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 CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
59 #define LOOP_TIMEOUT 100000
61 /* IO virtual address start page frame number */
62 #define IOVA_START_PFN (1)
63 #define IOVA_PFN(addr) ((addr) >> PAGE_SHIFT)
64 #define DMA_32BIT_PFN IOVA_PFN(DMA_BIT_MASK(32))
66 /* Reserved IOVA ranges */
67 #define MSI_RANGE_START (0xfee00000)
68 #define MSI_RANGE_END (0xfeefffff)
69 #define HT_RANGE_START (0xfd00000000ULL)
70 #define HT_RANGE_END (0xffffffffffULL)
73 * This bitmap is used to advertise the page sizes our hardware support
74 * to the IOMMU core, which will then use this information to split
75 * physically contiguous memory regions it is mapping into page sizes
78 * 512GB Pages are not supported due to a hardware bug
80 #define AMD_IOMMU_PGSIZES ((~0xFFFUL) & ~(2ULL << 38))
82 static DEFINE_RWLOCK(amd_iommu_devtable_lock);
84 /* List of all available dev_data structures */
85 static LIST_HEAD(dev_data_list);
86 static DEFINE_SPINLOCK(dev_data_list_lock);
88 LIST_HEAD(ioapic_map);
90 LIST_HEAD(acpihid_map);
93 * Domain for untranslated devices - only allocated
94 * if iommu=pt passed on kernel cmd line.
96 const struct iommu_ops amd_iommu_ops;
98 static ATOMIC_NOTIFIER_HEAD(ppr_notifier);
99 int amd_iommu_max_glx_val = -1;
101 static const struct dma_map_ops amd_iommu_dma_ops;
104 * This struct contains device specific data for the IOMMU
106 struct iommu_dev_data {
107 struct list_head list; /* For domain->dev_list */
108 struct list_head dev_data_list; /* For global dev_data_list */
109 struct protection_domain *domain; /* Domain the device is bound to */
110 u16 devid; /* PCI Device ID */
111 u16 alias; /* Alias Device ID */
112 bool iommu_v2; /* Device can make use of IOMMUv2 */
113 bool passthrough; /* Device is identity mapped */
117 } ats; /* ATS state */
118 bool pri_tlp; /* PASID TLB required for
120 u32 errata; /* Bitmap for errata to apply */
121 bool use_vapic; /* Enable device to use vapic mode */
123 struct ratelimit_state rs; /* Ratelimit IOPF messages */
127 * general struct to manage commands send to an IOMMU
133 struct kmem_cache *amd_iommu_irq_cache;
135 static void update_domain(struct protection_domain *domain);
136 static int protection_domain_init(struct protection_domain *domain);
137 static void detach_device(struct device *dev);
139 #define FLUSH_QUEUE_SIZE 256
141 struct flush_queue_entry {
142 unsigned long iova_pfn;
144 u64 counter; /* Flush counter when this entry was added to the queue */
148 struct flush_queue_entry *entries;
154 * Data container for a dma_ops specific protection domain
156 struct dma_ops_domain {
157 /* generic protection domain information */
158 struct protection_domain domain;
161 struct iova_domain iovad;
163 struct flush_queue __percpu *flush_queue;
166 * We need two counter here to be race-free wrt. IOTLB flushing and
167 * adding entries to the flush queue.
169 * The flush_start_cnt is incremented _before_ the IOTLB flush starts.
170 * New entries added to the flush ring-buffer get their 'counter' value
171 * from here. This way we can make sure that entries added to the queue
172 * (or other per-cpu queues of the same domain) while the TLB is about
173 * to be flushed are not considered to be flushed already.
175 atomic64_t flush_start_cnt;
178 * The flush_finish_cnt is incremented when an IOTLB flush is complete.
179 * This value is always smaller than flush_start_cnt. The queue_add
180 * function frees all IOVAs that have a counter value smaller than
181 * flush_finish_cnt. This makes sure that we only free IOVAs that are
182 * flushed out of the IOTLB of the domain.
184 atomic64_t flush_finish_cnt;
187 * Timer to make sure we don't keep IOVAs around unflushed
190 struct timer_list flush_timer;
191 atomic_t flush_timer_on;
194 static struct iova_domain reserved_iova_ranges;
195 static struct lock_class_key reserved_rbtree_key;
197 /****************************************************************************
201 ****************************************************************************/
203 static inline int match_hid_uid(struct device *dev,
204 struct acpihid_map_entry *entry)
206 const char *hid, *uid;
208 hid = acpi_device_hid(ACPI_COMPANION(dev));
209 uid = acpi_device_uid(ACPI_COMPANION(dev));
215 return strcmp(hid, entry->hid);
218 return strcmp(hid, entry->hid);
220 return (strcmp(hid, entry->hid) || strcmp(uid, entry->uid));
223 static inline u16 get_pci_device_id(struct device *dev)
225 struct pci_dev *pdev = to_pci_dev(dev);
227 return PCI_DEVID(pdev->bus->number, pdev->devfn);
230 static inline int get_acpihid_device_id(struct device *dev,
231 struct acpihid_map_entry **entry)
233 struct acpihid_map_entry *p;
235 list_for_each_entry(p, &acpihid_map, list) {
236 if (!match_hid_uid(dev, p)) {
245 static inline int get_device_id(struct device *dev)
250 devid = get_pci_device_id(dev);
252 devid = get_acpihid_device_id(dev, NULL);
257 static struct protection_domain *to_pdomain(struct iommu_domain *dom)
259 return container_of(dom, struct protection_domain, domain);
262 static struct dma_ops_domain* to_dma_ops_domain(struct protection_domain *domain)
264 BUG_ON(domain->flags != PD_DMA_OPS_MASK);
265 return container_of(domain, struct dma_ops_domain, domain);
268 static struct iommu_dev_data *alloc_dev_data(u16 devid)
270 struct iommu_dev_data *dev_data;
273 dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
277 dev_data->devid = devid;
279 spin_lock_irqsave(&dev_data_list_lock, flags);
280 list_add_tail(&dev_data->dev_data_list, &dev_data_list);
281 spin_unlock_irqrestore(&dev_data_list_lock, flags);
283 ratelimit_default_init(&dev_data->rs);
288 static struct iommu_dev_data *search_dev_data(u16 devid)
290 struct iommu_dev_data *dev_data;
293 spin_lock_irqsave(&dev_data_list_lock, flags);
294 list_for_each_entry(dev_data, &dev_data_list, dev_data_list) {
295 if (dev_data->devid == devid)
302 spin_unlock_irqrestore(&dev_data_list_lock, flags);
307 static int __last_alias(struct pci_dev *pdev, u16 alias, void *data)
309 *(u16 *)data = alias;
313 static u16 get_alias(struct device *dev)
315 struct pci_dev *pdev = to_pci_dev(dev);
316 u16 devid, ivrs_alias, pci_alias;
318 /* The callers make sure that get_device_id() does not fail here */
319 devid = get_device_id(dev);
320 ivrs_alias = amd_iommu_alias_table[devid];
321 pci_for_each_dma_alias(pdev, __last_alias, &pci_alias);
323 if (ivrs_alias == pci_alias)
329 * The IVRS is fairly reliable in telling us about aliases, but it
330 * can't know about every screwy device. If we don't have an IVRS
331 * reported alias, use the PCI reported alias. In that case we may
332 * still need to initialize the rlookup and dev_table entries if the
333 * alias is to a non-existent device.
335 if (ivrs_alias == devid) {
336 if (!amd_iommu_rlookup_table[pci_alias]) {
337 amd_iommu_rlookup_table[pci_alias] =
338 amd_iommu_rlookup_table[devid];
339 memcpy(amd_iommu_dev_table[pci_alias].data,
340 amd_iommu_dev_table[devid].data,
341 sizeof(amd_iommu_dev_table[pci_alias].data));
347 pr_info("AMD-Vi: Using IVRS reported alias %02x:%02x.%d "
348 "for device %s[%04x:%04x], kernel reported alias "
349 "%02x:%02x.%d\n", PCI_BUS_NUM(ivrs_alias), PCI_SLOT(ivrs_alias),
350 PCI_FUNC(ivrs_alias), dev_name(dev), pdev->vendor, pdev->device,
351 PCI_BUS_NUM(pci_alias), PCI_SLOT(pci_alias),
352 PCI_FUNC(pci_alias));
355 * If we don't have a PCI DMA alias and the IVRS alias is on the same
356 * bus, then the IVRS table may know about a quirk that we don't.
358 if (pci_alias == devid &&
359 PCI_BUS_NUM(ivrs_alias) == pdev->bus->number) {
360 pci_add_dma_alias(pdev, ivrs_alias & 0xff);
361 pr_info("AMD-Vi: Added PCI DMA alias %02x.%d for %s\n",
362 PCI_SLOT(ivrs_alias), PCI_FUNC(ivrs_alias),
369 static struct iommu_dev_data *find_dev_data(u16 devid)
371 struct iommu_dev_data *dev_data;
373 dev_data = search_dev_data(devid);
375 if (dev_data == NULL)
376 dev_data = alloc_dev_data(devid);
381 static struct iommu_dev_data *get_dev_data(struct device *dev)
383 return dev->archdata.iommu;
387 * Find or create an IOMMU group for a acpihid device.
389 static struct iommu_group *acpihid_device_group(struct device *dev)
391 struct acpihid_map_entry *p, *entry = NULL;
394 devid = get_acpihid_device_id(dev, &entry);
396 return ERR_PTR(devid);
398 list_for_each_entry(p, &acpihid_map, list) {
399 if ((devid == p->devid) && p->group)
400 entry->group = p->group;
404 entry->group = generic_device_group(dev);
406 iommu_group_ref_get(entry->group);
411 static bool pci_iommuv2_capable(struct pci_dev *pdev)
413 static const int caps[] = {
416 PCI_EXT_CAP_ID_PASID,
420 for (i = 0; i < 3; ++i) {
421 pos = pci_find_ext_capability(pdev, caps[i]);
429 static bool pdev_pri_erratum(struct pci_dev *pdev, u32 erratum)
431 struct iommu_dev_data *dev_data;
433 dev_data = get_dev_data(&pdev->dev);
435 return dev_data->errata & (1 << erratum) ? true : false;
439 * This function checks if the driver got a valid device from the caller to
440 * avoid dereferencing invalid pointers.
442 static bool check_device(struct device *dev)
446 if (!dev || !dev->dma_mask)
449 devid = get_device_id(dev);
453 /* Out of our scope? */
454 if (devid > amd_iommu_last_bdf)
457 if (amd_iommu_rlookup_table[devid] == NULL)
463 static void init_iommu_group(struct device *dev)
465 struct iommu_group *group;
467 group = iommu_group_get_for_dev(dev);
471 iommu_group_put(group);
474 static int iommu_init_device(struct device *dev)
476 struct iommu_dev_data *dev_data;
477 struct amd_iommu *iommu;
480 if (dev->archdata.iommu)
483 devid = get_device_id(dev);
487 iommu = amd_iommu_rlookup_table[devid];
489 dev_data = find_dev_data(devid);
493 dev_data->alias = get_alias(dev);
495 if (dev_is_pci(dev) && pci_iommuv2_capable(to_pci_dev(dev))) {
496 struct amd_iommu *iommu;
498 iommu = amd_iommu_rlookup_table[dev_data->devid];
499 dev_data->iommu_v2 = iommu->is_iommu_v2;
502 dev->archdata.iommu = dev_data;
504 iommu_device_link(&iommu->iommu, dev);
509 static void iommu_ignore_device(struct device *dev)
514 devid = get_device_id(dev);
518 alias = get_alias(dev);
520 memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
521 memset(&amd_iommu_dev_table[alias], 0, sizeof(struct dev_table_entry));
523 amd_iommu_rlookup_table[devid] = NULL;
524 amd_iommu_rlookup_table[alias] = NULL;
527 static void iommu_uninit_device(struct device *dev)
529 struct iommu_dev_data *dev_data;
530 struct amd_iommu *iommu;
533 devid = get_device_id(dev);
537 iommu = amd_iommu_rlookup_table[devid];
539 dev_data = search_dev_data(devid);
543 if (dev_data->domain)
546 iommu_device_unlink(&iommu->iommu, dev);
548 iommu_group_remove_device(dev);
554 * We keep dev_data around for unplugged devices and reuse it when the
555 * device is re-plugged - not doing so would introduce a ton of races.
