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);
92 #define FLUSH_QUEUE_SIZE 256
94 struct flush_queue_entry {
95 unsigned long iova_pfn;
97 struct dma_ops_domain *dma_dom;
103 struct flush_queue_entry *entries;
106 static DEFINE_PER_CPU(struct flush_queue, flush_queue);
108 static atomic_t queue_timer_on;
109 static struct timer_list queue_timer;
112 * Domain for untranslated devices - only allocated
113 * if iommu=pt passed on kernel cmd line.
115 const struct iommu_ops amd_iommu_ops;
117 static ATOMIC_NOTIFIER_HEAD(ppr_notifier);
118 int amd_iommu_max_glx_val = -1;
120 static const struct dma_map_ops amd_iommu_dma_ops;
123 * This struct contains device specific data for the IOMMU
125 struct iommu_dev_data {
126 struct list_head list; /* For domain->dev_list */
127 struct list_head dev_data_list; /* For global dev_data_list */
128 struct protection_domain *domain; /* Domain the device is bound to */
129 u16 devid; /* PCI Device ID */
130 u16 alias; /* Alias Device ID */
131 bool iommu_v2; /* Device can make use of IOMMUv2 */
132 bool passthrough; /* Device is identity mapped */
136 } ats; /* ATS state */
137 bool pri_tlp; /* PASID TLB required for
139 u32 errata; /* Bitmap for errata to apply */
140 bool use_vapic; /* Enable device to use vapic mode */
144 * general struct to manage commands send to an IOMMU
150 struct kmem_cache *amd_iommu_irq_cache;
152 static void update_domain(struct protection_domain *domain);
153 static int protection_domain_init(struct protection_domain *domain);
154 static void detach_device(struct device *dev);
157 * Data container for a dma_ops specific protection domain
159 struct dma_ops_domain {
160 /* generic protection domain information */
161 struct protection_domain domain;
164 struct iova_domain iovad;
167 static struct iova_domain reserved_iova_ranges;
168 static struct lock_class_key reserved_rbtree_key;
170 /****************************************************************************
174 ****************************************************************************/
176 static inline int match_hid_uid(struct device *dev,
177 struct acpihid_map_entry *entry)
179 const char *hid, *uid;
181 hid = acpi_device_hid(ACPI_COMPANION(dev));
182 uid = acpi_device_uid(ACPI_COMPANION(dev));
188 return strcmp(hid, entry->hid);
191 return strcmp(hid, entry->hid);
193 return (strcmp(hid, entry->hid) || strcmp(uid, entry->uid));
196 static inline u16 get_pci_device_id(struct device *dev)
198 struct pci_dev *pdev = to_pci_dev(dev);
200 return PCI_DEVID(pdev->bus->number, pdev->devfn);
203 static inline int get_acpihid_device_id(struct device *dev,
204 struct acpihid_map_entry **entry)
206 struct acpihid_map_entry *p;
208 list_for_each_entry(p, &acpihid_map, list) {
209 if (!match_hid_uid(dev, p)) {
218 static inline int get_device_id(struct device *dev)
223 devid = get_pci_device_id(dev);
225 devid = get_acpihid_device_id(dev, NULL);
230 static struct protection_domain *to_pdomain(struct iommu_domain *dom)
232 return container_of(dom, struct protection_domain, domain);
235 static struct dma_ops_domain* to_dma_ops_domain(struct protection_domain *domain)
237 BUG_ON(domain->flags != PD_DMA_OPS_MASK);
238 return container_of(domain, struct dma_ops_domain, domain);
241 static struct iommu_dev_data *alloc_dev_data(u16 devid)
243 struct iommu_dev_data *dev_data;
246 dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
250 dev_data->devid = devid;
252 spin_lock_irqsave(&dev_data_list_lock, flags);
253 list_add_tail(&dev_data->dev_data_list, &dev_data_list);
254 spin_unlock_irqrestore(&dev_data_list_lock, flags);
259 static struct iommu_dev_data *search_dev_data(u16 devid)
261 struct iommu_dev_data *dev_data;
264 spin_lock_irqsave(&dev_data_list_lock, flags);
265 list_for_each_entry(dev_data, &dev_data_list, dev_data_list) {
266 if (dev_data->devid == devid)
273 spin_unlock_irqrestore(&dev_data_list_lock, flags);
278 static int __last_alias(struct pci_dev *pdev, u16 alias, void *data)
280 *(u16 *)data = alias;
284 static u16 get_alias(struct device *dev)
286 struct pci_dev *pdev = to_pci_dev(dev);
287 u16 devid, ivrs_alias, pci_alias;
289 /* The callers make sure that get_device_id() does not fail here */
290 devid = get_device_id(dev);
291 ivrs_alias = amd_iommu_alias_table[devid];
292 pci_for_each_dma_alias(pdev, __last_alias, &pci_alias);
294 if (ivrs_alias == pci_alias)
300 * The IVRS is fairly reliable in telling us about aliases, but it
301 * can't know about every screwy device. If we don't have an IVRS
302 * reported alias, use the PCI reported alias. In that case we may
303 * still need to initialize the rlookup and dev_table entries if the
304 * alias is to a non-existent device.
306 if (ivrs_alias == devid) {
307 if (!amd_iommu_rlookup_table[pci_alias]) {
308 amd_iommu_rlookup_table[pci_alias] =
309 amd_iommu_rlookup_table[devid];
310 memcpy(amd_iommu_dev_table[pci_alias].data,
311 amd_iommu_dev_table[devid].data,
312 sizeof(amd_iommu_dev_table[pci_alias].data));
318 pr_info("AMD-Vi: Using IVRS reported alias %02x:%02x.%d "
319 "for device %s[%04x:%04x], kernel reported alias "
320 "%02x:%02x.%d\n", PCI_BUS_NUM(ivrs_alias), PCI_SLOT(ivrs_alias),
321 PCI_FUNC(ivrs_alias), dev_name(dev), pdev->vendor, pdev->device,
322 PCI_BUS_NUM(pci_alias), PCI_SLOT(pci_alias),
323 PCI_FUNC(pci_alias));
326 * If we don't have a PCI DMA alias and the IVRS alias is on the same
327 * bus, then the IVRS table may know about a quirk that we don't.
329 if (pci_alias == devid &&
330 PCI_BUS_NUM(ivrs_alias) == pdev->bus->number) {
331 pci_add_dma_alias(pdev, ivrs_alias & 0xff);
332 pr_info("AMD-Vi: Added PCI DMA alias %02x.%d for %s\n",
333 PCI_SLOT(ivrs_alias), PCI_FUNC(ivrs_alias),
340 static struct iommu_dev_data *find_dev_data(u16 devid)
342 struct iommu_dev_data *dev_data;
344 dev_data = search_dev_data(devid);
346 if (dev_data == NULL)
347 dev_data = alloc_dev_data(devid);
352 static struct iommu_dev_data *get_dev_data(struct device *dev)
354 return dev->archdata.iommu;
358 * Find or create an IOMMU group for a acpihid device.
360 static struct iommu_group *acpihid_device_group(struct device *dev)
362 struct acpihid_map_entry *p, *entry = NULL;
365 devid = get_acpihid_device_id(dev, &entry);
367 return ERR_PTR(devid);
369 list_for_each_entry(p, &acpihid_map, list) {
370 if ((devid == p->devid) && p->group)
371 entry->group = p->group;
375 entry->group = generic_device_group(dev);
377 iommu_group_ref_get(entry->group);
382 static bool pci_iommuv2_capable(struct pci_dev *pdev)
384 static const int caps[] = {
387 PCI_EXT_CAP_ID_PASID,
391 for (i = 0; i < 3; ++i) {
392 pos = pci_find_ext_capability(pdev, caps[i]);
400 static bool pdev_pri_erratum(struct pci_dev *pdev, u32 erratum)
402 struct iommu_dev_data *dev_data;
404 dev_data = get_dev_data(&pdev->dev);
406 return dev_data->errata & (1 << erratum) ? true : false;
410 * This function checks if the driver got a valid device from the caller to
411 * avoid dereferencing invalid pointers.
413 static bool check_device(struct device *dev)
417 if (!dev || !dev->dma_mask)
420 devid = get_device_id(dev);
424 /* Out of our scope? */
425 if (devid > amd_iommu_last_bdf)
428 if (amd_iommu_rlookup_table[devid] == NULL)
434 static void init_iommu_group(struct device *dev)
436 struct iommu_group *group;
438 group = iommu_group_get_for_dev(dev);
442 iommu_group_put(group);
445 static int iommu_init_device(struct device *dev)
447 struct iommu_dev_data *dev_data;
448 struct amd_iommu *iommu;
451 if (dev->archdata.iommu)
454 devid = get_device_id(dev);
458 iommu = amd_iommu_rlookup_table[devid];
460 dev_data = find_dev_data(devid);
464 dev_data->alias = get_alias(dev);
466 if (dev_is_pci(dev) && pci_iommuv2_capable(to_pci_dev(dev))) {
467 struct amd_iommu *iommu;
469 iommu = amd_iommu_rlookup_table[dev_data->devid];
470 dev_data->iommu_v2 = iommu->is_iommu_v2;
473 dev->archdata.iommu = dev_data;
475 iommu_device_link(&iommu->iommu, dev);
480 static void iommu_ignore_device(struct device *dev)
485 devid = get_device_id(dev);
489 alias = get_alias(dev);
491 memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
492 memset(&amd_iommu_dev_table[alias], 0, sizeof(struct dev_table_entry));
494 amd_iommu_rlookup_table[devid] = NULL;
495 amd_iommu_rlookup_table[alias] = NULL;
498 static void iommu_uninit_device(struct device *dev)
500 struct iommu_dev_data *dev_data;
501 struct amd_iommu *iommu;
504 devid = get_device_id(dev);
508 iommu = amd_iommu_rlookup_table[devid];
510 dev_data = search_dev_data(devid);
514 if (dev_data->domain)
517 iommu_device_unlink(&iommu->iommu, dev);
519 iommu_group_remove_device(dev);
525 * We keep dev_data around for unplugged devices and reuse it when the
526 * device is re-plugged - not doing so would introduce a ton of races.
