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 static const struct iommu_ops amd_iommu_ops;
117 static ATOMIC_NOTIFIER_HEAD(ppr_notifier);
118 int amd_iommu_max_glx_val = -1;
120 static 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);
380 static bool pci_iommuv2_capable(struct pci_dev *pdev)
382 static const int caps[] = {
385 PCI_EXT_CAP_ID_PASID,
389 for (i = 0; i < 3; ++i) {
390 pos = pci_find_ext_capability(pdev, caps[i]);
398 static bool pdev_pri_erratum(struct pci_dev *pdev, u32 erratum)
400 struct iommu_dev_data *dev_data;
402 dev_data = get_dev_data(&pdev->dev);
404 return dev_data->errata & (1 << erratum) ? true : false;
408 * This function checks if the driver got a valid device from the caller to
409 * avoid dereferencing invalid pointers.
411 static bool check_device(struct device *dev)
415 if (!dev || !dev->dma_mask)
418 devid = get_device_id(dev);
422 /* Out of our scope? */
423 if (devid > amd_iommu_last_bdf)
426 if (amd_iommu_rlookup_table[devid] == NULL)
432 static void init_iommu_group(struct device *dev)
434 struct iommu_group *group;
436 group = iommu_group_get_for_dev(dev);
440 iommu_group_put(group);
443 static int iommu_init_device(struct device *dev)
445 struct iommu_dev_data *dev_data;
448 if (dev->archdata.iommu)
451 devid = get_device_id(dev);
455 dev_data = find_dev_data(devid);
459 dev_data->alias = get_alias(dev);
461 if (dev_is_pci(dev) && pci_iommuv2_capable(to_pci_dev(dev))) {
462 struct amd_iommu *iommu;
464 iommu = amd_iommu_rlookup_table[dev_data->devid];
465 dev_data->iommu_v2 = iommu->is_iommu_v2;
468 dev->archdata.iommu = dev_data;
470 iommu_device_link(amd_iommu_rlookup_table[dev_data->devid]->iommu_dev,
476 static void iommu_ignore_device(struct device *dev)
481 devid = get_device_id(dev);
485 alias = get_alias(dev);
487 memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
488 memset(&amd_iommu_dev_table[alias], 0, sizeof(struct dev_table_entry));
490 amd_iommu_rlookup_table[devid] = NULL;
491 amd_iommu_rlookup_table[alias] = NULL;
494 static void iommu_uninit_device(struct device *dev)
497 struct iommu_dev_data *dev_data;
499 devid = get_device_id(dev);
503 dev_data = search_dev_data(devid);
507 if (dev_data->domain)
510 iommu_device_unlink(amd_iommu_rlookup_table[dev_data->devid]->iommu_dev,
513 iommu_group_remove_device(dev);
516 dev->archdata.dma_ops = NULL;
519 * We keep dev_data around for unplugged devices and reuse it when the
520 * device is re-plugged - not doing so would introduce a ton of races.
524 /****************************************************************************
526 * Interrupt handling functions
528 ****************************************************************************/
530 static void dump_dte_entry(u16 devid)
534 for (i = 0; i < 4; ++i)
535 pr_err("AMD-Vi: DTE[%d]: %016llx\n", i,
536 amd_iommu_dev_table[devid].data[i]);
539 static void dump_command(unsigned long phys_addr)
541 struct iommu_cmd *cmd = phys_to_virt(phys_addr);
544 for (i = 0; i < 4; ++i)
545 pr_err("AMD-Vi: CMD[%d]: %08x\n", i, cmd->data[i]);
548 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
550 int type, devid, domid, flags;
551 volatile u32 *event = __evt;
556 type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
557 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
558 domid = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
559 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
560 address = (u64)(((u64)event[3]) << 32) | event[2];
563 /* Did we hit the erratum? */
564 if (++count == LOOP_TIMEOUT) {
565 pr_err("AMD-Vi: No event written to event log\n");
572 printk(KERN_ERR "AMD-Vi: Event logged [");
575 case EVENT_TYPE_ILL_DEV:
576 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
577 "address=0x%016llx flags=0x%04x]\n",
578 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
580 dump_dte_entry(devid);
582 case EVENT_TYPE_IO_FAULT:
583 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
584 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
585 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
586 domid, address, flags);
588 case EVENT_TYPE_DEV_TAB_ERR:
589 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
590 "address=0x%016llx flags=0x%04x]\n",
591 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
594 case EVENT_TYPE_PAGE_TAB_ERR:
595 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
596 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
597 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
598 domid, address, flags);
600 case EVENT_TYPE_ILL_CMD:
601 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
602 dump_command(address);
604 case EVENT_TYPE_CMD_HARD_ERR:
605 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
606 "flags=0x%04x]\n", address, flags);
608 case EVENT_TYPE_IOTLB_INV_TO:
609 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
610 "address=0x%016llx]\n",
611 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
614 case EVENT_TYPE_INV_DEV_REQ:
615 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
616 "address=0x%016llx flags=0x%04x]\n",
617 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
621 printk(KERN_ERR "UNKNOWN type=0x%02x]\n", type);
624 memset(__evt, 0, 4 * sizeof(u32));
627 static void iommu_poll_events(struct amd_iommu *iommu)
631 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
632 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
634 while (head != tail) {
635 iommu_print_event(iommu, iommu->evt_buf + head);
636 head = (head + EVENT_ENTRY_SIZE) % EVT_BUFFER_SIZE;
639 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
642 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
644 struct amd_iommu_fault fault;
646 if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
647 pr_err_ratelimited("AMD-Vi: Unknown PPR request received\n");
651 fault.address = raw[1];
652 fault.pasid = PPR_PASID(raw[0]);
653 fault.device_id = PPR_DEVID(raw[0]);
654 fault.tag = PPR_TAG(raw[0]);
655 fault.flags = PPR_FLAGS(raw[0]);
657 atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
660 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
664 if (iommu->ppr_log == NULL)
667 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
668 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
670 while (head != tail) {
675 raw = (u64 *)(iommu->ppr_log + head);
678 * Hardware bug: Interrupt may arrive before the entry is
679 * written to memory. If this happens we need to wait for the
682 for (i = 0; i < LOOP_TIMEOUT; ++i) {
683 if (PPR_REQ_TYPE(raw[0]) != 0)
688 /* Avoid memcpy function-call overhead */
693 * To detect the hardware bug we need to clear the entry
696 raw[0] = raw[1] = 0UL;
698 /* Update head pointer of hardware ring-buffer */
699 head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
700 writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
702 /* Handle PPR entry */
703 iommu_handle_ppr_entry(iommu, entry);
705 /* Refresh ring-buffer information */
706 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
707 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
711 #ifdef CONFIG_IRQ_REMAP
712 static int (*iommu_ga_log_notifier)(u32);
714 int amd_iommu_register_ga_log_notifier(int (*notifier)(u32))
716 iommu_ga_log_notifier = notifier;
720 EXPORT_SYMBOL(amd_iommu_register_ga_log_notifier);
722 static void iommu_poll_ga_log(struct amd_iommu *iommu)
724 u32 head, tail, cnt = 0;
726 if (iommu->ga_log == NULL)
729 head = readl(iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
730 tail = readl(iommu->mmio_base + MMIO_GA_TAIL_OFFSET);
732 while (head != tail) {
736 raw = (u64 *)(iommu->ga_log + head);
739 /* Avoid memcpy function-call overhead */
742 /* Update head pointer of hardware ring-buffer */
743 head = (head + GA_ENTRY_SIZE) % GA_LOG_SIZE;
744 writel(head, iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
746 /* Handle GA entry */
747 switch (GA_REQ_TYPE(log_entry)) {
749 if (!iommu_ga_log_notifier)
752 pr_debug("AMD-Vi: %s: devid=%#x, ga_tag=%#x\n",
753 __func__, GA_DEVID(log_entry),
756 if (iommu_ga_log_notifier(GA_TAG(log_entry)) != 0)
757 pr_err("AMD-Vi: GA log notifier failed.\n");
764 #endif /* CONFIG_IRQ_REMAP */
766 #define AMD_IOMMU_INT_MASK \
767 (MMIO_STATUS_EVT_INT_MASK | \
768 MMIO_STATUS_PPR_INT_MASK | \
769 MMIO_STATUS_GALOG_INT_MASK)
771 irqreturn_t amd_iommu_int_thread(int irq, void *data)
773 struct amd_iommu *iommu = (struct amd_iommu *) data;
774 u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
776 while (status & AMD_IOMMU_INT_MASK) {
777 /* Enable EVT and PPR and GA interrupts again */
778 writel(AMD_IOMMU_INT_MASK,
779 iommu->mmio_base + MMIO_STATUS_OFFSET);
781 if (status & MMIO_STATUS_EVT_INT_MASK) {
782 pr_devel("AMD-Vi: Processing IOMMU Event Log\n");
783 iommu_poll_events(iommu);
786 if (status & MMIO_STATUS_PPR_INT_MASK) {
787 pr_devel("AMD-Vi: Processing IOMMU PPR Log\n");
788 iommu_poll_ppr_log(iommu);
791 #ifdef CONFIG_IRQ_REMAP
792 if (status & MMIO_STATUS_GALOG_INT_MASK) {
793 pr_devel("AMD-Vi: Processing IOMMU GA Log\n");
794 iommu_poll_ga_log(iommu);
799 * Hardware bug: ERBT1312
800 * When re-enabling interrupt (by writing 1
801 * to clear the bit), the hardware might also try to set
802 * the interrupt bit in the event status register.
803 * In this scenario, the bit will be set, and disable
804 * subsequent interrupts.
806 * Workaround: The IOMMU driver should read back the
807 * status register and check if the interrupt bits are cleared.
808 * If not, driver will need to go through the interrupt handler
809 * again and re-clear the bits
811 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
816 irqreturn_t amd_iommu_int_handler(int irq, void *data)
818 return IRQ_WAKE_THREAD;
821 /****************************************************************************
823 * IOMMU command queuing functions
825 ****************************************************************************/
827 static int wait_on_sem(volatile u64 *sem)
831 while (*sem == 0 && i < LOOP_TIMEOUT) {
836 if (i == LOOP_TIMEOUT) {
837 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
844 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
845 struct iommu_cmd *cmd,
850 target = iommu->cmd_buf + tail;
851 tail = (tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
853 /* Copy command to buffer */
854 memcpy(target, cmd, sizeof(*cmd));
856 /* Tell the IOMMU about it */
857 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
860 static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
862 WARN_ON(address & 0x7ULL);
864 memset(cmd, 0, sizeof(*cmd));
865 cmd->data[0] = lower_32_bits(__pa(address)) | CMD_COMPL_WAIT_STORE_MASK;
866 cmd->data[1] = upper_32_bits(__pa(address));
868 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
871 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
873 memset(cmd, 0, sizeof(*cmd));
874 cmd->data[0] = devid;
875 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
878 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
879 size_t size, u16 domid, int pde)
884 pages = iommu_num_pages(address, size, PAGE_SIZE);
889 * If we have to flush more than one page, flush all
890 * TLB entries for this domain
892 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
896 address &= PAGE_MASK;
898 memset(cmd, 0, sizeof(*cmd));
899 cmd->data[1] |= domid;
900 cmd->data[2] = lower_32_bits(address);
901 cmd->data[3] = upper_32_bits(address);
902 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
903 if (s) /* size bit - we flush more than one 4kb page */
904 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
905 if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
906 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
909 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
910 u64 address, size_t size)
915 pages = iommu_num_pages(address, size, PAGE_SIZE);
920 * If we have to flush more than one page, flush all
921 * TLB entries for this domain
923 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
927 address &= PAGE_MASK;
929 memset(cmd, 0, sizeof(*cmd));
930 cmd->data[0] = devid;
931 cmd->data[0] |= (qdep & 0xff) << 24;
932 cmd->data[1] = devid;
933 cmd->data[2] = lower_32_bits(address);
934 cmd->data[3] = upper_32_bits(address);
935 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
937 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
940 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, int pasid,
941 u64 address, bool size)
943 memset(cmd, 0, sizeof(*cmd));
945 address &= ~(0xfffULL);
947 cmd->data[0] = pasid;
948 cmd->data[1] = domid;
949 cmd->data[2] = lower_32_bits(address);
950 cmd->data[3] = upper_32_bits(address);
951 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
952 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
954 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
955 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
958 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, int pasid,
959 int qdep, u64 address, bool size)
961 memset(cmd, 0, sizeof(*cmd));
963 address &= ~(0xfffULL);
965 cmd->data[0] = devid;
966 cmd->data[0] |= ((pasid >> 8) & 0xff) << 16;
967 cmd->data[0] |= (qdep & 0xff) << 24;
968 cmd->data[1] = devid;
969 cmd->data[1] |= (pasid & 0xff) << 16;
970 cmd->data[2] = lower_32_bits(address);
971 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
972 cmd->data[3] = upper_32_bits(address);
974 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
975 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
978 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, int pasid,
979 int status, int tag, bool gn)
981 memset(cmd, 0, sizeof(*cmd));
983 cmd->data[0] = devid;
985 cmd->data[1] = pasid;
986 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
988 cmd->data[3] = tag & 0x1ff;
989 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
991 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
994 static void build_inv_all(struct iommu_cmd *cmd)
996 memset(cmd, 0, sizeof(*cmd));
997 CMD_SET_TYPE(cmd, CMD_INV_ALL);
1000 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
1002 memset(cmd, 0, sizeof(*cmd));
1003 cmd->data[0] = devid;
1004 CMD_SET_TYPE(cmd, CMD_INV_IRT);
1008 * Writes the command to the IOMMUs command buffer and informs the
1009 * hardware about the new command.