559 /****************************************************************************
561 * Interrupt handling functions
563 ****************************************************************************/
565 static void dump_dte_entry(u16 devid)
569 for (i = 0; i < 4; ++i)
570 pr_err("AMD-Vi: DTE[%d]: %016llx\n", i,
571 amd_iommu_dev_table[devid].data[i]);
574 static void dump_command(unsigned long phys_addr)
576 struct iommu_cmd *cmd = phys_to_virt(phys_addr);
579 for (i = 0; i < 4; ++i)
580 pr_err("AMD-Vi: CMD[%d]: %08x\n", i, cmd->data[i]);
583 static void amd_iommu_report_page_fault(u16 devid, u16 domain_id,
584 u64 address, int flags)
586 struct iommu_dev_data *dev_data = NULL;
587 struct pci_dev *pdev;
589 pdev = pci_get_bus_and_slot(PCI_BUS_NUM(devid), devid & 0xff);
591 dev_data = get_dev_data(&pdev->dev);
593 if (dev_data && __ratelimit(&dev_data->rs)) {
594 dev_err(&pdev->dev, "AMD-Vi: Event logged [IO_PAGE_FAULT domain=0x%04x address=0x%016llx flags=0x%04x]\n",
595 domain_id, address, flags);
596 } else if (printk_ratelimit()) {
597 pr_err("AMD-Vi: Event logged [IO_PAGE_FAULT device=%02x:%02x.%x domain=0x%04x address=0x%016llx flags=0x%04x]\n",
598 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
599 domain_id, address, flags);
606 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
608 int type, devid, domid, flags;
609 volatile u32 *event = __evt;
614 type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
615 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
616 domid = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
617 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
618 address = (u64)(((u64)event[3]) << 32) | event[2];
621 /* Did we hit the erratum? */
622 if (++count == LOOP_TIMEOUT) {
623 pr_err("AMD-Vi: No event written to event log\n");
630 if (type == EVENT_TYPE_IO_FAULT) {
631 amd_iommu_report_page_fault(devid, domid, address, flags);
634 printk(KERN_ERR "AMD-Vi: Event logged [");
638 case EVENT_TYPE_ILL_DEV:
639 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
640 "address=0x%016llx flags=0x%04x]\n",
641 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
643 dump_dte_entry(devid);
645 case EVENT_TYPE_DEV_TAB_ERR:
646 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
647 "address=0x%016llx flags=0x%04x]\n",
648 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
651 case EVENT_TYPE_PAGE_TAB_ERR:
652 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
653 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
654 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
655 domid, address, flags);
657 case EVENT_TYPE_ILL_CMD:
658 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
659 dump_command(address);
661 case EVENT_TYPE_CMD_HARD_ERR:
662 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
663 "flags=0x%04x]\n", address, flags);
665 case EVENT_TYPE_IOTLB_INV_TO:
666 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
667 "address=0x%016llx]\n",
668 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
671 case EVENT_TYPE_INV_DEV_REQ:
672 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
673 "address=0x%016llx flags=0x%04x]\n",
674 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
678 printk(KERN_ERR "UNKNOWN type=0x%02x]\n", type);
681 memset(__evt, 0, 4 * sizeof(u32));
684 static void iommu_poll_events(struct amd_iommu *iommu)
688 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
689 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
691 while (head != tail) {
692 iommu_print_event(iommu, iommu->evt_buf + head);
693 head = (head + EVENT_ENTRY_SIZE) % EVT_BUFFER_SIZE;
696 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
699 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
701 struct amd_iommu_fault fault;
703 if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
704 pr_err_ratelimited("AMD-Vi: Unknown PPR request received\n");
708 fault.address = raw[1];
709 fault.pasid = PPR_PASID(raw[0]);
710 fault.device_id = PPR_DEVID(raw[0]);
711 fault.tag = PPR_TAG(raw[0]);
712 fault.flags = PPR_FLAGS(raw[0]);
714 atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
717 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
721 if (iommu->ppr_log == NULL)
724 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
725 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
727 while (head != tail) {
732 raw = (u64 *)(iommu->ppr_log + head);
735 * Hardware bug: Interrupt may arrive before the entry is
736 * written to memory. If this happens we need to wait for the
739 for (i = 0; i < LOOP_TIMEOUT; ++i) {
740 if (PPR_REQ_TYPE(raw[0]) != 0)
745 /* Avoid memcpy function-call overhead */
750 * To detect the hardware bug we need to clear the entry
753 raw[0] = raw[1] = 0UL;
755 /* Update head pointer of hardware ring-buffer */
756 head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
757 writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
759 /* Handle PPR entry */
760 iommu_handle_ppr_entry(iommu, entry);
762 /* Refresh ring-buffer information */
763 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
764 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
768 #ifdef CONFIG_IRQ_REMAP
769 static int (*iommu_ga_log_notifier)(u32);
771 int amd_iommu_register_ga_log_notifier(int (*notifier)(u32))
773 iommu_ga_log_notifier = notifier;
777 EXPORT_SYMBOL(amd_iommu_register_ga_log_notifier);
779 static void iommu_poll_ga_log(struct amd_iommu *iommu)
781 u32 head, tail, cnt = 0;
783 if (iommu->ga_log == NULL)
786 head = readl(iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
787 tail = readl(iommu->mmio_base + MMIO_GA_TAIL_OFFSET);
789 while (head != tail) {
793 raw = (u64 *)(iommu->ga_log + head);
796 /* Avoid memcpy function-call overhead */
799 /* Update head pointer of hardware ring-buffer */
800 head = (head + GA_ENTRY_SIZE) % GA_LOG_SIZE;
801 writel(head, iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
803 /* Handle GA entry */
804 switch (GA_REQ_TYPE(log_entry)) {
806 if (!iommu_ga_log_notifier)
809 pr_debug("AMD-Vi: %s: devid=%#x, ga_tag=%#x\n",
810 __func__, GA_DEVID(log_entry),
813 if (iommu_ga_log_notifier(GA_TAG(log_entry)) != 0)
814 pr_err("AMD-Vi: GA log notifier failed.\n");
821 #endif /* CONFIG_IRQ_REMAP */
823 #define AMD_IOMMU_INT_MASK \
824 (MMIO_STATUS_EVT_INT_MASK | \
825 MMIO_STATUS_PPR_INT_MASK | \
826 MMIO_STATUS_GALOG_INT_MASK)
828 irqreturn_t amd_iommu_int_thread(int irq, void *data)
830 struct amd_iommu *iommu = (struct amd_iommu *) data;
831 u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
833 while (status & AMD_IOMMU_INT_MASK) {
834 /* Enable EVT and PPR and GA interrupts again */
835 writel(AMD_IOMMU_INT_MASK,
836 iommu->mmio_base + MMIO_STATUS_OFFSET);
838 if (status & MMIO_STATUS_EVT_INT_MASK) {
839 pr_devel("AMD-Vi: Processing IOMMU Event Log\n");
840 iommu_poll_events(iommu);
843 if (status & MMIO_STATUS_PPR_INT_MASK) {
844 pr_devel("AMD-Vi: Processing IOMMU PPR Log\n");
845 iommu_poll_ppr_log(iommu);
848 #ifdef CONFIG_IRQ_REMAP
849 if (status & MMIO_STATUS_GALOG_INT_MASK) {
850 pr_devel("AMD-Vi: Processing IOMMU GA Log\n");
851 iommu_poll_ga_log(iommu);
856 * Hardware bug: ERBT1312
857 * When re-enabling interrupt (by writing 1
858 * to clear the bit), the hardware might also try to set
859 * the interrupt bit in the event status register.
860 * In this scenario, the bit will be set, and disable
861 * subsequent interrupts.
863 * Workaround: The IOMMU driver should read back the
864 * status register and check if the interrupt bits are cleared.
865 * If not, driver will need to go through the interrupt handler
866 * again and re-clear the bits
868 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
873 irqreturn_t amd_iommu_int_handler(int irq, void *data)
875 return IRQ_WAKE_THREAD;
878 /****************************************************************************
880 * IOMMU command queuing functions
882 ****************************************************************************/
884 static int wait_on_sem(volatile u64 *sem)
888 while (*sem == 0 && i < LOOP_TIMEOUT) {
893 if (i == LOOP_TIMEOUT) {
894 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
901 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
902 struct iommu_cmd *cmd)
906 target = iommu->cmd_buf + iommu->cmd_buf_tail;
908 iommu->cmd_buf_tail += sizeof(*cmd);
909 iommu->cmd_buf_tail %= CMD_BUFFER_SIZE;
911 /* Copy command to buffer */
912 memcpy(target, cmd, sizeof(*cmd));
914 /* Tell the IOMMU about it */
915 writel(iommu->cmd_buf_tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
918 static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
920 WARN_ON(address & 0x7ULL);
922 memset(cmd, 0, sizeof(*cmd));
923 cmd->data[0] = lower_32_bits(__pa(address)) | CMD_COMPL_WAIT_STORE_MASK;
924 cmd->data[1] = upper_32_bits(__pa(address));
926 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
929 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
931 memset(cmd, 0, sizeof(*cmd));
932 cmd->data[0] = devid;
933 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
936 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
937 size_t size, u16 domid, int pde)
942 pages = iommu_num_pages(address, size, PAGE_SIZE);
947 * If we have to flush more than one page, flush all
948 * TLB entries for this domain
950 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
954 address &= PAGE_MASK;
956 memset(cmd, 0, sizeof(*cmd));
957 cmd->data[1] |= domid;
958 cmd->data[2] = lower_32_bits(address);
959 cmd->data[3] = upper_32_bits(address);
960 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
961 if (s) /* size bit - we flush more than one 4kb page */
962 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
963 if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
964 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
967 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
968 u64 address, size_t size)
973 pages = iommu_num_pages(address, size, PAGE_SIZE);
978 * If we have to flush more than one page, flush all
979 * TLB entries for this domain
981 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
985 address &= PAGE_MASK;
987 memset(cmd, 0, sizeof(*cmd));
988 cmd->data[0] = devid;
989 cmd->data[0] |= (qdep & 0xff) << 24;
990 cmd->data[1] = devid;
991 cmd->data[2] = lower_32_bits(address);
992 cmd->data[3] = upper_32_bits(address);
993 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
995 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
998 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, int pasid,
999 u64 address, bool size)
1001 memset(cmd, 0, sizeof(*cmd));
1003 address &= ~(0xfffULL);
1005 cmd->data[0] = pasid;
1006 cmd->data[1] = domid;
1007 cmd->data[2] = lower_32_bits(address);
1008 cmd->data[3] = upper_32_bits(address);
1009 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
1010 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
1012 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
1013 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
1016 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, int pasid,
1017 int qdep, u64 address, bool size)
1019 memset(cmd, 0, sizeof(*cmd));
1021 address &= ~(0xfffULL);
1023 cmd->data[0] = devid;
1024 cmd->data[0] |= ((pasid >> 8) & 0xff) << 16;
1025 cmd->data[0] |= (qdep & 0xff) << 24;
1026 cmd->data[1] = devid;
1027 cmd->data[1] |= (pasid & 0xff) << 16;
1028 cmd->data[2] = lower_32_bits(address);
1029 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
1030 cmd->data[3] = upper_32_bits(address);
1032 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
1033 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
1036 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, int pasid,
1037 int status, int tag, bool gn)
1039 memset(cmd, 0, sizeof(*cmd));
1041 cmd->data[0] = devid;
1043 cmd->data[1] = pasid;
1044 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
1046 cmd->data[3] = tag & 0x1ff;
1047 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
1049 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
1052 static void build_inv_all(struct iommu_cmd *cmd)
1054 memset(cmd, 0, sizeof(*cmd));
1055 CMD_SET_TYPE(cmd, CMD_INV_ALL);
1058 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
1060 memset(cmd, 0, sizeof(*cmd));
1061 cmd->data[0] = devid;
1062 CMD_SET_TYPE(cmd, CMD_INV_IRT);
1066 * Writes the command to the IOMMUs command buffer and informs the
1067 * hardware about the new command.
1069 static int __iommu_queue_command_sync(struct amd_iommu *iommu,
1070 struct iommu_cmd *cmd,
1073 unsigned int count = 0;
1074 u32 left, next_tail;
1076 next_tail = (iommu->cmd_buf_tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
1078 left = (iommu->cmd_buf_head - next_tail) % CMD_BUFFER_SIZE;
1081 /* Skip udelay() the first time around */
1083 if (count == LOOP_TIMEOUT) {
1084 pr_err("AMD-Vi: Command buffer timeout\n");
1091 /* Update head and recheck remaining space */
1092 iommu->cmd_buf_head = readl(iommu->mmio_base +
1093 MMIO_CMD_HEAD_OFFSET);
1098 copy_cmd_to_buffer(iommu, cmd);
1100 /* Do we need to make sure all commands are processed? */
1101 iommu->need_sync = sync;
1106 static int iommu_queue_command_sync(struct amd_iommu *iommu,
1107 struct iommu_cmd *cmd,
1110 unsigned long flags;
1113 spin_lock_irqsave(&iommu->lock, flags);
1114 ret = __iommu_queue_command_sync(iommu, cmd, sync);
1115 spin_unlock_irqrestore(&iommu->lock, flags);
1120 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
1122 return iommu_queue_command_sync(iommu, cmd, true);
1126 * This function queues a completion wait command into the command
1127 * buffer of an IOMMU
1129 static int iommu_completion_wait(struct amd_iommu *iommu)
1131 struct iommu_cmd cmd;
1132 unsigned long flags;
1135 if (!iommu->need_sync)
1139 build_completion_wait(&cmd, (u64)&iommu->cmd_sem);
1141 spin_lock_irqsave(&iommu->lock, flags);
1145 ret = __iommu_queue_command_sync(iommu, &cmd, false);
1149 ret = wait_on_sem(&iommu->cmd_sem);
1152 spin_unlock_irqrestore(&iommu->lock, flags);
1157 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
1159 struct iommu_cmd cmd;
1161 build_inv_dte(&cmd, devid);
1163 return iommu_queue_command(iommu, &cmd);
1166 static void iommu_flush_dte_all(struct amd_iommu *iommu)
1170 for (devid = 0; devid <= 0xffff; ++devid)
1171 iommu_flush_dte(iommu, devid);
1173 iommu_completion_wait(iommu);
1177 * This function uses heavy locking and may disable irqs for some time. But
1178 * this is no issue because it is only called during resume.