530 /****************************************************************************
532 * Interrupt handling functions
534 ****************************************************************************/
536 static void dump_dte_entry(u16 devid)
540 for (i = 0; i < 4; ++i)
541 pr_err("AMD-Vi: DTE[%d]: %016llx\n", i,
542 amd_iommu_dev_table[devid].data[i]);
545 static void dump_command(unsigned long phys_addr)
547 struct iommu_cmd *cmd = phys_to_virt(phys_addr);
550 for (i = 0; i < 4; ++i)
551 pr_err("AMD-Vi: CMD[%d]: %08x\n", i, cmd->data[i]);
554 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
556 int type, devid, domid, flags;
557 volatile u32 *event = __evt;
562 type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
563 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
564 domid = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
565 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
566 address = (u64)(((u64)event[3]) << 32) | event[2];
569 /* Did we hit the erratum? */
570 if (++count == LOOP_TIMEOUT) {
571 pr_err("AMD-Vi: No event written to event log\n");
578 printk(KERN_ERR "AMD-Vi: Event logged [");
581 case EVENT_TYPE_ILL_DEV:
582 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
583 "address=0x%016llx flags=0x%04x]\n",
584 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
586 dump_dte_entry(devid);
588 case EVENT_TYPE_IO_FAULT:
589 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
590 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
591 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
592 domid, address, flags);
594 case EVENT_TYPE_DEV_TAB_ERR:
595 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
596 "address=0x%016llx flags=0x%04x]\n",
597 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
600 case EVENT_TYPE_PAGE_TAB_ERR:
601 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
602 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
603 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
604 domid, address, flags);
606 case EVENT_TYPE_ILL_CMD:
607 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
608 dump_command(address);
610 case EVENT_TYPE_CMD_HARD_ERR:
611 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
612 "flags=0x%04x]\n", address, flags);
614 case EVENT_TYPE_IOTLB_INV_TO:
615 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
616 "address=0x%016llx]\n",
617 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
620 case EVENT_TYPE_INV_DEV_REQ:
621 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
622 "address=0x%016llx flags=0x%04x]\n",
623 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
627 printk(KERN_ERR "UNKNOWN type=0x%02x]\n", type);
630 memset(__evt, 0, 4 * sizeof(u32));
633 static void iommu_poll_events(struct amd_iommu *iommu)
637 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
638 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
640 while (head != tail) {
641 iommu_print_event(iommu, iommu->evt_buf + head);
642 head = (head + EVENT_ENTRY_SIZE) % EVT_BUFFER_SIZE;
645 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
648 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
650 struct amd_iommu_fault fault;
652 if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
653 pr_err_ratelimited("AMD-Vi: Unknown PPR request received\n");
657 fault.address = raw[1];
658 fault.pasid = PPR_PASID(raw[0]);
659 fault.device_id = PPR_DEVID(raw[0]);
660 fault.tag = PPR_TAG(raw[0]);
661 fault.flags = PPR_FLAGS(raw[0]);
663 atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
666 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
670 if (iommu->ppr_log == NULL)
673 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
674 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
676 while (head != tail) {
681 raw = (u64 *)(iommu->ppr_log + head);
684 * Hardware bug: Interrupt may arrive before the entry is
685 * written to memory. If this happens we need to wait for the
688 for (i = 0; i < LOOP_TIMEOUT; ++i) {
689 if (PPR_REQ_TYPE(raw[0]) != 0)
694 /* Avoid memcpy function-call overhead */
699 * To detect the hardware bug we need to clear the entry
702 raw[0] = raw[1] = 0UL;
704 /* Update head pointer of hardware ring-buffer */
705 head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
706 writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
708 /* Handle PPR entry */
709 iommu_handle_ppr_entry(iommu, entry);
711 /* Refresh ring-buffer information */
712 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
713 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
717 #ifdef CONFIG_IRQ_REMAP
718 static int (*iommu_ga_log_notifier)(u32);
720 int amd_iommu_register_ga_log_notifier(int (*notifier)(u32))
722 iommu_ga_log_notifier = notifier;
726 EXPORT_SYMBOL(amd_iommu_register_ga_log_notifier);
728 static void iommu_poll_ga_log(struct amd_iommu *iommu)
730 u32 head, tail, cnt = 0;
732 if (iommu->ga_log == NULL)
735 head = readl(iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
736 tail = readl(iommu->mmio_base + MMIO_GA_TAIL_OFFSET);
738 while (head != tail) {
742 raw = (u64 *)(iommu->ga_log + head);
745 /* Avoid memcpy function-call overhead */
748 /* Update head pointer of hardware ring-buffer */
749 head = (head + GA_ENTRY_SIZE) % GA_LOG_SIZE;
750 writel(head, iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
752 /* Handle GA entry */
753 switch (GA_REQ_TYPE(log_entry)) {
755 if (!iommu_ga_log_notifier)
758 pr_debug("AMD-Vi: %s: devid=%#x, ga_tag=%#x\n",
759 __func__, GA_DEVID(log_entry),
762 if (iommu_ga_log_notifier(GA_TAG(log_entry)) != 0)
763 pr_err("AMD-Vi: GA log notifier failed.\n");
770 #endif /* CONFIG_IRQ_REMAP */
772 #define AMD_IOMMU_INT_MASK \
773 (MMIO_STATUS_EVT_INT_MASK | \
774 MMIO_STATUS_PPR_INT_MASK | \
775 MMIO_STATUS_GALOG_INT_MASK)
777 irqreturn_t amd_iommu_int_thread(int irq, void *data)
779 struct amd_iommu *iommu = (struct amd_iommu *) data;
780 u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
782 while (status & AMD_IOMMU_INT_MASK) {
783 /* Enable EVT and PPR and GA interrupts again */
784 writel(AMD_IOMMU_INT_MASK,
785 iommu->mmio_base + MMIO_STATUS_OFFSET);
787 if (status & MMIO_STATUS_EVT_INT_MASK) {
788 pr_devel("AMD-Vi: Processing IOMMU Event Log\n");
789 iommu_poll_events(iommu);
792 if (status & MMIO_STATUS_PPR_INT_MASK) {
793 pr_devel("AMD-Vi: Processing IOMMU PPR Log\n");
794 iommu_poll_ppr_log(iommu);
797 #ifdef CONFIG_IRQ_REMAP
798 if (status & MMIO_STATUS_GALOG_INT_MASK) {
799 pr_devel("AMD-Vi: Processing IOMMU GA Log\n");
800 iommu_poll_ga_log(iommu);
805 * Hardware bug: ERBT1312
806 * When re-enabling interrupt (by writing 1
807 * to clear the bit), the hardware might also try to set
808 * the interrupt bit in the event status register.
809 * In this scenario, the bit will be set, and disable
810 * subsequent interrupts.
812 * Workaround: The IOMMU driver should read back the
813 * status register and check if the interrupt bits are cleared.
814 * If not, driver will need to go through the interrupt handler
815 * again and re-clear the bits
817 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
822 irqreturn_t amd_iommu_int_handler(int irq, void *data)
824 return IRQ_WAKE_THREAD;
827 /****************************************************************************
829 * IOMMU command queuing functions
831 ****************************************************************************/
833 static int wait_on_sem(volatile u64 *sem)
837 while (*sem == 0 && i < LOOP_TIMEOUT) {
842 if (i == LOOP_TIMEOUT) {
843 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
850 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
851 struct iommu_cmd *cmd,
856 target = iommu->cmd_buf + tail;
857 tail = (tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
859 /* Copy command to buffer */
860 memcpy(target, cmd, sizeof(*cmd));
862 /* Tell the IOMMU about it */
863 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
866 static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
868 WARN_ON(address & 0x7ULL);
870 memset(cmd, 0, sizeof(*cmd));
871 cmd->data[0] = lower_32_bits(__pa(address)) | CMD_COMPL_WAIT_STORE_MASK;
872 cmd->data[1] = upper_32_bits(__pa(address));
874 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
877 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
879 memset(cmd, 0, sizeof(*cmd));
880 cmd->data[0] = devid;
881 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
884 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
885 size_t size, u16 domid, int pde)
890 pages = iommu_num_pages(address, size, PAGE_SIZE);
895 * If we have to flush more than one page, flush all
896 * TLB entries for this domain
898 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
902 address &= PAGE_MASK;
904 memset(cmd, 0, sizeof(*cmd));
905 cmd->data[1] |= domid;
906 cmd->data[2] = lower_32_bits(address);
907 cmd->data[3] = upper_32_bits(address);
908 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
909 if (s) /* size bit - we flush more than one 4kb page */
910 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
911 if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
912 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
915 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
916 u64 address, size_t size)
921 pages = iommu_num_pages(address, size, PAGE_SIZE);
926 * If we have to flush more than one page, flush all
927 * TLB entries for this domain
929 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
933 address &= PAGE_MASK;
935 memset(cmd, 0, sizeof(*cmd));
936 cmd->data[0] = devid;
937 cmd->data[0] |= (qdep & 0xff) << 24;
938 cmd->data[1] = devid;
939 cmd->data[2] = lower_32_bits(address);
940 cmd->data[3] = upper_32_bits(address);
941 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
943 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
946 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, int pasid,
947 u64 address, bool size)
949 memset(cmd, 0, sizeof(*cmd));
951 address &= ~(0xfffULL);
953 cmd->data[0] = pasid;
954 cmd->data[1] = domid;
955 cmd->data[2] = lower_32_bits(address);
956 cmd->data[3] = upper_32_bits(address);
957 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
958 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
960 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
961 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
964 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, int pasid,
965 int qdep, u64 address, bool size)
967 memset(cmd, 0, sizeof(*cmd));
969 address &= ~(0xfffULL);
971 cmd->data[0] = devid;
972 cmd->data[0] |= ((pasid >> 8) & 0xff) << 16;
973 cmd->data[0] |= (qdep & 0xff) << 24;
974 cmd->data[1] = devid;
975 cmd->data[1] |= (pasid & 0xff) << 16;
976 cmd->data[2] = lower_32_bits(address);
977 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
978 cmd->data[3] = upper_32_bits(address);
980 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
981 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
984 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, int pasid,
985 int status, int tag, bool gn)
987 memset(cmd, 0, sizeof(*cmd));
989 cmd->data[0] = devid;
991 cmd->data[1] = pasid;
992 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
994 cmd->data[3] = tag & 0x1ff;
995 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
997 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
1000 static void build_inv_all(struct iommu_cmd *cmd)
1002 memset(cmd, 0, sizeof(*cmd));
1003 CMD_SET_TYPE(cmd, CMD_INV_ALL);
1006 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
1008 memset(cmd, 0, sizeof(*cmd));
1009 cmd->data[0] = devid;
1010 CMD_SET_TYPE(cmd, CMD_INV_IRT);
1014 * Writes the command to the IOMMUs command buffer and informs the
1015 * hardware about the new command.
1017 static int __iommu_queue_command_sync(struct amd_iommu *iommu,
1018 struct iommu_cmd *cmd,
1021 u32 left, tail, head, next_tail;
1025 head = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
1026 tail = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
1027 next_tail = (tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
1028 left = (head - next_tail) % CMD_BUFFER_SIZE;
1031 struct iommu_cmd sync_cmd;
1036 build_completion_wait(&sync_cmd, (u64)&iommu->cmd_sem);
1037 copy_cmd_to_buffer(iommu, &sync_cmd, tail);
1039 if ((ret = wait_on_sem(&iommu->cmd_sem)) != 0)
1045 copy_cmd_to_buffer(iommu, cmd, tail);
1047 /* We need to sync now to make sure all commands are processed */
1048 iommu->need_sync = sync;
1053 static int iommu_queue_command_sync(struct amd_iommu *iommu,
1054 struct iommu_cmd *cmd,
1057 unsigned long flags;
1060 spin_lock_irqsave(&iommu->lock, flags);
1061 ret = __iommu_queue_command_sync(iommu, cmd, sync);
1062 spin_unlock_irqrestore(&iommu->lock, flags);
1067 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
1069 return iommu_queue_command_sync(iommu, cmd, true);
1073 * This function queues a completion wait command into the command
1074 * buffer of an IOMMU
1076 static int iommu_completion_wait(struct amd_iommu *iommu)
1078 struct iommu_cmd cmd;
1079 unsigned long flags;
1082 if (!iommu->need_sync)
1086 build_completion_wait(&cmd, (u64)&iommu->cmd_sem);
1088 spin_lock_irqsave(&iommu->lock, flags);
1092 ret = __iommu_queue_command_sync(iommu, &cmd, false);
1096 ret = wait_on_sem(&iommu->cmd_sem);
1099 spin_unlock_irqrestore(&iommu->lock, flags);
1104 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
1106 struct iommu_cmd cmd;
1108 build_inv_dte(&cmd, devid);
1110 return iommu_queue_command(iommu, &cmd);
1113 static void iommu_flush_dte_all(struct amd_iommu *iommu)
1117 for (devid = 0; devid <= 0xffff; ++devid)
1118 iommu_flush_dte(iommu, devid);
1120 iommu_completion_wait(iommu);
1124 * This function uses heavy locking and may disable irqs for some time. But
1125 * this is no issue because it is only called during resume.
1127 static void iommu_flush_tlb_all(struct amd_iommu *iommu)
1131 for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
1132 struct iommu_cmd cmd;
1133 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1135 iommu_queue_command(iommu, &cmd);
1138 iommu_completion_wait(iommu);
1141 static void iommu_flush_all(struct amd_iommu *iommu)
1143 struct iommu_cmd cmd;
1145 build_inv_all(&cmd);
1147 iommu_queue_command(iommu, &cmd);
1148 iommu_completion_wait(iommu);
1151 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1153 struct iommu_cmd cmd;
1155 build_inv_irt(&cmd, devid);
1157 iommu_queue_command(iommu, &cmd);
1160 static void iommu_flush_irt_all(struct amd_iommu *iommu)
1164 for (devid = 0; devid <= MAX_DEV_TABLE_ENTRIES; devid++)
1165 iommu_flush_irt(iommu, devid);
1167 iommu_completion_wait(iommu);
1170 void iommu_flush_all_caches(struct amd_iommu *iommu)
1172 if (iommu_feature(iommu, FEATURE_IA)) {
1173 iommu_flush_all(iommu);
1175 iommu_flush_dte_all(iommu);
1176 iommu_flush_irt_all(iommu);
1177 iommu_flush_tlb_all(iommu);
1182 * Command send function for flushing on-device TLB
1184 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1185 u64 address, size_t size)
1187 struct amd_iommu *iommu;
1188 struct iommu_cmd cmd;
1191 qdep = dev_data->ats.qdep;
1192 iommu = amd_iommu_rlookup_table[dev_data->devid];
1194 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1196 return iommu_queue_command(iommu, &cmd);
1200 * Command send function for invalidating a device table entry
1202 static int device_flush_dte(struct iommu_dev_data *dev_data)
1204 struct amd_iommu *iommu;
1208 iommu = amd_iommu_rlookup_table[dev_data->devid];
1209 alias = dev_data->alias;
1211 ret = iommu_flush_dte(iommu, dev_data->devid);
1212 if (!ret && alias != dev_data->devid)
1213 ret = iommu_flush_dte(iommu, alias);
1217 if (dev_data->ats.enabled)
1218 ret = device_flush_iotlb(dev_data, 0, ~0UL);
1224 * TLB invalidation function which is called from the mapping functions.
1225 * It invalidates a single PTE if the range to flush is within a single
1226 * page. Otherwise it flushes the whole TLB of the IOMMU.