1011 static int iommu_queue_command_sync(struct amd_iommu *iommu,
1012 struct iommu_cmd *cmd,
1015 u32 left, tail, head, next_tail;
1016 unsigned long flags;
1019 spin_lock_irqsave(&iommu->lock, flags);
1021 head = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
1022 tail = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
1023 next_tail = (tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
1024 left = (head - next_tail) % CMD_BUFFER_SIZE;
1027 struct iommu_cmd sync_cmd;
1028 volatile u64 sem = 0;
1031 build_completion_wait(&sync_cmd, (u64)&sem);
1032 copy_cmd_to_buffer(iommu, &sync_cmd, tail);
1034 spin_unlock_irqrestore(&iommu->lock, flags);
1036 if ((ret = wait_on_sem(&sem)) != 0)
1042 copy_cmd_to_buffer(iommu, cmd, tail);
1044 /* We need to sync now to make sure all commands are processed */
1045 iommu->need_sync = sync;
1047 spin_unlock_irqrestore(&iommu->lock, flags);
1052 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
1054 return iommu_queue_command_sync(iommu, cmd, true);
1058 * This function queues a completion wait command into the command
1059 * buffer of an IOMMU
1061 static int iommu_completion_wait(struct amd_iommu *iommu)
1063 struct iommu_cmd cmd;
1064 volatile u64 sem = 0;
1067 if (!iommu->need_sync)
1070 build_completion_wait(&cmd, (u64)&sem);
1072 ret = iommu_queue_command_sync(iommu, &cmd, false);
1076 return wait_on_sem(&sem);
1079 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
1081 struct iommu_cmd cmd;
1083 build_inv_dte(&cmd, devid);
1085 return iommu_queue_command(iommu, &cmd);
1088 static void iommu_flush_dte_all(struct amd_iommu *iommu)
1092 for (devid = 0; devid <= 0xffff; ++devid)
1093 iommu_flush_dte(iommu, devid);
1095 iommu_completion_wait(iommu);
1099 * This function uses heavy locking and may disable irqs for some time. But
1100 * this is no issue because it is only called during resume.
1102 static void iommu_flush_tlb_all(struct amd_iommu *iommu)
1106 for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
1107 struct iommu_cmd cmd;
1108 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1110 iommu_queue_command(iommu, &cmd);
1113 iommu_completion_wait(iommu);
1116 static void iommu_flush_all(struct amd_iommu *iommu)
1118 struct iommu_cmd cmd;
1120 build_inv_all(&cmd);
1122 iommu_queue_command(iommu, &cmd);
1123 iommu_completion_wait(iommu);
1126 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1128 struct iommu_cmd cmd;
1130 build_inv_irt(&cmd, devid);
1132 iommu_queue_command(iommu, &cmd);
1135 static void iommu_flush_irt_all(struct amd_iommu *iommu)
1139 for (devid = 0; devid <= MAX_DEV_TABLE_ENTRIES; devid++)
1140 iommu_flush_irt(iommu, devid);
1142 iommu_completion_wait(iommu);
1145 void iommu_flush_all_caches(struct amd_iommu *iommu)
1147 if (iommu_feature(iommu, FEATURE_IA)) {
1148 iommu_flush_all(iommu);
1150 iommu_flush_dte_all(iommu);
1151 iommu_flush_irt_all(iommu);
1152 iommu_flush_tlb_all(iommu);
1157 * Command send function for flushing on-device TLB
1159 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1160 u64 address, size_t size)
1162 struct amd_iommu *iommu;
1163 struct iommu_cmd cmd;
1166 qdep = dev_data->ats.qdep;
1167 iommu = amd_iommu_rlookup_table[dev_data->devid];
1169 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1171 return iommu_queue_command(iommu, &cmd);
1175 * Command send function for invalidating a device table entry
1177 static int device_flush_dte(struct iommu_dev_data *dev_data)
1179 struct amd_iommu *iommu;
1183 iommu = amd_iommu_rlookup_table[dev_data->devid];
1184 alias = dev_data->alias;
1186 ret = iommu_flush_dte(iommu, dev_data->devid);
1187 if (!ret && alias != dev_data->devid)
1188 ret = iommu_flush_dte(iommu, alias);
1192 if (dev_data->ats.enabled)
1193 ret = device_flush_iotlb(dev_data, 0, ~0UL);
1199 * TLB invalidation function which is called from the mapping functions.
1200 * It invalidates a single PTE if the range to flush is within a single
1201 * page. Otherwise it flushes the whole TLB of the IOMMU.
1203 static void __domain_flush_pages(struct protection_domain *domain,
1204 u64 address, size_t size, int pde)
1206 struct iommu_dev_data *dev_data;
1207 struct iommu_cmd cmd;
1210 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1212 for (i = 0; i < amd_iommus_present; ++i) {
1213 if (!domain->dev_iommu[i])
1217 * Devices of this domain are behind this IOMMU
1218 * We need a TLB flush
1220 ret |= iommu_queue_command(amd_iommus[i], &cmd);
1223 list_for_each_entry(dev_data, &domain->dev_list, list) {
1225 if (!dev_data->ats.enabled)
1228 ret |= device_flush_iotlb(dev_data, address, size);
1234 static void domain_flush_pages(struct protection_domain *domain,
1235 u64 address, size_t size)
1237 __domain_flush_pages(domain, address, size, 0);
1240 /* Flush the whole IO/TLB for a given protection domain */
1241 static void domain_flush_tlb(struct protection_domain *domain)
1243 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
1246 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1247 static void domain_flush_tlb_pde(struct protection_domain *domain)
1249 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1252 static void domain_flush_complete(struct protection_domain *domain)
1256 for (i = 0; i < amd_iommus_present; ++i) {
1257 if (domain && !domain->dev_iommu[i])
1261 * Devices of this domain are behind this IOMMU
1262 * We need to wait for completion of all commands.
1264 iommu_completion_wait(amd_iommus[i]);
1270 * This function flushes the DTEs for all devices in domain
1272 static void domain_flush_devices(struct protection_domain *domain)
1274 struct iommu_dev_data *dev_data;
1276 list_for_each_entry(dev_data, &domain->dev_list, list)
1277 device_flush_dte(dev_data);
1280 /****************************************************************************
1282 * The functions below are used the create the page table mappings for
1283 * unity mapped regions.
1285 ****************************************************************************/
1288 * This function is used to add another level to an IO page table. Adding
1289 * another level increases the size of the address space by 9 bits to a size up
1292 static bool increase_address_space(struct protection_domain *domain,
1297 if (domain->mode == PAGE_MODE_6_LEVEL)
1298 /* address space already 64 bit large */
1301 pte = (void *)get_zeroed_page(gfp);
1305 *pte = PM_LEVEL_PDE(domain->mode,
1306 virt_to_phys(domain->pt_root));
1307 domain->pt_root = pte;
1309 domain->updated = true;
1314 static u64 *alloc_pte(struct protection_domain *domain,
1315 unsigned long address,
1316 unsigned long page_size,
1323 BUG_ON(!is_power_of_2(page_size));
1325 while (address > PM_LEVEL_SIZE(domain->mode))
1326 increase_address_space(domain, gfp);
1328 level = domain->mode - 1;
1329 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1330 address = PAGE_SIZE_ALIGN(address, page_size);
1331 end_lvl = PAGE_SIZE_LEVEL(page_size);
1333 while (level > end_lvl) {
1338 if (!IOMMU_PTE_PRESENT(__pte)) {
1339 page = (u64 *)get_zeroed_page(gfp);
1343 __npte = PM_LEVEL_PDE(level, virt_to_phys(page));
1345 /* pte could have been changed somewhere. */
1346 if (cmpxchg64(pte, __pte, __npte) != __pte) {
1347 free_page((unsigned long)page);
1352 /* No level skipping support yet */
1353 if (PM_PTE_LEVEL(*pte) != level)
1358 pte = IOMMU_PTE_PAGE(*pte);
1360 if (pte_page && level == end_lvl)
1363 pte = &pte[PM_LEVEL_INDEX(level, address)];
1370 * This function checks if there is a PTE for a given dma address. If
1371 * there is one, it returns the pointer to it.
1373 static u64 *fetch_pte(struct protection_domain *domain,
1374 unsigned long address,
1375 unsigned long *page_size)
1380 if (address > PM_LEVEL_SIZE(domain->mode))
1383 level = domain->mode - 1;
1384 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1385 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1390 if (!IOMMU_PTE_PRESENT(*pte))
1394 if (PM_PTE_LEVEL(*pte) == 7 ||
1395 PM_PTE_LEVEL(*pte) == 0)
1398 /* No level skipping support yet */
1399 if (PM_PTE_LEVEL(*pte) != level)
1404 /* Walk to the next level */
1405 pte = IOMMU_PTE_PAGE(*pte);
1406 pte = &pte[PM_LEVEL_INDEX(level, address)];
1407 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1410 if (PM_PTE_LEVEL(*pte) == 0x07) {
1411 unsigned long pte_mask;
1414 * If we have a series of large PTEs, make
1415 * sure to return a pointer to the first one.
1417 *page_size = pte_mask = PTE_PAGE_SIZE(*pte);
1418 pte_mask = ~((PAGE_SIZE_PTE_COUNT(pte_mask) << 3) - 1);
1419 pte = (u64 *)(((unsigned long)pte) & pte_mask);
1426 * Generic mapping functions. It maps a physical address into a DMA
1427 * address space. It allocates the page table pages if necessary.
1428 * In the future it can be extended to a generic mapping function
1429 * supporting all features of AMD IOMMU page tables like level skipping
1430 * and full 64 bit address spaces.