1180 static void iommu_flush_tlb_all(struct amd_iommu *iommu)
1184 for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
1185 struct iommu_cmd cmd;
1186 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1188 iommu_queue_command(iommu, &cmd);
1191 iommu_completion_wait(iommu);
1194 static void iommu_flush_all(struct amd_iommu *iommu)
1196 struct iommu_cmd cmd;
1198 build_inv_all(&cmd);
1200 iommu_queue_command(iommu, &cmd);
1201 iommu_completion_wait(iommu);
1204 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1206 struct iommu_cmd cmd;
1208 build_inv_irt(&cmd, devid);
1210 iommu_queue_command(iommu, &cmd);
1213 static void iommu_flush_irt_all(struct amd_iommu *iommu)
1217 for (devid = 0; devid <= MAX_DEV_TABLE_ENTRIES; devid++)
1218 iommu_flush_irt(iommu, devid);
1220 iommu_completion_wait(iommu);
1223 void iommu_flush_all_caches(struct amd_iommu *iommu)
1225 if (iommu_feature(iommu, FEATURE_IA)) {
1226 iommu_flush_all(iommu);
1228 iommu_flush_dte_all(iommu);
1229 iommu_flush_irt_all(iommu);
1230 iommu_flush_tlb_all(iommu);
1235 * Command send function for flushing on-device TLB
1237 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1238 u64 address, size_t size)
1240 struct amd_iommu *iommu;
1241 struct iommu_cmd cmd;
1244 qdep = dev_data->ats.qdep;
1245 iommu = amd_iommu_rlookup_table[dev_data->devid];
1247 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1249 return iommu_queue_command(iommu, &cmd);
1253 * Command send function for invalidating a device table entry
1255 static int device_flush_dte(struct iommu_dev_data *dev_data)
1257 struct amd_iommu *iommu;
1261 iommu = amd_iommu_rlookup_table[dev_data->devid];
1262 alias = dev_data->alias;
1264 ret = iommu_flush_dte(iommu, dev_data->devid);
1265 if (!ret && alias != dev_data->devid)
1266 ret = iommu_flush_dte(iommu, alias);
1270 if (dev_data->ats.enabled)
1271 ret = device_flush_iotlb(dev_data, 0, ~0UL);
1277 * TLB invalidation function which is called from the mapping functions.
1278 * It invalidates a single PTE if the range to flush is within a single
1279 * page. Otherwise it flushes the whole TLB of the IOMMU.
1281 static void __domain_flush_pages(struct protection_domain *domain,
1282 u64 address, size_t size, int pde)
1284 struct iommu_dev_data *dev_data;
1285 struct iommu_cmd cmd;
1288 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1290 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
1291 if (!domain->dev_iommu[i])
1295 * Devices of this domain are behind this IOMMU
1296 * We need a TLB flush
1298 ret |= iommu_queue_command(amd_iommus[i], &cmd);
1301 list_for_each_entry(dev_data, &domain->dev_list, list) {
1303 if (!dev_data->ats.enabled)
1306 ret |= device_flush_iotlb(dev_data, address, size);
1312 static void domain_flush_pages(struct protection_domain *domain,
1313 u64 address, size_t size)
1315 __domain_flush_pages(domain, address, size, 0);
1318 /* Flush the whole IO/TLB for a given protection domain */
1319 static void domain_flush_tlb(struct protection_domain *domain)
1321 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
1324 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1325 static void domain_flush_tlb_pde(struct protection_domain *domain)
1327 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1330 static void domain_flush_complete(struct protection_domain *domain)
1334 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
1335 if (domain && !domain->dev_iommu[i])
1339 * Devices of this domain are behind this IOMMU
1340 * We need to wait for completion of all commands.
1342 iommu_completion_wait(amd_iommus[i]);
1348 * This function flushes the DTEs for all devices in domain
1350 static void domain_flush_devices(struct protection_domain *domain)
1352 struct iommu_dev_data *dev_data;
1354 list_for_each_entry(dev_data, &domain->dev_list, list)
1355 device_flush_dte(dev_data);
1358 /****************************************************************************
1360 * The functions below are used the create the page table mappings for
1361 * unity mapped regions.
1363 ****************************************************************************/
1366 * This function is used to add another level to an IO page table. Adding
1367 * another level increases the size of the address space by 9 bits to a size up
1370 static bool increase_address_space(struct protection_domain *domain,
1375 if (domain->mode == PAGE_MODE_6_LEVEL)
1376 /* address space already 64 bit large */
1379 pte = (void *)get_zeroed_page(gfp);
1383 *pte = PM_LEVEL_PDE(domain->mode,
1384 virt_to_phys(domain->pt_root));
1385 domain->pt_root = pte;
1387 domain->updated = true;
1392 static u64 *alloc_pte(struct protection_domain *domain,
1393 unsigned long address,
1394 unsigned long page_size,
1401 BUG_ON(!is_power_of_2(page_size));
1403 while (address > PM_LEVEL_SIZE(domain->mode))
1404 increase_address_space(domain, gfp);
1406 level = domain->mode - 1;
1407 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1408 address = PAGE_SIZE_ALIGN(address, page_size);
1409 end_lvl = PAGE_SIZE_LEVEL(page_size);
1411 while (level > end_lvl) {
1416 if (!IOMMU_PTE_PRESENT(__pte)) {
1417 page = (u64 *)get_zeroed_page(gfp);
1421 __npte = PM_LEVEL_PDE(level, virt_to_phys(page));
1423 /* pte could have been changed somewhere. */
1424 if (cmpxchg64(pte, __pte, __npte) != __pte) {
1425 free_page((unsigned long)page);
1430 /* No level skipping support yet */
1431 if (PM_PTE_LEVEL(*pte) != level)
1436 pte = IOMMU_PTE_PAGE(*pte);
1438 if (pte_page && level == end_lvl)
1441 pte = &pte[PM_LEVEL_INDEX(level, address)];
1448 * This function checks if there is a PTE for a given dma address. If
1449 * there is one, it returns the pointer to it.
1451 static u64 *fetch_pte(struct protection_domain *domain,
1452 unsigned long address,
1453 unsigned long *page_size)
1458 if (address > PM_LEVEL_SIZE(domain->mode))
1461 level = domain->mode - 1;
1462 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1463 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1468 if (!IOMMU_PTE_PRESENT(*pte))
1472 if (PM_PTE_LEVEL(*pte) == 7 ||
1473 PM_PTE_LEVEL(*pte) == 0)
1476 /* No level skipping support yet */
1477 if (PM_PTE_LEVEL(*pte) != level)
1482 /* Walk to the next level */
1483 pte = IOMMU_PTE_PAGE(*pte);
1484 pte = &pte[PM_LEVEL_INDEX(level, address)];
1485 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1488 if (PM_PTE_LEVEL(*pte) == 0x07) {
1489 unsigned long pte_mask;
1492 * If we have a series of large PTEs, make
1493 * sure to return a pointer to the first one.
1495 *page_size = pte_mask = PTE_PAGE_SIZE(*pte);
1496 pte_mask = ~((PAGE_SIZE_PTE_COUNT(pte_mask) << 3) - 1);
1497 pte = (u64 *)(((unsigned long)pte) & pte_mask);
1504 * Generic mapping functions. It maps a physical address into a DMA
1505 * address space. It allocates the page table pages if necessary.
1506 * In the future it can be extended to a generic mapping function
1507 * supporting all features of AMD IOMMU page tables like level skipping
1508 * and full 64 bit address spaces.
1510 static int iommu_map_page(struct protection_domain *dom,
1511 unsigned long bus_addr,
1512 unsigned long phys_addr,
1513 unsigned long page_size,
1520 BUG_ON(!IS_ALIGNED(bus_addr, page_size));
1521 BUG_ON(!IS_ALIGNED(phys_addr, page_size));
1523 if (!(prot & IOMMU_PROT_MASK))
1526 count = PAGE_SIZE_PTE_COUNT(page_size);
1527 pte = alloc_pte(dom, bus_addr, page_size, NULL, gfp);
1532 for (i = 0; i < count; ++i)
1533 if (IOMMU_PTE_PRESENT(pte[i]))
1537 __pte = PAGE_SIZE_PTE(phys_addr, page_size);
1538 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_P | IOMMU_PTE_FC;
1540 __pte = phys_addr | IOMMU_PTE_P | IOMMU_PTE_FC;
1542 if (prot & IOMMU_PROT_IR)
1543 __pte |= IOMMU_PTE_IR;
1544 if (prot & IOMMU_PROT_IW)
1545 __pte |= IOMMU_PTE_IW;
1547 for (i = 0; i < count; ++i)
1555 static unsigned long iommu_unmap_page(struct protection_domain *dom,
1556 unsigned long bus_addr,
1557 unsigned long page_size)
1559 unsigned long long unmapped;
1560 unsigned long unmap_size;
1563 BUG_ON(!is_power_of_2(page_size));
1567 while (unmapped < page_size) {
1569 pte = fetch_pte(dom, bus_addr, &unmap_size);
1574 count = PAGE_SIZE_PTE_COUNT(unmap_size);
1575 for (i = 0; i < count; i++)
1579 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1580 unmapped += unmap_size;
1583 BUG_ON(unmapped && !is_power_of_2(unmapped));
1588 /****************************************************************************
1590 * The next functions belong to the address allocator for the dma_ops
1591 * interface functions.
1593 ****************************************************************************/
1596 static unsigned long dma_ops_alloc_iova(struct device *dev,
1597 struct dma_ops_domain *dma_dom,
1598 unsigned int pages, u64 dma_mask)
1600 unsigned long pfn = 0;
1602 pages = __roundup_pow_of_two(pages);
1604 if (dma_mask > DMA_BIT_MASK(32))
1605 pfn = alloc_iova_fast(&dma_dom->iovad, pages,
1606 IOVA_PFN(DMA_BIT_MASK(32)));
1609 pfn = alloc_iova_fast(&dma_dom->iovad, pages, IOVA_PFN(dma_mask));
1611 return (pfn << PAGE_SHIFT);
1614 static void dma_ops_free_iova(struct dma_ops_domain *dma_dom,
1615 unsigned long address,
1618 pages = __roundup_pow_of_two(pages);
1619 address >>= PAGE_SHIFT;
1621 free_iova_fast(&dma_dom->iovad, address, pages);
1624 /****************************************************************************
1626 * The next functions belong to the domain allocation. A domain is
1627 * allocated for every IOMMU as the default domain. If device isolation
1628 * is enabled, every device get its own domain. The most important thing
1629 * about domains is the page table mapping the DMA address space they
1632 ****************************************************************************/
1635 * This function adds a protection domain to the global protection domain list
1637 static void add_domain_to_list(struct protection_domain *domain)
1639 unsigned long flags;
1641 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1642 list_add(&domain->list, &amd_iommu_pd_list);
1643 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1647 * This function removes a protection domain to the global
1648 * protection domain list
1650 static void del_domain_from_list(struct protection_domain *domain)
1652 unsigned long flags;
1654 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1655 list_del(&domain->list);
1656 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1659 static u16 domain_id_alloc(void)
1661 unsigned long flags;
1664 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1665 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1667 if (id > 0 && id < MAX_DOMAIN_ID)
1668 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1671 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1676 static void domain_id_free(int id)
1678 unsigned long flags;
1680 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1681 if (id > 0 && id < MAX_DOMAIN_ID)
1682 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1683 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1686 #define DEFINE_FREE_PT_FN(LVL, FN) \
1687 static void free_pt_##LVL (unsigned long __pt) \
1695 for (i = 0; i < 512; ++i) { \
1696 /* PTE present? */ \
1697 if (!IOMMU_PTE_PRESENT(pt[i])) \
1701 if (PM_PTE_LEVEL(pt[i]) == 0 || \
1702 PM_PTE_LEVEL(pt[i]) == 7) \
1705 p = (unsigned long)IOMMU_PTE_PAGE(pt[i]); \
1708 free_page((unsigned long)pt); \
1711 DEFINE_FREE_PT_FN(l2, free_page)
1712 DEFINE_FREE_PT_FN(l3, free_pt_l2)
1713 DEFINE_FREE_PT_FN(l4, free_pt_l3)
1714 DEFINE_FREE_PT_FN(l5, free_pt_l4)
1715 DEFINE_FREE_PT_FN(l6, free_pt_l5)
1717 static void free_pagetable(struct protection_domain *domain)
1719 unsigned long root = (unsigned long)domain->pt_root;
1721 switch (domain->mode) {
1722 case PAGE_MODE_NONE:
1724 case PAGE_MODE_1_LEVEL:
1727 case PAGE_MODE_2_LEVEL:
1730 case PAGE_MODE_3_LEVEL:
1733 case PAGE_MODE_4_LEVEL:
1736 case PAGE_MODE_5_LEVEL:
1739 case PAGE_MODE_6_LEVEL:
1747 static void free_gcr3_tbl_level1(u64 *tbl)
1752 for (i = 0; i < 512; ++i) {
1753 if (!(tbl[i] & GCR3_VALID))
1756 ptr = __va(tbl[i] & PAGE_MASK);
1758 free_page((unsigned long)ptr);
1762 static void free_gcr3_tbl_level2(u64 *tbl)
1767 for (i = 0; i < 512; ++i) {
1768 if (!(tbl[i] & GCR3_VALID))
1771 ptr = __va(tbl[i] & PAGE_MASK);
1773 free_gcr3_tbl_level1(ptr);
1777 static void free_gcr3_table(struct protection_domain *domain)
1779 if (domain->glx == 2)
1780 free_gcr3_tbl_level2(domain->gcr3_tbl);
1781 else if (domain->glx == 1)
1782 free_gcr3_tbl_level1(domain->gcr3_tbl);
1784 BUG_ON(domain->glx != 0);
1786 free_page((unsigned long)domain->gcr3_tbl);
1789 static void dma_ops_domain_free_flush_queue(struct dma_ops_domain *dom)
1793 for_each_possible_cpu(cpu) {
1794 struct flush_queue *queue;
1796 queue = per_cpu_ptr(dom->flush_queue, cpu);
1797 kfree(queue->entries);
1800 free_percpu(dom->flush_queue);
1802 dom->flush_queue = NULL;
1805 static int dma_ops_domain_alloc_flush_queue(struct dma_ops_domain *dom)
1809 atomic64_set(&dom->flush_start_cnt, 0);
1810 atomic64_set(&dom->flush_finish_cnt, 0);
1812 dom->flush_queue = alloc_percpu(struct flush_queue);
1813 if (!dom->flush_queue)
1816 /* First make sure everything is cleared */
1817 for_each_possible_cpu(cpu) {
1818 struct flush_queue *queue;
1820 queue = per_cpu_ptr(dom->flush_queue, cpu);
1823 queue->entries = NULL;
1826 /* Now start doing the allocation */
1827 for_each_possible_cpu(cpu) {
1828 struct flush_queue *queue;
1830 queue = per_cpu_ptr(dom->flush_queue, cpu);
1831 queue->entries = kzalloc(FLUSH_QUEUE_SIZE * sizeof(*queue->entries),
1833 if (!queue->entries) {
1834 dma_ops_domain_free_flush_queue(dom);
1838 spin_lock_init(&queue->lock);
1844 static void dma_ops_domain_flush_tlb(struct dma_ops_domain *dom)
1846 atomic64_inc(&dom->flush_start_cnt);
1847 domain_flush_tlb(&dom->domain);
1848 domain_flush_complete(&dom->domain);
1849 atomic64_inc(&dom->flush_finish_cnt);
1852 static inline bool queue_ring_full(struct flush_queue *queue)
1854 assert_spin_locked(&queue->lock);
1856 return (((queue->tail + 1) % FLUSH_QUEUE_SIZE) == queue->head);
1859 #define queue_ring_for_each(i, q) \
1860 for (i = (q)->head; i != (q)->tail; i = (i + 1) % FLUSH_QUEUE_SIZE)
1862 static inline unsigned queue_ring_add(struct flush_queue *queue)
1864 unsigned idx = queue->tail;
1866 assert_spin_locked(&queue->lock);
1867 queue->tail = (idx + 1) % FLUSH_QUEUE_SIZE;
1872 static inline void queue_ring_remove_head(struct flush_queue *queue)
1874 assert_spin_locked(&queue->lock);
1875 queue->head = (queue->head + 1) % FLUSH_QUEUE_SIZE;
1878 static void queue_ring_free_flushed(struct dma_ops_domain *dom,
1879 struct flush_queue *queue)
1881 u64 counter = atomic64_read(&dom->flush_finish_cnt);
1884 queue_ring_for_each(idx, queue) {
1886 * This assumes that counter values in the ring-buffer are
1887 * monotonously rising.