1228 static void __domain_flush_pages(struct protection_domain *domain,
1229 u64 address, size_t size, int pde)
1231 struct iommu_dev_data *dev_data;
1232 struct iommu_cmd cmd;
1235 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1237 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
1238 if (!domain->dev_iommu[i])
1242 * Devices of this domain are behind this IOMMU
1243 * We need a TLB flush
1245 ret |= iommu_queue_command(amd_iommus[i], &cmd);
1248 list_for_each_entry(dev_data, &domain->dev_list, list) {
1250 if (!dev_data->ats.enabled)
1253 ret |= device_flush_iotlb(dev_data, address, size);
1259 static void domain_flush_pages(struct protection_domain *domain,
1260 u64 address, size_t size)
1262 __domain_flush_pages(domain, address, size, 0);
1265 /* Flush the whole IO/TLB for a given protection domain */
1266 static void domain_flush_tlb(struct protection_domain *domain)
1268 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
1271 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1272 static void domain_flush_tlb_pde(struct protection_domain *domain)
1274 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1277 static void domain_flush_complete(struct protection_domain *domain)
1281 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
1282 if (domain && !domain->dev_iommu[i])
1286 * Devices of this domain are behind this IOMMU
1287 * We need to wait for completion of all commands.
1289 iommu_completion_wait(amd_iommus[i]);
1295 * This function flushes the DTEs for all devices in domain
1297 static void domain_flush_devices(struct protection_domain *domain)
1299 struct iommu_dev_data *dev_data;
1301 list_for_each_entry(dev_data, &domain->dev_list, list)
1302 device_flush_dte(dev_data);
1305 /****************************************************************************
1307 * The functions below are used the create the page table mappings for
1308 * unity mapped regions.
1310 ****************************************************************************/
1313 * This function is used to add another level to an IO page table. Adding
1314 * another level increases the size of the address space by 9 bits to a size up
1317 static bool increase_address_space(struct protection_domain *domain,
1322 if (domain->mode == PAGE_MODE_6_LEVEL)
1323 /* address space already 64 bit large */
1326 pte = (void *)get_zeroed_page(gfp);
1330 *pte = PM_LEVEL_PDE(domain->mode,
1331 virt_to_phys(domain->pt_root));
1332 domain->pt_root = pte;
1334 domain->updated = true;
1339 static u64 *alloc_pte(struct protection_domain *domain,
1340 unsigned long address,
1341 unsigned long page_size,
1348 BUG_ON(!is_power_of_2(page_size));
1350 while (address > PM_LEVEL_SIZE(domain->mode))
1351 increase_address_space(domain, gfp);
1353 level = domain->mode - 1;
1354 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1355 address = PAGE_SIZE_ALIGN(address, page_size);
1356 end_lvl = PAGE_SIZE_LEVEL(page_size);
1358 while (level > end_lvl) {
1363 if (!IOMMU_PTE_PRESENT(__pte)) {
1364 page = (u64 *)get_zeroed_page(gfp);
1368 __npte = PM_LEVEL_PDE(level, virt_to_phys(page));
1370 /* pte could have been changed somewhere. */
1371 if (cmpxchg64(pte, __pte, __npte) != __pte) {
1372 free_page((unsigned long)page);
1377 /* No level skipping support yet */
1378 if (PM_PTE_LEVEL(*pte) != level)
1383 pte = IOMMU_PTE_PAGE(*pte);
1385 if (pte_page && level == end_lvl)
1388 pte = &pte[PM_LEVEL_INDEX(level, address)];
1395 * This function checks if there is a PTE for a given dma address. If
1396 * there is one, it returns the pointer to it.
1398 static u64 *fetch_pte(struct protection_domain *domain,
1399 unsigned long address,
1400 unsigned long *page_size)
1405 if (address > PM_LEVEL_SIZE(domain->mode))
1408 level = domain->mode - 1;
1409 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1410 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1415 if (!IOMMU_PTE_PRESENT(*pte))
1419 if (PM_PTE_LEVEL(*pte) == 7 ||
1420 PM_PTE_LEVEL(*pte) == 0)
1423 /* No level skipping support yet */
1424 if (PM_PTE_LEVEL(*pte) != level)
1429 /* Walk to the next level */
1430 pte = IOMMU_PTE_PAGE(*pte);
1431 pte = &pte[PM_LEVEL_INDEX(level, address)];
1432 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1435 if (PM_PTE_LEVEL(*pte) == 0x07) {
1436 unsigned long pte_mask;
1439 * If we have a series of large PTEs, make
1440 * sure to return a pointer to the first one.
1442 *page_size = pte_mask = PTE_PAGE_SIZE(*pte);
1443 pte_mask = ~((PAGE_SIZE_PTE_COUNT(pte_mask) << 3) - 1);
1444 pte = (u64 *)(((unsigned long)pte) & pte_mask);
1451 * Generic mapping functions. It maps a physical address into a DMA
1452 * address space. It allocates the page table pages if necessary.
1453 * In the future it can be extended to a generic mapping function
1454 * supporting all features of AMD IOMMU page tables like level skipping
1455 * and full 64 bit address spaces.
1457 static int iommu_map_page(struct protection_domain *dom,
1458 unsigned long bus_addr,
1459 unsigned long phys_addr,
1460 unsigned long page_size,
1467 BUG_ON(!IS_ALIGNED(bus_addr, page_size));
1468 BUG_ON(!IS_ALIGNED(phys_addr, page_size));
1470 if (!(prot & IOMMU_PROT_MASK))
1473 count = PAGE_SIZE_PTE_COUNT(page_size);
1474 pte = alloc_pte(dom, bus_addr, page_size, NULL, gfp);
1479 for (i = 0; i < count; ++i)
1480 if (IOMMU_PTE_PRESENT(pte[i]))
1484 __pte = PAGE_SIZE_PTE(phys_addr, page_size);
1485 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_P | IOMMU_PTE_FC;
1487 __pte = phys_addr | IOMMU_PTE_P | IOMMU_PTE_FC;
1489 if (prot & IOMMU_PROT_IR)
1490 __pte |= IOMMU_PTE_IR;
1491 if (prot & IOMMU_PROT_IW)
1492 __pte |= IOMMU_PTE_IW;
1494 for (i = 0; i < count; ++i)
1502 static unsigned long iommu_unmap_page(struct protection_domain *dom,
1503 unsigned long bus_addr,
1504 unsigned long page_size)
1506 unsigned long long unmapped;
1507 unsigned long unmap_size;
1510 BUG_ON(!is_power_of_2(page_size));
1514 while (unmapped < page_size) {
1516 pte = fetch_pte(dom, bus_addr, &unmap_size);
1521 count = PAGE_SIZE_PTE_COUNT(unmap_size);
1522 for (i = 0; i < count; i++)
1526 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1527 unmapped += unmap_size;
1530 BUG_ON(unmapped && !is_power_of_2(unmapped));
1535 /****************************************************************************
1537 * The next functions belong to the address allocator for the dma_ops
1538 * interface functions.
1540 ****************************************************************************/
1543 static unsigned long dma_ops_alloc_iova(struct device *dev,
1544 struct dma_ops_domain *dma_dom,
1545 unsigned int pages, u64 dma_mask)
1547 unsigned long pfn = 0;
1549 pages = __roundup_pow_of_two(pages);
1551 if (dma_mask > DMA_BIT_MASK(32))
1552 pfn = alloc_iova_fast(&dma_dom->iovad, pages,
1553 IOVA_PFN(DMA_BIT_MASK(32)));
1556 pfn = alloc_iova_fast(&dma_dom->iovad, pages, IOVA_PFN(dma_mask));
1558 return (pfn << PAGE_SHIFT);
1561 static void dma_ops_free_iova(struct dma_ops_domain *dma_dom,
1562 unsigned long address,
1565 pages = __roundup_pow_of_two(pages);
1566 address >>= PAGE_SHIFT;
1568 free_iova_fast(&dma_dom->iovad, address, pages);
1571 /****************************************************************************
1573 * The next functions belong to the domain allocation. A domain is
1574 * allocated for every IOMMU as the default domain. If device isolation
1575 * is enabled, every device get its own domain. The most important thing
1576 * about domains is the page table mapping the DMA address space they
1579 ****************************************************************************/
1582 * This function adds a protection domain to the global protection domain list
1584 static void add_domain_to_list(struct protection_domain *domain)
1586 unsigned long flags;
1588 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1589 list_add(&domain->list, &amd_iommu_pd_list);
1590 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1594 * This function removes a protection domain to the global
1595 * protection domain list
1597 static void del_domain_from_list(struct protection_domain *domain)
1599 unsigned long flags;
1601 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1602 list_del(&domain->list);
1603 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1606 static u16 domain_id_alloc(void)
1608 unsigned long flags;
1611 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1612 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1614 if (id > 0 && id < MAX_DOMAIN_ID)
1615 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1618 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1623 static void domain_id_free(int id)
1625 unsigned long flags;
1627 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1628 if (id > 0 && id < MAX_DOMAIN_ID)
1629 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1630 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1633 #define DEFINE_FREE_PT_FN(LVL, FN) \
1634 static void free_pt_##LVL (unsigned long __pt) \
1642 for (i = 0; i < 512; ++i) { \
1643 /* PTE present? */ \
1644 if (!IOMMU_PTE_PRESENT(pt[i])) \
1648 if (PM_PTE_LEVEL(pt[i]) == 0 || \
1649 PM_PTE_LEVEL(pt[i]) == 7) \
1652 p = (unsigned long)IOMMU_PTE_PAGE(pt[i]); \
1655 free_page((unsigned long)pt); \
1658 DEFINE_FREE_PT_FN(l2, free_page)
1659 DEFINE_FREE_PT_FN(l3, free_pt_l2)
1660 DEFINE_FREE_PT_FN(l4, free_pt_l3)
1661 DEFINE_FREE_PT_FN(l5, free_pt_l4)
1662 DEFINE_FREE_PT_FN(l6, free_pt_l5)
1664 static void free_pagetable(struct protection_domain *domain)
1666 unsigned long root = (unsigned long)domain->pt_root;
1668 switch (domain->mode) {
1669 case PAGE_MODE_NONE:
1671 case PAGE_MODE_1_LEVEL:
1674 case PAGE_MODE_2_LEVEL:
1677 case PAGE_MODE_3_LEVEL:
1680 case PAGE_MODE_4_LEVEL:
1683 case PAGE_MODE_5_LEVEL:
1686 case PAGE_MODE_6_LEVEL:
1694 static void free_gcr3_tbl_level1(u64 *tbl)
1699 for (i = 0; i < 512; ++i) {
1700 if (!(tbl[i] & GCR3_VALID))
1703 ptr = __va(tbl[i] & PAGE_MASK);
1705 free_page((unsigned long)ptr);
1709 static void free_gcr3_tbl_level2(u64 *tbl)
1714 for (i = 0; i < 512; ++i) {
1715 if (!(tbl[i] & GCR3_VALID))
1718 ptr = __va(tbl[i] & PAGE_MASK);
1720 free_gcr3_tbl_level1(ptr);
1724 static void free_gcr3_table(struct protection_domain *domain)
1726 if (domain->glx == 2)
1727 free_gcr3_tbl_level2(domain->gcr3_tbl);
1728 else if (domain->glx == 1)
1729 free_gcr3_tbl_level1(domain->gcr3_tbl);
1731 BUG_ON(domain->glx != 0);
1733 free_page((unsigned long)domain->gcr3_tbl);
1737 * Free a domain, only used if something went wrong in the
1738 * allocation path and we need to free an already allocated page table
1740 static void dma_ops_domain_free(struct dma_ops_domain *dom)
1745 del_domain_from_list(&dom->domain);
1747 put_iova_domain(&dom->iovad);
1749 free_pagetable(&dom->domain);
1752 domain_id_free(dom->domain.id);
1758 * Allocates a new protection domain usable for the dma_ops functions.
