table, index - c + 1);
3773 spin_unlock_irqrestore(&table->lock, flags);
3778 static int modify_irte_ga(u16 devid, int index, struct irte_ga *irte,
3779 struct amd_ir_data *data)
3781 struct irq_remap_table *table;
3782 struct amd_iommu *iommu;
3783 unsigned long flags;
3784 struct irte_ga *entry;
3786 iommu = amd_iommu_rlookup_table[devid];
3790 table = get_irq_table(devid, false);
3794 spin_lock_irqsave(&table->lock, flags);
3796 entry = (struct irte_ga *)table->table;
3797 entry = &entry[index];
3798 entry->lo.fields_remap.valid = 0;
3799 entry->hi.val = irte->hi.val;
3800 entry->lo.val = irte->lo.val;
3801 entry->lo.fields_remap.valid = 1;
3805 spin_unlock_irqrestore(&table->lock, flags);
3807 iommu_flush_irt(iommu, devid);
3808 iommu_completion_wait(iommu);
3813 static int modify_irte(u16 devid, int index, union irte *irte)
3815 struct irq_remap_table *table;
3816 struct amd_iommu *iommu;
3817 unsigned long flags;
3819 iommu = amd_iommu_rlookup_table[devid];
3823 table = get_irq_table(devid, false);
3827 spin_lock_irqsave(&table->lock, flags);
3828 table->table[index] = irte->val;
3829 spin_unlock_irqrestore(&table->lock, flags);
3831 iommu_flush_irt(iommu, devid);
3832 iommu_completion_wait(iommu);
3837 static void free_irte(u16 devid, int index)
3839 struct irq_remap_table *table;
3840 struct amd_iommu *iommu;
3841 unsigned long flags;
3843 iommu = amd_iommu_rlookup_table[devid];
3847 table = get_irq_table(devid, false);
3851 spin_lock_irqsave(&table->lock, flags);
3852 iommu->irte_ops->clear_allocated(table, index);
3853 spin_unlock_irqrestore(&table->lock, flags);
3855 iommu_flush_irt(iommu, devid);
3856 iommu_completion_wait(iommu);
3859 static void irte_prepare(void *entry,
3860 u32 delivery_mode, u32 dest_mode,
3861 u8 vector, u32 dest_apicid, int devid)
3863 union irte *irte = (union irte *) entry;
3866 irte->fields.vector = vector;
3867 irte->fields.int_type = delivery_mode;
3868 irte->fields.destination = dest_apicid;
3869 irte->fields.dm = dest_mode;
3870 irte->fields.valid = 1;
3873 static void irte_ga_prepare(void *entry,
3874 u32 delivery_mode, u32 dest_mode,
3875 u8 vector, u32 dest_apicid, int devid)
3877 struct irte_ga *irte = (struct irte_ga *) entry;
3881 irte->lo.fields_remap.int_type = delivery_mode;
3882 irte->lo.fields_remap.dm = dest_mode;
3883 irte->hi.fields.vector = vector;
3884 irte->lo.fields_remap.destination = dest_apicid;
3885 irte->lo.fields_remap.valid = 1;
3888 static void irte_activate(void *entry, u16 devid, u16 index)
3890 union irte *irte = (union irte *) entry;
3892 irte->fields.valid = 1;
3893 modify_irte(devid, index, irte);
3896 static void irte_ga_activate(void *entry, u16 devid, u16 index)
3898 struct irte_ga *irte = (struct irte_ga *) entry;
3900 irte->lo.fields_remap.valid = 1;
3901 modify_irte_ga(devid, index, irte, NULL);
3904 static void irte_deactivate(void *entry, u16 devid, u16 index)