1889 if (queue->entries[idx].counter >= counter)
1892 free_iova_fast(&dom->iovad,
1893 queue->entries[idx].iova_pfn,
1894 queue->entries[idx].pages);
1896 queue_ring_remove_head(queue);
1900 static void queue_add(struct dma_ops_domain *dom,
1901 unsigned long address, unsigned long pages)
1903 struct flush_queue *queue;
1904 unsigned long flags;
1907 pages = __roundup_pow_of_two(pages);
1908 address >>= PAGE_SHIFT;
1910 queue = get_cpu_ptr(dom->flush_queue);
1911 spin_lock_irqsave(&queue->lock, flags);
1914 * When ring-queue is full, flush the entries from the IOTLB so
1915 * that we can free all entries with queue_ring_free_flushed()
1918 if (queue_ring_full(queue))
1919 dma_ops_domain_flush_tlb(dom);
1921 queue_ring_free_flushed(dom, queue);
1923 idx = queue_ring_add(queue);
1925 queue->entries[idx].iova_pfn = address;
1926 queue->entries[idx].pages = pages;
1927 queue->entries[idx].counter = atomic64_read(&dom->flush_start_cnt);
1929 spin_unlock_irqrestore(&queue->lock, flags);
1931 if (atomic_cmpxchg(&dom->flush_timer_on, 0, 1) == 0)
1932 mod_timer(&dom->flush_timer, jiffies + msecs_to_jiffies(10));
1934 put_cpu_ptr(dom->flush_queue);
1937 static void queue_flush_timeout(unsigned long data)
1939 struct dma_ops_domain *dom = (struct dma_ops_domain *)data;
1942 atomic_set(&dom->flush_timer_on, 0);
1944 dma_ops_domain_flush_tlb(dom);
1946 for_each_possible_cpu(cpu) {
1947 struct flush_queue *queue;
1948 unsigned long flags;
1950 queue = per_cpu_ptr(dom->flush_queue, cpu);
1951 spin_lock_irqsave(&queue->lock, flags);
1952 queue_ring_free_flushed(dom, queue);
1953 spin_unlock_irqrestore(&queue->lock, flags);
1958 * Free a domain, only used if something went wrong in the
1959 * allocation path and we need to free an already allocated page table
1961 static void dma_ops_domain_free(struct dma_ops_domain *dom)
1966 del_domain_from_list(&dom->domain);
1968 if (timer_pending(&dom->flush_timer))
1969 del_timer(&dom->flush_timer);
1971 dma_ops_domain_free_flush_queue(dom);
1973 put_iova_domain(&dom->iovad);
1975 free_pagetable(&dom->domain);
1978 domain_id_free(dom->domain.id);
1984 * Allocates a new protection domain usable for the dma_ops functions.
1985 * It also initializes the page table and the address allocator data
1986 * structures required for the dma_ops interface
1988 static struct dma_ops_domain *dma_ops_domain_alloc(void)
1990 struct dma_ops_domain *dma_dom;
1992 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
1996 if (protection_domain_init(&dma_dom->domain))
1999 dma_dom->domain.mode = PAGE_MODE_3_LEVEL;
2000 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2001 dma_dom->domain.flags = PD_DMA_OPS_MASK;
2002 if (!dma_dom->domain.pt_root)
2005 init_iova_domain(&dma_dom->iovad, PAGE_SIZE,
2006 IOVA_START_PFN, DMA_32BIT_PFN);
2008 /* Initialize reserved ranges */
2009 copy_reserved_iova(&reserved_iova_ranges, &dma_dom->iovad);
2011 if (dma_ops_domain_alloc_flush_queue(dma_dom))
2014 setup_timer(&dma_dom->flush_timer, queue_flush_timeout,
2015 (unsigned long)dma_dom);
2017 atomic_set(&dma_dom->flush_timer_on, 0);
2019 add_domain_to_list(&dma_dom->domain);
2024 dma_ops_domain_free(dma_dom);
2030 * little helper function to check whether a given protection domain is a
2033 static bool dma_ops_domain(struct protection_domain *domain)
2035 return domain->flags & PD_DMA_OPS_MASK;
2038 static void set_dte_entry(u16 devid, struct protection_domain *domain, bool ats)
2043 if (domain->mode != PAGE_MODE_NONE)
2044 pte_root = virt_to_phys(domain->pt_root);
2046 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
2047 << DEV_ENTRY_MODE_SHIFT;
2048 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;
2050 flags = amd_iommu_dev_table[devid].data[1];
2053 flags |= DTE_FLAG_IOTLB;
2055 if (domain->flags & PD_IOMMUV2_MASK) {
2056 u64 gcr3 = __pa(domain->gcr3_tbl);
2057 u64 glx = domain->glx;
2060 pte_root |= DTE_FLAG_GV;
2061 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
2063 /* First mask out possible old values for GCR3 table */
2064 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
2067 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
2070 /* Encode GCR3 table into DTE */
2071 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
2074 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
2077 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
2082 flags &= ~(DTE_FLAG_SA | 0xffffULL);
2083 flags |= domain->id;
2085 amd_iommu_dev_table[devid].data[1] = flags;
2086 amd_iommu_dev_table[devid].data[0] = pte_root;
2089 static void clear_dte_entry(u16 devid)
2091 /* remove entry from the device table seen by the hardware */
2092 amd_iommu_dev_table[devid].data[0] = IOMMU_PTE_P | IOMMU_PTE_TV;
2093 amd_iommu_dev_table[devid].data[1] &= DTE_FLAG_MASK;
2095 amd_iommu_apply_erratum_63(devid);
2098 static void do_attach(struct iommu_dev_data *dev_data,
2099 struct protection_domain *domain)
2101 struct amd_iommu *iommu;
2105 iommu = amd_iommu_rlookup_table[dev_data->devid];
2106 alias = dev_data->alias;
2107 ats = dev_data->ats.enabled;
2109 /* Update data structures */
2110 dev_data->domain = domain;
2111 list_add(&dev_data->list, &domain->dev_list);
2113 /* Do reference counting */
2114 domain->dev_iommu[iommu->index] += 1;
2115 domain->dev_cnt += 1;
2117 /* Update device table */
2118 set_dte_entry(dev_data->devid, domain, ats);
2119 if (alias != dev_data->devid)
2120 set_dte_entry(alias, domain, ats);
2122 device_flush_dte(dev_data);
2125 static void do_detach(struct iommu_dev_data *dev_data)
2127 struct amd_iommu *iommu;
2131 * First check if the device is still attached. It might already
2132 * be detached from its domain because the generic
2133 * iommu_detach_group code detached it and we try again here in
2134 * our alias handling.
2136 if (!dev_data->domain)
2139 iommu = amd_iommu_rlookup_table[dev_data->devid];
2140 alias = dev_data->alias;
2142 /* decrease reference counters */
2143 dev_data->domain->dev_iommu[iommu->index] -= 1;
2144 dev_data->domain->dev_cnt -= 1;
2146 /* Update data structures */
2147 dev_data->domain = NULL;
2148 list_del(&dev_data->list);
2149 clear_dte_entry(dev_data->devid);
2150 if (alias != dev_data->devid)
2151 clear_dte_entry(alias);
2153 /* Flush the DTE entry */
2154 device_flush_dte(dev_data);
2158 * If a device is not yet associated with a domain, this function does
2159 * assigns it visible for the hardware
2161 static int __attach_device(struct iommu_dev_data *dev_data,
2162 struct protection_domain *domain)
2167 * Must be called with IRQs disabled. Warn here to detect early
2170 WARN_ON(!irqs_disabled());
2173 spin_lock(&domain->lock);
2176 if (dev_data->domain != NULL)
2179 /* Attach alias group root */
2180 do_attach(dev_data, domain);
2187 spin_unlock(&domain->lock);
2193 static void pdev_iommuv2_disable(struct pci_dev *pdev)
2195 pci_disable_ats(pdev);
2196 pci_disable_pri(pdev);
2197 pci_disable_pasid(pdev);
2200 /* FIXME: Change generic reset-function to do the same */
2201 static int pri_reset_while_enabled(struct pci_dev *pdev)
2206 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2210 pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
2211 control |= PCI_PRI_CTRL_RESET;
2212 pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
2217 static int pdev_iommuv2_enable(struct pci_dev *pdev)
2222 /* FIXME: Hardcode number of outstanding requests for now */
2224 if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
2226 reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
2228 /* Only allow access to user-accessible pages */
2229 ret = pci_enable_pasid(pdev, 0);
2233 /* First reset the PRI state of the device */
2234 ret = pci_reset_pri(pdev);
2239 ret = pci_enable_pri(pdev, reqs);
2244 ret = pri_reset_while_enabled(pdev);
2249 ret = pci_enable_ats(pdev, PAGE_SHIFT);
2256 pci_disable_pri(pdev);
2257 pci_disable_pasid(pdev);
2262 /* FIXME: Move this to PCI code */
2263 #define PCI_PRI_TLP_OFF (1 << 15)
2265 static bool pci_pri_tlp_required(struct pci_dev *pdev)
2270 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2274 pci_read_config_word(pdev, pos + PCI_PRI_STATUS, &status);
2276 return (status & PCI_PRI_TLP_OFF) ? true : false;
2280 * If a device is not yet associated with a domain, this function
2281 * assigns it visible for the hardware
2283 static int attach_device(struct device *dev,
2284 struct protection_domain *domain)
2286 struct pci_dev *pdev;
2287 struct iommu_dev_data *dev_data;
2288 unsigned long flags;
2291 dev_data = get_dev_data(dev);
2293 if (!dev_is_pci(dev))
2294 goto skip_ats_check;
2296 pdev = to_pci_dev(dev);
2297 if (domain->flags & PD_IOMMUV2_MASK) {
2298 if (!dev_data->passthrough)
2301 if (dev_data->iommu_v2) {
2302 if (pdev_iommuv2_enable(pdev) != 0)
2305 dev_data->ats.enabled = true;
2306 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2307 dev_data->pri_tlp = pci_pri_tlp_required(pdev);
2309 } else if (amd_iommu_iotlb_sup &&
2310 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
2311 dev_data->ats.enabled = true;
2312 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2316 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2317 ret = __attach_device(dev_data, domain);
2318 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2321 * We might boot into a crash-kernel here. The crashed kernel
2322 * left the caches in the IOMMU dirty. So we have to flush
2323 * here to evict all dirty stuff.