1759 * It also initializes the page table and the address allocator data
1760 * structures required for the dma_ops interface
1762 static struct dma_ops_domain *dma_ops_domain_alloc(void)
1764 struct dma_ops_domain *dma_dom;
1766 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
1770 if (protection_domain_init(&dma_dom->domain))
1773 dma_dom->domain.mode = PAGE_MODE_3_LEVEL;
1774 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
1775 dma_dom->domain.flags = PD_DMA_OPS_MASK;
1776 if (!dma_dom->domain.pt_root)
1779 init_iova_domain(&dma_dom->iovad, PAGE_SIZE,
1780 IOVA_START_PFN, DMA_32BIT_PFN);
1782 /* Initialize reserved ranges */
1783 copy_reserved_iova(&reserved_iova_ranges, &dma_dom->iovad);
1785 add_domain_to_list(&dma_dom->domain);
1790 dma_ops_domain_free(dma_dom);
1796 * little helper function to check whether a given protection domain is a
1799 static bool dma_ops_domain(struct protection_domain *domain)
1801 return domain->flags & PD_DMA_OPS_MASK;
1804 static void set_dte_entry(u16 devid, struct protection_domain *domain, bool ats)
1809 if (domain->mode != PAGE_MODE_NONE)
1810 pte_root = virt_to_phys(domain->pt_root);
1812 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
1813 << DEV_ENTRY_MODE_SHIFT;
1814 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;
1816 flags = amd_iommu_dev_table[devid].data[1];
1819 flags |= DTE_FLAG_IOTLB;
1821 if (domain->flags & PD_IOMMUV2_MASK) {
1822 u64 gcr3 = __pa(domain->gcr3_tbl);
1823 u64 glx = domain->glx;
1826 pte_root |= DTE_FLAG_GV;
1827 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
1829 /* First mask out possible old values for GCR3 table */
1830 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
1833 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
1836 /* Encode GCR3 table into DTE */
1837 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
1840 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
1843 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
1847 flags &= ~(0xffffUL);
1848 flags |= domain->id;
1850 amd_iommu_dev_table[devid].data[1] = flags;
1851 amd_iommu_dev_table[devid].data[0] = pte_root;
1854 static void clear_dte_entry(u16 devid)
1856 /* remove entry from the device table seen by the hardware */
1857 amd_iommu_dev_table[devid].data[0] = IOMMU_PTE_P | IOMMU_PTE_TV;
1858 amd_iommu_dev_table[devid].data[1] &= DTE_FLAG_MASK;
1860 amd_iommu_apply_erratum_63(devid);
1863 static void do_attach(struct iommu_dev_data *dev_data,
1864 struct protection_domain *domain)
1866 struct amd_iommu *iommu;
1870 iommu = amd_iommu_rlookup_table[dev_data->devid];
1871 alias = dev_data->alias;
1872 ats = dev_data->ats.enabled;
1874 /* Update data structures */
1875 dev_data->domain = domain;
1876 list_add(&dev_data->list, &domain->dev_list);
1878 /* Do reference counting */
1879 domain->dev_iommu[iommu->index] += 1;
1880 domain->dev_cnt += 1;
1882 /* Update device table */
1883 set_dte_entry(dev_data->devid, domain, ats);
1884 if (alias != dev_data->devid)
1885 set_dte_entry(alias, domain, ats);
1887 device_flush_dte(dev_data);
1890 static void do_detach(struct iommu_dev_data *dev_data)
1892 struct amd_iommu *iommu;
1896 * First check if the device is still attached. It might already
1897 * be detached from its domain because the generic
1898 * iommu_detach_group code detached it and we try again here in
1899 * our alias handling.
1901 if (!dev_data->domain)
1904 iommu = amd_iommu_rlookup_table[dev_data->devid];
1905 alias = dev_data->alias;
1907 /* decrease reference counters */
1908 dev_data->domain->dev_iommu[iommu->index] -= 1;
1909 dev_data->domain->dev_cnt -= 1;
1911 /* Update data structures */
1912 dev_data->domain = NULL;
1913 list_del(&dev_data->list);
1914 clear_dte_entry(dev_data->devid);
1915 if (alias != dev_data->devid)
1916 clear_dte_entry(alias);
1918 /* Flush the DTE entry */
1919 device_flush_dte(dev_data);
1923 * If a device is not yet associated with a domain, this function does
1924 * assigns it visible for the hardware
1926 static int __attach_device(struct iommu_dev_data *dev_data,
1927 struct protection_domain *domain)
1932 * Must be called with IRQs disabled. Warn here to detect early
1935 WARN_ON(!irqs_disabled());
1938 spin_lock(&domain->lock);
1941 if (dev_data->domain != NULL)
1944 /* Attach alias group root */
1945 do_attach(dev_data, domain);
1952 spin_unlock(&domain->lock);
1958 static void pdev_iommuv2_disable(struct pci_dev *pdev)
1960 pci_disable_ats(pdev);
1961 pci_disable_pri(pdev);
1962 pci_disable_pasid(pdev);
1965 /* FIXME: Change generic reset-function to do the same */
1966 static int pri_reset_while_enabled(struct pci_dev *pdev)
1971 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
1975 pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
1976 control |= PCI_PRI_CTRL_RESET;
1977 pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
1982 static int pdev_iommuv2_enable(struct pci_dev *pdev)
1987 /* FIXME: Hardcode number of outstanding requests for now */
1989 if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
1991 reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
1993 /* Only allow access to user-accessible pages */
1994 ret = pci_enable_pasid(pdev, 0);
1998 /* First reset the PRI state of the device */
1999 ret = pci_reset_pri(pdev);
2004 ret = pci_enable_pri(pdev, reqs);
2009 ret = pri_reset_while_enabled(pdev);
2014 ret = pci_enable_ats(pdev, PAGE_SHIFT);
2021 pci_disable_pri(pdev);
2022 pci_disable_pasid(pdev);
2027 /* FIXME: Move this to PCI code */
2028 #define PCI_PRI_TLP_OFF (1 << 15)
2030 static bool pci_pri_tlp_required(struct pci_dev *pdev)
2035 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2039 pci_read_config_word(pdev, pos + PCI_PRI_STATUS, &status);
2041 return (status & PCI_PRI_TLP_OFF) ? true : false;
2045 * If a device is not yet associated with a domain, this function
2046 * assigns it visible for the hardware
2048 static int attach_device(struct device *dev,
2049 struct protection_domain *domain)
2051 struct pci_dev *pdev;
2052 struct iommu_dev_data *dev_data;
2053 unsigned long flags;
2056 dev_data = get_dev_data(dev);
2058 if (!dev_is_pci(dev))
2059 goto skip_ats_check;
2061 pdev = to_pci_dev(dev);
2062 if (domain->flags & PD_IOMMUV2_MASK) {
2063 if (!dev_data->passthrough)
2066 if (dev_data->iommu_v2) {
2067 if (pdev_iommuv2_enable(pdev) != 0)
2070 dev_data->ats.enabled = true;
2071 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2072 dev_data->pri_tlp = pci_pri_tlp_required(pdev);
2074 } else if (amd_iommu_iotlb_sup &&
2075 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
2076 dev_data->ats.enabled = true;
2077 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2081 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2082 ret = __attach_device(dev_data, domain);
2083 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2086 * We might boot into a crash-kernel here. The crashed kernel
2087 * left the caches in the IOMMU dirty. So we have to flush
2088 * here to evict all dirty stuff.
2090 domain_flush_tlb_pde(domain);
2096 * Removes a device from a protection domain (unlocked)
2098 static void __detach_device(struct iommu_dev_data *dev_data)
2100 struct protection_domain *domain;
2103 * Must be called with IRQs disabled. Warn here to detect early
2106 WARN_ON(!irqs_disabled());
2108 if (WARN_ON(!dev_data->domain))
2111 domain = dev_data->domain;
2113 spin_lock(&domain->lock);
2115 do_detach(dev_data);
2117 spin_unlock(&domain->lock);
2121 * Removes a device from a protection domain (with devtable_lock held)
2123 static void detach_device(struct device *dev)
2125 struct protection_domain *domain;
2126 struct iommu_dev_data *dev_data;
2127 unsigned long flags;
2129 dev_data = get_dev_data(dev);
2130 domain = dev_data->domain;
2132 /* lock device table */
2133 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2134 __detach_device(dev_data);
2135 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2137 if (!dev_is_pci(dev))
2140 if (domain->flags & PD_IOMMUV2_MASK && dev_data->iommu_v2)
2141 pdev_iommuv2_disable(to_pci_dev(dev));
2142 else if (dev_data->ats.enabled)
2143 pci_disable_ats(to_pci_dev(dev));
2145 dev_data->ats.enabled = false;
2148 static int amd_iommu_add_device(struct device *dev)
2150 struct iommu_dev_data *dev_data;
2151 struct iommu_domain *domain;
2152 struct amd_iommu *iommu;
2155 if (!check_device(dev) || get_dev_data(dev))
2158 devid = get_device_id(dev);
2162 iommu = amd_iommu_rlookup_table[devid];
2164 ret = iommu_init_device(dev);
2166 if (ret != -ENOTSUPP)
2167 pr_err("Failed to initialize device %s - trying to proceed anyway\n",
2170 iommu_ignore_device(dev);
2171 dev->dma_ops = &nommu_dma_ops;
2174 init_iommu_group(dev);
2176 dev_data = get_dev_data(dev);
2180 if (iommu_pass_through || dev_data->iommu_v2)
2181 iommu_request_dm_for_dev(dev);
2183 /* Domains are initialized for this device - have a look what we ended up with */
2184 domain = iommu_get_domain_for_dev(dev);
2185 if (domain->type == IOMMU_DOMAIN_IDENTITY)
2186 dev_data->passthrough = true;
2188 dev->dma_ops = &amd_iommu_dma_ops;
2191 iommu_completion_wait(iommu);
2196 static void amd_iommu_remove_device(struct device *dev)
2198 struct amd_iommu *iommu;
2201 if (!check_device(dev))
2204 devid = get_device_id(dev);
2208 iommu = amd_iommu_rlookup_table[devid];
2210 iommu_uninit_device(dev);
2211 iommu_completion_wait(iommu);
2214 static struct iommu_group *amd_iommu_device_group(struct device *dev)
2216 if (dev_is_pci(dev))
2217 return pci_device_group(dev);
2219 return acpihid_device_group(dev);
2222 /*****************************************************************************
2224 * The next functions belong to the dma_ops mapping/unmapping code.
2226 *****************************************************************************/
2228 static void __queue_flush(struct flush_queue *queue)
2230 struct protection_domain *domain;
2231 unsigned long flags;
2234 /* First flush TLB of all known domains */
2235 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
2236 list_for_each_entry(domain, &amd_iommu_pd_list, list)
2237 domain_flush_tlb(domain);
2238 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
2240 /* Wait until flushes have completed */
2241 domain_flush_complete(NULL);
2243 for (idx = 0; idx < queue->next; ++idx) {
2244 struct flush_queue_entry *entry;
2246 entry = queue->entries + idx;
2248 free_iova_fast(&entry->dma_dom->iovad,
2252 /* Not really necessary, just to make sure we catch any bugs */
2253 entry->dma_dom = NULL;
2259 static void queue_flush_all(void)
2263 for_each_possible_cpu(cpu) {
2264 struct flush_queue *queue;
2265 unsigned long flags;
2267 queue = per_cpu_ptr(&flush_queue, cpu);
2268 spin_lock_irqsave(&queue->lock, flags);
2269 if (queue->next > 0)
2270 __queue_flush(queue);
2271 spin_unlock_irqrestore(&queue->lock, flags);
2275 static void queue_flush_timeout(unsigned long unsused)
2277 atomic_set(&queue_timer_on, 0);
2281 static void queue_add(struct dma_ops_domain *dma_dom,
2282 unsigned long address, unsigned long pages)
2284 struct flush_queue_entry *entry;
2285 struct flush_queue *queue;
2286 unsigned long flags;
2289 pages = __roundup_pow_of_two(pages);
2290 address >>= PAGE_SHIFT;
2292 queue = get_cpu_ptr(&flush_queue);
2293 spin_lock_irqsave(&queue->lock, flags);
2295 if (queue->next == FLUSH_QUEUE_SIZE)
2296 __queue_flush(queue);
2298 idx = queue->next++;
2299 entry = queue->entries + idx;
2301 entry->iova_pfn = address;
2302 entry->pages = pages;
2303 entry->dma_dom = dma_dom;
2305 spin_unlock_irqrestore(&queue->lock, flags);
2307 if (atomic_cmpxchg(&queue_timer_on, 0, 1) == 0)
2308 mod_timer(&queue_timer, jiffies + msecs_to_jiffies(10));
2310 put_cpu_ptr(&flush_queue);
2315 * In the dma_ops path we only have the struct device. This function
2316 * finds the corresponding IOMMU, the protection domain and the
2317 * requestor id for a given device.
2318 * If the device is not yet associated with a domain this is also done
2321 static struct protection_domain *get_domain(struct device *dev)
2323 struct protection_domain *domain;
2325 if (!check_device(dev))
2326 return ERR_PTR(-EINVAL);
2328 domain = get_dev_data(dev)->domain;
2329 if (!dma_ops_domain(domain))
2330 return ERR_PTR(-EBUSY);
2335 static void update_device_table(struct protection_domain *domain)
2337 struct iommu_dev_data *dev_data;
2339 list_for_each_entry(dev_data, &domain->dev_list, list) {
2340 set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled);
2342 if (dev_data->devid == dev_data->alias)
2345 /* There is an alias, update device table entry for it */
2346 set_dte_entry(dev_data->alias, domain, dev_data->ats.enabled);
2350 static void update_domain(struct protection_domain *domain)
2352 if (!domain->updated)
2355 update_device_table(domain);
2357 domain_flush_devices(domain);
2358 domain_flush_tlb_pde(domain);
2360 domain->updated = false;
2363 static int dir2prot(enum dma_data_direction direction)
2365 if (direction == DMA_TO_DEVICE)
2366 return IOMMU_PROT_IR;
2367 else if (direction == DMA_FROM_DEVICE)
2368 return IOMMU_PROT_IW;
2369 else if (direction == DMA_BIDIRECTIONAL)
2370 return IOMMU_PROT_IW | IOMMU_PROT_IR;
2375 * This function contains common code for mapping of a physically
2376 * contiguous memory region into DMA address space. It is used by all
2377 * mapping functions provided with this IOMMU driver.
2378 * Must be called with the domain lock held.