1432 static int iommu_map_page(struct protection_domain *dom,
1433 unsigned long bus_addr,
1434 unsigned long phys_addr,
1435 unsigned long page_size,
1442 BUG_ON(!IS_ALIGNED(bus_addr, page_size));
1443 BUG_ON(!IS_ALIGNED(phys_addr, page_size));
1445 if (!(prot & IOMMU_PROT_MASK))
1448 count = PAGE_SIZE_PTE_COUNT(page_size);
1449 pte = alloc_pte(dom, bus_addr, page_size, NULL, gfp);
1454 for (i = 0; i < count; ++i)
1455 if (IOMMU_PTE_PRESENT(pte[i]))
1459 __pte = PAGE_SIZE_PTE(phys_addr, page_size);
1460 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_P | IOMMU_PTE_FC;
1462 __pte = phys_addr | IOMMU_PTE_P | IOMMU_PTE_FC;
1464 if (prot & IOMMU_PROT_IR)
1465 __pte |= IOMMU_PTE_IR;
1466 if (prot & IOMMU_PROT_IW)
1467 __pte |= IOMMU_PTE_IW;
1469 for (i = 0; i < count; ++i)
1477 static unsigned long iommu_unmap_page(struct protection_domain *dom,
1478 unsigned long bus_addr,
1479 unsigned long page_size)
1481 unsigned long long unmapped;
1482 unsigned long unmap_size;
1485 BUG_ON(!is_power_of_2(page_size));
1489 while (unmapped < page_size) {
1491 pte = fetch_pte(dom, bus_addr, &unmap_size);
1496 count = PAGE_SIZE_PTE_COUNT(unmap_size);
1497 for (i = 0; i < count; i++)
1501 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1502 unmapped += unmap_size;
1505 BUG_ON(unmapped && !is_power_of_2(unmapped));
1510 /****************************************************************************
1512 * The next functions belong to the address allocator for the dma_ops
1513 * interface functions.
1515 ****************************************************************************/
1518 static unsigned long dma_ops_alloc_iova(struct device *dev,
1519 struct dma_ops_domain *dma_dom,
1520 unsigned int pages, u64 dma_mask)
1522 unsigned long pfn = 0;
1524 pages = __roundup_pow_of_two(pages);
1526 if (dma_mask > DMA_BIT_MASK(32))
1527 pfn = alloc_iova_fast(&dma_dom->iovad, pages,
1528 IOVA_PFN(DMA_BIT_MASK(32)));
1531 pfn = alloc_iova_fast(&dma_dom->iovad, pages, IOVA_PFN(dma_mask));
1533 return (pfn << PAGE_SHIFT);
1536 static void dma_ops_free_iova(struct dma_ops_domain *dma_dom,
1537 unsigned long address,
1540 pages = __roundup_pow_of_two(pages);
1541 address >>= PAGE_SHIFT;
1543 free_iova_fast(&dma_dom->iovad, address, pages);
1546 /****************************************************************************
1548 * The next functions belong to the domain allocation. A domain is
1549 * allocated for every IOMMU as the default domain. If device isolation
1550 * is enabled, every device get its own domain. The most important thing
1551 * about domains is the page table mapping the DMA address space they
1554 ****************************************************************************/
1557 * This function adds a protection domain to the global protection domain list
1559 static void add_domain_to_list(struct protection_domain *domain)
1561 unsigned long flags;
1563 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1564 list_add(&domain->list, &amd_iommu_pd_list);
1565 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1569 * This function removes a protection domain to the global
1570 * protection domain list
1572 static void del_domain_from_list(struct protection_domain *domain)
1574 unsigned long flags;
1576 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1577 list_del(&domain->list);
1578 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1581 static u16 domain_id_alloc(void)
1583 unsigned long flags;
1586 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1587 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1589 if (id > 0 && id < MAX_DOMAIN_ID)
1590 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1593 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1598 static void domain_id_free(int id)
1600 unsigned long flags;
1602 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1603 if (id > 0 && id < MAX_DOMAIN_ID)
1604 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1605 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1608 #define DEFINE_FREE_PT_FN(LVL, FN) \
1609 static void free_pt_##LVL (unsigned long __pt) \
1617 for (i = 0; i < 512; ++i) { \
1618 /* PTE present? */ \
1619 if (!IOMMU_PTE_PRESENT(pt[i])) \
1623 if (PM_PTE_LEVEL(pt[i]) == 0 || \
1624 PM_PTE_LEVEL(pt[i]) == 7) \
1627 p = (unsigned long)IOMMU_PTE_PAGE(pt[i]); \
1630 free_page((unsigned long)pt); \
1633 DEFINE_FREE_PT_FN(l2, free_page)
1634 DEFINE_FREE_PT_FN(l3, free_pt_l2)
1635 DEFINE_FREE_PT_FN(l4, free_pt_l3)
1636 DEFINE_FREE_PT_FN(l5, free_pt_l4)
1637 DEFINE_FREE_PT_FN(l6, free_pt_l5)
1639 static void free_pagetable(struct protection_domain *domain)
1641 unsigned long root = (unsigned long)domain->pt_root;
1643 switch (domain->mode) {
1644 case PAGE_MODE_NONE:
1646 case PAGE_MODE_1_LEVEL:
1649 case PAGE_MODE_2_LEVEL:
1652 case PAGE_MODE_3_LEVEL:
1655 case PAGE_MODE_4_LEVEL:
1658 case PAGE_MODE_5_LEVEL:
1661 case PAGE_MODE_6_LEVEL:
1669 static void free_gcr3_tbl_level1(u64 *tbl)
1674 for (i = 0; i < 512; ++i) {
1675 if (!(tbl[i] & GCR3_VALID))
1678 ptr = __va(tbl[i] & PAGE_MASK);
1680 free_page((unsigned long)ptr);
1684 static void free_gcr3_tbl_level2(u64 *tbl)
1689 for (i = 0; i < 512; ++i) {
1690 if (!(tbl[i] & GCR3_VALID))
1693 ptr = __va(tbl[i] & PAGE_MASK);
1695 free_gcr3_tbl_level1(ptr);
1699 static void free_gcr3_table(struct protection_domain *domain)
1701 if (domain->glx == 2)
1702 free_gcr3_tbl_level2(domain->gcr3_tbl);
1703 else if (domain->glx == 1)
1704 free_gcr3_tbl_level1(domain->gcr3_tbl);
1706 BUG_ON(domain->glx != 0);
1708 free_page((unsigned long)domain->gcr3_tbl);
1712 * Free a domain, only used if something went wrong in the
1713 * allocation path and we need to free an already allocated page table
1715 static void dma_ops_domain_free(struct dma_ops_domain *dom)
1720 del_domain_from_list(&dom->domain);
1722 put_iova_domain(&dom->iovad);
1724 free_pagetable(&dom->domain);
1727 domain_id_free(dom->domain.id);
1733 * Allocates a new protection domain usable for the dma_ops functions.
1734 * It also initializes the page table and the address allocator data
1735 * structures required for the dma_ops interface
1737 static struct dma_ops_domain *dma_ops_domain_alloc(void)
1739 struct dma_ops_domain *dma_dom;
1741 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
1745 if (protection_domain_init(&dma_dom->domain))
1748 dma_dom->domain.mode = PAGE_MODE_3_LEVEL;
1749 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
1750 dma_dom->domain.flags = PD_DMA_OPS_MASK;
1751 if (!dma_dom->domain.pt_root)
1754 init_iova_domain(&dma_dom->iovad, PAGE_SIZE,
1755 IOVA_START_PFN, DMA_32BIT_PFN);
1757 /* Initialize reserved ranges */
1758 copy_reserved_iova(&reserved_iova_ranges, &dma_dom->iovad);
1760 add_domain_to_list(&dma_dom->domain);
1765 dma_ops_domain_free(dma_dom);
1771 * little helper function to check whether a given protection domain is a
1774 static bool dma_ops_domain(struct protection_domain *domain)
1776 return domain->flags & PD_DMA_OPS_MASK;
1779 static void set_dte_entry(u16 devid, struct protection_domain *domain, bool ats)
1784 if (domain->mode != PAGE_MODE_NONE)
1785 pte_root = virt_to_phys(domain->pt_root);
1787 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
1788 << DEV_ENTRY_MODE_SHIFT;
1789 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;
1791 flags = amd_iommu_dev_table[devid].data[1];
1794 flags |= DTE_FLAG_IOTLB;
1796 if (domain->flags & PD_IOMMUV2_MASK) {
1797 u64 gcr3 = __pa(domain->gcr3_tbl);
1798 u64 glx = domain->glx;
1801 pte_root |= DTE_FLAG_GV;
1802 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
1804 /* First mask out possible old values for GCR3 table */
1805 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
1808 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
1811 /* Encode GCR3 table into DTE */
1812 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
1815 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
1818 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
1822 flags &= ~(0xffffUL);
1823 flags |= domain->id;
1825 amd_iommu_dev_table[devid].data[1] = flags;
1826 amd_iommu_dev_table[devid].data[0] = pte_root;
1829 static void clear_dte_entry(u16 devid)
1831 /* remove entry from the device table seen by the hardware */
1832 amd_iommu_dev_table[devid].data[0] = IOMMU_PTE_P | IOMMU_PTE_TV;
1833 amd_iommu_dev_table[devid].data[1] &= DTE_FLAG_MASK;
1835 amd_iommu_apply_erratum_63(devid);
1838 static void do_attach(struct iommu_dev_data *dev_data,
1839 struct protection_domain *domain)
1841 struct amd_iommu *iommu;
1845 iommu = amd_iommu_rlookup_table[dev_data->devid];
1846 alias = dev_data->alias;
1847 ats = dev_data->ats.enabled;
1849 /* Update data structures */
1850 dev_data->domain = domain;
1851 list_add(&dev_data->list, &domain->dev_list);
1853 /* Do reference counting */
1854 domain->dev_iommu[iommu->index] += 1;
1855 domain->dev_cnt += 1;
1857 /* Update device table */
1858 set_dte_entry(dev_data->devid, domain, ats);
1859 if (alias != dev_data->devid)
1860 set_dte_entry(alias, domain, ats);
1862 device_flush_dte(dev_data);
1865 static void do_detach(struct iommu_dev_data *dev_data)
1867 struct amd_iommu *iommu;
1871 * First check if the device is still attached. It might already
1872 * be detached from its domain because the generic
1873 * iommu_detach_group code detached it and we try again here in
1874 * our alias handling.
1876 if (!dev_data->domain)
1879 iommu = amd_iommu_rlookup_table[dev_data->devid];
1880 alias = dev_data->alias;
1882 /* decrease reference counters */
1883 dev_data->domain->dev_iommu[iommu->index] -= 1;
1884 dev_data->domain->dev_cnt -= 1;
1886 /* Update data structures */
1887 dev_data->domain = NULL;
1888 list_del(&dev_data->list);
1889 clear_dte_entry(dev_data->devid);
1890 if (alias != dev_data->devid)
1891 clear_dte_entry(alias);
1893 /* Flush the DTE entry */
1894 device_flush_dte(dev_data);
1898 * If a device is not yet associated with a domain, this function does
1899 * assigns it visible for the hardware
1901 static int __attach_device(struct iommu_dev_data *dev_data,
1902 struct protection_domain *domain)
1907 * Must be called with IRQs disabled. Warn here to detect early
1910 WARN_ON(!irqs_disabled());
1913 spin_lock(&domain->lock);
1916 if (dev_data->domain != NULL)
1919 /* Attach alias group root */
1920 do_attach(dev_data, domain);
1927 spin_unlock(&domain->lock);
1933 static void pdev_iommuv2_disable(struct pci_dev *pdev)
1935 pci_disable_ats(pdev);
1936 pci_disable_pri(pdev);
1937 pci_disable_pasid(pdev);
1940 /* FIXME: Change generic reset-function to do the same */
1941 static int pri_reset_while_enabled(struct pci_dev *pdev)
1946 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
1950 pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
1951 control |= PCI_PRI_CTRL_RESET;
1952 pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
1957 static int pdev_iommuv2_enable(struct pci_dev *pdev)
1962 /* FIXME: Hardcode number of outstanding requests for now */
1964 if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
1966 reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
1968 /* Only allow access to user-accessible pages */
1969 ret = pci_enable_pasid(pdev, 0);
1973 /* First reset the PRI state of the device */
1974 ret = pci_reset_pri(pdev);
1979 ret = pci_enable_pri(pdev, reqs);
1984 ret = pri_reset_while_enabled(pdev);
1989 ret = pci_enable_ats(pdev, PAGE_SHIFT);
1996 pci_disable_pri(pdev);
1997 pci_disable_pasid(pdev);
2002 /* FIXME: Move this to PCI code */
2003 #define PCI_PRI_TLP_OFF (1 << 15)
2005 static bool pci_pri_tlp_required(struct pci_dev *pdev)
2010 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2014 pci_read_config_word(pdev, pos + PCI_PRI_STATUS, &status);
2016 return (status & PCI_PRI_TLP_OFF) ? true : false;
2020 * If a device is not yet associated with a domain, this function
2021 * assigns it visible for the hardware
2023 static int attach_device(struct device *dev,
2024 struct protection_domain *domain)
2026 struct pci_dev *pdev;
2027 struct iommu_dev_data *dev_data;
2028 unsigned long flags;
2031 dev_data = get_dev_data(dev);
2033 if (!dev_is_pci(dev))
2034 goto skip_ats_check;
2036 pdev = to_pci_dev(dev);
2037 if (domain->flags & PD_IOMMUV2_MASK) {
2038 if (!dev_data->passthrough)
2041 if (dev_data->iommu_v2) {
2042 if (pdev_iommuv2_enable(pdev) != 0)
2045 dev_data->ats.enabled = true;
2046 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2047 dev_data->pri_tlp = pci_pri_tlp_required(pdev);
2049 } else if (amd_iommu_iotlb_sup &&
2050 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
2051 dev_data->ats.enabled = true;
2052 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2056 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2057 ret = __attach_device(dev_data, domain);
2058 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2061 * We might boot into a crash-kernel here. The crashed kernel
2062 * left the caches in the IOMMU dirty. So we have to flush
2063 * here to evict all dirty stuff.