2325 domain_flush_tlb_pde(domain);
2331 * Removes a device from a protection domain (unlocked)
2333 static void __detach_device(struct iommu_dev_data *dev_data)
2335 struct protection_domain *domain;
2338 * Must be called with IRQs disabled. Warn here to detect early
2341 WARN_ON(!irqs_disabled());
2343 if (WARN_ON(!dev_data->domain))
2346 domain = dev_data->domain;
2348 spin_lock(&domain->lock);
2350 do_detach(dev_data);
2352 spin_unlock(&domain->lock);
2356 * Removes a device from a protection domain (with devtable_lock held)
2358 static void detach_device(struct device *dev)
2360 struct protection_domain *domain;
2361 struct iommu_dev_data *dev_data;
2362 unsigned long flags;
2364 dev_data = get_dev_data(dev);
2365 domain = dev_data->domain;
2367 /* lock device table */
2368 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2369 __detach_device(dev_data);
2370 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2372 if (!dev_is_pci(dev))
2375 if (domain->flags & PD_IOMMUV2_MASK && dev_data->iommu_v2)
2376 pdev_iommuv2_disable(to_pci_dev(dev));
2377 else if (dev_data->ats.enabled)
2378 pci_disable_ats(to_pci_dev(dev));
2380 dev_data->ats.enabled = false;
2383 static int amd_iommu_add_device(struct device *dev)
2385 struct iommu_dev_data *dev_data;
2386 struct iommu_domain *domain;
2387 struct amd_iommu *iommu;
2390 if (!check_device(dev) || get_dev_data(dev))
2393 devid = get_device_id(dev);
2397 iommu = amd_iommu_rlookup_table[devid];
2399 ret = iommu_init_device(dev);
2401 if (ret != -ENOTSUPP)
2402 pr_err("Failed to initialize device %s - trying to proceed anyway\n",
2405 iommu_ignore_device(dev);
2406 dev->dma_ops = &nommu_dma_ops;
2409 init_iommu_group(dev);
2411 dev_data = get_dev_data(dev);
2415 if (iommu_pass_through || dev_data->iommu_v2)
2416 iommu_request_dm_for_dev(dev);
2418 /* Domains are initialized for this device - have a look what we ended up with */
2419 domain = iommu_get_domain_for_dev(dev);
2420 if (domain->type == IOMMU_DOMAIN_IDENTITY)
2421 dev_data->passthrough = true;
2423 dev->dma_ops = &amd_iommu_dma_ops;
2426 iommu_completion_wait(iommu);
2431 static void amd_iommu_remove_device(struct device *dev)
2433 struct amd_iommu *iommu;
2436 if (!check_device(dev))
2439 devid = get_device_id(dev);
2443 iommu = amd_iommu_rlookup_table[devid];
2445 iommu_uninit_device(dev);
2446 iommu_completion_wait(iommu);
2449 static struct iommu_group *amd_iommu_device_group(struct device *dev)
2451 if (dev_is_pci(dev))
2452 return pci_device_group(dev);
2454 return acpihid_device_group(dev);
2457 /*****************************************************************************
2459 * The next functions belong to the dma_ops mapping/unmapping code.
2461 *****************************************************************************/
2464 * In the dma_ops path we only have the struct device. This function
2465 * finds the corresponding IOMMU, the protection domain and the
2466 * requestor id for a given device.
2467 * If the device is not yet associated with a domain this is also done
2470 static struct protection_domain *get_domain(struct device *dev)
2472 struct protection_domain *domain;
2474 if (!check_device(dev))
2475 return ERR_PTR(-EINVAL);
2477 domain = get_dev_data(dev)->domain;
2478 if (!dma_ops_domain(domain))
2479 return ERR_PTR(-EBUSY);
2484 static void update_device_table(struct protection_domain *domain)
2486 struct iommu_dev_data *dev_data;
2488 list_for_each_entry(dev_data, &domain->dev_list, list) {
2489 set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled);
2491 if (dev_data->devid == dev_data->alias)
2494 /* There is an alias, update device table entry for it */
2495 set_dte_entry(dev_data->alias, domain, dev_data->ats.enabled);
2499 static void update_domain(struct protection_domain *domain)
2501 if (!domain->updated)
2504 update_device_table(domain);
2506 domain_flush_devices(domain);
2507 domain_flush_tlb_pde(domain);
2509 domain->updated = false;
2512 static int dir2prot(enum dma_data_direction direction)
2514 if (direction == DMA_TO_DEVICE)
2515 return IOMMU_PROT_IR;
2516 else if (direction == DMA_FROM_DEVICE)
2517 return IOMMU_PROT_IW;
2518 else if (direction == DMA_BIDIRECTIONAL)
2519 return IOMMU_PROT_IW | IOMMU_PROT_IR;
2524 * This function contains common code for mapping of a physically
2525 * contiguous memory region into DMA address space. It is used by all
2526 * mapping functions provided with this IOMMU driver.
2527 * Must be called with the domain lock held.
2529 static dma_addr_t __map_single(struct device *dev,
2530 struct dma_ops_domain *dma_dom,
2533 enum dma_data_direction direction,
2536 dma_addr_t offset = paddr & ~PAGE_MASK;
2537 dma_addr_t address, start, ret;
2542 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
2545 address = dma_ops_alloc_iova(dev, dma_dom, pages, dma_mask);
2546 if (address == DMA_ERROR_CODE)
2549 prot = dir2prot(direction);
2552 for (i = 0; i < pages; ++i) {
2553 ret = iommu_map_page(&dma_dom->domain, start, paddr,
2554 PAGE_SIZE, prot, GFP_ATOMIC);
2563 if (unlikely(amd_iommu_np_cache)) {
2564 domain_flush_pages(&dma_dom->domain, address, size);
2565 domain_flush_complete(&dma_dom->domain);
2573 for (--i; i >= 0; --i) {
2575 iommu_unmap_page(&dma_dom->domain, start, PAGE_SIZE);
2578 domain_flush_tlb(&dma_dom->domain);
2579 domain_flush_complete(&dma_dom->domain);
2581 dma_ops_free_iova(dma_dom, address, pages);
2583 return DMA_ERROR_CODE;
2587 * Does the reverse of the __map_single function. Must be called with
2588 * the domain lock held too
2590 static void __unmap_single(struct dma_ops_domain *dma_dom,
2591 dma_addr_t dma_addr,
2595 dma_addr_t flush_addr;
2596 dma_addr_t i, start;
2599 flush_addr = dma_addr;
2600 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
2601 dma_addr &= PAGE_MASK;
2604 for (i = 0; i < pages; ++i) {
2605 iommu_unmap_page(&dma_dom->domain, start, PAGE_SIZE);
2609 if (amd_iommu_unmap_flush) {
2610 dma_ops_free_iova(dma_dom, dma_addr, pages);
2611 domain_flush_tlb(&dma_dom->domain);
2612 domain_flush_complete(&dma_dom->domain);
2614 queue_add(dma_dom, dma_addr, pages);
2619 * The exported map_single function for dma_ops.
2621 static dma_addr_t map_page(struct device *dev, struct page *page,
2622 unsigned long offset, size_t size,
2623 enum dma_data_direction dir,
2624 unsigned long attrs)
2626 phys_addr_t paddr = page_to_phys(page) + offset;
2627 struct protection_domain *domain;
2628 struct dma_ops_domain *dma_dom;
2631 domain = get_domain(dev);
2632 if (PTR_ERR(domain) == -EINVAL)
2633 return (dma_addr_t)paddr;
2634 else if (IS_ERR(domain))
2635 return DMA_ERROR_CODE;
2637 dma_mask = *dev->dma_mask;
2638 dma_dom = to_dma_ops_domain(domain);
2640 return __map_single(dev, dma_dom, paddr, size, dir, dma_mask);
2644 * The exported unmap_single function for dma_ops.
2646 static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2647 enum dma_data_direction dir, unsigned long attrs)
2649 struct protection_domain *domain;
2650 struct dma_ops_domain *dma_dom;
2652 domain = get_domain(dev);
2656 dma_dom = to_dma_ops_domain(domain);
2658 __unmap_single(dma_dom, dma_addr, size, dir);
2661 static int sg_num_pages(struct device *dev,
2662 struct scatterlist *sglist,
2665 unsigned long mask, boundary_size;
2666 struct scatterlist *s;
2669 mask = dma_get_seg_boundary(dev);
2670 boundary_size = mask + 1 ? ALIGN(mask + 1, PAGE_SIZE) >> PAGE_SHIFT :
2671 1UL << (BITS_PER_LONG - PAGE_SHIFT);
2673 for_each_sg(sglist, s, nelems, i) {
2676 s->dma_address = npages << PAGE_SHIFT;
2677 p = npages % boundary_size;
2678 n = iommu_num_pages(sg_phys(s), s->length, PAGE_SIZE);
2679 if (p + n > boundary_size)
2680 npages += boundary_size - p;
2688 * The exported map_sg function for dma_ops (handles scatter-gather
2691 static int map_sg(struct device *dev, struct scatterlist *sglist,
2692 int nelems, enum dma_data_direction direction,
2693 unsigned long attrs)
2695 int mapped_pages = 0, npages = 0, prot = 0, i;
2696 struct protection_domain *domain;
2697 struct dma_ops_domain *dma_dom;
2698 struct scatterlist *s;
2699 unsigned long address;
2702 domain = get_domain(dev);
2706 dma_dom = to_dma_ops_domain(domain);
2707 dma_mask = *dev->dma_mask;
2709 npages = sg_num_pages(dev, sglist, nelems);
2711 address = dma_ops_alloc_iova(dev, dma_dom, npages, dma_mask);
2712 if (address == DMA_ERROR_CODE)
2715 prot = dir2prot(direction);
2717 /* Map all sg entries */
2718 for_each_sg(sglist, s, nelems, i) {
2719 int j, pages = iommu_num_pages(sg_phys(s), s->length, PAGE_SIZE);
2721 for (j = 0; j < pages; ++j) {
2722 unsigned long bus_addr, phys_addr;
2725 bus_addr = address + s->dma_address + (j << PAGE_SHIFT);
2726 phys_addr = (sg_phys(s) & PAGE_MASK) + (j << PAGE_SHIFT);
2727 ret = iommu_map_page(domain, bus_addr, phys_addr, PAGE_SIZE, prot, GFP_ATOMIC);
2735 /* Everything is mapped - write the right values into s->dma_address */
2736 for_each_sg(sglist, s, nelems, i) {
2737 s->dma_address += address + s->offset;
2738 s->dma_length = s->length;
2744 pr_err("%s: IOMMU mapping error in map_sg (io-pages: %d)\n",
2745 dev_name(dev), npages);
2747 for_each_sg(sglist, s, nelems, i) {
2748 int j, pages = iommu_num_pages(sg_phys(s), s->length, PAGE_SIZE);
2750 for (j = 0; j < pages; ++j) {
2751 unsigned long bus_addr;
2753 bus_addr = address + s->dma_address + (j << PAGE_SHIFT);
2754 iommu_unmap_page(domain, bus_addr, PAGE_SIZE);
2762 free_iova_fast(&dma_dom->iovad, address, npages);
2769 * The exported map_sg function for dma_ops (handles scatter-gather
2772 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2773 int nelems, enum dma_data_direction dir,
2774 unsigned long attrs)
2776 struct protection_domain *domain;
2777 struct dma_ops_domain *dma_dom;
2778 unsigned long startaddr;
2781 domain = get_domain(dev);
2785 startaddr = sg_dma_address(sglist) & PAGE_MASK;
2786 dma_dom = to_dma_ops_domain(domain);
2787 npages = sg_num_pages(dev, sglist, nelems);
2789 __unmap_single(dma_dom, startaddr, npages << PAGE_SHIFT, dir);
2793 * The exported alloc_coherent function for dma_ops.
2795 static void *alloc_coherent(struct device *dev, size_t size,
2796 dma_addr_t *dma_addr, gfp_t flag,
2797 unsigned long attrs)
2799 u64 dma_mask = dev->coherent_dma_mask;
2800 struct protection_domain *domain;
2801 struct dma_ops_domain *dma_dom;
2804 domain = get_domain(dev);
2805 if (PTR_ERR(domain) == -EINVAL) {
2806 page = alloc_pages(flag, get_order(size));
2807 *dma_addr = page_to_phys(page);
2808 return page_address(page);
2809 } else if (IS_ERR(domain))
2812 dma_dom = to_dma_ops_domain(domain);
2813 size = PAGE_ALIGN(size);
2814 dma_mask = dev->coherent_dma_mask;
2815 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
2818 page = alloc_pages(flag | __GFP_NOWARN, get_order(size));
2820 if (!gfpflags_allow_blocking(flag))
2823 page = dma_alloc_from_contiguous(dev, size >> PAGE_SHIFT,
2824 get_order(size), flag);
2830 dma_mask = *dev->dma_mask;
2832 *dma_addr = __map_single(dev, dma_dom, page_to_phys(page),
2833 size, DMA_BIDIRECTIONAL, dma_mask);
2835 if (*dma_addr == DMA_ERROR_CODE)
2838 return page_address(page);
2842 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2843 __free_pages(page, get_order(size));
2849 * The exported free_coherent function for dma_ops.
2851 static void free_coherent(struct device *dev, size_t size,
2852 void *virt_addr, dma_addr_t dma_addr,
2853 unsigned long attrs)
2855 struct protection_domain *domain;
2856 struct dma_ops_domain *dma_dom;
2859 page = virt_to_page(virt_addr);
2860 size = PAGE_ALIGN(size);
2862 domain = get_domain(dev);
2866 dma_dom = to_dma_ops_domain(domain);
2868 __unmap_single(dma_dom, dma_addr, size, DMA_BIDIRECTIONAL);
2871 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2872 __free_pages(page, get_order(size));
2876 * This function is called by the DMA layer to find out if we can handle a
2877 * particular device. It is part of the dma_ops.