2380 static dma_addr_t __map_single(struct device *dev,
2381 struct dma_ops_domain *dma_dom,
2384 enum dma_data_direction direction,
2387 dma_addr_t offset = paddr & ~PAGE_MASK;
2388 dma_addr_t address, start, ret;
2393 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
2396 address = dma_ops_alloc_iova(dev, dma_dom, pages, dma_mask);
2397 if (address == DMA_ERROR_CODE)
2400 prot = dir2prot(direction);
2403 for (i = 0; i < pages; ++i) {
2404 ret = iommu_map_page(&dma_dom->domain, start, paddr,
2405 PAGE_SIZE, prot, GFP_ATOMIC);
2414 if (unlikely(amd_iommu_np_cache)) {
2415 domain_flush_pages(&dma_dom->domain, address, size);
2416 domain_flush_complete(&dma_dom->domain);
2424 for (--i; i >= 0; --i) {
2426 iommu_unmap_page(&dma_dom->domain, start, PAGE_SIZE);
2429 domain_flush_tlb(&dma_dom->domain);
2430 domain_flush_complete(&dma_dom->domain);
2432 dma_ops_free_iova(dma_dom, address, pages);
2434 return DMA_ERROR_CODE;
2438 * Does the reverse of the __map_single function. Must be called with
2439 * the domain lock held too
2441 static void __unmap_single(struct dma_ops_domain *dma_dom,
2442 dma_addr_t dma_addr,
2446 dma_addr_t flush_addr;
2447 dma_addr_t i, start;
2450 flush_addr = dma_addr;
2451 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
2452 dma_addr &= PAGE_MASK;
2455 for (i = 0; i < pages; ++i) {
2456 iommu_unmap_page(&dma_dom->domain, start, PAGE_SIZE);
2460 if (amd_iommu_unmap_flush) {
2461 dma_ops_free_iova(dma_dom, dma_addr, pages);
2462 domain_flush_tlb(&dma_dom->domain);
2463 domain_flush_complete(&dma_dom->domain);
2465 queue_add(dma_dom, dma_addr, pages);
2470 * The exported map_single function for dma_ops.
2472 static dma_addr_t map_page(struct device *dev, struct page *page,
2473 unsigned long offset, size_t size,
2474 enum dma_data_direction dir,
2475 unsigned long attrs)
2477 phys_addr_t paddr = page_to_phys(page) + offset;
2478 struct protection_domain *domain;
2479 struct dma_ops_domain *dma_dom;
2482 domain = get_domain(dev);
2483 if (PTR_ERR(domain) == -EINVAL)
2484 return (dma_addr_t)paddr;
2485 else if (IS_ERR(domain))
2486 return DMA_ERROR_CODE;
2488 dma_mask = *dev->dma_mask;
2489 dma_dom = to_dma_ops_domain(domain);
2491 return __map_single(dev, dma_dom, paddr, size, dir, dma_mask);
2495 * The exported unmap_single function for dma_ops.
2497 static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2498 enum dma_data_direction dir, unsigned long attrs)
2500 struct protection_domain *domain;
2501 struct dma_ops_domain *dma_dom;
2503 domain = get_domain(dev);
2507 dma_dom = to_dma_ops_domain(domain);
2509 __unmap_single(dma_dom, dma_addr, size, dir);
2512 static int sg_num_pages(struct device *dev,
2513 struct scatterlist *sglist,
2516 unsigned long mask, boundary_size;
2517 struct scatterlist *s;
2520 mask = dma_get_seg_boundary(dev);
2521 boundary_size = mask + 1 ? ALIGN(mask + 1, PAGE_SIZE) >> PAGE_SHIFT :
2522 1UL << (BITS_PER_LONG - PAGE_SHIFT);
2524 for_each_sg(sglist, s, nelems, i) {
2527 s->dma_address = npages << PAGE_SHIFT;
2528 p = npages % boundary_size;
2529 n = iommu_num_pages(sg_phys(s), s->length, PAGE_SIZE);
2530 if (p + n > boundary_size)
2531 npages += boundary_size - p;
2539 * The exported map_sg function for dma_ops (handles scatter-gather
2542 static int map_sg(struct device *dev, struct scatterlist *sglist,
2543 int nelems, enum dma_data_direction direction,
2544 unsigned long attrs)
2546 int mapped_pages = 0, npages = 0, prot = 0, i;
2547 struct protection_domain *domain;
2548 struct dma_ops_domain *dma_dom;
2549 struct scatterlist *s;
2550 unsigned long address;
2553 domain = get_domain(dev);
2557 dma_dom = to_dma_ops_domain(domain);
2558 dma_mask = *dev->dma_mask;
2560 npages = sg_num_pages(dev, sglist, nelems);
2562 address = dma_ops_alloc_iova(dev, dma_dom, npages, dma_mask);
2563 if (address == DMA_ERROR_CODE)
2566 prot = dir2prot(direction);
2568 /* Map all sg entries */
2569 for_each_sg(sglist, s, nelems, i) {
2570 int j, pages = iommu_num_pages(sg_phys(s), s->length, PAGE_SIZE);
2572 for (j = 0; j < pages; ++j) {
2573 unsigned long bus_addr, phys_addr;
2576 bus_addr = address + s->dma_address + (j << PAGE_SHIFT);
2577 phys_addr = (sg_phys(s) & PAGE_MASK) + (j << PAGE_SHIFT);
2578 ret = iommu_map_page(domain, bus_addr, phys_addr, PAGE_SIZE, prot, GFP_ATOMIC);
2586 /* Everything is mapped - write the right values into s->dma_address */
2587 for_each_sg(sglist, s, nelems, i) {
2588 s->dma_address += address + s->offset;
2589 s->dma_length = s->length;
2595 pr_err("%s: IOMMU mapping error in map_sg (io-pages: %d)\n",
2596 dev_name(dev), npages);
2598 for_each_sg(sglist, s, nelems, i) {
2599 int j, pages = iommu_num_pages(sg_phys(s), s->length, PAGE_SIZE);
2601 for (j = 0; j < pages; ++j) {
2602 unsigned long bus_addr;
2604 bus_addr = address + s->dma_address + (j << PAGE_SHIFT);
2605 iommu_unmap_page(domain, bus_addr, PAGE_SIZE);
2613 free_iova_fast(&dma_dom->iovad, address, npages);
2620 * The exported map_sg function for dma_ops (handles scatter-gather
2623 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2624 int nelems, enum dma_data_direction dir,
2625 unsigned long attrs)
2627 struct protection_domain *domain;
2628 struct dma_ops_domain *dma_dom;
2629 unsigned long startaddr;
2632 domain = get_domain(dev);
2636 startaddr = sg_dma_address(sglist) & PAGE_MASK;
2637 dma_dom = to_dma_ops_domain(domain);
2638 npages = sg_num_pages(dev, sglist, nelems);
2640 __unmap_single(dma_dom, startaddr, npages << PAGE_SHIFT, dir);
2644 * The exported alloc_coherent function for dma_ops.
2646 static void *alloc_coherent(struct device *dev, size_t size,
2647 dma_addr_t *dma_addr, gfp_t flag,
2648 unsigned long attrs)
2650 u64 dma_mask = dev->coherent_dma_mask;
2651 struct protection_domain *domain;
2652 struct dma_ops_domain *dma_dom;
2655 domain = get_domain(dev);
2656 if (PTR_ERR(domain) == -EINVAL) {
2657 page = alloc_pages(flag, get_order(size));
2658 *dma_addr = page_to_phys(page);
2659 return page_address(page);
2660 } else if (IS_ERR(domain))
2663 dma_dom = to_dma_ops_domain(domain);
2664 size = PAGE_ALIGN(size);
2665 dma_mask = dev->coherent_dma_mask;
2666 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
2669 page = alloc_pages(flag | __GFP_NOWARN, get_order(size));
2671 if (!gfpflags_allow_blocking(flag))
2674 page = dma_alloc_from_contiguous(dev, size >> PAGE_SHIFT,
2675 get_order(size), flag);
2681 dma_mask = *dev->dma_mask;
2683 *dma_addr = __map_single(dev, dma_dom, page_to_phys(page),
2684 size, DMA_BIDIRECTIONAL, dma_mask);
2686 if (*dma_addr == DMA_ERROR_CODE)
2689 return page_address(page);
2693 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2694 __free_pages(page, get_order(size));
2700 * The exported free_coherent function for dma_ops.
2702 static void free_coherent(struct device *dev, size_t size,
2703 void *virt_addr, dma_addr_t dma_addr,
2704 unsigned long attrs)
2706 struct protection_domain *domain;
2707 struct dma_ops_domain *dma_dom;
2710 page = virt_to_page(virt_addr);
2711 size = PAGE_ALIGN(size);
2713 domain = get_domain(dev);
2717 dma_dom = to_dma_ops_domain(domain);
2719 __unmap_single(dma_dom, dma_addr, size, DMA_BIDIRECTIONAL);
2722 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2723 __free_pages(page, get_order(size));
2727 * This function is called by the DMA layer to find out if we can handle a
2728 * particular device. It is part of the dma_ops.
2730 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
2732 return check_device(dev);
2735 static const struct dma_map_ops amd_iommu_dma_ops = {
2736 .alloc = alloc_coherent,
2737 .free = free_coherent,
2738 .map_page = map_page,
2739 .unmap_page = unmap_page,
2741 .unmap_sg = unmap_sg,
2742 .dma_supported = amd_iommu_dma_supported,
2745 static int init_reserved_iova_ranges(void)
2747 struct pci_dev *pdev = NULL;
2750 init_iova_domain(&reserved_iova_ranges, PAGE_SIZE,
2751 IOVA_START_PFN, DMA_32BIT_PFN);
2753 lockdep_set_class(&reserved_iova_ranges.iova_rbtree_lock,
2754 &reserved_rbtree_key);
2756 /* MSI memory range */
2757 val = reserve_iova(&reserved_iova_ranges,
2758 IOVA_PFN(MSI_RANGE_START), IOVA_PFN(MSI_RANGE_END));
2760 pr_err("Reserving MSI range failed\n");
2764 /* HT memory range */
2765 val = reserve_iova(&reserved_iova_ranges,
2766 IOVA_PFN(HT_RANGE_START), IOVA_PFN(HT_RANGE_END));
2768 pr_err("Reserving HT range failed\n");
2773 * Memory used for PCI resources
2774 * FIXME: Check whether we can reserve the PCI-hole completly
2776 for_each_pci_dev(pdev) {
2779 for (i = 0; i < PCI_NUM_RESOURCES; ++i) {
2780 struct resource *r = &pdev->resource[i];
2782 if (!(r->flags & IORESOURCE_MEM))
2785 val = reserve_iova(&reserved_iova_ranges,
2789 pr_err("Reserve pci-resource range failed\n");
2798 int __init amd_iommu_init_api(void)
2800 int ret, cpu, err = 0;
2802 ret = iova_cache_get();
2806 ret = init_reserved_iova_ranges();
2810 for_each_possible_cpu(cpu) {
2811 struct flush_queue *queue = per_cpu_ptr(&flush_queue, cpu);
2813 queue->entries = kzalloc(FLUSH_QUEUE_SIZE *
2814 sizeof(*queue->entries),
2816 if (!queue->entries)
2819 spin_lock_init(&queue->lock);
2822 err = bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
2825 #ifdef CONFIG_ARM_AMBA
2826 err = bus_set_iommu(&amba_bustype, &amd_iommu_ops);
2830 err = bus_set_iommu(&platform_bus_type, &amd_iommu_ops);
2836 for_each_possible_cpu(cpu) {
2837 struct flush_queue *queue = per_cpu_ptr(&flush_queue, cpu);
2839 kfree(queue->entries);
2845 int __init amd_iommu_init_dma_ops(void)
2847 setup_timer(&queue_timer, queue_flush_timeout, 0);
2848 atomic_set(&queue_timer_on, 0);
2850 swiotlb = iommu_pass_through ? 1 : 0;
2854 * In case we don't initialize SWIOTLB (actually the common case
2855 * when AMD IOMMU is enabled), make sure there are global
2856 * dma_ops set as a fall-back for devices not handled by this
2857 * driver (for example non-PCI devices).
2860 dma_ops = &nommu_dma_ops;
2862 if (amd_iommu_unmap_flush)
2863 pr_info("AMD-Vi: IO/TLB flush on unmap enabled\n");
2865 pr_info("AMD-Vi: Lazy IO/TLB flushing enabled\n");
2871 /*****************************************************************************
2873 * The following functions belong to the exported interface of AMD IOMMU
2875 * This interface allows access to lower level functions of the IOMMU
2876 * like protection domain handling and assignement of devices to domains
2877 * which is not possible with the dma_ops interface.