2065 domain_flush_tlb_pde(domain);
2071 * Removes a device from a protection domain (unlocked)
2073 static void __detach_device(struct iommu_dev_data *dev_data)
2075 struct protection_domain *domain;
2078 * Must be called with IRQs disabled. Warn here to detect early
2081 WARN_ON(!irqs_disabled());
2083 if (WARN_ON(!dev_data->domain))
2086 domain = dev_data->domain;
2088 spin_lock(&domain->lock);
2090 do_detach(dev_data);
2092 spin_unlock(&domain->lock);
2096 * Removes a device from a protection domain (with devtable_lock held)
2098 static void detach_device(struct device *dev)
2100 struct protection_domain *domain;
2101 struct iommu_dev_data *dev_data;
2102 unsigned long flags;
2104 dev_data = get_dev_data(dev);
2105 domain = dev_data->domain;
2107 /* lock device table */
2108 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2109 __detach_device(dev_data);
2110 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2112 if (!dev_is_pci(dev))
2115 if (domain->flags & PD_IOMMUV2_MASK && dev_data->iommu_v2)
2116 pdev_iommuv2_disable(to_pci_dev(dev));
2117 else if (dev_data->ats.enabled)
2118 pci_disable_ats(to_pci_dev(dev));
2120 dev_data->ats.enabled = false;
2123 static int amd_iommu_add_device(struct device *dev)
2125 struct iommu_dev_data *dev_data;
2126 struct iommu_domain *domain;
2127 struct amd_iommu *iommu;
2130 if (!check_device(dev) || get_dev_data(dev))
2133 devid = get_device_id(dev);
2137 iommu = amd_iommu_rlookup_table[devid];
2139 ret = iommu_init_device(dev);
2141 if (ret != -ENOTSUPP)
2142 pr_err("Failed to initialize device %s - trying to proceed anyway\n",
2145 iommu_ignore_device(dev);
2146 dev->archdata.dma_ops = &nommu_dma_ops;
2149 init_iommu_group(dev);
2151 dev_data = get_dev_data(dev);
2155 if (iommu_pass_through || dev_data->iommu_v2)
2156 iommu_request_dm_for_dev(dev);
2158 /* Domains are initialized for this device - have a look what we ended up with */
2159 domain = iommu_get_domain_for_dev(dev);
2160 if (domain->type == IOMMU_DOMAIN_IDENTITY)
2161 dev_data->passthrough = true;
2163 dev->archdata.dma_ops = &amd_iommu_dma_ops;
2166 iommu_completion_wait(iommu);
2171 static void amd_iommu_remove_device(struct device *dev)
2173 struct amd_iommu *iommu;
2176 if (!check_device(dev))
2179 devid = get_device_id(dev);
2183 iommu = amd_iommu_rlookup_table[devid];
2185 iommu_uninit_device(dev);
2186 iommu_completion_wait(iommu);
2189 static struct iommu_group *amd_iommu_device_group(struct device *dev)
2191 if (dev_is_pci(dev))
2192 return pci_device_group(dev);
2194 return acpihid_device_group(dev);
2197 /*****************************************************************************
2199 * The next functions belong to the dma_ops mapping/unmapping code.
2201 *****************************************************************************/
2203 static void __queue_flush(struct flush_queue *queue)
2205 struct protection_domain *domain;
2206 unsigned long flags;
2209 /* First flush TLB of all known domains */
2210 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
2211 list_for_each_entry(domain, &amd_iommu_pd_list, list)
2212 domain_flush_tlb(domain);
2213 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
2215 /* Wait until flushes have completed */
2216 domain_flush_complete(NULL);
2218 for (idx = 0; idx < queue->next; ++idx) {
2219 struct flush_queue_entry *entry;
2221 entry = queue->entries + idx;
2223 free_iova_fast(&entry->dma_dom->iovad,
2227 /* Not really necessary, just to make sure we catch any bugs */
2228 entry->dma_dom = NULL;
2234 static void queue_flush_all(void)
2238 for_each_possible_cpu(cpu) {
2239 struct flush_queue *queue;
2240 unsigned long flags;
2242 queue = per_cpu_ptr(&flush_queue, cpu);
2243 spin_lock_irqsave(&queue->lock, flags);
2244 if (queue->next > 0)
2245 __queue_flush(queue);
2246 spin_unlock_irqrestore(&queue->lock, flags);
2250 static void queue_flush_timeout(unsigned long unsused)
2252 atomic_set(&queue_timer_on, 0);
2256 static void queue_add(struct dma_ops_domain *dma_dom,
2257 unsigned long address, unsigned long pages)
2259 struct flush_queue_entry *entry;
2260 struct flush_queue *queue;
2261 unsigned long flags;
2264 pages = __roundup_pow_of_two(pages);
2265 address >>= PAGE_SHIFT;
2267 queue = get_cpu_ptr(&flush_queue);
2268 spin_lock_irqsave(&queue->lock, flags);
2270 if (queue->next == FLUSH_QUEUE_SIZE)
2271 __queue_flush(queue);
2273 idx = queue->next++;
2274 entry = queue->entries + idx;
2276 entry->iova_pfn = address;
2277 entry->pages = pages;
2278 entry->dma_dom = dma_dom;
2280 spin_unlock_irqrestore(&queue->lock, flags);
2282 if (atomic_cmpxchg(&queue_timer_on, 0, 1) == 0)
2283 mod_timer(&queue_timer, jiffies + msecs_to_jiffies(10));
2285 put_cpu_ptr(&flush_queue);
2290 * In the dma_ops path we only have the struct device. This function
2291 * finds the corresponding IOMMU, the protection domain and the
2292 * requestor id for a given device.
2293 * If the device is not yet associated with a domain this is also done
2296 static struct protection_domain *get_domain(struct device *dev)
2298 struct protection_domain *domain;
2300 if (!check_device(dev))
2301 return ERR_PTR(-EINVAL);
2303 domain = get_dev_data(dev)->domain;
2304 if (!dma_ops_domain(domain))
2305 return ERR_PTR(-EBUSY);
2310 static void update_device_table(struct protection_domain *domain)
2312 struct iommu_dev_data *dev_data;
2314 list_for_each_entry(dev_data, &domain->dev_list, list) {
2315 set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled);
2317 if (dev_data->devid == dev_data->alias)
2320 /* There is an alias, update device table entry for it */
2321 set_dte_entry(dev_data->alias, domain, dev_data->ats.enabled);
2325 static void update_domain(struct protection_domain *domain)
2327 if (!domain->updated)
2330 update_device_table(domain);
2332 domain_flush_devices(domain);
2333 domain_flush_tlb_pde(domain);
2335 domain->updated = false;
2338 static int dir2prot(enum dma_data_direction direction)
2340 if (direction == DMA_TO_DEVICE)
2341 return IOMMU_PROT_IR;
2342 else if (direction == DMA_FROM_DEVICE)
2343 return IOMMU_PROT_IW;
2344 else if (direction == DMA_BIDIRECTIONAL)
2345 return IOMMU_PROT_IW | IOMMU_PROT_IR;
2350 * This function contains common code for mapping of a physically
2351 * contiguous memory region into DMA address space. It is used by all
2352 * mapping functions provided with this IOMMU driver.
2353 * Must be called with the domain lock held.
2355 static dma_addr_t __map_single(struct device *dev,
2356 struct dma_ops_domain *dma_dom,
2359 enum dma_data_direction direction,
2362 dma_addr_t offset = paddr & ~PAGE_MASK;
2363 dma_addr_t address, start, ret;
2368 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
2371 address = dma_ops_alloc_iova(dev, dma_dom, pages, dma_mask);
2372 if (address == DMA_ERROR_CODE)
2375 prot = dir2prot(direction);
2378 for (i = 0; i < pages; ++i) {
2379 ret = iommu_map_page(&dma_dom->domain, start, paddr,
2380 PAGE_SIZE, prot, GFP_ATOMIC);
2389 if (unlikely(amd_iommu_np_cache)) {
2390 domain_flush_pages(&dma_dom->domain, address, size);
2391 domain_flush_complete(&dma_dom->domain);
2399 for (--i; i >= 0; --i) {
2401 iommu_unmap_page(&dma_dom->domain, start, PAGE_SIZE);
2404 domain_flush_tlb(&dma_dom->domain);
2405 domain_flush_complete(&dma_dom->domain);
2407 dma_ops_free_iova(dma_dom, address, pages);
2409 return DMA_ERROR_CODE;
2413 * Does the reverse of the __map_single function. Must be called with
2414 * the domain lock held too
2416 static void __unmap_single(struct dma_ops_domain *dma_dom,
2417 dma_addr_t dma_addr,
2421 dma_addr_t flush_addr;
2422 dma_addr_t i, start;
2425 flush_addr = dma_addr;
2426 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
2427 dma_addr &= PAGE_MASK;
2430 for (i = 0; i < pages; ++i) {
2431 iommu_unmap_page(&dma_dom->domain, start, PAGE_SIZE);
2435 if (amd_iommu_unmap_flush) {
2436 dma_ops_free_iova(dma_dom, dma_addr, pages);
2437 domain_flush_tlb(&dma_dom->domain);
2438 domain_flush_complete(&dma_dom->domain);
2440 queue_add(dma_dom, dma_addr, pages);
2445 * The exported map_single function for dma_ops.
2447 static dma_addr_t map_page(struct device *dev, struct page *page,
2448 unsigned long offset, size_t size,
2449 enum dma_data_direction dir,
2450 unsigned long attrs)
2452 phys_addr_t paddr = page_to_phys(page) + offset;
2453 struct protection_domain *domain;
2454 struct dma_ops_domain *dma_dom;
2457 domain = get_domain(dev);
2458 if (PTR_ERR(domain) == -EINVAL)
2459 return (dma_addr_t)paddr;
2460 else if (IS_ERR(domain))
2461 return DMA_ERROR_CODE;
2463 dma_mask = *dev->dma_mask;
2464 dma_dom = to_dma_ops_domain(domain);
2466 return __map_single(dev, dma_dom, paddr, size, dir, dma_mask);
2470 * The exported unmap_single function for dma_ops.