2879 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
2881 return check_device(dev);
2884 static const struct dma_map_ops amd_iommu_dma_ops = {
2885 .alloc = alloc_coherent,
2886 .free = free_coherent,
2887 .map_page = map_page,
2888 .unmap_page = unmap_page,
2890 .unmap_sg = unmap_sg,
2891 .dma_supported = amd_iommu_dma_supported,
2894 static int init_reserved_iova_ranges(void)
2896 struct pci_dev *pdev = NULL;
2899 init_iova_domain(&reserved_iova_ranges, PAGE_SIZE,
2900 IOVA_START_PFN, DMA_32BIT_PFN);
2902 lockdep_set_class(&reserved_iova_ranges.iova_rbtree_lock,
2903 &reserved_rbtree_key);
2905 /* MSI memory range */
2906 val = reserve_iova(&reserved_iova_ranges,
2907 IOVA_PFN(MSI_RANGE_START), IOVA_PFN(MSI_RANGE_END));
2909 pr_err("Reserving MSI range failed\n");
2913 /* HT memory range */
2914 val = reserve_iova(&reserved_iova_ranges,
2915 IOVA_PFN(HT_RANGE_START), IOVA_PFN(HT_RANGE_END));
2917 pr_err("Reserving HT range failed\n");
2922 * Memory used for PCI resources
2923 * FIXME: Check whether we can reserve the PCI-hole completly
2925 for_each_pci_dev(pdev) {
2928 for (i = 0; i < PCI_NUM_RESOURCES; ++i) {
2929 struct resource *r = &pdev->resource[i];
2931 if (!(r->flags & IORESOURCE_MEM))
2934 val = reserve_iova(&reserved_iova_ranges,
2938 pr_err("Reserve pci-resource range failed\n");
2947 int __init amd_iommu_init_api(void)
2951 ret = iova_cache_get();
2955 ret = init_reserved_iova_ranges();
2959 err = bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
2962 #ifdef CONFIG_ARM_AMBA
2963 err = bus_set_iommu(&amba_bustype, &amd_iommu_ops);
2967 err = bus_set_iommu(&platform_bus_type, &amd_iommu_ops);
2974 int __init amd_iommu_init_dma_ops(void)
2976 swiotlb = iommu_pass_through ? 1 : 0;
2980 * In case we don't initialize SWIOTLB (actually the common case
2981 * when AMD IOMMU is enabled), make sure there are global
2982 * dma_ops set as a fall-back for devices not handled by this
2983 * driver (for example non-PCI devices).
2986 dma_ops = &nommu_dma_ops;
2988 if (amd_iommu_unmap_flush)
2989 pr_info("AMD-Vi: IO/TLB flush on unmap enabled\n");
2991 pr_info("AMD-Vi: Lazy IO/TLB flushing enabled\n");
2997 /*****************************************************************************
2999 * The following functions belong to the exported interface of AMD IOMMU
3001 * This interface allows access to lower level functions of the IOMMU
3002 * like protection domain handling and assignement of devices to domains
3003 * which is not possible with the dma_ops interface.
3005 *****************************************************************************/
3007 static void cleanup_domain(struct protection_domain *domain)
3009 struct iommu_dev_data *entry;
3010 unsigned long flags;
3012 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3014 while (!list_empty(&domain->dev_list)) {
3015 entry = list_first_entry(&domain->dev_list,
3016 struct iommu_dev_data, list);
3017 __detach_device(entry);
3020 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3023 static void protection_domain_free(struct protection_domain *domain)
3028 del_domain_from_list(domain);
3031 domain_id_free(domain->id);
3036 static int protection_domain_init(struct protection_domain *domain)
3038 spin_lock_init(&domain->lock);
3039 mutex_init(&domain->api_lock);
3040 domain->id = domain_id_alloc();
3043 INIT_LIST_HEAD(&domain->dev_list);
3048 static struct protection_domain *protection_domain_alloc(void)
3050 struct protection_domain *domain;
3052 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
3056 if (protection_domain_init(domain))
3059 add_domain_to_list(domain);
3069 static struct iommu_domain *amd_iommu_domain_alloc(unsigned type)
3071 struct protection_domain *pdomain;
3072 struct dma_ops_domain *dma_domain;
3075 case IOMMU_DOMAIN_UNMANAGED:
3076 pdomain = protection_domain_alloc();
3080 pdomain->mode = PAGE_MODE_3_LEVEL;
3081 pdomain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
3082 if (!pdomain->pt_root) {
3083 protection_domain_free(pdomain);
3087 pdomain->domain.geometry.aperture_start = 0;
3088 pdomain->domain.geometry.aperture_end = ~0ULL;
3089 pdomain->domain.geometry.force_aperture = true;
3092 case IOMMU_DOMAIN_DMA:
3093 dma_domain = dma_ops_domain_alloc();
3095 pr_err("AMD-Vi: Failed to allocate\n");
3098 pdomain = &dma_domain->domain;
3100 case IOMMU_DOMAIN_IDENTITY:
3101 pdomain = protection_domain_alloc();
3105 pdomain->mode = PAGE_MODE_NONE;
3111 return &pdomain->domain;
3114 static void amd_iommu_domain_free(struct iommu_domain *dom)
3116 struct protection_domain *domain;
3117 struct dma_ops_domain *dma_dom;
3119 domain = to_pdomain(dom);
3121 if (domain->dev_cnt > 0)
3122 cleanup_domain(domain);
3124 BUG_ON(domain->dev_cnt != 0);
3129 switch (dom->type) {
3130 case IOMMU_DOMAIN_DMA:
3131 /* Now release the domain */
3132 dma_dom = to_dma_ops_domain(domain);
3133 dma_ops_domain_free(dma_dom);
3136 if (domain->mode != PAGE_MODE_NONE)
3137 free_pagetable(domain);
3139 if (domain->flags & PD_IOMMUV2_MASK)
3140 free_gcr3_table(domain);
3142 protection_domain_free(domain);
3147 static void amd_iommu_detach_device(struct iommu_domain *dom,
3150 struct iommu_dev_data *dev_data = dev->archdata.iommu;
3151 struct amd_iommu *iommu;
3154 if (!check_device(dev))
3157 devid = get_device_id(dev);
3161 if (dev_data->domain != NULL)
3164 iommu = amd_iommu_rlookup_table[devid];
3168 #ifdef CONFIG_IRQ_REMAP
3169 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) &&
3170 (dom->type == IOMMU_DOMAIN_UNMANAGED))
3171 dev_data->use_vapic = 0;
3174 iommu_completion_wait(iommu);
3177 static int amd_iommu_attach_device(struct iommu_domain *dom,
3180 struct protection_domain *domain = to_pdomain(dom);
3181 struct iommu_dev_data *dev_data;
3182 struct amd_iommu *iommu;
3185 if (!check_device(dev))
3188 dev_data = dev->archdata.iommu;
3190 iommu = amd_iommu_rlookup_table[dev_data->devid];
3194 if (dev_data->domain)
3197 ret = attach_device(dev, domain);
3199 #ifdef CONFIG_IRQ_REMAP
3200 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
3201 if (dom->type == IOMMU_DOMAIN_UNMANAGED)
3202 dev_data->use_vapic = 1;
3204 dev_data->use_vapic = 0;
3208 iommu_completion_wait(iommu);
3213 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
3214 phys_addr_t paddr, size_t page_size, int iommu_prot)
3216 struct protection_domain *domain = to_pdomain(dom);
3220 if (domain->mode == PAGE_MODE_NONE)
3223 if (iommu_prot & IOMMU_READ)
3224 prot |= IOMMU_PROT_IR;
3225 if (iommu_prot & IOMMU_WRITE)
3226 prot |= IOMMU_PROT_IW;
3228 mutex_lock(&domain->api_lock);
3229 ret = iommu_map_page(domain, iova, paddr, page_size, prot, GFP_KERNEL);
3230 mutex_unlock(&domain->api_lock);
3235 static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
3238 struct protection_domain *domain = to_pdomain(dom);
3241 if (domain->mode == PAGE_MODE_NONE)
3244 mutex_lock(&domain->api_lock);
3245 unmap_size = iommu_unmap_page(domain, iova, page_size);
3246 mutex_unlock(&domain->api_lock);
3248 domain_flush_tlb_pde(domain);
3253 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
3256 struct protection_domain *domain = to_pdomain(dom);
3257 unsigned long offset_mask, pte_pgsize;
3260 if (domain->mode == PAGE_MODE_NONE)
3263 pte = fetch_pte(domain, iova, &pte_pgsize);
3265 if (!pte || !IOMMU_PTE_PRESENT(*pte))
3268 offset_mask = pte_pgsize - 1;
3269 __pte = *pte & PM_ADDR_MASK;
3271 return (__pte & ~offset_mask) | (iova & offset_mask);
3274 static bool amd_iommu_capable(enum iommu_cap cap)
3277 case IOMMU_CAP_CACHE_COHERENCY:
3279 case IOMMU_CAP_INTR_REMAP:
3280 return (irq_remapping_enabled == 1);
3281 case IOMMU_CAP_NOEXEC:
3288 static void amd_iommu_get_resv_regions(struct device *dev,
3289 struct list_head *head)
3291 struct iommu_resv_region *region;
3292 struct unity_map_entry *entry;
3295 devid = get_device_id(dev);
3299 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
3303 if (devid < entry->devid_start || devid > entry->devid_end)
3306 length = entry->address_end - entry->address_start;
3307 if (entry->prot & IOMMU_PROT_IR)
3309 if (entry->prot & IOMMU_PROT_IW)
3310 prot |= IOMMU_WRITE;
3312 region = iommu_alloc_resv_region(entry->address_start,
3316 pr_err("Out of memory allocating dm-regions for %s\n",
3320 list_add_tail(®ion->list, head);
3323 region = iommu_alloc_resv_region(MSI_RANGE_START,
3324 MSI_RANGE_END - MSI_RANGE_START + 1,
3328 list_add_tail(®ion->list, head);
3330 region = iommu_alloc_resv_region(HT_RANGE_START,
3331 HT_RANGE_END - HT_RANGE_START + 1,
3332 0, IOMMU_RESV_RESERVED);
3335 list_add_tail(®ion->list, head);
3338 static void amd_iommu_put_resv_regions(struct device *dev,
3339 struct list_head *head)
3341 struct iommu_resv_region *entry, *next;
3343 list_for_each_entry_safe(entry, next, head, list)
3347 static void amd_iommu_apply_resv_region(struct device *dev,
3348 struct iommu_domain *domain,
3349 struct iommu_resv_region *region)
3351 struct dma_ops_domain *dma_dom = to_dma_ops_domain(to_pdomain(domain));
3352 unsigned long start, end;
3354 start = IOVA_PFN(region->start);
3355 end = IOVA_PFN(region->start + region->length);
3357 WARN_ON_ONCE(reserve_iova(&dma_dom->iovad, start, end) == NULL);
3360 const struct iommu_ops amd_iommu_ops = {
3361 .capable = amd_iommu_capable,
3362 .domain_alloc = amd_iommu_domain_alloc,
3363 .domain_free = amd_iommu_domain_free,
3364 .attach_dev = amd_iommu_attach_device,
3365 .detach_dev = amd_iommu_detach_device,
3366 .map = amd_iommu_map,
3367 .unmap = amd_iommu_unmap,
3368 .map_sg = default_iommu_map_sg,
3369 .iova_to_phys = amd_iommu_iova_to_phys,
3370 .add_device = amd_iommu_add_device,
3371 .remove_device = amd_iommu_remove_device,
3372 .device_group = amd_iommu_device_group,
3373 .get_resv_regions = amd_iommu_get_resv_regions,
3374 .put_resv_regions = amd_iommu_put_resv_regions,
3375 .apply_resv_region = amd_iommu_apply_resv_region,
3376 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
3379 /*****************************************************************************
3381 * The next functions do a basic initialization of IOMMU for pass through
3384 * In passthrough mode the IOMMU is initialized and enabled but not used for
3385 * DMA-API translation.
3387 *****************************************************************************/
3389 /* IOMMUv2 specific functions */
3390 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
3392 return atomic_notifier_chain_register(&ppr_notifier, nb);
3394 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
3396 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
3398 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
3400 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
3402 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
3404 struct protection_domain *domain = to_pdomain(dom);
3405 unsigned long flags;
3407 spin_lock_irqsave(&domain->lock, flags);
3409 /* Update data structure */
3410 domain->mode = PAGE_MODE_NONE;
3411 domain->updated = true;
3413 /* Make changes visible to IOMMUs */
3414 update_domain(domain);
3416 /* Page-table is not visible to IOMMU anymore, so free it */
3417 free_pagetable(domain);
3419 spin_unlock_irqrestore(&domain->lock, flags);
3421 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
3423 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
3425 struct protection_domain *domain = to_pdomain(dom);
3426 unsigned long flags;
3429 if (pasids <= 0 || pasids > (PASID_MASK + 1))
3432 /* Number of GCR3 table levels required */
3433 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
3436 if (levels > amd_iommu_max_glx_val)
3439 spin_lock_irqsave(&domain->lock, flags);
3442 * Save us all sanity checks whether devices already in the
3443 * domain support IOMMUv2. Just force that the domain has no
3444 * devices attached when it is switched into IOMMUv2 mode.