3906 union irte *irte = (union irte *) entry;
3908 irte->fields.valid = 0;
3909 modify_irte(devid, index, irte);
3912 static void irte_ga_deactivate(void *entry, u16 devid, u16 index)
3914 struct irte_ga *irte = (struct irte_ga *) entry;
3916 irte->lo.fields_remap.valid = 0;
3917 modify_irte_ga(devid, index, irte, NULL);
3920 static void irte_set_affinity(void *entry, u16 devid, u16 index,
3921 u8 vector, u32 dest_apicid)
3923 union irte *irte = (union irte *) entry;
3925 irte->fields.vector = vector;
3926 irte->fields.destination = dest_apicid;
3927 modify_irte(devid, index, irte);
3930 static void irte_ga_set_affinity(void *entry, u16 devid, u16 index,
3931 u8 vector, u32 dest_apicid)
3933 struct irte_ga *irte = (struct irte_ga *) entry;
3934 struct iommu_dev_data *dev_data = search_dev_data(devid);
3936 if (!dev_data || !dev_data->use_vapic ||
3937 !irte->lo.fields_remap.guest_mode) {
3938 irte->hi.fields.vector = vector;
3939 irte->lo.fields_remap.destination = dest_apicid;
3940 modify_irte_ga(devid, index, irte, NULL);
3944 #define IRTE_ALLOCATED (~1U)
3945 static void irte_set_allocated(struct irq_remap_table *table, int index)
3947 table->table[index] = IRTE_ALLOCATED;
3950 static void irte_ga_set_allocated(struct irq_remap_table *table, int index)
3952 struct irte_ga *ptr = (struct irte_ga *)table->table;
3953 struct irte_ga *irte = &ptr[index];
3955 memset(&irte->lo.val, 0, sizeof(u64));
3956 memset(&irte->hi.val, 0, sizeof(u64));
3957 irte->hi.fields.vector = 0xff;
3960 static bool irte_is_allocated(struct irq_remap_table *table, int index)
3962 union irte *ptr = (union irte *)table->table;
3963 union irte *irte = &ptr[index];
3965 return irte->val != 0;
3968 static bool irte_ga_is_allocated(struct irq_remap_table *table, int index)
3970 struct irte_ga *ptr = (struct irte_ga *)table->table;
3971 struct irte_ga *irte = &ptr[index];
3973 return irte->hi.fields.vector != 0;
3976 static void irte_clear_allocated(struct irq_remap_table *table, int index)
3978 table->table[index] = 0;
3981 static void irte_ga_clear_allocated(struct irq_remap_table *table, int index)
3983 struct irte_ga *ptr = (struct irte_ga *)table->table;
3984 struct irte_ga *irte = &ptr[index];
3986 memset(&irte->lo.val, 0, sizeof(u64));
3987 memset(&irte->hi.val, 0, sizeof(u64));
3990 static int get_devid(struct irq_alloc_info *info)
3994 switch (info->type) {
3995 case X86_IRQ_ALLOC_TYPE_IOAPIC:
3996 devid = get_ioapic_devid(info->ioapic_id);
3998 case X86_IRQ_ALLOC_TYPE_HPET:
3999 devid = get_hpet_devid(info->hpet_id);
4001 case X86_IRQ_ALLOC_TYPE_MSI:
4002 case X86_IRQ_ALLOC_TYPE_MSIX:
4003 devid = get_device_id(&info->msi_dev->dev);
4013 static struct irq_domain *get_ir_irq_domain(struct irq_alloc_info *info)
4015 struct amd_iommu *iommu;
4021 devid = get_devid(info);
4023 iommu = amd_iommu_rlookup_table[devid];
4025 return iommu->ir_domain;
4031 static struct irq_domain *get_irq_domain(struct irq_alloc_info *info)