2879 *****************************************************************************/
2881 static void cleanup_domain(struct protection_domain *domain)
2883 struct iommu_dev_data *entry;
2884 unsigned long flags;
2886 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2888 while (!list_empty(&domain->dev_list)) {
2889 entry = list_first_entry(&domain->dev_list,
2890 struct iommu_dev_data, list);
2891 __detach_device(entry);
2894 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2897 static void protection_domain_free(struct protection_domain *domain)
2902 del_domain_from_list(domain);
2905 domain_id_free(domain->id);
2910 static int protection_domain_init(struct protection_domain *domain)
2912 spin_lock_init(&domain->lock);
2913 mutex_init(&domain->api_lock);
2914 domain->id = domain_id_alloc();
2917 INIT_LIST_HEAD(&domain->dev_list);
2922 static struct protection_domain *protection_domain_alloc(void)
2924 struct protection_domain *domain;
2926 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2930 if (protection_domain_init(domain))
2933 add_domain_to_list(domain);
2943 static struct iommu_domain *amd_iommu_domain_alloc(unsigned type)
2945 struct protection_domain *pdomain;
2946 struct dma_ops_domain *dma_domain;
2949 case IOMMU_DOMAIN_UNMANAGED:
2950 pdomain = protection_domain_alloc();
2954 pdomain->mode = PAGE_MODE_3_LEVEL;
2955 pdomain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2956 if (!pdomain->pt_root) {
2957 protection_domain_free(pdomain);
2961 pdomain->domain.geometry.aperture_start = 0;
2962 pdomain->domain.geometry.aperture_end = ~0ULL;
2963 pdomain->domain.geometry.force_aperture = true;
2966 case IOMMU_DOMAIN_DMA:
2967 dma_domain = dma_ops_domain_alloc();
2969 pr_err("AMD-Vi: Failed to allocate\n");
2972 pdomain = &dma_domain->domain;
2974 case IOMMU_DOMAIN_IDENTITY:
2975 pdomain = protection_domain_alloc();
2979 pdomain->mode = PAGE_MODE_NONE;
2985 return &pdomain->domain;
2988 static void amd_iommu_domain_free(struct iommu_domain *dom)
2990 struct protection_domain *domain;
2991 struct dma_ops_domain *dma_dom;
2993 domain = to_pdomain(dom);
2995 if (domain->dev_cnt > 0)
2996 cleanup_domain(domain);
2998 BUG_ON(domain->dev_cnt != 0);
3003 switch (dom->type) {
3004 case IOMMU_DOMAIN_DMA:
3006 * First make sure the domain is no longer referenced from the
3011 /* Now release the domain */
3012 dma_dom = to_dma_ops_domain(domain);
3013 dma_ops_domain_free(dma_dom);
3016 if (domain->mode != PAGE_MODE_NONE)
3017 free_pagetable(domain);
3019 if (domain->flags & PD_IOMMUV2_MASK)
3020 free_gcr3_table(domain);
3022 protection_domain_free(domain);
3027 static void amd_iommu_detach_device(struct iommu_domain *dom,
3030 struct iommu_dev_data *dev_data = dev->archdata.iommu;
3031 struct amd_iommu *iommu;
3034 if (!check_device(dev))
3037 devid = get_device_id(dev);
3041 if (dev_data->domain != NULL)
3044 iommu = amd_iommu_rlookup_table[devid];
3048 #ifdef CONFIG_IRQ_REMAP
3049 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) &&
3050 (dom->type == IOMMU_DOMAIN_UNMANAGED))
3051 dev_data->use_vapic = 0;
3054 iommu_completion_wait(iommu);
3057 static int amd_iommu_attach_device(struct iommu_domain *dom,
3060 struct protection_domain *domain = to_pdomain(dom);
3061 struct iommu_dev_data *dev_data;
3062 struct amd_iommu *iommu;
3065 if (!check_device(dev))
3068 dev_data = dev->archdata.iommu;
3070 iommu = amd_iommu_rlookup_table[dev_data->devid];
3074 if (dev_data->domain)
3077 ret = attach_device(dev, domain);
3079 #ifdef CONFIG_IRQ_REMAP
3080 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
3081 if (dom->type == IOMMU_DOMAIN_UNMANAGED)
3082 dev_data->use_vapic = 1;
3084 dev_data->use_vapic = 0;
3088 iommu_completion_wait(iommu);
3093 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
3094 phys_addr_t paddr, size_t page_size, int iommu_prot)
3096 struct protection_domain *domain = to_pdomain(dom);
3100 if (domain->mode == PAGE_MODE_NONE)
3103 if (iommu_prot & IOMMU_READ)
3104 prot |= IOMMU_PROT_IR;
3105 if (iommu_prot & IOMMU_WRITE)
3106 prot |= IOMMU_PROT_IW;
3108 mutex_lock(&domain->api_lock);
3109 ret = iommu_map_page(domain, iova, paddr, page_size, prot, GFP_KERNEL);
3110 mutex_unlock(&domain->api_lock);
3115 static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
3118 struct protection_domain *domain = to_pdomain(dom);
3121 if (domain->mode == PAGE_MODE_NONE)
3124 mutex_lock(&domain->api_lock);
3125 unmap_size = iommu_unmap_page(domain, iova, page_size);
3126 mutex_unlock(&domain->api_lock);
3128 domain_flush_tlb_pde(domain);
3133 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
3136 struct protection_domain *domain = to_pdomain(dom);
3137 unsigned long offset_mask, pte_pgsize;
3140 if (domain->mode == PAGE_MODE_NONE)
3143 pte = fetch_pte(domain, iova, &pte_pgsize);
3145 if (!pte || !IOMMU_PTE_PRESENT(*pte))
3148 offset_mask = pte_pgsize - 1;
3149 __pte = *pte & PM_ADDR_MASK;
3151 return (__pte & ~offset_mask) | (iova & offset_mask);
3154 static bool amd_iommu_capable(enum iommu_cap cap)
3157 case IOMMU_CAP_CACHE_COHERENCY:
3159 case IOMMU_CAP_INTR_REMAP:
3160 return (irq_remapping_enabled == 1);
3161 case IOMMU_CAP_NOEXEC:
3168 static void amd_iommu_get_resv_regions(struct device *dev,
3169 struct list_head *head)
3171 struct iommu_resv_region *region;
3172 struct unity_map_entry *entry;
3175 devid = get_device_id(dev);
3179 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
3183 if (devid < entry->devid_start || devid > entry->devid_end)
3186 length = entry->address_end - entry->address_start;
3187 if (entry->prot & IOMMU_PROT_IR)
3189 if (entry->prot & IOMMU_PROT_IW)
3190 prot |= IOMMU_WRITE;
3192 region = iommu_alloc_resv_region(entry->address_start,
3196 pr_err("Out of memory allocating dm-regions for %s\n",
3200 list_add_tail(®ion->list, head);
3203 region = iommu_alloc_resv_region(MSI_RANGE_START,
3204 MSI_RANGE_END - MSI_RANGE_START + 1,
3208 list_add_tail(®ion->list, head);
3210 region = iommu_alloc_resv_region(HT_RANGE_START,
3211 HT_RANGE_END - HT_RANGE_START + 1,
3212 0, IOMMU_RESV_RESERVED);
3215 list_add_tail(®ion->list, head);
3218 static void amd_iommu_put_resv_regions(struct device *dev,
3219 struct list_head *head)
3221 struct iommu_resv_region *entry, *next;
3223 list_for_each_entry_safe(entry, next, head, list)
3227 static void amd_iommu_apply_resv_region(struct device *dev,
3228 struct iommu_domain *domain,
3229 struct iommu_resv_region *region)
3231 struct dma_ops_domain *dma_dom = to_dma_ops_domain(to_pdomain(domain));
3232 unsigned long start, end;
3234 start = IOVA_PFN(region->start);
3235 end = IOVA_PFN(region->start + region->length);
3237 WARN_ON_ONCE(reserve_iova(&dma_dom->iovad, start, end) == NULL);
3240 const struct iommu_ops amd_iommu_ops = {
3241 .capable = amd_iommu_capable,
3242 .domain_alloc = amd_iommu_domain_alloc,
3243 .domain_free = amd_iommu_domain_free,
3244 .attach_dev = amd_iommu_attach_device,
3245 .detach_dev = amd_iommu_detach_device,
3246 .map = amd_iommu_map,
3247 .unmap = amd_iommu_unmap,
3248 .map_sg = default_iommu_map_sg,
3249 .iova_to_phys = amd_iommu_iova_to_phys,
3250 .add_device = amd_iommu_add_device,
3251 .remove_device = amd_iommu_remove_device,
3252 .device_group = amd_iommu_device_group,
3253 .get_resv_regions = amd_iommu_get_resv_regions,
3254 .put_resv_regions = amd_iommu_put_resv_regions,
3255 .apply_resv_region = amd_iommu_apply_resv_region,
3256 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
3259 /*****************************************************************************
3261 * The next functions do a basic initialization of IOMMU for pass through
3264 * In passthrough mode the IOMMU is initialized and enabled but not used for
3265 * DMA-API translation.
3267 *****************************************************************************/
3269 /* IOMMUv2 specific functions */
3270 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
3272 return atomic_notifier_chain_register(&ppr_notifier, nb);
3274 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
3276 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
3278 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
3280 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
3282 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
3284 struct protection_domain *domain = to_pdomain(dom);
3285 unsigned long flags;
3287 spin_lock_irqsave(&domain->lock, flags);
3289 /* Update data structure */
3290 domain->mode = PAGE_MODE_NONE;
3291 domain->updated = true;
3293 /* Make changes visible to IOMMUs */
3294 update_domain(domain);
3296 /* Page-table is not visible to IOMMU anymore, so free it */
3297 free_pagetable(domain);
3299 spin_unlock_irqrestore(&domain->lock, flags);
3301 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
3303 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
3305 struct protection_domain *domain = to_pdomain(dom);
3306 unsigned long flags;
3309 if (pasids <= 0 || pasids > (PASID_MASK + 1))
3312 /* Number of GCR3 table levels required */
3313 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
3316 if (levels > amd_iommu_max_glx_val)
3319 spin_lock_irqsave(&domain->lock, flags);
3322 * Save us all sanity checks whether devices already in the
3323 * domain support IOMMUv2. Just force that the domain has no
3324 * devices attached when it is switched into IOMMUv2 mode.