3447 if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
3451 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
3452 if (domain->gcr3_tbl == NULL)
3455 domain->glx = levels;
3456 domain->flags |= PD_IOMMUV2_MASK;
3457 domain->updated = true;
3459 update_domain(domain);
3464 spin_unlock_irqrestore(&domain->lock, flags);
3468 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
3470 static int __flush_pasid(struct protection_domain *domain, int pasid,
3471 u64 address, bool size)
3473 struct iommu_dev_data *dev_data;
3474 struct iommu_cmd cmd;
3477 if (!(domain->flags & PD_IOMMUV2_MASK))
3480 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
3483 * IOMMU TLB needs to be flushed before Device TLB to
3484 * prevent device TLB refill from IOMMU TLB
3486 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
3487 if (domain->dev_iommu[i] == 0)
3490 ret = iommu_queue_command(amd_iommus[i], &cmd);
3495 /* Wait until IOMMU TLB flushes are complete */
3496 domain_flush_complete(domain);
3498 /* Now flush device TLBs */
3499 list_for_each_entry(dev_data, &domain->dev_list, list) {
3500 struct amd_iommu *iommu;
3504 There might be non-IOMMUv2 capable devices in an IOMMUv2
3507 if (!dev_data->ats.enabled)
3510 qdep = dev_data->ats.qdep;
3511 iommu = amd_iommu_rlookup_table[dev_data->devid];
3513 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
3514 qdep, address, size);
3516 ret = iommu_queue_command(iommu, &cmd);
3521 /* Wait until all device TLBs are flushed */
3522 domain_flush_complete(domain);
3531 static int __amd_iommu_flush_page(struct protection_domain *domain, int pasid,
3534 return __flush_pasid(domain, pasid, address, false);
3537 int amd_iommu_flush_page(struct iommu_domain *dom, int pasid,
3540 struct protection_domain *domain = to_pdomain(dom);
3541 unsigned long flags;
3544 spin_lock_irqsave(&domain->lock, flags);
3545 ret = __amd_iommu_flush_page(domain, pasid, address);
3546 spin_unlock_irqrestore(&domain->lock, flags);
3550 EXPORT_SYMBOL(amd_iommu_flush_page);
3552 static int __amd_iommu_flush_tlb(struct protection_domain *domain, int pasid)
3554 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
3558 int amd_iommu_flush_tlb(struct iommu_domain *dom, int pasid)
3560 struct protection_domain *domain = to_pdomain(dom);
3561 unsigned long flags;
3564 spin_lock_irqsave(&domain->lock, flags);
3565 ret = __amd_iommu_flush_tlb(domain, pasid);
3566 spin_unlock_irqrestore(&domain->lock, flags);
3570 EXPORT_SYMBOL(amd_iommu_flush_tlb);
3572 static u64 *__get_gcr3_pte(u64 *root, int level, int pasid, bool alloc)
3579 index = (pasid >> (9 * level)) & 0x1ff;
3585 if (!(*pte & GCR3_VALID)) {
3589 root = (void *)get_zeroed_page(GFP_ATOMIC);
3593 *pte = __pa(root) | GCR3_VALID;
3596 root = __va(*pte & PAGE_MASK);
3604 static int __set_gcr3(struct protection_domain *domain, int pasid,
3609 if (domain->mode != PAGE_MODE_NONE)
3612 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3616 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3618 return __amd_iommu_flush_tlb(domain, pasid);
3621 static int __clear_gcr3(struct protection_domain *domain, int pasid)
3625 if (domain->mode != PAGE_MODE_NONE)
3628 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3634 return __amd_iommu_flush_tlb(domain, pasid);
3637 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, int pasid,
3640 struct protection_domain *domain = to_pdomain(dom);
3641 unsigned long flags;
3644 spin_lock_irqsave(&domain->lock, flags);
3645 ret = __set_gcr3(domain, pasid, cr3);
3646 spin_unlock_irqrestore(&domain->lock, flags);
3650 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3652 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, int pasid)
3654 struct protection_domain *domain = to_pdomain(dom);
3655 unsigned long flags;
3658 spin_lock_irqsave(&domain->lock, flags);
3659 ret = __clear_gcr3(domain, pasid);
3660 spin_unlock_irqrestore(&domain->lock, flags);
3664 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3666 int amd_iommu_complete_ppr(struct pci_dev *pdev, int pasid,
3667 int status, int tag)
3669 struct iommu_dev_data *dev_data;
3670 struct amd_iommu *iommu;
3671 struct iommu_cmd cmd;
3673 dev_data = get_dev_data(&pdev->dev);
3674 iommu = amd_iommu_rlookup_table[dev_data->devid];
3676 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3677 tag, dev_data->pri_tlp);
3679 return iommu_queue_command(iommu, &cmd);
3681 EXPORT_SYMBOL(amd_iommu_complete_ppr);
3683 struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3685 struct protection_domain *pdomain;
3687 pdomain = get_domain(&pdev->dev);
3688 if (IS_ERR(pdomain))
3691 /* Only return IOMMUv2 domains */
3692 if (!(pdomain->flags & PD_IOMMUV2_MASK))
3695 return &pdomain->domain;
3697 EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3699 void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3701 struct iommu_dev_data *dev_data;
3703 if (!amd_iommu_v2_supported())
3706 dev_data = get_dev_data(&pdev->dev);
3707 dev_data->errata |= (1 << erratum);
3709 EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3711 int amd_iommu_device_info(struct pci_dev *pdev,
3712 struct amd_iommu_device_info *info)
3717 if (pdev == NULL || info == NULL)
3720 if (!amd_iommu_v2_supported())
3723 memset(info, 0, sizeof(*info));
3725 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS);
3727 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3729 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3731 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3733 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3737 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3738 max_pasids = min(max_pasids, (1 << 20));
3740 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3741 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3743 features = pci_pasid_features(pdev);
3744 if (features & PCI_PASID_CAP_EXEC)
3745 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3746 if (features & PCI_PASID_CAP_PRIV)
3747 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3752 EXPORT_SYMBOL(amd_iommu_device_info);
3754 #ifdef CONFIG_IRQ_REMAP
3756 /*****************************************************************************
3758 * Interrupt Remapping Implementation
3760 *****************************************************************************/
3762 static struct irq_chip amd_ir_chip;
3764 #define DTE_IRQ_PHYS_ADDR_MASK (((1ULL << 45)-1) << 6)
3765 #define DTE_IRQ_REMAP_INTCTL (2ULL << 60)
3766 #define DTE_IRQ_TABLE_LEN (8ULL << 1)
3767 #define DTE_IRQ_REMAP_ENABLE 1ULL
3769 static void set_dte_irq_entry(u16 devid, struct irq_remap_table *table)
3773 dte = amd_iommu_dev_table[devid].data[2];
3774 dte &= ~DTE_IRQ_PHYS_ADDR_MASK;
3775 dte |= virt_to_phys(table->table);
3776 dte |= DTE_IRQ_REMAP_INTCTL;
3777 dte |= DTE_IRQ_TABLE_LEN;
3778 dte |= DTE_IRQ_REMAP_ENABLE;
3780 amd_iommu_dev_table[devid].data[2] = dte;
3783 static struct irq_remap_table *get_irq_table(u16 devid, bool ioapic)
3785 struct irq_remap_table *table = NULL;
3786 struct amd_iommu *iommu;
3787 unsigned long flags;
3790 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3792 iommu = amd_iommu_rlookup_table[devid];
3796 table = irq_lookup_table[devid];
3800 alias = amd_iommu_alias_table[devid];
3801 table = irq_lookup_table[alias];
3803 irq_lookup_table[devid] = table;
3804 set_dte_irq_entry(devid, table);
3805 iommu_flush_dte(iommu, devid);
3809 /* Nothing there yet, allocate new irq remapping table */
3810 table = kzalloc(sizeof(*table), GFP_ATOMIC);
3814 /* Initialize table spin-lock */
3815 spin_lock_init(&table->lock);
3818 /* Keep the first 32 indexes free for IOAPIC interrupts */
3819 table->min_index = 32;
3821 table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_ATOMIC);
3822 if (!table->table) {
3828 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
3829 memset(table->table, 0,
3830 MAX_IRQS_PER_TABLE * sizeof(u32));
3832 memset(table->table, 0,
3833 (MAX_IRQS_PER_TABLE * (sizeof(u64) * 2)));
3838 for (i = 0; i < 32; ++i)
3839 iommu->irte_ops->set_allocated(table, i);
3842 irq_lookup_table[devid] = table;
3843 set_dte_irq_entry(devid, table);
3844 iommu_flush_dte(iommu, devid);
3845 if (devid != alias) {
3846 irq_lookup_table[alias] = table;
3847 set_dte_irq_entry(alias, table);
3848 iommu_flush_dte(iommu, alias);
3852 iommu_completion_wait(iommu);
3855 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3860 static int alloc_irq_index(u16 devid, int count)
3862 struct irq_remap_table *table;
3863 unsigned long flags;
3865 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
3870 table = get_irq_table(devid, false);
3874 spin_lock_irqsave(&table->lock, flags);
3876 /* Scan table for free entries */
3877 for (c = 0, index = table->min_index;
3878 index < MAX_IRQS_PER_TABLE;
3880 if (!iommu->irte_ops->is_allocated(table, index))
3887 iommu->irte_ops->set_allocated(table, index - c + 1);
3897 spin_unlock_irqrestore(&table->lock, flags);
3902 static int modify_irte_ga(u16 devid, int index, struct irte_ga *irte,
3903 struct amd_ir_data *data)
3905 struct irq_remap_table *table;
3906 struct amd_iommu *iommu;
3907 unsigned long flags;
3908 struct irte_ga *entry;
3910 iommu = amd_iommu_rlookup_table[devid];
3914 table = get_irq_table(devid, false);
3918 spin_lock_irqsave(&table->lock, flags);
3920 entry = (struct irte_ga *)table->table;
3921 entry = &entry[index];
3922 entry->lo.fields_remap.valid = 0;
3923 entry->hi.val = irte->hi.val;
3924 entry->lo.val = irte->lo.val;
3925 entry->lo.fields_remap.valid = 1;
3929 spin_unlock_irqrestore(&table->lock, flags);
3931 iommu_flush_irt(iommu, devid);
3932 iommu_completion_wait(iommu);
3937 static int modify_irte(u16 devid, int index, union irte *irte)
3939 struct irq_remap_table *table;
3940 struct amd_iommu *iommu;
3941 unsigned long flags;
3943 iommu = amd_iommu_rlookup_table[devid];
3947 table = get_irq_table(devid, false);
3951 spin_lock_irqsave(&table->lock, flags);
3952 table->table[index] = irte->val;
3953 spin_unlock_irqrestore(&table->lock, flags);
3955 iommu_flush_irt(iommu, devid);
3956 iommu_completion_wait(iommu);
3961 static void free_irte(u16 devid, int index)
3963 struct irq_remap_table *table;
3964 struct amd_iommu *iommu;
3965 unsigned long flags;
3967 iommu = amd_iommu_rlookup_table[devid];
3971 table = get_irq_table(devid, false);
3975 spin_lock_irqsave(&table->lock, flags);
3976 iommu->irte_ops->clear_allocated(table, index);
3977 spin_unlock_irqrestore(&table->lock, flags);
3979 iommu_flush_irt(iommu, devid);
3980 iommu_completion_wait(iommu);
3983 static void irte_prepare(void *entry,
3984 u32 delivery_mode, u32 dest_mode,
3985 u8 vector, u32 dest_apicid, int devid)
3987 union irte *irte = (union irte *) entry;
3990 irte->fields.vector = vector;
3991 irte->fields.int_type = delivery_mode;
3992 irte->fields.destination = dest_apicid;
3993 irte->fields.dm = dest_mode;
3994 irte->fields.valid = 1;
3997 static void irte_ga_prepare(void *entry,
3998 u32 delivery_mode, u32 dest_mode,
3999 u8 vector, u32 dest_apicid, int devid)
4001 struct irte_ga *irte = (struct irte_ga *) entry;
4002 struct iommu_dev_data *dev_data = search_dev_data(devid);
4006 irte->lo.fields_remap.guest_mode = dev_data ? dev_data->use_vapic : 0;
4007 irte->lo.fields_remap.int_type = delivery_mode;
4008 irte->lo.fields_remap.dm = dest_mode;
4009 irte->hi.fields.vector = vector;
4010 irte->lo.fields_remap.destination = dest_apicid;
4011 irte->lo.fields_remap.valid = 1;
4014 static void irte_activate(void *entry, u16 devid, u16 index)
4016 union irte *irte = (union irte *) entry;
4018 irte->fields.valid = 1;
4019 modify_irte(devid, index, irte);
4022 static void irte_ga_activate(void *entry, u16 devid, u16 index)
4024 struct irte_ga *irte = (struct irte_ga *) entry;
4026 irte->lo.fields_remap.valid = 1;
4027 modify_irte_ga(devid, index, irte, NULL);
4030 static void irte_deactivate(void *entry, u16 devid, u16 index)
4032 union irte *irte = (union irte *) entry;
4034 irte->fields.valid = 0;
4035 modify_irte(devid, index, irte);
4038 static void irte_ga_deactivate(void *entry, u16 devid, u16 index)
4040 struct irte_ga *irte = (struct irte_ga *) entry;
4042 irte->lo.fields_remap.valid = 0;
4043 modify_irte_ga(devid, index, irte, NULL);
4046 static void irte_set_affinity(void *entry, u16 devid, u16 index,
4047 u8 vector, u32 dest_apicid)
4049 union irte *irte = (union irte *) entry;
4051 irte->fields.vector = vector;