4033 struct amd_iommu *iommu;
4039 switch (info->type) {
4040 case X86_IRQ_ALLOC_TYPE_MSI:
4041 case X86_IRQ_ALLOC_TYPE_MSIX:
4042 devid = get_device_id(&info->msi_dev->dev);
4046 iommu = amd_iommu_rlookup_table[devid];
4048 return iommu->msi_domain;
4057 struct irq_remap_ops amd_iommu_irq_ops = {
4058 .prepare = amd_iommu_prepare,
4059 .enable = amd_iommu_enable,
4060 .disable = amd_iommu_disable,
4061 .reenable = amd_iommu_reenable,
4062 .enable_faulting = amd_iommu_enable_faulting,
4063 .get_ir_irq_domain = get_ir_irq_domain,
4064 .get_irq_domain = get_irq_domain,
4067 static void irq_remapping_prepare_irte(struct amd_ir_data *data,
4068 struct irq_cfg *irq_cfg,
4069 struct irq_alloc_info *info,
4070 int devid, int index, int sub_handle)
4072 struct irq_2_irte *irte_info = &data->irq_2_irte;
4073 struct msi_msg *msg = &data->msi_entry;
4074 struct IO_APIC_route_entry *entry;
4075 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
4080 data->irq_2_irte.devid = devid;
4081 data->irq_2_irte.index = index + sub_handle;
4082 iommu->irte_ops->prepare(data->entry, apic->irq_delivery_mode,
4083 apic->irq_dest_mode, irq_cfg->vector,
4084 irq_cfg->dest_apicid, devid);
4086 switch (info->type) {
4087 case X86_IRQ_ALLOC_TYPE_IOAPIC:
4088 /* Setup IOAPIC entry */
4089 entry = info->ioapic_entry;
4090 info->ioapic_entry = NULL;
4091 memset(entry, 0, sizeof(*entry));
4092 entry->vector = index;
4094 entry->trigger = info->ioapic_trigger;
4095 entry->polarity = info->ioapic_polarity;
4096 /* Mask level triggered irqs. */
4097 if (info->ioapic_trigger)
4101 case X86_IRQ_ALLOC_TYPE_HPET:
4102 case X86_IRQ_ALLOC_TYPE_MSI:
4103 case X86_IRQ_ALLOC_TYPE_MSIX:
4104 msg->address_hi = MSI_ADDR_BASE_HI;
4105 msg->address_lo = MSI_ADDR_BASE_LO;
4106 msg->data = irte_info->index;
4115 struct amd_irte_ops irte_32_ops = {
4116 .prepare = irte_prepare,
4117 .activate = irte_activate,
4118 .deactivate = irte_deactivate,
4119 .set_affinity = irte_set_affinity,
4120 .set_allocated = irte_set_allocated,
4121 .is_allocated = irte_is_allocated,
4122 .clear_allocated = irte_clear_allocated,
4125 struct amd_irte_ops irte_128_ops = {
4126 .prepare = irte_ga_prepare,
4127 .activate = irte_ga_activate,
4128 .deactivate = irte_ga_deactivate,
4129 .set_affinity = irte_ga_set_affinity,
4130 .set_allocated = irte_ga_set_allocated,
4131 .is_allocated = irte_ga_is_allocated,
4132 .clear_allocated = irte_ga_clear_allocated,
4135 static int irq_remapping_alloc(struct irq_domain *domain, unsigned int virq,
4136 unsigned int nr_irqs, void *arg)
4138 struct irq_alloc_info *info = arg;
4139 struct irq_data *irq_data;
4140 struct amd_ir_data *data = NULL;
4141 struct irq_cfg *cfg;
4147 if (nr_irqs > 1 && info->type != X86_IRQ_ALLOC_TYPE_MSI &&
4148 info->type != X86_IRQ_ALLOC_TYPE_MSIX)
4152 * With IRQ remapping enabled, don't need contiguous CPU vectors
4153 * to support multiple MSI interrupts.