2472 static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2473 enum dma_data_direction dir, unsigned long attrs)
2475 struct protection_domain *domain;
2476 struct dma_ops_domain *dma_dom;
2478 domain = get_domain(dev);
2482 dma_dom = to_dma_ops_domain(domain);
2484 __unmap_single(dma_dom, dma_addr, size, dir);
2487 static int sg_num_pages(struct device *dev,
2488 struct scatterlist *sglist,
2491 unsigned long mask, boundary_size;
2492 struct scatterlist *s;
2495 mask = dma_get_seg_boundary(dev);
2496 boundary_size = mask + 1 ? ALIGN(mask + 1, PAGE_SIZE) >> PAGE_SHIFT :
2497 1UL << (BITS_PER_LONG - PAGE_SHIFT);
2499 for_each_sg(sglist, s, nelems, i) {
2502 s->dma_address = npages << PAGE_SHIFT;
2503 p = npages % boundary_size;
2504 n = iommu_num_pages(sg_phys(s), s->length, PAGE_SIZE);
2505 if (p + n > boundary_size)
2506 npages += boundary_size - p;
2514 * The exported map_sg function for dma_ops (handles scatter-gather
2517 static int map_sg(struct device *dev, struct scatterlist *sglist,
2518 int nelems, enum dma_data_direction direction,
2519 unsigned long attrs)
2521 int mapped_pages = 0, npages = 0, prot = 0, i;
2522 struct protection_domain *domain;
2523 struct dma_ops_domain *dma_dom;
2524 struct scatterlist *s;
2525 unsigned long address;
2528 domain = get_domain(dev);
2532 dma_dom = to_dma_ops_domain(domain);
2533 dma_mask = *dev->dma_mask;
2535 npages = sg_num_pages(dev, sglist, nelems);
2537 address = dma_ops_alloc_iova(dev, dma_dom, npages, dma_mask);
2538 if (address == DMA_ERROR_CODE)
2541 prot = dir2prot(direction);
2543 /* Map all sg entries */
2544 for_each_sg(sglist, s, nelems, i) {
2545 int j, pages = iommu_num_pages(sg_phys(s), s->length, PAGE_SIZE);
2547 for (j = 0; j < pages; ++j) {
2548 unsigned long bus_addr, phys_addr;
2551 bus_addr = address + s->dma_address + (j << PAGE_SHIFT);
2552 phys_addr = (sg_phys(s) & PAGE_MASK) + (j << PAGE_SHIFT);
2553 ret = iommu_map_page(domain, bus_addr, phys_addr, PAGE_SIZE, prot, GFP_ATOMIC);
2561 /* Everything is mapped - write the right values into s->dma_address */
2562 for_each_sg(sglist, s, nelems, i) {
2563 s->dma_address += address + s->offset;
2564 s->dma_length = s->length;
2570 pr_err("%s: IOMMU mapping error in map_sg (io-pages: %d)\n",
2571 dev_name(dev), npages);
2573 for_each_sg(sglist, s, nelems, i) {
2574 int j, pages = iommu_num_pages(sg_phys(s), s->length, PAGE_SIZE);
2576 for (j = 0; j < pages; ++j) {
2577 unsigned long bus_addr;
2579 bus_addr = address + s->dma_address + (j << PAGE_SHIFT);
2580 iommu_unmap_page(domain, bus_addr, PAGE_SIZE);
2588 free_iova_fast(&dma_dom->iovad, address, npages);
2595 * The exported map_sg function for dma_ops (handles scatter-gather
2598 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2599 int nelems, enum dma_data_direction dir,
2600 unsigned long attrs)
2602 struct protection_domain *domain;
2603 struct dma_ops_domain *dma_dom;
2604 unsigned long startaddr;
2607 domain = get_domain(dev);
2611 startaddr = sg_dma_address(sglist) & PAGE_MASK;
2612 dma_dom = to_dma_ops_domain(domain);
2613 npages = sg_num_pages(dev, sglist, nelems);
2615 __unmap_single(dma_dom, startaddr, npages << PAGE_SHIFT, dir);
2619 * The exported alloc_coherent function for dma_ops.
2621 static void *alloc_coherent(struct device *dev, size_t size,
2622 dma_addr_t *dma_addr, gfp_t flag,
2623 unsigned long attrs)
2625 u64 dma_mask = dev->coherent_dma_mask;
2626 struct protection_domain *domain;
2627 struct dma_ops_domain *dma_dom;
2630 domain = get_domain(dev);
2631 if (PTR_ERR(domain) == -EINVAL) {
2632 page = alloc_pages(flag, get_order(size));
2633 *dma_addr = page_to_phys(page);
2634 return page_address(page);
2635 } else if (IS_ERR(domain))
2638 dma_dom = to_dma_ops_domain(domain);
2639 size = PAGE_ALIGN(size);
2640 dma_mask = dev->coherent_dma_mask;
2641 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
2644 page = alloc_pages(flag | __GFP_NOWARN, get_order(size));
2646 if (!gfpflags_allow_blocking(flag))
2649 page = dma_alloc_from_contiguous(dev, size >> PAGE_SHIFT,
2656 dma_mask = *dev->dma_mask;
2658 *dma_addr = __map_single(dev, dma_dom, page_to_phys(page),
2659 size, DMA_BIDIRECTIONAL, dma_mask);
2661 if (*dma_addr == DMA_ERROR_CODE)
2664 return page_address(page);
2668 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2669 __free_pages(page, get_order(size));
2675 * The exported free_coherent function for dma_ops.
2677 static void free_coherent(struct device *dev, size_t size,
2678 void *virt_addr, dma_addr_t dma_addr,
2679 unsigned long attrs)
2681 struct protection_domain *domain;
2682 struct dma_ops_domain *dma_dom;
2685 page = virt_to_page(virt_addr);
2686 size = PAGE_ALIGN(size);
2688 domain = get_domain(dev);
2692 dma_dom = to_dma_ops_domain(domain);
2694 __unmap_single(dma_dom, dma_addr, size, DMA_BIDIRECTIONAL);
2697 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2698 __free_pages(page, get_order(size));
2702 * This function is called by the DMA layer to find out if we can handle a
2703 * particular device. It is part of the dma_ops.
2705 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
2707 return check_device(dev);
2710 static struct dma_map_ops amd_iommu_dma_ops = {
2711 .alloc = alloc_coherent,
2712 .free = free_coherent,
2713 .map_page = map_page,
2714 .unmap_page = unmap_page,
2716 .unmap_sg = unmap_sg,
2717 .dma_supported = amd_iommu_dma_supported,
2720 static int init_reserved_iova_ranges(void)
2722 struct pci_dev *pdev = NULL;
2725 init_iova_domain(&reserved_iova_ranges, PAGE_SIZE,
2726 IOVA_START_PFN, DMA_32BIT_PFN);
2728 lockdep_set_class(&reserved_iova_ranges.iova_rbtree_lock,
2729 &reserved_rbtree_key);
2731 /* MSI memory range */
2732 val = reserve_iova(&reserved_iova_ranges,
2733 IOVA_PFN(MSI_RANGE_START), IOVA_PFN(MSI_RANGE_END));
2735 pr_err("Reserving MSI range failed\n");
2739 /* HT memory range */
2740 val = reserve_iova(&reserved_iova_ranges,
2741 IOVA_PFN(HT_RANGE_START), IOVA_PFN(HT_RANGE_END));
2743 pr_err("Reserving HT range failed\n");
2748 * Memory used for PCI resources
2749 * FIXME: Check whether we can reserve the PCI-hole completly
2751 for_each_pci_dev(pdev) {
2754 for (i = 0; i < PCI_NUM_RESOURCES; ++i) {
2755 struct resource *r = &pdev->resource[i];
2757 if (!(r->flags & IORESOURCE_MEM))
2760 val = reserve_iova(&reserved_iova_ranges,
2764 pr_err("Reserve pci-resource range failed\n");
2773 int __init amd_iommu_init_api(void)
2775 int ret, cpu, err = 0;
2777 ret = iova_cache_get();
2781 ret = init_reserved_iova_ranges();
2785 for_each_possible_cpu(cpu) {
2786 struct flush_queue *queue = per_cpu_ptr(&flush_queue, cpu);
2788 queue->entries = kzalloc(FLUSH_QUEUE_SIZE *
2789 sizeof(*queue->entries),
2791 if (!queue->entries)
2794 spin_lock_init(&queue->lock);
2797 err = bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
2800 #ifdef CONFIG_ARM_AMBA
2801 err = bus_set_iommu(&amba_bustype, &amd_iommu_ops);
2805 err = bus_set_iommu(&platform_bus_type, &amd_iommu_ops);
2811 for_each_possible_cpu(cpu) {
2812 struct flush_queue *queue = per_cpu_ptr(&flush_queue, cpu);
2814 kfree(queue->entries);
2820 int __init amd_iommu_init_dma_ops(void)
2822 setup_timer(&queue_timer, queue_flush_timeout, 0);
2823 atomic_set(&queue_timer_on, 0);
2825 swiotlb = iommu_pass_through ? 1 : 0;
2829 * In case we don't initialize SWIOTLB (actually the common case
2830 * when AMD IOMMU is enabled), make sure there are global
2831 * dma_ops set as a fall-back for devices not handled by this
2832 * driver (for example non-PCI devices).
2835 dma_ops = &nommu_dma_ops;
2837 if (amd_iommu_unmap_flush)
2838 pr_info("AMD-Vi: IO/TLB flush on unmap enabled\n");
2840 pr_info("AMD-Vi: Lazy IO/TLB flushing enabled\n");
2846 /*****************************************************************************
2848 * The following functions belong to the exported interface of AMD IOMMU
2850 * This interface allows access to lower level functions of the IOMMU
2851 * like protection domain handling and assignement of devices to domains
2852 * which is not possible with the dma_ops interface.