4052 irte->fields.destination = dest_apicid;
4053 modify_irte(devid, index, irte);
4056 static void irte_ga_set_affinity(void *entry, u16 devid, u16 index,
4057 u8 vector, u32 dest_apicid)
4059 struct irte_ga *irte = (struct irte_ga *) entry;
4060 struct iommu_dev_data *dev_data = search_dev_data(devid);
4062 if (!dev_data || !dev_data->use_vapic) {
4063 irte->hi.fields.vector = vector;
4064 irte->lo.fields_remap.destination = dest_apicid;
4065 irte->lo.fields_remap.guest_mode = 0;
4066 modify_irte_ga(devid, index, irte, NULL);
4070 #define IRTE_ALLOCATED (~1U)
4071 static void irte_set_allocated(struct irq_remap_table *table, int index)
4073 table->table[index] = IRTE_ALLOCATED;
4076 static void irte_ga_set_allocated(struct irq_remap_table *table, int index)
4078 struct irte_ga *ptr = (struct irte_ga *)table->table;
4079 struct irte_ga *irte = &ptr[index];
4081 memset(&irte->lo.val, 0, sizeof(u64));
4082 memset(&irte->hi.val, 0, sizeof(u64));
4083 irte->hi.fields.vector = 0xff;
4086 static bool irte_is_allocated(struct irq_remap_table *table, int index)
4088 union irte *ptr = (union irte *)table->table;
4089 union irte *irte = &ptr[index];
4091 return irte->val != 0;
4094 static bool irte_ga_is_allocated(struct irq_remap_table *table, int index)
4096 struct irte_ga *ptr = (struct irte_ga *)table->table;
4097 struct irte_ga *irte = &ptr[index];
4099 return irte->hi.fields.vector != 0;
4102 static void irte_clear_allocated(struct irq_remap_table *table, int index)
4104 table->table[index] = 0;
4107 static void irte_ga_clear_allocated(struct irq_remap_table *table, int index)
4109 struct irte_ga *ptr = (struct irte_ga *)table->table;
4110 struct irte_ga *irte = &ptr[index];
4112 memset(&irte->lo.val, 0, sizeof(u64));
4113 memset(&irte->hi.val, 0, sizeof(u64));
4116 static int get_devid(struct irq_alloc_info *info)
4120 switch (info->type) {
4121 case X86_IRQ_ALLOC_TYPE_IOAPIC:
4122 devid = get_ioapic_devid(info->ioapic_id);
4124 case X86_IRQ_ALLOC_TYPE_HPET:
4125 devid = get_hpet_devid(info->hpet_id);
4127 case X86_IRQ_ALLOC_TYPE_MSI:
4128 case X86_IRQ_ALLOC_TYPE_MSIX:
4129 devid = get_device_id(&info->msi_dev->dev);
4139 static struct irq_domain *get_ir_irq_domain(struct irq_alloc_info *info)
4141 struct amd_iommu *iommu;
4147 devid = get_devid(info);
4149 iommu = amd_iommu_rlookup_table[devid];
4151 return iommu->ir_domain;
4157 static struct irq_domain *get_irq_domain(struct irq_alloc_info *info)
4159 struct amd_iommu *iommu;
4165 switch (info->type) {
4166 case X86_IRQ_ALLOC_TYPE_MSI:
4167 case X86_IRQ_ALLOC_TYPE_MSIX:
4168 devid = get_device_id(&info->msi_dev->dev);
4172 iommu = amd_iommu_rlookup_table[devid];
4174 return iommu->msi_domain;
4183 struct irq_remap_ops amd_iommu_irq_ops = {
4184 .prepare = amd_iommu_prepare,
4185 .enable = amd_iommu_enable,
4186 .disable = amd_iommu_disable,
4187 .reenable = amd_iommu_reenable,
4188 .enable_faulting = amd_iommu_enable_faulting,
4189 .get_ir_irq_domain = get_ir_irq_domain,
4190 .get_irq_domain = get_irq_domain,
4193 static void irq_remapping_prepare_irte(struct amd_ir_data *data,
4194 struct irq_cfg *irq_cfg,
4195 struct irq_alloc_info *info,
4196 int devid, int index, int sub_handle)
4198 struct irq_2_irte *irte_info = &data->irq_2_irte;
4199 struct msi_msg *msg = &data->msi_entry;
4200 struct IO_APIC_route_entry *entry;
4201 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
4206 data->irq_2_irte.devid = devid;
4207 data->irq_2_irte.index = index + sub_handle;
4208 iommu->irte_ops->prepare(data->entry, apic->irq_delivery_mode,
4209 apic->irq_dest_mode, irq_cfg->vector,
4210 irq_cfg->dest_apicid, devid);
4212 switch (info->type) {
4213 case X86_IRQ_ALLOC_TYPE_IOAPIC:
4214 /* Setup IOAPIC entry */
4215 entry = info->ioapic_entry;
4216 info->ioapic_entry = NULL;
4217 memset(entry, 0, sizeof(*entry));
4218 entry->vector = index;
4220 entry->trigger = info->ioapic_trigger;
4221 entry->polarity = info->ioapic_polarity;
4222 /* Mask level triggered irqs. */
4223 if (info->ioapic_trigger)
4227 case X86_IRQ_ALLOC_TYPE_HPET:
4228 case X86_IRQ_ALLOC_TYPE_MSI:
4229 case X86_IRQ_ALLOC_TYPE_MSIX:
4230 msg->address_hi = MSI_ADDR_BASE_HI;
4231 msg->address_lo = MSI_ADDR_BASE_LO;
4232 msg->data = irte_info->index;
4241 struct amd_irte_ops irte_32_ops = {
4242 .prepare = irte_prepare,
4243 .activate = irte_activate,
4244 .deactivate = irte_deactivate,
4245 .set_affinity = irte_set_affinity,
4246 .set_allocated = irte_set_allocated,
4247 .is_allocated = irte_is_allocated,
4248 .clear_allocated = irte_clear_allocated,
4251 struct amd_irte_ops irte_128_ops = {
4252 .prepare = irte_ga_prepare,
4253 .activate = irte_ga_activate,
4254 .deactivate = irte_ga_deactivate,
4255 .set_affinity = irte_ga_set_affinity,
4256 .set_allocated = irte_ga_set_allocated,
4257 .is_allocated = irte_ga_is_allocated,
4258 .clear_allocated = irte_ga_clear_allocated,
4261 static int irq_remapping_alloc(struct irq_domain *domain, unsigned int virq,
4262 unsigned int nr_irqs, void *arg)
4264 struct irq_alloc_info *info = arg;
4265 struct irq_data *irq_data;
4266 struct amd_ir_data *data = NULL;
4267 struct irq_cfg *cfg;
4273 if (nr_irqs > 1 && info->type != X86_IRQ_ALLOC_TYPE_MSI &&
4274 info->type != X86_IRQ_ALLOC_TYPE_MSIX)
4278 * With IRQ remapping enabled, don't need contiguous CPU vectors
4279 * to support multiple MSI interrupts.
4281 if (info->type == X86_IRQ_ALLOC_TYPE_MSI)
4282 info->flags &= ~X86_IRQ_ALLOC_CONTIGUOUS_VECTORS;
4284 devid = get_devid(info);
4288 ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
4292 if (info->type == X86_IRQ_ALLOC_TYPE_IOAPIC) {
4293 if (get_irq_table(devid, true))
4294 index = info->ioapic_pin;
4298 index = alloc_irq_index(devid, nr_irqs);
4301 pr_warn("Failed to allocate IRTE\n");
4303 goto out_free_parent;
4306 for (i = 0; i < nr_irqs; i++) {
4307 irq_data = irq_domain_get_irq_data(domain, virq + i);
4308 cfg = irqd_cfg(irq_data);
4309 if (!irq_data || !cfg) {
4315 data = kzalloc(sizeof(*data), GFP_KERNEL);
4319 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
4320 data->entry = kzalloc(sizeof(union irte), GFP_KERNEL);
4322 data->entry = kzalloc(sizeof(struct irte_ga),
4329 irq_data->hwirq = (devid << 16) + i;
4330 irq_data->chip_data = data;
4331 irq_data->chip = &amd_ir_chip;
4332 irq_remapping_prepare_irte(data, cfg, info, devid, index, i);
4333 irq_set_status_flags(virq + i, IRQ_MOVE_PCNTXT);
4339 for (i--; i >= 0; i--) {
4340 irq_data = irq_domain_get_irq_data(domain, virq + i);
4342 kfree(irq_data->chip_data);
4344 for (i = 0; i < nr_irqs; i++)
4345 free_irte(devid, index + i);
4347 irq_domain_free_irqs_common(domain, virq, nr_irqs);
4351 static void irq_remapping_free(struct irq_domain *domain, unsigned int virq,
4352 unsigned int nr_irqs)
4354 struct irq_2_irte *irte_info;
4355 struct irq_data *irq_data;
4356 struct amd_ir_data *data;
4359 for (i = 0; i < nr_irqs; i++) {
4360 irq_data = irq_domain_get_irq_data(domain, virq + i);
4361 if (irq_data && irq_data->chip_data) {
4362 data = irq_data->chip_data;
4363 irte_info = &data->irq_2_irte;
4364 free_irte(irte_info->devid, irte_info->index);
4369 irq_domain_free_irqs_common(domain, virq, nr_irqs);
4372 static void irq_remapping_activate(struct irq_domain *domain,
4373 struct irq_data *irq_data)
4375 struct amd_ir_data *data = irq_data->chip_data;
4376 struct irq_2_irte *irte_info = &data->irq_2_irte;
4377 struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4380 iommu->irte_ops->activate(data->entry, irte_info->devid,
4384 static void irq_remapping_deactivate(struct irq_domain *domain,
4385 struct irq_data *irq_data)
4387 struct amd_ir_data *data = irq_data->chip_data;
4388 struct irq_2_irte *irte_info = &data->irq_2_irte;
4389 struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4392 iommu->irte_ops->deactivate(data->entry, irte_info->devid,
4396 static const struct irq_domain_ops amd_ir_domain_ops = {
4397 .alloc = irq_remapping_alloc,
4398 .free = irq_remapping_free,
4399 .activate = irq_remapping_activate,
4400 .deactivate = irq_remapping_deactivate,
4403 static int amd_ir_set_vcpu_affinity(struct irq_data *data, void *vcpu_info)
4405 struct amd_iommu *iommu;
4406 struct amd_iommu_pi_data *pi_data = vcpu_info;
4407 struct vcpu_data *vcpu_pi_info = pi_data->vcpu_data;
4408 struct amd_ir_data *ir_data = data->chip_data;
4409 struct irte_ga *irte = (struct irte_ga *) ir_data->entry;
4410 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
4411 struct iommu_dev_data *dev_data = search_dev_data(irte_info->devid);
4414 * This device has never been set up for guest mode.
4415 * we should not modify the IRTE
4417 if (!dev_data || !dev_data->use_vapic)
4420 pi_data->ir_data = ir_data;
4423 * SVM tries to set up for VAPIC mode, but we are in
4424 * legacy mode. So, we force legacy mode instead.
4426 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
4427 pr_debug("AMD-Vi: %s: Fall back to using intr legacy remap\n",
4429 pi_data->is_guest_mode = false;
4432 iommu = amd_iommu_rlookup_table[irte_info->devid];
4436 pi_data->prev_ga_tag = ir_data->cached_ga_tag;
4437 if (pi_data->is_guest_mode) {
4439 irte->hi.fields.ga_root_ptr = (pi_data->base >> 12);
4440 irte->hi.fields.vector = vcpu_pi_info->vector;
4441 irte->lo.fields_vapic.guest_mode = 1;
4442 irte->lo.fields_vapic.ga_tag = pi_data->ga_tag;
4444 ir_data->cached_ga_tag = pi_data->ga_tag;
4447 struct irq_cfg *cfg = irqd_cfg(data);
4451 irte->hi.fields.vector = cfg->vector;
4452 irte->lo.fields_remap.guest_mode = 0;
4453 irte->lo.fields_remap.destination = cfg->dest_apicid;
4454 irte->lo.fields_remap.int_type = apic->irq_delivery_mode;
4455 irte->lo.fields_remap.dm = apic->irq_dest_mode;
4458 * This communicates the ga_tag back to the caller
4459 * so that it can do all the necessary clean up.
4461 ir_data->cached_ga_tag = 0;
4464 return modify_irte_ga(irte_info->devid, irte_info->index, irte, ir_data);
4467 static int amd_ir_set_affinity(struct irq_data *data,
4468 const struct cpumask *mask, bool force)
4470 struct amd_ir_data *ir_data = data->chip_data;
4471 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
4472 struct irq_cfg *cfg = irqd_cfg(data);
4473 struct irq_data *parent = data->parent_data;
4474 struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4480 ret = parent->chip->irq_set_affinity(parent, mask, force);
4481 if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
4485 * Atomically updates the IRTE with the new destination, vector
4486 * and flushes the interrupt entry cache.
4488 iommu->irte_ops->set_affinity(ir_data->entry, irte_info->devid,
4489 irte_info->index, cfg->vector, cfg->dest_apicid);
4492 * After this point, all the interrupts will start arriving
4493 * at the new destination. So, time to cleanup the previous
4494 * vector allocation.
4496 send_cleanup_vector(cfg);
4498 return IRQ_SET_MASK_OK_DONE;
4501 static void ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg)
4503 struct amd_ir_data *ir_data = irq_data->chip_data;
4505 *msg = ir_data->msi_entry;
4508 static struct irq_chip amd_ir_chip = {
4509 .irq_ack = ir_ack_apic_edge,
4510 .irq_set_affinity = amd_ir_set_affinity,
4511 .irq_set_vcpu_affinity = amd_ir_set_vcpu_affinity,
4512 .irq_compose_msi_msg = ir_compose_msi_msg,
4515 int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
4517 iommu->ir_domain = irq_domain_add_tree(NULL, &amd_ir_domain_ops, iommu);
4518 if (!iommu->ir_domain)
4521 iommu->ir_domain->parent = arch_get_ir_parent_domain();
4522 iommu->msi_domain = arch_create_msi_irq_domain(iommu->ir_domain);
4527 int amd_iommu_update_ga(int cpu, bool is_run, void *data)
4529 unsigned long flags;
4530 struct amd_iommu *iommu;
4531 struct irq_remap_table *irt;
4532 struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
4533 int devid = ir_data->irq_2_irte.devid;
4534 struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
4535 struct irte_ga *ref = (struct irte_ga *) ir_data->ref;
4537 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
4538 !ref || !entry || !entry->lo.fields_vapic.guest_mode)
4541 iommu = amd_iommu_rlookup_table[devid];
4545 irt = get_irq_table(devid, false);
4549 spin_lock_irqsave(&irt->lock, flags);
4551 if (ref->lo.fields_vapic.guest_mode) {
4553 ref->lo.fields_vapic.destination = cpu;
4554 ref->lo.fields_vapic.is_run = is_run;
4558 spin_unlock_irqrestore(&irt->lock, flags);
4560 iommu_flush_irt(iommu, devid);
4561 iommu_completion_wait(iommu);
4564 EXPORT_SYMBOL(amd_iommu_update_ga);