4155 if (info->type == X86_IRQ_ALLOC_TYPE_MSI)
4156 info->flags &= ~X86_IRQ_ALLOC_CONTIGUOUS_VECTORS;
4158 devid = get_devid(info);
4162 ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
4166 if (info->type == X86_IRQ_ALLOC_TYPE_IOAPIC) {
4167 if (get_irq_table(devid, true))
4168 index = info->ioapic_pin;
4172 index = alloc_irq_index(devid, nr_irqs);
4175 pr_warn("Failed to allocate IRTE\n");
4177 goto out_free_parent;
4180 for (i = 0; i < nr_irqs; i++) {
4181 irq_data = irq_domain_get_irq_data(domain, virq + i);
4182 cfg = irqd_cfg(irq_data);
4183 if (!irq_data || !cfg) {
4189 data = kzalloc(sizeof(*data), GFP_KERNEL);
4193 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
4194 data->entry = kzalloc(sizeof(union irte), GFP_KERNEL);
4196 data->entry = kzalloc(sizeof(struct irte_ga),
4203 irq_data->hwirq = (devid << 16) + i;
4204 irq_data->chip_data = data;
4205 irq_data->chip = &amd_ir_chip;
4206 irq_remapping_prepare_irte(data, cfg, info, devid, index, i);
4207 irq_set_status_flags(virq + i, IRQ_MOVE_PCNTXT);
4213 for (i--; i >= 0; i--) {
4214 irq_data = irq_domain_get_irq_data(domain, virq + i);
4216 kfree(irq_data->chip_data);
4218 for (i = 0; i < nr_irqs; i++)
4219 free_irte(devid, index + i);
4221 irq_domain_free_irqs_common(domain, virq, nr_irqs);
4225 static void irq_remapping_free(struct irq_domain *domain, unsigned int virq,
4226 unsigned int nr_irqs)
4228 struct irq_2_irte *irte_info;
4229 struct irq_data *irq_data;
4230 struct amd_ir_data *data;
4233 for (i = 0; i < nr_irqs; i++) {
4234 irq_data = irq_domain_get_irq_data(domain, virq + i);
4235 if (irq_data && irq_data->chip_data) {
4236 data = irq_data->chip_data;
4237 irte_info = &data->irq_2_irte;
4238 free_irte(irte_info->devid, irte_info->index);
4243 irq_domain_free_irqs_common(domain, virq, nr_irqs);
4246 static void irq_remapping_activate(struct irq_domain *domain,
4247 struct irq_data *irq_data)
4249 struct amd_ir_data *data = irq_data->chip_data;
4250 struct irq_2_irte *irte_info = &data->irq_2_irte;
4251 struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4254 iommu->irte_ops->activate(data->entry, irte_info->devid,
4258 static void irq_remapping_deactivate(struct irq_domain *domain,
4259 struct irq_data *irq_data)
4261 struct amd_ir_data *data = irq_data->chip_data;
4262 struct irq_2_irte *irte_info = &data->irq_2_irte;
4263 struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4266 iommu->irte_ops->deactivate(data->entry, irte_info->devid,
4270 static const struct irq_domain_ops amd_ir_domain_ops = {
4271 .alloc = irq_remapping_alloc,
4272 .free = irq_remapping_free,
4273 .activate = irq_remapping_activate,
4274 .deactivate = irq_remapping_deactivate,
4277 static int amd_ir_set_vcpu_affinity(struct irq_data *data, void *vcpu_info)
4279 struct amd_iommu *iommu;
4280 struct amd_iommu_pi_data *pi_data = vcpu_info;
4281 struct vcpu_data *vcpu_pi_info = pi_data->vcpu_data;
4282 struct amd_ir_data *ir_data = data->chip_data;
4283 struct irte_ga *irte = (struct irte_ga *) ir_data->entry;
4284 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
4285 struct iommu_dev_data *dev_data = search_dev_data(irte_info->devid);
4288 * This device has never been set up for guest mode.
4289 * we should not modify the IRTE
4291 if (!dev_data || !dev_data->use_vapic)
4294 pi_data->ir_data = ir_data;
4297 * SVM tries to set up for VAPIC mode, but we are in
4298 * legacy mode. So, we force legacy mode instead.