2854 *****************************************************************************/
2856 static void cleanup_domain(struct protection_domain *domain)
2858 struct iommu_dev_data *entry;
2859 unsigned long flags;
2861 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2863 while (!list_empty(&domain->dev_list)) {
2864 entry = list_first_entry(&domain->dev_list,
2865 struct iommu_dev_data, list);
2866 __detach_device(entry);
2869 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2872 static void protection_domain_free(struct protection_domain *domain)
2877 del_domain_from_list(domain);
2880 domain_id_free(domain->id);
2885 static int protection_domain_init(struct protection_domain *domain)
2887 spin_lock_init(&domain->lock);
2888 mutex_init(&domain->api_lock);
2889 domain->id = domain_id_alloc();
2892 INIT_LIST_HEAD(&domain->dev_list);
2897 static struct protection_domain *protection_domain_alloc(void)
2899 struct protection_domain *domain;
2901 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2905 if (protection_domain_init(domain))
2908 add_domain_to_list(domain);
2918 static struct iommu_domain *amd_iommu_domain_alloc(unsigned type)
2920 struct protection_domain *pdomain;
2921 struct dma_ops_domain *dma_domain;
2924 case IOMMU_DOMAIN_UNMANAGED:
2925 pdomain = protection_domain_alloc();
2929 pdomain->mode = PAGE_MODE_3_LEVEL;
2930 pdomain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2931 if (!pdomain->pt_root) {
2932 protection_domain_free(pdomain);
2936 pdomain->domain.geometry.aperture_start = 0;
2937 pdomain->domain.geometry.aperture_end = ~0ULL;
2938 pdomain->domain.geometry.force_aperture = true;
2941 case IOMMU_DOMAIN_DMA:
2942 dma_domain = dma_ops_domain_alloc();
2944 pr_err("AMD-Vi: Failed to allocate\n");
2947 pdomain = &dma_domain->domain;
2949 case IOMMU_DOMAIN_IDENTITY:
2950 pdomain = protection_domain_alloc();
2954 pdomain->mode = PAGE_MODE_NONE;
2960 return &pdomain->domain;
2963 static void amd_iommu_domain_free(struct iommu_domain *dom)
2965 struct protection_domain *domain;
2966 struct dma_ops_domain *dma_dom;
2968 domain = to_pdomain(dom);
2970 if (domain->dev_cnt > 0)
2971 cleanup_domain(domain);
2973 BUG_ON(domain->dev_cnt != 0);
2978 switch (dom->type) {
2979 case IOMMU_DOMAIN_DMA:
2981 * First make sure the domain is no longer referenced from the
2986 /* Now release the domain */
2987 dma_dom = to_dma_ops_domain(domain);
2988 dma_ops_domain_free(dma_dom);
2991 if (domain->mode != PAGE_MODE_NONE)
2992 free_pagetable(domain);
2994 if (domain->flags & PD_IOMMUV2_MASK)
2995 free_gcr3_table(domain);
2997 protection_domain_free(domain);
3002 static void amd_iommu_detach_device(struct iommu_domain *dom,
3005 struct iommu_dev_data *dev_data = dev->archdata.iommu;
3006 struct amd_iommu *iommu;
3009 if (!check_device(dev))
3012 devid = get_device_id(dev);
3016 if (dev_data->domain != NULL)
3019 iommu = amd_iommu_rlookup_table[devid];
3023 #ifdef CONFIG_IRQ_REMAP
3024 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) &&
3025 (dom->type == IOMMU_DOMAIN_UNMANAGED))
3026 dev_data->use_vapic = 0;
3029 iommu_completion_wait(iommu);
3032 static int amd_iommu_attach_device(struct iommu_domain *dom,
3035 struct protection_domain *domain = to_pdomain(dom);
3036 struct iommu_dev_data *dev_data;
3037 struct amd_iommu *iommu;
3040 if (!check_device(dev))
3043 dev_data = dev->archdata.iommu;
3045 iommu = amd_iommu_rlookup_table[dev_data->devid];
3049 if (dev_data->domain)
3052 ret = attach_device(dev, domain);
3054 #ifdef CONFIG_IRQ_REMAP
3055 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
3056 if (dom->type == IOMMU_DOMAIN_UNMANAGED)
3057 dev_data->use_vapic = 1;
3059 dev_data->use_vapic = 0;
3063 iommu_completion_wait(iommu);
3068 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
3069 phys_addr_t paddr, size_t page_size, int iommu_prot)
3071 struct protection_domain *domain = to_pdomain(dom);
3075 if (domain->mode == PAGE_MODE_NONE)
3078 if (iommu_prot & IOMMU_READ)
3079 prot |= IOMMU_PROT_IR;
3080 if (iommu_prot & IOMMU_WRITE)
3081 prot |= IOMMU_PROT_IW;
3083 mutex_lock(&domain->api_lock);
3084 ret = iommu_map_page(domain, iova, paddr, page_size, prot, GFP_KERNEL);
3085 mutex_unlock(&domain->api_lock);
3090 static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
3093 struct protection_domain *domain = to_pdomain(dom);
3096 if (domain->mode == PAGE_MODE_NONE)
3099 mutex_lock(&domain->api_lock);
3100 unmap_size = iommu_unmap_page(domain, iova, page_size);
3101 mutex_unlock(&domain->api_lock);
3103 domain_flush_tlb_pde(domain);
3108 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
3111 struct protection_domain *domain = to_pdomain(dom);
3112 unsigned long offset_mask, pte_pgsize;
3115 if (domain->mode == PAGE_MODE_NONE)
3118 pte = fetch_pte(domain, iova, &pte_pgsize);
3120 if (!pte || !IOMMU_PTE_PRESENT(*pte))
3123 offset_mask = pte_pgsize - 1;
3124 __pte = *pte & PM_ADDR_MASK;
3126 return (__pte & ~offset_mask) | (iova & offset_mask);
3129 static bool amd_iommu_capable(enum iommu_cap cap)
3132 case IOMMU_CAP_CACHE_COHERENCY:
3134 case IOMMU_CAP_INTR_REMAP:
3135 return (irq_remapping_enabled == 1);
3136 case IOMMU_CAP_NOEXEC:
3143 static void amd_iommu_get_dm_regions(struct device *dev,
3144 struct list_head *head)
3146 struct unity_map_entry *entry;
3149 devid = get_device_id(dev);
3153 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
3154 struct iommu_dm_region *region;
3156 if (devid < entry->devid_start || devid > entry->devid_end)
3159 region = kzalloc(sizeof(*region), GFP_KERNEL);
3161 pr_err("Out of memory allocating dm-regions for %s\n",
3166 region->start = entry->address_start;
3167 region->length = entry->address_end - entry->address_start;
3168 if (entry->prot & IOMMU_PROT_IR)
3169 region->prot |= IOMMU_READ;
3170 if (entry->prot & IOMMU_PROT_IW)
3171 region->prot |= IOMMU_WRITE;
3173 list_add_tail(®ion->list, head);
3177 static void amd_iommu_put_dm_regions(struct device *dev,
3178 struct list_head *head)
3180 struct iommu_dm_region *entry, *next;
3182 list_for_each_entry_safe(entry, next, head, list)
3186 static void amd_iommu_apply_dm_region(struct device *dev,
3187 struct iommu_domain *domain,
3188 struct iommu_dm_region *region)
3190 struct dma_ops_domain *dma_dom = to_dma_ops_domain(to_pdomain(domain));
3191 unsigned long start, end;
3193 start = IOVA_PFN(region->start);
3194 end = IOVA_PFN(region->start + region->length);
3196 WARN_ON_ONCE(reserve_iova(&dma_dom->iovad, start, end) == NULL);
3199 static const struct iommu_ops amd_iommu_ops = {
3200 .capable = amd_iommu_capable,
3201 .domain_alloc = amd_iommu_domain_alloc,
3202 .domain_free = amd_iommu_domain_free,
3203 .attach_dev = amd_iommu_attach_device,
3204 .detach_dev = amd_iommu_detach_device,
3205 .map = amd_iommu_map,
3206 .unmap = amd_iommu_unmap,
3207 .map_sg = default_iommu_map_sg,
3208 .iova_to_phys = amd_iommu_iova_to_phys,
3209 .add_device = amd_iommu_add_device,
3210 .remove_device = amd_iommu_remove_device,
3211 .device_group = amd_iommu_device_group,
3212 .get_dm_regions = amd_iommu_get_dm_regions,
3213 .put_dm_regions = amd_iommu_put_dm_regions,
3214 .apply_dm_region = amd_iommu_apply_dm_region,
3215 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
3218 /*****************************************************************************
3220 * The next functions do a basic initialization of IOMMU for pass through
3223 * In passthrough mode the IOMMU is initialized and enabled but not used for
3224 * DMA-API translation.
3226 *****************************************************************************/
3228 /* IOMMUv2 specific functions */
3229 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
3231 return atomic_notifier_chain_register(&ppr_notifier, nb);
3233 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
3235 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
3237 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
3239 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
3241 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
3243 struct protection_domain *domain = to_pdomain(dom);
3244 unsigned long flags;
3246 spin_lock_irqsave(&domain->lock, flags);
3248 /* Update data structure */
3249 domain->mode = PAGE_MODE_NONE;
3250 domain->updated = true;
3252 /* Make changes visible to IOMMUs */
3253 update_domain(domain);
3255 /* Page-table is not visible to IOMMU anymore, so free it */
3256 free_pagetable(domain);
3258 spin_unlock_irqrestore(&domain->lock, flags);
3260 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
3262 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
3264 struct protection_domain *domain = to_pdomain(dom);
3265 unsigned long flags;
3268 if (pasids <= 0 || pasids > (PASID_MASK + 1))
3271 /* Number of GCR3 table levels required */
3272 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
3275 if (levels > amd_iommu_max_glx_val)
3278 spin_lock_irqsave(&domain->lock, flags);
3281 * Save us all sanity checks whether devices already in the
3282 * domain support IOMMUv2. Just force that the domain has no
3283 * devices attached when it is switched into IOMMUv2 mode.
3286 if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
3290 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
3291 if (domain->gcr3_tbl == NULL)
3294 domain->glx = levels;
3295 domain->flags |= PD_IOMMUV2_MASK;
3296 domain->updated = true;
3298 update_domain(domain);
3303 spin_unlock_irqrestore(&domain->lock, flags);
3307 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
3309 static int __flush_pasid(struct protection_domain *domain, int pasid,
3310 u64 address, bool size)
3312 struct iommu_dev_data *dev_data;
3313 struct iommu_cmd cmd;
3316 if (!(domain->flags & PD_IOMMUV2_MASK))
3319 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
3322 * IOMMU TLB needs to be flushed before Device TLB to
3323 * prevent device TLB refill from IOMMU TLB
3325 for (i = 0; i < amd_iommus_present; ++i) {
3326 if (domain->dev_iommu[i] == 0)
3329 ret = iommu_queue_command(amd_iommus[i], &cmd);
3334 /* Wait until IOMMU TLB flushes are complete */
3335 domain_flush_complete(domain);
3337 /* Now flush device TLBs */
3338 list_for_each_entry(dev_data, &domain->dev_list, list) {
3339 struct amd_iommu *iommu;
3343 There might be non-IOMMUv2 capable devices in an IOMMUv2
3346 if (!dev_data->ats.enabled)
3349 qdep = dev_data->ats.qdep;
3350 iommu = amd_iommu_rlookup_table[dev_data->devid];
3352 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
3353 qdep, address, size);
3355 ret = iommu_queue_command(iommu, &cmd);
3360 /* Wait until all device TLBs are flushed */
3361 domain_flush_complete(domain);
3370 static int __amd_iommu_flush_page(struct protection_domain *domain, int pasid,
3373 return __flush_pasid(domain, pasid, address, false);
3376 int amd_iommu_flush_page(struct iommu_domain *dom, int pasid,
3379 struct protection_domain *domain = to_pdomain(dom);
3380 unsigned long flags;
3383 spin_lock_irqsave(&domain->lock, flags);
3384 ret = __amd_iommu_flush_page(domain, pasid, address);
3385 spin_unlock_irqrestore(&domain->lock, flags);
3389 EXPORT_SYMBOL(amd_iommu_flush_page);
3391 static int __amd_iommu_flush_tlb(struct protection_domain *domain, int pasid)
3393 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
3397 int amd_iommu_flush_tlb(struct iommu_domain *dom, int pasid)