3327 if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
3331 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
3332 if (domain->gcr3_tbl == NULL)
3335 domain->glx = levels;
3336 domain->flags |= PD_IOMMUV2_MASK;
3337 domain->updated = true;
3339 update_domain(domain);
3344 spin_unlock_irqrestore(&domain->lock, flags);
3348 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
3350 static int __flush_pasid(struct protection_domain *domain, int pasid,
3351 u64 address, bool size)
3353 struct iommu_dev_data *dev_data;
3354 struct iommu_cmd cmd;
3357 if (!(domain->flags & PD_IOMMUV2_MASK))
3360 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
3363 * IOMMU TLB needs to be flushed before Device TLB to
3364 * prevent device TLB refill from IOMMU TLB
3366 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
3367 if (domain->dev_iommu[i] == 0)
3370 ret = iommu_queue_command(amd_iommus[i], &cmd);
3375 /* Wait until IOMMU TLB flushes are complete */
3376 domain_flush_complete(domain);
3378 /* Now flush device TLBs */
3379 list_for_each_entry(dev_data, &domain->dev_list, list) {
3380 struct amd_iommu *iommu;
3384 There might be non-IOMMUv2 capable devices in an IOMMUv2
3387 if (!dev_data->ats.enabled)
3390 qdep = dev_data->ats.qdep;
3391 iommu = amd_iommu_rlookup_table[dev_data->devid];
3393 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
3394 qdep, address, size);
3396 ret = iommu_queue_command(iommu, &cmd);
3401 /* Wait until all device TLBs are flushed */
3402 domain_flush_complete(domain);
3411 static int __amd_iommu_flush_page(struct protection_domain *domain, int pasid,
3414 return __flush_pasid(domain, pasid, address, false);
3417 int amd_iommu_flush_page(struct iommu_domain *dom, int pasid,
3420 struct protection_domain *domain = to_pdomain(dom);
3421 unsigned long flags;
3424 spin_lock_irqsave(&domain->lock, flags);
3425 ret = __amd_iommu_flush_page(domain, pasid, address);
3426 spin_unlock_irqrestore(&domain->lock, flags);
3430 EXPORT_SYMBOL(amd_iommu_flush_page);
3432 static int __amd_iommu_flush_tlb(struct protection_domain *domain, int pasid)
3434 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
3438 int amd_iommu_flush_tlb(struct iommu_domain *dom, int pasid)
3440 struct protection_domain *domain = to_pdomain(dom);
3441 unsigned long flags;
3444 spin_lock_irqsave(&domain->lock, flags);
3445 ret = __amd_iommu_flush_tlb(domain, pasid);
3446 spin_unlock_irqrestore(&domain->lock, flags);
3450 EXPORT_SYMBOL(amd_iommu_flush_tlb);
3452 static u64 *__get_gcr3_pte(u64 *root, int level, int pasid, bool alloc)
3459 index = (pasid >> (9 * level)) & 0x1ff;
3465 if (!(*pte & GCR3_VALID)) {
3469 root = (void *)get_zeroed_page(GFP_ATOMIC);
3473 *pte = __pa(root) | GCR3_VALID;
3476 root = __va(*pte & PAGE_MASK);
3484 static int __set_gcr3(struct protection_domain *domain, int pasid,
3489 if (domain->mode != PAGE_MODE_NONE)
3492 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3496 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3498 return __amd_iommu_flush_tlb(domain, pasid);
3501 static int __clear_gcr3(struct protection_domain *domain, int pasid)
3505 if (domain->mode != PAGE_MODE_NONE)
3508 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3514 return __amd_iommu_flush_tlb(domain, pasid);
3517 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, int pasid,
3520 struct protection_domain *domain = to_pdomain(dom);
3521 unsigned long flags;
3524 spin_lock_irqsave(&domain->lock, flags);
3525 ret = __set_gcr3(domain, pasid, cr3);
3526 spin_unlock_irqrestore(&domain->lock, flags);
3530 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3532 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, int pasid)
3534 struct protection_domain *domain = to_pdomain(dom);
3535 unsigned long flags;
3538 spin_lock_irqsave(&domain->lock, flags);
3539 ret = __clear_gcr3(domain, pasid);
3540 spin_unlock_irqrestore(&domain->lock, flags);
3544 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3546 int amd_iommu_complete_ppr(struct pci_dev *pdev, int pasid,
3547 int status, int tag)
3549 struct iommu_dev_data *dev_data;
3550 struct amd_iommu *iommu;
3551 struct iommu_cmd cmd;
3553 dev_data = get_dev_data(&pdev->dev);
3554 iommu = amd_iommu_rlookup_table[dev_data->devid];
3556 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3557 tag, dev_data->pri_tlp);
3559 return iommu_queue_command(iommu, &cmd);
3561 EXPORT_SYMBOL(amd_iommu_complete_ppr);
3563 struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3565 struct protection_domain *pdomain;
3567 pdomain = get_domain(&pdev->dev);
3568 if (IS_ERR(pdomain))
3571 /* Only return IOMMUv2 domains */
3572 if (!(pdomain->flags & PD_IOMMUV2_MASK))
3575 return &pdomain->domain;
3577 EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3579 void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3581 struct iommu_dev_data *dev_data;
3583 if (!amd_iommu_v2_supported())
3586 dev_data = get_dev_data(&pdev->dev);
3587 dev_data->errata |= (1 << erratum);
3589 EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3591 int amd_iommu_device_info(struct pci_dev *pdev,
3592 struct amd_iommu_device_info *info)
3597 if (pdev == NULL || info == NULL)
3600 if (!amd_iommu_v2_supported())
3603 memset(info, 0, sizeof(*info));
3605 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS);
3607 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3609 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3611 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3613 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3617 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3618 max_pasids = min(max_pasids, (1 << 20));
3620 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3621 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3623 features = pci_pasid_features(pdev);
3624 if (features & PCI_PASID_CAP_EXEC)
3625 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3626 if (features & PCI_PASID_CAP_PRIV)
3627 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3632 EXPORT_SYMBOL(amd_iommu_device_info);
3634 #ifdef CONFIG_IRQ_REMAP
3636 /*****************************************************************************
3638 * Interrupt Remapping Implementation
3640 *****************************************************************************/
3642 static struct irq_chip amd_ir_chip;
3644 #define DTE_IRQ_PHYS_ADDR_MASK (((1ULL << 45)-1) << 6)
3645 #define DTE_IRQ_REMAP_INTCTL (2ULL << 60)
3646 #define DTE_IRQ_TABLE_LEN (8ULL << 1)
3647 #define DTE_IRQ_REMAP_ENABLE 1ULL
3649 static void set_dte_irq_entry(u16 devid, struct irq_remap_table *table)
3653 dte = amd_iommu_dev_table[devid].data[2];
3654 dte &= ~DTE_IRQ_PHYS_ADDR_MASK;
3655 dte |= virt_to_phys(table->table);
3656 dte |= DTE_IRQ_REMAP_INTCTL;
3657 dte |= DTE_IRQ_TABLE_LEN;
3658 dte |= DTE_IRQ_REMAP_ENABLE;
3660 amd_iommu_dev_table[devid].data[2] = dte;
3663 static struct irq_remap_table *get_irq_table(u16 devid, bool ioapic)
3665 struct irq_remap_table *table = NULL;
3666 struct amd_iommu *iommu;
3667 unsigned long flags;
3670 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3672 iommu = amd_iommu_rlookup_table[devid];
3676 table = irq_lookup_table[devid];
3680 alias = amd_iommu_alias_table[devid];
3681 table = irq_lookup_table[alias];
3683 irq_lookup_table[devid] = table;
3684 set_dte_irq_entry(devid, table);
3685 iommu_flush_dte(iommu, devid);
3689 /* Nothing there yet, allocate new irq remapping table */
3690 table = kzalloc(sizeof(*table), GFP_ATOMIC);
3694 /* Initialize table spin-lock */
3695 spin_lock_init(&table->lock);
3698 /* Keep the first 32 indexes free for IOAPIC interrupts */
3699 table->min_index = 32;
3701 table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_ATOMIC);
3702 if (!table->table) {
3708 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
3709 memset(table->table, 0,
3710 MAX_IRQS_PER_TABLE * sizeof(u32));
3712 memset(table->table, 0,
3713 (MAX_IRQS_PER_TABLE * (sizeof(u64) * 2)));
3718 for (i = 0; i < 32; ++i)
3719 iommu->irte_ops->set_allocated(table, i);
3722 irq_lookup_table[devid] = table;
3723 set_dte_irq_entry(devid, table);
3724 iommu_flush_dte(iommu, devid);
3725 if (devid != alias) {
3726 irq_lookup_table[alias] = table;
3727 set_dte_irq_entry(alias, table);
3728 iommu_flush_dte(iommu, alias);
3732 iommu_completion_wait(iommu);
3735 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3740 static int alloc_irq_index(u16 devid, int count)
3742 struct irq_remap_table *table;
3743 unsigned long flags;
3745 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
3750 table = get_irq_table(devid, false);
3754 spin_lock_irqsave(&table->lock, flags);
3756 /* Scan table for free entries */
3757 for (c = 0, index = table->min_index;
3758 index < MAX_IRQS_PER_TABLE;
3760 if (!iommu->irte_ops->is_allocated(table, index))
3767 iommu->irte_ops->set_allocated(table, index - c + 1);
3777 spin_unlock_irqrestore(&table->lock, flags);
3782 static int modify_irte_ga(u16 devid, int index, struct irte_ga *irte,
3783 struct amd_ir_data *data)
3785 struct irq_remap_table *table;
3786 struct amd_iommu *iommu;
3787 unsigned long flags;
3788 struct irte_ga *entry;
3790 iommu = amd_iommu_rlookup_table[devid];
3794 table = get_irq_table(devid, false);
3798 spin_lock_irqsave(&table->lock, flags);
3800 entry = (struct irte_ga *)table->table;
3801 entry = &entry[index];
3802 entry->lo.fields_remap.valid = 0;
3803 entry->hi.val = irte->hi.val;
3804 entry->lo.val = irte->lo.val;
3805 entry->lo.fields_remap.valid = 1;
3809 spin_unlock_irqrestore(&table->lock, flags);
3811 iommu_flush_irt(iommu, devid);
3812 iommu_completion_wait(iommu);
3817 static int modify_irte(u16 devid, int index, union irte *irte)
3819 struct irq_remap_table *table;
3820 struct amd_iommu *iommu;
3821 unsigned long flags;
3823 iommu = amd_iommu_rlookup_table[devid];
3827 table = get_irq_table(devid, false);
3831 spin_lock_irqsave(&table->lock, flags);
3832 table->table[index] = irte->val;
3833 spin_unlock_irqrestore(&table->lock, flags);
3835 iommu_flush_irt(iommu, devid);
3836 iommu_completion_wait(iommu);
3841 static void free_irte(u16 devid, int index)
3843 struct irq_remap_table *table;
3844 struct amd_iommu *iommu;
3845 unsigned long flags;
3847 iommu = amd_iommu_rlookup_table[devid];
3851 table = get_irq_table(devid, false);
3855 spin_lock_irqsave(&table->lock, flags);
3856 iommu->irte_ops->clear_allocated(table, index);
3857 spin_unlock_irqrestore(&table->lock, flags);
3859 iommu_flush_irt(iommu, devid);
3860 iommu_completion_wait(iommu);
3863 static void irte_prepare(void *entry,
3864 u32 delivery_mode, u32 dest_mode,
3865 u8 vector, u32 dest_apicid, int devid)
3867 union irte *irte = (union irte *) entry;
3870 irte->fields.vector = vector;
3871 irte->fields.int_type = delivery_mode;
3872 irte->fields.destination = dest_apicid;
3873 irte->fields.dm = dest_mode;
3874 irte->fields.valid = 1;
3877 static void irte_ga_prepare(void *entry,
3878 u32 delivery_mode, u32 dest_mode,
3879 u8 vector, u32 dest_apicid, int devid)
3881 struct irte_ga *irte = (struct irte_ga *) entry;
3882 struct iommu_dev_data *dev_data = search_dev_data(devid);
3886 irte->lo.fields_remap.guest_mode = dev_data ? dev_data->use_vapic : 0;
3887 irte->lo.fields_remap.int_type = delivery_mode;
3888 irte->lo.fields_remap.dm = dest_mode;
3889 irte->hi.fields.vector = vector;
3890 irte->lo.fields_remap.destination = dest_apicid;
3891 irte->lo.fields_remap.valid = 1;
3894 static void irte_activate(void *entry, u16 devid, u16 index)
3896 union irte *irte = (union irte *) entry;
3898 irte->fields.valid = 1;
3899 modify_irte(devid, index, irte);
3902 static void irte_ga_activate(void *entry, u16 devid, u16 index)
3904 struct irte_ga *irte = (struct irte_ga *) entry;
3906 irte->lo.fields_remap.valid = 1;
3907 modify_irte_ga(devid, index, irte, NULL);
3910 static void irte_deactivate(void *entry, u16 devid, u16 index)
3912 union irte *irte = (union irte *) entry;
3914 irte->fields.valid = 0;
3915 modify_irte(devid, index, irte);
3918 static void irte_ga_deactivate(void *entry, u16 devid, u16 index)
3920 struct irte_ga *irte = (struct irte_ga *) entry;
3922 irte->lo.fields_remap.valid = 0;
3923 modify_irte_ga(devid, index, irte, NULL);
3926 static void irte_set_affinity(void *entry, u16 devid, u16 index,
3927 u8 vector, u32 dest_apicid)
3929 union irte *irte = (union irte *) entry;
3931 irte->fields.vector = vector;
3932 irte->fields.destination = dest_apicid;
3933 modify_irte(devid, index, irte);