4300 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
4301 pr_debug("AMD-Vi: %s: Fall back to using intr legacy remap\n",
4303 pi_data->is_guest_mode = false;
4306 iommu = amd_iommu_rlookup_table[irte_info->devid];
4310 pi_data->prev_ga_tag = ir_data->cached_ga_tag;
4311 if (pi_data->is_guest_mode) {
4313 irte->hi.fields.ga_root_ptr = (pi_data->base >> 12);
4314 irte->hi.fields.vector = vcpu_pi_info->vector;
4315 irte->lo.fields_vapic.ga_log_intr = 1;
4316 irte->lo.fields_vapic.guest_mode = 1;
4317 irte->lo.fields_vapic.ga_tag = pi_data->ga_tag;
4319 ir_data->cached_ga_tag = pi_data->ga_tag;
4322 struct irq_cfg *cfg = irqd_cfg(data);
4326 irte->hi.fields.vector = cfg->vector;
4327 irte->lo.fields_remap.guest_mode = 0;
4328 irte->lo.fields_remap.destination = cfg->dest_apicid;
4329 irte->lo.fields_remap.int_type = apic->irq_delivery_mode;
4330 irte->lo.fields_remap.dm = apic->irq_dest_mode;
4333 * This communicates the ga_tag back to the caller
4334 * so that it can do all the necessary clean up.
4336 ir_data->cached_ga_tag = 0;
4339 return modify_irte_ga(irte_info->devid, irte_info->index, irte, ir_data);
4342 static int amd_ir_set_affinity(struct irq_data *data,
4343 const struct cpumask *mask, bool force)
4345 struct amd_ir_data *ir_data = data->chip_data;
4346 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
4347 struct irq_cfg *cfg = irqd_cfg(data);
4348 struct irq_data *parent = data->parent_data;
4349 struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4355 ret = parent->chip->irq_set_affinity(parent, mask, force);
4356 if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
4360 * Atomically updates the IRTE with the new destination, vector
4361 * and flushes the interrupt entry cache.
4363 iommu->irte_ops->set_affinity(ir_data->entry, irte_info->devid,
4364 irte_info->index, cfg->vector, cfg->dest_apicid);
4367 * After this point, all the interrupts will start arriving
4368 * at the new destination. So, time to cleanup the previous
4369 * vector allocation.
4371 send_cleanup_vector(cfg);
4373 return IRQ_SET_MASK_OK_DONE;
4376 static void ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg)
4378 struct amd_ir_data *ir_data = irq_data->chip_data;
4380 *msg = ir_data->msi_entry;
4383 static struct irq_chip amd_ir_chip = {
4385 .irq_ack = ir_ack_apic_edge,
4386 .irq_set_affinity = amd_ir_set_affinity,
4387 .irq_set_vcpu_affinity = amd_ir_set_vcpu_affinity,
4388 .irq_compose_msi_msg = ir_compose_msi_msg,
4391 int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
4393 struct fwnode_handle *fn;
4395 fn = irq_domain_alloc_named_id_fwnode("AMD-IR", iommu->index);
4398 iommu->ir_domain = irq_domain_create_tree(fn, &amd_ir_domain_ops, iommu);
4399 if (!iommu->ir_domain) {
4400 irq_domain_free_fwnode(fn);
4404 iommu->ir_domain->parent = arch_get_ir_parent_domain();
4405 iommu->msi_domain = arch_create_remap_msi_irq_domain(iommu->ir_domain,
4411 int amd_iommu_update_ga(int cpu, bool is_run, void *data)
4413 unsigned long flags;
4414 struct amd_iommu *iommu;
4415 struct irq_remap_table *irt;
4416 struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
4417 int devid = ir_data->irq_2_irte.devid;
4418 struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
4419 struct irte_ga *ref = (struct irte_ga *) ir_data->ref;
4421 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
4422 !ref || !entry || !entry->lo.fields_vapic.guest_mode)
4425 iommu = amd_iommu_rlookup_table[devid];
4429 irt = get_irq_table(devid, false);
4433 spin_lock_irqsave(&irt->lock, flags);
4435 if (ref->lo.fields_vapic.guest_mode) {
4437 ref->lo.fields_vapic.destination = cpu;
4438 ref->lo.fields_vapic.is_run = is_run;
4442 spin_unlock_irqrestore(&irt->lock, flags);
4444 iommu_flush_irt(iommu, devid);
4445 iommu_completion_wait(iommu);
4448 EXPORT_SYMBOL(amd_iommu_update_ga);