3399 struct protection_domain *domain = to_pdomain(dom);
3400 unsigned long flags;
3403 spin_lock_irqsave(&domain->lock, flags);
3404 ret = __amd_iommu_flush_tlb(domain, pasid);
3405 spin_unlock_irqrestore(&domain->lock, flags);
3409 EXPORT_SYMBOL(amd_iommu_flush_tlb);
3411 static u64 *__get_gcr3_pte(u64 *root, int level, int pasid, bool alloc)
3418 index = (pasid >> (9 * level)) & 0x1ff;
3424 if (!(*pte & GCR3_VALID)) {
3428 root = (void *)get_zeroed_page(GFP_ATOMIC);
3432 *pte = __pa(root) | GCR3_VALID;
3435 root = __va(*pte & PAGE_MASK);
3443 static int __set_gcr3(struct protection_domain *domain, int pasid,
3448 if (domain->mode != PAGE_MODE_NONE)
3451 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3455 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3457 return __amd_iommu_flush_tlb(domain, pasid);
3460 static int __clear_gcr3(struct protection_domain *domain, int pasid)
3464 if (domain->mode != PAGE_MODE_NONE)
3467 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3473 return __amd_iommu_flush_tlb(domain, pasid);
3476 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, int pasid,
3479 struct protection_domain *domain = to_pdomain(dom);
3480 unsigned long flags;
3483 spin_lock_irqsave(&domain->lock, flags);
3484 ret = __set_gcr3(domain, pasid, cr3);
3485 spin_unlock_irqrestore(&domain->lock, flags);
3489 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3491 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, int pasid)
3493 struct protection_domain *domain = to_pdomain(dom);
3494 unsigned long flags;
3497 spin_lock_irqsave(&domain->lock, flags);
3498 ret = __clear_gcr3(domain, pasid);
3499 spin_unlock_irqrestore(&domain->lock, flags);
3503 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3505 int amd_iommu_complete_ppr(struct pci_dev *pdev, int pasid,
3506 int status, int tag)
3508 struct iommu_dev_data *dev_data;
3509 struct amd_iommu *iommu;
3510 struct iommu_cmd cmd;
3512 dev_data = get_dev_data(&pdev->dev);
3513 iommu = amd_iommu_rlookup_table[dev_data->devid];
3515 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3516 tag, dev_data->pri_tlp);
3518 return iommu_queue_command(iommu, &cmd);
3520 EXPORT_SYMBOL(amd_iommu_complete_ppr);
3522 struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3524 struct protection_domain *pdomain;
3526 pdomain = get_domain(&pdev->dev);
3527 if (IS_ERR(pdomain))
3530 /* Only return IOMMUv2 domains */
3531 if (!(pdomain->flags & PD_IOMMUV2_MASK))
3534 return &pdomain->domain;
3536 EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3538 void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3540 struct iommu_dev_data *dev_data;
3542 if (!amd_iommu_v2_supported())
3545 dev_data = get_dev_data(&pdev->dev);
3546 dev_data->errata |= (1 << erratum);
3548 EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3550 int amd_iommu_device_info(struct pci_dev *pdev,
3551 struct amd_iommu_device_info *info)
3556 if (pdev == NULL || info == NULL)
3559 if (!amd_iommu_v2_supported())
3562 memset(info, 0, sizeof(*info));
3564 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS);
3566 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3568 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3570 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3572 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3576 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3577 max_pasids = min(max_pasids, (1 << 20));
3579 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3580 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3582 features = pci_pasid_features(pdev);
3583 if (features & PCI_PASID_CAP_EXEC)
3584 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3585 if (features & PCI_PASID_CAP_PRIV)
3586 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3591 EXPORT_SYMBOL(amd_iommu_device_info);
3593 #ifdef CONFIG_IRQ_REMAP
3595 /*****************************************************************************
3597 * Interrupt Remapping Implementation
3599 *****************************************************************************/
3601 static struct irq_chip amd_ir_chip;
3603 #define DTE_IRQ_PHYS_ADDR_MASK (((1ULL << 45)-1) << 6)
3604 #define DTE_IRQ_REMAP_INTCTL (2ULL << 60)
3605 #define DTE_IRQ_TABLE_LEN (8ULL << 1)
3606 #define DTE_IRQ_REMAP_ENABLE 1ULL
3608 static void set_dte_irq_entry(u16 devid, struct irq_remap_table *table)
3612 dte = amd_iommu_dev_table[devid].data[2];
3613 dte &= ~DTE_IRQ_PHYS_ADDR_MASK;
3614 dte |= virt_to_phys(table->table);
3615 dte |= DTE_IRQ_REMAP_INTCTL;
3616 dte |= DTE_IRQ_TABLE_LEN;
3617 dte |= DTE_IRQ_REMAP_ENABLE;
3619 amd_iommu_dev_table[devid].data[2] = dte;
3622 static struct irq_remap_table *get_irq_table(u16 devid, bool ioapic)
3624 struct irq_remap_table *table = NULL;
3625 struct amd_iommu *iommu;
3626 unsigned long flags;
3629 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3631 iommu = amd_iommu_rlookup_table[devid];
3635 table = irq_lookup_table[devid];
3639 alias = amd_iommu_alias_table[devid];
3640 table = irq_lookup_table[alias];
3642 irq_lookup_table[devid] = table;
3643 set_dte_irq_entry(devid, table);
3644 iommu_flush_dte(iommu, devid);
3648 /* Nothing there yet, allocate new irq remapping table */
3649 table = kzalloc(sizeof(*table), GFP_ATOMIC);
3653 /* Initialize table spin-lock */
3654 spin_lock_init(&table->lock);
3657 /* Keep the first 32 indexes free for IOAPIC interrupts */
3658 table->min_index = 32;
3660 table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_ATOMIC);
3661 if (!table->table) {
3667 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
3668 memset(table->table, 0,
3669 MAX_IRQS_PER_TABLE * sizeof(u32));
3671 memset(table->table, 0,
3672 (MAX_IRQS_PER_TABLE * (sizeof(u64) * 2)));
3677 for (i = 0; i < 32; ++i)
3678 iommu->irte_ops->set_allocated(table, i);
3681 irq_lookup_table[devid] = table;
3682 set_dte_irq_entry(devid, table);
3683 iommu_flush_dte(iommu, devid);
3684 if (devid != alias) {
3685 irq_lookup_table[alias] = table;
3686 set_dte_irq_entry(alias, table);
3687 iommu_flush_dte(iommu, alias);
3691 iommu_completion_wait(iommu);
3694 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3699 static int alloc_irq_index(u16 devid, int count)
3701 struct irq_remap_table *table;
3702 unsigned long flags;
3704 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
3709 table = get_irq_table(devid, false);
3713 spin_lock_irqsave(&table->lock, flags);
3715 /* Scan table for free entries */
3716 for (c = 0, index = table->min_index;
3717 index < MAX_IRQS_PER_TABLE;
3719 if (!iommu->irte_ops->is_allocated(table, index))
3726 iommu->irte_ops->set_allocated(table, index - c + 1);
3736 spin_unlock_irqrestore(&table->lock, flags);
3741 static int modify_irte_ga(u16 devid, int index, struct irte_ga *irte,
3742 struct amd_ir_data *data)
3744 struct irq_remap_table *table;
3745 struct amd_iommu *iommu;
3746 unsigned long flags;
3747 struct irte_ga *entry;
3749 iommu = amd_iommu_rlookup_table[devid];
3753 table = get_irq_table(devid, false);
3757 spin_lock_irqsave(&table->lock, flags);
3759 entry = (struct irte_ga *)table->table;
3760 entry = &entry[index];
3761 entry->lo.fields_remap.valid = 0;
3762 entry->hi.val = irte->hi.val;
3763 entry->lo.val = irte->lo.val;
3764 entry->lo.fields_remap.valid = 1;
3768 spin_unlock_irqrestore(&table->lock, flags);
3770 iommu_flush_irt(iommu, devid);
3771 iommu_completion_wait(iommu);
3776 static int modify_irte(u16 devid, int index, union irte *irte)
3778 struct irq_remap_table *table;
3779 struct amd_iommu *iommu;
3780 unsigned long flags;
3782 iommu = amd_iommu_rlookup_table[devid];
3786 table = get_irq_table(devid, false);
3790 spin_lock_irqsave(&table->lock, flags);
3791 table->table[index] = irte->val;
3792 spin_unlock_irqrestore(&table->lock, flags);
3794 iommu_flush_irt(iommu, devid);
3795 iommu_completion_wait(iommu);
3800 static void free_irte(u16 devid, int index)
3802 struct irq_remap_table *table;
3803 struct amd_iommu *iommu;
3804 unsigned long flags;
3806 iommu = amd_iommu_rlookup_table[devid];
3810 table = get_irq_table(devid, false);
3814 spin_lock_irqsave(&table->lock, flags);
3815 iommu->irte_ops->clear_allocated(table, index);
3816 spin_unlock_irqrestore(&table->lock, flags);
3818 iommu_flush_irt(iommu, devid);
3819 iommu_completion_wait(iommu);
3822 static void irte_prepare(void *entry,
3823 u32 delivery_mode, u32 dest_mode,
3824 u8 vector, u32 dest_apicid, int devid)
3826 union irte *irte = (union irte *) entry;
3829 irte->fields.vector = vector;
3830 irte->fields.int_type = delivery_mode;
3831 irte->fields.destination = dest_apicid;
3832 irte->fields.dm = dest_mode;
3833 irte->fields.valid = 1;
3836 static void irte_ga_prepare(void *entry,
3837 u32 delivery_mode, u32 dest_mode,
3838 u8 vector, u32 dest_apicid, int devid)
3840 struct irte_ga *irte = (struct irte_ga *) entry;
3841 struct iommu_dev_data *dev_data = search_dev_data(devid);
3845 irte->lo.fields_remap.guest_mode = dev_data ? dev_data->use_vapic : 0;
3846 irte->lo.fields_remap.int_type = delivery_mode;
3847 irte->lo.fields_remap.dm = dest_mode;
3848 irte->hi.fields.vector = vector;
3849 irte->lo.fields_remap.destination = dest_apicid;
3850 irte->lo.fields_remap.valid = 1;
3853 static void irte_activate(void *entry, u16 devid, u16 index)
3855 union irte *irte = (union irte *) entry;
3857 irte->fields.valid = 1;
3858 modify_irte(devid, index, irte);
3861 static void irte_ga_activate(void *entry, u16 devid, u16 index)
3863 struct irte_ga *irte = (struct irte_ga *) entry;
3865 irte->lo.fields_remap.valid = 1;
3866 modify_irte_ga(devid, index, irte, NULL);
3869 static void irte_deactivate(void *entry, u16 devid, u16 index)
3871 union irte *irte = (union irte *) entry;
3873 irte->fields.valid = 0;
3874 modify_irte(devid, index, irte);
3877 static void irte_ga_deactivate(void *entry, u16 devid, u16 index)
3879 struct irte_ga *irte = (struct irte_ga *) entry;
3881 irte->lo.fields_remap.valid = 0;
3882 modify_irte_ga(devid, index, irte, NULL);
3885 static void irte_set_affinity(void *entry, u16 devid, u16 index,
3886 u8 vector, u32 dest_apicid)
3888 union irte *irte = (union irte *) entry;
3890 irte->fields.vector = vector;
3891 irte->fields.destination = dest_apicid;
3892 modify_irte(devid, index, irte);
3895 static void irte_ga_set_affinity(void *entry, u16 devid, u16 index,
3896 u8 vector, u32 dest_apicid)
3898 struct irte_ga *irte = (struct irte_ga *) entry;
3899 struct iommu_dev_data *dev_data = search_dev_data(devid);
3901 if (!dev_data || !dev_data->use_vapic) {
3902 irte->hi.fields.vector = vector;
3903 irte->lo.fields_remap.destination = dest_apicid;
3904 irte->lo.fields_remap.guest_mode = 0;
3905 modify_irte_ga(devid, index, irte, NULL);
3909 #define IRTE_ALLOCATED (~1U)
3910 static void irte_set_allocated(struct irq_remap_table *table, int index)
3912 table->table[index] = IRTE_ALLOCATED;
3915 static void irte_ga_set_allocated(struct irq_remap_table *table, int index)
3917 struct irte_ga *ptr = (struct irte_ga *)table->table;
3918 struct irte_ga *irte = &ptr[index];
3920 memset(&irte->lo.val, 0, sizeof(u64));
3921 memset(&irte->hi.val, 0, sizeof(u64));
3922 irte->hi.fields.vector = 0xff;
3925 static bool irte_is_allocated(struct irq_remap_table *table, int index)
3927 union irte *ptr = (union irte *)table->table;