3936 static void irte_ga_set_affinity(void *entry, u16 devid, u16 index,
3937 u8 vector, u32 dest_apicid)
3939 struct irte_ga *irte = (struct irte_ga *) entry;
3940 struct iommu_dev_data *dev_data = search_dev_data(devid);
3942 if (!dev_data || !dev_data->use_vapic) {
3943 irte->hi.fields.vector = vector;
3944 irte->lo.fields_remap.destination = dest_apicid;
3945 irte->lo.fields_remap.guest_mode = 0;
3946 modify_irte_ga(devid, index, irte, NULL);
3950 #define IRTE_ALLOCATED (~1U)
3951 static void irte_set_allocated(struct irq_remap_table *table, int index)
3953 table->table[index] = IRTE_ALLOCATED;
3956 static void irte_ga_set_allocated(struct irq_remap_table *table, int index)
3958 struct irte_ga *ptr = (struct irte_ga *)table->table;
3959 struct irte_ga *irte = &ptr[index];
3961 memset(&irte->lo.val, 0, sizeof(u64));
3962 memset(&irte->hi.val, 0, sizeof(u64));
3963 irte->hi.fields.vector = 0xff;
3966 static bool irte_is_allocated(struct irq_remap_table *table, int index)
3968 union irte *ptr = (union irte *)table->table;
3969 union irte *irte = &ptr[index];
3971 return irte->val != 0;
3974 static bool irte_ga_is_allocated(struct irq_remap_table *table, int index)
3976 struct irte_ga *ptr = (struct irte_ga *)table->table;
3977 struct irte_ga *irte = &ptr[index];
3979 return irte->hi.fields.vector != 0;
3982 static void irte_clear_allocated(struct irq_remap_table *table, int index)
3984 table->table[index] = 0;
3987 static void irte_ga_clear_allocated(struct irq_remap_table *table, int index)
3989 struct irte_ga *ptr = (struct irte_ga *)table->table;
3990 struct irte_ga *irte = &ptr[index];
3992 memset(&irte->lo.val, 0, sizeof(u64));
3993 memset(&irte->hi.val, 0, sizeof(u64));
3996 static int get_devid(struct irq_alloc_info *info)
4000 switch (info->type) {
4001 case X86_IRQ_ALLOC_TYPE_IOAPIC:
4002 devid = get_ioapic_devid(info->ioapic_id);
4004 case X86_IRQ_ALLOC_TYPE_HPET:
4005 devid = get_hpet_devid(info->hpet_id);
4007 case X86_IRQ_ALLOC_TYPE_MSI:
4008 case X86_IRQ_ALLOC_TYPE_MSIX:
4009 devid = get_device_id(&info->msi_dev->dev);
4019 static struct irq_domain *get_ir_irq_domain(struct irq_alloc_info *info)
4021 struct amd_iommu *iommu;
4027 devid = get_devid(info);
4029 iommu = amd_iommu_rlookup_table[devid];
4031 return iommu->ir_domain;
4037 static struct irq_domain *get_irq_domain(struct irq_alloc_info *info)
4039 struct amd_iommu *iommu;
4045 switch (info->type) {
4046 case X86_IRQ_ALLOC_TYPE_MSI:
4047 case X86_IRQ_ALLOC_TYPE_MSIX:
4048 devid = get_device_id(&info->msi_dev->dev);
4052 iommu = amd_iommu_rlookup_table[devid];
4054 return iommu->msi_domain;
4063 struct irq_remap_ops amd_iommu_irq_ops = {
4064 .prepare = amd_iommu_prepare,
4065 .enable = amd_iommu_enable,
4066 .disable = amd_iommu_disable,
4067 .reenable = amd_iommu_reenable,
4068 .enable_faulting = amd_iommu_enable_faulting,
4069 .get_ir_irq_domain = get_ir_irq_domain,
4070 .get_irq_domain = get_irq_domain,
4073 static void irq_remapping_prepare_irte(struct amd_ir_data *data,
4074 struct irq_cfg *irq_cfg,
4075 struct irq_alloc_info *info,
4076 int devid, int index, int sub_handle)
4078 struct irq_2_irte *irte_info = &data->irq_2_irte;
4079 struct msi_msg *msg = &data->msi_entry;
4080 struct IO_APIC_route_entry *entry;
4081 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
4086 data->irq_2_irte.devid = devid;
4087 data->irq_2_irte.index = index + sub_handle;
4088 iommu->irte_ops->prepare(data->entry, apic->irq_delivery_mode,
4089 apic->irq_dest_mode, irq_cfg->vector,
4090 irq_cfg->dest_apicid, devid);
4092 switch (info->type) {
4093 case X86_IRQ_ALLOC_TYPE_IOAPIC:
4094 /* Setup IOAPIC entry */
4095 entry = info->ioapic_entry;
4096 info->ioapic_entry = NULL;
4097 memset(entry, 0, sizeof(*entry));
4098 entry->vector = index;
4100 entry->trigger = info->ioapic_trigger;
4101 entry->polarity = info->ioapic_polarity;
4102 /* Mask level triggered irqs. */
4103 if (info->ioapic_trigger)
4107 case X86_IRQ_ALLOC_TYPE_HPET:
4108 case X86_IRQ_ALLOC_TYPE_MSI:
4109 case X86_IRQ_ALLOC_TYPE_MSIX:
4110 msg->address_hi = MSI_ADDR_BASE_HI;
4111 msg->address_lo = MSI_ADDR_BASE_LO;
4112 msg->data = irte_info->index;
4121 struct amd_irte_ops irte_32_ops = {
4122 .prepare = irte_prepare,
4123 .activate = irte_activate,
4124 .deactivate = irte_deactivate,
4125 .set_affinity = irte_set_affinity,
4126 .set_allocated = irte_set_allocated,
4127 .is_allocated = irte_is_allocated,
4128 .clear_allocated = irte_clear_allocated,
4131 struct amd_irte_ops irte_128_ops = {
4132 .prepare = irte_ga_prepare,
4133 .activate = irte_ga_activate,
4134 .deactivate = irte_ga_deactivate,
4135 .set_affinity = irte_ga_set_affinity,
4136 .set_allocated = irte_ga_set_allocated,
4137 .is_allocated = irte_ga_is_allocated,
4138 .clear_allocated = irte_ga_clear_allocated,
4141 static int irq_remapping_alloc(struct irq_domain *domain, unsigned int virq,
4142 unsigned int nr_irqs, void *arg)
4144 struct irq_alloc_info *info = arg;
4145 struct irq_data *irq_data;
4146 struct amd_ir_data *data = NULL;
4147 struct irq_cfg *cfg;
4153 if (nr_irqs > 1 && info->type != X86_IRQ_ALLOC_TYPE_MSI &&
4154 info->type != X86_IRQ_ALLOC_TYPE_MSIX)
4158 * With IRQ remapping enabled, don't need contiguous CPU vectors
4159 * to support multiple MSI interrupts.
4161 if (info->type == X86_IRQ_ALLOC_TYPE_MSI)
4162 info->flags &= ~X86_IRQ_ALLOC_CONTIGUOUS_VECTORS;
4164 devid = get_devid(info);
4168 ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
4172 if (info->type == X86_IRQ_ALLOC_TYPE_IOAPIC) {
4173 if (get_irq_table(devid, true))
4174 index = info->ioapic_pin;
4178 index = alloc_irq_index(devid, nr_irqs);
4181 pr_warn("Failed to allocate IRTE\n");
4183 goto out_free_parent;
4186 for (i = 0; i < nr_irqs; i++) {
4187 irq_data = irq_domain_get_irq_data(domain, virq + i);
4188 cfg = irqd_cfg(irq_data);
4189 if (!irq_data || !cfg) {
4195 data = kzalloc(sizeof(*data), GFP_KERNEL);
4199 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
4200 data->entry = kzalloc(sizeof(union irte), GFP_KERNEL);
4202 data->entry = kzalloc(sizeof(struct irte_ga),
4209 irq_data->hwirq = (devid << 16) + i;
4210 irq_data->chip_data = data;
4211 irq_data->chip = &amd_ir_chip;
4212 irq_remapping_prepare_irte(data, cfg, info, devid, index, i);
4213 irq_set_status_flags(virq + i, IRQ_MOVE_PCNTXT);
4219 for (i--; i >= 0; i--) {
4220 irq_data = irq_domain_get_irq_data(domain, virq + i);
4222 kfree(irq_data->chip_data);
4224 for (i = 0; i < nr_irqs; i++)
4225 free_irte(devid, index + i);
4227 irq_domain_free_irqs_common(domain, virq, nr_irqs);
4231 static void irq_remapping_free(struct irq_domain *domain, unsigned int virq,
4232 unsigned int nr_irqs)
4234 struct irq_2_irte *irte_info;
4235 struct irq_data *irq_data;
4236 struct amd_ir_data *data;
4239 for (i = 0; i < nr_irqs; i++) {
4240 irq_data = irq_domain_get_irq_data(domain, virq + i);
4241 if (irq_data && irq_data->chip_data) {
4242 data = irq_data->chip_data;
4243 irte_info = &data->irq_2_irte;
4244 free_irte(irte_info->devid, irte_info->index);
4249 irq_domain_free_irqs_common(domain, virq, nr_irqs);
4252 static void irq_remapping_activate(struct irq_domain *domain,
4253 struct irq_data *irq_data)
4255 struct amd_ir_data *data = irq_data->chip_data;
4256 struct irq_2_irte *irte_info = &data->irq_2_irte;
4257 struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4260 iommu->irte_ops->activate(data->entry, irte_info->devid,
4264 static void irq_remapping_deactivate(struct irq_domain *domain,
4265 struct irq_data *irq_data)
4267 struct amd_ir_data *data = irq_data->chip_data;
4268 struct irq_2_irte *irte_info = &data->irq_2_irte;
4269 struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4272 iommu->irte_ops->deactivate(data->entry, irte_info->devid,
4276 static struct irq_domain_ops amd_ir_domain_ops = {
4277 .alloc = irq_remapping_alloc,
4278 .free = irq_remapping_free,
4279 .activate = irq_remapping_activate,
4280 .deactivate = irq_remapping_deactivate,
4283 static int amd_ir_set_vcpu_affinity(struct irq_data *data, void *vcpu_info)
4285 struct amd_iommu *iommu;
4286 struct amd_iommu_pi_data *pi_data = vcpu_info;
4287 struct vcpu_data *vcpu_pi_info = pi_data->vcpu_data;
4288 struct amd_ir_data *ir_data = data->chip_data;
4289 struct irte_ga *irte = (struct irte_ga *) ir_data->entry;
4290 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
4291 struct iommu_dev_data *dev_data = search_dev_data(irte_info->devid);
4294 * This device has never been set up for guest mode.
4295 * we should not modify the IRTE
4297 if (!dev_data || !dev_data->use_vapic)
4300 pi_data->ir_data = ir_data;
4303 * SVM tries to set up for VAPIC mode, but we are in
4304 * legacy mode. So, we force legacy mode instead.
4306 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
4307 pr_debug("AMD-Vi: %s: Fall back to using intr legacy remap\n",
4309 pi_data->is_guest_mode = false;
4312 iommu = amd_iommu_rlookup_table[irte_info->devid];
4316 pi_data->prev_ga_tag = ir_data->cached_ga_tag;
4317 if (pi_data->is_guest_mode) {
4319 irte->hi.fields.ga_root_ptr = (pi_data->base >> 12);
4320 irte->hi.fields.vector = vcpu_pi_info->vector;
4321 irte->lo.fields_vapic.guest_mode = 1;
4322 irte->lo.fields_vapic.ga_tag = pi_data->ga_tag;
4324 ir_data->cached_ga_tag = pi_data->ga_tag;
4327 struct irq_cfg *cfg = irqd_cfg(data);
4331 irte->hi.fields.vector = cfg->vector;
4332 irte->lo.fields_remap.guest_mode = 0;
4333 irte->lo.fields_remap.destination = cfg->dest_apicid;
4334 irte->lo.fields_remap.int_type = apic->irq_delivery_mode;
4335 irte->lo.fields_remap.dm = apic->irq_dest_mode;
4338 * This communicates the ga_tag back to the caller
4339 * so that it can do all the necessary clean up.
4341 ir_data->cached_ga_tag = 0;
4344 return modify_irte_ga(irte_info->devid, irte_info->index, irte, ir_data);
4347 static int amd_ir_set_affinity(struct irq_data *data,
4348 const struct cpumask *mask, bool force)
4350 struct amd_ir_data *ir_data = data->chip_data;
4351 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
4352 struct irq_cfg *cfg = irqd_cfg(data);
4353 struct irq_data *parent = data->parent_data;
4354 struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4360 ret = parent->chip->irq_set_affinity(parent, mask, force);
4361 if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
4365 * Atomically updates the IRTE with the new destination, vector
4366 * and flushes the interrupt entry cache.
4368 iommu->irte_ops->set_affinity(ir_data->entry, irte_info->devid,
4369 irte_info->index, cfg->vector, cfg->dest_apicid);
4372 * After this point, all the interrupts will start arriving
4373 * at the new destination. So, time to cleanup the previous
4374 * vector allocation.
4376 send_cleanup_vector(cfg);
4378 return IRQ_SET_MASK_OK_DONE;
4381 static void ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg)
4383 struct amd_ir_data *ir_data = irq_data->chip_data;
4385 *msg = ir_data->msi_entry;
4388 static struct irq_chip amd_ir_chip = {
4390 .irq_ack = ir_ack_apic_edge,
4391 .irq_set_affinity = amd_ir_set_affinity,
4392 .irq_set_vcpu_affinity = amd_ir_set_vcpu_affinity,
4393 .irq_compose_msi_msg = ir_compose_msi_msg,
4396 int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
4398 struct fwnode_handle *fn;
4400 fn = irq_domain_alloc_named_id_fwnode("AMD-IR", iommu->index);
4403 iommu->ir_domain = irq_domain_create_tree(fn, &amd_ir_domain_ops, iommu);
4404 irq_domain_free_fwnode(fn);
4405 if (!iommu->ir_domain)
4408 iommu->ir_domain->parent = arch_get_ir_parent_domain();
4409 iommu->msi_domain = arch_create_remap_msi_irq_domain(iommu->ir_domain,
4415 int amd_iommu_update_ga(int cpu, bool is_run, void *data)
4417 unsigned long flags;
4418 struct amd_iommu *iommu;
4419 struct irq_remap_table *irt;
4420 struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
4421 int devid = ir_data->irq_2_irte.devid;
4422 struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
4423 struct irte_ga *ref = (struct irte_ga *) ir_data->ref;
4425 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
4426 !ref || !entry || !entry->lo.fields_vapic.guest_mode)
4429 iommu = amd_iommu_rlookup_table[devid];
4433 irt = get_irq_table(devid, false);
4437 spin_lock_irqsave(&irt->lock, flags);
4439 if (ref->lo.fields_vapic.guest_mode) {
4441 ref->lo.fields_vapic.destination = cpu;
4442 ref->lo.fields_vapic.is_run = is_run;
4446 spin_unlock_irqrestore(&irt->lock, flags);
4448 iommu_flush_irt(iommu, devid);
4449 iommu_completion_wait(iommu);
4452 EXPORT_SYMBOL(amd_iommu_update_ga);