3928 union irte *irte = &ptr[index];
3930 return irte->val != 0;
3933 static bool irte_ga_is_allocated(struct irq_remap_table *table, int index)
3935 struct irte_ga *ptr = (struct irte_ga *)table->table;
3936 struct irte_ga *irte = &ptr[index];
3938 return irte->hi.fields.vector != 0;
3941 static void irte_clear_allocated(struct irq_remap_table *table, int index)
3943 table->table[index] = 0;
3946 static void irte_ga_clear_allocated(struct irq_remap_table *table, int index)
3948 struct irte_ga *ptr = (struct irte_ga *)table->table;
3949 struct irte_ga *irte = &ptr[index];
3951 memset(&irte->lo.val, 0, sizeof(u64));
3952 memset(&irte->hi.val, 0, sizeof(u64));
3955 static int get_devid(struct irq_alloc_info *info)
3959 switch (info->type) {
3960 case X86_IRQ_ALLOC_TYPE_IOAPIC:
3961 devid = get_ioapic_devid(info->ioapic_id);
3963 case X86_IRQ_ALLOC_TYPE_HPET:
3964 devid = get_hpet_devid(info->hpet_id);
3966 case X86_IRQ_ALLOC_TYPE_MSI:
3967 case X86_IRQ_ALLOC_TYPE_MSIX:
3968 devid = get_device_id(&info->msi_dev->dev);
3978 static struct irq_domain *get_ir_irq_domain(struct irq_alloc_info *info)
3980 struct amd_iommu *iommu;
3986 devid = get_devid(info);
3988 iommu = amd_iommu_rlookup_table[devid];
3990 return iommu->ir_domain;
3996 static struct irq_domain *get_irq_domain(struct irq_alloc_info *info)
3998 struct amd_iommu *iommu;
4004 switch (info->type) {
4005 case X86_IRQ_ALLOC_TYPE_MSI:
4006 case X86_IRQ_ALLOC_TYPE_MSIX:
4007 devid = get_device_id(&info->msi_dev->dev);
4011 iommu = amd_iommu_rlookup_table[devid];
4013 return iommu->msi_domain;
4022 struct irq_remap_ops amd_iommu_irq_ops = {
4023 .prepare = amd_iommu_prepare,
4024 .enable = amd_iommu_enable,
4025 .disable = amd_iommu_disable,
4026 .reenable = amd_iommu_reenable,
4027 .enable_faulting = amd_iommu_enable_faulting,
4028 .get_ir_irq_domain = get_ir_irq_domain,
4029 .get_irq_domain = get_irq_domain,
4032 static void irq_remapping_prepare_irte(struct amd_ir_data *data,
4033 struct irq_cfg *irq_cfg,
4034 struct irq_alloc_info *info,
4035 int devid, int index, int sub_handle)
4037 struct irq_2_irte *irte_info = &data->irq_2_irte;
4038 struct msi_msg *msg = &data->msi_entry;
4039 struct IO_APIC_route_entry *entry;
4040 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
4045 data->irq_2_irte.devid = devid;
4046 data->irq_2_irte.index = index + sub_handle;
4047 iommu->irte_ops->prepare(data->entry, apic->irq_delivery_mode,
4048 apic->irq_dest_mode, irq_cfg->vector,
4049 irq_cfg->dest_apicid, devid);
4051 switch (info->type) {
4052 case X86_IRQ_ALLOC_TYPE_IOAPIC:
4053 /* Setup IOAPIC entry */
4054 entry = info->ioapic_entry;
4055 info->ioapic_entry = NULL;
4056 memset(entry, 0, sizeof(*entry));
4057 entry->vector = index;
4059 entry->trigger = info->ioapic_trigger;
4060 entry->polarity = info->ioapic_polarity;
4061 /* Mask level triggered irqs. */
4062 if (info->ioapic_trigger)
4066 case X86_IRQ_ALLOC_TYPE_HPET:
4067 case X86_IRQ_ALLOC_TYPE_MSI:
4068 case X86_IRQ_ALLOC_TYPE_MSIX:
4069 msg->address_hi = MSI_ADDR_BASE_HI;
4070 msg->address_lo = MSI_ADDR_BASE_LO;
4071 msg->data = irte_info->index;
4080 struct amd_irte_ops irte_32_ops = {
4081 .prepare = irte_prepare,
4082 .activate = irte_activate,
4083 .deactivate = irte_deactivate,
4084 .set_affinity = irte_set_affinity,
4085 .set_allocated = irte_set_allocated,
4086 .is_allocated = irte_is_allocated,
4087 .clear_allocated = irte_clear_allocated,
4090 struct amd_irte_ops irte_128_ops = {
4091 .prepare = irte_ga_prepare,
4092 .activate = irte_ga_activate,
4093 .deactivate = irte_ga_deactivate,
4094 .set_affinity = irte_ga_set_affinity,
4095 .set_allocated = irte_ga_set_allocated,
4096 .is_allocated = irte_ga_is_allocated,
4097 .clear_allocated = irte_ga_clear_allocated,
4100 static int irq_remapping_alloc(struct irq_domain *domain, unsigned int virq,
4101 unsigned int nr_irqs, void *arg)
4103 struct irq_alloc_info *info = arg;
4104 struct irq_data *irq_data;
4105 struct amd_ir_data *data = NULL;
4106 struct irq_cfg *cfg;
4112 if (nr_irqs > 1 && info->type != X86_IRQ_ALLOC_TYPE_MSI &&
4113 info->type != X86_IRQ_ALLOC_TYPE_MSIX)
4117 * With IRQ remapping enabled, don't need contiguous CPU vectors
4118 * to support multiple MSI interrupts.
4120 if (info->type == X86_IRQ_ALLOC_TYPE_MSI)
4121 info->flags &= ~X86_IRQ_ALLOC_CONTIGUOUS_VECTORS;
4123 devid = get_devid(info);
4127 ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
4131 if (info->type == X86_IRQ_ALLOC_TYPE_IOAPIC) {
4132 if (get_irq_table(devid, true))
4133 index = info->ioapic_pin;
4137 index = alloc_irq_index(devid, nr_irqs);
4140 pr_warn("Failed to allocate IRTE\n");
4142 goto out_free_parent;
4145 for (i = 0; i < nr_irqs; i++) {
4146 irq_data = irq_domain_get_irq_data(domain, virq + i);
4147 cfg = irqd_cfg(irq_data);
4148 if (!irq_data || !cfg) {
4154 data = kzalloc(sizeof(*data), GFP_KERNEL);
4158 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
4159 data->entry = kzalloc(sizeof(union irte), GFP_KERNEL);
4161 data->entry = kzalloc(sizeof(struct irte_ga),
4168 irq_data->hwirq = (devid << 16) + i;
4169 irq_data->chip_data = data;
4170 irq_data->chip = &amd_ir_chip;
4171 irq_remapping_prepare_irte(data, cfg, info, devid, index, i);
4172 irq_set_status_flags(virq + i, IRQ_MOVE_PCNTXT);
4178 for (i--; i >= 0; i--) {
4179 irq_data = irq_domain_get_irq_data(domain, virq + i);
4181 kfree(irq_data->chip_data);
4183 for (i = 0; i < nr_irqs; i++)
4184 free_irte(devid, index + i);
4186 irq_domain_free_irqs_common(domain, virq, nr_irqs);
4190 static void irq_remapping_free(struct irq_domain *domain, unsigned int virq,
4191 unsigned int nr_irqs)
4193 struct irq_2_irte *irte_info;
4194 struct irq_data *irq_data;
4195 struct amd_ir_data *data;
4198 for (i = 0; i < nr_irqs; i++) {
4199 irq_data = irq_domain_get_irq_data(domain, virq + i);
4200 if (irq_data && irq_data->chip_data) {
4201 data = irq_data->chip_data;
4202 irte_info = &data->irq_2_irte;
4203 free_irte(irte_info->devid, irte_info->index);
4208 irq_domain_free_irqs_common(domain, virq, nr_irqs);
4211 static void irq_remapping_activate(struct irq_domain *domain,
4212 struct irq_data *irq_data)
4214 struct amd_ir_data *data = irq_data->chip_data;
4215 struct irq_2_irte *irte_info = &data->irq_2_irte;
4216 struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4219 iommu->irte_ops->activate(data->entry, irte_info->devid,
4223 static void irq_remapping_deactivate(struct irq_domain *domain,
4224 struct irq_data *irq_data)
4226 struct amd_ir_data *data = irq_data->chip_data;
4227 struct irq_2_irte *irte_info = &data->irq_2_irte;
4228 struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4231 iommu->irte_ops->deactivate(data->entry, irte_info->devid,
4235 static struct irq_domain_ops amd_ir_domain_ops = {
4236 .alloc = irq_remapping_alloc,
4237 .free = irq_remapping_free,
4238 .activate = irq_remapping_activate,
4239 .deactivate = irq_remapping_deactivate,
4242 static int amd_ir_set_vcpu_affinity(struct irq_data *data, void *vcpu_info)
4244 struct amd_iommu *iommu;
4245 struct amd_iommu_pi_data *pi_data = vcpu_info;
4246 struct vcpu_data *vcpu_pi_info = pi_data->vcpu_data;
4247 struct amd_ir_data *ir_data = data->chip_data;
4248 struct irte_ga *irte = (struct irte_ga *) ir_data->entry;
4249 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
4250 struct iommu_dev_data *dev_data = search_dev_data(irte_info->devid);
4253 * This device has never been set up for guest mode.
4254 * we should not modify the IRTE
4256 if (!dev_data || !dev_data->use_vapic)
4259 pi_data->ir_data = ir_data;
4262 * SVM tries to set up for VAPIC mode, but we are in
4263 * legacy mode. So, we force legacy mode instead.
4265 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
4266 pr_debug("AMD-Vi: %s: Fall back to using intr legacy remap\n",
4268 pi_data->is_guest_mode = false;
4271 iommu = amd_iommu_rlookup_table[irte_info->devid];
4275 pi_data->prev_ga_tag = ir_data->cached_ga_tag;
4276 if (pi_data->is_guest_mode) {
4278 irte->hi.fields.ga_root_ptr = (pi_data->base >> 12);
4279 irte->hi.fields.vector = vcpu_pi_info->vector;
4280 irte->lo.fields_vapic.guest_mode = 1;
4281 irte->lo.fields_vapic.ga_tag = pi_data->ga_tag;
4283 ir_data->cached_ga_tag = pi_data->ga_tag;
4286 struct irq_cfg *cfg = irqd_cfg(data);
4290 irte->hi.fields.vector = cfg->vector;
4291 irte->lo.fields_remap.guest_mode = 0;
4292 irte->lo.fields_remap.destination = cfg->dest_apicid;
4293 irte->lo.fields_remap.int_type = apic->irq_delivery_mode;
4294 irte->lo.fields_remap.dm = apic->irq_dest_mode;
4297 * This communicates the ga_tag back to the caller
4298 * so that it can do all the necessary clean up.
4300 ir_data->cached_ga_tag = 0;
4303 return modify_irte_ga(irte_info->devid, irte_info->index, irte, ir_data);
4306 static int amd_ir_set_affinity(struct irq_data *data,
4307 const struct cpumask *mask, bool force)
4309 struct amd_ir_data *ir_data = data->chip_data;
4310 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
4311 struct irq_cfg *cfg = irqd_cfg(data);
4312 struct irq_data *parent = data->parent_data;
4313 struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4319 ret = parent->chip->irq_set_affinity(parent, mask, force);
4320 if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
4324 * Atomically updates the IRTE with the new destination, vector
4325 * and flushes the interrupt entry cache.
4327 iommu->irte_ops->set_affinity(ir_data->entry, irte_info->devid,
4328 irte_info->index, cfg->vector, cfg->dest_apicid);
4331 * After this point, all the interrupts will start arriving
4332 * at the new destination. So, time to cleanup the previous
4333 * vector allocation.
4335 send_cleanup_vector(cfg);
4337 return IRQ_SET_MASK_OK_DONE;
4340 static void ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg)
4342 struct amd_ir_data *ir_data = irq_data->chip_data;
4344 *msg = ir_data->msi_entry;
4347 static struct irq_chip amd_ir_chip = {
4348 .irq_ack = ir_ack_apic_edge,
4349 .irq_set_affinity = amd_ir_set_affinity,
4350 .irq_set_vcpu_affinity = amd_ir_set_vcpu_affinity,
4351 .irq_compose_msi_msg = ir_compose_msi_msg,
4354 int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
4356 iommu->ir_domain = irq_domain_add_tree(NULL, &amd_ir_domain_ops, iommu);
4357 if (!iommu->ir_domain)
4360 iommu->ir_domain->parent = arch_get_ir_parent_domain();
4361 iommu->msi_domain = arch_create_msi_irq_domain(iommu->ir_domain);
4366 int amd_iommu_update_ga(int cpu, bool is_run, void *data)
4368 unsigned long flags;
4369 struct amd_iommu *iommu;
4370 struct irq_remap_table *irt;
4371 struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
4372 int devid = ir_data->irq_2_irte.devid;
4373 struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
4374 struct irte_ga *ref = (struct irte_ga *) ir_data->ref;
4376 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
4377 !ref || !entry || !entry->lo.fields_vapic.guest_mode)
4380 iommu = amd_iommu_rlookup_table[devid];
4384 irt = get_irq_table(devid, false);
4388 spin_lock_irqsave(&irt->lock, flags);
4390 if (ref->lo.fields_vapic.guest_mode) {
4392 ref->lo.fields_vapic.destination = cpu;
4393 ref->lo.fields_vapic.is_run = is_run;
4397 spin_unlock_irqrestore(&irt->lock, flags);
4399 iommu_flush_irt(iommu, devid);
4400 iommu_completion_wait(iommu);
4403 EXPORT_SYMBOL(amd_iommu_update_ga);